diff --git a/clients/gtest/known_bugs.yaml b/clients/gtest/known_bugs.yaml
index fa7592d3f..5ce181dad 100644
--- a/clients/gtest/known_bugs.yaml
+++ b/clients/gtest/known_bugs.yaml
@@ -3,7 +3,13 @@
 # Wildcards can be used for the function
 
 Known bugs:
+- { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 512, N: 512, K: 512, lda: 512, ldb: 512, ldc: 512, ldd: 512, alpha: 5.0, alphai: 0.0, beta: 0.0, betai: 0.0 }
+- { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 512, N: 512, K: 512, lda: 512, ldb: 512, ldc: 512, ldd: 512, alpha: 0.0, alphai: 0.0, beta: 3.0, betai: 0.0 }
+- { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 512, N: 512, K: 512, lda: 512, ldb: 512, ldc: 512, ldd: 512, alpha: 1.0, alphai: 0.0, beta: 3.0, betai: 0.0 }
+- { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 512, N: 512, K: 512, lda: 512, ldb: 512, ldc: 512, ldd: 512, alpha: 1.0, alphai: 0.0, beta: 1.0, betai: 0.0 }
 - { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 1024, N: 1024, K: 1024, lda: 1024, ldb: 1024, ldc: 1024, ldd: 1024, alpha: 5.0, alphai: 0.0, beta: 0.0, betai: 0.0 }
 - { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 1024, N: 1024, K: 1024, lda: 1024, ldb: 1024, ldc: 1024, ldd: 1024, alpha: 0.0, alphai: 0.0, beta: 3.0, betai: 0.0 }
 - { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 1024, N: 1024, K: 1024, lda: 1024, ldb: 1024, ldc: 1024, ldd: 1024, alpha: 1.0, alphai: 0.0, beta: 3.0, betai: 0.0 }
 - { function: "gemm_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 1024, N: 1024, K: 1024, lda: 1024, ldb: 1024, ldc: 1024, ldd: 1024, alpha: 1.0, alphai: 0.0, beta: 1.0, betai: 0.0 }
+- { function: "gemm_strided_batched_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 512, N: 512, K: 512, lda: 512, ldb: 512, ldc: 512, ldd: 512, alpha: -2.0, alphai: 0.0, beta: -3.0, betai: 0.0, batch_count: 3, stride_a: 262144, stride_b: 262144, stride_c: 262144, stride_d: 262144 }
+- { function: "gemm_strided_batched_ex", a_type: "bf16_r", b_type: "bf16_r", c_type: "bf16_r", d_type: "bf16_r", compute_type: "f32_r", transA: 'C', transB: 'N', M: 512, N: 512, K: 512, lda: 512, ldb: 512, ldc: 512, ldd: 512, alpha: 0.0, alphai: 0.0, beta: 1.0, betai: 0.0, batch_count: 3, stride_a: 262144, stride_b: 262144, stride_c: 262144, stride_d: 262144 }
diff --git a/library/src/blas3/Tensile/Logic/asm_full/arcturus_Cijk_Alik_Bljk_BBH.yaml b/library/src/blas3/Tensile/Logic/asm_full/arcturus_Cijk_Alik_Bljk_BBH.yaml
index 79e819a61..44b823d28 100644
--- a/library/src/blas3/Tensile/Logic/asm_full/arcturus_Cijk_Alik_Bljk_BBH.yaml
+++ b/library/src/blas3/Tensile/Logic/asm_full/arcturus_Cijk_Alik_Bljk_BBH.yaml
@@ -7,6 +7,7 @@
   ComplexConjugateA: false
   ComplexConjugateB: false
   ComputeDataType: 0
+  ConvolutionConfig: []
   DataType: 7
   DestDataType: 7
   HighPrecisionAccumulate: true
@@ -30,6 +31,7 @@
   NumIndicesSummation: 1
   OperationType: GEMM
   SetConstStrideA: []
+  SetConstStrideB: []
   SilentHighPrecisionAccumulate: false
   TLUA: false
   TLUB: false
@@ -60,6 +62,7 @@
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
     ExpandPointerSwap: false
@@ -150,6 +153,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -173,6 +177,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -232,6 +237,7 @@
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
     ExpandPointerSwap: false
@@ -322,6 +328,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -345,6 +352,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -404,6 +412,7 @@
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
     ExpandPointerSwap: false
@@ -494,6 +503,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -517,6 +527,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -576,6 +587,7 @@
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
     ExpandPointerSwap: false
@@ -666,6 +678,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -689,6 +702,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -748,6 +762,7 @@
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
     ExpandPointerSwap: false
@@ -838,6 +853,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -861,6 +877,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -906,59 +923,56 @@
     WorkGroupMappingType: B
     _staggerStrideShift: 3
   - AggressivePerfMode: 1
-    AssertFree0ElementMultiple: 8
+    AssertFree0ElementMultiple: 1
     AssertFree1ElementMultiple: 1
-    AssertMinApproxSize: 3
-    AssertSummationElementMultiple: 8
+    AssertMinApproxSize: 0
+    AssertSummationElementMultiple: 1
     AssignedDerivedParameters: false
     AssignedProblemIndependentDerivedParameters: true
-    BufferLoad: true
+    BufferLoad: false
     BufferStore: true
     CheckDimOverflow: 0
     CheckTensorDimAsserts: false
-    DepthU: 32
+    DepthU: 16
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
-    ExpandPointerSwap: true
-    FractionalLoad: 0
-    GlobalLoadVectorWidthA: 4
-    GlobalLoadVectorWidthB: 4
+    ExpandPointerSwap: false
+    FractionalLoad: false
+    GlobalLoadVectorWidthA: 1
+    GlobalLoadVectorWidthB: 1
     GlobalRead2A: true
     GlobalRead2B: true
     GlobalReadCoalesceGroupA: true
     GlobalReadCoalesceGroupB: true
     GlobalReadCoalesceVectorA: true
     GlobalReadCoalesceVectorB: true
-    GlobalReadVectorWidth: 4
+    GlobalReadVectorWidth: 1
     GlobalSplitU: 1
     GlobalSplitUSummationAssignmentRoundRobin: true
     GlobalSplitUWorkGroupMappingRoundRobin: false
-    GlobalWriteVectorWidth: 4
+    GlobalWriteVectorWidth: 1
     GuaranteeNoPartialA: true
     GuaranteeNoPartialB: true
-    ISA: [9, 0, 8]
+    ISA: [0, 0, 0]
     InnerUnroll: 1
     InterleaveAlpha: 0
-    KernelLanguage: Assembly
-    LSCA: 32
-    LSCB: 32
-    LSPA: 32
-    LSPB: 32
-    LVCA: 8
-    LVCB: 8
-    LVPA: 8
-    LVPB: 8
+    KernelLanguage: Source
+    LSCA: 16
+    LSCB: 16
+    LSPA: 16
+    LSPB: 16
+    LVCA: 16
+    LVCB: 16
+    LVPA: 16
+    LVPB: 16
     LdcEqualsLdd: true
-    LdsNumElements: 14336
-    LdsNumElementsAlignedA: 2048
-    LdsNumElementsAlignedB: 4096
+    LdsNumElements: 6144
     LdsOffsetA: 0
-    LdsOffsetA_Blk: 8192
     LdsOffsetB: 2048
-    LdsOffsetB_Blk: 10240
     LdsPadA: 0
     LdsPadB: 0
     LocalDotLayout: 1
@@ -971,11 +985,11 @@
     LocalWriteUseSgprB: false
     LoopDoWhile: false
     LoopTail: true
-    LoopUnroll: 32
-    MacroTile0: 64
-    MacroTile1: 128
-    MacroTileA: 64
-    MacroTileB: 128
+    LoopUnroll: 16
+    MacroTile0: 128
+    MacroTile1: 256
+    MacroTileA: 128
+    MacroTileB: 256
     MacroTileShapeMax: 64
     MacroTileShapeMin: 1
     MaxOccupancy: 40
@@ -984,14 +998,14 @@
     NonTemporalA: 0
     NonTemporalB: 0
     NonTemporalC: 0
-    NumElementsPerThread: 32
-    NumGlobalWriteVectorsPerThread: 8
-    NumLoadsA: 2
-    NumLoadsB: 4
+    NumElementsPerThread: 128
+    NumGlobalWriteVectorsPerThread: 128
+    NumLoadsA: 8
+    NumLoadsB: 16
     NumLoadsCoalescedA: 1
     NumLoadsCoalescedB: 1
-    NumLoadsPerpendicularA: 2
-    NumLoadsPerpendicularB: 4
+    NumLoadsPerpendicularA: 8
+    NumLoadsPerpendicularB: 16
     NumThreads: 256
     OptNoLoadLoop: 1
     PackBatchDims: 0
@@ -1006,7 +1020,7 @@
     PerformanceWaitLocation: -1
     PersistentKernel: 0
     PrefetchAcrossPersistent: 0
-    PrefetchGlobalRead: true
+    PrefetchGlobalRead: false
     PrefetchLocalRead: true
     ProblemType:
       AssignedDerivedParameters: true
@@ -1014,6 +1028,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -1037,6 +1052,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -1052,12 +1068,12 @@
       UseInitialStrides: false
       ZeroPadA: []
       ZeroPadB: []
-    ReplacementKernel: true
+    ReplacementKernel: false
     ScheduleGlobalRead: 1
     ScheduleIterAlg: 1
     ScheduleLocalWrite: 1
     SolutionIndex: 5
-    SolutionNameMin: Cijk_Alik_Bljk_BBH_MT64x128x32_SE_
+    SolutionNameMin: Cijk_Alik_Bljk_BBH_MT128x256x16_SE_
     StaggerU: 32
     StaggerUMapping: 0
     StaggerUStride: 256
@@ -1066,21 +1082,21 @@
     SubGroupA: 16
     SubGroupB: 16
     SuppressNoLoadLoop: false
-    ThreadTile: [4, 8]
-    ThreadTile0: 4
-    ThreadTile1: 8
-    ThreadTileA: 4
-    ThreadTileB: 8
+    ThreadTile: [8, 16]
+    ThreadTile0: 8
+    ThreadTile1: 16
+    ThreadTileA: 8
+    ThreadTileB: 16
     UnrollMemFence: false
-    UseSgprForGRO: 1
+    UseSgprForGRO: 0
     Valid: true
     VectorAtomicWidth: 1
     VectorStore: true
-    VectorWidth: 4
+    VectorWidth: 1
     WorkGroup: [16, 16, 1]
     WorkGroupMapping: 8
     WorkGroupMappingType: B
-    _staggerStrideShift: 2
+    _staggerStrideShift: 3
   - AggressivePerfMode: 1
     AssertFree0ElementMultiple: 1
     AssertFree1ElementMultiple: 1
@@ -1096,6 +1112,7 @@
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
     ExpandPointerSwap: false
@@ -1186,6 +1203,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -1209,6 +1227,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -1253,6 +1272,181 @@
     WorkGroupMapping: 8
     WorkGroupMappingType: B
     _staggerStrideShift: 3
+  - AggressivePerfMode: 1
+    AssertFree0ElementMultiple: 1
+    AssertFree1ElementMultiple: 1
+    AssertMinApproxSize: 0
+    AssertSummationElementMultiple: 1
+    AssignedDerivedParameters: false
+    AssignedProblemIndependentDerivedParameters: true
+    BufferLoad: false
+    BufferStore: true
+    CheckDimOverflow: 0
+    CheckTensorDimAsserts: false
+    DepthU: 16
+    DirectToLds: false
+    DirectToLdsA: false
+    DirectToLdsB: false
+    DisableAtomicFail: 0
+    DisableKernelPieces: 0
+    EdgeType: ShiftPtr
+    ExpandPointerSwap: false
+    FractionalLoad: false
+    GlobalLoadVectorWidthA: 1
+    GlobalLoadVectorWidthB: 1
+    GlobalRead2A: true
+    GlobalRead2B: true
+    GlobalReadCoalesceGroupA: true
+    GlobalReadCoalesceGroupB: true
+    GlobalReadCoalesceVectorA: true
+    GlobalReadCoalesceVectorB: true
+    GlobalReadVectorWidth: 1
+    GlobalSplitU: 1
+    GlobalSplitUSummationAssignmentRoundRobin: true
+    GlobalSplitUWorkGroupMappingRoundRobin: false
+    GlobalWriteVectorWidth: 1
+    GuaranteeNoPartialA: true
+    GuaranteeNoPartialB: true
+    ISA: [0, 0, 0]
+    InnerUnroll: 1
+    InterleaveAlpha: 0
+    KernelLanguage: Source
+    LSCA: 16
+    LSCB: 16
+    LSPA: 4
+    LSPB: 4
+    LVCA: 16
+    LVCB: 16
+    LVPA: 4
+    LVPB: 4
+    LdcEqualsLdd: true
+    LdsNumElements: 1280
+    LdsOffsetA: 0
+    LdsOffsetB: 256
+    LdsPadA: 0
+    LdsPadB: 0
+    LocalDotLayout: 1
+    LocalRead2A: true
+    LocalRead2B: true
+    LocalSplitU: 1
+    LocalWrite2A: true
+    LocalWrite2B: true
+    LocalWriteUseSgprA: false
+    LocalWriteUseSgprB: false
+    LoopDoWhile: false
+    LoopTail: true
+    LoopUnroll: 16
+    MacroTile0: 16
+    MacroTile1: 64
+    MacroTileA: 16
+    MacroTileB: 64
+    MacroTileShapeMax: 64
+    MacroTileShapeMin: 1
+    MaxOccupancy: 40
+    MaxVgprNumber: 256
+    MinVgprNumber: 0
+    NonTemporalA: 0
+    NonTemporalB: 0
+    NonTemporalC: 0
+    NumElementsPerThread: 16
+    NumGlobalWriteVectorsPerThread: 16
+    NumLoadsA: 4
+    NumLoadsB: 16
+    NumLoadsCoalescedA: 1
+    NumLoadsCoalescedB: 1
+    NumLoadsPerpendicularA: 4
+    NumLoadsPerpendicularB: 16
+    NumThreads: 64
+    OptNoLoadLoop: 1
+    PackBatchDims: 0
+    PackFreeDims: 1
+    PackGranularity: 2
+    PackedC0IdxChars: [I]
+    PackedC0IndicesX: [0]
+    PackedC1IdxChars: [J]
+    PackedC1IndicesX: [1]
+    PerformanceSyncLocation: -1
+    PerformanceWaitCount: -1
+    PerformanceWaitLocation: -1
+    PersistentKernel: 0
+    PrefetchAcrossPersistent: 0
+    PrefetchGlobalRead: false
+    PrefetchLocalRead: true
+    ProblemType:
+      AssignedDerivedParameters: true
+      Batched: true
+      ComplexConjugateA: false
+      ComplexConjugateB: false
+      ComputeDataType: 0
+      ConvolutionConfig: []
+      DataType: 7
+      DestDataType: 7
+      HighPrecisionAccumulate: true
+      Index0: 0
+      Index01A: 0
+      Index01B: 1
+      Index1: 1
+      IndexAssignmentsA: [3, 0, 2]
+      IndexAssignmentsB: [3, 1, 2]
+      IndexAssignmentsLD: [4, 5, 6, 7]
+      IndexUnroll: 3
+      IndexUnrollA: 0
+      IndexUnrollB: 0
+      IndicesBatch: [2]
+      IndicesFree: [0, 1]
+      IndicesSummation: [3]
+      NumIndicesBatch: 1
+      NumIndicesC: 3
+      NumIndicesFree: 2
+      NumIndicesLD: 4
+      NumIndicesSummation: 1
+      OperationType: GEMM
+      SetConstStrideA: []
+      SetConstStrideB: []
+      SilentHighPrecisionAccumulate: false
+      TLUA: false
+      TLUB: false
+      Tensor0: 0
+      Tensor1: 1
+      TileA: 0
+      TileAwareSelection: false
+      TileB: 1
+      TotalIndices: 4
+      TransposeA: true
+      TransposeB: false
+      UseBeta: true
+      UseInitialStrides: false
+      ZeroPadA: []
+      ZeroPadB: []
+    ReplacementKernel: false
+    ScheduleGlobalRead: 1
+    ScheduleIterAlg: 1
+    ScheduleLocalWrite: 1
+    SolutionIndex: 7
+    SolutionNameMin: Cijk_Alik_Bljk_BBH_MT16x64x16_SE_
+    StaggerU: 32
+    StaggerUMapping: 0
+    StaggerUStride: 256
+    SubGroup0: 8
+    SubGroup1: 8
+    SubGroupA: 8
+    SubGroupB: 8
+    SuppressNoLoadLoop: false
+    ThreadTile: [2, 8]
+    ThreadTile0: 2
+    ThreadTile1: 8
+    ThreadTileA: 2
+    ThreadTileB: 8
+    UnrollMemFence: false
+    UseSgprForGRO: 0
+    Valid: true
+    VectorAtomicWidth: 1
+    VectorStore: true
+    VectorWidth: 1
+    WorkGroup: [8, 8, 1]
+    WorkGroupMapping: 8
+    WorkGroupMappingType: B
+    _staggerStrideShift: 3
   - AggressivePerfMode: 1
     AssertFree0ElementMultiple: 8
     AssertFree1ElementMultiple: 1
@@ -1268,6 +1462,186 @@
     DirectToLds: false
     DirectToLdsA: false
     DirectToLdsB: false
+    DisableAtomicFail: 0
+    DisableKernelPieces: 0
+    EdgeType: ShiftPtr
+    ExpandPointerSwap: true
+    FractionalLoad: 0
+    GlobalLoadVectorWidthA: 4
+    GlobalLoadVectorWidthB: 4
+    GlobalRead2A: true
+    GlobalRead2B: true
+    GlobalReadCoalesceGroupA: true
+    GlobalReadCoalesceGroupB: true
+    GlobalReadCoalesceVectorA: true
+    GlobalReadCoalesceVectorB: true
+    GlobalReadVectorWidth: 4
+    GlobalSplitU: 1
+    GlobalSplitUSummationAssignmentRoundRobin: true
+    GlobalSplitUWorkGroupMappingRoundRobin: false
+    GlobalWriteVectorWidth: 4
+    GuaranteeNoPartialA: true
+    GuaranteeNoPartialB: true
+    ISA: [9, 0, 8]
+    InnerUnroll: 1
+    InterleaveAlpha: 0
+    KernelLanguage: Assembly
+    LSCA: 32
+    LSCB: 32
+    LSPA: 32
+    LSPB: 32
+    LVCA: 8
+    LVCB: 8
+    LVPA: 8
+    LVPB: 8
+    LdcEqualsLdd: true
+    LdsNumElements: 14336
+    LdsNumElementsAlignedA: 2048
+    LdsNumElementsAlignedB: 4096
+    LdsOffsetA: 0
+    LdsOffsetA_Blk: 8192
+    LdsOffsetB: 2048
+    LdsOffsetB_Blk: 10240
+    LdsPadA: 0
+    LdsPadB: 0
+    LocalDotLayout: 1
+    LocalRead2A: true
+    LocalRead2B: true
+    LocalSplitU: 1
+    LocalWrite2A: true
+    LocalWrite2B: true
+    LocalWriteUseSgprA: false
+    LocalWriteUseSgprB: false
+    LoopDoWhile: false
+    LoopTail: true
+    LoopUnroll: 32
+    MacroTile0: 64
+    MacroTile1: 128
+    MacroTileA: 64
+    MacroTileB: 128
+    MacroTileShapeMax: 64
+    MacroTileShapeMin: 1
+    MaxOccupancy: 40
+    MaxVgprNumber: 256
+    MinVgprNumber: 0
+    NonTemporalA: 0
+    NonTemporalB: 0
+    NonTemporalC: 0
+    NumElementsPerThread: 32
+    NumGlobalWriteVectorsPerThread: 8
+    NumLoadsA: 2
+    NumLoadsB: 4
+    NumLoadsCoalescedA: 1
+    NumLoadsCoalescedB: 1
+    NumLoadsPerpendicularA: 2
+    NumLoadsPerpendicularB: 4
+    NumThreads: 256
+    OptNoLoadLoop: 1
+    PackBatchDims: 0
+    PackFreeDims: 1
+    PackGranularity: 2
+    PackedC0IdxChars: [I]
+    PackedC0IndicesX: [0]
+    PackedC1IdxChars: [J]
+    PackedC1IndicesX: [1]
+    PerformanceSyncLocation: -1
+    PerformanceWaitCount: -1
+    PerformanceWaitLocation: -1
+    PersistentKernel: 0
+    PrefetchAcrossPersistent: 0
+    PrefetchGlobalRead: true
+    PrefetchLocalRead: true
+    ProblemType:
+      AssignedDerivedParameters: true
+      Batched: true
+      ComplexConjugateA: false
+      ComplexConjugateB: false
+      ComputeDataType: 0
+      ConvolutionConfig: []
+      DataType: 7
+      DestDataType: 7
+      HighPrecisionAccumulate: true
+      Index0: 0
+      Index01A: 0
+      Index01B: 1
+      Index1: 1
+      IndexAssignmentsA: [3, 0, 2]
+      IndexAssignmentsB: [3, 1, 2]
+      IndexAssignmentsLD: [4, 5, 6, 7]
+      IndexUnroll: 3
+      IndexUnrollA: 0
+      IndexUnrollB: 0
+      IndicesBatch: [2]
+      IndicesFree: [0, 1]
+      IndicesSummation: [3]
+      NumIndicesBatch: 1
+      NumIndicesC: 3
+      NumIndicesFree: 2
+      NumIndicesLD: 4
+      NumIndicesSummation: 1
+      OperationType: GEMM
+      SetConstStrideA: []
+      SetConstStrideB: []
+      SilentHighPrecisionAccumulate: false
+      TLUA: false
+      TLUB: false
+      Tensor0: 0
+      Tensor1: 1
+      TileA: 0
+      TileAwareSelection: false
+      TileB: 1
+      TotalIndices: 4
+      TransposeA: true
+      TransposeB: false
+      UseBeta: true
+      UseInitialStrides: false
+      ZeroPadA: []
+      ZeroPadB: []
+    ReplacementKernel: true
+    ScheduleGlobalRead: 1
+    ScheduleIterAlg: 1
+    ScheduleLocalWrite: 1
+    SolutionIndex: 8
+    SolutionNameMin: Cijk_Alik_Bljk_BBH_MT64x128x32_SE_
+    StaggerU: 32
+    StaggerUMapping: 0
+    StaggerUStride: 256
+    SubGroup0: 16
+    SubGroup1: 16
+    SubGroupA: 16
+    SubGroupB: 16
+    SuppressNoLoadLoop: false
+    ThreadTile: [4, 8]
+    ThreadTile0: 4
+    ThreadTile1: 8
+    ThreadTileA: 4
+    ThreadTileB: 8
+    UnrollMemFence: false
+    UseSgprForGRO: 1
+    Valid: true
+    VectorAtomicWidth: 1
+    VectorStore: true
+    VectorWidth: 4
+    WorkGroup: [16, 16, 1]
+    WorkGroupMapping: 8
+    WorkGroupMappingType: B
+    _staggerStrideShift: 2
+  - AggressivePerfMode: 1
+    AssertFree0ElementMultiple: 8
+    AssertFree1ElementMultiple: 1
+    AssertMinApproxSize: 3
+    AssertSummationElementMultiple: 8
+    AssignedDerivedParameters: false
+    AssignedProblemIndependentDerivedParameters: true
+    BufferLoad: true
+    BufferStore: true
+    CheckDimOverflow: 0
+    CheckTensorDimAsserts: false
+    DepthU: 32
+    DirectToLds: false
+    DirectToLdsA: false
+    DirectToLdsB: false
+    DisableAtomicFail: 0
     DisableKernelPieces: 0
     EdgeType: ShiftPtr
     ExpandPointerSwap: true
@@ -1362,6 +1736,7 @@
       ComplexConjugateA: false
       ComplexConjugateB: false
       ComputeDataType: 0
+      ConvolutionConfig: []
       DataType: 7
       DestDataType: 7
       HighPrecisionAccumulate: true
@@ -1385,6 +1760,7 @@
       NumIndicesSummation: 1
       OperationType: GEMM
       SetConstStrideA: []
+      SetConstStrideB: []
       SilentHighPrecisionAccumulate: false
       TLUA: false
       TLUB: false
@@ -1404,7 +1780,7 @@
     ScheduleGlobalRead: 1
     ScheduleIterAlg: 1
     ScheduleLocalWrite: 1
-    SolutionIndex: 7
+    SolutionIndex: 9
     SolutionNameMin: Cijk_Alik_Bljk_BBH_MT32x32x32_SE_
     StaggerU: 32
     StaggerUMapping: 0
@@ -1430,130 +1806,158 @@
     WorkGroupMappingType: B
     _staggerStrideShift: 2
 - [2, 3, 0, 1]
-- - - [959, 1024, 1, 1024]
-    - [3, 1055.49]
-  - - [960, 1023, 1, 1024]
-    - [3, 1071.67]
-  - - [960, 1024, 1, 1023]
-    - [3, 1069.84]
-  - - [960, 1024, 1, 1025]
-    - [3, 1077.07]
-  - - [960, 1025, 1, 1024]
-    - [2, 822.52]
-  - - [961, 1024, 1, 1024]
-    - [2, 813.489]
-  - - [1023, 1024, 1, 1024]
-    - [0, 865.973]
-  - - [1024, 1023, 1, 1024]
-    - [0, 865.805]
-  - - [1024, 1024, 1, 1023]
-    - [2, 866.42]
-  - - [1024, 1024, 1, 1025]
-    - [0, 865.43]
+- - - [512, 512, 1, 513]
+    - [0, 381.611]
+  - - [511, 512, 1, 512]
+    - [0, 379.091]
+  - - [2048, 2049, 1, 2048]
+    - [1, 1383.87]
+  - - [480, 511, 1, 512]
+    - [0, 357.991]
+  - - [64, 1023, 1, 1024]
+    - [6, 111.903]
+  - - [2879, 3072, 1, 3072]
+    - [3, 1661.36]
+  - - [480, 512, 1, 511]
+    - [0, 358.482]
+  - - [4096, 4096, 1, 4096]
+    - [8, 16868.2]
+  - - [480, 513, 1, 512]
+    - [0, 359.311]
+  - - [2880, 3072, 1, 3073]
+    - [3, 1658.3]
+  - - [2880, 3071, 1, 3072]
+    - [3, 1661.59]
+  - - [3840, 4097, 1, 4096]
+    - [1, 1597.42]
   - - [1024, 1025, 1, 1024]
-    - [0, 866.658]
+    - [2, 857.747]
+  - - [512, 512, 1, 511]
+    - [0, 380.902]
   - - [1025, 1024, 1, 1024]
-    - [0, 866.155]
-  - - [2039, 2048, 1, 2048]
-    - [1, 1376.9]
-  - - [2040, 2047, 1, 2048]
-    - [1, 1378.17]
-  - - [2040, 2048, 1, 2047]
-    - [1, 1372.13]
-  - - [2040, 2048, 1, 2049]
-    - [1, 1387.02]
-  - - [2040, 2049, 1, 2048]
-    - [1, 1380.06]
-  - - [2041, 2048, 1, 2048]
-    - [1, 1379.05]
+    - [2, 859.448]
+  - - [959, 1024, 1, 1024]
+    - [3, 1067.68]
+  - - [480, 512, 1, 513]
+    - [0, 358.656]
   - - [2047, 2048, 1, 2048]
-    - [1, 1384.61]
-  - - [2048, 2047, 1, 2048]
-    - [1, 1383.95]
-  - - [2048, 2048, 1, 2047]
-    - [1, 1374.29]
-  - - [2048, 2048, 1, 2049]
-    - [1, 1391.34]
-  - - [2048, 2049, 1, 2048]
-    - [1, 1384.69]
+    - [1, 1384.19]
+  - - [479, 512, 1, 512]
+    - [0, 357.293]
+  - - [3840, 4096, 1, 4096]
+    - [8, 17340.1]
   - - [2049, 2048, 1, 2048]
-    - [1, 1383.0]
-  - - [2999, 3072, 1, 3072]
-    - [1, 1614.3]
-  - - [3000, 3071, 1, 3072]
-    - [1, 1614.82]
-  - - [3000, 3072, 1, 3071]
-    - [1, 1600.23]
-  - - [3000, 3072, 1, 3073]
-    - [1, 1608.61]
-  - - [3000, 3073, 1, 3072]
-    - [1, 1614.31]
-  - - [3001, 3072, 1, 3072]
-    - [1, 1616.45]
-  - - [3071, 3072, 1, 3072]
-    - [1, 1651.37]
+    - [1, 1383.81]
+  - - [960, 1023, 1, 1024]
+    - [3, 1069.66]
+  - - [480, 512, 1, 512]
+    - [8, 2545.11]
+  - - [1024, 1024, 1, 1025]
+    - [2, 860.273]
+  - - [64, 1024, 1, 1023]
+    - [6, 113.741]
+  - - [3072, 3072, 1, 3072]
+    - [8, 15664.8]
+  - - [960, 1025, 1, 1024]
+    - [2, 815.013]
   - - [3072, 3071, 1, 3072]
-    - [1, 1652.19]
-  - - [3072, 3072, 1, 3071]
-    - [1, 1636.33]
+    - [1, 1650.83]
+  - - [1920, 2047, 1, 2048]
+    - [5, 1758.25]
+  - - [1921, 2048, 1, 2048]
+    - [1, 1295.9]
+  - - [4096, 4097, 1, 4096]
+    - [1, 1604.45]
+  - - [960, 1024, 1, 1025]
+    - [3, 1078.27]
+  - - [961, 1024, 1, 1024]
+    - [0, 805.424]
+  - - [4096, 4095, 1, 4096]
+    - [1, 1607.6]
+  - - [512, 512, 1, 512]
+    - [8, 3445.01]
   - - [3072, 3072, 1, 3073]
-    - [1, 1645.1]
+    - [1, 1645.65]
+  - - [3071, 3072, 1, 3072]
+    - [1, 1650.57]
+  - - [2880, 3073, 1, 3072]
+    - [1, 1551.9]
+  - - [2048, 2047, 1, 2048]
+    - [1, 1383.25]
+  - - [2880, 3072, 1, 3072]
+    - [8, 17010.8]
+  - - [63, 1024, 1, 1024]
+    - [6, 109.939]
+  - - [1920, 2049, 1, 2048]
+    - [1, 1306.24]
+  - - [960, 1024, 1, 1024]
+    - [8, 13827.4]
+  - - [1920, 2048, 1, 2048]
+    - [8, 15206.0]
+  - - [4096, 4096, 1, 4097]
+    - [1, 1627.47]
+  - - [2048, 2048, 1, 2048]
+    - [8, 13584.8]
+  - - [3072, 3072, 1, 3071]
+    - [1, 1639.2]
+  - - [3840, 4095, 1, 4096]
+    - [4, 1782.78]
+  - - [3840, 4096, 1, 4097]
+    - [4, 1872.84]
+  - - [3841, 4096, 1, 4096]
+    - [1, 1592.57]
+  - - [2048, 2048, 1, 2047]
+    - [1, 1375.46]
+  - - [481, 512, 1, 512]
+    - [0, 359.439]
+  - - [3839, 4096, 1, 4096]
+    - [4, 1789.24]
+  - - [64, 1025, 1, 1024]
+    - [6, 112.921]
   - - [3072, 3073, 1, 3072]
-    - [1, 1651.42]
-  - - [3073, 3072, 1, 3072]
-    - [1, 1652.4]
-  - - [4079, 4096, 1, 4096]
-    - [1, 1600.96]
-  - - [4080, 4095, 1, 4096]
-    - [1, 1600.47]
-  - - [4080, 4096, 1, 4095]
-    - [4, 1604.11]
-  - - [4080, 4096, 1, 4097]
-    - [1, 1621.16]
-  - - [4080, 4097, 1, 4096]
-    - [1, 1602.61]
-  - - [4081, 4096, 1, 4096]
-    - [1, 1599.52]
+    - [1, 1650.31]
+  - - [1919, 2048, 1, 2048]
+    - [5, 1758.38]
+  - - [2881, 3072, 1, 3072]
+    - [1, 1553.64]
   - - [4095, 4096, 1, 4096]
-    - [1, 1604.71]
-  - - [4096, 4095, 1, 4096]
-    - [1, 1604.32]
-  - - [4096, 4096, 1, 4095]
-    - [4, 1609.64]
-  - - [4096, 4096, 1, 4097]
-    - [1, 1625.9]
-  - - [4096, 4097, 1, 4096]
-    - [1, 1605.8]
+    - [1, 1605.24]
+  - - [1024, 1023, 1, 1024]
+    - [0, 859.2]
   - - [4097, 4096, 1, 4096]
-    - [1, 1603.85]
-  - - [960, 1024, 1, 1024]
-    - [5, 9007.14]
+    - [1, 1605.11]
+  - - [3840, 4096, 1, 4095]
+    - [4, 1872.71]
+  - - [512, 511, 1, 512]
+    - [0, 379.349]
+  - - [3073, 3072, 1, 3072]
+    - [1, 1652.19]
+  - - [4096, 4096, 1, 4095]
+    - [4, 1608.39]
+  - - [1024, 1024, 1, 1023]
+    - [0, 864.574]
   - - [1024, 1024, 1, 1024]
-    - [5, 7604.43]
-  - - [2040, 2048, 1, 2048]
-    - [5, 11066.3]
-  - - [2048, 2048, 1, 2048]
-    - [5, 14117.1]
-  - - [3000, 3072, 1, 3072]
-    - [5, 12651.4]
-  - - [3072, 3072, 1, 3072]
-    - [5, 12416.1]
-  - - [4080, 4096, 1, 4096]
-    - [5, 16848.0]
-  - - [4096, 4096, 1, 4096]
-    - [5, 16911.1]
-  - - [63, 1024, 1, 1024]
-    - [6, 111.559]
-  - - [64, 1023, 1, 1024]
-    - [6, 113.219]
-  - - [64, 1024, 1, 1023]
-    - [6, 114.051]
-  - - [64, 1024, 1, 1025]
-    - [6, 114.243]
-  - - [64, 1025, 1, 1024]
-    - [6, 112.937]
+    - [8, 8505.59]
+  - - [512, 513, 1, 512]
+    - [0, 379.887]
+  - - [960, 1024, 1, 1023]
+    - [3, 1078.56]
+  - - [513, 512, 1, 512]
+    - [0, 379.458]
   - - [65, 1024, 1, 1024]
-    - [6, 109.143]
+    - [7, 105.664]
+  - - [64, 1024, 1, 1025]
+    - [6, 113.532]
   - - [64, 1024, 1, 1024]
-    - [7, 1762.34]
+    - [9, 967.544]
+  - - [1920, 2048, 1, 2047]
+    - [5, 1773.88]
+  - - [2880, 3072, 1, 3071]
+    - [3, 1659.92]
+  - - [1920, 2048, 1, 2049]
+    - [5, 1786.73]
+  - - [1023, 1024, 1, 1024]
+    - [0, 862.294]
+  - - [2048, 2048, 1, 2049]
+    - [1, 1388.52]
 - null