Skip to content

Modify LocalRead.py to deal with LDS > 64K #614

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
amgddm merged 3 commits into from
Jul 18, 2025
Merged

Modify LocalRead.py to deal with LDS > 64K #614

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8c3e07d
Modify LocalRead.py to deal with LDS > 64K
amgddm Jul 11, 2025
51ebf92
Merge branch 'develop' into LDSTrOffset
amgddm Jul 17, 2025
70cda35
Merge branch 'develop' into LDSTrOffset
amgddm Jul 18, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 17 additions & 8 deletions projects/hipblaslt/tensilelite/Tensile/Components/LocalRead.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,6 +148,15 @@ def __call__(self, writer, kernel, bufferIdx, iui, epsi, tP):

class LocalReadMFMA(LocalRead):
kernel = {"EnableMatrixInstruction": True}

# LDS size is increased on gfx950. const offset is still 16-bit.
# this function handles both LDS size < 64K and LDS size >= 64K
def cal_offset_srcAddr(self, maxLDSConstOffset, tc, offset):
num = offset // maxLDSConstOffset
offset_val = offset - num * maxLDSConstOffset
srcAddr = vgpr("LocalReadAddr%s+%u" %(tc, num))
return offset_val, srcAddr


"""
Local Read: Do It A/B
Expand Down Expand Up @@ -234,6 +243,7 @@ def __call__(self, writer, kernel, bufferIdx, iui, epsi, tP):
blockStride = elementsPerBlockSMFMA * threadGroups
blockOffsetSMFMA = blockStride - elementsPerBlockSMFMA

maxLDSConstOffset = writer.states.regCaps["maxLDSConstOffset"]
valufIdx = 0
if enableLDSTr:
numberMTilesPerWave = kernel["MIWaveTile"][tile01]
Expand All @@ -242,20 +252,19 @@ def __call__(self, writer, kernel, bufferIdx, iui, epsi, tP):
offset_val = (tP["localReadOffset"]+MIWaveGroupShape[tile01]*tIdx) * tP["bpeDS"]
if (kernel["LdsBlockSizePerPad%s"%tc] != 0) and (kernel["LdsPad%s"%tc] != 0):
offset_val = offset_val + (offset_val // kernel["LdsBlockSizePerPad%s"%tc]) * kernel["LdsPad%s"%tc] * tP["bpeDS"]
paramList = []
paramList.append(int(offset_val))
ds = DSModifiers(na=1, offset=paramList[0])
offset, srcAddr = self.cal_offset_srcAddr(maxLDSConstOffset, tc, offset_val)
ds = DSModifiers(na=1, offset=offset)
LocalReadX = instruction.getInst(highBits)
destVgpr = vgpr("Valu%s_X%u_I%u+%u+0"%(tc,bufferIdx,iui, 4*tIdx), 2)
comment = "LDS Transpose"
valuiIdx = int(valufIdx)
localReadCode = imod.add(Module("LocalRead%s Valu%u"%(tc,valuiIdx)))
localReadCode.add(LocalReadX(dst=destVgpr, src=vgpr("LocalReadAddr%s"%tc), ds=ds, comment=comment))
localReadCode.add(LocalReadX(dst=destVgpr, src=srcAddr, ds=ds, comment=comment))
destVgpr = vgpr("Valu%s_X%u_I%u+%u+2"%(tc,bufferIdx,iui,4*tIdx), 2)
offset_val = paramList[0]+UnrollStride*inputPerThread;
paramList.append(int(offset_val))
ds = DSModifiers(na=1, offset=paramList[1])
localReadCode.add(LocalReadX(dst=destVgpr, src=vgpr("LocalReadAddr%s"%tc), ds=ds, comment=comment))
offset_val += UnrollStride*inputPerThread;
offset, srcAddr = self.cal_offset_srcAddr(maxLDSConstOffset, tc, offset_val)
Comment thread
b-shi marked this conversation as resolved.
Resolved
ds = DSModifiers(na=1, offset=offset)
localReadCode.add(LocalReadX(dst=destVgpr, src=srcAddr, ds=ds, comment=comment))
else:
for vIdx in range(0, numVectorsPerTile):
for eIdx in range(0, numReadsPerVector):
Expand Down
63 changes: 63 additions & 0 deletions projects/hipblaslt/tensilelite/Tensile/Tests/common/gemm/gfx950/lds_tr.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -786,3 +786,66 @@ BenchmarkProblems:
- Exact: [512, 512, 1, 1088]
- Exact: [519, 519, 1, 1120]
- Exact: [2053, 2053, 1, 2053]

########################################
# HSS - Test LDS > 64K
########################################
-
- OperationType: GEMM
DataType: H
DestDataType: S
ComputeDataType: S
HighPrecisionAccumulate: true
TransposeA: False
TransposeB: False
#ComplexConjugateA: False
#ComplexConjugateB: False
UseBeta: True
#UseBias: 1
Batched: True
#Activation: True
#ActivationHPA: True
- InitialSolutionParameters:
BenchmarkCommonParameters:
- KernelLanguage: [Assembly]
ForkParameters:
- 1LDSBuffer: [1] # Default Value: [0] # Range: [-1, 0, 1]
- ClusterLocalRead: [1] # Default Value: [1] # Range: [0, 1]
- DepthU: [128] # Default Value: [-1] # Range: [-1, 2, ... ,1023, 1024]
- DirectToVgprA: [0] # Default Value: [False] # Range: [False, True]
- DirectToVgprB: [0] # Default Value: [False] # Range: [False, True]
- GlobalReadVectorWidthA: [8] # Default Value: [-1] # Range: [-2, -1, ... ,8, 16]
- GlobalReadVectorWidthB: [8] # Default Value: [-1] # Range: [-2, -1, ... ,8, 16]
- GlobalSplitUAlgorithm: ["MultipleBuffer"] # Default Value: [MultipleBuffer] # Range: [SingleBuffer, MultipleBuffer, MultipleBufferSingleKernel]
- LocalReadVectorWidth: [-1] # Default Value: [-1] # Range: [-1, 1, ... ,8, 16]
- MIArchVgpr: [1] # Default Value: [False] # Range: [False, True]
- NonTemporalA: [0] # Default Value: [0] # Range: [0, 1, ... ,6, 7]
- NumElementsPerBatchStore: [16] # Default Value: [0] # Range: [-1, 0, ... ,254, 255]
- PrefetchGlobalRead: [2] # Default Value: [1] # Range: [0, 1, 2]
- PrefetchLocalRead: [1] # Default Value: [1] # Range: [0, 1, ... ,127, 128]
- SourceSwap: [1] # Default Value: [False] # Range: [False, True]
- StaggerU: [8] # Default Value: [32] # Range: [0, 2, ... ,32, 64]
- StorePriorityOpt: [1] # Default Value: [False] # Range: [False, True]
- StoreSyncOpt: [4] # Default Value: [0] # Range: [0, 1, ... ,254, 255]
- StreamK: [3] # Default Value: [0] # Range: [0, 1, ... ,2, 3]
- StreamKXCCMapping: [8] # Default Value: [0] # Range: [0, 2, ... ,7, 8]
- UseSgprForGRO: [-1] # Default Value: [-1] # Range: [-1, 0, 1]
- VectorWidthA: [-1] # Default Value: [-1] # Range: [-1, 1, ... ,6, 8]
- VectorWidthB: [-1] # Default Value: [-1] # Range: [-1, 1, ... ,6, 8]
- WaveSeparateGlobalReadA: [0] # Default Value: [0] # Range: [0, 1, 2]
- WaveSeparateGlobalReadB: [0] # Default Value: [0] # Range: [0, 1, 2]
- WorkGroupMapping: [16] # Default Value: [8] # Range: [-1024, -1023, ... ,1023, 1024]
- WorkGroupMappingXCC: [1] # Default Value: [1] # Range: [1, 2, ... ,16, 32]
- Groups:
- - NonTemporalC: 4 # Default Value: [0] # Range: [0, 1, ... ,6, 7]
NonTemporalD: 4 # Default Value: [0] # Range: [0, 1, ... ,6, 7]
- - MatrixInstruction: [32, 32, 16, 1, 1, 5, 2, 2, 2] #MT 320x128 - TT 5x64 - WG 64x4 - MIBlockM 1 -- sizes [[(40960, 192, 1, 256)]]
MIArchVgpr: 0 # Default Value: [False] # Range: [False, True]
- - ClusterLocalRead: 0 # Default Value: [1] # Range: [0, 1]
LDSTrInst: 1 # Default Value: [False] # Range: [False, True]
#- ClusterLocalRead: 1 # Default Value: [1] # Range: [0, 1]
#LDSTrInst: 0 # Default Value: [False] # Range: [False, True]
BenchmarkJoinParameters:
BenchmarkFinalParameters:
- ProblemSizes:
- Exact: [ 40960, 192, 1, 256 ]