Skip to content

Refine the method for caculating sparse mfma offset at tile section #2195

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
merged 1 commit into from
Jun 13, 2025
Merged

Refine the method for caculating sparse mfma offset at tile section #2195

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
47 changes: 29 additions & 18 deletions tensilelite/Tensile/KernelWriterAssembly.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6410,12 +6410,25 @@ def dataTypeNameAbbrevToInstType(abbrev: str, sourceSwap: bool = False) -> InstT
if numTmpSgpr == 4:
tmpSgprX3 = tmpSgprInfo.idx+3

sparseInputBlocksA = 1
sparseInputBlocksB = 1
if kernel["MIInputPerThread"] * kernel["ProblemType"]["DataTypeB"].numBytes() > 16 :
sparseInputBlocksA = 2 if kernel["ProblemType"]["Sparse"] == 2 else 1
sparseInputBlocksB = 2 if kernel["ProblemType"]["Sparse"] == 1 else 1

def findSparseOffset(isA:bool):
blocksPerTGroupSMFMA = 1
elementsPerBlockSMFMA = 1
blockOffsetSMFMA =1
if kernel["ProblemType"]["Sparse"] != 0:
if kernel["MIInputPerThread"] * kernel["ProblemType"]["DataTypeB"].numBytes() > 16: # double K
isSparseTrack = (kernel["ProblemType"]["Sparse"] == 1 and isA) or (kernel["ProblemType"]["Sparse"] == 2 and not isA)
# gfx950 sparse track only has one block for each thread group.
# TODO adjust this value for other arch.
blocksPerTGroupSMFMA = 1 if isSparseTrack else 2
if blocksPerTGroupSMFMA > 1:
threadGroups = kernel["MatrixInstK"] // kernel["MIInputPerThread"]
elementsPerBlockSMFMA = kernel["MIInputPerThread"] // blocksPerTGroupSMFMA
blockStride = elementsPerBlockSMFMA * threadGroups
blockOffsetSMFMA = blockStride - elementsPerBlockSMFMA
return blocksPerTGroupSMFMA, elementsPerBlockSMFMA, blockOffsetSMFMA

blocksPerTGroupSMFMAA, elementsPerBlockSMFMAA, blockOffsetSMFMAA = findSparseOffset(True)
blocksPerTGroupSMFMAB, elementsPerBlockSMFMAB, blockOffsetSMFMAB = findSparseOffset(False)
# replace 0 for differnet thread
if kernel["ProblemType"]["Sparse"] == 1 and numMIInput//8 >= 1:
vgprPerSet0Group = 1
Expand All @@ -6436,16 +6449,15 @@ def dataTypeNameAbbrevToInstType(abbrev: str, sourceSwap: bool = False) -> InstT
for group in range(0, numSet0GroupA):
if numSet0GroupA > 1 or (is_wmma_v2 and vgprPerInputA > 2):
if group == 0:
if kernel["ProblemType"]["Sparse"] == 1:
multiplyBy = numMIInput//sparseInputBlocksA
if kernel["ProblemType"]["Sparse"]:
multiplyBy = numMIInput//blocksPerTGroupSMFMAA
else:
multiplyBy = numMIInput//2 if vgprPerInputA == 8 else numMIInput
shiftK.add(vectorStaticMultiply(vgpr(kReg_first), vgpr(kReg_first), multiplyBy, tmpSgprInfo))
shiftK.add(VAddU32(vgpr(kReg), vgpr(kReg_first), 0, ""))
elif kernel["ProblemType"]["Sparse"] == 2 and group == 2 and sparseInputBlocksA == 2:
is8bits = kernel["ProblemType"]["DataType"].numBytes() == 1
strideK = 3 if kernel["MatrixInstK"]== 32 or (is8bits and kernel["MatrixInstK"]== 64) else 7
shiftK.add(VAddU32(vgpr(kReg), vgpr(kReg), numMIInput//numSet0GroupA*strideK, "add part of K"))
elif blocksPerTGroupSMFMAA == 2 and (group * vgprPerSet0Group) == (elementsPerBlockSMFMAA * numRegistersIn):
kIncA = blockOffsetSMFMAA + (numMIInput//numSet0GroupA) * max(group - 1, 0)
shiftK.add(VAddU32(vgpr(kReg), vgpr(kReg), kIncA, "add part of K"))
else:
kIncA = numMIInput//numSet0GroupA
if self.states.asmCaps["HasMFMA_f8f6f4"]:
Expand Down Expand Up @@ -6482,16 +6494,15 @@ def dataTypeNameAbbrevToInstType(abbrev: str, sourceSwap: bool = False) -> InstT
for group in range(0, numSet0GroupB):
if numSet0GroupB > 1 or (is_wmma_v2 and vgprPerInputB > 2):
if group == 0:
if kernel["ProblemType"]["Sparse"] == 1:
multiplyBy = numMIInput//sparseInputBlocksB
if kernel["ProblemType"]["Sparse"]:
multiplyBy = numMIInput//blocksPerTGroupSMFMAB
else:
multiplyBy = numMIInput//2 if vgprPerInputB == 8 else numMIInput
shiftK.add(vectorStaticMultiply(vgpr(kReg_first), vgpr(kReg_first), multiplyBy, tmpSgprInfo))
shiftK.add(VAddU32(vgpr(kReg), vgpr(kReg_first), 0, ""))
elif kernel["ProblemType"]["Sparse"] == 1 and group == 2 and sparseInputBlocksB == 2:
is8bits = kernel["ProblemType"]["DataType"].numBytes() == 1
strideK = 3 if kernel["MatrixInstK"]== 32 or (is8bits and kernel["MatrixInstK"]== 64) else 7
shiftK.add(VAddU32(vgpr(kReg), vgpr(kReg), numMIInput//numSet0GroupB*strideK, "add part of K"))
elif blocksPerTGroupSMFMAB == 2 and (group * vgprPerSet0Group) == (elementsPerBlockSMFMAB * numRegistersIn):
kIncB = blockOffsetSMFMAB + (numMIInput//numSet0GroupB) * max(group - 1, 0)
shiftK.add(VAddU32(vgpr(kReg), vgpr(kReg), kIncB, "add part of K"))
else:
kIncB = numMIInput//numSet0GroupB
if group == 2 and vgprPerInputB == 8:
Expand Down