Skip to content

Changing sgpr limits #2184

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

Closed
mahmoodw wants to merge 2 commits into from
Closed

Changing sgpr limits #2184

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
619364e
changing sgpr limits
Jun 4, 2025
7b85b73
Fix conflicts from merge
Jun 16, 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
33 changes: 21 additions & 12 deletions tensilelite/Tensile/Components/GSU.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -705,18 +705,27 @@ def GSUSynccodegen(self, writer, kernel, tmpVgpr, tmpVgprSize, tmpVgprDynamic, b

indices = list(range(0, kernel["ProblemType"]["NumIndicesC"]))
numDim = len(indices)
with writer.allocTmpSgpr(5) as tmpSgprInfo:
tmpSgpr = tmpSgprInfo.idx
module.addModuleAsFlatItems(writer.s_mul_u64_u32(sgpr(tmpSgpr+0), sgpr(tmpSgpr+1), sgpr("SizesFree+0"), 1, tmpVgpr, "Free0"))
for i in range(1, numDim):
module.add(SSubU32(dst=sgpr(tmpSgpr+4), src0=sgpr("SizesFree+%u"%i), src1=1, comment="Free%u" % i))
module.add(SMulI32(dst=sgpr(tmpSgpr+4), src0=sgpr(tmpSgpr+4), src1=1, comment="Free%u" % i))
module.addModuleAsFlatItems(writer.s_mul_u64_u32(sgpr(tmpSgpr+2), sgpr(tmpSgpr+3), sgpr(tmpSgpr+4), sgpr("StrideC%s"%writer.states.indexChars[i]), tmpVgpr, "Free%u" % i))
module.add(SAddU32(dst=sgpr(tmpSgpr+0), src0=sgpr(tmpSgpr+0), src1=sgpr(tmpSgpr+2), comment="Free%u" % i))
module.add(SAddCU32(dst=sgpr(tmpSgpr+1), src0=sgpr(tmpSgpr+1), src1=sgpr(tmpSgpr+3), comment="Free%u" % i))

bpetmp = int(writer.states.bpr * kernel["ProblemType"]["DestDataType"].numRegisters()) # self.states.bpeCinternal
module.add(SLShiftLeftB64(dst=sgpr(tmpS04,2), src=sgpr(tmpSgpr+0,2), shiftHex=log2(writer.states.bpeCexternal), comment="scale by bpe"))
tmpSgpr1 = writer.sgprPool.checkOut(2, preventOverflow=False)
tmpSgpr2 = writer.sgprPool.checkOut(1, preventOverflow=False)
tmpSgpr3 = writer.sgprPool.checkOut(1, preventOverflow=False)
tmpSgpr4 = writer.sgprPool.checkOut(1, preventOverflow=False)

module.addModuleAsFlatItems(writer.s_mul_u64_u32(sgpr(tmpSgpr1+0), sgpr(tmpSgpr1+1), sgpr("SizesFree+0"), 1, tmpVgpr, "Free0"))

for i in range(1, numDim):
module.add(SSubU32(dst=sgpr(tmpSgpr4), src0=sgpr("SizesFree+%u" % i), src1=1, comment="Free%u" % i))
module.add(SMulI32(dst=sgpr(tmpSgpr4), src0=sgpr(tmpSgpr4), src1=1, comment="Free%u" % i))
module.addModuleAsFlatItems(writer.s_mul_u64_u32(sgpr(tmpSgpr2), sgpr(tmpSgpr3), sgpr(tmpSgpr4), sgpr("StrideC%s" % writer.states.indexChars[i]), tmpVgpr, "Free%u" % i))
module.add(SAddU32(dst=sgpr(tmpSgpr1+0), src0=sgpr(tmpSgpr1+0), src1=sgpr(tmpSgpr2), comment="Free%u" % i))
module.add(SAddCU32(dst=sgpr(tmpSgpr1+1), src0=sgpr(tmpSgpr1+1), src1=sgpr(tmpSgpr3), comment="Free%u" % i))

bpetmp = int(writer.states.bpr * kernel["ProblemType"]["DestDataType"].numRegisters()) # self.states.bpeCinternal
module.add(SLShiftLeftB64(dst=sgpr(tmpS04, 2), src=sgpr(tmpSgpr1+0, 2), shiftHex=log2(writer.states.bpeCexternal), comment="scale by bpe"))

writer.sgprPool.checkIn(tmpSgpr1)
writer.sgprPool.checkIn(tmpSgpr2)
writer.sgprPool.checkIn(tmpSgpr3)
writer.sgprPool.checkIn(tmpSgpr4)

module.addSpaceLine()
#####################################cal synchronizer sum start#####################################
Expand Down
4 changes: 3 additions & 1 deletion tensilelite/Tensile/KernelWriterAssembly.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11614,7 +11614,9 @@ def getMBSKGSUTotal(self, kernel):

def setOccupancy(self, kernel):
# Use VGPR up to next occupancy threshold:
maxVgprs, occupancy = self.getMaxRegsForOccupancy(kernel["NumThreads"], self.vgprPool.size(), self.sgprPool.size(), \
# Account for additional temp sgprs that will be required for code gen, up to physical limits. +5 approximates upper end of required temp space for GSU sync
requiredSgprs = min(self.sgprPool.size() + 5, self.states.regCaps["MaxSgpr"])
Copy link
Copy Markdown
Collaborator

@jichangjichang jichangjichang Jun 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

"+5" should not be required for all kernels.

Copy link
Copy Markdown
Contributor Author

@mahmoodw mahmoodw Jun 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That's right, this is meant to be a temporary fix to unblock perf teams. A debt ticket will be issued to more accurately collect the high-water mark. Any suggestions or insight is appreciated.

maxVgprs, occupancy = self.getMaxRegsForOccupancy(kernel["NumThreads"], self.vgprPool.size(), requiredSgprs, \
self.getLdsSize(kernel), self.agprPool.size(), self.states.doubleVgpr)
# Set occupancy limit for register pools
# TODO: Support gfx12
Expand Down
Loading