Skip to content

Fix a verification fail with MXFP4 + non DTL #4715

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
nakajee merged 3 commits into from
Feb 24, 2026
Merged

Fix a verification fail with MXFP4 + non DTL #4715

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4e8baaf
Fix a verification fail with MXFP4 + non DTL
nakajee Feb 19, 2026
49b0fbc
Merge remote-tracking branch 'origin/gfx950_mx_rebase' into users/nak…
nakajee Feb 21, 2026
9f1dd94
Merge branch 'gfx950_mx_rebase' into users/nakajee/gfx950mxfp4_nondtl…
nakajee Feb 24, 2026
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
14 changes: 13 additions & 1 deletion projects/hipblaslt/tensilelite/Tensile/KernelWriter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1469,7 +1469,7 @@ def calculateRangeAndUpdateCounter(itemCounter, writeCounters, length):
iterCode.add(waitDsRead)
else:
if kernel["UnrollMajorLDSB"] and not (kernel["ProblemType"]["DataTypeB"].isAnyFloat8() and kernel["ConvertAfterDS"]):
if iteration == 0 and i == kernel["MIWaveTileA"]:
if iteration == 0 and i == (kernel["MIWaveTileA"] // kernel["numSubTiles"]):
# add 1 more waitcnt before using ds read data
waitCode2 = deepcopy(waitCode)
waitCode2.dscnt = localReadsIssuedInThisIter
Expand Down Expand Up @@ -5878,13 +5878,19 @@ def GNLCOInit(tc):
# reads Per Iteration
########################################
if kernel["EnableMatrixInstruction"]:
factorSubIterA = kernel["numSubTiles"]
if kernel["UnrollMajorLDSA"] or kernel["enableLDSTrA"]:
self.states.numReadsPerUnrollA = ceil(tensorParametersA["bpe"] * kernel["MIInputPerThreadA"] / int(tensorParametersA["localReadInstruction"].blockWidth * 4))
else:
self.states.numReadsPerUnrollA = kernel["MIInputPerThreadA"]
if kernel["ForceUnrollSubIter"] and self.states.numReadsPerUnrollA > 1:
self.states.numReadsPerUnrollA //= kernel["numSubTiles"]
factorSubIterA = 1
numA = kernel["InnerUnroll"]*(kernel["MIWaveTile"][0] * self.states.numReadsPerUnrollA) // tensorParametersA["localReadInstruction"].numOffsets
if self.states.lrvwTileA > 1:
numA = numA // kernel["VectorWidthA"]
if kernel["ForceUnrollSubIter"]:
numA = numA // factorSubIterA

if kernel["ProblemType"]["MXBlockA"]:
self.states.numReadsPerUnrollMXSA = 1
Expand All @@ -5900,13 +5906,19 @@ def GNLCOInit(tc):
tileM = kernel["MIWaveTile"][1] if kernel["ProblemType"]["Sparse"] == 2 else kernel["MIWaveTile"][0]
numM = kernel["InnerUnroll"]*(tileM * self.states.numReadsPerUnrollMetadata) // tensorParametersM["localReadInstruction"].numOffsets

factorSubIterB = kernel["numSubTiles"]
if kernel["UnrollMajorLDSB"] or kernel["enableLDSTrB"]:
self.states.numReadsPerUnrollB = ceil(tensorParametersB["bpe"] * kernel["MIInputPerThreadB"] / int(tensorParametersB["localReadInstruction"].blockWidth * 4))
else:
self.states.numReadsPerUnrollB = kernel["MIInputPerThreadB"]
if kernel["ForceUnrollSubIter"] and self.states.numReadsPerUnrollB > 1:
self.states.numReadsPerUnrollB //= kernel["numSubTiles"]
factorSubIterB = 1
numB = kernel["InnerUnroll"]*(kernel["MIWaveTile"][1] * self.states.numReadsPerUnrollB) // tensorParametersB["localReadInstruction"].numOffsets
if self.states.lrvwTileB > 1:
numB = numB // kernel["VectorWidthB"]
if kernel["ForceUnrollSubIter"]:
numB = numB // factorSubIterB

if kernel["ProblemType"]["MXBlockB"]:
self.states.numReadsPerUnrollMXSB = 1
Expand Down
Loading