Skip to content

[TensileLite] Fix MX FP4 scale data initialization & argument scoping #4917

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
pdhirajkumarprasad merged 2 commits into from
Feb 26, 2026
Merged

[TensileLite] Fix MX FP4 scale data initialization & argument scoping #4917

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
6a346d6
Fix failing tests
archana-ramalingam Feb 26, 2026
4ba510a
Fix tests
archana-ramalingam Feb 26, 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
19 changes: 9 additions & 10 deletions projects/hipblaslt/tensilelite/Tensile/Components/Signature.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -221,20 +221,19 @@ def __call__(self, writer) -> SignatureBase:
if kernel["ProblemType"]["UseScaleAB"]:
signature.addArg("AddressScaleA", SVK.SIG_GLOBALBUFFER, cptValueType, "generic")
signature.addArg("AddressScaleB", SVK.SIG_GLOBALBUFFER, cptValueType, "generic")
userArgumentsInfo.scaleASize += 8
userArgumentsInfo.scaleBSize += 8
userArgumentsInfo.scaleASize += 8
userArgumentsInfo.scaleBSize += 8
if kernel["ProblemType"]["UseScaleCD"]:
signature.addArg("AddressScaleC", SVK.SIG_GLOBALBUFFER, cptValueType, "generic")
signature.addArg("AddressScaleD", SVK.SIG_GLOBALBUFFER, cptValueType, "generic")
userArgumentsInfo.scaleCSize += 8
userArgumentsInfo.scaleDSize += 8
userArgumentsInfo.scaleCSize += 8
userArgumentsInfo.scaleDSize += 8

if kernel["ProblemType"]["UseScaleAlphaVec"]:
signature.addArg("AddressScaleAlphaVec", SVK.SIG_GLOBALBUFFER, cptValueType, "generic")
if kernel["ProblemType"]["UseScaleAlphaVec"] == 3:
userArgumentsInfo.factorDimSize =4

userArgumentsInfo.scaleAlphaVecSize += 8
userArgumentsInfo.scaleAlphaVecSize += 8

if writer.states.useBias != DataDirection.NONE:
signature.addArg("bias", SVK.SIG_GLOBALBUFFER, biasValueType, "generic") # Note: We append the data in ws_d
Expand All @@ -243,7 +242,7 @@ def __call__(self, writer) -> SignatureBase:
signature.addArg("StrideBias", SVK.SIG_VALUE, "u32")
if kernel["ProblemType"]["UseBias"] == 3:
userArgumentsInfo.factorDimSize = 4
userArgumentsInfo.biasSize += (8 + 4 + 4)
userArgumentsInfo.biasSize += (8 + 4 + 4)

if userArgumentsInfo.factorDimSize == 4:
signature.addArg("factorDim", SVK.SIG_VALUE, "u32")
Expand All @@ -252,9 +251,9 @@ def __call__(self, writer) -> SignatureBase:
signature.addArg( "E", SVK.SIG_GLOBALBUFFER, cptValueType, "generic")
for i in range(0, writer.states.e.numSgprStrides):
signature.addArg("StrideE%u"%i, SVK.SIG_VALUE, "u32")
userArgumentsInfo.eSize += 8
for i in range(0, writer.states.e.numSgprStrides):
userArgumentsInfo.eSize += 4
userArgumentsInfo.eSize += 8
for i in range(0, writer.states.e.numSgprStrides):
userArgumentsInfo.eSize += 4

if ((kernel["ProblemType"]["ActivationType"] != 'none') and kernel["ActivationFused"]):
if kernel["ProblemType"]["ActivationComputeDataType"].isHalf():
Expand Down
5 changes: 4 additions & 1 deletion projects/hipblaslt/tensilelite/client/src/DataInitialization.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1709,7 +1709,10 @@ namespace TensileLite
or i == ContractionProblemGemm::TENSOR::METADATA)
continue;

if(useMXGenerator && (i == ContractionProblemGemm::TENSOR::A || i == ContractionProblemGemm::TENSOR::B))
if(useMXGenerator && (i == ContractionProblemGemm::TENSOR::A
|| i == ContractionProblemGemm::TENSOR::B
|| i == ContractionProblemGemm::TENSOR::MXSA
|| i == ContractionProblemGemm::TENSOR::MXSB))
continue;

if(m_problemDependentData)
Expand Down
3 changes: 1 addition & 2 deletions projects/hipblaslt/tensilelite/src/ContractionSolution.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -809,8 +809,7 @@ namespace TensileLite

// NOTE: an assumption here is A & B must be both MX data types or non-MX data types.
// Mixing is not supported.
if(!problemType.useScaleAB.empty() or
(problemType.mxBlockA != 0 && problemType.mxBlockB != 0)) //kernel input data
if(!problemType.useScaleAB.empty())
{
args.template append<void const*>("scaleA", inputs.scaleA);
args.template append<void const*>("scaleB", inputs.scaleB);
Expand Down
Loading