Skip to content

fix(hipblaslt): avoid regenerating unchanged MX inputs on every solution #11963

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
minsukim-amd merged 1 commit into from
Sep 11, 2026
Merged

fix(hipblaslt): avoid regenerating unchanged MX inputs on every solution #11963

Show file tree
Hide file tree
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
5 changes: 5 additions & 0 deletions projects/hipblaslt/tensilelite/client/include/DataInitialization.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,7 @@
#include <mxDataGen.hpp>

#include <cstddef>
#include <map>
#include <random>

#include "RunListener.hpp"
Expand Down Expand Up @@ -1191,6 +1192,10 @@ namespace TensileLite
// hand back gpuInput.valid as-is rather than re-swizzling).
bool m_mxPreswizzledA = false;
bool m_mxPreswizzledB = false;

// Scale descriptor most recently swizzled into gpuInput.valid, per
// tensor index. The MX equivalent of g_swizzleCache.
std::map<size_t, TensorDescriptor> m_mxSwizzledDescriptor;
};

template <>
Expand Down
32 changes: 32 additions & 0 deletions projects/hipblaslt/tensilelite/client/src/DataInitialization.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2130,6 +2130,28 @@ namespace TensileLite
auto& pristineData = m_vdata[dataTensorEnum].pristine[dataDesc.dataType()];
auto& pristineScale = m_vdata[scaleTensorEnum].pristine[scaleEltType];

// prepareGPUInputs() runs once per solution, so without this
// the generator re-runs for every solution on identical input.
// Same idea as the non-MX initDescriptor check below.
//
// gfx1250 only: its scale layout comes from problem geometry, so
// the descriptors fully determine the generated bytes. GFX950
// preswizzles per solution matrix instruction and must not reuse.
bool const reuseSupported = (m_mxScaleLayout == MXScaleLayout::GFX1250);
bool const descriptorsUnchanged
= pristineData.initDescriptor.size() == 1
&& pristineScale.initDescriptor.size() == 1
&& pristineData.initDescriptor[0] == dataDesc
&& pristineScale.initDescriptor[0] == scaleDesc;

if(reuseSupported && descriptorsUnchanged)
return;

if(pristineData.initDescriptor.size() == 1)
pristineData.initDescriptor[0] = dataDesc;
if(pristineScale.initDescriptor.size() == 1)
pristineScale.initDescriptor[0] = scaleDesc;

// Element-size-aware byte stride: FP4 (0.5), FP6 (0.75), FP8 (1.0).
size_t dataBatchStrideBytes = 0;
size_t scaleBatchStrideBytes = 0;
Expand Down Expand Up @@ -2848,8 +2870,18 @@ namespace TensileLite
{
ptr = p.gpuInput.valid.get();
}
else if(auto mxIt = m_mxSwizzledDescriptor.find(i);
m_mxScaleLayout == MXScaleLayout::GFX1250
&& mxIt != m_mxSwizzledDescriptor.end() && mxIt->second == desc)
{
// Already swizzled, and initOneMXSide kept the source bytes
// unchanged, so skip the host-side permute.
ptr = p.gpuInput.valid.get();
}
else
{
m_mxSwizzledDescriptor[i] = desc;

// Fallback K-dimension swizzle for the case where
// initializeMXData did NOT pre-produce a swizzled
// scale into gpuInput.valid (e.g. m_mxScaleLayout is
Expand Down
Loading