[rocblas][tensile] Use kernel ISA during build with enforcement #2162

bstefanuk · 2025-10-17T22:55:23Z

Motivation

Performance regressions are found when adding --no-enumerate to the TensileCreateLibrary build. This PR re-implements the kernel ISA reliance from #2094 without needing to change logic files.

Technical Details

Rely only on the kernel's ISA during the build phase.
Add additional ISA enforcement given architecture details extracted from logic files.

Test Plan

Local performance testing for specific sizes
Comprehensive performance testing through gemmaiperf
Standard CI testing

Test Result

See CI results in this PR for standard pipeline checks.
Performance: tested on 6665 sizes using rocblas-bench on gfx950 (results below)
Performance: select sizes were evaluated on gfx942 and confirmed no performance change beyond +/-1%

Single precision NN

Stat	Result
Average (% speed up)	0.50
Median (% speed up)	0.01
Count Faster	3482
Count Slower	3161

Single precision TN

Stat	Result
Average (% speed up)	4.17
Median (%speed up)	-0.02
Count Faster	3042
Count Slower	3579

Complex double precision TN

Stat	Result
Average (% speed up)	0.18
Median (% speed up)	0.04
Count Faster	4452
Count Slower	2207

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

codecov-commenter · 2025-10-18T00:19:34Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

❗ There is a different number of reports uploaded between BASE (5f34bda) and HEAD (ebb8f13). Click for more details.

HEAD has 1 upload less than BASE

Flag BASE (5f34bda) HEAD (ebb8f13)

hipSPARSE 1 0

Additional details and impacted files

@@             Coverage Diff              @@
##           develop    #2162       +/-   ##
============================================
- Coverage    88.75%   67.19%   -21.55%     
============================================
  Files          301      362       +61     
  Lines        25607    50705    +25098     
  Branches         0     5708     +5708     
============================================
+ Hits         22725    34069    +11344     
- Misses        2882    13052    +10170     
- Partials         0     3584     +3584

Flag	Coverage Δ
hipSPARSE	`?`
rocBLAS	`67.19% <ø> (?)`

Flags with carried forward coverage won't be shown. Click here to find out more.
see 663 files with indirect coverage changes

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

…efanuk/rocm-libraries into bug/tensile-build-with-no-enumerate2

shared/tensile/Tensile/LibraryIO.py

.github/workflows/therock-ci-linux.yml

.github/workflows/therock-ci-windows.yml

bstefanuk · 2025-11-06T18:40:44Z

Testing failure assessment:

Failed tensile precheckin jobs for gfx908 a gfx942 are also present on develop build 50

Using gardener override to merge.

[rocblas][tensile] Use kernel ISA during build with enforcement (#2162) ## Motivation Performance regressions are found when adding --no-enumerate to the TensileCreateLibrary build. This PR re-implements the kernel ISA reliance from #2094 without needing to change logic files. ## Technical Details - Rely only on the kernel's ISA during the build phase. - Add additional ISA enforcement given architecture details extracted from logic files. ## Test Plan - Local performance testing for specific sizes - Comprehensive performance testing through gemmaiperf - Standard CI testing ## Test Result - See CI results in this PR for standard pipeline checks. - Performance: tested on 6665 sizes using `rocblas-bench` on gfx950 (results below) - Performance: select sizes were evaluated on gfx942 and confirmed no performance change beyond +/-1% `Single precision NN` Stat | Result -- | -- Average (% speed up) | 0.50 Median (% speed up) | 0.01 Count Faster | 3482 Count Slower | 3161 `Single precision TN` Stat | Result -- | -- Average (% speed up) | 4.17 Median (%speed up) | -0.02 Count Faster | 3042 Count Slower | 3579 `Complex double precision TN` Stat | Result -- | -- Average (% speed up) | 0.18 Median (% speed up) | 0.04 Count Faster | 4452 Count Slower | 2207 ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

## Motivation The patch [0008-Revert-remove-options-no-enumerate-966.patch](https://github.com/ROCm/TheRock/blob/77e4a8304c0544a7ee5779fcabcee265a00f38ba/patches/amd-mainline/rocm-libraries/0008-Revert-remove-options-no-enumerate-966.patch) can be removed now that `--no-enumerate` is no longer needed in tensile. (PR #2162). This PR allows the patch to be removed without breaking CI. ## Technical Details Use `rm -f` to allow the pipeline to continue even if the file is missing. ## Test Plan Low risk, standard CI testing is sufficient. ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

## Motivation Pick up rocWMMA changes for compatibility with TheRock build. ## Technical Details Deleted a patch that is no longer required due to: ROCm/rocm-libraries#2162  ## Test Plan  ## Test Result  ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

## Motivation The `gfx1103` logic files used an incorrect reference to `navi33`, which has been exposed due to the recent inclusion of new logic file consistency checks in #2162. ## Technical Details Update the schedule name in the rocblas logic files to properly map to `gfx1103` which is consistent with the architectureMap in Common.py ## Test Plan Local testing before and after this change shows the following outputs Command: ``` Tensile/bin/TensileCreateLibrary [...] --architecture=gfx1103 /path/to/Tensile/Logic/asm_full build_gfx1103 HIP ``` Result (before; develop) ``` ValueError: Architecture mismatch: gfx1103 does not match navi33. Review the library logic file ``` Result (after) ``` # Reading logic files: 32 thread(s), 144 tasks .............................. 100.0% (took 4.0 secs) # Generating kernels: 32 thread(s), 2690 tasks .............................. 100.0% (took 23.1 secs) # Compiling source kernels .............................. 100.0% (took 0.0 secs) ``` ## Submission Checklist - [x] Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

fix(tensile): use kernel ISA during build with enforcement

d3f2cb5

bstefanuk requested review from a team as code owners October 17, 2025 22:55

bstefanuk self-assigned this Oct 17, 2025

github-actions bot added shared: tensile project: rocblas labels Oct 17, 2025

assistant-librarian bot added the organization: ROCm label Oct 17, 2025

yoichiyoshida approved these changes Oct 20, 2025

View reviewed changes

bstefanuk mentioned this pull request Oct 22, 2025

[tensile] Rely only on kernel ISA with no enumerate #2094

Closed

1 task

bstefanuk added 2 commits October 29, 2025 22:32

fix: use kernel ISA not CurrentISA

6d3fbef

Merge branch 'bug/tensile-build-with-no-enumerate2' of github.com:bst…

201044e

…efanuk/rocm-libraries into bug/tensile-build-with-no-enumerate2

bstefanuk commented Oct 29, 2025

View reviewed changes

shared/tensile/Tensile/LibraryIO.py Show resolved Hide resolved

bstefanuk commented Oct 29, 2025

View reviewed changes

shared/tensile/Tensile/LibraryIO.py Outdated Show resolved Hide resolved

bstefanuk added 3 commits October 29, 2025 17:01

Apply suggestions from code review

b76b671

fix: add and enforce ISA to custom kernels

3347af5

ci: ignore tensile --no-enumerate patch

4c93921

bstefanuk requested a review from a team as a code owner October 30, 2025 17:13

github-actions bot added the github actions label Oct 30, 2025

bstefanuk added 2 commits October 30, 2025 17:18

ci: ignore tensile --no-enumerate patch 2

ce830fa

fix: benchmarking needs to use CurrentISA when creating solutions

1c6df22

bstefanuk added the ci:performance label Oct 30, 2025

fix: rename strixpoint to strixhalo

26297bf

geomin12 reviewed Oct 31, 2025

View reviewed changes

.github/workflows/therock-ci-linux.yml Outdated Show resolved Hide resolved

.github/workflows/therock-ci-windows.yml Outdated Show resolved Hide resolved

bstefanuk removed the ci:performance label Nov 3, 2025

ci: use simple remove instead of loop

ebb8f13

amcamd approved these changes Nov 5, 2025

View reviewed changes

bstefanuk merged commit b75ec66 into ROCm:develop Nov 6, 2025
61 of 65 checks passed

bstefanuk deleted the bug/tensile-build-with-no-enumerate2 branch November 6, 2025 18:42

TorreZuk mentioned this pull request Nov 6, 2025

Revert "remove options --no-enumerate (#966)" #1558

Closed

bsyrowik mentioned this pull request Nov 6, 2025

Bump rocm-libraries 20251106 ROCm/TheRock#2044

Merged

1 task

This was referenced Nov 7, 2025

[ci] Let rm pass for missing patch #2518

Merged

[rocblas] Fix schedule name for gfx1103 logics #2539

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[rocblas][tensile] Use kernel ISA during build with enforcement #2162

[rocblas][tensile] Use kernel ISA during build with enforcement #2162

Uh oh!

bstefanuk commented Oct 17, 2025 •

edited

Loading

Uh oh!

codecov-commenter commented Oct 18, 2025 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bstefanuk commented Nov 6, 2025 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

[rocblas][tensile] Use kernel ISA during build with enforcement #2162

[rocblas][tensile] Use kernel ISA during build with enforcement #2162

Uh oh!

Conversation

bstefanuk commented Oct 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

Uh oh!

codecov-commenter commented Oct 18, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bstefanuk commented Nov 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

bstefanuk commented Oct 17, 2025 •

edited

Loading

codecov-commenter commented Oct 18, 2025 •

edited

Loading

bstefanuk commented Nov 6, 2025 •

edited

Loading