Skip to content

[hipblaslt] tensilelite client validation threshold fix for tf32x3 and tf32x1#4890

Merged
carsonbrownlee merged 4 commits intodevelopfrom
users/cbrownle/tf32_accuracyfix
Feb 27, 2026
Merged

[hipblaslt] tensilelite client validation threshold fix for tf32x3 and tf32x1#4890
carsonbrownlee merged 4 commits intodevelopfrom
users/cbrownle/tf32_accuracyfix

Conversation

@carsonbrownlee
Copy link
Contributor

@carsonbrownlee carsonbrownlee commented Feb 25, 2026

Motivation

enable trig_init comparisons for tensilelite for tf32x3 and tf32x1.

Technical Details

Modify hardcoded thresholds in tensilelite validation to vary based on sqrt(k). Currently only implemented for float comparison for tf32x3 and tf32x1.

Test Plan

Tested with tensilelite yamls, modifying Tensile/Tests/common/gemm/gfx950/xfp32.yaml and ss_bss.yaml to have datainittypeA/B: 12/13. Only xfp32.yaml (tf32x3) test modified to use trig init as it makes more sense to test the lower bits by default. ss_bss.yaml not changed and needs to be modified to test.

Test Result

PASS

@carsonbrownlee carsonbrownlee changed the title tensilelite client validation threshold fix for tf32x3 and tf32x1 [hipblaslt] tensilelite client validation threshold fix for tf32x3 and tf32x1 Feb 25, 2026
@math-ci-webhook
Copy link

perfci run on commit 49fe485

math-ci run

@math-ci-webhook
Copy link

perfci run on commit 4cad8a2

math-ci run

@codecov-commenter
Copy link

codecov-commenter commented Feb 25, 2026

Codecov Report

✅ All modified and coverable lines are covered by tests.

❌ Your project status has failed because the head coverage (76.83%) is below the target coverage (80.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files
@@           Coverage Diff            @@
##           develop    #4890   +/-   ##
========================================
  Coverage    65.96%   65.96%           
========================================
  Files         1720     1720           
  Lines       267465   267465           
  Branches     37091    37091           
========================================
  Hits        176427   176427           
  Misses       75473    75473           
  Partials     15565    15565           
Flag Coverage Δ *Carryforward flag
hipBLAS 90.67% <ø> (ø) Carriedforward from 1082959
hipBLASLt 43.55% <ø> (ø)
hipCUB 82.38% <ø> (ø) Carriedforward from 1082959
hipDNN 80.91% <ø> (ø) Carriedforward from 1082959
hipFFT 55.93% <ø> (ø) Carriedforward from 1082959
hipRAND 76.12% <ø> (ø) Carriedforward from 1082959
hipSOLVER 68.81% <ø> (ø) Carriedforward from 1082959
hipSPARSE 84.70% <ø> (ø) Carriedforward from 1082959
rocBLAS 47.97% <ø> (ø) Carriedforward from 1082959
rocFFT 52.91% <ø> (ø) Carriedforward from 1082959
rocRAND 57.06% <ø> (ø) Carriedforward from 1082959
rocSOLVER 76.83% <ø> (ø) Carriedforward from 1082959
rocSPARSE 71.53% <ø> (ø) Carriedforward from 1082959

*This pull request uses carry forward flags. Click here to find out more.

🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
  • 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

@carsonbrownlee carsonbrownlee force-pushed the users/cbrownle/tf32_accuracyfix branch from dade002 to 1eab8ec Compare February 26, 2026 06:07
@math-ci-webhook
Copy link

perfci run on commit 891ff74

math-ci run

@msujon-AMD
Copy link
Collaborator

BTW, we may not need K in threshold for RND_INT initialization since we may not have any precision loss in down-conversion. I am not sure if we consider that as a special case.

@carsonbrownlee carsonbrownlee merged commit 240a74b into develop Feb 27, 2026
39 checks passed
@carsonbrownlee carsonbrownlee deleted the users/cbrownle/tf32_accuracyfix branch February 27, 2026 18:34
assistant-librarian bot pushed a commit to ROCm/hipBLASLt that referenced this pull request Feb 27, 2026
[hipblaslt] tensilelite client validation threshold fix for
 tf32x3 and tf32x1 (#4890)

## Motivation

enable trig_init comparisons for tensilelite for tf32x3 and tf32x1.

## Technical Details

Modify hardcoded thresholds in tensilelite validation to vary based on
sqrt(k). Currently only implemented for float comparison for tf32x3 and
tf32x1.

## Test Plan

Tested with tensilelite yamls, modifying
Tensile/Tests/common/gemm/gfx950/xfp32.yaml and ss_bss.yaml to have
datainittypeA/B: 12/13. Only xfp32.yaml (tf32x3) test modified to use
trig init as it makes more sense to test the lower bits by default.
ss_bss.yaml not changed and needs to be modified to test.

## Test Result

PASS
kokolchin pushed a commit to kokolchin/rocm-libraries that referenced this pull request Mar 4, 2026
…d tf32x1 (ROCm#4890)

## Motivation

enable trig_init comparisons for tensilelite for tf32x3 and tf32x1.

## Technical Details

Modify hardcoded thresholds in tensilelite validation to vary based on
sqrt(k). Currently only implemented for float comparison for tf32x3 and
tf32x1.

## Test Plan

Tested with tensilelite yamls, modifying
Tensile/Tests/common/gemm/gfx950/xfp32.yaml and ss_bss.yaml to have
datainittypeA/B: 12/13. Only xfp32.yaml (tf32x3) test modified to use
trig init as it makes more sense to test the lower bits by default.
ss_bss.yaml not changed and needs to be modified to test.

## Test Result

PASS
NaveenElumalaiAMD pushed a commit that referenced this pull request Mar 6, 2026
…d tf32x1 (#4890)

## Motivation

enable trig_init comparisons for tensilelite for tf32x3 and tf32x1.

## Technical Details

Modify hardcoded thresholds in tensilelite validation to vary based on
sqrt(k). Currently only implemented for float comparison for tf32x3 and
tf32x1.

## Test Plan

Tested with tensilelite yamls, modifying
Tensile/Tests/common/gemm/gfx950/xfp32.yaml and ss_bss.yaml to have
datainittypeA/B: 12/13. Only xfp32.yaml (tf32x3) test modified to use
trig init as it makes more sense to test the lower bits by default.
ss_bss.yaml not changed and needs to be modified to test.

## Test Result

PASS
jovanau pushed a commit to jovanau/rocm-libraries that referenced this pull request Mar 19, 2026
…d tf32x1 (ROCm#4890)

## Motivation

enable trig_init comparisons for tensilelite for tf32x3 and tf32x1.

## Technical Details

Modify hardcoded thresholds in tensilelite validation to vary based on
sqrt(k). Currently only implemented for float comparison for tf32x3 and
tf32x1.

## Test Plan

Tested with tensilelite yamls, modifying
Tensile/Tests/common/gemm/gfx950/xfp32.yaml and ss_bss.yaml to have
datainittypeA/B: 12/13. Only xfp32.yaml (tf32x3) test modified to use
trig init as it makes more sense to test the lower bits by default.
ss_bss.yaml not changed and needs to be modified to test.

## Test Result

PASS
johannes-graner pushed a commit that referenced this pull request Mar 20, 2026
…d tf32x1 (#4890)

## Motivation

enable trig_init comparisons for tensilelite for tf32x3 and tf32x1.

## Technical Details

Modify hardcoded thresholds in tensilelite validation to vary based on
sqrt(k). Currently only implemented for float comparison for tf32x3 and
tf32x1.

## Test Plan

Tested with tensilelite yamls, modifying
Tensile/Tests/common/gemm/gfx950/xfp32.yaml and ss_bss.yaml to have
datainittypeA/B: 12/13. Only xfp32.yaml (tf32x3) test modified to use
trig init as it makes more sense to test the lower bits by default.
ss_bss.yaml not changed and needs to be modified to test.

## Test Result

PASS
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants