Skip to content

Enable DirectToLds for MXSA/B and re-enable LdsPad for MXFP4 + DirectToLds#4683

Merged
nakajee merged 3 commits into
gfx950_mx_rebasefrom
users/nakajee/gfx950mxfp4_dtl_fix_2
Feb 21, 2026
Merged

Enable DirectToLds for MXSA/B and re-enable LdsPad for MXFP4 + DirectToLds#4683
nakajee merged 3 commits into
gfx950_mx_rebasefrom
users/nakajee/gfx950mxfp4_dtl_fix_2

Conversation

@nakajee
Copy link
Copy Markdown
Contributor

@nakajee nakajee commented Feb 19, 2026 •
edited
Loading

Motivation

Improve MXFP4 performance with DirectoLds

Technical Details

  • Enable DirectToLds for MXSA/B
  • Re-enable LdsPad for MXFP4 + DirectToLds
  • Added test cases to mx32f4_tn.yaml

Test Plan

Local test

Test Result

Local test passed

Submission Checklist

@nakajee nakajee requested review from b-shi and msujon-AMD February 19, 2026 06:36
@nakajee nakajee requested a review from a team as a code owner February 19, 2026 06:36
@nakajee
Copy link
Copy Markdown
Contributor Author

nakajee commented Feb 20, 2026

Added reject conditions for DirectToLdsMXSA/B.

@nakajee nakajee mentioned this pull request Feb 20, 2026
Merged
1 task
msujon-AMD
msujon-AMD reviewed Feb 20, 2026
View reviewed changes
Copy link
Copy Markdown
Collaborator

@msujon-AMD msujon-AMD left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

do we want to have any test case in yaml with DTL for mxf4?
It looks good to me otherwise.

@nakajee
Copy link
Copy Markdown
Contributor Author

nakajee commented Feb 20, 2026

do we want to have any test case in yaml with DTL for mxf4? It looks good to me otherwise.

I can add a test case for this if needed.
Do we have any mxfp4 test cases?

@msujon-AMD
Copy link
Copy Markdown
Collaborator

msujon-AMD commented Feb 20, 2026

do we want to have any test case in yaml with DTL for mxf4? It looks good to me otherwise.

I can add a test case for this if needed. Do we have any mxfp4 test cases?

this one? Tests/common/gemm/gfx950/mx32f4_tn.yaml

@nakajee
Copy link
Copy Markdown
Contributor Author

nakajee commented Feb 20, 2026

do we want to have any test case in yaml with DTL for mxf4? It looks good to me otherwise.

I can add a test case for this if needed. Do we have any mxfp4 test cases?

this one? Tests/common/gemm/gfx950/mx32f4_tn.yaml

I added some DTL test cases in this file.

NineKa added a commit that referenced this pull request Feb 20, 2026
@NineKa
port changes for calcualting LDS padding for mx datatypes from #4683
c946745
@nakajee nakajee merged commit e91ecf3 into gfx950_mx_rebase Feb 21, 2026
13 of 22 checks passed
@nakajee nakajee deleted the users/nakajee/gfx950mxfp4_dtl_fix_2 branch February 21, 2026 03:03
@talumbau talumbau mentioned this pull request Mar 11, 2026
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@msujon-AMD msujon-AMD msujon-AMD approved these changes

@b-shi b-shi Awaiting requested review from b-shi

Assignees

No one assigned

Labels

organization: ROCm project: hipblaslt project: hipsparselt

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@nakajee @msujon-AMD @NineKa