Skip to content

[XLA:GPU] Use DeviceDescription instead of hard-coding warp size as 32 #2938

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

Jump to bottom
Open
wants to merge 5 commits into
base: r2.18-rocm-enhanced
Choose a base branch
from
Open

[XLA:GPU] Use DeviceDescription instead of hard-coding warp size as 32 #2938

wants to merge 5 commits into from
+1,070 −536

Conversation

amd-jianli12
Copy link
Collaborator

We should query the hardware to discover its warp size.

PiperOrigin-RevId: 700787004

@majnemer @amd-jianli12
Use DeviceDescription instead of hard-coding warp size as 32
5e84717 
tensorflow/tf-build-actions@600513b [ROCm] Fix flaky gpu compiler test when building with rocm
tensorflow/tf-build-actions@a35cf48 [XLA:GPU] Use DeviceDescription instead of hard-coding warp size as 32
xla@e849446 [ROCm] Pass correct warp size to Triton pipeline
xla@3e7b0fe cherry-picked warp size passing to triton calls, and globally enabled warpsize=64
xla@750ad89 Fixes.
@amd-jianli12 amd-jianli12 force-pushed the r2.18-rocm-enhanced-warpsize branch from c76562c to 5e84717 Compare April 23, 2025 04:26
@amd-jianli12 amd-jianli12 requested a review from i-chaochen May 16, 2025 04:49
@zoranjovanovic-ns
Copy link

PR is very large; it changes 90 files.
There are some changes that are only related to code formatting.
Is it possible to remove formatting changes from this PR?
If it is really necessary to change 90 files, is it possible to split PR to several smaller PRs?

@amd-jianli12
Copy link
Collaborator Author

amd-jianli12 commented May 20, 2025 •
edited
Loading

PR is very large; it changes 90 files. There are some changes that are only related to code formatting. Is it possible to remove formatting changes from this PR? If it is really necessary to change 90 files, is it possible to split PR to several smaller PRs?

This is the original commit tensorflow@a35cf48 which has 75 files modified.

Yeah, definitely this is a quite large one for us already. And since the latest XLA is quite different from the one in tensorflow r2.18, the number of files modified increased during backporting.

Maybe I can split pure original commit (with conflicts) and modifications made during backporting (to solve conflicts and make it compile), but this will still leave us a commit contains at least 75 files modified plus a relatively smaller patch. Do you think this would help this case?

@zoranjovanovic-ns
Copy link

Do you think this would help this case?

Probably not too much, will try to review it as it is.

@amd-jianli12
Copy link
Collaborator Author

Do you think this would help this case?

Probably not too much, will try to review it as it is.

That would be tough work, sorry

@zoranjovanovic-ns
Copy link

Do you think this would help this case?

Probably not too much, will try to review it as it is.

That would be tough work, sorry

Is it possible to get that second commit (fixing errors) somewhere?
If it is not difficult to get it (otherwise I will continue like this).

@amd-jianli12
Copy link
Collaborator Author

amd-jianli12 commented May 20, 2025 •
edited
Loading

Do you think this would help this case?

Probably not too much, will try to review it as it is.

That would be tough work, sorry

Is it possible to get that second commit (fixing errors) somewhere? If it is not difficult to get it (otherwise I will continue like this).

Just looked through local branches but had no fortune. But no worries, I believe it won't take too long to do the rework. Let me do it, otherwise it might not be well organized for you to review.

@amd-jianli12
Copy link
Collaborator Author

Do you think this would help this case?

Probably not too much, will try to review it as it is.

That would be tough work, sorry

Is it possible to get that second commit (fixing errors) somewhere? If it is not difficult to get it (otherwise I will continue like this).

Just looked through local branches but had no fortune. But no worries, I believe it won't take too long to do the rework. Let me do it, otherwise it might not be well organized for you to review.

Done. Please help review under this reorganized PR: #2962

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

@i-chaochen i-chaochen Awaiting requested review from i-chaochen

@zoranjovanovic-ns zoranjovanovic-ns Awaiting requested review from zoranjovanovic-ns

@draganmladjenovic draganmladjenovic Awaiting requested review from draganmladjenovic

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@amd-jianli12 @zoranjovanovic-ns @majnemer