Skip to content

[FA performance] Improve the Q matrix load stategy #3966

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
mfrancepillois opened this issue Apr 18, 2025 · 1 comment · May be fixed by #3965
Open

[FA performance] Improve the Q matrix load stategy #3966

mfrancepillois opened this issue Apr 18, 2025 · 1 comment · May be fixed by #3965
Assignees
@mfrancepillois
Labels
codegen: attention performance
Milestone
4. [Performance] ...

Comments

@mfrancepillois
Copy link
Contributor

In the Cutlass implementation, the Q matrix is prefetched outside the loop but loaded inside the loop.
This improves FA performance by avoiding register spilling.
A similar strategy should be evaluated for FA in Triton.
@mfrancepillois mfrancepillois self-assigned this Apr 18, 2025
@mfrancepillois mfrancepillois linked a pull request Apr 18, 2025 that will close this issue
New pass Reduce variable liveness #3965
Open
@vlad-penkin vlad-penkin added this to the 4. [Performance] Core milestone Apr 28, 2025
@mfrancepillois
Copy link
Contributor Author

The performance assessment of the new pass is blocked by the performance regression on FA: #4239

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

@mfrancepillois mfrancepillois

Labels
codegen: attention performance
Projects
None yet
Milestone
4. [Performance] Core
Development

Successfully merging a pull request may close this issue.

New pass Reduce variable liveness intel/intel-xpu-backend-for-triton
2 participants
@vlad-penkin @mfrancepillois