-
Notifications
You must be signed in to change notification settings - Fork 63
[MatmulLoopPipeline]: Prefetch 2D loads #4051
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.
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
Conversation
Signed-off-by: Tiotto, Ettore <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Pull Request Overview
This PR adds a safeguard to ensure that only 2D tensor load operations are prefetched in the MatmulLoopPipeline, preventing generation of invalid prefetch operations that could lead to assertions or incorrect code generation.
- Updated comments to clarify caching behavior.
- Added a check to skip prefetching for non-2D tensor loads.
third_party/intel/lib/TritonIntelGPUTransforms/Pipeliner/MatmulLoopPipeline.cpp
Show resolved
Hide resolved
third_party/intel/lib/TritonIntelGPUTransforms/Pipeliner/MatmulLoopPipeline.cpp
Show resolved
Hide resolved
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If we are changing this behavior, we should have a test - is this in response to a particular bug or just something we're worried could occur?
Signed-off-by: Tiotto, Ettore <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks! Looks good.
I have added a new lit test. I discovered the problem while working on another PR (#3634) which was failing one of the CI tests.. |
Add a check to prefetch only 2D tensor loads loads. This avoid potential generation of invalid prefetch operations which would cause assertions in subsequent passes or lead to incorrect code generation.