Open
Conversation
Collaborator
Author
|
Here is the perf comparison on navi48 against mlir: |
Contributor
There was a problem hiding this comment.
Pull request overview
This PR extends the rewrite_reduce optimization pass to rewrite “skinny” dot ops (small M dimension) into an equivalent mul + reduce_sum formulation, while detecting and exempting attention-related dots (QK^T and softmaxV) to preserve attention fusion opportunities.
Changes:
- Add attention-pattern detection (via decomposed softmax matching) to identify dots that should not be rewritten.
- Rewrite
dotops with M (rows) ≤ 2 intounsqueeze + transpose + mul + reduce_sum + squeeze. - Expand
test/rewrite_reduce.cppwith new coverage for skinny-dot rewrites and attention/non-attention cases.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
src/rewrite_reduce.cpp |
Adds attention dot collection + a new find_dot matcher that rewrites small-M dots into mul/reduce, skipping attention dots. |
test/rewrite_reduce.cpp |
Adds/updates unit tests to validate skinny-dot rewrites and to ensure attention patterns are not rewritten. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+136
to
+142
| // If the b matrix is const foldable then make sure its a transposed layout unless its | ||
| // broadcasting | ||
| if(b_mat->can_eval() and not b_shape.transposed()) | ||
| { | ||
| b_mat = | ||
| m.insert_instruction(ins, make_op("layout", {{"permutation", permutation}}), b_mat); | ||
| } |
There was a problem hiding this comment.
The layout insertion for const-foldable b_mat is guarded only by !b_shape.transposed(), but the preceding comment says it should be skipped when b_mat is broadcasting. If b_shape.broadcasted() is true (e.g., a multibroadcasted literal), inserting layout can materialize the broadcasted tensor and significantly increase compile-time/memory usage. Consider adding and not b_shape.broadcasted() (or an equivalent check) to the condition so broadcasted constants keep their cheap broadcasted representation.
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## develop #4811 +/- ##
===========================================
+ Coverage 92.49% 92.50% +0.01%
===========================================
Files 583 583
Lines 29561 29621 +60
===========================================
+ Hits 27342 27399 +57
- Misses 2219 2222 +3
🚀 New features to boost your workflow:
|
Collaborator
Regressions detected 🔴 |
Collaborator
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
This algorithm is usually much faster than mfmas.
Technical Details
Since decoding attention also uses skinny gemms, it does skip the gemms used in attention.
Changelog Category
Add a
CHANGELOG.mdentry for any option other thanNot Applicable