Canonicalize triton::ReduceOp select(cmpf) patterns to min/max operations #361
Conversation
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
This patch adds a pattern rewrite that canonicalizes triton::ReduceOp bodies which use arith.select based on floating-point comparisons into equivalent arith min/max ops.
Author
|
@microsoft-github-policy-service agree company="Qualcomm" |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
This PR introduces a pattern rewrite that simplifies
triton::ReduceOpbodies containingarith.selectoperations based on floating-point comparisons. The transformation replaces select-based logic with equivalent min/max operations, improving IR canonicalization and enabling further optimization.Supported Transformations:
select(cmpf ogt a, b), a, b→arith.maxf(a, b)select(cmpf olt a, b), a, b→arith.minf(a, b)select((cmpf ogt a, b) || cmpf une a, a), a, b→arith.maximumf(a, b)select((cmpf olt a, b) || cmpf une a, a), a, b→arith.minimumf(a, b)Applying this transformation before the
ReduceOpconverter runs allows the natural lowering oftt.reducetolinalg.reduce; the motivational use case was from a torch-inductor spit out Triton kernel which consisted of this complex pattern in the reduction body and did not allow theReduceOpto get lowered to thelinalgdialect.