[Dispatch Creation] Don't fuse if there are no common parallel loops #22819
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
Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
kuhar
approved these changes
Dec 3, 2025
IanWood1
commented
Dec 4, 2025
compiler/src/iree/compiler/DispatchCreation/test/dispatch_linalg_ext_fusion.mlir
Outdated
Show resolved
Hide resolved
Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
Contributor
Author
|
CLIP dispatch count increased because this now gets split into 2 dispatches https://gist.github.com/IanWood1/73971d57bd8243f491291bfa4d4b8e36 |
Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
MaheshRavishankar
approved these changes
Dec 4, 2025
Collaborator
MaheshRavishankar
left a comment
MaheshRavishankar
left a comment
There was a problem hiding this comment.
The increase in clip dispatch is fine. This shouldnt have fused to start with.
IanWood1
added a commit
that referenced
this pull request
Dec 8, 2025
Fixes golden dispatch counts that were not updated in #22819 because the PR wasn't run with `ci-extra: test_torch`. ci-extra: test_torch Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
keshavvinayak01
pushed a commit
that referenced
this pull request
Jan 27, 2026
…22819) This change prevents fusion when the root op does not share parallel loops with its producers or consumers. Because we require the iteration space of reduction ops to match exactly, this can only occur when the non-root is all parallel. And when the non-root is all parallel, there will only ever be no common parallel loops if the root has no parallel loops. `newMap` is the mapping from the root op's to the non-root op's parallel loops. If there aren't any dims in the output, then there is no mapping between the loops. This preserves the old behavior introduced by #12661. Related #22815 --------- Signed-off-by: Ian Wood <ianwood@u.northwestern.edu> Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
keshavvinayak01
pushed a commit
that referenced
this pull request
Jan 27, 2026
Fixes golden dispatch counts that were not updated in #22819 because the PR wasn't run with `ci-extra: test_torch`. ci-extra: test_torch Signed-off-by: Ian Wood <ianwood@u.northwestern.edu> Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha@gmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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 change prevents fusion when the root op does not share parallel loops with it's producers or consumers. Because we require the iteration space of reduction ops to match exactly, this can only occur when the non-root is all parallel. And when the non-root is all parallel, there will only ever be no common parallel loops if the root has no parallel loops.
newMapis the mapping from the root op's to the non-root op's parallel loops. If there aren't any dims in the output, then there is no mapping between the loops.This preserves the old behavior introduced by #12661.
Related #22815