[GPU] Use affine.delinearize_index for MMA tiles and vector distribution #19228
Conversation
This file contains 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
5328767 to
5a8fa83
Compare
829c3d5 to
aaf6cf8
Compare
5a8fa83 to
291f570
Compare
aaf6cf8 to
3ccf6a4
Compare
Base automatically changed from
users/krzysz00/gpu-distribute-with-linearize
to
main
November 26, 2024 20:38
ba7bc66 to
d577300
Compare
31e58d0 to
5592a62
Compare
|
Update: staring at tests showed that I should go implement the value bounds op interfaces on the affine.delinearize_index and affine.linearize_index because there were some single-iteration loops that weren't getting eliminated. |
8e5ebb0 to
de7e313
Compare
0277d07 to
3ff65f0
Compare
12377f8 to
d8e695c
Compare
qedawkins
reviewed
Jan 27, 2025
compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUAttrs.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/IR/IREEGPUAttrs.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/test/distribute_mma_to_lanes.mlir
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/VectorExt/IR/VectorExtAttrs.cpp
Show resolved
Hide resolved
| /// | ||
| /// As a special case, dimensions with stride 0 are treated as size-1 | ||
| /// dimensions that are placed at the end of the delinearization, from where | ||
| /// they will canonicalize to 0. |
qedawkins
approved these changes
Jan 27, 2025
This commit updates some by-hand delinearizations in MMA tile generation and vector distribution to use `affine.delinearize_index` instead. The main tricky thing here is that a lot of that MMA code would use `(id / stride) % size`, whereas delinearize's outputs all have the form `(id % stride) / nextStride`. In all the cases at issue, we could use a utility to convert arrays of sizes and strides to a permutation on a delinearization basis. In order to not break existing tests, the trivial-loop detector had to be manually instrumented to support `delinearize_index` (and I got `util.assume.int` while I was there). (I suspect there're a few other cases, and that, long-term, that detector should be using one of the bounds interfaces, but that's not this PR)# This is a combination of 7 commits.
d8e695c to
88ade5d
Compare
Co-authored-by: Quinn Dawkins <quinn.dawkins@gmail.com>
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.
This commit updates some by-hand delinearizations in MMA tile generation and vector distribution to use
affine.delinearize_indexinstead.The main tricky thing here is that a lot of that MMA code would use
(id / stride) % size, whereas delinearize's outputs all have the form(id % stride) / nextStride. In all the cases at issue, we could use a utility to convert arrays of sizes and strides to a permutation on a delinearization basis.In order to not break existing tests, the trivial-loop detector had to be manually instrumented to support
delinearize_index(and I gotutil.assume.intwhile I was there). (I suspect there're a few other cases, and that, long-term, that detector should be using one of the bounds interfaces, but that's not this PR)