[Codegen][GPU] Add support for all other intrinsics to TileAndFuse #18179
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
qedawkins
requested review from
MaheshRavishankar,
kuhar,
Groverkss and
antiagainst
as code owners
qedawkins
force-pushed
the
expand_and_add_all_intrinsics
branch
from
8edf5b1 to
2166c03
Compare
Max191
approved these changes
Aug 12, 2024
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/Transforms.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/test/concretize_mma_shapes.mlir
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/test/distribute_mma_to_lanes.mlir
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/LLVMGPU/test/ROCDL/pipeline_tile_and_fuse.mlir
Show resolved
Hide resolved
|
@hanhanW was this the intended PR for this comment? |
no... let me delete it. I was confused where my comment is in llvm/llvm-project#102952 (comment).. Sorry about that. |
This adds the ConcretizeMmaShapes pass to the LLVMGPUTileAndFuse pipeline to add support for other intrinsic types, in particular MFMA and WMMA variants that require reshaping of the accumulator to match requirements of the layout. This also reworks the reshaping code to use SingleSubgroupLayout instead of VectorExt::PerDimLayoutAttr to drop an unneeded dialect dependency and also simplify the IR for cases where reshaping is not needed. In particular, when there is a unit `outer` dimension in a layout, no additional reshaping is needed so we can omit the reshapes in such cases. There is an option in the future to still do such reshaping so as to pre-swizzle the data needed for the MMA during the store to shared memory, but the details for how best to implement that are left as TODO.
qedawkins
force-pushed
the
expand_and_add_all_intrinsics
branch
from
2166c03 to
0a6a163
Compare
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 adds the ConcretizeMmaShapes pass to the LLVMGPUTileAndFuse pipeline to add support for other intrinsic types, in particular MFMA and WMMA variants that require reshaping of the accumulator to match requirements of the layout.
This also reworks the reshaping code to use SingleSubgroupLayout instead of VectorExt::PerDimLayoutAttr to drop an unneeded dialect dependency and also simplify the IR for cases where reshaping is not needed. In particular, when there is a unit
outerdimension in a layout, no additional reshaping is needed so we can omit the reshapes in such cases. There is an option in the future to still do such reshaping so as to pre-swizzle the data needed for the MMA during the store to shared memory, but the details for how best to implement that are left as TODO.