[GPU] Add pattern to fuse tensor.collapse_shape into forall producer #19295
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
|
PR is mostly ready, but I need to add more lit tests and improve the docs. EDIT: Ready now |
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
from
4eb7167 to
04c4f2b
Compare
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
from
04c4f2b to
5639b90
Compare
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
from
5639b90 to
688e8cb
Compare
pashu123
reviewed
Nov 28, 2024
| @@ -228,4 +228,38 @@ def FuseForallOp : Op<Transform_Dialect, "iree.fuse_forall", | |||
| let cppNamespace = "mlir::iree_compiler::IREE::transform_dialect"; | |||
| } | |||
|
|
|||
| def FuseCollapseShapeWithForallOp : Op<Transform_Dialect, "iree.fuse_collapse_shape_with_forall", | |||
There was a problem hiding this comment.
def FuseCollapseShapeIntoForallOp
qedawkins
requested changes
Dec 2, 2024
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/Transforms.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/Transforms.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/Transforms.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/TransformExtensions/IREEGPUExtensionsOps.td
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/Transforms.cpp
Outdated
Show resolved
Hide resolved
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
from
688e8cb to
102bc5a
Compare
|
This PR now depends on #19231. The tile_and_fuse pipeline test is failing because the pattern from that PR is needed here. |
qedawkins
reviewed
Dec 9, 2024
I am going to send another separate PR that this should rebase on (to fix the test failure), but the code in here shouldn't change, so it can be reviewed now. |
qedawkins
approved these changes
Dec 10, 2024
compiler/src/iree/compiler/Codegen/Dialect/GPU/Transforms/Transforms.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/LLVMGPU/test/ROCDL/pipeline_tile_and_fuse.mlir
Show resolved
Hide resolved
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
from
82d6046 to
2ba3856
Compare
|
Rebasing and preparing to land this PR now after finishing initial data tiling op codegen prototype. I'll wait a bit in case there are comments. |
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
from
e9ab180 to
6425acc
Compare
|
Based on #19730 now Edit: Rebased now |
Max191
added a commit
that referenced
this pull request
Jan 21, 2025
This PR moves the `SwapExpandShapeWithSlicePattern` to Codegen/Common/Transforms, and adds the pattern to the FuseAndHoistParallelLoops pass. This pattern is generally useful for tiling fusion, because it exposes more producer fusion opportunities when there are reshapes in the IR, but more specifically, it is useful in combination with the pattern introduced in #19295. That pattern creates an expanded parallel_insert_slice, and an expand_shape on the corresponding init block arg in the forall loop body. This makes the slice on the init argument lower dimensional than the parallel_insert_slice at the end. It is better for bufferization if these slices are the same, and this pattern makes that happen by bubbling the slice of the init arg up through the expand_shape, increasing the dimensionality to match the parallel_insert_slice. --------- Signed-off-by: Max Dawkins <[email protected]>
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
3 times, most recently
from
13fc8f9 to
da68bc2
Compare
Signed-off-by: Max Dawkins <[email protected]>
Max191
force-pushed
the
collapse-shape-forall-fusion
branch
from
da68bc2 to
8d74baa
Compare
Max191
added a commit
that referenced
this pull request
Jan 28, 2025
…19296) This PR adds a pattern to fuse a consumer tensor.extract_slice into a producer scf.forall op. The transform is added to FuseAndHoistParallelLoops, where it helps to fuse tensor.unpack ops with extract_slice semantics into producer loops. This is needed when targeting MFMA intrinsics for unaligned shapes, and also in generating code for unset encoding ops on GPU. This is a follow up to #19295, which has the complementing pattern for collapse_shape. The PR also adds a transform op to keep the long lit tests separate from the FuseAndHoistParallelLoop tests. --------- Signed-off-by: Max Dawkins <[email protected]> Signed-off-by: Max Dawkins <[email protected]> Co-authored-by: Max Dawkins <[email protected]>
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 PR adds a pattern to fuse a consumer tensor.collapse_shape into a producer scf.forall op. The transform is added to FuseAndHoistParallelLoops, where it helps to fuse tensor.unpack ops with extract_slice semantics into producer loops. This is needed when targeting MFMA intrinsics for unaligned shapes, and also in generating code for unset encoding ops on GPU.
The PR also adds a transform op to keep the long lit tests separate from the FuseAndHoistParallelLoop tests.