[CPU] Tile reduction dimensions for non-root reduction ops. #21500
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
hanhanW
commented
Jul 25, 2025
Comment on lines
421
to
425
| // Tile all the reduction ops for target vector sizes. It is a nop for | ||
| // rootOp because it is already tiled with the same tile sizes. It | ||
| // ensures that all the dimensions are tiled in all the reduction ops. | ||
| funcPassManager.addPass( | ||
| createLLVMCPUTilePass(static_cast<IREE::CPU::TilingLevel>(i))); |
Contributor
Author
There was a problem hiding this comment.
It is not NOP if the dimension size is not divisible by tile size, because it becomes dynamic shapes and the for loops can't be folded away.
I think we need to provide an option to LLVMCPUTilePass that allows only tiling on non-root ops.
Contributor
Author
hanhanW
force-pushed
the
issue-21359
branch
2 times, most recently
from
b37133e to
cd80291
Compare
In softmax dispatch, there are two reduction ops. Only the root op is tiled for reduction dimensions when we switched to LLVMCPUTileRootAndFuseInputOperandsPass. It results in large vector sizes in the other reduction op when `util.assume.hint` ops are present. We did not hit the issue in e2e tests because AnnotateDispatchAssumptions pass behaves differently. The value range is [0, 0] if the input is from `flow.tensor.dynamic_constant`. Fixes iree-org#21359 Signed-off-by: hanhanW <[email protected]>
Signed-off-by: hanhanW <[email protected]>
…generic` Signed-off-by: hanhanW <[email protected]>
hhkit
pushed a commit
to opencompl/iree
that referenced
this pull request
Aug 8, 2025
…#21500) The revision adds an option to skip root op in LLVMCPUTile pass, and uses it in multi level tiling pipeline. In softmax dispatch, there are two reduction ops. Only the root op is tiled for reduction dimensions when we switched to LLVMCPUTileRootAndFuseInputOperandsPass. It results in large vector sizes in the other reduction op when `util.assume.hint` ops are present. We did not hit the issue in e2e tests because AnnotateDispatchAssumptions pass behaves differently. The value range is [0, 0] if the input is from `flow.tensor.dynamic_constant`. Fixes iree-org#21359 --------- Signed-off-by: hanhanW <[email protected]>
keshavvinayak01
pushed a commit
to keshavvinayak01/iree
that referenced
this pull request
Sep 4, 2025
…#21500) The revision adds an option to skip root op in LLVMCPUTile pass, and uses it in multi level tiling pipeline. In softmax dispatch, there are two reduction ops. Only the root op is tiled for reduction dimensions when we switched to LLVMCPUTileRootAndFuseInputOperandsPass. It results in large vector sizes in the other reduction op when `util.assume.hint` ops are present. We did not hit the issue in e2e tests because AnnotateDispatchAssumptions pass behaves differently. The value range is [0, 0] if the input is from `flow.tensor.dynamic_constant`. Fixes iree-org#21359 --------- Signed-off-by: hanhanW <[email protected]> Signed-off-by: keshavvinayak01 <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
The revision adds an option to skip root op in LLVMCPUTile pass, and uses it in multi level tiling pipeline.
In softmax dispatch, there are two reduction ops. Only the root op is tiled for reduction dimensions when we switched to LLVMCPUTileRootAndFuseInputOperandsPass. It results in large vector sizes in the other reduction op when
util.assume.hintops are present.We did not hit the issue in e2e tests because AnnotateDispatchAssumptions pass behaves differently. The value range is [0, 0] if the input is from
flow.tensor.dynamic_constant.Fixes #21359