[AutoBump] Merge with fixes of f03a5762 (Dec 12) (136) #526
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
…ps and conversion patterns. (llvm#3759) This PR refactors TorchToTosa to separate the construction of legal/illegal ops and conversion patterns in their own functions: 1. populateTorchToTosaConversionLegalOps -- populate any ops that are legal after the conversion pass 2. populateTorchToTosaConversionIllegalOps -- populate any ops that are illegal after the conversion pass 3. populateTorchToTosaConversionPatterns -- populate the ops conversion patterns Currently the (il)legality of the ops that are (il)legal after the conversion pass runs is embedded within the conversion pattern. Our end goal is to write a new pass pipeline that converts `torch` ops to a mix of `tosa`, `linalg`, `tensor`, etc dialect ops. The reason we want to also emit `tosa` ops (instead of using the existing `TorchToLinalg` to emit `linalg`+`tensor`+...) is because some operations like `conv2d` encodes the padding behavior in the op in `tosa` unlike the `linalg` version -- this helps in lowering the `tosa.conv2d` to a custom implementation that does padding on the fly. To implement this new pipeline we need to be able to separate out the illegal `tosa` ops from the conversion pattern itself. Otherwise we will hit an issue for ops like `AtenMaxDimOp` which can be lowered to both `tosa` and `linalg + others` dialects. Not all `AtenMaxDimOp` can be lowered successfully to `tosa` as the implementation uses `tosa.reshape` which cannot handle multiple dynamic dimensions but the `TorchToLinalg` lowering can handle it. In the current behavior the pipeline will stop as soon as the existing `TorchToTosa` conversion runs as `AtenMaxDimOp` will be marked as an illegal op. Essentially we want to be able to control what the legality of the ops should be independent of the conversion pattern. This is also inline with the conversion patterns in the llvm-mlir repo such as https://github.com/llvm/llvm-project/blob/000e790be35b77a01872851646d54432a203542c/mlir/lib/Conversion/SCFToControlFlow/SCFToControlFlow.cpp#L718 "THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY."
…m#3916) The onnx output tensor has a shape of ((n, z)), where (n) is the number of dimensions in the input tensor and (z) is the number of non-zero elements2. This is different from PyTorch's default behavior, where the dimensions are reversed.
This commit adds the support for 1-d group convolution by transforming it into a 2-d group convolution which is already supported. This commit also refactors the unsqueeze and squeeze tensor utility. --------- Signed-off-by: Vivek Khandelwal <[email protected]>
…llvm#3918) We incorrectly relied on the fact that StableHLO registers the sparse tensor dialect, but when building for e.g. just LinAlg, the dependency was missing. This fixes this shortcoming. FIXES: llvm#3816
[AutoBump] Merge with 2c72a82 (Dec 13) (137)
Base automatically changed from
bump_to_5077090a
to
feature/backport_ea1_ops
February 11, 2025 08:38
cferry-AMD
approved these changes
Feb 11, 2025
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.
No description provided.