Arm backend: Cleanup dim-order and permute handling#19278
Merged
AdrianLundell merged 1 commit intopytorch:mainfrom May 5, 2026
Merged
Arm backend: Cleanup dim-order and permute handling#19278AdrianLundell merged 1 commit intopytorch:mainfrom
AdrianLundell merged 1 commit intopytorch:mainfrom
Conversation
- Replace u55 permute dimension check with a u55-only pass decomposing large permutes. This pass checks for support by compiling targeted permutes using Vela to ensure alignment between Executorch and Vela. - Remove passes and testing not required anymore after dim-order update. - Remove all outdated mention of dim-order in the arm backend. Signed-off-by: Adrian Lundell <adrian.lundell@arm.com> Change-Id: I098db4539179cb223b5c76683720e68c1bbecb8f
AdrianLundell
added
partner: arm
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19278
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ You can merge normally! (3 Unrelated Failures)As of commit a606e58 with merge base 69989b7 ( BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Collaborator
Author
|
No buck2 changes required according to codex |
Contributor
There was a problem hiding this comment.
Pull request overview
This PR removes the old Arm backend dim-order/memory-format machinery and shifts permute handling toward explicit graph rewrites, including a new U55-specific large-permute decomposition pass. It mainly refactors TOSA lowering/serialization paths, operator support checks, and the associated Arm backend tests.
Changes:
- Replace dim-order-based TOSA shape/constant handling with direct shape normalization and remove the old transpose/memory-format infrastructure.
- Add
DecomposePermuteForU55Passand update U55 permute/view/select expectations and tests around the new behavior. - Delete legacy passes/operators/tests that were only needed for the previous dim-order approach.
Reviewed changes
Copilot reviewed 33 out of 33 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
backends/arm/tosa/utils.py |
Renames shape helper to normalize_symint. |
backends/arm/tosa/mapping.py |
Drops dim-order from extracted tensor metadata and TosaArg. |
backends/arm/tosa/dialect/ops/transpose.py |
Deletes fake TOSA transpose dialect op. |
backends/arm/tosa/dialect/__init__.py |
Removes transpose dialect registration. |
backends/arm/test/tester/arm_tester.py |
Updates tester helper for new extract_tensor_meta signature. |
backends/arm/test/passes/test_to_tosa_memory_format.py |
Deletes tests for removed memory-format pass. |
backends/arm/test/passes/test_decompose_int16_activation_conv_pass.py |
Deletes tests for removed int16 conv decomposition pass. |
backends/arm/test/ops/test_view.py |
Renames test data sets and removes old U55 non-delegation cases. |
backends/arm/test/ops/test_select.py |
Removes old U55 non-delegation coverage. |
backends/arm/test/ops/test_permute.py |
Expands U55 permute coverage for large-shape decomposition path. |
backends/arm/test/misc/test_transpose_counts.py |
Updates expected transpose count for grouped conv channels-last case. |
backends/arm/process_node.py |
Removes dim-order-aware tensor serialization and uses normalized shapes directly. |
backends/arm/operators/op_while.py |
Stops consulting output dim-order when creating dummy loop outputs. |
backends/arm/operators/op_tosa_transpose.py |
Deletes backend visitor for removed TOSA transpose op. |
backends/arm/operators/op_tosa_shapes.py |
Serializes shape constants without dim-order remapping. |
backends/arm/operators/op_sum.py |
Uses raw reduction axis directly. |
backends/arm/operators/op_permute.py |
Removes dim-order permutation remapping logic. |
backends/arm/operators/op_cat.py |
Uses raw concat dimension directly. |
backends/arm/operators/op_any.py |
Uses raw reduction axis directly. |
backends/arm/operators/op_amin.py |
Uses raw reduction axis directly. |
backends/arm/operators/op_amax.py |
Uses raw reduction axis directly. |
backends/arm/operators/__init__.py |
Unregisters removed transpose visitor. |
backends/arm/operator_support/tosa_supported_operators.py |
Removes old U55 transpose/view support checks from factory. |
backends/arm/operator_support/ethos_u55_support.py |
Deletes legacy U55 view/permute support-check implementations. |
backends/arm/operator_support/convolution_support.py |
Simplifies transpose-conv U55 shape handling. |
backends/arm/_passes/to_tosa_memory_format_pass.py |
Deletes old dim-order/memory-format pass. |
backends/arm/_passes/insert_data_layout_casts_pass.py |
Removes dependency on deleted backend transpose op. |
backends/arm/_passes/decompose_permute_for_u55_pass.py |
Adds new U55 large-permute decomposition/probing pass. |
backends/arm/_passes/decompose_int16_activation_conv_pass.py |
Deletes old int16 activation conv decomposition pass. |
backends/arm/_passes/arm_pass_utils.py |
Removes output dim-order helper. |
backends/arm/_passes/arm_pass_manager.py |
Inserts new U55 permute pass and reorders slice rewriting. |
backends/arm/_passes/annotate_output_dim_order_pass.py |
Deletes old output dim-order annotation pass. |
backends/arm/_passes/__init__.py |
Updates exported pass list for removed/added passes. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
153
to
+156
| dtype = map_dtype(val.dtype) | ||
| shape = tuple(val.size()) | ||
|
|
||
| dim_order = tuple(range(len(shape))) | ||
| return (dtype, shape, dim_order) | ||
| return (dtype, shape) |
Comment on lines
+336
to
+339
| # This is a quick check to avoid the overhead of the Vela compilation in 99% of cases. | ||
| if not self._violates_u55_worst_case_constraint(input_shape): | ||
| return super().call_operator(op, args, kwargs, meta) | ||
|
|
Comment on lines
+202
to
205
| assert isinstance(tensor, torch.Tensor), ( | ||
| f"Expected lifted tensor constant '{node.name}' to be a torch.Tensor, got " | ||
| f"{type(tensor).__name__}" | ||
| ) |
Comment on lines
119
to
+126
| """Extract dtype, shape, and dimension order from FX metadata. | ||
|
|
||
| Args: | ||
| meta (dict): FX node ``meta`` containing a ``val`` FakeTensor (or tuple). | ||
|
|
||
| Returns: | ||
| tuple[ts.DType, tuple[int, ...], tuple[int, ...]]: Tuple containing | ||
| tensor dtype, shape, and dimension order. | ||
| tensor dtype and shape. |
| permutation and dtype to check wheter it is supported. | ||
| """ | ||
|
|
||
| if dtype not in (torch.int8, torch.bool, torch.int16): |
zingo
approved these changes
May 4, 2026
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
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.
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell