Add MaxPool1D decomposition pass support (#17022) - #21446
Open
Ninja91 wants to merge 1 commit into
Open
Conversation
Summary: Implement DecomposeMaxPool1dPass to enable MaxPool1D support on ARM backend by decomposing max_pool1d into unsqueeze_copy → max_pool2d → squeeze_copy. ## Implementation Strategy ### Decomposition Approach (Optimal for TOSA/Vela) The pass decomposes max_pool1d into max_pool2d via unsqueeze_copy/squeeze_copy operations: 1. unsqueeze_copy(dim=2): (N, C, L) → (N, C, 1, L) - add height dimension 2. max_pool2d: with adapted params [k]→[1,k], [s]→[1,s], [p]→[0,p], [d]→[1,d] 3. squeeze_copy(dims=[2]): (N, C, 1, L_out) → (N, C, L_out) - remove height dimension ### Why This Approach is Optimal 1. **unsqueeze_copy and squeeze_copy map to TOSA RESHAPE** which is zero-cost in Vela: - Classified as memory_only_ops (Reshape, Squeeze, ExpandDims, Identity) - Bypassed entirely when conditions met (NPU-produced, single consumer) - Tensor equivalence enables memory aliasing (same address) 2. **TFA Pipeline Placement (before quantization)**: - unsqueeze_copy.default is in _one_to_one_shared_input_qspec - squeeze_copy.dims is added to _one_to_one_shared_input_qspec - max_pool2d is in _one_to_one_shared_input_or_input_act_qspec - All get proper SharedQuantizationSpec from the annotator automatically 3. **Quantization Handling**: - Clear qparams on intermediate unsqueeze_copy and squeeze_copy ops (let annotator fill them) - Preserve original meta on max_pool2d for proper tracing - MAX_POOL2D doesn't need zero-point handling (unlike AVG_POOL2D) ### TOSA/Vela Constraints Validated - U55: Stride ≤3 ✓, Kernel ≤256x256 ✓ - U85: Extended stride support via accumulator save/restore - Dilation: Handled by separate DecomposeMaxPool2dPass if needed Reviewed By: 3l1 Differential Revision: D91760459
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21446
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 2 Unrelated FailuresAs of commit 501e2ea with merge base 32f56be ( NEW FAILURES - The following jobs have failed:
FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
This PR needs a
|
Contributor
There was a problem hiding this comment.
Pull request overview
This PR adds Arm-backend support for aten.max_pool1d by introducing a decomposition pass that rewrites MaxPool1D into an equivalent MaxPool2D sequence (unsqueeze_copy → max_pool2d → squeeze_copy) and wires it into the Arm pass pipelines and quantization annotator so the resulting intermediate ops are properly annotated.
Changes:
- Add
DecomposeMaxPool1dPassto rewriteaten.max_pool1dinto 2D pooling plus view-like reshapes. - Register the new pass in the Arm pass manager and
_passespackage exports. - Update Arm quantization annotation to treat
aten.squeeze_copy.dimsas a shared-input qspec op; enable the previously xfailed TOSA INT MaxPool1D test.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| backends/arm/test/ops/test_max_pool1d.py | Un-xfails the TOSA INT MaxPool1D decomposition test. |
| backends/arm/quantizer/quantization_annotator.py | Adds squeeze_copy.dims to shared-input qspec ops to support the new decomposition intermediates. |
| backends/arm/_passes/decompose_max_pool1d_pass.py | Introduces the MaxPool1D→MaxPool2D decomposition pass. |
| backends/arm/_passes/arm_pass_manager.py | Inserts the new decomposition pass into relevant Arm pipelines. |
| backends/arm/_passes/init.py | Exports DecomposeMaxPool1dPass from the passes package. |
Comments suppressed due to low confidence (1)
backends/arm/_passes/decompose_max_pool1d_pass.py:54
- This decomposition assumes the max_pool1d input is 3D (N, C, L). However, max_pool1d can also be called with 2D inputs (C, L); in that case
unsqueeze_copy(dim=2)produces a rank-3 tensor andmax_pool2dwill fail due to incorrect rank. Consider guarding on input rank (or explicitly handling 2D) before applying the rewrite.
# Extract and normalize arguments
x = args[0]
kernel_size = _normalize_to_list(args[1])
stride = _normalize_to_list(
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+15
to
+26
| def _normalize_to_list( | ||
| value: Optional[Union[int, List[int], tuple]], | ||
| default: Optional[List[int]] = None, | ||
| ) -> List[int]: | ||
| """Normalize parameter to list: handle None, int, tuple, list.""" | ||
| if value is None: | ||
| if default is None: | ||
| raise ValueError("Value cannot be None without a default") | ||
| return default | ||
| if isinstance(value, int): | ||
| return [value] | ||
| return list(value) |
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.
Summary:
Implement DecomposeMaxPool1dPass to enable MaxPool1D support on ARM backend
by decomposing max_pool1d into unsqueeze_copy → max_pool2d → squeeze_copy.
Implementation Strategy
Decomposition Approach (Optimal for TOSA/Vela)
The pass decomposes max_pool1d into max_pool2d via unsqueeze_copy/squeeze_copy
operations:
Why This Approach is Optimal
unsqueeze_copy and squeeze_copy map to TOSA RESHAPE which is zero-cost in Vela:
TFA Pipeline Placement (before quantization):
Quantization Handling:
TOSA/Vela Constraints Validated
Reviewed By: 3l1
Differential Revision: D91760459
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani