Cortex-M: Add depthwise conv2d operator #16233
Open
+636
−19
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/16233
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 3 Unrelated FailuresAs of commit 577364c with merge base 662b973 ( NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
rascani
force-pushed
the
cmsis_depthwise_conv
branch
from
db21fc0 to
3de2c83
Compare
added 7 commits
December 15, 2025 11:49
Add quantized depthwise convolution operator for the Cortex-M backend using CMSIS-NN's optimized arm_depthwise_conv_wrapper_s8 function. Key changes: - New op_quantized_depthwise_conv2d.cpp with CMSIS-NN implementation - Python operator registration in operators.py with reference implementation - Operator schema definition in operators.yaml - Updated ConvertToCortexMPass to automatically detect and route depthwise convolutions (where groups == input_channels) to the specialized operator - Comprehensive test coverage with 5 test cases covering different depthwise convolution scenarios (stride, padding, bias, depth multiplier) The implementation validates the depthwise constraint (groups must equal input channels) and supports NHWC layout, int8 quantization, per-channel requantization, and configurable stride/padding/dilation parameters.
…lidations Key changes: - Move depth_multiplier calculation from runtime to AOT pass (eliminates runtime division by computing depth_multiplier = output_channels / input_channels in the graph transformation pass) - Add critical defensive validations in validate_depthwise_conv2d_arguments(): * Validate IHWO weight layout (dimension 0 must be 1) * Validate dilation == 1 (CMSIS-NN constraint) * Validate depth_multiplier consistency with channel counts - Fix CMSIS-NN API usage: * Use arm_depthwise_conv_wrapper_s8_get_buffer_size() with correct parameters * Improve buffer allocation error handling with detailed error messages - Add _compute_depthwise_conv2d_output_shape() to read channels from correct dimension (dim 3 for IHWO layout vs dim 0 for OHWI) - Update operator schema to use depth_multiplier parameter instead of groups This ensures proper validation of CMSIS-NN constraints and moves computation to compile-time where possible.
CMSIS-NN arm_depthwise_conv_wrapper_s8 only supports batch size 1. Add validation in both AOT pass (fail during compilation) and runtime (defensive check). Add 6 test cases covering edge cases: - Combined stride/padding/bias - 1x1 kernels (common in mobile networks) - Higher depth_multiplier (4) - Asymmetric kernels (1x3) - Asymmetric stride/padding - Larger kernels (5x5) Fix depthwise_conv2d_stride test to use batch size 1.
rascani
force-pushed
the
cmsis_depthwise_conv
branch
from
3de2c83 to
577364c
Compare
mansnils
requested changes
Dec 16, 2025
| # Detect depthwise convolution: | ||
| # PyTorch depthwise weight is [out_ch, 1, H, W] where dimension 1 is 1 | ||
| # and groups == input_channels (groups > 1) | ||
| is_depthwise = weight_tensor.shape[1] == 1 and groups > 1 |
Collaborator
There was a problem hiding this comment.
I think groups could be 1 for a DW conv?
So a better condition is then, is_depthwise = (in_channels == groups) and (out_channels % in_channels) == 0
| # and groups == input_channels (groups > 1) | ||
| is_depthwise = weight_tensor.shape[1] == 1 and groups > 1 | ||
|
|
||
| if is_depthwise: |
Collaborator
There was a problem hiding this comment.
Here we actually have the benefit of choosing between a regular and DW conv. It is likely but not certain that the un-optimized CMSIS-NN DW conv or the one without any SIMD is less efficient that the corresponding CMSIS-NN conv. We don't know exactly until we measure. We could then add something like this for now with a TODO comment:
optimal_dw_conv_constraints = (
in_channels == out_channels and dilation == [1, 1]
) or in_channels == 1
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.
Summary
Add quantized depthwise convolution operator for the Cortex-M backend using CMSIS-NN's optimized arm_depthwise_conv_wrapper_s8 function.
Fixes #16105
Test plan