Add ARM embedded platform compatibility to all operators#17457
Add ARM embedded platform compatibility to all operators#17457Ninja91 wants to merge 2 commits intopytorch:mainfrom
Conversation
meta-cla
bot
added
the
CLA Signed
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/17457
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Cancelled JobAs of commit 806def5 with merge base 6e31609 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOB - The following job was cancelled. Please retry:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
There was a problem hiding this comment.
Pull request overview
Enable Cortex-M operator targets to build under ARM32 embedded configurations by constraining operator build targets to embedded CPU configs and widening visibility of required portable CPU utilities so Cortex-M ops can depend on them.
Changes:
- Extend visibility of selected portable CPU util libraries to allow
//executorch/backends/cortex_m/...dependencies. - Introduce
_ARM_EMBEDDED_PLATFORMSand applycompatible_withconstraints to Cortex-M operator libraries and the aggregatecortex_m_operatorstarget. - Expand the Cortex-M
OPERATORSlist to include all 15 operators declared inoperators.yaml.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
| kernels/portable/cpu/util/targets.bzl | Adds Cortex-M backend visibility to specific portable CPU util targets required by Cortex-M ops. |
| backends/cortex_m/ops/targets.bzl | Adds ARM embedded compatibility constraints and expands the set of Cortex-M operator targets emitted. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| visibility = ["PUBLIC"], | ||
| platforms = CXX, | ||
| compatible_with = _ARM_EMBEDDED_PLATFORMS, | ||
| exported_deps = all_op_targets, |
There was a problem hiding this comment.
Setting compatible_with = _ARM_EMBEDDED_PLATFORMS on cortex_m_operators (and on each op_* library) will make these targets incompatible with non-ARM embedded configurations. However, cortex_m_generated_lib(_headers) and the cortex_m unit tests depend on these targets and are not similarly constrained, which can cause target incompatibility/build failures in default (host) builds. Consider either propagating the same compatibility constraints to the generated-lib targets/tests, or narrowing the constraint to only the operators that truly require embedded-only deps (e.g., CMSIS-NN-based ops) so host builds/tests for pure-C++ ops still work.
Ninja91
force-pushed
the
export-D93254992
branch
from
62bc39f to
b772a64
Compare
Summary: Enable cortex_m operators to build for ARM embedded platforms (FVP simulation, on-device execution). ## Context This diff is required for D92456499 (use on_device_model.pte for CC pipeline offline eval). When the FVP benchmark runner builds for ARM embedded platforms, cortex_m operators must be compatible with those platforms. ## Changes 1. Added `_ARM_EMBEDDED_PLATFORMS` constant for ARM32 embedded targets (ovr_config//cpu:arm32-embedded, arm32-embedded-fpu) 2. Added `compatible_with = _ARM_EMBEDDED_PLATFORMS` to operator definitions 3. Expanded `OPERATORS` list to include all 15 cortex_m operators (matching operators.yaml) ## Operators Included The cortex_m operators handle quantization for ARM Cortex-M targets: - `quantize_per_tensor`, `dequantize_per_tensor` - Core quantization - `quantized_linear`, `quantized_conv2d`, `quantized_depthwise_conv2d` - Linear/Conv layers - `quantized_add`, `quantized_mul` - Element-wise operations - `quantized_avg_pool2d`, `quantized_max_pool2d`, `quantized_transpose_conv2d` - Pooling/Transpose - `softmax`, `transpose`, `pad`, `minimum`, `maximum` - Utility ops Requires D93254993 (visibility fix) to allow cortex_m operators to depend on portable kernel utilities. Differential Revision: D93254992
Ninja91
force-pushed
the
export-D93254992
branch
from
b772a64 to
7b62516
Compare
Summary: Add cortex_m backends to visibility for portable kernel utilities that cortex_m operators depend on. ## Context The cortex_m operators (`quantized_linear`, `quantized_conv2d`, `quantized_add`, etc.) require utilities from `//executorch/kernels/portable/cpu/util/` for kernel calculations, tensor operations, and padding. These utilities had restricted visibility that excluded `//executorch/backends/cortex_m/...`, causing build failures when building cortex_m operators for ARM embedded platforms. This diff is part of a stack enabling D92456499 (use `on_device_model.pte` for CC pipeline offline eval). ## Changes Updated visibility for three utility targets: | Utility | Purpose | |---------|---------| | `kernel_ops_util` | Kernel dimension calculations for quantized ops | | `copy_ops_util` | Tensor copy operations | | `padding_util` | Padding calculations for pooling operators | Differential Revision: D93254993
Summary: Enable cortex_m operators to build for ARM embedded platforms (FVP simulation, on-device execution). ## Context This diff is required for D92456499 (use on_device_model.pte for CC pipeline offline eval). When the FVP benchmark runner builds for ARM embedded platforms, cortex_m operators must be compatible with those platforms. ## Changes 1. Added `_ARM_EMBEDDED_PLATFORMS` constant for ARM32 embedded targets (ovr_config//cpu:arm32-embedded, arm32-embedded-fpu) 2. Added `compatible_with = _ARM_EMBEDDED_PLATFORMS` to operator definitions 3. Expanded `OPERATORS` list to include all 15 cortex_m operators (matching operators.yaml) ## Operators Included The cortex_m operators handle quantization for ARM Cortex-M targets: - `quantize_per_tensor`, `dequantize_per_tensor` - Core quantization - `quantized_linear`, `quantized_conv2d`, `quantized_depthwise_conv2d` - Linear/Conv layers - `quantized_add`, `quantized_mul` - Element-wise operations - `quantized_avg_pool2d`, `quantized_max_pool2d`, `quantized_transpose_conv2d` - Pooling/Transpose - `softmax`, `transpose`, `pad`, `minimum`, `maximum` - Utility ops Requires D93254993 (visibility fix) to allow cortex_m operators to depend on portable kernel utilities. Differential Revision: D93254992
Ninja91
force-pushed
the
export-D93254992
branch
from
7b62516 to
806def5
Compare
1 participant
Summary:
Enable cortex_m operators to build for ARM embedded platforms (FVP simulation, on-device execution).
Changes:
This requires the visibility fix in portable kernel utilities (parent diff) to allow
cortex_m operators to depend on kernel_ops_util, copy_ops_util, and padding_util.
Differential Revision: D93254992