Add optimized Arm64 BF16 _to_copy - #22493
Conversation
Summary: Add an optimized `_to_copy.out` kernel. On Arm64, contiguous FP32-to-BF16 and BF16-to-FP32 conversions process eight elements per NEON iteration and use the ExecuTorch threadpool above its grain size. Other dtype pairs, layouts, and platforms fall back to the portable implementation, which remains unchanged. Differential Revision: D118502073
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22493
Note: Links to docs will display an error until the docs builds have been completed. ❌ 5 New FailuresAs of commit 4ee4411 with merge base 4747ab7 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D118502073. |
This PR needs a
|
Summary: Add an optimized
_to_copy.outkernel. On Arm64, contiguous FP32-to-BF16 and BF16-to-FP32 conversions process eight elements per NEON iteration and use the ExecuTorch threadpool above its grain size. Other dtype pairs, layouts, and platforms fall back to the portable implementation, which remains unchanged.Differential Revision: D118502073