aarch64: apply the cherrypicked onednn PR-1768 #1717
Conversation
This is to improve the torch.compile() perf by 5.8x on AWS Graviton3 instances. This patching is required till PyTorch oneDNN is upgraded to v3.4.
snadampal
force-pushed
the
cherrypick_onednn1768
branch
from
1e01d5d to
32d9e8e
Compare
This is to improve the torch.compile() perf by 5.8x on AWS Graviton3 instances. This patching is required till PyTorch oneDNN is upgraded to v3.4.
|
Why have we landed this change? If it missed the deadline for oneDNN branch cut, then it should wait until the next release or needs to be cherry-picked into one |
| os.system("cd /pytorch/third_party/ideep/mkl-dnn && patch -p1 < /builder/mkldnn_fix/fix-xbyak-failure.patch") | ||
|
|
||
| print("Applying mkl-dnn patch to improve torch.compile() perf") | ||
| os.system("cd /pytorch/third_party/ideep/mkl-dnn && patch -p1 < /builder/mkldnn_fix/onednn-pr1768-aarch64-add-acl-sbgemm-inner-product-primitive.patch") # noqa: E501 |
There was a problem hiding this comment.
Not to author, reviewer: I don't know why those were accepted initially, but one should avoid using os.system as much as possible, as it gives no signal back on whether the underlying command succeeded or failed, this way making regression detection a much harder process. (i.e. one can add os.system("fokjhdsfkjhsef"); anywhere in this script and it will be just a no-op)
subprocess.check_call is a better way of doing that, for example like that:
builder/aarch64_linux/aarch64_wheel_ci_build.py
Lines 111 to 112 in ab5fc90
This way typo in patch name or failure to apply the patch will result in the failure
There was a problem hiding this comment.
Sounds good. Sorry about missing this one. I see this commit adresses exactly that: 0926610
4 participants
This is to improve the torch.compile() perf by 5.8x on AWS Graviton3 instances. This patching is required till PyTorch oneDNN is upgraded to v3.4.