Add AXPY oneDNN handler#33046
Conversation
|
Thanks for your contribution! |
|
Sorry to inform you that c7b8f49's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
* Compilation of this function is possible only when Paddle is not build with CUDA nor HIP.
|
CI-APPROVAL failure is because of need for approval to skip test: |
|
@lidanqing-intel PR-CI-APPROVAL failed because of use of |
|
mkldnn bf16 ops tests always use |
|
Hi, @arogowie-intel I cherry-picked your PR and it's been merged. You may close this. |
|
#33632 has merged this PR. So I close this one because we need to summarise merged PRs each week. |
PR types
Performance optimization
PR changes
OPs
Describe
This PR introduces oneDNN AXPY handler providing BLAS::AXPY interface. The use of this handler is enabled at this moment only for
bfloat16data type forsumoperator withSelectedRowsinput tensors. It gives almost 10% performance improvement on CPX machine.Additionally this PR speeds up SGD operator with Selected rows with AXPY handler.