[oneDNN] Optimize fused elementwise kernel #59663
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
LLee233
changed the title
|
Hi @yuanlehome, would you mind helping check this PR? Since the CI coverage seems not work... The coverage result didn't show up and I restarted the check but the result was as still. Thanks~ |
Okay, please wait for the latest CI results. |
Hi @yuanlehome, seems it's still no result... |
|
@xinyu-intel, @vivienfanghuagood, @yuanlehome, hi, would you mind helping review this PR? Thanks~ |
| const auto src_y_memory = handler.AcquireSecondSrcMemory(non_const_y); | ||
| const auto src_x_memory = | ||
| handler.swin_case ? (x.numel() == y.numel() | ||
| ? handler.AcquireExtendSrcMemory(non_const_x, 0) |
There was a problem hiding this comment.
This aligns with former PR #59421. Since we need to manually broadcast src1/src2, I hereby name such operations as "extend".
PR types
Bug fixes
PR changes
Others
Description
This PR aims for int8 case in #59252 when
config.enable_mkldnn_int8()is activated. For float & int8, paddle will go through different passes and hence different kernels. So for int8 case, the refreshed kernel can not utilizes what former PR #59421 optimizes.