MV-532: [MV-MIOpen] Benchmark & Port AvgPool #48
Conversation
This was discussed with the Team leader before. We still need to port Avgpool op separately. Please continue reviewing |
There was a problem hiding this comment.
Just partial review until I tried to run your driver and realized that benchmark result seems to not be correct. May I ask you to check benchmark result before I continue to review? Thank you. And please run make analyze after that because I also tried to run it and the PR produced some errors, please fix those errors too.
|
Please continue reviewing @littlecutebird @cognaiger9 |
| std::vector<size_t> ComputeStrides(std::vector<size_t> inputDim) const | ||
| { | ||
| if(!is_contiguous) | ||
| std::swap(inputDim.front(), inputDim.back()); | ||
| std::vector<size_t> strides(inputDim.size()); | ||
| strides.back() = 1; | ||
| for(int i = inputDim.size() - 2; i >= 0; --i) | ||
| strides[i] = strides[i + 1] * inputDim[i + 1]; | ||
| if(!is_contiguous) | ||
| std::swap(strides.front(), strides.back()); | ||
| return strides; | ||
| } | ||
| }; | ||
|
|
There was a problem hiding this comment.
@long10024070 maybe we should put this function into an util file in gtest instead of duplicating it in every gtest files
|
LGTM |
|
OK |
Added AvgPool 2D 3D forward and backward with new solvers and bfp16 supported.
Added driver test and gtest for AvgPool.
New API is guarded by MIOPEN_BETA_API macro.
Compared to ROCm pytorch: Link
Average over all cases:
AvgPool 2D