mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
Summary: This PR contains the ROCm contributions of last week: * documentation of pyHIPIFY data format originating from #8812 reviewing comments by ezyang * removal of most patch files from the `amd_build` directory and integration into the code base * enabling of previously disabled_features that do compile now * improvement to the static_cast feature in pyHIPIFY (it will only apply static_cast to kernel arguments, not launch arguments) * addition of two workarounds to pyHIPIFY for ROCm/HIP shortcomings: a) `__forceinline__` does not imply `static`, hence change to `__inline__`, b) `std::[exp,log,pow]` math functions cannot be selected in device code, use `::[exp,log,pow]` instead. Both of these workarounds will be removed once the issues are fixed upstream. Neither of these issues have surfaced on the CI but were reproduced internally. Pull Request resolved: https://github.com/pytorch/pytorch/pull/9432 Differential Revision: D8887441 Pulled By: ezyang fbshipit-source-id: 71cf5c6b13772a66d10be369a45ebf06e4e268e1 |
||
|---|---|---|
| .. | ||
| a_torch_cuda___init__.py.patch | ||