Consistently use TORCH_CUDA_API for all files that live in cuda targets. (#29158)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29158
My plan is to split out libtorch_cuda.so from libtorch.so. To do this,
I need accurate _API annotations for files in these directories.
I determined the correct set of annotations by looking at
tools/build_variables.py and making sure every file that was a member
of the libtorch_cuda/ATen-cu targets had these annotations. (torch-cpp-cuda
doesn't count since that's going to be where the stuff that has explicit
USE_CUDA lives, so it's going to be in a separate dynamic library).
As future work, it would be good to setup a lint rule to help people
understand what the correct _API annotation to use in a file is; it
would also be good to reorganize folder structure so that the library
structure is clearer.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Differential Revision: D18309593
Pulled By: ezyang
fbshipit-source-id: de710e721b6013a09dad17b35f9a358c95a91030