[Pytorch] Weaker memory ordering for c10::intrusive_ptr (#48221)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48221
load-acquire, acquire-release increment and decrement. (We
need acquire-release increment to make unique() and use_count()
reliable.) Note that this doesn't make a difference on x86, but we
should expect it to improve things on ARM and ARM64.
ghstack-source-id: 117065956
Test Plan: Careful review :)
Reviewed By: ezyang
Differential Revision: D24708209
fbshipit-source-id: 5e574115eee5c0a65047b638c5f9b1ec0124d04d