[ROCM] Navi21 Enablement 1 (#69942)
Summary:
This pr is the first in a series of prs that will introduce support for Navi 21 GPUs.
There is one change here which is that for ROCM we define an alternative num_threads function that returns a constant number 256 instead of warpsize dependent function used by CUDA.
cc jeffdaily sunway513 jithunnair-amd ROCmSupport KyleCZH
Pull Request resolved: https://github.com/pytorch/pytorch/pull/69942
Reviewed By: ejguan
Differential Revision: D33759380
Pulled By: malfet
fbshipit-source-id: e69b556aded15a66d11db256be727dbc91372b7d
(cherry picked from commit 76e4df0901a1941a64cc35952f17d3881cd60dc0)
