Automatically infer HCC_AMDGPU_TARGET

• build: Support CUPY_NVCC_GENERATE_CODE=current equivalent in ROCm to set HCC_AMDGPU_TARGET automatically
• runtime: Automatically infer HCC_AMDGPU_TARGET (this environment variable is only needed for hipcc case?)

[kmaehashi]

I skimmed hipcc code (it’s written in perl 😅 ), it seems the output from /opt/rocm/bin/rocm_agent_enumerator -t GPU (which prints arch for each GPU installed) is used when HCC_AMDGPU_TARGET is unset. I’m curious how hiprtc would behave, though.

[leofang]

• build: Support CUPY_NVCC_GENERATE_CODE=current equivalent in ROCm to set HCC_AMDGPU_TARGET automatically

I just did a quick test locally. It seems if only building for local use, as mentioned in the above linked comment hipcc can also infer HCC_AMDGPU_TARGET correctly. I confirmed this by removing the cub and thrust modules, unset both HCC_AMDGPU_TARGET and __HIP_PLATFORM_HCC__, and rebuild. It builds just fine.

After that, we can examine the binary by ROCm’s built-in tool (similar to cuobjdump on CUDA):

$ extractkernel -i cupy/cuda/thrust.cpython-37m-x86_64-linux-gnu.so

which generates three files, two of which are for my arch (gfx906):

$ cd cupy/cuda; ls -lh thrust.cp*
-rw-rw-r-- 1 leofang leofang 328K Feb 23 16:16 thrust.cpp
-rwxrwxr-x 1 leofang leofang 7.6M Feb 23 16:19 thrust.cpython-37m-x86_64-linux-gnu.so
-rw-rw-r-- 1 leofang leofang 6.2M Feb 23 16:20 thrust.cpython-37m-x86_64-linux-gnu.so-000.bundle
-rw-rw-r-- 1 leofang leofang 6.2M Feb 23 16:20 thrust.cpython-37m-x86_64-linux-gnu.so-000-gfx906.hsaco
-rw-rw-r-- 1 leofang leofang  73M Feb 23 16:20 thrust.cpython-37m-x86_64-linux-gnu.so-000-gfx906.isa

Could be relevant to #4720.