[PATCH] D99949: [AMDGPU][OpenMP] Add amdgpu-arch tool to list AMD GPUs installed

Thu Apr 15 09:16:19 PDT 2021

JonChesterfield added inline comments.

================
Comment at: clang/tools/amdgpu-arch/CMakeLists.txt:9
+
+find_package(hsa-runtime64 QUIET 1.2.0 HINTS ${CMAKE_INSTALL_PREFIX} PATHS /opt/rocm)
+if (NOT ${hsa-runtime64_FOUND})
----------------
t-tye wrote:
> JonChesterfield wrote:
> > JonChesterfield wrote:
> > > gregrodgers wrote:
> > > > What happens when /opt/rocm is not available?   Again, we need a cross-architecture mechanism to identify the offload-arch. 
> > > Exactly the same as the amdgpu plugin. The cmake detection is char for char identical. This will look in CMAKE_INSTALL_PREFIX, which is where I install these libs when using trunk, and falls back to /opt/rocm which seems to be convenient for some users.
> > Which may need revising at some point - I like installing hsa as if it was an llvm subcomponent, but other people might want a different convention. As long as we remember to change this file + amdgpu's cmake at the same time, all good.
> /opt/tocm will not work with the side-by-side ROCm installation which installs ROCm in directories with the version number. Should there be the ability to configure this?
If /opt/rocm isn't a safe out of the box option for finding rocm, let's remove that from here and the amdgpu plugin. Pre-existing hazard so doesn't need to block this patch.

I'm installing roct and rocr to the same CMAKE_INSTALL_PREFIX as llvm which is why the first clause works out.

How are the rocm components meant to find the corresponding pieces? If there's a rocm cmake install dir variable we could add that to the hints.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99949/new/

https://reviews.llvm.org/D99949