[PATCH] D128090: [Clang][OpenMP] Process multi-arch compilation options given via -march

[Joseph Huber via Phabricator via cfe-commits]

jhuber6 added a comment.

In D128090#3649579 <https://reviews.llvm.org/D128090#3649579>, @tra wrote:

> In D128090#3649235 <https://reviews.llvm.org/D128090#3649235>, @jhuber6 wrote:
>
>> Interesting, may be worthwhile to query that if it exists, though AMD does this with `amdgpu-arch` which has led to problems for me in the past. But even if it could be wrong it will still spit out an architecture that would run on at least one local card rather than zero.
>
> We do have existing precedent of `-march=native`, so it may make sense to introduce `--offload-arch=native`, though it may be a bit too ambiguous, considering that there may be more than one likely choice -- multiple GPUs, possibly a mix from different vendors. Perhaps we would need something more specific, like `--offload-arch=native-nvidia`.

That's an interesting idea, we could put that functionality in OpenMP using this method as well.

  -Xopenmp-target=nvptx64 -march=native

For CUDA / HIP we could also potentially just make users specify which toolchain the native applies to somewhat like `-Xopenmp-target`. Might be a more general solution than attaching it to the architecture string.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D128090/new/

https://reviews.llvm.org/D128090