[Openmp-commits] [PATCH] D101266: [OpenMP][CMake] Pass --cuda-path to regression tests.

Michael Kruse via Phabricator via Openmp-commits openmp-commits at lists.llvm.org
Mon Apr 26 17:35:53 PDT 2021


Meinersbur added a comment.

In D101266#2718200 <https://reviews.llvm.org/D101266#2718200>, @JonChesterfield wrote:

> It leaks an internal detail to the user.

It's only used during testing by llvm-lit whether the built was successful.

> Someone builds their application with openmp enabled, goes to run it, and it fails because it can't find the libraries that clang implicitly created dependencies on, even though clang knows where said dependencies are.

Again, no consequence to the end-user of clang. To compile a program, the user will have to pass `--cuda-path` and/or the environment variables manually (or use `module load`).

Or is your concern that when `check-openmp` succeeds, the user might think that they do not have to point to a CUDA installation anymore?

> Defaulting to the openmp that ships with clang

This is for the search path of `libdevice.so` and `cudart.so`, not `libomp.so` nor `libomptarget.so`.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D101266/new/

https://reviews.llvm.org/D101266



More information about the Openmp-commits mailing list