[Openmp-commits] [openmp] [OpenMP] Fix hyper barrier performance issue (PR #195473)

[Terry Wilmarth via Openmp-commits]

TerryLWilmarth wrote:

> > I think, aggressive active waiting mainly makes sense on otherwise empty systems. Yielding in the barrier can reduce context switches in the middle of computation.
> > A better option might be to allow changing the OpenMP wait policy at runtime (i.e. introduce omp_get/set_wait_policy calls), so that the application can encode the preferred behavior.
> 
> I agree. Also, simply using the already available kmp_set_blocktime inside the application gets you most of the way there without any changes to the runtime. But I agree it would be nice to have the API to set wait policy at runtime too.

I forgot -- there is also kmp_set_library.  Setting both at runtime to match the desired wait policy should work.

https://github.com/llvm/llvm-project/pull/195473