Christian Trott

icon company Sandia National Laboratories icon location Albuquerque, NM, USA icon location High Performance Computing developer with extensive experience in modern architectures (APUs, GPUs, MIC). Lead developer of the @kokkos programming model.

Results 519 comments of Christian Trott

mdspan converting constructors

Retest this please!

[Bug] Parallel for on Cuda not calling team_size_recommended

I think the extra stuff is actually missing in the first code place ...

[Bug] Parallel for on Cuda not calling team_size_recommended

Some of that scratch is in case you do a nested `parallel_reduce`.

Only copy data in UnorderedMap deep_copy

Retest this please

simd: allow wider vector width for 32 bit types

Windows CUDA wasn't passing because of restricitons of NVCC with MSVC. I think you may need to guard the test against Windows + CUDA

windows build with msvc: issue with openMP detection

I also posted this in the other thread: Are you trying to use a different OpenMP implementation, or just MSVCs? I think we need to add a configure time error...

Enhancement: Add a `Fencer` class for scope-guarding "no execution space instance provided" overloads

I want to think a bit more about this, but my initial reaction is that this isn't a thing I want in particular as a public interface. It also could...

allow `local_deep_copy` or `deep_copy` to work at team and thread level

Actually we should just make all of these things overloads of deep_copy with the corresponding policy arguments ... And there is a good reason for us to provide this because...

allow `local_deep_copy` or `deep_copy` to work at team and thread level

I took the freedom to change the title and make this a feature request.

Nightly unit test failure, OpenMP backend, gcc/8.3.0: openmp.reducers_int Kokkos::abort: Requested Team Size is too large!

Yeah and also the original PRs shouldn't have caused the issue. The only way how OpenMP fails TeamSize check is if you ask for more than occupancy isn't it?