Use Kokkos in functions struct replacing TBB #14633
Conversation
|
@masterleinad @kronbichler What would be a good setup to test this? I am happy to run some extended performance tests. |
If I configure with Kokkos_ENABLE_THREADS or Kokkos_ENABLE_OPENMP, right? Any preference between them or anything else we need to know? It would also be sensible to check if we see no overhead for an MPI-only run where one would not want vector operations to be threaded... |
Probably better to use OpenMP over threads. You want to set |
The benchmark in #2496 (comment) using |
|
/rebuild |
| Kokkos::View<T *, typename MemorySpace::kokkos_space> new_values) | ||
| : values_host_buffer(Kokkos::View<T *, Kokkos::HostSpace>("host buffer", 0)) | ||
| , values(new_values) |
There was a problem hiding this comment.
You're just copying new_values twice this way. Make it a const reference to avoid one of them.
There was a problem hiding this comment.
I don't need this constructor anymore anyway. It probably makes sense to replace MemorySpaceData with Kokkos::View at least internally in most places.
|
@masterleinad What do you want to do with this PR? |
I don't have plans to pursue this further at the moment. The performance results weren't that convincing when I initially opened this pull request. |
Currently, we are using TBB for most vector operations on the host and
Kokkosfor theDefaultmemory space. This pull request unifies the implementation by always usingKokkos. We should evaluate if the performance on the host is acceptable (ifKokkoswas configured with a parallel host execution space).