Skip to content

MPI performance updates and Kernel naming#158

Open
jacob-moore22 wants to merge 12 commits into
mainfrom
MPI_Updates
Open

MPI performance updates and Kernel naming#158
jacob-moore22 wants to merge 12 commits into
mainfrom
MPI_Updates

Conversation

@jacob-moore22

Copy link
Copy Markdown
Collaborator

This PR does 2 things, first, it attempts to improve the "safety" of the mpi types. MPI send/recv bookkeeping that was previously flat data members on MPICArrayKokkos is consolidated into a dedicated host-side buffer object, accessed through a Kokkos host view. This involved a few method updates to make the new access pattern work.

Second, it adds naming to the FOR_/DO_/RUN_* kernels. This optional trailing kernel name support to MATAR's top-level Kokkos parallel macros. When provided, the name is forwarded as the first argument to Kokkos::parallel_for / Kokkos::parallel_reduce, enabling Kokkos Tools, NVTX, rocprofiler, VTune, and similar profilers to attribute kernels with human-readable labels. When omitted, behavior is unchanged (Kokkos already defaults unnamed kernels to "").

Backward compatibility: All existing call sites compile and run unchanged. The name is purely optional.

Also, I added lots of unit test for the macros that now get ran with everything else.

Still lots of testing to do, namely with MPI and CUDA.

@jacob-moore22 jacob-moore22 self-assigned this Jun 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant