Skip to content

Add autotuning candidate filtering controls. Add MPI backend disablement option to autotune options struct.#154

Open
romerojosh wants to merge 2 commits into
mainfrom
autotune-candidate-filters
Open

Add autotuning candidate filtering controls. Add MPI backend disablement option to autotune options struct.#154
romerojosh wants to merge 2 commits into
mainfrom
autotune-candidate-filters

Conversation

@romerojosh

Copy link
Copy Markdown
Collaborator

This PR introduces environment variable based filters of candidate backends and process dimensions to be used during the autotuning process. The new variables are:

  • CUDECOMP_AUTOTUNE_TRANSPOSE_BACKENDS is a include/exclude list of transpose backends to try during autotuning.
  • CUDECOMP_AUTOTUNE_HALO_BACKENDS is a include/exclude list of halo backends to try during autotuning.
  • CUDECOMP_AUTOTUNE_P_ROW_RANGE/CUDECOMP_AUTOTUNE_P_COL_RANGE restrict the respective candidate process dimension values during autotuning.

See the updated docs for more detailed information on their use.

Additionally, this PR introduces a disable_mpi_backends flag in the autotuning options structs for users to programmatically disable MPI backends (to mirror the existing NCCL and NVSHMEM disablement flags).

Note that this PR (along with #119) have modified the library ABI since the last release due to changes to the grid descriptor configuration and now autotuning options struct. A follow up PR is planned to address this and make the ABI more stable moving forward starting in the next release.

Signed-off-by: romerojosh <joshr@nvidia.com>
Signed-off-by: romerojosh <joshr@nvidia.com>
@romerojosh

Copy link
Copy Markdown
Collaborator Author

/build

@github-actions

Copy link
Copy Markdown

🚀 Build workflow triggered! View run

@github-actions

Copy link
Copy Markdown

✅ Build workflow passed! View run

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