Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

coll/accelerator initial enhancements #12986

Open
wants to merge 2 commits into
base: main
Choose a base branch
from

Conversation

edgargabriel
Copy link
Member

This PR contains two enhancements to the coll/accelerator component:

  • use the device_id and the transfer direction for the accelerator.mem_copy operations: since this information is actually available, we might as well use it and give the accelerator components the ability to potential use this information for tuning purposes.

  • introduce the ability to register the component only for the operations that were provided by the end-user as an MCA parameter. This mimics closely the feature of the coll/ucc component, and the code is in fact based on the coll/ucc component code (with some minor modifications/adjustments).

These PRs are in preparation for additional work on the coll/accelerator component, I just want to avoid having a gigantic PR that is harder to digest.

make use of the correct device id associated with a device buffer
and provide the correct transfer type (DTOH or HTOD) for the accelerator
memcpy operations. This information can be potentially used for some
optimizations internally.

Signed-off-by: Edgar Gabriel <[email protected]>
This PR introduces the ability to register the component only to the
select functions specified by an MCA parameter string. The idea and the
code is based on the UCC component, and some of the bits might be moved
later to coll/base to make the mechanism more gnerally available to
other components as well.

Note, that the PR introduces the define statments for all MPI collective
operations, not just the ones support by the component at the moment,
since it is a bitmask based operation, and we anticipate to add support
for more collective operations into coll/accelerator shortly

Signed-off-by: Edgar Gabriel <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant