(cco) impl sdma transport in cco primitives#454
Open
QizhouZhang97 wants to merge 12 commits into
Open
Conversation
Contributor
Author
Contributor
Author
|
@Copilot |
QizhouZhang97
force-pushed
the
dev/sdma_adapt_main
branch
from
0feb744 to
882e559
Compare
Contributor
Author
Collaborator
|
@claude review |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
cco primitives support P2P and RDMA/IBGDA but lack an intra-node SDMA (copy-engine) path. SDMA offloads bulk intra-node GPU-to-GPU copies onto the dedicated DMA engines, freeing CUs for compute and improving bandwidth for large transfers. This PR adds SDMA as a first-class transport behind the cco device API.
Technical Details
New device-only ccoSdma session (guarded by HIPCC/CUDACC) built on ccoDevComm::sdma, addressing peers by LSA rank.
Exposes put / get / quiet, each templated on Coop:
ccoCoopThread — a single thread drives one queue (queueId).
ccoCoopWarp — a warp drives all queues, one lane per queue, splitting bytes across sdmaNumQueue.
put/get are non-blocking: each appends COPY + ATOMIC(signal++) on the queue and bumps the expected count; completion lands in the local signal pool and is awaited by quiet.
Fixes SdmaPutWarp slicing: each queue owns a contiguous slice and the last queue absorbs the remainder, so uneven sizes are fully covered.
Test Plan
tests/cpp/cco/test_sdma_put.cpp and test_sdma_get.cpp: thread- and warp-scope alltoall put/get across all ranks on an 8-GPU node, plus benchmark updates for SDMA.
Tests
on n09-355 & on banff 300, sdma put/get/put_mt passed