Skip to content

feat(cco): add cco_sdma_put_ptr_rsig__thread and cco_sdma_signal_slot bitcode - #786

Merged
jhchouuu merged 2 commits into
mainfrom
binding_cco_test
Oct 16, 2026
Merged

jhchouuu merged 2 commits into
mainfrom
binding_cco_test

Conversation

@binding7012

Copy link
Copy Markdown
Contributor

Motivation

add cco_sdma_put_ptr_rsig__thread and cco_sdma_signal_slot bitcode

@jhchouuu jhchouuu left a comment •

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@jhchouuu jhchouuu changed the title add cco_sdma_put_ptr_rsig__thread and cco_sdma_signal_slot bitcode feat(cco): add cco_sdma_put_ptr_rsig__thread and cco_sdma_signal_slot bitcode Oct 16, 2026
@jhchouuu
jhchouuu merged commit 217150f into main Oct 16, 2026
23 of 25 checks passed
@jhchouuu
jhchouuu deleted the binding_cco_test branch October 16, 2026 14:26
yangyuhuiling pushed a commit that referenced this pull request Oct 19, 2026
Picks up 11 commits, notably bba3dd8 (#790), which fixes test_peer_pool's
FailingEngine against the new CreatePlanBuilder interface. Without it the PR's
"mori umbp unit test" job fails to compile when run against current main, as
the /ci-rerun on 10-16 did; nothing in this PR touches UMBP.

No conflicts. The only file here near this PR's area is
src/cco/device/cco_device_wrapper.cpp (#786, new SDMA bitcode entry points),
which the AsyncLL kernels do not use.

Co-Authored-By: Claude <noreply@anthropic.com>
whx-sjtu added a commit to ROCm/ATOM that referenced this pull request Oct 19, 2026
70c93048 compiles every big-M tier at initialization into a disk cache
and takes its SDMA device functions from MORI (ROCm/mori#786). On this
implementation the 32K A/B gains +27.19% total throughput with GSM8K at
parity. Big-M rejects 16384 local rows, so the 8192-row tile cap stays.
Rebased onto main 53e4fb5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
jhchouuu added a commit that referenced this pull request Oct 20, 2026
- cco.hpp: add ccoSdma::putPtr / getPtr (put/get delegate to them) and
  a bounds-checked signalSlot.
- Wrapper: cco_sdma_{put,get}_ptr__<coop>(..., flags) carry every other
  template axis in flags (aggregate, signal-per-copy, local and remote
  signal, same-queue); a constant flags folds to one instantiation.
  Add cco_sdma_wait_signal and cco_local_ptr. Existing symbols unchanged.
- FlyDSL and Triton Sdma: put / get / put_ptr / get_ptr / wait_signal /
  signal_slot with C++'s template arguments as keywords; signal= stays
  compatible. FlyDSL binds every registered function.
- Register #786's put_ptr_rsig / signal_slot wrappers.
- Tests: argument types checked against the bitcode, reverse symbol check,
  constant-flags folding check, GPU signal matrix (Triton) and a FlyDSL
  signals test.
eppaneamd added a commit to ROCm/aiter that referenced this pull request Oct 30, 2026
The "late rank" case ran the small-batch kernel (ag_rs, ar) and SpRsNorm.
Big-M takes the same limit and has its own launch and error word, but no
case made one of its ranks late. In each mode the case now also runs its
steps with 64 local tokens through big-M (bigm_min_tokens = 64 * world).

Big-M links MORI's SDMA device functions (ROCm/mori#786), so this part,
like the "big-M" case, needs a MORI that provides them.

Signed-off-by: Rony Leppänen <rony.leppanen@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
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.

2 participants