Repository navigation
[cudax][cuco] Broadcast one slot index in cooperative insert_and_find - #12385
Merged
PointKernel merged 3 commits intoNov 4, 2026
Merged
Conversation
Scan the bucket without break, take the expected slot from the bucket, and broadcast a single combined slot index instead of separate bucket and intra-bucket indices. Forming the slot pointer from the storage base after the shuffle keeps it in global address space.
PointKernel
added
cudax
PointKernel
added
cudax
Contributor
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
Assume slot addresses are not in local memory before the CAS so the atomic_ref local-memory fallback folds early and the expected value no longer spills on sm_90. The assumption is skipped for nvcc and NVRTC before 13.1, which miscompile __isLocal.
PointKernel
changed the title
PointKernel
changed the title
Member
Author
|
/ok to test |
Contributor
🥳 CI Workflow Results🟩 Finished in 1h 14m: Pass: 100%/75 | Total: 15h 41m | Max: 1h 10m | Hits: 74%/65573See results here. |
Contributor
|
No actionable comments were generated in the recent review.
 |
| Layer / File(s) | Summary |
|---|---|
Scan and pointer selection cudax/include/cuda/experimental/__cuco/detail/open_addressing/open_addressing_ref_impl.cuh |
The scan initializes its candidate from bucket slot zero and uses intra-bucket index zero as the default. It records the first qualifying slot, then shuffles the combined index to form pointers for matching slots and insertion. |
Possibly related PRs
- NVIDIA/cccl#12355: It applies the combined-slot-index and post-shuffle pointer pattern to cooperative
findin the shared implementation. - NVIDIA/cccl#11869: It added cooperative insertion to the shared open-addressing implementation changed here.
Suggested reviewers: srinivasyadav18, sleeepyjack, gevtushenko
Change: Refactor
Merge Risk: 
4cc1c
-
No actionable issue is established for the changed insert-and-find behavior; the PR is mergeable after normal checks.
sleeepyjack
approved these changes
Nov 4, 2026
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
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.
Description
Related to #12382.
This PR improves the cooperative
insert_and_findcodegen shared byfixed_capacity_mapandfixed_capacity_set. The bucket scan no longer breaks out of the loop, and the winning lane's slot index is broadcast with one shuffle instead of two, with the slot pointer formed from the storage base so it stays in global address space.The cooperative device
insert_and_findis now on par with cuCollections, e.g. 357 → 308 SASS instructions on sm_120 and 362 → 332 on sm_90 forint32keys with CG size 4.Checklist