Skip to content

[WIP] Share RAFT selection kernel instantiations - #2889

Draft
divyegala wants to merge 4 commits into
codex/dedup-function-deps-2889from
codex/dedup-shared-selection
Draft

divyegala wants to merge 4 commits into
codex/dedup-function-deps-2889from
codex/dedup-shared-selection

Conversation

@divyegala

@divyegala divyegala commented Oct 22, 2026 •

Copy link
Copy Markdown
Contributor

Move dense, CSR, radix, segmented-sort and warp-sort selection launchers into separate owners while preserving the existing typed callers.

The existing generate_inst_matrix expands ordinary JSON matrices directly into separate .cu files, with one concrete owning instantiation per generated translation unit. Launcher declarations stay in headers; host definitions and the corresponding device kernels stay in CUDA sources. Shared lower kernels have their own owners.

Prerequisites: #2937, #2913.

Diff against this PR’s parent: 21 files, +2046/−21 lines.

Validation: native JSON/CMake generation checked on this exact branch and its dependency closure; each generated owner has one owning instantiation. Complete-library builds, device-kernel census, and focused runtime checks are tracked in #2888.

Binary savings: the previous figures measured the grouped-owner implementation and are retired for this revised layout. Per-PR CUDA 12.9 and 13.3 complete-library remeasurement is pending; increased fatbin-container overhead means the old figures must not be carried forward.

@copy-pr-bot

copy-pr-bot Bot commented Oct 22, 2026

Copy link
Copy Markdown

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.

@divyegala
divyegala added this pull request to stack #2893 October 22, 2026 00:46
@divyegala
divyegala removed this pull request from stack #2893 October 22, 2026 00:57
@divyegala divyegala changed the title [WIP] Reuse shared selection launchers in CAGRA and IVF-PQ [WIP] Own shared dense and sparse RAFT selection host dispatch Oct 22, 2026
@divyegala
divyegala force-pushed the codex/dedup-shared-selection branch 2 times, most recently from ca1f869 to c683e82 Compare October 23, 2026 02:20
@divyegala divyegala changed the title [WIP] Own shared dense and sparse RAFT selection host dispatch [WIP] Share dense and sparse RAFT selection kernels Oct 23, 2026
@divyegala
divyegala changed the base branch from main to codex/dedup-cuda-radix-sorting October 23, 2026 02:20
@divyegala
divyegala force-pushed the codex/dedup-shared-selection branch from c683e82 to 04b09e3 Compare October 23, 2026 07:45
@divyegala divyegala changed the title [WIP] Share dense and sparse RAFT selection kernels [WIP] Share RAFT selection kernel instantiations Oct 23, 2026
@divyegala
divyegala changed the base branch from codex/dedup-cuda-radix-sorting to codex/dedup-function-deps-2889 October 23, 2026 07:46
@copy-pr-bot

copy-pr-bot Bot commented Oct 23, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: No status

Development

Successfully merging this pull request may close these issues.

1 participant