Skip to content

perf(scan): register-raking items-per-thread GPU prefix scan - #108

Open
shreyas-omkar wants to merge 1 commit into
JuliaGPU:mainfrom
shreyas-omkar:sh/accumulate-raking
Open

perf(scan): register-raking items-per-thread GPU prefix scan#108
shreyas-omkar wants to merge 1 commit into
JuliaGPU:mainfrom
shreyas-omkar:sh/accumulate-raking

Conversation

@shreyas-omkar

Copy link
Copy Markdown
Member

No description provided.

Each thread serial-scans its contiguous run of ITEMS elements in registers,
the block_size run-totals are scanned once with a Blelloch tree, and each
thread folds its exclusive block-prefix back in. A block covers
block_size*ITEMS elements without deepening the tree, so raising ITEMS cuts
the block count (and the per-block prefix array) at the same barrier cost.
Default 2 keeps current behaviour; opt into 4 via `_scan_items_per_thread`.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
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.

1 participant