Skip to content

[Docs] Add user-guide page for qd.simt.grid.* primitives#641

Draft
hughperkins wants to merge 1 commit intomainfrom
hp/docs-grid
Draft

[Docs] Add user-guide page for qd.simt.grid.* primitives#641
hughperkins wants to merge 1 commit intomainfrom
hp/docs-grid

Conversation

@hughperkins
Copy link
Copy Markdown
Collaborator

Summary

New user-guide page docs/source/user_guide/grid.md documenting the grid tier — currently a single op, qd.simt.grid.memfence().

Companion to the in-flight subgroup (#639), block (#638), and atomics (#640) docs; this is one of a series of tier-by-tier user-guide pages. Independent — no cross-links to unmerged docs.

Includes:

  • Backend support matrix (CUDA-only today; AMDGPU and SPIR-V tracked as future work).
  • Why there is no grid.sync() and what to do instead (kernel relaunch).
  • Barrier-vs-fence framing carried down the tier ladder (subgroup → block → grid).
  • Worked example: cross-block reduction with the producer-fence + consumer-fence idiom that Onesweep / decoupled-look-back scans rely on.
  • Performance and portability notes — cost scales with the cache-invalidation domain; not a substitute for atomics.

Adds grid to the SIMT-primitives toctree in index.md.

Test plan

  • cd docs && make html builds without warnings.
  • Visual check rendered page on RTD preview.

Made with Cursor

Documents qd.simt.grid.memfence() — currently the sole public op in
the qd.simt.grid namespace. Covers semantics (device-scope memory
fence, no thread convergence), backend support (CUDA only today),
the producer-fence + consumer-fence pattern that decoupled-look-back
scans and Onesweep build on, and how to pick between subgroup /
block / grid scopes.

Also surfaces the asymmetry: there is no qd.simt.grid.sync()
(grid-scope barrier) — full thread synchronization across blocks
requires a kernel relaunch.

Adds grid.md to the SIMT-primitives toctree.
@github-actions
Copy link
Copy Markdown

github-actions Bot commented May 7, 2026

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