[tpn]

Superseded by #12088, which now owns this implementation within its complete feature family. The original commits, branch and validation evidence remain available. This draft is closed unmerged.

---

Historical scope and validation below refer to the original source heads.

Rust kernels can load and store block tiles through the public cuda.coop API, with inferred element types and item counts:

let mut items = coop::ThreadData::<_, ITEMS_PER_THREAD>::new();
coop::load(coop::this_block(), input, &mut items, Default::default());
coop::store(coop::this_block(), &mut output, &items, Default::default());

Direct operations support ten numeric types on complete one-dimensional blocks. Options supply counts, offsets and typed defaults. Load tracks initialized slots; Store accepts a shared payload and preserves its values and validity. Scalar Store supplies one item per thread.

This update aligns the compiler pins and lockfiles with Rust 1.99.1 and corrects three expected diagnostic strings. The API, provider and compiler-case programs are unchanged.

Owning validation passes: host tests, strict lints/docs and provider MSRV 1.85; native tests covering 16 and 100 requests; the 1,906-value example, 925 public cases and seven offset cases; four sanitizer commands and seven expected traps. All 13 compiler controls pass (four accepts, nine rejects). Executable-source checks transfer across the final diagnostic-only change; the controls ran on the corrected commit. Host CI and all upstream checks pass on this commit.

Builds on #12091 and #12092. The PR diff against main is cumulative. Immutable comparisons above R02 and above R03 each also include the other prerequisite.

[copy-pr-bot]

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.