[Rust] Add physical and logical Warp Load/Store by tpn · Pull Request #12113 · NVIDIA/cccl

tpn · 2026-10-21T01:03:23Z

Superseded by #12088, which now owns this implementation within its complete feature family.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.

Physical and logical warps can use Direct, Striped, Vectorize and Transpose Load/Store through the group-first API. Logical widths are 1, 2, 4, 8, 16 and 32. Sibling groups may choose different counts and defaults while preserving checked ownership and payload validity. Transpose uses independent planned scratch per group and permutes values and validity on mutable Store; the other algorithms preserve the payload. The guide includes generic kernels and finite sibling loops that reconverge before wider operations when completion requirements are met.

Depends on both explicit storage #12105 and Block algorithms #12106, with the pinned compiler owned by cuda-rust #1589. This is a source-branch join: the comparison above storage also includes Block algorithms; the comparison above Block algorithms also includes storage. Neither isolates Warp alone. GitHub's main-based diff includes all prerequisites.

Consolidates #12114, #12115 and #12116, retaining their commits and accepted source, including the corrected API overview.

Original qualification covers host/docs/native checks, compiler controls, standalone consumers, Normal/Full logical-Warp execution, numerical GPU checks, sanitizers and expected traps. Executable sections, ABI, launch facts, scratch and actual stack/local traffic were inspected. Full-profile cuobjdump --dump-elf still fails on retained debug locations; complete DWARF validity remains unverified. Instruction differences remain, and no performance equivalence is claimed. The accepted endpoint passed host CI; checks for this updated draft are tracked separately.

Create a host-only Cargo workspace that specializes selected CUB forms through NVRTC and reports exact scratch layouts with their LTO output. Bind specialization identities to source, headers and toolkit bytes. Include a runnable provider example, Rust guide and CPU checks. Signed-off-by: Trent Nelson <trent@trent.me>

Install the pinned toolchain without updating rustup itself. Signed-off-by: Trent Nelson <trent@trent.me>

Serve native specialization batches through the versioned Oxide boundary and exercise Load and Store through a standalone Cargo consumer with a generic helper dependency. Signed-off-by: Trent Nelson <trent@trent.me>

Signed-off-by: Trent Nelson <trent@trent.me>

Reuse the same helper sites with 64- and 128-thread kernels and retain one device static across their emitted bodies. Check each launch plan and preserve fixed external-symbol diagnostics. Signed-off-by: Trent Nelson <trent@trent.me>

Select fixture output locations from raw CUDA coordinates and exercise payload initialization, access and copying at both item counts. Annotate compile-fail examples and clarify hierarchy ordering in the guide. Signed-off-by: Trent Nelson <trent@trent.me>

Signed-off-by: Trent Nelson <trent@trent.me> # Conflicts: # rust/Cargo.lock

Signed-off-by: Trent Nelson <trent@trent.me>