corpus

The kernel corpus

Real, small, self-contained cuda-oxide kernels — the tuning subjects. Fixtures, not products: minimal and readable beats clever.

kernel	why it is here	expected gate behaviour
reduce-flip	the README's warp_id() guard — the case the product exists for	known-flip: disqualified at block_x > 32, admitted at 32
reduce-stable	the same reduction with unconditional barriers	known-stable: nothing disqualified, ever
matmul-tiled	genuinely large tile/block space, constraint-heavy	clean
scan-block	barrier-dense (2 per doubling round) in a uniform loop	clean — a false positive here poisons every scan
histogram	shared-memory capacity pressure up to 48 KB (RC004, --cc)	clean at default budget
stencil-1d	register/unroll trade-offs, launch_bounds as a spec dim	clean — no barriers at all

Structure

Every kernel is a standalone, device-only lib crate:

• Not a workspace member — the main workspace builds and tests with no GPU, no CUDA SDK, and no cuda-oxide checkout.
• Depends on cuda-device alone, via a path dependency on a sibling checkout (../../../cuda-oxide). cuda-core/cuda-host pull in cuda-bindings, whose build script needs cuda.h — that would kill the laptop prune path (measured in docs/research-baseline.md).
• kernel.toml declares the tuning space: dimensions (launch-role or spec-role) and constraints. launchbound space reads it.
• src/params.rs holds the compile-time parameters at repo defaults. The tuner rewrites this file per candidate in a scratch copy; the repo copy never changes.
• needs_cc records the minimum compute capability (cuda-oxide emits .target sm_80 PTX, so 8.0 across the board today).

Prerequisite

git clone https://github.com/NVlabs/cuda-oxide ../cuda-oxide   # sibling of this repo
git -C ../cuda-oxide checkout 50d07314                          # the pinned commit