fix(cuda): pass the plane index to the 16-bpc PSNR kernel

[lusoris]

Summary

CUDA chroma PSNR above 8 bpc was a luma number. Found by the twin-drift sweep (ids 7, 80), confirmed by measurement before touching code.

Netflix 576x324 pair, frame 0:

	psnr_y	psnr_cb	psnr_cr
CPU, 10 bpc	34.786288	39.255496	41.375212
CUDA, 10 bpc before	34.786288	34.358981	34.358981
CUDA, 12 bpc before	34.792654	34.365347	34.365347
CUDA, 10/12 bpc after	identical to CPU on all three planes

psnr_cb == psnr_cr, both luma-like: the chroma dispatches were measuring a chroma-sized top-left window of the luma plane. psnr_cuda_dispatch has always passed plane as the sixth kernel argument for both bit depths; the 8-bpc kernel declares it and indexes data[plane], the 16-bpc kernel was declared without it and hard-coded index 0. cuLaunchKernel silently ignores the surplus argument, so nothing failed loudly. SYCL and HIP are correct (checked at 10 bpc).

Every PSNR parity fixture was 8-bit with flat chroma — both sides sit at the psnr_max sentinel, so a wrong-plane chroma read could never show. The fixture is now bit-depth generic (8-bit kept byte-identical, widened above 8 bpc with non-flat, ref/dist-different chroma) and registered again at 10 bpc.

• test_cuda_psnr_parity 3/3, test_cuda_psnr_parity_10bit 3/3 on an RTX 4090.

Type

• fix — bug fix
• test — test-only (new gate)
• sycl / cuda / simd — backend-specific (CUDA)

Checklist

• Commits follow Conventional Commits.
• make format && make lint is green locally (pre-commit run --files clean; check-state-md-rows OK; ADR index in sync).
• Unit tests pass: meson test -C build — the touched targets on CUDA, listed above.
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2 — equivalent done directly: CPU vs CUDA identical to six decimals on all planes at 10 and 12 bpc.
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap — only the CUDA 16-bpc kernel was wrong; SYCL and HIP verified correct at 10 bpc.
• If I added a new .c / .h — no new source files.
• If this is a breaking change — not breaking; the previous values were wrong.
• If this PR adds an ADR, the row lives in docs/adr/_index_fragments/ — in sync.

Bug-status hygiene (ADR-0165)

• docs/state.md updated — T-CUDA-PSNR-16BPC-CHROMA-READS-LUMA-2026-09-07 closed.

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• If I believe a golden value must change — none does; the golden gate is CPU-only and 8-bit.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: one measurement and one signature diff, both recorded in ADR-1215's Context.
• Decision matrix — ## Alternatives considered in ADR-1215.
• AGENTS.md invariant note — added to core/src/cuda/AGENTS.md.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/fixed/cuda-psnr-16bpc-plane-argument.md.
• Rebase note — docs/rebase-notes.md, "ADR-1215 — per-plane CUDA kernels must take the plane index".

Reproducer

Y=python/test/resource/yuv
for f in psnr psnr_cuda; do
  vmaf -r $Y/src01_hrc00_576x324.yuv420p10le.yuv -d $Y/src01_hrc01_576x324.yuv420p10le.yuv \
       --width 576 --height 324 --pixel_format 420 --bitdepth 10 --frame_cnt 1 --no_prediction \
       --feature $f --output /dev/stdout --json | grep -oE '"psnr_c[br]": *[0-9.]+'
done
meson test -C build test_cuda_psnr_parity_10bit

On master the CUDA line prints psnr_cb = psnr_cr ≈ 34.36 against the CPU's 39.26 / 41.38; on this branch they match.

Known follow-ups

• The sweep flagged (id 85, ids 31–33) that the psnr_hvs GPU round-trip scales 9- and 11-bit input by 16 in the kernel while the host divides by a different factor — an unusual-bit-depth defect, tracked separately.

🤖 Generated with Claude Code