Research-0912: Pixel-format edge coverage audit¶

• Date: 2026-05-31
• Author: Claude (Anthropic) on behalf of Lusoris
• Related: ADR-0912

Context¶

Most libvmaf unit tests under core/test/ allocate VmafPicture instances at (VMAF_PIX_FMT_YUV420P, 8). The chroma-stride and bit-depth-shift code paths that diverge for 4:2:2 / 4:4:4 / 10-bit / 12-bit / 16-bit inputs are largely unexercised at the unit-test layer — regressions there are only caught downstream (Python harness or cross-backend parity gate).

Method¶

1. Enumerate every test that references a non-default pixel format or bit depth:

grep -rIn -E '(VMAF_PIX_FMT_YUV4|yuv4(22|44)|10le|12le|16le)' core/test
grep -rIn -E 'vmaf_picture_alloc\([^,]+,[^,]+,\s*(10|12|16)' core/test

1. For each registered CPU extractor (psnr, ssim, ms_ssim, ciede, cambi, psnr_hvs, adm, vif, motion), assess whether any existing test reaches its init / extract path with each of {YUV422P, YUV444P} × {8, 10, 12, 16}.

Findings¶

Coverage matrix before this PR (CPU extractor surface)¶

Notes:

• "✓ smoke" = at least one end-to-end smoke test wired into the fast suite at the extractor's public surface.
• "math only" = inner helper exercised via #include of the .c file (test_ciede.c only calls ciede2000(LABColor, ...)).
• "(full-model only)" = exercised by test_pic_preallocation::test_picture_pool_yuv444 which runs the full vmaf_v0.6.1 model on YUV444P 8-bit but does not isolate a single extractor.
• "(test_picture only)" = referenced by an allocation-shape test that does not run an extractor.

Gap intersections by severity¶

S0 — High-value, low-cost (this PR):

1. PSNR on 4:2:2 8-bit — exercises the ss_hor=1, ss_ver=0 chroma-stride path; no other CPU test runs an extractor on YUV422P input.
2. PSNR on 4:4:4 10-bit — exercises HBD scoring with chroma plane dimensions equal to luma; the only existing HBD smoke uses 4:2:0 16-bit at 2×2.
3. PSNR on 4:2:0 12-bit — fills the 12-bit hole (the most common HDR bit depth in production pipelines).
4. SSIM on 4:2:2 8-bit — exercises asymmetric chroma geometry under the integer SSIM decimation / window machinery.
5. CIEDE on 4:2:2 8-bit — exercises ciede.c::init chroma upscaling (scale_chroma_planes + scratch YUV444 picture allocation) that the 4:4:4 fast-path bypasses entirely.

S1 — Worth filing as follow-ups:

• MS-SSIM at 10-bit / 12-bit.
• ADM at 10-bit (HDR pipelines feed the model with HBD ADM).
• VIF at 10-bit.
• psnr_hvs on 4:4:4 (no chroma path, but the JPEG-DCT block loop should be smoked).
• motion / motion_v2 at 10-bit (allocates 16-bit blur scratch pictures regardless of input bpc, but the upcast path isn't smoked).

S2 — Specialised; leave alone:

• cambi already covers HBD via the 10-bit YUV400 tests in test_cambi.c.
• tad, lpips, dists, transnet_v2, mobilesal, fastdvdnet_pre, ssimulacra2 — research-grade extractors; pix-fmt edge testing is lower priority than fork-original audits already on backlog.

Decision¶

Plug the S0 cluster in one fast test (test_pixel_format_edge_coverage.c, 5 cases, ~40 ms wall-clock combined). S1 / S2 can land as follow-up bundles using the same pattern.

Reproducer¶

git worktree add -b test/pixel-format-edge-coverage \
    /tmp/wt-pixfmt origin/master
cd /tmp/wt-pixfmt
meson setup core/build-cpu core \
    -Denable_cuda=false -Denable_sycl=false \
    -Denable_hip=false -Denable_metal=disabled
ninja -C core/build-cpu test/test_pixel_format_edge_coverage
core/build-cpu/test/test_pixel_format_edge_coverage
# → "5 tests run, 5 passed"
meson test -C core/build-cpu --suite=fast
# → "Ok: 50  Fail: 0"