refactor(simd): bring the integer motion SIMD kernels to the lint and HISS standard (ADR-1142)

[lusoris]

Summary

First of four PRs for the 16 files that still had no SPDX-License-Identifier line because they carried baselined debt. This one covers the integer motion SIMD kernels: core/src/feature/x86/motion_avx2.c, core/src/feature/x86/motion_avx512.c and core/src/feature/arm64/motion_neon.c are at the lint and HISS standard, carry their SPDX line, and no bit of any score moves.

	cpu	cuda	hip	sycl	arm64
x86/motion_avx2.c, clang-tidy	7 → 0	7 → 0	7 → 0	7 → 0	n/a
x86/motion_avx512.c, clang-tidy	18 → 0	18 → 0	18 → 0	18 → 0	n/a
arm64/motion_neon.c, clang-tidy	n/a	n/a	n/a	n/a	2 → 0
lane total in scripts/ci/tidy-baseline-<lane>.json	310 → 285	662 → 637	664 → 639	766 → 741	558 → 556

HISS: ten rows gone (three in motion_avx2.c, six in motion_avx512.c, one in motion_neon.c, all "function over 60 lines"); .standards-baseline.json 242 → 232, recorded with praetorctl baseline --record; the README count follows. Uncited NOLINTs: 0, and the three files have no NOLINT at all.

SPDX: BSD-2-Clause-Patent on all three. scripts/dev/relicense_fork_files.py --list classifies them upstream-path (the files exist in Netflix/vmaf), so they keep Netflix's terms (ADR-1250).

What changed

Each motion_score_pipeline_{8,16}_* keeps its name and signature and is a row loop:

Stage	AVX2	AVX-512
vertical pass of one row; returns whether any y_row value is non-zero	y_conv_row_{8,16}_avx2 over y_conv16_8_avx2 / y_conv8_16_avx2	y_conv_row_{8,16}_avx512 over y_conv32_8_avx512 / y_conv16_16_avx512
horizontal pass + abs + SAD of one row	x_conv_row_sad_avx2 over x_conv_abs8_avx2, x_conv_abs_scalar, hsum_epi32_avx2	x_conv_row_sad_avx512 over x_conv_abs16_avx512, x_conv_abs_scalar

The three test-only kernels of motion_avx512.c (y_convolution_8_avx512, y_convolution_16_avx512, x_convolution_16_avx512) share filter5_epu32_avx512() and filter5_scalar(), and x_convolution_16_neon is three passes over two helpers. sad_avx512 is untouched.

Every statement that computes moved whole: the widths of the accumulators, the order of the widening additions, the rounding constants, the logical _mm256_srlv_epi64 of the AVX2 16-bit path and the arithmetic _mm512_srav_epi64 of the AVX-512 one are as they were. All arithmetic in these files is integer.

The findings, by check:

Check	Count	Fixed by
readability-function-size	10	the split
misc-use-internal-linkage	9	each file includes its own header, so the exported functions are checked against their declarations
readability-isolate-declaration	4	one declaration per line
readability-braces-around-statements	4	the edge loop of y_convolution_8_avx512 is one braced helper

Not one bit moves

• Kernels, old against new in one process. Master's and this PR's objects linked into one harness under different names, compared on generated input (noise, extremes, flat, near-equal pictures; widths 3 to 140 and 576, 1919, 1920, 3840; 8, 10, 12 and 16 bit; the returned SAD and the whole y_row scratch; the destination buffers of the three convolution kernels including their padding): 286 952 comparisons, 0 mismatches, with GCC 16, clang 22 and icx 2026.0. x_convolution_16_neon under qemu-aarch64: 3 920 comparisons, 0 mismatches, with aarch64 GCC and clang.

• Scores. motion, motion=debug=true, motion_v2, float_motion and the default model at --precision max, this PR against master 513d2a6fc:

  Build	Fixtures	Dispatch	Cases identical	Values
  x86-64 GCC	Netflix 576x324 at 8, 10, 12, 16 bit; both 1080p checkerboards; BBB 3840x2160, 8 frames	scalar, AVX2, AVX-512	105 of 105	5 964
  aarch64 GCC under qemu	the same without BBB, 6 frames	scalar, NEON	48 of 48	1 050

• Netflix golden gate. x86-64 GCC: 271 passed, 12 skipped. aarch64 GCC under qemu: 271 passed, 12 skipped (86 minutes; the aarch64 binary was built from this PR's sources on 25fe95d73, one rebase before the pushed head, so master's two later aarch64 changes, refactor(feature): bring the NEON float ADM kernels to the lint and HISS standard (ADR-1142) #1865 and the NEON psnr_hvs fix, are not in that binary). x_convolution_16_neon, the one aarch64 function this PR changes, is called by core/test/test_motion_neon.c only; no extractor reaches it.

• Unit tests. --suite=fast on the CPU build: 244 of 244 (includes test_motion_avx512_parity, test_motion_v2_simd, test_integer_motion_coverage). Under qemu: test_motion_neon and test_motion_pipeline_neon pass.

• Speed. The helpers are static inline and the two AVX2 pipelines are the only symbols of their object file, as before. One process, 1920x1080, minimum of 300 calls, old then new, load 56 on the host:

  	GCC	clang	icx
  pipeline_8_avx2	0.962 → 0.948 ms	0.860 → 0.890	0.891 → 0.929
  pipeline_16_avx2 (10 bit)	1.479 → 1.449	1.345 → 1.395	1.480 → 1.475
  pipeline_8_avx512	0.517 → 0.533	0.565 → 0.519	0.592 → 0.624
  pipeline_16_avx512 (10 bit)	0.742 → 0.755	0.792 → 0.755	0.842 → 0.903

  Between 0.92x and 1.07x with no direction; not a timing claim on a host this loaded.

• GPU twins. No header changes; feature/integer_motion.h is untouched, so the CUDA, HIP and SYCL motion twins compile the same text.

• scripts/dev/preflight.sh --stage msvcism: pass.

Lane notes

• sycl lane: on master the two x86 files cannot be measured there at all (T-SYCL-TIDY-OVERRIDING-OPTION-2026-10-02, fixed by fix(ci): silence clang's fp-override driver note so the SYCL tidy lane measures the x86 SIMD units #1867). The numbers above were taken with that fix's flag (-Wno-overriding-option) passed to clang-tidy.
• Host false positives (misc-static-assert on glibc 2.44): none in these three files.
• All five lanes were measured on this host (cross aarch64-linux-gnu for arm64); nothing was left unmeasured.

Type

• feat — new feature
• fix — bug fix
• perf — performance improvement
• refactor — no behavior change
• docs — documentation only
• test — test-only
• build / ci — tooling / infra
• port — cherry-pick from upstream Netflix/vmaf
• sycl / cuda / simd — backend-specific

Checklist

• Commits follow Conventional Commits (the commit-msg hook enforces this).
• make format && make lint is green locally. (clang-format and the commit hooks; clang-tidy 0 for the three files in every lane that reads them.)
• Unit tests pass: python3 scripts/ci/run_meson_test.py -- -C build. (--suite=fast on a CPU build: 244 of 244.)
• If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2. (Bit-identical: old and new kernels compared directly, and scores under every dispatch.)
• If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap under "Known follow-ups" below. (No arithmetic changes; nothing for a twin to follow.)
• If I added a new .c / .cpp / .cu / .h / .hpp, it has the appropriate license header (see CONTRIBUTING.md). (None added; the SPDX line of the three files added.)
• If this is a breaking change, the commit message uses ! or BREAKING CHANGE: and the migration path is documented below. (Not breaking.)
• If this PR adds an ADR, the ADR row lives in docs/adr/_index_fragments/<NNNN-slug>.md. (No ADR: ADR-1142 is the rule.)

Bug-status hygiene (ADR-0165)

• docs/state.md updated in this PR. no state delta: refactor to the lint and HISS standard; no row names these files' debt.

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, I have explained why below AND pinged @lusoris for a CODEOWNERS exception. (None changes.)

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: the findings and the split are described here and in the rebase note.
• Decision matrix — no alternatives: only-one-way fix. ADR-1142 is the rule; the split follows each pipeline's own stages.
• AGENTS.md invariant note — core/src/feature/x86/AGENTS.d/motion.md, "Pipelines are row loops over inlined stages".
• Reproducer / smoke-test command — pasted below under "Reproducer".
• CHANGELOG fragment — changelog.d/changed/std-motion-simd.md.
• Rebase note — docs/rebase-notes.md, "Integer motion SIMD kernels split into stages".

Reproducer

meson setup /tmp/tidy-cpu core -Denable_cuda=false -Denable_sycl=false -Db_lto=false
ninja -C /tmp/tidy-cpu
python3 scripts/ci/write-compile-commands.py --build-dir /tmp/tidy-cpu
python3 scripts/ci/tidy-ratchet.py --lane cpu --build-dir /tmp/tidy-cpu \
  --only core/src/feature/x86/motion_avx2.c --only core/src/feature/x86/motion_avx512.c   # 0 warnings
praetorctl baseline --verify      # 232 recorded
for mask in 0 48 56; do           # AVX-512, AVX2, scalar
  /tmp/tidy-cpu/tools/vmaf -r python/test/resource/yuv/src01_hrc00_576x324.yuv \
    -d python/test/resource/yuv/src01_hrc01_576x324.yuv -w 576 -h 324 -p 420 -b 8 \
    --no_prediction --feature motion=debug=true --feature motion_v2 --cpumask $mask \
    --precision max --json -o /dev/stdout -q | sha256sum
done
python3 scripts/ci/run_meson_test.py -- -C /tmp/tidy-cpu --suite=fast   # 244 of 244
make test-netflix-golden          # 271 passed, 12 skipped
make test-netflix-golden-arm64    # 271 passed, 12 skipped

Known follow-ups

• The other 13 of my 16 files: the extractor sources and their tests, the CLI tool files, and compat/python-vmaf/__init__.py, one PR each.