Research digest: AVX-512 motion parity test coverage gap (2026-05-29)¶

Summary¶

Prior to this PR, the six AVX-512 motion kernels had no direct unit tests. The gap was identified during the SIMD intrinsics code review initiated in the 2026-05-29 session.

Findings¶

Coverage gap¶

test_motion_v2_simd.c exercised motion_score_pipeline_16_avx2 via the AVX2 dispatch gate. The corresponding AVX-512 functions in motion_v2_avx512.c and motion_avx512.c were not referenced by any test file. The Netflix golden-data gate (python/test/) operates at full-frame granularity and has no ability to isolate a per-kernel regression.

Shift correctness audit¶

The AVX2 path used _mm256_srlv_epi64 (logical right shift) for the >> bpc step in motion_score_pipeline_16_avx2. The AVX-512 twin uses _mm512_srav_epi64 (arithmetic right shift), which matches the scalar C >> on int64_t exactly. The negative-diff adversarial fixture in the new test confirms this is correct on the host.

Scalar reference strategy¶

y_convolution_8, y_convolution_16, x_convolution_16, and sad_c have static linkage in integer_motion.c. Local scalar reimplementations are provided in the test file, mirroring the existing precedent in test_motion_v2_simd.c for the pipeline functions.

Conclusion¶

The ten test cases added by this PR provide direct bit-exactness coverage for all six AVX-512 motion kernels and serve as a permanent regression guard. No kernel-level divergences were found during the audit.