Skip to content

ADR-0984: Port Netflix Upstream May–Jun 2026 (5 commits)

Status

Accepted

Date

2026-06-01

Context

Five upstream Netflix/vmaf commits from May–June 2026 were ported to the fork, each path-remapped from libvmaf/ to core/ (per ADR-0700). The commits were applied in chronological order on branch port/upstream-netflix-may-jun-2026.

Commits ported

  1. e4b93c6ed (2026-05-08) — tools/vmaf: enable direct read by default Removes the #ifdef USE_DIRECT_READ compile-time guard from vmaf.c, making video_input_fetch_into_vmaf_picture() the only input path. Supersedes the compile-time flag introduced by ADR-0600.

  2. a4a1492d3 (2026-05-25) — libvmaf: replace integer_motion with pipelined v2 variant, rename Merges the pipelined v2 implementation into integer_motion.c, removes the separate integer_motion_v2.c CPU extractor, and deletes the merged motion_v2_avx2.{c,h} / motion_v2_avx512.{c,h} SIMD files. In this fork, the v2 SIMD files are retained under their original names because the GPU backends (CUDA, SYCL, HIP, Metal) expose vmaf_fex_integer_motion_v2_* extractors that depend on the motion_score_pipeline_{8,16} pipeline. The CPU vmaf_fex_integer_motion_v2 extractor is removed from the feature_extractor_list[]; the GPU variants remain.

  3. c2155d6cd (2026-06-01) — libvmaf: add 2160p@1.5H CSF support Adds adm_ref_display_height == 2160 && adm_norm_view_dist == 1.5 branches to barten_csf_tools.h, routing them to the existing BLENDED_CSF_1080_3H tables (1.52160 == 3.01080 == 56.55 ppd). Adds pinning tests.

  4. 9a078011c (2026-06-01) — ADM: enable adm_decouple_s123 SIMD dispatch and fix AVX2 kh_msb blend Uncomments s->adm_decouple_s123 = adm_decouple_s123_avx2/avx512 in integer_adm.c::init(). Fixes a kh/kv/kd MSB blend bug in the AVX2 path (const_32768 → abs_oh_epi32) that caused small float-feature drift on 10-bit content. Adds matching kh_shift_epi32 zero-clear on the small-value branch in both AVX2 and AVX-512.

  5. 30f472b14 (2026-06-01) — Speed_chroma: vectorize compute_covariance (AVX2 + AVX-512) Adds a compute_cov_kernel_fn function-pointer in SpeedState and dispatches to new compute_cov_kernel_avx2 / compute_cov_kernel_avx512 implementations in new files x86/speed_avx2.{c,h} and x86/speed_avx512.{c,h}. Scalar path retained as fallback.

Decision

All 5 commits ported with path remapping. Fork-local adaptations:

  • Commit #1: copy_picture_data() and finish_unread_picture() helper removed (no longer called). Call sites updated.
  • Commit #2: CPU vmaf_fex_integer_motion_v2 removed from feature_extractor_list[]; GPU _v2_* variants kept. integer_motion_v2.c and motion_v2_avx2/avx512.{c,h} kept for GPU build paths. motion_avx2/avx512.{c,h} updated to the pipeline implementation. Upstream precision reductions in python/test/ golden assertions NOT applied (CLAUDE.md §12 r1 / correctness-first rule).
  • Commit #3–5: applied cleanly.

Alternatives considered

  • Reject commit #2 (integer_motion rename): the GPU backends reference the v2 name; a full rename would require touching all GPU files. Accepted partial port instead.
  • Apply upstream precision relaxations in python/test/: rejected — they weaken the golden assertion gate (CLAUDE.md §12 r1).

Consequences

  • tools/vmaf CLI always uses direct-read input path (no USE_DIRECT_READ flag).
  • vmaf_fex_integer_motion is now the pipelined implementation on CPU.
  • 2160p at 1.5H viewing distance is a valid CSF configuration.
  • ADM adm_decouple_s123 SIMD dispatch is enabled; AVX2 kh_msb bug fixed.
  • Speed_chroma compute_covariance is SIMD-accelerated on x86.

References

  • Upstream commits: e4b93c6ed, a4a1492d3, c2155d6cd, 9a078011c, 30f472b14
  • ADR-0600: previous USE_DIRECT_READ port (now superseded for CPU path)
  • ADR-0138, ADR-0139: SIMD bit-exactness invariants
  • ADR-0700: libvmaf/ → core/ path remap
  • CLAUDE.md §12 r1: never modify Netflix golden assertions