Skip to content

ADR-0158: Netflix#1486 "Port motion updates" — verified present in fork

  • Status: Accepted
  • Date: 2026-04-24
  • Deciders: Lusoris, Claude (Anthropic)
  • Tags: upstream-port, motion, netflix-upstream, verification

Context

Backlog item T4-1 called for porting Netflix upstream PR #1486 "Port motion updates", merged upstream on 2026-04-20 as two commits:

  1. a44e5e6 — libvmaf/feature: port motion updates, bugfix for edge mirroring. Adds motion_max_val option (default 10000.0; clips motion scores above), fixes the edge_8 mirroring bug (height - (i_tap - height + 1) → height - (i_tap - height + 2)) across the scalar + AVX2 + AVX-512 paths, adds VMAF_integer_feature_motion3_score output in extract_force_zero, and moves SIMD dispatch out of init.

  2. 62f47d5 — loosen assertion precision for motion mirror bugfix. Netflix golden assertAlmostEqual updates across python/test/{feature_extractor_test.py,quality_runner_test.py, vmafexec_feature_extractor_test.py,vmafexec_test.py} — the numerical-gate coordinated with commit 1.

A verification pass confirmed that both commits are already fully present in the fork's master:

The fork's port didn't land as a single PR matching Netflix#1486 1:1. Fork PR #45 (9371a0aa feat(libvmaf/feature): port upstream motion updates (Netflix PR #1486)) was opened but never merged; its substance was folded into master incrementally via later motion3 / blend / five-frame-window / moving-average commits that went through separate PRs. The end state is equivalent: every line the upstream patch would add is in the fork today.

Why this matters: T4-1 on the backlog was sized M (motion metric math + cross-backend-diff + Netflix golden gate). Re-porting would be a no-op at best and a conflict-resolution exercise at worst. The verification + paper trail is the right closing move.

Decision

Close T4-1 as "verified present; no port PR needed." This ADR + a rebase-notes entry document the coverage so that future /sync-upstream runs don't re-discover Netflix#1486 as pending.

No code change. No CHANGELOG entry (no user-visible delta — everything upstream shipped is already in the fork's released binaries). Backlog item flipped to done with a verification reference.

Alternatives considered

Option Pros Cons Why not chosen
Doc-only close (this ADR) Paper trail for future sync-upstream; no code churn; honest representation of current state Small docs-only PR adds a commit to master Chosen via popup — next /sync-upstream gets the coverage context without re-auditing
Open a port PR anyway Keeps 1:1 commit-to-port mapping across upstream/fork No-op diff; pure churn; conflict resolution with existing motion3 / blend work Rejected — wasted cycles
In-session close (no PR) Zero commits No tracked paper trail; future sync-upstream re-checks #1486 from scratch Rejected via popup
Supersede fork PR #45 Would close the unmerged PR explicitly PR #45 is already stale (last activity >6 months); GitHub auto-closes on push reference Not needed — the branch is long-dead

Consequences

  • Positive:
  • /sync-upstream won't re-flag Netflix#1486 as a pending port.
  • Future rebase-audits have a cite for "this upstream PR is already absorbed via fork-native work, no cherry-pick required."
  • Backlog tier T4 shrinks by one.
  • Negative: None.
  • Neutral / follow-ups:
  • Fork-local motion extensions (five-frame-window, moving-average, blend, fps_weight) are additions on top of Netflix#1486; future upstream changes to motion extractor internals may conflict with them. That invariant is noted in core/src/feature/AGENTS.md under the existing motion_v2 NEON entry.
  • The pre-existing readability-braces-around-statements warnings on integer_motion.h:38-48 (noticed during verification) are upstream style carried forward — ADR-0141 would catch them on the next touch of that header. Not a blocker here since nothing changed.

Verification

  • ninja -C build && meson test -C build → 35/35 pass (no delta vs pre-verification master).
  • Markers verified present via:
grep -n "height - (i_tap - height + 2)\|motion_max_val\|VMAF_integer_feature_motion3_score" \
    core/src/feature/integer_motion.c \
    core/src/feature/alias.c \
    core/src/feature/x86/motion_avx2.c \
    core/src/feature/x86/motion_avx512.c
  • 73-triple golden-value match on python tests verified by programmatic AST-style scan (details in subagent report).
  • git status → clean; this PR contributes no code hunks.

References

  • Upstream PR: Netflix/vmaf#1486 ("Port motion updates"), MERGED 2026-04-20.
  • Upstream commits: a44e5e6, 62f47d5 (in Netflix/vmaf@master).
  • Fork PR #45 (never merged; 9371a0aa) — tracked the original port attempt; substance landed via incremental fork-native commits instead.
  • ADR-0024 — Netflix golden gate.
  • ADR-0142 — Netflix-authority carve-out for coordinated golden-value updates; would apply if a real port PR were needed here.
  • ADR-0145 — fork-local motion_v2_neon.c; unrelated to this upstream PR but touches the same file family on rebase.
  • rebase-notes 0051 — sync-upstream invariants.
  • Backlog: .workingdir2/BACKLOG.md T4-1.
  • User direction 2026-04-24 popup: "T4-1 Port motion updates Netflix#1486" → "Doc-only PR (Recommended)".