Skip to content

Research-0089: Upstream sync coverage — 2026-05-08

  • Status: Draft
  • Date: 2026-05-08
  • Author: @Lusoris
  • Tags: upstream-sync, port, fork-local

Question

Where does the lusoris fork stand vs Netflix/vmaf upstream as of 2026-05-08 — what's recently been added upstream that we don't have, what's already silently ported, what's noise?

Method

Per the /sync-upstream skill's port-only-topology procedure:

  1. Pre-flight: git merge-base master upstream/master returns empty → port-only confirmed (expected for the fork's cherry-pick-then-squash workflow). No bare merge attempted.
  2. Pass 1 (subject match): scan the last 50 upstream commits; classify as PORTED if the subject line appears verbatim on fork master. Result: 0 / 49 ported by subject. That's expected: the fork's port-history uses different conventions (feat: port upstream X (...)), not a verbatim copy of the upstream subject.
  3. Pass 2 (content-hash similarity): for each UNPORTED commit, extract the discriminating identifiers it introduces (function defs, type defs, macros — filter out common keywords) and probe fork master for their presence. ≥ 80 % present → PORTED-SILENTLY. 50–80 % → PARTIAL. < 50 % → genuinely UNPORTED. < 3 idents in the commit → SKIP-doc-or-format (heuristic limit; needs human triage).

Findings

Status Count Action
UNPORTED 2 Real gaps; recommend /port-upstream-commit <sha>.
PARTIAL (50–80 %) 5 Likely incremental optimisations on top of code we already have; eyeball before porting.
PORTED-SILENTLY 1 Already in fork. Commit subject in our master doesn't cite the upstream SHA — maintainer can decide whether to backfill the citation.
SKIP-doc-or-format 41 Heuristic limit; needs human triage by file area. Mix of motion_v2 option additions, cambi SIMD refactors, python-test cleanup, doc updates, merge commits.
Total surveyed 49

UNPORTED — recommend /port-upstream-commit

  • 984f281f — feature/cambi: fuse sliding-window subtract+add into uh_slide/uh_slide_edge
  • 9fad7317 — feature/cambi: factor 2D SAT recurrence into 1D row prefix-sum + column add

Both are cambi performance optimisations. Cambi has CUDA + SYCL + SIMD twins on the fork, so each /port-upstream-commit call must either update every twin or list the gap in Known follow-ups (per CLAUDE.md §12 r12 + ADR-0141).

PARTIAL — incremental on top of present code

  • 1091b0c1 — feature/cambi: add decimate_avx2
  • 933cccb4 — feature/cambi: frame-level calc_c_values dispatch
  • 767a6780 — feature/cambi: refactor calculate_c_values_row, add calculate_c_values_row_avx2
  • bd278ea6 — feature/cambi: add filter_mode_avx2
  • 41bacc83 — feature/cambi: move shared code to cambi.h

These add new AVX2 paths or refactor for the AVX2 dispatch. The fork already has feature/x86/cambi_avx2.c (different content) so each needs careful three-way merge. Recommend porting after the UNPORTED two.

PORTED-SILENTLY — already in fork

  • c17dd898 — libvmaf/motion_v2: add motion_max_val

Fork master has motion_max_val semantics but doesn't cite this SHA in its commit subject. Optional cleanup: backfill the citation in the fork's commit message, or note the silent port in docs/rebase-notes.md.

SKIP-doc-or-format — needs human triage

41 commits the heuristic couldn't classify because they don't introduce new top-level identifiers. The heuristic is the limit here, not the commits' importance. A few are substantive:

  • 4e469601 — libvmaf/motion_v2: port remaining options — option plumbing that almost certainly isn't on the fork yet (we ported motion_v2 selectively).
  • 856d3835 — libvmaf/motion_v2: fix mirroring behavior, since a44e5e61 — a real bug fix on a feature we ship.
  • a2b59b77 — libvmaf/motion_v2: add motion_five_frame_window — new option; needs porting if we want feature parity.
  • 77474251, 8c60dc9e, d655cefe, 721569bc — cambi refactors / option additions / docs.
  • 662fb9ce — python: replace polling-based workfile synchronization with semaphores — fork's python/vmaf/workspace/ diverges, so this needs a careful three-way before porting.
  • The python/test/MyTestCase series (10+ commits) — Netflix's test harness adoption. Touches files the fork's Netflix-golden-data gate pins. Port carefully — the assertion values are immutable.

The remaining ~20 SKIP rows are Merge pull request #... commits that flatten the upstream PR boundaries; not portable in isolation.

Recommendation

Port in this order, one PR per /port-upstream-commit invocation (per the skill's guidance — bulk merge is forbidden under port-only topology):

  1. 856d3835 (motion_v2 mirroring fix) — bug fix, smallest blast radius, validates the workflow.
  2. a2b59b77 (motion_five_frame_window) — new option; trivial to wire if the predecessor commits are already present.
  3. 4e469601 (motion_v2 port remaining options) — option-API surface; affects libvmaf public API. ADR-0186 ffmpeg-patches replay is required (§14 hard rule).
  4. 984f281f + 9fad7317 (cambi UNPORTED algorithmic optimisations) — paired; must update CUDA / SYCL / SIMD twins.
  5. PARTIAL cambi commits (5) — only after the two UNPORTED ones above land, because they build on the same dispatch table.

Defer indefinitely:

  • All 20+ Merge pull request commits — captured by the squash- ports of the underlying changes once those land.
  • Python MyTestCase series — touches Netflix-golden-tied test files; the fork's per-test reformatting cost outweighs the benefit while the golden-gate is the source of numerical truth.

References