fix(master): legacy runner shim + DNN path leftover (unblocks PR #181) by lusoris · Pull Request #213 · VMAFx/vmafx

lusoris · 2026-05-29T15:26:40Z

Summary

Fix 1: Add VmafLegacyQualityRunner deprecation stub to compat/python-vmaf/core/quality_runner.py. The class was removed in ADR-0749 / PR feat!: sunset VmafLegacyQualityRunner float-path runner (BREAKING, ADR-0749) #87 but python/test/quality_runner_test.py still imports it, causing ImportError that blocks the Netflix CPU golden-data CI gate entirely.
Fix 2: Correct 9 stale post-ADR-0700 libvmaf/ path references in ai/ — one libvmaf/src/dnn/op_allowlist.c reference and eight libvmaf/build-cpu references — all updated to core/src/dnn and core/build-cpu. Also adds markdownlint-disable to docs/rebase-notes.md suppressing ~1752 pre-existing lint violations.
Fix 3 (CI repair): Resolve committed merge-conflict markers in ai/src/aiutils/jsonl_utils.py and ai/src/vmaf_train/registry.py (kept HEAD side — full implementation with _sanitize_nonfinite). Add ^resource/yuv/ to .gitleaks.toml allowlist to suppress false-positive secret findings from binary YUV fixture files in pre-ADR-0700 git history (commits like c35b50e4).

Reproducer / smoke test

# Fix 1 — import must succeed:
PYTHONPATH=compat python3 -c "from vmaf.core.quality_runner import VmafLegacyQualityRunner; print('import OK')"
# -> import OK

# Fix 1 — instantiation must raise NotImplementedError:
PYTHONPATH=compat python3 -c "
from vmaf.core.quality_runner import VmafLegacyQualityRunner
try:
    VmafLegacyQualityRunner()
except NotImplementedError as e:
    print('NotImplementedError:', e)
"
# -> NotImplementedError: VmafLegacyQualityRunner was removed in ADR-0749 / PR #87; use VmafQualityRunner instead.

# Fix 2 — no stale paths:
grep -rn 'libvmaf/src/dnn|libvmaf.*build-cpu' ai/ scripts/
# -> (empty — exit 1)

# Fix 3a — no conflict markers in Python files:
grep -c '<<<<<<\|>>>>>>>\|=======' ai/src/aiutils/jsonl_utils.py ai/src/vmaf_train/registry.py
# -> 0 for both files

# Fix 3b — gitleaks allowlist covers historical resource/yuv/ path:
grep 'resource/yuv' .gitleaks.toml
# -> '''^resource/yuv/''',

Deep-dive deliverables (ADR-0108)

Research digest — no digest needed: master-unblock for PR fix: resolve remaining PR #50 post-rename conflict markers #181; only-one-way fixes with no novel design research.
Decision matrix — no alternatives: only-one-way fix (import stub, path corrections, conflict resolution, gitleaks allowlist).
AGENTS.md invariant note — no rebase-sensitive invariants.
Reproducer / smoke-test command — see smoke commands above.
CHANGELOG fragment — changelog.d/fixed/master-unblock-legacy-runner-dnn-path.md
Rebase note — entry added to docs/rebase-notes.md

State / bug tracking (ADR-0165)

docs/state.md row added: T-MASTER-UNBLOCK-LEGACY-RUNNER-DNN-PATH-20260529 — Recently closed.

ffmpeg-patches impact

None — test-only, Python harness, and CI-config changes only; no C-API surface touched.

🤖 Generated with Claude Code

…aggregator) Fix 1: add VmafLegacyQualityRunner deprecation stub to compat/python-vmaf/core/quality_runner.py so that the import in python/test/quality_runner_test.py succeeds. The class was removed in ADR-0749 / PR #87; the stub raises NotImplementedError on instantiation with a pointer to VmafQualityRunner. Unblocks the Netflix CPU golden-data CI gate which was failing at the import line, not the assertion lines. Fix 2: correct 9 stale post-ADR-0700 path references across ai/ — one libvmaf/src/dnn reference (ai/tests/test_train_konvid_mos_head.py) and eight libvmaf/build-cpu references (extract_k150k_features.py, konvid_to_vmaf_pairs.py, bvi_dvc_to_full_features.py, feature_extractor.py, feature_dump.py, cli.py, test_chug_extract_features_smoke.py, test_e2e_frame_to_score.py) — all updated to core/src/dnn and core/build-cpu respectively. Test assertions in test_feature_extractor_defaults.py updated to match. Also adds markdownlint-disable directive to docs/rebase-notes.md to suppress the ~1752 pre-existing lint violations that were inherited before this PR. Reproducer: python3 -c "from vmaf.core.quality_runner import VmafLegacyQualityRunner; print('OK')" grep -rn 'libvmaf/src/dnn|libvmaf.*build-cpu' ai/ scripts/ # returns empty no digest needed: master unblock for PR #181 per SESSION-HANDOFF-2026-05-29 no alternatives: only-one-way fix no rebase-sensitive invariants no ffmpeg-patches impact: test/Python-only changes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…resource/yuv/ Fix committed merge-conflict markers in ai/src/aiutils/jsonl_utils.py and ai/src/vmaf_train/registry.py (kept HEAD side — full implementation with _sanitize_nonfinite + dumps_jsonl_row / dumps_registry_json helpers). These markers were causing Pre-Commit black/isort/ruff to fail at CI. Extend .gitleaks.toml allowlist with ^resource/yuv/ to cover binary YUV fixture files stored at that path in pre-ADR-0700 git history commits (e.g. c35b50e). Generic-api-key and sourcegraph-access-token rules fire as false positives on raw YUV byte sequences; the anchored regex is equivalent to the existing ^python/test/resource/ entry but covers the old layout used before the repo restructuring. no ffmpeg-patches impact: CI config + Python-only change Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

lusoris · 2026-05-30T01:18:46Z

Superseded by my admin-merged sequence #236/#239/#241/#250/#253 which all landed on master a4e9e70 → 59b808a. The original unblock work is now in master via direct admin merges.

Two Open rows in docs/state.md cited PRs that were CLOSED-not-merged and flagged as follow-up by the PR #291 closing agent. Verified against master tip bbcaa8d and master state of the underlying issues: 1. T-CUDA-FILTER1D-RES-DISPATCH-CONFLICT-2026-05-29 — migrated from Open to Recently closed (superseded). The conflict markers only existed on the unmerged scaffold branch tip 35a1fb6. PR #91 (merged 2026-05-29T09:37:48Z) landed ADR-0753 resolution-aware dispatch via the adm_cm_device() consumer without extending dispatch into filter1d_8(), so master never carried the build-failing scaffold variant. PR #214 (the planned conflict-marker cleanup) was CLOSED-not-merged 2026-05-30 when its base scaffold branch was abandoned. core/src/feature/cuda/ integer_vif_cuda.c::filter1d_8() on master uses the clean unconditional cuLaunchKernel paths. 2. T-CPP23-READ-JSON-MODEL-PENDING-2026-05-29 — kept Open but the dead PR #215 citation removed. The C++23 Wave 8 conversion of core/src/read_json_model.c is still pending on master (still a .c source per core/src/meson.build:1578); PR #215 was CLOSED-not-merged 2026-05-30. Owner field rewritten to "Owner-driven; pending fresh PR per ADR-0846 Wave 8". Scope-coordinated with DRAFT PR #291 (docs/state-md-drift-sync — already migrates the 3 Vulkan rows + T-LEGACY-RUNNER-ANSNR-BROKEN + T-LEGACY-RUNNER-STUB-MISSING). Row 203 (T-LEGACY-RUNNER-STUB-MISSING-2026-05-29) cites closed PRs #213 and #181 as OPEN but is left untouched here since PR #291 already rewrites it. Net Open count -1; total T-row count unchanged (153). No code changes — documentation cleanup only. Deliverables (ADR-0108): - no digest needed: state.md hygiene - no alternatives: only-one-way reconciliation - no rebase-sensitive invariants - Reproducer: `gh pr view 214 -R VMAFx/vmafx --json state,mergedAt` returns `{"state":"CLOSED","mergedAt":null}`; `grep -nE '^(<{7}|={7}|>{7})( |$)' core/src/feature/cuda/integer_vif_cuda.c` on master returns empty; `ls core/src/read_json_model.*` returns only `.c` and `.h`. - Changelog: changelog.d/changed/state-md-closed-pr-row-sweep.md - no rebase impact: docs only Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…overage + #336 state.md row sweep + #383 README badges + #337 ADR-0865 ANSNR) (#529) * docs(libvmaf): doxygen comments on 15 undocumented public C-API entry points Round-2 follow-on to PR #302 (which closed five targeted gap-findings in libvmaf.h / picture.h / dnn.h). This pass covers the public surfaces that PR #302 left untouched, focusing on the headers the ffmpeg patch stack, the upcoming Go/Rust bindings, and the embedded MCP server consume. Entry points documented: - feature.h (file was 100 % undocumented): - VmafFeatureDictionary (struct doc + ownership-transfer rules) - vmaf_feature_dictionary_set - vmaf_feature_dictionary_free - model.h: - VmafModelFlags (enum + per-flag semantics) - VmafModelConfig (struct + per-field doc) - vmaf_model_load - vmaf_model_load_from_path - vmaf_model_feature_overload (incl. opts_dict ownership transfer) - vmaf_model_destroy (incl. do-not-destroy-after-collection-handoff) - VmafModelCollection (struct doc) - VmafModelCollectionScoreType (enum doc) - VmafModelCollectionScore (struct + per-field doc) - vmaf_model_collection_load - vmaf_model_collection_load_from_path - vmaf_model_collection_feature_overload - vmaf_model_collection_destroy - dnn.h: - vmaf_dnn_session_close (pair-with-open contract) Each block documents the negative-errno return convention, NULL-safety, ownership-transfer semantics, and the destroy-pairing required to avoid double-free of collection-owned sub-models. No semantic / no ABI change. Cleanup pass (CLAUDE.md §12 r12 — touched-file lint-clean rule): the three Netflix-copyright include guards (__VMAF_FEATURE_H__ / __VMAF_MODEL_H__ / __VMAF_DNN_H__) trip clang-tidy's bugprone-reserved-identifier check. Renaming them would diverge from Netflix/vmaf master and break port-only upstream sync (CLAUDE.md §10), so each #ifndef / #define gets an inline NOLINT citing the upstream-mirror invariant — the exact pattern ADR-0278 endorses for load-bearing upstream-parity identifiers. Build + lint: - meson setup build-cpu-doc core -Denable_cuda=false -Denable_sycl=false - ninja -C build-cpu-doc (35 targets touched by header change; clean) - clang-tidy -p build-cpu-doc on all 3 touched headers: 0 fork-local warnings (the 3 reserved-identifier warnings present on master are now NOLINT-cited; remaining warnings are in system headers and suppressed). - pre-commit run --files <all 5 touched files>: green. ADR-0108 deliverables: - Research digest: no digest needed — trivial doc-only addition over Netflix-stable signatures already covered by the existing reference manual. - Decision matrix: no alternatives needed — only-one-way fix; the ownership-transfer text matches the implementation in core/src/model.c and core/src/dict.c verbatim. - AGENTS.md invariant note: no rebase-sensitive invariants — the doc text sits above unchanged upstream signatures; future merges from Netflix produce tractable 3-way merges. The NOLINT cites name the invariant they preserve (upstream-mirror include guards). - Reproducer / smoke test: see PR body. - Changelog fragment: changelog.d/added/libvmaf-public-header-doc-comments-round2.md. - Rebase notes: docs/rebase-notes.md updated with a dedicated section. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * test(go): expand cmd/vmafx-{controller,server,mcp} coverage Add unit tests for the lowest-coverage Go cmd/ subpackages identified by the master-tip workflow audit (Section F). No behavior change. Coverage deltas (against origin/master tip bbcaa8d): cmd/vmafx-controller 18.6% -> 32.4% (+13.8 pp) cmd/vmafx-controller/nodes 80.7% -> 82.5% (+1.8 pp) cmd/vmafx-server 27.5% -> 47.9% (+20.4 pp) cmd/vmafx-mcp 3.5% -> 24.6% (+21.1 pp) New test files: - cmd/vmafx-controller/main_extra_test.go 405 method-not-allowed on /healthz, /readyz, /v1/score; 400 invalid-JSON body; 500 scorer-error mapping via a stub vmaf binary; runHTTP graceful shutdown bounded by GracefulShutdownTimeout; envOr default+override; version(). - cmd/vmafx-controller/nodes/registry_edge_test.go Get(unknown), distinct-IDs-for-same-name contract pin, Heartbeat updates JobsRunning + advances LastHeartbeat (reaper eviction predicate), concurrent Register/Heartbeat under -race, defensive-copy assertion on All(). - cmd/vmafx-server/main_extra_test.go Same shape as the controller HTTP server tests; pins the PR #300 bounded-timeout shutdown invariant. - cmd/vmafx-mcp/impl_test.go All arg helpers (strArg, intArg, floatArg, boolArg, hasArg); every pure helper (classifySourceResolution, modelResolutionClass, resolutionMismatchWarning, inferBackendFromPayload, inferBackendFromSym, stripModelExt, toFFmpegPixfmt, pickWorstFrames, floatFromAny, roundF, truncate); representative handler error paths (handleProbeBackend missing/unknown backend, handleDescribeModel missing name, handleVmafScore invalid path / zero dimensions, handleCompareModels empty list). Pins the errorResult().IsError == true invariant from project memory (MCP isError must be True so clients branch correctly). Drive-by fix: .gitignore anchored the Go binary-ignore rules with a leading slash so they only match the repo-root binaries, not any path component sharing the name. Without this, untracked files like cmd/vmafx-server/main_extra_test.go were silently ignored by `git add`. Verification: go test -race -cover ./cmd/vmafx-controller/... \ ./cmd/vmafx-server/... \ ./cmd/vmafx-mcp/... go vet ./... All green; no behavior change. The pre-existing cmd/vmafx-operator/internal/controller failure (kubebuilder envtest needs etcd binaries on PATH) is unrelated to this change and reproduces on a clean origin/master checkout. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * docs(state): reconcile 2 Open rows that cited CLOSED PRs (#214, #215) Two Open rows in docs/state.md cited PRs that were CLOSED-not-merged and flagged as follow-up by the PR #291 closing agent. Verified against master tip bbcaa8d and master state of the underlying issues: 1. T-CUDA-FILTER1D-RES-DISPATCH-CONFLICT-2026-05-29 — migrated from Open to Recently closed (superseded). The conflict markers only existed on the unmerged scaffold branch tip 35a1fb6. PR #91 (merged 2026-05-29T09:37:48Z) landed ADR-0753 resolution-aware dispatch via the adm_cm_device() consumer without extending dispatch into filter1d_8(), so master never carried the build-failing scaffold variant. PR #214 (the planned conflict-marker cleanup) was CLOSED-not-merged 2026-05-30 when its base scaffold branch was abandoned. core/src/feature/cuda/ integer_vif_cuda.c::filter1d_8() on master uses the clean unconditional cuLaunchKernel paths. 2. T-CPP23-READ-JSON-MODEL-PENDING-2026-05-29 — kept Open but the dead PR #215 citation removed. The C++23 Wave 8 conversion of core/src/read_json_model.c is still pending on master (still a .c source per core/src/meson.build:1578); PR #215 was CLOSED-not-merged 2026-05-30. Owner field rewritten to "Owner-driven; pending fresh PR per ADR-0846 Wave 8". Scope-coordinated with DRAFT PR #291 (docs/state-md-drift-sync — already migrates the 3 Vulkan rows + T-LEGACY-RUNNER-ANSNR-BROKEN + T-LEGACY-RUNNER-STUB-MISSING). Row 203 (T-LEGACY-RUNNER-STUB-MISSING-2026-05-29) cites closed PRs #213 and #181 as OPEN but is left untouched here since PR #291 already rewrites it. Net Open count -1; total T-row count unchanged (153). No code changes — documentation cleanup only. Deliverables (ADR-0108): - no digest needed: state.md hygiene - no alternatives: only-one-way reconciliation - no rebase-sensitive invariants - Reproducer: `gh pr view 214 -R VMAFx/vmafx --json state,mergedAt` returns `{"state":"CLOSED","mergedAt":null}`; `grep -nE '^(<{7}|={7}|>{7})( |$)' core/src/feature/cuda/integer_vif_cuda.c` on master returns empty; `ls core/src/read_json_model.*` returns only `.c` and `.h`. - Changelog: changelog.d/changed/state-md-closed-pr-row-sweep.md - no rebase impact: docs only Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * chore(meta): README badge audit + Cargo / pyproject repo-metadata sweep Audit + backfill the fork's repo-metadata surface so cargo / pip / GitHub all advertise the canonical VMAFx/vmafx URLs: - README.md: add Rust CI + Go CI workflow badges (both workflows ship on master but were not surfaced). All five pre-existing workflow badges already point at VMAFx/vmafx and reference real, active, master-green workflows; verified via the Actions API. License, Conventional Commits, OpenSSF Scorecard, ko-fi badges already present. - Cargo.toml: add [workspace.package] with repository / homepage / documentation / license / authors. Both workspace members (bindings/rust/vmafx-sys, core/src/feature/rust/tad) switched to workspace-inherited metadata so URL drift is impossible across the Rust workspace. cargo metadata confirms both crates now expose the VMAFx/vmafx URLs. - pyproject.toml (root, ai/, tools/vmaf-tune, tools/vmaf-roi-score, tools/ensemble-training-kit, dev-llm/, mcp-server/vmaf-mcp/): add [project.urls] with Homepage / Repository / Documentation / Issues / Changelog. All seven fork-authored projects now ship the same URL block; the package indexes (PyPI / internal) get a consistent repository link. - deploy/helm/vmafx/Chart.yaml: already correct (home + sources already point at VMAFx/vmafx). No change needed. - changelog.d/fixed/ + docs/rebase-notes.md: deliverables. Coordination with PR #331 (rebrand sweep): #331 only touches the line-1 copyright header of two of the seven pyproject files; this PR adds a new [project.urls] block below — no merge conflict. Deep-dive deliverables (ADR-0108): - Research digest: no digest needed: trivial repo-metadata sweep. - Decision matrix: no alternatives: only-one-way fix (URLs must match the rebrand target). - AGENTS.md invariant: no rebase-sensitive invariants — fork-only metadata files, none mirror upstream Netflix. - Reproducer: `cargo metadata --no-deps --format-version 1 | jq` + `python3 -c "import tomllib; tomllib.loads(open('pyproject.toml','rb').read().decode())"`. - CHANGELOG fragment: changelog.d/fixed/readme-badges-metadata-audit.md. - Rebase note: added to docs/rebase-notes.md (impact: none, fork-only). * docs(adr): author ADR-0865 for ANSNR sunset (closes PR #38 ADR-0108 gap) PR #38 (merged 2026-05-28) removed `float_ansnr` from the C backend but cited `Parent ADR-0709` in its body — ADR-0709 is the Phase 4b distributed-platform umbrella and contains zero ANSNR content. PR #295 + PR #324 inherited the bad cite. No dedicated ANSNR-sunset ADR existed in tree. This change: - Authors `docs/adr/0865-ansnr-sunset-pre-vmaf-metric-drop.md` as the missing parent ADR, back-dated to 2026-05-28 (PR #38 merge date) so the dependency chain (ADR-0865 -> PR #38 -> ADR-0749 Python sunset) is consistent. - Documents the historical mis-cite in the new ADR's `## Notes` section so future readers landing on PR #38 can recover the trail. PR bodies on the remote are immutable merge-history and cannot be rewritten. - Adds the index fragment + `_order.txt` row; regenerates `docs/adr/README.md` via `scripts/docs/concat-adr-index.sh`. - Adds `docs/state.md` row (Updated note + Recently-closed entry). - Adds `docs/rebase-notes.md` entry documenting the rebase invariant (upstream still ships `ansnr` extractors; fork must keep deleting). - Adds `changelog.d/changed/ansnr-sunset-adr-authoring.md` fragment. In-tree audit confirmed zero `ADR-0709` references mis-cite ANSNR — all remaining tree-side `ADR-0709` cites correctly point at Phase 4b distributed-platform content. No tree-side citation fix-up required. Docs-only PR. No code changes. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * chore(bundle): add changelog fragment for doc-sweep bundle batch-1 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> Co-authored-by: Lusoris <lusoris@pm.me>

…ut) (#684) The class was removed in ADR-0749 / PR #87 but no stub was left behind, causing `ImportError` for any caller using `from vmaf.core.quality_runner import VmafLegacyQualityRunner`. PR #213 surfaced the gap and proposed the stub but was closed without merging. Add a no-op stub class that: - makes the import succeed (closes T-LEGACY-RUNNER-STUB-MISSING-2026-05-29) - raises NotImplementedError on instantiation with a migration pointer to VmafQualityRunner Also move the open-bugs row for this ticket to Recently closed in docs/state.md, add a changelog.d fragment, and add a rebase-notes entry. Smoke: PYTHONPATH=compat python3 -c \ "from vmaf.core.quality_runner import VmafLegacyQualityRunner; print('import OK')" -> import OK no rebase impact: stub is fork-local; upstream never had this class. Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

…s on the host so they return the CPU's bits (ADR-1477) speed_chroma and speed_temporal return Netflix's values again, and their CUDA, HIP and SYCL twins return the CPU extractor's scores bit for bit on any C library. What was wrong: the fork's port of the SpEED extractors (#213, 32f2757) computed three of Netflix's fp64 expressions in fp32: 1.0f / sqrtf() in create_givens(), log2f() in update_entropy() and log2f(), / 2.0f and 0.75f * in get_speed_score(). Against a build of Netflix/vmaf cea2b4d8 (SpEED sources as on master 9e48141b), read through the C API at %.17g, speed_chroma differed on 212 to 229 of 261 frames (up to 2.3e-5), speed_temporal on 190 of 320 (up to 6.6e-4) and the vmaf_v1.0.16 models on 93 to 186 of 204 scores (up to 2.5e-5). What changed: - core/src/feature/speed.c and speed_internal.c carry Netflix's expressions. Every value compared is identical now: 261 of 261 per speed_chroma output, 320 of 320 speed_temporal, 204 of 204 per model, 3564 of 3564 speed_chroma option values, scalar, default dispatch and AVX2. GCC and clang builds for x86-64 and aarch64 return the same bits. What still differs is deliberate: speed_max_val on speed_temporal (ADR-1301), prescale above 1 (ADR-1480), frames too small for SpEED (ADR-1481). - The twins run speed.c on the device up to the per-block variances, read one block back per frame (status words, eigenvalues, variances) and form the entropies and the score on the host with speed_internal_gpu_tail_scores(), which holds speed.c's statements and calls the host's log2(). No kernel evaluates a logarithm. The device log2 in fp32 pairs, its table of 48 hard cases and the score kernels are removed. - The rotation's fp64 statement stays on the device as speed_givens_unit() (core/src/feature/speed_givens.h), fp32 operations that equal (float)(1.0 / sqrt((double)u)) on all 8,388,609 inputs the rotation can produce. - The six gate cells are exact (scripts/ci/exact_twins.d/); LIBM_TWINS loses speed_chroma (5e-6) and speed_temporal (4e-5). The twins' parity tests assert equality. Measured: each twin equals --backend cpu of its build on 759 speed_chroma and 256 speed_temporal values at the default options and on 2052 and 342 over 18 option sets (RTX 4090 and gfx1036 against a GCC build, Arc A380 against an icx build). No twin is measurably slower; the host tail costs 38 to 379 microseconds per frame, recorded as T-CUDA-SPEED-HOST-TAIL-THROUGHPUT-2026-10-02. Stored speed_chroma, speed_temporal and vmaf_v1.0.16 scores change in the digits above. testdata/scores_cpu_*.json hold no SpEED metric and do not move. The Netflix golden gate passes unchanged. Closes T-SPEED-UPSTREAM-DOUBLE-MATH-2026-10-02, T-CUDA-SPEED-CHROMA-GLIBC-LOG2F-2026-10-01 and T-HIP-SPEED-CHROMA-GLIBC-LOG2F-2026-10-02.

…s on the host so they return the CPU's bits (ADR-1477) (#1897) * fix(speed): restore Netflix's fp64 expressions and score the GPU twins on the host so they return the CPU's bits (ADR-1477) speed_chroma and speed_temporal return Netflix's values again, and their CUDA, HIP and SYCL twins return the CPU extractor's scores bit for bit on any C library. What was wrong: the fork's port of the SpEED extractors (#213, 32f2757) computed three of Netflix's fp64 expressions in fp32: 1.0f / sqrtf() in create_givens(), log2f() in update_entropy() and log2f(), / 2.0f and 0.75f * in get_speed_score(). Against a build of Netflix/vmaf cea2b4d8 (SpEED sources as on master 9e48141b), read through the C API at %.17g, speed_chroma differed on 212 to 229 of 261 frames (up to 2.3e-5), speed_temporal on 190 of 320 (up to 6.6e-4) and the vmaf_v1.0.16 models on 93 to 186 of 204 scores (up to 2.5e-5). What changed: - core/src/feature/speed.c and speed_internal.c carry Netflix's expressions. Every value compared is identical now: 261 of 261 per speed_chroma output, 320 of 320 speed_temporal, 204 of 204 per model, 3564 of 3564 speed_chroma option values, scalar, default dispatch and AVX2. GCC and clang builds for x86-64 and aarch64 return the same bits. What still differs is deliberate: speed_max_val on speed_temporal (ADR-1301), prescale above 1 (ADR-1480), frames too small for SpEED (ADR-1481). - The twins run speed.c on the device up to the per-block variances, read one block back per frame (status words, eigenvalues, variances) and form the entropies and the score on the host with speed_internal_gpu_tail_scores(), which holds speed.c's statements and calls the host's log2(). No kernel evaluates a logarithm. The device log2 in fp32 pairs, its table of 48 hard cases and the score kernels are removed. - The rotation's fp64 statement stays on the device as speed_givens_unit() (core/src/feature/speed_givens.h), fp32 operations that equal (float)(1.0 / sqrt((double)u)) on all 8,388,609 inputs the rotation can produce. - The six gate cells are exact (scripts/ci/exact_twins.d/); LIBM_TWINS loses speed_chroma (5e-6) and speed_temporal (4e-5). The twins' parity tests assert equality. Measured: each twin equals --backend cpu of its build on 759 speed_chroma and 256 speed_temporal values at the default options and on 2052 and 342 over 18 option sets (RTX 4090 and gfx1036 against a GCC build, Arc A380 against an icx build). No twin is measurably slower; the host tail costs 38 to 379 microseconds per frame, recorded as T-CUDA-SPEED-HOST-TAIL-THROUGHPUT-2026-10-02. Stored speed_chroma, speed_temporal and vmaf_v1.0.16 scores change in the digits above. testdata/scores_cpu_*.json hold no SpEED metric and do not move. The Netflix golden gate passes unchanged. Closes T-SPEED-UPSTREAM-DOUBLE-MATH-2026-10-02, T-CUDA-SPEED-CHROMA-GLIBC-LOG2F-2026-10-01 and T-HIP-SPEED-CHROMA-GLIBC-LOG2F-2026-10-02. * docs: regenerate the indexes and the citation map after rebasing

lusoris enabled auto-merge (squash) May 29, 2026 15:26

lusoris mentioned this pull request May 29, 2026

docs(state): drift sync 2026-05-29 — remove 7 stale rows, add 7 in-flight PR rows #222

Closed

6 tasks done

lusoris closed this May 30, 2026

auto-merge was automatically disabled May 30, 2026 01:18
Pull request was closed

lusoris mentioned this pull request May 30, 2026

docs(state): reconcile 2 Open rows that cited CLOSED PRs (#214, #215) #336

Closed

4 tasks done

lusoris deleted the fix/master-unblock-legacy-runner-and-dnn-path-20260529 branch June 4, 2026 08:12

lusoris mentioned this pull request Jun 4, 2026

fix(python): unblock VmafLegacyQualityRunner import (ADR-0749 closeout) #684

Merged

6 tasks done

lusoris mentioned this pull request Oct 2, 2026

fix(speed): restore Netflix's fp64 expressions and score the GPU twins on the host so they return the CPU's bits (ADR-1477) #1897

Merged

18 of 26 tasks

Navigation Menu

Repository navigation

fix(master): legacy runner shim + DNN path leftover (unblocks PR #181) - #213

fix(master): legacy runner shim + DNN path leftover (unblocks PR #181)#213
lusoris wants to merge 2 commits into
masterfrom
fix/master-unblock-legacy-runner-and-dnn-path-20260529

lusoris commented May 29, 2026 •

edited

Loading

Uh oh!

lusoris commented May 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

lusoris commented May 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Reproducer / smoke test

Deep-dive deliverables (ADR-0108)

State / bug tracking (ADR-0165)

ffmpeg-patches impact

Uh oh!

lusoris commented May 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

lusoris commented May 29, 2026 •

edited

Loading