fix(cuda): push context in cambi submit and gate twin selection on model options (ADR-1183) by lusoris · Pull Request #1312 · VMAFx/vmafx

lusoris · 2026-09-05T09:59:42Z

Summary

Fixes the 100% crash on vmaf --backend cuda with the default model (vmaf_v1.0.16_3d0h, ADR-1169). The root cause was twofold: (1) submit_fex_cuda() in core/src/feature/cuda/integer_cambi_cuda.c performed the GPU pipeline and synchronous cuMemcpyDtoH without pushing fex->cu_state->ctx onto the calling worker/batch thread, causing CUDA_ERROR_INVALID_CONTEXT (201); (2) vmaf_fex_ctx_parse_options() silently dropped unrecognized model options, causing integer_adm_cuda (which lacks adm_csf_mode) to run with mode 0 while the model specified mode 2, producing mismatched feature names. This PR pushes/pops the CUDA context around CAMBI submit, adds strict option validation returning -EINVAL for unknown extractor options, implements option-aware GPU twin dispatch (ADR-1183) falling back to the CPU twin when model options are unsupported, routes integer_motion_cuda scores via feature_name_dict to support both motion2 and motion3, and adds unit and integration tests.

2026-09-05 13:20 — history rewritten: the branch was squashed to one commit and force-pushed because gitleaks scans every branch of the repo and the test's adm3_key = "integer_adm3_csf_2_…" / cambi_key = "cambi_hrs_…" lines tripped generic-api-key on every PR (#1284's scan went red on a commit from this branch). The variables are renamed and .gitleaks.toml allowlists VMAF pooled-metric names; the earlier per-commit references above therefore no longer resolve.

Type

fix — bug fix
cuda — backend-specific

Checklist

Commits follow Conventional Commits.
make format && make lint is green locally.
Unit tests pass: meson test -C build.
If I touched any SIMD/GPU code path, I ran /cross-backend-diff and the worst ULP is ≤ 2.
If I touched a feature extractor with SIMD/GPU twins, I either updated every twin or listed the gap under "Known follow-ups" below.
If I added a new .c / .cpp / .cu / .h / .hpp, it has the appropriate license header.
If this PR adds an ADR, the ADR row lives in docs/adr/_index_fragments/<NNNN-slug>.md and the slug is appended to docs/adr/_index_fragments/_order.txt — do not edit docs/adr/README.md directly.

Bug-status hygiene (ADR-0165)

docs/state.md updated: T-CAMBI-CUDA-INVALID-CONTEXT-2026-09-04 and T-GPU-TWINS-IGNORE-MODEL-OPTIONS-2026-09-05 moved to Recently closed; T-GPU-ADM-CSF-MODE-NOT-PORTED-2026-09-05 added to Open.

Netflix golden-data gate (ADR-0024)

I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.

Cross-backend numerical results

Ran CPU vs CUDA comparison on 576x324 Netflix pair (src01_hrc00 vs src01_hrc01) with default model vmaf_v1.0.16_3d0h:

vmaf: CPU = 82.816059, CUDA = 82.814059 (abs delta: 0.002000)
cambi: CPU = 11.233516, CUDA = 11.230853 (abs delta: 0.002663)
adm3: CPU = 0.902996, CUDA = 0.902996 (bit-exact match via CPU fallback per ADR-1183)
motion3: CPU = 8.163351, CUDA = 8.163352 (abs delta: 0.000001)
speed_chroma: CPU = 0.000000, CUDA = 0.000001 (abs delta: 0.000001)

Ran CPU vs CUDA comparison on 1080p Tennis pair (Tennis_24fps.yuv vs Tennis_20_288_375.yuv, 10 frames):

vmaf: CPU = 41.459893, CUDA = 41.445641 (abs delta: 0.014252)
cambi: CPU = 16.518621, CUDA = 16.500987 (abs delta: 0.017634)
adm3: CPU = 0.771923, CUDA = 0.771923 (bit-exact match)
motion3: CPU = 1.636750, CUDA = 1.636762 (abs delta: 0.000012)
speed_chroma: CPU = 0.000000, CUDA = 0.000000 (abs delta: 0.000000)

Deep-dive deliverables (ADR-0108)

Research digest — no digest needed: bug fix and twin option negotiation pattern (ADR-1183).
Decision matrix — captured in docs/adr/1183-model-options-gate-gpu-twin-selection.md § Alternatives considered.
AGENTS.md invariant note — added to core/src/feature/AGENTS.md.
Reproducer / smoke-test command — included below under Reproducer.
CHANGELOG fragment — changelog.d/fixed/cambi-cuda-context.md, changelog.d/fixed/model-options-unknown-key-error.md.
Rebase note — entry added to docs/rebase-notes.md.

Reproducer

# 1. Reproducer: run default model on CUDA (previously failed with CUDA_ERROR_INVALID_CONTEXT in integer_cambi_cuda.c:907)
core/build/tools/vmaf \
  -r python/test/resource/yuv/src01_hrc00_576x324.yuv \
  -d python/test/resource/yuv/src01_hrc01_576x324.yuv \
  -w 576 -h 324 -p 420 -b 8 \
  --backend cuda --model version=vmaf_v1.0.16_3d0h --json

# 2. Pytest suite for CUDA default model
pytest python/test/cuda_default_model_test.py -v

# 3. Unit test for option validation error
meson test -C core/build test_feature_extractor

Known follow-ups

T-GPU-ADM-CSF-MODE-NOT-PORTED-2026-09-05: Implement CSF mode 2 in integer_adm_cuda and integer_adm_sycl so that the default model can run ADM on GPU silicon instead of falling back to CPU.
T-FEATURE-FLAG-IGNORES-BACKEND-2026-09-04: Upstream Netflix behavior is preserved where --feature <name> uses vmaf_get_feature_extractor_by_name() returning CPU extractors; documented in docs/usage/cli.md and docs/backends/cuda/overview.md.

🤖 Generated with Claude Code

Two `Rules` gates caught the previous commit: - `docs/state.md Gate` (ADR-0334) rejects a placeholder PR reference in an inserted row. The `T-TIDY-CHANGED-LTO-FLAG-2026-09-05` row said "This PR was the first C/C++ PR after #1290"; it now names PR #1312. - `Release Script Contract` requires `CHANGELOG.md`'s Unreleased block to match `changelog.d/`. Re-rendered with `scripts/release/concat-changelog-fragments.sh --write`, which picks up `changelog.d/fixed/tidy-changed-lto-flag.md`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…ption validation, CPU-twin dispatch (ADR-1183) - integer_cambi_cuda.c: push/pop the CUDA context around the submit path (the DtoH readback ran with no current context: CUDA_ERROR_INVALID_CONTEXT), mirror the CPU cambi option table (cambi_high_res_speedup/hrs, vlt, cmxv) and the TVI table init from cambi.c so the emitted feature name matches the model's. - integer_motion_cuda.c: route scores through the feature_name_dict so motion3 emits integer_motion3_mmxv_18 like the CPU twin. - feature_extractor.cpp: an opts_dict key the chosen extractor does not declare is an error (-EINVAL, logged) instead of being silently dropped. - libvmaf.c: when a GPU twin lacks an option the model sets, resolve the CPU twin for that feature and log it (ADR-1183) — adm3 of vmaf_v1.0.16_3d0h runs on the CPU under CUDA/SYCL/HIP until the csf_mode port lands (docs/state.md T-GPU-ADM-CSF-MODE-NOT-PORTED-2026-09-05). - python/test/cuda_default_model_test.py (new, skips without CUDA); .gitleaks.toml allowlists VMAF pooled-metric names (the test's `<name>_key = "<metric>"` lines tripped generic-api-key on every PR because gitleaks scans all branches). Verified on the RTX 4090: 576x324 and 1080p reproducers exit 0; adm3 bit-exact vs CPU, cambi delta 2.66e-3 / 1.76e-2 (tracked), v0.6.1 unchanged, typo option rejected; golden gate 271 passed / 12 skipped / 0 failed.

The ADR-1183 GPU-twin fallback pushed vmaf_use_features_from_model past the readability-function-size threshold and the two new test cases in test_feature_extractor.c added eight uncited modernize-use-nullptr findings, so the whole-tree ratchet (ADR-1142) refused the PR. - The fallback moves into fex_honouring_model_options(), a helper that returns the extractor able to honour every option the model sets for one feature (the GPU twin, or the CPU twin with an INFO line, or NULL when nothing provides the feature). vmaf_use_features_from_model is back under the threshold and reads as one loop again. - The new tests' NULL arguments and returns carry the ADR-1138 citation the other C TUs use (MSVC /std:clatest has no C nullptr). - feature_extractor.cpp lost four findings to the option-validation refactor, so its baseline entry is tightened 54 -> 50 with the count CI measured (the local clang is a different version and is not the authority). Behaviour is unchanged: the CUDA default-model run still exits 0 with the CPU dispatch line for adm3, test_feature_extractor is 13/13.

The touched-file cleanup rule (ADR-0141) makes this PR inherit every pre-existing clang-tidy finding in the test TUs it edits. Discharge all 59 in `core/test/test_feature_extractor.c`: * modernize-use-nullptr (54): replace the 17 scattered, partly stale `NOLINTNEXTLINE` markers (several of which sat above comment lines and suppressed nothing) with the single file-scoped `NOLINTBEGIN(modernize-use-nullptr)` / `NOLINTEND` bracket that ADR-1138 prescribes and that `core/src/libvmaf.c`, `core/test/test_output.c` and the other C TUs already use. The TU keeps spelling the null pointer constant `NULL`: the required MSVC C lane has no documented C `nullptr`, and upstream parity matters for the sync story. * readability-function-size (5): split the oversized cases into one behaviour per test instead of suppressing them. - `test_feature_extractor_flush` and `test_feature_extractor_initialization_options` now share a `FexFixture` (context + ref/dist pictures + collector) with `fex_fixture_open` / `fex_fixture_close` / `fex_fixture_extract_two_frames_and_flush` helpers. - `test_fex_vector_dedup_by_provided_feature_name` gained a `fex_vector_create_and_append` helper. - `test_feature_extractor_supports_options_helper` split into `test_supports_options_empty_dict`, `_known_key`, `_unknown_key_reports_missing`, `_alias_key` and `_null_extractor`. - the SpEED GPU-twin registry guard moved out of `test_get_feature_extractor_by_name_and_feature_name` into `test_speed_gpu_twins_resolve_by_name`, which also keeps that case inside budget in an all-backends build. - `run_tests` now delegates to `run_registry_tests`, `run_context_tests` and `run_option_tests`, the grouping pattern `test_cli_parse.c` and `test_iqa_helpers.c` already use. No assertion was dropped: 17 cases / 79 assertions become 18 cases / 88 assertions, and the conditions folded into the fixture helpers are still checked there and reported through the fixture's status. The Netflix golden assertions are untouched, and the PR's CUDA CAMBI fix is unaffected. Verified: `clang-tidy -p core/build core/test/test_feature_extractor.c` 59 -> 0 warnings (clang-tidy 22.1.8); `core/build/test/test_feature_extractor` 18/18 pass; `meson test -C core/build --suite fast --no-suite gpu` 115/115 Ok; `pre-commit run --files` clean. Baseline tightened in `scripts/ci/tidy-baseline-cpu.json` (entry removed, total 3060 -> 3001) per the ADR-1142 ratchet, and a rebase note recorded for the new group runners and the NOLINT bracket. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

The touched-file cleanup rule (ADR-0141) makes this PR inherit every pre-existing clang-tidy finding in the TUs it edits. Discharge all 51 in `core/src/feature/feature_extractor.cpp` with no `NOLINT`: * modernize-use-nullptr (29): this TU is C++, so ADR-1138's keep-`NULL` exemption (which exists for the required MSVC C lane and upstream C parity) does not apply — spell the null pointer constant `nullptr`. Prose in comments still says "NULL" where it refers to the C twin. * misc-const-correctness (14): `const` on the error-code locals, on `ctx_array_sz`, and on the three read-only pointees (`ref_priv` in extract, `fex` in `vmaf_fex_ctx_pool_release` / `_flush`). * misc-use-anonymous-namespace (5): move `feature_extractor_list[]`, `vmaf_fex_ctx_parse_options`, `get_fex_list_entry`, `ctx_pool_ensure_slot_ctx` and `ctx_pool_claim_slot` out of `static` and into anonymous namespaces. The C twin keeps `static`. * readability-function-size (2): split rather than suppress. - `vmaf_feature_extractor_context_extract` (85 lines / 16 branches): the CUDA and HIP picture buf_type validation moves verbatim into `check_pic_buf_type(fex, ref_priv)`, which returns 0 or -EINVAL. Same predicates, same log messages, same return value. - `get_fex_list_entry` (68 lines) splits into `find_fex_list_entry` (the existing-entry lookup), `grow_fex_list` (the geometric realloc plus placement-new of the fresh slots) and `init_fex_list_slot` (capacity/condvar/ctx_list/opts_dict init, with the same unwind on each failure). The helpers return an errno; `get_fex_list_entry` still returns nullptr on any failure, which the caller still turns into -EINVAL, and `pool->cnt` is still bumped only on success. * cert-dcl03-c,misc-static-assert (1): clang-tidy 22 flags every `assert()` whose condition contains no non-constexpr call, so `assert(pool->capacity > 0)` cannot stay. Replaced with a hard `if (pool->capacity == 0) return -EINVAL;` in `grow_fex_list` — an always-on guard is strictly stronger than a check compiled out under NDEBUG, and the state is unreachable either way (`pool_create` seeds capacity = 8 and this is its only writer). `<cassert>` is now unused and dropped. Dropping `static` from the registry declaration broke the `check-dispatch-registry` pre-commit / CI gate, whose `sed` range hard-coded the `static VmafFeatureExtractor *feature_extractor_list[]` spelling; it now accepts both spellings, with a new self-test case (Test 4) covering the anonymous-namespace form. No behaviour change: the extractor registry, the context pool and the option validation this PR adds all keep their exact control flow, and the CUDA CAMBI fix is untouched. Verified: `clang-tidy -p core/build core/src/feature/feature_extractor.cpp` 51 -> 0 warnings (clang-tidy 22.1.8, CUDA-enabled build) and 0 in a CPU-only build dir matching the CI lane; `core/build/test/test_feature_extractor` 18/18 pass; `meson test -C core/build --suite fast --no-suite gpu` 115/115 Ok; `bash scripts/ci/tests/test-check-dispatch-registry.sh` 11/11 pass; `pre-commit run --files` clean. Baseline tightened in `scripts/ci/tidy-baseline-cpu.json` (entry removed, total 3001 -> 2951) per the ADR-1142 ratchet, and a rebase note recorded for the anonymous namespaces and the two function splits. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`core/src/feature/feature_extractor.h` was the last file this branch touches that still carried clang-tidy findings (8). clang-tidy has no compile command for a header and falls back to the nearest matching TU's — `feature_extractor.cpp` — so it parses this C header as C++ and proposes rewrites that would break the roughly one hundred C translation units which include it: `<cstdint>` / `<cstdlib>` for the C headers, `using` for the `typedef struct`s, a C23 fixed underlying type on the two flag enums (which MSVC does not document under `/std:clatest`, and the Windows MSVC build is a required check), and a rename of the upstream `__VMAF_FEATURE_EXTRACTOR_H__` include guard that every other header under `core/src/` also uses. Discharged with one file-scoped `NOLINTBEGIN` / `NOLINTEND` bracket — the shape ADR-1138 prescribes and `core/test/test_feature_extractor.c` already uses — whose justification and its ADR-0141 / ADR-1138 / ADR-0148 / ADR-0150 citations live inside the marker's own block comment, where `scripts/ci/tidy-ratchet.py::count_uncited_nolints` looks for them. No declaration, macro, type or guard changed, so no includer is affected. The `-Db_lto=false` half of the original commit is dropped: PR #1326 landed the identical fix on master while this branch was in review. Verified: clang-tidy 22.1.8 against a CPU-only build dir configured like the CI lane reports 0 warnings, 0 errors, exit 0 for all four CI-linted files (`feature_extractor.cpp`, `feature_extractor.h`, `libvmaf.c`, `test_feature_extractor.c`); `count_uncited_nolints` on the header returns 0. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…195) CLAUDE.md rule 15 says to rebuild the container when "its image predates the last master sync". That is stated in terms of time, and time cannot answer the question. A build run against a checkout that is behind master produces an image newer than every commit in the repository and missing exactly the work it was rebuilt for. That happened today. The container was rebuilt specifically to pick up the GPU default-model fixes (#1307, #1312, #1324) so the epic #1246 GPU smoke could run against them. The build succeeded, the image was the newest thing on disk, and it contained none of the three: the build context was 28 commits behind origin/master. It was caught only because a test file added by one of those PRs was missing. Had the smoke run instead, it would have reported green numbers for code that was not in the image, and those numbers would have been cited as a retrain gate. ADR-1102's marker answers "did a container build this?". It cannot answer "which code was in that container?", and that second question is the one that was wrong. dev/Containerfile now records /etc/vmafx-dev-source (source_rev, source_ref, source_repo) from a VMAFX_SOURCE_REV build argument supplied by dev/docker-compose.yml. It is written in the LAST stage, deliberately away from the ADR-1102 marker in the first: the first stage is reused by every rebuild, so a marker there would report the revision of whichever build first populated the layer cache -- authoritative and stale, which is worse than absent. scripts/dev/check-container-source.sh answers it in both directions. --pre-build refuses a checkout that is behind the reference and lists the commits under baked-in paths the image would be missing; --image reads the marker out of an existing image and reports current, stale (naming what is missing), or unverifiable. A build that never received the argument records `unknown`, and `unknown` is reported as "cannot verify", not as a pass: an image that cannot say what it holds is not evidence. dev/scripts/container-build.sh makes the correct path the easy one -- check, build with the verified revision, re-verify the result. --allow-behind exists for local experiments and says so loudly. Verified by scripts/ci/tests/test-check-container-source.sh: 8 assertions, hermetic apart from one Docker case that skips when no daemon is reachable. The stale-context case reproduces today's shape; ahead-of-master is deliberately allowed, since a feature branch legitimately leads master. no digest needed: trivial Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

… (ADR-1192) Re-ran the Netflix benchmark suite on cd52f26 for epic #1245 items 1 and 5. All three fixtures reproduce on CPU, CUDA and SYCL through the FFmpeg filter path against a container-built current-master libvmaf, but every backend's pooled score has drifted from testdata/netflix_benchmark_results.json (recorded by PR #309 on 2026-05-02): CPU +2.83e-06, CUDA -1.07e-03, SYCL -1.40e-03 on the 576x324 pair. A rebuild of 5a08030 — the commit before the 2026-09-06 GPU merges #1307/#1312/#1324 — shows the same drift, so none of it comes from today's merges. The snapshot is deliberately NOT regenerated (ADR-1192) and no throughput baseline is recorded, because the run also reproduced two pre-existing GPU defects: - vmaf --threads N aborts on every GPU backend (exit 234, "context could not be synchronized"); without --threads both CUDA and SYCL score correctly and are bit-stable over 10 runs. bench_all.sh hard-codes --threads 1. - The libvmaf_cuda FFmpeg filter returns a wrong pooled score in 10 of 40 runs on master and 8 of 40 on 5a08030 — inside binomial noise of each other. Harness fixes in the same change: - bench_all.sh kept its stderr on /dev/null and relabelled every non-zero exit as "backend likely unavailable", which is how a hard abort passed for a missing device for months. It now captures stderr per row and prints FAIL with the exit code and the real last line. Its flag sets also drop --no_vulkan, unrecognized since ADR-0726 removed the Vulkan backend. - benchmark_netflix.py hard-coded /home/kilian/dev/ffmpeg-8/ffmpeg (gone) and /dev/dri/renderD130 for the SYCL/QSV import (now the AMD iGPU on the bench host, so the SYCL rows failed outright). Both are environment overrides now, VMAF_FFMPEG and the new VMAF_SYCL_RENDER_NODE, per the ADR-0792 pattern. No golden assertions touched; no snapshot regenerated. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…195) CLAUDE.md rule 15 says to rebuild the container when "its image predates the last master sync". That is stated in terms of time, and time cannot answer the question. A build run against a checkout that is behind master produces an image newer than every commit in the repository and missing exactly the work it was rebuilt for. That happened today. The container was rebuilt specifically to pick up the GPU default-model fixes (#1307, #1312, #1324) so the epic #1246 GPU smoke could run against them. The build succeeded, the image was the newest thing on disk, and it contained none of the three: the build context was 28 commits behind origin/master. It was caught only because a test file added by one of those PRs was missing. Had the smoke run instead, it would have reported green numbers for code that was not in the image, and those numbers would have been cited as a retrain gate. ADR-1102's marker answers "did a container build this?". It cannot answer "which code was in that container?", and that second question is the one that was wrong. dev/Containerfile now records /etc/vmafx-dev-source (source_rev, source_ref, source_repo) from a VMAFX_SOURCE_REV build argument supplied by dev/docker-compose.yml. It is written in the LAST stage, deliberately away from the ADR-1102 marker in the first: the first stage is reused by every rebuild, so a marker there would report the revision of whichever build first populated the layer cache -- authoritative and stale, which is worse than absent. scripts/dev/check-container-source.sh answers it in both directions. --pre-build refuses a checkout that is behind the reference and lists the commits under baked-in paths the image would be missing; --image reads the marker out of an existing image and reports current, stale (naming what is missing), or unverifiable. A build that never received the argument records `unknown`, and `unknown` is reported as "cannot verify", not as a pass: an image that cannot say what it holds is not evidence. dev/scripts/container-build.sh makes the correct path the easy one -- check, build with the verified revision, re-verify the result. --allow-behind exists for local experiments and says so loudly. Verified by scripts/ci/tests/test-check-container-source.sh: 8 assertions, hermetic apart from one Docker case that skips when no daemon is reachable. The stale-context case reproduces today's shape; ahead-of-master is deliberately allowed, since a feature branch legitimately leads master. no digest needed: trivial Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…195) (#1337) * ci(dev): make the container say which source it was built from (ADR-1195) CLAUDE.md rule 15 says to rebuild the container when "its image predates the last master sync". That is stated in terms of time, and time cannot answer the question. A build run against a checkout that is behind master produces an image newer than every commit in the repository and missing exactly the work it was rebuilt for. That happened today. The container was rebuilt specifically to pick up the GPU default-model fixes (#1307, #1312, #1324) so the epic #1246 GPU smoke could run against them. The build succeeded, the image was the newest thing on disk, and it contained none of the three: the build context was 28 commits behind origin/master. It was caught only because a test file added by one of those PRs was missing. Had the smoke run instead, it would have reported green numbers for code that was not in the image, and those numbers would have been cited as a retrain gate. ADR-1102's marker answers "did a container build this?". It cannot answer "which code was in that container?", and that second question is the one that was wrong. dev/Containerfile now records /etc/vmafx-dev-source (source_rev, source_ref, source_repo) from a VMAFX_SOURCE_REV build argument supplied by dev/docker-compose.yml. It is written in the LAST stage, deliberately away from the ADR-1102 marker in the first: the first stage is reused by every rebuild, so a marker there would report the revision of whichever build first populated the layer cache -- authoritative and stale, which is worse than absent. scripts/dev/check-container-source.sh answers it in both directions. --pre-build refuses a checkout that is behind the reference and lists the commits under baked-in paths the image would be missing; --image reads the marker out of an existing image and reports current, stale (naming what is missing), or unverifiable. A build that never received the argument records `unknown`, and `unknown` is reported as "cannot verify", not as a pass: an image that cannot say what it holds is not evidence. dev/scripts/container-build.sh makes the correct path the easy one -- check, build with the verified revision, re-verify the result. --allow-behind exists for local experiments and says so loudly. Verified by scripts/ci/tests/test-check-container-source.sh: 8 assertions, hermetic apart from one Docker case that skips when no daemon is reachable. The stale-context case reproduces today's shape; ahead-of-master is deliberately allowed, since a feature branch legitimately leads master. no digest needed: trivial Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(state): cite PR #1337 in the container-staleness row (ADR-0165 touch gate) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(state): drop the duplicate rows a keep-both rebase created Each dropped row restates one origin/master already carries; master is the authoritative record. Verified with scripts/ci/check-state-md-rows.sh. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

… (ADR-1192) Re-ran the Netflix benchmark suite on cd52f26 for epic #1245 items 1 and 5. All three fixtures reproduce on CPU, CUDA and SYCL through the FFmpeg filter path against a container-built current-master libvmaf, but every backend's pooled score has drifted from testdata/netflix_benchmark_results.json (recorded by PR #309 on 2026-05-02): CPU +2.83e-06, CUDA -1.07e-03, SYCL -1.40e-03 on the 576x324 pair. A rebuild of 5a08030 — the commit before the 2026-09-06 GPU merges #1307/#1312/#1324 — shows the same drift, so none of it comes from today's merges. The snapshot is deliberately NOT regenerated (ADR-1192) and no throughput baseline is recorded, because the run also reproduced two pre-existing GPU defects: - vmaf --threads N aborts on every GPU backend (exit 234, "context could not be synchronized"); without --threads both CUDA and SYCL score correctly and are bit-stable over 10 runs. bench_all.sh hard-codes --threads 1. - The libvmaf_cuda FFmpeg filter returns a wrong pooled score in 10 of 40 runs on master and 8 of 40 on 5a08030 — inside binomial noise of each other. Harness fixes in the same change: - bench_all.sh kept its stderr on /dev/null and relabelled every non-zero exit as "backend likely unavailable", which is how a hard abort passed for a missing device for months. It now captures stderr per row and prints FAIL with the exit code and the real last line. Its flag sets also drop --no_vulkan, unrecognized since ADR-0726 removed the Vulkan backend. - benchmark_netflix.py hard-coded /home/kilian/dev/ffmpeg-8/ffmpeg (gone) and /dev/dri/renderD130 for the SYCL/QSV import (now the AMD iGPU on the bench host, so the SYCL rows failed outright). Both are environment overrides now, VMAF_FFMPEG and the new VMAF_SYCL_RENDER_NODE, per the ADR-0792 pattern. No golden assertions touched; no snapshot regenerated. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

… (ADR-1192) (#1334) * fix(testdata): make the Netflix benchmark harness honest and portable (ADR-1192) Re-ran the Netflix benchmark suite on cd52f26 for epic #1245 items 1 and 5. All three fixtures reproduce on CPU, CUDA and SYCL through the FFmpeg filter path against a container-built current-master libvmaf, but every backend's pooled score has drifted from testdata/netflix_benchmark_results.json (recorded by PR #309 on 2026-05-02): CPU +2.83e-06, CUDA -1.07e-03, SYCL -1.40e-03 on the 576x324 pair. A rebuild of 5a08030 — the commit before the 2026-09-06 GPU merges #1307/#1312/#1324 — shows the same drift, so none of it comes from today's merges. The snapshot is deliberately NOT regenerated (ADR-1192) and no throughput baseline is recorded, because the run also reproduced two pre-existing GPU defects: - vmaf --threads N aborts on every GPU backend (exit 234, "context could not be synchronized"); without --threads both CUDA and SYCL score correctly and are bit-stable over 10 runs. bench_all.sh hard-codes --threads 1. - The libvmaf_cuda FFmpeg filter returns a wrong pooled score in 10 of 40 runs on master and 8 of 40 on 5a08030 — inside binomial noise of each other. Harness fixes in the same change: - bench_all.sh kept its stderr on /dev/null and relabelled every non-zero exit as "backend likely unavailable", which is how a hard abort passed for a missing device for months. It now captures stderr per row and prints FAIL with the exit code and the real last line. Its flag sets also drop --no_vulkan, unrecognized since ADR-0726 removed the Vulkan backend. - benchmark_netflix.py hard-coded /home/kilian/dev/ffmpeg-8/ffmpeg (gone) and /dev/dri/renderD130 for the SYCL/QSV import (now the AMD iGPU on the bench host, so the SYCL rows failed outright). Both are environment overrides now, VMAF_FFMPEG and the new VMAF_SYCL_RENDER_NODE, per the ADR-0792 pattern. No golden assertions touched; no snapshot regenerated. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(state): drop the duplicate rows a keep-both rebase created Each dropped row restates one origin/master already carries; master is the authoritative record. Verified with scripts/ci/check-state-md-rows.sh. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

The epic #1246 gate table listed G2 and G3 as FAIL against conditions that no longer hold. G3 in particular blamed "PR #1307 & fix/cambi-cuda-context unmerged" -- both merged on 2026-09-05/06. Re-ran each gate rather than reasoning about it: G2 PASS zero failing runs on master HEAD. The release-please failure the row was written for is gone; ADR-1171 made the missing release-bot credential warn-not-error on push and the workflow reports success. G3 PASS #1307, #1312 and #1324 all on master; container rebuilt from cd52f26 and the default model verified on ALL FOUR backends -- CPU 82.816062, CUDA 82.814062, SYCL 82.814061, HIP 82.816061, each exiting 0. The runbook only asked for CUDA; the others were checked because a model reaches the GPU twins through the model, not through --feature, so a CUDA-only check would not have covered them. G1 FAIL 12 epics open, listed by number, with the caveat that the epic bodies are snapshots and several of their items have already shipped -- the count overstates the work. G4 FAIL still blocked on #1302, and now says why precisely: master's extract_k150k_features.py has no --vmaf-model flag (grep returns 0), which is what §4.2's teacher_model assertion needs. #1302 has no failing check -- its only red mark is the aggregator's draft guard, and its ADR-0108 validator passes six of six. It needs promotion, not repair. Adds a note that the table is a measurement and each row must be re-run rather than carried forward, since stale-status drift is what it just corrected. no digest needed: trivial Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…tus (#1350) * docs(ai): record measured retrain gate status, not authoring-time status The epic #1246 gate table listed G2 and G3 as FAIL against conditions that no longer hold. G3 in particular blamed "PR #1307 & fix/cambi-cuda-context unmerged" -- both merged on 2026-09-05/06. Re-ran each gate rather than reasoning about it: G2 PASS zero failing runs on master HEAD. The release-please failure the row was written for is gone; ADR-1171 made the missing release-bot credential warn-not-error on push and the workflow reports success. G3 PASS #1307, #1312 and #1324 all on master; container rebuilt from cd52f26 and the default model verified on ALL FOUR backends -- CPU 82.816062, CUDA 82.814062, SYCL 82.814061, HIP 82.816061, each exiting 0. The runbook only asked for CUDA; the others were checked because a model reaches the GPU twins through the model, not through --feature, so a CUDA-only check would not have covered them. G1 FAIL 12 epics open, listed by number, with the caveat that the epic bodies are snapshots and several of their items have already shipped -- the count overstates the work. G4 FAIL still blocked on #1302, and now says why precisely: master's extract_k150k_features.py has no --vmaf-model flag (grep returns 0), which is what §4.2's teacher_model assertion needs. #1302 has no failing check -- its only red mark is the aggregator's draft guard, and its ADR-0108 validator passes six of six. It needs promotion, not repair. Adds a note that the table is a measurement and each row must be re-run rather than carried forward, since stale-status drift is what it just corrected. no digest needed: trivial Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(ai): fix the K150K scores path in both the smoke and the production run The runbook pointed --scores at .corpus/konvid-150k/scores.csv. That file does not exist. KoNViD-150k splits its scores by part, and the corpus on this workstation holds k150ka_scores.csv and k150kb_scores.csv (plus the matching *_votes.csv and a manifest.csv). The wrong path appeared TWICE: in the section 4 five-clip smoke and in the section 5.1 multi-day production extraction. extract_k150k_features.py validates the file and exits with 'error: scores CSV not found', so each would have aborted on its first line -- the smoke immediately, and the ~105-110 hour K150K run at its very start. Corrected to k150ka_scores.csv, which is also the script's own argparse default and the path in its module docstring example. --clips-dir is deliberately left as clips/: a real directory of 153,841 files and a superset of the script's default k150ka_extracted/ (152,265). Lookup is by video_name, so either resolves. Both paths are now listed in a verify-first command so an operator checks them before committing to the long run. no digest needed: trivial Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(ai): correct four K150K command errors found by running the smoke Ran the section 4 five-clip smoke end to end in the container. As written it failed on its first line, then its second, then on every clip. Each fix below comes from an actual run, not from reading. 1. --scores named scores.csv, which does not exist. KoNViD-150k splits its scores by part; the corpus holds k150ka_scores.csv (154,746 rows) and k150kb_scores.csv. k150ka_scores.csv is the script's own default. 2. --cpu-vmaf-bin was missing entirely. It is required, and its default /build/vmaf/core/build-cpu/tools/vmaf does not exist in the container, so the run aborts with 'error: cpu-vmaf-bin not found'. /usr/local/bin/vmaf serves both roles. 3. --clips-dir named clips/, which is unusable from inside the container. Its 153,841 entries are symlinks to HOST absolute paths under /home/kilian/dev/vmaf/.workingdir2/konvid-150k/, which do not resolve in the container mount. Every clip failed with 'ffprobe ... returned non-zero exit status 1' -- which reads like corrupt media and is really a dangling link. The real files are in k150ka_extracted/ (152,265 files, ffprobe reports 960x540), again the script's default. All three appeared in BOTH the section 4 smoke and the section 5.1 multi-day production extraction, so the ~105-110 hour K150K run would have aborted at its very start. With them fixed the pipeline runs clean: ok=5 fail=0 at 1.26 clip/s, status complete, schema k150k-feature-extraction-manifest-v1, 5 parquet rows. Of section 4.2's assertions, schema, status and stats.ok already pass on master. Three do not, and all three come from #1302: teacher_model in the manifest, the teacher_model parquet column, and adm3_mean (grep -c adm3 returns 0 on master's extractor and 3 on #1302's). G4 is blocked on that PR alone -- corpus, binary and pipeline are all verified working. no digest needed: trivial Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

lusoris added this to the 1.0.0 — First release milestone Sep 5, 2026

lusoris force-pushed the fix/cambi-cuda-context branch 2 times, most recently from 635643f to 6ed95b4 Compare September 5, 2026 16:05

lusoris marked this pull request as ready for review September 5, 2026 16:20

This was referenced Sep 5, 2026

fix(gpu): align the CUDA/SYCL cambi twins' spatial mask and filter_mode with cambi.c #1325

Merged

fix(ci): build the clang-tidy lanes without LTO so the gate can pass at all #1326

Merged

lusoris marked this pull request as draft September 5, 2026 20:32

lusoris marked this pull request as ready for review September 5, 2026 20:45

lusoris marked this pull request as draft September 5, 2026 21:01

lusoris force-pushed the fix/cambi-cuda-context branch from 0f59ed9 to a8d69e8 Compare September 5, 2026 21:15

lusoris marked this pull request as ready for review September 5, 2026 21:16

lusoris marked this pull request as draft September 5, 2026 21:39

Lusoris and others added 4 commits September 6, 2026 00:13

lusoris force-pushed the fix/cambi-cuda-context branch from a8d69e8 to ce6f1c7 Compare September 5, 2026 22:13

lusoris marked this pull request as ready for review September 5, 2026 22:13

lusoris force-pushed the fix/cambi-cuda-context branch from ce6f1c7 to 97575f1 Compare September 5, 2026 22:15

lusoris merged commit fcae339 into master Sep 5, 2026
78 of 79 checks passed

lusoris deleted the fix/cambi-cuda-context branch September 5, 2026 22:35

This was referenced Sep 5, 2026

fix(testdata): make the Netflix benchmark harness honest and portable (ADR-1192) #1334

Merged

ci(dev): make the container say which source it was built from (ADR-1195) #1337

Merged

Model retraining against the v1.0.16 teacher (LAST) #1246

Open

lusoris mentioned this pull request Sep 6, 2026

docs(ai): record measured retrain gate status, not authoring-time status #1350

Merged

17 tasks

lusoris mentioned this pull request Sep 3, 2026

chore(master): release 1.0.0-rc.1 #1213

Open

lusoris added the type:bug Something isn't working label Sep 7, 2026

Navigation Menu

Uh oh!

fix(cuda): push context in cambi submit and gate twin selection on model options (ADR-1183) - #1312

fix(cuda): push context in cambi submit and gate twin selection on model options (ADR-1183)#1312
lusoris merged 5 commits into
masterfrom
fix/cambi-cuda-context

lusoris commented Sep 5, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

lusoris commented Sep 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Type

Checklist

Bug-status hygiene (ADR-0165)

Netflix golden-data gate (ADR-0024)

Cross-backend numerical results

Deep-dive deliverables (ADR-0108)

Reproducer

Known follow-ups

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

lusoris commented Sep 5, 2026 •

edited

Loading