Repository navigation
fix: 6 real-code-fix bundle (matrix-v2 follow-up + NEON FMA re-dispatch) - #853
Merged
Merged
Conversation
…-1057) PR #695 / ADR-1057 reverted the float-ADM NEON dispatch because vmlaq_laneq_f32 is a hardware FMA intrinsic that produces a 1-ULP divergence from the scalar reference; #pragma clang fp contract(off) alone cannot suppress intrinsic-encoded FMA. Fix: extract the DWT2 kernel into float_adm_dwt2_neon.c compiled in its own static library (arm64_adm_dwt2_neon_lib) with -ffp-contract=off. Replace every vmlaq_laneq_f32(acc, v, coeff, lane) with the explicit two-step vaddq_f32(acc, vmulq_laneq_f32(v, coeff, lane)), emitting separate fmul + fadd instructions that match the scalar reference rounding. Belt-and-suspenders: add #pragma clang fp contract(off) at TU level (Clang) and __attribute__((optimize("-ffp-contract=off"))) on the function definition (GCC) to protect the scalar tail. Wire the dispatch in adm.c::adm_dwt2_dispatch(): selects the NEON kernel when VMAF_ARM_CPU_FLAG_NEON is set, falls back to adm_dwt2_s() otherwise. The #define adm_dwt2 adm_dwt2_dispatch replaces the direct scalar macro. Verified: aarch64-linux-gnu-gcc -c compiles all three TUs cleanly. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…tching conclude_adm_cm offsets
The adm_cm_aim_line_kernel<rows_per_thread> for scale-0 int16 AIM path
incorrectly used inline_s0_csf_a() to compute the signal, which applies
a >>15 pre-shift before returning. The threshold path (conclude_adm_cm)
and the shift_sub[] constant array {10, 10, 12} are calibrated for the
unshifted signal magnitude (abs(rfactor * a_val)), producing near-zero
AIM scores because the signal was ~32768x too small.
Fix mirrors the DLM kernel pattern (adm_cm_line_kernel lines 349-355):
use inline_s0_decouple_r() to obtain r_val, load the distorted band
value t_val via __ldg(), compute a_val = t_val - r_val, then form the
signal as abs(int32_t(i_rfactor[blockIdx.z] * a_val)) without the >>15
shift. No API or header changes; CUDA-only internal kernel fix.
Build verified: meson setup core/build-aim-fix -Denable_cuda=true &&
ninja -C core/build-aim-fix (exit 0, adm_cm PTX compiled without errors).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ice before return (host-device race) vmaf_cuda_picture_download_async issues a cuMemcpy2DAsync on the picture's per-picture CUstream (CU_STREAM_NON_BLOCKING, allocated in vmaf_cuda_picture_alloc per ADR-0378). The function returned immediately without waiting for the DMA to complete, so CPU-side feature extractors (motion, adm, vif, etc.) racing in on the same thread pool read partially-written host buffers — producing VMAF scores of ~69-71 instead of ~100 on identical-input pairs. Fix: after the async download succeeds, call cu_f->cuStreamSynchronize(vmaf_cuda_picture_get_stream(pic)) to drain the per-picture stream before returning. The synchronize is O(1) overhead on the critical path (the DMA is already in-flight; this just adds a CPU-side wait) and is consistent with the cuStreamSynchronize call in vmaf_cuda_picture_free and the cuStreamSynchronize in flush_context_cuda. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…to integer_* JSON paths Modern libvmaf emits pooled_metrics keys with an ``integer_`` prefix for the integer-pipeline feature extractors (e.g. ``integer_adm2`` instead of ``adm2``, ``integer_vif_scale0`` instead of ``vif_scale0``). ``parse_feature_aggregates`` was looking up the bare canonical names directly, so every corpus row's 12 per-feature columns (adm2_mean, vif_scale[0..3]_mean, motion2_mean and their _std counterparts) were silently NaN against any real libvmaf output. Fix: add ``_CANONICAL_TO_POOLED_KEY`` mapping in score.py and update ``parse_feature_aggregates`` to resolve each canonical name through the map first, falling back to the bare name for non-integer features (cambi, future additions) and for synthetic test fixtures that use bare keys. Regression guard: add test_parse_feature_aggregates_integer_keys.py with 8 cases covering real integer_* payloads, stddev presence/absence, bare-key fallback, partial presence, empty pooled_metrics, and backward compat with the existing synthetic-fixture shape used by test_corpus_schema_v3.py. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…path _run_recommend_from_corpus unconditionally called the point-estimate recommend() path, silently ignoring args.with_uncertainty. When --from-corpus and --with-uncertainty are combined, the function now routes through pick_target_vmaf_with_uncertainty (the interval-aware predictor from ADR-0279), emitting decision= and visited= fields in human-readable output. --with-uncertainty + --target-bitrate is unsupported (no interval-aware bitrate predicate exists); the function now warns on stderr and falls back to the point-estimate path rather than silently ignoring the flag. Tests added to test_recommend_uncertainty.py covering: - plain vs uncertainty-aware output differ when tight interval fires - --json output emits the winning row correctly on the uncertainty path - --with-uncertainty + --target-bitrate warns and falls back gracefully Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… golden gate The script only fetched the src01 576x324 normal pair. The Containerfile (dev/Containerfile, added in PR #850) calls this script at build time, but the two 1920x1080 checkerboard pairs required by CLAUDE.md §8 golden gate (python/test/quality_runner_test.py::test_run_vmaf_runner_checkerboard) were never downloaded, so the checkerboard golden tests always failed inside the container. Add the three missing YUV files to the FIXTURES array with md5sums verified 2026-06-08 against Netflix/vmaf_resource HEAD: - checkerboard_1920_1080_10_3_0_0.yuv ad14c75d1897e7f2bc72a882c32e49a6 - checkerboard_1920_1080_10_3_1_0.yuv 0e290566458d800c534ab18103619d43 - checkerboard_1920_1080_10_3_10_0.yuv 289119e1168ab656ac79487df8d307b9 Local verify: script run against a fresh tmp dir — all 5 fixtures fetched and md5-verified; re-run with all files present — all 5 report "ok" (no network I/O). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
lusoris
marked this pull request as ready for review
June 8, 2026 17:56
26 of 31 tasks
5 of 9 tasks
lusoris
added a commit
that referenced
this pull request
Sep 6, 2026
The ledger row for T-NEON-FMA-FLOAT-ADM-DWT2-2026-06-06 was self-contradictory: a 2026-06-27 note inside the T-MASTER-CI-TSAN-ARM-GOLDEN-2026-06-27 row said the bug "stays Open" for a correct re-attempt, while its Recently-closed row already recorded a 2026-08-30 closure — and that closure narrative described only the integer-path dropped-tap defect and cited a branch name (`fix/adr-1057-arm-fma-drift`) instead of a commit. Re-verified against origin/master, with no defective code left to point at: - The float-ADM follow-up the row asked for landed in `a6c4dfffb` (PR #853): the kernel lives in a dedicated non-contracting TU `core/src/feature/arm64/float_adm_dwt2_neon.c`, built with `-ffp-contract=off` (`core/src/meson.build`), and `adm_dwt2_dispatch()` in `core/src/feature/adm.c` calls `float_adm_dwt2_neon()` under `VMAF_ARM_CPU_FLAG_NEON`. The scalar `adm_dwt2_s` carries the matching function-scoped guard, so both sides of the comparison are non-contracting and the 1-ULP FMA gap that motivated the row cannot arise. - The integer `idx < 3` dropped filter tap in `adm_dwt2_8_neon` was fixed by `a013c1410` (PR #1134) and hardened to bit-exactness by `89a8e3258` (PR #1154); `6d61106ed` (PR #1156) added NEON parity tests for the remaining uncovered kernels. - `core/test/test_float_adm_dwt2_neon.c` gates NEON-vs-scalar bit-exactness by bit pattern and is registered in the default `fast` suite. The row stays in "Recently closed" (exactly one row for the id), rewritten in past tense with real commit shas, and the stale "stays Open" note is marked superseded so the next session does not re-investigate. Documentation only — no code or behaviour change. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
lusoris
added a commit
that referenced
this pull request
Sep 6, 2026
The ledger row for T-NEON-FMA-FLOAT-ADM-DWT2-2026-06-06 was self-contradictory: a 2026-06-27 note inside the T-MASTER-CI-TSAN-ARM-GOLDEN-2026-06-27 row said the bug "stays Open" for a correct re-attempt, while its Recently-closed row already recorded a 2026-08-30 closure — and that closure narrative described only the integer-path dropped-tap defect and cited a branch name (`fix/adr-1057-arm-fma-drift`) instead of a commit. Re-verified against origin/master, with no defective code left to point at: - The float-ADM follow-up the row asked for landed in `a6c4dfffb` (PR #853): the kernel lives in a dedicated non-contracting TU `core/src/feature/arm64/float_adm_dwt2_neon.c`, built with `-ffp-contract=off` (`core/src/meson.build`), and `adm_dwt2_dispatch()` in `core/src/feature/adm.c` calls `float_adm_dwt2_neon()` under `VMAF_ARM_CPU_FLAG_NEON`. The scalar `adm_dwt2_s` carries the matching function-scoped guard, so both sides of the comparison are non-contracting and the 1-ULP FMA gap that motivated the row cannot arise. - The integer `idx < 3` dropped filter tap in `adm_dwt2_8_neon` was fixed by `a013c1410` (PR #1134) and hardened to bit-exactness by `89a8e3258` (PR #1154); `6d61106ed` (PR #1156) added NEON parity tests for the remaining uncovered kernels. - `core/test/test_float_adm_dwt2_neon.c` gates NEON-vs-scalar bit-exactness by bit pattern and is registered in the default `fast` suite. The row stays in "Recently closed" (exactly one row for the id), rewritten in past tense with real commit shas, and the stale "stays Open" note is marked superseded so the next session does not re-investigate. Documentation only — no code or behaviour change. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
lusoris
added a commit
that referenced
this pull request
Sep 6, 2026
The ledger row for T-NEON-FMA-FLOAT-ADM-DWT2-2026-06-06 was self-contradictory: a 2026-06-27 note inside the T-MASTER-CI-TSAN-ARM-GOLDEN-2026-06-27 row said the bug "stays Open" for a correct re-attempt, while its Recently-closed row already recorded a 2026-08-30 closure — and that closure narrative described only the integer-path dropped-tap defect and cited a branch name (`fix/adr-1057-arm-fma-drift`) instead of a commit. Re-verified against origin/master, with no defective code left to point at: - The float-ADM follow-up the row asked for landed in `a6c4dfffb` (PR #853): the kernel lives in a dedicated non-contracting TU `core/src/feature/arm64/float_adm_dwt2_neon.c`, built with `-ffp-contract=off` (`core/src/meson.build`), and `adm_dwt2_dispatch()` in `core/src/feature/adm.c` calls `float_adm_dwt2_neon()` under `VMAF_ARM_CPU_FLAG_NEON`. The scalar `adm_dwt2_s` carries the matching function-scoped guard, so both sides of the comparison are non-contracting and the 1-ULP FMA gap that motivated the row cannot arise. - The integer `idx < 3` dropped filter tap in `adm_dwt2_8_neon` was fixed by `a013c1410` (PR #1134) and hardened to bit-exactness by `89a8e3258` (PR #1154); `6d61106ed` (PR #1156) added NEON parity tests for the remaining uncovered kernels. - `core/test/test_float_adm_dwt2_neon.c` gates NEON-vs-scalar bit-exactness by bit pattern and is registered in the default `fast` suite. The row stays in "Recently closed" (exactly one row for the id), rewritten in past tense with real commit shas, and the stale "stays Open" note is marked superseded so the next session does not re-investigate. Documentation only — no code or behaviour change. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
lusoris
added a commit
that referenced
this pull request
Sep 7, 2026
The ledger row for T-NEON-FMA-FLOAT-ADM-DWT2-2026-06-06 was self-contradictory: a 2026-06-27 note inside the T-MASTER-CI-TSAN-ARM-GOLDEN-2026-06-27 row said the bug "stays Open" for a correct re-attempt, while its Recently-closed row already recorded a 2026-08-30 closure — and that closure narrative described only the integer-path dropped-tap defect and cited a branch name (`fix/adr-1057-arm-fma-drift`) instead of a commit. Re-verified against origin/master, with no defective code left to point at: - The float-ADM follow-up the row asked for landed in `a6c4dfffb` (PR #853): the kernel lives in a dedicated non-contracting TU `core/src/feature/arm64/float_adm_dwt2_neon.c`, built with `-ffp-contract=off` (`core/src/meson.build`), and `adm_dwt2_dispatch()` in `core/src/feature/adm.c` calls `float_adm_dwt2_neon()` under `VMAF_ARM_CPU_FLAG_NEON`. The scalar `adm_dwt2_s` carries the matching function-scoped guard, so both sides of the comparison are non-contracting and the 1-ULP FMA gap that motivated the row cannot arise. - The integer `idx < 3` dropped filter tap in `adm_dwt2_8_neon` was fixed by `a013c1410` (PR #1134) and hardened to bit-exactness by `89a8e3258` (PR #1154); `6d61106ed` (PR #1156) added NEON parity tests for the remaining uncovered kernels. - `core/test/test_float_adm_dwt2_neon.c` gates NEON-vs-scalar bit-exactness by bit pattern and is registered in the default `fast` suite. The row stays in "Recently closed" (exactly one row for the id), rewritten in past tense with real commit shas, and the stale "stays Open" note is marked superseded so the next session does not re-investigate. Documentation only — no code or behaviour change. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
lusoris
added a commit
that referenced
this pull request
Sep 7, 2026
…#1331) The ledger row for T-NEON-FMA-FLOAT-ADM-DWT2-2026-06-06 was self-contradictory: a 2026-06-27 note inside the T-MASTER-CI-TSAN-ARM-GOLDEN-2026-06-27 row said the bug "stays Open" for a correct re-attempt, while its Recently-closed row already recorded a 2026-08-30 closure — and that closure narrative described only the integer-path dropped-tap defect and cited a branch name (`fix/adr-1057-arm-fma-drift`) instead of a commit. Re-verified against origin/master, with no defective code left to point at: - The float-ADM follow-up the row asked for landed in `a6c4dfffb` (PR #853): the kernel lives in a dedicated non-contracting TU `core/src/feature/arm64/float_adm_dwt2_neon.c`, built with `-ffp-contract=off` (`core/src/meson.build`), and `adm_dwt2_dispatch()` in `core/src/feature/adm.c` calls `float_adm_dwt2_neon()` under `VMAF_ARM_CPU_FLAG_NEON`. The scalar `adm_dwt2_s` carries the matching function-scoped guard, so both sides of the comparison are non-contracting and the 1-ULP FMA gap that motivated the row cannot arise. - The integer `idx < 3` dropped filter tap in `adm_dwt2_8_neon` was fixed by `a013c1410` (PR #1134) and hardened to bit-exactness by `89a8e3258` (PR #1154); `6d61106ed` (PR #1156) added NEON parity tests for the remaining uncovered kernels. - `core/test/test_float_adm_dwt2_neon.c` gates NEON-vs-scalar bit-exactness by bit pattern and is registered in the default `fast` suite. The row stays in "Recently closed" (exactly one row for the id), rewritten in past tense with real commit shas, and the stale "stays Open" note is marked superseded so the next session does not re-investigate. Documentation only — no code or behaviour change. Co-authored-by: Lusoris <lusoris@pm.me> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Bundles 6 real production-code fixes:
Test plan
Deep-dive deliverables (ADR-0108)
state.md touch