Prevent concurrent fixture producers from replacing files held open by
external decoders. Add a four-consumer Windows sharing regression.
Co-Authored-By: Codex (GPT-6) <noreply@openai.com>
Stateless VA-API decoding with FilmCraft's own H.264 host side: the
software decoder's SPS/PPS/slice-header parsers and its DPB (now generic
over the picture type in filmcraft-h264) drive the GPU, so reference
marking, reference lists and output order are shared and the pictures are
bit-exact with the software decoder. libva is loaded at run time; no libva
or driver means software decoding as before. frame_num gaps (open-GOP
seeks) get the software decoder's stand-in frames. HybridDecoder keeps the
transparent software fallback.
Tests: hardware parity on 7 H.264 streams incl. 1080p/2160p, seeks,
flushes, forced failures, fuzzing, concurrency; front end against a
recording stand-in on every OS; ABI checks against va.h. Docs: platform
README, ADR 0001 addendum, architecture, performance (H.264 2160p CPU per
frame 417 -> 11 ms on Iris Xe), roadmap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With MEM1-MEM3 twelve clips on a timeline still took 5.1 GB, only 2.5 GB of
it live: each decoded picture allocated three planes, each eviction freed
them, and the system allocator kept the freed pages.
filmcraft_frame::pool keeps the planes of frames the GOP caches evict (up to
192 MiB per sample type; a frame something else still holds is left alone).
The H.264 decoder, through a plane allocator hook that keeps the crate free
of dependencies, and the VideoToolbox path take their output planes from it.
perf.stats reports decode.planePoolMB and decode.planesReused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Mutation fuzzing found a corrupt H.264 transport stream whose chroma DC levels overflowed
`chroma_dc_dequant_420` ("attempt to multiply with overflow" in debug builds, a decoder worker
panic). The 4x4 / 8x8 / DC dequantisers now use wrapping arithmetic, which is what release
builds always computed. Regression test `huge_levels_dequantise_without_overflow_panics`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- I_PCM samples, deblocking lines, transform rows, block copies: fixed-size views through
`get()` + `first_chunk(_mut)` instead of slice + `try_into().unwrap()`. A truncated I_PCM
macroblock is `Error::Invalid` from the read itself.
- Decoder: a missing parameter set or a slice without a started picture is an error, not an
`expect`; the output queue pop no longer expects.
- CAVLC table builder parses its literal codes without `from_str_radix().unwrap()`.
- `#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented,
clippy::todo, clippy::unreachable)]` outside tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Interleaved before/after bench (decode, playback incl. 1/4 and draft cases,
4K cold seeks) and decoder cycle counts in docs/performance.md; threading,
draft mode and 4K numbers in crates/h264/README.md; bench cases in
docs/testing.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The column pass and the add/clip of idct4/idct8 work on whole rows so they
vectorise (same integer steps in the spec's row-then-column order; checked
against the per-column formulation on random blocks). Docs: draft decoding in
architecture.md and perf.stats fields in control-protocol.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Decoder::set_draft: non-reference pictures skip the deblocking filter and are
flagged Picture::draft; all other pictures stay bit-exact (conformance test,
single- and frame-threaded). Conformance fixtures now also decode with 3
threads. Decoder::threads / default_threads for perf.stats.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Deblocking filters all lines of an edge at once with per-lane masks (bit-exact
with the per-line filter, checked on random edges); CABAC decode_decision /
bypass select instead of branching. 4K single-thread CPU -19 %, output
identical to ffmpeg.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Runs each crate's ignored generate_fixtures test (h264, hevc, isobmff,
matroska, prores) - the same generators the oracle tests use - and
prints made/cached/skipped per fixture plus a summary. Optional crate
filter: cargo xtask fixtures hevc prores.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New dev-dependency-only crate crates/testkit (L0, no FilmCraft deps):
FILMCRAFT_FFMPEG/FILMCRAFT_FFPROBE, then PATH (incl. .exe), then known
dirs; SKIPPED lines and FILMCRAFT_REQUIRE_ORACLES=1 to make skips fail;
require_ffmpeg!/require_ffprobe!/require_oracles!; fixtures_dir() under
<workspace>/target/fixtures with per-thread temp names and atomic
generation. Oracle helpers in aac, opus, h264, hevc, h264enc, isobmff,
matroska, prores and export migrated (codecs left on its local lookup).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- CABAC engine keeps codIOffset scaled in a 64-bit register with 32-bit
refills instead of re-reading the bitstream on every renormalisation
- significance map collected as a position list; context increments from
per-category tables
- fixed-size row copies for motion-compensation windows and block copies
- explicit weighted prediction (P_Skip and B_Skip bi-prediction) synthetic
test, cross-checked with ffmpeg
- crates/h264/README.md: features, API, threading model, fixture matrix,
performance, known gaps and the plan for High 10 / 4:2:2
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- hand-built CAVLC and CABAC bitstreams (test-only CABAC encoder) with I_PCM
and P_Skip macroblocks exercising long-term references, MMCO 1-6,
frame_num gaps and reference list modification (idc 0 and 2); expected
output is known exactly and cross-checked with ffmpeg when available
- avcC configuration + length-prefixed samples with pts round trip
- corrupted / truncated / garbage input never panics or deadlocks
(single- and multi-threaded)
- Decoder::stats() coverage counters and an ignored coverage report test
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- frame-level parallelism: the calling thread handles headers, POC, DPB and
reference lists; each picture decodes as a job on a rayon pool (feature
'threads') and publishes deblocked macroblock rows into a shared Frame;
motion compensation and co-located lookups block per row on references
still being decoded (safe Rust, OnceLock per MB row); outputs are handed
out once their job has finished
- row-pipelined deblocking inside a picture (row r filtered after row r+1 is
reconstructed), picture buffer recycling, panic/error isolation per job
- faster luma/chroma interpolation (row-wise filters, window fetch), direct
interpolation into the picture for unweighted single-list blocks, merged
motion compensation for uniform direct macroblocks, 8x8 direct derivation
with direct_8x8_inference, vectorisable weighted prediction, deblocking
with per-edge parameters and cached reference identities
- conformance tests now decode every fixture single- and multi-threaded and
check pts/POC output order; ignored perf test (1080p High CRF20, 120 frames)
1080p: ~100 fps single-threaded, ~600 fps with 14 threads (M4 Pro, heavily
loaded machine).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- CAVLC residual parsing (coeff_token / total_zeros / run_before tables taken
from the spec text, level decoding with escape codes), macroblock layer for
I/P/B slices (mb_type, sub_mb_type, mb_pred, sub_mb_pred, me(v) cbp,
mb_qp_delta), I_PCM
- intra prediction: Intra4x4 (9 modes), Intra8x8 with reference filtering,
Intra16x16, chroma; constrained_intra_pred
- dequantisation with flat/custom scaling matrices; 4x4, 8x8, luma DC
Hadamard and 2x2 chroma DC inverse transforms
- inter prediction: median / directional mv prediction, P_Skip, B spatial and
temporal direct (direct_8x8_inference), B_Skip, 6-tap luma and bilinear
chroma interpolation, explicit and implicit weighted prediction
- reference lists (init + modification), DPB with sliding window, all MMCOs,
frame_num gaps, bumping / max_num_reorder_frames output order, IDR and
no_output_of_prior_pics
- public Decoder/Picture API (Annex-B and avcC input, pts passes through
reordering)
Deblocking (M2.5) follows in the next commit, so the bit-exact fixtures in
this commit are the no-deblock ones. CABAC slices return Unsupported.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New crate filmcraft-h264 (L0, depends only on filmcraft-bitstream):
- SPS incl. VUI/HRD, scaling lists with fall-back rules A/B, frame cropping,
high-profile chroma_format_idc / bit depth fields, DPB size from level
- PPS incl. transform_8x8_mode, scaling lists, slice groups (parse only)
- slice header: all fields incl. ref_pic_list_modification, pred_weight_table,
dec_ref_pic_marking
- POC types 0/1/2 (incl. MMCO5 handling and frame_num gap hooks)
- default scaling / QPc / deblocking tables from the public ITU-T spec
- ffmpeg/libx264 fixture generator for tests (skips when ffmpeg is absent)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
