CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.4.0] - 2026-04-19

Added

• MBAFF (Macroblock-Adaptive Frame-Field) interlaced support — the decoder now handles interlaced H.264 content encoded with mb_adaptive_frame_field_flag. Both frame-coded and field-coded MB pairs are fully supported:
  • Pair-based MB addressing with mb_field_decoding_flag decode (CABAC contexts 70-72 and CAVLC 1-bit)
  • MBAFF neighbor derivation per spec Tables 6-3/6-4 with y-coordinate remapping for all 4 frame/field mode combinations
  • MVy scaling at cross frame/field boundaries (×2 field→frame, /2 frame→field)
  • All CABAC/CAVLC context functions updated for MBAFF pair-based addressing
  • Field-coded pixel layout with doubled stride and field-line offset
  • Field-aware motion compensation via luma_mc_stride with per-field reference buffer offset for top/bottom field access
  • Field reference lists: num_ref_idx doubled for field-coded MBs, ref_idx maps same-parity and opposite-parity fields
  • Field-coded CABAC significance/last coefficient contexts (ctxIdx 277+/338+ per spec Table 9-34)
  • Field-coded CABAC neighbor addressing: bottom field MBs use same-field MB from above pair (not top of current pair)
  • MBAFF-aware deblocking filter with pair-based iteration and correct left/above neighbor addressing
  • MBAFF-aware POC computation: min(TopFieldOrderCnt, BottomFieldOrderCnt) for POC types 0 and 1
• 15 new MBAFF-specific byte-exact tests covering CAVLC/CABAC, I/P/B slices, frame-coded and field-coded pairs, deblocking, High profile 8x8 DCT, and implicit weighted bi-prediction. Total test count: 178.

Fixed

• POC computation for pic_order_cnt_type 0 and 1 now correctly returns min(TopFieldOrderCnt, BottomFieldOrderCnt) instead of just TopFieldOrderCnt. Fixes implicit weighted bi-prediction for MBAFF B-frames with non-zero delta_pic_order_cnt_bottom.
• Above-right MV neighbor for MBAFF bottom MBs no longer reads from undecoded pairs due to mb_slice_id initialization matching this_slice_id=0. Falls back to above-left correctly.
• Fuzzing robustness: fixed multiple integer overflow panics found by fuzzing on malformed bitstreams:
  • IDCT 4x4 and 8x8 add/sub overflows (use wrapping arithmetic)
  • POC type 1 multiplication overflow and MSB computation overflow
  • POC shift overflow (clamp shift amount)
  • Dequant DC multiply overflow
  • SPS scaling list delta add overflow
  • Deblocking filter multiply overflow
  • SPS width overflow
• Bounds checking: fixed multiple out-of-bounds panics on malformed bitstreams:
  • Chroma MC reference plane boundary checks
  • CABAC I_PCM frame buffer index validation
  • Bitstream reader position validation
  • Empty reference list safe indexing
  • CAVLC coefficient position underflow guard
  • Missing bounds/range validation on values from malformed bitstreams

[0.3.0] - 2026-04-09

Added

• OrderedDecoder — wraps Decoder with a built-in reorder buffer that emits frames in display order automatically. Handles IDR boundary tracking internally, eliminating a common pitfall where callers would tag the last B-frame of a GOP with the wrong GOP id and produce visible glitches at scene cuts. Recommended for most users.
• AVCC (length-prefixed) NAL parser for MP4/MKV containers:
  • nal::parse_avcc(data, length_size) — parse a length-prefixed sample
  • nal::parse_avcc_config(avcc_box) — parse an MP4 avcC configuration record and extract SPS/PPS NALs
  • nal::AvccConfig struct exposing length_size, sps_nals, pps_nals
• NEON SIMD acceleration on aarch64 for luma half-pel filters (half_pel_h, half_pel_v, quarter-pel paths) and chroma bilinear interpolation. Requires no opt-in — enabled automatically on aarch64.
• Comprehensive doc comments on the entire public API. Crate-level docs, module-level intros, and runnable examples render on docs.rs.
• Test for OrderedDecoder validating display-order output against the manual sort baseline on a real preset_medium stream.

Changed

• Performance: 1080p decode improved from ~48 fps to ~67 fps (28% speedup) on Apple Silicon thanks to NEON luma half-pel intrinsics. 1080p @ 60 fps target achieved.
• Performance: 720p P-only improved from ~186 fps to ~278 fps; 720p with B-frames from ~190 fps to ~299 fps.
• README.md restructured to lead with OrderedDecoder as the recommended entry point. The lower-level Decoder is documented as an advanced option for callers who need raw decode order.
• parse_annex_b and AVCC parsers now share a common parse_nal_bytes helper for NAL header parsing and emulation-prevention removal.

Fixed

• Build clean on stable Rust with no clippy warnings.

[0.2.0] - 2026-04-04

Added

• Initial release on crates.io. Pure Rust H.264 decoder supporting Baseline, Main, and High profile (8-bit 4:2:0 progressive) with both CAVLC and CABAC entropy coding.
• Full support for I/P/B-frame decoding, multi-reference prediction with ref_pic_list_modification, multi-slice frames, weighted prediction (explicit and implicit), long-term references, MMCO ops 1-6, deblocking filter, intra prediction (I4x4/I8x8/I16x16, all 9 modes for 4x4 and 8x8), spatial and temporal direct mode for B-slices.
• Annex B bitstream parser (nal::parse_annex_b) with zero-copy emulation prevention removal.
• 119 unit tests, including 73 byte-exact stream tests against FFmpeg, with coverage up to 1080p (CABAC, CABAC+deblock, CAVLC) using SHA-256 hash comparison to keep test data small.
• License: dual MIT / Apache-2.0.

Performance

• ~48 fps at 1080p (mandelbrot content, x264 --preset medium bframes=3 ref=2)
• ~186 fps at 720p P-only / ~190 fps at 720p with B-frames
• See BENCHMARK.md for detailed profile breakdown and methodology.

Not implemented

• Interlaced coding (MBAFF, field pictures)
• High 10 / 4:2:2 / 4:4:4 profiles
• SP/SI slice types
• Slice groups / FMO