Changelog

v1.4.1 — 2026-10-15

Changed

• Listing an MKV’s keyframes reads its Cues instead of the whole file, with the same list: a 4.2 GB file on a network share answers in about half a second instead of 35 s. Files without complete Cues are still read in full.

v1.4.0 — 2026-10-15

Added

• The README shows the PITO MD wordmark at its Licence section.

Changed

• An MKV or WebM file opens for video from its headers and Cues instead of reading every cluster: a 4.4 GB file on a network share shows its first frame in about 110 ms, reading 6.3 MB, and callers opening the same file at once share one read.
• Media times count from the file’s start, and Probe carries that start, so a stream that starts late lines up with the others.

Fixed

• VP9 sources export on the GPU through VA-API with or without an overlay; files whose frames keep several references used to fail with “no free VA-API surface for VP9”.
• The ac3 feature builds without hevc, vp9, gif, webp or opus.

v1.3.0 — 2026-10-14

Added

• AC-3 and E-AC-3 sound in MKV and MP4 decodes natively, ported from oxideav-ac3.
• MP3, MP2 and MP1 sound in MKV, MP4 and MOV decodes natively, ported from minimp3.
• VP9 sources export to VP9 in MP4 with smart cuts: the middles are copied and only the frames before each piece’s first keyframe are encoded.
• pcod frames turns a folder of PNG frames and an optional WAV into an H.264 and AAC MP4 with faststart, tagged BT.709.
• PNG decode, ported from zune-png, and a WAV reader.
• On VA-API with Mesa, an export with an overlay stays on the GPU: decoded, overlaid by a video-processing blend and encoded without a copy to system memory.
• Every Release carries a docs bundle for the PITO MD docs site, with guides and a generated command reference.

Changed

• The brand now reads PITO MD.
• A caption or watermark layer costs what its painted area costs: about 1 ms per change at 1080p instead of 12 ms or more, with the same output.
• Exports decode on the GPU whenever they encode on the GPU, smart-cut edges included.
• VA-API frame readback waits until a frame is no longer a reference, so it no longer holds up decoding.
• The README shows a section, a chart and a picture per feature.
• Codec gains Mp3, Ac3 and Eac3, and CodecConfig gains Vp9, so code that matches them exhaustively needs new arms.

Fixed

• export_frames stops after Canvas.frames frames instead of running until the frame callback ends.

v1.2.1 — 2026-10-12

Fixed

• An audio range or stream from an MKV PCM track starts on the right sample: a block whose millisecond timestamp rounds to the start could make the seek land after it and shift the range by up to one block.

v1.2.0 — 2026-10-12

Exports from the app’s own frames at any size, a public blend without copies, export progress that follows real work, GPU stills about 3x faster, and the audit’s second hardening wave.

Added

• export_frames writes caller-made frames at their own size and frame rate to MP4, with the same GPU-first H.264 routing as export and premixed AAC sample-exact against the video.
• Blender is public: kept across frames with the same Arc<Rgba> overlay, it blends without copying it.

Changed

• Export progress counts copied frames as they’re written and encoded frames as they’re encoded, then the final file pass, so it rises steadily to 1.0.
• GPU keyframe stills read the decoded VA-API picture directly instead of through the streaming staging path: time to first frame on the GPU about 3x faster (10.5 to 3.6 ms at 1440p in a same-run comparison).
• HEVC pictures above level 6.2’s luma size (8192x4352) are refused, and VA-API builds only the tile boundaries it sends.

Fixed

• MKV resync never rescans bytes it has scanned, so a damaged file can’t make a seek take quadratic time; warnings are capped.
• MP4 sample tables are bounded by their chunk table and the file’s bytes across tracks.
• packet on the MP4 and MKV readers looks samples up without indexing; H.264 slice-group maps are capped by the picture size; EBML floats of an odd size are ignored.

v1.1.1 — 2026-10-12

Hardening for files from anywhere: malformed headers are refused instead of crashing or allocating without bound, and every parser is fuzzed on CI.

Added

• media::Pcm, the PCM sample converter the readers use.
• Fuzzing on GitHub CI for the MP4 and MKV readers, probe, the H.264, HEVC and VP9 decoders, AAC, Opus and PCM: a short run on every push and a longer one nightly.

Fixed

• H.264 crop offsets are checked without wrapping, so a picture is never sliced outside its planes; the HEVC conformance window always stays inside the picture.
• An indexed MKV open keeps at most 2^26 samples per track and no more samples than half the file’s bytes; laced frames need at least one byte each.
• Fragmented MP4 runs of empty samples are bounded.

v1.1.0 — 2026-10-12

One home for each shared public type. Every old path still compiles, marked deprecated.

Added

• media::length_prefixed_to_annexb, media::threads and mp4::AvcConfig::parameter_sets_annexb, which pcod now uses instead of its own copies.

Changed

• Rgba, Rate and Bitstream each live once in media. filter::Rgba, webp::Rgba, h264::Rate, gpu::Rate, h264::Bitstream and gpu::Bitstream are deprecated aliases of them, so existing code keeps compiling and values are interchangeable.
• media::Rate and media::Bitstream derive Hash, and media::Rgba has stride().

v1.0.5 — 2026-10-12

Recordings of 100 GB and more: memory that doesn’t grow with the file, fast probes with or without Cues, and exports that use about a quarter of the memory.

Added

• audio_mix_stream streams the mix of every audio track, with the same rules as audio_mix.

Changed

• probe of an MKV without Cues reads its head and tail: milliseconds instead of a pass over the whole file. Probe.frames is None there too; frame_count gives the count.
• keyframes and frame_count read block headers in one sequential pass with bounded memory, and seeks in a Cue-less MKV bisect its clusters.
• The full MKV index the media front builds holds video blocks only, about half the memory.
• Exports write each re-encoded edge as it finishes, and a track mix streams in one-second steps, so a long export’s peak memory stays near 0.2 GB.
• Audio reads use large sequential windows, at disk speed.

Fixed

• An exported MP4’s mvhd, tkhd, mdhd and edit-list durations reach the last frame’s end; they could be two frames short.
• Every feature builds on its own.

v1.0.4 — 2026-10-12

Exports of MKV recordings smart-cut, every join is exact, and probing a large MKV reads only its head.

Added

• frame_count gives a file’s exact frame count on demand.
• Exported.reasons says why a piece or a whole export was re-encoded instead of copied.

Changed

• MKV sources smart-cut into MP4: the stream between cuts is copied, and only the frames around each cut are re-encoded.
• probe of an MKV with Cues reads its head, Cues, first blocks and last cluster instead of every block: a few milliseconds whatever the file size. Probe.frames is None there; frame_count gives the exact count.
• VA-API H.264 and HEVC refuse pictures below 352x240, which then run on the CPU: smaller ones faulted AMD’s video engine.

Fixed

• The ffmpeg route’s join gives every part its exact duration, so a long edit can’t drift; copied MKV parts start on the right frame, tiny parts open, and the audio mix trims to exact sample counts.
• A track mix ends on the video’s last sample instead of up to one AAC frame past it, and reads MKV audio without ffmpeg.

v1.0.3 — 2026-10-12

Long MKV recordings open, play audio and give stills without reading the whole file.

Added

• audio_stream streams a track’s audio in sample-aligned blocks from any start time, with the same resample and downmix as audio_pcm_range, from MP4 and MKV (PCM, AAC, Opus) or through ffmpeg.
• frames_at_seek with Seek::Exact gives the frame at the asked time, decoding forward from its keyframe; Seek::Keyframe keeps the fast keyframe still.
• export_with_audio takes the app’s own premixed audio, one buffer per piece, and writes it as AAC, sample-exact against the video.
• first_frame, frames_at and filmstrips read MKV and WebM, seeking through their Cues.

Changed

• Audio and stills open an MKV from its head (SeekHead, Tracks, Cues) instead of indexing every block: about 0.1 ms instead of hundreds of milliseconds to seconds on long recordings.

Fixed

• A range that runs past the end of a track returns only the samples the track has, so a buffer shorter than asked means the track ended; before, it was padded with silence.

v1.0.2 — 2026-10-12

Native audio from every container we read, and a simpler crate: about 7,200 lines fewer, with every output byte-identical to v1.0.1.

Added

• MKV, WebM, MP4 and MOV audio in PCM (s16, s24, s32, f32, f64), AAC and Opus decodes natively through audio_pcm, audio_pcm_range and audio_mix.
• AudioTrack::native, Probe::video_native and Probe::unsupported say which streams a file holds that pito-codecs doesn’t decode, by codec name.
• A Formats table in the README, with the formats it doesn’t read.

Changed

• The GPU routes parse H.264 and HEVC with the CPU decoders’ own parameter sets, slices and DPB, including their overflow checks.
• The H.264, VP9, Opus and AAC encoders use their decoders’ spec tables and helpers instead of copies; one CABAC table set serves H.264 and HEVC.
• One AVX2 check, one reader cache, one smart-cut plan, one route picker and one remux planner across the crate.

Fixed

• AAC and Opus in MKV line up with the source to the sample (MKV’s millisecond block times had put them up to 24 frames off).

v1.0.1 — 2026-10-12

H.264, HEVC and VP9 decode speed-ups and MKV and WebM playback, with one benchmark run on this code behind every chart and number in the README.

Added

• MKV and WebM play through decode_video and decode_video_gpu, on VA-API or the CPU.

Changed

• VP9 decode: 8-bit streams decode into 8-bit planes, and motion vector candidates read their neighbours in place.
• VP9 decode: AVX2 loop filter specialised by width and direction, with its decisions per cell pair shared by U and V.
• VP9 VA-API decode starts each shown frame’s read-back copy early.
• H.264 decode: inter prediction compiled per ISA, with one row wait per block.
• H.264 decode: spatial direct reads colocated motion only when it matters.
• HEVC: 10-bit output rows are copied as bytes on little-endian targets.
• bench: --only takes a full row id.
• README: every Speed chart comes before the table, and the README, CHANGELOG and SECURITY.md say what the code does today.

Fixed

• VP9: a failing tile or loop-filter worker becomes an error, never a panic.
• VP9 lints clean on arm64 targets.

v1.0.0 — 2026-10-11

Every format in the README’s Status decodes and encodes, on Linux, Windows and macOS. Opus speech encoding (SILK and hybrid) is experimental.

Added

• pcod is a real binary (behind the cli feature), released for Linux x86-64, Windows x86-64 and macOS on Apple Silicon, with SHA256SUMS, LICENSE, README and NOTICE in each archive.
• The Opus SILK encoder, experimental: it decodes clean in libopus but doesn’t sound as good as libopus.
• An “Against ffmpeg” summary in the README, with every win and loss from one benchmark run.
• GPU playback through the media front: decode_video_gpu keeps VA-API frames on the GPU, scaled there, exported as DRM PRIME, with CPU frames when the route falls back.
• Ranged audio and a track mix in the media front: audio_pcm_range and audio_mix.
• VA-API VP9 decode, bit-exact against ffmpeg.
• WebP lossy (VP8) encode, and WebP lossless, alpha and animation encode.
• The Opus encoder’s CELT pitch prefilter, tf analysis and spread decision.
• Filters: BGRA and BGRX, interpolated chroma, chroma siting, NEON kernels.
• CI on Linux x64 and arm64, Windows and macOS, with VideoToolbox running on the macOS runner, and a check that a tag matches the crate’s version.

Changed

• H.264 decode: AVX2 deblocking, motion compensation and weighted and bi-prediction into the planes.
• HEVC decode: AVX2 dequantisation, batched luma deblocking, CTB rows published in two parts and row-aligned slices decoded in parallel.
• H.264 encode: mbtree lookahead, graded scene cuts and calibrated CRF.
• Errors name the file they failed on, and every example and pcod print one line and exit 1 instead of panicking.
• Damaged MP4, MKV, AAC, Opus, Ogg, GIF and WebP input returns an error instead of panicking.
• H.264 encode: one slice per frame on frame threads with row-published references.
• VA-API encode: rate control starts with a full one-second buffer, and per-frame quality matches ffmpeg’s VA-API encoders.
• VA-API read-back decode copies into host memory ahead of time, faster than ffmpeg at 1080p and 1440p.
• VP9 decode: AVX2 motion compensation, loop filter and inverse transforms, faster entropy decoding.
• The README’s routes are the GPU, then our CPU code, with their fallbacks; INSPIRATION.md says why; the charts are drawn per unit with paired bars and a centred ratio chart.

Fixed

• VideoToolbox outputs pictures in display order with the right timestamps.
• The GPU playback test expects GPU surfaces only on VA-API.
• WebP VP8 token counts.

v0.1.2 — 2026-10-11

Fixed

• The crate’s version is 0.1.2, matching its tag, so pcod --version reports it. Otherwise the same as v0.1.1.

v0.1.1 — 2026-10-11

Changed

• The command line is pcod, and the README GIF shows it.
• Opus encode: VBR by default, 5 s chunks encoded in parallel and byte-identical to streaming, an Ogg writer.
• H.264 decode: macroblock state in a four-row ring, one colocated row lookup per macroblock.
• The bench has a --quick mode.

Fixed

• Tests write shared media under per-process names, so parallel test processes don’t race.
• Media Foundation and VideoToolbox refuse VP9 and Opus instead of failing to build.

v0.1.0 — 2026-10-11

Added

• The first release: H.264 and H.265 decode and encode on the CPU and through VA-API (H.265 encode on the GPU only), VP9 decode and a first VP9 encoder, AAC-LC, Opus decode and CELT encode, MP4/MOV and MKV/WebM, GIF, WebP decode, filters and a media front for apps.