miniav_tools_ffmpeg 0.5.8

FFmpeg-backed codecs/muxers for miniav_tools (libavcodec/libavformat).

Changelog

0.5.8

• Increment downstream deps
• Increment downstream deps
• Increment downstream deps

0.5.7

• Version alignment with the miniav_tools 0.5.7 family release.

0.5.6

• Host the container demuxer on spawn, so it runs on an isolate natively and a Web Worker on web through one API.

0.5.5

• IsolateDemuxer is now hosted by package:spawn. The handshake, the pending-request map, id correlation, the [op, id, ...] reply protocol and the bounded-close-then-kill teardown were all hand-rolled here and are now the worker library's job. Behaviour is unchanged - the full demuxer roundtrip suite passes untouched, including the starved-live-worker shutdown and the stalled-mid-probe open timeout.

  One thing deliberately did NOT move: closing the byte pipe BEFORE closing the worker. Killing an isolate does not preempt a blocking native call, so only the code that owns the pipe can free a starved av_read_frame. No worker library can do that for us.

  New src/demux_protocol.dart gives TrackInfo and EncodedPacket a byte encoding, because spawn carries portable values or WireMessages rather than deep-copying arbitrary objects. That is more work than an isolate send, and it is what makes the protocol legible to a non-Dart host later.

  Adds a dependency on spawn.

• Native av_log forwarding moved from a NativeCallable to a Dart native port (shim ABI 20 -> 21, new exports miniav_shim_init_dart_api / miniav_shim_set_log_port; tool/shim_c/dart_dl/ vendors the BSD-licensed Dart SDK dynamic-linking API). av_log_set_callback is a PROCESS-GLOBAL registry: a NativeCallable installed there is owned by ONE isolate, and when that isolate exits the VM deletes the trampoline while FFmpeg keeps the pointer — the next log line from a codec thread aborted the whole process (runtime_entry.cc: Callback invoked after it has been deleted, typically under avcodec_open2). A whole-suite dart test hit it every run, because every test FILE is a separate isolate in ONE VM process; dart test now completes. setFfmpegLogCallback gained an optional level: argument and registers a ReceivePort per isolate. Semantics unchanged and still documented: PROCESS-GLOBAL, LAST WRITER WINS — with several isolates registered only the most recent receives lines. Trap: the message is carried as raw bytes, not a Dart_CObject string, because FFmpeg emits Latin-1 filenames on Windows and Dart_CObject_kString requires valid UTF-8. The v8/v9 function-pointer exports remain for non-Dart embedders; Dart no longer binds them.

• FfmpegDemuxer now fills VideoTrackInfo.rotationDegrees from the container's display matrix (shim ABI 18 -> 20, new exports miniav_shim_stream_rotation_degrees / _set_rotation_degrees). It previously reported 0 for every container, so a 90-degree phone MP4/MKV/WebM played back sideways. Value is degrees CLOCKWISE in {0, 90, 180, 270}; decoded frames stay in coded orientation, so width/height remain the coded dimensions and the consumer applies the turn. Only a plain quadrant turn is reported: flips, scales and arbitrary transforms read as 0, matching miniav_tools_codecs' MP4 demuxer, so the two never disagree about the same file. Trap: av_display_rotation_get cannot be used for this — it derives the angle from atan2, so a mirrored (pure horizontal flip) clip comes back as 180 and would render upside down. Second trap: FFmpeg 7 removed AVStream.side_data / av_stream_get_side_data, so stream side data lives in codecpar->coded_side_data.

• FfmpegMuxer now WRITES VideoTrackInfo.rotationDegrees as a container display matrix, so a rotation read by the demuxer survives a remux instead of being silently dropped (it plays sideways otherwise). Set before avformat_write_header, after codecpar is populated — avcodec_parameters_from_context overwrites the side data. A rotation that is not a multiple of 90, or a missing shim, throws rather than writing an upright file.

• hook/build.dart registers tool/shim_c/** as build dependencies. Without them an edited shim.c kept serving the previously built DLL, and a stale ABI makes FfmpegShim.tryLoad() return null — which takes the whole backend down (every open throws "shim not loadable"), not just the new export.

0.5.4

• Internal pins are now caret ranges, not exact versions. Exact pins made every patch cascade: publishing miniav_tools_platform_interface 0.5.3 made the already-published miniav_tools 0.5.3 and miniav_tools_ffmpeg 0.5.3 unsatisfiable next to it, because they pinned 0.5.2 exactly and nothing in the set could move independently. dart pub publish warns about this. release.py sync now normalises to caret so it cannot recur.
• Republished: 0.5.3 shipped pinning miniav_tools_platform_interface: 0.5.2, which cannot resolve alongside any package needing 0.5.3.

0.5.3

• Declare the new PlatformEncoder GPU-input capabilities. FfmpegD3d11HwEncoder reports both supportsD3d11SharedHandleInput and supportsD3d11TextureInput; the CPU-fed encoders report neither. Behaviour is unchanged — this replaces the recorder's platform is FfmpegD3d11HwEncoder check with the same answer, expressed as a capability other encoders can give.

• Demuxer + live-stream muxing for miniav_player stream/file playback (shim ABI 14 -> 15):

  • FfmpegDemuxer (was a throwing scaffold) + IsolateDemuxer — file / bytes / live byte-stream → EncodedPackets with pts/dts/duration rescaled to microseconds (inverse of the muxer's rescale). createDemuxer returns the isolate host by default (av_read_frame blocks on live input and must be off the UI isolate); backendOptions {'sw_isolate': '0'} = in-isolate (file/bytes only). durationUs / isSeekable / keyframe seek.
  • Shim blocking byte pipe (miniav_shim_bytepipe_*) — feeds libavformat from a live Stream<List<int>> with pause/resume backpressure; closing the pipe unblocks a natively-stuck av_read_frame (the only way — Isolate.kill can't preempt synchronous FFI).
  • Seekable in-memory IO: miniav_shim_open_input_bytes (read+seek AVIO over a C-owned copy, so moov-at-end MP4 probes) and ..._memsink_* (seekable output for BytesMuxerOutput so +faststart works). FfmpegMuxer now supports BytesMuxerOutput (seekable memsink) and CallbackMuxerOutput (streaming byte-pipe, streamable containers).
  • Container.fmp4 now sets real fragmentation movflags (+frag_keyframe+empty_moov+default_base_moof + optional frag_duration) — init segment up-front, live-streamable.
  • Open-timeout robustness: a live stream stalling mid-probe surfaces a CodecInitException (default 15 s, backendOptions {'open_timeout_ms'}) instead of hanging open() forever.

• Decode side for the miniav_player work (shim ABI 13 -> 14):

  • FfmpegAudioDecoder — AAC/Opus/MP3/Vorbis/FLAC -> interleaved f32 (planar formats interleaved during the mandatory AVFrame copy-out; stream rate/channels read per-frame via new shim getters miniav_shim_frame_sample_rate / _nb_channels).
  • miniav_shim_codec_set_extradata: DecoderConfig.extraData / AudioDecoderConfig.extraData now actually reach the codec context (avcC / hvcC / AAC ASC / OpusHead) instead of being ignored.
  • IsolateVideoDecoder / IsolateAudioDecoder — worker-isolate decoder hosts mirroring IsolateSoftwareEncoder's TransferableTypedData protocol; createDecoder/createAudioDecoder return them by default (backendOptions {'sw_isolate': '0'} keeps the in-isolate path).
  • Fixed FfmpegSoftwareDecoder pts: decoded frame.pts is already in microseconds (packets are fed us with no time_base), so the old * (1e6/30) scaling corrupted timestamps; now passthrough with a delta-extrapolating AV_NOPTS fallback.

0.5.2

• Isolate-hosted CPU-fed hardware encoder (fixes the recording freeze on the software/CPU-fed path). IsolateSoftwareEncoder.open gained hwVendorOrder / requireHardware: the worker now tries each CPU-fed hardware vendor (looping FfmpegHwEncoder.openWith) before the software encoder, all on the worker isolate. Two things this unblocks:
  • QSV / MediaFoundation now initialise. Their COM objects require the MTA apartment; Flutter's UI isolate is STA (miniav_shim_ensure_mta → RPC_E_CHANGED_MODE), so on the UI isolate they failed at avcodec_open2. The worker runs on a fresh OS thread that enters MTA, so e.g. h264_mf and h264_qsv finally open. (Verified on an AMD iGPU box: the worker opens h264_mf where AMF fails its NV12 pool and NVENC has no device.)
  • No UI freeze. The blocking avcodec_open2 probes and the per-frame encode both run off the calling (UI) isolate; frames cross as TransferableTypedData (~1 ms at 720p). Previously the Stage-A hardware open ran synchronously on the UI isolate. FfmpegBackend.createEncoder now routes the post-zero-copy path through this isolate encoder (removing the UI-isolate Stage-A open) and derives the vendor order from the capture adapter via the new hwVendorOrderForDevice (AMD→MF-before-AMF to dodge AMF's silent-black on AMD, Intel→QSV, NVIDIA→ NVENC). IsolateSoftwareEncoder reports its actual encoder's acceptsYuv420pPlanes and an activeEncoderDescription. The sw_isolate:'0' escape hatch keeps the legacy in-isolate path.

0.5.1

• D3D11 HW-encoder vendor order now follows the injected device's adapter. FfmpegD3d11HwEncoder.open with an injected ID3D11Device (the recorder's zero-copy path) previously always tried the global default order (NVENC → AMF → QSV → MF), so on an AMD/Intel device it wasted an NVENC avcodec_open2 — "OpenEncodeSessionEx failed: no encode device" — before falling through to the native vendor. It now derives the order from the device's DXGI vendor (AMD→AMF, Intel→QSV, NVIDIA→NVENC, + MediaFoundation fallback), so the native encoder is tried first and no cross-vendor attempt is made. The adapter-aware helper already backed the probe/warm-up paths; this wires it into open() too. vendorOrder is now optional; passing it explicitly still overrides.

0.5.0

• Software encoding no longer freezes the app: createEncoder's software fallback now returns IsolateSoftwareEncoder, which hosts the exact same FfmpegSoftwareEncoder on a long-lived worker isolate. The synchronous libav encode (tens of ms/frame at 720p+) runs off the calling (UI) isolate; frames cross as TransferableTypedData (~1 ms), packets come back the same way, and codec extradata (SPS/PPS) is captured at open. Opt out with backendOptions: {'sw_isolate': '0'} (returns the classic in-isolate encoder, e.g. for tests that wire its FfmpegEncoderBridge directly).
• AMF (AMD) D3D11 zero-copy: BGRA input to h264_amf/hevc_amf is broken on real AMD hardware — AMD iGPUs reject BGRA frames at avcodec_send_frame ("Unknown error"), and some driver combos accept them but silently encode black. AMF now always uses the NV12 + D3D11 VideoProcessor pool (AMF's native input format; same fixed-function path QSV/MF use).
• NV12 hwframes-pool creation is now adaptive: bind-flag sets are tried from richest (Intel's SR|RT|DECODER|VIDEO_ENCODER) down to minimal (SR|RT), re-allocating the frames context per attempt — AMD rejects DECODER|RENDER_TARGET combinations with E_INVALIDARG. If no set works (some adapters cannot create NV12 texture arrays with RENDER_TARGET at all), the vendor is reported cleanly unavailable and encoding falls back to the next vendor / CPU — correct output instead of black video.

0.4.10

• Software encoder (FfmpegSoftwareEncoder) now accepts pre-converted planar YUV420P frames (FrameSource.yuv420p) directly — the planes are copied straight into the AVFrame with no internal RGBA→YUV conversion. Advertised via acceptsYuv420pPlanes => true. Lets the recorder convert RGBA→YUV420P on the GPU and skip the per-pixel Dart conversion on the software-encode path. The HW encoders (d3d11, generic hw/NV12, nvenc) keep their existing RGBA/NV12 input (acceptsYuv420pPlanes => false).

0.4.9

• Licensing: switch to the BtbN LGPL FFmpeg build (kFfmpegLicense). The previous GPL build (-gpl-shared) links libx264/libx265 and makes the whole binary GPLv2+, imposing copyleft on downstream products. The LGPL build keeps libav* under LGPL-2.1 (safe for proprietary dynamic linking). The cache dir is namespaced per-licence (latest-lgpl) so the change forces a fresh download instead of reusing a cached GPL install.
• Consequences of LGPL (no libx264/libx265): software H.264 now comes from libopenh264 (BSD) and Windows MediaFoundation h264_mf; the MF spec no longer forces hw_encoding=1, so its software MFT works as a true CPU fallback. There is no software HEVC — createEncoder falls back to a downscaled (≤4096px) H.264 stream when a >4096px or HEVC request has no hardware encoder. AV1 (SVT-AV1) and VP9 (libvpx) remain available at any resolution.

0.4.8

• Increment to keep in step with others.

0.4.7

• Remove the inert auto-register top-level final and fix the docs: importing the package never actually registered the backend (Dart top-level finals are lazy and nothing read it). Call registerFfmpegBackend() explicitly (idempotent); miniav_recorder does it in warmup() and start().

0.4.6

• Replace all direct stderr.writeln diagnostics with a levelled log hook: setFfmpegToolsLogCallback / setFfmpegToolsLogLevel (uses MiniAVLogLevel). Default sink is print — dart:io stdio writes crash console-less Windows GUI apps with an uncatchable async FileSystemException (errno 6), which also masked the real error when the FFmpeg auto-download failed.
• README: document FFmpeg warmup (ensureFFmpegLoaded, MiniAVTools.warmup(), FfmpegDownloader.ensureFfmpeg), correct the cache paths (%LOCALAPPDATA%\miniav_tools\ffmpeg etc. — not ~/.miniav_tools_ffmpeg), and describe the new logging hook.

0.4.5

• Version bump for coordinated release with miniav_recorder 0.4.5.

0.4.4

• fixing recorder loopback drift

0.4.3

• audio data issue, increment miniav

0.4.2

• fix timing issue

0.4.1

• fix audio timing, add frame duplication

0.4.0

• fix frame rate scheduling

0.3.12

• fused shader cache fix

0.3.11

• Use GPU until we cant.

0.3.9

• AMF Fix, fix unknown audio error

0.3.8

• Fix vendor Order

0.3.7

• fix NV12 path

0.3.6

• fix cpu path

0.3.5

• fix recorder sync drift

0.3.4

• attempt fix resolution issue

0.3.3

• fix precheck

0.3.2

• fix property

0.3.1

• fix scaling crazy, attempt fix other HW encoders

0.3.0

• fix recorder logging, Tier A path, deps to 1.5.0

0.2.21

• increments minigpu to 1.4.15

0.2.20

• increments minigpu to 1.4.14

0.2.19

• increments minigpu to 1.4.12

0.2.18

• increments minigpu to 1.4.11

0.2.17

• increments minigpu to 1.4.9

0.2.16

• increments minigpu to 1.4.8, hopefully fix cpu fallback

0.2.15

• increments minigpu to 1.4.7

0.2.14

• fixes unicode, increments minigpu to 1.4.7

0.2.12

• Increment minigpu to 1.4.6

0.2.11

0.2.9

0.2.8

• add RecorderLogSource.minigpu: routes native minigpu/Dawn log lines through the unified Recorder log callback; Recorder.minigpuLevelFor public helper for tests; 12 new tests in log_level_test.dart

0.2.7

• fix FormatException on non-UTF-8 bytes in MiniAV log callback: use Utf8Decoder(allowMalformed: true) instead of toDartString()
• Increment minigpu to 1.4.4

0.2.6

• fix: add missing dart:convert import for Utf8Decoder (compilation error introduced in 0.2.5).

0.2.5

• fix: FfmpegShim.setFfmpegLogCallback now uses Utf8Decoder(allowMalformed: true) when decoding native C strings, so FFmpeg log messages containing non-UTF-8 bytes (Latin-1 filenames, Windows-1252 error strings, etc.) no longer throw FormatException: Unexpected extension byte.

0.2.4

• fix: FfmpegShim.tryLoad() no longer poisons its cache when called before FFmpeg has been loaded. The shim DLL imports avcodec/avutil; if those aren't loaded yet on Windows, the shim load fails and previously the failure was cached forever — breaking the audio encoder when Recorder.setLogLevel/setLogCallback was called early. tryLoad now skips silently (without caching) until FFmpeg is loaded.

0.2.3

• FFmpeg log forwarding shim bridge (miniav_shim_set_ffmpeg_log_callback, miniav_shim_set_ffmpeg_log_level); FfmpegShim.setFfmpegLogLevel, setFfmpegLogCallback; ABI bumped to 8; export FfmpegShim from barrel
• add unified Recorder.setLogLevel and Recorder.setLogCallback routing all native logs (MiniAV + FFmpeg) through a single Dart callback

0.2.1

• fixes dawn find issue

0.2.0

• add more quality control, fix ffmpeg usage issue

0.1.9

• fixes timestamp issues

0.1.8

• recorder scaling, warmup feature

0.1.7

• adds transform effects

0.1.6

• adds clip buffer

0.1.5

• fix loopback issue

0.1.4

• fix loopback issue, add tests

0.1.3

• update with fixes

0.1.2

• recorder sync and multi files

0.1.1

• updated to latest miniav/minigpu deps

0.1.0

• Initial release. FFmpeg-backed encoder, decoder, muxer and demuxer registered against miniav_tools_platform_interface.
• Software encode for H.264 / HEVC / VP9 / VP8 / AV1 / MJPEG / ProRes via libavcodec.
• Stage A hardware encode (CPU frame in, encoded packet out) using NVENC / AMF / QSV / VideoToolbox / MediaFoundation / V4L2 M2M.
• Stage B zero-copy D3D11 encode on Windows: takes an ID3D11Texture2D NT shared handle directly without any PCIe transfer or colour conversion.
  • FfmpegD3d11HwEncoder.openWith(cfg, vendor, existingD3d11Device, sourceTextureFormat) lets callers inject their own ID3D11Device* so the encoder runs on the same DXGI adapter as an external GPU API (Dawn / WebGPU). Required for cross-API NT-handle sharing — different adapters fail with E_INVALIDARG.
  • sourceTextureFormat selects bgra (default; matches DXGI screen capture and minigpu SharedOutputTexture) or rgba. The hwframes pool is allocated with this sw_format so CopySubresourceRegion from the caller's texture stays in the same DXGI type group.
• MP4 / Matroska / MPEG-TS muxer with byte-buffer, file and streaming callback outputs. Container-level fragmenting for low-latency streaming.
• Auto-download of FFmpeg shared libraries (BtbN GPL builds) on first run; set MINIAV_TOOLS_FFMPEG_NO_AUTODOWNLOAD=1 to opt out.