Module mf

Windows Media Foundation backend (IMFSourceReader). Native software decode, no external deps, instant seeks.

Contract (see media/mod.rs):

• probe(path) -> Asset with duration, size, fps, every audio stream (language/title if available).
• open_video(path) -> VideoSource producing top-down RGBA8 at the requested size (RGB32 output type + MF_SOURCE_READER_ENABLE_ADVANCED_VIDEO_PROCESSING; handle negative stride / bottom-up; scale to (w,h) — either via the reader’s output type or a simple resize in Rust).
• open_audio(path, stream) -> AudioSource producing stereo f32 @ 48 kHz (PCM float output type; stream = Nth audio stream in container order; resample/upmix in Rust if MF refuses). Seeking: SetCurrentPosition then decode forward until pts >= t. Sequential reads must not seek. COM: MFStartup / CoInitializeEx per thread (decoders live on the thread that created them; they are Send in the sense that a thread creates and owns them — mark types unsafe impl Send if needed, see ARCHITECTURE.md).

Structs

MfAudio 🔒

MfVideo 🔒

Resamp 🔒

Linear resampler state carried across decoded buffers.

SharedDeviceManager 🔒

IMFDXGIDeviceManager is documented as safe to share and call from multiple threads (that’s its purpose: every decoder MFT on every thread opens a handle to the same device through it).

StreamInfo 🔒

Constants

AUDIO_FWD_FRAMES 🔒

Audio: read forward (no seek) when the target is less than this many output frames past the FIFO.

AUDIO_PREROLL_FRAMES 🔒

Audio: seek this many output frames early (imprecise MP3 seeks), discard by timestamp.

DXVA_MIN_PIXELS 🔒

DXVA pays for itself only on big frames: a hardware decode is a GPU round-trip per frame (submit, decode, video-process, sync, copy back), which costs more than just software-decoding anything SD-sized. Attach the D3D manager at 720p and up.

GUID_NULL 🔒

HNS 🔒

100 ns ticks per second (MF time base).

MAX_READS 🔒

Safety cap on ReadSample calls per request (no busy loops on a misbehaving source).

STABLE_REQS 🔒

Consecutive identical size requests before the decoder’s output type follows them (see ensure_decode_size).

VIDEO_FWD_HNS 🔒

Video: read forward (no seek) when the target is less than this far ahead of the cached frame.

Functions

audio_format 🔒

(sample rate, channels) of the reader’s current output type.

bilinear 🔒

box_down 🔒

Area-average downscale. xs/ys are the source edges per destination column/row (len = n+1).

codec_name 🔒

MF subtype -> ffmpeg-style codec name (cheap; “” when unknown).

convert_rows 🔒 ^⚠

Copy h rows of w BGRX pixels starting at top (row stride pitch, signed) into top-down RGBA.

duration_secs 🔒

dxgi_device_manager 🔒

D3D11 device + DXGI device manager for hardware-accelerated decode (DXVA). Built once per process; every source reader attaches to the same manager. None when the adapter/driver can’t do it — callers fall back to the source reader’s normal software decode, so this is purely a speed opt-in.

err 🔒

frame_rate 🔒

get_string 🔒

hns 🔒

init 🔒

Process-wide MFStartup + per-thread COM init. Cheap; called on every open.

is_mp4 🔒

MF’s MPEG-4 source (mp4/mov/m4a: MIME */mp4, video/quicktime).

open_audio

open_reader 🔒

open_video

probe

push_audio 🔒

Append src (interleaved ch channels at rate) to fifo as interleaved stereo @ SAMPLE_RATE. Mono is duplicated, >2 channels keep L/R. Non-48k input is linearly resampled.

streams 🔒

Every stream in container order (ffmpeg’s 0:v:N / 0:a:N). Language/title are best effort via the presentation descriptor (descriptor index == reader stream index).