Merge pull request #4007 from FredApps/fix/778-matroska-cold-seek

fix(media): preserve Matroska audio position on cold seek
This commit is contained in:
FredApps
2026-11-21 09:25:39 +07:00
committed by GitHub
2 changed files with 90 additions and 6 deletions
+15 -6
View File
@@ -173,22 +173,31 @@ impl ContainerFile {
}
ContainerFile::Matroska(f) => {
let track = &f.tracks[track_idx];
// Indexed PTS already subtracts CodecDelay rounded to the container clock.
// Translate between that clock and the warm path's whole-frame delay.
let tick_ns = i128::from(f.info.timestamp_scale.max(1));
let delay_ns = i128::from(track.codec_delay_ns);
let indexed_delay_ns = ((delay_ns + tick_ns / 2) / tick_ns) * tick_ns;
let delay_frames = audio_frames_from_ns(delay_ns, native_rate);
let preroll_ns = i128::from(track.seek_pre_roll_ns.max(80_000_000));
let target_ns = (i128::from(start.max(0)) * 1_000_000_000) / i128::from(native_rate) + delay_ns;
let seek_ns = (target_ns - preroll_ns).max(0);
let t_track = track.ns_to_ticks(i64::try_from(seek_ns).unwrap_or(0));
let target_ns = ((i128::from(start.max(0)) + delay_frames) * 1_000_000_000) / i128::from(native_rate);
let seek_ns = (target_ns - preroll_ns - indexed_delay_ns).max(0);
let t_track = track.ns_to_ticks(i64::try_from(seek_ns).unwrap_or(i64::MAX));
let sample_idx = track.sample_at_pts(t_track).unwrap_or(0);
let seek_idx = sample_idx.saturating_sub(2);
let sample_pts = track.samples.get(seek_idx).map_or(0, |s| s.pts);
let sample_ns = i128::from(track.ticks_to_ns(sample_pts)) - delay_ns;
let decoded_pos = i64::try_from((sample_ns * i128::from(native_rate)) / 1_000_000_000).unwrap_or(0);
let sample_ns = i128::from(track.ticks_to_ns(sample_pts)) + indexed_delay_ns;
let decoded_pos = i64::try_from(audio_frames_from_ns(sample_ns, native_rate) - delay_frames).unwrap_or(0);
(seek_idx, decoded_pos)
}
}
}
}
fn audio_frames_from_ns(ns: i128, rate: u32) -> i128 {
(ns * i128::from(rate)) / 1_000_000_000
}
fn open_audio_tracks<I>(container: &ContainerFile, source: &dyn ByteSource, candidates: I) -> Option<(Vec<AudioTrackEntry>, Vec<AudioStreamInfo>)>
where
I: IntoIterator<Item = (usize, usize, CodecParams, u64)>,
@@ -209,7 +218,7 @@ where
let dec = AudioTrackDecoder::new(&params, &refs).ok()?;
let native_rate = dec.sample_rate().max(1);
let channels = u32::from(dec.layout().channels()).max(1);
let delay_samples = i64::try_from((i128::from(codec_delay_ns) * i128::from(native_rate)) / 1_000_000_000).unwrap_or(0);
let delay_samples = i64::try_from(audio_frames_from_ns(i128::from(codec_delay_ns), native_rate)).unwrap_or(0);
audio_infos.push(AudioStreamInfo {
sample_rate: native_rate,
channels,
+75
View File
@@ -1592,6 +1592,81 @@ fn filmcraft_fallback_handles_mjpeg_mov() {
assert_eq!((img.width, img.height), (160, 120));
}
/// A cold seek and uninterrupted playback must cross the same audible boundary. The PCM is
/// contiguous, so the test isolates reset positioning from separately handled timestamp gaps.
#[test]
fn matroska_cold_seek_applies_codec_delay_once() {
assert_matroska_cold_seek_codec_delay(8_250_000); // Exactly 396 samples, but 8 indexed ticks.
}
#[test]
fn matroska_cold_seek_matches_fractional_codec_delay_rounding() {
assert_matroska_cold_seek_codec_delay(8_260_000); // 396.48 samples, also 8 indexed ticks.
}
fn assert_matroska_cold_seek_codec_delay(delay_ns: i64) {
use craft_media::matroska::{self, MkvWriter, MuxOptions, TrackKind, TrackSpec};
use craft_video::prores::{ChromaFormat, Encoder, Frame, Profile};
const RATE: u32 = 48_000;
const PACKET_FRAMES: usize = 4_800;
const PACKET_NS: i64 = 100_000_000;
const CHANGE_PACKET: i64 = 32;
const HALF_WINDOW: usize = 192;
let profile = Profile::P4444;
let black = Frame::new(16, 16, ChromaFormat::Yuv444, 10, false).unwrap();
let picture = Encoder::new(profile, 16, 16).encode(&black).unwrap();
let mut video = TrackSpec::new(TrackKind::Video, "V_PRORES");
video.video_size = Some((16, 16));
video.codec_private = profile.fourcc().to_vec();
video.default_duration_ns = Some(4_000_000_000);
let mut audio = TrackSpec::new(TrackKind::Audio, "A_PCM/INT/LIT");
audio.audio = Some((f64::from(RATE), 1, Some(16)));
audio.codec_delay_ns = delay_ns as u64;
audio.default_duration_ns = Some(PACKET_NS as u64);
let mut mux = MkvWriter::new(std::io::Cursor::new(Vec::new()), vec![video, audio], MuxOptions::default()).unwrap();
mux.write_frame(0, 0, true, &picture, None).unwrap();
let silent = vec![0; PACKET_FRAMES * 2];
let signal = 16_384i16.to_le_bytes().repeat(PACKET_FRAMES);
for n in 0..40 {
let pcm = if n < CHANGE_PACKET { &silent } else { &signal };
mux.write_frame(1, n * PACKET_NS - delay_ns, true, pcm, None).unwrap();
}
let bytes = mux.finish().unwrap().into_inner();
let mkv = matroska::open(bytes.as_slice()).unwrap();
assert_eq!(mkv.tracks[1].samples[0].pts, -8);
// Distinct files and pools keep the two decoder states independent.
let dir = tempfile::tempdir().unwrap();
let warm_path = dir.path().join("sequential.mkv");
let cold_path = dir.path().join("cold.mkv");
std::fs::write(&warm_path, &bytes).unwrap();
std::fs::write(&cold_path, &bytes).unwrap();
let warm = probe(&warm_path).unwrap();
let cold = probe(&cold_path).unwrap();
for f in [&warm, &cold] {
assert!(f.has_audio);
assert_eq!(f.codec, "ProRes", "exercise the shared Matroska route, not legacy fallback");
}
let warm_pool = MediaPool::new();
for n in 0..CHANGE_PACKET - 1 {
let t = Tick::from_units(n * PACKET_FRAMES as i64, i64::from(RATE));
let _ = warm_pool.audio_samples(&warm, t, PACKET_FRAMES, RATE);
}
let boundary = CHANGE_PACKET * PACKET_FRAMES as i64 - delay_ns * i64::from(RATE) / 1_000_000_000;
let t = Tick::from_units(boundary - HALF_WINDOW as i64, i64::from(RATE));
let sequential = warm_pool.audio_samples(&warm, t, 2 * HALF_WINDOW, RATE);
assert_eq!(sequential.len(), 4 * HALF_WINDOW);
assert!(sequential[..2 * HALF_WINDOW].iter().all(|s| s.abs() < 1e-6));
assert!(sequential[2 * HALF_WINDOW..].iter().all(|s| s.abs() > 0.1));
let seek = MediaPool::new().audio_samples(&cold, t, 2 * HALF_WINDOW, RATE);
assert_eq!(seek, sequential, "cold seek must not shift the signal boundary");
}
#[test]
fn matroska_opus_audio_timing_and_seek() {
let Some(p) = try_ffmpeg_fixture(