From 0c846af54055b9915c6c68617cd28176d5f84351 Mon Sep 17 00:00:00 2001 From: Carl Hetherington Date: Mon, 14 Mar 2022 20:29:44 +0100 Subject: Fix incorrectly-timed emission of silence padding causing buffer fill (#2217). On initialisation or after seek we insert silence corresponding to a positive delay in an audio stream. Previously this inserted silence was done at time 0, so that after a seek to time T the silent frames would come out of the audio merger at time 0 and then the player would fill the space up to time T with silence. If T was far enough along this would fill the audio buffers without there being any video. --- src/lib/audio_decoder.cc | 52 +++++++++++++++++++++++++----------------------- 1 file changed, 27 insertions(+), 25 deletions(-) (limited to 'src/lib/audio_decoder.cc') diff --git a/src/lib/audio_decoder.cc b/src/lib/audio_decoder.cc index 2d02043b5..664a56c2a 100644 --- a/src/lib/audio_decoder.cc +++ b/src/lib/audio_decoder.cc @@ -69,22 +69,20 @@ AudioDecoder::emit (shared_ptr film, AudioStreamPtr stream, shared_p time += ContentTime::from_seconds (_content->delay() / 1000.0); } - auto reset = false; - if (_positions[stream] == 0) { - /* This is the first data we have received since initialisation or seek. Set - the position based on the ContentTime that was given. After this first time - we just count samples unless the timestamp is more than slack_frames away - from where we think it should be. This is because ContentTimes seem to be - slightly unreliable from FFmpegDecoder (i.e. not sample accurate), but we still - need to obey them sometimes otherwise we get sync problems such as #1833. - */ - if (_content->delay() > 0) { - /* Insert silence to give the delay */ - silence (_content->delay ()); - } - reset = true; - } else if (std::abs(_positions[stream] - time.frames_round(resampled_rate)) > slack_frames) { - reset = true; + /* first_since_seek is set to true if this is the first data we have + received since initialisation or seek. We'll set the position based + on the ContentTime that was given. After this first time we just + count samples unless the timestamp is more than slack_frames away + from where we think it should be. This is because ContentTimes seem + to be slightly unreliable from FFmpegDecoder (i.e. not sample + accurate), but we still need to obey them sometimes otherwise we get + sync problems such as #1833. + */ + + auto const first_since_seek = _positions[stream] == 0; + auto const need_reset = !first_since_seek && (std::abs(_positions[stream] - time.frames_round(resampled_rate)) > slack_frames); + + if (need_reset) { LOG_GENERAL ( "Reset audio position: was %1, new data at %2, slack: %3 frames", _positions[stream], @@ -93,10 +91,14 @@ AudioDecoder::emit (shared_ptr film, AudioStreamPtr stream, shared_p ); } - if (reset) { + if (first_since_seek || need_reset) { _positions[stream] = time.frames_round (resampled_rate); } + if (first_since_seek && _content->delay() > 0) { + silence (stream, _content->delay()); + } + shared_ptr resampler; auto i = _resamplers.find(stream); if (i != _resamplers.end()) { @@ -183,18 +185,18 @@ AudioDecoder::flush () if (_content->delay() < 0) { /* Finish off with the gap caused by the delay */ - silence (-_content->delay ()); + for (auto stream: _content->streams()) { + silence (stream, -_content->delay()); + } } } void -AudioDecoder::silence (int milliseconds) +AudioDecoder::silence (AudioStreamPtr stream, int milliseconds) { - for (auto i: _content->streams()) { - int const samples = ContentTime::from_seconds(milliseconds / 1000.0).frames_round(i->frame_rate()); - auto silence = make_shared(i->channels(), samples); - silence->make_silent (); - Data (i, ContentAudio (silence, _positions[i])); - } + int const samples = ContentTime::from_seconds(milliseconds / 1000.0).frames_round(stream->frame_rate()); + auto silence = make_shared(stream->channels(), samples); + silence->make_silent (); + Data (stream, ContentAudio(silence, _positions[stream])); } -- cgit v1.2.3