X-Git-Url: https://main.carlh.net/gitweb/?a=blobdiff_plain;f=src%2Flib%2Fffmpeg_decoder.cc;h=7c846349f466d225cd190601446a917bd0e02eab;hb=54c0378137091de604b8c0d6d98959b55b0cddb6;hp=42be8227e95e57ebd0481e5b500566b1fd6cafb4;hpb=74fe68e5895654e27a7cf8097917c1e95fa89519;p=dcpomatic.git diff --git a/src/lib/ffmpeg_decoder.cc b/src/lib/ffmpeg_decoder.cc index 42be8227e..7c846349f 100644 --- a/src/lib/ffmpeg_decoder.cc +++ b/src/lib/ffmpeg_decoder.cc @@ -23,7 +23,6 @@ #include #include -#include #include #include #include @@ -43,22 +42,24 @@ extern "C" { #include "filter_graph.h" #include "audio_buffers.h" #include "ffmpeg_content.h" -#include "image_proxy.h" +#include "raw_image_proxy.h" #include "film.h" +#include "timer.h" #include "i18n.h" #define LOG_GENERAL(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_GENERAL); #define LOG_ERROR(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_ERROR); -#define LOG_WARNING(...) _video_content->film()->log()->log (__VA_ARGS__, Log::TYPE_WARNING); +#define LOG_WARNING_NC(...) _video_content->film()->log()->log (__VA_ARGS__, Log::TYPE_WARNING); +#define LOG_WARNING(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_WARNING); using std::cout; using std::string; using std::vector; -using std::stringstream; using std::list; using std::min; using std::pair; +using std::make_pair; using boost::shared_ptr; using boost::optional; using boost::dynamic_pointer_cast; @@ -96,6 +97,14 @@ FFmpegDecoder::FFmpegDecoder (shared_ptr c, shared_ptr _pts_offset = - c->audio_stream()->first_audio.get(); } + /* If _pts_offset is positive we would be pushing things from a -ve PTS to be played. + I don't think we ever want to do that, as it seems things at -ve PTS are not meant + to be seen (use for alignment bars etc.); see mantis #418. + */ + if (_pts_offset > ContentTime ()) { + _pts_offset = ContentTime (); + } + /* Now adjust both so that the video pts starts on a frame */ if (have_video && have_audio) { ContentTime first_video = c->first_video().get() + _pts_offset; @@ -127,21 +136,25 @@ FFmpegDecoder::pass () { int r = av_read_frame (_format_context, &_packet); - if (r < 0) { + /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame + has pretty-much succeeded (and hence generated data which should be processed). + Hence it makes sense to continue here in that case. + */ + if (r < 0 && r != AVERROR_INVALIDDATA) { if (r != AVERROR_EOF) { /* Maybe we should fail here, but for now we'll just finish off instead */ char buf[256]; av_strerror (r, buf, sizeof(buf)); LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), buf, r); } - + flush (); return true; } int const si = _packet.stream_index; - - if (si == _video_stream) { + + if (si == _video_stream && !_ignore_video) { decode_video_packet (); } else if (_ffmpeg_content->audio_stream() && _ffmpeg_content->audio_stream()->uses_index (_format_context, si)) { decode_audio_packet (); @@ -159,13 +172,14 @@ FFmpegDecoder::pass () shared_ptr FFmpegDecoder::deinterleave_audio (uint8_t** data, int size) { - assert (_ffmpeg_content->audio_channels()); - assert (bytes_per_audio_sample()); + DCPOMATIC_ASSERT (_ffmpeg_content->audio_channels()); + DCPOMATIC_ASSERT (bytes_per_audio_sample()); /* Deinterleave and convert to float */ - assert ((size % (bytes_per_audio_sample() * _ffmpeg_content->audio_channels())) == 0); - + /* total_samples and frames will be rounded down here, so if there are stray samples at the end + of the block that do not form a complete sample or frame they will be dropped. + */ int const total_samples = size / bytes_per_audio_sample(); int const frames = total_samples / _ffmpeg_content->audio_channels(); shared_ptr audio (new AudioBuffers (_ffmpeg_content->audio_channels(), frames)); @@ -282,75 +296,24 @@ FFmpegDecoder::bytes_per_audio_sample () const return av_get_bytes_per_sample (audio_sample_format ()); } -int -FFmpegDecoder::minimal_run (boost::function, optional, int)> finished) -{ - int frames_read = 0; - optional last_video; - optional last_audio; - - while (!finished (last_video, last_audio, frames_read)) { - int r = av_read_frame (_format_context, &_packet); - if (r < 0) { - /* We should flush our decoders here, possibly yielding a few more frames, - but the consequence of having to do that is too hideous to contemplate. - Instead we give up and say that you can't seek too close to the end - of a file. - */ - return frames_read; - } - - ++frames_read; - - double const time_base = av_q2d (_format_context->streams[_packet.stream_index]->time_base); - - if (_packet.stream_index == _video_stream) { - - av_frame_unref (_frame); - - int got_picture = 0; - r = avcodec_decode_video2 (video_codec_context(), _frame, &got_picture, &_packet); - if (r >= 0 && got_picture) { - last_video = ContentTime::from_seconds (av_frame_get_best_effort_timestamp (_frame) * time_base) + _pts_offset; - } - - } else if (_ffmpeg_content->audio_stream() && _ffmpeg_content->audio_stream()->uses_index (_format_context, _packet.stream_index)) { - AVPacket copy_packet = _packet; - while (copy_packet.size > 0) { - - int got_frame; - r = avcodec_decode_audio4 (audio_codec_context(), _frame, &got_frame, &_packet); - if (r >= 0 && got_frame) { - last_audio = ContentTime::from_seconds (av_frame_get_best_effort_timestamp (_frame) * time_base) + _pts_offset; - } - - copy_packet.data += r; - copy_packet.size -= r; - } - } - - av_free_packet (&_packet); - } - - return frames_read; -} - -bool -FFmpegDecoder::seek_overrun_finished (ContentTime seek, optional last_video, optional last_audio) const +void +FFmpegDecoder::seek (ContentTime time, bool accurate) { - return (last_video && last_video.get() >= seek) || (last_audio && last_audio.get() >= seek); -} + VideoDecoder::seek (time, accurate); + AudioDecoder::seek (time, accurate); + + /* If we are doing an `accurate' seek, we need to use pre-roll, as + we don't really know what the seek will give us. + */ -bool -FFmpegDecoder::seek_final_finished (int n, int done) const -{ - return n == done; -} + ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0); + time -= pre_roll; -void -FFmpegDecoder::seek_and_flush (ContentTime t) -{ - ContentTime const u = t - _pts_offset; + /* XXX: it seems debatable whether PTS should be used here... + http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html + */ + + ContentTime const u = time - _pts_offset; int64_t s = u.seconds() / av_q2d (_format_context->streams[_video_stream]->time_base); if (_ffmpeg_content->audio_stream ()) { @@ -359,12 +322,6 @@ FFmpegDecoder::seek_and_flush (ContentTime t) ); } - /* Ridiculous empirical hack */ - s--; - if (s < 0) { - s = 0; - } - av_seek_frame (_format_context, _video_stream, s, 0); avcodec_flush_buffers (video_codec_context()); @@ -376,40 +333,6 @@ FFmpegDecoder::seek_and_flush (ContentTime t) } } -void -FFmpegDecoder::seek (ContentTime time, bool accurate) -{ - VideoDecoder::seek (time, accurate); - AudioDecoder::seek (time, accurate); - - /* If we are doing an accurate seek, our initial shot will be 2s (2 being - a number plucked from the air) earlier than we want to end up. The loop below - will hopefully then step through to where we want to be. - */ - - ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0); - ContentTime initial_seek = time - pre_roll; - if (initial_seek < ContentTime (0)) { - initial_seek = ContentTime (0); - } - - /* Initial seek time in the video stream's timebase */ - - seek_and_flush (initial_seek); - - if (!accurate) { - /* That'll do */ - return; - } - - int const N = minimal_run (boost::bind (&FFmpegDecoder::seek_overrun_finished, this, time, _1, _2)); - - seek_and_flush (initial_seek); - if (N > 0) { - minimal_run (boost::bind (&FFmpegDecoder::seek_final_finished, this, N - 1, _3)); - } -} - void FFmpegDecoder::decode_audio_packet () { @@ -422,11 +345,21 @@ FFmpegDecoder::decode_audio_packet () while (copy_packet.size > 0) { int frame_finished; - int const decode_result = avcodec_decode_audio4 (audio_codec_context(), _frame, &frame_finished, ©_packet); - + int decode_result = avcodec_decode_audio4 (audio_codec_context(), _frame, &frame_finished, ©_packet); if (decode_result < 0) { - LOG_ERROR ("avcodec_decode_audio4 failed (%1)", decode_result); - return; + /* avcodec_decode_audio4 can sometimes return an error even though it has decoded + some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA + if it overreads the auxiliary data. ffplay carries on if frame_finished is true, + even in the face of such an error, so I think we should too. + + Returning from the method here caused mantis #352. + */ + LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result); + + /* Fudge decode_result so that we come out of the while loop when + we've processed this data. + */ + decode_result = copy_packet.size; } if (frame_finished) { @@ -481,11 +414,11 @@ FFmpegDecoder::decode_video_packet () if (i->second != AV_NOPTS_VALUE) { double const pts = i->second * av_q2d (_format_context->streams[_video_stream]->time_base) + _pts_offset.seconds (); video ( - shared_ptr (new RawImageProxy (image, _video_content->film()->log())), + shared_ptr (new RawImageProxy (image)), rint (pts * _ffmpeg_content->video_frame_rate ()) ); } else { - LOG_WARNING ("Dropping frame without PTS"); + LOG_WARNING_NC ("Dropping frame without PTS"); } } @@ -565,8 +498,8 @@ FFmpegDecoder::decode_subtitle_packet () avsubtitle_free (&sub); } -bool -FFmpegDecoder::has_subtitle_during (ContentTimePeriod p) const +list +FFmpegDecoder::subtitles_during (ContentTimePeriod p, bool starting) const { - return _ffmpeg_content->has_subtitle_during (p); + return _ffmpeg_content->subtitles_during (p, starting); }