Merge master.
[dcpomatic.git] / src / lib / ffmpeg_decoder.cc
index f7bb979441870c707454df55e23ed15c94dfc926..d251a37446ff96e2e49d514575b7719e53f44b6c 100644 (file)
@@ -1,7 +1,5 @@
-/* -*- c-basic-offset: 8; default-tab-width: 8; -*- */
-
 /*
-    Copyright (C) 2012 Carl Hetherington <cth@carlh.net>
+    Copyright (C) 2012-2014 Carl Hetherington <cth@carlh.net>
 
     This program is free software; you can redistribute it and/or modify
     it under the terms of the GNU General Public License as published by
 #include <iomanip>
 #include <iostream>
 #include <stdint.h>
-#include <boost/lexical_cast.hpp>
+#include <sndfile.h>
 extern "C" {
-#include <tiffio.h>
 #include <libavcodec/avcodec.h>
 #include <libavformat/avformat.h>
-#include <libswscale/swscale.h>
-#include <libpostproc/postprocess.h>
 }
-#include <sndfile.h>
-#include "film.h"
-#include "format.h"
-#include "transcoder.h"
-#include "job.h"
 #include "filter.h"
 #include "exceptions.h"
 #include "image.h"
@@ -49,8 +39,9 @@ extern "C" {
 #include "log.h"
 #include "ffmpeg_decoder.h"
 #include "filter_graph.h"
-#include "subtitle.h"
 #include "audio_buffers.h"
+#include "ffmpeg_content.h"
+#include "image_proxy.h"
 
 #include "i18n.h"
 
@@ -60,169 +51,83 @@ using std::vector;
 using std::stringstream;
 using std::list;
 using std::min;
+using std::pair;
 using boost::shared_ptr;
 using boost::optional;
 using boost::dynamic_pointer_cast;
-using libdcp::Size;
-
-boost::mutex FFmpegDecoder::_mutex;
-
-FFmpegDecoder::FFmpegDecoder (shared_ptr<const Film> f, shared_ptr<const FFmpegContent> c, bool video, bool audio, bool subtitles)
-       : Decoder (f)
-       , VideoDecoder (f, c)
-       , AudioDecoder (f, c)
-       , _ffmpeg_content (c)
-       , _format_context (0)
-       , _video_stream (-1)
-       , _frame (0)
-       , _video_codec_context (0)
-       , _video_codec (0)
-       , _audio_codec_context (0)
-       , _audio_codec (0)
+using dcp::Size;
+
+FFmpegDecoder::FFmpegDecoder (shared_ptr<const FFmpegContent> c, shared_ptr<Log> log)
+       : VideoDecoder (c)
+       , AudioDecoder (c)
+       , FFmpeg (c)
+       , _log (log)
        , _subtitle_codec_context (0)
        , _subtitle_codec (0)
-       , _decode_video (video)
-       , _decode_audio (audio)
-       , _decode_subtitles (subtitles)
 {
-       setup_general ();
-       setup_video ();
-       setup_audio ();
        setup_subtitle ();
-}
-
-FFmpegDecoder::~FFmpegDecoder ()
-{
-       boost::mutex::scoped_lock lm (_mutex);
-       
-       if (_audio_codec_context) {
-               avcodec_close (_audio_codec_context);
-       }
-
-       if (_video_codec_context) {
-               avcodec_close (_video_codec_context);
-       }
-
-       if (_subtitle_codec_context) {
-               avcodec_close (_subtitle_codec_context);
-       }
-
-       av_free (_frame);
-       
-       avformat_close_input (&_format_context);
-}      
-
-void
-FFmpegDecoder::setup_general ()
-{
-       av_register_all ();
-
-       if (avformat_open_input (&_format_context, _ffmpeg_content->file().string().c_str(), 0, 0) < 0) {
-               throw OpenFileError (_ffmpeg_content->file().string ());
-       }
 
-       if (avformat_find_stream_info (_format_context, 0) < 0) {
-               throw DecodeError (_("could not find stream information"));
-       }
+       /* Audio and video frame PTS values may not start with 0.  We want
+          to fiddle them so that:
 
-       /* Find video, audio and subtitle streams */
+          1.  One of them starts at time 0.
+          2.  The first video PTS value ends up on a frame boundary.
 
-       for (uint32_t i = 0; i < _format_context->nb_streams; ++i) {
-               AVStream* s = _format_context->streams[i];
-               if (s->codec->codec_type == AVMEDIA_TYPE_VIDEO) {
-                       _video_stream = i;
-               } else if (s->codec->codec_type == AVMEDIA_TYPE_AUDIO) {
+          Then we remove big initial gaps in PTS and we allow our
+          insertion of black frames to work.
 
-                       /* This is a hack; sometimes it seems that _audio_codec_context->channel_layout isn't set up,
-                          so bodge it here.  No idea why we should have to do this.
-                       */
+          We will do pts_to_use = pts_from_ffmpeg + pts_offset;
+       */
 
-                       if (s->codec->channel_layout == 0) {
-                               s->codec->channel_layout = av_get_default_channel_layout (s->codec->channels);
-                       }
-                       
-                       _audio_streams.push_back (
-                               shared_ptr<FFmpegAudioStream> (
-                                       new FFmpegAudioStream (stream_name (s), i, s->codec->sample_rate, s->codec->channels)
-                                       )
-                               );
+       bool const have_video = c->first_video();
+       bool const have_audio = c->audio_stream () && c->audio_stream()->first_audio;
 
-               } else if (s->codec->codec_type == AVMEDIA_TYPE_SUBTITLE) {
-                       _subtitle_streams.push_back (shared_ptr<FFmpegSubtitleStream> (new FFmpegSubtitleStream (stream_name (s), i)));
-               }
-       }
+       /* First, make one of them start at 0 */
 
-       if (_video_stream < 0) {
-               throw DecodeError (N_("could not find video stream"));
+       if (have_audio && have_video) {
+               _pts_offset = - min (c->first_video().get(), c->audio_stream()->first_audio.get());
+       } else if (have_video) {
+               _pts_offset = - c->first_video().get();
+       } else if (have_audio) {
+               _pts_offset = - c->audio_stream()->first_audio.get();
        }
 
-       _frame = avcodec_alloc_frame ();
-       if (_frame == 0) {
-               throw DecodeError (N_("could not allocate frame"));
+       /* 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;
+               ContentTime const old_first_video = first_video;
+               _pts_offset += first_video.round_up (c->video_frame_rate ()) - old_first_video;
        }
 }
 
-void
-FFmpegDecoder::setup_video ()
+FFmpegDecoder::~FFmpegDecoder ()
 {
        boost::mutex::scoped_lock lm (_mutex);
-       
-       _video_codec_context = _format_context->streams[_video_stream]->codec;
-       _video_codec = avcodec_find_decoder (_video_codec_context->codec_id);
 
-       if (_video_codec == 0) {
-               throw DecodeError (_("could not find video decoder"));
-       }
-
-       if (avcodec_open2 (_video_codec_context, _video_codec, 0) < 0) {
-               throw DecodeError (N_("could not open video decoder"));
+       if (_subtitle_codec_context) {
+               avcodec_close (_subtitle_codec_context);
        }
 }
 
 void
-FFmpegDecoder::setup_audio ()
+FFmpegDecoder::flush ()
 {
-       boost::mutex::scoped_lock lm (_mutex);
+       /* Get any remaining frames */
        
-       if (!_ffmpeg_content->audio_stream ()) {
-               return;
-       }
-
-       _audio_codec_context = _format_context->streams[_ffmpeg_content->audio_stream()->id]->codec;
-       _audio_codec = avcodec_find_decoder (_audio_codec_context->codec_id);
-
-       if (_audio_codec == 0) {
-               throw DecodeError (_("could not find audio decoder"));
-       }
-
-       if (avcodec_open2 (_audio_codec_context, _audio_codec, 0) < 0) {
-               throw DecodeError (N_("could not open audio decoder"));
-       }
-}
-
-void
-FFmpegDecoder::setup_subtitle ()
-{
-       boost::mutex::scoped_lock lm (_mutex);
+       _packet.data = 0;
+       _packet.size = 0;
        
-       if (!_ffmpeg_content->subtitle_stream() || _ffmpeg_content->subtitle_stream()->id >= int (_format_context->nb_streams)) {
-               return;
-       }
-
-       _subtitle_codec_context = _format_context->streams[_ffmpeg_content->subtitle_stream()->id]->codec;
-       _subtitle_codec = avcodec_find_decoder (_subtitle_codec_context->codec_id);
-
-       if (_subtitle_codec == 0) {
-               throw DecodeError (_("could not find subtitle decoder"));
-       }
+       /* XXX: should we reset _packet.data and size after each *_decode_* call? */
        
-       if (avcodec_open2 (_subtitle_codec_context, _subtitle_codec, 0) < 0) {
-               throw DecodeError (N_("could not open subtitle decoder"));
+       while (decode_video_packet ()) {}
+       
+       if (_ffmpeg_content->audio_stream()) {
+               decode_audio_packet ();
+               AudioDecoder::flush ();
        }
 }
 
-
-void
+bool
 FFmpegDecoder::pass ()
 {
        int r = av_read_frame (_format_context, &_packet);
@@ -232,58 +137,25 @@ FFmpegDecoder::pass ()
                        /* Maybe we should fail here, but for now we'll just finish off instead */
                        char buf[256];
                        av_strerror (r, buf, sizeof(buf));
-                       shared_ptr<const Film> film = _film.lock ();
-                       assert (film);
-                       film->log()->log (String::compose (N_("error on av_read_frame (%1) (%2)"), buf, r));
-               }
-
-               /* Get any remaining frames */
-               
-               _packet.data = 0;
-               _packet.size = 0;
-               
-               /* XXX: should we reset _packet.data and size after each *_decode_* call? */
-               
-               if (_decode_video) {
-                       while (decode_video_packet ());
+                       _log->log (String::compose (N_("error on av_read_frame (%1) (%2)"), buf, r));
                }
 
-               if (_ffmpeg_content->audio_stream() && _decode_audio) {
-                       decode_audio_packet ();
-               }
-                       
-               return;
+               flush ();
+               return true;
        }
 
-       avcodec_get_frame_defaults (_frame);
-
-       if (_packet.stream_index == _video_stream && _decode_video) {
+       int const si = _packet.stream_index;
+       
+       if (si == _video_stream) {
                decode_video_packet ();
-       } else if (_ffmpeg_content->audio_stream() && _packet.stream_index == _ffmpeg_content->audio_stream()->id && _decode_audio) {
+       } else if (_ffmpeg_content->audio_stream() && _ffmpeg_content->audio_stream()->uses_index (_format_context, si)) {
                decode_audio_packet ();
-       } else if (_ffmpeg_content->subtitle_stream() && _packet.stream_index == _ffmpeg_content->subtitle_stream()->id && _decode_subtitles) {
-
-               int got_subtitle;
-               AVSubtitle sub;
-               if (avcodec_decode_subtitle2 (_subtitle_codec_context, &sub, &got_subtitle, &_packet) && got_subtitle) {
-                       /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
-                          indicate that the previous subtitle should stop.
-                       */
-                       if (sub.num_rects > 0) {
-                               shared_ptr<TimedSubtitle> ts;
-                               try {
-                                       subtitle (shared_ptr<TimedSubtitle> (new TimedSubtitle (sub)));
-                               } catch (...) {
-                                       /* some problem with the subtitle; we probably didn't understand it */
-                               }
-                       } else {
-                               subtitle (shared_ptr<TimedSubtitle> ());
-                       }
-                       avsubtitle_free (&sub);
-               }
+       } else if (_ffmpeg_content->subtitle_stream() && _ffmpeg_content->subtitle_stream()->uses_index (_format_context, si)) {
+               decode_subtitle_packet ();
        }
 
        av_free_packet (&_packet);
+       return false;
 }
 
 /** @param data pointer to array of pointers to buffers.
@@ -304,6 +176,23 @@ FFmpegDecoder::deinterleave_audio (uint8_t** data, int size)
        shared_ptr<AudioBuffers> audio (new AudioBuffers (_ffmpeg_content->audio_channels(), frames));
 
        switch (audio_sample_format()) {
+       case AV_SAMPLE_FMT_U8:
+       {
+               uint8_t* p = reinterpret_cast<uint8_t *> (data[0]);
+               int sample = 0;
+               int channel = 0;
+               for (int i = 0; i < total_samples; ++i) {
+                       audio->data(channel)[sample] = float(*p++) / (1 << 23);
+
+                       ++channel;
+                       if (channel == _ffmpeg_content->audio_channels()) {
+                               channel = 0;
+                               ++sample;
+                       }
+               }
+       }
+       break;
+       
        case AV_SAMPLE_FMT_S16:
        {
                int16_t* p = reinterpret_cast<int16_t *> (data[0]);
@@ -382,152 +271,148 @@ FFmpegDecoder::deinterleave_audio (uint8_t** data, int size)
        return audio;
 }
 
-float
-FFmpegDecoder::video_frame_rate () const
-{
-       AVStream* s = _format_context->streams[_video_stream];
-
-       if (s->avg_frame_rate.num && s->avg_frame_rate.den) {
-               return av_q2d (s->avg_frame_rate);
-       }
-
-       return av_q2d (s->r_frame_rate);
-}
-
 AVSampleFormat
 FFmpegDecoder::audio_sample_format () const
 {
-       if (_audio_codec_context == 0) {
+       if (!_ffmpeg_content->audio_stream()) {
                return (AVSampleFormat) 0;
        }
        
-       return _audio_codec_context->sample_fmt;
+       return audio_codec_context()->sample_fmt;
 }
 
-libdcp::Size
-FFmpegDecoder::video_size () const
+int
+FFmpegDecoder::bytes_per_audio_sample () const
 {
-       return libdcp::Size (_video_codec_context->width, _video_codec_context->height);
+       return av_get_bytes_per_sample (audio_sample_format ());
 }
 
-string
-FFmpegDecoder::stream_name (AVStream* s) const
+int
+FFmpegDecoder::minimal_run (boost::function<bool (optional<ContentTime>, optional<ContentTime>, int)> finished)
 {
-       stringstream n;
-
-       if (s->metadata) {
-               AVDictionaryEntry const * lang = av_dict_get (s->metadata, N_("language"), 0, 0);
-               if (lang) {
-                       n << lang->value;
+       int frames_read = 0;
+       optional<ContentTime> last_video;
+       optional<ContentTime> 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;
                }
-               
-               AVDictionaryEntry const * title = av_dict_get (s->metadata, N_("title"), 0, 0);
-               if (title) {
-                       if (!n.str().empty()) {
-                               n << N_(" ");
+
+               ++frames_read;
+
+               double const time_base = av_q2d (_format_context->streams[_packet.stream_index]->time_base);
+
+               if (_packet.stream_index == _video_stream) {
+
+                       avcodec_get_frame_defaults (_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;
                        }
-                       n << title->value;
-               }
-       }
 
-       if (n.str().empty()) {
-               n << N_("unknown");
+               } 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 n.str ();
+       return frames_read;
 }
 
-int
-FFmpegDecoder::bytes_per_audio_sample () const
+bool
+FFmpegDecoder::seek_overrun_finished (ContentTime seek, optional<ContentTime> last_video, optional<ContentTime> last_audio) const
 {
-       return av_get_bytes_per_sample (audio_sample_format ());
+       return (last_video && last_video.get() >= seek) || (last_audio && last_audio.get() >= seek);
 }
 
-void
-FFmpegDecoder::seek (Time t)
+bool
+FFmpegDecoder::seek_final_finished (int n, int done) const
 {
-       do_seek (t, false, false);
+       return n == done;
 }
 
 void
-FFmpegDecoder::seek_back ()
+FFmpegDecoder::seek_and_flush (ContentTime t)
 {
-       if (next() < 2.5) {
-               return;
+       ContentTime const u = t - _pts_offset;
+       int64_t s = u.seconds() / av_q2d (_format_context->streams[_video_stream]->time_base);
+
+       if (_ffmpeg_content->audio_stream ()) {
+               s = min (
+                       s, int64_t (u.seconds() / av_q2d (_ffmpeg_content->audio_stream()->stream(_format_context)->time_base))
+                       );
        }
-       
-       do_seek (next() - 2.5 * TIME_HZ / video_frame_rate(), true, true);
-}
 
-void
-FFmpegDecoder::seek_forward ()
-{
-       if (next() >= (video_length() - video_frame_rate())) {
-               return;
+       /* Ridiculous empirical hack */
+       s--;
+       if (s < 0) {
+               s = 0;
        }
-       
-       do_seek (next() - 0.5 * TIME_HZ / video_frame_rate(), true, true);
-}
 
-void
-FFmpegDecoder::do_seek (Time t, bool backwards, bool accurate)
-{
-       int64_t const vt = t / (av_q2d (_format_context->streams[_video_stream]->time_base) * TIME_HZ);
-       av_seek_frame (_format_context, _video_stream, vt, backwards ? AVSEEK_FLAG_BACKWARD : 0);
+       av_seek_frame (_format_context, _video_stream, s, 0);
 
-       avcodec_flush_buffers (_video_codec_context);
+       avcodec_flush_buffers (video_codec_context());
+       if (audio_codec_context ()) {
+               avcodec_flush_buffers (audio_codec_context ());
+       }
        if (_subtitle_codec_context) {
                avcodec_flush_buffers (_subtitle_codec_context);
        }
-
-       if (accurate) {
-               while (1) {
-                       int r = av_read_frame (_format_context, &_packet);
-                       if (r < 0) {
-                               return;
-                       }
-                       
-                       avcodec_get_frame_defaults (_frame);
-                       
-                       if (_packet.stream_index == _video_stream) {
-                               int finished = 0;
-                               int const r = avcodec_decode_video2 (_video_codec_context, _frame, &finished, &_packet);
-                               if (r >= 0 && finished) {
-                                       int64_t const bet = av_frame_get_best_effort_timestamp (_frame);
-                                       if (bet > vt) {
-                                               break;
-                                       }
-                               }
-                       }
-                       
-                       av_free_packet (&_packet);
-               }
-       }
-
-       return;
 }
 
 void
-FFmpegDecoder::film_changed (Film::Property p)
+FFmpegDecoder::seek (ContentTime time, bool accurate)
 {
-       switch (p) {
-       case Film::FILTERS:
-       {
-               boost::mutex::scoped_lock lm (_filter_graphs_mutex);
-               _filter_graphs.clear ();
+       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);
        }
-       break;
 
-       default:
-               break;
+       /* Initial seek time in the video stream's timebase */
+
+       seek_and_flush (initial_seek);
+
+       if (!accurate) {
+               /* That'll do */
+               return;
        }
-}
 
-/** @return Length (in video frames) according to our content's header */
-ContentVideoFrame
-FFmpegDecoder::video_length () const
-{
-       return (double(_format_context->duration) / AV_TIME_BASE) * video_frame_rate();
+       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
@@ -538,29 +423,32 @@ FFmpegDecoder::decode_audio_packet ()
        */
        
        AVPacket copy_packet = _packet;
-
+       
        while (copy_packet.size > 0) {
 
                int frame_finished;
-               int const decode_result = avcodec_decode_audio4 (_audio_codec_context, _frame, &frame_finished, &copy_packet);
-               if (decode_result >= 0) {
-                       if (frame_finished) {
-                       
-                               /* Where we are in the source, in seconds */
-                               double const source_pts_seconds = av_q2d (_format_context->streams[copy_packet.stream_index]->time_base)
-                                       * av_frame_get_best_effort_timestamp(_frame);
-                               
-                               int const data_size = av_samples_get_buffer_size (
-                                       0, _audio_codec_context->channels, _frame->nb_samples, audio_sample_format (), 1
-                                       );
-                               
-                               assert (_audio_codec_context->channels == _ffmpeg_content->audio_channels());
-                               Audio (deinterleave_audio (_frame->data, data_size), source_pts_seconds);
-                       }
+               int const decode_result = avcodec_decode_audio4 (audio_codec_context(), _frame, &frame_finished, &copy_packet);
+
+               if (decode_result < 0) {
+                       _log->log (String::compose ("avcodec_decode_audio4 failed (%1)", decode_result));
+                       return;
+               }
+
+               if (frame_finished) {
+                       ContentTime const ct = ContentTime::from_seconds (
+                               av_frame_get_best_effort_timestamp (_frame) *
+                               av_q2d (_ffmpeg_content->audio_stream()->stream (_format_context)->time_base))
+                               + _pts_offset;
                        
-                       copy_packet.data += decode_result;
-                       copy_packet.size -= decode_result;
+                       int const data_size = av_samples_get_buffer_size (
+                               0, audio_codec_context()->channels, _frame->nb_samples, audio_sample_format (), 1
+                               );
+
+                       audio (deinterleave_audio (_frame->data, data_size), ct);
                }
+                       
+               copy_packet.data += decode_result;
+               copy_packet.size -= decode_result;
        }
 }
 
@@ -568,53 +456,144 @@ bool
 FFmpegDecoder::decode_video_packet ()
 {
        int frame_finished;
-       if (avcodec_decode_video2 (_video_codec_context, _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
+       if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
                return false;
        }
-               
+
        boost::mutex::scoped_lock lm (_filter_graphs_mutex);
 
        shared_ptr<FilterGraph> graph;
        
        list<shared_ptr<FilterGraph> >::iterator i = _filter_graphs.begin();
-       while (i != _filter_graphs.end() && !(*i)->can_process (libdcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
+       while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
                ++i;
        }
 
        if (i == _filter_graphs.end ()) {
-               graph.reset (new FilterGraph (_film, this, libdcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
+               graph.reset (new FilterGraph (_ffmpeg_content, dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
                _filter_graphs.push_back (graph);
-
-               shared_ptr<const Film> film = _film.lock ();
-               assert (film);
-               film->log()->log (String::compose (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format));
+               _log->log (String::compose (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format));
        } else {
                graph = *i;
        }
 
+       list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
 
-       list<shared_ptr<Image> > images = graph->process (_frame);
-       
-       for (list<shared_ptr<Image> >::iterator i = images.begin(); i != images.end(); ++i) {
-               int64_t const bet = av_frame_get_best_effort_timestamp (_frame);
-               if (bet != AV_NOPTS_VALUE) {
-                       /* XXX: may need to insert extra frames / remove frames here ...
-                          (as per old Matcher)
-                       */
-                       Time const t = bet * av_q2d (_format_context->streams[_video_stream]->time_base) * TIME_HZ;
-                       video (*i, false, t);
+       for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
+
+               shared_ptr<Image> image = i->first;
+               
+               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<ImageProxy> (new RawImageProxy (image)), rint (pts * _ffmpeg_content->video_frame_rate ()));
                } else {
-                       shared_ptr<const Film> film = _film.lock ();
-                       assert (film);
-                       film->log()->log ("Dropping frame without PTS");
+                       _log->log ("Dropping frame without PTS");
                }
        }
 
        return true;
 }
 
-Time
-FFmpegDecoder::next () const
+       
+void
+FFmpegDecoder::setup_subtitle ()
 {
-       return min (_next_video, _next_audio);
+       boost::mutex::scoped_lock lm (_mutex);
+       
+       if (!_ffmpeg_content->subtitle_stream()) {
+               return;
+       }
+
+       _subtitle_codec_context = _ffmpeg_content->subtitle_stream()->stream(_format_context)->codec;
+       if (_subtitle_codec_context == 0) {
+               throw DecodeError (N_("could not find subtitle stream"));
+       }
+
+       _subtitle_codec = avcodec_find_decoder (_subtitle_codec_context->codec_id);
+
+       if (_subtitle_codec == 0) {
+               throw DecodeError (N_("could not find subtitle decoder"));
+       }
+       
+       if (avcodec_open2 (_subtitle_codec_context, _subtitle_codec, 0) < 0) {
+               throw DecodeError (N_("could not open subtitle decoder"));
+       }
+}
+
+void
+FFmpegDecoder::decode_subtitle_packet ()
+{
+       int got_subtitle;
+       AVSubtitle sub;
+       if (avcodec_decode_subtitle2 (_subtitle_codec_context, &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
+               return;
+       }
+
+       /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
+          indicate that the previous subtitle should stop.
+       */
+       if (sub.num_rects <= 0) {
+               image_subtitle (ContentTime (), ContentTime (), shared_ptr<Image> (), dcpomatic::Rect<double> ());
+               return;
+       } else if (sub.num_rects > 1) {
+               throw DecodeError (_("multi-part subtitles not yet supported"));
+       }
+               
+       /* Subtitle PTS (within the source, not taking into account any of the
+          source that we may have chopped off for the DCP)
+       */
+       ContentTime packet_time = ContentTime::from_seconds (static_cast<double> (sub.pts) / AV_TIME_BASE) + _pts_offset;
+       
+       /* hence start time for this sub */
+       ContentTime const from = packet_time + ContentTime::from_seconds (sub.start_display_time / 1e3);
+       ContentTime const to = packet_time + ContentTime::from_seconds (sub.end_display_time / 1e3);
+
+       AVSubtitleRect const * rect = sub.rects[0];
+
+       if (rect->type != SUBTITLE_BITMAP) {
+               /* XXX */
+               // throw DecodeError (_("non-bitmap subtitles not yet supported"));
+               return;
+       }
+
+       /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
+          G, third B, fourth A.
+       */
+       shared_ptr<Image> image (new Image (PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
+
+       /* Start of the first line in the subtitle */
+       uint8_t* sub_p = rect->pict.data[0];
+       /* sub_p looks up into a BGRA palette which is here
+          (i.e. first byte B, second G, third R, fourth A)
+       */
+       uint32_t const * palette = (uint32_t *) rect->pict.data[1];
+       /* Start of the output data */
+       uint32_t* out_p = (uint32_t *) image->data()[0];
+
+       for (int y = 0; y < rect->h; ++y) {
+               uint8_t* sub_line_p = sub_p;
+               uint32_t* out_line_p = out_p;
+               for (int x = 0; x < rect->w; ++x) {
+                       uint32_t const p = palette[*sub_line_p++];
+                       *out_line_p++ = ((p & 0xff) << 16) | (p & 0xff00) | ((p & 0xff0000) >> 16) | (p & 0xff000000);
+               }
+               sub_p += rect->pict.linesize[0];
+               out_p += image->stride()[0] / sizeof (uint32_t);
+       }
+
+       dcp::Size const vs = _ffmpeg_content->video_size ();
+
+       image_subtitle (
+               from,
+               to,
+               image,
+               dcpomatic::Rect<double> (
+                       static_cast<double> (rect->x) / vs.width,
+                       static_cast<double> (rect->y) / vs.height,
+                       static_cast<double> (rect->w) / vs.width,
+                       static_cast<double> (rect->h) / vs.height
+                       )
+               );
+       
+       avsubtitle_free (&sub);
 }