2 Copyright (C) 2012-2014 Carl Hetherington <cth@carlh.net>
4 This program is free software; you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 2 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program; if not, write to the Free Software
16 Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20 /** @file src/ffmpeg_decoder.cc
21 * @brief A decoder using FFmpeg to decode content.
32 #include <libavcodec/avcodec.h>
33 #include <libavformat/avformat.h>
37 #include "exceptions.h"
41 #include "ffmpeg_decoder.h"
42 #include "filter_graph.h"
43 #include "audio_buffers.h"
44 #include "ffmpeg_content.h"
45 #include "image_proxy.h"
49 #define LOG_GENERAL(...) film->log()->log (String::compose (__VA_ARGS__), Log::TYPE_GENERAL);
50 #define LOG_ERROR(...) film->log()->log (String::compose (__VA_ARGS__), Log::TYPE_ERROR);
51 #define LOG_WARNING(...) film->log()->log (__VA_ARGS__, Log::TYPE_WARNING);
56 using std::stringstream;
60 using boost::shared_ptr;
61 using boost::optional;
62 using boost::dynamic_pointer_cast;
65 FFmpegDecoder::FFmpegDecoder (shared_ptr<const Film> f, shared_ptr<const FFmpegContent> c, bool video, bool audio)
71 , _subtitle_codec_context (0)
73 , _decode_video (video)
74 , _decode_audio (audio)
76 , _just_sought (false)
80 /* Audio and video frame PTS values may not start with 0. We want
81 to fiddle them so that:
83 1. One of them starts at time 0.
84 2. The first video PTS value ends up on a frame boundary.
86 Then we remove big initial gaps in PTS and we allow our
87 insertion of black frames to work.
90 audio_pts_to_use = audio_pts_from_ffmpeg + pts_offset;
91 video_pts_to_use = video_pts_from_ffmpeg + pts_offset;
94 bool const have_video = video && c->first_video();
95 bool const have_audio = audio && c->audio_stream() && c->audio_stream()->first_audio;
97 /* First, make one of them start at 0 */
99 if (have_audio && have_video) {
100 _pts_offset = - min (c->first_video().get(), c->audio_stream()->first_audio.get());
101 } else if (have_video) {
102 _pts_offset = - c->first_video().get();
103 } else if (have_audio) {
104 _pts_offset = - c->audio_stream()->first_audio.get();
107 /* Now adjust both so that the video pts starts on a frame */
108 if (have_video && have_audio) {
109 double first_video = c->first_video().get() + _pts_offset;
110 double const old_first_video = first_video;
112 /* Round the first video up to a frame boundary */
113 if (fabs (rint (first_video * c->video_frame_rate()) - first_video * c->video_frame_rate()) > 1e-6) {
114 first_video = ceil (first_video * c->video_frame_rate()) / c->video_frame_rate ();
117 _pts_offset += first_video - old_first_video;
121 FFmpegDecoder::~FFmpegDecoder ()
123 boost::mutex::scoped_lock lm (_mutex);
125 if (_subtitle_codec_context) {
126 avcodec_close (_subtitle_codec_context);
131 FFmpegDecoder::flush ()
133 /* Get any remaining frames */
138 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
141 while (decode_video_packet ()) {}
144 if (_ffmpeg_content->audio_stream() && _decode_audio) {
145 decode_audio_packet ();
148 /* Stop us being asked for any more data */
149 _video_position = _ffmpeg_content->video_length_after_3d_combine ();
150 _audio_position = _ffmpeg_content->audio_length ();
154 FFmpegDecoder::pass ()
156 int r = av_read_frame (_format_context, &_packet);
159 if (r != AVERROR_EOF) {
160 /* Maybe we should fail here, but for now we'll just finish off instead */
162 av_strerror (r, buf, sizeof(buf));
163 shared_ptr<const Film> film = _film.lock ();
165 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), buf, r);
172 shared_ptr<const Film> film = _film.lock ();
175 int const si = _packet.stream_index;
177 if (si == _video_stream && _decode_video) {
178 decode_video_packet ();
179 } else if (_ffmpeg_content->audio_stream() && _ffmpeg_content->audio_stream()->uses_index (_format_context, si) && _decode_audio) {
180 decode_audio_packet ();
181 } else if (_ffmpeg_content->subtitle_stream() && _ffmpeg_content->subtitle_stream()->uses_index (_format_context, si) && film->with_subtitles ()) {
182 decode_subtitle_packet ();
185 av_free_packet (&_packet);
188 /** @param data pointer to array of pointers to buffers.
189 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
191 shared_ptr<AudioBuffers>
192 FFmpegDecoder::deinterleave_audio (uint8_t** data, int size)
194 assert (_ffmpeg_content->audio_channels());
195 assert (bytes_per_audio_sample());
197 /* Deinterleave and convert to float */
199 assert ((size % (bytes_per_audio_sample() * _ffmpeg_content->audio_channels())) == 0);
201 int const total_samples = size / bytes_per_audio_sample();
202 int const frames = total_samples / _ffmpeg_content->audio_channels();
203 shared_ptr<AudioBuffers> audio (new AudioBuffers (_ffmpeg_content->audio_channels(), frames));
205 switch (audio_sample_format()) {
206 case AV_SAMPLE_FMT_U8:
208 uint8_t* p = reinterpret_cast<uint8_t *> (data[0]);
211 for (int i = 0; i < total_samples; ++i) {
212 audio->data(channel)[sample] = float(*p++) / (1 << 23);
215 if (channel == _ffmpeg_content->audio_channels()) {
223 case AV_SAMPLE_FMT_S16:
225 int16_t* p = reinterpret_cast<int16_t *> (data[0]);
228 for (int i = 0; i < total_samples; ++i) {
229 audio->data(channel)[sample] = float(*p++) / (1 << 15);
232 if (channel == _ffmpeg_content->audio_channels()) {
240 case AV_SAMPLE_FMT_S16P:
242 int16_t** p = reinterpret_cast<int16_t **> (data);
243 for (int i = 0; i < _ffmpeg_content->audio_channels(); ++i) {
244 for (int j = 0; j < frames; ++j) {
245 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
251 case AV_SAMPLE_FMT_S32:
253 int32_t* p = reinterpret_cast<int32_t *> (data[0]);
256 for (int i = 0; i < total_samples; ++i) {
257 audio->data(channel)[sample] = static_cast<float>(*p++) / (1 << 31);
260 if (channel == _ffmpeg_content->audio_channels()) {
268 case AV_SAMPLE_FMT_FLT:
270 float* p = reinterpret_cast<float*> (data[0]);
273 for (int i = 0; i < total_samples; ++i) {
274 audio->data(channel)[sample] = *p++;
277 if (channel == _ffmpeg_content->audio_channels()) {
285 case AV_SAMPLE_FMT_FLTP:
287 float** p = reinterpret_cast<float**> (data);
288 for (int i = 0; i < _ffmpeg_content->audio_channels(); ++i) {
289 memcpy (audio->data(i), p[i], frames * sizeof(float));
295 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format())));
302 FFmpegDecoder::audio_sample_format () const
304 if (!_ffmpeg_content->audio_stream()) {
305 return (AVSampleFormat) 0;
308 return audio_codec_context()->sample_fmt;
312 FFmpegDecoder::bytes_per_audio_sample () const
314 return av_get_bytes_per_sample (audio_sample_format ());
318 FFmpegDecoder::seek (VideoContent::Frame frame, bool accurate)
320 double const time_base = av_q2d (_format_context->streams[_video_stream]->time_base);
322 /* If we are doing an accurate seek, our initial shot will be 5 frames (5 being
323 a number plucked from the air) earlier than we want to end up. The loop below
324 will hopefully then step through to where we want to be.
336 /* Initial seek time in the stream's timebase */
337 int64_t const initial_vt = ((initial / _ffmpeg_content->video_frame_rate()) - _pts_offset) / time_base;
339 av_seek_frame (_format_context, _video_stream, initial_vt, AVSEEK_FLAG_BACKWARD);
341 avcodec_flush_buffers (video_codec_context());
342 if (_subtitle_codec_context) {
343 avcodec_flush_buffers (_subtitle_codec_context);
346 /* This !accurate is piling hack upon hack; setting _just_sought to true
347 even with accurate == true defeats our attempt to align the start
348 of the video and audio. Here we disable that defeat when accurate == true
349 i.e. when we are making a DCP rather than just previewing one.
350 Ewww. This should be gone in 2.0.
356 _video_position = frame;
358 if (frame == 0 || !accurate) {
359 /* We're already there, or we're as close as we need to be */
364 int r = av_read_frame (_format_context, &_packet);
369 if (_packet.stream_index != _video_stream) {
370 av_free_packet (&_packet);
375 r = avcodec_decode_video2 (video_codec_context(), _frame, &finished, &_packet);
376 if (r >= 0 && finished) {
377 _video_position = rint (
378 (av_frame_get_best_effort_timestamp (_frame) * time_base + _pts_offset) * _ffmpeg_content->video_frame_rate()
381 if (_video_position >= (frame - 1)) {
382 av_free_packet (&_packet);
387 av_free_packet (&_packet);
390 /* _video_position should be the next thing to be emitted, which will the one after the thing
397 FFmpegDecoder::decode_audio_packet ()
399 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
403 AVPacket copy_packet = _packet;
405 while (copy_packet.size > 0) {
408 int const decode_result = avcodec_decode_audio4 (audio_codec_context(), _frame, &frame_finished, ©_packet);
409 if (decode_result < 0) {
410 shared_ptr<const Film> film = _film.lock ();
412 LOG_ERROR ("avcodec_decode_audio4 failed (%1)", decode_result);
416 if (frame_finished) {
418 if (_audio_position == 0) {
419 /* Where we are in the source, in seconds */
420 double const pts = av_q2d (_format_context->streams[copy_packet.stream_index]->time_base)
421 * av_frame_get_best_effort_timestamp(_frame) + _pts_offset;
424 /* Emit some silence */
425 int64_t frames = pts * _ffmpeg_content->content_audio_frame_rate ();
427 int64_t const this_time = min (frames, (int64_t) _ffmpeg_content->content_audio_frame_rate() / 2);
429 shared_ptr<AudioBuffers> silence (
430 new AudioBuffers (_ffmpeg_content->audio_channels(), this_time)
433 silence->make_silent ();
434 audio (silence, _audio_position);
440 int const data_size = av_samples_get_buffer_size (
441 0, audio_codec_context()->channels, _frame->nb_samples, audio_sample_format (), 1
444 audio (deinterleave_audio (_frame->data, data_size), _audio_position);
447 copy_packet.data += decode_result;
448 copy_packet.size -= decode_result;
453 FFmpegDecoder::decode_video_packet ()
456 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
460 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
462 shared_ptr<FilterGraph> graph;
464 list<shared_ptr<FilterGraph> >::iterator i = _filter_graphs.begin();
465 while (i != _filter_graphs.end() && !(*i)->can_process (libdcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
469 if (i == _filter_graphs.end ()) {
470 shared_ptr<const Film> film = _film.lock ();
473 graph.reset (new FilterGraph (_ffmpeg_content, libdcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
474 _filter_graphs.push_back (graph);
476 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
481 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
483 shared_ptr<const Film> film = _film.lock ();
486 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
488 shared_ptr<Image> image = i->first;
490 if (i->second != AV_NOPTS_VALUE) {
492 double const pts = i->second * av_q2d (_format_context->streams[_video_stream]->time_base) + _pts_offset;
495 /* We just did a seek, so disable any attempts to correct for where we
498 _video_position = rint (pts * _ffmpeg_content->video_frame_rate ());
499 _just_sought = false;
502 double const next = _video_position / _ffmpeg_content->video_frame_rate();
503 double const one_frame = 1 / _ffmpeg_content->video_frame_rate ();
504 double delta = pts - next;
506 while (delta > one_frame) {
507 /* This PTS is more than one frame forward in time of where we think we should be; emit
511 /* XXX: I think this should be a copy of the last frame... */
512 boost::shared_ptr<Image> black (
514 static_cast<AVPixelFormat> (_frame->format),
515 libdcp::Size (video_codec_context()->width, video_codec_context()->height),
520 shared_ptr<const Film> film = _film.lock ();
523 black->make_black ();
524 video (shared_ptr<ImageProxy> (new RawImageProxy (image, film->log())), false, _video_position);
528 if (delta > -one_frame) {
529 /* This PTS is within a frame of being right; emit this (otherwise it will be dropped) */
530 video (shared_ptr<ImageProxy> (new RawImageProxy (image, film->log())), false, _video_position);
534 LOG_WARNING ("Dropping frame without PTS");
543 FFmpegDecoder::setup_subtitle ()
545 boost::mutex::scoped_lock lm (_mutex);
547 if (!_ffmpeg_content->subtitle_stream()) {
551 _subtitle_codec_context = _ffmpeg_content->subtitle_stream()->stream(_format_context)->codec;
552 if (_subtitle_codec_context == 0) {
553 throw DecodeError (N_("could not find subtitle stream"));
556 _subtitle_codec = avcodec_find_decoder (_subtitle_codec_context->codec_id);
558 if (_subtitle_codec == 0) {
559 throw DecodeError (N_("could not find subtitle decoder"));
562 if (avcodec_open2 (_subtitle_codec_context, _subtitle_codec, 0) < 0) {
563 throw DecodeError (N_("could not open subtitle decoder"));
568 FFmpegDecoder::done () const
570 bool const vd = !_decode_video || (_video_position >= _ffmpeg_content->video_length());
571 bool const ad = !_decode_audio || !_ffmpeg_content->audio_stream() || (_audio_position >= _ffmpeg_content->audio_length());
576 FFmpegDecoder::decode_subtitle_packet ()
580 if (avcodec_decode_subtitle2 (_subtitle_codec_context, &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
584 /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
585 indicate that the previous subtitle should stop.
587 if (sub.num_rects <= 0) {
588 subtitle (shared_ptr<Image> (), dcpomatic::Rect<double> (), 0, 0);
590 } else if (sub.num_rects > 1) {
591 throw DecodeError (_("multi-part subtitles not yet supported"));
594 /* Subtitle PTS in seconds (within the source, not taking into account any of the
595 source that we may have chopped off for the DCP)
597 double const packet_time = (static_cast<double> (sub.pts ) / AV_TIME_BASE) + _pts_offset;
599 /* hence start time for this sub */
600 Time const from = (packet_time + (double (sub.start_display_time) / 1e3)) * TIME_HZ;
601 Time const to = (packet_time + (double (sub.end_display_time) / 1e3)) * TIME_HZ;
603 AVSubtitleRect const * rect = sub.rects[0];
605 if (rect->type != SUBTITLE_BITMAP) {
606 throw DecodeError (_("non-bitmap subtitles not yet supported"));
609 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
610 G, third B, fourth A.
612 shared_ptr<Image> image (new Image (PIX_FMT_RGBA, libdcp::Size (rect->w, rect->h), true));
614 /* Start of the first line in the subtitle */
615 uint8_t* sub_p = rect->pict.data[0];
616 /* sub_p looks up into a BGRA palette which is here
617 (i.e. first byte B, second G, third R, fourth A)
619 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
620 /* Start of the output data */
621 uint32_t* out_p = (uint32_t *) image->data()[0];
623 for (int y = 0; y < rect->h; ++y) {
624 uint8_t* sub_line_p = sub_p;
625 uint32_t* out_line_p = out_p;
626 for (int x = 0; x < rect->w; ++x) {
627 uint32_t const p = palette[*sub_line_p++];
628 *out_line_p++ = ((p & 0xff) << 16) | (p & 0xff00) | ((p & 0xff0000) >> 16) | (p & 0xff000000);
630 sub_p += rect->pict.linesize[0];
631 out_p += image->stride()[0] / sizeof (uint32_t);
634 libdcp::Size const vs = _ffmpeg_content->video_size ();
638 dcpomatic::Rect<double> (
639 static_cast<double> (rect->x) / vs.width,
640 static_cast<double> (rect->y) / vs.height,
641 static_cast<double> (rect->w) / vs.width,
642 static_cast<double> (rect->h) / vs.height
649 avsubtitle_free (&sub);