2 Copyright (C) 2012-2015 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.
25 #include "exceptions.h"
29 #include "ffmpeg_decoder.h"
30 #include "ffmpeg_audio_stream.h"
31 #include "ffmpeg_subtitle_stream.h"
32 #include "filter_graph.h"
33 #include "audio_buffers.h"
34 #include "ffmpeg_content.h"
35 #include "raw_image_proxy.h"
37 #include "compose.hpp"
39 #include <libavcodec/avcodec.h>
40 #include <libavformat/avformat.h>
42 #include <boost/foreach.hpp>
50 #define LOG_GENERAL(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_GENERAL);
51 #define LOG_ERROR(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_ERROR);
52 #define LOG_WARNING_NC(...) _video_content->film()->log()->log (__VA_ARGS__, Log::TYPE_WARNING);
53 #define LOG_WARNING(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_WARNING);
61 using boost::shared_ptr;
64 FFmpegDecoder::FFmpegDecoder (shared_ptr<const FFmpegContent> c, shared_ptr<Log> log)
71 /* Audio and video frame PTS values may not start with 0. We want
72 to fiddle them so that:
74 1. One of them starts at time 0.
75 2. The first video PTS value ends up on a frame boundary.
77 Then we remove big initial gaps in PTS and we allow our
78 insertion of black frames to work.
81 audio_pts_to_use = audio_pts_from_ffmpeg + pts_offset;
82 video_pts_to_use = video_pts_from_ffmpeg + pts_offset;
85 /* First, make one of them start at 0 */
87 vector<shared_ptr<FFmpegAudioStream> > streams = c->ffmpeg_audio_streams ();
89 _pts_offset = ContentTime::min ();
91 if (c->first_video ()) {
92 _pts_offset = - c->first_video().get ();
95 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, streams) {
97 _pts_offset = max (_pts_offset, - i->first_audio.get ());
101 /* If _pts_offset is positive we would be pushing things from a -ve PTS to be played.
102 I don't think we ever want to do that, as it seems things at -ve PTS are not meant
103 to be seen (use for alignment bars etc.); see mantis #418.
105 if (_pts_offset > ContentTime ()) {
106 _pts_offset = ContentTime ();
109 /* Now adjust so that the video pts starts on a frame */
110 if (c->first_video ()) {
111 ContentTime first_video = c->first_video().get() + _pts_offset;
112 ContentTime const old_first_video = first_video;
113 _pts_offset += first_video.round_up (c->video_frame_rate ()) - old_first_video;
118 FFmpegDecoder::flush ()
120 /* Get any remaining frames */
125 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
127 while (decode_video_packet ()) {}
129 decode_audio_packet ();
130 AudioDecoder::flush ();
134 FFmpegDecoder::pass ()
136 int r = av_read_frame (_format_context, &_packet);
138 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
139 has pretty-much succeeded (and hence generated data which should be processed).
140 Hence it makes sense to continue here in that case.
142 if (r < 0 && r != AVERROR_INVALIDDATA) {
143 if (r != AVERROR_EOF) {
144 /* Maybe we should fail here, but for now we'll just finish off instead */
146 av_strerror (r, buf, sizeof(buf));
147 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), buf, r);
154 int const si = _packet.stream_index;
155 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
157 if (si == _video_stream && !_ignore_video) {
158 decode_video_packet ();
159 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index (_format_context, si)) {
160 decode_subtitle_packet ();
162 decode_audio_packet ();
165 av_free_packet (&_packet);
169 /** @param data pointer to array of pointers to buffers.
170 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
172 shared_ptr<AudioBuffers>
173 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream, uint8_t** data, int size)
175 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
177 /* Deinterleave and convert to float */
179 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
180 of the block that do not form a complete sample or frame they will be dropped.
182 int const total_samples = size / bytes_per_audio_sample (stream);
183 int const frames = total_samples / stream->channels();
184 shared_ptr<AudioBuffers> audio (new AudioBuffers (stream->channels(), frames));
186 switch (audio_sample_format (stream)) {
187 case AV_SAMPLE_FMT_U8:
189 uint8_t* p = reinterpret_cast<uint8_t *> (data[0]);
192 for (int i = 0; i < total_samples; ++i) {
193 audio->data(channel)[sample] = float(*p++) / (1 << 23);
196 if (channel == stream->channels()) {
204 case AV_SAMPLE_FMT_S16:
206 int16_t* p = reinterpret_cast<int16_t *> (data[0]);
209 for (int i = 0; i < total_samples; ++i) {
210 audio->data(channel)[sample] = float(*p++) / (1 << 15);
213 if (channel == stream->channels()) {
221 case AV_SAMPLE_FMT_S16P:
223 int16_t** p = reinterpret_cast<int16_t **> (data);
224 for (int i = 0; i < stream->channels(); ++i) {
225 for (int j = 0; j < frames; ++j) {
226 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
232 case AV_SAMPLE_FMT_S32:
234 int32_t* p = reinterpret_cast<int32_t *> (data[0]);
237 for (int i = 0; i < total_samples; ++i) {
238 audio->data(channel)[sample] = static_cast<float>(*p++) / (1 << 31);
241 if (channel == stream->channels()) {
249 case AV_SAMPLE_FMT_FLT:
251 float* p = reinterpret_cast<float*> (data[0]);
254 for (int i = 0; i < total_samples; ++i) {
255 audio->data(channel)[sample] = *p++;
258 if (channel == stream->channels()) {
266 case AV_SAMPLE_FMT_FLTP:
268 float** p = reinterpret_cast<float**> (data);
269 for (int i = 0; i < stream->channels(); ++i) {
270 memcpy (audio->data(i), p[i], frames * sizeof(float));
276 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
283 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
285 return stream->stream (_format_context)->codec->sample_fmt;
289 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
291 return av_get_bytes_per_sample (audio_sample_format (stream));
295 FFmpegDecoder::seek (ContentTime time, bool accurate)
297 VideoDecoder::seek (time, accurate);
298 AudioDecoder::seek (time, accurate);
299 SubtitleDecoder::seek (time, accurate);
301 /* If we are doing an `accurate' seek, we need to use pre-roll, as
302 we don't really know what the seek will give us.
305 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
308 /* XXX: it seems debatable whether PTS should be used here...
309 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
312 ContentTime u = time - _pts_offset;
313 if (u < ContentTime ()) {
316 av_seek_frame (_format_context, _video_stream, u.seconds() / av_q2d (_format_context->streams[_video_stream]->time_base), AVSEEK_FLAG_BACKWARD);
318 avcodec_flush_buffers (video_codec_context());
320 /* XXX: should be flushing audio buffers? */
322 if (subtitle_codec_context ()) {
323 avcodec_flush_buffers (subtitle_codec_context ());
328 FFmpegDecoder::decode_audio_packet ()
330 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
334 AVPacket copy_packet = _packet;
336 /* XXX: inefficient */
337 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
338 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
339 while (stream != streams.end () && !(*stream)->uses_index (_format_context, copy_packet.stream_index)) {
343 if (stream == streams.end ()) {
344 /* The packet's stream may not be an audio one; just ignore it in this method if so */
348 while (copy_packet.size > 0) {
351 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
352 if (decode_result < 0) {
353 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
354 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
355 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
356 even in the face of such an error, so I think we should too.
358 Returning from the method here caused mantis #352.
360 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
362 /* Fudge decode_result so that we come out of the while loop when
363 we've processed this data.
365 decode_result = copy_packet.size;
368 if (frame_finished) {
369 ContentTime const ct = ContentTime::from_seconds (
370 av_frame_get_best_effort_timestamp (_frame) *
371 av_q2d ((*stream)->stream (_format_context)->time_base))
374 int const data_size = av_samples_get_buffer_size (
375 0, (*stream)->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (*stream), 1
378 audio (*stream, deinterleave_audio (*stream, _frame->data, data_size), ct);
381 copy_packet.data += decode_result;
382 copy_packet.size -= decode_result;
387 FFmpegDecoder::decode_video_packet ()
390 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
394 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
396 shared_ptr<FilterGraph> graph;
398 list<shared_ptr<FilterGraph> >::iterator i = _filter_graphs.begin();
399 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
403 if (i == _filter_graphs.end ()) {
404 graph.reset (new FilterGraph (_ffmpeg_content, dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
405 _filter_graphs.push_back (graph);
406 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
411 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
413 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
415 shared_ptr<Image> image = i->first;
417 if (i->second != AV_NOPTS_VALUE) {
418 double const pts = i->second * av_q2d (_format_context->streams[_video_stream]->time_base) + _pts_offset.seconds ();
420 shared_ptr<ImageProxy> (new RawImageProxy (image)),
421 rint (pts * _ffmpeg_content->video_frame_rate ())
424 LOG_WARNING_NC ("Dropping frame without PTS");
432 FFmpegDecoder::decode_subtitle_packet ()
436 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
440 if (sub.num_rects <= 0) {
441 /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
442 indicate that the previous subtitle should stop. We can ignore it here.
445 } else if (sub.num_rects > 1) {
446 throw DecodeError (_("multi-part subtitles not yet supported"));
449 /* Subtitle PTS (within the source, not taking into account any of the
450 source that we may have chopped off for the DCP).
452 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
453 ContentTimePeriod period;
454 period.from = sub_period.from + _pts_offset;
456 /* We already know the subtitle period `to' time */
457 period.to = sub_period.to.get() + _pts_offset;
459 /* We have to look up the `to' time in the stream's records */
460 period.to = ffmpeg_content()->subtitle_stream()->find_subtitle_to (sub_period.from);
463 AVSubtitleRect const * rect = sub.rects[0];
465 switch (rect->type) {
468 case SUBTITLE_BITMAP:
469 decode_bitmap_subtitle (rect, period);
472 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
475 cout << "XXX: SUBTITLE_ASS " << rect->ass << "\n";
479 avsubtitle_free (&sub);
482 list<ContentTimePeriod>
483 FFmpegDecoder::image_subtitles_during (ContentTimePeriod p, bool starting) const
485 return _ffmpeg_content->subtitles_during (p, starting);
488 list<ContentTimePeriod>
489 FFmpegDecoder::text_subtitles_during (ContentTimePeriod, bool) const
491 return list<ContentTimePeriod> ();
495 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTimePeriod period)
497 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
498 G, third B, fourth A.
500 shared_ptr<Image> image (new Image (PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
502 /* Start of the first line in the subtitle */
503 uint8_t* sub_p = rect->pict.data[0];
504 /* sub_p looks up into a BGRA palette which is here
505 (i.e. first byte B, second G, third R, fourth A)
507 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
508 /* Start of the output data */
509 uint32_t* out_p = (uint32_t *) image->data()[0];
511 for (int y = 0; y < rect->h; ++y) {
512 uint8_t* sub_line_p = sub_p;
513 uint32_t* out_line_p = out_p;
514 for (int x = 0; x < rect->w; ++x) {
515 uint32_t const p = palette[*sub_line_p++];
516 *out_line_p++ = ((p & 0xff) << 16) | (p & 0xff00) | ((p & 0xff0000) >> 16) | (p & 0xff000000);
518 sub_p += rect->pict.linesize[0];
519 out_p += image->stride()[0] / sizeof (uint32_t);
522 dcp::Size const vs = _ffmpeg_content->video_size ();
523 dcpomatic::Rect<double> const scaled_rect (
524 static_cast<double> (rect->x) / vs.width,
525 static_cast<double> (rect->y) / vs.height,
526 static_cast<double> (rect->w) / vs.width,
527 static_cast<double> (rect->h) / vs.height
530 image_subtitle (period, image, scaled_rect);