2 Copyright (C) 2012-2018 Carl Hetherington <cth@carlh.net>
4 This file is part of DCP-o-matic.
6 DCP-o-matic is free software; you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 2 of the License, or
9 (at your option) any later version.
11 DCP-o-matic is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with DCP-o-matic. If not, see <http://www.gnu.org/licenses/>.
21 /** @file src/ffmpeg_decoder.cc
22 * @brief A decoder using FFmpeg to decode content.
26 #include "exceptions.h"
30 #include "dcpomatic_log.h"
31 #include "ffmpeg_decoder.h"
32 #include "text_decoder.h"
33 #include "ffmpeg_audio_stream.h"
34 #include "ffmpeg_subtitle_stream.h"
35 #include "video_filter_graph.h"
36 #include "audio_buffers.h"
37 #include "ffmpeg_content.h"
38 #include "raw_image_proxy.h"
39 #include "video_decoder.h"
41 #include "audio_decoder.h"
42 #include "compose.hpp"
43 #include "text_content.h"
44 #include "audio_content.h"
45 #include <dcp/subtitle_string.h>
46 #include <sub/ssa_reader.h>
47 #include <sub/subtitle.h>
48 #include <sub/collect.h>
50 #include <libavcodec/avcodec.h>
51 #include <libavformat/avformat.h>
53 #include <boost/foreach.hpp>
54 #include <boost/algorithm/string.hpp>
70 using boost::shared_ptr;
71 using boost::is_any_of;
73 using boost::optional;
74 using boost::dynamic_pointer_cast;
76 using namespace dcpomatic;
78 FFmpegDecoder::FFmpegDecoder (shared_ptr<const Film> film, shared_ptr<const FFmpegContent> c, bool fast)
81 , _have_current_subtitle (false)
83 if (c->video && c->video->use()) {
84 video.reset (new VideoDecoder (this, c));
85 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate(film));
86 /* It doesn't matter what size or pixel format this is, it just needs to be black */
87 _black_image.reset (new Image (AV_PIX_FMT_RGB24, dcp::Size (128, 128), true));
88 _black_image->make_black ();
90 _pts_offset = ContentTime ();
94 audio.reset (new AudioDecoder (this, c->audio, fast));
98 /* XXX: this time here should be the time of the first subtitle, not 0 */
99 text.push_back (shared_ptr<TextDecoder> (new TextDecoder (this, c->only_text(), ContentTime())));
102 _next_time.resize (_format_context->nb_streams);
106 FFmpegDecoder::flush ()
108 /* Get any remaining frames */
113 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
115 while (video && decode_video_packet()) {}
118 decode_audio_packet ();
121 /* Make sure all streams are the same length and round up to the next video frame */
123 FrameRateChange const frc = film()->active_frame_rate_change(_ffmpeg_content->position());
124 ContentTime full_length (_ffmpeg_content->full_length(film()), frc);
125 full_length = full_length.ceil (frc.source);
127 double const vfr = _ffmpeg_content->video_frame_rate().get();
128 Frame const f = full_length.frames_round (vfr);
129 Frame v = video->position(film()).get_value_or(ContentTime()).frames_round(vfr) + 1;
131 video->emit (film(), shared_ptr<const ImageProxy> (new RawImageProxy (_black_image)), v);
136 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, _ffmpeg_content->ffmpeg_audio_streams ()) {
137 ContentTime a = audio->stream_position(film(), i);
138 /* Unfortunately if a is 0 that really means that we don't know the stream position since
139 there has been no data on it since the last seek. In this case we'll just do nothing
140 here. I'm not sure if that's the right idea.
142 if (a > ContentTime()) {
143 while (a < full_length) {
144 ContentTime to_do = min (full_length - a, ContentTime::from_seconds (0.1));
145 shared_ptr<AudioBuffers> silence (new AudioBuffers (i->channels(), to_do.frames_ceil (i->frame_rate())));
146 silence->make_silent ();
147 audio->emit (film(), i, silence, a);
159 FFmpegDecoder::pass ()
161 #ifdef DCPOMATIC_VARIANT_SWAROOP
162 if (_ffmpeg_content->encrypted() && !_ffmpeg_content->kdm()) {
167 int r = av_read_frame (_format_context, &_packet);
169 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
170 has pretty-much succeeded (and hence generated data which should be processed).
171 Hence it makes sense to continue here in that case.
173 if (r < 0 && r != AVERROR_INVALIDDATA) {
174 if (r != AVERROR_EOF) {
175 /* Maybe we should fail here, but for now we'll just finish off instead */
177 av_strerror (r, buf, sizeof(buf));
178 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
185 int const si = _packet.stream_index;
186 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
188 if (_video_stream && si == _video_stream.get() && video && !video->ignore()) {
189 decode_video_packet ();
190 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index(_format_context, si) && !only_text()->ignore()) {
191 decode_subtitle_packet ();
193 decode_audio_packet ();
196 av_packet_unref (&_packet);
200 /** @param data pointer to array of pointers to buffers.
201 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
203 shared_ptr<AudioBuffers>
204 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
206 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
208 int const size = av_samples_get_buffer_size (
209 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
212 /* Deinterleave and convert to float */
214 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
215 of the block that do not form a complete sample or frame they will be dropped.
217 int const total_samples = size / bytes_per_audio_sample (stream);
218 int const channels = stream->channels();
219 int const frames = total_samples / channels;
220 shared_ptr<AudioBuffers> audio (new AudioBuffers (channels, frames));
221 float** data = audio->data();
223 switch (audio_sample_format (stream)) {
224 case AV_SAMPLE_FMT_U8:
226 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
229 for (int i = 0; i < total_samples; ++i) {
230 data[channel][sample] = float(*p++) / (1 << 23);
233 if (channel == channels) {
241 case AV_SAMPLE_FMT_S16:
243 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
246 for (int i = 0; i < total_samples; ++i) {
247 data[channel][sample] = float(*p++) / (1 << 15);
250 if (channel == channels) {
258 case AV_SAMPLE_FMT_S16P:
260 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
261 for (int i = 0; i < channels; ++i) {
262 for (int j = 0; j < frames; ++j) {
263 data[i][j] = static_cast<float>(p[i][j]) / (1 << 15);
269 case AV_SAMPLE_FMT_S32:
271 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
274 for (int i = 0; i < total_samples; ++i) {
275 data[channel][sample] = static_cast<float>(*p++) / 2147483648;
278 if (channel == channels) {
286 case AV_SAMPLE_FMT_S32P:
288 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
289 for (int i = 0; i < channels; ++i) {
290 for (int j = 0; j < frames; ++j) {
291 data[i][j] = static_cast<float>(p[i][j]) / 2147483648;
297 case AV_SAMPLE_FMT_FLT:
299 float* p = reinterpret_cast<float*> (_frame->data[0]);
302 for (int i = 0; i < total_samples; ++i) {
303 data[channel][sample] = *p++;
306 if (channel == channels) {
314 case AV_SAMPLE_FMT_FLTP:
316 float** p = reinterpret_cast<float**> (_frame->data);
317 /* Sometimes there aren't as many channels in the _frame as in the stream */
318 for (int i = 0; i < _frame->channels; ++i) {
319 memcpy (data[i], p[i], frames * sizeof(float));
321 for (int i = _frame->channels; i < channels; ++i) {
322 audio->make_silent (i);
328 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
335 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
337 return stream->stream (_format_context)->codec->sample_fmt;
341 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
343 return av_get_bytes_per_sample (audio_sample_format (stream));
347 FFmpegDecoder::seek (ContentTime time, bool accurate)
349 Decoder::seek (time, accurate);
351 /* If we are doing an `accurate' seek, we need to use pre-roll, as
352 we don't really know what the seek will give us.
355 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
358 /* XXX: it seems debatable whether PTS should be used here...
359 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
362 optional<int> stream;
365 stream = _video_stream;
367 DCPOMATIC_ASSERT (_ffmpeg_content->audio);
368 shared_ptr<FFmpegAudioStream> s = dynamic_pointer_cast<FFmpegAudioStream> (_ffmpeg_content->audio->stream ());
370 stream = s->index (_format_context);
374 DCPOMATIC_ASSERT (stream);
376 ContentTime u = time - _pts_offset;
377 if (u < ContentTime ()) {
383 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
388 /* Force re-creation of filter graphs to reset them and hence to make sure
389 they don't have any pre-seek frames knocking about.
391 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
392 _filter_graphs.clear ();
395 if (video_codec_context ()) {
396 avcodec_flush_buffers (video_codec_context());
399 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, ffmpeg_content()->ffmpeg_audio_streams()) {
400 avcodec_flush_buffers (i->stream(_format_context)->codec);
403 if (subtitle_codec_context ()) {
404 avcodec_flush_buffers (subtitle_codec_context ());
407 _have_current_subtitle = false;
411 FFmpegDecoder::decode_audio_packet ()
413 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
417 AVPacket copy_packet = _packet;
418 int const stream_index = copy_packet.stream_index;
420 /* XXX: inefficient */
421 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
422 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
423 while (stream != streams.end () && !(*stream)->uses_index (_format_context, stream_index)) {
427 if (stream == streams.end ()) {
428 /* The packet's stream may not be an audio one; just ignore it in this method if so */
432 while (copy_packet.size > 0) {
435 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
436 if (decode_result < 0) {
437 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
438 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
439 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
440 even in the face of such an error, so I think we should too.
442 Returning from the method here caused mantis #352.
444 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
446 /* Fudge decode_result so that we come out of the while loop when
447 we've processed this data.
449 decode_result = copy_packet.size;
452 if (frame_finished) {
453 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
456 if (_frame->pts == AV_NOPTS_VALUE && _next_time[stream_index]) {
457 /* In some streams we see not every frame coming through with a timestamp; for those
458 that have AV_NOPTS_VALUE we need to work out the timestamp ourselves. This is
459 particularly noticeable with TrueHD streams (see #1111).
461 ct = *_next_time[stream_index];
463 ct = ContentTime::from_seconds (
464 av_frame_get_best_effort_timestamp (_frame) *
465 av_q2d ((*stream)->stream (_format_context)->time_base))
469 _next_time[stream_index] = ct + ContentTime::from_frames(data->frames(), (*stream)->frame_rate());
471 if (ct < ContentTime ()) {
472 /* Discard audio data that comes before time 0 */
473 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
474 data->move (data->frames() - remove, remove, 0);
475 data->set_frames (data->frames() - remove);
476 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
479 if (ct < ContentTime()) {
481 "Crazy timestamp %1 for %2 samples in stream %3 packet pts %4 (ts=%5 tb=%6, off=%7)",
484 copy_packet.stream_index,
486 av_frame_get_best_effort_timestamp(_frame),
487 av_q2d((*stream)->stream(_format_context)->time_base),
488 to_string(_pts_offset)
492 /* Give this data provided there is some, and its time is sane */
493 if (ct >= ContentTime() && data->frames() > 0) {
494 audio->emit (film(), *stream, data, ct);
498 copy_packet.data += decode_result;
499 copy_packet.size -= decode_result;
504 FFmpegDecoder::decode_video_packet ()
506 DCPOMATIC_ASSERT (_video_stream);
509 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
513 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
515 shared_ptr<VideoFilterGraph> graph;
517 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
518 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
522 if (i == _filter_graphs.end ()) {
523 dcp::Fraction vfr (lrint(_ffmpeg_content->video_frame_rate().get() * 1000), 1000);
524 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format, vfr));
525 graph->setup (_ffmpeg_content->filters ());
526 _filter_graphs.push_back (graph);
527 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
532 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
534 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
536 shared_ptr<Image> image = i->first;
538 if (i->second != AV_NOPTS_VALUE) {
539 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
543 shared_ptr<ImageProxy> (new RawImageProxy (image)),
544 llrint(pts * _ffmpeg_content->active_video_frame_rate(film()))
547 LOG_WARNING_NC ("Dropping frame without PTS");
555 FFmpegDecoder::decode_subtitle_packet ()
559 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
563 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
564 if (_have_current_subtitle) {
565 if (_current_subtitle_to) {
566 only_text()->emit_stop (min(*_current_subtitle_to, subtitle_period(sub).from + _pts_offset));
568 only_text()->emit_stop (subtitle_period(sub).from + _pts_offset);
570 _have_current_subtitle = false;
573 if (sub.num_rects <= 0) {
574 /* Nothing new in this subtitle */
578 /* Subtitle PTS (within the source, not taking into account any of the
579 source that we may have chopped off for the DCP).
581 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
583 from = sub_period.from + _pts_offset;
585 _current_subtitle_to = *sub_period.to + _pts_offset;
587 _current_subtitle_to = optional<ContentTime>();
588 _have_current_subtitle = true;
591 for (unsigned int i = 0; i < sub.num_rects; ++i) {
592 AVSubtitleRect const * rect = sub.rects[i];
594 switch (rect->type) {
597 case SUBTITLE_BITMAP:
598 decode_bitmap_subtitle (rect, from);
601 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
604 decode_ass_subtitle (rect->ass, from);
609 if (_current_subtitle_to) {
610 only_text()->emit_stop (*_current_subtitle_to);
613 avsubtitle_free (&sub);
617 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTime from)
619 /* Note BGRA is expressed little-endian, so the first byte in the word is B, second
620 G, third R, fourth A.
622 shared_ptr<Image> image (new Image (AV_PIX_FMT_BGRA, dcp::Size (rect->w, rect->h), true));
624 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
625 /* Start of the first line in the subtitle */
626 uint8_t* sub_p = rect->pict.data[0];
627 /* sub_p looks up into a BGRA palette which is here
628 (i.e. first byte B, second G, third R, fourth A)
630 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
632 /* Start of the first line in the subtitle */
633 uint8_t* sub_p = rect->data[0];
634 /* sub_p looks up into a BGRA palette which is at rect->data[1].
635 (first byte B, second G, third R, fourth A)
638 /* And the stream has a map of those palette colours to colours
639 chosen by the user; created a `mapped' palette from those settings.
641 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
642 vector<RGBA> mapped_palette (rect->nb_colors);
643 uint8_t const * palette = rect->data[1];
644 for (int i = 0; i < rect->nb_colors; ++i) {
645 RGBA c (palette[2], palette[1], palette[0], palette[3]);
646 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
647 if (j != colour_map.end ()) {
648 mapped_palette[i] = j->second;
650 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
651 it is from a project that was created before this stuff was added. Just use the
652 colour straight from the original palette.
654 mapped_palette[i] = c;
659 /* Start of the output data */
660 uint8_t* out_p = image->data()[0];
662 for (int y = 0; y < rect->h; ++y) {
663 uint8_t* sub_line_p = sub_p;
664 uint8_t* out_line_p = out_p;
665 for (int x = 0; x < rect->w; ++x) {
666 RGBA const p = mapped_palette[*sub_line_p++];
672 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
673 sub_p += rect->pict.linesize[0];
675 sub_p += rect->linesize[0];
677 out_p += image->stride()[0];
680 int target_width = subtitle_codec_context()->width;
681 if (target_width == 0 && video_codec_context()) {
682 /* subtitle_codec_context()->width == 0 has been seen in the wild but I don't
683 know if it's supposed to mean something from FFmpeg's point of view.
685 target_width = video_codec_context()->width;
687 int target_height = subtitle_codec_context()->height;
688 if (target_height == 0 && video_codec_context()) {
689 target_height = video_codec_context()->height;
691 DCPOMATIC_ASSERT (target_width);
692 DCPOMATIC_ASSERT (target_height);
693 dcpomatic::Rect<double> const scaled_rect (
694 static_cast<double> (rect->x) / target_width,
695 static_cast<double> (rect->y) / target_height,
696 static_cast<double> (rect->w) / target_width,
697 static_cast<double> (rect->h) / target_height
700 only_text()->emit_bitmap_start (from, image, scaled_rect);
704 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTime from)
706 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
707 produces a single format of Dialogue: lines...
712 for (size_t i = 0; i < ass.length(); ++i) {
713 if (commas < 9 && ass[i] == ',') {
715 } else if (commas == 9) {
724 sub::RawSubtitle base;
725 list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (
728 _ffmpeg_content->video->size().width,
729 _ffmpeg_content->video->size().height
732 BOOST_FOREACH (sub::Subtitle const & i, sub::collect<list<sub::Subtitle> > (raw)) {
733 only_text()->emit_plain_start (from, i);