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>
71 using boost::shared_ptr;
72 using boost::is_any_of;
74 using boost::optional;
75 using boost::dynamic_pointer_cast;
78 FFmpegDecoder::FFmpegDecoder (shared_ptr<const Film> film, shared_ptr<const FFmpegContent> c, bool fast)
81 , _have_current_subtitle (false)
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()).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 int r = av_read_frame (_format_context, &_packet);
163 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
164 has pretty-much succeeded (and hence generated data which should be processed).
165 Hence it makes sense to continue here in that case.
167 if (r < 0 && r != AVERROR_INVALIDDATA) {
168 if (r != AVERROR_EOF) {
169 /* Maybe we should fail here, but for now we'll just finish off instead */
171 av_strerror (r, buf, sizeof(buf));
172 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
179 int const si = _packet.stream_index;
180 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
182 if (_video_stream && si == _video_stream.get() && !video->ignore()) {
183 decode_video_packet ();
184 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index(_format_context, si) && !only_text()->ignore()) {
185 decode_subtitle_packet ();
187 decode_audio_packet ();
190 av_packet_unref (&_packet);
194 /** @param data pointer to array of pointers to buffers.
195 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
197 shared_ptr<AudioBuffers>
198 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
200 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
202 int const size = av_samples_get_buffer_size (
203 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
206 /* Deinterleave and convert to float */
208 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
209 of the block that do not form a complete sample or frame they will be dropped.
211 int const total_samples = size / bytes_per_audio_sample (stream);
212 int const channels = stream->channels();
213 int const frames = total_samples / channels;
214 shared_ptr<AudioBuffers> audio (new AudioBuffers (channels, frames));
215 float** data = audio->data();
217 switch (audio_sample_format (stream)) {
218 case AV_SAMPLE_FMT_U8:
220 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
223 for (int i = 0; i < total_samples; ++i) {
224 data[channel][sample] = float(*p++) / (1 << 23);
227 if (channel == channels) {
235 case AV_SAMPLE_FMT_S16:
237 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
240 for (int i = 0; i < total_samples; ++i) {
241 data[channel][sample] = float(*p++) / (1 << 15);
244 if (channel == channels) {
252 case AV_SAMPLE_FMT_S16P:
254 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
255 for (int i = 0; i < channels; ++i) {
256 for (int j = 0; j < frames; ++j) {
257 data[i][j] = static_cast<float>(p[i][j]) / (1 << 15);
263 case AV_SAMPLE_FMT_S32:
265 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
268 for (int i = 0; i < total_samples; ++i) {
269 data[channel][sample] = static_cast<float>(*p++) / 2147483648;
272 if (channel == channels) {
280 case AV_SAMPLE_FMT_S32P:
282 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
283 for (int i = 0; i < channels; ++i) {
284 for (int j = 0; j < frames; ++j) {
285 data[i][j] = static_cast<float>(p[i][j]) / 2147483648;
291 case AV_SAMPLE_FMT_FLT:
293 float* p = reinterpret_cast<float*> (_frame->data[0]);
296 for (int i = 0; i < total_samples; ++i) {
297 data[channel][sample] = *p++;
300 if (channel == channels) {
308 case AV_SAMPLE_FMT_FLTP:
310 float** p = reinterpret_cast<float**> (_frame->data);
311 /* Sometimes there aren't as many channels in the _frame as in the stream */
312 for (int i = 0; i < _frame->channels; ++i) {
313 memcpy (data[i], p[i], frames * sizeof(float));
315 for (int i = _frame->channels; i < channels; ++i) {
316 audio->make_silent (i);
322 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
329 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
331 return stream->stream (_format_context)->codec->sample_fmt;
335 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
337 return av_get_bytes_per_sample (audio_sample_format (stream));
341 FFmpegDecoder::seek (ContentTime time, bool accurate)
343 Decoder::seek (time, accurate);
345 /* If we are doing an `accurate' seek, we need to use pre-roll, as
346 we don't really know what the seek will give us.
349 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
352 /* XXX: it seems debatable whether PTS should be used here...
353 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
356 optional<int> stream;
359 stream = _video_stream;
361 shared_ptr<FFmpegAudioStream> s = dynamic_pointer_cast<FFmpegAudioStream> (_ffmpeg_content->audio->stream ());
363 stream = s->index (_format_context);
367 DCPOMATIC_ASSERT (stream);
369 ContentTime u = time - _pts_offset;
370 if (u < ContentTime ()) {
376 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
381 /* Force re-creation of filter graphs to reset them and hence to make sure
382 they don't have any pre-seek frames knocking about.
384 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
385 _filter_graphs.clear ();
388 if (video_codec_context ()) {
389 avcodec_flush_buffers (video_codec_context());
392 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, ffmpeg_content()->ffmpeg_audio_streams()) {
393 avcodec_flush_buffers (i->stream(_format_context)->codec);
396 if (subtitle_codec_context ()) {
397 avcodec_flush_buffers (subtitle_codec_context ());
400 _have_current_subtitle = false;
404 FFmpegDecoder::decode_audio_packet ()
406 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
410 AVPacket copy_packet = _packet;
411 int const stream_index = copy_packet.stream_index;
413 /* XXX: inefficient */
414 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
415 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
416 while (stream != streams.end () && !(*stream)->uses_index (_format_context, stream_index)) {
420 if (stream == streams.end ()) {
421 /* The packet's stream may not be an audio one; just ignore it in this method if so */
425 while (copy_packet.size > 0) {
428 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
429 if (decode_result < 0) {
430 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
431 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
432 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
433 even in the face of such an error, so I think we should too.
435 Returning from the method here caused mantis #352.
437 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
439 /* Fudge decode_result so that we come out of the while loop when
440 we've processed this data.
442 decode_result = copy_packet.size;
445 if (frame_finished) {
446 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
449 if (_frame->pts == AV_NOPTS_VALUE && _next_time[stream_index]) {
450 /* In some streams we see not every frame coming through with a timestamp; for those
451 that have AV_NOPTS_VALUE we need to work out the timestamp ourselves. This is
452 particularly noticeable with TrueHD streams (see #1111).
454 ct = *_next_time[stream_index];
456 ct = ContentTime::from_seconds (
457 av_frame_get_best_effort_timestamp (_frame) *
458 av_q2d ((*stream)->stream (_format_context)->time_base))
462 _next_time[stream_index] = ct + ContentTime::from_frames(data->frames(), (*stream)->frame_rate());
464 if (ct < ContentTime ()) {
465 /* Discard audio data that comes before time 0 */
466 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
467 data->move (data->frames() - remove, remove, 0);
468 data->set_frames (data->frames() - remove);
469 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
472 if (ct < ContentTime()) {
474 "Crazy timestamp %1 for %2 samples in stream %3 packet pts %4 (ts=%5 tb=%6, off=%7)",
477 copy_packet.stream_index,
479 av_frame_get_best_effort_timestamp(_frame),
480 av_q2d((*stream)->stream(_format_context)->time_base),
481 to_string(_pts_offset)
485 /* Give this data provided there is some, and its time is sane */
486 if (ct >= ContentTime() && data->frames() > 0) {
487 audio->emit (film(), *stream, data, ct);
491 copy_packet.data += decode_result;
492 copy_packet.size -= decode_result;
497 FFmpegDecoder::decode_video_packet ()
499 DCPOMATIC_ASSERT (_video_stream);
502 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
506 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
508 shared_ptr<VideoFilterGraph> graph;
510 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
511 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
515 if (i == _filter_graphs.end ()) {
516 dcp::Fraction vfr (lrint(_ffmpeg_content->video_frame_rate().get() * 1000), 1000);
517 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format, vfr));
518 graph->setup (_ffmpeg_content->filters ());
519 _filter_graphs.push_back (graph);
520 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
525 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
527 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
529 shared_ptr<Image> image = i->first;
531 if (i->second != AV_NOPTS_VALUE) {
532 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
536 shared_ptr<ImageProxy> (new RawImageProxy (image)),
537 llrint(pts * _ffmpeg_content->active_video_frame_rate(film()))
540 LOG_WARNING_NC ("Dropping frame without PTS");
548 FFmpegDecoder::decode_subtitle_packet ()
552 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
556 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
557 if (_have_current_subtitle) {
558 if (_current_subtitle_to) {
559 only_text()->emit_stop (min(*_current_subtitle_to, subtitle_period(sub).from + _pts_offset));
561 only_text()->emit_stop (subtitle_period(sub).from + _pts_offset);
563 _have_current_subtitle = false;
566 if (sub.num_rects <= 0) {
567 /* Nothing new in this subtitle */
571 /* Subtitle PTS (within the source, not taking into account any of the
572 source that we may have chopped off for the DCP).
574 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
576 from = sub_period.from + _pts_offset;
578 _current_subtitle_to = *sub_period.to + _pts_offset;
580 _current_subtitle_to = optional<ContentTime>();
581 _have_current_subtitle = true;
584 for (unsigned int i = 0; i < sub.num_rects; ++i) {
585 AVSubtitleRect const * rect = sub.rects[i];
587 switch (rect->type) {
590 case SUBTITLE_BITMAP:
591 decode_bitmap_subtitle (rect, from);
594 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
597 decode_ass_subtitle (rect->ass, from);
602 if (_current_subtitle_to) {
603 only_text()->emit_stop (*_current_subtitle_to);
606 avsubtitle_free (&sub);
610 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTime from)
612 /* Note BGRA is expressed little-endian, so the first byte in the word is B, second
613 G, third R, fourth A.
615 shared_ptr<Image> image (new Image (AV_PIX_FMT_BGRA, dcp::Size (rect->w, rect->h), true));
617 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
618 /* Start of the first line in the subtitle */
619 uint8_t* sub_p = rect->pict.data[0];
620 /* sub_p looks up into a BGRA palette which is here
621 (i.e. first byte B, second G, third R, fourth A)
623 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
625 /* Start of the first line in the subtitle */
626 uint8_t* sub_p = rect->data[0];
627 /* sub_p looks up into a BGRA palette which is at rect->data[1].
628 (first byte B, second G, third R, fourth A)
631 /* And the stream has a map of those palette colours to colours
632 chosen by the user; created a `mapped' palette from those settings.
634 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
635 vector<RGBA> mapped_palette (rect->nb_colors);
636 uint8_t const * palette = rect->data[1];
637 for (int i = 0; i < rect->nb_colors; ++i) {
638 RGBA c (palette[2], palette[1], palette[0], palette[3]);
639 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
640 if (j != colour_map.end ()) {
641 mapped_palette[i] = j->second;
643 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
644 it is from a project that was created before this stuff was added. Just use the
645 colour straight from the original palette.
647 mapped_palette[i] = c;
652 /* Start of the output data */
653 uint8_t* out_p = image->data()[0];
655 for (int y = 0; y < rect->h; ++y) {
656 uint8_t* sub_line_p = sub_p;
657 uint8_t* out_line_p = out_p;
658 for (int x = 0; x < rect->w; ++x) {
659 RGBA const p = mapped_palette[*sub_line_p++];
665 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
666 sub_p += rect->pict.linesize[0];
668 sub_p += rect->linesize[0];
670 out_p += image->stride()[0];
673 int target_width = subtitle_codec_context()->width;
674 if (target_width == 0 && video_codec_context()) {
675 /* subtitle_codec_context()->width == 0 has been seen in the wild but I don't
676 know if it's supposed to mean something from FFmpeg's point of view.
678 target_width = video_codec_context()->width;
680 int target_height = subtitle_codec_context()->height;
681 if (target_height == 0 && video_codec_context()) {
682 target_height = video_codec_context()->height;
684 DCPOMATIC_ASSERT (target_width);
685 DCPOMATIC_ASSERT (target_height);
686 dcpomatic::Rect<double> const scaled_rect (
687 static_cast<double> (rect->x) / target_width,
688 static_cast<double> (rect->y) / target_height,
689 static_cast<double> (rect->w) / target_width,
690 static_cast<double> (rect->h) / target_height
693 only_text()->emit_bitmap_start (from, image, scaled_rect);
697 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTime from)
699 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
700 produces a single format of Dialogue: lines...
705 for (size_t i = 0; i < ass.length(); ++i) {
706 if (commas < 9 && ass[i] == ',') {
708 } else if (commas == 9) {
717 sub::RawSubtitle base;
718 list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (
721 _ffmpeg_content->video->size().width,
722 _ffmpeg_content->video->size().height
725 BOOST_FOREACH (sub::Subtitle const & i, sub::collect<list<sub::Subtitle> > (raw)) {
726 only_text()->emit_plain_start (from, i);