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 DCPOMATIC_ASSERT (_frame->channels <= channels);
318 /* Sometimes there aren't as many channels in the _frame as in the stream */
319 for (int i = 0; i < _frame->channels; ++i) {
320 memcpy (data[i], p[i], frames * sizeof(float));
322 for (int i = _frame->channels; i < channels; ++i) {
323 audio->make_silent (i);
329 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
336 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
338 return stream->stream (_format_context)->codec->sample_fmt;
342 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
344 return av_get_bytes_per_sample (audio_sample_format (stream));
348 FFmpegDecoder::seek (ContentTime time, bool accurate)
350 Decoder::seek (time, accurate);
352 /* If we are doing an `accurate' seek, we need to use pre-roll, as
353 we don't really know what the seek will give us.
356 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
359 /* XXX: it seems debatable whether PTS should be used here...
360 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
363 optional<int> stream;
366 stream = _video_stream;
368 DCPOMATIC_ASSERT (_ffmpeg_content->audio);
369 shared_ptr<FFmpegAudioStream> s = dynamic_pointer_cast<FFmpegAudioStream> (_ffmpeg_content->audio->stream ());
371 stream = s->index (_format_context);
375 DCPOMATIC_ASSERT (stream);
377 ContentTime u = time - _pts_offset;
378 if (u < ContentTime ()) {
384 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
389 /* Force re-creation of filter graphs to reset them and hence to make sure
390 they don't have any pre-seek frames knocking about.
392 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
393 _filter_graphs.clear ();
396 if (video_codec_context ()) {
397 avcodec_flush_buffers (video_codec_context());
400 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, ffmpeg_content()->ffmpeg_audio_streams()) {
401 avcodec_flush_buffers (i->stream(_format_context)->codec);
404 if (subtitle_codec_context ()) {
405 avcodec_flush_buffers (subtitle_codec_context ());
408 _have_current_subtitle = false;
412 FFmpegDecoder::decode_audio_packet ()
414 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
418 AVPacket copy_packet = _packet;
419 int const stream_index = copy_packet.stream_index;
421 /* XXX: inefficient */
422 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
423 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
424 while (stream != streams.end () && !(*stream)->uses_index (_format_context, stream_index)) {
428 if (stream == streams.end ()) {
429 /* The packet's stream may not be an audio one; just ignore it in this method if so */
433 while (copy_packet.size > 0) {
436 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
437 if (decode_result < 0) {
438 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
439 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
440 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
441 even in the face of such an error, so I think we should too.
443 Returning from the method here caused mantis #352.
445 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
447 /* Fudge decode_result so that we come out of the while loop when
448 we've processed this data.
450 decode_result = copy_packet.size;
453 if (frame_finished) {
454 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
457 if (_frame->pts == AV_NOPTS_VALUE && _next_time[stream_index]) {
458 /* In some streams we see not every frame coming through with a timestamp; for those
459 that have AV_NOPTS_VALUE we need to work out the timestamp ourselves. This is
460 particularly noticeable with TrueHD streams (see #1111).
462 ct = *_next_time[stream_index];
464 ct = ContentTime::from_seconds (
465 av_frame_get_best_effort_timestamp (_frame) *
466 av_q2d ((*stream)->stream (_format_context)->time_base))
470 _next_time[stream_index] = ct + ContentTime::from_frames(data->frames(), (*stream)->frame_rate());
472 if (ct < ContentTime ()) {
473 /* Discard audio data that comes before time 0 */
474 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
475 data->move (data->frames() - remove, remove, 0);
476 data->set_frames (data->frames() - remove);
477 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
480 if (ct < ContentTime()) {
482 "Crazy timestamp %1 for %2 samples in stream %3 packet pts %4 (ts=%5 tb=%6, off=%7)",
485 copy_packet.stream_index,
487 av_frame_get_best_effort_timestamp(_frame),
488 av_q2d((*stream)->stream(_format_context)->time_base),
489 to_string(_pts_offset)
493 /* Give this data provided there is some, and its time is sane */
494 if (ct >= ContentTime() && data->frames() > 0) {
495 audio->emit (film(), *stream, data, ct);
499 copy_packet.data += decode_result;
500 copy_packet.size -= decode_result;
505 FFmpegDecoder::decode_video_packet ()
507 DCPOMATIC_ASSERT (_video_stream);
510 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
514 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
516 shared_ptr<VideoFilterGraph> graph;
518 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
519 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
523 if (i == _filter_graphs.end ()) {
524 dcp::Fraction vfr (lrint(_ffmpeg_content->video_frame_rate().get() * 1000), 1000);
525 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format, vfr));
526 graph->setup (_ffmpeg_content->filters ());
527 _filter_graphs.push_back (graph);
528 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
533 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
535 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
537 shared_ptr<Image> image = i->first;
539 if (i->second != AV_NOPTS_VALUE) {
540 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
544 shared_ptr<ImageProxy> (new RawImageProxy (image)),
545 llrint(pts * _ffmpeg_content->active_video_frame_rate(film()))
548 LOG_WARNING_NC ("Dropping frame without PTS");
556 FFmpegDecoder::decode_subtitle_packet ()
560 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
564 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
565 if (_have_current_subtitle) {
566 if (_current_subtitle_to) {
567 only_text()->emit_stop (min(*_current_subtitle_to, subtitle_period(sub).from + _pts_offset));
569 only_text()->emit_stop (subtitle_period(sub).from + _pts_offset);
571 _have_current_subtitle = false;
574 if (sub.num_rects <= 0) {
575 /* Nothing new in this subtitle */
579 /* Subtitle PTS (within the source, not taking into account any of the
580 source that we may have chopped off for the DCP).
582 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
584 from = sub_period.from + _pts_offset;
586 _current_subtitle_to = *sub_period.to + _pts_offset;
588 _current_subtitle_to = optional<ContentTime>();
589 _have_current_subtitle = true;
592 for (unsigned int i = 0; i < sub.num_rects; ++i) {
593 AVSubtitleRect const * rect = sub.rects[i];
595 switch (rect->type) {
598 case SUBTITLE_BITMAP:
599 decode_bitmap_subtitle (rect, from);
602 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
605 decode_ass_subtitle (rect->ass, from);
610 if (_current_subtitle_to) {
611 only_text()->emit_stop (*_current_subtitle_to);
614 avsubtitle_free (&sub);
618 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTime from)
620 /* Note BGRA is expressed little-endian, so the first byte in the word is B, second
621 G, third R, fourth A.
623 shared_ptr<Image> image (new Image (AV_PIX_FMT_BGRA, dcp::Size (rect->w, rect->h), true));
625 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
626 /* Start of the first line in the subtitle */
627 uint8_t* sub_p = rect->pict.data[0];
628 /* sub_p looks up into a BGRA palette which is here
629 (i.e. first byte B, second G, third R, fourth A)
631 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
633 /* Start of the first line in the subtitle */
634 uint8_t* sub_p = rect->data[0];
635 /* sub_p looks up into a BGRA palette which is at rect->data[1].
636 (first byte B, second G, third R, fourth A)
639 /* And the stream has a map of those palette colours to colours
640 chosen by the user; created a `mapped' palette from those settings.
642 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
643 vector<RGBA> mapped_palette (rect->nb_colors);
644 uint8_t const * palette = rect->data[1];
645 for (int i = 0; i < rect->nb_colors; ++i) {
646 RGBA c (palette[2], palette[1], palette[0], palette[3]);
647 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
648 if (j != colour_map.end ()) {
649 mapped_palette[i] = j->second;
651 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
652 it is from a project that was created before this stuff was added. Just use the
653 colour straight from the original palette.
655 mapped_palette[i] = c;
660 /* Start of the output data */
661 uint8_t* out_p = image->data()[0];
663 for (int y = 0; y < rect->h; ++y) {
664 uint8_t* sub_line_p = sub_p;
665 uint8_t* out_line_p = out_p;
666 for (int x = 0; x < rect->w; ++x) {
667 RGBA const p = mapped_palette[*sub_line_p++];
673 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
674 sub_p += rect->pict.linesize[0];
676 sub_p += rect->linesize[0];
678 out_p += image->stride()[0];
681 int target_width = subtitle_codec_context()->width;
682 if (target_width == 0 && video_codec_context()) {
683 /* subtitle_codec_context()->width == 0 has been seen in the wild but I don't
684 know if it's supposed to mean something from FFmpeg's point of view.
686 target_width = video_codec_context()->width;
688 int target_height = subtitle_codec_context()->height;
689 if (target_height == 0 && video_codec_context()) {
690 target_height = video_codec_context()->height;
692 DCPOMATIC_ASSERT (target_width);
693 DCPOMATIC_ASSERT (target_height);
694 dcpomatic::Rect<double> const scaled_rect (
695 static_cast<double> (rect->x) / target_width,
696 static_cast<double> (rect->y) / target_height,
697 static_cast<double> (rect->w) / target_width,
698 static_cast<double> (rect->h) / target_height
701 only_text()->emit_bitmap_start (from, image, scaled_rect);
705 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTime from)
707 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
708 produces a single format of Dialogue: lines...
713 for (size_t i = 0; i < ass.length(); ++i) {
714 if (commas < 9 && ass[i] == ',') {
716 } else if (commas == 9) {
725 sub::RawSubtitle base;
726 list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (
729 _ffmpeg_content->video->size().width,
730 _ffmpeg_content->video->size().height
733 BOOST_FOREACH (sub::Subtitle const & i, sub::collect<list<sub::Subtitle> > (raw)) {
734 only_text()->emit_plain_start (from, i);