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 "frame_interval_checker.h"
46 #include <dcp/subtitle_string.h>
47 #include <sub/ssa_reader.h>
48 #include <sub/subtitle.h>
49 #include <sub/collect.h>
51 #include <libavcodec/avcodec.h>
52 #include <libavformat/avformat.h>
54 #include <boost/algorithm/string.hpp>
70 using std::shared_ptr;
71 using std::make_shared;
72 using boost::is_any_of;
74 using boost::optional;
75 using std::dynamic_pointer_cast;
77 using namespace dcpomatic;
80 FFmpegDecoder::FFmpegDecoder (shared_ptr<const Film> film, shared_ptr<const FFmpegContent> c, bool fast)
83 , _have_current_subtitle (false)
85 if (c->video && c->video->use()) {
86 video = make_shared<VideoDecoder>(this, c);
87 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate(film));
88 /* It doesn't matter what size or pixel format this is, it just needs to be black */
89 _black_image.reset (new Image (AV_PIX_FMT_RGB24, dcp::Size (128, 128), true));
90 _black_image->make_black ();
96 audio = make_shared<AudioDecoder>(this, c->audio, fast);
100 /* XXX: this time here should be the time of the first subtitle, not 0 */
101 text.push_back (make_shared<TextDecoder>(this, c->only_text(), ContentTime()));
104 _next_time.resize (_format_context->nb_streams);
109 FFmpegDecoder::flush ()
111 /* Get any remaining frames */
116 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
118 while (video && decode_video_packet()) {}
121 decode_audio_packet ();
124 /* Make sure all streams are the same length and round up to the next video frame */
126 auto const frc = film()->active_frame_rate_change(_ffmpeg_content->position());
127 ContentTime full_length (_ffmpeg_content->full_length(film()), frc);
128 full_length = full_length.ceil (frc.source);
130 double const vfr = _ffmpeg_content->video_frame_rate().get();
131 auto const f = full_length.frames_round (vfr);
132 auto v = video->position(film()).get_value_or(ContentTime()).frames_round(vfr) + 1;
134 video->emit (film(), shared_ptr<const ImageProxy> (new RawImageProxy (_black_image)), v);
139 for (auto i: _ffmpeg_content->ffmpeg_audio_streams ()) {
140 auto a = audio->stream_position(film(), i);
141 /* Unfortunately if a is 0 that really means that we don't know the stream position since
142 there has been no data on it since the last seek. In this case we'll just do nothing
143 here. I'm not sure if that's the right idea.
145 if (a > ContentTime()) {
146 while (a < full_length) {
147 auto to_do = min (full_length - a, ContentTime::from_seconds (0.1));
148 auto silence = make_shared<AudioBuffers>(i->channels(), to_do.frames_ceil (i->frame_rate()));
149 silence->make_silent ();
150 audio->emit (film(), i, silence, a, true);
163 FFmpegDecoder::pass ()
165 int r = av_read_frame (_format_context, &_packet);
167 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
168 has pretty-much succeeded (and hence generated data which should be processed).
169 Hence it makes sense to continue here in that case.
171 if (r < 0 && r != AVERROR_INVALIDDATA) {
172 if (r != AVERROR_EOF) {
173 /* Maybe we should fail here, but for now we'll just finish off instead */
175 av_strerror (r, buf, sizeof(buf));
176 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
183 int const si = _packet.stream_index;
184 auto fc = _ffmpeg_content;
186 if (_video_stream && si == _video_stream.get() && video && !video->ignore()) {
187 decode_video_packet ();
188 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index(_format_context, si) && !only_text()->ignore()) {
189 decode_subtitle_packet ();
191 decode_audio_packet ();
194 av_packet_unref (&_packet);
199 /** @param data pointer to array of pointers to buffers.
200 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
202 shared_ptr<AudioBuffers>
203 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
205 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
207 DCPOMATIC_DISABLE_WARNINGS
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
211 DCPOMATIC_ENABLE_WARNINGS
213 /* XXX: can't we just use _frame->nb_samples directly here? */
214 /* XXX: can't we use swr_convert() to do the format conversion? */
216 /* Deinterleave and convert to float */
218 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
219 of the block that do not form a complete sample or frame they will be dropped.
221 int const total_samples = size / bytes_per_audio_sample (stream);
222 int const channels = stream->channels();
223 int const frames = total_samples / channels;
224 auto audio = make_shared<AudioBuffers>(channels, frames);
225 auto data = audio->data();
227 switch (audio_sample_format (stream)) {
228 case AV_SAMPLE_FMT_U8:
230 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
233 for (int i = 0; i < total_samples; ++i) {
234 data[channel][sample] = float(*p++) / (1 << 23);
237 if (channel == channels) {
245 case AV_SAMPLE_FMT_S16:
247 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
250 for (int i = 0; i < total_samples; ++i) {
251 data[channel][sample] = float(*p++) / (1 << 15);
254 if (channel == channels) {
262 case AV_SAMPLE_FMT_S16P:
264 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
265 for (int i = 0; i < channels; ++i) {
266 for (int j = 0; j < frames; ++j) {
267 data[i][j] = static_cast<float>(p[i][j]) / (1 << 15);
273 case AV_SAMPLE_FMT_S32:
275 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
278 for (int i = 0; i < total_samples; ++i) {
279 data[channel][sample] = static_cast<float>(*p++) / 2147483648;
282 if (channel == channels) {
290 case AV_SAMPLE_FMT_S32P:
292 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
293 for (int i = 0; i < channels; ++i) {
294 for (int j = 0; j < frames; ++j) {
295 data[i][j] = static_cast<float>(p[i][j]) / 2147483648;
301 case AV_SAMPLE_FMT_FLT:
303 float* p = reinterpret_cast<float*> (_frame->data[0]);
306 for (int i = 0; i < total_samples; ++i) {
307 data[channel][sample] = *p++;
310 if (channel == channels) {
318 case AV_SAMPLE_FMT_FLTP:
320 float** p = reinterpret_cast<float**> (_frame->data);
321 DCPOMATIC_ASSERT (_frame->channels <= channels);
322 /* Sometimes there aren't as many channels in the _frame as in the stream */
323 for (int i = 0; i < _frame->channels; ++i) {
324 memcpy (data[i], p[i], frames * sizeof(float));
326 for (int i = _frame->channels; i < channels; ++i) {
327 audio->make_silent (i);
333 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
341 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
343 DCPOMATIC_DISABLE_WARNINGS
344 return stream->stream (_format_context)->codec->sample_fmt;
345 DCPOMATIC_ENABLE_WARNINGS
350 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
352 return av_get_bytes_per_sample (audio_sample_format (stream));
357 FFmpegDecoder::seek (ContentTime time, bool accurate)
359 Decoder::seek (time, accurate);
361 /* If we are doing an `accurate' seek, we need to use pre-roll, as
362 we don't really know what the seek will give us.
365 auto pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
368 /* XXX: it seems debatable whether PTS should be used here...
369 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
372 optional<int> stream;
375 stream = _video_stream;
377 DCPOMATIC_ASSERT (_ffmpeg_content->audio);
378 auto s = dynamic_pointer_cast<FFmpegAudioStream>(_ffmpeg_content->audio->stream());
380 stream = s->index (_format_context);
384 DCPOMATIC_ASSERT (stream);
386 auto u = time - _pts_offset;
387 if (u < ContentTime ()) {
393 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
398 /* Force re-creation of filter graphs to reset them and hence to make sure
399 they don't have any pre-seek frames knocking about.
401 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
402 _filter_graphs.clear ();
405 if (video_codec_context ()) {
406 avcodec_flush_buffers (video_codec_context());
409 DCPOMATIC_DISABLE_WARNINGS
410 for (auto i: ffmpeg_content()->ffmpeg_audio_streams()) {
411 avcodec_flush_buffers (i->stream(_format_context)->codec);
413 DCPOMATIC_ENABLE_WARNINGS
415 if (subtitle_codec_context ()) {
416 avcodec_flush_buffers (subtitle_codec_context ());
419 _have_current_subtitle = false;
421 for (auto& i: _next_time) {
422 i = optional<ContentTime>();
427 shared_ptr<FFmpegAudioStream>
428 FFmpegDecoder::audio_stream_from_index (int index) const
430 /* XXX: inefficient */
431 auto streams = ffmpeg_content()->ffmpeg_audio_streams();
432 auto stream = streams.begin();
433 while (stream != streams.end() && !(*stream)->uses_index(_format_context, index)) {
437 if (stream == streams.end ()) {
446 FFmpegDecoder::decode_audio_packet ()
448 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
452 AVPacket copy_packet = _packet;
453 int const stream_index = copy_packet.stream_index;
455 auto stream = audio_stream_from_index (stream_index);
457 /* The packet's stream may not be an audio one; just ignore it in this method if so */
461 DCPOMATIC_DISABLE_WARNINGS
462 while (copy_packet.size > 0) {
465 int decode_result = avcodec_decode_audio4 (stream->stream(_format_context)->codec, _frame, &frame_finished, ©_packet);
466 if (decode_result < 0) {
467 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
468 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
469 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
470 even in the face of such an error, so I think we should too.
472 Returning from the method here caused mantis #352.
474 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
476 /* Fudge decode_result so that we come out of the while loop when
477 we've processed this data.
479 decode_result = copy_packet.size;
482 if (frame_finished) {
483 auto data = deinterleave_audio (stream);
486 if (_frame->pts == AV_NOPTS_VALUE) {
487 /* In some streams we see not every frame coming through with a timestamp; for those
488 that have AV_NOPTS_VALUE we need to work out the timestamp ourselves. This is
489 particularly noticeable with TrueHD streams (see #1111).
491 if (_next_time[stream_index]) {
492 ct = *_next_time[stream_index];
495 ct = ContentTime::from_seconds (
496 av_frame_get_best_effort_timestamp (_frame) *
497 av_q2d (stream->stream(_format_context)->time_base))
501 _next_time[stream_index] = ct + ContentTime::from_frames(data->frames(), stream->frame_rate());
503 if (ct < ContentTime()) {
504 /* Discard audio data that comes before time 0 */
505 auto const remove = min (int64_t(data->frames()), (-ct).frames_ceil(double(stream->frame_rate())));
506 data->move (data->frames() - remove, remove, 0);
507 data->set_frames (data->frames() - remove);
508 ct += ContentTime::from_frames (remove, stream->frame_rate());
511 if (ct < ContentTime()) {
513 "Crazy timestamp %1 for %2 samples in stream %3 packet pts %4 (ts=%5 tb=%6, off=%7)",
516 copy_packet.stream_index,
518 av_frame_get_best_effort_timestamp(_frame),
519 av_q2d(stream->stream(_format_context)->time_base),
520 to_string(_pts_offset)
523 DCPOMATIC_ENABLE_WARNINGS
525 /* Give this data provided there is some, and its time is sane */
526 if (ct >= ContentTime() && data->frames() > 0) {
527 audio->emit (film(), stream, data, ct);
531 copy_packet.data += decode_result;
532 copy_packet.size -= decode_result;
538 FFmpegDecoder::decode_video_packet ()
540 DCPOMATIC_ASSERT (_video_stream);
543 DCPOMATIC_DISABLE_WARNINGS
544 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
547 DCPOMATIC_ENABLE_WARNINGS
549 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
551 shared_ptr<VideoFilterGraph> graph;
553 auto i = _filter_graphs.begin();
554 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
558 if (i == _filter_graphs.end ()) {
559 dcp::Fraction vfr (lrint(_ffmpeg_content->video_frame_rate().get() * 1000), 1000);
560 graph = make_shared<VideoFilterGraph>(dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format, vfr);
561 graph->setup (_ffmpeg_content->filters ());
562 _filter_graphs.push_back (graph);
563 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
568 auto images = graph->process (_frame);
570 for (auto const& i: images) {
572 auto image = i.first;
574 if (i.second != AV_NOPTS_VALUE) {
575 double const pts = i.second * av_q2d(_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds();
579 make_shared<RawImageProxy>(image),
580 llrint(pts * _ffmpeg_content->active_video_frame_rate(film()))
583 LOG_WARNING_NC ("Dropping frame without PTS");
592 FFmpegDecoder::decode_subtitle_packet ()
596 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
600 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
601 if (_have_current_subtitle) {
602 if (_current_subtitle_to) {
603 only_text()->emit_stop (min(*_current_subtitle_to, subtitle_period(sub).from + _pts_offset));
605 only_text()->emit_stop (subtitle_period(sub).from + _pts_offset);
607 _have_current_subtitle = false;
610 if (sub.num_rects <= 0) {
611 /* Nothing new in this subtitle */
615 /* Subtitle PTS (within the source, not taking into account any of the
616 source that we may have chopped off for the DCP).
618 auto sub_period = subtitle_period (sub);
620 from = sub_period.from + _pts_offset;
622 _current_subtitle_to = *sub_period.to + _pts_offset;
624 _current_subtitle_to = optional<ContentTime>();
625 _have_current_subtitle = true;
628 for (unsigned int i = 0; i < sub.num_rects; ++i) {
629 auto const rect = sub.rects[i];
631 switch (rect->type) {
634 case SUBTITLE_BITMAP:
635 decode_bitmap_subtitle (rect, from);
638 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
641 decode_ass_subtitle (rect->ass, from);
646 if (_current_subtitle_to) {
647 only_text()->emit_stop (*_current_subtitle_to);
650 avsubtitle_free (&sub);
655 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTime from)
657 /* Note BGRA is expressed little-endian, so the first byte in the word is B, second
658 G, third R, fourth A.
660 auto image = make_shared<Image>(AV_PIX_FMT_BGRA, dcp::Size (rect->w, rect->h), true);
662 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
663 /* Start of the first line in the subtitle */
664 auto sub_p = rect->pict.data[0];
665 /* sub_p looks up into a BGRA palette which is at rect->pict.data[1];
666 (i.e. first byte B, second G, third R, fourth A)
668 auto const palette = rect->pict.data[1];
670 /* Start of the first line in the subtitle */
671 auto sub_p = rect->data[0];
672 /* sub_p looks up into a BGRA palette which is at rect->data[1].
673 (first byte B, second G, third R, fourth A)
675 auto const* palette = rect->data[1];
677 /* And the stream has a map of those palette colours to colours
678 chosen by the user; created a `mapped' palette from those settings.
680 auto colour_map = ffmpeg_content()->subtitle_stream()->colours();
681 vector<RGBA> mapped_palette (rect->nb_colors);
682 for (int i = 0; i < rect->nb_colors; ++i) {
683 RGBA c (palette[2], palette[1], palette[0], palette[3]);
684 auto j = colour_map.find (c);
685 if (j != colour_map.end ()) {
686 mapped_palette[i] = j->second;
688 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
689 it is from a project that was created before this stuff was added. Just use the
690 colour straight from the original palette.
692 mapped_palette[i] = c;
697 /* Start of the output data */
698 auto out_p = image->data()[0];
700 for (int y = 0; y < rect->h; ++y) {
701 auto sub_line_p = sub_p;
702 auto out_line_p = out_p;
703 for (int x = 0; x < rect->w; ++x) {
704 auto const p = mapped_palette[*sub_line_p++];
710 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
711 sub_p += rect->pict.linesize[0];
713 sub_p += rect->linesize[0];
715 out_p += image->stride()[0];
718 int target_width = subtitle_codec_context()->width;
719 if (target_width == 0 && video_codec_context()) {
720 /* subtitle_codec_context()->width == 0 has been seen in the wild but I don't
721 know if it's supposed to mean something from FFmpeg's point of view.
723 target_width = video_codec_context()->width;
725 int target_height = subtitle_codec_context()->height;
726 if (target_height == 0 && video_codec_context()) {
727 target_height = video_codec_context()->height;
729 DCPOMATIC_ASSERT (target_width);
730 DCPOMATIC_ASSERT (target_height);
731 dcpomatic::Rect<double> const scaled_rect (
732 static_cast<double>(rect->x) / target_width,
733 static_cast<double>(rect->y) / target_height,
734 static_cast<double>(rect->w) / target_width,
735 static_cast<double>(rect->h) / target_height
738 only_text()->emit_bitmap_start (from, image, scaled_rect);
743 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTime from)
745 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
746 produces a single format of Dialogue: lines...
751 for (size_t i = 0; i < ass.length(); ++i) {
752 if (commas < 9 && ass[i] == ',') {
754 } else if (commas == 9) {
763 sub::RawSubtitle base;
764 auto raw = sub::SSAReader::parse_line (
767 _ffmpeg_content->video->size().width,
768 _ffmpeg_content->video->size().height
771 for (auto const& i: sub::collect<vector<sub::Subtitle>>(raw)) {
772 only_text()->emit_plain_start (from, i);