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/>.
22 /** @file src/ffmpeg_decoder.cc
23 * @brief A decoder using FFmpeg to decode content.
27 #include "audio_buffers.h"
28 #include "audio_content.h"
29 #include "audio_decoder.h"
30 #include "compose.hpp"
31 #include "dcpomatic_log.h"
32 #include "exceptions.h"
33 #include "ffmpeg_audio_stream.h"
34 #include "ffmpeg_content.h"
35 #include "ffmpeg_decoder.h"
36 #include "ffmpeg_subtitle_stream.h"
39 #include "frame_interval_checker.h"
42 #include "raw_image_proxy.h"
43 #include "text_content.h"
44 #include "text_decoder.h"
46 #include "video_decoder.h"
47 #include "video_filter_graph.h"
48 #include <dcp/subtitle_string.h>
49 #include <sub/ssa_reader.h>
50 #include <sub/subtitle.h>
51 #include <sub/collect.h>
53 #include <libavcodec/avcodec.h>
54 #include <libavformat/avformat.h>
56 #include <boost/algorithm/string.hpp>
66 using std::dynamic_pointer_cast;
67 using std::make_shared;
69 using std::shared_ptr;
72 using boost::optional;
74 using namespace dcpomatic;
77 FFmpegDecoder::FFmpegDecoder (shared_ptr<const Film> film, shared_ptr<const FFmpegContent> c, bool fast)
81 if (c->video && c->video->use()) {
82 video = make_shared<VideoDecoder>(this, c);
83 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate(film));
84 /* It doesn't matter what size or pixel format this is, it just needs to be black */
85 _black_image = make_shared<Image>(AV_PIX_FMT_RGB24, dcp::Size (128, 128), Image::Alignment::PADDED);
86 _black_image->make_black ();
92 audio = make_shared<AudioDecoder>(this, c->audio, fast);
96 /* XXX: this time here should be the time of the first subtitle, not 0 */
97 text.push_back (make_shared<TextDecoder>(this, c->only_text(), ContentTime()));
100 for (auto i: c->ffmpeg_audio_streams()) {
101 _next_time[i] = boost::optional<dcpomatic::ContentTime>();
107 FFmpegDecoder::flush ()
109 /* Flush video and audio once */
111 bool did_something = false;
113 if (decode_and_process_video_packet(nullptr)) {
114 did_something = true;
118 for (auto i: ffmpeg_content()->ffmpeg_audio_streams()) {
119 auto context = _codec_context[i->index(_format_context)];
120 int r = avcodec_send_packet (context, nullptr);
121 if (r < 0 && r != AVERROR_EOF) {
122 /* EOF can happen if we've already sent a flush packet */
123 throw DecodeError (N_("avcodec_send_packet"), N_("FFmpegDecoder::flush"), r);
125 r = avcodec_receive_frame (context, audio_frame(i));
127 process_audio_frame (i);
128 did_something = true;
133 /* We want to be called again */
137 /* Make sure all streams are the same length and round up to the next video frame */
139 auto const frc = film()->active_frame_rate_change(_ffmpeg_content->position());
140 ContentTime full_length (_ffmpeg_content->full_length(film()), frc);
141 full_length = full_length.ceil (frc.source);
143 double const vfr = _ffmpeg_content->video_frame_rate().get();
144 auto const f = full_length.frames_round (vfr);
145 auto v = video->position(film()).get_value_or(ContentTime()).frames_round(vfr) + 1;
147 video->emit (film(), shared_ptr<const ImageProxy> (new RawImageProxy (_black_image)), v);
152 for (auto i: _ffmpeg_content->ffmpeg_audio_streams ()) {
153 auto a = audio->stream_position(film(), i);
154 /* Unfortunately if a is 0 that really means that we don't know the stream position since
155 there has been no data on it since the last seek. In this case we'll just do nothing
156 here. I'm not sure if that's the right idea.
158 if (a > ContentTime()) {
159 while (a < full_length) {
160 auto to_do = min (full_length - a, ContentTime::from_seconds (0.1));
161 auto silence = make_shared<AudioBuffers>(i->channels(), to_do.frames_ceil (i->frame_rate()));
162 silence->make_silent ();
163 audio->emit (film(), i, silence, a, true);
178 FFmpegDecoder::pass ()
180 auto packet = av_packet_alloc();
181 DCPOMATIC_ASSERT (packet);
183 int r = av_read_frame (_format_context, packet);
185 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
186 has pretty-much succeeded (and hence generated data which should be processed).
187 Hence it makes sense to continue here in that case.
189 if (r < 0 && r != AVERROR_INVALIDDATA) {
190 if (r != AVERROR_EOF) {
191 /* Maybe we should fail here, but for now we'll just finish off instead */
193 av_strerror (r, buf, sizeof(buf));
194 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
197 av_packet_free (&packet);
201 int const si = packet->stream_index;
202 auto fc = _ffmpeg_content;
204 if (_video_stream && si == _video_stream.get() && video && !video->ignore()) {
205 decode_and_process_video_packet (packet);
206 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index(_format_context, si) && !only_text()->ignore()) {
207 decode_and_process_subtitle_packet (packet);
209 decode_and_process_audio_packet (packet);
212 av_packet_free (&packet);
217 /** @param data pointer to array of pointers to buffers.
218 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
220 shared_ptr<AudioBuffers>
221 FFmpegDecoder::deinterleave_audio (AVFrame* frame)
223 auto format = static_cast<AVSampleFormat>(frame->format);
225 /* XXX: can't we use swr_convert() to do the format conversion? */
227 int const channels = frame->channels;
228 int const frames = frame->nb_samples;
229 int const total_samples = frames * channels;
230 auto audio = make_shared<AudioBuffers>(channels, frames);
231 auto data = audio->data();
234 case AV_SAMPLE_FMT_U8:
236 auto p = reinterpret_cast<uint8_t *> (frame->data[0]);
239 for (int i = 0; i < total_samples; ++i) {
240 data[channel][sample] = float(*p++) / (1 << 23);
243 if (channel == channels) {
251 case AV_SAMPLE_FMT_S16:
253 auto p = reinterpret_cast<int16_t *> (frame->data[0]);
256 for (int i = 0; i < total_samples; ++i) {
257 data[channel][sample] = float(*p++) / (1 << 15);
260 if (channel == channels) {
268 case AV_SAMPLE_FMT_S16P:
270 auto p = reinterpret_cast<int16_t **> (frame->data);
271 for (int i = 0; i < channels; ++i) {
272 for (int j = 0; j < frames; ++j) {
273 data[i][j] = static_cast<float>(p[i][j]) / (1 << 15);
279 case AV_SAMPLE_FMT_S32:
281 auto p = reinterpret_cast<int32_t *> (frame->data[0]);
284 for (int i = 0; i < total_samples; ++i) {
285 data[channel][sample] = static_cast<float>(*p++) / 2147483648;
288 if (channel == channels) {
296 case AV_SAMPLE_FMT_S32P:
298 auto p = reinterpret_cast<int32_t **> (frame->data);
299 for (int i = 0; i < channels; ++i) {
300 for (int j = 0; j < frames; ++j) {
301 data[i][j] = static_cast<float>(p[i][j]) / 2147483648;
307 case AV_SAMPLE_FMT_FLT:
309 auto p = reinterpret_cast<float*> (frame->data[0]);
312 for (int i = 0; i < total_samples; ++i) {
313 data[channel][sample] = *p++;
316 if (channel == channels) {
324 case AV_SAMPLE_FMT_FLTP:
326 auto p = reinterpret_cast<float**> (frame->data);
327 DCPOMATIC_ASSERT (frame->channels <= channels);
328 /* Sometimes there aren't as many channels in the frame as in the stream */
329 for (int i = 0; i < frame->channels; ++i) {
330 memcpy (data[i], p[i], frames * sizeof(float));
332 for (int i = frame->channels; i < channels; ++i) {
333 audio->make_silent (i);
339 throw DecodeError (String::compose(_("Unrecognised audio sample format (%1)"), static_cast<int>(format)));
347 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
349 return static_cast<AVSampleFormat>(stream->stream(_format_context)->codecpar->format);
354 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
356 return av_get_bytes_per_sample (audio_sample_format (stream));
361 FFmpegDecoder::seek (ContentTime time, bool accurate)
363 Decoder::seek (time, accurate);
365 /* If we are doing an `accurate' seek, we need to use pre-roll, as
366 we don't really know what the seek will give us.
369 auto pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
372 /* XXX: it seems debatable whether PTS should be used here...
373 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
376 optional<int> stream;
379 stream = _video_stream;
381 DCPOMATIC_ASSERT (_ffmpeg_content->audio);
382 auto s = dynamic_pointer_cast<FFmpegAudioStream>(_ffmpeg_content->audio->stream());
384 stream = s->index (_format_context);
388 DCPOMATIC_ASSERT (stream);
390 auto u = time - _pts_offset;
391 if (u < ContentTime ()) {
397 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
402 /* Force re-creation of filter graphs to reset them and hence to make sure
403 they don't have any pre-seek frames knocking about.
405 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
406 _filter_graphs.clear ();
409 if (video_codec_context ()) {
410 avcodec_flush_buffers (video_codec_context());
413 for (auto i: ffmpeg_content()->ffmpeg_audio_streams()) {
414 avcodec_flush_buffers (_codec_context[i->index(_format_context)]);
417 if (subtitle_codec_context ()) {
418 avcodec_flush_buffers (subtitle_codec_context ());
421 _have_current_subtitle = false;
423 for (auto& i: _next_time) {
424 i.second = boost::optional<dcpomatic::ContentTime>();
429 shared_ptr<FFmpegAudioStream>
430 FFmpegDecoder::audio_stream_from_index (int index) const
432 /* XXX: inefficient */
433 auto streams = ffmpeg_content()->ffmpeg_audio_streams();
434 auto stream = streams.begin();
435 while (stream != streams.end() && !(*stream)->uses_index(_format_context, index)) {
439 if (stream == streams.end ()) {
448 FFmpegDecoder::process_audio_frame (shared_ptr<FFmpegAudioStream> stream)
450 auto frame = audio_frame (stream);
451 auto data = deinterleave_audio (frame);
454 if (frame->pts == AV_NOPTS_VALUE) {
455 /* In some streams we see not every frame coming through with a timestamp; for those
456 that have AV_NOPTS_VALUE we need to work out the timestamp ourselves. This is
457 particularly noticeable with TrueHD streams (see #1111).
459 if (_next_time[stream]) {
460 ct = *_next_time[stream];
463 ct = ContentTime::from_seconds (
464 frame->best_effort_timestamp *
465 av_q2d (stream->stream(_format_context)->time_base))
469 _next_time[stream] = ct + ContentTime::from_frames(data->frames(), stream->frame_rate());
471 if (ct < ContentTime()) {
472 /* Discard audio data that comes before time 0 */
473 auto 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 (ts=%4 tb=%5, off=%6)",
485 frame->best_effort_timestamp,
486 av_q2d(stream->stream(_format_context)->time_base),
487 to_string(_pts_offset)
491 /* Give this data provided there is some, and its time is sane */
492 if (ct >= ContentTime() && data->frames() > 0) {
493 audio->emit (film(), stream, data, ct);
499 FFmpegDecoder::decode_and_process_audio_packet (AVPacket* packet)
501 auto stream = audio_stream_from_index (packet->stream_index);
506 auto context = _codec_context[stream->index(_format_context)];
507 auto frame = audio_frame (stream);
509 int r = avcodec_send_packet (context, packet);
511 LOG_WARNING("avcodec_send_packet returned %1 for an audio packet", r);
514 r = avcodec_receive_frame (context, frame);
515 if (r == AVERROR(EAGAIN)) {
516 /* More input is required */
520 /* We choose to be relaxed here about other errors; it seems that there may be valid
521 * data to decode even if an error occurred. #352 may be related (though this was
522 * when we were using an old version of the FFmpeg API).
524 process_audio_frame (stream);
530 FFmpegDecoder::decode_and_process_video_packet (AVPacket* packet)
532 DCPOMATIC_ASSERT (_video_stream);
534 auto context = video_codec_context();
536 int r = avcodec_send_packet (context, packet);
538 LOG_WARNING("avcodec_send_packet returned %1 for a video packet", r);
541 r = avcodec_receive_frame (context, _video_frame);
542 if (r == AVERROR(EAGAIN) || r == AVERROR_EOF || (r < 0 && !packet)) {
543 /* More input is required, no more frames are coming, or we are flushing and there was
544 * some error which we just want to ignore.
548 throw DecodeError (N_("avcodec_receive_frame"), N_("FFmpeg::decode_and_process_video_packet"), r);
551 /* We assume we'll only get one frame here, which I think is safe */
553 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
555 shared_ptr<VideoFilterGraph> graph;
557 auto i = _filter_graphs.begin();
558 while (i != _filter_graphs.end() && !(*i)->can_process(dcp::Size(_video_frame->width, _video_frame->height), (AVPixelFormat) _video_frame->format)) {
562 if (i == _filter_graphs.end ()) {
563 dcp::Fraction vfr (lrint(_ffmpeg_content->video_frame_rate().get() * 1000), 1000);
564 graph = make_shared<VideoFilterGraph>(dcp::Size(_video_frame->width, _video_frame->height), (AVPixelFormat) _video_frame->format, vfr);
565 graph->setup (_ffmpeg_content->filters ());
566 _filter_graphs.push_back (graph);
567 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _video_frame->width, _video_frame->height, _video_frame->format);
572 auto images = graph->process (_video_frame);
574 for (auto const& i: images) {
576 auto image = i.first;
578 if (i.second != AV_NOPTS_VALUE) {
579 double const pts = i.second * av_q2d(_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds();
583 make_shared<RawImageProxy>(image),
584 llrint(pts * _ffmpeg_content->active_video_frame_rate(film()))
587 LOG_WARNING_NC ("Dropping frame without PTS");
596 FFmpegDecoder::decode_and_process_subtitle_packet (AVPacket* packet)
600 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, packet) < 0 || !got_subtitle) {
604 auto sub_period = subtitle_period (packet, ffmpeg_content()->subtitle_stream()->stream(_format_context), sub);
606 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
607 if (_have_current_subtitle) {
608 if (_current_subtitle_to) {
609 only_text()->emit_stop (min(*_current_subtitle_to, sub_period.from + _pts_offset));
611 only_text()->emit_stop (sub_period.from + _pts_offset);
613 _have_current_subtitle = false;
616 if (sub.num_rects <= 0) {
617 /* Nothing new in this subtitle */
618 avsubtitle_free (&sub);
622 /* Subtitle PTS (within the source, not taking into account any of the
623 source that we may have chopped off for the DCP).
626 from = sub_period.from + _pts_offset;
628 _current_subtitle_to = *sub_period.to + _pts_offset;
630 _current_subtitle_to = optional<ContentTime>();
631 _have_current_subtitle = true;
634 for (unsigned int i = 0; i < sub.num_rects; ++i) {
635 auto const rect = sub.rects[i];
637 switch (rect->type) {
640 case SUBTITLE_BITMAP:
641 process_bitmap_subtitle (rect, from);
644 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
647 process_ass_subtitle (rect->ass, from);
652 if (_current_subtitle_to) {
653 only_text()->emit_stop (*_current_subtitle_to);
656 avsubtitle_free (&sub);
661 FFmpegDecoder::process_bitmap_subtitle (AVSubtitleRect const * rect, ContentTime from)
663 /* Note BGRA is expressed little-endian, so the first byte in the word is B, second
664 G, third R, fourth A.
666 auto image = make_shared<Image>(AV_PIX_FMT_BGRA, dcp::Size (rect->w, rect->h), Image::Alignment::PADDED);
668 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
669 /* Start of the first line in the subtitle */
670 auto sub_p = rect->pict.data[0];
671 /* sub_p looks up into a BGRA palette which is at rect->pict.data[1];
672 (i.e. first byte B, second G, third R, fourth A)
674 auto const palette = rect->pict.data[1];
676 /* Start of the first line in the subtitle */
677 auto sub_p = rect->data[0];
678 /* sub_p looks up into a BGRA palette which is at rect->data[1].
679 (first byte B, second G, third R, fourth A)
681 auto const* palette = rect->data[1];
683 /* And the stream has a map of those palette colours to colours
684 chosen by the user; created a `mapped' palette from those settings.
686 auto colour_map = ffmpeg_content()->subtitle_stream()->colours();
687 vector<RGBA> mapped_palette (rect->nb_colors);
688 for (int i = 0; i < rect->nb_colors; ++i) {
689 RGBA c (palette[2], palette[1], palette[0], palette[3]);
690 auto j = colour_map.find (c);
691 if (j != colour_map.end ()) {
692 mapped_palette[i] = j->second;
694 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
695 it is from a project that was created before this stuff was added. Just use the
696 colour straight from the original palette.
698 mapped_palette[i] = c;
703 /* Start of the output data */
704 auto out_p = image->data()[0];
706 for (int y = 0; y < rect->h; ++y) {
707 auto sub_line_p = sub_p;
708 auto out_line_p = out_p;
709 for (int x = 0; x < rect->w; ++x) {
710 auto const p = mapped_palette[*sub_line_p++];
716 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
717 sub_p += rect->pict.linesize[0];
719 sub_p += rect->linesize[0];
721 out_p += image->stride()[0];
724 int target_width = subtitle_codec_context()->width;
725 if (target_width == 0 && video_codec_context()) {
726 /* subtitle_codec_context()->width == 0 has been seen in the wild but I don't
727 know if it's supposed to mean something from FFmpeg's point of view.
729 target_width = video_codec_context()->width;
731 int target_height = subtitle_codec_context()->height;
732 if (target_height == 0 && video_codec_context()) {
733 target_height = video_codec_context()->height;
735 DCPOMATIC_ASSERT (target_width);
736 DCPOMATIC_ASSERT (target_height);
737 dcpomatic::Rect<double> const scaled_rect (
738 static_cast<double>(rect->x) / target_width,
739 static_cast<double>(rect->y) / target_height,
740 static_cast<double>(rect->w) / target_width,
741 static_cast<double>(rect->h) / target_height
744 only_text()->emit_bitmap_start (from, image, scaled_rect);
749 FFmpegDecoder::process_ass_subtitle (string ass, ContentTime from)
751 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
752 produces a single format of Dialogue: lines...
757 for (size_t i = 0; i < ass.length(); ++i) {
758 if (commas < 9 && ass[i] == ',') {
760 } else if (commas == 9) {
769 sub::RawSubtitle base;
770 auto raw = sub::SSAReader::parse_line (
773 _ffmpeg_content->video->size().width,
774 _ffmpeg_content->video->size().height
777 for (auto const& i: sub::collect<vector<sub::Subtitle>>(raw)) {
778 only_text()->emit_plain_start (from, i);