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)
80 , _filter_graphs(c->filters(), dcp::Fraction(lrint(_ffmpeg_content->video_frame_rate().get_value_or(24) * 1000), 1000))
82 if (c->video && c->video->use()) {
83 video = make_shared<VideoDecoder>(this, c);
84 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate(film));
85 /* It doesn't matter what size or pixel format this is, it just needs to be black */
86 _black_image = make_shared<Image>(AV_PIX_FMT_RGB24, dcp::Size (128, 128), Image::Alignment::PADDED);
87 _black_image->make_black ();
93 audio = make_shared<AudioDecoder>(this, c->audio, fast);
97 text.push_back (make_shared<TextDecoder>(this, c->only_text()));
98 /* XXX: we should be calling maybe_set_position() on this TextDecoder, but we can't easily find
99 * the time of the first subtitle at this point.
103 for (auto i: c->ffmpeg_audio_streams()) {
104 _next_time[i] = boost::optional<dcpomatic::ContentTime>();
109 FFmpegDecoder::FlushResult
110 FFmpegDecoder::flush ()
112 LOG_DEBUG_PLAYER("Flush FFmpeg decoder: current state %1", static_cast<int>(_flush_state));
114 switch (_flush_state) {
115 case FlushState::CODECS:
116 if (flush_codecs() == FlushResult::DONE) {
117 LOG_DEBUG_PLAYER_NC("Finished flushing codecs");
118 _flush_state = FlushState::AUDIO_DECODER;
121 case FlushState::AUDIO_DECODER:
125 LOG_DEBUG_PLAYER_NC("Finished flushing audio decoder");
126 _flush_state = FlushState::FILL;
128 case FlushState::FILL:
129 if (flush_fill() == FlushResult::DONE) {
130 LOG_DEBUG_PLAYER_NC("Finished flushing fills");
131 return FlushResult::DONE;
136 return FlushResult::AGAIN;
140 /** @return true if we have finished flushing the codecs */
141 FFmpegDecoder::FlushResult
142 FFmpegDecoder::flush_codecs()
144 bool did_something = false;
146 if (decode_and_process_video_packet(nullptr)) {
147 did_something = true;
151 for (auto i: ffmpeg_content()->ffmpeg_audio_streams()) {
152 auto context = _codec_context[i->index(_format_context)];
153 int r = avcodec_send_packet (context, nullptr);
154 if (r < 0 && r != AVERROR_EOF) {
155 /* EOF can happen if we've already sent a flush packet */
156 throw DecodeError (N_("avcodec_send_packet"), N_("FFmpegDecoder::flush"), r);
158 r = avcodec_receive_frame (context, audio_frame(i));
160 process_audio_frame (i);
161 did_something = true;
165 return did_something ? FlushResult::AGAIN : FlushResult::DONE;
169 FFmpegDecoder::FlushResult
170 FFmpegDecoder::flush_fill()
172 /* Make sure all streams are the same length and round up to the next video frame */
174 bool did_something = false;
176 auto const frc = film()->active_frame_rate_change(_ffmpeg_content->position());
177 ContentTime full_length (_ffmpeg_content->full_length(film()), frc);
178 full_length = full_length.ceil (frc.source);
179 if (video && !video->ignore()) {
180 double const vfr = _ffmpeg_content->video_frame_rate().get();
181 auto const f = full_length.frames_round (vfr);
182 auto const v = video->position(film()).get_value_or(ContentTime()).frames_round(vfr) + 1;
184 video->emit(film(), make_shared<const RawImageProxy>(_black_image), v);
185 did_something = true;
189 if (audio && !audio->ignore()) {
190 for (auto i: _ffmpeg_content->ffmpeg_audio_streams ()) {
191 auto const a = audio->stream_position(film(), i);
192 /* Unfortunately if a is 0 that really means that we don't know the stream position since
193 there has been no data on it since the last seek. In this case we'll just do nothing
194 here. I'm not sure if that's the right idea.
196 if (a > ContentTime() && a < full_length) {
197 LOG_DEBUG_PLAYER("Flush inserts silence at %1", to_string(a));
198 auto to_do = min (full_length - a, ContentTime::from_seconds (0.1));
199 auto silence = make_shared<AudioBuffers>(i->channels(), to_do.frames_ceil (i->frame_rate()));
200 silence->make_silent ();
201 audio->emit (film(), i, silence, a, true);
202 did_something = true;
207 return did_something ? FlushResult::AGAIN : FlushResult::DONE;
212 FFmpegDecoder::pass ()
214 auto packet = av_packet_alloc();
215 DCPOMATIC_ASSERT (packet);
217 int r = av_read_frame (_format_context, packet);
219 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
220 has pretty-much succeeded (and hence generated data which should be processed).
221 Hence it makes sense to continue here in that case.
223 if (r < 0 && r != AVERROR_INVALIDDATA) {
224 LOG_DEBUG_PLAYER("FFpmegDecoder::pass flushes because av_read_frame returned %1", r);
225 if (r != AVERROR_EOF) {
226 /* Maybe we should fail here, but for now we'll just finish off instead */
228 av_strerror (r, buf, sizeof(buf));
229 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
232 av_packet_free (&packet);
233 return flush() == FlushResult::DONE;
236 int const si = packet->stream_index;
237 auto fc = _ffmpeg_content;
239 if (_video_stream && si == _video_stream.get() && video && !video->ignore()) {
240 decode_and_process_video_packet (packet);
241 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index(_format_context, si) && !only_text()->ignore()) {
242 decode_and_process_subtitle_packet (packet);
244 decode_and_process_audio_packet (packet);
247 av_packet_free (&packet);
252 /** @param data pointer to array of pointers to buffers.
253 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
256 shared_ptr<AudioBuffers>
257 deinterleave_audio(AVFrame* frame)
259 auto format = static_cast<AVSampleFormat>(frame->format);
261 /* XXX: can't we use swr_convert() to do the format conversion? */
263 int const channels = frame->ch_layout.nb_channels;
264 int const frames = frame->nb_samples;
265 int const total_samples = frames * channels;
266 auto audio = make_shared<AudioBuffers>(channels, frames);
267 auto data = audio->data();
274 case AV_SAMPLE_FMT_U8:
276 auto p = reinterpret_cast<uint8_t *> (frame->data[0]);
279 for (int i = 0; i < total_samples; ++i) {
280 data[channel][sample] = float(*p++) / (1 << 23);
283 if (channel == channels) {
291 case AV_SAMPLE_FMT_S16:
293 auto p = reinterpret_cast<int16_t *> (frame->data[0]);
296 for (int i = 0; i < total_samples; ++i) {
297 data[channel][sample] = float(*p++) / (1 << 15);
300 if (channel == channels) {
308 case AV_SAMPLE_FMT_S16P:
310 auto p = reinterpret_cast<int16_t **> (frame->data);
311 for (int i = 0; i < channels; ++i) {
312 for (int j = 0; j < frames; ++j) {
313 data[i][j] = static_cast<float>(p[i][j]) / (1 << 15);
319 case AV_SAMPLE_FMT_S32:
321 auto p = reinterpret_cast<int32_t *> (frame->data[0]);
324 for (int i = 0; i < total_samples; ++i) {
325 data[channel][sample] = static_cast<float>(*p++) / 2147483648;
328 if (channel == channels) {
336 case AV_SAMPLE_FMT_S32P:
338 auto p = reinterpret_cast<int32_t **> (frame->data);
339 for (int i = 0; i < channels; ++i) {
340 for (int j = 0; j < frames; ++j) {
341 data[i][j] = static_cast<float>(p[i][j]) / 2147483648;
347 case AV_SAMPLE_FMT_FLT:
349 auto p = reinterpret_cast<float*> (frame->data[0]);
352 for (int i = 0; i < total_samples; ++i) {
353 data[channel][sample] = *p++;
356 if (channel == channels) {
364 case AV_SAMPLE_FMT_FLTP:
366 auto p = reinterpret_cast<float**> (frame->data);
367 for (int i = 0; i < channels; ++i) {
368 memcpy (data[i], p[i], frames * sizeof(float));
374 throw DecodeError (String::compose(_("Unrecognised audio sample format (%1)"), static_cast<int>(format)));
382 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
384 return static_cast<AVSampleFormat>(stream->stream(_format_context)->codecpar->format);
389 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
391 return av_get_bytes_per_sample (audio_sample_format (stream));
396 FFmpegDecoder::seek (ContentTime time, bool accurate)
398 Decoder::seek (time, accurate);
400 /* If we are doing an `accurate' seek, we need to use pre-roll, as
401 we don't really know what the seek will give us.
404 auto pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
407 /* XXX: it seems debatable whether PTS should be used here...
408 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
411 optional<int> stream;
414 stream = _video_stream;
416 DCPOMATIC_ASSERT (_ffmpeg_content->audio);
417 auto s = dynamic_pointer_cast<FFmpegAudioStream>(_ffmpeg_content->audio->stream());
419 stream = s->index (_format_context);
423 DCPOMATIC_ASSERT (stream);
425 auto u = time - _pts_offset;
426 if (u < ContentTime ()) {
432 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
436 /* Force re-creation of filter graphs to reset them and hence to make sure
437 they don't have any pre-seek frames knocking about.
439 _filter_graphs.clear();
441 if (video_codec_context ()) {
442 avcodec_flush_buffers (video_codec_context());
445 for (auto i: ffmpeg_content()->ffmpeg_audio_streams()) {
446 avcodec_flush_buffers (_codec_context[i->index(_format_context)]);
449 if (subtitle_codec_context ()) {
450 avcodec_flush_buffers (subtitle_codec_context ());
453 _have_current_subtitle = false;
455 for (auto& i: _next_time) {
456 i.second = boost::optional<dcpomatic::ContentTime>();
461 shared_ptr<FFmpegAudioStream>
462 FFmpegDecoder::audio_stream_from_index (int index) const
464 /* XXX: inefficient */
465 auto streams = ffmpeg_content()->ffmpeg_audio_streams();
466 auto stream = streams.begin();
467 while (stream != streams.end() && !(*stream)->uses_index(_format_context, index)) {
471 if (stream == streams.end ()) {
480 FFmpegDecoder::process_audio_frame (shared_ptr<FFmpegAudioStream> stream)
482 auto frame = audio_frame (stream);
483 auto data = deinterleave_audio(frame);
485 auto const time_base = stream->stream(_format_context)->time_base;
488 if (frame->pts == AV_NOPTS_VALUE) {
489 /* In some streams we see not every frame coming through with a timestamp; for those
490 that have AV_NOPTS_VALUE we need to work out the timestamp ourselves. This is
491 particularly noticeable with TrueHD streams (see #1111).
493 if (_next_time[stream]) {
494 ct = *_next_time[stream];
497 ct = ContentTime::from_seconds (
498 frame->best_effort_timestamp *
502 "Process audio with timestamp %1 (BET %2, timebase %3/%4, (PTS offset %5)",
504 frame->best_effort_timestamp,
507 to_string(_pts_offset)
511 _next_time[stream] = ct + ContentTime::from_frames(data->frames(), stream->frame_rate());
513 if (ct < ContentTime()) {
514 /* Discard audio data that comes before time 0 */
515 auto const remove = min (int64_t(data->frames()), (-ct).frames_ceil(double(stream->frame_rate())));
516 data->move (data->frames() - remove, remove, 0);
517 data->set_frames (data->frames() - remove);
518 ct += ContentTime::from_frames (remove, stream->frame_rate());
521 if (ct < ContentTime()) {
523 "Crazy timestamp %1 for %2 samples in stream %3 (ts=%4 tb=%5, off=%6)",
527 frame->best_effort_timestamp,
529 to_string(_pts_offset)
533 /* Give this data provided there is some, and its time is sane */
534 if (ct >= ContentTime() && data->frames() > 0) {
535 audio->emit (film(), stream, data, ct);
541 FFmpegDecoder::decode_and_process_audio_packet (AVPacket* packet)
543 auto stream = audio_stream_from_index (packet->stream_index);
548 auto context = _codec_context[stream->index(_format_context)];
549 auto frame = audio_frame (stream);
551 LOG_DEBUG_PLAYER("Send audio packet on stream %1", stream->index(_format_context));
552 int r = avcodec_send_packet (context, packet);
554 LOG_WARNING("avcodec_send_packet returned %1 for an audio packet", r);
557 r = avcodec_receive_frame (context, frame);
558 if (r == AVERROR(EAGAIN)) {
559 /* More input is required */
560 LOG_DEBUG_PLAYER_NC("EAGAIN after trying to receive audio frame");
564 /* We choose to be relaxed here about other errors; it seems that there may be valid
565 * data to decode even if an error occurred. #352 may be related (though this was
566 * when we were using an old version of the FFmpeg API).
568 process_audio_frame (stream);
574 FFmpegDecoder::decode_and_process_video_packet (AVPacket* packet)
576 DCPOMATIC_ASSERT (_video_stream);
578 auto context = video_codec_context();
580 bool pending = false;
582 int r = avcodec_send_packet (context, packet);
584 LOG_WARNING("avcodec_send_packet returned %1 for a video packet", r);
587 /* EAGAIN means we should call avcodec_receive_frame and then re-send the same packet */
588 pending = r == AVERROR(EAGAIN);
591 r = avcodec_receive_frame (context, _video_frame);
592 if (r == AVERROR(EAGAIN) || r == AVERROR_EOF || (r < 0 && !packet)) {
593 /* More input is required, no more frames are coming, or we are flushing and there was
594 * some error which we just want to ignore.
598 throw DecodeError (N_("avcodec_receive_frame"), N_("FFmpeg::decode_and_process_video_packet"), r);
601 process_video_frame ();
610 FFmpegDecoder::process_video_frame ()
612 auto graph = _filter_graphs.get(dcp::Size(_video_frame->width, _video_frame->height), static_cast<AVPixelFormat>(_video_frame->format));
613 auto images = graph->process (_video_frame);
615 for (auto const& i: images) {
617 auto image = i.first;
619 if (i.second != AV_NOPTS_VALUE) {
620 double const pts = i.second * av_q2d(_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds();
624 make_shared<RawImageProxy>(image),
625 llrint(pts * _ffmpeg_content->active_video_frame_rate(film()))
628 LOG_WARNING_NC ("Dropping frame without PTS");
635 FFmpegDecoder::decode_and_process_subtitle_packet (AVPacket* packet)
637 auto context = subtitle_codec_context();
644 if (avcodec_decode_subtitle2(context, &sub, &got_subtitle, packet) < 0 || !got_subtitle) {
648 auto sub_period = subtitle_period (packet, ffmpeg_content()->subtitle_stream()->stream(_format_context), sub);
650 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
651 if (_have_current_subtitle) {
652 if (_current_subtitle_to) {
653 only_text()->emit_stop (min(*_current_subtitle_to, sub_period.from + _pts_offset));
655 only_text()->emit_stop (sub_period.from + _pts_offset);
657 _have_current_subtitle = false;
660 if (sub.num_rects <= 0) {
661 /* Nothing new in this subtitle */
662 avsubtitle_free (&sub);
666 /* Subtitle PTS (within the source, not taking into account any of the
667 source that we may have chopped off for the DCP).
670 from = sub_period.from + _pts_offset;
672 _current_subtitle_to = *sub_period.to + _pts_offset;
674 _current_subtitle_to = optional<ContentTime>();
675 _have_current_subtitle = true;
678 ContentBitmapText bitmap_text(from);
679 for (unsigned int i = 0; i < sub.num_rects; ++i) {
680 auto const rect = sub.rects[i];
682 switch (rect->type) {
685 case SUBTITLE_BITMAP:
686 bitmap_text.subs.push_back(process_bitmap_subtitle(rect));
689 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
692 process_ass_subtitle (rect->ass, from);
697 if (!bitmap_text.subs.empty()) {
698 only_text()->emit_bitmap_start(bitmap_text);
701 if (_current_subtitle_to) {
702 only_text()->emit_stop (*_current_subtitle_to);
705 avsubtitle_free (&sub);
710 FFmpegDecoder::process_bitmap_subtitle (AVSubtitleRect const * rect)
712 /* Note BGRA is expressed little-endian, so the first byte in the word is B, second
713 G, third R, fourth A.
715 auto image = make_shared<Image>(AV_PIX_FMT_BGRA, dcp::Size (rect->w, rect->h), Image::Alignment::PADDED);
717 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
718 /* Start of the first line in the subtitle */
719 auto sub_p = rect->pict.data[0];
720 /* sub_p looks up into a BGRA palette which is at rect->pict.data[1];
721 (i.e. first byte B, second G, third R, fourth A)
723 auto const palette = rect->pict.data[1];
725 /* Start of the first line in the subtitle */
726 auto sub_p = rect->data[0];
727 /* sub_p looks up into a BGRA palette which is at rect->data[1].
728 (first byte B, second G, third R, fourth A)
730 auto const* palette = rect->data[1];
732 /* And the stream has a map of those palette colours to colours
733 chosen by the user; created a `mapped' palette from those settings.
735 auto colour_map = ffmpeg_content()->subtitle_stream()->colours();
736 vector<RGBA> mapped_palette (rect->nb_colors);
737 for (int i = 0; i < rect->nb_colors; ++i) {
738 RGBA c (palette[2], palette[1], palette[0], palette[3]);
739 auto j = colour_map.find (c);
740 if (j != colour_map.end ()) {
741 mapped_palette[i] = j->second;
743 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
744 it is from a project that was created before this stuff was added. Just use the
745 colour straight from the original palette.
747 mapped_palette[i] = c;
752 /* Start of the output data */
753 auto out_p = image->data()[0];
755 for (int y = 0; y < rect->h; ++y) {
756 auto sub_line_p = sub_p;
757 auto out_line_p = out_p;
758 for (int x = 0; x < rect->w; ++x) {
759 auto const p = mapped_palette[*sub_line_p++];
765 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
766 sub_p += rect->pict.linesize[0];
768 sub_p += rect->linesize[0];
770 out_p += image->stride()[0];
773 int target_width = subtitle_codec_context()->width;
774 if (target_width == 0 && video_codec_context()) {
775 /* subtitle_codec_context()->width == 0 has been seen in the wild but I don't
776 know if it's supposed to mean something from FFmpeg's point of view.
778 target_width = video_codec_context()->width;
780 int target_height = subtitle_codec_context()->height;
781 if (target_height == 0 && video_codec_context()) {
782 target_height = video_codec_context()->height;
784 DCPOMATIC_ASSERT (target_width);
785 DCPOMATIC_ASSERT (target_height);
786 dcpomatic::Rect<double> const scaled_rect (
787 static_cast<double>(rect->x) / target_width,
788 static_cast<double>(rect->y) / target_height,
789 static_cast<double>(rect->w) / target_width,
790 static_cast<double>(rect->h) / target_height
793 return { image, scaled_rect };
798 FFmpegDecoder::process_ass_subtitle (string ass, ContentTime from)
800 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
801 produces a single format of Dialogue: lines...
806 for (size_t i = 0; i < ass.length(); ++i) {
807 if (commas < 9 && ass[i] == ',') {
809 } else if (commas == 9) {
818 sub::RawSubtitle base;
819 auto raw = sub::SSAReader::parse_line (
822 _ffmpeg_content->video->size().width,
823 _ffmpeg_content->video->size().height,
827 for (auto const& i: sub::collect<vector<sub::Subtitle>>(raw)) {
828 only_text()->emit_plain_start (from, i);