2 Copyright (C) 2012-2016 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 "ffmpeg_decoder.h"
31 #include "subtitle_decoder.h"
32 #include "ffmpeg_audio_stream.h"
33 #include "ffmpeg_subtitle_stream.h"
34 #include "video_filter_graph.h"
35 #include "audio_buffers.h"
36 #include "ffmpeg_content.h"
37 #include "raw_image_proxy.h"
38 #include "video_decoder.h"
40 #include "audio_decoder.h"
41 #include "compose.hpp"
42 #include "subtitle_content.h"
43 #include "audio_content.h"
44 #include <dcp/subtitle_string.h>
45 #include <sub/ssa_reader.h>
46 #include <sub/subtitle.h>
47 #include <sub/collect.h>
49 #include <libavcodec/avcodec.h>
50 #include <libavformat/avformat.h>
52 #include <boost/foreach.hpp>
53 #include <boost/algorithm/string.hpp>
61 #define LOG_GENERAL(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_GENERAL);
62 #define LOG_ERROR(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_ERROR);
63 #define LOG_WARNING_NC(...) _log->log (__VA_ARGS__, LogEntry::TYPE_WARNING);
64 #define LOG_WARNING(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_WARNING);
74 using boost::shared_ptr;
75 using boost::is_any_of;
77 using boost::optional;
78 using boost::dynamic_pointer_cast;
81 FFmpegDecoder::FFmpegDecoder (shared_ptr<const FFmpegContent> c, shared_ptr<Log> log)
84 , _have_current_subtitle (false)
87 video.reset (new VideoDecoder (this, c, log));
88 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate());
89 /* It doesn't matter what size or pixel format this is, it just needs to be black */
90 _black_image.reset (new Image (AV_PIX_FMT_RGB24, dcp::Size (128, 128), true));
91 _black_image->make_black ();
93 _pts_offset = ContentTime ();
97 audio.reset (new AudioDecoder (this, c->audio, log));
101 subtitle.reset (new SubtitleDecoder (this, c->subtitle, log));
104 _next_time.resize (_format_context->nb_streams);
108 FFmpegDecoder::flush ()
110 /* Get any remaining frames */
115 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
117 while (video && decode_video_packet ()) {}
120 decode_audio_packet ();
123 /* Make sure all streams are the same length and round up to the next video frame */
125 FrameRateChange const frc = _ffmpeg_content->film()->active_frame_rate_change(_ffmpeg_content->position());
126 ContentTime full_length (_ffmpeg_content->full_length(), frc);
127 full_length = full_length.ceil (frc.source);
129 double const vfr = _ffmpeg_content->video_frame_rate().get();
130 Frame const f = full_length.frames_round (vfr);
131 Frame v = video->position().frames_round (vfr);
133 video->emit (shared_ptr<const ImageProxy> (new RawImageProxy (_black_image)), v);
138 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, _ffmpeg_content->ffmpeg_audio_streams ()) {
139 ContentTime a = audio->stream_position(i);
140 while (a < full_length) {
141 ContentTime to_do = min (full_length - a, ContentTime::from_seconds (0.1));
142 shared_ptr<AudioBuffers> silence (new AudioBuffers (i->channels(), to_do.frames_ceil (i->frame_rate())));
143 silence->make_silent ();
144 audio->emit (i, silence, a);
155 FFmpegDecoder::pass ()
157 int r = av_read_frame (_format_context, &_packet);
159 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
160 has pretty-much succeeded (and hence generated data which should be processed).
161 Hence it makes sense to continue here in that case.
163 if (r < 0 && r != AVERROR_INVALIDDATA) {
164 if (r != AVERROR_EOF) {
165 /* Maybe we should fail here, but for now we'll just finish off instead */
167 av_strerror (r, buf, sizeof(buf));
168 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
175 int const si = _packet.stream_index;
176 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
178 if (_video_stream && si == _video_stream.get() && !video->ignore()) {
179 decode_video_packet ();
180 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index(_format_context, si) && !subtitle->ignore()) {
181 decode_subtitle_packet ();
183 decode_audio_packet ();
186 av_packet_unref (&_packet);
190 /** @param data pointer to array of pointers to buffers.
191 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
193 shared_ptr<AudioBuffers>
194 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
196 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
198 int const size = av_samples_get_buffer_size (
199 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
202 /* Deinterleave and convert to float */
204 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
205 of the block that do not form a complete sample or frame they will be dropped.
207 int const total_samples = size / bytes_per_audio_sample (stream);
208 int const channels = stream->channels();
209 int const frames = total_samples / channels;
210 shared_ptr<AudioBuffers> audio (new AudioBuffers (channels, frames));
211 float** data = audio->data();
213 switch (audio_sample_format (stream)) {
214 case AV_SAMPLE_FMT_U8:
216 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
219 for (int i = 0; i < total_samples; ++i) {
220 data[channel][sample] = float(*p++) / (1 << 23);
223 if (channel == channels) {
231 case AV_SAMPLE_FMT_S16:
233 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
236 for (int i = 0; i < total_samples; ++i) {
237 data[channel][sample] = float(*p++) / (1 << 15);
240 if (channel == channels) {
248 case AV_SAMPLE_FMT_S16P:
250 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
251 for (int i = 0; i < channels; ++i) {
252 for (int j = 0; j < frames; ++j) {
253 data[i][j] = static_cast<float>(p[i][j]) / (1 << 15);
259 case AV_SAMPLE_FMT_S32:
261 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
264 for (int i = 0; i < total_samples; ++i) {
265 data[channel][sample] = static_cast<float>(*p++) / 2147483648;
268 if (channel == channels) {
276 case AV_SAMPLE_FMT_S32P:
278 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
279 for (int i = 0; i < channels; ++i) {
280 for (int j = 0; j < frames; ++j) {
281 data[i][j] = static_cast<float>(p[i][j]) / 2147483648;
287 case AV_SAMPLE_FMT_FLT:
289 float* p = reinterpret_cast<float*> (_frame->data[0]);
292 for (int i = 0; i < total_samples; ++i) {
293 data[channel][sample] = *p++;
296 if (channel == channels) {
304 case AV_SAMPLE_FMT_FLTP:
306 float** p = reinterpret_cast<float**> (_frame->data);
307 /* Sometimes there aren't as many channels in the _frame as in the stream */
308 for (int i = 0; i < _frame->channels; ++i) {
309 memcpy (data[i], p[i], frames * sizeof(float));
311 for (int i = _frame->channels; i < channels; ++i) {
312 audio->make_silent (i);
318 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
325 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
327 return stream->stream (_format_context)->codec->sample_fmt;
331 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
333 return av_get_bytes_per_sample (audio_sample_format (stream));
337 FFmpegDecoder::seek (ContentTime time, bool accurate)
339 Decoder::seek (time, accurate);
341 /* If we are doing an `accurate' seek, we need to use pre-roll, as
342 we don't really know what the seek will give us.
345 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
348 /* XXX: it seems debatable whether PTS should be used here...
349 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
352 optional<int> stream;
355 stream = _video_stream;
357 shared_ptr<FFmpegAudioStream> s = dynamic_pointer_cast<FFmpegAudioStream> (_ffmpeg_content->audio->stream ());
359 stream = s->index (_format_context);
363 DCPOMATIC_ASSERT (stream);
365 ContentTime u = time - _pts_offset;
366 if (u < ContentTime ()) {
372 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
376 if (video_codec_context ()) {
377 avcodec_flush_buffers (video_codec_context());
380 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, ffmpeg_content()->ffmpeg_audio_streams()) {
381 avcodec_flush_buffers (i->stream(_format_context)->codec);
384 if (subtitle_codec_context ()) {
385 avcodec_flush_buffers (subtitle_codec_context ());
388 _have_current_subtitle = false;
392 FFmpegDecoder::decode_audio_packet ()
394 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
398 AVPacket copy_packet = _packet;
399 int const stream_index = copy_packet.stream_index;
401 /* XXX: inefficient */
402 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
403 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
404 while (stream != streams.end () && !(*stream)->uses_index (_format_context, stream_index)) {
408 if (stream == streams.end ()) {
409 /* The packet's stream may not be an audio one; just ignore it in this method if so */
413 while (copy_packet.size > 0) {
416 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
417 if (decode_result < 0) {
418 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
419 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
420 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
421 even in the face of such an error, so I think we should too.
423 Returning from the method here caused mantis #352.
425 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
427 /* Fudge decode_result so that we come out of the while loop when
428 we've processed this data.
430 decode_result = copy_packet.size;
433 if (frame_finished) {
434 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
437 if (_frame->pts == AV_NOPTS_VALUE && _next_time[stream_index]) {
438 /* In some streams we see not every frame coming through with a timestamp; for those
439 that have AV_NOPTS_VALUE we need to work out the timestamp ourselves. This is
440 particularly noticeable with TrueHD streams (see #1111).
442 ct = *_next_time[stream_index];
444 ct = ContentTime::from_seconds (
445 av_frame_get_best_effort_timestamp (_frame) *
446 av_q2d ((*stream)->stream (_format_context)->time_base))
450 _next_time[stream_index] = ct + ContentTime::from_frames(data->frames(), (*stream)->frame_rate());
452 if (ct < ContentTime ()) {
453 /* Discard audio data that comes before time 0 */
454 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
455 data->move (data->frames() - remove, remove, 0);
456 data->set_frames (data->frames() - remove);
457 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
460 if (ct < ContentTime()) {
462 "Crazy timestamp %1 for %2 samples in stream %3 packet pts %4 (ts=%5 tb=%6, off=%7)",
465 copy_packet.stream_index,
467 av_frame_get_best_effort_timestamp(_frame),
468 av_q2d((*stream)->stream(_format_context)->time_base),
469 to_string(_pts_offset)
473 /* Give this data provided there is some, and its time is sane */
474 if (ct >= ContentTime() && data->frames() > 0) {
475 audio->emit (*stream, data, ct);
479 copy_packet.data += decode_result;
480 copy_packet.size -= decode_result;
485 FFmpegDecoder::decode_video_packet ()
487 DCPOMATIC_ASSERT (_video_stream);
490 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
494 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
496 shared_ptr<VideoFilterGraph> graph;
498 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
499 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
503 if (i == _filter_graphs.end ()) {
504 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
505 graph->setup (_ffmpeg_content->filters ());
506 _filter_graphs.push_back (graph);
507 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
512 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
514 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
516 shared_ptr<Image> image = i->first;
518 if (i->second != AV_NOPTS_VALUE) {
519 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
521 shared_ptr<ImageProxy> (new RawImageProxy (image)),
522 llrint (pts * _ffmpeg_content->active_video_frame_rate ())
525 LOG_WARNING_NC ("Dropping frame without PTS");
533 FFmpegDecoder::decode_subtitle_packet ()
537 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
541 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
542 if (_have_current_subtitle) {
543 if (_current_subtitle_to) {
544 subtitle->emit_stop (min(*_current_subtitle_to, subtitle_period(sub).from + _pts_offset));
546 subtitle->emit_stop (subtitle_period(sub).from + _pts_offset);
548 _have_current_subtitle = false;
551 if (sub.num_rects <= 0) {
552 /* Nothing new in this subtitle */
556 /* Subtitle PTS (within the source, not taking into account any of the
557 source that we may have chopped off for the DCP).
559 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
561 from = sub_period.from + _pts_offset;
562 _have_current_subtitle = true;
564 _current_subtitle_to = *sub_period.to + _pts_offset;
567 for (unsigned int i = 0; i < sub.num_rects; ++i) {
568 AVSubtitleRect const * rect = sub.rects[i];
570 switch (rect->type) {
573 case SUBTITLE_BITMAP:
574 decode_bitmap_subtitle (rect, from);
577 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
580 decode_ass_subtitle (rect->ass, from);
585 avsubtitle_free (&sub);
589 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTime from)
591 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
592 G, third B, fourth A.
594 shared_ptr<Image> image (new Image (AV_PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
596 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
597 /* Start of the first line in the subtitle */
598 uint8_t* sub_p = rect->pict.data[0];
599 /* sub_p looks up into a BGRA palette which is here
600 (i.e. first byte B, second G, third R, fourth A)
602 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
604 /* Start of the first line in the subtitle */
605 uint8_t* sub_p = rect->data[0];
606 /* sub_p looks up into a BGRA palette which is here
607 (i.e. first byte B, second G, third R, fourth A)
609 uint32_t const * palette = (uint32_t *) rect->data[1];
611 /* And the stream has a map of those palette colours to colours
612 chosen by the user; created a `mapped' palette from those settings.
614 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
615 vector<RGBA> mapped_palette (rect->nb_colors);
616 for (int i = 0; i < rect->nb_colors; ++i) {
617 RGBA c ((palette[i] & 0xff0000) >> 16, (palette[i] & 0xff00) >> 8, palette[i] & 0xff, (palette[i] & 0xff000000) >> 24);
618 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
619 if (j != colour_map.end ()) {
620 mapped_palette[i] = j->second;
622 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
623 it is from a project that was created before this stuff was added. Just use the
624 colour straight from the original palette.
626 mapped_palette[i] = c;
630 /* Start of the output data */
631 uint32_t* out_p = (uint32_t *) image->data()[0];
633 for (int y = 0; y < rect->h; ++y) {
634 uint8_t* sub_line_p = sub_p;
635 uint32_t* out_line_p = out_p;
636 for (int x = 0; x < rect->w; ++x) {
637 RGBA const p = mapped_palette[*sub_line_p++];
638 /* XXX: this seems to be wrong to me (isn't the output image RGBA?) but it looks right on screen */
639 *out_line_p++ = (p.a << 24) | (p.r << 16) | (p.g << 8) | p.b;
641 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
642 sub_p += rect->pict.linesize[0];
644 sub_p += rect->linesize[0];
646 out_p += image->stride()[0] / sizeof (uint32_t);
649 int const target_width = subtitle_codec_context()->width;
650 int const target_height = subtitle_codec_context()->height;
651 dcpomatic::Rect<double> const scaled_rect (
652 static_cast<double> (rect->x) / target_width,
653 static_cast<double> (rect->y) / target_height,
654 static_cast<double> (rect->w) / target_width,
655 static_cast<double> (rect->h) / target_height
658 subtitle->emit_image_start (from, image, scaled_rect);
662 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTime from)
664 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
665 produces a single format of Dialogue: lines...
669 split (bits, ass, is_any_of (","));
670 if (bits.size() < 10) {
674 sub::RawSubtitle base;
675 list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (
678 _ffmpeg_content->video->size().width,
679 _ffmpeg_content->video->size().height
682 BOOST_FOREACH (sub::Subtitle const & i, sub::collect<list<sub::Subtitle> > (raw)) {
683 subtitle->emit_text_start (from, i);