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());
90 _pts_offset = ContentTime ();
94 audio.reset (new AudioDecoder (this, c->audio, log));
98 subtitle.reset (new SubtitleDecoder (this, c->subtitle, log));
103 FFmpegDecoder::flush ()
105 /* Get any remaining frames */
110 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
112 while (video && decode_video_packet ()) {}
115 decode_audio_packet ();
120 FFmpegDecoder::pass ()
122 int r = av_read_frame (_format_context, &_packet);
124 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
125 has pretty-much succeeded (and hence generated data which should be processed).
126 Hence it makes sense to continue here in that case.
128 if (r < 0 && r != AVERROR_INVALIDDATA) {
129 if (r != AVERROR_EOF) {
130 /* Maybe we should fail here, but for now we'll just finish off instead */
132 av_strerror (r, buf, sizeof(buf));
133 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
140 int const si = _packet.stream_index;
141 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
143 if (_video_stream && si == _video_stream.get() && !video->ignore()) {
144 decode_video_packet ();
145 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index (_format_context, si)) {
146 decode_subtitle_packet ();
148 decode_audio_packet ();
151 av_packet_unref (&_packet);
155 /** @param data pointer to array of pointers to buffers.
156 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
158 shared_ptr<AudioBuffers>
159 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
161 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
163 int const size = av_samples_get_buffer_size (
164 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
167 /* Deinterleave and convert to float */
169 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
170 of the block that do not form a complete sample or frame they will be dropped.
172 int const total_samples = size / bytes_per_audio_sample (stream);
173 int const frames = total_samples / stream->channels();
174 shared_ptr<AudioBuffers> audio (new AudioBuffers (stream->channels(), frames));
176 switch (audio_sample_format (stream)) {
177 case AV_SAMPLE_FMT_U8:
179 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
182 for (int i = 0; i < total_samples; ++i) {
183 audio->data(channel)[sample] = float(*p++) / (1 << 23);
186 if (channel == stream->channels()) {
194 case AV_SAMPLE_FMT_S16:
196 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
199 for (int i = 0; i < total_samples; ++i) {
200 audio->data(channel)[sample] = float(*p++) / (1 << 15);
203 if (channel == stream->channels()) {
211 case AV_SAMPLE_FMT_S16P:
213 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
214 for (int i = 0; i < stream->channels(); ++i) {
215 for (int j = 0; j < frames; ++j) {
216 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
222 case AV_SAMPLE_FMT_S32:
224 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
227 for (int i = 0; i < total_samples; ++i) {
228 audio->data(channel)[sample] = static_cast<float>(*p++) / 2147483648;
231 if (channel == stream->channels()) {
239 case AV_SAMPLE_FMT_S32P:
241 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
242 for (int i = 0; i < stream->channels(); ++i) {
243 for (int j = 0; j < frames; ++j) {
244 audio->data(i)[j] = static_cast<float>(p[i][j]) / 2147483648;
250 case AV_SAMPLE_FMT_FLT:
252 float* p = reinterpret_cast<float*> (_frame->data[0]);
255 for (int i = 0; i < total_samples; ++i) {
256 audio->data(channel)[sample] = *p++;
259 if (channel == stream->channels()) {
267 case AV_SAMPLE_FMT_FLTP:
269 float** p = reinterpret_cast<float**> (_frame->data);
270 /* Sometimes there aren't as many channels in the _frame as in the stream */
271 for (int i = 0; i < _frame->channels; ++i) {
272 memcpy (audio->data(i), p[i], frames * sizeof(float));
274 for (int i = _frame->channels; i < stream->channels(); ++i) {
275 audio->make_silent (i);
281 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
288 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
290 return stream->stream (_format_context)->codec->sample_fmt;
294 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
296 return av_get_bytes_per_sample (audio_sample_format (stream));
300 FFmpegDecoder::seek (ContentTime time, bool accurate)
302 Decoder::seek (time, accurate);
304 /* If we are doing an `accurate' seek, we need to use pre-roll, as
305 we don't really know what the seek will give us.
308 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
311 /* XXX: it seems debatable whether PTS should be used here...
312 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
315 optional<int> stream;
318 stream = _video_stream;
320 shared_ptr<FFmpegAudioStream> s = dynamic_pointer_cast<FFmpegAudioStream> (_ffmpeg_content->audio->stream ());
322 stream = s->index (_format_context);
326 DCPOMATIC_ASSERT (stream);
328 ContentTime u = time - _pts_offset;
329 if (u < ContentTime ()) {
335 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
339 if (video_codec_context ()) {
340 avcodec_flush_buffers (video_codec_context());
343 BOOST_FOREACH (shared_ptr<FFmpegAudioStream> i, ffmpeg_content()->ffmpeg_audio_streams()) {
344 avcodec_flush_buffers (i->stream(_format_context)->codec);
347 if (subtitle_codec_context ()) {
348 avcodec_flush_buffers (subtitle_codec_context ());
351 _have_current_subtitle = false;
355 FFmpegDecoder::decode_audio_packet ()
357 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
361 AVPacket copy_packet = _packet;
363 /* XXX: inefficient */
364 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
365 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
366 while (stream != streams.end () && !(*stream)->uses_index (_format_context, copy_packet.stream_index)) {
370 if (stream == streams.end ()) {
371 /* The packet's stream may not be an audio one; just ignore it in this method if so */
375 while (copy_packet.size > 0) {
378 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
379 if (decode_result < 0) {
380 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
381 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
382 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
383 even in the face of such an error, so I think we should too.
385 Returning from the method here caused mantis #352.
387 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
389 /* Fudge decode_result so that we come out of the while loop when
390 we've processed this data.
392 decode_result = copy_packet.size;
395 if (frame_finished) {
396 ContentTime ct = ContentTime::from_seconds (
397 av_frame_get_best_effort_timestamp (_frame) *
398 av_q2d ((*stream)->stream (_format_context)->time_base))
401 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
403 if (ct < ContentTime ()) {
404 /* Discard audio data that comes before time 0 */
405 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
406 data->move (data->frames() - remove, remove, 0);
407 data->set_frames (data->frames() - remove);
408 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
411 if (ct < ContentTime()) {
412 LOG_WARNING ("Crazy timestamp %1", to_string (ct));
415 /* Give this data provided there is some, and its time is sane */
416 if (ct >= ContentTime() && data->frames() > 0) {
417 audio->emit (*stream, data, ct);
421 copy_packet.data += decode_result;
422 copy_packet.size -= decode_result;
427 FFmpegDecoder::decode_video_packet ()
429 DCPOMATIC_ASSERT (_video_stream);
432 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
436 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
438 shared_ptr<VideoFilterGraph> graph;
440 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
441 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
445 if (i == _filter_graphs.end ()) {
446 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
447 graph->setup (_ffmpeg_content->filters ());
448 _filter_graphs.push_back (graph);
449 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
454 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
456 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
458 shared_ptr<Image> image = i->first;
460 if (i->second != AV_NOPTS_VALUE) {
461 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
463 shared_ptr<ImageProxy> (new RawImageProxy (image)),
464 llrint (pts * _ffmpeg_content->active_video_frame_rate ())
467 LOG_WARNING_NC ("Dropping frame without PTS");
475 FFmpegDecoder::decode_subtitle_packet ()
479 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
483 /* Stop any current subtitle, either at the time it was supposed to stop, or now if now is sooner */
484 if (_have_current_subtitle) {
485 if (_current_subtitle_to) {
486 subtitle->emit_stop (min(*_current_subtitle_to, subtitle_period(sub).from + _pts_offset));
488 subtitle->emit_stop (subtitle_period(sub).from + _pts_offset);
490 _have_current_subtitle = false;
493 if (sub.num_rects <= 0) {
494 /* Nothing new in this subtitle */
498 /* Subtitle PTS (within the source, not taking into account any of the
499 source that we may have chopped off for the DCP).
501 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
503 from = sub_period.from + _pts_offset;
504 _have_current_subtitle = true;
506 _current_subtitle_to = *sub_period.to + _pts_offset;
509 for (unsigned int i = 0; i < sub.num_rects; ++i) {
510 AVSubtitleRect const * rect = sub.rects[i];
512 switch (rect->type) {
515 case SUBTITLE_BITMAP:
516 decode_bitmap_subtitle (rect, from);
519 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
522 decode_ass_subtitle (rect->ass, from);
527 avsubtitle_free (&sub);
531 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTime from)
533 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
534 G, third B, fourth A.
536 shared_ptr<Image> image (new Image (AV_PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
538 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
539 /* Start of the first line in the subtitle */
540 uint8_t* sub_p = rect->pict.data[0];
541 /* sub_p looks up into a BGRA palette which is here
542 (i.e. first byte B, second G, third R, fourth A)
544 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
546 /* Start of the first line in the subtitle */
547 uint8_t* sub_p = rect->data[0];
548 /* sub_p looks up into a BGRA palette which is here
549 (i.e. first byte B, second G, third R, fourth A)
551 uint32_t const * palette = (uint32_t *) rect->data[1];
553 /* And the stream has a map of those palette colours to colours
554 chosen by the user; created a `mapped' palette from those settings.
556 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
557 vector<RGBA> mapped_palette (rect->nb_colors);
558 for (int i = 0; i < rect->nb_colors; ++i) {
559 RGBA c ((palette[i] & 0xff0000) >> 16, (palette[i] & 0xff00) >> 8, palette[i] & 0xff, (palette[i] & 0xff000000) >> 24);
560 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
561 if (j != colour_map.end ()) {
562 mapped_palette[i] = j->second;
564 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
565 it is from a project that was created before this stuff was added. Just use the
566 colour straight from the original palette.
568 mapped_palette[i] = c;
572 /* Start of the output data */
573 uint32_t* out_p = (uint32_t *) image->data()[0];
575 for (int y = 0; y < rect->h; ++y) {
576 uint8_t* sub_line_p = sub_p;
577 uint32_t* out_line_p = out_p;
578 for (int x = 0; x < rect->w; ++x) {
579 RGBA const p = mapped_palette[*sub_line_p++];
580 /* XXX: this seems to be wrong to me (isn't the output image RGBA?) but it looks right on screen */
581 *out_line_p++ = (p.a << 24) | (p.r << 16) | (p.g << 8) | p.b;
583 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
584 sub_p += rect->pict.linesize[0];
586 sub_p += rect->linesize[0];
588 out_p += image->stride()[0] / sizeof (uint32_t);
591 int const target_width = subtitle_codec_context()->width;
592 int const target_height = subtitle_codec_context()->height;
593 dcpomatic::Rect<double> const scaled_rect (
594 static_cast<double> (rect->x) / target_width,
595 static_cast<double> (rect->y) / target_height,
596 static_cast<double> (rect->w) / target_width,
597 static_cast<double> (rect->h) / target_height
600 subtitle->emit_image_start (from, image, scaled_rect);
604 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTime from)
606 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
607 produces a single format of Dialogue: lines...
611 split (bits, ass, is_any_of (","));
612 if (bits.size() < 10) {
616 sub::RawSubtitle base;
617 list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (base, bits[9]);
619 BOOST_FOREACH (sub::Subtitle const & i, sub::collect<list<sub::Subtitle> > (raw)) {
620 subtitle->emit_text_start (from, i);