2 Copyright (C) 2012-2016 Carl Hetherington <cth@carlh.net>
4 This program is free software; you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 2 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program; if not, write to the Free Software
16 Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20 /** @file src/ffmpeg_decoder.cc
21 * @brief A decoder using FFmpeg to decode content.
25 #include "exceptions.h"
29 #include "ffmpeg_decoder.h"
30 #include "subtitle_decoder.h"
31 #include "ffmpeg_audio_stream.h"
32 #include "ffmpeg_subtitle_stream.h"
33 #include "video_filter_graph.h"
34 #include "audio_buffers.h"
35 #include "ffmpeg_content.h"
36 #include "raw_image_proxy.h"
37 #include "video_decoder.h"
39 #include "md5_digester.h"
40 #include "audio_decoder.h"
41 #include "compose.hpp"
42 #include <dcp/subtitle_string.h>
43 #include <sub/ssa_reader.h>
44 #include <sub/subtitle.h>
45 #include <sub/collect.h>
47 #include <libavcodec/avcodec.h>
48 #include <libavformat/avformat.h>
50 #include <boost/foreach.hpp>
51 #include <boost/algorithm/string.hpp>
59 #define LOG_GENERAL(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_GENERAL);
60 #define LOG_ERROR(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_ERROR);
61 #define LOG_WARNING_NC(...) _log->log (__VA_ARGS__, LogEntry::TYPE_WARNING);
62 #define LOG_WARNING(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_WARNING);
72 using boost::shared_ptr;
73 using boost::is_any_of;
77 FFmpegDecoder::FFmpegDecoder (shared_ptr<const FFmpegContent> c, shared_ptr<Log> log, bool fast)
82 video.reset (new VideoDecoder (this, c, log));
83 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate());
85 _pts_offset = ContentTime ();
89 audio.reset (new AudioDecoder (this, c->audio, fast, log));
97 bind (&FFmpegDecoder::image_subtitles_during, this, _1, _2),
98 bind (&FFmpegDecoder::text_subtitles_during, this, _1, _2)
105 FFmpegDecoder::flush ()
107 /* Get any remaining frames */
112 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
114 while (video && decode_video_packet ()) {}
117 decode_audio_packet ();
123 FFmpegDecoder::pass (PassReason reason, bool accurate)
125 int r = av_read_frame (_format_context, &_packet);
127 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
128 has pretty-much succeeded (and hence generated data which should be processed).
129 Hence it makes sense to continue here in that case.
131 if (r < 0 && r != AVERROR_INVALIDDATA) {
132 if (r != AVERROR_EOF) {
133 /* Maybe we should fail here, but for now we'll just finish off instead */
135 av_strerror (r, buf, sizeof(buf));
136 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), buf, r);
143 int const si = _packet.stream_index;
144 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
146 if (_video_stream && si == _video_stream.get() && !video->ignore() && (accurate || reason != PASS_REASON_SUBTITLE)) {
147 decode_video_packet ();
148 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index (_format_context, si)) {
149 decode_subtitle_packet ();
150 } else if (accurate || reason != PASS_REASON_SUBTITLE) {
151 decode_audio_packet ();
154 av_packet_unref (&_packet);
158 /** @param data pointer to array of pointers to buffers.
159 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
161 shared_ptr<AudioBuffers>
162 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
164 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
166 int const size = av_samples_get_buffer_size (
167 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
170 /* Deinterleave and convert to float */
172 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
173 of the block that do not form a complete sample or frame they will be dropped.
175 int const total_samples = size / bytes_per_audio_sample (stream);
176 int const frames = total_samples / stream->channels();
177 shared_ptr<AudioBuffers> audio (new AudioBuffers (stream->channels(), frames));
179 switch (audio_sample_format (stream)) {
180 case AV_SAMPLE_FMT_U8:
182 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
185 for (int i = 0; i < total_samples; ++i) {
186 audio->data(channel)[sample] = float(*p++) / (1 << 23);
189 if (channel == stream->channels()) {
197 case AV_SAMPLE_FMT_S16:
199 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
202 for (int i = 0; i < total_samples; ++i) {
203 audio->data(channel)[sample] = float(*p++) / (1 << 15);
206 if (channel == stream->channels()) {
214 case AV_SAMPLE_FMT_S16P:
216 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
217 for (int i = 0; i < stream->channels(); ++i) {
218 for (int j = 0; j < frames; ++j) {
219 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
225 case AV_SAMPLE_FMT_S32:
227 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
230 for (int i = 0; i < total_samples; ++i) {
231 audio->data(channel)[sample] = static_cast<float>(*p++) / (1 << 31);
234 if (channel == stream->channels()) {
242 case AV_SAMPLE_FMT_S32P:
244 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
245 for (int i = 0; i < stream->channels(); ++i) {
246 for (int j = 0; j < frames; ++j) {
247 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 31);
253 case AV_SAMPLE_FMT_FLT:
255 float* p = reinterpret_cast<float*> (_frame->data[0]);
258 for (int i = 0; i < total_samples; ++i) {
259 audio->data(channel)[sample] = *p++;
262 if (channel == stream->channels()) {
270 case AV_SAMPLE_FMT_FLTP:
272 float** p = reinterpret_cast<float**> (_frame->data);
273 /* Sometimes there aren't as many channels in the _frame as in the stream */
274 for (int i = 0; i < _frame->channels; ++i) {
275 memcpy (audio->data(i), p[i], frames * sizeof(float));
277 for (int i = _frame->channels; i < stream->channels(); ++i) {
278 audio->make_silent (i);
284 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
291 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
293 return stream->stream (_format_context)->codec->sample_fmt;
297 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
299 return av_get_bytes_per_sample (audio_sample_format (stream));
303 FFmpegDecoder::seek (ContentTime time, bool accurate)
305 video->seek (time, accurate);
306 audio->seek (time, accurate);
307 subtitle->seek (time, accurate);
309 /* If we are doing an `accurate' seek, we need to use pre-roll, as
310 we don't really know what the seek will give us.
313 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
316 /* XXX: it seems debatable whether PTS should be used here...
317 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
320 DCPOMATIC_ASSERT (_video_stream);
322 ContentTime u = time - _pts_offset;
323 if (u < ContentTime ()) {
329 u.seconds() / av_q2d (_format_context->streams[_video_stream.get()]->time_base),
333 avcodec_flush_buffers (video_codec_context());
335 /* XXX: should be flushing audio buffers? */
337 if (subtitle_codec_context ()) {
338 avcodec_flush_buffers (subtitle_codec_context ());
343 FFmpegDecoder::decode_audio_packet ()
345 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
349 AVPacket copy_packet = _packet;
351 /* XXX: inefficient */
352 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
353 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
354 while (stream != streams.end () && !(*stream)->uses_index (_format_context, copy_packet.stream_index)) {
358 if (stream == streams.end ()) {
359 /* The packet's stream may not be an audio one; just ignore it in this method if so */
363 while (copy_packet.size > 0) {
366 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
367 if (decode_result < 0) {
368 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
369 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
370 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
371 even in the face of such an error, so I think we should too.
373 Returning from the method here caused mantis #352.
375 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
377 /* Fudge decode_result so that we come out of the while loop when
378 we've processed this data.
380 decode_result = copy_packet.size;
383 if (frame_finished) {
384 ContentTime ct = ContentTime::from_seconds (
385 av_frame_get_best_effort_timestamp (_frame) *
386 av_q2d ((*stream)->stream (_format_context)->time_base))
389 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
391 if (ct < ContentTime ()) {
392 /* Discard audio data that comes before time 0 */
393 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
394 data->move (remove, 0, data->frames() - remove);
395 data->set_frames (data->frames() - remove);
396 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
399 if (data->frames() > 0) {
400 audio->give (*stream, data, ct);
404 copy_packet.data += decode_result;
405 copy_packet.size -= decode_result;
410 FFmpegDecoder::decode_video_packet ()
412 DCPOMATIC_ASSERT (_video_stream);
415 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
419 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
421 shared_ptr<VideoFilterGraph> graph;
423 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
424 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
428 if (i == _filter_graphs.end ()) {
429 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
430 graph->setup (_ffmpeg_content->filters ());
431 _filter_graphs.push_back (graph);
432 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
437 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
439 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
441 shared_ptr<Image> image = i->first;
443 if (i->second != AV_NOPTS_VALUE) {
444 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
446 shared_ptr<ImageProxy> (new RawImageProxy (image)),
447 llrint (pts * _ffmpeg_content->active_video_frame_rate ())
450 LOG_WARNING_NC ("Dropping frame without PTS");
458 FFmpegDecoder::decode_subtitle_packet ()
462 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
466 if (sub.num_rects <= 0) {
467 /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
468 indicate that the previous subtitle should stop. We can ignore it here.
473 /* Subtitle PTS (within the source, not taking into account any of the
474 source that we may have chopped off for the DCP).
476 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
477 ContentTimePeriod period;
478 period.from = sub_period.from + _pts_offset;
480 /* We already know the subtitle period `to' time */
481 period.to = sub_period.to.get() + _pts_offset;
483 /* We have to look up the `to' time in the stream's records */
484 period.to = ffmpeg_content()->subtitle_stream()->find_subtitle_to (subtitle_id (sub));
487 for (unsigned int i = 0; i < sub.num_rects; ++i) {
488 AVSubtitleRect const * rect = sub.rects[i];
490 switch (rect->type) {
493 case SUBTITLE_BITMAP:
494 decode_bitmap_subtitle (rect, period);
497 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
500 decode_ass_subtitle (rect->ass, period);
505 avsubtitle_free (&sub);
508 list<ContentTimePeriod>
509 FFmpegDecoder::image_subtitles_during (ContentTimePeriod p, bool starting) const
511 return _ffmpeg_content->image_subtitles_during (p, starting);
514 list<ContentTimePeriod>
515 FFmpegDecoder::text_subtitles_during (ContentTimePeriod p, bool starting) const
517 return _ffmpeg_content->text_subtitles_during (p, starting);
521 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTimePeriod period)
523 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
524 G, third B, fourth A.
526 shared_ptr<Image> image (new Image (AV_PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
528 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
529 /* Start of the first line in the subtitle */
530 uint8_t* sub_p = rect->pict.data[0];
531 /* sub_p looks up into a BGRA palette which is here
532 (i.e. first byte B, second G, third R, fourth A)
534 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
536 /* Start of the first line in the subtitle */
537 uint8_t* sub_p = rect->data[0];
538 /* sub_p looks up into a BGRA palette which is here
539 (i.e. first byte B, second G, third R, fourth A)
541 uint32_t const * palette = (uint32_t *) rect->data[1];
543 /* And the stream has a map of those palette colours to colours
544 chosen by the user; created a `mapped' palette from those settings.
546 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
547 vector<RGBA> mapped_palette (rect->nb_colors);
548 for (int i = 0; i < rect->nb_colors; ++i) {
549 RGBA c ((palette[i] & 0xff0000) >> 16, (palette[i] & 0xff00) >> 8, palette[i] & 0xff, (palette[i] & 0xff000000) >> 24);
550 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
551 if (j != colour_map.end ()) {
552 mapped_palette[i] = j->second;
554 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
555 it is from a project that was created before this stuff was added. Just use the
556 colour straight from the original palette.
558 mapped_palette[i] = c;
562 /* Start of the output data */
563 uint32_t* out_p = (uint32_t *) image->data()[0];
565 for (int y = 0; y < rect->h; ++y) {
566 uint8_t* sub_line_p = sub_p;
567 uint32_t* out_line_p = out_p;
568 for (int x = 0; x < rect->w; ++x) {
569 RGBA const p = mapped_palette[*sub_line_p++];
570 /* XXX: this seems to be wrong to me (isn't the output image RGBA?) but it looks right on screen */
571 *out_line_p++ = (p.a << 24) | (p.r << 16) | (p.g << 8) | p.b;
573 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
574 sub_p += rect->pict.linesize[0];
576 sub_p += rect->linesize[0];
578 out_p += image->stride()[0] / sizeof (uint32_t);
581 dcp::Size const vs = _ffmpeg_content->video->size ();
582 dcpomatic::Rect<double> const scaled_rect (
583 static_cast<double> (rect->x) / vs.width,
584 static_cast<double> (rect->y) / vs.height,
585 static_cast<double> (rect->w) / vs.width,
586 static_cast<double> (rect->h) / vs.height
589 subtitle->give_image (period, image, scaled_rect);
593 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTimePeriod period)
595 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
596 produces a single format of Dialogue: lines...
600 split (bits, ass, is_any_of (","));
601 if (bits.size() < 10) {
605 sub::RawSubtitle base;
606 list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (base, bits[9]);
607 list<sub::Subtitle> subs = sub::collect<list<sub::Subtitle> > (raw);
609 /* XXX: lots of this is copied from TextSubtitle; there should probably be some sharing */
611 /* Highest line index in this subtitle */
613 BOOST_FOREACH (sub::Subtitle i, subs) {
614 BOOST_FOREACH (sub::Line j, i.lines) {
615 DCPOMATIC_ASSERT (j.vertical_position.reference && j.vertical_position.reference.get() == sub::TOP_OF_SUBTITLE);
616 DCPOMATIC_ASSERT (j.vertical_position.line);
617 highest = max (highest, j.vertical_position.line.get());
621 list<dcp::SubtitleString> ss;
623 BOOST_FOREACH (sub::Subtitle i, sub::collect<list<sub::Subtitle> > (sub::SSAReader::parse_line (base, bits[9]))) {
624 BOOST_FOREACH (sub::Line j, i.lines) {
625 BOOST_FOREACH (sub::Block k, j.blocks) {
627 dcp::SubtitleString (
628 boost::optional<string> (),
631 dcp::Colour (255, 255, 255),
632 /* 48pt is 1/22nd of the screen height */
635 dcp::Time (i.from.seconds(), 1000),
636 dcp::Time (i.to.seconds(), 1000),
639 /* This 1.015 is an arbitrary value to lift the bottom sub off the bottom
640 of the screen a bit to a pleasing degree.
642 1.015 - ((1 + highest - j.vertical_position.line.get()) * 1.5 / 22),
646 static_cast<dcp::Effect> (0),
647 dcp::Colour (255, 255, 255),
656 subtitle->give_text (period, ss);