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 <dcp/subtitle_string.h>
44 #include <sub/ssa_reader.h>
45 #include <sub/subtitle.h>
46 #include <sub/collect.h>
48 #include <libavcodec/avcodec.h>
49 #include <libavformat/avformat.h>
51 #include <boost/foreach.hpp>
52 #include <boost/algorithm/string.hpp>
53 #include <boost/make_shared.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::make_shared;
76 using boost::is_any_of;
80 FFmpegDecoder::FFmpegDecoder (shared_ptr<const FFmpegContent> c, shared_ptr<Log> log, bool fast)
85 video.reset (new VideoDecoder (this, c, log));
86 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate());
88 _pts_offset = ContentTime ();
92 audio.reset (new AudioDecoder (this, c->audio, fast, log));
100 bind (&FFmpegDecoder::image_subtitles_during, this, _1, _2),
101 bind (&FFmpegDecoder::text_subtitles_during, this, _1, _2)
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 ();
126 FFmpegDecoder::pass (PassReason reason, bool accurate)
128 int r = av_read_frame (_format_context, &_packet);
130 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
131 has pretty-much succeeded (and hence generated data which should be processed).
132 Hence it makes sense to continue here in that case.
134 if (r < 0 && r != AVERROR_INVALIDDATA) {
135 if (r != AVERROR_EOF) {
136 /* Maybe we should fail here, but for now we'll just finish off instead */
138 av_strerror (r, buf, sizeof(buf));
139 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), buf, r);
146 int const si = _packet.stream_index;
147 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
149 if (_video_stream && si == _video_stream.get() && !video->ignore() && (accurate || reason != PASS_REASON_SUBTITLE)) {
150 decode_video_packet ();
151 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index (_format_context, si)) {
152 decode_subtitle_packet ();
153 } else if (accurate || reason != PASS_REASON_SUBTITLE) {
154 decode_audio_packet ();
157 av_packet_unref (&_packet);
161 /** @param data pointer to array of pointers to buffers.
162 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
164 shared_ptr<AudioBuffers>
165 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
167 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
169 int const size = av_samples_get_buffer_size (
170 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
173 /* Deinterleave and convert to float */
175 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
176 of the block that do not form a complete sample or frame they will be dropped.
178 int const total_samples = size / bytes_per_audio_sample (stream);
179 int const frames = total_samples / stream->channels();
180 shared_ptr<AudioBuffers> audio = make_shared<AudioBuffers> (stream->channels(), frames);
182 switch (audio_sample_format (stream)) {
183 case AV_SAMPLE_FMT_U8:
185 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
188 for (int i = 0; i < total_samples; ++i) {
189 audio->data(channel)[sample] = float(*p++) / (1 << 23);
192 if (channel == stream->channels()) {
200 case AV_SAMPLE_FMT_S16:
202 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
205 for (int i = 0; i < total_samples; ++i) {
206 audio->data(channel)[sample] = float(*p++) / (1 << 15);
209 if (channel == stream->channels()) {
217 case AV_SAMPLE_FMT_S16P:
219 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
220 for (int i = 0; i < stream->channels(); ++i) {
221 for (int j = 0; j < frames; ++j) {
222 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
228 case AV_SAMPLE_FMT_S32:
230 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
233 for (int i = 0; i < total_samples; ++i) {
234 audio->data(channel)[sample] = static_cast<float>(*p++) / 2147483648;
237 if (channel == stream->channels()) {
245 case AV_SAMPLE_FMT_S32P:
247 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
248 for (int i = 0; i < stream->channels(); ++i) {
249 for (int j = 0; j < frames; ++j) {
250 audio->data(i)[j] = static_cast<float>(p[i][j]) / 2147483648;
256 case AV_SAMPLE_FMT_FLT:
258 float* p = reinterpret_cast<float*> (_frame->data[0]);
261 for (int i = 0; i < total_samples; ++i) {
262 audio->data(channel)[sample] = *p++;
265 if (channel == stream->channels()) {
273 case AV_SAMPLE_FMT_FLTP:
275 float** p = reinterpret_cast<float**> (_frame->data);
276 /* Sometimes there aren't as many channels in the _frame as in the stream */
277 for (int i = 0; i < _frame->channels; ++i) {
278 memcpy (audio->data(i), p[i], frames * sizeof(float));
280 for (int i = _frame->channels; i < stream->channels(); ++i) {
281 audio->make_silent (i);
287 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
294 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
296 return stream->stream (_format_context)->codec->sample_fmt;
300 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
302 return av_get_bytes_per_sample (audio_sample_format (stream));
306 FFmpegDecoder::seek (ContentTime time, bool accurate)
309 video->seek (time, accurate);
313 audio->seek (time, accurate);
317 subtitle->seek (time, accurate);
320 /* If we are doing an `accurate' seek, we need to use pre-roll, as
321 we don't really know what the seek will give us.
324 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
327 /* XXX: it seems debatable whether PTS should be used here...
328 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
331 DCPOMATIC_ASSERT (_video_stream);
333 ContentTime u = time - _pts_offset;
334 if (u < ContentTime ()) {
340 u.seconds() / av_q2d (_format_context->streams[_video_stream.get()]->time_base),
344 avcodec_flush_buffers (video_codec_context());
346 /* XXX: should be flushing audio buffers? */
348 if (subtitle_codec_context ()) {
349 avcodec_flush_buffers (subtitle_codec_context ());
354 FFmpegDecoder::decode_audio_packet ()
356 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
360 AVPacket copy_packet = _packet;
362 /* XXX: inefficient */
363 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
364 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
365 while (stream != streams.end () && !(*stream)->uses_index (_format_context, copy_packet.stream_index)) {
369 if (stream == streams.end ()) {
370 /* The packet's stream may not be an audio one; just ignore it in this method if so */
374 while (copy_packet.size > 0) {
377 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
378 if (decode_result < 0) {
379 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
380 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
381 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
382 even in the face of such an error, so I think we should too.
384 Returning from the method here caused mantis #352.
386 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
388 /* Fudge decode_result so that we come out of the while loop when
389 we've processed this data.
391 decode_result = copy_packet.size;
394 if (frame_finished) {
395 ContentTime ct = ContentTime::from_seconds (
396 av_frame_get_best_effort_timestamp (_frame) *
397 av_q2d ((*stream)->stream (_format_context)->time_base))
400 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
402 if (ct < ContentTime ()) {
403 /* Discard audio data that comes before time 0 */
404 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
405 data->move (remove, 0, data->frames() - remove);
406 data->set_frames (data->frames() - remove);
407 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
410 if (data->frames() > 0) {
411 audio->give (*stream, data, ct);
415 copy_packet.data += decode_result;
416 copy_packet.size -= decode_result;
421 FFmpegDecoder::decode_video_packet ()
423 DCPOMATIC_ASSERT (_video_stream);
426 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
430 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
432 shared_ptr<VideoFilterGraph> graph;
434 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
435 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
439 if (i == _filter_graphs.end ()) {
440 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
441 graph->setup (_ffmpeg_content->filters ());
442 _filter_graphs.push_back (graph);
443 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
448 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
450 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
452 shared_ptr<Image> image = i->first;
454 if (i->second != AV_NOPTS_VALUE) {
455 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
457 make_shared<RawImageProxy> (image),
458 llrint (pts * _ffmpeg_content->active_video_frame_rate ())
461 LOG_WARNING_NC ("Dropping frame without PTS");
469 FFmpegDecoder::decode_subtitle_packet ()
473 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
477 if (sub.num_rects <= 0) {
478 /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
479 indicate that the previous subtitle should stop. We can ignore it here.
484 /* Subtitle PTS (within the source, not taking into account any of the
485 source that we may have chopped off for the DCP).
487 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
488 ContentTimePeriod period;
489 period.from = sub_period.from + _pts_offset;
491 /* We already know the subtitle period `to' time */
492 period.to = sub_period.to.get() + _pts_offset;
494 /* We have to look up the `to' time in the stream's records */
495 period.to = ffmpeg_content()->subtitle_stream()->find_subtitle_to (subtitle_id (sub));
498 for (unsigned int i = 0; i < sub.num_rects; ++i) {
499 AVSubtitleRect const * rect = sub.rects[i];
501 switch (rect->type) {
504 case SUBTITLE_BITMAP:
505 decode_bitmap_subtitle (rect, period);
508 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
511 decode_ass_subtitle (rect->ass, period);
516 avsubtitle_free (&sub);
519 list<ContentTimePeriod>
520 FFmpegDecoder::image_subtitles_during (ContentTimePeriod p, bool starting) const
522 return _ffmpeg_content->image_subtitles_during (p, starting);
525 list<ContentTimePeriod>
526 FFmpegDecoder::text_subtitles_during (ContentTimePeriod p, bool starting) const
528 return _ffmpeg_content->text_subtitles_during (p, starting);
532 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTimePeriod period)
534 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
535 G, third B, fourth A.
537 shared_ptr<Image> image = make_shared<Image> (AV_PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true);
539 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
540 /* Start of the first line in the subtitle */
541 uint8_t* sub_p = rect->pict.data[0];
542 /* sub_p looks up into a BGRA palette which is here
543 (i.e. first byte B, second G, third R, fourth A)
545 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
547 /* Start of the first line in the subtitle */
548 uint8_t* sub_p = rect->data[0];
549 /* sub_p looks up into a BGRA palette which is here
550 (i.e. first byte B, second G, third R, fourth A)
552 uint32_t const * palette = (uint32_t *) rect->data[1];
554 /* And the stream has a map of those palette colours to colours
555 chosen by the user; created a `mapped' palette from those settings.
557 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
558 vector<RGBA> mapped_palette (rect->nb_colors);
559 for (int i = 0; i < rect->nb_colors; ++i) {
560 RGBA c ((palette[i] & 0xff0000) >> 16, (palette[i] & 0xff00) >> 8, palette[i] & 0xff, (palette[i] & 0xff000000) >> 24);
561 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
562 if (j != colour_map.end ()) {
563 mapped_palette[i] = j->second;
565 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
566 it is from a project that was created before this stuff was added. Just use the
567 colour straight from the original palette.
569 mapped_palette[i] = c;
573 /* Start of the output data */
574 uint32_t* out_p = (uint32_t *) image->data()[0];
576 for (int y = 0; y < rect->h; ++y) {
577 uint8_t* sub_line_p = sub_p;
578 uint32_t* out_line_p = out_p;
579 for (int x = 0; x < rect->w; ++x) {
580 RGBA const p = mapped_palette[*sub_line_p++];
581 /* XXX: this seems to be wrong to me (isn't the output image RGBA?) but it looks right on screen */
582 *out_line_p++ = (p.a << 24) | (p.r << 16) | (p.g << 8) | p.b;
584 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
585 sub_p += rect->pict.linesize[0];
587 sub_p += rect->linesize[0];
589 out_p += image->stride()[0] / sizeof (uint32_t);
592 dcp::Size const vs = _ffmpeg_content->video->size ();
593 dcpomatic::Rect<double> const scaled_rect (
594 static_cast<double> (rect->x) / vs.width,
595 static_cast<double> (rect->y) / vs.height,
596 static_cast<double> (rect->w) / vs.width,
597 static_cast<double> (rect->h) / vs.height
600 subtitle->give_image (period, image, scaled_rect);
604 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTimePeriod period)
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->give_text (period, i);