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)
86 video.reset (new VideoDecoder (this, c, log));
87 _pts_offset = pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->active_video_frame_rate());
89 _pts_offset = ContentTime ();
93 audio.reset (new AudioDecoder (this, c->audio, log));
101 bind (&FFmpegDecoder::image_subtitles_during, this, _1, _2),
102 bind (&FFmpegDecoder::text_subtitles_during, this, _1, _2)
109 FFmpegDecoder::flush ()
111 /* Get any remaining frames */
116 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
118 while (video && decode_video_packet ()) {}
121 decode_audio_packet ();
127 FFmpegDecoder::pass (PassReason reason, bool accurate)
129 int r = av_read_frame (_format_context, &_packet);
131 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
132 has pretty-much succeeded (and hence generated data which should be processed).
133 Hence it makes sense to continue here in that case.
135 if (r < 0 && r != AVERROR_INVALIDDATA) {
136 if (r != AVERROR_EOF) {
137 /* Maybe we should fail here, but for now we'll just finish off instead */
139 av_strerror (r, buf, sizeof(buf));
140 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), &buf[0], r);
147 int const si = _packet.stream_index;
148 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
150 if (_video_stream && si == _video_stream.get() && !video->ignore() && (accurate || reason != PASS_REASON_SUBTITLE)) {
151 decode_video_packet ();
152 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index (_format_context, si)) {
153 decode_subtitle_packet ();
154 } else if (accurate || reason != PASS_REASON_SUBTITLE) {
155 decode_audio_packet ();
158 av_packet_unref (&_packet);
162 /** @param data pointer to array of pointers to buffers.
163 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
165 shared_ptr<AudioBuffers>
166 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
168 DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
170 int const size = av_samples_get_buffer_size (
171 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
174 /* Deinterleave and convert to float */
176 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
177 of the block that do not form a complete sample or frame they will be dropped.
179 int const total_samples = size / bytes_per_audio_sample (stream);
180 int const frames = total_samples / stream->channels();
181 shared_ptr<AudioBuffers> audio (new AudioBuffers (stream->channels(), frames));
183 switch (audio_sample_format (stream)) {
184 case AV_SAMPLE_FMT_U8:
186 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
189 for (int i = 0; i < total_samples; ++i) {
190 audio->data(channel)[sample] = float(*p++) / (1 << 23);
193 if (channel == stream->channels()) {
201 case AV_SAMPLE_FMT_S16:
203 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
206 for (int i = 0; i < total_samples; ++i) {
207 audio->data(channel)[sample] = float(*p++) / (1 << 15);
210 if (channel == stream->channels()) {
218 case AV_SAMPLE_FMT_S16P:
220 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
221 for (int i = 0; i < stream->channels(); ++i) {
222 for (int j = 0; j < frames; ++j) {
223 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
229 case AV_SAMPLE_FMT_S32:
231 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
234 for (int i = 0; i < total_samples; ++i) {
235 audio->data(channel)[sample] = static_cast<float>(*p++) / 2147483648;
238 if (channel == stream->channels()) {
246 case AV_SAMPLE_FMT_S32P:
248 int32_t** p = reinterpret_cast<int32_t **> (_frame->data);
249 for (int i = 0; i < stream->channels(); ++i) {
250 for (int j = 0; j < frames; ++j) {
251 audio->data(i)[j] = static_cast<float>(p[i][j]) / 2147483648;
257 case AV_SAMPLE_FMT_FLT:
259 float* p = reinterpret_cast<float*> (_frame->data[0]);
262 for (int i = 0; i < total_samples; ++i) {
263 audio->data(channel)[sample] = *p++;
266 if (channel == stream->channels()) {
274 case AV_SAMPLE_FMT_FLTP:
276 float** p = reinterpret_cast<float**> (_frame->data);
277 /* Sometimes there aren't as many channels in the _frame as in the stream */
278 for (int i = 0; i < _frame->channels; ++i) {
279 memcpy (audio->data(i), p[i], frames * sizeof(float));
281 for (int i = _frame->channels; i < stream->channels(); ++i) {
282 audio->make_silent (i);
288 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
295 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
297 return stream->stream (_format_context)->codec->sample_fmt;
301 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
303 return av_get_bytes_per_sample (audio_sample_format (stream));
307 FFmpegDecoder::seek (ContentTime time, bool accurate)
310 video->seek (time, accurate);
314 audio->seek (time, accurate);
318 subtitle->seek (time, accurate);
321 /* If we are doing an `accurate' seek, we need to use pre-roll, as
322 we don't really know what the seek will give us.
325 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
328 /* XXX: it seems debatable whether PTS should be used here...
329 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
332 optional<int> stream;
335 stream = _video_stream;
337 shared_ptr<FFmpegAudioStream> s = dynamic_pointer_cast<FFmpegAudioStream> (_ffmpeg_content->audio->stream ());
339 stream = s->index (_format_context);
343 DCPOMATIC_ASSERT (stream);
345 ContentTime u = time - _pts_offset;
346 if (u < ContentTime ()) {
352 u.seconds() / av_q2d (_format_context->streams[stream.get()]->time_base),
356 if (video_codec_context ()) {
357 avcodec_flush_buffers (video_codec_context());
360 /* XXX: should be flushing audio buffers? */
362 if (subtitle_codec_context ()) {
363 avcodec_flush_buffers (subtitle_codec_context ());
368 FFmpegDecoder::decode_audio_packet ()
370 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
374 AVPacket copy_packet = _packet;
376 /* XXX: inefficient */
377 vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
378 vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
379 while (stream != streams.end () && !(*stream)->uses_index (_format_context, copy_packet.stream_index)) {
383 if (stream == streams.end ()) {
384 /* The packet's stream may not be an audio one; just ignore it in this method if so */
388 while (copy_packet.size > 0) {
391 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, ©_packet);
392 if (decode_result < 0) {
393 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
394 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
395 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
396 even in the face of such an error, so I think we should too.
398 Returning from the method here caused mantis #352.
400 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
402 /* Fudge decode_result so that we come out of the while loop when
403 we've processed this data.
405 decode_result = copy_packet.size;
408 if (frame_finished) {
409 ContentTime ct = ContentTime::from_seconds (
410 av_frame_get_best_effort_timestamp (_frame) *
411 av_q2d ((*stream)->stream (_format_context)->time_base))
414 shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
416 if (ct < ContentTime ()) {
417 /* Discard audio data that comes before time 0 */
418 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
419 data->move (remove, 0, data->frames() - remove);
420 data->set_frames (data->frames() - remove);
421 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
424 if (ct < ContentTime()) {
425 LOG_WARNING ("Crazy timestamp %s", to_string (ct));
428 update_position (ct);
430 /* Give this data provided there is some, and its time is sane */
431 if (ct >= ContentTime() && data->frames() > 0) {
432 audio->give (*stream, data, ct);
436 copy_packet.data += decode_result;
437 copy_packet.size -= decode_result;
442 FFmpegDecoder::decode_video_packet ()
444 DCPOMATIC_ASSERT (_video_stream);
447 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
451 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
453 shared_ptr<VideoFilterGraph> graph;
455 list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
456 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
460 if (i == _filter_graphs.end ()) {
461 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
462 graph->setup (_ffmpeg_content->filters ());
463 _filter_graphs.push_back (graph);
464 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
469 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
471 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
473 shared_ptr<Image> image = i->first;
475 if (i->second != AV_NOPTS_VALUE) {
476 double const pts = i->second * av_q2d (_format_context->streams[_video_stream.get()]->time_base) + _pts_offset.seconds ();
478 shared_ptr<ImageProxy> (new RawImageProxy (image)),
479 llrint (pts * _ffmpeg_content->active_video_frame_rate ())
481 update_position (ContentTime::from_seconds (pts));
483 LOG_WARNING_NC ("Dropping frame without PTS");
491 FFmpegDecoder::decode_subtitle_packet ()
495 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
499 if (sub.num_rects <= 0) {
500 /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
501 indicate that the previous subtitle should stop. We can ignore it here.
506 /* Subtitle PTS (within the source, not taking into account any of the
507 source that we may have chopped off for the DCP).
509 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
510 ContentTimePeriod period;
511 period.from = sub_period.from + _pts_offset;
512 update_position (period.from);
514 /* We already know the subtitle period `to' time */
515 period.to = sub_period.to.get() + _pts_offset;
517 /* We have to look up the `to' time in the stream's records */
518 period.to = ffmpeg_content()->subtitle_stream()->find_subtitle_to (subtitle_id (sub));
521 for (unsigned int i = 0; i < sub.num_rects; ++i) {
522 AVSubtitleRect const * rect = sub.rects[i];
524 switch (rect->type) {
527 case SUBTITLE_BITMAP:
528 decode_bitmap_subtitle (rect, period);
531 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
534 decode_ass_subtitle (rect->ass, period);
539 avsubtitle_free (&sub);
542 list<ContentTimePeriod>
543 FFmpegDecoder::image_subtitles_during (ContentTimePeriod p, bool starting) const
545 return _ffmpeg_content->image_subtitles_during (p, starting);
548 list<ContentTimePeriod>
549 FFmpegDecoder::text_subtitles_during (ContentTimePeriod p, bool starting) const
551 return _ffmpeg_content->text_subtitles_during (p, starting);
555 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTimePeriod period)
557 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
558 G, third B, fourth A.
560 shared_ptr<Image> image (new Image (AV_PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
562 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
563 /* Start of the first line in the subtitle */
564 uint8_t* sub_p = rect->pict.data[0];
565 /* sub_p looks up into a BGRA palette which is here
566 (i.e. first byte B, second G, third R, fourth A)
568 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
570 /* Start of the first line in the subtitle */
571 uint8_t* sub_p = rect->data[0];
572 /* sub_p looks up into a BGRA palette which is here
573 (i.e. first byte B, second G, third R, fourth A)
575 uint32_t const * palette = (uint32_t *) rect->data[1];
577 /* And the stream has a map of those palette colours to colours
578 chosen by the user; created a `mapped' palette from those settings.
580 map<RGBA, RGBA> colour_map = ffmpeg_content()->subtitle_stream()->colours ();
581 vector<RGBA> mapped_palette (rect->nb_colors);
582 for (int i = 0; i < rect->nb_colors; ++i) {
583 RGBA c ((palette[i] & 0xff0000) >> 16, (palette[i] & 0xff00) >> 8, palette[i] & 0xff, (palette[i] & 0xff000000) >> 24);
584 map<RGBA, RGBA>::const_iterator j = colour_map.find (c);
585 if (j != colour_map.end ()) {
586 mapped_palette[i] = j->second;
588 /* This colour was not found in the FFmpegSubtitleStream's colour map; probably because
589 it is from a project that was created before this stuff was added. Just use the
590 colour straight from the original palette.
592 mapped_palette[i] = c;
596 /* Start of the output data */
597 uint32_t* out_p = (uint32_t *) image->data()[0];
599 for (int y = 0; y < rect->h; ++y) {
600 uint8_t* sub_line_p = sub_p;
601 uint32_t* out_line_p = out_p;
602 for (int x = 0; x < rect->w; ++x) {
603 RGBA const p = mapped_palette[*sub_line_p++];
604 /* XXX: this seems to be wrong to me (isn't the output image RGBA?) but it looks right on screen */
605 *out_line_p++ = (p.a << 24) | (p.r << 16) | (p.g << 8) | p.b;
607 #ifdef DCPOMATIC_HAVE_AVSUBTITLERECT_PICT
608 sub_p += rect->pict.linesize[0];
610 sub_p += rect->linesize[0];
612 out_p += image->stride()[0] / sizeof (uint32_t);
615 int const target_width = subtitle_codec_context()->width;
616 int const target_height = subtitle_codec_context()->height;
617 dcpomatic::Rect<double> const scaled_rect (
618 static_cast<double> (rect->x) / target_width,
619 static_cast<double> (rect->y) / target_height,
620 static_cast<double> (rect->w) / target_width,
621 static_cast<double> (rect->h) / target_height
624 subtitle->give_image (period, image, scaled_rect);
628 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTimePeriod period)
630 /* We have no styles and no Format: line, so I'm assuming that FFmpeg
631 produces a single format of Dialogue: lines...
635 split (bits, ass, is_any_of (","));
636 if (bits.size() < 10) {
640 sub::RawSubtitle base;
641 list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (base, bits[9]);
643 BOOST_FOREACH (sub::Subtitle const & i, sub::collect<list<sub::Subtitle> > (raw)) {
644 subtitle->give_text (period, i);
649 FFmpegDecoder::update_position (ContentTime p)
651 /* _position should err on the side of being too big, as then there is less
652 chance that we will erroneously decide not to seek when _position > request.
657 _position = max (*_position, p);