src/lib/ffmpeg_decoder.cc

   1 /*
   2     Copyright (C) 2012-2016 Carl Hetherington <cth@carlh.net>
   3
   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.
   8
   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.
  13
  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.
  17
  18 */
  19
  20 /** @file  src/ffmpeg_decoder.cc
  21  *  @brief A decoder using FFmpeg to decode content.
  22  */
  23
  24 #include "filter.h"
  25 #include "exceptions.h"
  26 #include "image.h"
  27 #include "util.h"
  28 #include "log.h"
  29 #include "ffmpeg_decoder.h"
  30 #include "ffmpeg_audio_stream.h"
  31 #include "ffmpeg_subtitle_stream.h"
  32 #include "video_filter_graph.h"
  33 #include "audio_buffers.h"
  34 #include "ffmpeg_content.h"
  35 #include "raw_image_proxy.h"
  36 #include "film.h"
  37 #include "md5_digester.h"
  38 #include "compose.hpp"
  39 #include <dcp/subtitle_string.h>
  40 #include <sub/ssa_reader.h>
  41 #include <sub/subtitle.h>
  42 #include <sub/collect.h>
  43 extern "C" {
  44 #include <libavcodec/avcodec.h>
  45 #include <libavformat/avformat.h>
  46 }
  47 #include <boost/foreach.hpp>
  48 #include <boost/algorithm/string.hpp>
  49 #include <vector>
  50 #include <iomanip>
  51 #include <iostream>
  52 #include <stdint.h>
  53
  54 #include "i18n.h"
  55
  56 #define LOG_GENERAL(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_GENERAL);
  57 #define LOG_ERROR(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_ERROR);
  58 #define LOG_WARNING_NC(...) _log->log (__VA_ARGS__, LogEntry::TYPE_WARNING);
  59 #define LOG_WARNING(...) _log->log (String::compose (__VA_ARGS__), LogEntry::TYPE_WARNING);
  60
  61 using std::cout;
  62 using std::string;
  63 using std::vector;
  64 using std::list;
  65 using std::min;
  66 using std::pair;
  67 using std::max;
  68 using boost::shared_ptr;
  69 using boost::is_any_of;
  70 using boost::split;
  71 using dcp::Size;
  72
  73 FFmpegDecoder::FFmpegDecoder (shared_ptr<const FFmpegContent> c, shared_ptr<Log> log, bool fast)
  74         : VideoDecoder (c)
  75         , AudioDecoder (c, fast)
  76         , SubtitleDecoder (c)
  77         , FFmpeg (c)
  78         , _log (log)
  79         , _pts_offset (pts_offset (c->ffmpeg_audio_streams(), c->first_video(), c->video_frame_rate()))
  80 {
  81
  82 }
  83
  84 void
  85 FFmpegDecoder::flush ()
  86 {
  87         /* Get any remaining frames */
  88
  89         _packet.data = 0;
  90         _packet.size = 0;
  91
  92         /* XXX: should we reset _packet.data and size after each *_decode_* call? */
  93
  94         while (decode_video_packet ()) {}
  95
  96         decode_audio_packet ();
  97         AudioDecoder::flush ();
  98 }
  99
 100 bool
 101 FFmpegDecoder::pass (PassReason reason, bool accurate)
 102 {
 103         int r = av_read_frame (_format_context, &_packet);
 104
 105         /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
 106            has pretty-much succeeded (and hence generated data which should be processed).
 107            Hence it makes sense to continue here in that case.
 108         */
 109         if (r < 0 && r != AVERROR_INVALIDDATA) {
 110                 if (r != AVERROR_EOF) {
 111                         /* Maybe we should fail here, but for now we'll just finish off instead */
 112                         char buf[256];
 113                         av_strerror (r, buf, sizeof(buf));
 114                         LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), buf, r);
 115                 }
 116
 117                 flush ();
 118                 return true;
 119         }
 120
 121         int const si = _packet.stream_index;
 122         shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
 123
 124         if (si == _video_stream && !_ignore_video && (accurate || reason != PASS_REASON_SUBTITLE)) {
 125                 decode_video_packet ();
 126         } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index (_format_context, si)) {
 127                 decode_subtitle_packet ();
 128         } else if (accurate || reason != PASS_REASON_SUBTITLE) {
 129                 decode_audio_packet ();
 130         }
 131
 132         av_free_packet (&_packet);
 133         return false;
 134 }
 135
 136 /** @param data pointer to array of pointers to buffers.
 137  *  Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
 138  */
 139 shared_ptr<AudioBuffers>
 140 FFmpegDecoder::deinterleave_audio (shared_ptr<FFmpegAudioStream> stream) const
 141 {
 142         DCPOMATIC_ASSERT (bytes_per_audio_sample (stream));
 143
 144         int const size = av_samples_get_buffer_size (
 145                 0, stream->stream(_format_context)->codec->channels, _frame->nb_samples, audio_sample_format (stream), 1
 146                 );
 147
 148         /* Deinterleave and convert to float */
 149
 150         /* total_samples and frames will be rounded down here, so if there are stray samples at the end
 151            of the block that do not form a complete sample or frame they will be dropped.
 152         */
 153         int const total_samples = size / bytes_per_audio_sample (stream);
 154         int const frames = total_samples / stream->channels();
 155         shared_ptr<AudioBuffers> audio (new AudioBuffers (stream->channels(), frames));
 156
 157         switch (audio_sample_format (stream)) {
 158         case AV_SAMPLE_FMT_U8:
 159         {
 160                 uint8_t* p = reinterpret_cast<uint8_t *> (_frame->data[0]);
 161                 int sample = 0;
 162                 int channel = 0;
 163                 for (int i = 0; i < total_samples; ++i) {
 164                         audio->data(channel)[sample] = float(*p++) / (1 << 23);
 165
 166                         ++channel;
 167                         if (channel == stream->channels()) {
 168                                 channel = 0;
 169                                 ++sample;
 170                         }
 171                 }
 172         }
 173         break;
 174
 175         case AV_SAMPLE_FMT_S16:
 176         {
 177                 int16_t* p = reinterpret_cast<int16_t *> (_frame->data[0]);
 178                 int sample = 0;
 179                 int channel = 0;
 180                 for (int i = 0; i < total_samples; ++i) {
 181                         audio->data(channel)[sample] = float(*p++) / (1 << 15);
 182
 183                         ++channel;
 184                         if (channel == stream->channels()) {
 185                                 channel = 0;
 186                                 ++sample;
 187                         }
 188                 }
 189         }
 190         break;
 191
 192         case AV_SAMPLE_FMT_S16P:
 193         {
 194                 int16_t** p = reinterpret_cast<int16_t **> (_frame->data);
 195                 for (int i = 0; i < stream->channels(); ++i) {
 196                         for (int j = 0; j < frames; ++j) {
 197                                 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
 198                         }
 199                 }
 200         }
 201         break;
 202
 203         case AV_SAMPLE_FMT_S32:
 204         {
 205                 int32_t* p = reinterpret_cast<int32_t *> (_frame->data[0]);
 206                 int sample = 0;
 207                 int channel = 0;
 208                 for (int i = 0; i < total_samples; ++i) {
 209                         audio->data(channel)[sample] = static_cast<float>(*p++) / (1 << 31);
 210
 211                         ++channel;
 212                         if (channel == stream->channels()) {
 213                                 channel = 0;
 214                                 ++sample;
 215                         }
 216                 }
 217         }
 218         break;
 219
 220         case AV_SAMPLE_FMT_FLT:
 221         {
 222                 float* p = reinterpret_cast<float*> (_frame->data[0]);
 223                 int sample = 0;
 224                 int channel = 0;
 225                 for (int i = 0; i < total_samples; ++i) {
 226                         audio->data(channel)[sample] = *p++;
 227
 228                         ++channel;
 229                         if (channel == stream->channels()) {
 230                                 channel = 0;
 231                                 ++sample;
 232                         }
 233                 }
 234         }
 235         break;
 236
 237         case AV_SAMPLE_FMT_FLTP:
 238         {
 239                 float** p = reinterpret_cast<float**> (_frame->data);
 240                 /* Sometimes there aren't as many channels in the _frame as in the stream */
 241                 for (int i = 0; i < _frame->channels; ++i) {
 242                         memcpy (audio->data(i), p[i], frames * sizeof(float));
 243                 }
 244                 for (int i = _frame->channels; i < stream->channels(); ++i) {
 245                         audio->make_silent (i);
 246                 }
 247         }
 248         break;
 249
 250         default:
 251                 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format (stream))));
 252         }
 253
 254         return audio;
 255 }
 256
 257 AVSampleFormat
 258 FFmpegDecoder::audio_sample_format (shared_ptr<FFmpegAudioStream> stream) const
 259 {
 260         return stream->stream (_format_context)->codec->sample_fmt;
 261 }
 262
 263 int
 264 FFmpegDecoder::bytes_per_audio_sample (shared_ptr<FFmpegAudioStream> stream) const
 265 {
 266         return av_get_bytes_per_sample (audio_sample_format (stream));
 267 }
 268
 269 void
 270 FFmpegDecoder::seek (ContentTime time, bool accurate)
 271 {
 272         VideoDecoder::seek (time, accurate);
 273         AudioDecoder::seek (time, accurate);
 274         SubtitleDecoder::seek (time, accurate);
 275
 276         /* If we are doing an `accurate' seek, we need to use pre-roll, as
 277            we don't really know what the seek will give us.
 278         */
 279
 280         ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
 281         time -= pre_roll;
 282
 283         /* XXX: it seems debatable whether PTS should be used here...
 284            http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
 285         */
 286
 287         ContentTime u = time - _pts_offset;
 288         if (u < ContentTime ()) {
 289                 u = ContentTime ();
 290         }
 291         av_seek_frame (_format_context, _video_stream, u.seconds() / av_q2d (_format_context->streams[_video_stream]->time_base), AVSEEK_FLAG_BACKWARD);
 292
 293         avcodec_flush_buffers (video_codec_context());
 294
 295         /* XXX: should be flushing audio buffers? */
 296
 297         if (subtitle_codec_context ()) {
 298                 avcodec_flush_buffers (subtitle_codec_context ());
 299         }
 300 }
 301
 302 void
 303 FFmpegDecoder::decode_audio_packet ()
 304 {
 305         /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
 306            several times.
 307         */
 308
 309         AVPacket copy_packet = _packet;
 310
 311         /* XXX: inefficient */
 312         vector<shared_ptr<FFmpegAudioStream> > streams = ffmpeg_content()->ffmpeg_audio_streams ();
 313         vector<shared_ptr<FFmpegAudioStream> >::const_iterator stream = streams.begin ();
 314         while (stream != streams.end () && !(*stream)->uses_index (_format_context, copy_packet.stream_index)) {
 315                 ++stream;
 316         }
 317
 318         if (stream == streams.end ()) {
 319                 /* The packet's stream may not be an audio one; just ignore it in this method if so */
 320                 return;
 321         }
 322
 323         while (copy_packet.size > 0) {
 324
 325                 int frame_finished;
 326                 int decode_result = avcodec_decode_audio4 ((*stream)->stream (_format_context)->codec, _frame, &frame_finished, &copy_packet);
 327                 if (decode_result < 0) {
 328                         /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
 329                            some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
 330                            if it overreads the auxiliary data.  ffplay carries on if frame_finished is true,
 331                            even in the face of such an error, so I think we should too.
 332
 333                            Returning from the method here caused mantis #352.
 334                         */
 335                         LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
 336
 337                         /* Fudge decode_result so that we come out of the while loop when
 338                            we've processed this data.
 339                         */
 340                         decode_result = copy_packet.size;
 341                 }
 342
 343                 if (frame_finished) {
 344                         ContentTime ct = ContentTime::from_seconds (
 345                                 av_frame_get_best_effort_timestamp (_frame) *
 346                                 av_q2d ((*stream)->stream (_format_context)->time_base))
 347                                 + _pts_offset;
 348
 349                         shared_ptr<AudioBuffers> data = deinterleave_audio (*stream);
 350
 351                         if (ct < ContentTime ()) {
 352                                 /* Discard audio data that comes before time 0 */
 353                                 Frame const remove = min (int64_t (data->frames()), (-ct).frames_ceil(double((*stream)->frame_rate ())));
 354                                 data->move (remove, 0, data->frames() - remove);
 355                                 data->set_frames (data->frames() - remove);
 356                                 ct += ContentTime::from_frames (remove, (*stream)->frame_rate ());
 357                         }
 358
 359                         if (data->frames() > 0) {
 360                                 audio (*stream, data, ct);
 361                         }
 362                 }
 363
 364                 copy_packet.data += decode_result;
 365                 copy_packet.size -= decode_result;
 366         }
 367 }
 368
 369 bool
 370 FFmpegDecoder::decode_video_packet ()
 371 {
 372         int frame_finished;
 373         if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
 374                 return false;
 375         }
 376
 377         boost::mutex::scoped_lock lm (_filter_graphs_mutex);
 378
 379         shared_ptr<VideoFilterGraph> graph;
 380
 381         list<shared_ptr<VideoFilterGraph> >::iterator i = _filter_graphs.begin();
 382         while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
 383                 ++i;
 384         }
 385
 386         if (i == _filter_graphs.end ()) {
 387                 graph.reset (new VideoFilterGraph (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
 388                 graph->setup (_ffmpeg_content->filters ());
 389                 _filter_graphs.push_back (graph);
 390                 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
 391         } else {
 392                 graph = *i;
 393         }
 394
 395         list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
 396
 397         for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
 398
 399                 shared_ptr<Image> image = i->first;
 400
 401                 if (i->second != AV_NOPTS_VALUE) {
 402                         double const pts = i->second * av_q2d (_format_context->streams[_video_stream]->time_base) + _pts_offset.seconds ();
 403                         video (
 404                                 shared_ptr<ImageProxy> (new RawImageProxy (image)),
 405                                 llrint (pts * _ffmpeg_content->video_frame_rate ())
 406                                 );
 407                 } else {
 408                         LOG_WARNING_NC ("Dropping frame without PTS");
 409                 }
 410         }
 411
 412         return true;
 413 }
 414
 415 void
 416 FFmpegDecoder::decode_subtitle_packet ()
 417 {
 418         int got_subtitle;
 419         AVSubtitle sub;
 420         if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
 421                 return;
 422         }
 423
 424         if (sub.num_rects <= 0) {
 425                 /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
 426                    indicate that the previous subtitle should stop.  We can ignore it here.
 427                 */
 428                 return;
 429         }
 430
 431         /* Subtitle PTS (within the source, not taking into account any of the
 432            source that we may have chopped off for the DCP).
 433         */
 434         FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
 435         ContentTimePeriod period;
 436         period.from = sub_period.from + _pts_offset;
 437         if (sub_period.to) {
 438                 /* We already know the subtitle period `to' time */
 439                 period.to = sub_period.to.get() + _pts_offset;
 440         } else {
 441                 /* We have to look up the `to' time in the stream's records */
 442                 period.to = ffmpeg_content()->subtitle_stream()->find_subtitle_to (subtitle_id (sub));
 443         }
 444
 445         for (unsigned int i = 0; i < sub.num_rects; ++i) {
 446                 AVSubtitleRect const * rect = sub.rects[i];
 447
 448                 switch (rect->type) {
 449                 case SUBTITLE_NONE:
 450                         break;
 451                 case SUBTITLE_BITMAP:
 452                         decode_bitmap_subtitle (rect, period);
 453                         break;
 454                 case SUBTITLE_TEXT:
 455                         cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
 456                         break;
 457                 case SUBTITLE_ASS:
 458                         decode_ass_subtitle (rect->ass, period);
 459                         break;
 460                 }
 461         }
 462
 463         avsubtitle_free (&sub);
 464 }
 465
 466 list<ContentTimePeriod>
 467 FFmpegDecoder::image_subtitles_during (ContentTimePeriod p, bool starting) const
 468 {
 469         return _ffmpeg_content->image_subtitles_during (p, starting);
 470 }
 471
 472 list<ContentTimePeriod>
 473 FFmpegDecoder::text_subtitles_during (ContentTimePeriod p, bool starting) const
 474 {
 475         return _ffmpeg_content->text_subtitles_during (p, starting);
 476 }
 477
 478 void
 479 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTimePeriod period)
 480 {
 481         /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
 482            G, third B, fourth A.
 483         */
 484         shared_ptr<Image> image (new Image (AV_PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
 485
 486         /* Start of the first line in the subtitle */
 487         uint8_t* sub_p = rect->pict.data[0];
 488         /* sub_p looks up into a BGRA palette which is here
 489            (i.e. first byte B, second G, third R, fourth A)
 490         */
 491         uint32_t const * palette = (uint32_t *) rect->pict.data[1];
 492         /* Start of the output data */
 493         uint32_t* out_p = (uint32_t *) image->data()[0];
 494
 495         for (int y = 0; y < rect->h; ++y) {
 496                 uint8_t* sub_line_p = sub_p;
 497                 uint32_t* out_line_p = out_p;
 498                 for (int x = 0; x < rect->w; ++x) {
 499                         uint32_t const p = palette[*sub_line_p++];
 500                         *out_line_p++ = ((p & 0xff) << 16) | (p & 0xff00) | ((p & 0xff0000) >> 16) | (p & 0xff000000);
 501                 }
 502                 sub_p += rect->pict.linesize[0];
 503                 out_p += image->stride()[0] / sizeof (uint32_t);
 504         }
 505
 506         dcp::Size const vs = _ffmpeg_content->video_size ();
 507         dcpomatic::Rect<double> const scaled_rect (
 508                 static_cast<double> (rect->x) / vs.width,
 509                 static_cast<double> (rect->y) / vs.height,
 510                 static_cast<double> (rect->w) / vs.width,
 511                 static_cast<double> (rect->h) / vs.height
 512                 );
 513
 514         image_subtitle (period, image, scaled_rect);
 515 }
 516
 517 void
 518 FFmpegDecoder::decode_ass_subtitle (string ass, ContentTimePeriod period)
 519 {
 520         /* We have no styles and no Format: line, so I'm assuming that FFmpeg
 521            produces a single format of Dialogue: lines...
 522         */
 523
 524         vector<string> bits;
 525         split (bits, ass, is_any_of (","));
 526         if (bits.size() < 10) {
 527                 return;
 528         }
 529
 530         sub::RawSubtitle base;
 531         list<sub::RawSubtitle> raw = sub::SSAReader::parse_line (base, bits[9]);
 532         list<sub::Subtitle> subs = sub::collect<list<sub::Subtitle> > (raw);
 533
 534         /* XXX: lots of this is copied from TextSubtitle; there should probably be some sharing */
 535
 536         /* Highest line index in this subtitle */
 537         int highest = 0;
 538         BOOST_FOREACH (sub::Subtitle i, subs) {
 539                 BOOST_FOREACH (sub::Line j, i.lines) {
 540                         DCPOMATIC_ASSERT (j.vertical_position.reference && j.vertical_position.reference.get() == sub::TOP_OF_SUBTITLE);
 541                         DCPOMATIC_ASSERT (j.vertical_position.line);
 542                         highest = max (highest, j.vertical_position.line.get());
 543                 }
 544         }
 545
 546         list<dcp::SubtitleString> ss;
 547
 548         BOOST_FOREACH (sub::Subtitle i, sub::collect<list<sub::Subtitle> > (sub::SSAReader::parse_line (base, bits[9]))) {
 549                 BOOST_FOREACH (sub::Line j, i.lines) {
 550                         BOOST_FOREACH (sub::Block k, j.blocks) {
 551                                 ss.push_back (
 552                                         dcp::SubtitleString (
 553                                                 boost::optional<string> (),
 554                                                 k.italic,
 555                                                 dcp::Colour (255, 255, 255),
 556                                                 60,
 557                                                 1,
 558                                                 dcp::Time (i.from.seconds(), 1000),
 559                                                 dcp::Time (i.to.seconds(), 1000),
 560                                                 0,
 561                                                 dcp::HALIGN_CENTER,
 562                                                 /* This 1.015 is an arbitrary value to lift the bottom sub off the bottom
 563                                                    of the screen a bit to a pleasing degree.
 564                                                 */
 565                                                 1.015 - ((1 + highest - j.vertical_position.line.get()) * 1.5 / 22),
 566                                                 dcp::VALIGN_TOP,
 567                                                 k.text,
 568                                                 static_cast<dcp::Effect> (0),
 569                                                 dcp::Colour (255, 255, 255),
 570                                                 dcp::Time (),
 571                                                 dcp::Time ()
 572                                                 )
 573                                         );
 574                         }
 575                 }
 576         }
 577
 578         text_subtitle (period, ss);
 579 }