2 Copyright (C) 2012-2014 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.
31 #include <libavcodec/avcodec.h>
32 #include <libavformat/avformat.h>
35 #include "exceptions.h"
39 #include "ffmpeg_decoder.h"
40 #include "ffmpeg_audio_stream.h"
41 #include "ffmpeg_subtitle_stream.h"
42 #include "filter_graph.h"
43 #include "audio_buffers.h"
44 #include "ffmpeg_content.h"
45 #include "raw_image_proxy.h"
51 #define LOG_GENERAL(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_GENERAL);
52 #define LOG_ERROR(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_ERROR);
53 #define LOG_WARNING_NC(...) _video_content->film()->log()->log (__VA_ARGS__, Log::TYPE_WARNING);
54 #define LOG_WARNING(...) _video_content->film()->log()->log (String::compose (__VA_ARGS__), Log::TYPE_WARNING);
63 using boost::shared_ptr;
64 using boost::optional;
65 using boost::dynamic_pointer_cast;
68 FFmpegDecoder::FFmpegDecoder (shared_ptr<const FFmpegContent> c, shared_ptr<Log> log)
75 /* Audio and video frame PTS values may not start with 0. We want
76 to fiddle them so that:
78 1. One of them starts at time 0.
79 2. The first video PTS value ends up on a frame boundary.
81 Then we remove big initial gaps in PTS and we allow our
82 insertion of black frames to work.
84 We will do pts_to_use = pts_from_ffmpeg + pts_offset;
87 bool const have_video = c->first_video();
88 bool const have_audio = c->audio_stream () && c->audio_stream()->first_audio;
90 /* First, make one of them start at 0 */
92 if (have_audio && have_video) {
93 _pts_offset = - min (c->first_video().get(), c->audio_stream()->first_audio.get());
94 } else if (have_video) {
95 _pts_offset = - c->first_video().get();
96 } else if (have_audio) {
97 _pts_offset = - c->audio_stream()->first_audio.get();
100 /* If _pts_offset is positive we would be pushing things from a -ve PTS to be played.
101 I don't think we ever want to do that, as it seems things at -ve PTS are not meant
102 to be seen (use for alignment bars etc.); see mantis #418.
104 if (_pts_offset > ContentTime ()) {
105 _pts_offset = ContentTime ();
108 /* Now adjust both so that the video pts starts on a frame */
109 if (have_video && have_audio) {
110 ContentTime first_video = c->first_video().get() + _pts_offset;
111 ContentTime const old_first_video = first_video;
112 _pts_offset += first_video.round_up (c->video_frame_rate ()) - old_first_video;
117 FFmpegDecoder::flush ()
119 /* Get any remaining frames */
124 /* XXX: should we reset _packet.data and size after each *_decode_* call? */
126 while (decode_video_packet ()) {}
128 if (_ffmpeg_content->audio_stream()) {
129 decode_audio_packet ();
130 AudioDecoder::flush ();
135 FFmpegDecoder::pass (PassReason reason)
137 int r = av_read_frame (_format_context, &_packet);
139 /* AVERROR_INVALIDDATA can apparently be returned sometimes even when av_read_frame
140 has pretty-much succeeded (and hence generated data which should be processed).
141 Hence it makes sense to continue here in that case.
143 if (r < 0 && r != AVERROR_INVALIDDATA) {
144 if (r != AVERROR_EOF) {
145 /* Maybe we should fail here, but for now we'll just finish off instead */
147 av_strerror (r, buf, sizeof(buf));
148 LOG_ERROR (N_("error on av_read_frame (%1) (%2)"), buf, r);
155 int const si = _packet.stream_index;
156 shared_ptr<const FFmpegContent> fc = _ffmpeg_content;
158 if (si == _video_stream && !_ignore_video && reason != PASS_REASON_SUBTITLE) {
159 decode_video_packet ();
160 } else if (fc->audio_stream() && fc->audio_stream()->uses_index (_format_context, si) && reason != PASS_REASON_SUBTITLE) {
161 decode_audio_packet ();
162 } else if (fc->subtitle_stream() && fc->subtitle_stream()->uses_index (_format_context, si)) {
163 decode_subtitle_packet ();
166 av_free_packet (&_packet);
170 /** @param data pointer to array of pointers to buffers.
171 * Only the first buffer will be used for non-planar data, otherwise there will be one per channel.
173 shared_ptr<AudioBuffers>
174 FFmpegDecoder::deinterleave_audio (uint8_t** data, int size)
176 DCPOMATIC_ASSERT (_ffmpeg_content->audio_channels());
177 DCPOMATIC_ASSERT (bytes_per_audio_sample());
179 /* Deinterleave and convert to float */
181 /* total_samples and frames will be rounded down here, so if there are stray samples at the end
182 of the block that do not form a complete sample or frame they will be dropped.
184 int const total_samples = size / bytes_per_audio_sample();
185 int const frames = total_samples / _ffmpeg_content->audio_channels();
186 shared_ptr<AudioBuffers> audio (new AudioBuffers (_ffmpeg_content->audio_channels(), frames));
188 switch (audio_sample_format()) {
189 case AV_SAMPLE_FMT_U8:
191 uint8_t* p = reinterpret_cast<uint8_t *> (data[0]);
194 for (int i = 0; i < total_samples; ++i) {
195 audio->data(channel)[sample] = float(*p++) / (1 << 23);
198 if (channel == _ffmpeg_content->audio_channels()) {
206 case AV_SAMPLE_FMT_S16:
208 int16_t* p = reinterpret_cast<int16_t *> (data[0]);
211 for (int i = 0; i < total_samples; ++i) {
212 audio->data(channel)[sample] = float(*p++) / (1 << 15);
215 if (channel == _ffmpeg_content->audio_channels()) {
223 case AV_SAMPLE_FMT_S16P:
225 int16_t** p = reinterpret_cast<int16_t **> (data);
226 for (int i = 0; i < _ffmpeg_content->audio_channels(); ++i) {
227 for (int j = 0; j < frames; ++j) {
228 audio->data(i)[j] = static_cast<float>(p[i][j]) / (1 << 15);
234 case AV_SAMPLE_FMT_S32:
236 int32_t* p = reinterpret_cast<int32_t *> (data[0]);
239 for (int i = 0; i < total_samples; ++i) {
240 audio->data(channel)[sample] = static_cast<float>(*p++) / (1 << 31);
243 if (channel == _ffmpeg_content->audio_channels()) {
251 case AV_SAMPLE_FMT_FLT:
253 float* p = reinterpret_cast<float*> (data[0]);
256 for (int i = 0; i < total_samples; ++i) {
257 audio->data(channel)[sample] = *p++;
260 if (channel == _ffmpeg_content->audio_channels()) {
268 case AV_SAMPLE_FMT_FLTP:
270 float** p = reinterpret_cast<float**> (data);
271 for (int i = 0; i < _ffmpeg_content->audio_channels(); ++i) {
272 memcpy (audio->data(i), p[i], frames * sizeof(float));
278 throw DecodeError (String::compose (_("Unrecognised audio sample format (%1)"), static_cast<int> (audio_sample_format())));
285 FFmpegDecoder::audio_sample_format () const
287 if (!_ffmpeg_content->audio_stream()) {
288 return (AVSampleFormat) 0;
291 return audio_codec_context()->sample_fmt;
295 FFmpegDecoder::bytes_per_audio_sample () const
297 return av_get_bytes_per_sample (audio_sample_format ());
301 FFmpegDecoder::seek (ContentTime time, bool accurate)
303 VideoDecoder::seek (time, accurate);
304 AudioDecoder::seek (time, accurate);
305 SubtitleDecoder::seek (time, accurate);
307 /* If we are doing an `accurate' seek, we need to use pre-roll, as
308 we don't really know what the seek will give us.
311 ContentTime pre_roll = accurate ? ContentTime::from_seconds (2) : ContentTime (0);
314 /* XXX: it seems debatable whether PTS should be used here...
315 http://www.mjbshaw.com/2012/04/seeking-in-ffmpeg-know-your-timestamp.html
318 ContentTime const u = time - _pts_offset;
319 av_seek_frame (_format_context, _video_stream, u.seconds() / av_q2d (_format_context->streams[_video_stream]->time_base), 0);
321 avcodec_flush_buffers (video_codec_context());
322 if (audio_codec_context ()) {
323 avcodec_flush_buffers (audio_codec_context ());
325 if (subtitle_codec_context ()) {
326 avcodec_flush_buffers (subtitle_codec_context ());
331 FFmpegDecoder::decode_audio_packet ()
333 /* Audio packets can contain multiple frames, so we may have to call avcodec_decode_audio4
337 AVPacket copy_packet = _packet;
339 while (copy_packet.size > 0) {
342 int decode_result = avcodec_decode_audio4 (audio_codec_context(), _frame, &frame_finished, ©_packet);
343 if (decode_result < 0) {
344 /* avcodec_decode_audio4 can sometimes return an error even though it has decoded
345 some valid data; for example dca_subframe_footer can return AVERROR_INVALIDDATA
346 if it overreads the auxiliary data. ffplay carries on if frame_finished is true,
347 even in the face of such an error, so I think we should too.
349 Returning from the method here caused mantis #352.
351 LOG_WARNING ("avcodec_decode_audio4 failed (%1)", decode_result);
353 /* Fudge decode_result so that we come out of the while loop when
354 we've processed this data.
356 decode_result = copy_packet.size;
359 if (frame_finished) {
360 ContentTime const ct = ContentTime::from_seconds (
361 av_frame_get_best_effort_timestamp (_frame) *
362 av_q2d (_ffmpeg_content->audio_stream()->stream (_format_context)->time_base))
365 int const data_size = av_samples_get_buffer_size (
366 0, audio_codec_context()->channels, _frame->nb_samples, audio_sample_format (), 1
369 audio (deinterleave_audio (_frame->data, data_size), ct);
372 copy_packet.data += decode_result;
373 copy_packet.size -= decode_result;
378 FFmpegDecoder::decode_video_packet ()
381 if (avcodec_decode_video2 (video_codec_context(), _frame, &frame_finished, &_packet) < 0 || !frame_finished) {
385 boost::mutex::scoped_lock lm (_filter_graphs_mutex);
387 shared_ptr<FilterGraph> graph;
389 list<shared_ptr<FilterGraph> >::iterator i = _filter_graphs.begin();
390 while (i != _filter_graphs.end() && !(*i)->can_process (dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format)) {
394 if (i == _filter_graphs.end ()) {
395 graph.reset (new FilterGraph (_ffmpeg_content, dcp::Size (_frame->width, _frame->height), (AVPixelFormat) _frame->format));
396 _filter_graphs.push_back (graph);
397 LOG_GENERAL (N_("New graph for %1x%2, pixel format %3"), _frame->width, _frame->height, _frame->format);
402 list<pair<shared_ptr<Image>, int64_t> > images = graph->process (_frame);
404 for (list<pair<shared_ptr<Image>, int64_t> >::iterator i = images.begin(); i != images.end(); ++i) {
406 shared_ptr<Image> image = i->first;
408 if (i->second != AV_NOPTS_VALUE) {
409 double const pts = i->second * av_q2d (_format_context->streams[_video_stream]->time_base) + _pts_offset.seconds ();
411 shared_ptr<ImageProxy> (new RawImageProxy (image)),
412 rint (pts * _ffmpeg_content->video_frame_rate ())
415 LOG_WARNING_NC ("Dropping frame without PTS");
423 FFmpegDecoder::decode_subtitle_packet ()
427 if (avcodec_decode_subtitle2 (subtitle_codec_context(), &sub, &got_subtitle, &_packet) < 0 || !got_subtitle) {
431 if (sub.num_rects <= 0) {
432 /* Sometimes we get an empty AVSubtitle, which is used by some codecs to
433 indicate that the previous subtitle should stop. We can ignore it here.
436 } else if (sub.num_rects > 1) {
437 throw DecodeError (_("multi-part subtitles not yet supported"));
440 /* Subtitle PTS (within the source, not taking into account any of the
441 source that we may have chopped off for the DCP).
443 FFmpegSubtitlePeriod sub_period = subtitle_period (sub);
444 ContentTimePeriod period;
445 period.from = sub_period.from + _pts_offset;
447 /* We already know the subtitle period `to' time */
448 period.to = sub_period.to.get() + _pts_offset;
450 /* We have to look up the `to' time in the stream's records */
451 period.to = ffmpeg_content()->subtitle_stream()->find_subtitle_to (sub_period.from);
454 AVSubtitleRect const * rect = sub.rects[0];
456 switch (rect->type) {
459 case SUBTITLE_BITMAP:
460 decode_bitmap_subtitle (rect, period);
463 cout << "XXX: SUBTITLE_TEXT " << rect->text << "\n";
466 cout << "XXX: SUBTITLE_ASS " << rect->ass << "\n";
470 avsubtitle_free (&sub);
473 list<ContentTimePeriod>
474 FFmpegDecoder::image_subtitles_during (ContentTimePeriod p, bool starting) const
476 return _ffmpeg_content->subtitles_during (p, starting);
479 list<ContentTimePeriod>
480 FFmpegDecoder::text_subtitles_during (ContentTimePeriod, bool) const
482 return list<ContentTimePeriod> ();
486 FFmpegDecoder::decode_bitmap_subtitle (AVSubtitleRect const * rect, ContentTimePeriod period)
488 /* Note RGBA is expressed little-endian, so the first byte in the word is R, second
489 G, third B, fourth A.
491 shared_ptr<Image> image (new Image (PIX_FMT_RGBA, dcp::Size (rect->w, rect->h), true));
493 /* Start of the first line in the subtitle */
494 uint8_t* sub_p = rect->pict.data[0];
495 /* sub_p looks up into a BGRA palette which is here
496 (i.e. first byte B, second G, third R, fourth A)
498 uint32_t const * palette = (uint32_t *) rect->pict.data[1];
499 /* Start of the output data */
500 uint32_t* out_p = (uint32_t *) image->data()[0];
502 for (int y = 0; y < rect->h; ++y) {
503 uint8_t* sub_line_p = sub_p;
504 uint32_t* out_line_p = out_p;
505 for (int x = 0; x < rect->w; ++x) {
506 uint32_t const p = palette[*sub_line_p++];
507 *out_line_p++ = ((p & 0xff) << 16) | (p & 0xff00) | ((p & 0xff0000) >> 16) | (p & 0xff000000);
509 sub_p += rect->pict.linesize[0];
510 out_p += image->stride()[0] / sizeof (uint32_t);
513 dcp::Size const vs = _ffmpeg_content->video_size ();
514 dcpomatic::Rect<double> const scaled_rect (
515 static_cast<double> (rect->x) / vs.width,
516 static_cast<double> (rect->y) / vs.height,
517 static_cast<double> (rect->w) / vs.width,
518 static_cast<double> (rect->h) / vs.height
521 image_subtitle (period, image, scaled_rect);