1 /* $Id: decavcodec.c,v 1.6 2005/03/06 04:08:54 titer Exp $
3 This file is part of the HandBrake source code.
4 Homepage: <http://handbrake.fr/>.
5 It may be used under the terms of the GNU General Public License. */
7 /* This module is Handbrake's interface to the ffmpeg decoder library
8 (libavcodec & small parts of libavformat). It contains four Handbrake
11 decavcodec connects HB to an ffmpeg audio decoder
12 decavcodecv connects HB to an ffmpeg video decoder
14 (Two different routines are needed because the ffmpeg library
15 has different decoder calling conventions for audio & video.
16 The audio decoder should have had its name changed to "decavcodeca"
17 but I got lazy.) These work objects are self-contained & follow all
18 of HB's conventions for a decoder module. They can be used like
19 any other HB decoder (deca52, decmpeg2, etc.).
21 decavcodecai "internal" (incestuous?) version of decavcodec
22 decavcodecvi "internal" (incestuous?) version of decavcodecv
24 These routine are functionally equivalent to the routines above but
25 can only be used by the ffmpeg-based stream reader in libhb/stream.c.
26 The reason they exist is because the ffmpeg library leaves some of
27 the information needed by the decoder in the AVStream (the data
28 structure used by the stream reader) and we need to retrieve it
29 to successfully decode frames. But in HB the reader and decoder
30 modules are in completely separate threads and nothing goes between
31 them but hb_buffers containing frames to be decoded. I.e., there's
32 no easy way for the ffmpeg stream reader to pass a pointer to its
33 AVStream over to the ffmpeg video or audio decoder. So the *i work
34 objects use a private back door to the stream reader to get access
35 to the AVStream (routines hb_ffmpeg_avstream and hb_ffmpeg_context)
36 and the codec_param passed to these work objects is the key to this
37 back door (it's basically an index that allows the correct AVStream
40 The normal & *i objects share a lot of code (the basic frame decoding
41 and bitstream info code is factored out into subroutines that can be
42 called by either) but the top level routines of the *i objects
43 (decavcodecviWork, decavcodecviInfo, etc.) are different because:
44 1) they *have* to use the AVCodecContext that's contained in the
45 reader's AVStream rather than just allocating & using their own,
46 2) the Info routines have access to stuff kept in the AVStream in addition
47 to stuff kept in the AVCodecContext. This shouldn't be necessary but
48 crucial information like video frame rate that should be in the
49 AVCodecContext is either missing or wrong in the version of ffmpeg
50 we're currently using.
52 A consequence of the above is that the non-i work objects *can't* use
53 information from the AVStream because there isn't one - they get their
54 data from either the dvd reader or the mpeg reader, not the ffmpeg stream
55 reader. That means that they have to make up for deficiencies in the
56 AVCodecContext info by using stuff kept in the HB "title" struct. It
57 also means that ffmpeg codecs that randomly scatter state needed by
58 the decoder across both the AVCodecContext & the AVStream (e.g., the
59 VC1 decoder) can't easily be used by the HB mpeg stream reader.
65 #include "libavcodec/audioconvert.h"
67 static int decavcodecInit( hb_work_object_t *, hb_job_t * );
68 static int decavcodecWork( hb_work_object_t *, hb_buffer_t **, hb_buffer_t ** );
69 static void decavcodecClose( hb_work_object_t * );
70 static int decavcodecInfo( hb_work_object_t *, hb_work_info_t * );
71 static int decavcodecBSInfo( hb_work_object_t *, const hb_buffer_t *, hb_work_info_t * );
73 hb_work_object_t hb_decavcodec =
76 "MPGA decoder (libavcodec)",
86 // there are nheap items on the heap indexed 1..nheap (i.e., top of
87 // heap is 1). The 0th slot is unused - a marker is put there to check
88 // for overwrite errs.
89 int64_t h[HEAP_SIZE+1];
93 struct hb_work_private_s
96 AVCodecContext *context;
97 AVCodecParserContext *parser;
99 double duration; // frame duration (for video)
100 double pts_next; // next pts we expect to generate
101 int64_t pts; // (video) pts passing from parser to decoder
102 int64_t chap_time; // time of next chap mark (if new_chap != 0)
103 int new_chap; // output chapter mark pending
106 uint32_t decode_errors;
107 int brokenByMicrosoft; // video stream may contain packed b-frames
108 hb_buffer_t* delayq[HEAP_SIZE];
111 struct SwsContext *sws_context; // if we have to rescale or convert color space
112 hb_downmix_t *downmix;
113 hb_sample_t *downmix_buffer;
116 static void decodeAudio( hb_audio_t * audio, hb_work_private_t *pv, uint8_t *data, int size );
117 static hb_buffer_t *link_buf_list( hb_work_private_t *pv );
120 static int64_t heap_pop( pts_heap_t *heap )
124 if ( heap->nheap <= 0 )
129 // return the top of the heap then put the bottom element on top,
130 // decrease the heap size by one & rebalence the heap.
133 int64_t v = heap->h[heap->nheap--];
135 int child = parent << 1;
136 while ( child <= heap->nheap )
138 // find the smallest of the two children of parent
139 if (child < heap->nheap && heap->h[child] > heap->h[child+1] )
142 if (v <= heap->h[child])
143 // new item is smaller than either child so it's the new parent.
146 // smallest child is smaller than new item so move it up then
147 // check its children.
148 int64_t hp = heap->h[child];
149 heap->h[parent] = hp;
157 static void heap_push( pts_heap_t *heap, int64_t v )
159 if ( heap->nheap < HEAP_SIZE )
164 // stick the new value on the bottom of the heap then bubble it
165 // up to its correct spot.
166 int child = heap->nheap;
168 int parent = child >> 1;
169 if (heap->h[parent] <= v)
172 int64_t hp = heap->h[parent];
180 /***********************************************************************
181 * hb_work_decavcodec_init
182 ***********************************************************************
184 **********************************************************************/
185 static int decavcodecInit( hb_work_object_t * w, hb_job_t * job )
189 hb_work_private_t * pv = calloc( 1, sizeof( hb_work_private_t ) );
190 w->private_data = pv;
193 pv->list = hb_list_init();
195 int codec_id = w->codec_param;
198 codec_id = CODEC_ID_MP2;
200 codec = avcodec_find_decoder( codec_id );
201 pv->parser = av_parser_init( codec_id );
203 pv->context = avcodec_alloc_context();
204 hb_avcodec_open( pv->context, codec );
206 if ( w->audio != NULL &&
207 hb_need_downmix( w->audio->config.in.channel_layout,
208 w->audio->config.out.mixdown) )
210 pv->downmix = hb_downmix_init(w->audio->config.in.channel_layout,
211 w->audio->config.out.mixdown);
212 hb_downmix_set_chan_map( pv->downmix, &hb_smpte_chan_map, &hb_qt_chan_map );
213 if ( w->audio->config.out.codec == HB_ACODEC_LAME )
214 hb_downmix_adjust_level( pv->downmix );
220 /***********************************************************************
222 ***********************************************************************
224 **********************************************************************/
225 static void decavcodecClose( hb_work_object_t * w )
227 hb_work_private_t * pv = w->private_data;
231 if ( pv->job && pv->context && pv->context->codec )
233 hb_log( "%s-decoder done: %u frames, %u decoder errors, %u drops",
234 pv->context->codec->name, pv->nframes, pv->decode_errors,
237 if ( pv->sws_context )
239 sws_freeContext( pv->sws_context );
243 av_parser_close(pv->parser);
245 if ( pv->context && pv->context->codec )
247 hb_avcodec_close( pv->context );
251 hb_list_close( &pv->list );
255 av_free( pv->buffer );
260 hb_downmix_close( &(pv->downmix) );
262 if ( pv->downmix_buffer )
264 free( pv->downmix_buffer );
265 pv->downmix_buffer = NULL;
268 w->private_data = NULL;
272 /***********************************************************************
274 ***********************************************************************
276 **********************************************************************/
277 static int decavcodecWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
278 hb_buffer_t ** buf_out )
280 hb_work_private_t * pv = w->private_data;
281 hb_buffer_t * in = *buf_in;
285 /* EOF on input stream - send it downstream & say that we're done */
293 if ( in->start < -1 && pv->pts_next <= 0 )
295 // discard buffers that start before video time 0
299 // if the packet has a timestamp use it
300 if ( in->start != -1 )
302 pv->pts_next = in->start;
306 for ( pos = 0; pos < in->size; pos += len )
308 uint8_t *parser_output_buffer;
309 int parser_output_buffer_len;
310 int64_t cur = pv->pts_next;
312 if ( pv->parser != NULL )
314 len = av_parser_parse2( pv->parser, pv->context,
315 &parser_output_buffer, &parser_output_buffer_len,
316 in->data + pos, in->size - pos, cur, cur, AV_NOPTS_VALUE );
320 parser_output_buffer = in->data;
321 len = parser_output_buffer_len = in->size;
323 if (parser_output_buffer_len)
325 // set the duration on every frame since the stream format can
326 // change (it shouldn't but there's no way to guarantee it).
327 // duration is a scaling factor to go from #bytes in the decoded
328 // frame to frame time (in 90KHz mpeg ticks). 'channels' converts
329 // total samples to per-channel samples. 'sample_rate' converts
330 // per-channel samples to seconds per sample and the 90000
331 // is mpeg ticks per second.
332 if ( pv->context->sample_rate && pv->context->channels )
334 pv->duration = 90000. /
335 (double)( pv->context->sample_rate * pv->context->channels );
337 decodeAudio( w->audio, pv, parser_output_buffer, parser_output_buffer_len );
340 *buf_out = link_buf_list( pv );
344 static int decavcodecInfo( hb_work_object_t *w, hb_work_info_t *info )
346 hb_work_private_t *pv = w->private_data;
348 memset( info, 0, sizeof(*info) );
350 if ( pv && pv->context )
352 AVCodecContext *context = pv->context;
353 info->bitrate = context->bit_rate;
354 info->rate = context->time_base.num;
355 info->rate_base = context->time_base.den;
356 info->profile = context->profile;
357 info->level = context->level;
363 static int decavcodecBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
364 hb_work_info_t *info )
366 hb_work_private_t *pv = w->private_data;
369 memset( info, 0, sizeof(*info) );
371 if ( pv && pv->context )
373 return decavcodecInfo( w, info );
376 // We should parse the bitstream to find its parameters but for right
377 // now we just return dummy values if there's a codec that will handle it.
378 AVCodec *codec = avcodec_find_decoder( w->codec_param? w->codec_param :
382 // there's no ffmpeg codec for this audio type - give up
386 static char codec_name[64];
387 info->name = strncpy( codec_name, codec->name, sizeof(codec_name)-1 );
389 AVCodecParserContext *parser = av_parser_init( codec->id );
390 AVCodecContext *context = avcodec_alloc_context();
391 hb_avcodec_open( context, codec );
392 uint8_t *buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
393 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
394 unsigned char *pbuffer;
395 int pos, pbuffer_size;
397 while ( buf && !ret )
400 while ( pos < buf->size )
406 len = av_parser_parse2( parser, context, &pbuffer,
407 &pbuffer_size, buf->data + pos,
408 buf->size - pos, buf->start,
409 buf->start, AV_NOPTS_VALUE );
414 len = pbuffer_size = buf->size;
417 if ( pbuffer_size > 0 )
420 av_init_packet( &avp );
422 avp.size = pbuffer_size;
424 len = avcodec_decode_audio3( context, (int16_t*)buffer,
426 if ( len > 0 && context->sample_rate > 0 )
428 info->bitrate = context->bit_rate;
429 info->rate = context->sample_rate;
431 info->channel_layout =
432 hb_ff_layout_xlat(context->channel_layout,
443 if ( parser != NULL )
444 av_parser_close( parser );
445 hb_avcodec_close( context );
449 /* -------------------------------------------------------------
450 * General purpose video decoder using libavcodec
453 static uint8_t *copy_plane( uint8_t *dst, uint8_t* src, int dstride, int sstride,
456 if ( dstride == sstride )
458 memcpy( dst, src, dstride * h );
459 return dst + dstride * h;
461 int lbytes = dstride <= sstride? dstride : sstride;
464 memcpy( dst, src, lbytes );
471 // copy one video frame into an HB buf. If the frame isn't in our color space
472 // or at least one of its dimensions is odd, use sws_scale to convert/rescale it.
473 // Otherwise just copy the bits.
474 static hb_buffer_t *copy_frame( hb_work_private_t *pv, AVFrame *frame )
476 AVCodecContext *context = pv->context;
480 // if the dimensions are odd, drop the lsb since h264 requires that
481 // both width and height be even.
482 w = ( context->width >> 1 ) << 1;
483 h = ( context->height >> 1 ) << 1;
487 w = pv->job->title->width;
488 h = pv->job->title->height;
490 hb_buffer_t *buf = hb_video_buffer_init( w, h );
491 uint8_t *dst = buf->data;
493 if ( context->pix_fmt != PIX_FMT_YUV420P || w != context->width ||
494 h != context->height )
496 // have to convert to our internal color space and/or rescale
498 avpicture_fill( &dstpic, dst, PIX_FMT_YUV420P, w, h );
500 if ( ! pv->sws_context )
502 pv->sws_context = sws_getContext( context->width, context->height, context->pix_fmt,
503 w, h, PIX_FMT_YUV420P,
504 SWS_LANCZOS|SWS_ACCURATE_RND,
507 sws_scale( pv->sws_context, frame->data, frame->linesize, 0, h,
508 dstpic.data, dstpic.linesize );
512 dst = copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
513 w = (w + 1) >> 1; h = (h + 1) >> 1;
514 dst = copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
515 dst = copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
520 static int get_frame_buf( AVCodecContext *context, AVFrame *frame )
522 hb_work_private_t *pv = context->opaque;
523 frame->pts = pv->pts;
525 return avcodec_default_get_buffer( context, frame );
528 static void log_chapter( hb_work_private_t *pv, int chap_num, int64_t pts )
530 hb_chapter_t *c = hb_list_item( pv->job->title->list_chapter, chap_num - 1 );
533 hb_log( "%s: \"%s\" (%d) at frame %u time %"PRId64,
534 pv->context->codec->name, c->title, chap_num, pv->nframes, pts );
538 hb_log( "%s: Chapter %d at frame %u time %"PRId64,
539 pv->context->codec->name, chap_num, pv->nframes, pts );
543 static void flushDelayQueue( hb_work_private_t *pv )
546 int slot = pv->nframes & (HEAP_SIZE-1);
548 // flush all the video packets left on our timestamp-reordering delay q
549 while ( ( buf = pv->delayq[slot] ) != NULL )
551 buf->start = heap_pop( &pv->pts_heap );
552 hb_list_add( pv->list, buf );
553 pv->delayq[slot] = NULL;
554 slot = ( slot + 1 ) & (HEAP_SIZE-1);
558 static int decodeFrame( hb_work_private_t *pv, uint8_t *data, int size )
560 int got_picture, oldlevel = 0;
564 if ( global_verbosity_level <= 1 )
566 oldlevel = av_log_get_level();
567 av_log_set_level( AV_LOG_QUIET );
570 av_init_packet( &avp );
573 if ( avcodec_decode_video2( pv->context, &frame, &got_picture, &avp ) < 0 )
577 if ( global_verbosity_level <= 1 )
579 av_log_set_level( oldlevel );
583 // ffmpeg makes it hard to attach a pts to a frame. if the MPEG ES
584 // packet had a pts we handed it to av_parser_parse (if the packet had
585 // no pts we set it to -1 but before the parse we can't distinguish between
586 // the start of a video frame with no pts & an intermediate packet of
587 // some frame which never has a pts). we hope that when parse returns
588 // the frame to us the pts we originally handed it will be in parser->pts.
589 // we put this pts into pv->pts so that when a avcodec_decode_video
590 // finally gets around to allocating an AVFrame to hold the decoded
591 // frame we can stuff that pts into the frame. if all of these relays
592 // worked at this point frame.pts should hold the frame's pts from the
593 // original data stream or -1 if it didn't have one. in the latter case
594 // we generate the next pts in sequence for it.
595 double frame_dur = pv->duration;
596 if ( frame_dur <= 0 )
598 frame_dur = 90000. * (double)pv->context->time_base.num /
599 (double)pv->context->time_base.den;
600 pv->duration = frame_dur;
602 if ( frame.repeat_pict )
604 frame_dur += frame.repeat_pict * frame_dur * 0.5;
606 // XXX Unlike every other video decoder, the Raw decoder doesn't
607 // use the standard buffer allocation routines so we never
608 // get to put a PTS in the frame. Do it now.
609 if ( pv->context->codec_id == CODEC_ID_RAWVIDEO )
614 // If there was no pts for this frame, assume constant frame rate
615 // video & estimate the next frame time from the last & duration.
616 double pts = frame.pts;
621 pv->pts_next = pts + frame_dur;
625 // if we're doing a scan or this content couldn't have been broken
626 // by Microsoft we don't worry about timestamp reordering
627 if ( ! pv->job || ! pv->brokenByMicrosoft )
629 buf = copy_frame( pv, &frame );
631 hb_list_add( pv->list, buf );
636 // XXX This following probably addresses a libavcodec bug but I don't
637 // see an easy fix so we workaround it here.
639 // The M$ 'packed B-frames' atrocity results in decoded frames with
640 // the wrong timestamp. E.g., if there are 2 b-frames the timestamps
641 // we see here will be "2 3 1 5 6 4 ..." instead of "1 2 3 4 5 6".
642 // The frames are actually delivered in the right order but with
643 // the wrong timestamp. To get the correct timestamp attached to
644 // each frame we have a delay queue (longer than the max number of
645 // b-frames) & a sorting heap for the timestamps. As each frame
646 // comes out of the decoder the oldest frame in the queue is removed
647 // and associated with the smallest timestamp. Then the new frame is
648 // added to the queue & its timestamp is pushed on the heap.
649 // This does nothing if the timestamps are correct (i.e., the video
650 // uses a codec that Micro$oft hasn't broken yet) but the frames
651 // get timestamped correctly even when M$ has munged them.
653 // remove the oldest picture from the frame queue (if any) &
654 // give it the smallest timestamp from our heap. The queue size
655 // is a power of two so we get the slot of the oldest by masking
656 // the frame count & this will become the slot of the newest
657 // once we've removed & processed the oldest.
658 int slot = pv->nframes & (HEAP_SIZE-1);
659 if ( ( buf = pv->delayq[slot] ) != NULL )
661 buf->start = heap_pop( &pv->pts_heap );
663 if ( pv->new_chap && buf->start >= pv->chap_time )
665 buf->new_chap = pv->new_chap;
668 log_chapter( pv, buf->new_chap, buf->start );
670 else if ( pv->nframes == 0 )
672 log_chapter( pv, pv->job->chapter_start, buf->start );
674 hb_list_add( pv->list, buf );
677 // add the new frame to the delayq & push its timestamp on the heap
678 pv->delayq[slot] = copy_frame( pv, &frame );
679 heap_push( &pv->pts_heap, pts );
687 static void decodeVideo( hb_work_private_t *pv, uint8_t *data, int size,
688 int64_t pts, int64_t dts )
691 * The following loop is a do..while because we need to handle both
692 * data & the flush at the end (signaled by size=0). At the end there's
693 * generally a frame in the parser & one or more frames in the decoder
694 * (depending on the bframes setting).
700 int len = av_parser_parse2( pv->parser, pv->context, &pout, &pout_len,
701 data + pos, size - pos, pts, dts, AV_NOPTS_VALUE );
706 pv->pts = pv->parser->pts;
707 decodeFrame( pv, pout, pout_len );
709 } while ( pos < size );
711 /* the stuff above flushed the parser, now flush the decoder */
714 while ( decodeFrame( pv, NULL, 0 ) )
717 flushDelayQueue( pv );
721 static hb_buffer_t *link_buf_list( hb_work_private_t *pv )
723 hb_buffer_t *head = hb_list_item( pv->list, 0 );
727 hb_list_rem( pv->list, head );
729 hb_buffer_t *last = head, *buf;
731 while ( ( buf = hb_list_item( pv->list, 0 ) ) != NULL )
733 hb_list_rem( pv->list, buf );
742 static int decavcodecvInit( hb_work_object_t * w, hb_job_t * job )
745 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
746 w->private_data = pv;
748 pv->list = hb_list_init();
750 int codec_id = w->codec_param;
751 pv->parser = av_parser_init( codec_id );
752 pv->context = avcodec_alloc_context2( CODEC_TYPE_VIDEO );
754 /* we have to wrap ffmpeg's get_buffer to be able to set the pts (?!) */
755 pv->context->opaque = pv;
756 pv->context->get_buffer = get_frame_buf;
761 static int next_hdr( hb_buffer_t *in, int offset )
763 uint8_t *dat = in->data;
764 uint16_t last2 = 0xffff;
765 for ( ; in->size - offset > 1; ++offset )
767 if ( last2 == 0 && dat[offset] == 0x01 )
768 // found an mpeg start code
771 last2 = ( last2 << 8 ) | dat[offset];
777 static int find_hdr( hb_buffer_t *in, int offset, uint8_t hdr_type )
779 if ( in->size - offset < 4 )
780 // not enough room for an mpeg start code
783 for ( ; ( offset = next_hdr( in, offset ) ) >= 0; ++offset )
785 if ( in->data[offset+3] == hdr_type )
792 static int setup_extradata( hb_work_object_t *w, hb_buffer_t *in )
794 hb_work_private_t *pv = w->private_data;
796 // we can't call the avstream funcs but the read_header func in the
797 // AVInputFormat may set up some state in the AVContext. In particular
798 // vc1t_read_header allocates 'extradata' to deal with header issues
799 // related to Microsoft's bizarre engineering notions. We alloc a chunk
800 // of space to make vc1 work then associate the codec with the context.
801 if ( w->codec_param != CODEC_ID_VC1 )
803 // we haven't been inflicted with M$ - allocate a little space as
804 // a marker and return success.
805 pv->context->extradata_size = 16;
806 pv->context->extradata = av_malloc(pv->context->extradata_size);
810 // find the start and and of the sequence header
812 if ( ( shdr = find_hdr( in, 0, 0x0f ) ) < 0 )
814 // didn't find start of seq hdr
817 if ( ( shdr_end = next_hdr( in, shdr + 4 ) ) < 0 )
823 // find the start and and of the entry point header
825 if ( ( ehdr = find_hdr( in, 0, 0x0e ) ) < 0 )
827 // didn't find start of entry point hdr
830 if ( ( ehdr_end = next_hdr( in, ehdr + 4 ) ) < 0 )
836 // found both headers - allocate an extradata big enough to hold both
837 // then copy them into it.
838 pv->context->extradata_size = shdr_end + ehdr_end;
839 pv->context->extradata = av_malloc(pv->context->extradata_size + 8);
840 memcpy( pv->context->extradata, in->data + shdr, shdr_end );
841 memcpy( pv->context->extradata + shdr_end, in->data + ehdr, ehdr_end );
842 memset( pv->context->extradata + shdr_end + ehdr_end, 0, 8);
846 static int decavcodecvWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
847 hb_buffer_t ** buf_out )
849 hb_work_private_t *pv = w->private_data;
850 hb_buffer_t *in = *buf_in;
851 int64_t pts = AV_NOPTS_VALUE;
856 /* if we got an empty buffer signaling end-of-stream send it downstream */
859 decodeVideo( pv, in->data, in->size, pts, dts );
860 hb_list_add( pv->list, in );
861 *buf_out = link_buf_list( pv );
865 // if this is the first frame open the codec (we have to wait for the
866 // first frame because of M$ VC1 braindamage).
867 if ( pv->context->extradata_size == 0 )
869 if ( setup_extradata( w, in ) )
871 // we didn't find the headers needed to set up extradata.
872 // the codec will abort if we open it so just free the buf
873 // and hope we eventually get the info we need.
874 hb_buffer_close( &in );
877 AVCodec *codec = avcodec_find_decoder( w->codec_param );
878 // There's a mis-feature in ffmpeg that causes the context to be
879 // incorrectly initialized the 1st time avcodec_open is called.
880 // If you close it and open a 2nd time, it finishes the job.
881 hb_avcodec_open( pv->context, codec );
882 hb_avcodec_close( pv->context );
883 hb_avcodec_open( pv->context, codec );
889 dts = in->renderOffset;
893 pv->new_chap = in->new_chap;
894 pv->chap_time = pts >= 0? pts : pv->pts_next;
896 decodeVideo( pv, in->data, in->size, pts, dts );
897 hb_buffer_close( &in );
898 *buf_out = link_buf_list( pv );
902 static int decavcodecvInfo( hb_work_object_t *w, hb_work_info_t *info )
904 hb_work_private_t *pv = w->private_data;
906 memset( info, 0, sizeof(*info) );
908 if ( pv && pv->context )
910 AVCodecContext *context = pv->context;
911 info->bitrate = context->bit_rate;
912 info->width = context->width;
913 info->height = context->height;
915 /* ffmpeg gives the frame rate in frames per second while HB wants
916 * it in units of the 27MHz MPEG clock. */
917 info->rate = 27000000;
918 info->rate_base = (int64_t)context->time_base.num * 27000000LL /
919 context->time_base.den;
920 if ( context->ticks_per_frame > 1 )
922 // for ffmpeg 0.5 & later, the H.264 & MPEG-2 time base is
923 // field rate rather than frame rate so convert back to frames.
924 info->rate_base *= context->ticks_per_frame;
927 info->pixel_aspect_width = context->sample_aspect_ratio.num;
928 info->pixel_aspect_height = context->sample_aspect_ratio.den;
930 /* Sometimes there's no pixel aspect set in the source ffmpeg context
931 * which appears to come from the video stream. In that case,
932 * try the pixel aspect in AVStream (which appears to come from
933 * the container). Else assume a 1:1 PAR. */
934 if ( info->pixel_aspect_width == 0 ||
935 info->pixel_aspect_height == 0 )
937 AVStream *st = hb_ffmpeg_avstream( w->codec_param );
938 info->pixel_aspect_width = st->sample_aspect_ratio.num ?
939 st->sample_aspect_ratio.num : 1;
940 info->pixel_aspect_height = st->sample_aspect_ratio.den ?
941 st->sample_aspect_ratio.den : 1;
943 /* ffmpeg returns the Pixel Aspect Ratio (PAR). Handbrake wants the
944 * Display Aspect Ratio so we convert by scaling by the Storage
945 * Aspect Ratio (w/h). We do the calc in floating point to get the
947 info->aspect = (double)info->pixel_aspect_width *
948 (double)context->width /
949 (double)info->pixel_aspect_height /
950 (double)context->height;
952 info->profile = context->profile;
953 info->level = context->level;
954 info->name = context->codec->name;
960 static int decavcodecvBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
961 hb_work_info_t *info )
966 hb_work_object_t hb_decavcodecv =
969 "Video decoder (libavcodec)",
978 // This is a special decoder for ffmpeg streams. The ffmpeg stream reader
979 // includes a parser and passes information from the parser to the decoder
980 // via a codec context kept in the AVStream of the reader's AVFormatContext.
981 // We *have* to use that codec context to decode the stream or we'll get
982 // garbage. ffmpeg_title_scan put a cookie that can be used to get to that
983 // codec context in our codec_param.
985 // this routine gets the appropriate context pointer from the ffmpeg
986 // stream reader. it can't be called until we get the first buffer because
987 // we can't guarantee that reader will be called before the our init
988 // routine and if our init is called first we'll get a pointer to the
989 // old scan stream (which has already been closed).
990 static void init_ffmpeg_context( hb_work_object_t *w )
992 hb_work_private_t *pv = w->private_data;
993 pv->context = hb_ffmpeg_context( w->codec_param );
995 // during scan the decoder gets closed & reopened which will
996 // close the codec so reopen it if it's not there
997 if ( ! pv->context->codec )
999 AVCodec *codec = avcodec_find_decoder( pv->context->codec_id );
1000 hb_avcodec_open( pv->context, codec );
1002 // set up our best guess at the frame duration.
1003 // the frame rate in the codec is usually bogus but it's sometimes
1004 // ok in the stream.
1005 AVStream *st = hb_ffmpeg_avstream( w->codec_param );
1007 if ( st->nb_frames && st->duration )
1009 // compute the average frame duration from the total number
1010 // of frames & the total duration.
1011 pv->duration = ( (double)st->duration * (double)st->time_base.num ) /
1012 ( (double)st->nb_frames * (double)st->time_base.den );
1016 // XXX We don't have a frame count or duration so try to use the
1017 // far less reliable time base info in the stream.
1018 // Because the time bases are so screwed up, we only take values
1019 // in the range 8fps - 64fps.
1021 if ( st->time_base.num * 64 > st->time_base.den &&
1022 st->time_base.den > st->time_base.num * 8 )
1026 else if ( st->r_frame_rate.den * 64 > st->r_frame_rate.num &&
1027 st->r_frame_rate.num > st->r_frame_rate.den * 8 )
1029 tb.num = st->r_frame_rate.den;
1030 tb.den = st->r_frame_rate.num;
1034 tb.num = 1001; /*XXX*/
1035 tb.den = 24000; /*XXX*/
1037 pv->duration = (double)tb.num / (double)tb.den;
1039 pv->duration *= 90000.;
1041 // we have to wrap ffmpeg's get_buffer to be able to set the pts (?!)
1042 pv->context->opaque = pv;
1043 pv->context->get_buffer = get_frame_buf;
1045 // avi, mkv and possibly mp4 containers can contain the M$ VFW packed
1046 // b-frames abortion that messes up frame ordering and timestamps.
1047 // XXX ffmpeg knows which streams are broken but doesn't expose the
1048 // info externally. We should patch ffmpeg to add a flag to the
1049 // codec context for this but until then we mark all ffmpeg streams
1051 pv->brokenByMicrosoft = 1;
1054 static void prepare_ffmpeg_buffer( hb_buffer_t * in )
1056 // ffmpeg requires an extra 8 bytes of zero at the end of the buffer and
1057 // will seg fault in odd, data dependent ways if it's not there. (my guess
1058 // is this is a case of a local performance optimization creating a global
1059 // performance degradation since all the time wasted by extraneous data
1060 // copies & memory zeroing has to be huge compared to the minor reduction
1061 // in inner-loop instructions this affords - modern cpus bottleneck on
1062 // memory bandwidth not instruction bandwidth).
1063 if ( in->size + FF_INPUT_BUFFER_PADDING_SIZE > in->alloc )
1065 // have to realloc to add the padding
1066 hb_buffer_realloc( in, in->size + FF_INPUT_BUFFER_PADDING_SIZE );
1068 memset( in->data + in->size, 0, FF_INPUT_BUFFER_PADDING_SIZE );
1071 static int decavcodecviInit( hb_work_object_t * w, hb_job_t * job )
1074 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
1075 w->private_data = pv;
1077 pv->list = hb_list_init();
1081 if ( w->audio != NULL &&
1082 hb_need_downmix( w->audio->config.in.channel_layout,
1083 w->audio->config.out.mixdown) )
1085 pv->downmix = hb_downmix_init(w->audio->config.in.channel_layout,
1086 w->audio->config.out.mixdown);
1087 hb_downmix_set_chan_map( pv->downmix, &hb_smpte_chan_map, &hb_qt_chan_map );
1088 if ( w->audio->config.out.codec == HB_ACODEC_LAME )
1089 hb_downmix_adjust_level( pv->downmix );
1095 static int decavcodecviWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
1096 hb_buffer_t ** buf_out )
1098 hb_work_private_t *pv = w->private_data;
1099 hb_buffer_t *in = *buf_in;
1102 /* if we got an empty buffer signaling end-of-stream send it downstream */
1103 if ( in->size == 0 )
1105 /* flush any frames left in the decoder */
1106 while ( pv->context && decodeFrame( pv, NULL, 0 ) )
1109 flushDelayQueue( pv );
1110 hb_list_add( pv->list, in );
1111 *buf_out = link_buf_list( pv );
1112 return HB_WORK_DONE;
1115 if ( ! pv->context )
1117 init_ffmpeg_context( w );
1120 int64_t pts = in->start;
1123 // use the first timestamp as our 'next expected' pts
1124 if ( pv->pts_next < 0 )
1133 pv->new_chap = in->new_chap;
1134 pv->chap_time = pts >= 0? pts : pv->pts_next;
1136 prepare_ffmpeg_buffer( in );
1137 decodeFrame( pv, in->data, in->size );
1138 hb_buffer_close( &in );
1139 *buf_out = link_buf_list( pv );
1143 static int decavcodecviInfo( hb_work_object_t *w, hb_work_info_t *info )
1145 if ( decavcodecvInfo( w, info ) )
1147 hb_work_private_t *pv = w->private_data;
1148 if ( ! pv->context )
1150 init_ffmpeg_context( w );
1152 // we have the frame duration in units of the 90KHz pts clock but
1153 // need it in units of the 27MHz MPEG clock. */
1154 info->rate = 27000000;
1155 info->rate_base = pv->duration * 300.;
1161 static void decodeAudio( hb_audio_t * audio, hb_work_private_t *pv, uint8_t *data, int size )
1163 AVCodecContext *context = pv->context;
1166 while ( pos < size )
1168 int16_t *buffer = pv->buffer;
1169 if ( buffer == NULL )
1171 pv->buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
1172 buffer = pv->buffer;
1176 av_init_packet( &avp );
1177 avp.data = data + pos;
1178 avp.size = size - pos;
1180 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
1182 int len = avcodec_decode_audio3( context, buffer, &out_size, &avp );
1190 // We require signed 16-bit ints for the output format. If
1191 // we got something different convert it.
1192 if ( context->sample_fmt != SAMPLE_FMT_S16 )
1194 // Note: av_audio_convert seems to be a work-in-progress but
1195 // looks like it will eventually handle general audio
1196 // mixdowns which would allow us much more flexibility
1197 // in handling multichannel audio in HB. If we were doing
1198 // anything more complicated than a one-for-one format
1199 // conversion we'd probably want to cache the converter
1200 // context in the pv.
1201 int isamp = av_get_bits_per_sample_format( context->sample_fmt ) / 8;
1202 AVAudioConvert *ctx = av_audio_convert_alloc( SAMPLE_FMT_S16, 1,
1203 context->sample_fmt, 1,
1205 // get output buffer size (in 2-byte samples) then malloc a buffer
1206 nsamples = out_size / isamp;
1207 buffer = av_malloc( nsamples * 2 );
1209 // we're doing straight sample format conversion which behaves as if
1210 // there were only one channel.
1211 const void * const ibuf[6] = { pv->buffer };
1212 void * const obuf[6] = { buffer };
1213 const int istride[6] = { isamp };
1214 const int ostride[6] = { 2 };
1216 av_audio_convert( ctx, obuf, ostride, ibuf, istride, nsamples );
1217 av_audio_convert_free( ctx );
1221 nsamples = out_size / 2;
1228 pv->downmix_buffer = realloc(pv->downmix_buffer, nsamples * sizeof(hb_sample_t));
1231 for( i = 0; i < nsamples; ++i )
1233 pv->downmix_buffer[i] = buffer[i];
1236 int n_ch_samples = nsamples / context->channels;
1237 int channels = HB_AMIXDOWN_GET_DISCRETE_CHANNEL_COUNT(audio->config.out.mixdown);
1239 buf = hb_buffer_init( n_ch_samples * channels * sizeof(float) );
1240 hb_sample_t *samples = (hb_sample_t *)buf->data;
1241 hb_downmix(pv->downmix, samples, pv->downmix_buffer, n_ch_samples);
1245 buf = hb_buffer_init( nsamples * sizeof(float) );
1246 float *fl32 = (float *)buf->data;
1248 for( i = 0; i < nsamples; ++i )
1250 fl32[i] = buffer[i];
1252 int n_ch_samples = nsamples / context->channels;
1253 hb_layout_remap( &hb_smpte_chan_map, &hb_qt_chan_map,
1254 audio->config.in.channel_layout,
1255 fl32, n_ch_samples );
1258 double pts = pv->pts_next;
1260 pts += nsamples * pv->duration;
1264 hb_list_add( pv->list, buf );
1266 // if we allocated a buffer for sample format conversion, free it
1267 if ( buffer != pv->buffer )
1275 static int decavcodecaiWork( hb_work_object_t *w, hb_buffer_t **buf_in,
1276 hb_buffer_t **buf_out )
1278 if ( (*buf_in)->size <= 0 )
1280 /* EOF on input stream - send it downstream & say that we're done */
1283 return HB_WORK_DONE;
1286 hb_work_private_t *pv = w->private_data;
1288 if ( (*buf_in)->start < -1 && pv->pts_next <= 0 )
1290 // discard buffers that start before video time 0
1295 if ( ! pv->context )
1297 init_ffmpeg_context( w );
1298 // duration is a scaling factor to go from #bytes in the decoded
1299 // frame to frame time (in 90KHz mpeg ticks). 'channels' converts
1300 // total samples to per-channel samples. 'sample_rate' converts
1301 // per-channel samples to seconds per sample and the 90000
1302 // is mpeg ticks per second.
1303 pv->duration = 90000. /
1304 (double)( pv->context->sample_rate * pv->context->channels );
1306 hb_buffer_t *in = *buf_in;
1308 // if the packet has a timestamp use it if we don't have a timestamp yet
1309 // or if there's been a timing discontinuity of more than 100ms.
1310 if ( in->start >= 0 &&
1311 ( pv->pts_next < 0 || ( in->start - pv->pts_next ) > 90*100 ) )
1313 pv->pts_next = in->start;
1315 prepare_ffmpeg_buffer( in );
1316 decodeAudio( w->audio, pv, in->data, in->size );
1317 *buf_out = link_buf_list( pv );
1322 hb_work_object_t hb_decavcodecvi =
1325 "Video decoder (ffmpeg streams)",
1333 hb_work_object_t hb_decavcodecai =
1336 "Audio decoder (ffmpeg streams)",