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 );
218 /***********************************************************************
220 ***********************************************************************
222 **********************************************************************/
223 static void decavcodecClose( hb_work_object_t * w )
225 hb_work_private_t * pv = w->private_data;
229 if ( pv->job && pv->context && pv->context->codec )
231 hb_log( "%s-decoder done: %u frames, %u decoder errors, %u drops",
232 pv->context->codec->name, pv->nframes, pv->decode_errors,
235 if ( pv->sws_context )
237 sws_freeContext( pv->sws_context );
241 av_parser_close(pv->parser);
243 if ( pv->context && pv->context->codec )
245 hb_avcodec_close( pv->context );
249 hb_list_close( &pv->list );
253 av_free( pv->buffer );
258 hb_downmix_close( &(pv->downmix) );
260 if ( pv->downmix_buffer )
262 free( pv->downmix_buffer );
263 pv->downmix_buffer = NULL;
266 w->private_data = NULL;
270 /***********************************************************************
272 ***********************************************************************
274 **********************************************************************/
275 static int decavcodecWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
276 hb_buffer_t ** buf_out )
278 hb_work_private_t * pv = w->private_data;
279 hb_buffer_t * in = *buf_in;
283 /* EOF on input stream - send it downstream & say that we're done */
291 if ( in->start < -1 && pv->pts_next <= 0 )
293 // discard buffers that start before video time 0
297 // if the packet has a timestamp use it
298 if ( in->start != -1 )
300 pv->pts_next = in->start;
304 for ( pos = 0; pos < in->size; pos += len )
306 uint8_t *parser_output_buffer;
307 int parser_output_buffer_len;
308 int64_t cur = pv->pts_next;
310 if ( pv->parser != NULL )
312 len = av_parser_parse2( pv->parser, pv->context,
313 &parser_output_buffer, &parser_output_buffer_len,
314 in->data + pos, in->size - pos, cur, cur, AV_NOPTS_VALUE );
318 parser_output_buffer = in->data;
319 len = parser_output_buffer_len = in->size;
321 if (parser_output_buffer_len)
323 // set the duration on every frame since the stream format can
324 // change (it shouldn't but there's no way to guarantee it).
325 // duration is a scaling factor to go from #bytes in the decoded
326 // frame to frame time (in 90KHz mpeg ticks). 'channels' converts
327 // total samples to per-channel samples. 'sample_rate' converts
328 // per-channel samples to seconds per sample and the 90000
329 // is mpeg ticks per second.
330 if ( pv->context->sample_rate && pv->context->channels )
332 pv->duration = 90000. /
333 (double)( pv->context->sample_rate * pv->context->channels );
335 decodeAudio( w->audio, pv, parser_output_buffer, parser_output_buffer_len );
338 *buf_out = link_buf_list( pv );
342 static int decavcodecInfo( hb_work_object_t *w, hb_work_info_t *info )
344 hb_work_private_t *pv = w->private_data;
346 memset( info, 0, sizeof(*info) );
348 if ( pv && pv->context )
350 AVCodecContext *context = pv->context;
351 info->bitrate = context->bit_rate;
352 info->rate = context->time_base.num;
353 info->rate_base = context->time_base.den;
354 info->profile = context->profile;
355 info->level = context->level;
361 static int decavcodecBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
362 hb_work_info_t *info )
364 hb_work_private_t *pv = w->private_data;
367 memset( info, 0, sizeof(*info) );
369 if ( pv && pv->context )
371 return decavcodecInfo( w, info );
374 // We should parse the bitstream to find its parameters but for right
375 // now we just return dummy values if there's a codec that will handle it.
376 AVCodec *codec = avcodec_find_decoder( w->codec_param? w->codec_param :
380 // there's no ffmpeg codec for this audio type - give up
384 static char codec_name[64];
385 info->name = strncpy( codec_name, codec->name, sizeof(codec_name)-1 );
387 AVCodecParserContext *parser = av_parser_init( codec->id );
388 AVCodecContext *context = avcodec_alloc_context();
389 hb_avcodec_open( context, codec );
390 uint8_t *buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
391 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
392 unsigned char *pbuffer;
393 int pos, pbuffer_size;
395 while ( buf && !ret )
398 while ( pos < buf->size )
404 len = av_parser_parse2( parser, context, &pbuffer,
405 &pbuffer_size, buf->data + pos,
406 buf->size - pos, buf->start,
407 buf->start, AV_NOPTS_VALUE );
412 len = pbuffer_size = buf->size;
415 if ( pbuffer_size > 0 )
418 av_init_packet( &avp );
420 avp.size = pbuffer_size;
422 len = avcodec_decode_audio3( context, (int16_t*)buffer,
424 if ( len > 0 && context->sample_rate > 0 )
426 info->bitrate = context->bit_rate;
427 info->rate = context->sample_rate;
429 info->channel_layout =
430 hb_ff_layout_xlat(context->channel_layout,
441 if ( parser != NULL )
442 av_parser_close( parser );
443 hb_avcodec_close( context );
447 /* -------------------------------------------------------------
448 * General purpose video decoder using libavcodec
451 static uint8_t *copy_plane( uint8_t *dst, uint8_t* src, int dstride, int sstride,
454 if ( dstride == sstride )
456 memcpy( dst, src, dstride * h );
457 return dst + dstride * h;
459 int lbytes = dstride <= sstride? dstride : sstride;
462 memcpy( dst, src, lbytes );
469 // copy one video frame into an HB buf. If the frame isn't in our color space
470 // or at least one of its dimensions is odd, use sws_scale to convert/rescale it.
471 // Otherwise just copy the bits.
472 static hb_buffer_t *copy_frame( hb_work_private_t *pv, AVFrame *frame )
474 AVCodecContext *context = pv->context;
478 // if the dimensions are odd, drop the lsb since h264 requires that
479 // both width and height be even.
480 w = ( context->width >> 1 ) << 1;
481 h = ( context->height >> 1 ) << 1;
485 w = pv->job->title->width;
486 h = pv->job->title->height;
488 hb_buffer_t *buf = hb_video_buffer_init( w, h );
489 uint8_t *dst = buf->data;
491 if ( context->pix_fmt != PIX_FMT_YUV420P || w != context->width ||
492 h != context->height )
494 // have to convert to our internal color space and/or rescale
496 avpicture_fill( &dstpic, dst, PIX_FMT_YUV420P, w, h );
498 if ( ! pv->sws_context )
500 pv->sws_context = sws_getContext( context->width, context->height, context->pix_fmt,
501 w, h, PIX_FMT_YUV420P,
502 SWS_LANCZOS|SWS_ACCURATE_RND,
505 sws_scale( pv->sws_context, frame->data, frame->linesize, 0, h,
506 dstpic.data, dstpic.linesize );
510 dst = copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
511 w = (w + 1) >> 1; h = (h + 1) >> 1;
512 dst = copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
513 dst = copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
518 static int get_frame_buf( AVCodecContext *context, AVFrame *frame )
520 hb_work_private_t *pv = context->opaque;
521 frame->pts = pv->pts;
523 return avcodec_default_get_buffer( context, frame );
526 static int reget_frame_buf( AVCodecContext *context, AVFrame *frame )
528 hb_work_private_t *pv = context->opaque;
529 frame->pts = pv->pts;
531 return avcodec_default_reget_buffer( context, frame );
534 static void log_chapter( hb_work_private_t *pv, int chap_num, int64_t pts )
541 c = hb_list_item( pv->job->title->list_chapter, chap_num - 1 );
544 hb_log( "%s: \"%s\" (%d) at frame %u time %"PRId64,
545 pv->context->codec->name, c->title, chap_num, pv->nframes, pts );
549 hb_log( "%s: Chapter %d at frame %u time %"PRId64,
550 pv->context->codec->name, chap_num, pv->nframes, pts );
554 static void flushDelayQueue( hb_work_private_t *pv )
557 int slot = pv->nframes & (HEAP_SIZE-1);
559 // flush all the video packets left on our timestamp-reordering delay q
560 while ( ( buf = pv->delayq[slot] ) != NULL )
562 buf->start = heap_pop( &pv->pts_heap );
563 hb_list_add( pv->list, buf );
564 pv->delayq[slot] = NULL;
565 slot = ( slot + 1 ) & (HEAP_SIZE-1);
570 * Decodes a video frame from the specified raw packet data ('data', 'size', 'sequence').
571 * The output of this function is stored in 'pv->list', which contains a list
572 * of zero or more decoded packets.
574 * The returned packets are guaranteed to have their timestamps in the correct order,
575 * even if the original packets decoded by libavcodec have misordered timestamps,
576 * due to the use of 'packed B-frames'.
578 * Internally the set of decoded packets may be buffered in 'pv->delayq'
579 * until enough packets have been decoded so that the timestamps can be
580 * correctly rewritten, if this is necessary.
582 static int decodeFrame( hb_work_private_t *pv, uint8_t *data, int size, int sequence )
584 int got_picture, oldlevel = 0;
588 if ( global_verbosity_level <= 1 )
590 oldlevel = av_log_get_level();
591 av_log_set_level( AV_LOG_QUIET );
594 av_init_packet( &avp );
597 if ( avcodec_decode_video2( pv->context, &frame, &got_picture, &avp ) < 0 )
601 if ( global_verbosity_level <= 1 )
603 av_log_set_level( oldlevel );
607 // ffmpeg makes it hard to attach a pts to a frame. if the MPEG ES
608 // packet had a pts we handed it to av_parser_parse (if the packet had
609 // no pts we set it to -1 but before the parse we can't distinguish between
610 // the start of a video frame with no pts & an intermediate packet of
611 // some frame which never has a pts). we hope that when parse returns
612 // the frame to us the pts we originally handed it will be in parser->pts.
613 // we put this pts into pv->pts so that when a avcodec_decode_video
614 // finally gets around to allocating an AVFrame to hold the decoded
615 // frame we can stuff that pts into the frame. if all of these relays
616 // worked at this point frame.pts should hold the frame's pts from the
617 // original data stream or -1 if it didn't have one. in the latter case
618 // we generate the next pts in sequence for it.
619 double frame_dur = pv->duration;
620 if ( frame_dur <= 0 )
622 frame_dur = 90000. * (double)pv->context->time_base.num /
623 (double)pv->context->time_base.den;
624 pv->duration = frame_dur;
626 if ( frame.repeat_pict )
628 frame_dur += frame.repeat_pict * frame_dur * 0.5;
630 // XXX Unlike every other video decoder, the Raw decoder doesn't
631 // use the standard buffer allocation routines so we never
632 // get to put a PTS in the frame. Do it now.
633 if ( pv->context->codec_id == CODEC_ID_RAWVIDEO )
638 // If there was no pts for this frame, assume constant frame rate
639 // video & estimate the next frame time from the last & duration.
640 double pts = frame.pts;
645 pv->pts_next = pts + frame_dur;
649 // if we're doing a scan or this content couldn't have been broken
650 // by Microsoft we don't worry about timestamp reordering
651 if ( ! pv->job || ! pv->brokenByMicrosoft )
653 buf = copy_frame( pv, &frame );
655 buf->sequence = sequence;
656 if ( pv->new_chap && buf->start >= pv->chap_time )
658 buf->new_chap = pv->new_chap;
661 log_chapter( pv, buf->new_chap, buf->start );
663 else if ( pv->nframes == 0 && pv->job )
665 log_chapter( pv, pv->job->chapter_start, buf->start );
667 hb_list_add( pv->list, buf );
672 // XXX This following probably addresses a libavcodec bug but I don't
673 // see an easy fix so we workaround it here.
675 // The M$ 'packed B-frames' atrocity results in decoded frames with
676 // the wrong timestamp. E.g., if there are 2 b-frames the timestamps
677 // we see here will be "2 3 1 5 6 4 ..." instead of "1 2 3 4 5 6".
678 // The frames are actually delivered in the right order but with
679 // the wrong timestamp. To get the correct timestamp attached to
680 // each frame we have a delay queue (longer than the max number of
681 // b-frames) & a sorting heap for the timestamps. As each frame
682 // comes out of the decoder the oldest frame in the queue is removed
683 // and associated with the smallest timestamp. Then the new frame is
684 // added to the queue & its timestamp is pushed on the heap.
685 // This does nothing if the timestamps are correct (i.e., the video
686 // uses a codec that Micro$oft hasn't broken yet) but the frames
687 // get timestamped correctly even when M$ has munged them.
689 // remove the oldest picture from the frame queue (if any) &
690 // give it the smallest timestamp from our heap. The queue size
691 // is a power of two so we get the slot of the oldest by masking
692 // the frame count & this will become the slot of the newest
693 // once we've removed & processed the oldest.
694 int slot = pv->nframes & (HEAP_SIZE-1);
695 if ( ( buf = pv->delayq[slot] ) != NULL )
697 buf->start = heap_pop( &pv->pts_heap );
699 if ( pv->new_chap && buf->start >= pv->chap_time )
701 buf->new_chap = pv->new_chap;
704 log_chapter( pv, buf->new_chap, buf->start );
706 else if ( pv->nframes == 0 && pv->job )
708 log_chapter( pv, pv->job->chapter_start, buf->start );
710 hb_list_add( pv->list, buf );
713 // add the new frame to the delayq & push its timestamp on the heap
714 buf = copy_frame( pv, &frame );
715 buf->sequence = sequence;
716 pv->delayq[slot] = buf;
717 heap_push( &pv->pts_heap, pts );
725 static void decodeVideo( hb_work_private_t *pv, uint8_t *data, int size, int sequence,
726 int64_t pts, int64_t dts )
729 * The following loop is a do..while because we need to handle both
730 * data & the flush at the end (signaled by size=0). At the end there's
731 * generally a frame in the parser & one or more frames in the decoder
732 * (depending on the bframes setting).
738 int len = av_parser_parse2( pv->parser, pv->context, &pout, &pout_len,
739 data + pos, size - pos, pts, dts, AV_NOPTS_VALUE );
744 pv->pts = pv->parser->pts;
745 decodeFrame( pv, pout, pout_len, sequence );
747 } while ( pos < size );
749 /* the stuff above flushed the parser, now flush the decoder */
752 while ( decodeFrame( pv, NULL, 0, sequence ) )
755 flushDelayQueue( pv );
760 * Removes all packets from 'pv->list', links them together into
761 * a linked-list, and returns the first packet in the list.
763 static hb_buffer_t *link_buf_list( hb_work_private_t *pv )
765 hb_buffer_t *head = hb_list_item( pv->list, 0 );
769 hb_list_rem( pv->list, head );
771 hb_buffer_t *last = head, *buf;
773 while ( ( buf = hb_list_item( pv->list, 0 ) ) != NULL )
775 hb_list_rem( pv->list, buf );
784 static int decavcodecvInit( hb_work_object_t * w, hb_job_t * job )
787 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
788 w->private_data = pv;
790 pv->list = hb_list_init();
792 int codec_id = w->codec_param;
793 pv->parser = av_parser_init( codec_id );
794 pv->context = avcodec_alloc_context2( CODEC_TYPE_VIDEO );
796 /* we have to wrap ffmpeg's get_buffer to be able to set the pts (?!) */
797 pv->context->opaque = pv;
798 pv->context->get_buffer = get_frame_buf;
799 pv->context->reget_buffer = reget_frame_buf;
804 static int next_hdr( hb_buffer_t *in, int offset )
806 uint8_t *dat = in->data;
807 uint16_t last2 = 0xffff;
808 for ( ; in->size - offset > 1; ++offset )
810 if ( last2 == 0 && dat[offset] == 0x01 )
811 // found an mpeg start code
814 last2 = ( last2 << 8 ) | dat[offset];
820 static int find_hdr( hb_buffer_t *in, int offset, uint8_t hdr_type )
822 if ( in->size - offset < 4 )
823 // not enough room for an mpeg start code
826 for ( ; ( offset = next_hdr( in, offset ) ) >= 0; ++offset )
828 if ( in->data[offset+3] == hdr_type )
835 static int setup_extradata( hb_work_object_t *w, hb_buffer_t *in )
837 hb_work_private_t *pv = w->private_data;
839 // we can't call the avstream funcs but the read_header func in the
840 // AVInputFormat may set up some state in the AVContext. In particular
841 // vc1t_read_header allocates 'extradata' to deal with header issues
842 // related to Microsoft's bizarre engineering notions. We alloc a chunk
843 // of space to make vc1 work then associate the codec with the context.
844 if ( w->codec_param != CODEC_ID_VC1 )
846 // we haven't been inflicted with M$ - allocate a little space as
847 // a marker and return success.
848 pv->context->extradata_size = 16;
849 pv->context->extradata = av_malloc(pv->context->extradata_size);
853 // find the start and and of the sequence header
855 if ( ( shdr = find_hdr( in, 0, 0x0f ) ) < 0 )
857 // didn't find start of seq hdr
860 if ( ( shdr_end = next_hdr( in, shdr + 4 ) ) < 0 )
866 // find the start and and of the entry point header
868 if ( ( ehdr = find_hdr( in, 0, 0x0e ) ) < 0 )
870 // didn't find start of entry point hdr
873 if ( ( ehdr_end = next_hdr( in, ehdr + 4 ) ) < 0 )
879 // found both headers - allocate an extradata big enough to hold both
880 // then copy them into it.
881 pv->context->extradata_size = shdr_end + ehdr_end;
882 pv->context->extradata = av_malloc(pv->context->extradata_size + 8);
883 memcpy( pv->context->extradata, in->data + shdr, shdr_end );
884 memcpy( pv->context->extradata + shdr_end, in->data + ehdr, ehdr_end );
885 memset( pv->context->extradata + shdr_end + ehdr_end, 0, 8);
889 static int decavcodecvWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
890 hb_buffer_t ** buf_out )
892 hb_work_private_t *pv = w->private_data;
893 hb_buffer_t *in = *buf_in;
894 int64_t pts = AV_NOPTS_VALUE;
899 /* if we got an empty buffer signaling end-of-stream send it downstream */
902 decodeVideo( pv, in->data, in->size, in->sequence, pts, dts );
903 hb_list_add( pv->list, in );
904 *buf_out = link_buf_list( pv );
908 // if this is the first frame open the codec (we have to wait for the
909 // first frame because of M$ VC1 braindamage).
910 if ( pv->context->extradata_size == 0 )
912 if ( setup_extradata( w, in ) )
914 // we didn't find the headers needed to set up extradata.
915 // the codec will abort if we open it so just free the buf
916 // and hope we eventually get the info we need.
917 hb_buffer_close( &in );
920 AVCodec *codec = avcodec_find_decoder( w->codec_param );
921 // There's a mis-feature in ffmpeg that causes the context to be
922 // incorrectly initialized the 1st time avcodec_open is called.
923 // If you close it and open a 2nd time, it finishes the job.
924 hb_avcodec_open( pv->context, codec );
925 hb_avcodec_close( pv->context );
926 hb_avcodec_open( pv->context, codec );
932 dts = in->renderOffset;
936 pv->new_chap = in->new_chap;
937 pv->chap_time = pts >= 0? pts : pv->pts_next;
939 decodeVideo( pv, in->data, in->size, in->sequence, pts, dts );
940 hb_buffer_close( &in );
941 *buf_out = link_buf_list( pv );
945 static int decavcodecvInfo( hb_work_object_t *w, hb_work_info_t *info )
947 hb_work_private_t *pv = w->private_data;
949 memset( info, 0, sizeof(*info) );
951 if ( pv && pv->context )
953 AVCodecContext *context = pv->context;
954 info->bitrate = context->bit_rate;
955 info->width = context->width;
956 info->height = context->height;
958 /* ffmpeg gives the frame rate in frames per second while HB wants
959 * it in units of the 27MHz MPEG clock. */
960 info->rate = 27000000;
961 info->rate_base = (int64_t)context->time_base.num * 27000000LL /
962 context->time_base.den;
963 if ( context->ticks_per_frame > 1 )
965 // for ffmpeg 0.5 & later, the H.264 & MPEG-2 time base is
966 // field rate rather than frame rate so convert back to frames.
967 info->rate_base *= context->ticks_per_frame;
970 info->pixel_aspect_width = context->sample_aspect_ratio.num;
971 info->pixel_aspect_height = context->sample_aspect_ratio.den;
973 /* Sometimes there's no pixel aspect set in the source ffmpeg context
974 * which appears to come from the video stream. In that case,
975 * try the pixel aspect in AVStream (which appears to come from
976 * the container). Else assume a 1:1 PAR. */
977 if ( info->pixel_aspect_width == 0 ||
978 info->pixel_aspect_height == 0 )
980 AVStream *st = hb_ffmpeg_avstream( w->codec_param );
981 info->pixel_aspect_width = st->sample_aspect_ratio.num ?
982 st->sample_aspect_ratio.num : 1;
983 info->pixel_aspect_height = st->sample_aspect_ratio.den ?
984 st->sample_aspect_ratio.den : 1;
986 /* ffmpeg returns the Pixel Aspect Ratio (PAR). Handbrake wants the
987 * Display Aspect Ratio so we convert by scaling by the Storage
988 * Aspect Ratio (w/h). We do the calc in floating point to get the
990 info->aspect = (double)info->pixel_aspect_width *
991 (double)context->width /
992 (double)info->pixel_aspect_height /
993 (double)context->height;
995 info->profile = context->profile;
996 info->level = context->level;
997 info->name = context->codec->name;
1003 static int decavcodecvBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
1004 hb_work_info_t *info )
1009 hb_work_object_t hb_decavcodecv =
1012 "Video decoder (libavcodec)",
1021 // This is a special decoder for ffmpeg streams. The ffmpeg stream reader
1022 // includes a parser and passes information from the parser to the decoder
1023 // via a codec context kept in the AVStream of the reader's AVFormatContext.
1024 // We *have* to use that codec context to decode the stream or we'll get
1025 // garbage. ffmpeg_title_scan put a cookie that can be used to get to that
1026 // codec context in our codec_param.
1028 // this routine gets the appropriate context pointer from the ffmpeg
1029 // stream reader. it can't be called until we get the first buffer because
1030 // we can't guarantee that reader will be called before the our init
1031 // routine and if our init is called first we'll get a pointer to the
1032 // old scan stream (which has already been closed).
1033 static void init_ffmpeg_context( hb_work_object_t *w )
1035 hb_work_private_t *pv = w->private_data;
1036 pv->context = hb_ffmpeg_context( w->codec_param );
1038 // during scan the decoder gets closed & reopened which will
1039 // close the codec so reopen it if it's not there
1040 if ( ! pv->context->codec )
1042 AVCodec *codec = avcodec_find_decoder( pv->context->codec_id );
1043 hb_avcodec_open( pv->context, codec );
1045 // set up our best guess at the frame duration.
1046 // the frame rate in the codec is usually bogus but it's sometimes
1047 // ok in the stream.
1048 AVStream *st = hb_ffmpeg_avstream( w->codec_param );
1050 if ( st->nb_frames && st->duration )
1052 // compute the average frame duration from the total number
1053 // of frames & the total duration.
1054 pv->duration = ( (double)st->duration * (double)st->time_base.num ) /
1055 ( (double)st->nb_frames * (double)st->time_base.den );
1059 // XXX We don't have a frame count or duration so try to use the
1060 // far less reliable time base info in the stream.
1061 // Because the time bases are so screwed up, we only take values
1062 // in the range 8fps - 64fps.
1064 if ( st->avg_frame_rate.den * 64 > st->avg_frame_rate.num &&
1065 st->avg_frame_rate.num > st->avg_frame_rate.den * 8 )
1067 tb.num = st->avg_frame_rate.den;
1068 tb.den = st->avg_frame_rate.num;
1070 else if ( st->time_base.num * 64 > st->time_base.den &&
1071 st->time_base.den > st->time_base.num * 8 )
1075 else if ( st->r_frame_rate.den * 64 > st->r_frame_rate.num &&
1076 st->r_frame_rate.num > st->r_frame_rate.den * 8 )
1078 tb.num = st->r_frame_rate.den;
1079 tb.den = st->r_frame_rate.num;
1083 tb.num = 1001; /*XXX*/
1084 tb.den = 24000; /*XXX*/
1086 pv->duration = (double)tb.num / (double)tb.den;
1088 pv->duration *= 90000.;
1090 // we have to wrap ffmpeg's get_buffer to be able to set the pts (?!)
1091 pv->context->opaque = pv;
1092 pv->context->get_buffer = get_frame_buf;
1093 pv->context->reget_buffer = reget_frame_buf;
1095 // avi, mkv and possibly mp4 containers can contain the M$ VFW packed
1096 // b-frames abortion that messes up frame ordering and timestamps.
1097 // XXX ffmpeg knows which streams are broken but doesn't expose the
1098 // info externally. We should patch ffmpeg to add a flag to the
1099 // codec context for this but until then we mark all ffmpeg streams
1101 pv->brokenByMicrosoft = 1;
1104 static void prepare_ffmpeg_buffer( hb_buffer_t * in )
1106 // ffmpeg requires an extra 8 bytes of zero at the end of the buffer and
1107 // will seg fault in odd, data dependent ways if it's not there. (my guess
1108 // is this is a case of a local performance optimization creating a global
1109 // performance degradation since all the time wasted by extraneous data
1110 // copies & memory zeroing has to be huge compared to the minor reduction
1111 // in inner-loop instructions this affords - modern cpus bottleneck on
1112 // memory bandwidth not instruction bandwidth).
1113 if ( in->size + FF_INPUT_BUFFER_PADDING_SIZE > in->alloc )
1115 // have to realloc to add the padding
1116 hb_buffer_realloc( in, in->size + FF_INPUT_BUFFER_PADDING_SIZE );
1118 memset( in->data + in->size, 0, FF_INPUT_BUFFER_PADDING_SIZE );
1121 static int decavcodecviInit( hb_work_object_t * w, hb_job_t * job )
1124 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
1125 w->private_data = pv;
1127 pv->list = hb_list_init();
1131 if ( w->audio != NULL &&
1132 hb_need_downmix( w->audio->config.in.channel_layout,
1133 w->audio->config.out.mixdown) )
1135 pv->downmix = hb_downmix_init(w->audio->config.in.channel_layout,
1136 w->audio->config.out.mixdown);
1137 hb_downmix_set_chan_map( pv->downmix, &hb_smpte_chan_map, &hb_qt_chan_map );
1143 static int decavcodecviWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
1144 hb_buffer_t ** buf_out )
1146 hb_work_private_t *pv = w->private_data;
1147 hb_buffer_t *in = *buf_in;
1150 /* if we got an empty buffer signaling end-of-stream send it downstream */
1151 if ( in->size == 0 )
1153 /* flush any frames left in the decoder */
1154 while ( pv->context && decodeFrame( pv, NULL, 0, in->sequence ) )
1157 flushDelayQueue( pv );
1158 hb_list_add( pv->list, in );
1159 *buf_out = link_buf_list( pv );
1160 return HB_WORK_DONE;
1163 if ( ! pv->context )
1165 init_ffmpeg_context( w );
1168 int64_t pts = in->start;
1171 // use the first timestamp as our 'next expected' pts
1172 if ( pv->pts_next < 0 )
1181 pv->new_chap = in->new_chap;
1182 pv->chap_time = pts >= 0? pts : pv->pts_next;
1184 prepare_ffmpeg_buffer( in );
1185 decodeFrame( pv, in->data, in->size, in->sequence );
1186 hb_buffer_close( &in );
1187 *buf_out = link_buf_list( pv );
1191 static int decavcodecviInfo( hb_work_object_t *w, hb_work_info_t *info )
1193 if ( decavcodecvInfo( w, info ) )
1195 hb_work_private_t *pv = w->private_data;
1196 if ( ! pv->context )
1198 init_ffmpeg_context( w );
1200 // we have the frame duration in units of the 90KHz pts clock but
1201 // need it in units of the 27MHz MPEG clock. */
1202 info->rate = 27000000;
1203 info->rate_base = pv->duration * 300.;
1209 static void decodeAudio( hb_audio_t * audio, hb_work_private_t *pv, uint8_t *data, int size )
1211 AVCodecContext *context = pv->context;
1213 int loop_limit = 256;
1215 while ( pos < size )
1217 int16_t *buffer = pv->buffer;
1218 if ( buffer == NULL )
1220 pv->buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
1221 buffer = pv->buffer;
1225 av_init_packet( &avp );
1226 avp.data = data + pos;
1227 avp.size = size - pos;
1229 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
1231 int len = avcodec_decode_audio3( context, buffer, &out_size, &avp );
1238 if ( !(loop_limit--) )
1247 // We require signed 16-bit ints for the output format. If
1248 // we got something different convert it.
1249 if ( context->sample_fmt != SAMPLE_FMT_S16 )
1251 // Note: av_audio_convert seems to be a work-in-progress but
1252 // looks like it will eventually handle general audio
1253 // mixdowns which would allow us much more flexibility
1254 // in handling multichannel audio in HB. If we were doing
1255 // anything more complicated than a one-for-one format
1256 // conversion we'd probably want to cache the converter
1257 // context in the pv.
1258 int isamp = av_get_bits_per_sample_format( context->sample_fmt ) / 8;
1259 AVAudioConvert *ctx = av_audio_convert_alloc( SAMPLE_FMT_S16, 1,
1260 context->sample_fmt, 1,
1262 // get output buffer size (in 2-byte samples) then malloc a buffer
1263 nsamples = out_size / isamp;
1264 buffer = av_malloc( nsamples * 2 );
1266 // we're doing straight sample format conversion which behaves as if
1267 // there were only one channel.
1268 const void * const ibuf[6] = { pv->buffer };
1269 void * const obuf[6] = { buffer };
1270 const int istride[6] = { isamp };
1271 const int ostride[6] = { 2 };
1273 av_audio_convert( ctx, obuf, ostride, ibuf, istride, nsamples );
1274 av_audio_convert_free( ctx );
1278 nsamples = out_size / 2;
1285 pv->downmix_buffer = realloc(pv->downmix_buffer, nsamples * sizeof(hb_sample_t));
1288 for( i = 0; i < nsamples; ++i )
1290 pv->downmix_buffer[i] = buffer[i];
1293 int n_ch_samples = nsamples / context->channels;
1294 int channels = HB_AMIXDOWN_GET_DISCRETE_CHANNEL_COUNT(audio->config.out.mixdown);
1296 buf = hb_buffer_init( n_ch_samples * channels * sizeof(float) );
1297 hb_sample_t *samples = (hb_sample_t *)buf->data;
1298 hb_downmix(pv->downmix, samples, pv->downmix_buffer, n_ch_samples);
1302 buf = hb_buffer_init( nsamples * sizeof(float) );
1303 float *fl32 = (float *)buf->data;
1305 for( i = 0; i < nsamples; ++i )
1307 fl32[i] = buffer[i];
1309 int n_ch_samples = nsamples / context->channels;
1310 hb_layout_remap( &hb_smpte_chan_map, &hb_qt_chan_map,
1311 audio->config.in.channel_layout,
1312 fl32, n_ch_samples );
1315 double pts = pv->pts_next;
1317 pts += nsamples * pv->duration;
1321 hb_list_add( pv->list, buf );
1323 // if we allocated a buffer for sample format conversion, free it
1324 if ( buffer != pv->buffer )
1332 static int decavcodecaiWork( hb_work_object_t *w, hb_buffer_t **buf_in,
1333 hb_buffer_t **buf_out )
1335 if ( (*buf_in)->size <= 0 )
1337 /* EOF on input stream - send it downstream & say that we're done */
1340 return HB_WORK_DONE;
1343 hb_work_private_t *pv = w->private_data;
1345 if ( (*buf_in)->start < -1 && pv->pts_next <= 0 )
1347 // discard buffers that start before video time 0
1352 if ( ! pv->context )
1354 init_ffmpeg_context( w );
1355 // duration is a scaling factor to go from #bytes in the decoded
1356 // frame to frame time (in 90KHz mpeg ticks). 'channels' converts
1357 // total samples to per-channel samples. 'sample_rate' converts
1358 // per-channel samples to seconds per sample and the 90000
1359 // is mpeg ticks per second.
1360 pv->duration = 90000. /
1361 (double)( pv->context->sample_rate * pv->context->channels );
1363 hb_buffer_t *in = *buf_in;
1365 // if the packet has a timestamp use it if we don't have a timestamp yet
1366 // or if there's been a timing discontinuity of more than 100ms.
1367 if ( in->start >= 0 &&
1368 ( pv->pts_next < 0 || ( in->start - pv->pts_next ) > 90*100 ) )
1370 pv->pts_next = in->start;
1372 prepare_ffmpeg_buffer( in );
1373 decodeAudio( w->audio, pv, in->data, in->size );
1374 *buf_out = link_buf_list( pv );
1379 hb_work_object_t hb_decavcodecvi =
1382 "Video decoder (ffmpeg streams)",
1390 hb_work_object_t hb_decavcodecai =
1393 "Audio decoder (ffmpeg streams)",