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.
64 #include "libavcodec/audioconvert.h"
66 static int decavcodecInit( hb_work_object_t *, hb_job_t * );
67 static int decavcodecWork( hb_work_object_t *, hb_buffer_t **, hb_buffer_t ** );
68 static void decavcodecClose( hb_work_object_t * );
69 static int decavcodecInfo( hb_work_object_t *, hb_work_info_t * );
70 static int decavcodecBSInfo( hb_work_object_t *, const hb_buffer_t *, hb_work_info_t * );
72 hb_work_object_t hb_decavcodec =
75 "MPGA decoder (libavcodec)",
85 // there are nheap items on the heap indexed 1..nheap (i.e., top of
86 // heap is 1). The 0th slot is unused - a marker is put there to check
87 // for overwrite errs.
88 int64_t h[HEAP_SIZE+1];
92 struct hb_work_private_s
95 AVCodecContext *context;
96 AVCodecParserContext *parser;
98 double duration; // frame duration (for video)
99 double pts_next; // next pts we expect to generate
100 int64_t pts; // (video) pts passing from parser to decoder
101 int64_t chap_time; // time of next chap mark (if new_chap != 0)
102 int new_chap; // output chapter mark pending
105 uint32_t decode_errors;
106 int brokenByMicrosoft; // video stream may contain packed b-frames
107 hb_buffer_t* delayq[HEAP_SIZE];
110 struct SwsContext *sws_context; // if we have to rescale or convert color space
113 static int64_t heap_pop( pts_heap_t *heap )
117 if ( heap->nheap <= 0 )
122 // return the top of the heap then put the bottom element on top,
123 // decrease the heap size by one & rebalence the heap.
126 int64_t v = heap->h[heap->nheap--];
128 int child = parent << 1;
129 while ( child <= heap->nheap )
131 // find the smallest of the two children of parent
132 if (child < heap->nheap && heap->h[child] > heap->h[child+1] )
135 if (v <= heap->h[child])
136 // new item is smaller than either child so it's the new parent.
139 // smallest child is smaller than new item so move it up then
140 // check its children.
141 int64_t hp = heap->h[child];
142 heap->h[parent] = hp;
150 static void heap_push( pts_heap_t *heap, int64_t v )
152 if ( heap->nheap < HEAP_SIZE )
157 // stick the new value on the bottom of the heap then bubble it
158 // up to its correct spot.
159 int child = heap->nheap;
161 int parent = child >> 1;
162 if (heap->h[parent] <= v)
165 int64_t hp = heap->h[parent];
173 /***********************************************************************
174 * hb_work_decavcodec_init
175 ***********************************************************************
177 **********************************************************************/
178 static int decavcodecInit( hb_work_object_t * w, hb_job_t * job )
182 hb_work_private_t * pv = calloc( 1, sizeof( hb_work_private_t ) );
183 w->private_data = pv;
187 int codec_id = w->codec_param;
190 codec_id = CODEC_ID_MP2;
192 codec = avcodec_find_decoder( codec_id );
193 pv->parser = av_parser_init( codec_id );
195 pv->context = avcodec_alloc_context();
196 hb_avcodec_open( pv->context, codec );
201 /***********************************************************************
203 ***********************************************************************
205 **********************************************************************/
206 static void decavcodecClose( hb_work_object_t * w )
208 hb_work_private_t * pv = w->private_data;
212 if ( pv->job && pv->context && pv->context->codec )
214 hb_log( "%s-decoder done: %u frames, %u decoder errors, %u drops",
215 pv->context->codec->name, pv->nframes, pv->decode_errors,
218 if ( pv->sws_context )
220 sws_freeContext( pv->sws_context );
224 av_parser_close(pv->parser);
226 if ( pv->context && pv->context->codec )
228 hb_avcodec_close( pv->context );
232 hb_list_close( &pv->list );
236 av_free( pv->buffer );
240 w->private_data = NULL;
244 /***********************************************************************
246 ***********************************************************************
248 **********************************************************************/
249 static int decavcodecWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
250 hb_buffer_t ** buf_out )
252 hb_work_private_t * pv = w->private_data;
253 hb_buffer_t * in = *buf_in, * buf, * last = NULL;
254 int pos, len, out_size, i, uncompressed_len;
257 unsigned char *parser_output_buffer;
258 int parser_output_buffer_len;
260 if ( (*buf_in)->size <= 0 )
262 /* EOF on input stream - send it downstream & say that we're done */
270 if ( in->start < -1 && pv->pts_next <= 0 )
272 // discard buffers that start before video time 0
276 cur = ( in->start < 0 )? pv->pts_next : in->start;
278 bufaligned = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
280 while( pos < in->size )
282 len = av_parser_parse2( pv->parser, pv->context,
283 &parser_output_buffer, &parser_output_buffer_len,
284 in->data + pos, in->size - pos, cur, cur, AV_NOPTS_VALUE );
286 uncompressed_len = 0;
287 if (parser_output_buffer_len)
290 av_init_packet( &avp );
291 avp.data = parser_output_buffer;
292 avp.size = parser_output_buffer_len;
294 out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
295 uncompressed_len = avcodec_decode_audio3( pv->context, bufaligned, &out_size, &avp );
302 buf = hb_buffer_init( 2 * out_size );
304 int sample_size_in_bytes = 2; // Default to 2 bytes
305 switch (pv->context->sample_fmt)
308 sample_size_in_bytes = 2;
310 /* We should handle other formats here - but that needs additional format conversion work below */
311 /* For now we'll just report the error and try to carry on */
313 hb_log("decavcodecWork - Unknown Sample Format from avcodec_decode_audio (%d) !", pv->context->sample_fmt);
318 buf->stop = cur + 90000 * ( out_size / (sample_size_in_bytes * pv->context->channels) ) /
319 pv->context->sample_rate;
323 fl32 = (float *) buf->data;
324 for( i = 0; i < out_size / 2; i++ )
331 last = last->next = buf;
335 *buf_out = last = buf;
344 av_free( bufaligned );
348 static int decavcodecInfo( hb_work_object_t *w, hb_work_info_t *info )
350 hb_work_private_t *pv = w->private_data;
352 memset( info, 0, sizeof(*info) );
354 if ( pv && pv->context )
356 AVCodecContext *context = pv->context;
357 info->bitrate = context->bit_rate;
358 info->rate = context->time_base.num;
359 info->rate_base = context->time_base.den;
360 info->profile = context->profile;
361 info->level = context->level;
367 static const int chan2layout[] = {
368 HB_INPUT_CH_LAYOUT_MONO, // We should allow no audio really.
369 HB_INPUT_CH_LAYOUT_MONO,
370 HB_INPUT_CH_LAYOUT_STEREO,
371 HB_INPUT_CH_LAYOUT_2F1R,
372 HB_INPUT_CH_LAYOUT_2F2R,
373 HB_INPUT_CH_LAYOUT_3F2R,
374 HB_INPUT_CH_LAYOUT_4F2R,
375 HB_INPUT_CH_LAYOUT_STEREO,
376 HB_INPUT_CH_LAYOUT_STEREO,
379 static int decavcodecBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
380 hb_work_info_t *info )
382 hb_work_private_t *pv = w->private_data;
385 memset( info, 0, sizeof(*info) );
387 if ( pv && pv->context )
389 return decavcodecInfo( w, info );
392 // We should parse the bitstream to find its parameters but for right
393 // now we just return dummy values if there's a codec that will handle it.
394 AVCodec *codec = avcodec_find_decoder( w->codec_param? w->codec_param :
398 // there's no ffmpeg codec for this audio type - give up
402 static char codec_name[64];
403 info->name = strncpy( codec_name, codec->name, sizeof(codec_name)-1 );
405 AVCodecParserContext *parser = av_parser_init( codec->id );
406 AVCodecContext *context = avcodec_alloc_context();
407 hb_avcodec_open( context, codec );
408 uint8_t *buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
409 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
410 unsigned char *pbuffer;
411 int pos = 0, pbuffer_size;
413 while ( pos < buf->size )
415 int len = av_parser_parse2( parser, context, &pbuffer, &pbuffer_size,
416 buf->data + pos, buf->size - pos,
417 buf->start, buf->start, AV_NOPTS_VALUE );
419 if ( pbuffer_size > 0 )
422 av_init_packet( &avp );
424 avp.size = pbuffer_size;
426 len = avcodec_decode_audio3( context, (int16_t*)buffer, &out_size, &avp );
427 if ( len > 0 && context->sample_rate > 0 )
429 info->bitrate = context->bit_rate;
430 info->rate = context->sample_rate;
432 info->channel_layout = chan2layout[context->channels & 7];
439 av_parser_close( parser );
440 hb_avcodec_close( context );
444 /* -------------------------------------------------------------
445 * General purpose video decoder using libavcodec
448 static uint8_t *copy_plane( uint8_t *dst, uint8_t* src, int dstride, int sstride,
451 if ( dstride == sstride )
453 memcpy( dst, src, dstride * h );
454 return dst + dstride * h;
456 int lbytes = dstride <= sstride? dstride : sstride;
459 memcpy( dst, src, lbytes );
466 // copy one video frame into an HB buf. If the frame isn't in our color space
467 // or at least one of its dimensions is odd, use sws_scale to convert/rescale it.
468 // Otherwise just copy the bits.
469 static hb_buffer_t *copy_frame( hb_work_private_t *pv, AVFrame *frame )
471 AVCodecContext *context = pv->context;
475 // if the dimensions are odd, drop the lsb since h264 requires that
476 // both width and height be even.
477 w = ( context->width >> 1 ) << 1;
478 h = ( context->height >> 1 ) << 1;
482 w = pv->job->title->width;
483 h = pv->job->title->height;
485 hb_buffer_t *buf = hb_video_buffer_init( w, h );
486 uint8_t *dst = buf->data;
488 if ( context->pix_fmt != PIX_FMT_YUV420P || w != context->width ||
489 h != context->height )
491 // have to convert to our internal color space and/or rescale
493 avpicture_fill( &dstpic, dst, PIX_FMT_YUV420P, w, h );
495 if ( ! pv->sws_context )
497 pv->sws_context = sws_getContext( context->width, context->height, context->pix_fmt,
498 w, h, PIX_FMT_YUV420P,
499 SWS_LANCZOS|SWS_ACCURATE_RND,
502 sws_scale( pv->sws_context, frame->data, frame->linesize, 0, h,
503 dstpic.data, dstpic.linesize );
507 dst = copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
508 w = (w + 1) >> 1; h = (h + 1) >> 1;
509 dst = copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
510 dst = copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
515 static int get_frame_buf( AVCodecContext *context, AVFrame *frame )
517 hb_work_private_t *pv = context->opaque;
518 frame->pts = pv->pts;
520 return avcodec_default_get_buffer( context, frame );
523 static void log_chapter( hb_work_private_t *pv, int chap_num, int64_t pts )
525 hb_chapter_t *c = hb_list_item( pv->job->title->list_chapter, chap_num - 1 );
528 hb_log( "%s: \"%s\" (%d) at frame %u time %"PRId64,
529 pv->context->codec->name, c->title, chap_num, pv->nframes, pts );
533 hb_log( "%s: Chapter %d at frame %u time %"PRId64,
534 pv->context->codec->name, chap_num, pv->nframes, pts );
538 static void flushDelayQueue( hb_work_private_t *pv )
541 int slot = pv->nframes & (HEAP_SIZE-1);
543 // flush all the video packets left on our timestamp-reordering delay q
544 while ( ( buf = pv->delayq[slot] ) != NULL )
546 buf->start = heap_pop( &pv->pts_heap );
547 hb_list_add( pv->list, buf );
548 pv->delayq[slot] = NULL;
549 slot = ( slot + 1 ) & (HEAP_SIZE-1);
553 static int decodeFrame( hb_work_private_t *pv, uint8_t *data, int size )
555 int got_picture, oldlevel = 0;
559 if ( global_verbosity_level <= 1 )
561 oldlevel = av_log_get_level();
562 av_log_set_level( AV_LOG_QUIET );
565 av_init_packet( &avp );
568 if ( avcodec_decode_video2( pv->context, &frame, &got_picture, &avp ) < 0 )
572 if ( global_verbosity_level <= 1 )
574 av_log_set_level( oldlevel );
578 // ffmpeg makes it hard to attach a pts to a frame. if the MPEG ES
579 // packet had a pts we handed it to av_parser_parse (if the packet had
580 // no pts we set it to -1 but before the parse we can't distinguish between
581 // the start of a video frame with no pts & an intermediate packet of
582 // some frame which never has a pts). we hope that when parse returns
583 // the frame to us the pts we originally handed it will be in parser->pts.
584 // we put this pts into pv->pts so that when a avcodec_decode_video
585 // finally gets around to allocating an AVFrame to hold the decoded
586 // frame we can stuff that pts into the frame. if all of these relays
587 // worked at this point frame.pts should hold the frame's pts from the
588 // original data stream or -1 if it didn't have one. in the latter case
589 // we generate the next pts in sequence for it.
590 double frame_dur = pv->duration;
591 if ( frame_dur <= 0 )
593 frame_dur = 90000. * (double)pv->context->time_base.num /
594 (double)pv->context->time_base.den;
595 pv->duration = frame_dur;
597 if ( frame.repeat_pict )
599 frame_dur += frame.repeat_pict * frame_dur * 0.5;
601 // XXX Unlike every other video decoder, the Raw decoder doesn't
602 // use the standard buffer allocation routines so we never
603 // get to put a PTS in the frame. Do it now.
604 if ( pv->context->codec_id == CODEC_ID_RAWVIDEO )
609 // If there was no pts for this frame, assume constant frame rate
610 // video & estimate the next frame time from the last & duration.
611 double pts = frame.pts;
616 pv->pts_next = pts + frame_dur;
620 // if we're doing a scan or this content couldn't have been broken
621 // by Microsoft we don't worry about timestamp reordering
622 if ( ! pv->job || ! pv->brokenByMicrosoft )
624 buf = copy_frame( pv, &frame );
626 hb_list_add( pv->list, buf );
631 // XXX This following probably addresses a libavcodec bug but I don't
632 // see an easy fix so we workaround it here.
634 // The M$ 'packed B-frames' atrocity results in decoded frames with
635 // the wrong timestamp. E.g., if there are 2 b-frames the timestamps
636 // we see here will be "2 3 1 5 6 4 ..." instead of "1 2 3 4 5 6".
637 // The frames are actually delivered in the right order but with
638 // the wrong timestamp. To get the correct timestamp attached to
639 // each frame we have a delay queue (longer than the max number of
640 // b-frames) & a sorting heap for the timestamps. As each frame
641 // comes out of the decoder the oldest frame in the queue is removed
642 // and associated with the smallest timestamp. Then the new frame is
643 // added to the queue & its timestamp is pushed on the heap.
644 // This does nothing if the timestamps are correct (i.e., the video
645 // uses a codec that Micro$oft hasn't broken yet) but the frames
646 // get timestamped correctly even when M$ has munged them.
648 // remove the oldest picture from the frame queue (if any) &
649 // give it the smallest timestamp from our heap. The queue size
650 // is a power of two so we get the slot of the oldest by masking
651 // the frame count & this will become the slot of the newest
652 // once we've removed & processed the oldest.
653 int slot = pv->nframes & (HEAP_SIZE-1);
654 if ( ( buf = pv->delayq[slot] ) != NULL )
656 buf->start = heap_pop( &pv->pts_heap );
658 if ( pv->new_chap && buf->start >= pv->chap_time )
660 buf->new_chap = pv->new_chap;
663 log_chapter( pv, buf->new_chap, buf->start );
665 else if ( pv->nframes == 0 )
667 log_chapter( pv, pv->job->chapter_start, buf->start );
669 hb_list_add( pv->list, buf );
672 // add the new frame to the delayq & push its timestamp on the heap
673 pv->delayq[slot] = copy_frame( pv, &frame );
674 heap_push( &pv->pts_heap, pts );
682 static void decodeVideo( hb_work_private_t *pv, uint8_t *data, int size,
683 int64_t pts, int64_t dts )
686 * The following loop is a do..while because we need to handle both
687 * data & the flush at the end (signaled by size=0). At the end there's
688 * generally a frame in the parser & one or more frames in the decoder
689 * (depending on the bframes setting).
695 int len = av_parser_parse2( pv->parser, pv->context, &pout, &pout_len,
696 data + pos, size - pos, pts, dts, AV_NOPTS_VALUE );
701 pv->pts = pv->parser->pts;
702 decodeFrame( pv, pout, pout_len );
704 } while ( pos < size );
706 /* the stuff above flushed the parser, now flush the decoder */
709 while ( decodeFrame( pv, NULL, 0 ) )
712 flushDelayQueue( pv );
716 static hb_buffer_t *link_buf_list( hb_work_private_t *pv )
718 hb_buffer_t *head = hb_list_item( pv->list, 0 );
722 hb_list_rem( pv->list, head );
724 hb_buffer_t *last = head, *buf;
726 while ( ( buf = hb_list_item( pv->list, 0 ) ) != NULL )
728 hb_list_rem( pv->list, buf );
737 static int decavcodecvInit( hb_work_object_t * w, hb_job_t * job )
740 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
741 w->private_data = pv;
743 pv->list = hb_list_init();
745 int codec_id = w->codec_param;
746 pv->parser = av_parser_init( codec_id );
747 pv->context = avcodec_alloc_context2( CODEC_TYPE_VIDEO );
749 /* we have to wrap ffmpeg's get_buffer to be able to set the pts (?!) */
750 pv->context->opaque = pv;
751 pv->context->get_buffer = get_frame_buf;
756 static int next_hdr( hb_buffer_t *in, int offset )
758 uint8_t *dat = in->data;
759 uint16_t last2 = 0xffff;
760 for ( ; in->size - offset > 1; ++offset )
762 if ( last2 == 0 && dat[offset] == 0x01 )
763 // found an mpeg start code
766 last2 = ( last2 << 8 ) | dat[offset];
772 static int find_hdr( hb_buffer_t *in, int offset, uint8_t hdr_type )
774 if ( in->size - offset < 4 )
775 // not enough room for an mpeg start code
778 for ( ; ( offset = next_hdr( in, offset ) ) >= 0; ++offset )
780 if ( in->data[offset+3] == hdr_type )
787 static int setup_extradata( hb_work_object_t *w, hb_buffer_t *in )
789 hb_work_private_t *pv = w->private_data;
791 // we can't call the avstream funcs but the read_header func in the
792 // AVInputFormat may set up some state in the AVContext. In particular
793 // vc1t_read_header allocates 'extradata' to deal with header issues
794 // related to Microsoft's bizarre engineering notions. We alloc a chunk
795 // of space to make vc1 work then associate the codec with the context.
796 if ( w->codec_param != CODEC_ID_VC1 )
798 // we haven't been inflicted with M$ - allocate a little space as
799 // a marker and return success.
800 pv->context->extradata_size = 16;
801 pv->context->extradata = av_malloc(pv->context->extradata_size);
805 // find the start and and of the sequence header
807 if ( ( shdr = find_hdr( in, 0, 0x0f ) ) < 0 )
809 // didn't find start of seq hdr
812 if ( ( shdr_end = next_hdr( in, shdr + 4 ) ) < 0 )
818 // find the start and and of the entry point header
820 if ( ( ehdr = find_hdr( in, 0, 0x0e ) ) < 0 )
822 // didn't find start of entry point hdr
825 if ( ( ehdr_end = next_hdr( in, ehdr + 4 ) ) < 0 )
831 // found both headers - allocate an extradata big enough to hold both
832 // then copy them into it.
833 pv->context->extradata_size = shdr_end + ehdr_end;
834 pv->context->extradata = av_malloc(pv->context->extradata_size + 8);
835 memcpy( pv->context->extradata, in->data + shdr, shdr_end );
836 memcpy( pv->context->extradata + shdr_end, in->data + ehdr, ehdr_end );
837 memset( pv->context->extradata + shdr_end + ehdr_end, 0, 8);
841 static int decavcodecvWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
842 hb_buffer_t ** buf_out )
844 hb_work_private_t *pv = w->private_data;
845 hb_buffer_t *in = *buf_in;
846 int64_t pts = AV_NOPTS_VALUE;
851 /* if we got an empty buffer signaling end-of-stream send it downstream */
854 decodeVideo( pv, in->data, in->size, pts, dts );
855 hb_list_add( pv->list, in );
856 *buf_out = link_buf_list( pv );
860 // if this is the first frame open the codec (we have to wait for the
861 // first frame because of M$ VC1 braindamage).
862 if ( pv->context->extradata_size == 0 )
864 if ( setup_extradata( w, in ) )
866 // we didn't find the headers needed to set up extradata.
867 // the codec will abort if we open it so just free the buf
868 // and hope we eventually get the info we need.
869 hb_buffer_close( &in );
872 AVCodec *codec = avcodec_find_decoder( w->codec_param );
873 // There's a mis-feature in ffmpeg that causes the context to be
874 // incorrectly initialized the 1st time avcodec_open is called.
875 // If you close it and open a 2nd time, it finishes the job.
876 hb_avcodec_open( pv->context, codec );
877 hb_avcodec_close( pv->context );
878 hb_avcodec_open( pv->context, codec );
884 dts = in->renderOffset;
888 pv->new_chap = in->new_chap;
889 pv->chap_time = pts >= 0? pts : pv->pts_next;
891 decodeVideo( pv, in->data, in->size, pts, dts );
892 hb_buffer_close( &in );
893 *buf_out = link_buf_list( pv );
897 static int decavcodecvInfo( hb_work_object_t *w, hb_work_info_t *info )
899 hb_work_private_t *pv = w->private_data;
901 memset( info, 0, sizeof(*info) );
903 if ( pv && pv->context )
905 AVCodecContext *context = pv->context;
906 info->bitrate = context->bit_rate;
907 info->width = context->width;
908 info->height = context->height;
910 /* ffmpeg gives the frame rate in frames per second while HB wants
911 * it in units of the 27MHz MPEG clock. */
912 info->rate = 27000000;
913 info->rate_base = (int64_t)context->time_base.num * 27000000LL /
914 context->time_base.den;
915 if ( context->ticks_per_frame > 1 )
917 // for ffmpeg 0.5 & later, the H.264 & MPEG-2 time base is
918 // field rate rather than frame rate so convert back to frames.
919 info->rate_base *= context->ticks_per_frame;
922 /* Sometimes there's no pixel aspect set in the source. In that case,
923 assume a 1:1 PAR. Otherwise, preserve the source PAR. */
924 info->pixel_aspect_width = context->sample_aspect_ratio.num ?
925 context->sample_aspect_ratio.num : 1;
926 info->pixel_aspect_height = context->sample_aspect_ratio.den ?
927 context->sample_aspect_ratio.den : 1;
929 /* ffmpeg returns the Pixel Aspect Ratio (PAR). Handbrake wants the
930 * Display Aspect Ratio so we convert by scaling by the Storage
931 * Aspect Ratio (w/h). We do the calc in floating point to get the
933 info->aspect = (double)info->pixel_aspect_width *
934 (double)context->width /
935 (double)info->pixel_aspect_height /
936 (double)context->height;
938 info->profile = context->profile;
939 info->level = context->level;
940 info->name = context->codec->name;
946 static int decavcodecvBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
947 hb_work_info_t *info )
952 hb_work_object_t hb_decavcodecv =
955 "Video decoder (libavcodec)",
964 // This is a special decoder for ffmpeg streams. The ffmpeg stream reader
965 // includes a parser and passes information from the parser to the decoder
966 // via a codec context kept in the AVStream of the reader's AVFormatContext.
967 // We *have* to use that codec context to decode the stream or we'll get
968 // garbage. ffmpeg_title_scan put a cookie that can be used to get to that
969 // codec context in our codec_param.
971 // this routine gets the appropriate context pointer from the ffmpeg
972 // stream reader. it can't be called until we get the first buffer because
973 // we can't guarantee that reader will be called before the our init
974 // routine and if our init is called first we'll get a pointer to the
975 // old scan stream (which has already been closed).
976 static void init_ffmpeg_context( hb_work_object_t *w )
978 hb_work_private_t *pv = w->private_data;
979 pv->context = hb_ffmpeg_context( w->codec_param );
981 // during scan the decoder gets closed & reopened which will
982 // close the codec so reopen it if it's not there
983 if ( ! pv->context->codec )
985 AVCodec *codec = avcodec_find_decoder( pv->context->codec_id );
986 hb_avcodec_open( pv->context, codec );
988 // set up our best guess at the frame duration.
989 // the frame rate in the codec is usually bogus but it's sometimes
991 AVStream *st = hb_ffmpeg_avstream( w->codec_param );
993 if ( st->nb_frames && st->duration )
995 // compute the average frame duration from the total number
996 // of frames & the total duration.
997 pv->duration = ( (double)st->duration * (double)st->time_base.num ) /
998 ( (double)st->nb_frames * (double)st->time_base.den );
1002 // XXX We don't have a frame count or duration so try to use the
1003 // far less reliable time base info in the stream.
1004 // Because the time bases are so screwed up, we only take values
1005 // in the range 8fps - 64fps.
1007 if ( st->time_base.num * 64 > st->time_base.den &&
1008 st->time_base.den > st->time_base.num * 8 )
1012 else if ( st->r_frame_rate.den * 64 > st->r_frame_rate.num &&
1013 st->r_frame_rate.num > st->r_frame_rate.den * 8 )
1015 tb.num = st->r_frame_rate.den;
1016 tb.den = st->r_frame_rate.num;
1020 tb.num = 1001; /*XXX*/
1021 tb.den = 24000; /*XXX*/
1023 pv->duration = (double)tb.num / (double)tb.den;
1025 pv->duration *= 90000.;
1027 // we have to wrap ffmpeg's get_buffer to be able to set the pts (?!)
1028 pv->context->opaque = pv;
1029 pv->context->get_buffer = get_frame_buf;
1031 // avi, mkv and possibly mp4 containers can contain the M$ VFW packed
1032 // b-frames abortion that messes up frame ordering and timestamps.
1033 // XXX ffmpeg knows which streams are broken but doesn't expose the
1034 // info externally. We should patch ffmpeg to add a flag to the
1035 // codec context for this but until then we mark all ffmpeg streams
1037 pv->brokenByMicrosoft = 1;
1040 static void prepare_ffmpeg_buffer( hb_buffer_t * in )
1042 // ffmpeg requires an extra 8 bytes of zero at the end of the buffer and
1043 // will seg fault in odd, data dependent ways if it's not there. (my guess
1044 // is this is a case of a local performance optimization creating a global
1045 // performance degradation since all the time wasted by extraneous data
1046 // copies & memory zeroing has to be huge compared to the minor reduction
1047 // in inner-loop instructions this affords - modern cpus bottleneck on
1048 // memory bandwidth not instruction bandwidth).
1049 if ( in->size + FF_INPUT_BUFFER_PADDING_SIZE > in->alloc )
1051 // have to realloc to add the padding
1052 hb_buffer_realloc( in, in->size + FF_INPUT_BUFFER_PADDING_SIZE );
1054 memset( in->data + in->size, 0, FF_INPUT_BUFFER_PADDING_SIZE );
1057 static int decavcodecviInit( hb_work_object_t * w, hb_job_t * job )
1060 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
1061 w->private_data = pv;
1063 pv->list = hb_list_init();
1069 static int decavcodecviWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
1070 hb_buffer_t ** buf_out )
1072 hb_work_private_t *pv = w->private_data;
1073 if ( ! pv->context )
1075 init_ffmpeg_context( w );
1077 hb_buffer_t *in = *buf_in;
1080 /* if we got an empty buffer signaling end-of-stream send it downstream */
1081 if ( in->size == 0 )
1083 /* flush any frames left in the decoder */
1084 while ( decodeFrame( pv, NULL, 0 ) )
1087 flushDelayQueue( pv );
1088 hb_list_add( pv->list, in );
1089 *buf_out = link_buf_list( pv );
1090 return HB_WORK_DONE;
1093 int64_t pts = in->start;
1096 // use the first timestamp as our 'next expected' pts
1097 if ( pv->pts_next < 0 )
1106 pv->new_chap = in->new_chap;
1107 pv->chap_time = pts >= 0? pts : pv->pts_next;
1109 prepare_ffmpeg_buffer( in );
1110 decodeFrame( pv, in->data, in->size );
1111 hb_buffer_close( &in );
1112 *buf_out = link_buf_list( pv );
1116 static int decavcodecviInfo( hb_work_object_t *w, hb_work_info_t *info )
1118 if ( decavcodecvInfo( w, info ) )
1120 hb_work_private_t *pv = w->private_data;
1121 if ( ! pv->context )
1123 init_ffmpeg_context( w );
1125 // we have the frame duration in units of the 90KHz pts clock but
1126 // need it in units of the 27MHz MPEG clock. */
1127 info->rate = 27000000;
1128 info->rate_base = pv->duration * 300.;
1134 static void decodeAudio( hb_work_private_t *pv, uint8_t *data, int size )
1136 AVCodecContext *context = pv->context;
1139 while ( pos < size )
1141 int16_t *buffer = pv->buffer;
1142 if ( buffer == NULL )
1144 pv->buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
1145 buffer = pv->buffer;
1149 av_init_packet( &avp );
1150 avp.data = data + pos;
1151 avp.size = size - pos;
1153 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
1154 int len = avcodec_decode_audio3( context, buffer, &out_size, &avp );
1162 // We require signed 16-bit ints for the output format. If
1163 // we got something different convert it.
1164 if ( context->sample_fmt != SAMPLE_FMT_S16 )
1166 // Note: av_audio_convert seems to be a work-in-progress but
1167 // looks like it will eventually handle general audio
1168 // mixdowns which would allow us much more flexibility
1169 // in handling multichannel audio in HB. If we were doing
1170 // anything more complicated than a one-for-one format
1171 // conversion we'd probably want to cache the converter
1172 // context in the pv.
1173 int isamp = av_get_bits_per_sample_format( context->sample_fmt ) / 8;
1174 AVAudioConvert *ctx = av_audio_convert_alloc( SAMPLE_FMT_S16, 1,
1175 context->sample_fmt, 1,
1177 // get output buffer size (in 2-byte samples) then malloc a buffer
1178 out_size = ( out_size * 2 ) / isamp;
1179 buffer = av_malloc( out_size );
1181 // we're doing straight sample format conversion which behaves as if
1182 // there were only one channel.
1183 const void * const ibuf[6] = { pv->buffer };
1184 void * const obuf[6] = { buffer };
1185 const int istride[6] = { isamp };
1186 const int ostride[6] = { 2 };
1188 av_audio_convert( ctx, obuf, ostride, ibuf, istride, out_size >> 1 );
1189 av_audio_convert_free( ctx );
1191 hb_buffer_t *buf = hb_buffer_init( 2 * out_size );
1193 // convert from bytes to total samples
1196 double pts = pv->pts_next;
1198 pts += out_size * pv->duration;
1202 float *fl32 = (float *)buf->data;
1204 for( i = 0; i < out_size; ++i )
1206 fl32[i] = buffer[i];
1208 hb_list_add( pv->list, buf );
1210 // if we allocated a buffer for sample format conversion, free it
1211 if ( buffer != pv->buffer )
1219 static int decavcodecaiWork( hb_work_object_t *w, hb_buffer_t **buf_in,
1220 hb_buffer_t **buf_out )
1222 if ( (*buf_in)->size <= 0 )
1224 /* EOF on input stream - send it downstream & say that we're done */
1227 return HB_WORK_DONE;
1230 hb_work_private_t *pv = w->private_data;
1232 if ( (*buf_in)->start < -1 && pv->pts_next <= 0 )
1234 // discard buffers that start before video time 0
1239 if ( ! pv->context )
1241 init_ffmpeg_context( w );
1242 // duration is a scaling factor to go from #bytes in the decoded
1243 // frame to frame time (in 90KHz mpeg ticks). 'channels' converts
1244 // total samples to per-channel samples. 'sample_rate' converts
1245 // per-channel samples to seconds per sample and the 90000
1246 // is mpeg ticks per second.
1247 pv->duration = 90000. /
1248 (double)( pv->context->sample_rate * pv->context->channels );
1250 hb_buffer_t *in = *buf_in;
1252 // if the packet has a timestamp use it if we don't have a timestamp yet
1253 // or if there's been a timing discontinuity of more than 100ms.
1254 if ( in->start >= 0 &&
1255 ( pv->pts_next < 0 || ( in->start - pv->pts_next ) > 90*100 ) )
1257 pv->pts_next = in->start;
1259 prepare_ffmpeg_buffer( in );
1260 decodeAudio( pv, in->data, in->size );
1261 *buf_out = link_buf_list( pv );
1266 hb_work_object_t hb_decavcodecvi =
1269 "Video decoder (ffmpeg streams)",
1277 hb_work_object_t hb_decavcodecai =
1280 "Audio decoder (ffmpeg streams)",