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"
66 #include "../contrib/ffmpeg/libavcodec/audioconvert.h"
68 static int decavcodecInit( hb_work_object_t *, hb_job_t * );
69 static int decavcodecWork( hb_work_object_t *, hb_buffer_t **, hb_buffer_t ** );
70 static void decavcodecClose( hb_work_object_t * );
71 static int decavcodecInfo( hb_work_object_t *, hb_work_info_t * );
72 static int decavcodecBSInfo( hb_work_object_t *, const hb_buffer_t *, hb_work_info_t * );
74 hb_work_object_t hb_decavcodec =
77 "MPGA decoder (libavcodec)",
87 // there are nheap items on the heap indexed 1..nheap (i.e., top of
88 // heap is 1). The 0th slot is unused - a marker is put there to check
89 // for overwrite errs.
90 int64_t h[HEAP_SIZE+1];
94 struct hb_work_private_s
97 AVCodecContext *context;
98 AVCodecParserContext *parser;
100 double duration; // frame duration (for video)
101 double pts_next; // next pts we expect to generate
102 int64_t pts; // (video) pts passing from parser to decoder
103 int64_t chap_time; // time of next chap mark (if new_chap != 0)
104 int new_chap; // output chapter mark pending
107 uint32_t decode_errors;
108 int brokenByMicrosoft; // video stream may contain packed b-frames
109 hb_buffer_t* delayq[HEAP_SIZE];
112 struct SwsContext *sws_context; // if we have to rescale or convert color space
115 static int64_t heap_pop( pts_heap_t *heap )
119 if ( heap->nheap <= 0 )
124 // return the top of the heap then put the bottom element on top,
125 // decrease the heap size by one & rebalence the heap.
128 int64_t v = heap->h[heap->nheap--];
130 int child = parent << 1;
131 while ( child <= heap->nheap )
133 // find the smallest of the two children of parent
134 if (child < heap->nheap && heap->h[child] > heap->h[child+1] )
137 if (v <= heap->h[child])
138 // new item is smaller than either child so it's the new parent.
141 // smallest child is smaller than new item so move it up then
142 // check its children.
143 int64_t hp = heap->h[child];
144 heap->h[parent] = hp;
152 static void heap_push( pts_heap_t *heap, int64_t v )
154 if ( heap->nheap < HEAP_SIZE )
159 // stick the new value on the bottom of the heap then bubble it
160 // up to its correct spot.
161 int child = heap->nheap;
163 int parent = child >> 1;
164 if (heap->h[parent] <= v)
167 int64_t hp = heap->h[parent];
175 /***********************************************************************
176 * hb_work_decavcodec_init
177 ***********************************************************************
179 **********************************************************************/
180 static int decavcodecInit( hb_work_object_t * w, hb_job_t * job )
184 hb_work_private_t * pv = calloc( 1, sizeof( hb_work_private_t ) );
185 w->private_data = pv;
189 int codec_id = w->codec_param;
192 codec_id = CODEC_ID_MP2;
194 codec = avcodec_find_decoder( codec_id );
195 pv->parser = av_parser_init( codec_id );
197 pv->context = avcodec_alloc_context();
198 hb_avcodec_open( pv->context, codec );
203 /***********************************************************************
205 ***********************************************************************
207 **********************************************************************/
208 static void decavcodecClose( hb_work_object_t * w )
210 hb_work_private_t * pv = w->private_data;
214 if ( pv->job && pv->context && pv->context->codec )
216 hb_log( "%s-decoder done: %u frames, %u decoder errors, %u drops",
217 pv->context->codec->name, pv->nframes, pv->decode_errors,
220 if ( pv->sws_context )
222 sws_freeContext( pv->sws_context );
226 av_parser_close(pv->parser);
228 if ( pv->context && pv->context->codec )
230 hb_avcodec_close( pv->context );
234 hb_list_close( &pv->list );
242 w->private_data = NULL;
246 /***********************************************************************
248 ***********************************************************************
250 **********************************************************************/
251 static int decavcodecWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
252 hb_buffer_t ** buf_out )
254 hb_work_private_t * pv = w->private_data;
255 hb_buffer_t * in = *buf_in, * buf, * last = NULL;
256 int pos, len, out_size, i, uncompressed_len;
257 short buffer[AVCODEC_MAX_AUDIO_FRAME_SIZE];
259 unsigned char *parser_output_buffer;
260 int parser_output_buffer_len;
262 if ( (*buf_in)->size <= 0 )
264 /* EOF on input stream - send it downstream & say that we're done */
272 if ( in->start < -1 && pv->pts_next <= 0 )
274 // discard buffers that start before video time 0
278 cur = ( in->start < 0 )? pv->pts_next : in->start;
281 while( pos < in->size )
283 len = av_parser_parse( pv->parser, pv->context,
284 &parser_output_buffer, &parser_output_buffer_len,
285 in->data + pos, in->size - pos, cur, cur );
287 uncompressed_len = 0;
288 if (parser_output_buffer_len)
290 out_size = sizeof(buffer);
291 uncompressed_len = avcodec_decode_audio2( pv->context, buffer,
293 parser_output_buffer,
294 parser_output_buffer_len );
301 buf = hb_buffer_init( 2 * out_size );
303 int sample_size_in_bytes = 2; // Default to 2 bytes
304 switch (pv->context->sample_fmt)
307 sample_size_in_bytes = 2;
309 /* We should handle other formats here - but that needs additional format conversion work below */
310 /* For now we'll just report the error and try to carry on */
312 hb_log("decavcodecWork - Unknown Sample Format from avcodec_decode_audio (%d) !", pv->context->sample_fmt);
317 buf->stop = cur + 90000 * ( out_size / (sample_size_in_bytes * pv->context->channels) ) /
318 pv->context->sample_rate;
322 fl32 = (float *) buf->data;
323 for( i = 0; i < out_size / 2; i++ )
330 last = last->next = buf;
334 *buf_out = last = buf;
346 static int decavcodecInfo( hb_work_object_t *w, hb_work_info_t *info )
348 hb_work_private_t *pv = w->private_data;
350 memset( info, 0, sizeof(*info) );
352 if ( pv && pv->context )
354 AVCodecContext *context = pv->context;
355 info->bitrate = context->bit_rate;
356 info->rate = context->time_base.num;
357 info->rate_base = context->time_base.den;
358 info->profile = context->profile;
359 info->level = context->level;
365 static const int chan2layout[] = {
366 HB_INPUT_CH_LAYOUT_MONO, // We should allow no audio really.
367 HB_INPUT_CH_LAYOUT_MONO,
368 HB_INPUT_CH_LAYOUT_STEREO,
369 HB_INPUT_CH_LAYOUT_2F1R,
370 HB_INPUT_CH_LAYOUT_2F2R,
371 HB_INPUT_CH_LAYOUT_3F2R,
372 HB_INPUT_CH_LAYOUT_4F2R,
373 HB_INPUT_CH_LAYOUT_STEREO,
374 HB_INPUT_CH_LAYOUT_STEREO,
377 static int decavcodecBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
378 hb_work_info_t *info )
380 hb_work_private_t *pv = w->private_data;
383 memset( info, 0, sizeof(*info) );
385 if ( pv && pv->context )
387 return decavcodecInfo( w, info );
390 // We should parse the bitstream to find its parameters but for right
391 // now we just return dummy values if there's a codec that will handle it.
392 AVCodec *codec = avcodec_find_decoder( w->codec_param? w->codec_param :
396 // there's no ffmpeg codec for this audio type - give up
400 static char codec_name[64];
401 info->name = strncpy( codec_name, codec->name, sizeof(codec_name)-1 );
403 AVCodecParserContext *parser = av_parser_init( codec->id );
404 AVCodecContext *context = avcodec_alloc_context();
405 hb_avcodec_open( context, codec );
407 uint8_t *buffer = memalign(16, AVCODEC_MAX_AUDIO_FRAME_SIZE);
409 uint8_t *buffer = malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
411 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
412 unsigned char *pbuffer;
413 int pos = 0, pbuffer_size;
415 while ( pos < buf->size )
417 int len = av_parser_parse( parser, context, &pbuffer, &pbuffer_size,
418 buf->data + pos, buf->size - pos,
419 buf->start, buf->start );
421 if ( pbuffer_size > 0 )
423 len = avcodec_decode_audio2( context, (int16_t*)buffer, &out_size,
424 pbuffer, pbuffer_size );
425 if ( len > 0 && context->sample_rate > 0 )
427 info->bitrate = context->bit_rate;
428 info->rate = context->sample_rate;
430 info->channel_layout = chan2layout[context->channels & 7];
437 av_parser_close( parser );
438 hb_avcodec_close( context );
442 /* -------------------------------------------------------------
443 * General purpose video decoder using libavcodec
446 static uint8_t *copy_plane( uint8_t *dst, uint8_t* src, int dstride, int sstride,
449 if ( dstride == sstride )
451 memcpy( dst, src, dstride * h );
452 return dst + dstride * h;
454 int lbytes = dstride <= sstride? dstride : sstride;
457 memcpy( dst, src, lbytes );
464 // copy one video frame into an HB buf. If the frame isn't in our color space
465 // or at least one of its dimensions is odd, use sws_scale to convert/rescale it.
466 // Otherwise just copy the bits.
467 static hb_buffer_t *copy_frame( hb_work_private_t *pv, AVFrame *frame )
469 AVCodecContext *context = pv->context;
473 // if the dimensions are odd, drop the lsb since h264 requires that
474 // both width and height be even.
475 w = ( context->width >> 1 ) << 1;
476 h = ( context->height >> 1 ) << 1;
480 w = pv->job->title->width;
481 h = pv->job->title->height;
483 hb_buffer_t *buf = hb_video_buffer_init( w, h );
484 uint8_t *dst = buf->data;
486 if ( context->pix_fmt != PIX_FMT_YUV420P || w != context->width ||
487 h != context->height )
489 // have to convert to our internal color space and/or rescale
491 avpicture_fill( &dstpic, dst, PIX_FMT_YUV420P, w, h );
493 if ( ! pv->sws_context )
495 pv->sws_context = sws_getContext( context->width, context->height, context->pix_fmt,
496 w, h, PIX_FMT_YUV420P,
497 SWS_LANCZOS|SWS_ACCURATE_RND,
500 sws_scale( pv->sws_context, frame->data, frame->linesize, 0, h,
501 dstpic.data, dstpic.linesize );
505 dst = copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
506 w = (w + 1) >> 1; h = (h + 1) >> 1;
507 dst = copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
508 dst = copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
513 static int get_frame_buf( AVCodecContext *context, AVFrame *frame )
515 hb_work_private_t *pv = context->opaque;
516 frame->pts = pv->pts;
518 return avcodec_default_get_buffer( context, frame );
521 static void log_chapter( hb_work_private_t *pv, int chap_num, int64_t pts )
523 hb_chapter_t *c = hb_list_item( pv->job->title->list_chapter, chap_num - 1 );
526 hb_log( "%s: \"%s\" (%d) at frame %u time %lld",
527 pv->context->codec->name, c->title, chap_num, pv->nframes, pts );
531 hb_log( "%s: Chapter %d at frame %u time %lld",
532 pv->context->codec->name, chap_num, pv->nframes, pts );
536 static void flushDelayQueue( hb_work_private_t *pv )
539 int slot = pv->nframes & (HEAP_SIZE-1);
541 // flush all the video packets left on our timestamp-reordering delay q
542 while ( ( buf = pv->delayq[slot] ) != NULL )
544 buf->start = heap_pop( &pv->pts_heap );
545 hb_list_add( pv->list, buf );
546 pv->delayq[slot] = NULL;
547 slot = ( slot + 1 ) & (HEAP_SIZE-1);
551 static int decodeFrame( hb_work_private_t *pv, uint8_t *data, int size )
556 if ( avcodec_decode_video( pv->context, &frame, &got_picture, data, size ) < 0 )
562 // ffmpeg makes it hard to attach a pts to a frame. if the MPEG ES
563 // packet had a pts we handed it to av_parser_parse (if the packet had
564 // no pts we set it to -1 but before the parse we can't distinguish between
565 // the start of a video frame with no pts & an intermediate packet of
566 // some frame which never has a pts). we hope that when parse returns
567 // the frame to us the pts we originally handed it will be in parser->pts.
568 // we put this pts into pv->pts so that when a avcodec_decode_video
569 // finally gets around to allocating an AVFrame to hold the decoded
570 // frame we can stuff that pts into the frame. if all of these relays
571 // worked at this point frame.pts should hold the frame's pts from the
572 // original data stream or -1 if it didn't have one. in the latter case
573 // we generate the next pts in sequence for it.
574 double frame_dur = pv->duration;
575 if ( frame_dur <= 0 )
577 frame_dur = 90000. * (double)pv->context->time_base.num /
578 (double)pv->context->time_base.den;
579 pv->duration = frame_dur;
581 if ( frame.repeat_pict )
583 frame_dur += frame.repeat_pict * frame_dur * 0.5;
585 // If there was no pts for this frame, assume constant frame rate
586 // video & estimate the next frame time from the last & duration.
587 double pts = frame.pts;
592 pv->pts_next = pts + frame_dur;
596 // if we're doing a scan or this content couldn't have been broken
597 // by Microsoft we don't worry about timestamp reordering
598 if ( ! pv->job || ! pv->brokenByMicrosoft )
600 buf = copy_frame( pv, &frame );
602 hb_list_add( pv->list, buf );
607 // XXX This following probably addresses a libavcodec bug but I don't
608 // see an easy fix so we workaround it here.
610 // The M$ 'packed B-frames' atrocity results in decoded frames with
611 // the wrong timestamp. E.g., if there are 2 b-frames the timestamps
612 // we see here will be "2 3 1 5 6 4 ..." instead of "1 2 3 4 5 6".
613 // The frames are actually delivered in the right order but with
614 // the wrong timestamp. To get the correct timestamp attached to
615 // each frame we have a delay queue (longer than the max number of
616 // b-frames) & a sorting heap for the timestamps. As each frame
617 // comes out of the decoder the oldest frame in the queue is removed
618 // and associated with the smallest timestamp. Then the new frame is
619 // added to the queue & its timestamp is pushed on the heap.
620 // This does nothing if the timestamps are correct (i.e., the video
621 // uses a codec that Micro$oft hasn't broken yet) but the frames
622 // get timestamped correctly even when M$ has munged them.
624 // remove the oldest picture from the frame queue (if any) &
625 // give it the smallest timestamp from our heap. The queue size
626 // is a power of two so we get the slot of the oldest by masking
627 // the frame count & this will become the slot of the newest
628 // once we've removed & processed the oldest.
629 int slot = pv->nframes & (HEAP_SIZE-1);
630 if ( ( buf = pv->delayq[slot] ) != NULL )
632 buf->start = heap_pop( &pv->pts_heap );
634 if ( pv->new_chap && buf->start >= pv->chap_time )
636 buf->new_chap = pv->new_chap;
639 log_chapter( pv, buf->new_chap, buf->start );
641 else if ( pv->nframes == 0 )
643 log_chapter( pv, pv->job->chapter_start, buf->start );
645 hb_list_add( pv->list, buf );
648 // add the new frame to the delayq & push its timestamp on the heap
649 pv->delayq[slot] = copy_frame( pv, &frame );
650 heap_push( &pv->pts_heap, pts );
658 static void decodeVideo( hb_work_private_t *pv, uint8_t *data, int size,
659 int64_t pts, int64_t dts )
662 * The following loop is a do..while because we need to handle both
663 * data & the flush at the end (signaled by size=0). At the end there's
664 * generally a frame in the parser & one or more frames in the decoder
665 * (depending on the bframes setting).
671 int len = av_parser_parse( pv->parser, pv->context, &pout, &pout_len,
672 data + pos, size - pos, pts, dts );
677 pv->pts = pv->parser->pts;
678 decodeFrame( pv, pout, pout_len );
680 } while ( pos < size );
682 /* the stuff above flushed the parser, now flush the decoder */
685 while ( decodeFrame( pv, NULL, 0 ) )
688 flushDelayQueue( pv );
692 static hb_buffer_t *link_buf_list( hb_work_private_t *pv )
694 hb_buffer_t *head = hb_list_item( pv->list, 0 );
698 hb_list_rem( pv->list, head );
700 hb_buffer_t *last = head, *buf;
702 while ( ( buf = hb_list_item( pv->list, 0 ) ) != NULL )
704 hb_list_rem( pv->list, buf );
713 static int decavcodecvInit( hb_work_object_t * w, hb_job_t * job )
716 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
717 w->private_data = pv;
719 pv->list = hb_list_init();
721 int codec_id = w->codec_param;
722 pv->parser = av_parser_init( codec_id );
723 pv->context = avcodec_alloc_context2( CODEC_TYPE_VIDEO );
725 /* we have to wrap ffmpeg's get_buffer to be able to set the pts (?!) */
726 pv->context->opaque = pv;
727 pv->context->get_buffer = get_frame_buf;
732 static int next_hdr( hb_buffer_t *in, int offset )
734 uint8_t *dat = in->data;
735 uint16_t last2 = 0xffff;
736 for ( ; in->size - offset > 1; ++offset )
738 if ( last2 == 0 && dat[offset] == 0x01 )
739 // found an mpeg start code
742 last2 = ( last2 << 8 ) | dat[offset];
748 static int find_hdr( hb_buffer_t *in, int offset, uint8_t hdr_type )
750 if ( in->size - offset < 4 )
751 // not enough room for an mpeg start code
754 for ( ; ( offset = next_hdr( in, offset ) ) >= 0; ++offset )
756 if ( in->data[offset+3] == hdr_type )
763 static int setup_extradata( hb_work_object_t *w, hb_buffer_t *in )
765 hb_work_private_t *pv = w->private_data;
767 // we can't call the avstream funcs but the read_header func in the
768 // AVInputFormat may set up some state in the AVContext. In particular
769 // vc1t_read_header allocates 'extradata' to deal with header issues
770 // related to Microsoft's bizarre engineering notions. We alloc a chunk
771 // of space to make vc1 work then associate the codec with the context.
772 if ( w->codec_param != CODEC_ID_VC1 )
774 // we haven't been inflicted with M$ - allocate a little space as
775 // a marker and return success.
776 pv->context->extradata_size = 16;
777 pv->context->extradata = av_malloc(pv->context->extradata_size);
781 // find the start and and of the sequence header
783 if ( ( shdr = find_hdr( in, 0, 0x0f ) ) < 0 )
785 // didn't find start of seq hdr
788 if ( ( shdr_end = next_hdr( in, shdr + 4 ) ) < 0 )
794 // find the start and and of the entry point header
796 if ( ( ehdr = find_hdr( in, 0, 0x0e ) ) < 0 )
798 // didn't find start of entry point hdr
801 if ( ( ehdr_end = next_hdr( in, ehdr + 4 ) ) < 0 )
807 // found both headers - allocate an extradata big enough to hold both
808 // then copy them into it.
809 pv->context->extradata_size = shdr_end + ehdr_end;
810 pv->context->extradata = av_malloc(pv->context->extradata_size + 8);
811 memcpy( pv->context->extradata, in->data + shdr, shdr_end );
812 memcpy( pv->context->extradata + shdr_end, in->data + ehdr, ehdr_end );
813 memset( pv->context->extradata + shdr_end + ehdr_end, 0, 8);
817 static int decavcodecvWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
818 hb_buffer_t ** buf_out )
820 hb_work_private_t *pv = w->private_data;
821 hb_buffer_t *in = *buf_in;
822 int64_t pts = AV_NOPTS_VALUE;
827 /* if we got an empty buffer signaling end-of-stream send it downstream */
830 decodeVideo( pv, in->data, in->size, pts, dts );
831 hb_list_add( pv->list, in );
832 *buf_out = link_buf_list( pv );
836 // if this is the first frame open the codec (we have to wait for the
837 // first frame because of M$ VC1 braindamage).
838 if ( pv->context->extradata_size == 0 )
840 if ( setup_extradata( w, in ) )
842 // we didn't find the headers needed to set up extradata.
843 // the codec will abort if we open it so just free the buf
844 // and hope we eventually get the info we need.
845 hb_buffer_close( &in );
848 AVCodec *codec = avcodec_find_decoder( w->codec_param );
849 // There's a mis-feature in ffmpeg that causes the context to be
850 // incorrectly initialized the 1st time avcodec_open is called.
851 // If you close it and open a 2nd time, it finishes the job.
852 hb_avcodec_open( pv->context, codec );
853 hb_avcodec_close( pv->context );
854 hb_avcodec_open( pv->context, codec );
860 dts = in->renderOffset;
864 pv->new_chap = in->new_chap;
865 pv->chap_time = pts >= 0? pts : pv->pts_next;
867 decodeVideo( pv, in->data, in->size, pts, dts );
868 hb_buffer_close( &in );
869 *buf_out = link_buf_list( pv );
873 static int decavcodecvInfo( hb_work_object_t *w, hb_work_info_t *info )
875 hb_work_private_t *pv = w->private_data;
877 memset( info, 0, sizeof(*info) );
879 if ( pv && pv->context )
881 AVCodecContext *context = pv->context;
882 info->bitrate = context->bit_rate;
883 info->width = context->width;
884 info->height = context->height;
886 /* ffmpeg gives the frame rate in frames per second while HB wants
887 * it in units of the 27MHz MPEG clock. */
888 info->rate = 27000000;
889 info->rate_base = (int64_t)context->time_base.num * 27000000LL /
890 context->time_base.den;
892 /* Sometimes there's no pixel aspect set in the source. In that case,
893 assume a 1:1 PAR. Otherwise, preserve the source PAR. */
894 info->pixel_aspect_width = context->sample_aspect_ratio.num ?
895 context->sample_aspect_ratio.num : 1;
896 info->pixel_aspect_height = context->sample_aspect_ratio.den ?
897 context->sample_aspect_ratio.den : 1;
899 /* ffmpeg returns the Pixel Aspect Ratio (PAR). Handbrake wants the
900 * Display Aspect Ratio so we convert by scaling by the Storage
901 * Aspect Ratio (w/h). We do the calc in floating point to get the
903 info->aspect = (double)info->pixel_aspect_width *
904 (double)context->width /
905 (double)info->pixel_aspect_height /
906 (double)context->height;
908 info->profile = context->profile;
909 info->level = context->level;
910 info->name = context->codec->name;
916 static int decavcodecvBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
917 hb_work_info_t *info )
922 hb_work_object_t hb_decavcodecv =
925 "Video decoder (libavcodec)",
934 // This is a special decoder for ffmpeg streams. The ffmpeg stream reader
935 // includes a parser and passes information from the parser to the decoder
936 // via a codec context kept in the AVStream of the reader's AVFormatContext.
937 // We *have* to use that codec context to decode the stream or we'll get
938 // garbage. ffmpeg_title_scan put a cookie that can be used to get to that
939 // codec context in our codec_param.
941 // this routine gets the appropriate context pointer from the ffmpeg
942 // stream reader. it can't be called until we get the first buffer because
943 // we can't guarantee that reader will be called before the our init
944 // routine and if our init is called first we'll get a pointer to the
945 // old scan stream (which has already been closed).
946 static void init_ffmpeg_context( hb_work_object_t *w )
948 hb_work_private_t *pv = w->private_data;
949 pv->context = hb_ffmpeg_context( w->codec_param );
951 // during scan the decoder gets closed & reopened which will
952 // close the codec so reopen it if it's not there
953 if ( ! pv->context->codec )
955 AVCodec *codec = avcodec_find_decoder( pv->context->codec_id );
956 hb_avcodec_open( pv->context, codec );
958 // set up our best guess at the frame duration.
959 // the frame rate in the codec is usually bogus but it's sometimes
961 AVStream *st = hb_ffmpeg_avstream( w->codec_param );
963 if ( st->nb_frames && st->duration )
965 // compute the average frame duration from the total number
966 // of frames & the total duration.
967 pv->duration = ( (double)st->duration * (double)st->time_base.num ) /
968 ( (double)st->nb_frames * (double)st->time_base.den );
972 // XXX We don't have a frame count or duration so try to use the
973 // far less reliable time base info in the stream.
974 // Because the time bases are so screwed up, we only take values
975 // in the range 8fps - 64fps.
977 if ( st->time_base.num * 64 > st->time_base.den &&
978 st->time_base.den > st->time_base.num * 8 )
982 else if ( st->r_frame_rate.den * 64 > st->r_frame_rate.num &&
983 st->r_frame_rate.num > st->r_frame_rate.den * 8 )
985 tb.num = st->r_frame_rate.den;
986 tb.den = st->r_frame_rate.num;
990 tb.num = 1001; /*XXX*/
991 tb.den = 24000; /*XXX*/
993 pv->duration = (double)tb.num / (double)tb.den;
995 pv->duration *= 90000.;
997 // we have to wrap ffmpeg's get_buffer to be able to set the pts (?!)
998 pv->context->opaque = pv;
999 pv->context->get_buffer = get_frame_buf;
1001 // avi, mkv and possibly mp4 containers can contain the M$ VFW packed
1002 // b-frames abortion that messes up frame ordering and timestamps.
1003 // XXX ffmpeg knows which streams are broken but doesn't expose the
1004 // info externally. We should patch ffmpeg to add a flag to the
1005 // codec context for this but until then we mark all ffmpeg streams
1007 pv->brokenByMicrosoft = 1;
1010 static void prepare_ffmpeg_buffer( hb_buffer_t * in )
1012 // ffmpeg requires an extra 8 bytes of zero at the end of the buffer and
1013 // will seg fault in odd, data dependent ways if it's not there. (my guess
1014 // is this is a case of a local performance optimization creating a global
1015 // performance degradation since all the time wasted by extraneous data
1016 // copies & memory zeroing has to be huge compared to the minor reduction
1017 // in inner-loop instructions this affords - modern cpus bottleneck on
1018 // memory bandwidth not instruction bandwidth).
1019 if ( in->size + FF_INPUT_BUFFER_PADDING_SIZE > in->alloc )
1021 // have to realloc to add the padding
1022 hb_buffer_realloc( in, in->size + FF_INPUT_BUFFER_PADDING_SIZE );
1024 memset( in->data + in->size, 0, FF_INPUT_BUFFER_PADDING_SIZE );
1027 static int decavcodecviInit( hb_work_object_t * w, hb_job_t * job )
1030 hb_work_private_t *pv = calloc( 1, sizeof( hb_work_private_t ) );
1031 w->private_data = pv;
1033 pv->list = hb_list_init();
1039 static int decavcodecviWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
1040 hb_buffer_t ** buf_out )
1042 hb_work_private_t *pv = w->private_data;
1043 if ( ! pv->context )
1045 init_ffmpeg_context( w );
1047 hb_buffer_t *in = *buf_in;
1050 /* if we got an empty buffer signaling end-of-stream send it downstream */
1051 if ( in->size == 0 )
1053 /* flush any frames left in the decoder */
1054 while ( decodeFrame( pv, NULL, 0 ) )
1057 flushDelayQueue( pv );
1058 hb_list_add( pv->list, in );
1059 *buf_out = link_buf_list( pv );
1060 return HB_WORK_DONE;
1063 int64_t pts = in->start;
1066 // use the first timestamp as our 'next expected' pts
1067 if ( pv->pts_next < 0 )
1076 pv->new_chap = in->new_chap;
1077 pv->chap_time = pts >= 0? pts : pv->pts_next;
1079 prepare_ffmpeg_buffer( in );
1080 decodeFrame( pv, in->data, in->size );
1081 hb_buffer_close( &in );
1082 *buf_out = link_buf_list( pv );
1086 static int decavcodecviInfo( hb_work_object_t *w, hb_work_info_t *info )
1088 if ( decavcodecvInfo( w, info ) )
1090 hb_work_private_t *pv = w->private_data;
1091 if ( ! pv->context )
1093 init_ffmpeg_context( w );
1095 // we have the frame duration in units of the 90KHz pts clock but
1096 // need it in units of the 27MHz MPEG clock. */
1097 info->rate = 27000000;
1098 info->rate_base = pv->duration * 300.;
1104 static void decodeAudio( hb_work_private_t *pv, uint8_t *data, int size )
1106 AVCodecContext *context = pv->context;
1109 while ( pos < size )
1111 int16_t *buffer = pv->buffer;
1112 if ( buffer == NULL )
1114 // XXX ffmpeg bug workaround
1115 // malloc a buffer for the audio decode. On an x86, ffmpeg
1116 // uses mmx/sse instructions on this buffer without checking
1117 // that it's 16 byte aligned and this will cause an abort if
1118 // the buffer is allocated on our stack. Rather than doing
1119 // complicated, machine dependent alignment here we use the
1120 // fact that malloc returns an aligned pointer on most architectures.
1123 // Cygwin's malloc doesn't appear to return 16-byte aligned memory so use memalign instead.
1124 pv->buffer = memalign(16, AVCODEC_MAX_AUDIO_FRAME_SIZE);
1126 pv->buffer = malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
1129 buffer = pv->buffer;
1131 int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
1132 int len = avcodec_decode_audio2( context, buffer, &out_size,
1133 data + pos, size - pos );
1141 // We require signed 16-bit ints for the output format. If
1142 // we got something different convert it.
1143 if ( context->sample_fmt != SAMPLE_FMT_S16 )
1145 // Note: av_audio_convert seems to be a work-in-progress but
1146 // looks like it will eventually handle general audio
1147 // mixdowns which would allow us much more flexibility
1148 // in handling multichannel audio in HB. If we were doing
1149 // anything more complicated than a one-for-one format
1150 // conversion we'd probably want to cache the converter
1151 // context in the pv.
1152 int isamp = av_get_bits_per_sample_format( context->sample_fmt ) / 8;
1153 AVAudioConvert *ctx = av_audio_convert_alloc( SAMPLE_FMT_S16, 1,
1154 context->sample_fmt, 1,
1156 // get output buffer size (in 2-byte samples) then malloc a buffer
1157 out_size = ( out_size * 2 ) / isamp;
1158 buffer = malloc( out_size );
1160 // we're doing straight sample format conversion which behaves as if
1161 // there were only one channel.
1162 const void * const ibuf[6] = { pv->buffer };
1163 void * const obuf[6] = { buffer };
1164 const int istride[6] = { isamp };
1165 const int ostride[6] = { 2 };
1167 av_audio_convert( ctx, obuf, ostride, ibuf, istride, out_size >> 1 );
1168 av_audio_convert_free( ctx );
1170 hb_buffer_t *buf = hb_buffer_init( 2 * out_size );
1172 // convert from bytes to total samples
1175 double pts = pv->pts_next;
1177 pts += out_size * pv->duration;
1181 float *fl32 = (float *)buf->data;
1183 for( i = 0; i < out_size; ++i )
1185 fl32[i] = buffer[i];
1187 hb_list_add( pv->list, buf );
1189 // if we allocated a buffer for sample format conversion, free it
1190 if ( buffer != pv->buffer )
1198 static int decavcodecaiWork( hb_work_object_t *w, hb_buffer_t **buf_in,
1199 hb_buffer_t **buf_out )
1201 if ( (*buf_in)->size <= 0 )
1203 /* EOF on input stream - send it downstream & say that we're done */
1206 return HB_WORK_DONE;
1209 hb_work_private_t *pv = w->private_data;
1211 if ( (*buf_in)->start < -1 && pv->pts_next <= 0 )
1213 // discard buffers that start before video time 0
1218 if ( ! pv->context )
1220 init_ffmpeg_context( w );
1221 // duration is a scaling factor to go from #bytes in the decoded
1222 // frame to frame time (in 90KHz mpeg ticks). 'channels' converts
1223 // total samples to per-channel samples. 'sample_rate' converts
1224 // per-channel samples to seconds per sample and the 90000
1225 // is mpeg ticks per second.
1226 pv->duration = 90000. /
1227 (double)( pv->context->sample_rate * pv->context->channels );
1229 hb_buffer_t *in = *buf_in;
1231 // if the packet has a timestamp use it if we don't have a timestamp yet
1232 // or if there's been a timing discontinuity of more than 100ms.
1233 if ( in->start >= 0 &&
1234 ( pv->pts_next < 0 || ( in->start - pv->pts_next ) > 90*100 ) )
1236 pv->pts_next = in->start;
1238 prepare_ffmpeg_buffer( in );
1239 decodeAudio( pv, in->data, in->size );
1240 *buf_out = link_buf_list( pv );
1245 hb_work_object_t hb_decavcodecvi =
1248 "Video decoder (ffmpeg streams)",
1256 hb_work_object_t hb_decavcodecai =
1259 "Audio decoder (ffmpeg streams)",