*/
#include "hb.h"
-
-#include "libavcodec/avcodec.h"
-#include "libavformat/avformat.h"
+#include "hbffmpeg.h"
+#include "libavcodec/audioconvert.h"
static int decavcodecInit( hb_work_object_t *, hb_job_t * );
static int decavcodecWork( hb_work_object_t *, hb_buffer_t **, hb_buffer_t ** );
decavcodecBSInfo
};
+#define HEAP_SIZE 8
+typedef struct {
+ // there are nheap items on the heap indexed 1..nheap (i.e., top of
+ // heap is 1). The 0th slot is unused - a marker is put there to check
+ // for overwrite errs.
+ int64_t h[HEAP_SIZE+1];
+ int nheap;
+} pts_heap_t;
+
struct hb_work_private_s
{
- hb_job_t *job;
- AVCodecContext *context;
+ hb_job_t *job;
+ AVCodecContext *context;
AVCodecParserContext *parser;
- hb_list_t *list;
- double pts_next; // next pts we expect to generate
- int64_t pts; // (video) pts passing from parser to decoder
- int64_t chap_time; // time of next chap mark (if new_chap != 0)
- int new_chap;
- int ignore_pts; // workaround M$ bugs
- int nframes;
- int ndrops;
- double duration; // frame duration (for video)
+ hb_list_t *list;
+ double duration; // frame duration (for video)
+ double pts_next; // next pts we expect to generate
+ int64_t pts; // (video) pts passing from parser to decoder
+ int64_t chap_time; // time of next chap mark (if new_chap != 0)
+ int new_chap; // output chapter mark pending
+ uint32_t nframes;
+ uint32_t ndrops;
+ uint32_t decode_errors;
+ int brokenByMicrosoft; // video stream may contain packed b-frames
+ hb_buffer_t* delayq[HEAP_SIZE];
+ pts_heap_t pts_heap;
+ void* buffer;
+ struct SwsContext *sws_context; // if we have to rescale or convert color space
};
+static int64_t heap_pop( pts_heap_t *heap )
+{
+ int64_t result;
+
+ if ( heap->nheap <= 0 )
+ {
+ return -1;
+ }
+
+ // return the top of the heap then put the bottom element on top,
+ // decrease the heap size by one & rebalence the heap.
+ result = heap->h[1];
+
+ int64_t v = heap->h[heap->nheap--];
+ int parent = 1;
+ int child = parent << 1;
+ while ( child <= heap->nheap )
+ {
+ // find the smallest of the two children of parent
+ if (child < heap->nheap && heap->h[child] > heap->h[child+1] )
+ ++child;
+
+ if (v <= heap->h[child])
+ // new item is smaller than either child so it's the new parent.
+ break;
+
+ // smallest child is smaller than new item so move it up then
+ // check its children.
+ int64_t hp = heap->h[child];
+ heap->h[parent] = hp;
+ parent = child;
+ child = parent << 1;
+ }
+ heap->h[parent] = v;
+ return result;
+}
+
+static void heap_push( pts_heap_t *heap, int64_t v )
+{
+ if ( heap->nheap < HEAP_SIZE )
+ {
+ ++heap->nheap;
+ }
+
+ // stick the new value on the bottom of the heap then bubble it
+ // up to its correct spot.
+ int child = heap->nheap;
+ while (child > 1) {
+ int parent = child >> 1;
+ if (heap->h[parent] <= v)
+ break;
+ // move parent down
+ int64_t hp = heap->h[parent];
+ heap->h[child] = hp;
+ child = parent;
+ }
+ heap->h[child] = v;
+}
/***********************************************************************
/*XXX*/
if ( codec_id == 0 )
codec_id = CODEC_ID_MP2;
+
codec = avcodec_find_decoder( codec_id );
pv->parser = av_parser_init( codec_id );
pv->context = avcodec_alloc_context();
- avcodec_open( pv->context, codec );
+ hb_avcodec_open( pv->context, codec );
return 0;
}
static void decavcodecClose( hb_work_object_t * w )
{
hb_work_private_t * pv = w->private_data;
- if ( pv->parser )
- {
- av_parser_close(pv->parser);
- }
- if ( pv->context && pv->context->codec )
- {
- avcodec_close( pv->context );
- }
- if ( pv->list )
+
+ if ( pv )
{
- hb_list_close( &pv->list );
+ if ( pv->job && pv->context && pv->context->codec )
+ {
+ hb_log( "%s-decoder done: %u frames, %u decoder errors, %u drops",
+ pv->context->codec->name, pv->nframes, pv->decode_errors,
+ pv->ndrops );
+ }
+ if ( pv->sws_context )
+ {
+ sws_freeContext( pv->sws_context );
+ }
+ if ( pv->parser )
+ {
+ av_parser_close(pv->parser);
+ }
+ if ( pv->context && pv->context->codec )
+ {
+ hb_avcodec_close( pv->context );
+ }
+ if ( pv->list )
+ {
+ hb_list_close( &pv->list );
+ }
+ if ( pv->buffer )
+ {
+ av_free( pv->buffer );
+ pv->buffer = NULL;
+ }
+ free( pv );
+ w->private_data = NULL;
}
}
hb_work_private_t * pv = w->private_data;
hb_buffer_t * in = *buf_in, * buf, * last = NULL;
int pos, len, out_size, i, uncompressed_len;
- short buffer[AVCODEC_MAX_AUDIO_FRAME_SIZE];
+ short* bufaligned;
uint64_t cur;
unsigned char *parser_output_buffer;
int parser_output_buffer_len;
*buf_out = NULL;
+ if ( in->start < -1 && pv->pts_next <= 0 )
+ {
+ // discard buffers that start before video time 0
+ return HB_WORK_OK;
+ }
+
cur = ( in->start < 0 )? pv->pts_next : in->start;
+ bufaligned = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
pos = 0;
while( pos < in->size )
{
- len = av_parser_parse( pv->parser, pv->context,
- &parser_output_buffer, &parser_output_buffer_len,
- in->data + pos, in->size - pos, cur, cur );
+ len = av_parser_parse2( pv->parser, pv->context,
+ &parser_output_buffer, &parser_output_buffer_len,
+ in->data + pos, in->size - pos, cur, cur, AV_NOPTS_VALUE );
out_size = 0;
uncompressed_len = 0;
if (parser_output_buffer_len)
{
- out_size = sizeof(buffer);
- uncompressed_len = avcodec_decode_audio2( pv->context, buffer,
- &out_size,
- parser_output_buffer,
- parser_output_buffer_len );
+ AVPacket avp;
+ av_init_packet( &avp );
+ avp.data = parser_output_buffer;
+ avp.size = parser_output_buffer_len;
+
+ out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
+ uncompressed_len = avcodec_decode_audio3( pv->context, bufaligned, &out_size, &avp );
}
if( out_size )
{
pv->context->sample_rate;
cur = buf->stop;
- s16 = buffer;
+ s16 = bufaligned;
fl32 = (float *) buf->data;
for( i = 0; i < out_size / 2; i++ )
{
pv->pts_next = cur;
+ av_free( bufaligned );
return HB_WORK_OK;
}
return 0;
}
+static const int chan2layout[] = {
+ HB_INPUT_CH_LAYOUT_MONO, // We should allow no audio really.
+ HB_INPUT_CH_LAYOUT_MONO,
+ HB_INPUT_CH_LAYOUT_STEREO,
+ HB_INPUT_CH_LAYOUT_2F1R,
+ HB_INPUT_CH_LAYOUT_2F2R,
+ HB_INPUT_CH_LAYOUT_3F2R,
+ HB_INPUT_CH_LAYOUT_4F2R,
+ HB_INPUT_CH_LAYOUT_STEREO,
+ HB_INPUT_CH_LAYOUT_STEREO,
+};
+
static int decavcodecBSInfo( hb_work_object_t *w, const hb_buffer_t *buf,
hb_work_info_t *info )
{
hb_work_private_t *pv = w->private_data;
+ int ret = 0;
memset( info, 0, sizeof(*info) );
// now we just return dummy values if there's a codec that will handle it.
AVCodec *codec = avcodec_find_decoder( w->codec_param? w->codec_param :
CODEC_ID_MP2 );
- if ( codec )
+ if ( ! codec )
{
- static char codec_name[64];
+ // there's no ffmpeg codec for this audio type - give up
+ return -1;
+ }
- info->name = strncpy( codec_name, codec->name, sizeof(codec_name)-1 );
- info->bitrate = 384000;
- info->rate = 48000;
- info->rate_base = 1;
- info->channel_layout = HB_INPUT_CH_LAYOUT_STEREO;
- return 1;
+ static char codec_name[64];
+ info->name = strncpy( codec_name, codec->name, sizeof(codec_name)-1 );
+
+ AVCodecParserContext *parser = av_parser_init( codec->id );
+ AVCodecContext *context = avcodec_alloc_context();
+ hb_avcodec_open( context, codec );
+ uint8_t *buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
+ int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
+ unsigned char *pbuffer;
+ int pos = 0, pbuffer_size;
+
+ while ( pos < buf->size )
+ {
+ int len = av_parser_parse2( parser, context, &pbuffer, &pbuffer_size,
+ buf->data + pos, buf->size - pos,
+ buf->start, buf->start, AV_NOPTS_VALUE );
+ pos += len;
+ if ( pbuffer_size > 0 )
+ {
+ AVPacket avp;
+ av_init_packet( &avp );
+ avp.data = pbuffer;
+ avp.size = pbuffer_size;
+
+ len = avcodec_decode_audio3( context, (int16_t*)buffer, &out_size, &avp );
+ if ( len > 0 && context->sample_rate > 0 )
+ {
+ info->bitrate = context->bit_rate;
+ info->rate = context->sample_rate;
+ info->rate_base = 1;
+ info->channel_layout = chan2layout[context->channels & 7];
+ ret = 1;
+ break;
+ }
+ }
}
- return -1;
+ av_free( buffer );
+ av_parser_close( parser );
+ hb_avcodec_close( context );
+ return ret;
}
/* -------------------------------------------------------------
return dst;
}
-/* Note: assumes frame format is PIX_FMT_YUV420P */
-static hb_buffer_t *copy_frame( AVCodecContext *context, AVFrame *frame )
+// copy one video frame into an HB buf. If the frame isn't in our color space
+// or at least one of its dimensions is odd, use sws_scale to convert/rescale it.
+// Otherwise just copy the bits.
+static hb_buffer_t *copy_frame( hb_work_private_t *pv, AVFrame *frame )
{
- int w = context->width, h = context->height;
- hb_buffer_t *buf = hb_buffer_init( w * h * 3 / 2 );
+ AVCodecContext *context = pv->context;
+ int w, h;
+ if ( ! pv->job )
+ {
+ // if the dimensions are odd, drop the lsb since h264 requires that
+ // both width and height be even.
+ w = ( context->width >> 1 ) << 1;
+ h = ( context->height >> 1 ) << 1;
+ }
+ else
+ {
+ w = pv->job->title->width;
+ h = pv->job->title->height;
+ }
+ hb_buffer_t *buf = hb_video_buffer_init( w, h );
uint8_t *dst = buf->data;
- dst = copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
- w >>= 1; h >>= 1;
- dst = copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
- dst = copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
+ if ( context->pix_fmt != PIX_FMT_YUV420P || w != context->width ||
+ h != context->height )
+ {
+ // have to convert to our internal color space and/or rescale
+ AVPicture dstpic;
+ avpicture_fill( &dstpic, dst, PIX_FMT_YUV420P, w, h );
+ if ( ! pv->sws_context )
+ {
+ pv->sws_context = sws_getContext( context->width, context->height, context->pix_fmt,
+ w, h, PIX_FMT_YUV420P,
+ SWS_LANCZOS|SWS_ACCURATE_RND,
+ NULL, NULL, NULL );
+ }
+ sws_scale( pv->sws_context, frame->data, frame->linesize, 0, h,
+ dstpic.data, dstpic.linesize );
+ }
+ else
+ {
+ dst = copy_plane( dst, frame->data[0], w, frame->linesize[0], h );
+ w = (w + 1) >> 1; h = (h + 1) >> 1;
+ dst = copy_plane( dst, frame->data[1], w, frame->linesize[1], h );
+ dst = copy_plane( dst, frame->data[2], w, frame->linesize[2], h );
+ }
return buf;
}
hb_work_private_t *pv = context->opaque;
frame->pts = pv->pts;
pv->pts = -1;
-
return avcodec_default_get_buffer( context, frame );
}
static void log_chapter( hb_work_private_t *pv, int chap_num, int64_t pts )
{
hb_chapter_t *c = hb_list_item( pv->job->title->list_chapter, chap_num - 1 );
- hb_log( "%s: \"%s\" (%d) at frame %u time %lld", pv->context->codec->name,
- c->title, chap_num, pv->nframes, pts );
+ if ( c && c->title )
+ {
+ hb_log( "%s: \"%s\" (%d) at frame %u time %lld",
+ pv->context->codec->name, c->title, chap_num, pv->nframes, pts );
+ }
+ else
+ {
+ hb_log( "%s: Chapter %d at frame %u time %lld",
+ pv->context->codec->name, chap_num, pv->nframes, pts );
+ }
+}
+
+static void flushDelayQueue( hb_work_private_t *pv )
+{
+ hb_buffer_t *buf;
+ int slot = pv->nframes & (HEAP_SIZE-1);
+
+ // flush all the video packets left on our timestamp-reordering delay q
+ while ( ( buf = pv->delayq[slot] ) != NULL )
+ {
+ buf->start = heap_pop( &pv->pts_heap );
+ hb_list_add( pv->list, buf );
+ pv->delayq[slot] = NULL;
+ slot = ( slot + 1 ) & (HEAP_SIZE-1);
+ }
}
static int decodeFrame( hb_work_private_t *pv, uint8_t *data, int size )
{
- int got_picture;
+ int got_picture, oldlevel = 0;
AVFrame frame;
+ AVPacket avp;
- avcodec_decode_video( pv->context, &frame, &got_picture, data, size );
+ if ( global_verbosity_level <= 1 )
+ {
+ oldlevel = av_log_get_level();
+ av_log_set_level( AV_LOG_QUIET );
+ }
+
+ av_init_packet( &avp );
+ avp.data = data;
+ avp.size = size;
+ if ( avcodec_decode_video2( pv->context, &frame, &got_picture, &avp ) < 0 )
+ {
+ ++pv->decode_errors;
+ }
+ if ( global_verbosity_level <= 1 )
+ {
+ av_log_set_level( oldlevel );
+ }
if( got_picture )
{
// ffmpeg makes it hard to attach a pts to a frame. if the MPEG ES
// worked at this point frame.pts should hold the frame's pts from the
// original data stream or -1 if it didn't have one. in the latter case
// we generate the next pts in sequence for it.
+ double frame_dur = pv->duration;
+ if ( frame_dur <= 0 )
+ {
+ frame_dur = 90000. * (double)pv->context->time_base.num /
+ (double)pv->context->time_base.den;
+ pv->duration = frame_dur;
+ }
+ if ( frame.repeat_pict )
+ {
+ frame_dur += frame.repeat_pict * frame_dur * 0.5;
+ }
+ // XXX Unlike every other video decoder, the Raw decoder doesn't
+ // use the standard buffer allocation routines so we never
+ // get to put a PTS in the frame. Do it now.
+ if ( pv->context->codec_id == CODEC_ID_RAWVIDEO )
+ {
+ frame.pts = pv->pts;
+ pv->pts = -1;
+ }
+ // If there was no pts for this frame, assume constant frame rate
+ // video & estimate the next frame time from the last & duration.
double pts = frame.pts;
if ( pts < 0 )
{
pts = pv->pts_next;
}
- if ( pv->duration == 0 )
- {
- pv->duration = 90000. * pv->context->time_base.num /
- pv->context->time_base.den;
- }
- double frame_dur = pv->duration;
- frame_dur += frame.repeat_pict * frame_dur * 0.5;
pv->pts_next = pts + frame_dur;
- hb_buffer_t *buf = copy_frame( pv->context, &frame );
- buf->start = pts;
+ hb_buffer_t *buf;
+
+ // if we're doing a scan or this content couldn't have been broken
+ // by Microsoft we don't worry about timestamp reordering
+ if ( ! pv->job || ! pv->brokenByMicrosoft )
+ {
+ buf = copy_frame( pv, &frame );
+ buf->start = pts;
+ hb_list_add( pv->list, buf );
+ ++pv->nframes;
+ return got_picture;
+ }
- if ( pv->new_chap && buf->start >= pv->chap_time )
+ // XXX This following probably addresses a libavcodec bug but I don't
+ // see an easy fix so we workaround it here.
+ //
+ // The M$ 'packed B-frames' atrocity results in decoded frames with
+ // the wrong timestamp. E.g., if there are 2 b-frames the timestamps
+ // we see here will be "2 3 1 5 6 4 ..." instead of "1 2 3 4 5 6".
+ // The frames are actually delivered in the right order but with
+ // the wrong timestamp. To get the correct timestamp attached to
+ // each frame we have a delay queue (longer than the max number of
+ // b-frames) & a sorting heap for the timestamps. As each frame
+ // comes out of the decoder the oldest frame in the queue is removed
+ // and associated with the smallest timestamp. Then the new frame is
+ // added to the queue & its timestamp is pushed on the heap.
+ // This does nothing if the timestamps are correct (i.e., the video
+ // uses a codec that Micro$oft hasn't broken yet) but the frames
+ // get timestamped correctly even when M$ has munged them.
+
+ // remove the oldest picture from the frame queue (if any) &
+ // give it the smallest timestamp from our heap. The queue size
+ // is a power of two so we get the slot of the oldest by masking
+ // the frame count & this will become the slot of the newest
+ // once we've removed & processed the oldest.
+ int slot = pv->nframes & (HEAP_SIZE-1);
+ if ( ( buf = pv->delayq[slot] ) != NULL )
{
- buf->new_chap = pv->new_chap;
- pv->new_chap = 0;
- pv->chap_time = 0;
- if ( pv->job )
+ buf->start = heap_pop( &pv->pts_heap );
+
+ if ( pv->new_chap && buf->start >= pv->chap_time )
{
+ buf->new_chap = pv->new_chap;
+ pv->new_chap = 0;
+ pv->chap_time = 0;
log_chapter( pv, buf->new_chap, buf->start );
}
+ else if ( pv->nframes == 0 )
+ {
+ log_chapter( pv, pv->job->chapter_start, buf->start );
+ }
+ hb_list_add( pv->list, buf );
}
- else if ( pv->job && pv->nframes == 0 )
- {
- log_chapter( pv, pv->job->chapter_start, buf->start );
- }
- hb_list_add( pv->list, buf );
+
+ // add the new frame to the delayq & push its timestamp on the heap
+ pv->delayq[slot] = copy_frame( pv, &frame );
+ heap_push( &pv->pts_heap, pts );
+
++pv->nframes;
}
+
return got_picture;
}
do {
uint8_t *pout;
int pout_len;
- int len = av_parser_parse( pv->parser, pv->context, &pout, &pout_len,
- data + pos, size - pos, pts, dts );
+ int len = av_parser_parse2( pv->parser, pv->context, &pout, &pout_len,
+ data + pos, size - pos, pts, dts, AV_NOPTS_VALUE );
pos += len;
if ( pout_len > 0 )
} while ( pos < size );
/* the stuff above flushed the parser, now flush the decoder */
- while ( size == 0 && decodeFrame( pv, NULL, 0 ) )
+ if ( size <= 0 )
{
+ while ( decodeFrame( pv, NULL, 0 ) )
+ {
+ }
+ flushDelayQueue( pv );
}
}
pv->context->opaque = pv;
pv->context->get_buffer = get_frame_buf;
- AVCodec *codec = avcodec_find_decoder( codec_id );
+ return 0;
+}
+
+static int next_hdr( hb_buffer_t *in, int offset )
+{
+ uint8_t *dat = in->data;
+ uint16_t last2 = 0xffff;
+ for ( ; in->size - offset > 1; ++offset )
+ {
+ if ( last2 == 0 && dat[offset] == 0x01 )
+ // found an mpeg start code
+ return offset - 2;
+
+ last2 = ( last2 << 8 ) | dat[offset];
+ }
+
+ return -1;
+}
+
+static int find_hdr( hb_buffer_t *in, int offset, uint8_t hdr_type )
+{
+ if ( in->size - offset < 4 )
+ // not enough room for an mpeg start code
+ return -1;
+
+ for ( ; ( offset = next_hdr( in, offset ) ) >= 0; ++offset )
+ {
+ if ( in->data[offset+3] == hdr_type )
+ // found it
+ break;
+ }
+ return offset;
+}
+
+static int setup_extradata( hb_work_object_t *w, hb_buffer_t *in )
+{
+ hb_work_private_t *pv = w->private_data;
// we can't call the avstream funcs but the read_header func in the
// AVInputFormat may set up some state in the AVContext. In particular
// vc1t_read_header allocates 'extradata' to deal with header issues
// related to Microsoft's bizarre engineering notions. We alloc a chunk
// of space to make vc1 work then associate the codec with the context.
- pv->context->extradata_size = 32;
- pv->context->extradata = av_malloc(pv->context->extradata_size);
- avcodec_open( pv->context, codec );
+ if ( w->codec_param != CODEC_ID_VC1 )
+ {
+ // we haven't been inflicted with M$ - allocate a little space as
+ // a marker and return success.
+ pv->context->extradata_size = 16;
+ pv->context->extradata = av_malloc(pv->context->extradata_size);
+ return 0;
+ }
+
+ // find the start and and of the sequence header
+ int shdr, shdr_end;
+ if ( ( shdr = find_hdr( in, 0, 0x0f ) ) < 0 )
+ {
+ // didn't find start of seq hdr
+ return 1;
+ }
+ if ( ( shdr_end = next_hdr( in, shdr + 4 ) ) < 0 )
+ {
+ shdr_end = in->size;
+ }
+ shdr_end -= shdr;
+ // find the start and and of the entry point header
+ int ehdr, ehdr_end;
+ if ( ( ehdr = find_hdr( in, 0, 0x0e ) ) < 0 )
+ {
+ // didn't find start of entry point hdr
+ return 1;
+ }
+ if ( ( ehdr_end = next_hdr( in, ehdr + 4 ) ) < 0 )
+ {
+ ehdr_end = in->size;
+ }
+ ehdr_end -= ehdr;
+
+ // found both headers - allocate an extradata big enough to hold both
+ // then copy them into it.
+ pv->context->extradata_size = shdr_end + ehdr_end;
+ pv->context->extradata = av_malloc(pv->context->extradata_size + 8);
+ memcpy( pv->context->extradata, in->data + shdr, shdr_end );
+ memcpy( pv->context->extradata + shdr_end, in->data + ehdr, ehdr_end );
+ memset( pv->context->extradata + shdr_end + ehdr_end, 0, 8);
return 0;
}
{
hb_work_private_t *pv = w->private_data;
hb_buffer_t *in = *buf_in;
- int64_t pts = -1;
+ int64_t pts = AV_NOPTS_VALUE;
int64_t dts = pts;
*buf_in = NULL;
decodeVideo( pv, in->data, in->size, pts, dts );
hb_list_add( pv->list, in );
*buf_out = link_buf_list( pv );
- hb_log( "%s done: %d frames", pv->context->codec->name, pv->nframes );
return HB_WORK_DONE;
}
+ // if this is the first frame open the codec (we have to wait for the
+ // first frame because of M$ VC1 braindamage).
+ if ( pv->context->extradata_size == 0 )
+ {
+ if ( setup_extradata( w, in ) )
+ {
+ // we didn't find the headers needed to set up extradata.
+ // the codec will abort if we open it so just free the buf
+ // and hope we eventually get the info we need.
+ hb_buffer_close( &in );
+ return HB_WORK_OK;
+ }
+ AVCodec *codec = avcodec_find_decoder( w->codec_param );
+ // There's a mis-feature in ffmpeg that causes the context to be
+ // incorrectly initialized the 1st time avcodec_open is called.
+ // If you close it and open a 2nd time, it finishes the job.
+ hb_avcodec_open( pv->context, codec );
+ hb_avcodec_close( pv->context );
+ hb_avcodec_open( pv->context, codec );
+ }
+
if( in->start >= 0 )
{
pts = in->start;
info->rate = 27000000;
info->rate_base = (int64_t)context->time_base.num * 27000000LL /
context->time_base.den;
+ if ( context->ticks_per_frame > 1 )
+ {
+ // for ffmpeg 0.5 & later, the H.264 & MPEG-2 time base is
+ // field rate rather than frame rate so convert back to frames.
+ info->rate_base *= context->ticks_per_frame;
+ }
/* Sometimes there's no pixel aspect set in the source. In that case,
assume a 1:1 PAR. Otherwise, preserve the source PAR. */
if ( ! pv->context->codec )
{
AVCodec *codec = avcodec_find_decoder( pv->context->codec_id );
- avcodec_open( pv->context, codec );
+ hb_avcodec_open( pv->context, codec );
}
// set up our best guess at the frame duration.
// the frame rate in the codec is usually bogus but it's sometimes
// ok in the stream.
AVStream *st = hb_ffmpeg_avstream( w->codec_param );
- AVRational tb;
- // XXX because the time bases are so screwed up, we only take values
- // in the range 8fps - 64fps.
- if ( st->time_base.num * 64 > st->time_base.den &&
- st->time_base.den > st->time_base.num * 8 )
- {
- tb = st->time_base;
- }
- else if ( st->codec->time_base.num * 64 > st->codec->time_base.den &&
- st->codec->time_base.den > st->codec->time_base.num * 8 )
- {
- tb = st->codec->time_base;
- }
- else if ( st->r_frame_rate.den * 64 > st->r_frame_rate.num &&
- st->r_frame_rate.num > st->r_frame_rate.den * 8 )
+
+ if ( st->nb_frames && st->duration )
{
- tb.num = st->r_frame_rate.den;
- tb.den = st->r_frame_rate.num;
+ // compute the average frame duration from the total number
+ // of frames & the total duration.
+ pv->duration = ( (double)st->duration * (double)st->time_base.num ) /
+ ( (double)st->nb_frames * (double)st->time_base.den );
}
else
{
- tb.num = 1001; /*XXX*/
- tb.den = 30000; /*XXX*/
+ // XXX We don't have a frame count or duration so try to use the
+ // far less reliable time base info in the stream.
+ // Because the time bases are so screwed up, we only take values
+ // in the range 8fps - 64fps.
+ AVRational tb;
+ if ( st->time_base.num * 64 > st->time_base.den &&
+ st->time_base.den > st->time_base.num * 8 )
+ {
+ tb = st->time_base;
+ }
+ else if ( st->r_frame_rate.den * 64 > st->r_frame_rate.num &&
+ st->r_frame_rate.num > st->r_frame_rate.den * 8 )
+ {
+ tb.num = st->r_frame_rate.den;
+ tb.den = st->r_frame_rate.num;
+ }
+ else
+ {
+ tb.num = 1001; /*XXX*/
+ tb.den = 24000; /*XXX*/
+ }
+ pv->duration = (double)tb.num / (double)tb.den;
}
- pv->duration = 90000. * tb.num / tb.den;
+ pv->duration *= 90000.;
// we have to wrap ffmpeg's get_buffer to be able to set the pts (?!)
pv->context->opaque = pv;
pv->context->get_buffer = get_frame_buf;
+
+ // avi, mkv and possibly mp4 containers can contain the M$ VFW packed
+ // b-frames abortion that messes up frame ordering and timestamps.
+ // XXX ffmpeg knows which streams are broken but doesn't expose the
+ // info externally. We should patch ffmpeg to add a flag to the
+ // codec context for this but until then we mark all ffmpeg streams
+ // as suspicious.
+ pv->brokenByMicrosoft = 1;
}
static void prepare_ffmpeg_buffer( hb_buffer_t * in )
w->private_data = pv;
pv->job = job;
pv->list = hb_list_init();
-
+ pv->pts_next = -1;
+ pv->pts = -1;
return 0;
}
if ( ! pv->context )
{
init_ffmpeg_context( w );
-
- switch ( pv->context->codec_id )
- {
- // These are the only formats whose timestamps we'll believe.
- // All others are treated as CFR (i.e., we take the first timestamp
- // then generate all the others from the frame rate). The reason for
- // this is that the M$ encoders are so frigging buggy with garbage
- // like packed b-frames (vfw divx mpeg4) that believing their timestamps
- // results in discarding more than half the video frames because they'll
- // be out of sequence (and attempting to reseqence them doesn't work
- // because it's the timestamps that are wrong, not the decoded frame
- // order). All hail Redmond, ancestral home of the rich & stupid.
- case CODEC_ID_MPEG2VIDEO:
- case CODEC_ID_RAWVIDEO:
- case CODEC_ID_H264:
- case CODEC_ID_VC1:
- break;
-
- default:
- pv->ignore_pts = 1;
- break;
- }
}
hb_buffer_t *in = *buf_in;
- int64_t pts = -1;
-
*buf_in = NULL;
/* if we got an empty buffer signaling end-of-stream send it downstream */
while ( decodeFrame( pv, NULL, 0 ) )
{
}
+ flushDelayQueue( pv );
hb_list_add( pv->list, in );
*buf_out = link_buf_list( pv );
- hb_log( "%s done: %d frames %d drops", pv->context->codec->name,
- pv->nframes, pv->ndrops );
return HB_WORK_DONE;
}
- if( in->start >= 0 )
+ int64_t pts = in->start;
+ if( pts >= 0 )
{
// use the first timestamp as our 'next expected' pts
- if ( pv->pts_next <= 0 )
+ if ( pv->pts_next < 0 )
{
- pv->pts_next = in->start;
- }
-
- if ( ! pv->ignore_pts )
- {
- pts = in->start;
- if ( pv->pts > 0 )
- {
- hb_log( "overwriting pts %lld with %lld (diff %d)",
- pv->pts, pts, pts - pv->pts );
- }
- if ( pv->pts_next - pts >= pv->duration )
- {
- // this frame starts more than a frame time before where
- // the nominal frame rate says it should - drop it.
- // log the first 10 drops so we'll know what's going on.
- if ( pv->ndrops++ < 10 )
- {
- hb_log( "time reversal next %.0f pts %lld (diff %g)",
- pv->pts_next, pts, pv->pts_next - pts );
- }
- hb_buffer_close( &in );
- return HB_WORK_OK;
- }
- pv->pts = pts;
+ pv->pts_next = pts;
}
+ pv->pts = pts;
}
if ( in->new_chap )
{
if ( decavcodecvInfo( w, info ) )
{
- // There are at least three different video frame rates in ffmpeg:
- // - time_base in the AVStream
- // - time_base in the AVCodecContext
- // - r_frame_rate in the AVStream
- // There's no guidence on which if any of these to believe but the
- // routine compute_frame_duration tries the stream first then the codec.
- // In general the codec time base seems bogus & the stream time base is
- // ok except for wmv's where the stream time base is also bogus but
- // r_frame_rate is sometimes ok & sometimes a random number.
- AVStream *st = hb_ffmpeg_avstream( w->codec_param );
- AVRational tb;
- // XXX because the time bases are so screwed up, we only take values
- // in the range 8fps - 64fps.
- if ( st->time_base.num * 64 > st->time_base.den &&
- st->time_base.den > st->time_base.num * 8 )
+ hb_work_private_t *pv = w->private_data;
+ if ( ! pv->context )
{
- tb = st->time_base;
+ init_ffmpeg_context( w );
}
- else if ( st->codec->time_base.num * 64 > st->codec->time_base.den &&
- st->codec->time_base.den > st->codec->time_base.num * 8 )
- {
- tb = st->codec->time_base;
- }
- else if ( st->r_frame_rate.den * 64 > st->r_frame_rate.num &&
- st->r_frame_rate.num > st->r_frame_rate.den * 8 )
- {
- tb.num = st->r_frame_rate.den;
- tb.den = st->r_frame_rate.num;
- }
- else
- {
- tb.num = 1001; /*XXX*/
- tb.den = 30000; /*XXX*/
- }
-
- // ffmpeg gives the frame rate in frames per second while HB wants
- // it in units of the 27MHz MPEG clock. */
+ // we have the frame duration in units of the 90KHz pts clock but
+ // need it in units of the 27MHz MPEG clock. */
info->rate = 27000000;
- info->rate_base = (int64_t)tb.num * 27000000LL / tb.den;
+ info->rate_base = pv->duration * 300.;
return 1;
}
return 0;
while ( pos < size )
{
- int16_t buffer[AVCODEC_MAX_AUDIO_FRAME_SIZE];
- int out_size = sizeof(buffer);
- int len = avcodec_decode_audio2( context, buffer, &out_size,
- data + pos, size - pos );
+ int16_t *buffer = pv->buffer;
+ if ( buffer == NULL )
+ {
+ pv->buffer = av_malloc( AVCODEC_MAX_AUDIO_FRAME_SIZE );
+ buffer = pv->buffer;
+ }
+
+ AVPacket avp;
+ av_init_packet( &avp );
+ avp.data = data + pos;
+ avp.size = size - pos;
+
+ int out_size = AVCODEC_MAX_AUDIO_FRAME_SIZE;
+ int len = avcodec_decode_audio3( context, buffer, &out_size, &avp );
if ( len <= 0 )
{
return;
pos += len;
if( out_size > 0 )
{
+ // We require signed 16-bit ints for the output format. If
+ // we got something different convert it.
+ if ( context->sample_fmt != SAMPLE_FMT_S16 )
+ {
+ // Note: av_audio_convert seems to be a work-in-progress but
+ // looks like it will eventually handle general audio
+ // mixdowns which would allow us much more flexibility
+ // in handling multichannel audio in HB. If we were doing
+ // anything more complicated than a one-for-one format
+ // conversion we'd probably want to cache the converter
+ // context in the pv.
+ int isamp = av_get_bits_per_sample_format( context->sample_fmt ) / 8;
+ AVAudioConvert *ctx = av_audio_convert_alloc( SAMPLE_FMT_S16, 1,
+ context->sample_fmt, 1,
+ NULL, 0 );
+ // get output buffer size (in 2-byte samples) then malloc a buffer
+ out_size = ( out_size * 2 ) / isamp;
+ buffer = av_malloc( out_size );
+
+ // we're doing straight sample format conversion which behaves as if
+ // there were only one channel.
+ const void * const ibuf[6] = { pv->buffer };
+ void * const obuf[6] = { buffer };
+ const int istride[6] = { isamp };
+ const int ostride[6] = { 2 };
+
+ av_audio_convert( ctx, obuf, ostride, ibuf, istride, out_size >> 1 );
+ av_audio_convert_free( ctx );
+ }
hb_buffer_t *buf = hb_buffer_init( 2 * out_size );
+ // convert from bytes to total samples
+ out_size >>= 1;
+
double pts = pv->pts_next;
buf->start = pts;
- out_size >>= 1;
pts += out_size * pv->duration;
buf->stop = pts;
pv->pts_next = pts;
fl32[i] = buffer[i];
}
hb_list_add( pv->list, buf );
+
+ // if we allocated a buffer for sample format conversion, free it
+ if ( buffer != pv->buffer )
+ {
+ av_free( buffer );
+ }
}
}
}
}
hb_work_private_t *pv = w->private_data;
+
+ if ( (*buf_in)->start < -1 && pv->pts_next <= 0 )
+ {
+ // discard buffers that start before video time 0
+ *buf_out = NULL;
+ return HB_WORK_OK;
+ }
+
if ( ! pv->context )
{
init_ffmpeg_context( w );
+ // duration is a scaling factor to go from #bytes in the decoded
+ // frame to frame time (in 90KHz mpeg ticks). 'channels' converts
+ // total samples to per-channel samples. 'sample_rate' converts
+ // per-channel samples to seconds per sample and the 90000
+ // is mpeg ticks per second.
pv->duration = 90000. /
(double)( pv->context->sample_rate * pv->context->channels );
}
hb_buffer_t *in = *buf_in;
+ // if the packet has a timestamp use it if we don't have a timestamp yet
+ // or if there's been a timing discontinuity of more than 100ms.
if ( in->start >= 0 &&
( pv->pts_next < 0 || ( in->start - pv->pts_next ) > 90*100 ) )
{