1 /* $Id: render.c,v 1.17 2005/04/14 17:37:54 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. */
10 struct hb_work_private_s
14 struct SwsContext * context;
16 AVPicture pic_tmp_crop;
17 AVPicture pic_tmp_out;
18 hb_buffer_t * buf_scale;
19 hb_fifo_t * subtitle_queue;
20 hb_fifo_t * delay_queue;
23 uint64_t last_start[4];
24 uint64_t last_stop[4];
25 uint64_t lost_time[4];
26 uint64_t total_lost_time;
27 uint64_t total_gained_time;
30 int count_frames; // frames output so far
31 double frame_rate; // 90KHz ticks per frame (for CFR/PFR)
32 double out_last_stop; // where last frame ended (for CFR/PFR)
33 int drops; // frames dropped (for CFR/PFR)
34 int dups; // frames duped (for CFR/PFR)
37 int renderInit( hb_work_object_t *, hb_job_t * );
38 int renderWork( hb_work_object_t *, hb_buffer_t **, hb_buffer_t ** );
39 void renderClose( hb_work_object_t * );
41 hb_work_object_t hb_render =
53 * Utility function that finds where the U is in the YUV sub-picture
55 * The Y data is at the top, followed by U and V, but the U and V
56 * are half the width of the Y, i.e. each chroma element covers 2x2
59 static uint8_t *getU(uint8_t *data, int width, int height, int x, int y)
61 return(&data[(y>>1) * ((width+1)>>1) + (x>>1) + width*height]);
64 static uint8_t *getV(uint8_t *data, int width, int height, int x, int y)
66 int w2 = (width+1) >> 1, h2 = (height+1) >> 1;
67 return(&data[(y>>1) * w2 + (x>>1) + width*height + w2*h2]);
70 static void ApplySub( hb_job_t * job, hb_buffer_t * buf,
73 hb_buffer_t * sub = *_sub;
74 hb_title_t * title = job->title;
75 int i, j, offset_top, offset_left, margin_top, margin_percent;
76 uint8_t * lum, * alpha, * out, * sub_chromaU, * sub_chromaV;
79 * Percent of height of picture that form a margin that subtitles
80 * should not be displayed within.
90 * If necessary, move the subtitle so it is not in a cropped zone.
91 * When it won't fit, we center it so we lose as much on both ends.
92 * Otherwise we try to leave a 20px or 2% margin around it.
94 margin_top = ( ( title->height - job->crop[0] - job->crop[1] ) *
95 margin_percent ) / 100;
100 * A maximum margin of 20px regardless of height of the picture.
105 if( sub->height > title->height - job->crop[0] - job->crop[1] -
109 * The subtitle won't fit in the cropped zone, so center
110 * it vertically so we fit in as much as we can.
112 offset_top = job->crop[0] + ( title->height - job->crop[0] -
113 job->crop[1] - sub->height ) / 2;
115 else if( sub->y < job->crop[0] + margin_top )
118 * The subtitle fits in the cropped zone, but is currently positioned
119 * within our top margin, so move it outside of our margin.
121 offset_top = job->crop[0] + margin_top;
123 else if( sub->y > title->height - job->crop[1] - margin_top - sub->height )
126 * The subtitle fits in the cropped zone, and is not within the top
127 * margin but is within the bottom margin, so move it to be above
130 offset_top = title->height - job->crop[1] - margin_top - sub->height;
135 * The subtitle is fine where it is.
140 if( sub->width > title->width - job->crop[2] - job->crop[3] - 40 )
141 offset_left = job->crop[2] + ( title->width - job->crop[2] -
142 job->crop[3] - sub->width ) / 2;
143 else if( sub->x < job->crop[2] + 20 )
144 offset_left = job->crop[2] + 20;
145 else if( sub->x > title->width - job->crop[3] - 20 - sub->width )
146 offset_left = title->width - job->crop[3] - 20 - sub->width;
148 offset_left = sub->x;
151 alpha = lum + sub->width * sub->height;
152 sub_chromaU = alpha + sub->width * sub->height;
153 sub_chromaV = sub_chromaU + sub->width * sub->height;
155 out = buf->data + offset_top * title->width + offset_left;
157 for( i = 0; i < sub->height; i++ )
159 if( offset_top + i >= 0 && offset_top + i < title->height )
161 for( j = 0; j < sub->width; j++ )
163 if( offset_left + j >= 0 && offset_left + j < title->width )
165 uint8_t *chromaU, *chromaV;
168 * Merge the luminance and alpha with the picture
170 out[j] = ( (uint16_t) out[j] * ( 16 - (uint16_t) alpha[j] ) +
171 (uint16_t) lum[j] * (uint16_t) alpha[j] ) >> 4;
173 * Set the chroma (colour) based on whether there is
174 * any alpha at all. Don't try to blend with the picture.
176 chromaU = getU(buf->data, title->width, title->height,
177 offset_left+j, offset_top+i);
179 chromaV = getV(buf->data, title->width, title->height,
180 offset_left+j, offset_top+i);
185 * Add the chroma from the sub-picture, as this is
186 * not a transparent element.
188 *chromaU = sub_chromaU[j];
189 *chromaV = sub_chromaV[j];
197 sub_chromaU += sub->width;
198 sub_chromaV += sub->width;
202 hb_buffer_close( _sub );
205 // delete the buffer 'out' from the chain of buffers whose head is 'buf_out'.
206 // out & buf_out must be non-null (checks prior to anywhere this routine
207 // can be called guarantee this) and out must be on buf_out's chain
208 // (all places that call this get 'out' while traversing the chain).
209 // 'out' is freed and its predecessor is returned.
210 static hb_buffer_t *delete_buffer_from_chain( hb_buffer_t **buf_out, hb_buffer_t *out)
212 hb_buffer_t *succ = out->next;
213 hb_buffer_t *pred = *buf_out;
216 // we're deleting the first buffer
221 // target isn't the first buf so search for its predecessor.
222 while ( pred->next != out )
226 // found 'out' - remove it from the chain
230 hb_buffer_close( &out );
234 // insert buffer 'succ' after buffer chain element 'pred'.
235 // caller must guarantee that 'pred' and 'succ' are non-null.
236 static hb_buffer_t *insert_buffer_in_chain( hb_buffer_t *pred, hb_buffer_t *succ )
238 succ->next = pred->next;
243 // This section of the code implements video frame rate control.
244 // Since filters are allowed to duplicate and drop frames (which
245 // changes the timing), this has to be the last thing done in render.
247 // There are three options, selected by the value of job->cfr:
248 // 0 - Variable Frame Rate (VFR) or 'same as source': frame times
250 // 1 - Constant Frame Rate (CFR): Frame timings are adjusted so that all
251 // frames are exactly job->vrate_base ticks apart. Frames are dropped
252 // or duplicated if necessary to maintain this spacing.
253 // 2 - Peak Frame Rate (PFR): job->vrate_base is treated as the peak
254 // average frame rate. I.e., the average frame rate (current frame
255 // end time divided by number of frames so far) is never allowed to be
256 // greater than job->vrate_base and frames are dropped if necessary
257 // to keep the average under this value. Other than those drops, frame
258 // times are left alone.
261 static void adjust_frame_rate( hb_work_private_t *pv, hb_buffer_t **buf_out )
263 hb_buffer_t *out = *buf_out;
265 while ( out && out->size > 0 )
267 // this frame has to start where the last one stopped.
268 out->start = pv->out_last_stop;
270 // compute where this frame would stop if the frame rate were constant
271 // (this is our target stopping time for CFR and earliest possible
272 // stopping time for PFR).
273 double cfr_stop = pv->frame_rate * ( pv->count_frames + 1 );
275 if ( cfr_stop - (double)out->stop >= pv->frame_rate )
277 // This frame stops a frame time or more in the past - drop it
278 // but don't lose its chapter mark.
281 pv->chapter_time = out->start;
282 pv->chapter_val = out->new_chap;
285 out = delete_buffer_from_chain( buf_out, out );
289 // at this point we know that this frame doesn't push the average
290 // rate over the limit so we just pass it on for PFR. For CFR we're
291 // going to return it (with its start & stop times modified) and
292 // we may have to dup it.
294 if ( pv->job->cfr > 1 )
296 // PFR - we're going to keep the frame but may need to
297 // adjust it's stop time to meet the average rate constraint.
298 if ( out->stop <= cfr_stop )
300 out->stop = cfr_stop;
305 // we're doing CFR so we have to either trim some time from a
306 // buffer that ends too far in the future or, if the buffer is
307 // two or more frame times long, split it into multiple pieces,
308 // each of which is a frame time long.
309 double excess_dur = (double)out->stop - cfr_stop;
310 out->stop = cfr_stop;
311 for ( ; excess_dur >= pv->frame_rate; excess_dur -= cfr_stop )
313 /* next frame too far ahead - dup current frame */
314 hb_buffer_t *dup = hb_buffer_init( out->size );
315 memcpy( dup->data, out->data, out->size );
316 hb_buffer_copy_settings( dup, out );
318 dup->start = cfr_stop;
319 cfr_stop += pv->frame_rate;
320 dup->stop = cfr_stop;
321 out = insert_buffer_in_chain( out, dup );
326 pv->out_last_stop = out->stop;
331 int renderWork( hb_work_object_t * w, hb_buffer_t ** buf_in,
332 hb_buffer_t ** buf_out )
334 hb_work_private_t * pv = w->private_data;
335 hb_job_t * job = pv->job;
336 hb_title_t * title = job->title;
337 hb_buffer_t * in = *buf_in, * buf_tmp_in = *buf_in;
338 hb_buffer_t * ivtc_buffer = NULL;
342 hb_buffer_t *head = NULL, *tail = NULL, *next;
345 /* If the input buffer is end of stream, send out an empty one
346 * to the next stage as well. To avoid losing the contents of
347 * the delay queue connect the buffers in the delay queue in
348 * the correct order, and add the end of stream buffer to the
351 while( ( next = hb_fifo_get( pv->delay_queue ) ) != NULL )
354 /* We can't use the given time stamps. Previous frames
355 might already have been extended, throwing off the
356 raw values fed to render.c. Instead, their
357 stop and start times are stored in arrays.
358 The 4th cached frame will be the to use.
359 If it needed its duration extended to make up
360 lost time, it will have happened above. */
361 next->start = pv->last_start[counter];
362 next->stop = pv->last_stop[counter--];
378 adjust_frame_rate( pv, buf_out );
388 * During the indepth_scan ditch the buffers here before applying filters or attempting to
391 if( job->indepth_scan )
397 /* Push subtitles onto queue just in case we need to delay a frame */
400 hb_fifo_push( pv->subtitle_queue, in->sub );
404 hb_fifo_push( pv->subtitle_queue, hb_buffer_init(0) );
407 /* If there's a chapter mark remember it in case we delay or drop its frame */
410 pv->chapter_time = in->start;
411 pv->chapter_val = in->new_chap;
415 /* Setup render buffer */
416 hb_buffer_t * buf_render = hb_video_buffer_init( job->width, job->height );
421 int filter_count = hb_list_count( job->filters );
424 for( i = 0; i < filter_count; i++ )
426 hb_filter_object_t * filter = hb_list_item( job->filters, i );
433 hb_buffer_t * buf_tmp_out = NULL;
435 int result = filter->work( buf_tmp_in,
440 filter->private_data );
443 * FILTER_OK: set temp buffer to filter buffer, continue
444 * FILTER_DELAY: set temp buffer to NULL, abort
445 * FILTER_DROP: set temp buffer to NULL, pop subtitle, abort
446 * FILTER_FAILED: leave temp buffer alone, continue
448 if( result == FILTER_OK )
450 buf_tmp_in = buf_tmp_out;
452 else if( result == FILTER_DELAY )
457 else if( result == FILTER_DROP )
459 /* We need to compensate for the time lost by dropping this frame.
460 Spread its duration out in quarters, because usually dropped frames
461 maintain a 1-out-of-5 pattern and this spreads it out amongst the remaining ones.
462 Store these in the lost_time array, which has 4 slots in it.
463 Because not every frame duration divides evenly by 4, and we can't lose the
464 remainder, we have to go through an awkward process to preserve it in the 4th array index. */
465 uint64_t temp_duration = buf_tmp_out->stop - buf_tmp_out->start;
466 pv->lost_time[0] += (temp_duration / 4);
467 pv->lost_time[1] += (temp_duration / 4);
468 pv->lost_time[2] += (temp_duration / 4);
469 pv->lost_time[3] += ( temp_duration - (temp_duration / 4) - (temp_duration / 4) - (temp_duration / 4) );
471 pv->total_lost_time += temp_duration;
472 pv->dropped_frames++;
474 /* Pop the frame's subtitle and dispose of it. */
475 hb_buffer_t * subpicture_list = hb_fifo_get( pv->subtitle_queue );
476 hb_buffer_t * subpicture;
477 hb_buffer_t * subpicture_next;
478 for ( subpicture = subpicture_list; subpicture; subpicture = subpicture_next )
480 subpicture_next = subpicture->next_subpicture;
481 hb_buffer_close( &subpicture );
491 /* Cache frame start and stop times, so we can renumber
492 time stamps if dropping frames for VFR. */
494 for( i = 3; i >= 1; i-- )
496 pv->last_start[i] = pv->last_start[i-1];
497 pv->last_stop[i] = pv->last_stop[i-1];
500 /* In order to make sure we have continuous time stamps, store
501 the current frame's duration as starting when the last one stopped. */
502 pv->last_start[0] = pv->last_stop[1];
503 pv->last_stop[0] = pv->last_start[0] + (buf_tmp_in->stop - buf_tmp_in->start);
506 /* Apply subtitles */
509 hb_buffer_t * subpicture_list = hb_fifo_get( pv->subtitle_queue );
510 hb_buffer_t * subpicture;
511 hb_buffer_t * subpicture_next;
512 for ( subpicture = subpicture_list; subpicture; subpicture = subpicture_next )
514 subpicture_next = subpicture->next_subpicture;
515 ApplySub( job, buf_tmp_in, &subpicture );
519 /* Apply crop/scale if specified */
520 if( buf_tmp_in && pv->context )
522 avpicture_fill( &pv->pic_tmp_in, buf_tmp_in->data,
524 title->width, title->height );
526 avpicture_fill( &pv->pic_tmp_out, buf_render->data,
528 job->width, job->height );
530 // Crop; this alters the pointer to the data to point to the correct place for cropped frame
531 av_picture_crop( &pv->pic_tmp_crop, &pv->pic_tmp_in, PIX_FMT_YUV420P,
532 job->crop[0], job->crop[2] );
534 // Scale pic_crop into pic_render according to the context set up in renderInit
535 sws_scale(pv->context,
536 pv->pic_tmp_crop.data, pv->pic_tmp_crop.linesize,
537 0, title->height - (job->crop[0] + job->crop[1]),
538 pv->pic_tmp_out.data, pv->pic_tmp_out.linesize);
540 hb_buffer_copy_settings( buf_render, buf_tmp_in );
542 buf_tmp_in = buf_render;
545 /* Set output to render buffer */
546 (*buf_out) = buf_render;
548 if( buf_tmp_in == NULL )
550 /* Teardown and cleanup buffers if we are emitting NULL */
551 if( buf_in && *buf_in )
553 hb_buffer_close( buf_in );
556 if( buf_out && *buf_out )
558 hb_buffer_close( buf_out );
562 else if( buf_tmp_in != buf_render )
564 /* Copy temporary results and settings into render buffer */
565 memcpy( buf_render->data, buf_tmp_in->data, buf_render->size );
566 hb_buffer_copy_settings( buf_render, buf_tmp_in );
571 hb_fifo_push( pv->delay_queue, *buf_out );
576 * Keep the last three frames in our queue, this ensures that we have the last
577 * two always in there should we need to rewrite the durations on them.
580 if( hb_fifo_size( pv->delay_queue ) >= 4 )
582 *buf_out = hb_fifo_get( pv->delay_queue );
587 /* The current frame exists. That means it hasn't been dropped by a filter.
588 Make it accessible as ivtc_buffer so we can edit its duration if needed. */
589 ivtc_buffer = *buf_out;
591 if( pv->lost_time[3] > 0 )
594 * A frame's been dropped earlier by VFR detelecine.
595 * Gotta make up the lost time. This will also
596 * slow down the video.
597 * The dropped frame's has to be accounted for, so
598 * divvy it up amongst the 4 frames left behind.
599 * This is what the delay_queue is for;
600 * telecined sequences start 2 frames before
601 * the dropped frame, so to slow down the right
602 * ones you need a 2 frame delay between
603 * reading input and writing output.
606 /* We want to extend the outputted frame's duration by the value
607 stored in the 4th slot of the lost_time array. Because we need
608 to adjust all the values in the array so they're contiguous,
609 extend the duration inside the array first, before applying
610 it to the current frame buffer. */
611 pv->last_stop[3] += pv->lost_time[3];
613 /* Log how much time has been added back in to the video. */
614 pv->total_gained_time += pv->lost_time[3];
616 /* We've pulled the 4th value from the lost_time array
617 and added it to the last_stop array's 4th slot. Now, rotate the
618 lost_time array so the 4th slot now holds the 3rd's value, and
619 so on down the line, and set the 0 index to a value of 0. */
621 for( i=2; i >= 0; i--)
623 pv->lost_time[i+1] = pv->lost_time[i];
625 pv->lost_time[0] = 0;
627 /* Log how many frames have had their durations extended. */
628 pv->extended_frames++;
631 /* We can't use the given time stamps. Previous frames
632 might already have been extended, throwing off the
633 raw values fed to render.c. Instead, their
634 stop and start times are stored in arrays.
635 The 4th cached frame will be the to use.
636 If it needed its duration extended to make up
637 lost time, it will have happened above. */
638 ivtc_buffer->start = pv->last_start[3];
639 ivtc_buffer->stop = pv->last_stop[3];
641 /* Set the 3rd cached frame to start when this one stops,
642 and so on down the line. If any of them need to be
643 extended as well to make up lost time, it'll be handled
644 on the next loop through the renderer. */
646 for (i = 2; i >= 0; i--)
648 int temp_duration = pv->last_stop[i] - pv->last_start[i];
649 pv->last_start[i] = pv->last_stop[i+1];
650 pv->last_stop[i] = pv->last_start[i] + temp_duration;
653 /* If we have a pending chapter mark and this frame is at
654 or after the time of the mark, mark this frame & clear
656 if( pv->chapter_time && pv->chapter_time <= ivtc_buffer->start )
658 ivtc_buffer->new_chap = pv->chapter_val;
659 pv->chapter_time = 0;
664 if ( buf_out && *buf_out && job->cfr )
666 adjust_frame_rate( pv, buf_out );
671 void renderClose( hb_work_object_t * w )
673 hb_work_private_t * pv = w->private_data;
677 hb_log("render: %d frames output, %d dropped and %d duped for CFR/PFR",
678 pv->count_frames, pv->drops, pv->dups );
681 hb_interjob_t * interjob = hb_interjob_get( w->private_data->job->h );
683 /* Preserve dropped frame count for more accurate framerates in 2nd passes. */
684 interjob->render_dropped = pv->dropped_frames;
686 hb_log("render: lost time: %"PRId64" (%i frames)", pv->total_lost_time, pv->dropped_frames);
687 hb_log("render: gained time: %"PRId64" (%i frames) (%"PRId64" not accounted for)", pv->total_gained_time, pv->extended_frames, pv->total_lost_time - pv->total_gained_time);
688 if (pv->dropped_frames)
689 hb_log("render: average dropped frame duration: %"PRId64, (pv->total_lost_time / pv->dropped_frames) );
691 /* Cleanup subtitle queue */
692 if( pv->subtitle_queue )
694 hb_fifo_close( &pv->subtitle_queue );
697 if( pv->delay_queue )
699 hb_fifo_close( &pv->delay_queue );
702 /* Cleanup render work structure */
704 w->private_data = NULL;
707 int renderInit( hb_work_object_t * w, hb_job_t * job )
709 /* Allocate new private work object */
710 hb_work_private_t * pv = calloc( 1, sizeof( hb_work_private_t ) );
712 w->private_data = pv;
715 swsflags = SWS_LANCZOS | SWS_ACCURATE_RND;
717 /* Get title and title size */
718 hb_title_t * title = job->title;
720 /* If crop or scale is specified, setup rescale context */
721 if( job->crop[0] || job->crop[1] || job->crop[2] || job->crop[3] ||
722 job->width != title->width || job->height != title->height )
724 pv->context = hb_sws_get_context(title->width - (job->crop[2] + job->crop[3]),
725 title->height - (job->crop[0] + job->crop[1]),
727 job->width, job->height, PIX_FMT_YUV420P,
731 /* Setup FIFO queue for subtitle cache */
732 pv->subtitle_queue = hb_fifo_init( 8, 1 );
733 pv->delay_queue = hb_fifo_init( 8, 1 );
735 /* VFR IVTC needs a bunch of time-keeping variables to track
736 how many frames are dropped, how many are extended, what the
737 last 4 start and stop times were (so they can be modified),
738 how much time has been lost and gained overall, how much time
739 the latest 4 frames should be extended by, and where chapter
740 markers are (so they can be saved if their frames are dropped.) */
741 pv->dropped_frames = 0;
742 pv->extended_frames = 0;
743 pv->last_start[0] = 0;
744 pv->last_stop[0] = 0;
745 pv->total_lost_time = 0;
746 pv->total_gained_time = 0;
747 pv->lost_time[0] = 0; pv->lost_time[1] = 0; pv->lost_time[2] = 0; pv->lost_time[3] = 0;
748 pv->chapter_time = 0;
751 pv->frame_rate = (double)job->vrate_base * (1./300.);
754 /* TODO: Move to work.c? */
757 int filter_count = hb_list_count( job->filters );
760 for( i = 0; i < filter_count; i++ )
762 hb_filter_object_t * filter = hb_list_item( job->filters, i );
764 if( !filter ) continue;
766 filter->private_data = filter->init( PIX_FMT_YUV420P,