1 /*
2  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 
12 #include "vpx_config.h"
13 #include "vp8_rtcd.h"
14 #include "encodemb.h"
15 #include "encodemv.h"
16 #include "vp8/common/common.h"
17 #include "onyx_int.h"
18 #include "vp8/common/extend.h"
19 #include "vp8/common/entropymode.h"
20 #include "vp8/common/quant_common.h"
21 #include "segmentation.h"
22 #include "vp8/common/setupintrarecon.h"
23 #include "encodeintra.h"
24 #include "vp8/common/reconinter.h"
25 #include "rdopt.h"
26 #include "pickinter.h"
27 #include "vp8/common/findnearmv.h"
28 #include <stdio.h>
29 #include <limits.h>
30 #include "vp8/common/invtrans.h"
31 #include "vpx_ports/vpx_timer.h"
32 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
33 #include "bitstream.h"
34 #endif
35 #include "encodeframe.h"
36 
37 extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t) ;
38 extern void vp8_calc_ref_frame_costs(int *ref_frame_cost,
39                                      int prob_intra,
40                                      int prob_last,
41                                      int prob_garf
42                                     );
43 extern void vp8_convert_rfct_to_prob(VP8_COMP *const cpi);
44 extern void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex);
45 extern void vp8_auto_select_speed(VP8_COMP *cpi);
46 extern void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
47                                       MACROBLOCK *x,
48                                       MB_ROW_COMP *mbr_ei,
49                                       int count);
50 static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x );
51 
52 #ifdef MODE_STATS
53 unsigned int inter_y_modes[10] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
54 unsigned int inter_uv_modes[4] = {0, 0, 0, 0};
55 unsigned int inter_b_modes[15]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
56 unsigned int y_modes[5]   = {0, 0, 0, 0, 0};
57 unsigned int uv_modes[4]  = {0, 0, 0, 0};
58 unsigned int b_modes[14]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
59 #endif
60 
61 
62 /* activity_avg must be positive, or flat regions could get a zero weight
63  *  (infinite lambda), which confounds analysis.
64  * This also avoids the need for divide by zero checks in
65  *  vp8_activity_masking().
66  */
67 #define VP8_ACTIVITY_AVG_MIN (64)
68 
69 /* This is used as a reference when computing the source variance for the
70  *  purposes of activity masking.
71  * Eventually this should be replaced by custom no-reference routines,
72  *  which will be faster.
73  */
74 static const unsigned char VP8_VAR_OFFS[16]=
75 {
76     128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128
77 };
78 
79 
80 /* Original activity measure from Tim T's code. */
tt_activity_measure(VP8_COMP * cpi,MACROBLOCK * x)81 static unsigned int tt_activity_measure( VP8_COMP *cpi, MACROBLOCK *x )
82 {
83     unsigned int act;
84     unsigned int sse;
85     /* TODO: This could also be done over smaller areas (8x8), but that would
86      *  require extensive changes elsewhere, as lambda is assumed to be fixed
87      *  over an entire MB in most of the code.
88      * Another option is to compute four 8x8 variances, and pick a single
89      *  lambda using a non-linear combination (e.g., the smallest, or second
90      *  smallest, etc.).
91      */
92     (void)cpi;
93     act =  vp8_variance16x16(x->src.y_buffer,
94                     x->src.y_stride, VP8_VAR_OFFS, 0, &sse);
95     act = act<<4;
96 
97     /* If the region is flat, lower the activity some more. */
98     if (act < 8<<12)
99         act = act < 5<<12 ? act : 5<<12;
100 
101     return act;
102 }
103 
104 /* Stub for alternative experimental activity measures. */
alt_activity_measure(VP8_COMP * cpi,MACROBLOCK * x,int use_dc_pred)105 static unsigned int alt_activity_measure( VP8_COMP *cpi,
106                                           MACROBLOCK *x, int use_dc_pred )
107 {
108     return vp8_encode_intra(cpi,x, use_dc_pred);
109 }
110 
111 
112 /* Measure the activity of the current macroblock
113  * What we measure here is TBD so abstracted to this function
114  */
115 #define ALT_ACT_MEASURE 1
mb_activity_measure(VP8_COMP * cpi,MACROBLOCK * x,int mb_row,int mb_col)116 static unsigned int mb_activity_measure( VP8_COMP *cpi, MACROBLOCK *x,
117                                   int mb_row, int mb_col)
118 {
119     unsigned int mb_activity;
120 
121     if  ( ALT_ACT_MEASURE )
122     {
123         int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
124 
125         /* Or use and alternative. */
126         mb_activity = alt_activity_measure( cpi, x, use_dc_pred );
127     }
128     else
129     {
130         /* Original activity measure from Tim T's code. */
131         mb_activity = tt_activity_measure( cpi, x );
132     }
133 
134     if ( mb_activity < VP8_ACTIVITY_AVG_MIN )
135         mb_activity = VP8_ACTIVITY_AVG_MIN;
136 
137     return mb_activity;
138 }
139 
140 /* Calculate an "average" mb activity value for the frame */
141 #define ACT_MEDIAN 0
calc_av_activity(VP8_COMP * cpi,int64_t activity_sum)142 static void calc_av_activity( VP8_COMP *cpi, int64_t activity_sum )
143 {
144 #if ACT_MEDIAN
145     /* Find median: Simple n^2 algorithm for experimentation */
146     {
147         unsigned int median;
148         unsigned int i,j;
149         unsigned int * sortlist;
150         unsigned int tmp;
151 
152         /* Create a list to sort to */
153         CHECK_MEM_ERROR(sortlist,
154                         vpx_calloc(sizeof(unsigned int),
155                         cpi->common.MBs));
156 
157         /* Copy map to sort list */
158         vpx_memcpy( sortlist, cpi->mb_activity_map,
159                     sizeof(unsigned int) * cpi->common.MBs );
160 
161 
162         /* Ripple each value down to its correct position */
163         for ( i = 1; i < cpi->common.MBs; i ++ )
164         {
165             for ( j = i; j > 0; j -- )
166             {
167                 if ( sortlist[j] < sortlist[j-1] )
168                 {
169                     /* Swap values */
170                     tmp = sortlist[j-1];
171                     sortlist[j-1] = sortlist[j];
172                     sortlist[j] = tmp;
173                 }
174                 else
175                     break;
176             }
177         }
178 
179         /* Even number MBs so estimate median as mean of two either side. */
180         median = ( 1 + sortlist[cpi->common.MBs >> 1] +
181                    sortlist[(cpi->common.MBs >> 1) + 1] ) >> 1;
182 
183         cpi->activity_avg = median;
184 
185         vpx_free(sortlist);
186     }
187 #else
188     /* Simple mean for now */
189     cpi->activity_avg = (unsigned int)(activity_sum/cpi->common.MBs);
190 #endif
191 
192     if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN)
193         cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;
194 
195     /* Experimental code: return fixed value normalized for several clips */
196     if  ( ALT_ACT_MEASURE )
197         cpi->activity_avg = 100000;
198 }
199 
200 #define USE_ACT_INDEX   0
201 #define OUTPUT_NORM_ACT_STATS   0
202 
203 #if USE_ACT_INDEX
204 /* Calculate and activity index for each mb */
calc_activity_index(VP8_COMP * cpi,MACROBLOCK * x)205 static void calc_activity_index( VP8_COMP *cpi, MACROBLOCK *x )
206 {
207     VP8_COMMON *const cm = & cpi->common;
208     int mb_row, mb_col;
209 
210     int64_t act;
211     int64_t a;
212     int64_t b;
213 
214 #if OUTPUT_NORM_ACT_STATS
215     FILE *f = fopen("norm_act.stt", "a");
216     fprintf(f, "\n%12d\n", cpi->activity_avg );
217 #endif
218 
219     /* Reset pointers to start of activity map */
220     x->mb_activity_ptr = cpi->mb_activity_map;
221 
222     /* Calculate normalized mb activity number. */
223     for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
224     {
225         /* for each macroblock col in image */
226         for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
227         {
228             /* Read activity from the map */
229             act = *(x->mb_activity_ptr);
230 
231             /* Calculate a normalized activity number */
232             a = act + 4*cpi->activity_avg;
233             b = 4*act + cpi->activity_avg;
234 
235             if ( b >= a )
236                 *(x->activity_ptr) = (int)((b + (a>>1))/a) - 1;
237             else
238                 *(x->activity_ptr) = 1 - (int)((a + (b>>1))/b);
239 
240 #if OUTPUT_NORM_ACT_STATS
241             fprintf(f, " %6d", *(x->mb_activity_ptr));
242 #endif
243             /* Increment activity map pointers */
244             x->mb_activity_ptr++;
245         }
246 
247 #if OUTPUT_NORM_ACT_STATS
248         fprintf(f, "\n");
249 #endif
250 
251     }
252 
253 #if OUTPUT_NORM_ACT_STATS
254     fclose(f);
255 #endif
256 
257 }
258 #endif
259 
260 /* Loop through all MBs. Note activity of each, average activity and
261  * calculate a normalized activity for each
262  */
build_activity_map(VP8_COMP * cpi)263 static void build_activity_map( VP8_COMP *cpi )
264 {
265     MACROBLOCK *const x = & cpi->mb;
266     MACROBLOCKD *xd = &x->e_mbd;
267     VP8_COMMON *const cm = & cpi->common;
268 
269 #if ALT_ACT_MEASURE
270     YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
271     int recon_yoffset;
272     int recon_y_stride = new_yv12->y_stride;
273 #endif
274 
275     int mb_row, mb_col;
276     unsigned int mb_activity;
277     int64_t activity_sum = 0;
278 
279     /* for each macroblock row in image */
280     for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
281     {
282 #if ALT_ACT_MEASURE
283         /* reset above block coeffs */
284         xd->up_available = (mb_row != 0);
285         recon_yoffset = (mb_row * recon_y_stride * 16);
286 #endif
287         /* for each macroblock col in image */
288         for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
289         {
290 #if ALT_ACT_MEASURE
291             xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
292             xd->left_available = (mb_col != 0);
293             recon_yoffset += 16;
294 #endif
295             /* Copy current mb to a buffer */
296             vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
297 
298             /* measure activity */
299             mb_activity = mb_activity_measure( cpi, x, mb_row, mb_col );
300 
301             /* Keep frame sum */
302             activity_sum += mb_activity;
303 
304             /* Store MB level activity details. */
305             *x->mb_activity_ptr = mb_activity;
306 
307             /* Increment activity map pointer */
308             x->mb_activity_ptr++;
309 
310             /* adjust to the next column of source macroblocks */
311             x->src.y_buffer += 16;
312         }
313 
314 
315         /* adjust to the next row of mbs */
316         x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
317 
318 #if ALT_ACT_MEASURE
319         /* extend the recon for intra prediction */
320         vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16,
321                           xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
322 #endif
323 
324     }
325 
326     /* Calculate an "average" MB activity */
327     calc_av_activity(cpi, activity_sum);
328 
329 #if USE_ACT_INDEX
330     /* Calculate an activity index number of each mb */
331     calc_activity_index( cpi, x );
332 #endif
333 
334 }
335 
336 /* Macroblock activity masking */
vp8_activity_masking(VP8_COMP * cpi,MACROBLOCK * x)337 void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x)
338 {
339 #if USE_ACT_INDEX
340     x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
341     x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
342     x->errorperbit += (x->errorperbit==0);
343 #else
344     int64_t a;
345     int64_t b;
346     int64_t act = *(x->mb_activity_ptr);
347 
348     /* Apply the masking to the RD multiplier. */
349     a = act + (2*cpi->activity_avg);
350     b = (2*act) + cpi->activity_avg;
351 
352     x->rdmult = (unsigned int)(((int64_t)x->rdmult*b + (a>>1))/a);
353     x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
354     x->errorperbit += (x->errorperbit==0);
355 #endif
356 
357     /* Activity based Zbin adjustment */
358     adjust_act_zbin(cpi, x);
359 }
360 
361 static
encode_mb_row(VP8_COMP * cpi,VP8_COMMON * cm,int mb_row,MACROBLOCK * x,MACROBLOCKD * xd,TOKENEXTRA ** tp,int * segment_counts,int * totalrate)362 void encode_mb_row(VP8_COMP *cpi,
363                    VP8_COMMON *cm,
364                    int mb_row,
365                    MACROBLOCK  *x,
366                    MACROBLOCKD *xd,
367                    TOKENEXTRA **tp,
368                    int *segment_counts,
369                    int *totalrate)
370 {
371     int recon_yoffset, recon_uvoffset;
372     int mb_col;
373     int ref_fb_idx = cm->lst_fb_idx;
374     int dst_fb_idx = cm->new_fb_idx;
375     int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
376     int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
377     int map_index = (mb_row * cpi->common.mb_cols);
378 
379 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
380     const int num_part = (1 << cm->multi_token_partition);
381     TOKENEXTRA * tp_start = cpi->tok;
382     vp8_writer *w;
383 #endif
384 
385 #if CONFIG_MULTITHREAD
386     const int nsync = cpi->mt_sync_range;
387     const int rightmost_col = cm->mb_cols + nsync;
388     volatile const int *last_row_current_mb_col;
389     volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
390 
391     if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
392         last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
393     else
394         last_row_current_mb_col = &rightmost_col;
395 #endif
396 
397 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
398     if(num_part > 1)
399         w= &cpi->bc[1 + (mb_row % num_part)];
400     else
401         w = &cpi->bc[1];
402 #endif
403 
404     /* reset above block coeffs */
405     xd->above_context = cm->above_context;
406 
407     xd->up_available = (mb_row != 0);
408     recon_yoffset = (mb_row * recon_y_stride * 16);
409     recon_uvoffset = (mb_row * recon_uv_stride * 8);
410 
411     cpi->tplist[mb_row].start = *tp;
412     /* printf("Main mb_row = %d\n", mb_row); */
413 
414     /* Distance of Mb to the top & bottom edges, specified in 1/8th pel
415      * units as they are always compared to values that are in 1/8th pel
416      */
417     xd->mb_to_top_edge = -((mb_row * 16) << 3);
418     xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
419 
420     /* Set up limit values for vertical motion vector components
421      * to prevent them extending beyond the UMV borders
422      */
423     x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
424     x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
425                         + (VP8BORDERINPIXELS - 16);
426 
427     /* Set the mb activity pointer to the start of the row. */
428     x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
429 
430     /* for each macroblock col in image */
431     for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
432     {
433 
434 #if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
435         *tp = cpi->tok;
436 #endif
437         /* Distance of Mb to the left & right edges, specified in
438          * 1/8th pel units as they are always compared to values
439          * that are in 1/8th pel units
440          */
441         xd->mb_to_left_edge = -((mb_col * 16) << 3);
442         xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
443 
444         /* Set up limit values for horizontal motion vector components
445          * to prevent them extending beyond the UMV borders
446          */
447         x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
448         x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16)
449                             + (VP8BORDERINPIXELS - 16);
450 
451         xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
452         xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
453         xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
454         xd->left_available = (mb_col != 0);
455 
456         x->rddiv = cpi->RDDIV;
457         x->rdmult = cpi->RDMULT;
458 
459         /* Copy current mb to a buffer */
460         vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
461 
462 #if CONFIG_MULTITHREAD
463         if (cpi->b_multi_threaded != 0)
464         {
465             *current_mb_col = mb_col - 1; /* set previous MB done */
466 
467             if ((mb_col & (nsync - 1)) == 0)
468             {
469                 while (mb_col > (*last_row_current_mb_col - nsync))
470                 {
471                     x86_pause_hint();
472                     thread_sleep(0);
473                 }
474             }
475         }
476 #endif
477 
478         if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
479             vp8_activity_masking(cpi, x);
480 
481         /* Is segmentation enabled */
482         /* MB level adjustment to quantizer */
483         if (xd->segmentation_enabled)
484         {
485             /* Code to set segment id in xd->mbmi.segment_id for current MB
486              * (with range checking)
487              */
488             if (cpi->segmentation_map[map_index+mb_col] <= 3)
489                 xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index+mb_col];
490             else
491                 xd->mode_info_context->mbmi.segment_id = 0;
492 
493             vp8cx_mb_init_quantizer(cpi, x, 1);
494         }
495         else
496             /* Set to Segment 0 by default */
497             xd->mode_info_context->mbmi.segment_id = 0;
498 
499         x->active_ptr = cpi->active_map + map_index + mb_col;
500 
501         if (cm->frame_type == KEY_FRAME)
502         {
503             *totalrate += vp8cx_encode_intra_macroblock(cpi, x, tp);
504 #ifdef MODE_STATS
505             y_modes[xd->mbmi.mode] ++;
506 #endif
507         }
508         else
509         {
510             *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
511 
512 #ifdef MODE_STATS
513             inter_y_modes[xd->mbmi.mode] ++;
514 
515             if (xd->mbmi.mode == SPLITMV)
516             {
517                 int b;
518 
519                 for (b = 0; b < xd->mbmi.partition_count; b++)
520                 {
521                     inter_b_modes[x->partition->bmi[b].mode] ++;
522                 }
523             }
524 
525 #endif
526 
527             /* Special case code for cyclic refresh
528              * If cyclic update enabled then copy xd->mbmi.segment_id; (which
529              * may have been updated based on mode during
530              * vp8cx_encode_inter_macroblock()) back into the global
531              * segmentation map
532              */
533             if ((cpi->current_layer == 0) &&
534                 (cpi->cyclic_refresh_mode_enabled &&
535                  xd->segmentation_enabled))
536             {
537                 cpi->segmentation_map[map_index+mb_col] = xd->mode_info_context->mbmi.segment_id;
538 
539                 /* If the block has been refreshed mark it as clean (the
540                  * magnitude of the -ve influences how long it will be before
541                  * we consider another refresh):
542                  * Else if it was coded (last frame 0,0) and has not already
543                  * been refreshed then mark it as a candidate for cleanup
544                  * next time (marked 0) else mark it as dirty (1).
545                  */
546                 if (xd->mode_info_context->mbmi.segment_id)
547                     cpi->cyclic_refresh_map[map_index+mb_col] = -1;
548                 else if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
549                 {
550                     if (cpi->cyclic_refresh_map[map_index+mb_col] == 1)
551                         cpi->cyclic_refresh_map[map_index+mb_col] = 0;
552                 }
553                 else
554                     cpi->cyclic_refresh_map[map_index+mb_col] = 1;
555 
556             }
557         }
558 
559         cpi->tplist[mb_row].stop = *tp;
560 
561 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
562         /* pack tokens for this MB */
563         {
564             int tok_count = *tp - tp_start;
565             pack_tokens(w, tp_start, tok_count);
566         }
567 #endif
568         /* Increment pointer into gf usage flags structure. */
569         x->gf_active_ptr++;
570 
571         /* Increment the activity mask pointers. */
572         x->mb_activity_ptr++;
573 
574         /* adjust to the next column of macroblocks */
575         x->src.y_buffer += 16;
576         x->src.u_buffer += 8;
577         x->src.v_buffer += 8;
578 
579         recon_yoffset += 16;
580         recon_uvoffset += 8;
581 
582         /* Keep track of segment usage */
583         segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
584 
585         /* skip to next mb */
586         xd->mode_info_context++;
587         x->partition_info++;
588         xd->above_context++;
589     }
590 
591     /* extend the recon for intra prediction */
592     vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
593                         xd->dst.y_buffer + 16,
594                         xd->dst.u_buffer + 8,
595                         xd->dst.v_buffer + 8);
596 
597 #if CONFIG_MULTITHREAD
598     if (cpi->b_multi_threaded != 0)
599         *current_mb_col = rightmost_col;
600 #endif
601 
602     /* this is to account for the border */
603     xd->mode_info_context++;
604     x->partition_info++;
605 }
606 
init_encode_frame_mb_context(VP8_COMP * cpi)607 static void init_encode_frame_mb_context(VP8_COMP *cpi)
608 {
609     MACROBLOCK *const x = & cpi->mb;
610     VP8_COMMON *const cm = & cpi->common;
611     MACROBLOCKD *const xd = & x->e_mbd;
612 
613     /* GF active flags data structure */
614     x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
615 
616     /* Activity map pointer */
617     x->mb_activity_ptr = cpi->mb_activity_map;
618 
619     x->act_zbin_adj = 0;
620 
621     x->partition_info = x->pi;
622 
623     xd->mode_info_context = cm->mi;
624     xd->mode_info_stride = cm->mode_info_stride;
625 
626     xd->frame_type = cm->frame_type;
627 
628     /* reset intra mode contexts */
629     if (cm->frame_type == KEY_FRAME)
630         vp8_init_mbmode_probs(cm);
631 
632     /* Copy data over into macro block data structures. */
633     x->src = * cpi->Source;
634     xd->pre = cm->yv12_fb[cm->lst_fb_idx];
635     xd->dst = cm->yv12_fb[cm->new_fb_idx];
636 
637     /* set up frame for intra coded blocks */
638     vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);
639 
640     vp8_build_block_offsets(x);
641 
642     xd->mode_info_context->mbmi.mode = DC_PRED;
643     xd->mode_info_context->mbmi.uv_mode = DC_PRED;
644 
645     xd->left_context = &cm->left_context;
646 
647     x->mvc = cm->fc.mvc;
648 
649     vpx_memset(cm->above_context, 0,
650                sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);
651 
652     /* Special case treatment when GF and ARF are not sensible options
653      * for reference
654      */
655     if (cpi->ref_frame_flags == VP8_LAST_FRAME)
656         vp8_calc_ref_frame_costs(x->ref_frame_cost,
657                                  cpi->prob_intra_coded,255,128);
658     else if ((cpi->oxcf.number_of_layers > 1) &&
659                (cpi->ref_frame_flags == VP8_GOLD_FRAME))
660         vp8_calc_ref_frame_costs(x->ref_frame_cost,
661                                  cpi->prob_intra_coded,1,255);
662     else if ((cpi->oxcf.number_of_layers > 1) &&
663                 (cpi->ref_frame_flags == VP8_ALTR_FRAME))
664         vp8_calc_ref_frame_costs(x->ref_frame_cost,
665                                  cpi->prob_intra_coded,1,1);
666     else
667         vp8_calc_ref_frame_costs(x->ref_frame_cost,
668                                  cpi->prob_intra_coded,
669                                  cpi->prob_last_coded,
670                                  cpi->prob_gf_coded);
671 
672     xd->fullpixel_mask = 0xffffffff;
673     if(cm->full_pixel)
674         xd->fullpixel_mask = 0xfffffff8;
675 
676     vp8_zero(x->coef_counts);
677     vp8_zero(x->ymode_count);
678     vp8_zero(x->uv_mode_count)
679     x->prediction_error = 0;
680     x->intra_error = 0;
681     vp8_zero(x->count_mb_ref_frame_usage);
682 }
683 
sum_coef_counts(MACROBLOCK * x,MACROBLOCK * x_thread)684 static void sum_coef_counts(MACROBLOCK *x, MACROBLOCK *x_thread)
685 {
686     int i = 0;
687     do
688     {
689         int j = 0;
690         do
691         {
692             int k = 0;
693             do
694             {
695                 /* at every context */
696 
697                 /* calc probs and branch cts for this frame only */
698                 int t = 0;      /* token/prob index */
699 
700                 do
701                 {
702                     x->coef_counts [i][j][k][t] +=
703                         x_thread->coef_counts [i][j][k][t];
704                 }
705                 while (++t < ENTROPY_NODES);
706             }
707             while (++k < PREV_COEF_CONTEXTS);
708         }
709         while (++j < COEF_BANDS);
710     }
711     while (++i < BLOCK_TYPES);
712 }
713 
vp8_encode_frame(VP8_COMP * cpi)714 void vp8_encode_frame(VP8_COMP *cpi)
715 {
716     int mb_row;
717     MACROBLOCK *const x = & cpi->mb;
718     VP8_COMMON *const cm = & cpi->common;
719     MACROBLOCKD *const xd = & x->e_mbd;
720     TOKENEXTRA *tp = cpi->tok;
721     int segment_counts[MAX_MB_SEGMENTS];
722     int totalrate;
723 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
724     BOOL_CODER * bc = &cpi->bc[1]; /* bc[0] is for control partition */
725     const int num_part = (1 << cm->multi_token_partition);
726 #endif
727 
728     vpx_memset(segment_counts, 0, sizeof(segment_counts));
729     totalrate = 0;
730 
731     if (cpi->compressor_speed == 2)
732     {
733         if (cpi->oxcf.cpu_used < 0)
734             cpi->Speed = -(cpi->oxcf.cpu_used);
735         else
736             vp8_auto_select_speed(cpi);
737     }
738 
739     /* Functions setup for all frame types so we can use MC in AltRef */
740     if(!cm->use_bilinear_mc_filter)
741     {
742         xd->subpixel_predict        = vp8_sixtap_predict4x4;
743         xd->subpixel_predict8x4     = vp8_sixtap_predict8x4;
744         xd->subpixel_predict8x8     = vp8_sixtap_predict8x8;
745         xd->subpixel_predict16x16   = vp8_sixtap_predict16x16;
746     }
747     else
748     {
749         xd->subpixel_predict        = vp8_bilinear_predict4x4;
750         xd->subpixel_predict8x4     = vp8_bilinear_predict8x4;
751         xd->subpixel_predict8x8     = vp8_bilinear_predict8x8;
752         xd->subpixel_predict16x16   = vp8_bilinear_predict16x16;
753     }
754 
755     cpi->mb.skip_true_count = 0;
756     cpi->tok_count = 0;
757 
758 #if 0
759     /* Experimental code */
760     cpi->frame_distortion = 0;
761     cpi->last_mb_distortion = 0;
762 #endif
763 
764     xd->mode_info_context = cm->mi;
765 
766     vp8_zero(cpi->mb.MVcount);
767 
768     vp8cx_frame_init_quantizer(cpi);
769 
770     vp8_initialize_rd_consts(cpi, x,
771                              vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
772 
773     vp8cx_initialize_me_consts(cpi, cm->base_qindex);
774 
775     if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
776     {
777         /* Initialize encode frame context. */
778         init_encode_frame_mb_context(cpi);
779 
780         /* Build a frame level activity map */
781         build_activity_map(cpi);
782     }
783 
784     /* re-init encode frame context. */
785     init_encode_frame_mb_context(cpi);
786 
787 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
788     {
789         int i;
790         for(i = 0; i < num_part; i++)
791         {
792             vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
793                     cpi->partition_d_end[i + 1]);
794             bc[i].error = &cm->error;
795         }
796     }
797 
798 #endif
799 
800     {
801         struct vpx_usec_timer  emr_timer;
802         vpx_usec_timer_start(&emr_timer);
803 
804 #if CONFIG_MULTITHREAD
805         if (cpi->b_multi_threaded)
806         {
807             int i;
808 
809             vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei,
810                                       cpi->encoding_thread_count);
811 
812             for (i = 0; i < cm->mb_rows; i++)
813                 cpi->mt_current_mb_col[i] = -1;
814 
815             for (i = 0; i < cpi->encoding_thread_count; i++)
816             {
817                 sem_post(&cpi->h_event_start_encoding[i]);
818             }
819 
820             for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
821             {
822                 vp8_zero(cm->left_context)
823 
824 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
825                 tp = cpi->tok;
826 #else
827                 tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
828 #endif
829 
830                 encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
831 
832                 /* adjust to the next row of mbs */
833                 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
834                 x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
835                 x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
836 
837                 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
838                 x->partition_info  += xd->mode_info_stride * cpi->encoding_thread_count;
839                 x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
840 
841                 if(mb_row == cm->mb_rows - 1)
842                 {
843                     sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
844                 }
845             }
846 
847             sem_wait(&cpi->h_event_end_encoding); /* wait for other threads to finish */
848 
849             for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
850             {
851                 cpi->tok_count += (unsigned int)
852                   (cpi->tplist[mb_row].stop - cpi->tplist[mb_row].start);
853             }
854 
855             if (xd->segmentation_enabled)
856             {
857                 int j;
858 
859                 if (xd->segmentation_enabled)
860                 {
861                     for (i = 0; i < cpi->encoding_thread_count; i++)
862                     {
863                         for (j = 0; j < 4; j++)
864                             segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
865                     }
866                 }
867             }
868 
869             for (i = 0; i < cpi->encoding_thread_count; i++)
870             {
871                 int mode_count;
872                 int c_idx;
873                 totalrate += cpi->mb_row_ei[i].totalrate;
874 
875                 cpi->mb.skip_true_count += cpi->mb_row_ei[i].mb.skip_true_count;
876 
877                 for(mode_count = 0; mode_count < VP8_YMODES; mode_count++)
878                     cpi->mb.ymode_count[mode_count] +=
879                         cpi->mb_row_ei[i].mb.ymode_count[mode_count];
880 
881                 for(mode_count = 0; mode_count < VP8_UV_MODES; mode_count++)
882                     cpi->mb.uv_mode_count[mode_count] +=
883                         cpi->mb_row_ei[i].mb.uv_mode_count[mode_count];
884 
885                 for(c_idx = 0; c_idx < MVvals; c_idx++)
886                 {
887                     cpi->mb.MVcount[0][c_idx] +=
888                         cpi->mb_row_ei[i].mb.MVcount[0][c_idx];
889                     cpi->mb.MVcount[1][c_idx] +=
890                         cpi->mb_row_ei[i].mb.MVcount[1][c_idx];
891                 }
892 
893                 cpi->mb.prediction_error +=
894                     cpi->mb_row_ei[i].mb.prediction_error;
895                 cpi->mb.intra_error += cpi->mb_row_ei[i].mb.intra_error;
896 
897                 for(c_idx = 0; c_idx < MAX_REF_FRAMES; c_idx++)
898                     cpi->mb.count_mb_ref_frame_usage[c_idx] +=
899                         cpi->mb_row_ei[i].mb.count_mb_ref_frame_usage[c_idx];
900 
901                 for(c_idx = 0; c_idx < MAX_ERROR_BINS; c_idx++)
902                     cpi->mb.error_bins[c_idx] +=
903                         cpi->mb_row_ei[i].mb.error_bins[c_idx];
904 
905                 /* add up counts for each thread */
906                 sum_coef_counts(x, &cpi->mb_row_ei[i].mb);
907             }
908 
909         }
910         else
911 #endif
912         {
913 
914             /* for each macroblock row in image */
915             for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
916             {
917                 vp8_zero(cm->left_context)
918 
919 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
920                 tp = cpi->tok;
921 #endif
922 
923                 encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
924 
925                 /* adjust to the next row of mbs */
926                 x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
927                 x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
928                 x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
929             }
930 
931             cpi->tok_count = (unsigned int)(tp - cpi->tok);
932         }
933 
934 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
935         {
936             int i;
937             for(i = 0; i < num_part; i++)
938             {
939                 vp8_stop_encode(&bc[i]);
940                 cpi->partition_sz[i+1] = bc[i].pos;
941             }
942         }
943 #endif
944 
945         vpx_usec_timer_mark(&emr_timer);
946         cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
947     }
948 
949 
950     // Work out the segment probabilities if segmentation is enabled
951     // and needs to be updated
952     if (xd->segmentation_enabled && xd->update_mb_segmentation_map)
953     {
954         int tot_count;
955         int i;
956 
957         /* Set to defaults */
958         vpx_memset(xd->mb_segment_tree_probs, 255 , sizeof(xd->mb_segment_tree_probs));
959 
960         tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] + segment_counts[3];
961 
962         if (tot_count)
963         {
964             xd->mb_segment_tree_probs[0] = ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
965 
966             tot_count = segment_counts[0] + segment_counts[1];
967 
968             if (tot_count > 0)
969             {
970                 xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
971             }
972 
973             tot_count = segment_counts[2] + segment_counts[3];
974 
975             if (tot_count > 0)
976                 xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
977 
978             /* Zero probabilities not allowed */
979             for (i = 0; i < MB_FEATURE_TREE_PROBS; i ++)
980             {
981                 if (xd->mb_segment_tree_probs[i] == 0)
982                     xd->mb_segment_tree_probs[i] = 1;
983             }
984         }
985     }
986 
987     /* projected_frame_size in units of BYTES */
988     cpi->projected_frame_size = totalrate >> 8;
989 
990     /* Make a note of the percentage MBs coded Intra. */
991     if (cm->frame_type == KEY_FRAME)
992     {
993         cpi->this_frame_percent_intra = 100;
994     }
995     else
996     {
997         int tot_modes;
998 
999         tot_modes = cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME]
1000                     + cpi->mb.count_mb_ref_frame_usage[LAST_FRAME]
1001                     + cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME]
1002                     + cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
1003 
1004         if (tot_modes)
1005             cpi->this_frame_percent_intra =
1006                 cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
1007 
1008     }
1009 
1010 #if ! CONFIG_REALTIME_ONLY
1011     /* Adjust the projected reference frame usage probability numbers to
1012      * reflect what we have just seen. This may be useful when we make
1013      * multiple iterations of the recode loop rather than continuing to use
1014      * values from the previous frame.
1015      */
1016     if ((cm->frame_type != KEY_FRAME) && ((cpi->oxcf.number_of_layers > 1) ||
1017         (!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame)))
1018     {
1019       vp8_convert_rfct_to_prob(cpi);
1020     }
1021 #endif
1022 }
vp8_setup_block_ptrs(MACROBLOCK * x)1023 void vp8_setup_block_ptrs(MACROBLOCK *x)
1024 {
1025     int r, c;
1026     int i;
1027 
1028     for (r = 0; r < 4; r++)
1029     {
1030         for (c = 0; c < 4; c++)
1031         {
1032             x->block[r*4+c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
1033         }
1034     }
1035 
1036     for (r = 0; r < 2; r++)
1037     {
1038         for (c = 0; c < 2; c++)
1039         {
1040             x->block[16 + r*2+c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
1041         }
1042     }
1043 
1044 
1045     for (r = 0; r < 2; r++)
1046     {
1047         for (c = 0; c < 2; c++)
1048         {
1049             x->block[20 + r*2+c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
1050         }
1051     }
1052 
1053     x->block[24].src_diff = x->src_diff + 384;
1054 
1055 
1056     for (i = 0; i < 25; i++)
1057     {
1058         x->block[i].coeff = x->coeff + i * 16;
1059     }
1060 }
1061 
vp8_build_block_offsets(MACROBLOCK * x)1062 void vp8_build_block_offsets(MACROBLOCK *x)
1063 {
1064     int block = 0;
1065     int br, bc;
1066 
1067     vp8_build_block_doffsets(&x->e_mbd);
1068 
1069     /* y blocks */
1070     x->thismb_ptr = &x->thismb[0];
1071     for (br = 0; br < 4; br++)
1072     {
1073         for (bc = 0; bc < 4; bc++)
1074         {
1075             BLOCK *this_block = &x->block[block];
1076             this_block->base_src = &x->thismb_ptr;
1077             this_block->src_stride = 16;
1078             this_block->src = 4 * br * 16 + 4 * bc;
1079             ++block;
1080         }
1081     }
1082 
1083     /* u blocks */
1084     for (br = 0; br < 2; br++)
1085     {
1086         for (bc = 0; bc < 2; bc++)
1087         {
1088             BLOCK *this_block = &x->block[block];
1089             this_block->base_src = &x->src.u_buffer;
1090             this_block->src_stride = x->src.uv_stride;
1091             this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1092             ++block;
1093         }
1094     }
1095 
1096     /* v blocks */
1097     for (br = 0; br < 2; br++)
1098     {
1099         for (bc = 0; bc < 2; bc++)
1100         {
1101             BLOCK *this_block = &x->block[block];
1102             this_block->base_src = &x->src.v_buffer;
1103             this_block->src_stride = x->src.uv_stride;
1104             this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1105             ++block;
1106         }
1107     }
1108 }
1109 
sum_intra_stats(VP8_COMP * cpi,MACROBLOCK * x)1110 static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x)
1111 {
1112     const MACROBLOCKD *xd = & x->e_mbd;
1113     const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
1114     const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;
1115 
1116 #ifdef MODE_STATS
1117     const int is_key = cpi->common.frame_type == KEY_FRAME;
1118 
1119     ++ (is_key ? uv_modes : inter_uv_modes)[uvm];
1120 
1121     if (m == B_PRED)
1122     {
1123         unsigned int *const bct = is_key ? b_modes : inter_b_modes;
1124 
1125         int b = 0;
1126 
1127         do
1128         {
1129             ++ bct[xd->block[b].bmi.mode];
1130         }
1131         while (++b < 16);
1132     }
1133 
1134 #endif
1135     (void)cpi;
1136     ++x->ymode_count[m];
1137     ++x->uv_mode_count[uvm];
1138 
1139 }
1140 
1141 /* Experimental stub function to create a per MB zbin adjustment based on
1142  * some previously calculated measure of MB activity.
1143  */
adjust_act_zbin(VP8_COMP * cpi,MACROBLOCK * x)1144 static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x )
1145 {
1146 #if USE_ACT_INDEX
1147     x->act_zbin_adj = *(x->mb_activity_ptr);
1148 #else
1149     int64_t a;
1150     int64_t b;
1151     int64_t act = *(x->mb_activity_ptr);
1152 
1153     /* Apply the masking to the RD multiplier. */
1154     a = act + 4*cpi->activity_avg;
1155     b = 4*act + cpi->activity_avg;
1156 
1157     if ( act > cpi->activity_avg )
1158         x->act_zbin_adj = (int)(((int64_t)b + (a>>1))/a) - 1;
1159     else
1160         x->act_zbin_adj = 1 - (int)(((int64_t)a + (b>>1))/b);
1161 #endif
1162 }
1163 
vp8cx_encode_intra_macroblock(VP8_COMP * cpi,MACROBLOCK * x,TOKENEXTRA ** t)1164 int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
1165                                   TOKENEXTRA **t)
1166 {
1167     MACROBLOCKD *xd = &x->e_mbd;
1168     int rate;
1169 
1170     if (cpi->sf.RD && cpi->compressor_speed != 2)
1171         vp8_rd_pick_intra_mode(x, &rate);
1172     else
1173         vp8_pick_intra_mode(x, &rate);
1174 
1175     if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1176     {
1177         adjust_act_zbin( cpi, x );
1178         vp8_update_zbin_extra(cpi, x);
1179     }
1180 
1181     if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED)
1182         vp8_encode_intra4x4mby(x);
1183     else
1184         vp8_encode_intra16x16mby(x);
1185 
1186     vp8_encode_intra16x16mbuv(x);
1187 
1188     sum_intra_stats(cpi, x);
1189 
1190     vp8_tokenize_mb(cpi, x, t);
1191 
1192     if (xd->mode_info_context->mbmi.mode != B_PRED)
1193         vp8_inverse_transform_mby(xd);
1194 
1195     vp8_dequant_idct_add_uv_block
1196                     (xd->qcoeff+16*16, xd->dequant_uv,
1197                      xd->dst.u_buffer, xd->dst.v_buffer,
1198                      xd->dst.uv_stride, xd->eobs+16);
1199     return rate;
1200 }
1201 #ifdef SPEEDSTATS
1202 extern int cnt_pm;
1203 #endif
1204 
1205 extern void vp8_fix_contexts(MACROBLOCKD *x);
1206 
vp8cx_encode_inter_macroblock(VP8_COMP * cpi,MACROBLOCK * x,TOKENEXTRA ** t,int recon_yoffset,int recon_uvoffset,int mb_row,int mb_col)1207 int vp8cx_encode_inter_macroblock
1208 (
1209     VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
1210     int recon_yoffset, int recon_uvoffset,
1211     int mb_row, int mb_col
1212 )
1213 {
1214     MACROBLOCKD *const xd = &x->e_mbd;
1215     int intra_error = 0;
1216     int rate;
1217     int distortion;
1218 
1219     x->skip = 0;
1220 
1221     if (xd->segmentation_enabled)
1222         x->encode_breakout = cpi->segment_encode_breakout[xd->mode_info_context->mbmi.segment_id];
1223     else
1224         x->encode_breakout = cpi->oxcf.encode_breakout;
1225 
1226 #if CONFIG_TEMPORAL_DENOISING
1227     /* Reset the best sse mode/mv for each macroblock. */
1228     x->best_reference_frame = INTRA_FRAME;
1229     x->best_zeromv_reference_frame = INTRA_FRAME;
1230     x->best_sse_inter_mode = 0;
1231     x->best_sse_mv.as_int = 0;
1232     x->need_to_clamp_best_mvs = 0;
1233 #endif
1234 
1235     if (cpi->sf.RD)
1236     {
1237         int zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
1238 
1239         /* Are we using the fast quantizer for the mode selection? */
1240         if(cpi->sf.use_fastquant_for_pick)
1241         {
1242             x->quantize_b      = vp8_fast_quantize_b;
1243             x->quantize_b_pair = vp8_fast_quantize_b_pair;
1244 
1245             /* the fast quantizer does not use zbin_extra, so
1246              * do not recalculate */
1247             x->zbin_mode_boost_enabled = 0;
1248         }
1249         vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1250                                &distortion, &intra_error);
1251 
1252         /* switch back to the regular quantizer for the encode */
1253         if (cpi->sf.improved_quant)
1254         {
1255             x->quantize_b      = vp8_regular_quantize_b;
1256             x->quantize_b_pair = vp8_regular_quantize_b_pair;
1257         }
1258 
1259         /* restore cpi->zbin_mode_boost_enabled */
1260         x->zbin_mode_boost_enabled = zbin_mode_boost_enabled;
1261 
1262     }
1263     else
1264     {
1265         vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1266                             &distortion, &intra_error, mb_row, mb_col);
1267     }
1268 
1269     x->prediction_error += distortion;
1270     x->intra_error += intra_error;
1271 
1272     if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1273     {
1274         /* Adjust the zbin based on this MB rate. */
1275         adjust_act_zbin( cpi, x );
1276     }
1277 
1278 #if 0
1279     /* Experimental RD code */
1280     cpi->frame_distortion += distortion;
1281     cpi->last_mb_distortion = distortion;
1282 #endif
1283 
1284     /* MB level adjutment to quantizer setup */
1285     if (xd->segmentation_enabled)
1286     {
1287         /* If cyclic update enabled */
1288         if (cpi->current_layer == 0 && cpi->cyclic_refresh_mode_enabled)
1289         {
1290             /* Clear segment_id back to 0 if not coded (last frame 0,0) */
1291             if ((xd->mode_info_context->mbmi.segment_id == 1) &&
1292                 ((xd->mode_info_context->mbmi.ref_frame != LAST_FRAME) || (xd->mode_info_context->mbmi.mode != ZEROMV)))
1293             {
1294                 xd->mode_info_context->mbmi.segment_id = 0;
1295 
1296                 /* segment_id changed, so update */
1297                 vp8cx_mb_init_quantizer(cpi, x, 1);
1298             }
1299         }
1300     }
1301 
1302     {
1303         /* Experimental code.
1304          * Special case for gf and arf zeromv modes, for 1 temporal layer.
1305          * Increase zbin size to supress noise.
1306          */
1307         x->zbin_mode_boost = 0;
1308         if (x->zbin_mode_boost_enabled)
1309         {
1310             if ( xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME )
1311             {
1312                 if (xd->mode_info_context->mbmi.mode == ZEROMV)
1313                 {
1314                     if (xd->mode_info_context->mbmi.ref_frame != LAST_FRAME &&
1315                         cpi->oxcf.number_of_layers == 1)
1316                         x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
1317                     else
1318                         x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
1319                 }
1320                 else if (xd->mode_info_context->mbmi.mode == SPLITMV)
1321                     x->zbin_mode_boost = 0;
1322                 else
1323                     x->zbin_mode_boost = MV_ZBIN_BOOST;
1324             }
1325         }
1326 
1327         /* The fast quantizer doesn't use zbin_extra, only do so with
1328          * the regular quantizer. */
1329         if (cpi->sf.improved_quant)
1330             vp8_update_zbin_extra(cpi, x);
1331     }
1332 
1333     x->count_mb_ref_frame_usage[xd->mode_info_context->mbmi.ref_frame] ++;
1334 
1335     if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
1336     {
1337         vp8_encode_intra16x16mbuv(x);
1338 
1339         if (xd->mode_info_context->mbmi.mode == B_PRED)
1340         {
1341             vp8_encode_intra4x4mby(x);
1342         }
1343         else
1344         {
1345             vp8_encode_intra16x16mby(x);
1346         }
1347 
1348         sum_intra_stats(cpi, x);
1349     }
1350     else
1351     {
1352         int ref_fb_idx;
1353 
1354         if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
1355             ref_fb_idx = cpi->common.lst_fb_idx;
1356         else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
1357             ref_fb_idx = cpi->common.gld_fb_idx;
1358         else
1359             ref_fb_idx = cpi->common.alt_fb_idx;
1360 
1361         xd->pre.y_buffer = cpi->common.yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
1362         xd->pre.u_buffer = cpi->common.yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
1363         xd->pre.v_buffer = cpi->common.yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
1364 
1365         if (!x->skip)
1366         {
1367             vp8_encode_inter16x16(x);
1368         }
1369         else
1370             vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
1371                                            xd->dst.u_buffer, xd->dst.v_buffer,
1372                                            xd->dst.y_stride, xd->dst.uv_stride);
1373 
1374     }
1375 
1376     if (!x->skip)
1377     {
1378         vp8_tokenize_mb(cpi, x, t);
1379 
1380         if (xd->mode_info_context->mbmi.mode != B_PRED)
1381             vp8_inverse_transform_mby(xd);
1382 
1383         vp8_dequant_idct_add_uv_block
1384                         (xd->qcoeff+16*16, xd->dequant_uv,
1385                          xd->dst.u_buffer, xd->dst.v_buffer,
1386                          xd->dst.uv_stride, xd->eobs+16);
1387     }
1388     else
1389     {
1390         /* always set mb_skip_coeff as it is needed by the loopfilter */
1391         xd->mode_info_context->mbmi.mb_skip_coeff = 1;
1392 
1393         if (cpi->common.mb_no_coeff_skip)
1394         {
1395             x->skip_true_count ++;
1396             vp8_fix_contexts(xd);
1397         }
1398         else
1399         {
1400             vp8_stuff_mb(cpi, x, t);
1401         }
1402     }
1403 
1404     return rate;
1405 }
1406