• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include "vpx_config.h"
12 #include "vp8_rtcd.h"
13 #include "./vpx_dsp_rtcd.h"
14 #include "bitstream.h"
15 #include "encodemb.h"
16 #include "encodemv.h"
17 #if CONFIG_MULTITHREAD
18 #include "ethreading.h"
19 #endif
20 #include "vp8/common/common.h"
21 #include "onyx_int.h"
22 #include "vp8/common/extend.h"
23 #include "vp8/common/entropymode.h"
24 #include "vp8/common/quant_common.h"
25 #include "segmentation.h"
26 #include "vp8/common/setupintrarecon.h"
27 #include "encodeintra.h"
28 #include "vp8/common/reconinter.h"
29 #include "rdopt.h"
30 #include "pickinter.h"
31 #include "vp8/common/findnearmv.h"
32 #include <stdio.h>
33 #include <limits.h>
34 #include "vp8/common/invtrans.h"
35 #include "vpx_ports/vpx_timer.h"
36 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
37 #include "bitstream.h"
38 #endif
39 #include "encodeframe.h"
40 
41 extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
42 static void adjust_act_zbin(VP8_COMP *cpi, MACROBLOCK *x);
43 
44 #ifdef MODE_STATS
45 unsigned int inter_y_modes[10] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
46 unsigned int inter_uv_modes[4] = { 0, 0, 0, 0 };
47 unsigned int inter_b_modes[15] = {
48   0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
49 };
50 unsigned int y_modes[5] = { 0, 0, 0, 0, 0 };
51 unsigned int uv_modes[4] = { 0, 0, 0, 0 };
52 unsigned int b_modes[14] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
53 #endif
54 
55 /* activity_avg must be positive, or flat regions could get a zero weight
56  *  (infinite lambda), which confounds analysis.
57  * This also avoids the need for divide by zero checks in
58  *  vp8_activity_masking().
59  */
60 #define VP8_ACTIVITY_AVG_MIN (64)
61 
62 /* This is used as a reference when computing the source variance for the
63  *  purposes of activity masking.
64  * Eventually this should be replaced by custom no-reference routines,
65  *  which will be faster.
66  */
67 static const unsigned char VP8_VAR_OFFS[16] = {
68   128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
69 };
70 
71 /* Original activity measure from Tim T's code. */
tt_activity_measure(VP8_COMP * cpi,MACROBLOCK * x)72 static unsigned int tt_activity_measure(VP8_COMP *cpi, MACROBLOCK *x) {
73   unsigned int act;
74   unsigned int sse;
75   (void)cpi;
76   /* TODO: This could also be done over smaller areas (8x8), but that would
77    *  require extensive changes elsewhere, as lambda is assumed to be fixed
78    *  over an entire MB in most of the code.
79    * Another option is to compute four 8x8 variances, and pick a single
80    *  lambda using a non-linear combination (e.g., the smallest, or second
81    *  smallest, etc.).
82    */
83   act = vpx_variance16x16(x->src.y_buffer, x->src.y_stride, VP8_VAR_OFFS, 0,
84                           &sse);
85   act = act << 4;
86 
87   /* If the region is flat, lower the activity some more. */
88   if (act < 8 << 12) act = act < 5 << 12 ? act : 5 << 12;
89 
90   return act;
91 }
92 
93 /* Stub for alternative experimental activity measures. */
alt_activity_measure(VP8_COMP * cpi,MACROBLOCK * x,int use_dc_pred)94 static unsigned int alt_activity_measure(VP8_COMP *cpi, MACROBLOCK *x,
95                                          int use_dc_pred) {
96   return vp8_encode_intra(cpi, x, use_dc_pred);
97 }
98 
99 /* Measure the activity of the current macroblock
100  * What we measure here is TBD so abstracted to this function
101  */
102 #define ALT_ACT_MEASURE 1
mb_activity_measure(VP8_COMP * cpi,MACROBLOCK * x,int mb_row,int mb_col)103 static unsigned int mb_activity_measure(VP8_COMP *cpi, MACROBLOCK *x,
104                                         int mb_row, int mb_col) {
105   unsigned int mb_activity;
106 
107   if (ALT_ACT_MEASURE) {
108     int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
109 
110     /* Or use and alternative. */
111     mb_activity = alt_activity_measure(cpi, x, use_dc_pred);
112   } else {
113     /* Original activity measure from Tim T's code. */
114     mb_activity = tt_activity_measure(cpi, x);
115   }
116 
117   if (mb_activity < VP8_ACTIVITY_AVG_MIN) mb_activity = VP8_ACTIVITY_AVG_MIN;
118 
119   return mb_activity;
120 }
121 
122 /* Calculate an "average" mb activity value for the frame */
123 #define ACT_MEDIAN 0
calc_av_activity(VP8_COMP * cpi,int64_t activity_sum)124 static void calc_av_activity(VP8_COMP *cpi, int64_t activity_sum) {
125 #if ACT_MEDIAN
126   /* Find median: Simple n^2 algorithm for experimentation */
127   {
128     unsigned int median;
129     unsigned int i, j;
130     unsigned int *sortlist;
131     unsigned int tmp;
132 
133     /* Create a list to sort to */
134     CHECK_MEM_ERROR(sortlist,
135                     vpx_calloc(sizeof(unsigned int), cpi->common.MBs));
136 
137     /* Copy map to sort list */
138     memcpy(sortlist, cpi->mb_activity_map,
139            sizeof(unsigned int) * cpi->common.MBs);
140 
141     /* Ripple each value down to its correct position */
142     for (i = 1; i < cpi->common.MBs; ++i) {
143       for (j = i; j > 0; j--) {
144         if (sortlist[j] < sortlist[j - 1]) {
145           /* Swap values */
146           tmp = sortlist[j - 1];
147           sortlist[j - 1] = sortlist[j];
148           sortlist[j] = tmp;
149         } else
150           break;
151       }
152     }
153 
154     /* Even number MBs so estimate median as mean of two either side. */
155     median = (1 + sortlist[cpi->common.MBs >> 1] +
156               sortlist[(cpi->common.MBs >> 1) + 1]) >>
157              1;
158 
159     cpi->activity_avg = median;
160 
161     vpx_free(sortlist);
162   }
163 #else
164   /* Simple mean for now */
165   cpi->activity_avg = (unsigned int)(activity_sum / cpi->common.MBs);
166 #endif
167 
168   if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN) {
169     cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;
170   }
171 
172   /* Experimental code: return fixed value normalized for several clips */
173   if (ALT_ACT_MEASURE) cpi->activity_avg = 100000;
174 }
175 
176 #define USE_ACT_INDEX 0
177 #define OUTPUT_NORM_ACT_STATS 0
178 
179 #if USE_ACT_INDEX
180 /* Calculate and activity index for each mb */
calc_activity_index(VP8_COMP * cpi,MACROBLOCK * x)181 static void calc_activity_index(VP8_COMP *cpi, MACROBLOCK *x) {
182   VP8_COMMON *const cm = &cpi->common;
183   int mb_row, mb_col;
184 
185   int64_t act;
186   int64_t a;
187   int64_t b;
188 
189 #if OUTPUT_NORM_ACT_STATS
190   FILE *f = fopen("norm_act.stt", "a");
191   fprintf(f, "\n%12d\n", cpi->activity_avg);
192 #endif
193 
194   /* Reset pointers to start of activity map */
195   x->mb_activity_ptr = cpi->mb_activity_map;
196 
197   /* Calculate normalized mb activity number. */
198   for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
199     /* for each macroblock col in image */
200     for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
201       /* Read activity from the map */
202       act = *(x->mb_activity_ptr);
203 
204       /* Calculate a normalized activity number */
205       a = act + 4 * cpi->activity_avg;
206       b = 4 * act + cpi->activity_avg;
207 
208       if (b >= a)
209         *(x->activity_ptr) = (int)((b + (a >> 1)) / a) - 1;
210       else
211         *(x->activity_ptr) = 1 - (int)((a + (b >> 1)) / b);
212 
213 #if OUTPUT_NORM_ACT_STATS
214       fprintf(f, " %6d", *(x->mb_activity_ptr));
215 #endif
216       /* Increment activity map pointers */
217       x->mb_activity_ptr++;
218     }
219 
220 #if OUTPUT_NORM_ACT_STATS
221     fprintf(f, "\n");
222 #endif
223   }
224 
225 #if OUTPUT_NORM_ACT_STATS
226   fclose(f);
227 #endif
228 }
229 #endif
230 
231 /* Loop through all MBs. Note activity of each, average activity and
232  * calculate a normalized activity for each
233  */
build_activity_map(VP8_COMP * cpi)234 static void build_activity_map(VP8_COMP *cpi) {
235   MACROBLOCK *const x = &cpi->mb;
236   MACROBLOCKD *xd = &x->e_mbd;
237   VP8_COMMON *const cm = &cpi->common;
238 
239 #if ALT_ACT_MEASURE
240   YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
241   int recon_yoffset;
242   int recon_y_stride = new_yv12->y_stride;
243 #endif
244 
245   int mb_row, mb_col;
246   unsigned int mb_activity;
247   int64_t activity_sum = 0;
248 
249   /* for each macroblock row in image */
250   for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
251 #if ALT_ACT_MEASURE
252     /* reset above block coeffs */
253     xd->up_available = (mb_row != 0);
254     recon_yoffset = (mb_row * recon_y_stride * 16);
255 #endif
256     /* for each macroblock col in image */
257     for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
258 #if ALT_ACT_MEASURE
259       xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
260       xd->left_available = (mb_col != 0);
261       recon_yoffset += 16;
262 #endif
263       /* Copy current mb to a buffer */
264       vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
265 
266       /* measure activity */
267       mb_activity = mb_activity_measure(cpi, x, mb_row, mb_col);
268 
269       /* Keep frame sum */
270       activity_sum += mb_activity;
271 
272       /* Store MB level activity details. */
273       *x->mb_activity_ptr = mb_activity;
274 
275       /* Increment activity map pointer */
276       x->mb_activity_ptr++;
277 
278       /* adjust to the next column of source macroblocks */
279       x->src.y_buffer += 16;
280     }
281 
282     /* adjust to the next row of mbs */
283     x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
284 
285 #if ALT_ACT_MEASURE
286     /* extend the recon for intra prediction */
287     vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16, xd->dst.u_buffer + 8,
288                       xd->dst.v_buffer + 8);
289 #endif
290   }
291 
292   /* Calculate an "average" MB activity */
293   calc_av_activity(cpi, activity_sum);
294 
295 #if USE_ACT_INDEX
296   /* Calculate an activity index number of each mb */
297   calc_activity_index(cpi, x);
298 #endif
299 }
300 
301 /* Macroblock activity masking */
vp8_activity_masking(VP8_COMP * cpi,MACROBLOCK * x)302 void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x) {
303 #if USE_ACT_INDEX
304   x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
305   x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
306   x->errorperbit += (x->errorperbit == 0);
307 #else
308   int64_t a;
309   int64_t b;
310   int64_t act = *(x->mb_activity_ptr);
311 
312   /* Apply the masking to the RD multiplier. */
313   a = act + (2 * cpi->activity_avg);
314   b = (2 * act) + cpi->activity_avg;
315 
316   x->rdmult = (unsigned int)(((int64_t)x->rdmult * b + (a >> 1)) / a);
317   x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
318   x->errorperbit += (x->errorperbit == 0);
319 #endif
320 
321   /* Activity based Zbin adjustment */
322   adjust_act_zbin(cpi, x);
323 }
324 
encode_mb_row(VP8_COMP * cpi,VP8_COMMON * cm,int mb_row,MACROBLOCK * x,MACROBLOCKD * xd,TOKENEXTRA ** tp,int * segment_counts,int * totalrate)325 static void encode_mb_row(VP8_COMP *cpi, VP8_COMMON *cm, int mb_row,
326                           MACROBLOCK *x, MACROBLOCKD *xd, TOKENEXTRA **tp,
327                           int *segment_counts, int *totalrate) {
328   int recon_yoffset, recon_uvoffset;
329   int mb_col;
330   int ref_fb_idx = cm->lst_fb_idx;
331   int dst_fb_idx = cm->new_fb_idx;
332   int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
333   int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
334   int map_index = (mb_row * cpi->common.mb_cols);
335 
336 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
337   const int num_part = (1 << cm->multi_token_partition);
338   TOKENEXTRA *tp_start = cpi->tok;
339   vp8_writer *w;
340 #endif
341 
342 #if CONFIG_MULTITHREAD
343   const int nsync = cpi->mt_sync_range;
344   vpx_atomic_int rightmost_col = VPX_ATOMIC_INIT(cm->mb_cols + nsync);
345   const vpx_atomic_int *last_row_current_mb_col;
346   vpx_atomic_int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
347 
348   if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) != 0 && mb_row != 0) {
349     last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
350   } else {
351     last_row_current_mb_col = &rightmost_col;
352   }
353 #endif
354 
355 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
356   if (num_part > 1)
357     w = &cpi->bc[1 + (mb_row % num_part)];
358   else
359     w = &cpi->bc[1];
360 #endif
361 
362   /* reset above block coeffs */
363   xd->above_context = cm->above_context;
364 
365   xd->up_available = (mb_row != 0);
366   recon_yoffset = (mb_row * recon_y_stride * 16);
367   recon_uvoffset = (mb_row * recon_uv_stride * 8);
368 
369   cpi->tplist[mb_row].start = *tp;
370   /* printf("Main mb_row = %d\n", mb_row); */
371 
372   /* Distance of Mb to the top & bottom edges, specified in 1/8th pel
373    * units as they are always compared to values that are in 1/8th pel
374    */
375   xd->mb_to_top_edge = -((mb_row * 16) << 3);
376   xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
377 
378   /* Set up limit values for vertical motion vector components
379    * to prevent them extending beyond the UMV borders
380    */
381   x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
382   x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
383 
384   /* Set the mb activity pointer to the start of the row. */
385   x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
386 
387   /* for each macroblock col in image */
388   for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
389 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
390     *tp = cpi->tok;
391 #endif
392     /* Distance of Mb to the left & right edges, specified in
393      * 1/8th pel units as they are always compared to values
394      * that are in 1/8th pel units
395      */
396     xd->mb_to_left_edge = -((mb_col * 16) << 3);
397     xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
398 
399     /* Set up limit values for horizontal motion vector components
400      * to prevent them extending beyond the UMV borders
401      */
402     x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
403     x->mv_col_max =
404         ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
405 
406     xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
407     xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
408     xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
409     xd->left_available = (mb_col != 0);
410 
411     x->rddiv = cpi->RDDIV;
412     x->rdmult = cpi->RDMULT;
413 
414     /* Copy current mb to a buffer */
415     vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
416 
417 #if CONFIG_MULTITHREAD
418     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) != 0) {
419       if (((mb_col - 1) % nsync) == 0) {
420         vpx_atomic_store_release(current_mb_col, mb_col - 1);
421       }
422 
423       if (mb_row && !(mb_col & (nsync - 1))) {
424         vp8_atomic_spin_wait(mb_col, last_row_current_mb_col, nsync);
425       }
426     }
427 #endif
428 
429     if (cpi->oxcf.tuning == VP8_TUNE_SSIM) vp8_activity_masking(cpi, x);
430 
431     /* Is segmentation enabled */
432     /* MB level adjustment to quantizer */
433     if (xd->segmentation_enabled) {
434       /* Code to set segment id in xd->mbmi.segment_id for current MB
435        * (with range checking)
436        */
437       if (cpi->segmentation_map[map_index + mb_col] <= 3) {
438         xd->mode_info_context->mbmi.segment_id =
439             cpi->segmentation_map[map_index + mb_col];
440       } else {
441         xd->mode_info_context->mbmi.segment_id = 0;
442       }
443 
444       vp8cx_mb_init_quantizer(cpi, x, 1);
445     } else {
446       /* Set to Segment 0 by default */
447       xd->mode_info_context->mbmi.segment_id = 0;
448     }
449 
450     x->active_ptr = cpi->active_map + map_index + mb_col;
451 
452     if (cm->frame_type == KEY_FRAME) {
453       *totalrate += vp8cx_encode_intra_macroblock(cpi, x, tp);
454 #ifdef MODE_STATS
455       y_modes[xd->mbmi.mode]++;
456 #endif
457     } else {
458       *totalrate += vp8cx_encode_inter_macroblock(
459           cpi, x, tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
460 
461 #ifdef MODE_STATS
462       inter_y_modes[xd->mbmi.mode]++;
463 
464       if (xd->mbmi.mode == SPLITMV) {
465         int b;
466 
467         for (b = 0; b < xd->mbmi.partition_count; ++b) {
468           inter_b_modes[x->partition->bmi[b].mode]++;
469         }
470       }
471 
472 #endif
473 
474       // Keep track of how many (consecutive) times a  block is coded
475       // as ZEROMV_LASTREF, for base layer frames.
476       // Reset to 0 if its coded as anything else.
477       if (cpi->current_layer == 0) {
478         if (xd->mode_info_context->mbmi.mode == ZEROMV &&
479             xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
480           // Increment, check for wrap-around.
481           if (cpi->consec_zero_last[map_index + mb_col] < 255) {
482             cpi->consec_zero_last[map_index + mb_col] += 1;
483           }
484           if (cpi->consec_zero_last_mvbias[map_index + mb_col] < 255) {
485             cpi->consec_zero_last_mvbias[map_index + mb_col] += 1;
486           }
487         } else {
488           cpi->consec_zero_last[map_index + mb_col] = 0;
489           cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
490         }
491         if (x->zero_last_dot_suppress) {
492           cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
493         }
494       }
495 
496       /* Special case code for cyclic refresh
497        * If cyclic update enabled then copy xd->mbmi.segment_id; (which
498        * may have been updated based on mode during
499        * vp8cx_encode_inter_macroblock()) back into the global
500        * segmentation map
501        */
502       if ((cpi->current_layer == 0) &&
503           (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)) {
504         cpi->segmentation_map[map_index + mb_col] =
505             xd->mode_info_context->mbmi.segment_id;
506 
507         /* If the block has been refreshed mark it as clean (the
508          * magnitude of the -ve influences how long it will be before
509          * we consider another refresh):
510          * Else if it was coded (last frame 0,0) and has not already
511          * been refreshed then mark it as a candidate for cleanup
512          * next time (marked 0) else mark it as dirty (1).
513          */
514         if (xd->mode_info_context->mbmi.segment_id) {
515           cpi->cyclic_refresh_map[map_index + mb_col] = -1;
516         } else if ((xd->mode_info_context->mbmi.mode == ZEROMV) &&
517                    (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)) {
518           if (cpi->cyclic_refresh_map[map_index + mb_col] == 1) {
519             cpi->cyclic_refresh_map[map_index + mb_col] = 0;
520           }
521         } else {
522           cpi->cyclic_refresh_map[map_index + mb_col] = 1;
523         }
524       }
525     }
526 
527     cpi->tplist[mb_row].stop = *tp;
528 
529 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
530     /* pack tokens for this MB */
531     {
532       int tok_count = *tp - tp_start;
533       vp8_pack_tokens(w, tp_start, tok_count);
534     }
535 #endif
536     /* Increment pointer into gf usage flags structure. */
537     x->gf_active_ptr++;
538 
539     /* Increment the activity mask pointers. */
540     x->mb_activity_ptr++;
541 
542     /* adjust to the next column of macroblocks */
543     x->src.y_buffer += 16;
544     x->src.u_buffer += 8;
545     x->src.v_buffer += 8;
546 
547     recon_yoffset += 16;
548     recon_uvoffset += 8;
549 
550     /* Keep track of segment usage */
551     segment_counts[xd->mode_info_context->mbmi.segment_id]++;
552 
553     /* skip to next mb */
554     xd->mode_info_context++;
555     x->partition_info++;
556     xd->above_context++;
557   }
558 
559   /* extend the recon for intra prediction */
560   vp8_extend_mb_row(&cm->yv12_fb[dst_fb_idx], xd->dst.y_buffer + 16,
561                     xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
562 
563 #if CONFIG_MULTITHREAD
564   if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) != 0) {
565     vpx_atomic_store_release(current_mb_col,
566                              vpx_atomic_load_acquire(&rightmost_col));
567   }
568 #endif
569 
570   /* this is to account for the border */
571   xd->mode_info_context++;
572   x->partition_info++;
573 }
574 
init_encode_frame_mb_context(VP8_COMP * cpi)575 static void init_encode_frame_mb_context(VP8_COMP *cpi) {
576   MACROBLOCK *const x = &cpi->mb;
577   VP8_COMMON *const cm = &cpi->common;
578   MACROBLOCKD *const xd = &x->e_mbd;
579 
580   /* GF active flags data structure */
581   x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
582 
583   /* Activity map pointer */
584   x->mb_activity_ptr = cpi->mb_activity_map;
585 
586   x->act_zbin_adj = 0;
587 
588   x->partition_info = x->pi;
589 
590   xd->mode_info_context = cm->mi;
591   xd->mode_info_stride = cm->mode_info_stride;
592 
593   xd->frame_type = cm->frame_type;
594 
595   /* reset intra mode contexts */
596   if (cm->frame_type == KEY_FRAME) vp8_init_mbmode_probs(cm);
597 
598   /* Copy data over into macro block data structures. */
599   x->src = *cpi->Source;
600   xd->pre = cm->yv12_fb[cm->lst_fb_idx];
601   xd->dst = cm->yv12_fb[cm->new_fb_idx];
602 
603   /* set up frame for intra coded blocks */
604   vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);
605 
606   vp8_build_block_offsets(x);
607 
608   xd->mode_info_context->mbmi.mode = DC_PRED;
609   xd->mode_info_context->mbmi.uv_mode = DC_PRED;
610 
611   xd->left_context = &cm->left_context;
612 
613   x->mvc = cm->fc.mvc;
614 
615   memset(cm->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);
616 
617   /* Special case treatment when GF and ARF are not sensible options
618    * for reference
619    */
620   if (cpi->ref_frame_flags == VP8_LAST_FRAME) {
621     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded, 255,
622                              128);
623   } else if ((cpi->oxcf.number_of_layers > 1) &&
624              (cpi->ref_frame_flags == VP8_GOLD_FRAME)) {
625     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded, 1, 255);
626   } else if ((cpi->oxcf.number_of_layers > 1) &&
627              (cpi->ref_frame_flags == VP8_ALTR_FRAME)) {
628     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded, 1, 1);
629   } else {
630     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded,
631                              cpi->prob_last_coded, cpi->prob_gf_coded);
632   }
633 
634   xd->fullpixel_mask = 0xffffffff;
635   if (cm->full_pixel) xd->fullpixel_mask = 0xfffffff8;
636 
637   vp8_zero(x->coef_counts);
638   vp8_zero(x->ymode_count);
639   vp8_zero(x->uv_mode_count) x->prediction_error = 0;
640   x->intra_error = 0;
641   vp8_zero(x->count_mb_ref_frame_usage);
642 }
643 
644 #if CONFIG_MULTITHREAD
sum_coef_counts(MACROBLOCK * x,MACROBLOCK * x_thread)645 static void sum_coef_counts(MACROBLOCK *x, MACROBLOCK *x_thread) {
646   int i = 0;
647   do {
648     int j = 0;
649     do {
650       int k = 0;
651       do {
652         /* at every context */
653 
654         /* calc probs and branch cts for this frame only */
655         int t = 0; /* token/prob index */
656 
657         do {
658           x->coef_counts[i][j][k][t] += x_thread->coef_counts[i][j][k][t];
659         } while (++t < ENTROPY_NODES);
660       } while (++k < PREV_COEF_CONTEXTS);
661     } while (++j < COEF_BANDS);
662   } while (++i < BLOCK_TYPES);
663 }
664 #endif  // CONFIG_MULTITHREAD
665 
vp8_encode_frame(VP8_COMP * cpi)666 void vp8_encode_frame(VP8_COMP *cpi) {
667   int mb_row;
668   MACROBLOCK *const x = &cpi->mb;
669   VP8_COMMON *const cm = &cpi->common;
670   MACROBLOCKD *const xd = &x->e_mbd;
671   TOKENEXTRA *tp = cpi->tok;
672   int segment_counts[MAX_MB_SEGMENTS];
673   int totalrate;
674 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
675   BOOL_CODER *bc = &cpi->bc[1]; /* bc[0] is for control partition */
676   const int num_part = (1 << cm->multi_token_partition);
677 #endif
678 
679   memset(segment_counts, 0, sizeof(segment_counts));
680   totalrate = 0;
681 
682   if (cpi->compressor_speed == 2) {
683     if (cpi->oxcf.cpu_used < 0) {
684       cpi->Speed = -(cpi->oxcf.cpu_used);
685     } else {
686       vp8_auto_select_speed(cpi);
687     }
688   }
689 
690   /* Functions setup for all frame types so we can use MC in AltRef */
691   if (!cm->use_bilinear_mc_filter) {
692     xd->subpixel_predict = vp8_sixtap_predict4x4;
693     xd->subpixel_predict8x4 = vp8_sixtap_predict8x4;
694     xd->subpixel_predict8x8 = vp8_sixtap_predict8x8;
695     xd->subpixel_predict16x16 = vp8_sixtap_predict16x16;
696   } else {
697     xd->subpixel_predict = vp8_bilinear_predict4x4;
698     xd->subpixel_predict8x4 = vp8_bilinear_predict8x4;
699     xd->subpixel_predict8x8 = vp8_bilinear_predict8x8;
700     xd->subpixel_predict16x16 = vp8_bilinear_predict16x16;
701   }
702 
703   cpi->mb.skip_true_count = 0;
704   cpi->tok_count = 0;
705 
706 #if 0
707     /* Experimental code */
708     cpi->frame_distortion = 0;
709     cpi->last_mb_distortion = 0;
710 #endif
711 
712   xd->mode_info_context = cm->mi;
713 
714   vp8_zero(cpi->mb.MVcount);
715 
716   vp8cx_frame_init_quantizer(cpi);
717 
718   vp8_initialize_rd_consts(cpi, x,
719                            vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
720 
721   vp8cx_initialize_me_consts(cpi, cm->base_qindex);
722 
723   if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
724     /* Initialize encode frame context. */
725     init_encode_frame_mb_context(cpi);
726 
727     /* Build a frame level activity map */
728     build_activity_map(cpi);
729   }
730 
731   /* re-init encode frame context. */
732   init_encode_frame_mb_context(cpi);
733 
734 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
735   {
736     int i;
737     for (i = 0; i < num_part; ++i) {
738       vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
739                        cpi->partition_d_end[i + 1]);
740       bc[i].error = &cm->error;
741     }
742   }
743 
744 #endif
745 
746   {
747     struct vpx_usec_timer emr_timer;
748     vpx_usec_timer_start(&emr_timer);
749 
750 #if CONFIG_MULTITHREAD
751     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded)) {
752       int i;
753 
754       vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei,
755                                 cpi->encoding_thread_count);
756 
757       for (i = 0; i < cm->mb_rows; ++i)
758         vpx_atomic_store_release(&cpi->mt_current_mb_col[i], -1);
759 
760       for (i = 0; i < cpi->encoding_thread_count; ++i) {
761         sem_post(&cpi->h_event_start_encoding[i]);
762       }
763 
764       for (mb_row = 0; mb_row < cm->mb_rows;
765            mb_row += (cpi->encoding_thread_count + 1)) {
766         vp8_zero(cm->left_context)
767 
768 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
769             tp = cpi->tok;
770 #else
771             tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
772 #endif
773 
774         encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
775 
776         /* adjust to the next row of mbs */
777         x->src.y_buffer +=
778             16 * x->src.y_stride * (cpi->encoding_thread_count + 1) -
779             16 * cm->mb_cols;
780         x->src.u_buffer +=
781             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
782             8 * cm->mb_cols;
783         x->src.v_buffer +=
784             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
785             8 * cm->mb_cols;
786 
787         xd->mode_info_context +=
788             xd->mode_info_stride * cpi->encoding_thread_count;
789         x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
790         x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
791       }
792       /* Wait for all the threads to finish. */
793       for (i = 0; i < cpi->encoding_thread_count; ++i) {
794         sem_wait(&cpi->h_event_end_encoding[i]);
795       }
796 
797       for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
798         cpi->tok_count += (unsigned int)(cpi->tplist[mb_row].stop -
799                                          cpi->tplist[mb_row].start);
800       }
801 
802       if (xd->segmentation_enabled) {
803         int j;
804 
805         if (xd->segmentation_enabled) {
806           for (i = 0; i < cpi->encoding_thread_count; ++i) {
807             for (j = 0; j < 4; ++j) {
808               segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
809             }
810           }
811         }
812       }
813 
814       for (i = 0; i < cpi->encoding_thread_count; ++i) {
815         int mode_count;
816         int c_idx;
817         totalrate += cpi->mb_row_ei[i].totalrate;
818 
819         cpi->mb.skip_true_count += cpi->mb_row_ei[i].mb.skip_true_count;
820 
821         for (mode_count = 0; mode_count < VP8_YMODES; ++mode_count) {
822           cpi->mb.ymode_count[mode_count] +=
823               cpi->mb_row_ei[i].mb.ymode_count[mode_count];
824         }
825 
826         for (mode_count = 0; mode_count < VP8_UV_MODES; ++mode_count) {
827           cpi->mb.uv_mode_count[mode_count] +=
828               cpi->mb_row_ei[i].mb.uv_mode_count[mode_count];
829         }
830 
831         for (c_idx = 0; c_idx < MVvals; ++c_idx) {
832           cpi->mb.MVcount[0][c_idx] += cpi->mb_row_ei[i].mb.MVcount[0][c_idx];
833           cpi->mb.MVcount[1][c_idx] += cpi->mb_row_ei[i].mb.MVcount[1][c_idx];
834         }
835 
836         cpi->mb.prediction_error += cpi->mb_row_ei[i].mb.prediction_error;
837         cpi->mb.intra_error += cpi->mb_row_ei[i].mb.intra_error;
838 
839         for (c_idx = 0; c_idx < MAX_REF_FRAMES; ++c_idx) {
840           cpi->mb.count_mb_ref_frame_usage[c_idx] +=
841               cpi->mb_row_ei[i].mb.count_mb_ref_frame_usage[c_idx];
842         }
843 
844         for (c_idx = 0; c_idx < MAX_ERROR_BINS; ++c_idx) {
845           cpi->mb.error_bins[c_idx] += cpi->mb_row_ei[i].mb.error_bins[c_idx];
846         }
847 
848         /* add up counts for each thread */
849         sum_coef_counts(x, &cpi->mb_row_ei[i].mb);
850       }
851 
852     } else
853 #endif  // CONFIG_MULTITHREAD
854     {
855 
856       /* for each macroblock row in image */
857       for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
858         vp8_zero(cm->left_context)
859 
860 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
861             tp = cpi->tok;
862 #endif
863 
864         encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
865 
866         /* adjust to the next row of mbs */
867         x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
868         x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
869         x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
870       }
871 
872       cpi->tok_count = (unsigned int)(tp - cpi->tok);
873     }
874 
875 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
876     {
877       int i;
878       for (i = 0; i < num_part; ++i) {
879         vp8_stop_encode(&bc[i]);
880         cpi->partition_sz[i + 1] = bc[i].pos;
881       }
882     }
883 #endif
884 
885     vpx_usec_timer_mark(&emr_timer);
886     cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
887   }
888 
889   // Work out the segment probabilities if segmentation is enabled
890   // and needs to be updated
891   if (xd->segmentation_enabled && xd->update_mb_segmentation_map) {
892     int tot_count;
893     int i;
894 
895     /* Set to defaults */
896     memset(xd->mb_segment_tree_probs, 255, sizeof(xd->mb_segment_tree_probs));
897 
898     tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] +
899                 segment_counts[3];
900 
901     if (tot_count) {
902       xd->mb_segment_tree_probs[0] =
903           ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
904 
905       tot_count = segment_counts[0] + segment_counts[1];
906 
907       if (tot_count > 0) {
908         xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
909       }
910 
911       tot_count = segment_counts[2] + segment_counts[3];
912 
913       if (tot_count > 0) {
914         xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
915       }
916 
917       /* Zero probabilities not allowed */
918       for (i = 0; i < MB_FEATURE_TREE_PROBS; ++i) {
919         if (xd->mb_segment_tree_probs[i] == 0) xd->mb_segment_tree_probs[i] = 1;
920       }
921     }
922   }
923 
924   /* projected_frame_size in units of BYTES */
925   cpi->projected_frame_size = totalrate >> 8;
926 
927   /* Make a note of the percentage MBs coded Intra. */
928   if (cm->frame_type == KEY_FRAME) {
929     cpi->this_frame_percent_intra = 100;
930   } else {
931     int tot_modes;
932 
933     tot_modes = cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] +
934                 cpi->mb.count_mb_ref_frame_usage[LAST_FRAME] +
935                 cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME] +
936                 cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
937 
938     if (tot_modes) {
939       cpi->this_frame_percent_intra =
940           cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
941     }
942   }
943 
944 #if !CONFIG_REALTIME_ONLY
945   /* Adjust the projected reference frame usage probability numbers to
946    * reflect what we have just seen. This may be useful when we make
947    * multiple iterations of the recode loop rather than continuing to use
948    * values from the previous frame.
949    */
950   if ((cm->frame_type != KEY_FRAME) &&
951       ((cpi->oxcf.number_of_layers > 1) ||
952        (!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame))) {
953     vp8_convert_rfct_to_prob(cpi);
954   }
955 #endif
956 }
vp8_setup_block_ptrs(MACROBLOCK * x)957 void vp8_setup_block_ptrs(MACROBLOCK *x) {
958   int r, c;
959   int i;
960 
961   for (r = 0; r < 4; ++r) {
962     for (c = 0; c < 4; ++c) {
963       x->block[r * 4 + c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
964     }
965   }
966 
967   for (r = 0; r < 2; ++r) {
968     for (c = 0; c < 2; ++c) {
969       x->block[16 + r * 2 + c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
970     }
971   }
972 
973   for (r = 0; r < 2; ++r) {
974     for (c = 0; c < 2; ++c) {
975       x->block[20 + r * 2 + c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
976     }
977   }
978 
979   x->block[24].src_diff = x->src_diff + 384;
980 
981   for (i = 0; i < 25; ++i) {
982     x->block[i].coeff = x->coeff + i * 16;
983   }
984 }
985 
vp8_build_block_offsets(MACROBLOCK * x)986 void vp8_build_block_offsets(MACROBLOCK *x) {
987   int block = 0;
988   int br, bc;
989 
990   vp8_build_block_doffsets(&x->e_mbd);
991 
992   /* y blocks */
993   x->thismb_ptr = &x->thismb[0];
994   for (br = 0; br < 4; ++br) {
995     for (bc = 0; bc < 4; ++bc) {
996       BLOCK *this_block = &x->block[block];
997       this_block->base_src = &x->thismb_ptr;
998       this_block->src_stride = 16;
999       this_block->src = 4 * br * 16 + 4 * bc;
1000       ++block;
1001     }
1002   }
1003 
1004   /* u blocks */
1005   for (br = 0; br < 2; ++br) {
1006     for (bc = 0; bc < 2; ++bc) {
1007       BLOCK *this_block = &x->block[block];
1008       this_block->base_src = &x->src.u_buffer;
1009       this_block->src_stride = x->src.uv_stride;
1010       this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1011       ++block;
1012     }
1013   }
1014 
1015   /* v blocks */
1016   for (br = 0; br < 2; ++br) {
1017     for (bc = 0; bc < 2; ++bc) {
1018       BLOCK *this_block = &x->block[block];
1019       this_block->base_src = &x->src.v_buffer;
1020       this_block->src_stride = x->src.uv_stride;
1021       this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1022       ++block;
1023     }
1024   }
1025 }
1026 
sum_intra_stats(VP8_COMP * cpi,MACROBLOCK * x)1027 static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x) {
1028   const MACROBLOCKD *xd = &x->e_mbd;
1029   const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
1030   const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;
1031 
1032 #ifdef MODE_STATS
1033   const int is_key = cpi->common.frame_type == KEY_FRAME;
1034 
1035   ++(is_key ? uv_modes : inter_uv_modes)[uvm];
1036 
1037   if (m == B_PRED) {
1038     unsigned int *const bct = is_key ? b_modes : inter_b_modes;
1039 
1040     int b = 0;
1041 
1042     do {
1043       ++bct[xd->block[b].bmi.mode];
1044     } while (++b < 16);
1045   }
1046 
1047 #else
1048   (void)cpi;
1049 #endif
1050 
1051   ++x->ymode_count[m];
1052   ++x->uv_mode_count[uvm];
1053 }
1054 
1055 /* Experimental stub function to create a per MB zbin adjustment based on
1056  * some previously calculated measure of MB activity.
1057  */
adjust_act_zbin(VP8_COMP * cpi,MACROBLOCK * x)1058 static void adjust_act_zbin(VP8_COMP *cpi, MACROBLOCK *x) {
1059 #if USE_ACT_INDEX
1060   x->act_zbin_adj = *(x->mb_activity_ptr);
1061 #else
1062   int64_t a;
1063   int64_t b;
1064   int64_t act = *(x->mb_activity_ptr);
1065 
1066   /* Apply the masking to the RD multiplier. */
1067   a = act + 4 * cpi->activity_avg;
1068   b = 4 * act + cpi->activity_avg;
1069 
1070   if (act > cpi->activity_avg) {
1071     x->act_zbin_adj = (int)(((int64_t)b + (a >> 1)) / a) - 1;
1072   } else {
1073     x->act_zbin_adj = 1 - (int)(((int64_t)a + (b >> 1)) / b);
1074   }
1075 #endif
1076 }
1077 
vp8cx_encode_intra_macroblock(VP8_COMP * cpi,MACROBLOCK * x,TOKENEXTRA ** t)1078 int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
1079                                   TOKENEXTRA **t) {
1080   MACROBLOCKD *xd = &x->e_mbd;
1081   int rate;
1082 
1083   if (cpi->sf.RD && cpi->compressor_speed != 2) {
1084     vp8_rd_pick_intra_mode(x, &rate);
1085   } else {
1086     vp8_pick_intra_mode(x, &rate);
1087   }
1088 
1089   if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
1090     adjust_act_zbin(cpi, x);
1091     vp8_update_zbin_extra(cpi, x);
1092   }
1093 
1094   if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED) {
1095     vp8_encode_intra4x4mby(x);
1096   } else {
1097     vp8_encode_intra16x16mby(x);
1098   }
1099 
1100   vp8_encode_intra16x16mbuv(x);
1101 
1102   sum_intra_stats(cpi, x);
1103 
1104   vp8_tokenize_mb(cpi, x, t);
1105 
1106   if (xd->mode_info_context->mbmi.mode != B_PRED) vp8_inverse_transform_mby(xd);
1107 
1108   vp8_dequant_idct_add_uv_block(xd->qcoeff + 16 * 16, xd->dequant_uv,
1109                                 xd->dst.u_buffer, xd->dst.v_buffer,
1110                                 xd->dst.uv_stride, xd->eobs + 16);
1111   return rate;
1112 }
1113 #ifdef SPEEDSTATS
1114 extern int cnt_pm;
1115 #endif
1116 
1117 extern void vp8_fix_contexts(MACROBLOCKD *x);
1118 
vp8cx_encode_inter_macroblock(VP8_COMP * cpi,MACROBLOCK * x,TOKENEXTRA ** t,int recon_yoffset,int recon_uvoffset,int mb_row,int mb_col)1119 int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
1120                                   int recon_yoffset, int recon_uvoffset,
1121                                   int mb_row, int mb_col) {
1122   MACROBLOCKD *const xd = &x->e_mbd;
1123   int intra_error = 0;
1124   int rate;
1125   int distortion;
1126 
1127   x->skip = 0;
1128 
1129   if (xd->segmentation_enabled) {
1130     x->encode_breakout =
1131         cpi->segment_encode_breakout[xd->mode_info_context->mbmi.segment_id];
1132   } else {
1133     x->encode_breakout = cpi->oxcf.encode_breakout;
1134   }
1135 
1136 #if CONFIG_TEMPORAL_DENOISING
1137   /* Reset the best sse mode/mv for each macroblock. */
1138   x->best_reference_frame = INTRA_FRAME;
1139   x->best_zeromv_reference_frame = INTRA_FRAME;
1140   x->best_sse_inter_mode = 0;
1141   x->best_sse_mv.as_int = 0;
1142   x->need_to_clamp_best_mvs = 0;
1143 #endif
1144 
1145   if (cpi->sf.RD) {
1146     int zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
1147 
1148     /* Are we using the fast quantizer for the mode selection? */
1149     if (cpi->sf.use_fastquant_for_pick) {
1150       x->quantize_b = vp8_fast_quantize_b;
1151 
1152       /* the fast quantizer does not use zbin_extra, so
1153        * do not recalculate */
1154       x->zbin_mode_boost_enabled = 0;
1155     }
1156     vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1157                            &distortion, &intra_error, mb_row, mb_col);
1158 
1159     /* switch back to the regular quantizer for the encode */
1160     if (cpi->sf.improved_quant) {
1161       x->quantize_b = vp8_regular_quantize_b;
1162     }
1163 
1164     /* restore cpi->zbin_mode_boost_enabled */
1165     x->zbin_mode_boost_enabled = zbin_mode_boost_enabled;
1166 
1167   } else {
1168     vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1169                         &distortion, &intra_error, mb_row, mb_col);
1170   }
1171 
1172   x->prediction_error += distortion;
1173   x->intra_error += intra_error;
1174 
1175   if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
1176     /* Adjust the zbin based on this MB rate. */
1177     adjust_act_zbin(cpi, x);
1178   }
1179 
1180 #if 0
1181     /* Experimental RD code */
1182     cpi->frame_distortion += distortion;
1183     cpi->last_mb_distortion = distortion;
1184 #endif
1185 
1186   /* MB level adjutment to quantizer setup */
1187   if (xd->segmentation_enabled) {
1188     /* If cyclic update enabled */
1189     if (cpi->current_layer == 0 && cpi->cyclic_refresh_mode_enabled) {
1190       /* Clear segment_id back to 0 if not coded (last frame 0,0) */
1191       if ((xd->mode_info_context->mbmi.segment_id == 1) &&
1192           ((xd->mode_info_context->mbmi.ref_frame != LAST_FRAME) ||
1193            (xd->mode_info_context->mbmi.mode != ZEROMV))) {
1194         xd->mode_info_context->mbmi.segment_id = 0;
1195 
1196         /* segment_id changed, so update */
1197         vp8cx_mb_init_quantizer(cpi, x, 1);
1198       }
1199     }
1200   }
1201 
1202   {
1203     /* Experimental code.
1204      * Special case for gf and arf zeromv modes, for 1 temporal layer.
1205      * Increase zbin size to supress noise.
1206      */
1207     x->zbin_mode_boost = 0;
1208     if (x->zbin_mode_boost_enabled) {
1209       if (xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME) {
1210         if (xd->mode_info_context->mbmi.mode == ZEROMV) {
1211           if (xd->mode_info_context->mbmi.ref_frame != LAST_FRAME &&
1212               cpi->oxcf.number_of_layers == 1) {
1213             x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
1214           } else {
1215             x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
1216           }
1217         } else if (xd->mode_info_context->mbmi.mode == SPLITMV) {
1218           x->zbin_mode_boost = 0;
1219         } else {
1220           x->zbin_mode_boost = MV_ZBIN_BOOST;
1221         }
1222       }
1223     }
1224 
1225     /* The fast quantizer doesn't use zbin_extra, only do so with
1226      * the regular quantizer. */
1227     if (cpi->sf.improved_quant) vp8_update_zbin_extra(cpi, x);
1228   }
1229 
1230   x->count_mb_ref_frame_usage[xd->mode_info_context->mbmi.ref_frame]++;
1231 
1232   if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) {
1233     vp8_encode_intra16x16mbuv(x);
1234 
1235     if (xd->mode_info_context->mbmi.mode == B_PRED) {
1236       vp8_encode_intra4x4mby(x);
1237     } else {
1238       vp8_encode_intra16x16mby(x);
1239     }
1240 
1241     sum_intra_stats(cpi, x);
1242   } else {
1243     int ref_fb_idx;
1244 
1245     if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
1246       ref_fb_idx = cpi->common.lst_fb_idx;
1247     } else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME) {
1248       ref_fb_idx = cpi->common.gld_fb_idx;
1249     } else {
1250       ref_fb_idx = cpi->common.alt_fb_idx;
1251     }
1252 
1253     xd->pre.y_buffer = cpi->common.yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
1254     xd->pre.u_buffer =
1255         cpi->common.yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
1256     xd->pre.v_buffer =
1257         cpi->common.yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
1258 
1259     if (!x->skip) {
1260       vp8_encode_inter16x16(x);
1261     } else {
1262       vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer, xd->dst.u_buffer,
1263                                          xd->dst.v_buffer, xd->dst.y_stride,
1264                                          xd->dst.uv_stride);
1265     }
1266   }
1267 
1268   if (!x->skip) {
1269     vp8_tokenize_mb(cpi, x, t);
1270 
1271     if (xd->mode_info_context->mbmi.mode != B_PRED) {
1272       vp8_inverse_transform_mby(xd);
1273     }
1274 
1275     vp8_dequant_idct_add_uv_block(xd->qcoeff + 16 * 16, xd->dequant_uv,
1276                                   xd->dst.u_buffer, xd->dst.v_buffer,
1277                                   xd->dst.uv_stride, xd->eobs + 16);
1278   } else {
1279     /* always set mb_skip_coeff as it is needed by the loopfilter */
1280     xd->mode_info_context->mbmi.mb_skip_coeff = 1;
1281 
1282     if (cpi->common.mb_no_coeff_skip) {
1283       x->skip_true_count++;
1284       vp8_fix_contexts(xd);
1285     } else {
1286       vp8_stuff_mb(cpi, x, t);
1287     }
1288   }
1289 
1290   return rate;
1291 }
1292