• 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] = { 128, 128, 128, 128, 128, 128,
68                                                 128, 128, 128, 128, 128, 128,
69                                                 128, 128, 128, 128 };
70 
71 /* Original activity measure from Tim T's code. */
tt_activity_measure(MACROBLOCK * x)72 static unsigned int tt_activity_measure(MACROBLOCK *x) {
73   unsigned int act;
74   unsigned int sse;
75   /* TODO: This could also be done over smaller areas (8x8), but that would
76    *  require extensive changes elsewhere, as lambda is assumed to be fixed
77    *  over an entire MB in most of the code.
78    * Another option is to compute four 8x8 variances, and pick a single
79    *  lambda using a non-linear combination (e.g., the smallest, or second
80    *  smallest, etc.).
81    */
82   act = vpx_variance16x16(x->src.y_buffer, x->src.y_stride, VP8_VAR_OFFS, 0,
83                           &sse);
84   act = act << 4;
85 
86   /* If the region is flat, lower the activity some more. */
87   if (act < 8 << 12) act = act < 5 << 12 ? act : 5 << 12;
88 
89   return act;
90 }
91 
92 /* Measure the activity of the current macroblock
93  * What we measure here is TBD so abstracted to this function
94  */
95 #define ALT_ACT_MEASURE 1
mb_activity_measure(MACROBLOCK * x,int mb_row,int mb_col)96 static unsigned int mb_activity_measure(MACROBLOCK *x, int mb_row, int mb_col) {
97   unsigned int mb_activity;
98 
99   if (ALT_ACT_MEASURE) {
100     int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
101 
102     /* Or use an alternative. */
103     mb_activity = vp8_encode_intra(x, use_dc_pred);
104   } else {
105     /* Original activity measure from Tim T's code. */
106     mb_activity = tt_activity_measure(x);
107   }
108 
109   if (mb_activity < VP8_ACTIVITY_AVG_MIN) mb_activity = VP8_ACTIVITY_AVG_MIN;
110 
111   return mb_activity;
112 }
113 
114 /* Calculate an "average" mb activity value for the frame */
115 #define ACT_MEDIAN 0
calc_av_activity(VP8_COMP * cpi,int64_t activity_sum)116 static void calc_av_activity(VP8_COMP *cpi, int64_t activity_sum) {
117 #if ACT_MEDIAN
118   /* Find median: Simple n^2 algorithm for experimentation */
119   {
120     unsigned int median;
121     unsigned int i, j;
122     unsigned int *sortlist;
123     unsigned int tmp;
124 
125     /* Create a list to sort to */
126     CHECK_MEM_ERROR(sortlist,
127                     vpx_calloc(sizeof(unsigned int), cpi->common.MBs));
128 
129     /* Copy map to sort list */
130     memcpy(sortlist, cpi->mb_activity_map,
131            sizeof(unsigned int) * cpi->common.MBs);
132 
133     /* Ripple each value down to its correct position */
134     for (i = 1; i < cpi->common.MBs; ++i) {
135       for (j = i; j > 0; j--) {
136         if (sortlist[j] < sortlist[j - 1]) {
137           /* Swap values */
138           tmp = sortlist[j - 1];
139           sortlist[j - 1] = sortlist[j];
140           sortlist[j] = tmp;
141         } else
142           break;
143       }
144     }
145 
146     /* Even number MBs so estimate median as mean of two either side. */
147     median = (1 + sortlist[cpi->common.MBs >> 1] +
148               sortlist[(cpi->common.MBs >> 1) + 1]) >>
149              1;
150 
151     cpi->activity_avg = median;
152 
153     vpx_free(sortlist);
154   }
155 #else
156   /* Simple mean for now */
157   cpi->activity_avg = (unsigned int)(activity_sum / cpi->common.MBs);
158 #endif
159 
160   if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN) {
161     cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;
162   }
163 
164   /* Experimental code: return fixed value normalized for several clips */
165   if (ALT_ACT_MEASURE) cpi->activity_avg = 100000;
166 }
167 
168 #define USE_ACT_INDEX 0
169 #define OUTPUT_NORM_ACT_STATS 0
170 
171 #if USE_ACT_INDEX
172 /* Calculate and activity index for each mb */
calc_activity_index(VP8_COMP * cpi,MACROBLOCK * x)173 static void calc_activity_index(VP8_COMP *cpi, MACROBLOCK *x) {
174   VP8_COMMON *const cm = &cpi->common;
175   int mb_row, mb_col;
176 
177   int64_t act;
178   int64_t a;
179   int64_t b;
180 
181 #if OUTPUT_NORM_ACT_STATS
182   FILE *f = fopen("norm_act.stt", "a");
183   fprintf(f, "\n%12d\n", cpi->activity_avg);
184 #endif
185 
186   /* Reset pointers to start of activity map */
187   x->mb_activity_ptr = cpi->mb_activity_map;
188 
189   /* Calculate normalized mb activity number. */
190   for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
191     /* for each macroblock col in image */
192     for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
193       /* Read activity from the map */
194       act = *(x->mb_activity_ptr);
195 
196       /* Calculate a normalized activity number */
197       a = act + 4 * cpi->activity_avg;
198       b = 4 * act + cpi->activity_avg;
199 
200       if (b >= a)
201         *(x->activity_ptr) = (int)((b + (a >> 1)) / a) - 1;
202       else
203         *(x->activity_ptr) = 1 - (int)((a + (b >> 1)) / b);
204 
205 #if OUTPUT_NORM_ACT_STATS
206       fprintf(f, " %6d", *(x->mb_activity_ptr));
207 #endif
208       /* Increment activity map pointers */
209       x->mb_activity_ptr++;
210     }
211 
212 #if OUTPUT_NORM_ACT_STATS
213     fprintf(f, "\n");
214 #endif
215   }
216 
217 #if OUTPUT_NORM_ACT_STATS
218   fclose(f);
219 #endif
220 }
221 #endif
222 
223 /* Loop through all MBs. Note activity of each, average activity and
224  * calculate a normalized activity for each
225  */
build_activity_map(VP8_COMP * cpi)226 static void build_activity_map(VP8_COMP *cpi) {
227   MACROBLOCK *const x = &cpi->mb;
228   MACROBLOCKD *xd = &x->e_mbd;
229   VP8_COMMON *const cm = &cpi->common;
230 
231 #if ALT_ACT_MEASURE
232   YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
233   int recon_yoffset;
234   int recon_y_stride = new_yv12->y_stride;
235 #endif
236 
237   int mb_row, mb_col;
238   unsigned int mb_activity;
239   int64_t activity_sum = 0;
240 
241   /* for each macroblock row in image */
242   for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
243 #if ALT_ACT_MEASURE
244     /* reset above block coeffs */
245     xd->up_available = (mb_row != 0);
246     recon_yoffset = (mb_row * recon_y_stride * 16);
247 #endif
248     /* for each macroblock col in image */
249     for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
250 #if ALT_ACT_MEASURE
251       xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
252       xd->left_available = (mb_col != 0);
253       recon_yoffset += 16;
254 #endif
255       /* Copy current mb to a buffer */
256       vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
257 
258       /* measure activity */
259       mb_activity = mb_activity_measure(x, mb_row, mb_col);
260 
261       /* Keep frame sum */
262       activity_sum += mb_activity;
263 
264       /* Store MB level activity details. */
265       *x->mb_activity_ptr = mb_activity;
266 
267       /* Increment activity map pointer */
268       x->mb_activity_ptr++;
269 
270       /* adjust to the next column of source macroblocks */
271       x->src.y_buffer += 16;
272     }
273 
274     /* adjust to the next row of mbs */
275     x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
276 
277 #if ALT_ACT_MEASURE
278     /* extend the recon for intra prediction */
279     vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16, xd->dst.u_buffer + 8,
280                       xd->dst.v_buffer + 8);
281 #endif
282   }
283 
284   /* Calculate an "average" MB activity */
285   calc_av_activity(cpi, activity_sum);
286 
287 #if USE_ACT_INDEX
288   /* Calculate an activity index number of each mb */
289   calc_activity_index(cpi, x);
290 #endif
291 }
292 
293 /* Macroblock activity masking */
vp8_activity_masking(VP8_COMP * cpi,MACROBLOCK * x)294 void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x) {
295 #if USE_ACT_INDEX
296   x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
297   x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
298   x->errorperbit += (x->errorperbit == 0);
299 #else
300   int64_t a;
301   int64_t b;
302   int64_t act = *(x->mb_activity_ptr);
303 
304   /* Apply the masking to the RD multiplier. */
305   a = act + (2 * cpi->activity_avg);
306   b = (2 * act) + cpi->activity_avg;
307 
308   x->rdmult = (unsigned int)(((int64_t)x->rdmult * b + (a >> 1)) / a);
309   x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
310   x->errorperbit += (x->errorperbit == 0);
311 #endif
312 
313   /* Activity based Zbin adjustment */
314   adjust_act_zbin(cpi, x);
315 }
316 
encode_mb_row(VP8_COMP * cpi,VP8_COMMON * cm,int mb_row,MACROBLOCK * x,MACROBLOCKD * xd,TOKENEXTRA ** tp,int * segment_counts,int * totalrate)317 static void encode_mb_row(VP8_COMP *cpi, VP8_COMMON *cm, int mb_row,
318                           MACROBLOCK *x, MACROBLOCKD *xd, TOKENEXTRA **tp,
319                           int *segment_counts, int *totalrate) {
320   int recon_yoffset, recon_uvoffset;
321   int mb_col;
322   int ref_fb_idx = cm->lst_fb_idx;
323   int dst_fb_idx = cm->new_fb_idx;
324   int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
325   int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
326   int map_index = (mb_row * cpi->common.mb_cols);
327 
328 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
329   const int num_part = (1 << cm->multi_token_partition);
330   TOKENEXTRA *tp_start = cpi->tok;
331   vp8_writer *w;
332 #endif
333 
334 #if CONFIG_MULTITHREAD
335   const int nsync = cpi->mt_sync_range;
336   vpx_atomic_int rightmost_col = VPX_ATOMIC_INIT(cm->mb_cols + nsync);
337   const vpx_atomic_int *last_row_current_mb_col;
338   vpx_atomic_int *current_mb_col = NULL;
339 
340   if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) != 0) {
341     current_mb_col = &cpi->mt_current_mb_col[mb_row];
342   }
343   if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) != 0 && mb_row != 0) {
344     last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
345   } else {
346     last_row_current_mb_col = &rightmost_col;
347   }
348 #endif
349 
350 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
351   if (num_part > 1)
352     w = &cpi->bc[1 + (mb_row % num_part)];
353   else
354     w = &cpi->bc[1];
355 #endif
356 
357   /* reset above block coeffs */
358   xd->above_context = cm->above_context;
359 
360   xd->up_available = (mb_row != 0);
361   recon_yoffset = (mb_row * recon_y_stride * 16);
362   recon_uvoffset = (mb_row * recon_uv_stride * 8);
363 
364   cpi->tplist[mb_row].start = *tp;
365   /* printf("Main mb_row = %d\n", mb_row); */
366 
367   /* Distance of Mb to the top & bottom edges, specified in 1/8th pel
368    * units as they are always compared to values that are in 1/8th pel
369    */
370   xd->mb_to_top_edge = -((mb_row * 16) << 3);
371   xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
372 
373   /* Set up limit values for vertical motion vector components
374    * to prevent them extending beyond the UMV borders
375    */
376   x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
377   x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16) + (VP8BORDERINPIXELS - 16);
378 
379   /* Set the mb activity pointer to the start of the row. */
380   x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
381 
382   /* for each macroblock col in image */
383   for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
384 #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
385     *tp = cpi->tok;
386 #endif
387     /* Distance of Mb to the left & right edges, specified in
388      * 1/8th pel units as they are always compared to values
389      * that are in 1/8th pel units
390      */
391     xd->mb_to_left_edge = -((mb_col * 16) << 3);
392     xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
393 
394     /* Set up limit values for horizontal motion vector components
395      * to prevent them extending beyond the UMV borders
396      */
397     x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
398     x->mv_col_max =
399         ((cm->mb_cols - 1 - mb_col) * 16) + (VP8BORDERINPIXELS - 16);
400 
401     xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
402     xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
403     xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
404     xd->left_available = (mb_col != 0);
405 
406     x->rddiv = cpi->RDDIV;
407     x->rdmult = cpi->RDMULT;
408 
409     /* Copy current mb to a buffer */
410     vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
411 
412 #if CONFIG_MULTITHREAD
413     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) != 0) {
414       if (((mb_col - 1) % nsync) == 0) {
415         vpx_atomic_store_release(current_mb_col, mb_col - 1);
416       }
417 
418       if (mb_row && !(mb_col & (nsync - 1))) {
419         vp8_atomic_spin_wait(mb_col, last_row_current_mb_col, nsync);
420       }
421     }
422 #endif
423 
424     if (cpi->oxcf.tuning == VP8_TUNE_SSIM) vp8_activity_masking(cpi, x);
425 
426     /* Is segmentation enabled */
427     /* MB level adjustment to quantizer */
428     if (xd->segmentation_enabled) {
429       /* Code to set segment id in xd->mbmi.segment_id for current MB
430        * (with range checking)
431        */
432       if (cpi->segmentation_map[map_index + mb_col] <= 3) {
433         xd->mode_info_context->mbmi.segment_id =
434             cpi->segmentation_map[map_index + mb_col];
435       } else {
436         xd->mode_info_context->mbmi.segment_id = 0;
437       }
438 
439       vp8cx_mb_init_quantizer(cpi, x, 1);
440     } else {
441       /* Set to Segment 0 by default */
442       xd->mode_info_context->mbmi.segment_id = 0;
443     }
444 
445     x->active_ptr = cpi->active_map + map_index + mb_col;
446 
447     if (cm->frame_type == KEY_FRAME) {
448       *totalrate += vp8cx_encode_intra_macroblock(cpi, x, tp);
449 #ifdef MODE_STATS
450       y_modes[xd->mbmi.mode]++;
451 #endif
452     } else {
453       *totalrate += vp8cx_encode_inter_macroblock(
454           cpi, x, tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
455 
456 #ifdef MODE_STATS
457       inter_y_modes[xd->mbmi.mode]++;
458 
459       if (xd->mbmi.mode == SPLITMV) {
460         int b;
461 
462         for (b = 0; b < xd->mbmi.partition_count; ++b) {
463           inter_b_modes[x->partition->bmi[b].mode]++;
464         }
465       }
466 
467 #endif
468 
469       // Keep track of how many (consecutive) times a  block is coded
470       // as ZEROMV_LASTREF, for base layer frames.
471       // Reset to 0 if its coded as anything else.
472       if (cpi->current_layer == 0) {
473         if (xd->mode_info_context->mbmi.mode == ZEROMV &&
474             xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
475           // Increment, check for wrap-around.
476           if (cpi->consec_zero_last[map_index + mb_col] < 255) {
477             cpi->consec_zero_last[map_index + mb_col] += 1;
478           }
479           if (cpi->consec_zero_last_mvbias[map_index + mb_col] < 255) {
480             cpi->consec_zero_last_mvbias[map_index + mb_col] += 1;
481           }
482         } else {
483           cpi->consec_zero_last[map_index + mb_col] = 0;
484           cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
485         }
486         if (x->zero_last_dot_suppress) {
487           cpi->consec_zero_last_mvbias[map_index + mb_col] = 0;
488         }
489       }
490 
491       /* Special case code for cyclic refresh
492        * If cyclic update enabled then copy xd->mbmi.segment_id; (which
493        * may have been updated based on mode during
494        * vp8cx_encode_inter_macroblock()) back into the global
495        * segmentation map
496        */
497       if ((cpi->current_layer == 0) &&
498           (cpi->cyclic_refresh_mode_enabled && xd->segmentation_enabled)) {
499         cpi->segmentation_map[map_index + mb_col] =
500             xd->mode_info_context->mbmi.segment_id;
501 
502         /* If the block has been refreshed mark it as clean (the
503          * magnitude of the -ve influences how long it will be before
504          * we consider another refresh):
505          * Else if it was coded (last frame 0,0) and has not already
506          * been refreshed then mark it as a candidate for cleanup
507          * next time (marked 0) else mark it as dirty (1).
508          */
509         if (xd->mode_info_context->mbmi.segment_id) {
510           cpi->cyclic_refresh_map[map_index + mb_col] = -1;
511         } else if ((xd->mode_info_context->mbmi.mode == ZEROMV) &&
512                    (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)) {
513           if (cpi->cyclic_refresh_map[map_index + mb_col] == 1) {
514             cpi->cyclic_refresh_map[map_index + mb_col] = 0;
515           }
516         } else {
517           cpi->cyclic_refresh_map[map_index + mb_col] = 1;
518         }
519       }
520     }
521 
522     cpi->tplist[mb_row].stop = *tp;
523 
524 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
525     /* pack tokens for this MB */
526     {
527       int tok_count = *tp - tp_start;
528       vp8_pack_tokens(w, tp_start, tok_count);
529     }
530 #endif
531     /* Increment pointer into gf usage flags structure. */
532     x->gf_active_ptr++;
533 
534     /* Increment the activity mask pointers. */
535     x->mb_activity_ptr++;
536 
537     /* adjust to the next column of macroblocks */
538     x->src.y_buffer += 16;
539     x->src.u_buffer += 8;
540     x->src.v_buffer += 8;
541 
542     recon_yoffset += 16;
543     recon_uvoffset += 8;
544 
545     /* Keep track of segment usage */
546     segment_counts[xd->mode_info_context->mbmi.segment_id]++;
547 
548     /* skip to next mb */
549     xd->mode_info_context++;
550     x->partition_info++;
551     xd->above_context++;
552   }
553 
554   /* extend the recon for intra prediction */
555   vp8_extend_mb_row(&cm->yv12_fb[dst_fb_idx], xd->dst.y_buffer + 16,
556                     xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
557 
558 #if CONFIG_MULTITHREAD
559   if (vpx_atomic_load_acquire(&cpi->b_multi_threaded) != 0) {
560     vpx_atomic_store_release(current_mb_col,
561                              vpx_atomic_load_acquire(&rightmost_col));
562   }
563 #endif
564 
565   /* this is to account for the border */
566   xd->mode_info_context++;
567   x->partition_info++;
568 }
569 
init_encode_frame_mb_context(VP8_COMP * cpi)570 static void init_encode_frame_mb_context(VP8_COMP *cpi) {
571   MACROBLOCK *const x = &cpi->mb;
572   VP8_COMMON *const cm = &cpi->common;
573   MACROBLOCKD *const xd = &x->e_mbd;
574 
575   /* GF active flags data structure */
576   x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
577 
578   /* Activity map pointer */
579   x->mb_activity_ptr = cpi->mb_activity_map;
580 
581   x->act_zbin_adj = 0;
582 
583   x->partition_info = x->pi;
584 
585   xd->mode_info_context = cm->mi;
586   xd->mode_info_stride = cm->mode_info_stride;
587 
588   xd->frame_type = cm->frame_type;
589 
590   /* reset intra mode contexts */
591   if (cm->frame_type == KEY_FRAME) vp8_init_mbmode_probs(cm);
592 
593   /* Copy data over into macro block data structures. */
594   x->src = *cpi->Source;
595   xd->pre = cm->yv12_fb[cm->lst_fb_idx];
596   xd->dst = cm->yv12_fb[cm->new_fb_idx];
597 
598   /* set up frame for intra coded blocks */
599   vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);
600 
601   vp8_build_block_offsets(x);
602 
603   xd->mode_info_context->mbmi.mode = DC_PRED;
604   xd->mode_info_context->mbmi.uv_mode = DC_PRED;
605 
606   xd->left_context = &cm->left_context;
607 
608   x->mvc = cm->fc.mvc;
609 
610   memset(cm->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);
611 
612   /* Special case treatment when GF and ARF are not sensible options
613    * for reference
614    */
615   if (cpi->ref_frame_flags == VP8_LAST_FRAME) {
616     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded, 255,
617                              128);
618   } else if ((cpi->oxcf.number_of_layers > 1) &&
619              (cpi->ref_frame_flags == VP8_GOLD_FRAME)) {
620     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded, 1, 255);
621   } else if ((cpi->oxcf.number_of_layers > 1) &&
622              (cpi->ref_frame_flags == VP8_ALTR_FRAME)) {
623     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded, 1, 1);
624   } else {
625     vp8_calc_ref_frame_costs(x->ref_frame_cost, cpi->prob_intra_coded,
626                              cpi->prob_last_coded, cpi->prob_gf_coded);
627   }
628 
629   xd->fullpixel_mask = ~0;
630   if (cm->full_pixel) xd->fullpixel_mask = ~7;
631 
632   vp8_zero(x->coef_counts);
633   vp8_zero(x->ymode_count);
634   vp8_zero(x->uv_mode_count);
635   x->prediction_error = 0;
636   x->intra_error = 0;
637   vp8_zero(x->count_mb_ref_frame_usage);
638 }
639 
640 #if CONFIG_MULTITHREAD
sum_coef_counts(MACROBLOCK * x,MACROBLOCK * x_thread)641 static void sum_coef_counts(MACROBLOCK *x, MACROBLOCK *x_thread) {
642   int i = 0;
643   do {
644     int j = 0;
645     do {
646       int k = 0;
647       do {
648         /* at every context */
649 
650         /* calc probs and branch cts for this frame only */
651         int t = 0; /* token/prob index */
652 
653         do {
654           x->coef_counts[i][j][k][t] += x_thread->coef_counts[i][j][k][t];
655         } while (++t < ENTROPY_NODES);
656       } while (++k < PREV_COEF_CONTEXTS);
657     } while (++j < COEF_BANDS);
658   } while (++i < BLOCK_TYPES);
659 }
660 #endif  // CONFIG_MULTITHREAD
661 
vp8_encode_frame(VP8_COMP * cpi)662 void vp8_encode_frame(VP8_COMP *cpi) {
663   int mb_row;
664   MACROBLOCK *const x = &cpi->mb;
665   VP8_COMMON *const cm = &cpi->common;
666   MACROBLOCKD *const xd = &x->e_mbd;
667   TOKENEXTRA *tp = cpi->tok;
668   int segment_counts[MAX_MB_SEGMENTS];
669   int totalrate;
670 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
671   BOOL_CODER *bc = &cpi->bc[1]; /* bc[0] is for control partition */
672   const int num_part = (1 << cm->multi_token_partition);
673 #endif
674 
675   memset(segment_counts, 0, sizeof(segment_counts));
676   totalrate = 0;
677 
678   if (cpi->compressor_speed == 2) {
679     if (cpi->oxcf.cpu_used < 0) {
680       cpi->Speed = -(cpi->oxcf.cpu_used);
681     } else {
682       vp8_auto_select_speed(cpi);
683     }
684   }
685 
686   /* Functions setup for all frame types so we can use MC in AltRef */
687   if (!cm->use_bilinear_mc_filter) {
688     xd->subpixel_predict = vp8_sixtap_predict4x4;
689     xd->subpixel_predict8x4 = vp8_sixtap_predict8x4;
690     xd->subpixel_predict8x8 = vp8_sixtap_predict8x8;
691     xd->subpixel_predict16x16 = vp8_sixtap_predict16x16;
692   } else {
693     xd->subpixel_predict = vp8_bilinear_predict4x4;
694     xd->subpixel_predict8x4 = vp8_bilinear_predict8x4;
695     xd->subpixel_predict8x8 = vp8_bilinear_predict8x8;
696     xd->subpixel_predict16x16 = vp8_bilinear_predict16x16;
697   }
698 
699   cpi->mb.skip_true_count = 0;
700   cpi->tok_count = 0;
701 
702 #if 0
703     /* Experimental code */
704     cpi->frame_distortion = 0;
705     cpi->last_mb_distortion = 0;
706 #endif
707 
708   xd->mode_info_context = cm->mi;
709 
710   vp8_zero(cpi->mb.MVcount);
711 
712   vp8cx_frame_init_quantizer(cpi);
713 
714   vp8_initialize_rd_consts(cpi, x,
715                            vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
716 
717   vp8cx_initialize_me_consts(cpi, cm->base_qindex);
718 
719   if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
720     /* Initialize encode frame context. */
721     init_encode_frame_mb_context(cpi);
722 
723     /* Build a frame level activity map */
724     build_activity_map(cpi);
725   }
726 
727   /* re-init encode frame context. */
728   init_encode_frame_mb_context(cpi);
729 
730 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
731   {
732     int i;
733     for (i = 0; i < num_part; ++i) {
734       vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
735                        cpi->partition_d_end[i + 1]);
736       bc[i].error = &cm->error;
737     }
738   }
739 
740 #endif
741 
742   {
743     struct vpx_usec_timer emr_timer;
744     vpx_usec_timer_start(&emr_timer);
745 
746 #if CONFIG_MULTITHREAD
747     if (vpx_atomic_load_acquire(&cpi->b_multi_threaded)) {
748       int i;
749 
750       vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei,
751                                 cpi->encoding_thread_count);
752 
753       for (i = 0; i < cm->mb_rows; ++i)
754         vpx_atomic_store_release(&cpi->mt_current_mb_col[i], -1);
755 
756       for (i = 0; i < cpi->encoding_thread_count; ++i) {
757         sem_post(&cpi->h_event_start_encoding[i]);
758       }
759 
760       for (mb_row = 0; mb_row < cm->mb_rows;
761            mb_row += (cpi->encoding_thread_count + 1)) {
762         vp8_zero(cm->left_context);
763 
764 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
765         tp = cpi->tok;
766 #else
767         tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
768 #endif
769 
770         encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
771 
772         /* adjust to the next row of mbs */
773         x->src.y_buffer +=
774             16 * x->src.y_stride * (cpi->encoding_thread_count + 1) -
775             16 * cm->mb_cols;
776         x->src.u_buffer +=
777             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
778             8 * cm->mb_cols;
779         x->src.v_buffer +=
780             8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) -
781             8 * cm->mb_cols;
782 
783         xd->mode_info_context +=
784             xd->mode_info_stride * cpi->encoding_thread_count;
785         x->partition_info += xd->mode_info_stride * cpi->encoding_thread_count;
786         x->gf_active_ptr += cm->mb_cols * cpi->encoding_thread_count;
787       }
788       /* Wait for all the threads to finish. */
789       for (i = 0; i < cpi->encoding_thread_count; ++i) {
790         sem_wait(&cpi->h_event_end_encoding[i]);
791       }
792 
793       for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
794         cpi->tok_count += (unsigned int)(cpi->tplist[mb_row].stop -
795                                          cpi->tplist[mb_row].start);
796       }
797 
798       if (xd->segmentation_enabled) {
799         int j;
800 
801         if (xd->segmentation_enabled) {
802           for (i = 0; i < cpi->encoding_thread_count; ++i) {
803             for (j = 0; j < 4; ++j) {
804               segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
805             }
806           }
807         }
808       }
809 
810       for (i = 0; i < cpi->encoding_thread_count; ++i) {
811         int mode_count;
812         int c_idx;
813         totalrate += cpi->mb_row_ei[i].totalrate;
814 
815         cpi->mb.skip_true_count += cpi->mb_row_ei[i].mb.skip_true_count;
816 
817         for (mode_count = 0; mode_count < VP8_YMODES; ++mode_count) {
818           cpi->mb.ymode_count[mode_count] +=
819               cpi->mb_row_ei[i].mb.ymode_count[mode_count];
820         }
821 
822         for (mode_count = 0; mode_count < VP8_UV_MODES; ++mode_count) {
823           cpi->mb.uv_mode_count[mode_count] +=
824               cpi->mb_row_ei[i].mb.uv_mode_count[mode_count];
825         }
826 
827         for (c_idx = 0; c_idx < MVvals; ++c_idx) {
828           cpi->mb.MVcount[0][c_idx] += cpi->mb_row_ei[i].mb.MVcount[0][c_idx];
829           cpi->mb.MVcount[1][c_idx] += cpi->mb_row_ei[i].mb.MVcount[1][c_idx];
830         }
831 
832         cpi->mb.prediction_error += cpi->mb_row_ei[i].mb.prediction_error;
833         cpi->mb.intra_error += cpi->mb_row_ei[i].mb.intra_error;
834 
835         for (c_idx = 0; c_idx < MAX_REF_FRAMES; ++c_idx) {
836           cpi->mb.count_mb_ref_frame_usage[c_idx] +=
837               cpi->mb_row_ei[i].mb.count_mb_ref_frame_usage[c_idx];
838         }
839 
840         for (c_idx = 0; c_idx < MAX_ERROR_BINS; ++c_idx) {
841           cpi->mb.error_bins[c_idx] += cpi->mb_row_ei[i].mb.error_bins[c_idx];
842         }
843 
844         /* add up counts for each thread */
845         sum_coef_counts(x, &cpi->mb_row_ei[i].mb);
846       }
847 
848     } else
849 #endif  // CONFIG_MULTITHREAD
850     {
851 
852       /* for each macroblock row in image */
853       for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
854         vp8_zero(cm->left_context);
855 
856 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
857         tp = cpi->tok;
858 #endif
859 
860         encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
861 
862         /* adjust to the next row of mbs */
863         x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
864         x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
865         x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
866       }
867 
868       cpi->tok_count = (unsigned int)(tp - cpi->tok);
869     }
870 
871 #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
872     {
873       int i;
874       for (i = 0; i < num_part; ++i) {
875         vp8_stop_encode(&bc[i]);
876         cpi->partition_sz[i + 1] = bc[i].pos;
877       }
878     }
879 #endif
880 
881     vpx_usec_timer_mark(&emr_timer);
882     cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
883   }
884 
885   // Work out the segment probabilities if segmentation is enabled
886   // and needs to be updated
887   if (xd->segmentation_enabled && xd->update_mb_segmentation_map) {
888     int tot_count;
889     int i;
890 
891     /* Set to defaults */
892     memset(xd->mb_segment_tree_probs, 255, sizeof(xd->mb_segment_tree_probs));
893 
894     tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] +
895                 segment_counts[3];
896 
897     if (tot_count) {
898       xd->mb_segment_tree_probs[0] =
899           ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
900 
901       tot_count = segment_counts[0] + segment_counts[1];
902 
903       if (tot_count > 0) {
904         xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
905       }
906 
907       tot_count = segment_counts[2] + segment_counts[3];
908 
909       if (tot_count > 0) {
910         xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
911       }
912 
913       /* Zero probabilities not allowed */
914       for (i = 0; i < MB_FEATURE_TREE_PROBS; ++i) {
915         if (xd->mb_segment_tree_probs[i] == 0) xd->mb_segment_tree_probs[i] = 1;
916       }
917     }
918   }
919 
920   /* projected_frame_size in units of BYTES */
921   cpi->projected_frame_size = totalrate >> 8;
922 
923   /* Make a note of the percentage MBs coded Intra. */
924   if (cm->frame_type == KEY_FRAME) {
925     cpi->this_frame_percent_intra = 100;
926   } else {
927     int tot_modes;
928 
929     tot_modes = cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] +
930                 cpi->mb.count_mb_ref_frame_usage[LAST_FRAME] +
931                 cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME] +
932                 cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
933 
934     if (tot_modes) {
935       cpi->this_frame_percent_intra =
936           cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
937     }
938   }
939 
940 #if !CONFIG_REALTIME_ONLY
941   /* Adjust the projected reference frame usage probability numbers to
942    * reflect what we have just seen. This may be useful when we make
943    * multiple iterations of the recode loop rather than continuing to use
944    * values from the previous frame.
945    */
946   if ((cm->frame_type != KEY_FRAME) &&
947       ((cpi->oxcf.number_of_layers > 1) ||
948        (!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame))) {
949     vp8_convert_rfct_to_prob(cpi);
950   }
951 #endif
952 }
vp8_setup_block_ptrs(MACROBLOCK * x)953 void vp8_setup_block_ptrs(MACROBLOCK *x) {
954   int r, c;
955   int i;
956 
957   for (r = 0; r < 4; ++r) {
958     for (c = 0; c < 4; ++c) {
959       x->block[r * 4 + c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
960     }
961   }
962 
963   for (r = 0; r < 2; ++r) {
964     for (c = 0; c < 2; ++c) {
965       x->block[16 + r * 2 + c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
966     }
967   }
968 
969   for (r = 0; r < 2; ++r) {
970     for (c = 0; c < 2; ++c) {
971       x->block[20 + r * 2 + c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
972     }
973   }
974 
975   x->block[24].src_diff = x->src_diff + 384;
976 
977   for (i = 0; i < 25; ++i) {
978     x->block[i].coeff = x->coeff + i * 16;
979   }
980 }
981 
vp8_build_block_offsets(MACROBLOCK * x)982 void vp8_build_block_offsets(MACROBLOCK *x) {
983   int block = 0;
984   int br, bc;
985 
986   vp8_build_block_doffsets(&x->e_mbd);
987 
988   /* y blocks */
989   x->thismb_ptr = &x->thismb[0];
990   for (br = 0; br < 4; ++br) {
991     for (bc = 0; bc < 4; ++bc) {
992       BLOCK *this_block = &x->block[block];
993       this_block->base_src = &x->thismb_ptr;
994       this_block->src_stride = 16;
995       this_block->src = 4 * br * 16 + 4 * bc;
996       ++block;
997     }
998   }
999 
1000   /* u blocks */
1001   for (br = 0; br < 2; ++br) {
1002     for (bc = 0; bc < 2; ++bc) {
1003       BLOCK *this_block = &x->block[block];
1004       this_block->base_src = &x->src.u_buffer;
1005       this_block->src_stride = x->src.uv_stride;
1006       this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1007       ++block;
1008     }
1009   }
1010 
1011   /* v blocks */
1012   for (br = 0; br < 2; ++br) {
1013     for (bc = 0; bc < 2; ++bc) {
1014       BLOCK *this_block = &x->block[block];
1015       this_block->base_src = &x->src.v_buffer;
1016       this_block->src_stride = x->src.uv_stride;
1017       this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1018       ++block;
1019     }
1020   }
1021 }
1022 
sum_intra_stats(VP8_COMP * cpi,MACROBLOCK * x)1023 static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x) {
1024   const MACROBLOCKD *xd = &x->e_mbd;
1025   const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
1026   const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;
1027 
1028 #ifdef MODE_STATS
1029   const int is_key = cpi->common.frame_type == KEY_FRAME;
1030 
1031   ++(is_key ? uv_modes : inter_uv_modes)[uvm];
1032 
1033   if (m == B_PRED) {
1034     unsigned int *const bct = is_key ? b_modes : inter_b_modes;
1035 
1036     int b = 0;
1037 
1038     do {
1039       ++bct[xd->block[b].bmi.mode];
1040     } while (++b < 16);
1041   }
1042 
1043 #else
1044   (void)cpi;
1045 #endif
1046 
1047   ++x->ymode_count[m];
1048   ++x->uv_mode_count[uvm];
1049 }
1050 
1051 /* Experimental stub function to create a per MB zbin adjustment based on
1052  * some previously calculated measure of MB activity.
1053  */
adjust_act_zbin(VP8_COMP * cpi,MACROBLOCK * x)1054 static void adjust_act_zbin(VP8_COMP *cpi, MACROBLOCK *x) {
1055 #if USE_ACT_INDEX
1056   x->act_zbin_adj = *(x->mb_activity_ptr);
1057 #else
1058   int64_t a;
1059   int64_t b;
1060   int64_t act = *(x->mb_activity_ptr);
1061 
1062   /* Apply the masking to the RD multiplier. */
1063   a = act + 4 * cpi->activity_avg;
1064   b = 4 * act + cpi->activity_avg;
1065 
1066   if (act > cpi->activity_avg) {
1067     x->act_zbin_adj = (int)(((int64_t)b + (a >> 1)) / a) - 1;
1068   } else {
1069     x->act_zbin_adj = 1 - (int)(((int64_t)a + (b >> 1)) / b);
1070   }
1071 #endif
1072 }
1073 
vp8cx_encode_intra_macroblock(VP8_COMP * cpi,MACROBLOCK * x,TOKENEXTRA ** t)1074 int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
1075                                   TOKENEXTRA **t) {
1076   MACROBLOCKD *xd = &x->e_mbd;
1077   int rate;
1078 
1079   if (cpi->sf.RD && cpi->compressor_speed != 2) {
1080     vp8_rd_pick_intra_mode(x, &rate);
1081   } else {
1082     vp8_pick_intra_mode(x, &rate);
1083   }
1084 
1085   if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
1086     adjust_act_zbin(cpi, x);
1087     vp8_update_zbin_extra(cpi, x);
1088   }
1089 
1090   if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED) {
1091     vp8_encode_intra4x4mby(x);
1092   } else {
1093     vp8_encode_intra16x16mby(x);
1094   }
1095 
1096   vp8_encode_intra16x16mbuv(x);
1097 
1098   sum_intra_stats(cpi, x);
1099 
1100   vp8_tokenize_mb(cpi, x, t);
1101 
1102   if (xd->mode_info_context->mbmi.mode != B_PRED) vp8_inverse_transform_mby(xd);
1103 
1104   vp8_dequant_idct_add_uv_block(xd->qcoeff + 16 * 16, xd->dequant_uv,
1105                                 xd->dst.u_buffer, xd->dst.v_buffer,
1106                                 xd->dst.uv_stride, xd->eobs + 16);
1107   return rate;
1108 }
1109 #ifdef SPEEDSTATS
1110 extern int cnt_pm;
1111 #endif
1112 
1113 extern void vp8_fix_contexts(MACROBLOCKD *x);
1114 
vp8cx_encode_inter_macroblock(VP8_COMP * cpi,MACROBLOCK * x,TOKENEXTRA ** t,int recon_yoffset,int recon_uvoffset,int mb_row,int mb_col)1115 int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
1116                                   int recon_yoffset, int recon_uvoffset,
1117                                   int mb_row, int mb_col) {
1118   MACROBLOCKD *const xd = &x->e_mbd;
1119   int intra_error = 0;
1120   int rate;
1121   int distortion;
1122 
1123   x->skip = 0;
1124 
1125   if (xd->segmentation_enabled) {
1126     x->encode_breakout =
1127         cpi->segment_encode_breakout[xd->mode_info_context->mbmi.segment_id];
1128   } else {
1129     x->encode_breakout = cpi->oxcf.encode_breakout;
1130   }
1131 
1132 #if CONFIG_TEMPORAL_DENOISING
1133   /* Reset the best sse mode/mv for each macroblock. */
1134   x->best_reference_frame = INTRA_FRAME;
1135   x->best_zeromv_reference_frame = INTRA_FRAME;
1136   x->best_sse_inter_mode = 0;
1137   x->best_sse_mv.as_int = 0;
1138   x->need_to_clamp_best_mvs = 0;
1139 #endif
1140 
1141   if (cpi->sf.RD) {
1142     int zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
1143 
1144     /* Are we using the fast quantizer for the mode selection? */
1145     if (cpi->sf.use_fastquant_for_pick) {
1146       x->quantize_b = vp8_fast_quantize_b;
1147 
1148       /* the fast quantizer does not use zbin_extra, so
1149        * do not recalculate */
1150       x->zbin_mode_boost_enabled = 0;
1151     }
1152     vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1153                            &distortion, &intra_error, mb_row, mb_col);
1154 
1155     /* switch back to the regular quantizer for the encode */
1156     if (cpi->sf.improved_quant) {
1157       x->quantize_b = vp8_regular_quantize_b;
1158     }
1159 
1160     /* restore cpi->zbin_mode_boost_enabled */
1161     x->zbin_mode_boost_enabled = zbin_mode_boost_enabled;
1162 
1163   } else {
1164     vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1165                         &distortion, &intra_error, mb_row, mb_col);
1166   }
1167 
1168   x->prediction_error += distortion;
1169   x->intra_error += intra_error;
1170 
1171   if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
1172     /* Adjust the zbin based on this MB rate. */
1173     adjust_act_zbin(cpi, x);
1174   }
1175 
1176 #if 0
1177     /* Experimental RD code */
1178     cpi->frame_distortion += distortion;
1179     cpi->last_mb_distortion = distortion;
1180 #endif
1181 
1182   /* MB level adjutment to quantizer setup */
1183   if (xd->segmentation_enabled) {
1184     /* If cyclic update enabled */
1185     if (cpi->current_layer == 0 && cpi->cyclic_refresh_mode_enabled) {
1186       /* Clear segment_id back to 0 if not coded (last frame 0,0) */
1187       if ((xd->mode_info_context->mbmi.segment_id == 1) &&
1188           ((xd->mode_info_context->mbmi.ref_frame != LAST_FRAME) ||
1189            (xd->mode_info_context->mbmi.mode != ZEROMV))) {
1190         xd->mode_info_context->mbmi.segment_id = 0;
1191 
1192         /* segment_id changed, so update */
1193         vp8cx_mb_init_quantizer(cpi, x, 1);
1194       }
1195     }
1196   }
1197 
1198   {
1199     /* Experimental code.
1200      * Special case for gf and arf zeromv modes, for 1 temporal layer.
1201      * Increase zbin size to supress noise.
1202      */
1203     x->zbin_mode_boost = 0;
1204     if (x->zbin_mode_boost_enabled) {
1205       if (xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME) {
1206         if (xd->mode_info_context->mbmi.mode == ZEROMV) {
1207           if (xd->mode_info_context->mbmi.ref_frame != LAST_FRAME &&
1208               cpi->oxcf.number_of_layers == 1) {
1209             x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
1210           } else {
1211             x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
1212           }
1213         } else if (xd->mode_info_context->mbmi.mode == SPLITMV) {
1214           x->zbin_mode_boost = 0;
1215         } else {
1216           x->zbin_mode_boost = MV_ZBIN_BOOST;
1217         }
1218       }
1219     }
1220 
1221     /* The fast quantizer doesn't use zbin_extra, only do so with
1222      * the regular quantizer. */
1223     if (cpi->sf.improved_quant) vp8_update_zbin_extra(cpi, x);
1224   }
1225 
1226   x->count_mb_ref_frame_usage[xd->mode_info_context->mbmi.ref_frame]++;
1227 
1228   if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) {
1229     vp8_encode_intra16x16mbuv(x);
1230 
1231     if (xd->mode_info_context->mbmi.mode == B_PRED) {
1232       vp8_encode_intra4x4mby(x);
1233     } else {
1234       vp8_encode_intra16x16mby(x);
1235     }
1236 
1237     sum_intra_stats(cpi, x);
1238   } else {
1239     int ref_fb_idx;
1240 
1241     if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
1242       ref_fb_idx = cpi->common.lst_fb_idx;
1243     } else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME) {
1244       ref_fb_idx = cpi->common.gld_fb_idx;
1245     } else {
1246       ref_fb_idx = cpi->common.alt_fb_idx;
1247     }
1248 
1249     xd->pre.y_buffer = cpi->common.yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
1250     xd->pre.u_buffer =
1251         cpi->common.yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
1252     xd->pre.v_buffer =
1253         cpi->common.yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
1254 
1255     if (!x->skip) {
1256       vp8_encode_inter16x16(x);
1257     } else {
1258       vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer, xd->dst.u_buffer,
1259                                          xd->dst.v_buffer, xd->dst.y_stride,
1260                                          xd->dst.uv_stride);
1261     }
1262   }
1263 
1264   if (!x->skip) {
1265     vp8_tokenize_mb(cpi, x, t);
1266 
1267     if (xd->mode_info_context->mbmi.mode != B_PRED) {
1268       vp8_inverse_transform_mby(xd);
1269     }
1270 
1271     vp8_dequant_idct_add_uv_block(xd->qcoeff + 16 * 16, xd->dequant_uv,
1272                                   xd->dst.u_buffer, xd->dst.v_buffer,
1273                                   xd->dst.uv_stride, xd->eobs + 16);
1274   } else {
1275     /* always set mb_skip_coeff as it is needed by the loopfilter */
1276     xd->mode_info_context->mbmi.mb_skip_coeff = 1;
1277 
1278     if (cpi->common.mb_no_coeff_skip) {
1279       x->skip_true_count++;
1280       vp8_fix_contexts(xd);
1281     } else {
1282       vp8_stuff_mb(cpi, x, t);
1283     }
1284   }
1285 
1286   return rate;
1287 }
1288