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 "./vpx_dsp_rtcd.h"
15#include "encodemb.h"
16#include "encodemv.h"
17#include "vp8/common/common.h"
18#include "onyx_int.h"
19#include "vp8/common/extend.h"
20#include "vp8/common/entropymode.h"
21#include "vp8/common/quant_common.h"
22#include "segmentation.h"
23#include "vp8/common/setupintrarecon.h"
24#include "encodeintra.h"
25#include "vp8/common/reconinter.h"
26#include "rdopt.h"
27#include "pickinter.h"
28#include "vp8/common/findnearmv.h"
29#include <stdio.h>
30#include <limits.h>
31#include "vp8/common/invtrans.h"
32#include "vpx_ports/vpx_timer.h"
33#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
34#include "bitstream.h"
35#endif
36#include "encodeframe.h"
37
38extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t) ;
39extern void vp8_calc_ref_frame_costs(int *ref_frame_cost,
40                                     int prob_intra,
41                                     int prob_last,
42                                     int prob_garf
43                                    );
44extern void vp8_convert_rfct_to_prob(VP8_COMP *const cpi);
45extern void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex);
46extern void vp8_auto_select_speed(VP8_COMP *cpi);
47extern void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
48                                      MACROBLOCK *x,
49                                      MB_ROW_COMP *mbr_ei,
50                                      int count);
51static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x );
52
53#ifdef MODE_STATS
54unsigned int inter_y_modes[10] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
55unsigned int inter_uv_modes[4] = {0, 0, 0, 0};
56unsigned int inter_b_modes[15]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
57unsigned int y_modes[5]   = {0, 0, 0, 0, 0};
58unsigned int uv_modes[4]  = {0, 0, 0, 0};
59unsigned int b_modes[14]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
60#endif
61
62
63/* activity_avg must be positive, or flat regions could get a zero weight
64 *  (infinite lambda), which confounds analysis.
65 * This also avoids the need for divide by zero checks in
66 *  vp8_activity_masking().
67 */
68#define VP8_ACTIVITY_AVG_MIN (64)
69
70/* This is used as a reference when computing the source variance for the
71 *  purposes of activity masking.
72 * Eventually this should be replaced by custom no-reference routines,
73 *  which will be faster.
74 */
75static const unsigned char VP8_VAR_OFFS[16]=
76{
77    128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128
78};
79
80
81/* Original activity measure from Tim T's code. */
82static unsigned int tt_activity_measure( VP8_COMP *cpi, MACROBLOCK *x )
83{
84    unsigned int act;
85    unsigned int sse;
86    (void)cpi;
87    /* TODO: This could also be done over smaller areas (8x8), but that would
88     *  require extensive changes elsewhere, as lambda is assumed to be fixed
89     *  over an entire MB in most of the code.
90     * Another option is to compute four 8x8 variances, and pick a single
91     *  lambda using a non-linear combination (e.g., the smallest, or second
92     *  smallest, etc.).
93     */
94    act =  vpx_variance16x16(x->src.y_buffer,
95                    x->src.y_stride, VP8_VAR_OFFS, 0, &sse);
96    act = act<<4;
97
98    /* If the region is flat, lower the activity some more. */
99    if (act < 8<<12)
100        act = act < 5<<12 ? act : 5<<12;
101
102    return act;
103}
104
105/* Stub for alternative experimental activity measures. */
106static unsigned int alt_activity_measure( VP8_COMP *cpi,
107                                          MACROBLOCK *x, int use_dc_pred )
108{
109    return vp8_encode_intra(cpi,x, use_dc_pred);
110}
111
112
113/* Measure the activity of the current macroblock
114 * What we measure here is TBD so abstracted to this function
115 */
116#define ALT_ACT_MEASURE 1
117static unsigned int mb_activity_measure( VP8_COMP *cpi, MACROBLOCK *x,
118                                  int mb_row, int mb_col)
119{
120    unsigned int mb_activity;
121
122    if  ( ALT_ACT_MEASURE )
123    {
124        int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
125
126        /* Or use and alternative. */
127        mb_activity = alt_activity_measure( cpi, x, use_dc_pred );
128    }
129    else
130    {
131        /* Original activity measure from Tim T's code. */
132        mb_activity = tt_activity_measure( cpi, x );
133    }
134
135    if ( mb_activity < VP8_ACTIVITY_AVG_MIN )
136        mb_activity = VP8_ACTIVITY_AVG_MIN;
137
138    return mb_activity;
139}
140
141/* Calculate an "average" mb activity value for the frame */
142#define ACT_MEDIAN 0
143static void calc_av_activity( VP8_COMP *cpi, int64_t activity_sum )
144{
145#if ACT_MEDIAN
146    /* Find median: Simple n^2 algorithm for experimentation */
147    {
148        unsigned int median;
149        unsigned int i,j;
150        unsigned int * sortlist;
151        unsigned int tmp;
152
153        /* Create a list to sort to */
154        CHECK_MEM_ERROR(sortlist,
155                        vpx_calloc(sizeof(unsigned int),
156                        cpi->common.MBs));
157
158        /* Copy map to sort list */
159        memcpy( sortlist, cpi->mb_activity_map,
160                sizeof(unsigned int) * cpi->common.MBs );
161
162
163        /* Ripple each value down to its correct position */
164        for ( i = 1; i < cpi->common.MBs; i ++ )
165        {
166            for ( j = i; j > 0; j -- )
167            {
168                if ( sortlist[j] < sortlist[j-1] )
169                {
170                    /* Swap values */
171                    tmp = sortlist[j-1];
172                    sortlist[j-1] = sortlist[j];
173                    sortlist[j] = tmp;
174                }
175                else
176                    break;
177            }
178        }
179
180        /* Even number MBs so estimate median as mean of two either side. */
181        median = ( 1 + sortlist[cpi->common.MBs >> 1] +
182                   sortlist[(cpi->common.MBs >> 1) + 1] ) >> 1;
183
184        cpi->activity_avg = median;
185
186        vpx_free(sortlist);
187    }
188#else
189    /* Simple mean for now */
190    cpi->activity_avg = (unsigned int)(activity_sum/cpi->common.MBs);
191#endif
192
193    if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN)
194        cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;
195
196    /* Experimental code: return fixed value normalized for several clips */
197    if  ( ALT_ACT_MEASURE )
198        cpi->activity_avg = 100000;
199}
200
201#define USE_ACT_INDEX   0
202#define OUTPUT_NORM_ACT_STATS   0
203
204#if USE_ACT_INDEX
205/* Calculate and activity index for each mb */
206static void calc_activity_index( VP8_COMP *cpi, MACROBLOCK *x )
207{
208    VP8_COMMON *const cm = & cpi->common;
209    int mb_row, mb_col;
210
211    int64_t act;
212    int64_t a;
213    int64_t b;
214
215#if OUTPUT_NORM_ACT_STATS
216    FILE *f = fopen("norm_act.stt", "a");
217    fprintf(f, "\n%12d\n", cpi->activity_avg );
218#endif
219
220    /* Reset pointers to start of activity map */
221    x->mb_activity_ptr = cpi->mb_activity_map;
222
223    /* Calculate normalized mb activity number. */
224    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
225    {
226        /* for each macroblock col in image */
227        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
228        {
229            /* Read activity from the map */
230            act = *(x->mb_activity_ptr);
231
232            /* Calculate a normalized activity number */
233            a = act + 4*cpi->activity_avg;
234            b = 4*act + cpi->activity_avg;
235
236            if ( b >= a )
237                *(x->activity_ptr) = (int)((b + (a>>1))/a) - 1;
238            else
239                *(x->activity_ptr) = 1 - (int)((a + (b>>1))/b);
240
241#if OUTPUT_NORM_ACT_STATS
242            fprintf(f, " %6d", *(x->mb_activity_ptr));
243#endif
244            /* Increment activity map pointers */
245            x->mb_activity_ptr++;
246        }
247
248#if OUTPUT_NORM_ACT_STATS
249        fprintf(f, "\n");
250#endif
251
252    }
253
254#if OUTPUT_NORM_ACT_STATS
255    fclose(f);
256#endif
257
258}
259#endif
260
261/* Loop through all MBs. Note activity of each, average activity and
262 * calculate a normalized activity for each
263 */
264static void build_activity_map( VP8_COMP *cpi )
265{
266    MACROBLOCK *const x = & cpi->mb;
267    MACROBLOCKD *xd = &x->e_mbd;
268    VP8_COMMON *const cm = & cpi->common;
269
270#if ALT_ACT_MEASURE
271    YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
272    int recon_yoffset;
273    int recon_y_stride = new_yv12->y_stride;
274#endif
275
276    int mb_row, mb_col;
277    unsigned int mb_activity;
278    int64_t activity_sum = 0;
279
280    /* for each macroblock row in image */
281    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
282    {
283#if ALT_ACT_MEASURE
284        /* reset above block coeffs */
285        xd->up_available = (mb_row != 0);
286        recon_yoffset = (mb_row * recon_y_stride * 16);
287#endif
288        /* for each macroblock col in image */
289        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
290        {
291#if ALT_ACT_MEASURE
292            xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
293            xd->left_available = (mb_col != 0);
294            recon_yoffset += 16;
295#endif
296            /* Copy current mb to a buffer */
297            vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
298
299            /* measure activity */
300            mb_activity = mb_activity_measure( cpi, x, mb_row, mb_col );
301
302            /* Keep frame sum */
303            activity_sum += mb_activity;
304
305            /* Store MB level activity details. */
306            *x->mb_activity_ptr = mb_activity;
307
308            /* Increment activity map pointer */
309            x->mb_activity_ptr++;
310
311            /* adjust to the next column of source macroblocks */
312            x->src.y_buffer += 16;
313        }
314
315
316        /* adjust to the next row of mbs */
317        x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
318
319#if ALT_ACT_MEASURE
320        /* extend the recon for intra prediction */
321        vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16,
322                          xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
323#endif
324
325    }
326
327    /* Calculate an "average" MB activity */
328    calc_av_activity(cpi, activity_sum);
329
330#if USE_ACT_INDEX
331    /* Calculate an activity index number of each mb */
332    calc_activity_index( cpi, x );
333#endif
334
335}
336
337/* Macroblock activity masking */
338void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x)
339{
340#if USE_ACT_INDEX
341    x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
342    x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
343    x->errorperbit += (x->errorperbit==0);
344#else
345    int64_t a;
346    int64_t b;
347    int64_t act = *(x->mb_activity_ptr);
348
349    /* Apply the masking to the RD multiplier. */
350    a = act + (2*cpi->activity_avg);
351    b = (2*act) + cpi->activity_avg;
352
353    x->rdmult = (unsigned int)(((int64_t)x->rdmult*b + (a>>1))/a);
354    x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
355    x->errorperbit += (x->errorperbit==0);
356#endif
357
358    /* Activity based Zbin adjustment */
359    adjust_act_zbin(cpi, x);
360}
361
362static
363void encode_mb_row(VP8_COMP *cpi,
364                   VP8_COMMON *cm,
365                   int mb_row,
366                   MACROBLOCK  *x,
367                   MACROBLOCKD *xd,
368                   TOKENEXTRA **tp,
369                   int *segment_counts,
370                   int *totalrate)
371{
372    int recon_yoffset, recon_uvoffset;
373    int mb_col;
374    int ref_fb_idx = cm->lst_fb_idx;
375    int dst_fb_idx = cm->new_fb_idx;
376    int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
377    int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
378    int map_index = (mb_row * cpi->common.mb_cols);
379
380#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
381    const int num_part = (1 << cm->multi_token_partition);
382    TOKENEXTRA * tp_start = cpi->tok;
383    vp8_writer *w;
384#endif
385
386#if CONFIG_MULTITHREAD
387    const int nsync = cpi->mt_sync_range;
388    const int rightmost_col = cm->mb_cols + nsync;
389    volatile const int *last_row_current_mb_col;
390    volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
391
392    if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
393        last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
394    else
395        last_row_current_mb_col = &rightmost_col;
396#endif
397
398#if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
399    if(num_part > 1)
400        w= &cpi->bc[1 + (mb_row % num_part)];
401    else
402        w = &cpi->bc[1];
403#endif
404
405    /* reset above block coeffs */
406    xd->above_context = cm->above_context;
407
408    xd->up_available = (mb_row != 0);
409    recon_yoffset = (mb_row * recon_y_stride * 16);
410    recon_uvoffset = (mb_row * recon_uv_stride * 8);
411
412    cpi->tplist[mb_row].start = *tp;
413    /* printf("Main mb_row = %d\n", mb_row); */
414
415    /* Distance of Mb to the top & bottom edges, specified in 1/8th pel
416     * units as they are always compared to values that are in 1/8th pel
417     */
418    xd->mb_to_top_edge = -((mb_row * 16) << 3);
419    xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
420
421    /* Set up limit values for vertical motion vector components
422     * to prevent them extending beyond the UMV borders
423     */
424    x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
425    x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
426                        + (VP8BORDERINPIXELS - 16);
427
428    /* Set the mb activity pointer to the start of the row. */
429    x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
430
431    /* for each macroblock col in image */
432    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
433    {
434
435#if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
436        *tp = cpi->tok;
437#endif
438        /* Distance of Mb to the left & right edges, specified in
439         * 1/8th pel units as they are always compared to values
440         * that are in 1/8th pel units
441         */
442        xd->mb_to_left_edge = -((mb_col * 16) << 3);
443        xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
444
445        /* Set up limit values for horizontal motion vector components
446         * to prevent them extending beyond the UMV borders
447         */
448        x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
449        x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16)
450                            + (VP8BORDERINPIXELS - 16);
451
452        xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
453        xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
454        xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
455        xd->left_available = (mb_col != 0);
456
457        x->rddiv = cpi->RDDIV;
458        x->rdmult = cpi->RDMULT;
459
460        /* Copy current mb to a buffer */
461        vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
462
463#if CONFIG_MULTITHREAD
464        if (cpi->b_multi_threaded != 0)
465        {
466            *current_mb_col = mb_col - 1; /* set previous MB done */
467
468            if ((mb_col & (nsync - 1)) == 0)
469            {
470                while (mb_col > (*last_row_current_mb_col - nsync))
471                {
472                    x86_pause_hint();
473                    thread_sleep(0);
474                }
475            }
476        }
477#endif
478
479        if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
480            vp8_activity_masking(cpi, x);
481
482        /* Is segmentation enabled */
483        /* MB level adjustment to quantizer */
484        if (xd->segmentation_enabled)
485        {
486            /* Code to set segment id in xd->mbmi.segment_id for current MB
487             * (with range checking)
488             */
489            if (cpi->segmentation_map[map_index+mb_col] <= 3)
490                xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index+mb_col];
491            else
492                xd->mode_info_context->mbmi.segment_id = 0;
493
494            vp8cx_mb_init_quantizer(cpi, x, 1);
495        }
496        else
497            /* Set to Segment 0 by default */
498            xd->mode_info_context->mbmi.segment_id = 0;
499
500        x->active_ptr = cpi->active_map + map_index + mb_col;
501
502        if (cm->frame_type == KEY_FRAME)
503        {
504            *totalrate += vp8cx_encode_intra_macroblock(cpi, x, tp);
505#ifdef MODE_STATS
506            y_modes[xd->mbmi.mode] ++;
507#endif
508        }
509        else
510        {
511            *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
512
513#ifdef MODE_STATS
514            inter_y_modes[xd->mbmi.mode] ++;
515
516            if (xd->mbmi.mode == SPLITMV)
517            {
518                int b;
519
520                for (b = 0; b < xd->mbmi.partition_count; b++)
521                {
522                    inter_b_modes[x->partition->bmi[b].mode] ++;
523                }
524            }
525
526#endif
527
528            // Keep track of how many (consecutive) times a  block is coded
529            // as ZEROMV_LASTREF, for base layer frames.
530            // Reset to 0 if its coded as anything else.
531            if (cpi->current_layer == 0) {
532              if (xd->mode_info_context->mbmi.mode == ZEROMV &&
533                  xd->mode_info_context->mbmi.ref_frame == LAST_FRAME) {
534                // Increment, check for wrap-around.
535                if (cpi->consec_zero_last[map_index+mb_col] < 255)
536                  cpi->consec_zero_last[map_index+mb_col] += 1;
537                if (cpi->consec_zero_last_mvbias[map_index+mb_col] < 255)
538                  cpi->consec_zero_last_mvbias[map_index+mb_col] += 1;
539              } else {
540                cpi->consec_zero_last[map_index+mb_col] = 0;
541                cpi->consec_zero_last_mvbias[map_index+mb_col] = 0;
542              }
543              if (x->zero_last_dot_suppress)
544                cpi->consec_zero_last_mvbias[map_index+mb_col] = 0;
545            }
546
547            /* Special case code for cyclic refresh
548             * If cyclic update enabled then copy xd->mbmi.segment_id; (which
549             * may have been updated based on mode during
550             * vp8cx_encode_inter_macroblock()) back into the global
551             * segmentation map
552             */
553            if ((cpi->current_layer == 0) &&
554                (cpi->cyclic_refresh_mode_enabled &&
555                 xd->segmentation_enabled))
556            {
557                cpi->segmentation_map[map_index+mb_col] = xd->mode_info_context->mbmi.segment_id;
558
559                /* If the block has been refreshed mark it as clean (the
560                 * magnitude of the -ve influences how long it will be before
561                 * we consider another refresh):
562                 * Else if it was coded (last frame 0,0) and has not already
563                 * been refreshed then mark it as a candidate for cleanup
564                 * next time (marked 0) else mark it as dirty (1).
565                 */
566                if (xd->mode_info_context->mbmi.segment_id)
567                    cpi->cyclic_refresh_map[map_index+mb_col] = -1;
568                else if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
569                {
570                    if (cpi->cyclic_refresh_map[map_index+mb_col] == 1)
571                        cpi->cyclic_refresh_map[map_index+mb_col] = 0;
572                }
573                else
574                    cpi->cyclic_refresh_map[map_index+mb_col] = 1;
575
576            }
577        }
578
579        cpi->tplist[mb_row].stop = *tp;
580
581#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
582        /* pack tokens for this MB */
583        {
584            int tok_count = *tp - tp_start;
585            vp8_pack_tokens(w, tp_start, tok_count);
586        }
587#endif
588        /* Increment pointer into gf usage flags structure. */
589        x->gf_active_ptr++;
590
591        /* Increment the activity mask pointers. */
592        x->mb_activity_ptr++;
593
594        /* adjust to the next column of macroblocks */
595        x->src.y_buffer += 16;
596        x->src.u_buffer += 8;
597        x->src.v_buffer += 8;
598
599        recon_yoffset += 16;
600        recon_uvoffset += 8;
601
602        /* Keep track of segment usage */
603        segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
604
605        /* skip to next mb */
606        xd->mode_info_context++;
607        x->partition_info++;
608        xd->above_context++;
609    }
610
611    /* extend the recon for intra prediction */
612    vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
613                        xd->dst.y_buffer + 16,
614                        xd->dst.u_buffer + 8,
615                        xd->dst.v_buffer + 8);
616
617#if CONFIG_MULTITHREAD
618    if (cpi->b_multi_threaded != 0)
619        *current_mb_col = rightmost_col;
620#endif
621
622    /* this is to account for the border */
623    xd->mode_info_context++;
624    x->partition_info++;
625}
626
627static void init_encode_frame_mb_context(VP8_COMP *cpi)
628{
629    MACROBLOCK *const x = & cpi->mb;
630    VP8_COMMON *const cm = & cpi->common;
631    MACROBLOCKD *const xd = & x->e_mbd;
632
633    /* GF active flags data structure */
634    x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
635
636    /* Activity map pointer */
637    x->mb_activity_ptr = cpi->mb_activity_map;
638
639    x->act_zbin_adj = 0;
640
641    x->partition_info = x->pi;
642
643    xd->mode_info_context = cm->mi;
644    xd->mode_info_stride = cm->mode_info_stride;
645
646    xd->frame_type = cm->frame_type;
647
648    /* reset intra mode contexts */
649    if (cm->frame_type == KEY_FRAME)
650        vp8_init_mbmode_probs(cm);
651
652    /* Copy data over into macro block data structures. */
653    x->src = * cpi->Source;
654    xd->pre = cm->yv12_fb[cm->lst_fb_idx];
655    xd->dst = cm->yv12_fb[cm->new_fb_idx];
656
657    /* set up frame for intra coded blocks */
658    vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);
659
660    vp8_build_block_offsets(x);
661
662    xd->mode_info_context->mbmi.mode = DC_PRED;
663    xd->mode_info_context->mbmi.uv_mode = DC_PRED;
664
665    xd->left_context = &cm->left_context;
666
667    x->mvc = cm->fc.mvc;
668
669    memset(cm->above_context, 0, sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);
670
671    /* Special case treatment when GF and ARF are not sensible options
672     * for reference
673     */
674    if (cpi->ref_frame_flags == VP8_LAST_FRAME)
675        vp8_calc_ref_frame_costs(x->ref_frame_cost,
676                                 cpi->prob_intra_coded,255,128);
677    else if ((cpi->oxcf.number_of_layers > 1) &&
678               (cpi->ref_frame_flags == VP8_GOLD_FRAME))
679        vp8_calc_ref_frame_costs(x->ref_frame_cost,
680                                 cpi->prob_intra_coded,1,255);
681    else if ((cpi->oxcf.number_of_layers > 1) &&
682                (cpi->ref_frame_flags == VP8_ALTR_FRAME))
683        vp8_calc_ref_frame_costs(x->ref_frame_cost,
684                                 cpi->prob_intra_coded,1,1);
685    else
686        vp8_calc_ref_frame_costs(x->ref_frame_cost,
687                                 cpi->prob_intra_coded,
688                                 cpi->prob_last_coded,
689                                 cpi->prob_gf_coded);
690
691    xd->fullpixel_mask = 0xffffffff;
692    if(cm->full_pixel)
693        xd->fullpixel_mask = 0xfffffff8;
694
695    vp8_zero(x->coef_counts);
696    vp8_zero(x->ymode_count);
697    vp8_zero(x->uv_mode_count)
698    x->prediction_error = 0;
699    x->intra_error = 0;
700    vp8_zero(x->count_mb_ref_frame_usage);
701}
702
703#if CONFIG_MULTITHREAD
704static void sum_coef_counts(MACROBLOCK *x, MACROBLOCK *x_thread)
705{
706    int i = 0;
707    do
708    {
709        int j = 0;
710        do
711        {
712            int k = 0;
713            do
714            {
715                /* at every context */
716
717                /* calc probs and branch cts for this frame only */
718                int t = 0;      /* token/prob index */
719
720                do
721                {
722                    x->coef_counts [i][j][k][t] +=
723                        x_thread->coef_counts [i][j][k][t];
724                }
725                while (++t < ENTROPY_NODES);
726            }
727            while (++k < PREV_COEF_CONTEXTS);
728        }
729        while (++j < COEF_BANDS);
730    }
731    while (++i < BLOCK_TYPES);
732}
733#endif  // CONFIG_MULTITHREAD
734
735void vp8_encode_frame(VP8_COMP *cpi)
736{
737    int mb_row;
738    MACROBLOCK *const x = & cpi->mb;
739    VP8_COMMON *const cm = & cpi->common;
740    MACROBLOCKD *const xd = & x->e_mbd;
741    TOKENEXTRA *tp = cpi->tok;
742    int segment_counts[MAX_MB_SEGMENTS];
743    int totalrate;
744#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
745    BOOL_CODER * bc = &cpi->bc[1]; /* bc[0] is for control partition */
746    const int num_part = (1 << cm->multi_token_partition);
747#endif
748
749    memset(segment_counts, 0, sizeof(segment_counts));
750    totalrate = 0;
751
752    if (cpi->compressor_speed == 2)
753    {
754        if (cpi->oxcf.cpu_used < 0)
755            cpi->Speed = -(cpi->oxcf.cpu_used);
756        else
757            vp8_auto_select_speed(cpi);
758    }
759
760    /* Functions setup for all frame types so we can use MC in AltRef */
761    if(!cm->use_bilinear_mc_filter)
762    {
763        xd->subpixel_predict        = vp8_sixtap_predict4x4;
764        xd->subpixel_predict8x4     = vp8_sixtap_predict8x4;
765        xd->subpixel_predict8x8     = vp8_sixtap_predict8x8;
766        xd->subpixel_predict16x16   = vp8_sixtap_predict16x16;
767    }
768    else
769    {
770        xd->subpixel_predict        = vp8_bilinear_predict4x4;
771        xd->subpixel_predict8x4     = vp8_bilinear_predict8x4;
772        xd->subpixel_predict8x8     = vp8_bilinear_predict8x8;
773        xd->subpixel_predict16x16   = vp8_bilinear_predict16x16;
774    }
775
776    cpi->mb.skip_true_count = 0;
777    cpi->tok_count = 0;
778
779#if 0
780    /* Experimental code */
781    cpi->frame_distortion = 0;
782    cpi->last_mb_distortion = 0;
783#endif
784
785    xd->mode_info_context = cm->mi;
786
787    vp8_zero(cpi->mb.MVcount);
788
789    vp8cx_frame_init_quantizer(cpi);
790
791    vp8_initialize_rd_consts(cpi, x,
792                             vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
793
794    vp8cx_initialize_me_consts(cpi, cm->base_qindex);
795
796    if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
797    {
798        /* Initialize encode frame context. */
799        init_encode_frame_mb_context(cpi);
800
801        /* Build a frame level activity map */
802        build_activity_map(cpi);
803    }
804
805    /* re-init encode frame context. */
806    init_encode_frame_mb_context(cpi);
807
808#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
809    {
810        int i;
811        for(i = 0; i < num_part; i++)
812        {
813            vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
814                    cpi->partition_d_end[i + 1]);
815            bc[i].error = &cm->error;
816        }
817    }
818
819#endif
820
821    {
822        struct vpx_usec_timer  emr_timer;
823        vpx_usec_timer_start(&emr_timer);
824
825#if CONFIG_MULTITHREAD
826        if (cpi->b_multi_threaded)
827        {
828            int i;
829
830            vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei,
831                                      cpi->encoding_thread_count);
832
833            for (i = 0; i < cm->mb_rows; i++)
834                cpi->mt_current_mb_col[i] = -1;
835
836            for (i = 0; i < cpi->encoding_thread_count; i++)
837            {
838                sem_post(&cpi->h_event_start_encoding[i]);
839            }
840
841            for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
842            {
843                vp8_zero(cm->left_context)
844
845#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
846                tp = cpi->tok;
847#else
848                tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
849#endif
850
851                encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
852
853                /* adjust to the next row of mbs */
854                x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
855                x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
856                x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
857
858                xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
859                x->partition_info  += xd->mode_info_stride * cpi->encoding_thread_count;
860                x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
861
862                if(mb_row == cm->mb_rows - 1)
863                {
864                    sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
865                }
866            }
867
868            sem_wait(&cpi->h_event_end_encoding); /* wait for other threads to finish */
869
870            for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
871            {
872                cpi->tok_count += (unsigned int)
873                  (cpi->tplist[mb_row].stop - cpi->tplist[mb_row].start);
874            }
875
876            if (xd->segmentation_enabled)
877            {
878                int j;
879
880                if (xd->segmentation_enabled)
881                {
882                    for (i = 0; i < cpi->encoding_thread_count; i++)
883                    {
884                        for (j = 0; j < 4; j++)
885                            segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
886                    }
887                }
888            }
889
890            for (i = 0; i < cpi->encoding_thread_count; i++)
891            {
892                int mode_count;
893                int c_idx;
894                totalrate += cpi->mb_row_ei[i].totalrate;
895
896                cpi->mb.skip_true_count += cpi->mb_row_ei[i].mb.skip_true_count;
897
898                for(mode_count = 0; mode_count < VP8_YMODES; mode_count++)
899                    cpi->mb.ymode_count[mode_count] +=
900                        cpi->mb_row_ei[i].mb.ymode_count[mode_count];
901
902                for(mode_count = 0; mode_count < VP8_UV_MODES; mode_count++)
903                    cpi->mb.uv_mode_count[mode_count] +=
904                        cpi->mb_row_ei[i].mb.uv_mode_count[mode_count];
905
906                for(c_idx = 0; c_idx < MVvals; c_idx++)
907                {
908                    cpi->mb.MVcount[0][c_idx] +=
909                        cpi->mb_row_ei[i].mb.MVcount[0][c_idx];
910                    cpi->mb.MVcount[1][c_idx] +=
911                        cpi->mb_row_ei[i].mb.MVcount[1][c_idx];
912                }
913
914                cpi->mb.prediction_error +=
915                    cpi->mb_row_ei[i].mb.prediction_error;
916                cpi->mb.intra_error += cpi->mb_row_ei[i].mb.intra_error;
917
918                for(c_idx = 0; c_idx < MAX_REF_FRAMES; c_idx++)
919                    cpi->mb.count_mb_ref_frame_usage[c_idx] +=
920                        cpi->mb_row_ei[i].mb.count_mb_ref_frame_usage[c_idx];
921
922                for(c_idx = 0; c_idx < MAX_ERROR_BINS; c_idx++)
923                    cpi->mb.error_bins[c_idx] +=
924                        cpi->mb_row_ei[i].mb.error_bins[c_idx];
925
926                /* add up counts for each thread */
927                sum_coef_counts(x, &cpi->mb_row_ei[i].mb);
928            }
929
930        }
931        else
932#endif  // CONFIG_MULTITHREAD
933        {
934
935            /* for each macroblock row in image */
936            for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
937            {
938                vp8_zero(cm->left_context)
939
940#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
941                tp = cpi->tok;
942#endif
943
944                encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
945
946                /* adjust to the next row of mbs */
947                x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
948                x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
949                x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
950            }
951
952            cpi->tok_count = (unsigned int)(tp - cpi->tok);
953        }
954
955#if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
956        {
957            int i;
958            for(i = 0; i < num_part; i++)
959            {
960                vp8_stop_encode(&bc[i]);
961                cpi->partition_sz[i+1] = bc[i].pos;
962            }
963        }
964#endif
965
966        vpx_usec_timer_mark(&emr_timer);
967        cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
968    }
969
970
971    // Work out the segment probabilities if segmentation is enabled
972    // and needs to be updated
973    if (xd->segmentation_enabled && xd->update_mb_segmentation_map)
974    {
975        int tot_count;
976        int i;
977
978        /* Set to defaults */
979        memset(xd->mb_segment_tree_probs, 255 , sizeof(xd->mb_segment_tree_probs));
980
981        tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] + segment_counts[3];
982
983        if (tot_count)
984        {
985            xd->mb_segment_tree_probs[0] = ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
986
987            tot_count = segment_counts[0] + segment_counts[1];
988
989            if (tot_count > 0)
990            {
991                xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
992            }
993
994            tot_count = segment_counts[2] + segment_counts[3];
995
996            if (tot_count > 0)
997                xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
998
999            /* Zero probabilities not allowed */
1000            for (i = 0; i < MB_FEATURE_TREE_PROBS; i ++)
1001            {
1002                if (xd->mb_segment_tree_probs[i] == 0)
1003                    xd->mb_segment_tree_probs[i] = 1;
1004            }
1005        }
1006    }
1007
1008    /* projected_frame_size in units of BYTES */
1009    cpi->projected_frame_size = totalrate >> 8;
1010
1011    /* Make a note of the percentage MBs coded Intra. */
1012    if (cm->frame_type == KEY_FRAME)
1013    {
1014        cpi->this_frame_percent_intra = 100;
1015    }
1016    else
1017    {
1018        int tot_modes;
1019
1020        tot_modes = cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME]
1021                    + cpi->mb.count_mb_ref_frame_usage[LAST_FRAME]
1022                    + cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME]
1023                    + cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
1024
1025        if (tot_modes)
1026            cpi->this_frame_percent_intra =
1027                cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
1028
1029    }
1030
1031#if ! CONFIG_REALTIME_ONLY
1032    /* Adjust the projected reference frame usage probability numbers to
1033     * reflect what we have just seen. This may be useful when we make
1034     * multiple iterations of the recode loop rather than continuing to use
1035     * values from the previous frame.
1036     */
1037    if ((cm->frame_type != KEY_FRAME) && ((cpi->oxcf.number_of_layers > 1) ||
1038        (!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame)))
1039    {
1040      vp8_convert_rfct_to_prob(cpi);
1041    }
1042#endif
1043}
1044void vp8_setup_block_ptrs(MACROBLOCK *x)
1045{
1046    int r, c;
1047    int i;
1048
1049    for (r = 0; r < 4; r++)
1050    {
1051        for (c = 0; c < 4; c++)
1052        {
1053            x->block[r*4+c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
1054        }
1055    }
1056
1057    for (r = 0; r < 2; r++)
1058    {
1059        for (c = 0; c < 2; c++)
1060        {
1061            x->block[16 + r*2+c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
1062        }
1063    }
1064
1065
1066    for (r = 0; r < 2; r++)
1067    {
1068        for (c = 0; c < 2; c++)
1069        {
1070            x->block[20 + r*2+c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
1071        }
1072    }
1073
1074    x->block[24].src_diff = x->src_diff + 384;
1075
1076
1077    for (i = 0; i < 25; i++)
1078    {
1079        x->block[i].coeff = x->coeff + i * 16;
1080    }
1081}
1082
1083void vp8_build_block_offsets(MACROBLOCK *x)
1084{
1085    int block = 0;
1086    int br, bc;
1087
1088    vp8_build_block_doffsets(&x->e_mbd);
1089
1090    /* y blocks */
1091    x->thismb_ptr = &x->thismb[0];
1092    for (br = 0; br < 4; br++)
1093    {
1094        for (bc = 0; bc < 4; bc++)
1095        {
1096            BLOCK *this_block = &x->block[block];
1097            this_block->base_src = &x->thismb_ptr;
1098            this_block->src_stride = 16;
1099            this_block->src = 4 * br * 16 + 4 * bc;
1100            ++block;
1101        }
1102    }
1103
1104    /* u blocks */
1105    for (br = 0; br < 2; br++)
1106    {
1107        for (bc = 0; bc < 2; bc++)
1108        {
1109            BLOCK *this_block = &x->block[block];
1110            this_block->base_src = &x->src.u_buffer;
1111            this_block->src_stride = x->src.uv_stride;
1112            this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1113            ++block;
1114        }
1115    }
1116
1117    /* v blocks */
1118    for (br = 0; br < 2; br++)
1119    {
1120        for (bc = 0; bc < 2; bc++)
1121        {
1122            BLOCK *this_block = &x->block[block];
1123            this_block->base_src = &x->src.v_buffer;
1124            this_block->src_stride = x->src.uv_stride;
1125            this_block->src = 4 * br * this_block->src_stride + 4 * bc;
1126            ++block;
1127        }
1128    }
1129}
1130
1131static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x)
1132{
1133    const MACROBLOCKD *xd = & x->e_mbd;
1134    const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
1135    const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;
1136
1137#ifdef MODE_STATS
1138    const int is_key = cpi->common.frame_type == KEY_FRAME;
1139
1140    ++ (is_key ? uv_modes : inter_uv_modes)[uvm];
1141
1142    if (m == B_PRED)
1143    {
1144        unsigned int *const bct = is_key ? b_modes : inter_b_modes;
1145
1146        int b = 0;
1147
1148        do
1149        {
1150            ++ bct[xd->block[b].bmi.mode];
1151        }
1152        while (++b < 16);
1153    }
1154
1155#else
1156    (void)cpi;
1157#endif
1158
1159    ++x->ymode_count[m];
1160    ++x->uv_mode_count[uvm];
1161
1162}
1163
1164/* Experimental stub function to create a per MB zbin adjustment based on
1165 * some previously calculated measure of MB activity.
1166 */
1167static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x )
1168{
1169#if USE_ACT_INDEX
1170    x->act_zbin_adj = *(x->mb_activity_ptr);
1171#else
1172    int64_t a;
1173    int64_t b;
1174    int64_t act = *(x->mb_activity_ptr);
1175
1176    /* Apply the masking to the RD multiplier. */
1177    a = act + 4*cpi->activity_avg;
1178    b = 4*act + cpi->activity_avg;
1179
1180    if ( act > cpi->activity_avg )
1181        x->act_zbin_adj = (int)(((int64_t)b + (a>>1))/a) - 1;
1182    else
1183        x->act_zbin_adj = 1 - (int)(((int64_t)a + (b>>1))/b);
1184#endif
1185}
1186
1187int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
1188                                  TOKENEXTRA **t)
1189{
1190    MACROBLOCKD *xd = &x->e_mbd;
1191    int rate;
1192
1193    if (cpi->sf.RD && cpi->compressor_speed != 2)
1194        vp8_rd_pick_intra_mode(x, &rate);
1195    else
1196        vp8_pick_intra_mode(x, &rate);
1197
1198    if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1199    {
1200        adjust_act_zbin( cpi, x );
1201        vp8_update_zbin_extra(cpi, x);
1202    }
1203
1204    if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED)
1205        vp8_encode_intra4x4mby(x);
1206    else
1207        vp8_encode_intra16x16mby(x);
1208
1209    vp8_encode_intra16x16mbuv(x);
1210
1211    sum_intra_stats(cpi, x);
1212
1213    vp8_tokenize_mb(cpi, x, t);
1214
1215    if (xd->mode_info_context->mbmi.mode != B_PRED)
1216        vp8_inverse_transform_mby(xd);
1217
1218    vp8_dequant_idct_add_uv_block
1219                    (xd->qcoeff+16*16, xd->dequant_uv,
1220                     xd->dst.u_buffer, xd->dst.v_buffer,
1221                     xd->dst.uv_stride, xd->eobs+16);
1222    return rate;
1223}
1224#ifdef SPEEDSTATS
1225extern int cnt_pm;
1226#endif
1227
1228extern void vp8_fix_contexts(MACROBLOCKD *x);
1229
1230int vp8cx_encode_inter_macroblock
1231(
1232    VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
1233    int recon_yoffset, int recon_uvoffset,
1234    int mb_row, int mb_col
1235)
1236{
1237    MACROBLOCKD *const xd = &x->e_mbd;
1238    int intra_error = 0;
1239    int rate;
1240    int distortion;
1241
1242    x->skip = 0;
1243
1244    if (xd->segmentation_enabled)
1245        x->encode_breakout = cpi->segment_encode_breakout[xd->mode_info_context->mbmi.segment_id];
1246    else
1247        x->encode_breakout = cpi->oxcf.encode_breakout;
1248
1249#if CONFIG_TEMPORAL_DENOISING
1250    /* Reset the best sse mode/mv for each macroblock. */
1251    x->best_reference_frame = INTRA_FRAME;
1252    x->best_zeromv_reference_frame = INTRA_FRAME;
1253    x->best_sse_inter_mode = 0;
1254    x->best_sse_mv.as_int = 0;
1255    x->need_to_clamp_best_mvs = 0;
1256#endif
1257
1258    if (cpi->sf.RD)
1259    {
1260        int zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
1261
1262        /* Are we using the fast quantizer for the mode selection? */
1263        if(cpi->sf.use_fastquant_for_pick)
1264        {
1265            x->quantize_b      = vp8_fast_quantize_b;
1266
1267            /* the fast quantizer does not use zbin_extra, so
1268             * do not recalculate */
1269            x->zbin_mode_boost_enabled = 0;
1270        }
1271        vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1272                               &distortion, &intra_error, mb_row, mb_col);
1273
1274        /* switch back to the regular quantizer for the encode */
1275        if (cpi->sf.improved_quant)
1276        {
1277            x->quantize_b      = vp8_regular_quantize_b;
1278        }
1279
1280        /* restore cpi->zbin_mode_boost_enabled */
1281        x->zbin_mode_boost_enabled = zbin_mode_boost_enabled;
1282
1283    }
1284    else
1285    {
1286        vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
1287                            &distortion, &intra_error, mb_row, mb_col);
1288    }
1289
1290    x->prediction_error += distortion;
1291    x->intra_error += intra_error;
1292
1293    if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
1294    {
1295        /* Adjust the zbin based on this MB rate. */
1296        adjust_act_zbin( cpi, x );
1297    }
1298
1299#if 0
1300    /* Experimental RD code */
1301    cpi->frame_distortion += distortion;
1302    cpi->last_mb_distortion = distortion;
1303#endif
1304
1305    /* MB level adjutment to quantizer setup */
1306    if (xd->segmentation_enabled)
1307    {
1308        /* If cyclic update enabled */
1309        if (cpi->current_layer == 0 && cpi->cyclic_refresh_mode_enabled)
1310        {
1311            /* Clear segment_id back to 0 if not coded (last frame 0,0) */
1312            if ((xd->mode_info_context->mbmi.segment_id == 1) &&
1313                ((xd->mode_info_context->mbmi.ref_frame != LAST_FRAME) || (xd->mode_info_context->mbmi.mode != ZEROMV)))
1314            {
1315                xd->mode_info_context->mbmi.segment_id = 0;
1316
1317                /* segment_id changed, so update */
1318                vp8cx_mb_init_quantizer(cpi, x, 1);
1319            }
1320        }
1321    }
1322
1323    {
1324        /* Experimental code.
1325         * Special case for gf and arf zeromv modes, for 1 temporal layer.
1326         * Increase zbin size to supress noise.
1327         */
1328        x->zbin_mode_boost = 0;
1329        if (x->zbin_mode_boost_enabled)
1330        {
1331            if ( xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME )
1332            {
1333                if (xd->mode_info_context->mbmi.mode == ZEROMV)
1334                {
1335                    if (xd->mode_info_context->mbmi.ref_frame != LAST_FRAME &&
1336                        cpi->oxcf.number_of_layers == 1)
1337                        x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
1338                    else
1339                        x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
1340                }
1341                else if (xd->mode_info_context->mbmi.mode == SPLITMV)
1342                    x->zbin_mode_boost = 0;
1343                else
1344                    x->zbin_mode_boost = MV_ZBIN_BOOST;
1345            }
1346        }
1347
1348        /* The fast quantizer doesn't use zbin_extra, only do so with
1349         * the regular quantizer. */
1350        if (cpi->sf.improved_quant)
1351            vp8_update_zbin_extra(cpi, x);
1352    }
1353
1354    x->count_mb_ref_frame_usage[xd->mode_info_context->mbmi.ref_frame] ++;
1355
1356    if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
1357    {
1358        vp8_encode_intra16x16mbuv(x);
1359
1360        if (xd->mode_info_context->mbmi.mode == B_PRED)
1361        {
1362            vp8_encode_intra4x4mby(x);
1363        }
1364        else
1365        {
1366            vp8_encode_intra16x16mby(x);
1367        }
1368
1369        sum_intra_stats(cpi, x);
1370    }
1371    else
1372    {
1373        int ref_fb_idx;
1374
1375        if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
1376            ref_fb_idx = cpi->common.lst_fb_idx;
1377        else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
1378            ref_fb_idx = cpi->common.gld_fb_idx;
1379        else
1380            ref_fb_idx = cpi->common.alt_fb_idx;
1381
1382        xd->pre.y_buffer = cpi->common.yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
1383        xd->pre.u_buffer = cpi->common.yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
1384        xd->pre.v_buffer = cpi->common.yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
1385
1386        if (!x->skip)
1387        {
1388            vp8_encode_inter16x16(x);
1389        }
1390        else
1391            vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
1392                                           xd->dst.u_buffer, xd->dst.v_buffer,
1393                                           xd->dst.y_stride, xd->dst.uv_stride);
1394
1395    }
1396
1397    if (!x->skip)
1398    {
1399        vp8_tokenize_mb(cpi, x, t);
1400
1401        if (xd->mode_info_context->mbmi.mode != B_PRED)
1402            vp8_inverse_transform_mby(xd);
1403
1404        vp8_dequant_idct_add_uv_block
1405                        (xd->qcoeff+16*16, xd->dequant_uv,
1406                         xd->dst.u_buffer, xd->dst.v_buffer,
1407                         xd->dst.uv_stride, xd->eobs+16);
1408    }
1409    else
1410    {
1411        /* always set mb_skip_coeff as it is needed by the loopfilter */
1412        xd->mode_info_context->mbmi.mb_skip_coeff = 1;
1413
1414        if (cpi->common.mb_no_coeff_skip)
1415        {
1416            x->skip_true_count ++;
1417            vp8_fix_contexts(xd);
1418        }
1419        else
1420        {
1421            vp8_stuff_mb(cpi, x, t);
1422        }
1423    }
1424
1425    return rate;
1426}
1427