michael@0: /*
michael@0:  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS.  All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: 
michael@0: #include "vpx_config.h"
michael@0: #include "vp8_rtcd.h"
michael@0: #include "encodemb.h"
michael@0: #include "encodemv.h"
michael@0: #include "vp8/common/common.h"
michael@0: #include "onyx_int.h"
michael@0: #include "vp8/common/extend.h"
michael@0: #include "vp8/common/entropymode.h"
michael@0: #include "vp8/common/quant_common.h"
michael@0: #include "segmentation.h"
michael@0: #include "vp8/common/setupintrarecon.h"
michael@0: #include "encodeintra.h"
michael@0: #include "vp8/common/reconinter.h"
michael@0: #include "rdopt.h"
michael@0: #include "pickinter.h"
michael@0: #include "vp8/common/findnearmv.h"
michael@0: #include <stdio.h>
michael@0: #include <limits.h>
michael@0: #include "vp8/common/invtrans.h"
michael@0: #include "vpx_ports/vpx_timer.h"
michael@0: #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
michael@0: #include "bitstream.h"
michael@0: #endif
michael@0: #include "encodeframe.h"
michael@0: 
michael@0: extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t) ;
michael@0: extern void vp8_calc_ref_frame_costs(int *ref_frame_cost,
michael@0:                                      int prob_intra,
michael@0:                                      int prob_last,
michael@0:                                      int prob_garf
michael@0:                                     );
michael@0: extern void vp8_convert_rfct_to_prob(VP8_COMP *const cpi);
michael@0: extern void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex);
michael@0: extern void vp8_auto_select_speed(VP8_COMP *cpi);
michael@0: extern void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
michael@0:                                       MACROBLOCK *x,
michael@0:                                       MB_ROW_COMP *mbr_ei,
michael@0:                                       int count);
michael@0: static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x );
michael@0: 
michael@0: #ifdef MODE_STATS
michael@0: unsigned int inter_y_modes[10] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
michael@0: unsigned int inter_uv_modes[4] = {0, 0, 0, 0};
michael@0: unsigned int inter_b_modes[15]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
michael@0: unsigned int y_modes[5]   = {0, 0, 0, 0, 0};
michael@0: unsigned int uv_modes[4]  = {0, 0, 0, 0};
michael@0: unsigned int b_modes[14]  = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
michael@0: #endif
michael@0: 
michael@0: 
michael@0: /* activity_avg must be positive, or flat regions could get a zero weight
michael@0:  *  (infinite lambda), which confounds analysis.
michael@0:  * This also avoids the need for divide by zero checks in
michael@0:  *  vp8_activity_masking().
michael@0:  */
michael@0: #define VP8_ACTIVITY_AVG_MIN (64)
michael@0: 
michael@0: /* This is used as a reference when computing the source variance for the
michael@0:  *  purposes of activity masking.
michael@0:  * Eventually this should be replaced by custom no-reference routines,
michael@0:  *  which will be faster.
michael@0:  */
michael@0: static const unsigned char VP8_VAR_OFFS[16]=
michael@0: {
michael@0:     128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128
michael@0: };
michael@0: 
michael@0: 
michael@0: /* Original activity measure from Tim T's code. */
michael@0: static unsigned int tt_activity_measure( VP8_COMP *cpi, MACROBLOCK *x )
michael@0: {
michael@0:     unsigned int act;
michael@0:     unsigned int sse;
michael@0:     /* TODO: This could also be done over smaller areas (8x8), but that would
michael@0:      *  require extensive changes elsewhere, as lambda is assumed to be fixed
michael@0:      *  over an entire MB in most of the code.
michael@0:      * Another option is to compute four 8x8 variances, and pick a single
michael@0:      *  lambda using a non-linear combination (e.g., the smallest, or second
michael@0:      *  smallest, etc.).
michael@0:      */
michael@0:     act =  vp8_variance16x16(x->src.y_buffer,
michael@0:                     x->src.y_stride, VP8_VAR_OFFS, 0, &sse);
michael@0:     act = act<<4;
michael@0: 
michael@0:     /* If the region is flat, lower the activity some more. */
michael@0:     if (act < 8<<12)
michael@0:         act = act < 5<<12 ? act : 5<<12;
michael@0: 
michael@0:     return act;
michael@0: }
michael@0: 
michael@0: /* Stub for alternative experimental activity measures. */
michael@0: static unsigned int alt_activity_measure( VP8_COMP *cpi,
michael@0:                                           MACROBLOCK *x, int use_dc_pred )
michael@0: {
michael@0:     return vp8_encode_intra(cpi,x, use_dc_pred);
michael@0: }
michael@0: 
michael@0: 
michael@0: /* Measure the activity of the current macroblock
michael@0:  * What we measure here is TBD so abstracted to this function
michael@0:  */
michael@0: #define ALT_ACT_MEASURE 1
michael@0: static unsigned int mb_activity_measure( VP8_COMP *cpi, MACROBLOCK *x,
michael@0:                                   int mb_row, int mb_col)
michael@0: {
michael@0:     unsigned int mb_activity;
michael@0: 
michael@0:     if  ( ALT_ACT_MEASURE )
michael@0:     {
michael@0:         int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
michael@0: 
michael@0:         /* Or use and alternative. */
michael@0:         mb_activity = alt_activity_measure( cpi, x, use_dc_pred );
michael@0:     }
michael@0:     else
michael@0:     {
michael@0:         /* Original activity measure from Tim T's code. */
michael@0:         mb_activity = tt_activity_measure( cpi, x );
michael@0:     }
michael@0: 
michael@0:     if ( mb_activity < VP8_ACTIVITY_AVG_MIN )
michael@0:         mb_activity = VP8_ACTIVITY_AVG_MIN;
michael@0: 
michael@0:     return mb_activity;
michael@0: }
michael@0: 
michael@0: /* Calculate an "average" mb activity value for the frame */
michael@0: #define ACT_MEDIAN 0
michael@0: static void calc_av_activity( VP8_COMP *cpi, int64_t activity_sum )
michael@0: {
michael@0: #if ACT_MEDIAN
michael@0:     /* Find median: Simple n^2 algorithm for experimentation */
michael@0:     {
michael@0:         unsigned int median;
michael@0:         unsigned int i,j;
michael@0:         unsigned int * sortlist;
michael@0:         unsigned int tmp;
michael@0: 
michael@0:         /* Create a list to sort to */
michael@0:         CHECK_MEM_ERROR(sortlist,
michael@0:                         vpx_calloc(sizeof(unsigned int),
michael@0:                         cpi->common.MBs));
michael@0: 
michael@0:         /* Copy map to sort list */
michael@0:         vpx_memcpy( sortlist, cpi->mb_activity_map,
michael@0:                     sizeof(unsigned int) * cpi->common.MBs );
michael@0: 
michael@0: 
michael@0:         /* Ripple each value down to its correct position */
michael@0:         for ( i = 1; i < cpi->common.MBs; i ++ )
michael@0:         {
michael@0:             for ( j = i; j > 0; j -- )
michael@0:             {
michael@0:                 if ( sortlist[j] < sortlist[j-1] )
michael@0:                 {
michael@0:                     /* Swap values */
michael@0:                     tmp = sortlist[j-1];
michael@0:                     sortlist[j-1] = sortlist[j];
michael@0:                     sortlist[j] = tmp;
michael@0:                 }
michael@0:                 else
michael@0:                     break;
michael@0:             }
michael@0:         }
michael@0: 
michael@0:         /* Even number MBs so estimate median as mean of two either side. */
michael@0:         median = ( 1 + sortlist[cpi->common.MBs >> 1] +
michael@0:                    sortlist[(cpi->common.MBs >> 1) + 1] ) >> 1;
michael@0: 
michael@0:         cpi->activity_avg = median;
michael@0: 
michael@0:         vpx_free(sortlist);
michael@0:     }
michael@0: #else
michael@0:     /* Simple mean for now */
michael@0:     cpi->activity_avg = (unsigned int)(activity_sum/cpi->common.MBs);
michael@0: #endif
michael@0: 
michael@0:     if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN)
michael@0:         cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;
michael@0: 
michael@0:     /* Experimental code: return fixed value normalized for several clips */
michael@0:     if  ( ALT_ACT_MEASURE )
michael@0:         cpi->activity_avg = 100000;
michael@0: }
michael@0: 
michael@0: #define USE_ACT_INDEX   0
michael@0: #define OUTPUT_NORM_ACT_STATS   0
michael@0: 
michael@0: #if USE_ACT_INDEX
michael@0: /* Calculate and activity index for each mb */
michael@0: static void calc_activity_index( VP8_COMP *cpi, MACROBLOCK *x )
michael@0: {
michael@0:     VP8_COMMON *const cm = & cpi->common;
michael@0:     int mb_row, mb_col;
michael@0: 
michael@0:     int64_t act;
michael@0:     int64_t a;
michael@0:     int64_t b;
michael@0: 
michael@0: #if OUTPUT_NORM_ACT_STATS
michael@0:     FILE *f = fopen("norm_act.stt", "a");
michael@0:     fprintf(f, "\n%12d\n", cpi->activity_avg );
michael@0: #endif
michael@0: 
michael@0:     /* Reset pointers to start of activity map */
michael@0:     x->mb_activity_ptr = cpi->mb_activity_map;
michael@0: 
michael@0:     /* Calculate normalized mb activity number. */
michael@0:     for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
michael@0:     {
michael@0:         /* for each macroblock col in image */
michael@0:         for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
michael@0:         {
michael@0:             /* Read activity from the map */
michael@0:             act = *(x->mb_activity_ptr);
michael@0: 
michael@0:             /* Calculate a normalized activity number */
michael@0:             a = act + 4*cpi->activity_avg;
michael@0:             b = 4*act + cpi->activity_avg;
michael@0: 
michael@0:             if ( b >= a )
michael@0:                 *(x->activity_ptr) = (int)((b + (a>>1))/a) - 1;
michael@0:             else
michael@0:                 *(x->activity_ptr) = 1 - (int)((a + (b>>1))/b);
michael@0: 
michael@0: #if OUTPUT_NORM_ACT_STATS
michael@0:             fprintf(f, " %6d", *(x->mb_activity_ptr));
michael@0: #endif
michael@0:             /* Increment activity map pointers */
michael@0:             x->mb_activity_ptr++;
michael@0:         }
michael@0: 
michael@0: #if OUTPUT_NORM_ACT_STATS
michael@0:         fprintf(f, "\n");
michael@0: #endif
michael@0: 
michael@0:     }
michael@0: 
michael@0: #if OUTPUT_NORM_ACT_STATS
michael@0:     fclose(f);
michael@0: #endif
michael@0: 
michael@0: }
michael@0: #endif
michael@0: 
michael@0: /* Loop through all MBs. Note activity of each, average activity and
michael@0:  * calculate a normalized activity for each
michael@0:  */
michael@0: static void build_activity_map( VP8_COMP *cpi )
michael@0: {
michael@0:     MACROBLOCK *const x = & cpi->mb;
michael@0:     MACROBLOCKD *xd = &x->e_mbd;
michael@0:     VP8_COMMON *const cm = & cpi->common;
michael@0: 
michael@0: #if ALT_ACT_MEASURE
michael@0:     YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
michael@0:     int recon_yoffset;
michael@0:     int recon_y_stride = new_yv12->y_stride;
michael@0: #endif
michael@0: 
michael@0:     int mb_row, mb_col;
michael@0:     unsigned int mb_activity;
michael@0:     int64_t activity_sum = 0;
michael@0: 
michael@0:     /* for each macroblock row in image */
michael@0:     for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
michael@0:     {
michael@0: #if ALT_ACT_MEASURE
michael@0:         /* reset above block coeffs */
michael@0:         xd->up_available = (mb_row != 0);
michael@0:         recon_yoffset = (mb_row * recon_y_stride * 16);
michael@0: #endif
michael@0:         /* for each macroblock col in image */
michael@0:         for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
michael@0:         {
michael@0: #if ALT_ACT_MEASURE
michael@0:             xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
michael@0:             xd->left_available = (mb_col != 0);
michael@0:             recon_yoffset += 16;
michael@0: #endif
michael@0:             /* Copy current mb to a buffer */
michael@0:             vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
michael@0: 
michael@0:             /* measure activity */
michael@0:             mb_activity = mb_activity_measure( cpi, x, mb_row, mb_col );
michael@0: 
michael@0:             /* Keep frame sum */
michael@0:             activity_sum += mb_activity;
michael@0: 
michael@0:             /* Store MB level activity details. */
michael@0:             *x->mb_activity_ptr = mb_activity;
michael@0: 
michael@0:             /* Increment activity map pointer */
michael@0:             x->mb_activity_ptr++;
michael@0: 
michael@0:             /* adjust to the next column of source macroblocks */
michael@0:             x->src.y_buffer += 16;
michael@0:         }
michael@0: 
michael@0: 
michael@0:         /* adjust to the next row of mbs */
michael@0:         x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
michael@0: 
michael@0: #if ALT_ACT_MEASURE
michael@0:         /* extend the recon for intra prediction */
michael@0:         vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16,
michael@0:                           xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
michael@0: #endif
michael@0: 
michael@0:     }
michael@0: 
michael@0:     /* Calculate an "average" MB activity */
michael@0:     calc_av_activity(cpi, activity_sum);
michael@0: 
michael@0: #if USE_ACT_INDEX
michael@0:     /* Calculate an activity index number of each mb */
michael@0:     calc_activity_index( cpi, x );
michael@0: #endif
michael@0: 
michael@0: }
michael@0: 
michael@0: /* Macroblock activity masking */
michael@0: void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x)
michael@0: {
michael@0: #if USE_ACT_INDEX
michael@0:     x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
michael@0:     x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
michael@0:     x->errorperbit += (x->errorperbit==0);
michael@0: #else
michael@0:     int64_t a;
michael@0:     int64_t b;
michael@0:     int64_t act = *(x->mb_activity_ptr);
michael@0: 
michael@0:     /* Apply the masking to the RD multiplier. */
michael@0:     a = act + (2*cpi->activity_avg);
michael@0:     b = (2*act) + cpi->activity_avg;
michael@0: 
michael@0:     x->rdmult = (unsigned int)(((int64_t)x->rdmult*b + (a>>1))/a);
michael@0:     x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
michael@0:     x->errorperbit += (x->errorperbit==0);
michael@0: #endif
michael@0: 
michael@0:     /* Activity based Zbin adjustment */
michael@0:     adjust_act_zbin(cpi, x);
michael@0: }
michael@0: 
michael@0: static
michael@0: void encode_mb_row(VP8_COMP *cpi,
michael@0:                    VP8_COMMON *cm,
michael@0:                    int mb_row,
michael@0:                    MACROBLOCK  *x,
michael@0:                    MACROBLOCKD *xd,
michael@0:                    TOKENEXTRA **tp,
michael@0:                    int *segment_counts,
michael@0:                    int *totalrate)
michael@0: {
michael@0:     int recon_yoffset, recon_uvoffset;
michael@0:     int mb_col;
michael@0:     int ref_fb_idx = cm->lst_fb_idx;
michael@0:     int dst_fb_idx = cm->new_fb_idx;
michael@0:     int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
michael@0:     int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
michael@0:     int map_index = (mb_row * cpi->common.mb_cols);
michael@0: 
michael@0: #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
michael@0:     const int num_part = (1 << cm->multi_token_partition);
michael@0:     TOKENEXTRA * tp_start = cpi->tok;
michael@0:     vp8_writer *w;
michael@0: #endif
michael@0: 
michael@0: #if CONFIG_MULTITHREAD
michael@0:     const int nsync = cpi->mt_sync_range;
michael@0:     const int rightmost_col = cm->mb_cols + nsync;
michael@0:     volatile const int *last_row_current_mb_col;
michael@0:     volatile int *current_mb_col = &cpi->mt_current_mb_col[mb_row];
michael@0: 
michael@0:     if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
michael@0:         last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
michael@0:     else
michael@0:         last_row_current_mb_col = &rightmost_col;
michael@0: #endif
michael@0: 
michael@0: #if (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
michael@0:     if(num_part > 1)
michael@0:         w= &cpi->bc[1 + (mb_row % num_part)];
michael@0:     else
michael@0:         w = &cpi->bc[1];
michael@0: #endif
michael@0: 
michael@0:     /* reset above block coeffs */
michael@0:     xd->above_context = cm->above_context;
michael@0: 
michael@0:     xd->up_available = (mb_row != 0);
michael@0:     recon_yoffset = (mb_row * recon_y_stride * 16);
michael@0:     recon_uvoffset = (mb_row * recon_uv_stride * 8);
michael@0: 
michael@0:     cpi->tplist[mb_row].start = *tp;
michael@0:     /* printf("Main mb_row = %d\n", mb_row); */
michael@0: 
michael@0:     /* Distance of Mb to the top & bottom edges, specified in 1/8th pel
michael@0:      * units as they are always compared to values that are in 1/8th pel
michael@0:      */
michael@0:     xd->mb_to_top_edge = -((mb_row * 16) << 3);
michael@0:     xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
michael@0: 
michael@0:     /* Set up limit values for vertical motion vector components
michael@0:      * to prevent them extending beyond the UMV borders
michael@0:      */
michael@0:     x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
michael@0:     x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
michael@0:                         + (VP8BORDERINPIXELS - 16);
michael@0: 
michael@0:     /* Set the mb activity pointer to the start of the row. */
michael@0:     x->mb_activity_ptr = &cpi->mb_activity_map[map_index];
michael@0: 
michael@0:     /* for each macroblock col in image */
michael@0:     for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
michael@0:     {
michael@0: 
michael@0: #if  (CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING)
michael@0:         *tp = cpi->tok;
michael@0: #endif
michael@0:         /* Distance of Mb to the left & right edges, specified in
michael@0:          * 1/8th pel units as they are always compared to values
michael@0:          * that are in 1/8th pel units
michael@0:          */
michael@0:         xd->mb_to_left_edge = -((mb_col * 16) << 3);
michael@0:         xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;
michael@0: 
michael@0:         /* Set up limit values for horizontal motion vector components
michael@0:          * to prevent them extending beyond the UMV borders
michael@0:          */
michael@0:         x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
michael@0:         x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16)
michael@0:                             + (VP8BORDERINPIXELS - 16);
michael@0: 
michael@0:         xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
michael@0:         xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
michael@0:         xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
michael@0:         xd->left_available = (mb_col != 0);
michael@0: 
michael@0:         x->rddiv = cpi->RDDIV;
michael@0:         x->rdmult = cpi->RDMULT;
michael@0: 
michael@0:         /* Copy current mb to a buffer */
michael@0:         vp8_copy_mem16x16(x->src.y_buffer, x->src.y_stride, x->thismb, 16);
michael@0: 
michael@0: #if CONFIG_MULTITHREAD
michael@0:         if (cpi->b_multi_threaded != 0)
michael@0:         {
michael@0:             *current_mb_col = mb_col - 1; /* set previous MB done */
michael@0: 
michael@0:             if ((mb_col & (nsync - 1)) == 0)
michael@0:             {
michael@0:                 while (mb_col > (*last_row_current_mb_col - nsync))
michael@0:                 {
michael@0:                     x86_pause_hint();
michael@0:                     thread_sleep(0);
michael@0:                 }
michael@0:             }
michael@0:         }
michael@0: #endif
michael@0: 
michael@0:         if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
michael@0:             vp8_activity_masking(cpi, x);
michael@0: 
michael@0:         /* Is segmentation enabled */
michael@0:         /* MB level adjustment to quantizer */
michael@0:         if (xd->segmentation_enabled)
michael@0:         {
michael@0:             /* Code to set segment id in xd->mbmi.segment_id for current MB
michael@0:              * (with range checking)
michael@0:              */
michael@0:             if (cpi->segmentation_map[map_index+mb_col] <= 3)
michael@0:                 xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index+mb_col];
michael@0:             else
michael@0:                 xd->mode_info_context->mbmi.segment_id = 0;
michael@0: 
michael@0:             vp8cx_mb_init_quantizer(cpi, x, 1);
michael@0:         }
michael@0:         else
michael@0:             /* Set to Segment 0 by default */
michael@0:             xd->mode_info_context->mbmi.segment_id = 0;
michael@0: 
michael@0:         x->active_ptr = cpi->active_map + map_index + mb_col;
michael@0: 
michael@0:         if (cm->frame_type == KEY_FRAME)
michael@0:         {
michael@0:             *totalrate += vp8cx_encode_intra_macroblock(cpi, x, tp);
michael@0: #ifdef MODE_STATS
michael@0:             y_modes[xd->mbmi.mode] ++;
michael@0: #endif
michael@0:         }
michael@0:         else
michael@0:         {
michael@0:             *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset, mb_row, mb_col);
michael@0: 
michael@0: #ifdef MODE_STATS
michael@0:             inter_y_modes[xd->mbmi.mode] ++;
michael@0: 
michael@0:             if (xd->mbmi.mode == SPLITMV)
michael@0:             {
michael@0:                 int b;
michael@0: 
michael@0:                 for (b = 0; b < xd->mbmi.partition_count; b++)
michael@0:                 {
michael@0:                     inter_b_modes[x->partition->bmi[b].mode] ++;
michael@0:                 }
michael@0:             }
michael@0: 
michael@0: #endif
michael@0: 
michael@0:             /* Special case code for cyclic refresh
michael@0:              * If cyclic update enabled then copy xd->mbmi.segment_id; (which
michael@0:              * may have been updated based on mode during
michael@0:              * vp8cx_encode_inter_macroblock()) back into the global
michael@0:              * segmentation map
michael@0:              */
michael@0:             if ((cpi->current_layer == 0) &&
michael@0:                 (cpi->cyclic_refresh_mode_enabled &&
michael@0:                  xd->segmentation_enabled))
michael@0:             {
michael@0:                 cpi->segmentation_map[map_index+mb_col] = xd->mode_info_context->mbmi.segment_id;
michael@0: 
michael@0:                 /* If the block has been refreshed mark it as clean (the
michael@0:                  * magnitude of the -ve influences how long it will be before
michael@0:                  * we consider another refresh):
michael@0:                  * Else if it was coded (last frame 0,0) and has not already
michael@0:                  * been refreshed then mark it as a candidate for cleanup
michael@0:                  * next time (marked 0) else mark it as dirty (1).
michael@0:                  */
michael@0:                 if (xd->mode_info_context->mbmi.segment_id)
michael@0:                     cpi->cyclic_refresh_map[map_index+mb_col] = -1;
michael@0:                 else if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
michael@0:                 {
michael@0:                     if (cpi->cyclic_refresh_map[map_index+mb_col] == 1)
michael@0:                         cpi->cyclic_refresh_map[map_index+mb_col] = 0;
michael@0:                 }
michael@0:                 else
michael@0:                     cpi->cyclic_refresh_map[map_index+mb_col] = 1;
michael@0: 
michael@0:             }
michael@0:         }
michael@0: 
michael@0:         cpi->tplist[mb_row].stop = *tp;
michael@0: 
michael@0: #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
michael@0:         /* pack tokens for this MB */
michael@0:         {
michael@0:             int tok_count = *tp - tp_start;
michael@0:             pack_tokens(w, tp_start, tok_count);
michael@0:         }
michael@0: #endif
michael@0:         /* Increment pointer into gf usage flags structure. */
michael@0:         x->gf_active_ptr++;
michael@0: 
michael@0:         /* Increment the activity mask pointers. */
michael@0:         x->mb_activity_ptr++;
michael@0: 
michael@0:         /* adjust to the next column of macroblocks */
michael@0:         x->src.y_buffer += 16;
michael@0:         x->src.u_buffer += 8;
michael@0:         x->src.v_buffer += 8;
michael@0: 
michael@0:         recon_yoffset += 16;
michael@0:         recon_uvoffset += 8;
michael@0: 
michael@0:         /* Keep track of segment usage */
michael@0:         segment_counts[xd->mode_info_context->mbmi.segment_id] ++;
michael@0: 
michael@0:         /* skip to next mb */
michael@0:         xd->mode_info_context++;
michael@0:         x->partition_info++;
michael@0:         xd->above_context++;
michael@0:     }
michael@0: 
michael@0:     /* extend the recon for intra prediction */
michael@0:     vp8_extend_mb_row( &cm->yv12_fb[dst_fb_idx],
michael@0:                         xd->dst.y_buffer + 16,
michael@0:                         xd->dst.u_buffer + 8,
michael@0:                         xd->dst.v_buffer + 8);
michael@0: 
michael@0: #if CONFIG_MULTITHREAD
michael@0:     if (cpi->b_multi_threaded != 0)
michael@0:         *current_mb_col = rightmost_col;
michael@0: #endif
michael@0: 
michael@0:     /* this is to account for the border */
michael@0:     xd->mode_info_context++;
michael@0:     x->partition_info++;
michael@0: }
michael@0: 
michael@0: static void init_encode_frame_mb_context(VP8_COMP *cpi)
michael@0: {
michael@0:     MACROBLOCK *const x = & cpi->mb;
michael@0:     VP8_COMMON *const cm = & cpi->common;
michael@0:     MACROBLOCKD *const xd = & x->e_mbd;
michael@0: 
michael@0:     /* GF active flags data structure */
michael@0:     x->gf_active_ptr = (signed char *)cpi->gf_active_flags;
michael@0: 
michael@0:     /* Activity map pointer */
michael@0:     x->mb_activity_ptr = cpi->mb_activity_map;
michael@0: 
michael@0:     x->act_zbin_adj = 0;
michael@0: 
michael@0:     x->partition_info = x->pi;
michael@0: 
michael@0:     xd->mode_info_context = cm->mi;
michael@0:     xd->mode_info_stride = cm->mode_info_stride;
michael@0: 
michael@0:     xd->frame_type = cm->frame_type;
michael@0: 
michael@0:     /* reset intra mode contexts */
michael@0:     if (cm->frame_type == KEY_FRAME)
michael@0:         vp8_init_mbmode_probs(cm);
michael@0: 
michael@0:     /* Copy data over into macro block data structures. */
michael@0:     x->src = * cpi->Source;
michael@0:     xd->pre = cm->yv12_fb[cm->lst_fb_idx];
michael@0:     xd->dst = cm->yv12_fb[cm->new_fb_idx];
michael@0: 
michael@0:     /* set up frame for intra coded blocks */
michael@0:     vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);
michael@0: 
michael@0:     vp8_build_block_offsets(x);
michael@0: 
michael@0:     xd->mode_info_context->mbmi.mode = DC_PRED;
michael@0:     xd->mode_info_context->mbmi.uv_mode = DC_PRED;
michael@0: 
michael@0:     xd->left_context = &cm->left_context;
michael@0: 
michael@0:     x->mvc = cm->fc.mvc;
michael@0: 
michael@0:     vpx_memset(cm->above_context, 0,
michael@0:                sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);
michael@0: 
michael@0:     /* Special case treatment when GF and ARF are not sensible options
michael@0:      * for reference
michael@0:      */
michael@0:     if (cpi->ref_frame_flags == VP8_LAST_FRAME)
michael@0:         vp8_calc_ref_frame_costs(x->ref_frame_cost,
michael@0:                                  cpi->prob_intra_coded,255,128);
michael@0:     else if ((cpi->oxcf.number_of_layers > 1) &&
michael@0:                (cpi->ref_frame_flags == VP8_GOLD_FRAME))
michael@0:         vp8_calc_ref_frame_costs(x->ref_frame_cost,
michael@0:                                  cpi->prob_intra_coded,1,255);
michael@0:     else if ((cpi->oxcf.number_of_layers > 1) &&
michael@0:                 (cpi->ref_frame_flags == VP8_ALTR_FRAME))
michael@0:         vp8_calc_ref_frame_costs(x->ref_frame_cost,
michael@0:                                  cpi->prob_intra_coded,1,1);
michael@0:     else
michael@0:         vp8_calc_ref_frame_costs(x->ref_frame_cost,
michael@0:                                  cpi->prob_intra_coded,
michael@0:                                  cpi->prob_last_coded,
michael@0:                                  cpi->prob_gf_coded);
michael@0: 
michael@0:     xd->fullpixel_mask = 0xffffffff;
michael@0:     if(cm->full_pixel)
michael@0:         xd->fullpixel_mask = 0xfffffff8;
michael@0: 
michael@0:     vp8_zero(x->coef_counts);
michael@0:     vp8_zero(x->ymode_count);
michael@0:     vp8_zero(x->uv_mode_count)
michael@0:     x->prediction_error = 0;
michael@0:     x->intra_error = 0;
michael@0:     vp8_zero(x->count_mb_ref_frame_usage);
michael@0: }
michael@0: 
michael@0: static void sum_coef_counts(MACROBLOCK *x, MACROBLOCK *x_thread)
michael@0: {
michael@0:     int i = 0;
michael@0:     do
michael@0:     {
michael@0:         int j = 0;
michael@0:         do
michael@0:         {
michael@0:             int k = 0;
michael@0:             do
michael@0:             {
michael@0:                 /* at every context */
michael@0: 
michael@0:                 /* calc probs and branch cts for this frame only */
michael@0:                 int t = 0;      /* token/prob index */
michael@0: 
michael@0:                 do
michael@0:                 {
michael@0:                     x->coef_counts [i][j][k][t] +=
michael@0:                         x_thread->coef_counts [i][j][k][t];
michael@0:                 }
michael@0:                 while (++t < ENTROPY_NODES);
michael@0:             }
michael@0:             while (++k < PREV_COEF_CONTEXTS);
michael@0:         }
michael@0:         while (++j < COEF_BANDS);
michael@0:     }
michael@0:     while (++i < BLOCK_TYPES);
michael@0: }
michael@0: 
michael@0: void vp8_encode_frame(VP8_COMP *cpi)
michael@0: {
michael@0:     int mb_row;
michael@0:     MACROBLOCK *const x = & cpi->mb;
michael@0:     VP8_COMMON *const cm = & cpi->common;
michael@0:     MACROBLOCKD *const xd = & x->e_mbd;
michael@0:     TOKENEXTRA *tp = cpi->tok;
michael@0:     int segment_counts[MAX_MB_SEGMENTS];
michael@0:     int totalrate;
michael@0: #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
michael@0:     BOOL_CODER * bc = &cpi->bc[1]; /* bc[0] is for control partition */
michael@0:     const int num_part = (1 << cm->multi_token_partition);
michael@0: #endif
michael@0: 
michael@0:     vpx_memset(segment_counts, 0, sizeof(segment_counts));
michael@0:     totalrate = 0;
michael@0: 
michael@0:     if (cpi->compressor_speed == 2)
michael@0:     {
michael@0:         if (cpi->oxcf.cpu_used < 0)
michael@0:             cpi->Speed = -(cpi->oxcf.cpu_used);
michael@0:         else
michael@0:             vp8_auto_select_speed(cpi);
michael@0:     }
michael@0: 
michael@0:     /* Functions setup for all frame types so we can use MC in AltRef */
michael@0:     if(!cm->use_bilinear_mc_filter)
michael@0:     {
michael@0:         xd->subpixel_predict        = vp8_sixtap_predict4x4;
michael@0:         xd->subpixel_predict8x4     = vp8_sixtap_predict8x4;
michael@0:         xd->subpixel_predict8x8     = vp8_sixtap_predict8x8;
michael@0:         xd->subpixel_predict16x16   = vp8_sixtap_predict16x16;
michael@0:     }
michael@0:     else
michael@0:     {
michael@0:         xd->subpixel_predict        = vp8_bilinear_predict4x4;
michael@0:         xd->subpixel_predict8x4     = vp8_bilinear_predict8x4;
michael@0:         xd->subpixel_predict8x8     = vp8_bilinear_predict8x8;
michael@0:         xd->subpixel_predict16x16   = vp8_bilinear_predict16x16;
michael@0:     }
michael@0: 
michael@0:     cpi->mb.skip_true_count = 0;
michael@0:     cpi->tok_count = 0;
michael@0: 
michael@0: #if 0
michael@0:     /* Experimental code */
michael@0:     cpi->frame_distortion = 0;
michael@0:     cpi->last_mb_distortion = 0;
michael@0: #endif
michael@0: 
michael@0:     xd->mode_info_context = cm->mi;
michael@0: 
michael@0:     vp8_zero(cpi->mb.MVcount);
michael@0: 
michael@0:     vp8cx_frame_init_quantizer(cpi);
michael@0: 
michael@0:     vp8_initialize_rd_consts(cpi, x,
michael@0:                              vp8_dc_quant(cm->base_qindex, cm->y1dc_delta_q));
michael@0: 
michael@0:     vp8cx_initialize_me_consts(cpi, cm->base_qindex);
michael@0: 
michael@0:     if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
michael@0:     {
michael@0:         /* Initialize encode frame context. */
michael@0:         init_encode_frame_mb_context(cpi);
michael@0: 
michael@0:         /* Build a frame level activity map */
michael@0:         build_activity_map(cpi);
michael@0:     }
michael@0: 
michael@0:     /* re-init encode frame context. */
michael@0:     init_encode_frame_mb_context(cpi);
michael@0: 
michael@0: #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
michael@0:     {
michael@0:         int i;
michael@0:         for(i = 0; i < num_part; i++)
michael@0:         {
michael@0:             vp8_start_encode(&bc[i], cpi->partition_d[i + 1],
michael@0:                     cpi->partition_d_end[i + 1]);
michael@0:             bc[i].error = &cm->error;
michael@0:         }
michael@0:     }
michael@0: 
michael@0: #endif
michael@0: 
michael@0:     {
michael@0:         struct vpx_usec_timer  emr_timer;
michael@0:         vpx_usec_timer_start(&emr_timer);
michael@0: 
michael@0: #if CONFIG_MULTITHREAD
michael@0:         if (cpi->b_multi_threaded)
michael@0:         {
michael@0:             int i;
michael@0: 
michael@0:             vp8cx_init_mbrthread_data(cpi, x, cpi->mb_row_ei,
michael@0:                                       cpi->encoding_thread_count);
michael@0: 
michael@0:             for (i = 0; i < cm->mb_rows; i++)
michael@0:                 cpi->mt_current_mb_col[i] = -1;
michael@0: 
michael@0:             for (i = 0; i < cpi->encoding_thread_count; i++)
michael@0:             {
michael@0:                 sem_post(&cpi->h_event_start_encoding[i]);
michael@0:             }
michael@0: 
michael@0:             for (mb_row = 0; mb_row < cm->mb_rows; mb_row += (cpi->encoding_thread_count + 1))
michael@0:             {
michael@0:                 vp8_zero(cm->left_context)
michael@0: 
michael@0: #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
michael@0:                 tp = cpi->tok;
michael@0: #else
michael@0:                 tp = cpi->tok + mb_row * (cm->mb_cols * 16 * 24);
michael@0: #endif
michael@0: 
michael@0:                 encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
michael@0: 
michael@0:                 /* adjust to the next row of mbs */
michael@0:                 x->src.y_buffer += 16 * x->src.y_stride * (cpi->encoding_thread_count + 1) - 16 * cm->mb_cols;
michael@0:                 x->src.u_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
michael@0:                 x->src.v_buffer +=  8 * x->src.uv_stride * (cpi->encoding_thread_count + 1) - 8 * cm->mb_cols;
michael@0: 
michael@0:                 xd->mode_info_context += xd->mode_info_stride * cpi->encoding_thread_count;
michael@0:                 x->partition_info  += xd->mode_info_stride * cpi->encoding_thread_count;
michael@0:                 x->gf_active_ptr   += cm->mb_cols * cpi->encoding_thread_count;
michael@0: 
michael@0:                 if(mb_row == cm->mb_rows - 1)
michael@0:                 {
michael@0:                     sem_post(&cpi->h_event_end_encoding); /* signal frame encoding end */
michael@0:                 }
michael@0:             }
michael@0: 
michael@0:             sem_wait(&cpi->h_event_end_encoding); /* wait for other threads to finish */
michael@0: 
michael@0:             for (mb_row = 0; mb_row < cm->mb_rows; mb_row ++)
michael@0:             {
michael@0:                 cpi->tok_count += (unsigned int)
michael@0:                   (cpi->tplist[mb_row].stop - cpi->tplist[mb_row].start);
michael@0:             }
michael@0: 
michael@0:             if (xd->segmentation_enabled)
michael@0:             {
michael@0:                 int j;
michael@0: 
michael@0:                 if (xd->segmentation_enabled)
michael@0:                 {
michael@0:                     for (i = 0; i < cpi->encoding_thread_count; i++)
michael@0:                     {
michael@0:                         for (j = 0; j < 4; j++)
michael@0:                             segment_counts[j] += cpi->mb_row_ei[i].segment_counts[j];
michael@0:                     }
michael@0:                 }
michael@0:             }
michael@0: 
michael@0:             for (i = 0; i < cpi->encoding_thread_count; i++)
michael@0:             {
michael@0:                 int mode_count;
michael@0:                 int c_idx;
michael@0:                 totalrate += cpi->mb_row_ei[i].totalrate;
michael@0: 
michael@0:                 cpi->mb.skip_true_count += cpi->mb_row_ei[i].mb.skip_true_count;
michael@0: 
michael@0:                 for(mode_count = 0; mode_count < VP8_YMODES; mode_count++)
michael@0:                     cpi->mb.ymode_count[mode_count] +=
michael@0:                         cpi->mb_row_ei[i].mb.ymode_count[mode_count];
michael@0: 
michael@0:                 for(mode_count = 0; mode_count < VP8_UV_MODES; mode_count++)
michael@0:                     cpi->mb.uv_mode_count[mode_count] +=
michael@0:                         cpi->mb_row_ei[i].mb.uv_mode_count[mode_count];
michael@0: 
michael@0:                 for(c_idx = 0; c_idx < MVvals; c_idx++)
michael@0:                 {
michael@0:                     cpi->mb.MVcount[0][c_idx] +=
michael@0:                         cpi->mb_row_ei[i].mb.MVcount[0][c_idx];
michael@0:                     cpi->mb.MVcount[1][c_idx] +=
michael@0:                         cpi->mb_row_ei[i].mb.MVcount[1][c_idx];
michael@0:                 }
michael@0: 
michael@0:                 cpi->mb.prediction_error +=
michael@0:                     cpi->mb_row_ei[i].mb.prediction_error;
michael@0:                 cpi->mb.intra_error += cpi->mb_row_ei[i].mb.intra_error;
michael@0: 
michael@0:                 for(c_idx = 0; c_idx < MAX_REF_FRAMES; c_idx++)
michael@0:                     cpi->mb.count_mb_ref_frame_usage[c_idx] +=
michael@0:                         cpi->mb_row_ei[i].mb.count_mb_ref_frame_usage[c_idx];
michael@0: 
michael@0:                 for(c_idx = 0; c_idx < MAX_ERROR_BINS; c_idx++)
michael@0:                     cpi->mb.error_bins[c_idx] +=
michael@0:                         cpi->mb_row_ei[i].mb.error_bins[c_idx];
michael@0: 
michael@0:                 /* add up counts for each thread */
michael@0:                 sum_coef_counts(x, &cpi->mb_row_ei[i].mb);
michael@0:             }
michael@0: 
michael@0:         }
michael@0:         else
michael@0: #endif
michael@0:         {
michael@0: 
michael@0:             /* for each macroblock row in image */
michael@0:             for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
michael@0:             {
michael@0:                 vp8_zero(cm->left_context)
michael@0: 
michael@0: #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
michael@0:                 tp = cpi->tok;
michael@0: #endif
michael@0: 
michael@0:                 encode_mb_row(cpi, cm, mb_row, x, xd, &tp, segment_counts, &totalrate);
michael@0: 
michael@0:                 /* adjust to the next row of mbs */
michael@0:                 x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;
michael@0:                 x->src.u_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
michael@0:                 x->src.v_buffer += 8 * x->src.uv_stride - 8 * cm->mb_cols;
michael@0:             }
michael@0: 
michael@0:             cpi->tok_count = (unsigned int)(tp - cpi->tok);
michael@0:         }
michael@0: 
michael@0: #if CONFIG_REALTIME_ONLY & CONFIG_ONTHEFLY_BITPACKING
michael@0:         {
michael@0:             int i;
michael@0:             for(i = 0; i < num_part; i++)
michael@0:             {
michael@0:                 vp8_stop_encode(&bc[i]);
michael@0:                 cpi->partition_sz[i+1] = bc[i].pos;
michael@0:             }
michael@0:         }
michael@0: #endif
michael@0: 
michael@0:         vpx_usec_timer_mark(&emr_timer);
michael@0:         cpi->time_encode_mb_row += vpx_usec_timer_elapsed(&emr_timer);
michael@0:     }
michael@0: 
michael@0: 
michael@0:     // Work out the segment probabilities if segmentation is enabled
michael@0:     // and needs to be updated
michael@0:     if (xd->segmentation_enabled && xd->update_mb_segmentation_map)
michael@0:     {
michael@0:         int tot_count;
michael@0:         int i;
michael@0: 
michael@0:         /* Set to defaults */
michael@0:         vpx_memset(xd->mb_segment_tree_probs, 255 , sizeof(xd->mb_segment_tree_probs));
michael@0: 
michael@0:         tot_count = segment_counts[0] + segment_counts[1] + segment_counts[2] + segment_counts[3];
michael@0: 
michael@0:         if (tot_count)
michael@0:         {
michael@0:             xd->mb_segment_tree_probs[0] = ((segment_counts[0] + segment_counts[1]) * 255) / tot_count;
michael@0: 
michael@0:             tot_count = segment_counts[0] + segment_counts[1];
michael@0: 
michael@0:             if (tot_count > 0)
michael@0:             {
michael@0:                 xd->mb_segment_tree_probs[1] = (segment_counts[0] * 255) / tot_count;
michael@0:             }
michael@0: 
michael@0:             tot_count = segment_counts[2] + segment_counts[3];
michael@0: 
michael@0:             if (tot_count > 0)
michael@0:                 xd->mb_segment_tree_probs[2] = (segment_counts[2] * 255) / tot_count;
michael@0: 
michael@0:             /* Zero probabilities not allowed */
michael@0:             for (i = 0; i < MB_FEATURE_TREE_PROBS; i ++)
michael@0:             {
michael@0:                 if (xd->mb_segment_tree_probs[i] == 0)
michael@0:                     xd->mb_segment_tree_probs[i] = 1;
michael@0:             }
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     /* projected_frame_size in units of BYTES */
michael@0:     cpi->projected_frame_size = totalrate >> 8;
michael@0: 
michael@0:     /* Make a note of the percentage MBs coded Intra. */
michael@0:     if (cm->frame_type == KEY_FRAME)
michael@0:     {
michael@0:         cpi->this_frame_percent_intra = 100;
michael@0:     }
michael@0:     else
michael@0:     {
michael@0:         int tot_modes;
michael@0: 
michael@0:         tot_modes = cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME]
michael@0:                     + cpi->mb.count_mb_ref_frame_usage[LAST_FRAME]
michael@0:                     + cpi->mb.count_mb_ref_frame_usage[GOLDEN_FRAME]
michael@0:                     + cpi->mb.count_mb_ref_frame_usage[ALTREF_FRAME];
michael@0: 
michael@0:         if (tot_modes)
michael@0:             cpi->this_frame_percent_intra =
michael@0:                 cpi->mb.count_mb_ref_frame_usage[INTRA_FRAME] * 100 / tot_modes;
michael@0: 
michael@0:     }
michael@0: 
michael@0: #if ! CONFIG_REALTIME_ONLY
michael@0:     /* Adjust the projected reference frame usage probability numbers to
michael@0:      * reflect what we have just seen. This may be useful when we make
michael@0:      * multiple iterations of the recode loop rather than continuing to use
michael@0:      * values from the previous frame.
michael@0:      */
michael@0:     if ((cm->frame_type != KEY_FRAME) && ((cpi->oxcf.number_of_layers > 1) ||
michael@0:         (!cm->refresh_alt_ref_frame && !cm->refresh_golden_frame)))
michael@0:     {
michael@0:       vp8_convert_rfct_to_prob(cpi);
michael@0:     }
michael@0: #endif
michael@0: }
michael@0: void vp8_setup_block_ptrs(MACROBLOCK *x)
michael@0: {
michael@0:     int r, c;
michael@0:     int i;
michael@0: 
michael@0:     for (r = 0; r < 4; r++)
michael@0:     {
michael@0:         for (c = 0; c < 4; c++)
michael@0:         {
michael@0:             x->block[r*4+c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     for (r = 0; r < 2; r++)
michael@0:     {
michael@0:         for (c = 0; c < 2; c++)
michael@0:         {
michael@0:             x->block[16 + r*2+c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
michael@0:         }
michael@0:     }
michael@0: 
michael@0: 
michael@0:     for (r = 0; r < 2; r++)
michael@0:     {
michael@0:         for (c = 0; c < 2; c++)
michael@0:         {
michael@0:             x->block[20 + r*2+c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     x->block[24].src_diff = x->src_diff + 384;
michael@0: 
michael@0: 
michael@0:     for (i = 0; i < 25; i++)
michael@0:     {
michael@0:         x->block[i].coeff = x->coeff + i * 16;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void vp8_build_block_offsets(MACROBLOCK *x)
michael@0: {
michael@0:     int block = 0;
michael@0:     int br, bc;
michael@0: 
michael@0:     vp8_build_block_doffsets(&x->e_mbd);
michael@0: 
michael@0:     /* y blocks */
michael@0:     x->thismb_ptr = &x->thismb[0];
michael@0:     for (br = 0; br < 4; br++)
michael@0:     {
michael@0:         for (bc = 0; bc < 4; bc++)
michael@0:         {
michael@0:             BLOCK *this_block = &x->block[block];
michael@0:             this_block->base_src = &x->thismb_ptr;
michael@0:             this_block->src_stride = 16;
michael@0:             this_block->src = 4 * br * 16 + 4 * bc;
michael@0:             ++block;
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     /* u blocks */
michael@0:     for (br = 0; br < 2; br++)
michael@0:     {
michael@0:         for (bc = 0; bc < 2; bc++)
michael@0:         {
michael@0:             BLOCK *this_block = &x->block[block];
michael@0:             this_block->base_src = &x->src.u_buffer;
michael@0:             this_block->src_stride = x->src.uv_stride;
michael@0:             this_block->src = 4 * br * this_block->src_stride + 4 * bc;
michael@0:             ++block;
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     /* v blocks */
michael@0:     for (br = 0; br < 2; br++)
michael@0:     {
michael@0:         for (bc = 0; bc < 2; bc++)
michael@0:         {
michael@0:             BLOCK *this_block = &x->block[block];
michael@0:             this_block->base_src = &x->src.v_buffer;
michael@0:             this_block->src_stride = x->src.uv_stride;
michael@0:             this_block->src = 4 * br * this_block->src_stride + 4 * bc;
michael@0:             ++block;
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x)
michael@0: {
michael@0:     const MACROBLOCKD *xd = & x->e_mbd;
michael@0:     const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
michael@0:     const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;
michael@0: 
michael@0: #ifdef MODE_STATS
michael@0:     const int is_key = cpi->common.frame_type == KEY_FRAME;
michael@0: 
michael@0:     ++ (is_key ? uv_modes : inter_uv_modes)[uvm];
michael@0: 
michael@0:     if (m == B_PRED)
michael@0:     {
michael@0:         unsigned int *const bct = is_key ? b_modes : inter_b_modes;
michael@0: 
michael@0:         int b = 0;
michael@0: 
michael@0:         do
michael@0:         {
michael@0:             ++ bct[xd->block[b].bmi.mode];
michael@0:         }
michael@0:         while (++b < 16);
michael@0:     }
michael@0: 
michael@0: #endif
michael@0: 
michael@0:     ++x->ymode_count[m];
michael@0:     ++x->uv_mode_count[uvm];
michael@0: 
michael@0: }
michael@0: 
michael@0: /* Experimental stub function to create a per MB zbin adjustment based on
michael@0:  * some previously calculated measure of MB activity.
michael@0:  */
michael@0: static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x )
michael@0: {
michael@0: #if USE_ACT_INDEX
michael@0:     x->act_zbin_adj = *(x->mb_activity_ptr);
michael@0: #else
michael@0:     int64_t a;
michael@0:     int64_t b;
michael@0:     int64_t act = *(x->mb_activity_ptr);
michael@0: 
michael@0:     /* Apply the masking to the RD multiplier. */
michael@0:     a = act + 4*cpi->activity_avg;
michael@0:     b = 4*act + cpi->activity_avg;
michael@0: 
michael@0:     if ( act > cpi->activity_avg )
michael@0:         x->act_zbin_adj = (int)(((int64_t)b + (a>>1))/a) - 1;
michael@0:     else
michael@0:         x->act_zbin_adj = 1 - (int)(((int64_t)a + (b>>1))/b);
michael@0: #endif
michael@0: }
michael@0: 
michael@0: int vp8cx_encode_intra_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
michael@0:                                   TOKENEXTRA **t)
michael@0: {
michael@0:     MACROBLOCKD *xd = &x->e_mbd;
michael@0:     int rate;
michael@0: 
michael@0:     if (cpi->sf.RD && cpi->compressor_speed != 2)
michael@0:         vp8_rd_pick_intra_mode(x, &rate);
michael@0:     else
michael@0:         vp8_pick_intra_mode(x, &rate);
michael@0: 
michael@0:     if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
michael@0:     {
michael@0:         adjust_act_zbin( cpi, x );
michael@0:         vp8_update_zbin_extra(cpi, x);
michael@0:     }
michael@0: 
michael@0:     if (x->e_mbd.mode_info_context->mbmi.mode == B_PRED)
michael@0:         vp8_encode_intra4x4mby(x);
michael@0:     else
michael@0:         vp8_encode_intra16x16mby(x);
michael@0: 
michael@0:     vp8_encode_intra16x16mbuv(x);
michael@0: 
michael@0:     sum_intra_stats(cpi, x);
michael@0: 
michael@0:     vp8_tokenize_mb(cpi, x, t);
michael@0: 
michael@0:     if (xd->mode_info_context->mbmi.mode != B_PRED)
michael@0:         vp8_inverse_transform_mby(xd);
michael@0: 
michael@0:     vp8_dequant_idct_add_uv_block
michael@0:                     (xd->qcoeff+16*16, xd->dequant_uv,
michael@0:                      xd->dst.u_buffer, xd->dst.v_buffer,
michael@0:                      xd->dst.uv_stride, xd->eobs+16);
michael@0:     return rate;
michael@0: }
michael@0: #ifdef SPEEDSTATS
michael@0: extern int cnt_pm;
michael@0: #endif
michael@0: 
michael@0: extern void vp8_fix_contexts(MACROBLOCKD *x);
michael@0: 
michael@0: int vp8cx_encode_inter_macroblock
michael@0: (
michael@0:     VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
michael@0:     int recon_yoffset, int recon_uvoffset,
michael@0:     int mb_row, int mb_col
michael@0: )
michael@0: {
michael@0:     MACROBLOCKD *const xd = &x->e_mbd;
michael@0:     int intra_error = 0;
michael@0:     int rate;
michael@0:     int distortion;
michael@0: 
michael@0:     x->skip = 0;
michael@0: 
michael@0:     if (xd->segmentation_enabled)
michael@0:         x->encode_breakout = cpi->segment_encode_breakout[xd->mode_info_context->mbmi.segment_id];
michael@0:     else
michael@0:         x->encode_breakout = cpi->oxcf.encode_breakout;
michael@0: 
michael@0: #if CONFIG_TEMPORAL_DENOISING
michael@0:     /* Reset the best sse mode/mv for each macroblock. */
michael@0:     x->best_reference_frame = INTRA_FRAME;
michael@0:     x->best_zeromv_reference_frame = INTRA_FRAME;
michael@0:     x->best_sse_inter_mode = 0;
michael@0:     x->best_sse_mv.as_int = 0;
michael@0:     x->need_to_clamp_best_mvs = 0;
michael@0: #endif
michael@0: 
michael@0:     if (cpi->sf.RD)
michael@0:     {
michael@0:         int zbin_mode_boost_enabled = x->zbin_mode_boost_enabled;
michael@0: 
michael@0:         /* Are we using the fast quantizer for the mode selection? */
michael@0:         if(cpi->sf.use_fastquant_for_pick)
michael@0:         {
michael@0:             x->quantize_b      = vp8_fast_quantize_b;
michael@0:             x->quantize_b_pair = vp8_fast_quantize_b_pair;
michael@0: 
michael@0:             /* the fast quantizer does not use zbin_extra, so
michael@0:              * do not recalculate */
michael@0:             x->zbin_mode_boost_enabled = 0;
michael@0:         }
michael@0:         vp8_rd_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
michael@0:                                &distortion, &intra_error);
michael@0: 
michael@0:         /* switch back to the regular quantizer for the encode */
michael@0:         if (cpi->sf.improved_quant)
michael@0:         {
michael@0:             x->quantize_b      = vp8_regular_quantize_b;
michael@0:             x->quantize_b_pair = vp8_regular_quantize_b_pair;
michael@0:         }
michael@0: 
michael@0:         /* restore cpi->zbin_mode_boost_enabled */
michael@0:         x->zbin_mode_boost_enabled = zbin_mode_boost_enabled;
michael@0: 
michael@0:     }
michael@0:     else
michael@0:     {
michael@0:         vp8_pick_inter_mode(cpi, x, recon_yoffset, recon_uvoffset, &rate,
michael@0:                             &distortion, &intra_error, mb_row, mb_col);
michael@0:     }
michael@0: 
michael@0:     x->prediction_error += distortion;
michael@0:     x->intra_error += intra_error;
michael@0: 
michael@0:     if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
michael@0:     {
michael@0:         /* Adjust the zbin based on this MB rate. */
michael@0:         adjust_act_zbin( cpi, x );
michael@0:     }
michael@0: 
michael@0: #if 0
michael@0:     /* Experimental RD code */
michael@0:     cpi->frame_distortion += distortion;
michael@0:     cpi->last_mb_distortion = distortion;
michael@0: #endif
michael@0: 
michael@0:     /* MB level adjutment to quantizer setup */
michael@0:     if (xd->segmentation_enabled)
michael@0:     {
michael@0:         /* If cyclic update enabled */
michael@0:         if (cpi->current_layer == 0 && cpi->cyclic_refresh_mode_enabled)
michael@0:         {
michael@0:             /* Clear segment_id back to 0 if not coded (last frame 0,0) */
michael@0:             if ((xd->mode_info_context->mbmi.segment_id == 1) &&
michael@0:                 ((xd->mode_info_context->mbmi.ref_frame != LAST_FRAME) || (xd->mode_info_context->mbmi.mode != ZEROMV)))
michael@0:             {
michael@0:                 xd->mode_info_context->mbmi.segment_id = 0;
michael@0: 
michael@0:                 /* segment_id changed, so update */
michael@0:                 vp8cx_mb_init_quantizer(cpi, x, 1);
michael@0:             }
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     {
michael@0:         /* Experimental code.
michael@0:          * Special case for gf and arf zeromv modes, for 1 temporal layer.
michael@0:          * Increase zbin size to supress noise.
michael@0:          */
michael@0:         x->zbin_mode_boost = 0;
michael@0:         if (x->zbin_mode_boost_enabled)
michael@0:         {
michael@0:             if ( xd->mode_info_context->mbmi.ref_frame != INTRA_FRAME )
michael@0:             {
michael@0:                 if (xd->mode_info_context->mbmi.mode == ZEROMV)
michael@0:                 {
michael@0:                     if (xd->mode_info_context->mbmi.ref_frame != LAST_FRAME &&
michael@0:                         cpi->oxcf.number_of_layers == 1)
michael@0:                         x->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
michael@0:                     else
michael@0:                         x->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
michael@0:                 }
michael@0:                 else if (xd->mode_info_context->mbmi.mode == SPLITMV)
michael@0:                     x->zbin_mode_boost = 0;
michael@0:                 else
michael@0:                     x->zbin_mode_boost = MV_ZBIN_BOOST;
michael@0:             }
michael@0:         }
michael@0: 
michael@0:         /* The fast quantizer doesn't use zbin_extra, only do so with
michael@0:          * the regular quantizer. */
michael@0:         if (cpi->sf.improved_quant)
michael@0:             vp8_update_zbin_extra(cpi, x);
michael@0:     }
michael@0: 
michael@0:     x->count_mb_ref_frame_usage[xd->mode_info_context->mbmi.ref_frame] ++;
michael@0: 
michael@0:     if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME)
michael@0:     {
michael@0:         vp8_encode_intra16x16mbuv(x);
michael@0: 
michael@0:         if (xd->mode_info_context->mbmi.mode == B_PRED)
michael@0:         {
michael@0:             vp8_encode_intra4x4mby(x);
michael@0:         }
michael@0:         else
michael@0:         {
michael@0:             vp8_encode_intra16x16mby(x);
michael@0:         }
michael@0: 
michael@0:         sum_intra_stats(cpi, x);
michael@0:     }
michael@0:     else
michael@0:     {
michael@0:         int ref_fb_idx;
michael@0: 
michael@0:         if (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME)
michael@0:             ref_fb_idx = cpi->common.lst_fb_idx;
michael@0:         else if (xd->mode_info_context->mbmi.ref_frame == GOLDEN_FRAME)
michael@0:             ref_fb_idx = cpi->common.gld_fb_idx;
michael@0:         else
michael@0:             ref_fb_idx = cpi->common.alt_fb_idx;
michael@0: 
michael@0:         xd->pre.y_buffer = cpi->common.yv12_fb[ref_fb_idx].y_buffer + recon_yoffset;
michael@0:         xd->pre.u_buffer = cpi->common.yv12_fb[ref_fb_idx].u_buffer + recon_uvoffset;
michael@0:         xd->pre.v_buffer = cpi->common.yv12_fb[ref_fb_idx].v_buffer + recon_uvoffset;
michael@0: 
michael@0:         if (!x->skip)
michael@0:         {
michael@0:             vp8_encode_inter16x16(x);
michael@0:         }
michael@0:         else
michael@0:             vp8_build_inter16x16_predictors_mb(xd, xd->dst.y_buffer,
michael@0:                                            xd->dst.u_buffer, xd->dst.v_buffer,
michael@0:                                            xd->dst.y_stride, xd->dst.uv_stride);
michael@0: 
michael@0:     }
michael@0: 
michael@0:     if (!x->skip)
michael@0:     {
michael@0:         vp8_tokenize_mb(cpi, x, t);
michael@0: 
michael@0:         if (xd->mode_info_context->mbmi.mode != B_PRED)
michael@0:             vp8_inverse_transform_mby(xd);
michael@0: 
michael@0:         vp8_dequant_idct_add_uv_block
michael@0:                         (xd->qcoeff+16*16, xd->dequant_uv,
michael@0:                          xd->dst.u_buffer, xd->dst.v_buffer,
michael@0:                          xd->dst.uv_stride, xd->eobs+16);
michael@0:     }
michael@0:     else
michael@0:     {
michael@0:         /* always set mb_skip_coeff as it is needed by the loopfilter */
michael@0:         xd->mode_info_context->mbmi.mb_skip_coeff = 1;
michael@0: 
michael@0:         if (cpi->common.mb_no_coeff_skip)
michael@0:         {
michael@0:             x->skip_true_count ++;
michael@0:             vp8_fix_contexts(xd);
michael@0:         }
michael@0:         else
michael@0:         {
michael@0:             vp8_stuff_mb(cpi, x, t);
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     return rate;
michael@0: }