encodeframe.c

/*
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */


#include "vpx_ports/config.h"
#include "encodemb.h"
#include "encodemv.h"
#include "vp8/common/common.h"
#include "onyx_int.h"
#include "vp8/common/extend.h"
#include "vp8/common/entropymode.h"
#include "vp8/common/quant_common.h"
#include "segmentation.h"
#include "vp8/common/setupintrarecon.h"
#include "encodeintra.h"
#include "vp8/common/reconinter.h"
#include "rdopt.h"
#include "pickinter.h"
#include "vp8/common/findnearmv.h"
#include "vp8/common/reconintra.h"
#include "vp8/common/seg_common.h"
#include <stdio.h>
#include <math.h>
#include <limits.h>
#include "vp8/common/subpixel.h"
#include "vpx_ports/vpx_timer.h"
#include "vp8/common/pred_common.h"

//#if CONFIG_SEGFEATURES
//#define DBG_PRNT_SEGMAP 1

#if CONFIG_RUNTIME_CPU_DETECT
#define RTCD(x)     &cpi->common.rtcd.x
#define IF_RTCD(x)  (x)
#else
#define RTCD(x)     NULL
#define IF_RTCD(x)  NULL
#endif

#ifdef ENC_DEBUG
int enc_debug=0;
int mb_row_debug, mb_col_debug;
#endif

extern void vp8_stuff_mb(VP8_COMP *cpi, MACROBLOCKD *x, TOKENEXTRA **t) ;

extern void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex);
extern void vp8_auto_select_speed(VP8_COMP *cpi);
extern void vp8cx_init_mbrthread_data(VP8_COMP *cpi,
                                      MACROBLOCK *x,
                                      MB_ROW_COMP *mbr_ei,
                                      int mb_row,
                                      int count);
void vp8_build_block_offsets(MACROBLOCK *x);
void vp8_setup_block_ptrs(MACROBLOCK *x);
int vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t, int recon_yoffset, int recon_uvoffset);
int vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t);
static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x );



#ifdef MODE_STATS
unsigned int inter_y_modes[MB_MODE_COUNT] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
unsigned int inter_uv_modes[VP8_UV_MODES] = {0, 0, 0, 0};
unsigned int inter_b_modes[B_MODE_COUNT] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
unsigned int y_modes[VP8_YMODES] = {0, 0, 0, 0, 0, 0};
unsigned int i8x8_modes[VP8_I8X8_MODES]={0  };
unsigned int uv_modes[VP8_UV_MODES] = {0, 0, 0, 0};
unsigned int uv_modes_y[VP8_YMODES][VP8_UV_MODES]=
{
{0, 0, 0, 0},
{0, 0, 0, 0},
{0, 0, 0, 0},
{0, 0, 0, 0},
{0, 0, 0, 0},
{0, 0, 0, 0}
};
unsigned int b_modes[B_MODE_COUNT] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
#endif


/* activity_avg must be positive, or flat regions could get a zero weight
 *  (infinite lambda), which confounds analysis.
 * This also avoids the need for divide by zero checks in
 *  vp8_activity_masking().
 */
#define VP8_ACTIVITY_AVG_MIN (64)

/* This is used as a reference when computing the source variance for the
 *  purposes of activity masking.
 * Eventually this should be replaced by custom no-reference routines,
 *  which will be faster.
 */
static const unsigned char VP8_VAR_OFFS[16]=
{
    128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128
};



#if CONFIG_T8X8

//INTRA mode transform size
//When all three criteria are off the default is 4x4
//#define INTRA_VARIANCE_ENTROPY_CRITERIA
#define INTRA_WTD_SSE_ENTROPY_CRITERIA
//#define INTRA_TEST_8X8_ONLY
//
//INTER mode transform size
//When all three criteria are off the default is 4x4
//#define INTER_VARIANCE_ENTROPY_CRITERIA
#define INTER_WTD_SSE_ENTROPY_CRITERIA
//#define INTER_TEST_8X8_ONLY

double variance_Block(short *b1, int pitch, int dimension)
{
    short ip[8][8]={{0}};
    short *b = b1;
    int i, j = 0;
    double mean = 0.0, variance = 0.0;
    for (i = 0; i < dimension; i++)
    {
        for (j = 0; j < dimension; j++)
        {
            ip[i][j] = b[j];
            mean += ip[i][j];
        }
        b += pitch;
    }
    mean /= (dimension*dimension);

    for (i = 0; i < dimension; i++)
    {
        for (j = 0; j < dimension; j++)
        {
            variance += (ip[i][j]-mean)*(ip[i][j]-mean);
        }
    }
    variance /= (dimension*dimension);
    return variance;
}

double mean_Block(short *b, int pitch, int dimension)
{
    short ip[8][8]={{0}};
    int i, j = 0;
    double mean = 0;
    for (i = 0; i < dimension; i++)
    {
        for (j = 0; j < dimension; j++)
        {
            ip[i][j] = b[j];
            mean += ip[i][j];
        }
        b += pitch;
    }
    mean /= (dimension*dimension);

    return mean;
}

int SSE_Block(short *b, int pitch, int dimension)
{
    int i, j, sse_block = 0;
    for (i = 0; i < dimension; i++)
    {
        for (j = 0; j < dimension; j++)
        {
            sse_block += b[j]*b[j];
        }
        b += pitch;
    }
   return sse_block;
}

double Compute_Variance_Entropy(MACROBLOCK *x)
{
    double variance_8[4] = {0.0, 0.0, 0.0, 0.0}, sum_var = 0.0, all_entropy = 0.0;
    variance_8[0] = variance_Block(x->block[0].src_diff, 16, 8);
    variance_8[1] = variance_Block(x->block[2].src_diff, 16, 8);
    variance_8[2] = variance_Block(x->block[8].src_diff, 16, 8);
    variance_8[3] = variance_Block(x->block[10].src_diff, 16, 8);
    sum_var = variance_8[0] + variance_8[1] + variance_8[2] + variance_8[3];
    if(sum_var)
    {
      int i;
      for(i = 0; i <4; i++)
      {
        if(variance_8[i])
        {
          variance_8[i] /= sum_var;
          all_entropy -= variance_8[i]*log(variance_8[i]);
        }
      }
    }
    return (all_entropy /log(2));
}

double Compute_Wtd_SSE_SubEntropy(MACROBLOCK *x)
{
    double variance_8[4] = {0.0, 0.0, 0.0, 0.0};
    double entropy_8[4] = {0.0, 0.0, 0.0, 0.0};
    double sse_1, sse_2, sse_3, sse_4, sse_0;
    int i;
    for (i=0;i<3;i+=2)
    {
      sse_0 = SSE_Block(x->block[i].src_diff, 16, 8);
      if(sse_0)
      {
        sse_1 = SSE_Block(x->block[i].src_diff, 16, 4)/sse_0;
        sse_2 = SSE_Block(x->block[i+1].src_diff, 16, 4)/sse_0;
        sse_3 = SSE_Block(x->block[i+4].src_diff, 16, 4)/sse_0;
        sse_4 = SSE_Block(x->block[i+5].src_diff, 16, 4)/sse_0;
        variance_8[i]= variance_Block(x->block[i].src_diff, 16, 8);
        if(sse_1 && sse_2 && sse_3 && sse_4)
        entropy_8[i]= (-sse_1*log(sse_1)
                       -sse_2*log(sse_2)
                       -sse_3*log(sse_3)
                       -sse_4*log(sse_4))/log(2);
      }
    }
    for (i=8;i<11;i+=2)
    {
      if(sse_0)
      {
        sse_0 = SSE_Block(x->block[i].src_diff, 16, 8);
        sse_1 = SSE_Block(x->block[i].src_diff, 16, 4)/sse_0;
        sse_2 = SSE_Block(x->block[i+1].src_diff, 16, 4)/sse_0;
        sse_3 = SSE_Block(x->block[i+4].src_diff, 16, 4)/sse_0;
        sse_4 = SSE_Block(x->block[i+5].src_diff, 16, 4)/sse_0;
        variance_8[i-7]= variance_Block(x->block[i].src_diff, 16, 8);
        if(sse_1 && sse_2 && sse_3 && sse_4)
        entropy_8[i-7]= (-sse_1*log(sse_1)
                         -sse_2*log(sse_2)
                         -sse_3*log(sse_3)
                         -sse_4*log(sse_4))/log(2);
      }
    }

    if(variance_8[0]+variance_8[1]+variance_8[2]+variance_8[3])
      return (entropy_8[0]*variance_8[0]+
              entropy_8[1]*variance_8[1]+
              entropy_8[2]*variance_8[2]+
              entropy_8[3]*variance_8[3])/
             (variance_8[0]+
              variance_8[1]+
              variance_8[2]+
              variance_8[3]);
    else
      return 0;
}

int vp8_8x8_selection_intra(MACROBLOCK *x)
{
#ifdef INTRA_VARIANCE_ENTROPY_CRITERIA
    return (Compute_Variance_Entropy(x) > 1.2);
#elif defined(INTRA_WTD_SSE_ENTROPY_CRITERIA)
    return (Compute_Wtd_SSE_SubEntropy(x) > 1.2);
#elif defined(INTRA_TEST_8X8_ONLY)
    return 1;
#else
    return 0; //when all criteria are off use the default 4x4 only
#endif
}

int vp8_8x8_selection_inter(MACROBLOCK *x)
{
#ifdef INTER_VARIANCE_ENTROPY_CRITERIA
    return (Compute_Variance_Entropy(x) > 1.5);
#elif defined(INTER_WTD_SSE_ENTROPY_CRITERIA)
    return (Compute_Wtd_SSE_SubEntropy(x) > 1.5);
#elif defined(INTER_TEST_8X8_ONLY)
    return 1;
#else
    return 0; //when all criteria are off use the default 4x4 only
#endif
}

#endif

// Original activity measure from Tim T's code.
static unsigned int tt_activity_measure( VP8_COMP *cpi, MACROBLOCK *x )
{
    unsigned int act;
    unsigned int sse;
    /* TODO: This could also be done over smaller areas (8x8), but that would
     *  require extensive changes elsewhere, as lambda is assumed to be fixed
     *  over an entire MB in most of the code.
     * Another option is to compute four 8x8 variances, and pick a single
     *  lambda using a non-linear combination (e.g., the smallest, or second
     *  smallest, etc.).
     */
    act =     VARIANCE_INVOKE(&cpi->rtcd.variance, var16x16)(x->src.y_buffer,
                    x->src.y_stride, VP8_VAR_OFFS, 0, &sse);
    act = act<<4;

    /* If the region is flat, lower the activity some more. */
    if (act < 8<<12)
        act = act < 5<<12 ? act : 5<<12;

    return act;
}

// Stub for alternative experimental activity measures.
static unsigned int alt_activity_measure( VP8_COMP *cpi,
                                          MACROBLOCK *x, int use_dc_pred )
{
    return vp8_encode_intra(cpi,x, use_dc_pred);
}


// Measure the activity of the current macroblock
// What we measure here is TBD so abstracted to this function
#define ALT_ACT_MEASURE 1
static unsigned int mb_activity_measure( VP8_COMP *cpi, MACROBLOCK *x,
                                  int mb_row, int mb_col)
{
    unsigned int mb_activity;

    if  ( ALT_ACT_MEASURE )
    {
        int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);

        // Or use and alternative.
        mb_activity = alt_activity_measure( cpi, x, use_dc_pred );
    }
    else
    {
        // Original activity measure from Tim T's code.
        mb_activity = tt_activity_measure( cpi, x );
    }

    if ( mb_activity < VP8_ACTIVITY_AVG_MIN )
        mb_activity = VP8_ACTIVITY_AVG_MIN;

    return mb_activity;
}

// Calculate an "average" mb activity value for the frame
#define ACT_MEDIAN 0
static void calc_av_activity( VP8_COMP *cpi, int64_t activity_sum )
{
#if ACT_MEDIAN
    // Find median: Simple n^2 algorithm for experimentation
    {
        unsigned int median;
        unsigned int i,j;
        unsigned int * sortlist;
        unsigned int tmp;

        // Create a list to sort to
        CHECK_MEM_ERROR(sortlist,
                        vpx_calloc(sizeof(unsigned int),
                        cpi->common.MBs));

        // Copy map to sort list
        vpx_memcpy( sortlist, cpi->mb_activity_map,
                    sizeof(unsigned int) * cpi->common.MBs );


        // Ripple each value down to its correct position
        for ( i = 1; i < cpi->common.MBs; i ++ )
        {
            for ( j = i; j > 0; j -- )
            {
                if ( sortlist[j] < sortlist[j-1] )
                {
                    // Swap values
                    tmp = sortlist[j-1];
                    sortlist[j-1] = sortlist[j];
                    sortlist[j] = tmp;
                }
                else
                    break;
            }
        }

        // Even number MBs so estimate median as mean of two either side.
        median = ( 1 + sortlist[cpi->common.MBs >> 1] +
                   sortlist[(cpi->common.MBs >> 1) + 1] ) >> 1;

        cpi->activity_avg = median;

        vpx_free(sortlist);
    }
#else
    // Simple mean for now
    cpi->activity_avg = (unsigned int)(activity_sum/cpi->common.MBs);
#endif

    if (cpi->activity_avg < VP8_ACTIVITY_AVG_MIN)
        cpi->activity_avg = VP8_ACTIVITY_AVG_MIN;

    // Experimental code: return fixed value normalized for several clips
    if  ( ALT_ACT_MEASURE )
        cpi->activity_avg = 100000;
}

#define USE_ACT_INDEX   0
#define OUTPUT_NORM_ACT_STATS   0

#if USE_ACT_INDEX
// Calculate and activity index for each mb
static void calc_activity_index( VP8_COMP *cpi, MACROBLOCK *x )
{
    VP8_COMMON *const cm = & cpi->common;
    int mb_row, mb_col;

    int64_t act;
    int64_t a;
    int64_t b;

#if OUTPUT_NORM_ACT_STATS
    FILE *f = fopen("norm_act.stt", "a");
    fprintf(f, "\n%12d\n", cpi->activity_avg );
#endif

    // Reset pointers to start of activity map
    x->mb_activity_ptr = cpi->mb_activity_map;

    // Calculate normalized mb activity number.
    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
    {
        // for each macroblock col in image
        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
        {
            // Read activity from the map
            act = *(x->mb_activity_ptr);

            // Calculate a normalized activity number
            a = act + 4*cpi->activity_avg;
            b = 4*act + cpi->activity_avg;

            if ( b >= a )
                *(x->activity_ptr) = (int)((b + (a>>1))/a) - 1;
            else
                *(x->activity_ptr) = 1 - (int)((a + (b>>1))/b);

#if OUTPUT_NORM_ACT_STATS
            fprintf(f, " %6d", *(x->mb_activity_ptr));
#endif
            // Increment activity map pointers
            x->mb_activity_ptr++;
        }

#if OUTPUT_NORM_ACT_STATS
        fprintf(f, "\n");
#endif

    }

#if OUTPUT_NORM_ACT_STATS
    fclose(f);
#endif

}
#endif

// Loop through all MBs. Note activity of each, average activity and
// calculate a normalized activity for each
static void build_activity_map( VP8_COMP *cpi )
{
    MACROBLOCK *const x = & cpi->mb;
    MACROBLOCKD *xd = &x->e_mbd;
    VP8_COMMON *const cm = & cpi->common;

#if ALT_ACT_MEASURE
    YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
    int recon_yoffset;
    int recon_y_stride = new_yv12->y_stride;
#endif

    int mb_row, mb_col;
    unsigned int mb_activity;
    int64_t activity_sum = 0;

    // for each macroblock row in image
    for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
    {
#if ALT_ACT_MEASURE
        // reset above block coeffs
        xd->up_available = (mb_row != 0);
        recon_yoffset = (mb_row * recon_y_stride * 16);
#endif
        // for each macroblock col in image
        for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
        {
#if ALT_ACT_MEASURE
            xd->dst.y_buffer = new_yv12->y_buffer + recon_yoffset;
            xd->left_available = (mb_col != 0);
            recon_yoffset += 16;
#endif
            //Copy current mb to a buffer
            RECON_INVOKE(&xd->rtcd->recon, copy16x16)(x->src.y_buffer, x->src.y_stride, x->thismb, 16);

            // measure activity
            mb_activity = mb_activity_measure( cpi, x, mb_row, mb_col );

            // Keep frame sum
            activity_sum += mb_activity;

            // Store MB level activity details.
            *x->mb_activity_ptr = mb_activity;

            // Increment activity map pointer
            x->mb_activity_ptr++;

            // adjust to the next column of source macroblocks
            x->src.y_buffer += 16;
        }


        // adjust to the next row of mbs
        x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;

#if ALT_ACT_MEASURE
        //extend the recon for intra prediction
        vp8_extend_mb_row(new_yv12, xd->dst.y_buffer + 16,
                          xd->dst.u_buffer + 8, xd->dst.v_buffer + 8);
#endif

    }

    // Calculate an "average" MB activity
    calc_av_activity(cpi, activity_sum);

#if USE_ACT_INDEX
    // Calculate an activity index number of each mb
    calc_activity_index( cpi, x );
#endif

}

// Macroblock activity masking
void vp8_activity_masking(VP8_COMP *cpi, MACROBLOCK *x)
{
#if USE_ACT_INDEX
    x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
    x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
    x->errorperbit += (x->errorperbit==0);
#else
    int64_t a;
    int64_t b;
    int64_t act = *(x->mb_activity_ptr);

    // Apply the masking to the RD multiplier.
    a = act + (2*cpi->activity_avg);
    b = (2*act) + cpi->activity_avg;

    x->rdmult = (unsigned int)(((int64_t)x->rdmult*b + (a>>1))/a);
    x->errorperbit = x->rdmult * 100 /(110 * x->rddiv);
    x->errorperbit += (x->errorperbit==0);
#endif

    // Activity based Zbin adjustment
    adjust_act_zbin(cpi, x);
}

#if CONFIG_SUPERBLOCKS
static
void encode_sb_row (VP8_COMP *cpi,
                   VP8_COMMON *cm,
                   int mbrow,
                   MACROBLOCK  *x,
                   MACROBLOCKD *xd,
                   TOKENEXTRA **tp,
                   int *totalrate)
{
    int i;
    int map_index;
    int mb_row, mb_col;
    int recon_yoffset, recon_uvoffset;
    int ref_fb_idx = cm->lst_fb_idx;
    int dst_fb_idx = cm->new_fb_idx;
    int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
    int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
    int row_delta[4] = {-1,  0, +1,  0};
    int col_delta[4] = {+1, +1, -1, +1};
    int sb_cols = (cm->mb_cols + 1)>>1;
    int sb_col;
    ENTROPY_CONTEXT_PLANES left_context[2];

    vpx_memset (left_context, 0, sizeof(left_context));

    // TODO put NULL into MB rows that have no tokens?
    cpi->tplist[mbrow].start = *tp;

    x->src.y_buffer -= 16 * (col_delta[0] + row_delta[0]*x->src.y_stride);
    x->src.u_buffer -= 8  * (col_delta[0] + row_delta[0]*x->src.uv_stride);
    x->src.v_buffer -= 8  * (col_delta[0] + row_delta[0]*x->src.uv_stride);
    mb_row = mbrow - row_delta[0];
    mb_col = 0     - col_delta[0];

    for (sb_col=0; sb_col<sb_cols; sb_col++)
   {
        /* Encode MBs within the SB in raster order */
        for ( i=0; i<4; i++ )
        {
            int offset_extended = row_delta[(i+1) & 0x3] *
                                  xd->mode_info_stride + col_delta[(i+1) & 0x3];
            int offset_unextended = row_delta[(i+1) & 0x3] *
                                    cm->mb_cols + col_delta[(i+1) & 0x3];
           int dy = row_delta[i];
            int dx = col_delta[i];

            mb_row += dy;
            mb_col += dx;

            x->src.y_buffer += 16 * (dx + dy*x->src.y_stride);
            x->src.u_buffer += 8  * (dx + dy*x->src.uv_stride);
            x->src.v_buffer += 8  * (dx + dy*x->src.uv_stride);

            if ((mb_row >= cm->mb_rows) || (mb_col >= cm->mb_cols))
            {
                // Skip on to the next MB
                x->gf_active_ptr      += offset_unextended;
                x->partition_info     += offset_extended;
                xd->mode_info_context += offset_extended;

                xd->prev_mode_info_context += offset_extended;

                assert((xd->prev_mode_info_context - cpi->common.prev_mip)
                    ==(xd->mode_info_context - cpi->common.mip));
                continue;
            }

            // Copy in the appropriate left context
            vpx_memcpy (&cm->left_context,
                        &left_context[(i>>1) & 0x1],
                        sizeof(ENTROPY_CONTEXT_PLANES));

            map_index = (mb_row * cpi->common.mb_cols) + mb_col;
            x->mb_activity_ptr = &cpi->mb_activity_map[map_index];

            // reset above block coeffs
            xd->above_context = cm->above_context + mb_col;

            // Distance of Mb to the top & bottom edges, specified in 1/8th pel
            // units as they are always compared to values in 1/8th pel units
            xd->mb_to_top_edge = -((mb_row * 16) << 3);
            xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;

            // Set up limit values for motion vector components
            // to prevent them extending beyond the UMV borders
            x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
            x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
                                + (VP8BORDERINPIXELS - 16);
            x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
            x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16)
                                + (VP8BORDERINPIXELS - 16);

            // Distance of Mb to the left & right edges, specified in
            // 1/8th pel units as they are always compared to values
            // that are in 1/8th pel units
            xd->mb_to_left_edge = -((mb_col * 16) << 3);
            xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;

            xd->up_available = (mb_row != 0);
            xd->left_available = (mb_col != 0);

            recon_yoffset = (mb_row * recon_y_stride * 16) + (mb_col * 16);
            recon_uvoffset = (mb_row * recon_uv_stride * 8) + (mb_col * 8);

            xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
            xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
            xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;

            x->rddiv = cpi->RDDIV;
            x->rdmult = cpi->RDMULT;

            // Copy current mb to a buffer
            RECON_INVOKE(&xd->rtcd->recon, copy16x16)(x->src.y_buffer,
                                                      x->src.y_stride,
                                                      x->thismb, 16);

            if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
                vp8_activity_masking(cpi, x);

            // Is segmentation enabled
            if (xd->segmentation_enabled)
            {
                // Code to set segment id in xd->mbmi.segment_id
                if (cpi->segmentation_map[map_index] <= 3)
                    xd->mode_info_context->mbmi.segment_id =
                                  cpi->segmentation_map[map_index];
                else
                    xd->mode_info_context->mbmi.segment_id = 0;

                vp8cx_mb_init_quantizer(cpi, x);
            }
            else
                // Set to Segment 0 by default
                xd->mode_info_context->mbmi.segment_id = 0;

            x->active_ptr = cpi->active_map + map_index;

            if (cm->frame_type == KEY_FRAME)
            {
                *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
                //Note the encoder may have changed the segment_id

#ifdef MODE_STATS
                y_modes[xd->mode_info_context->mbmi.mode] ++;
#endif
            }
            else
            {
                *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp,
                                             recon_yoffset, recon_uvoffset);
                //Note the encoder may have changed the segment_id

#ifdef MODE_STATS
                inter_y_modes[xd->mode_info_context->mbmi.mode] ++;

                if (xd->mode_info_context->mbmi.mode == SPLITMV)
                {
                    int b;

                    for (b = 0; b < x->partition_info->count; b++)
                    {
                        inter_b_modes[x->partition_info->bmi[b].mode] ++;
                   }
                }

#endif

                // Count of last ref frame 0,0 usage
                if ((xd->mode_info_context->mbmi.mode == ZEROMV) &&
                    (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
                    cpi->inter_zz_count ++;

                // Actions required if segmentation enabled
                if ( xd->segmentation_enabled )
                {
                    // Special case code for cyclic refresh
                    // If cyclic update enabled then copy xd->mbmi.segment_id;
                    // (which may have been updated based on mode during
                    // vp8cx_encode_inter_macroblock()) back into the global
                    // segmentation map
                    if (cpi->cyclic_refresh_mode_enabled)
                    {
                        cpi->segmentation_map[map_index] =
                            xd->mode_info_context->mbmi.segment_id;

                        // If the block has been refreshed mark it as clean (the
                        // magnitude of the -ve influences how long it will be
                        // before we consider another refresh):
                        // Else if it was coded (last frame 0,0) and has not
                        // already been refreshed then mark it as a candidate
                        // for cleanup next time (marked 0)
                        // else mark it as dirty (1).
                        if (xd->mode_info_context->mbmi.segment_id)
                            cpi->cyclic_refresh_map[map_index] = -1;

                        else if ((xd->mode_info_context->mbmi.mode == ZEROMV) &&
                                 (xd->mode_info_context->mbmi.ref_frame ==
                                  LAST_FRAME))
                        {
                            if (cpi->cyclic_refresh_map[map_index] == 1)
                                cpi->cyclic_refresh_map[map_index] = 0;
                        }
                        else
                            cpi->cyclic_refresh_map[map_index] = 1;
                    }
                }
            }

            // TODO Make sure partitioning works with this new scheme
            cpi->tplist[mbrow].stop = *tp;

            // Copy back updated left context
            vpx_memcpy (&left_context[(i>>1) & 0x1],
                        &cm->left_context,
                        sizeof(ENTROPY_CONTEXT_PLANES));

            // skip to next mb
            x->gf_active_ptr      += offset_unextended;
            x->partition_info     += offset_extended;
            xd->mode_info_context += offset_extended;

            xd->prev_mode_info_context += offset_extended;

            assert((xd->prev_mode_info_context - cpi->common.prev_mip)
                ==(xd->mode_info_context - cpi->common.mip));
        }
    }

    // Intra-pred modes requiring top-right data have been disabled,
    // so we don't need this:
    // extend the recon for intra prediction
    /*vp8_extend_mb_row(
        &cm->yv12_fb[dst_fb_idx],
        xd->dst.y_buffer + 16,
        xd->dst.u_buffer + 8,
        xd->dst.v_buffer + 8);*/

    // this is to account for the border
    xd->prev_mode_info_context += 1 - (cm->mb_cols & 0x1) + xd->mode_info_stride;
    xd->mode_info_context += 1 - (cm->mb_cols & 0x1) + xd->mode_info_stride;
    x->partition_info     += 1 - (cm->mb_cols & 0x1) + xd->mode_info_stride;
    x->gf_active_ptr      += cm->mb_cols - (cm->mb_cols & 0x1);

//#if CONFIG_SEGFEATURES
// debug output
#if DBG_PRNT_SEGMAP
    {
        FILE *statsfile;
        statsfile = fopen("segmap2.stt", "a");
        fprintf(statsfile, "\n" );
        fclose(statsfile);
    }
#endif
}
#else
static
void encode_mb_row(VP8_COMP *cpi,
                   VP8_COMMON *cm,
                   int mb_row,
                   MACROBLOCK  *x,
                   MACROBLOCKD *xd,
                   TOKENEXTRA **tp,
                   int *totalrate)
{
    int recon_yoffset, recon_uvoffset;
    int mb_col;
    int ref_fb_idx = cm->lst_fb_idx;
    int dst_fb_idx = cm->new_fb_idx;
    int recon_y_stride = cm->yv12_fb[ref_fb_idx].y_stride;
    int recon_uv_stride = cm->yv12_fb[ref_fb_idx].uv_stride;
    int map_index = (mb_row * cpi->common.mb_cols);

#if CONFIG_MULTITHREAD
    const int nsync = cpi->mt_sync_range;
    const int rightmost_col = cm->mb_cols - 1;
    volatile const int *last_row_current_mb_col;

    if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
        last_row_current_mb_col = &cpi->mt_current_mb_col[mb_row - 1];
    else
        last_row_current_mb_col = &rightmost_col;
#endif
    // Reset the left context
    vp8_zero(cm->left_context)

    // reset above block coeffs
    xd->above_context = cm->above_context;

    xd->up_available = (mb_row != 0);
    recon_yoffset = (mb_row * recon_y_stride * 16);
    recon_uvoffset = (mb_row * recon_uv_stride * 8);

    cpi->tplist[mb_row].start = *tp;
    //printf("Main mb_row = %d\n", mb_row);

    // Distance of Mb to the top & bottom edges, specified in 1/8th pel
    // units as they are always compared to values that are in 1/8th pel units
    xd->mb_to_top_edge = -((mb_row * 16) << 3);
    xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;

    // Set up limit values for vertical motion vector components
    // to prevent them extending beyond the UMV borders
    x->mv_row_min = -((mb_row * 16) + (VP8BORDERINPIXELS - 16));
    x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
                        + (VP8BORDERINPIXELS - 16);

    // Set the mb activity pointer to the start of the row.
    x->mb_activity_ptr = &cpi->mb_activity_map[map_index];

    // for each macroblock col in image
    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
    {
#ifdef ENC_DEBUG
        enc_debug = (cpi->common.current_video_frame ==1 && mb_row==4 && mb_col==0);
        mb_col_debug=mb_col;
        mb_row_debug=mb_row;
#endif
        // Distance of Mb to the left & right edges, specified in
        // 1/8th pel units as they are always compared to values
        // that are in 1/8th pel units
        xd->mb_to_left_edge = -((mb_col * 16) << 3);
        xd->mb_to_right_edge = ((cm->mb_cols - 1 - mb_col) * 16) << 3;

        // Set up limit values for horizontal motion vector components
        // to prevent them extending beyond the UMV borders
        x->mv_col_min = -((mb_col * 16) + (VP8BORDERINPIXELS - 16));
        x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16)
                            + (VP8BORDERINPIXELS - 16);

        xd->dst.y_buffer = cm->yv12_fb[dst_fb_idx].y_buffer + recon_yoffset;
        xd->dst.u_buffer = cm->yv12_fb[dst_fb_idx].u_buffer + recon_uvoffset;
        xd->dst.v_buffer = cm->yv12_fb[dst_fb_idx].v_buffer + recon_uvoffset;
        xd->left_available = (mb_col != 0);

        x->rddiv = cpi->RDDIV;
        x->rdmult = cpi->RDMULT;

        //Copy current mb to a buffer
        RECON_INVOKE(&xd->rtcd->recon, copy16x16)(x->src.y_buffer, x->src.y_stride, x->thismb, 16);

#if CONFIG_MULTITHREAD
        if ((cpi->b_multi_threaded != 0) && (mb_row != 0))
        {
            if ((mb_col & (nsync - 1)) == 0)
            {
                while (mb_col > (*last_row_current_mb_col - nsync)
                        && (*last_row_current_mb_col) != (cm->mb_cols - 1))
                {
                    x86_pause_hint();
                    thread_sleep(0);
                }
            }
        }
#endif

        if(cpi->oxcf.tuning == VP8_TUNE_SSIM)
            vp8_activity_masking(cpi, x);

        // Is segmentation enabled
        if (xd->segmentation_enabled)
        {
            // Code to set segment id in xd->mbmi.segment_id
            if (cpi->segmentation_map[map_index+mb_col] <= 3)
                xd->mode_info_context->mbmi.segment_id = cpi->segmentation_map[map_index+mb_col];
            else
                xd->mode_info_context->mbmi.segment_id = 0;

            vp8cx_mb_init_quantizer(cpi, x);
        }
        else
            // Set to Segment 0 by default
            xd->mode_info_context->mbmi.segment_id = 0;

        x->active_ptr = cpi->active_map + map_index + mb_col;

        if (cm->frame_type == KEY_FRAME)
        {
            *totalrate += vp8cx_encode_intra_macro_block(cpi, x, tp);
            //Note the encoder may have changed the segment_id

#ifdef MODE_STATS
            y_modes[xd->mode_info_context->mbmi.mode] ++;
#endif
        }
        else
        {
            *totalrate += vp8cx_encode_inter_macroblock(cpi, x, tp, recon_yoffset, recon_uvoffset);
            //Note the encoder may have changed the segment_id

#ifdef MODE_STATS
            inter_y_modes[xd->mode_info_context->mbmi.mode] ++;

            if (xd->mode_info_context->mbmi.mode == SPLITMV)
            {
                int b;

                for (b = 0; b < x->partition_info->count; b++)
                {
                    inter_b_modes[x->partition_info->bmi[b].mode] ++;
                }
            }

#endif

            // Count of last ref frame 0,0 usage
            if ((xd->mode_info_context->mbmi.mode == ZEROMV) && (xd->mode_info_context->mbmi.ref_frame == LAST_FRAME))
                cpi->inter_zz_count ++;

            // Actions required if segmentation enabled
            if ( xd->segmentation_enabled )
            {
                // Special case code for cyclic refresh
                // If cyclic update enabled then copy xd->mbmi.segment_id;
                // (which may have been updated based on mode during
                // vp8cx_encode_inter_macroblock()) back into the global
                // segmentation map
                if (cpi->cyclic_refresh_mode_enabled)
                {
                    cpi->segmentation_map[map_index+mb_col] =
                        xd->mode_info_context->mbmi.segment_id;

                    // If the block has been refreshed mark it as clean (the
                    // magnitude of the -ve influences how long it will be
                    // before we consider another refresh):
                    // Else if it was coded (last frame 0,0) and has not
                    // already been refreshed then mark it as a candidate
                    // for cleanup next time (marked 0)
                    // else mark it as dirty (1).
                    if (xd->mode_info_context->mbmi.segment_id)
                        cpi->cyclic_refresh_map[map_index+mb_col] = -1;

                    else if ((xd->mode_info_context->mbmi.mode == ZEROMV) &&
                             (xd->mode_info_context->mbmi.ref_frame ==
                              LAST_FRAME))