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 "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"

//#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);
extern int vp8cx_pick_mode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x,
                                            int recon_yoffset,
                                            int recon_uvoffset);
void vp8_build_block_offsets(MACROBLOCK *x);
void vp8_setup_block_ptrs(MACROBLOCK *x);
void vp8cx_encode_inter_macroblock(VP8_COMP *cpi, MACROBLOCK *x, TOKENEXTRA **t,
                                   int recon_yoffset, int recon_uvoffset,
                                   int output_enabled);
void vp8cx_encode_intra_macro_block(VP8_COMP *cpi, MACROBLOCK *x,
                                    TOKENEXTRA **t, int output_enabled);
static void adjust_act_zbin( VP8_COMP *cpi, MACROBLOCK *x );



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


// 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);
}

static void update_state (VP8_COMP *cpi, MACROBLOCK *x, PICK_MODE_CONTEXT *ctx)
{
    int i;
    MACROBLOCKD *xd = &x->e_mbd;
    MODE_INFO *mi = &ctx->mic;
    int mb_mode = mi->mbmi.mode;
    int mb_mode_index = ctx->best_mode_index;

#if CONFIG_DEBUG
    assert (mb_mode < MB_MODE_COUNT);
    assert (mb_mode_index < MAX_MODES);
    assert (mi->mbmi.ref_frame < MAX_REF_FRAMES);
#endif

    // Restore the coding context of the MB to that that was in place
    // when the mode was picked for it
    vpx_memcpy(xd->mode_info_context, mi, sizeof(MODE_INFO));

    if (mb_mode == B_PRED)
    {
        for (i = 0; i < 16; i++)
        {
            xd->block[i].bmi.as_mode = xd->mode_info_context->bmi[i].as_mode;
            assert (xd->block[i].bmi.as_mode.first < MB_MODE_COUNT);
        }
    }
    else if (mb_mode == I8X8_PRED)
    {
        for (i = 0; i < 16; i++)
        {
            xd->block[i].bmi = xd->mode_info_context->bmi[i];
        }
    }
    else if (mb_mode == SPLITMV)
    {
        vpx_memcpy(x->partition_info, &ctx->partition_info,
                   sizeof(PARTITION_INFO));

        xd->mode_info_context->mbmi.mv.as_int =
                                      x->partition_info->bmi[15].mv.as_int;
        xd->mode_info_context->mbmi.second_mv.as_int =
                                        x->partition_info->bmi[15].second_mv.as_int;
    }

    if (cpi->common.frame_type == KEY_FRAME)
    {
        // Restore the coding modes to that held in the coding context
        //if (mb_mode == B_PRED)
        //    for (i = 0; i < 16; i++)
        //    {
        //        xd->block[i].bmi.as_mode =
        //                          xd->mode_info_context->bmi[i].as_mode;
        //        assert(xd->mode_info_context->bmi[i].as_mode < MB_MODE_COUNT);
        //    }
    }
    else
    {
/*
        // Reduce the activation RD thresholds for the best choice mode
        if ((cpi->rd_baseline_thresh[mb_mode_index] > 0) &&
            (cpi->rd_baseline_thresh[mb_mode_index] < (INT_MAX >> 2)))
        {
            int best_adjustment = (cpi->rd_thresh_mult[mb_mode_index] >> 2);

            cpi->rd_thresh_mult[mb_mode_index] =
                    (cpi->rd_thresh_mult[mb_mode_index]
                     >= (MIN_THRESHMULT + best_adjustment)) ?
                            cpi->rd_thresh_mult[mb_mode_index] - best_adjustment :
                            MIN_THRESHMULT;
            cpi->rd_threshes[mb_mode_index] =
                    (cpi->rd_baseline_thresh[mb_mode_index] >> 7)
                    * cpi->rd_thresh_mult[mb_mode_index];

        }
*/
        // Note how often each mode chosen as best
        cpi->mode_chosen_counts[mb_mode_index]++;

        rd_update_mvcount(cpi, x, &ctx->best_ref_mv, &ctx->second_best_ref_mv);

        cpi->prediction_error += ctx->distortion;
        cpi->intra_error += ctx->intra_error;
    }
}

static void pick_mb_modes (VP8_COMP *cpi,
                           VP8_COMMON *cm,
                           int mb_row,
                           int mb_col,
                           MACROBLOCK  *x,
                           MACROBLOCKD *xd,
                           TOKENEXTRA **tp,
                           int *totalrate)
{
    int i;
    int map_index;
    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;
    ENTROPY_CONTEXT_PLANES left_context[2];
    ENTROPY_CONTEXT_PLANES above_context[2];
    ENTROPY_CONTEXT_PLANES *initial_above_context_ptr = cm->above_context
                                                        + mb_col;

    // Offsets to move pointers from MB to MB within a SB in raster order
    int row_delta[4] = { 0, +1,  0, -1};
    int col_delta[4] = {+1, -1, +1, +1};

    /* Function should not modify L & A contexts; save and restore on exit */
    vpx_memcpy (left_context,
                cpi->left_context,
                sizeof(left_context));
    vpx_memcpy (above_context,
                initial_above_context_ptr,
                sizeof(above_context));

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

        // TODO Many of the index items here can be computed more efficiently!

        if ((mb_row >= cm->mb_rows) || (mb_col >= cm->mb_cols))
        {
            // MB lies outside frame, move on
            mb_row += dy;
            mb_col += dx;

            // Update pointers
            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);

            x->gf_active_ptr += offset_unextended;
            x->partition_info += offset_extended;
            xd->mode_info_context += offset_extended;
            xd->prev_mode_info_context += offset_extended;
#if CONFIG_DEBUG
            assert((xd->prev_mode_info_context - cpi->common.prev_mip) ==
                   (xd->mode_info_context - cpi->common.mip));
#endif
            continue;
        }

        // Index of the MB in the SB 0..3
        xd->mb_index = i;

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

        // set above context pointer
        xd->above_context = cm->above_context + mb_col;

        // Restore the appropriate left context depending on which
        // row in the SB the MB is situated
        vpx_memcpy (&cm->left_context,
                    &cpi->left_context[i>>1],
                    sizeof(ENTROPY_CONTEXT_PLANES));

        // Set up distance of MB to edge of frame in 1/8th pel units
        xd->mb_to_top_edge    = -((mb_row * 16) << 3);
        xd->mb_to_left_edge   = -((mb_col * 16) << 3);
        xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
        xd->mb_to_right_edge  = ((cm->mb_cols - 1 - mb_col) * 16) << 3;

        // Set up limit values for MV components to prevent them from
        // extending beyond the UMV borders assuming 16x16 block size
        x->mv_row_min = -((mb_row * 16) + VP8BORDERINPIXELS - INTERP_EXTEND);
        x->mv_col_min = -((mb_col * 16) + VP8BORDERINPIXELS - INTERP_EXTEND);
        x->mv_row_max = ((cm->mb_rows - mb_row) * 16 +
                         (VP8BORDERINPIXELS - 16 - INTERP_EXTEND));
        x->mv_col_max = ((cm->mb_cols - mb_col) * 16 +
                         (VP8BORDERINPIXELS - 16 - INTERP_EXTEND));

        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;

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

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

        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;

        /* force 4x4 transform for mode selection */
        xd->mode_info_context->mbmi.txfm_size = TX_4X4; // TODO IS this right??

        cpi->update_context = 0;    // TODO Do we need this now??

        // Find best coding mode & reconstruct the MB so it is available
        // as a predictor for MBs that follow in the SB
        if (cm->frame_type == KEY_FRAME)
        {
            *totalrate += vp8_rd_pick_intra_mode(cpi, x);

            // Save the coding context
            vpx_memcpy (&x->mb_context[i].mic, xd->mode_info_context,
                        sizeof(MODE_INFO));

            // Dummy encode, do not do the tokenization
            vp8cx_encode_intra_macro_block(cpi, x, tp, 0);
            //Note the encoder may have changed the segment_id
        }
        else
        {
            int seg_id;

            if (xd->segmentation_enabled && cpi->seg0_cnt > 0 &&
                !segfeature_active( xd, 0, SEG_LVL_REF_FRAME ) &&
                segfeature_active( xd, 1, SEG_LVL_REF_FRAME ) &&
                check_segref(xd, 1, INTRA_FRAME)  +
                check_segref(xd, 1, LAST_FRAME)   +
                check_segref(xd, 1, GOLDEN_FRAME) +
                check_segref(xd, 1, ALTREF_FRAME) == 1)
            {
                cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
            }
            else
            {
                cpi->seg0_progress = (((mb_col & ~1) * 2 + (mb_row & ~1) * cm->mb_cols + i) << 16) / cm->MBs;
            }

            *totalrate += vp8cx_pick_mode_inter_macroblock(cpi, x,
                                                           recon_yoffset,
                                                           recon_uvoffset);

            // Dummy encode, do not do the tokenization
            vp8cx_encode_inter_macroblock(cpi, x, tp,
                                         recon_yoffset, recon_uvoffset, 0);

            seg_id = xd->mode_info_context->mbmi.segment_id;
            if (cpi->mb.e_mbd.segmentation_enabled && seg_id == 0)
            {
                cpi->seg0_idx++;
            }
            if (!xd->segmentation_enabled ||
                !segfeature_active( xd, seg_id, SEG_LVL_REF_FRAME ) ||
                check_segref(xd, seg_id, INTRA_FRAME)  +
                check_segref(xd, seg_id, LAST_FRAME)   +
                check_segref(xd, seg_id, GOLDEN_FRAME) +
                check_segref(xd, seg_id, ALTREF_FRAME) > 1)
            {
                // Get the prediction context and status
                int pred_flag = get_pred_flag( xd, PRED_REF );
                int pred_context = get_pred_context( cm, xd, PRED_REF );

                // Count prediction success
                cpi->ref_pred_count[pred_context][pred_flag]++;
            }
        }

        // Keep a copy of the updated left context
        vpx_memcpy (&cpi->left_context[i>>1],
                    &cm->left_context,
                    sizeof(ENTROPY_CONTEXT_PLANES));

        // Next MB
        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);

        x->gf_active_ptr += offset_unextended;
        x->partition_info += offset_extended;
        xd->mode_info_context += offset_extended;
        xd->prev_mode_info_context += offset_extended;

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

    /* Restore L & A coding context to those in place on entry */
    vpx_memcpy (cpi->left_context,
                left_context,
                sizeof(left_context));
    vpx_memcpy (initial_above_context_ptr,
                above_context,
                sizeof(above_context));
}

static void encode_sb ( VP8_COMP *cpi,
                        VP8_COMMON *cm,
                        int mbrow,
                        int mbcol,
                        MACROBLOCK  *x,
                        MACROBLOCKD *xd,
                        TOKENEXTRA **tp )
{
    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] = { 0, +1,  0, -1};
    int col_delta[4] = {+1, -1, +1, +1};

    mb_row = mbrow;
    mb_col = mbcol;

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

        if ((mb_row >= cm->mb_rows) || (mb_col >= cm->mb_cols))
        {
            // MB lies outside frame, move on
            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);

            x->gf_active_ptr      += offset_unextended;
            x->partition_info     += offset_extended;
            xd->mode_info_context += offset_extended;
            xd->prev_mode_info_context += offset_extended;

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

        xd->mb_index = i;

#ifdef ENC_DEBUG
        enc_debug = (cpi->common.current_video_frame == 0 &&
                     mb_row==0 && mb_col==0);
        mb_col_debug=mb_col;
        mb_row_debug=mb_row;
#endif

        // Restore MB state to that when it was picked
#if CONFIG_SUPERBLOCKS
        if (x->encode_as_sb)
            update_state (cpi, x, &x->sb_context[i]);
        else
#endif
            update_state (cpi, x, &x->mb_context[i]);

        // Copy in the appropriate left context
        vpx_memcpy (&cm->left_context,
                    &cpi->left_context[i>>1],
                    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;

        // Set up distance of MB to edge of the frame in 1/8th pel units
        xd->mb_to_top_edge    = -((mb_row * 16) << 3);
        xd->mb_to_left_edge   = -((mb_col * 16) << 3);
        xd->mb_to_bottom_edge = ((cm->mb_rows - 1 - mb_row) * 16) << 3;
        xd->mb_to_right_edge  = ((cm->mb_cols - 1 - mb_col) * 16) << 3;

        // Set up limit values for MV components to prevent them from
        // extending beyond the UMV borders assuming 16x16 block size
        x->mv_row_min = -((mb_row * 16) + VP8BORDERINPIXELS - INTERP_EXTEND);
        x->mv_col_min = -((mb_col * 16) + VP8BORDERINPIXELS - INTERP_EXTEND);
        x->mv_row_max = ((cm->mb_rows - mb_row) * 16 +
                         (VP8BORDERINPIXELS - 16 - INTERP_EXTEND));
        x->mv_col_max = ((cm->mb_cols - mb_col) * 16 +
                         (VP8BORDERINPIXELS - 16 - INTERP_EXTEND));

#if CONFIG_SUPERBLOCKS
        // Set up limit values for MV components to prevent them from
        // extending beyond the UMV borders assuming 32x32 block size
        x->mv_row_min_sb = -((mb_row * 16) + VP8BORDERINPIXELS - INTERP_EXTEND);
        x->mv_col_min_sb = -((mb_col * 16) + VP8BORDERINPIXELS - INTERP_EXTEND);
        x->mv_row_max_sb = ((cm->mb_rows - mb_row) * 16 +
                         (VP8BORDERINPIXELS - 32 - INTERP_EXTEND));
        x->mv_col_max_sb = ((cm->mb_cols - mb_col) * 16 +
                         (VP8BORDERINPIXELS - 32 - INTERP_EXTEND));
#endif

        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;

        // Copy current MB to a work 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;

        cpi->update_context = 0;

        if (cm->frame_type == KEY_FRAME)
        {
            vp8cx_encode_intra_macro_block(cpi, x, tp, 1);
            //Note the encoder may have changed the segment_id

#ifdef MODE_STATS
            y_modes[xd->mode_info_context->mbmi.mode] ++;
#endif
        }
        else
        {
            vp8cx_encode_inter_macroblock(cpi, x, tp,
                                         recon_yoffset, recon_uvoffset, 1);
            //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 ++;
        }

        // TODO Partitioning is broken!
        cpi->tplist[mb_row].stop = *tp;

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

        // Next MB
        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);

        x->gf_active_ptr      += offset_unextended;
        x->partition_info     += offset_extended;
        xd->mode_info_context += offset_extended;
        xd->prev_mode_info_context += offset_extended;

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

    // debug output
#if DBG_PRNT_SEGMAP
        {
            FILE *statsfile;
            statsfile = fopen("segmap2.stt", "a");
            fprintf(statsfile, "\n" );
            fclose(statsfile);
        }
    #endif
}

static
void encode_sb_row ( VP8_COMP *cpi,
                     VP8_COMMON *cm,
                     int mb_row,
                     MACROBLOCK  *x,
                     MACROBLOCKD *xd,
                     TOKENEXTRA **tp,
                     int *totalrate )
{
    int mb_col;
    int mb_cols = cm->mb_cols;

    // Initialize the left context for the new SB row
    vpx_memset (cpi->left_context, 0, sizeof(cpi->left_context));
    vpx_memset (&cm->left_context, 0, sizeof(ENTROPY_CONTEXT_PLANES));

    // Code each SB in the row
    for (mb_col=0; mb_col<mb_cols; mb_col+=2)
    {
        int mb_rate = 0;
#if CONFIG_SUPERBLOCKS
        int sb_rate = INT_MAX;
#endif

#if CONFIG_DEBUG
        MODE_INFO *mic = xd->mode_info_context;
        PARTITION_INFO *pi = x->partition_info;
        signed char  *gfa = x->gf_active_ptr;
        unsigned char *yb = x->src.y_buffer;
        unsigned char *ub = x->src.u_buffer;
        unsigned char *vb = x->src.v_buffer;
#endif

        // Pick modes assuming the SB is coded as 4 independent MBs
        pick_mb_modes (cpi, cm, mb_row, mb_col, x, xd, tp, &mb_rate);

        x->src.y_buffer -= 32;
        x->src.u_buffer -= 16;
        x->src.v_buffer -= 16;

        x->gf_active_ptr -= 2;
        x->partition_info -= 2;
        xd->mode_info_context -= 2;
        xd->prev_mode_info_context -= 2;

#if CONFIG_DEBUG
        assert (x->gf_active_ptr == gfa);
        assert (x->partition_info == pi);
        assert (xd->mode_info_context == mic);
        assert (x->src.y_buffer == yb);
        assert (x->src.u_buffer == ub);
        assert (x->src.v_buffer == vb);
#endif

#if CONFIG_SUPERBLOCKS
        // Pick a mode assuming that it applies all 4 of the MBs in the SB
        pick_sb_modes(cpi, cm, mb_row, mb_col, x, xd, &sb_rate);

        // Decide whether to encode as a SB or 4xMBs
        if(sb_rate < mb_rate)
        {
            x->encode_as_sb = 1;
            *totalrate += sb_rate;
        }
        else
#endif
        {
            x->encode_as_sb = 0;
            *totalrate += mb_rate;
        }

        // Encode SB using best computed mode(s)
        encode_sb (cpi, cm, mb_row, mb_col, x, xd, tp);

#if CONFIG_DEBUG
        assert (x->gf_active_ptr == gfa+2);
        assert (x->partition_info == pi+2);
        assert (xd->mode_info_context == mic+2);
        assert (x->src.y_buffer == yb+32);
        assert (x->src.u_buffer == ub+16);
        assert (x->src.v_buffer == vb+16);