else
  {
    for (j=j0; j<j1; j++)
      for (i=i0; i<i1; i++)
      {
        if (mode==0)
        {
          pdir  =direct_pdir[img->block_x+i][img->block_y+j];
          fwref =direct_ref_idx[LIST_0][img->block_x+i][img->block_y+j];
          bwref =direct_ref_idx[LIST_1][img->block_x+i][img->block_y+j];
        }
        
        if ((pdir==0 || pdir==2))
        {
          enc_picture->mv[LIST_0][img->block_x+i][img->block_y+j][0] = img->all_mv[i][j][LIST_0][fwref][mode][0];
          enc_picture->mv[LIST_0][img->block_x+i][img->block_y+j][1] = img->all_mv[i][j][LIST_0][fwref][mode][1];
          enc_picture->ref_idx[LIST_0][img->block_x+i][img->block_y+j] = fwref;
          enc_picture->ref_pic_id[LIST_0][img->block_x+i][img->block_y+j] = enc_picture->ref_pic_num[LIST_0+list_offset][enc_picture->ref_idx[LIST_0][img->block_x+i][img->block_y+j]];
        }
        else
        {
          enc_picture->mv[LIST_0][img->block_x+i][img->block_y+j][0] = 0;
          enc_picture->mv[LIST_0][img->block_x+i][img->block_y+j][1] = 0;
          enc_picture->ref_idx[LIST_0][img->block_x+i][img->block_y+j] = -1;
          enc_picture->ref_pic_id[LIST_0][img->block_x+i][img->block_y+j] = -1;
        }

        if ((pdir==1 || pdir==2))
        {
          enc_picture->mv[LIST_1][img->block_x+i][img->block_y+j][0] = img->all_mv[i][j][LIST_1][bwref][mode][0];
          enc_picture->mv[LIST_1][img->block_x+i][img->block_y+j][1] = img->all_mv[i][j][LIST_1][bwref][mode][1];
          enc_picture->ref_idx[LIST_1][img->block_x+i][img->block_y+j] = bwref;
          enc_picture->ref_pic_id[LIST_1][img->block_x+i][img->block_y+j] = enc_picture->ref_pic_num[LIST_1+list_offset][enc_picture->ref_idx[LIST_1][img->block_x+i][img->block_y+j]];
        }
        else
        {
          enc_picture->mv[LIST_1][img->block_x+i][img->block_y+j][0] = 0;
          enc_picture->mv[LIST_1][img->block_x+i][img->block_y+j][1] = 0;
          enc_picture->ref_idx[LIST_1][img->block_x+i][img->block_y+j] = -1;
          enc_picture->ref_pic_id[LIST_1][img->block_x+i][img->block_y+j] = -1;
        }
      }
  }
}

/*! 
 *************************************************************************************
 * \brief
 *    skip macroblock field inference
 * \return
 *    inferred field flag
 *************************************************************************************
 */
int field_flag_inference()
{
  int mb_field;
  Macroblock *currMB = &img->mb_data[img->current_mb_nr];

  if (currMB->mbAvailA)
  {
    mb_field = img->mb_data[currMB->mbAddrA].mb_field;
  }
  else
  {
    // check top macroblock pair
    if (currMB->mbAvailB)
    {
      mb_field = img->mb_data[currMB->mbAddrB].mb_field;
    }
    else
      mb_field = 0;
  }

  return mb_field;
}

/*! 
 *************************************************************************************
 * \brief
 *    Mode Decision for a macroblock
 *************************************************************************************
 */
 void encode_one_macroblock ()
 {
   static const int  b8_mode_table[6]  = {0, 4, 5, 6, 7};         // DO NOT CHANGE ORDER !!!
   static const int  mb_mode_table[7]  = {0, 1, 2, 3, P8x8, I16MB, I4MB}; // DO NOT CHANGE ORDER !!!
   
   int         valid[MAXMODE];
   int         rerun, block, index, mode, i0, i1, j0, j1, pdir, ref, i, j, k, ctr16x16, dummy;
   double      qp, lambda_mode, lambda_motion, min_rdcost, rdcost = 0, max_rdcost=1e30;
   int         lambda_motion_factor;
   int         fw_mcost, bw_mcost, bid_mcost, mcost, max_mcost=(1<<30);
   int         curr_cbp_blk, cnt_nonz = 0, best_cnt_nonz = 0, best_fw_ref = 0, best_pdir;
   int         cost=0;
   int         min_cost = max_mcost, min_cost8x8, cost8x8, cost_direct=0, have_direct=0, i16mode;
   int         intra1 = 0;
   
   int         intra       = (((img->type==P_SLICE||img->type==SP_SLICE) && img->mb_y==img->mb_y_upd && img->mb_y_upd!=img->mb_y_intra) || img->type==I_SLICE);
   int         spframe     = (img->type==SP_SLICE);
   int         siframe     = (img->type==SI_SLICE);
   int         bframe      = (img->type==B_SLICE);
   int         runs        = (input->RestrictRef==1 && input->rdopt==2 && (img->type==P_SLICE || img->type==SP_SLICE || (img->type==B_SLICE && img->nal_reference_idc>0)) ? 2 : 1);
   
   int         checkref    = (input->rdopt && input->RestrictRef && (img->type==P_SLICE || img->type==SP_SLICE));
   Macroblock* currMB      = &img->mb_data[img->current_mb_nr];
   Macroblock* prevMB      = img->current_mb_nr ? &img->mb_data[img->current_mb_nr-1]:NULL ;
   
   int     **ipredmodes = img->ipredmode;
   int     best_bw_ref = -1;
   int     ******allmvs = img->all_mv;

   
   int  l,list_offset;

   int curr_mb_field = ((img->MbaffFrameFlag)&&(currMB->mb_field));

   // find out the correct list offsets
   if (curr_mb_field)
   {
     if(img->current_mb_nr%2)
       list_offset = 4; // bottom field mb
     else
       list_offset = 2; // top field mb
   }
   else
   {
     list_offset = 0;  // no mb aff or frame mb
   }

   if(input->FMEnable)
     decide_intrabk_SAD();
   
   intra |= RandomIntra (img->current_mb_nr);    // Forced Pseudo-Random Intra

   //===== SET VALID MODES =====
   valid[I4MB]   = 1;
   valid[I16MB]  = 1;
   
   valid[0]      = (!intra );
   valid[1]      = (!intra && input->InterSearch16x16);
   valid[2]      = (!intra && input->InterSearch16x8);
   valid[3]      = (!intra && input->InterSearch8x16);
   valid[4]      = (!intra && input->InterSearch8x8);
   valid[5]      = (!intra && input->InterSearch8x4);
   valid[6]      = (!intra && input->InterSearch4x8);
   valid[7]      = (!intra && input->InterSearch4x4);
   valid[P8x8]   = (valid[4] || valid[5] || valid[6] || valid[7]);
   valid[12]     = (siframe);

   if (!img->MbaffFrameFlag)
   {
     for (l=0+list_offset;l<(2+list_offset);l++)
     {
       for(k = 0; k < listXsize[l]; k++)
       {
         listX[l][k]->chroma_vector_adjustment= 0;
         if(img->structure == TOP_FIELD && img->structure != listX[l][k]->structure)
           listX[l][k]->chroma_vector_adjustment = -2;
         if(img->structure == BOTTOM_FIELD && img->structure != listX[l][k]->structure)
           listX[l][k]->chroma_vector_adjustment = 2;
       }
     }
   }
   else
   {
     if (curr_mb_field)
     {
       for (l=0+list_offset;l<(2+list_offset);l++)
       {
         for(k = 0; k < listXsize[l]; k++)
         {
           listX[l][k]->chroma_vector_adjustment= 0;
           if(img->current_mb_nr % 2 == 0 && listX[l][k]->structure == BOTTOM_FIELD)
             listX[l][k]->chroma_vector_adjustment = -2;
           if(img->current_mb_nr % 2 == 1 && listX[l][k]->structure == TOP_FIELD)
             listX[l][k]->chroma_vector_adjustment = 2;
         }
       }
     }
     else
     {
       for (l=0+list_offset;l<(2+list_offset);l++)
       {
         for(k = 0; k < listXsize[l]; k++)
         {
           listX[l][k]->chroma_vector_adjustment= 0;
         }
       }
     }
     
   }
   //===== SET LAGRANGE PARAMETERS =====
   if (input->rdopt)
   {
     qp = (double)img->qp - SHIFT_QP;

     if (input->successive_Bframe>0)
       lambda_mode   = 0.68 * pow (2, qp/3.0) * (img->type==B_SLICE? max(2.00,min(4.00,(qp / 6.0))):spframe?max(1.4,min(3.0,(qp / 12.0))):1.0);  
     else
       lambda_mode   = 0.85 * pow (2, qp/3.0) * (img->type==B_SLICE? 4.0:spframe?max(1.4,min(3.0,(qp / 12.0))):1.0);  

     lambda_motion = sqrt (lambda_mode);
   }
   else
   {
     lambda_mode = lambda_motion = QP2QUANT[max(0,img->qp-SHIFT_QP)];
   }
   lambda_motion_factor = LAMBDA_FACTOR (lambda_motion);
   
   
   for (rerun=0; rerun<runs; rerun++)
   {
     if (runs==2)
     {
       if (rerun==0)   input->rdopt=1;
       else            input->rdopt=2;
     }
     
     // reset chroma intra predictor to default
     currMB->c_ipred_mode = DC_PRED_8;
     
     if (!intra)
     {
       //===== set direct motion vectors =====
       if (bframe)
       {
         Get_Direct_Motion_Vectors ();
       }
       
       //===== MOTION ESTIMATION FOR 16x16, 16x8, 8x16 BLOCKS =====
       for (min_cost=1<<20, best_mode=1, mode=1; mode<4; mode++)
       {
         if (valid[mode])
         {
           for (cost=0, block=0; block<(mode==1?1:2); block++)
           {
             PartitionMotionSearch (mode, block, lambda_motion);

             //--- set 4x4 block indizes (for getting MV) ---
             j = (block==1 && mode==2 ? 2 : 0);
             i = (block==1 && mode==3 ? 2 : 0);

             //--- get cost and reference frame for forward prediction ---
             for (fw_mcost=max_mcost, ref=0; ref<listXsize[LIST_0+list_offset]; ref++)
             {
               if (!checkref || ref==0 || CheckReliabilityOfRef (block, LIST_0, ref, mode))
               {
                 mcost  = (input->rdopt ? REF_COST (lambda_motion_factor, ref, LIST_0 + list_offset) : (int)(2*lambda_motion*min(ref,1)));

                 mcost += motion_cost[mode][LIST_0][ref][block];
                 if (mcost < fw_mcost)
                 {
                   fw_mcost    = mcost;
                   best_fw_ref = ref;
                 }
               }
             }

             if (bframe)
             {
               //--- get cost for bidirectional prediction ---
               for (bw_mcost=max_mcost, ref=0; ref<listXsize[LIST_1 + list_offset]; ref++)
               {
                 mcost  = (input->rdopt ? REF_COST (lambda_motion_factor, ref, LIST_1 + list_offset) : (int)(2*lambda_motion*min(ref,1)));
                 mcost += motion_cost[mode][LIST_1][ref][block];
                 if (mcost < bw_mcost)
                 {
                   bw_mcost    = mcost;
                   best_bw_ref = ref;
                 }
               }

               // search bidirectional between best forward and ref_idx=0 backward
               bid_mcost  = (input->rdopt ? (REF_COST (lambda_motion_factor, best_fw_ref,LIST_0+list_offset)+REF_COST (lambda_motion_factor, 0,LIST_1+list_offset)) : (int)(2*lambda_motion*min(best_fw_ref,1)));
               bid_mcost += BIDPartitionCost (mode, block, best_fw_ref, 0, lambda_motion_factor);

               //--- get prediction direction ----
               if (fw_mcost<=bw_mcost && fw_mcost<=bid_mcost)
               {
                 best_pdir = 0;
                 best_bw_ref = 0;
                 cost += fw_mcost;
               }
               else if (bw_mcost<=fw_mcost && bw_mcost<=bid_mcost)
               {
                 best_pdir = 1;
                 cost += bw_mcost;
                 best_fw_ref = 0;
               }
               else
               {
                 best_pdir = 2;
                 cost += bid_mcost;
                 best_bw_ref = 0;
               }
             }
             else // if (bframe)
             {
               best_pdir  = 0;
               cost      += fw_mcost;
             }

             if (mode==1)
             {
               if (best_pdir==1)
               {
                 for (j=0; j<4; j++)
                 {
                   for (i=0; i<4; i++)
                   {
                     enc_picture->ref_idx[LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = -1;
                     enc_picture->ref_pic_id [LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = -1; 
                     enc_picture->mv[LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][0] = 0;
                     enc_picture->mv[LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][1] = 0;
                   }
                 }
               }
               else
               {
                 for (j=0; j<4; j++)
                 {
                   for (i=0; i<4; i++)
                   {
                     enc_picture->ref_idx[LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = best_fw_ref;
                     enc_picture->ref_pic_id [LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = enc_picture->ref_pic_num[LIST_0 + list_offset][enc_picture->ref_idx[LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j]];  
                     enc_picture->mv[LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][0] = img->all_mv[i][j][LIST_0][best_fw_ref][mode][0];
                     enc_picture->mv[LIST_0][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][1] = img->all_mv[i][j][LIST_0][best_fw_ref][mode][1];
                   }
                 }
               }

               if (bframe)
               {
                 if (best_pdir==0)
                 {
                   for (j=0; j<4; j++)
                   {
                     for (i=0; i<4; i++)
                     {
                       enc_picture->ref_idx[LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = -1;
                       enc_picture->ref_pic_id [LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = -1;   
                       enc_picture->mv[LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][0] = 0;
                       enc_picture->mv[LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][1] = 0;
                     }
                   }
                 }
                 else
                 {
                   for (j=0; j<4; j++)
                   {
                     for (i=0; i<4; i++)
                     {
                       enc_picture->ref_idx[LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = best_bw_ref;
                       enc_picture->ref_pic_id [LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j] = enc_picture->ref_pic_num[LIST_1 + list_offset][enc_picture->ref_idx[LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j]];
                       if(best_bw_ref>=0)
                       {
                         enc_picture->mv[LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][0] = img->all_mv[i][j][LIST_1][best_bw_ref][mode][0];
                         enc_picture->mv[LIST_1][img->block_x+(block&1)*2+i][img->block_y+(block&2)+j][1] = img->all_mv[i][j][LIST_1][best_bw_ref][mode][1];
                       }
                     }
                   }
                 }
               }
             }
             else if (mode==2)
             {
               for (j=0; j<2; j++)
               {
                 for (i=0; i<4; i++)
                 {
                   if (best_pdir==1)
                   {
                     enc_picture->ref_idx[LIST_0][img->block_x+i][img->block_y+block*2+j] = -1;
                     enc_picture->ref_pic_id [LIST_0][img->block_x+i][img->block_y+block*2+j] = -1;
                     enc_picture->mv[LIST_0][img->block_x+i][img->block_y+block*2+j][0] = 0;
                     enc_picture->mv[LIST_0][img->block_x+i][img->block_y+block*2+j][1] = 0;
                   }
                   else
                   {
                     enc_picture->ref_idx[LIST_0][img->block_x+i][img->block_y+block*2+j] = best_fw_ref;
                     enc_picture->ref_pic_id [LIST_0][img->block_x+i][img->block_y+block*2+j] = enc_picture->ref_pic_num[LIST_0 + list_offset][enc_picture->ref_idx[LIST_0][img->block_x+i][img->block_y+block*2+j]];
                     enc_picture->mv[LIST_0][img->block_x+i][img->block_y+block*2+j][0] = img->all_mv[i][j+block*2][LIST_0][best_fw_ref][mode][0];
                     enc_picture->mv[LIST_0][img->block_x+i][img->block_y+block*2+j][1] = img->all_mv[i][j+block*2][LIST_0][best_fw_ref][mode][1];
                   }

                   if (bframe)
                   {
                     if (best_pdir==0)
                     {
                       enc_picture->ref_idx[LIST_1][img->block_x+i][img->block_y+block*2+j] = -1;
                       enc_picture->ref_pic_id [LIST_1][img->block_x+i][img->block_y+block*2+j] = -1;
                       enc_picture->mv[LIST_1][img->