enc_picture->mv[LIST_1][bxr][by][1] = 0;
        }
        else if (pdir8==1) // backward
        {
          enc_picture->mv[LIST_0][bxr][by][0] = 0;
          enc_picture->mv[LIST_0][bxr][by][1] = 0;

          enc_picture->mv[LIST_1][bxr][by][0] = all_mv[i][j][LIST_1][bw_ref][mode8][0];
          enc_picture->mv[LIST_1][bxr][by][1] = all_mv[i][j][LIST_1][bw_ref][mode8][1];
        }
        else if (pdir8==2) // bidir
        {
          enc_picture->mv[LIST_0][bxr][by][0] = all_mv [i][j][LIST_0][ ref][mode8][0];
          enc_picture->mv[LIST_0][bxr][by][1] = all_mv [i][j][LIST_0][ ref][mode8][1];


          enc_picture->mv[LIST_1][bxr][by][0] = all_mv[i][j][LIST_1][bw_ref][mode8][0];
          enc_picture->mv[LIST_1][bxr][by][1] = all_mv[i][j][LIST_1][bw_ref][mode8][1];
        }
        else
        {
          error("invalid direction mode", 255);
        }
      }
  }

  // copy all the motion vectors into rdopt structure
  // Can simplify this by copying the MV's of the best mode (TBD)
  if(img->MbaffFrameFlag)
  {
    for(i=0;i<4;i++)
      for(j=0;j<4;j++)
        for (k=0;k<2;k++)
          for(l=0;l<img->max_num_references;l++)
            for(m=0;m<9;m++)
            {
              rdopt->all_mv [i][j][k][l][m][0]  = all_mv [i][j][k][l][m][0];
              rdopt->pred_mv[i][j][k][l][m][0]  = pred_mv[i][j][k][l][m][0];

              rdopt->all_mv [i][j][k][l][m][1]  = all_mv [i][j][k][l][m][1];
              rdopt->pred_mv[i][j][k][l][m][1]  = pred_mv[i][j][k][l][m][1];
            }
  }
}



/*!
 *************************************************************************************
 * \brief
 *    R-D Cost for a macroblock
 *************************************************************************************
 */
int RDCost_for_macroblocks (double   lambda,      // <-- lagrange multiplier
                        int      mode,        // <-- modus (0-COPY/DIRECT, 1-16x16, 2-16x8, 3-8x16, 4-8x8(+), 5-Intra4x4, 6-Intra16x16)
                        double*  min_rdcost)  // <-> minimum rate-distortion cost
{
  int         i, j, k; //, k, ****ip4;
  int         i16mode, rate=0;
  int64       distortion=0;
  double      rdcost;
  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         bframe    = (img->type==B_SLICE);
  int         tmp_cc;
  int         use_of_cc =  (img->type!=I_SLICE &&  input->symbol_mode!=CABAC);
  int         cc_rate, dummy;
  int         cr_cbp=0, uv;

  //=====
  //=====  SET REFERENCE FRAMES AND BLOCK MODES
  //=====
  SetModesAndRefframeForBlocks (mode);

  //=====
  //=====  GET COEFFICIENTS, RECONSTRUCTIONS, CBP
  //=====
  if (bframe && mode==0)
  {
    int block_x=img->pix_x>>2;
    int block_y=img->pix_y>>2;
    for (i=0;i<4;i++)
      for (j=0;j<4;j++)
        if (direct_pdir[block_x+i][block_y+j]<0)
          return 0;
  }

  if (mode<P8x8)
  {
    LumaResidualCoding ();
    if(mode==0 && currMB->cbp==0 && currMB->luma_transform_size_8x8_flag==1) //for B_skip, luma_transform_size_8x8_flag=0 only
      return 0;
  }
  else if (mode==P8x8)
  {
    SetCoeffAndReconstruction8x8 (currMB);
  }
  else if (mode==I4MB)
  {
    currMB->cbp = Mode_Decision_for_Intra4x4Macroblock (lambda, &dummy);

    // Residue Color Transform
    if(img->residue_transform_flag)
    {
      for(i=0; i<2; i++)
      for(j=0; j<4; j++)
      for(k=0; k<4; k++)
        if(cbp_chroma_block[i][j][k])
          cr_cbp = 2;

      for(uv=0; uv<2; uv++)
        cr_cbp = dct_chroma_DC(uv, cr_cbp);

      currMB->cbp += (cr_cbp<<4);
    }
  }
  else if (mode==I16MB)
  {
    Intra16x16_Mode_Decision  (currMB, &i16mode);
  }
  else if(mode==I8MB){
    currMB->cbp = Mode_Decision_for_new_Intra8x8Macroblock(lambda, &dummy);

    // Residue Color Transform
    if(img->residue_transform_flag)
    {
      for(i=0; i<2; i++)
      for(j=0; j<4; j++)
      for(k=0; k<4; k++)
        if(cbp_chroma_block[i][j][k])
          cr_cbp = 2;

      for(uv=0; uv<2; uv++)
        cr_cbp = dct_chroma_DC(uv, cr_cbp);

      currMB->cbp += (cr_cbp<<4);
    }
  }

  if (input->rdopt==2 && img->type!=B_SLICE)
  {
    // We need the reconstructed prediction residue for the simulated decoders.
    compute_residue_mb (mode==I16MB?i16mode:-1);
  }

  //Rate control
  if (mode == I16MB)
  {
    for(i=0; i<16; i++)
    for(j=0; j<16; j++)
      pred[j][i] = img->mprr_2[i16mode][j][i];
  }else
  {
    for(i=0; i<16; i++)
    for(j=0; j<16; j++)
      pred[j][i] = img->mpr[i][j];
  }

  img->i16offset = 0;
  dummy = 0;
  if ((!(img->residue_transform_flag && (mode==I4MB || mode==I16MB || mode==I8MB))) && img->yuv_format!=YUV400)
    ChromaResidualCoding (&dummy);
  if (mode==I16MB)     img->i16offset = I16Offset  (currMB->cbp, i16mode);

  //=====
  //=====   GET DISTORTION
  //=====
  // LUMA
  if (input->rdopt==2 && img->type!=B_SLICE)
  {
    for (k=0; k<input->NoOfDecoders ;k++)
    {
      decode_one_mb (k, currMB);
      for (j=0; j<MB_BLOCK_SIZE; j++)
      for (i=img->opix_x; i<img->opix_x+MB_BLOCK_SIZE; i++)
      {
        distortion += img->quad [imgY_org[img->opix_y+j][i] - decs->decY[k][img->opix_y+j][i]];
      }
    }
    distortion /= input->NoOfDecoders;
  }
  else
  {
    for (j=0; j<16; j++)
    for (i=img->opix_x; i<img->opix_x+16; i++)
    {
      distortion += img->quad [imgY_org[j+img->opix_y][i] - enc_picture->imgY[img->pix_y+j][i]];
    }
  }

  if (img->yuv_format != YUV400)
  {
    // CHROMA
    for (j=0; j<img->mb_cr_size_y; j++)
    for (i=img->opix_c_x; i<img->opix_c_x+img->mb_cr_size_x; i++)
    {
      distortion += img->quad [imgUV_org[0][j+img->opix_c_y][i] - enc_picture->imgUV[0][img->pix_c_y+j][i]];
      distortion += img->quad [imgUV_org[1][j+img->opix_c_y][i] - enc_picture->imgUV[1][img->pix_c_y+j][i]];
    }
  }


  //=====   S T O R E   C O D I N G   S T A T E   =====
  //---------------------------------------------------
  store_coding_state (cs_cm);

  //=====
  //=====   GET RATE
  //=====
  //----- macroblock header -----
  if (use_of_cc)
  {
    if (currMB->mb_type!=0 || (bframe && currMB->cbp!=0))
    {
      // cod counter and macroblock mode are written ==> do not consider code counter
      tmp_cc = img->cod_counter;
      rate   = writeMBLayer (1);
      ue_linfo (tmp_cc, dummy, &cc_rate, &dummy);
      rate  -= cc_rate;
      img->cod_counter = tmp_cc;
    }
    else
    {
      // cod counter is just increased  ==> get additional rate
      ue_linfo (img->cod_counter+1, dummy, &rate,    &dummy);
      ue_linfo (img->cod_counter,   dummy, &cc_rate, &dummy);
      rate -= cc_rate;
    }
  }
  else
  {
    rate = writeMBLayer (1);
  }

  //=====   R E S T O R E   C O D I N G   S T A T E   =====
  //-------------------------------------------------------
  reset_coding_state (cs_cm);

  rdcost = (double)distortion + lambda * (double)rate;

  if (rdcost >= *min_rdcost ||
      ((img->qp + img->bitdepth_luma_qp_scale)==0 && img->lossless_qpprime_flag==1 && distortion!=0))
  {
    return 0;
  }


  if ((img->MbaffFrameFlag) && (mode ? 0: ((img->type == B_SLICE) ? !currMB->cbp:1)))  // AFF and current is skip
  {
    if (img->current_mb_nr%2) //bottom
    {
      if (prevMB->mb_type ? 0:((img->type == B_SLICE) ? !prevMB->cbp:1)) //top is skip
      {
        if (!(field_flag_inference() == currMB->mb_field)) //skip only allowed when correct inference
          return 0;
      }
    }
  }

  //=====   U P D A T E   M I N I M U M   C O S T   =====
  //-----------------------------------------------------
  *min_rdcost = rdcost;
  return 1;
}





/*!
 *************************************************************************************
 * \brief
 *    Store macroblock parameters
 *************************************************************************************
 */
void store_macroblock_parameters (int mode)
{
  int  i, j, k, ****i4p, ***i3p;
  Macroblock *currMB  = &img->mb_data[img->current_mb_nr];
  int        bframe   = (img->type==B_SLICE);

  //--- store best mode ---
  best_mode = mode;
  best_c_imode = currMB->c_ipred_mode;
  best_i16offset = img->i16offset;
  for (i=0; i<4; i++)
  {
    b8mode[i]   = currMB->b8mode[i];
    b8pdir[i]   = currMB->b8pdir[i];
  }
  // Residue Color Transform
  for (k=0, j=img->block_y; j<img->block_y+4; j++)
  for (i=img->block_x; i<img->block_x+4; i++, k++)
  {
    b4_ipredmode[k] = img->ipredmode[i][j];
    b4_intra_pred_modes[k] = currMB->intra_pred_modes[k];
  }
  for(i = k = 0; i < 4; i++)
  for(j = 0; j < 4; j++,k++)
  {
    b8_ipredmode8x8[i][j] = img->ipredmode8x8[img->block_x+i][img->block_y+j];
    //only every fourth value relevant
    b8_intra_pred_modes8x8[k] = currMB->intra_pred_modes8x8[k];
  }

  //--- reconstructed blocks ----
  for (j=0; j<16; j++)
  for (i=0; i<16; i++)
  {
    rec_mbY[j][i] = enc_picture->imgY[img->pix_y+j][img->pix_x+i];
  }

  if (img->yuv_format != YUV400)
  {
    for (j=0; j<img->mb_cr_size_y; j++)
    for (i=0; i<img->mb_cr_size_x; i++)
    {
      rec_mbU[j][i] = enc_picture->imgUV[0][img->pix_c_y+j][img->pix_c_x+i];
      rec_mbV[j][i] = enc_picture->imgUV[1][img->pix_c_y+j][img->pix_c_x+i];
    }
  }


  //--- store results of decoders ---
  if (input->rdopt==2 && img->type!=B_SLICE)
  {
    for (k=0; k<input->NoOfDecoders; k++)
    {
      for (j=img->pix_y; j<img->pix_y+16; j++)
      for (i=img->pix_x; i<img->pix_x+16; i++)
      {
        // Keep the decoded values of each MB for updating the ref frames
        decs->decY_best[k][j][i] = decs->decY[k][j][i];
      }
    }
  }

  //--- coeff, cbp, kac ---
  if (mode || bframe)
  {
    i4p=cofAC; cofAC=img->cofAC; img->cofAC=i4p;
    i3p=cofDC; cofDC=img->cofDC; img->cofDC=i3p;
    cbp     = currMB->cbp;
    cbp_blk = currMB->cbp_blk;
  }
  else
  {
    cbp_blk = cbp = 0;
  }

  //--- store transform size ---
  luma_transform_size_8x8_flag = currMB->luma_transform_size_8x8_flag;
//  NoMbPartLessThan8x8Flag  = currMB->NoMbPartLessThan8x8Flag;   //DEL-VG-29072004


  for (j=0; j<4; j++)
  for (i=0; i<4; i++)
  {
    frefframe[j][i] = enc_picture->ref_idx[LIST_0][img->block_x+i][img->block_y+j];

    if (bframe)
    {
      brefframe[j][i] = enc_picture->ref_idx[LIST_1][img->block_x+i][img->block_y+j];
    }
  }

}


/*!
 *************************************************************************************
 * \brief
 *    Set stored macroblock parameters
 *************************************************************************************
 */
void set_stored_macroblock_parameters ()
{
  int  i, j, k, ****i4p, ***i3p,l;
  Macroblock  *currMB  = &img->mb_data[img->current_mb_nr];
  int         mode     = best_mode;
  int         bframe   = (img->type==B_SLICE);
  int         **ipredmodes = img->ipredmode;

  imgpel        **imgY  = enc_picture->imgY;
  imgpel       ***imgUV = enc_picture->imgUV;

  int        list_offset   = ((img->MbaffFrameFlag)&&(currMB->mb_field))? img->current_mb_nr%2 ? 4 : 2 : 0;

  //===== reconstruction values =====
  for (j=0; j<16; j++)
  for (i=0; i<16; i++)
  {
    imgY[img->pix_y+j][img->pix_x+i] = rec_mbY[j][i];

    if(img->MbaffFrameFlag)
      rdopt->rec_mbY[j][i]       = rec_mbY[j][i];
  }

  if (img->yuv_format != YUV400)
  {
    for (j=0; j<img->mb_cr_size_y; j++)
      for (i=0; i<img->mb_cr_size_x; i++)
      {
        imgUV[0][img->pix_c_y+j][img->pix_c_x+i] = rec_mbU[j][i];
        imgUV[1][img->pix_c_y+j][img->pix_c_x+i] = rec_mbV[j][i];
        
        if(img->MbaffFrameFlag)
        {
          rdopt->rec_mbU[j][i]           = rec_mbU[j][i];
          rdopt->rec_mbV[j][i]           = rec_mbV[j]