me_umhex.c

the newest JM software by h.264 JVT official reference model.

  //early termination alogrithm, refer to JVT-G016
  EARLY_TERMINATION;


  //sec_step: //Unsymmetrical-cross search
  iXMinNow = best_x;
  iYMinNow = best_y;

  for(i = 1; i < search_range; i+=2)
  {
    search_step = i;
    cand_x = iXMinNow + search_step;
    cand_y = iYMinNow ;
    SEARCH_ONE_PIXEL_BIPRED;
    cand_x = iXMinNow - search_step;
    cand_y = iYMinNow ;
    SEARCH_ONE_PIXEL_BIPRED;
  }

  for(i = 1; i < (search_range/2);i+=2)
  {
    search_step = i;
    cand_x = iXMinNow ;
    cand_y = iYMinNow + search_step;
    SEARCH_ONE_PIXEL_BIPRED;
    cand_x = iXMinNow ;
    cand_y = iYMinNow - search_step;
    SEARCH_ONE_PIXEL_BIPRED;
  }
  //early termination alogrithm, refer to JVT-G016
  EARLY_TERMINATION;

  //third_step:     // Uneven Multi-Hexagon-grid Search
  iXMinNow = best_x;
  iYMinNow = best_y;
  //sub step1: 5x5 square search
  for(pos=1;pos<25;pos++)
  {
    cand_x = iXMinNow + spiral_search_x[pos];
    cand_y = iYMinNow + spiral_search_y[pos];
    SEARCH_ONE_PIXEL_BIPRED;
  }

  //early termination alogrithm, refer to JVT-G016
  EARLY_TERMINATION;      //added back by xxz

  //sub step2: multi-grid-hexagon-search
  memcpy(temp_Big_Hexagon_x,Big_Hexagon_x,64);
  memcpy(temp_Big_Hexagon_y,Big_Hexagon_y,64);
  for(i=1;i<=(params->search_range>>2); i++)
  {

    for (m = 0; m < 16; m++)
    {
      cand_x = iXMinNow + temp_Big_Hexagon_x[m];
      cand_y = iYMinNow + temp_Big_Hexagon_y[m];
      temp_Big_Hexagon_x[m] += Big_Hexagon_x[m];
      temp_Big_Hexagon_y[m] += Big_Hexagon_y[m];

      SEARCH_ONE_PIXEL_BIPRED;
    }
    if(min_mcost < ET_Thred)
    {
      goto terminate_step;

    }
  }
  //fourth step: Local Refinement: Extended Hexagon-based Search
fourth_1_step:

  for(i=0; i < search_range; i++)
  {
    iXMinNow = best_x;
    iYMinNow = best_y;
    for (m = 0; m < 6; m++)
    {
      cand_x = iXMinNow + Hexagon_x[m];
      cand_y = iYMinNow + Hexagon_y[m];
      SEARCH_ONE_PIXEL_BIPRED;
    }
    if(best_x == iXMinNow && best_y == iYMinNow)
      break;
  }
fourth_2_step:

  for(i = 0; i < search_range; i++)
  {
    iXMinNow = best_x;
    iYMinNow = best_y;
    for (m = 0; m < 4; m++)
    {
      cand_x = iXMinNow + Diamond_x[m];
      cand_y = iYMinNow + Diamond_y[m];
      SEARCH_ONE_PIXEL_BIPRED;
    }
    if(best_x == iXMinNow && best_y == iYMinNow)
      break;
  }

terminate_step:
  for (i=0; i < (blocksize_x>>2); i++)
  {
    for (j=0; j < (blocksize_y>>2); j++)
    {
      if(list == 0)
      {
        fastme_l0_cost_bipred[blocktype][(img->pix_y>>2)+block_y+j][(img->pix_x>>2)+block_x+i] = min_mcost;
      }
      else
      {
        fastme_l1_cost_bipred[blocktype][(img->pix_y>>2)+block_y+j][(img->pix_x>>2)+block_x+i] = min_mcost;
      }
    }
  }

  mv1->mv_x = best_x - pic_pix_x;
  mv1->mv_y = best_y - pic_pix_y;

  return min_mcost;
}

/*!
 ************************************************************************
 * \brief
 *    Set motion vector predictor
 ************************************************************************
 */
void UMHEXSetMotionVectorPredictor (Macroblock *currMB, 
                                    short  pmv[2],
                                    char   **refPic,
                                    short  ***tmp_mv,
                                    short  ref_frame,
                                    int    list,
                                    int    mb_x,
                                    int    mb_y,
                                    int    blockshape_x,
                                    int    blockshape_y,
                                    int    *search_range)
{
  int mv_a, mv_b, mv_c, pred_vec=0;
  int mvPredType, rFrameL, rFrameU, rFrameUR;
  int hv;

  PixelPos block_a, block_b, block_c, block_d;

  // added for bipred mode
  int *** fastme_l0_cost_flag = (bipred_flag ? fastme_l0_cost_bipred:fastme_l0_cost);
  int *** fastme_l1_cost_flag = (bipred_flag ? fastme_l1_cost_bipred:fastme_l1_cost);

  //Dynamic Search Range

  int dsr_temp_search_range[2];
  int dsr_mv_avail, dsr_mv_max, dsr_mv_sum, dsr_small_search_range;
  int *mb_size = img->mb_size[IS_LUMA];

  // neighborhood SAD init
  SAD_a=0;
  SAD_b=0;
  SAD_c=0;
  SAD_d=0;

  get4x4Neighbour(currMB, mb_x - 1           , mb_y    , mb_size, &block_a);
  get4x4Neighbour(currMB, mb_x               , mb_y - 1, mb_size, &block_b);
  get4x4Neighbour(currMB, mb_x + blockshape_x, mb_y - 1, mb_size, &block_c);
  get4x4Neighbour(currMB, mb_x - 1           , mb_y - 1, mb_size, &block_d);

  if (mb_y > 0)
  {
    if (mb_x < 8)  // first column of 8x8 blocks
    {
      if (mb_y==8)
      {
        if (blockshape_x == 16)      block_c.available  = 0;
      }
      else
      {
        if (mb_x+blockshape_x == 8)  block_c.available = 0;
      }
    }
    else
    {
      if (mb_x+blockshape_x == 16)   block_c.available = 0;
    }
  }

  if (!block_c.available)
  {
    block_c=block_d;
  }

  mvPredType = MVPRED_MEDIAN;

  if (!img->MbaffFrameFlag)
  {
    rFrameL    = block_a.available    ? refPic[block_a.pos_y][block_a.pos_x] : -1;
    rFrameU    = block_b.available    ? refPic[block_b.pos_y][block_b.pos_x] : -1;
    rFrameUR   = block_c.available    ? refPic[block_c.pos_y][block_c.pos_x] : -1;
  }
  else
  {
    if (img->mb_data[img->current_mb_nr].mb_field)
    {
      rFrameL  = block_a.available
        ? (img->mb_data[block_a.mb_addr].mb_field
        ? refPic[block_a.pos_y][block_a.pos_x]
        : refPic[block_a.pos_y][block_a.pos_x] * 2) : -1;
      rFrameU  = block_b.available
        ? (img->mb_data[block_b.mb_addr].mb_field
        ? refPic[block_b.pos_y][block_b.pos_x]
        : refPic[block_b.pos_y][block_b.pos_x] * 2) : -1;
      rFrameUR = block_c.available
        ? (img->mb_data[block_c.mb_addr].mb_field
        ? refPic[block_c.pos_y][block_c.pos_x]
        : refPic[block_c.pos_y][block_c.pos_x] * 2) : -1;
    }
    else
      {
      rFrameL = block_a.available
        ? (img->mb_data[block_a.mb_addr].mb_field
        ? refPic[block_a.pos_y][block_a.pos_x] >>1
        : refPic[block_a.pos_y][block_a.pos_x]) : -1;
      rFrameU    = block_b.available    ?
        img->mb_data[block_b.mb_addr].mb_field ?
        refPic[block_b.pos_y][block_b.pos_x] >>1:
      refPic[block_b.pos_y][block_b.pos_x] :
      -1;
      rFrameUR    = block_c.available    ?
        img->mb_data[block_c.mb_addr].mb_field ?
        refPic[block_c.pos_y][block_c.pos_x] >>1:
      refPic[block_c.pos_y][block_c.pos_x] :
      -1;
    }
  }

  /* Prediction if only one of the neighbors uses the reference frame
  * we are checking
  */
  if(rFrameL == ref_frame && rFrameU != ref_frame && rFrameUR != ref_frame)       mvPredType = MVPRED_L;
  else if(rFrameL != ref_frame && rFrameU == ref_frame && rFrameUR != ref_frame)  mvPredType = MVPRED_U;
  else if(rFrameL != ref_frame && rFrameU != ref_frame && rFrameUR == ref_frame)  mvPredType = MVPRED_UR;
  // Directional predictions
  if(blockshape_x == 8 && blockshape_y == 16)
  {
    if(mb_x == 0)
    {
      if(rFrameL == ref_frame)
        mvPredType = MVPRED_L;
    }
    else
    {
      if( rFrameUR == ref_frame)
        mvPredType = MVPRED_UR;
    }
  }
  else if(blockshape_x == 16 && blockshape_y == 8)
  {
    if(mb_y == 0)
    {
      if(rFrameU == ref_frame)
        mvPredType = MVPRED_U;
    }
    else
    {
      if(rFrameL == ref_frame)
        mvPredType = MVPRED_L;
    }
  }

  // neighborhood SAD prediction
  if((params->UMHexDSR == 1 || params->BiPredMotionEstimation == 1))
  {
    SAD_a = block_a.available ? ((list==1) ? (fastme_l1_cost_flag[UMHEX_blocktype][block_a.pos_y][block_a.pos_x]) : (fastme_l0_cost_flag[UMHEX_blocktype][block_a.pos_y][block_a.pos_x])) : 0;
    SAD_b = block_b.available ? ((list==1) ? (fastme_l1_cost_flag[UMHEX_blocktype][block_b.pos_y][block_b.pos_x]) : (fastme_l0_cost_flag[UMHEX_blocktype][block_b.pos_y][block_b.pos_x])) : 0;
    SAD_d = block_d.available ? ((list==1) ? (fastme_l1_cost_flag[UMHEX_blocktype][block_d.pos_y][block_d.pos_x]) : (fastme_l0_cost_flag[UMHEX_blocktype][block_d.pos_y][block_d.pos_x])) : 0;
    SAD_c = block_c.available ? ((list==1) ? (fastme_l1_cost_flag[UMHEX_blocktype][block_c.pos_y][block_c.pos_x]) : (fastme_l0_cost_flag[UMHEX_blocktype][block_c.pos_y][block_c.pos_x])) : SAD_d;
  }
  for (hv=0; hv < 2; hv++)
  {
    if (!img->MbaffFrameFlag || hv==0)
    {
      mv_a = block_a.available  ? tmp_mv[block_a.pos_y][block_a.pos_x][hv] : 0;
      mv_b = block_b.available  ? tmp_mv[block_b.pos_y][block_b.pos_x][hv] : 0;
      mv_c = block_c.available  ? tmp_mv[block_c.pos_y][block_c.pos_x][hv] : 0;
    }
    else
    {
      if (img->mb_data[img->current_mb_nr].mb_field)
      {
        mv_a = block_a.available  ? img->mb_data[block_a.mb_addr].mb_field
          ? tmp_mv[block_a.pos_y][block_a.pos_x][hv]
          : tmp_mv[block_a.pos_y][block_a.pos_x][hv] / 2
          : 0;
        mv_b = block_b.available  ? img->mb_data[block_b.mb_addr].mb_field
          ? tmp_mv[block_b.pos_y][block_b.pos_x][hv]
          : tmp_mv[block_b.pos_y][block_b.pos_x][hv] / 2
          : 0;
        mv_c = block_c.available  ? img->mb_data[block_c.mb_addr].mb_field
          ? tmp_mv[block_c.pos_y][block_c.pos_x][hv]
          : tmp_mv[block_c.pos_y][block_c.pos_x][hv] / 2
          : 0;
      }
      else
      {
        mv_a = block_a.available  ? img->mb_data[block_a.mb_addr].mb_field
          ? tmp_mv[block_a.pos_y][block_a.pos_x][hv] * 2
          : tmp_mv[block_a.pos_y][block_a.pos_x][hv]
          : 0;
        mv_b = block_b.available  ? img->mb_data[block_b.mb_addr].mb_field
          ? tmp_mv[block_b.pos_y][block_b.pos_x][hv] * 2
          : tmp_mv[block_b.pos_y][block_b.pos_x][hv]
          : 0;
        mv_c = block_c.available  ? img->mb_data[block_c.mb_addr].mb_field
          ? tmp_mv[block_c.pos_y][block_c.pos_x][hv] * 2
          : tmp_mv[block_c.pos_y][block_c.pos_x][hv]
          : 0;
      }
    }

    switch (mvPredType)
    {
    case MVPRED_MEDIAN:
      if(!(block_b.available || block_c.available))
      {
        pred_vec = mv_a;
      }
      else
      {
        pred_vec = mv_a+mv_b+mv_c-imin(mv_a,imin(mv_b,mv_c))-imax(mv_a,imax(mv_b,mv_c));
      }
      break;
    case MVPRED_L:
      pred_vec = mv_a;
      break;
    case MVPRED_U:
      pred_vec = mv_b;
      break;
    case MVPRED_UR:
      pred_vec = mv_c;
      break;
    default:
      break;
    }

    pmv[hv] = pred_vec;
    //Dynamic Search Range
    if (params->UMHexDSR)
    {
      dsr_mv_avail=block_a.available+block_b.available+block_c.available;
      if(dsr_mv_avail < 2)
      {
        dsr_temp_search_range[hv] = params->search_range;
      }
      else
      {
        dsr_mv_max = imax(iabs(mv_a),imax(iabs(mv_b),iabs(mv_c)));
        dsr_mv_sum = (iabs(mv_a)+iabs(mv_b)+iabs(mv_c));
        if(dsr_mv_sum == 0) dsr_small_search_range = (params->search_range + 4) >> 3;
        else if(dsr_mv_sum > 3 ) dsr_small_search_range = (params->search_range + 2) >>2;
        else dsr_small_search_range = (3*params->search_range + 8) >> 4;
        dsr_temp_search_range[hv]=imin(params->search_range,imax(dsr_small_search_range,dsr_mv_max<<1));
        if(imax(SAD_a,imax(SAD_b,SAD_c)) > Threshold_DSR_MB[UMHEX_blocktype])
          dsr_temp_search_range[hv] = params->search_range;
      }
    }
  }

  //Dynamic Search Range
  if (params->UMHexDSR) 
  {
    dsr_new_search_range = imax(dsr_temp_search_range[0],dsr_temp_search_range[1]);

    if      (params->full_search == 2) 
      *search_range = dsr_new_search_range;
    else if (params->full_search == 1) 
      *search_range = dsr_new_search_range /  (imin(ref_frame,1)+1);
    else                              
      *search_range = dsr_new_search_range / ((imin(ref_frame,1)+1) * imin(2,BlockType_LUT[(blockshape_y >> 2) - 1][(blockshape_x >> 2) - 1]));
  }
}