fast_me.c

H.264 codec source code

        x2  = x1 + 1;
        x3  = x2 + 1;
        d   = diff;

        orig_line = orig_pic [y0];    
        ref_line  = get_line (ref_pic, ry0, rx0, img_height, img_width);
        *d++      = orig_line[x0] - *(ref_line     );
        *d++      = orig_line[x1] - *(ref_line + 4 );
        *d++      = orig_line[x2] - *(ref_line + 8 );
        *d++      = orig_line[x3] - *(ref_line + 12);

        orig_line = orig_pic [y1];    
        ref_line  = get_line (ref_pic, ry4, rx0, img_height, img_width);
        *d++      = orig_line[x0] - *(ref_line     );
        *d++      = orig_line[x1] - *(ref_line + 4 );
        *d++      = orig_line[x2] - *(ref_line + 8 );
        *d++      = orig_line[x3] - *(ref_line + 12);

        orig_line = orig_pic [y2];
        ref_line  = get_line (ref_pic, ry8, rx0, img_height, img_width);
        *d++      = orig_line[x0] - *(ref_line     );
        *d++      = orig_line[x1] - *(ref_line += 4 );
        *d++      = orig_line[x2] - *(ref_line += 4 );
        *d++      = orig_line[x3] - *(ref_line += 4);

        orig_line = orig_pic [y3];    
        ref_line  = get_line (ref_pic, ry12, rx0, img_height, img_width);
        *d++      = orig_line[x0] - *(ref_line     );
        *d++      = orig_line[x1] - *(ref_line += 4);
        *d++      = orig_line[x2] - *(ref_line += 4);
        *d        = orig_line[x3] - *(ref_line += 4);

      if (!useABT)
      {
        if ((mcost += SATD (diff, input->hadamard)) > min_mcost)
        {
          abort_search = 1;
          break;
        }
      }
      else  // copy diff to curr_diff for ABT SATD calculation
      {
          i = (x0&0x7) +  (x0>7) * 64 + y_offset;
          for(k=0, j=y0; j<BLOCK_SIZE + y0; j++, k+=BLOCK_SIZE)
            memcpy(&(c_diff[i + ((j&0x7)<<3)]), &diff[k], BLOCK_SIZE*sizeof(int));
      }
    }
  }
  
  if(useABT)
  {
    mcost += find_SATD (c_diff, blocktype);
  }

  return mcost;
}


int                                                   //  ==> minimum motion cost after search
FastSubPelBlockMotionSearch (pel_t**   orig_pic,      // <--  original pixel values for the AxB block
                             short     ref,           // <--  reference frame (0... or -1 (backward))
                             int       list,
                             int       pic_pix_x,     // <--  absolute x-coordinate of regarded AxB block
                             int       pic_pix_y,     // <--  absolute y-coordinate of regarded AxB block
                             int       blocktype,     // <--  block type (1-16x16 ... 7-4x4)
                             short     pred_mv_x,     // <--  motion vector predictor (x) in sub-pel units
                             short     pred_mv_y,     // <--  motion vector predictor (y) in sub-pel units
                             short*    mv_x,          // <--> in: search center (x) / out: motion vector (x) - in pel units
                             short*    mv_y,          // <--> in: search center (y) / out: motion vector (y) - in pel units
                             int       search_pos2,   // <--  search positions for    half-pel search  (default: 9)
                             int       search_pos4,   // <--  search positions for quarter-pel search  (default: 9)
                             int       min_mcost,     // <--  minimum motion cost (cost for center or huge value)
                             int       lambda_factor,
                             int       useABT)        // <--  lagrangian parameter for determining motion cost
{
  static int Diamond_x[4] = {-1, 0, 1, 0};
  static int Diamond_y[4] = {0, 1, 0, -1};
  int   mcost;
  int   cand_mv_x, cand_mv_y;
  
  int   list_offset   = ((img->MbaffFrameFlag)&&(img->mb_data[img->current_mb_nr].mb_field))? img->current_mb_nr%2 ? 4 : 2 : 0;
  StorablePicture *ref_picture = listX[list+list_offset][ref];
  
  int   mv_shift        = 0;
  int   blocksize_x     = input->blc_size[blocktype][0];
  int   blocksize_y     = input->blc_size[blocktype][1];
  int   pic4_pix_x      = ((pic_pix_x + IMG_PAD_SIZE)<< 2);
  int   pic4_pix_y      = ((pic_pix_y + IMG_PAD_SIZE)<< 2);
  short max_pos_x4      = ((ref_picture->size_x - blocksize_x + 2*IMG_PAD_SIZE)<<2);
  short max_pos_y4      = ((ref_picture->size_y - blocksize_y + 2*IMG_PAD_SIZE)<<2);
  
  int   search_range_dynamic,iXMinNow,iYMinNow,i;
  int   m,currmv_x = 0,currmv_y = 0;
  int   pred_frac_mv_x,pred_frac_mv_y,abort_search;
  int   mv_cost; 
  
  int   pred_frac_up_mv_x, pred_frac_up_mv_y;
  
  if ((pic4_pix_x + *mv_x > 1) && (pic4_pix_x + *mv_x < max_pos_x4 - 1) &&
      (pic4_pix_y + *mv_y > 1) && (pic4_pix_y + *mv_y < max_pos_y4 - 1)   )
  {
    get_line = FastLine4X;
  }
  else
  {
    get_line = UMVLine4X;    
  }

  search_range_dynamic = 3;
  pred_frac_mv_x = (pred_mv_x - *mv_x)%4;
  pred_frac_mv_y = (pred_mv_y - *mv_y)%4; 
  
  pred_frac_up_mv_x = (pred_MV_uplayer[0] - *mv_x)%4;
  pred_frac_up_mv_y = (pred_MV_uplayer[1] - *mv_y)%4;
  
  
  memset(SearchState[0],0,(2*search_range_dynamic+1)*(2*search_range_dynamic+1));
  
  if(input->hadamard)
  {
    cand_mv_x = *mv_x;    
    cand_mv_y = *mv_y;    
    mv_cost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);    
    mcost = AddUpSADQuarter(pic_pix_x,pic_pix_y,blocksize_x,blocksize_y,cand_mv_x + pic4_pix_x,cand_mv_y + pic4_pix_y,ref_picture,orig_pic,mv_cost,min_mcost,useABT, blocktype);
    SearchState[search_range_dynamic][search_range_dynamic] = 1;
    if (mcost < min_mcost)
    {
      min_mcost = mcost;
      currmv_x = cand_mv_x;
      currmv_y = cand_mv_y; 
    }
  }
  else
  {
    SearchState[search_range_dynamic][search_range_dynamic] = 1;
    currmv_x = *mv_x;
    currmv_y = *mv_y; 
  }
  
  if(pred_frac_mv_x!=0 || pred_frac_mv_y!=0)
  {
    cand_mv_x = *mv_x + pred_frac_mv_x;    
    cand_mv_y = *mv_y + pred_frac_mv_y;    
    mv_cost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);    
    mcost = AddUpSADQuarter(pic_pix_x,pic_pix_y,blocksize_x,blocksize_y,cand_mv_x + pic4_pix_x, cand_mv_y + pic4_pix_y,ref_picture,orig_pic,mv_cost,min_mcost,useABT, blocktype);
    SearchState[cand_mv_y -*mv_y + search_range_dynamic][cand_mv_x - *mv_x + search_range_dynamic] = 1;
    if (mcost < min_mcost)
    {
      min_mcost = mcost;
      currmv_x = cand_mv_x;
      currmv_y = cand_mv_y; 
    }
  }
  
  
  iXMinNow = currmv_x;
  iYMinNow = currmv_y;
  for(i=0;i<search_range_dynamic;i++) 
  {
    abort_search=1;
    for (m = 0; m < 4; m++)
    {
      cand_mv_x = iXMinNow + Diamond_x[m];    
      cand_mv_y = iYMinNow + Diamond_y[m]; 
      
      if(abs(cand_mv_x - *mv_x) <=search_range_dynamic && abs(cand_mv_y - *mv_y)<= search_range_dynamic)
      {
        if(!SearchState[cand_mv_y -*mv_y+ search_range_dynamic][cand_mv_x -*mv_x+ search_range_dynamic])
        {
          mv_cost = MV_COST (lambda_factor, mv_shift, cand_mv_x, cand_mv_y, pred_mv_x, pred_mv_y);    
          mcost = AddUpSADQuarter(pic_pix_x,pic_pix_y,blocksize_x,blocksize_y,cand_mv_x + pic4_pix_x, cand_mv_y + pic4_pix_y,ref_picture,orig_pic,mv_cost,min_mcost,useABT, blocktype);
          SearchState[cand_mv_y - *mv_y + search_range_dynamic][cand_mv_x - *mv_x + search_range_dynamic] = 1;
          if (mcost < min_mcost)
          {
            min_mcost = mcost;
            currmv_x = cand_mv_x;
            currmv_y = cand_mv_y; 
            abort_search = 0; 
            
          }
        }
      }
    }
    iXMinNow = currmv_x;
    iYMinNow = currmv_y;
    if(abort_search)
      break;
  }
  
  *mv_x = currmv_x;
  *mv_y = currmv_y;
  
  //===== return minimum motion cost =====
  return min_mcost;
}

/*!
 ************************************************************************
 * \brief
 * Functions for SAD prediction of intra block cases.
 * 1. void   decide_intrabk_SAD() judges the block coding type(intra/inter) 
 *    of neibouring blocks
 * 2. void skip_intrabk_SAD() set the SAD to zero if neigouring block coding 
 *    type is intra
 * \date
 *    2003.4
 ************************************************************************
 */
void   decide_intrabk_SAD()
{
  if (img->type != I_SLICE)
  {
    if (img->pix_x == 0 && img->pix_y == 0)
    {
      flag_intra_SAD = 0;
    }
    else if (img->pix_x == 0)
    {
      flag_intra_SAD = flag_intra[(img->pix_x)>>4];
    }
    else if (img->pix_y == 0)
    {
      flag_intra_SAD = flag_intra[((img->pix_x)>>4)-1];
    }
    else 
    {
      flag_intra_SAD = ((flag_intra[(img->pix_x)>>4])||(flag_intra[((img->pix_x)>>4)-1])||(flag_intra[((img->pix_x)>>4)+1])) ;
    }
  }
  return;
}

void skip_intrabk_SAD(int best_mode, int ref_max)
{
  int i,j,k, ref;
  if (img->number > 0) 
    flag_intra[(img->pix_x)>>4] = (best_mode == 9 || best_mode == 10) ? 1:0;
  if (img->type != I_SLICE  && (best_mode == 9 || best_mode == 10))
  {
    for (i=0; i < 4; i++)
    {
      for (j=0; j < 4; j++)
      {
        for (k=0; k < 9;k++)
        {
	  fastme_l0_cost[k][j][i] = 0;
	  fastme_l1_cost[k][j][i] = 0;
          for (ref=0; ref<ref_max;ref++)
          {
            fastme_ref_cost[ref][k][j][i] = 0;
          }
        }
      }
    }
  
  }
  return;
}


void setup_FME(short ref, int list, int block_y, int block_x, int blocktype, short   ******all_mv)
{
  int N_Bframe=0;
  int n_Bframe=0;
  int temp_blocktype = 0;
  int indication_blocktype[8]={0,0,1,1,2,4,4,5};
  N_Bframe = input->successive_Bframe;
  n_Bframe =(N_Bframe) ? (frame_ctr[B_SLICE]%(N_Bframe+1)): 0;


  /**************************** MV prediction **********************/ 
  //MV uplayer prediction
  if (blocktype>1) 
  {
    temp_blocktype = indication_blocktype[blocktype];
    pred_MV_uplayer[0] = all_mv[block_y][block_x][list][ref][temp_blocktype][0];
    pred_MV_uplayer[1] = all_mv[block_y][block_x][list][ref][temp_blocktype][1];

  }


  //MV ref-frame prediction
  pred_MV_ref_flag = 0;
  if(list==0)
  {
    if (img->field_picture) 
    {
      if ( ref > 1)
      {
        pred_MV_ref[0] = all_mv[block_y][block_x][0][ref-2][blocktype][0];
        pred_MV_ref[0] = (int)(pred_MV_ref[0]*((ref>>1)+1)/(float)((ref>>1)));
        pred_MV_ref[1] = all_mv[block_y][block_x][0][ref-2][blocktype][1];
        pred_MV_ref[1] = (int)(pred_MV_ref[1]*((ref>>1)+1)/(float)((ref>>1)));
        pred_MV_ref_flag = 1;
      }
      if (img->type == B_SLICE &&  (ref==0 || ref==1) )
      {
        pred_MV_ref[0] =(int) (all_mv[block_y][block_x][1][0][blocktype][0]*(-n_Bframe)/(N_Bframe-n_Bframe+1.0f));
        pred_MV_ref[1] =(int) (all_mv[block_y][block_x][1][0][blocktype][1]*(-n_Bframe)/(N_Bframe-n_Bframe+1.0f));
        pred_MV_ref_flag = 1;
      }
    }
    else //frame case
    {
      if ( ref > 0)
      {
        pred_MV_ref[0] = all_mv[block_y][block_x][0][ref-1][blocktype][0];
        pred_MV_ref[0] = (int)(pred_MV_ref[0]*(ref+1)/(float)(ref));
        pred_MV_ref[1] = all_mv[block_y][block_x][0][ref-1][blocktype][1];
        pred_MV_ref[1] = (int)(pred_MV_ref[1]*(ref+1)/(float)(ref));
        pred_MV_ref_flag = 1;
      }
      if (img->type == B_SLICE && (ref==0)) //B frame forward prediction, first ref
      {
        pred_MV_ref[0] =(int) (all_mv[block_y][block_x][1][0][blocktype][0]*(-n_Bframe)/(N_Bframe-n_Bframe+1.0f));
        pred_MV_ref[1] =(int) (all_mv[block_y][block_x][1][0][blocktype][1]*(-n_Bframe)/(N_Bframe-n_Bframe+1.0f));
        pred_MV_ref_flag = 1;
      }
    }
  }
  /******************************SAD prediction**********************************/
  if (list==0 && ref>0)  //pred_SAD_ref
  {

    if (flag_intra_SAD) //add this for irregular motion
    {
      pred_SAD = 0;
    }
    else 
    {
      if (img->field_picture)
      {
        if (ref > 1)
        {
          pred_SAD = fastme_ref_cost[ref-2][blocktype][block_y][block_x];
        }
        else
        {
          pred_SAD = fastme_ref_cost[0][blocktype][block_y][block_x];
        }
      }
      else
      {
        pred_SAD = fastme_ref_cost[ref-1][blocktype][block_y][block_x];
      }

    }
  }
  else if (blocktype>1)  // pred_SAD_uplayer
  {
    if (flag_intra_SAD) 
    {
      pred_SAD = 0;
    }