mc_prediction.c

H.264编码实现

          x1 = tmp_res[j + 1];
          x2 = tmp_res[j + 2];
          x3 = tmp_res[j + 3];
          x4 = tmp_res[j + 4];
          x5 = tmp_res[j + 5];
          orig_line = block[j];

          for (i = 0; i < hor_block_size; i++)
          {
            result  = (*x0++ + *x5++) * COEF[0]
                    + (*x1++ + *x4++) * COEF[1]
                    + (*x2++ + *x3++) * COEF[2];

            *(orig_line++) = (imgpel) iClip1(max_imgpel_value, ((result+512)>>10));
          }
        }

        if ((dy&1) == 1)
        {
          jj = 2 + (dy>>1);
          for (j = 0; j < ver_block_size; j++)
          {            
            tmp_line  = tmp_res[jj++];
            orig_line = block[j];
            for (i = 0; i < hor_block_size; i++)
            {
              *orig_line = (imgpel) ((*orig_line + iClip1(max_imgpel_value, ((*(tmp_line++) + 16) >> 5)) + 1 )>> 1);
              orig_line++;
            }
          }
        }
      }
      else if (dy == 2)
      {  /* Horizontal & vertical interpolation */
        p0 = &(cur_imgY[ -2][x_pos - 2]);
        for (j = 0; j < ver_block_size; j++)
        {                    
          p1 = p0 + shift_x;
          p2 = p1 + shift_x;
          p3 = p2 + shift_x;
          p4 = p3 + shift_x;
          p5 = p4 + shift_x;
          tmp_line  = tmp_res[j];

          for (i = 0; i < hor_block_size + 5; i++)
          {
            *(tmp_line++)  = (*(p0++) + *(p5++)) * COEF[0]
                           + (*(p1++) + *(p4++)) * COEF[1]
                           + (*(p2++) + *(p3++)) * COEF[2];
          }
          p0 = p1 - (hor_block_size + 5);
        }

        for (j = 0; j < ver_block_size; j++)
        {
          orig_line = block[j];
          x0 = tmp_res[j];
          x1 = x0 + 1;
          x2 = x1 + 1;
          x3 = x2 + 1;
          x4 = x3 + 1;
          x5 = x4 + 1;

          for (i = 0; i < hor_block_size; i++)
          {
            result  = (*(x0++) + *(x5++)) * COEF[0]
                    + (*(x1++) + *(x4++)) * COEF[1]
                    + (*(x2++) + *(x3++)) * COEF[2];

            *(orig_line++) = (imgpel) iClip1(max_imgpel_value, ((result + 512)>>10));
          }
        }

        if ((dx&1) == 1)
        {
          for (j = 0; j < ver_block_size; j++)
          {
            tmp_line  = &tmp_res[j][2 + (dx>>1)];
            orig_line = block[j];
            for (i = 0; i < hor_block_size; i++)
            {
              *orig_line = (imgpel) ((*orig_line + iClip1(max_imgpel_value, ((*(tmp_line++) + 16)>>5))+1)>>1);
              orig_line ++;
            }
          }
        }
      }
      else
      {  /* Diagonal interpolation */
        jj = (dy == 1 ? 0 : 1);

        for (j = 0; j < ver_block_size; j++)
        {
          p0 = &cur_imgY[jj++][x_pos - 2];
          p1 = p0 + 1;
          p2 = p1 + 1;
          p3 = p2 + 1;
          p4 = p3 + 1;
          p5 = p4 + 1;

          orig_line = block[j];

          for (i = 0; i < hor_block_size; i++)
          {        
            result  = (*(p0++) + *(p5++)) * COEF[0]
                    + (*(p1++) + *(p4++)) * COEF[1]
                    + (*(p2++) + *(p3++)) * COEF[2];

            *(orig_line++) = (imgpel) iClip1(max_imgpel_value, ((result + 16)>>5));
          }
        }

        p0 = &(cur_imgY[-2][(dx == 1 ? x_pos : x_pos + 1)]);
        for (j = 0; j < ver_block_size; j++)
        {        
          p1 = p0 + shift_x;
          p2 = p1 + shift_x;
          p3 = p2 + shift_x;
          p4 = p3 + shift_x;
          p5 = p4 + shift_x;
          orig_line = block[j];

          for (i = 0; i < hor_block_size; i++)
          {
            result  = (*(p0++) + *(p5++)) * COEF[0]
                    + (*(p1++) + *(p4++)) * COEF[1]
                    + (*(p2++) + *(p3++)) * COEF[2];

            *orig_line = (imgpel) ((*orig_line + iClip1(max_imgpel_value, ((result + 16) >> 5)) + 1) >> 1);
            orig_line++;
          }
          p0 = p1 - hor_block_size ;
        }      
      }
    }
  }
  else // unsafe positions
  {
    if (dx == 0 && dy == 0)
    {  /* fullpel position */
      for (j = 0; j < ver_block_size; j++)
      {
        cur_lineY = cur_imgY[iClip3(0, maxold_y, y_pos + j)];
        orig_line = block[j];
        for (i = 0; i < hor_block_size; i++)
        {
          *(orig_line++) = cur_lineY[iClip3(0, maxold_x, x_pos + i )];
        }
      }
    }
    else
    { /* other positions */

      if (dy == 0)
      { /* No vertical interpolation */
        tmp_pos = x_pos - 2;
        for (i = 0; i < hor_block_size; i++)
        {        
          ipos_m2 = iClip3(0, maxold_x, tmp_pos++);
          ipos_m1 = iClip3(0, maxold_x, tmp_pos++);
          ipos    = iClip3(0, maxold_x, tmp_pos++);
          ipos_p1 = iClip3(0, maxold_x, tmp_pos++);
          ipos_p2 = iClip3(0, maxold_x, tmp_pos++);
          ipos_p3 = iClip3(0, maxold_x, tmp_pos  );
          tmp_pos -= 4;          

          for (j = 0; j < ver_block_size; j++)
          {
            cur_lineY = cur_imgY[iClip3(0,maxold_y,y_pos+j)];

            result  = (cur_lineY[ipos_m2] + cur_lineY[ipos_p3]) * COEF[0];
            result += (cur_lineY[ipos_m1] + cur_lineY[ipos_p2]) * COEF[1];
            result += (cur_lineY[ipos   ] + cur_lineY[ipos_p1]) * COEF[2];

            block[j][i] = (imgpel) iClip1(max_imgpel_value, ((result + 16)>>5));
          }
        }

        if ((dx&1) == 1)
        {
          for (j = 0; j < ver_block_size; j++)
          {
            cur_lineY = cur_imgY[iClip3(0,maxold_y,y_pos+j)];
            orig_line = block[j];
            for (i = 0; i < hor_block_size; i++)
            {
              *orig_line = (imgpel) ((*orig_line + cur_lineY[iClip3(0, maxold_x, x_pos + i + (dx >> 1))] + 1 ) >> 1);
              orig_line++;
            }
          }
        }
      }
      else if (dx == 0)
      {  /* No horizontal interpolation */
        tmp_pos = y_pos - 2;
        for (j = 0; j < ver_block_size; j++)
        { 
          p0 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p1 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p2 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p3 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p4 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p5 = cur_imgY[iClip3(0, maxold_y, tmp_pos  )];
          
          tmp_pos -= 4;
          orig_line = block[j];

          for (i = 0; i < hor_block_size; i++)
          {
            pres_x = iClip3(0,maxold_x,x_pos+i);

            result  = (p0[pres_x] + p5[pres_x]) * COEF[0];
            result += (p1[pres_x] + p4[pres_x]) * COEF[1];
            result += (p2[pres_x] + p3[pres_x]) * COEF[2];
            *(orig_line++) = (imgpel) iClip1(max_imgpel_value, ((result+16)>>5));
          }
        }

        if ((dy&1) == 1)
        {
          for (j = 0; j < ver_block_size; j++)
          {
            cur_lineY = cur_imgY[iClip3(0,maxold_y,y_pos+j+(dy>>1))];
            orig_line = block[j];
            for (i = 0; i < hor_block_size; i++)
            {
              *orig_line = (imgpel) ((*orig_line + cur_lineY[iClip3(0, maxold_x, x_pos + i)] + 1 ) >> 1);
              orig_line++;
            }
          }
        }
      }
      else if (dx == 2)
      {  /* Vertical & horizontal interpolation */
        tmp_pos = x_pos - 2;
        for (i = 0; i < hor_block_size; i++)
        {        
          ipos_m2 = iClip3(0, maxold_x, tmp_pos++);
          ipos_m1 = iClip3(0, maxold_x, tmp_pos++);
          ipos    = iClip3(0, maxold_x, tmp_pos++);
          ipos_p1 = iClip3(0, maxold_x, tmp_pos++);
          ipos_p2 = iClip3(0, maxold_x, tmp_pos++);
          ipos_p3 = iClip3(0, maxold_x, tmp_pos  );
          tmp_pos -= 4;

          for (j = 0; j < ver_block_size + 5; j++)
          {
            cur_lineY = cur_imgY[iClip3(0,maxold_y,y_pos + j - 2)];

            tmp_res[j][i]  = (cur_lineY[ipos_m2] + cur_lineY[ipos_p3]) * COEF[0];
            tmp_res[j][i] += (cur_lineY[ipos_m1] + cur_lineY[ipos_p2]) * COEF[1];
            tmp_res[j][i] += (cur_lineY[ipos   ] + cur_lineY[ipos_p1]) * COEF[2];
          }
        }

        for (j = 0; j < ver_block_size; j++)
        {
          x0 = tmp_res[j    ];
          x1 = tmp_res[j + 1];
          x2 = tmp_res[j + 2];
          x3 = tmp_res[j + 3];
          x4 = tmp_res[j + 4];
          x5 = tmp_res[j + 5];
          orig_line = block[j];

          for (i = 0; i < hor_block_size; i++)
          {
            result  = (*x0++ + *x5++) * COEF[0]
                    + (*x1++ + *x4++) * COEF[1]
                    + (*x2++ + *x3++) * COEF[2];

            *(orig_line++) = (imgpel) iClip1(max_imgpel_value, ((result+512)>>10));
          }
        }

        if ((dy&1) == 1)
        {
          for (j = 0; j < ver_block_size; j++)
          {
            tmp_line  = tmp_res[j + 2 + (dy>>1)];            
            orig_line = block[j];
            for (i = 0; i < hor_block_size; i++)
            {
              *orig_line = (imgpel) ((*orig_line + iClip1(max_imgpel_value, ((*(tmp_line++) + 16) >> 5)) + 1 )>>1);
              orig_line++;
            }
          }
        }
      }
      else if (dy == 2)
      {  /* Horizontal & vertical interpolation */

        tmp_pos = y_pos - 2;
        for (j = 0; j < ver_block_size; j++)
        {
          p0 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p1 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p2 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p3 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p4 = cur_imgY[iClip3(0, maxold_y, tmp_pos++)];
          p5 = cur_imgY[iClip3(0, maxold_y, tmp_pos  )];

          tmp_pos -= 4;

          for (i = 0; i < hor_block_size + 5; i++)
          {
            pres_x = iClip3(0,maxold_x, x_pos + i - 2);
            result  = (p0[pres_x] + p5[pres_x])*COEF[0]
                    + (p1[pres_x] + p4[pres_x])*COEF[1]
                    + (p2[pres_x] + p3[pres_x])*COEF[2];
            tmp_res[j][i] = result;
          }
        }

        for (j = 0; j < ver_block_size; j++)
        {
          orig_line = block[j];
          tmp_line  = tmp_res[j];
          x0 = tmp_res[j];
          x1 = x0 + 1;