mc_prediction.c

H.264基于baseline解码器的C++实现源码

        }
      }
      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;
          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 */
        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,(dy == 1 ? y_pos+j : y_pos+j+1))];

            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));
          }
        }

        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 = dx == 1 ? x_pos+i : x_pos+i+1;
            pres_x = iClip3(0, maxold_x, pres_x);

            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) ((*orig_line + iClip1(max_imgpel_value, ((result+16)>>5)) + 1 ) >> 1);
            orig_line++;
          }
        }      
      }
    }
  }
}

void get_block_chroma(int uv, StorablePicture *curr_ref, int x_pos, int y_pos, int hor_block_size, int ver_block_size, ImageParameters *img, imgpel **block)
{
  int subpel_x    = img->mb_cr_size_x == 8 ? 7 : 3;
  int subpel_y    = img->mb_cr_size_y == 8 ? 7 : 3;
  int shiftpel_x  = img->mb_cr_size_x == 8 ? 3 : 2;
  int shiftpel_y  = img->mb_cr_size_y == 8 ? 3 : 2;
  int total_scale = shiftpel_x + shiftpel_y;

  int dx = (x_pos & subpel_x);
  int dy = (y_pos & subpel_y);
  int dxcur = (subpel_x + 1 - dx);
  int dycur = (subpel_y + 1 - dy);

  int w00 = dxcur * dycur;
  int w01 = dxcur * dy;
  int w10 = dx * dycur;
  int w11 = dx * dy;

  int i, j;
  int maxold_x = dec_picture->size_x_cr_m1;
  int maxold_y = (dec_picture->motion.mb_field[img->current_mb_nr]) ? (dec_picture->size_y_cr >> 1) - 1 : dec_picture->size_y_cr_m1;
  int result;
  
  static imgpel **cur_img, *blk_line;
  static imgpel *cur_line, *cur_line_p1;
  int tmp_pos;
  static int ipos, ipos_p1;
  int    max_imgpel_value = img->max_imgpel_value_comp[uv + 1];


  if (curr_ref == no_reference_picture && img->framepoc < img->recovery_poc)
  {
    printf("list[ref_frame] is equal to 'no reference picture' before RAP\n");

    /* fill the block with sample value 128 */
    for (j = 0; j < ver_block_size; j++)
      for (i = 0; i < hor_block_size; i++)
        block[j][i] = 128;
    return;
  }

  cur_img = curr_ref->imgUV[uv];

  x_pos = x_pos >> shiftpel_x;
  y_pos = y_pos >> shiftpel_y;

  if ((y_pos >= 0) && (y_pos < maxold_y - ver_block_size) && (x_pos >= 0) && (x_pos < maxold_x - hor_block_size))
  {
    if (dx == 0 && dy == 0)
    {  /* fullpel position */
      for (j = 0; j < ver_block_size; j++)
      {        
        memcpy(&(block[j][0]), &(cur_img[ y_pos + j ][x_pos]), hor_block_size * sizeof(imgpel));
      }
    }
    else if (dx == 0)
    { 
      for (j = 0; j < ver_block_size; j++)
      {
        cur_line    = &cur_img[y_pos + j    ][x_pos];
        cur_line_p1 = &cur_img[y_pos + j + 1][x_pos];
        blk_line = block[j];

        for (i = 0; i < hor_block_size; i++)
        {
          result = (w00 * *cur_line++ + w01 * *cur_line_p1++);
          *(blk_line++) = (imgpel) iClip1(max_imgpel_value, rshift_rnd_sf(result, total_scale));
        }
      }
    }
    else if (dy == 0)
    { 
      for (j = 0; j < ver_block_size; j++)
      {
        cur_line    = &cur_img[y_pos + j][x_pos];
        cur_line_p1 = cur_line + 1;
        blk_line = block[j];

        for (i = 0; i < hor_block_size; i++)
        {
          result = (w00 * *cur_line++ + w10 * *cur_line_p1++);
          *(blk_line++) = (imgpel) iClip1(max_imgpel_value, rshift_rnd_sf(result, total_scale));
        }
      }
    }
    else
    { /* other positions */
      for (j = 0; j < ver_block_size; j++)
      {
        cur_line    = &cur_img[y_pos + j    ][x_pos];
        cur_line_p1 = &cur_img[y_pos + j + 1][x_pos];
        blk_line = block[j];

        for (i = 0; i < hor_block_size; i++)
        {
          result  = (w00 * *(cur_line++) + w01 * *(cur_line_p1++));
          result += (w10 * *(cur_line  ) + w11 * *(cur_line_p1  ));
          *(blk_line++) = (imgpel) iClip1(max_imgpel_value, rshift_rnd_sf(result, total_scale));
        }
      }
    }
  }