block.c

H.264视频编解码的标准测试模型

        for(i=0; i<4; i++)
          m1[i]=(m5[i] * invlevelscale[qp_rem][0][0])<<(qp_per-5);
      }

      img->m7[0][0] = m1[0];
      img->m7[0][4] = m1[1];
      img->m7[4][0] = m1[2];
      img->m7[4][4] = m1[3];
    }
  }
  else if(yuv == YUV422)
  {
    //================== CHROMA DC YUV422 ===================
    //transform DC coeff
    //horizontal
    
    //pick out DC coeff
    for (j=0; j < img->mb_cr_size_y; j+=BLOCK_SIZE)
    {
      for (i=0; i < img->mb_cr_size_x; i+=BLOCK_SIZE)
        m3[i>>2][j>>2]= img->m7[j][i];
    } 
    //horizontal
    if(!lossless_qpprime)
    {
      m4[0][0] = m3[0][0] + m3[1][0];
      m4[0][1] = m3[0][1] + m3[1][1];
      m4[0][2] = m3[0][2] + m3[1][2];
      m4[0][3] = m3[0][3] + m3[1][3];
      
      m4[1][0] = m3[0][0] - m3[1][0];
      m4[1][1] = m3[0][1] - m3[1][1];
      m4[1][2] = m3[0][2] - m3[1][2];
      m4[1][3] = m3[0][3] - m3[1][3];
      
      // vertical
      for (i=0;i<2;i++)
      {
        m5[0] = m4[i][0] + m4[i][3];
        m5[1] = m4[i][1] + m4[i][2];
        m5[2] = m4[i][1] - m4[i][2];
        m5[3] = m4[i][0] - m4[i][3];
        
        m4[i][0] = (m5[0] + m5[1]);
        m4[i][2] = (m5[0] - m5[1]);
        m4[i][1] = (m5[3] + m5[2]);
        m4[i][3] = (m5[3] - m5[2]);
      }
    }
    
    run=-1;
    scan_pos=0;
    
    //quant of chroma DC-coeffs
    for (coeff_ctr=0;coeff_ctr<8;coeff_ctr++)
    {
      i=SCAN_YUV422[coeff_ctr][0];
      j=SCAN_YUV422[coeff_ctr][1];
      
      run++;

      if(lossless_qpprime)
      {
        level = absm(m3[i][j]);
        m4[i][j]=m3[i][j];
      }
      else 
        level =(absm(m4[i][j]) * levelscale[qp_rem_dc][0][0] + (leveloffset[qp_per_dc][0][0]*2)) >> (q_bits_422+1);

      if (level != 0)
      {
        //YUV422
        currMB->cbp_blk |= 0xff0000 << (uv << 3) ;   // if one of the DC levels is != 0 set the
        cr_cbp=max(1,cr_cbp);												   // coded-bit all 4 4x4 blocks (bit 16-31 or 32-47) //YUV444
        DCcoded = 1 ;
        
        DCLevel[scan_pos] = sign(level,m4[i][j]);
        DCRun  [scan_pos] = run;
        ++scan_pos;
        run=-1;
      }
      if(!lossless_qpprime)
        m3[i][j]=sign(level,m4[i][j]);
    }
    DCLevel[scan_pos]=0;

    //inverse DC transform
    //horizontal
    if(!lossless_qpprime)
    {
      m4[0][0] = m3[0][0] + m3[1][0];
      m4[0][1] = m3[0][1] + m3[1][1];
      m4[0][2] = m3[0][2] + m3[1][2];
      m4[0][3] = m3[0][3] + m3[1][3];
      
      m4[1][0] = m3[0][0] - m3[1][0];
      m4[1][1] = m3[0][1] - m3[1][1];
      m4[1][2] = m3[0][2] - m3[1][2];
      m4[1][3] = m3[0][3] - m3[1][3];      
      
      // vertical
      for (i=0;i<2;i++)
      {       
        m6[0]=m4[i][0]+m4[i][2];
        m6[1]=m4[i][0]-m4[i][2];
        m6[2]=m4[i][1]-m4[i][3];
        m6[3]=m4[i][1]+m4[i][3];
        
        if(qp_per_dc<4)
        {          
          img->m7[0 ][i*4]=((((m6[0]+m6[3])*invlevelscale[qp_rem_dc][0][0]+(1<<(3-qp_per_dc)))>>(4-qp_per_dc))+2)>>2;
          img->m7[4 ][i*4]=((((m6[1]+m6[2])*invlevelscale[qp_rem_dc][0][0]+(1<<(3-qp_per_dc)))>>(4-qp_per_dc))+2)>>2;
          img->m7[8 ][i*4]=((((m6[1]-m6[2])*invlevelscale[qp_rem_dc][0][0]+(1<<(3-qp_per_dc)))>>(4-qp_per_dc))+2)>>2;
          img->m7[12][i*4]=((((m6[0]-m6[3])*invlevelscale[qp_rem_dc][0][0]+(1<<(3-qp_per_dc)))>>(4-qp_per_dc))+2)>>2;
        }
        else
        {
          img->m7[0 ][i*4]=((((m6[0]+m6[3])*invlevelscale[qp_rem_dc][0][0])<<(qp_per_dc-4))+2)>>2;
          img->m7[4 ][i*4]=((((m6[1]+m6[2])*invlevelscale[qp_rem_dc][0][0])<<(qp_per_dc-4))+2)>>2;
          img->m7[8 ][i*4]=((((m6[1]-m6[2])*invlevelscale[qp_rem_dc][0][0])<<(qp_per_dc-4))+2)>>2;
          img->m7[12][i*4]=((((m6[0]-m6[3])*invlevelscale[qp_rem_dc][0][0])<<(qp_per_dc-4))+2)>>2;
        }
      }//for (i=0;i<2;i++)    
    }
  }
  else if(yuv == YUV444)
  {
    //================== CHROMA DC YUV444 ===================
    //transform DC coeff
    //pick out DC coeff
    for (j=0; j < img->mb_cr_size_y; j+=BLOCK_SIZE)
    {
      for (i=0; i < img->mb_cr_size_x; i+=BLOCK_SIZE)
        m4[i>>2][j>>2]= img->m7[j][i];
    }
    
    //horizontal
    for (j=0;j<4 && !lossless_qpprime;j++)
    {
      m5[0] = m4[0][j] + m4[3][j];
      m5[1] = m4[1][j] + m4[2][j];
      m5[2] = m4[1][j] - m4[2][j];
      m5[3] = m4[0][j] - m4[3][j];
      
      m4[0][j]=m5[0]+m5[1];
      m4[2][j]=m5[0]-m5[1];
      m4[1][j]=m5[3]+m5[2];
      m4[3][j]=m5[3]-m5[2];
    }
    // vertical
    for (i=0;i<4 && !lossless_qpprime;i++)
    {
      m5[0] = m4[i][0] + m4[i][3];
      m5[1] = m4[i][1] + m4[i][2];
      m5[2] = m4[i][1] - m4[i][2];
      m5[3] = m4[i][0] - m4[i][3];

      m4[i][0]=(m5[0]+m5[1])>>1;
      m4[i][2]=(m5[0]-m5[1])>>1;
      m4[i][1]=(m5[3]+m5[2])>>1;
      m4[i][3]=(m5[3]-m5[2])>>1;
    }

    run=-1;
    scan_pos=0;
    
    //quant of chroma DC-coeffs
    for (coeff_ctr=0;coeff_ctr<16;coeff_ctr++)
    {
      i=SNGL_SCAN[coeff_ctr][0];
      j=SNGL_SCAN[coeff_ctr][1];
      
      run++;
      
      if(lossless_qpprime)
        level = absm(m4[i][j]);
      else 
        level =(absm(m4[i][j]) * levelscale[qp_rem][0][0] + (leveloffset[qp_per][0][0]*2)) >> (q_bits+1);
      
      if (level != 0)
      {
        //YUV444
        currMB->cbp_blk |= ((int64)0xffff0000) << (uv << 4) ;   // if one of the DC levels is != 0 set the
        cr_cbp=max(1,cr_cbp);												   // coded-bit all 4 4x4 blocks (bit 16-31 or 32-47) //YUV444
        DCcoded = 1 ;
        
        DCLevel[scan_pos] = sign(level,m4[i][j]);
        DCRun  [scan_pos] = run;
        ++scan_pos;
        run=-1;
      }
      if(!lossless_qpprime)
        m4[i][j]=sign(level,m4[i][j]);
    }
    DCLevel[scan_pos]=0;

    // inverse DC transform
    //horizontal
    for (j=0;j<4 && !lossless_qpprime;j++)
    {     
      m6[0] = m4[0][j] + m4[2][j];
      m6[1] = m4[0][j] - m4[2][j];
      m6[2] = m4[1][j] - m4[3][j];
      m6[3] = m4[1][j] + m4[3][j];
      
      m4[0][j] = m6[0] + m6[3];
      m4[1][j] = m6[1] + m6[2];
      m4[2][j] = m6[1] - m6[2];
      m4[3][j] = m6[0] - m6[3];
    }
    
    //vertical
    for (i=0;i<4 && !lossless_qpprime;i++)
    {
      m6[0]=m4[i][0]+m4[i][2];
      m6[1]=m4[i][0]-m4[i][2];
      m6[2]=m4[i][1]-m4[i][3];
      m6[3]=m4[i][1]+m4[i][3];

      if(qp_per<4)
      {
        img->m7[0 ][i*4] = ((((m6[0] + m6[3])*invlevelscale[qp_rem][0][0]+(1<<(3-qp_per)))>>(4-qp_per))+2)>>2;
        img->m7[4 ][i*4] = ((((m6[1] + m6[2])*invlevelscale[qp_rem][0][0]+(1<<(3-qp_per)))>>(4-qp_per))+2)>>2;
        img->m7[8 ][i*4] = ((((m6[1] - m6[2])*invlevelscale[qp_rem][0][0]+(1<<(3-qp_per)))>>(4-qp_per))+2)>>2;
        img->m7[12][i*4] = ((((m6[0] - m6[3])*invlevelscale[qp_rem][0][0]+(1<<(3-qp_per)))>>(4-qp_per))+2)>>2;
      }
      else
      {
        img->m7[0 ][i*4] = ((((m6[0]+m6[3])*invlevelscale[qp_rem][0][0])<<(qp_per-4))+2)>>2;
        img->m7[4 ][i*4] = ((((m6[1]+m6[2])*invlevelscale[qp_rem][0][0])<<(qp_per-4))+2)>>2;
        img->m7[8 ][i*4] = ((((m6[1]-m6[2])*invlevelscale[qp_rem][0][0])<<(qp_per-4))+2)>>2;
        img->m7[12][i*4] = ((((m6[0]-m6[3])*invlevelscale[qp_rem][0][0])<<(qp_per-4))+2)>>2;
      }
    }
  }

  //     Quant of chroma AC-coeffs.
  coeff_cost=0;
  cr_cbp_tmp=0;

  for (b8=0; b8 < (img->num_blk8x8_uv >> 1); b8++)
  {
    for (b4=0; b4 < 4; b4++)
    {
      n1 = hor_offset[yuv][b8][b4];
      n2 = ver_offset[yuv][b8][b4];
      ACLevel = img->cofAC[4+b8+uv_scale][b4][0];
      ACRun   = img->cofAC[4+b8+uv_scale][b4][1];
      run=-1;
      scan_pos=0;

      for (coeff_ctr=1; coeff_ctr < 16; coeff_ctr++)// start change rd_quant
      {

        i=pos_scan[coeff_ctr][0];
        j=pos_scan[coeff_ctr][1];

        ++run;
        ilev=0;

        if(lossless_qpprime)
          level = absm(img->m7[n2+j][n1+i]);
        else 
          level=(absm(img->m7[n2+j][n1+i])*levelscale[qp_rem][i][j]+leveloffset[qp_per][i][j])>>q_bits;

        if (img->AdaptiveRounding)
        {
          if (lossless_qpprime || level == 0 )
          {
            img->fadjust4x4Cr[intra][uv][n2+j][n1+i] = 0;
          }
          else
          {
            img->fadjust4x4Cr[intra][uv][n2+j][n1+i] = 
              (AdaptRndWeight * (absm(img->m7[n2+j][n1+i]) * levelscale[qp_rem][i][j] - (level << q_bits)) + (1<< (q_bits))) >> (q_bits + 1); 
          }          
        }

        if (level  != 0)
        {
          currMB->cbp_blk |= ((int64)1) << cbp_blk_chroma[b8 + uv_scale][b4];
          if (level > 1 || lossless_qpprime)
            coeff_cost += MAX_VALUE;                // set high cost, shall not be discarded
          else
            coeff_cost += COEFF_COST[input->disthres][run];

          cr_cbp_tmp=2;
          ACLevel[scan_pos] = sign(level,img->m7[n2+j][n1+i]);
          ACRun  [scan_pos] = run;
          ++scan_pos;
          run=-1;

          level=sign(level, img->m7[n2+j][n1+i]);
          if(lossless_qpprime)
          {
            ilev = level;
          }
          else if(qp_per<4)
          {
            ilev=(level*invlevelscale[qp_rem][i][j]+(1<<(3-qp_per)))>>(4-qp_per);
          }
          else
          {
            ilev=(level*invlevelscale[qp_rem][i][j])<<(qp_per-4);
          }
        }
        if(!lossless_qpprime)
          img->m7[n2+j][n1+i]=ilev;
      }
      ACLevel[scan_pos] = 0;
    }
  }

  // * reset chroma coeffs
  if(coeff_cost < _CHROMA_COEFF_COST_ && !lossless_qpprime)
  {
    cr_cbp_tmp = 0 ;
    
    for (b8=0; b8 < (img->num_blk8x8_uv >> 1); b8++)
    {
      for (b4=0; b4 < 4; b4++)
      {
        n1 = hor_offset[yuv][b8][b4];
        n2 = ver_offset[yuv][b8][b4];
        ACLevel = img->cofAC[4+b8+uv_scale][b4][0];
        ACRun   = img->cofAC[4+b8+uv_scale][b4][1];
        if( DCcoded == 0) 
          currMB->cbp_blk &= ~((int64)cbpblk_pattern[yuv] << (uv << (1+yuv)));  // if no chroma DC's: then reset coded-bits of this chroma subblock
        
        ACLevel[0] = 0;
        for (coeff_ctr=1; coeff_ctr < 16; coeff_ctr++)// ac coeff
        {
          i=pos_scan[coeff_ctr][0];
          j=pos_scan[coeff_ctr][1];
          
          img->m7[n2+j][n1+i]=0;
          ACLevel[coeff_ctr] = 0;
        }
      }
    }
  }

  if(cr_cbp_tmp==2)   
    cr_cbp = 2;
  
  //     IDCT.
  //     Horizontal.
  for (n2=0; n2 < img->mb_cr_size_y && !lossless_qpprime; n2 += BLOCK_SIZE)
  {
    for (n1=0; n1 < img->mb_cr_size_x; n1 += BLOCK_SIZE)
    {
      for (j=0; j < BLOCK_SIZE; j++)
      {
        j2 = n2 + j;
        for (i=0; i < BLOCK_SIZE; i++)
        {
          m5[i]=img->m7[j2][n1+i];
        }

        m6[0] = (m5[0]     +  m5[2]);
        m6[1] = (m5[0]     -  m5[2]);
        m6[2] = (m5[1]>>1) -  m5[3];
        m6[3] =  m5[1]     + (m5[3]>>1);

        img->m7[j2][n1  ] = m6[0] + m6[3];
        img->m7[j2][n1+1] = m6[1] + m6[2];
        img->m7[j2][n1+2] = m6[1] - m6[2];
        img->m7[j2][n1+3] = m6[0] - m6[3];
      }

      //     Vertical.
      for (i=0; i < BLOCK_SIZE && !lossless_qpprime; i++)
      {
        i1 = n1 + i;
        for (j=0; j < BLOCK_SIZE; j++)
        {
          m5[j]=img->m7[n2+j][i1];
        }
        m6[0]=(m5[0]+m5[2]);
        m6[1]=(m5[0]-m5[2]);
        m6[2]=(m5[1]>>1)-m5[3];
        m6[3]=m5[1]+(m5[3]>>1);

          // Residue Color Transform
        if (!img->residue_transform_flag)
        {
          img->m7[n2  ][i1] = min(img->max_imgpel_value_uv,max(0,(m6[0]+m6[3]+((long)img->mpr[n2  ][i1] << DQ_BITS)+DQ_ROUND)>>DQ_BITS));
          img->m7[n2+1][i1] = min(img->max_imgpel_value_uv,max(0,(m6[1]+m6[2]+((long)img->mpr[n2+1][i1] << DQ_BITS)+DQ_ROUND)>>DQ_BITS));
          img->m7[n2+2][i1] = min(img->max_imgpel_value_uv,max(0,(m6[1]-m6[2]+((long)img->mpr[n2+2][i1] << DQ_BITS)+DQ_ROUND)>>DQ_BITS));
          img->m7[n2+3][i1] = min(img->max_imgpel_value_uv,max(0,(m6[0]-m6[3]+((long)img->mpr[n2+3][i1] << DQ_BITS)+DQ_ROUND)>>DQ_BITS));
        } 
        else 
        {
          if(lossless_qpprime)
          {
            img->m7[n2  ][i1] = m6[0]+m6[3];
            img->m7[n2+1][i1] = m6[1]+m6[2];
            img->m7[n2+2][i1] = m6[1]-m6[2];
            img->m7[n2+3][i1] = m6[0]-m6[3];
          }
          else
          {
            img->m7[n2  ][i1] = (m6[0]+m6[3]+DQ_ROUND)>>DQ_BITS;
            img->m7[n2+1][i1] = (m6[1]+m6[2]+DQ_ROUND)>>DQ_BITS;
            img->m7[n2+2][i1] = (m6[1]-m6[2]+DQ_ROUND)>>DQ_BITS;
            img->m7[n2+3][i1] = (m6[0]-m6[3]+DQ_ROUND)>>DQ_BITS;
          }
        }
      }
    }
  }

  //  Decoded block moved to memory
  if (!img->residue_transform_flag)
  {
    for (j=0; j < img->mb_cr_size_y; j++)
    {
      pix_c_y = img->pix_c_y+j;
      for (i=0; i < img->mb_cr_size_x; i++)
      {
        pix_c_x = img->pix_c_x+i;
        if(lossless_qpprime)
          enc_picture->imgUV[uv][pix_c_y][pix_c_x]= img->m7[j][i]+img->mpr[j][i];
        else
          enc_picture->imgUV[uv][pix_c_y][pix_c_x]= img->m7[j][i];
      }
    }
  }
  return cr_cbp;
}


// Residue Color Transform
int dct_chroma4x4(int uv, int b8, int b4)
{
  int sign(int a,int b);