mv-search.c

本源码是H.26L标准的Visual C++源代码

    iy0=pic4_pix_x+s_pos_x;
    jy0=pic4_pix_y+s_pos_y;

    if (lambda)
      current_inter_sad = MVCostLambda (mv_mul==1?0:1, lambda,
                                        predicted_x, predicted_y, s_pos_x, s_pos_y);
    else
      current_inter_sad = MVCost (4, (mv_mul==1)?4:8, blocktype, img->qp, predicted_x, predicted_y, s_pos_x, s_pos_y);

    for (vec1_x=0; vec1_x < blockshape_x; vec1_x += 4)
    {
      for (vec1_y=0; vec1_y < blockshape_y; vec1_y += 4)
      {
        for (i=0; i < BLOCK_SIZE; i++)
        {
          vec2_x=i+vec1_x;
          i22=iy0+vec2_x*4;
          for (j=0; j < BLOCK_SIZE; j++)
          {
            vec2_y=j+vec1_y;
            m7[i][j]=mo[vec2_y][vec2_x]-PelY_14 (CurrRefPic, jy0+vec2_y*4, i22);
          }
        }
        current_inter_sad += find_sad(input->hadamard, m7);
      }
    }
    if (current_inter_sad < best_inter_sad)
    {
      // Vectors are saved in all_mv[] to be able to assign correct vectors to each block after mode selection.
      // tmp_mv[] is a 'temporary' assignment of vectors to be used to estimate 'bit cost' in vector prediction.

      *s_pos_x1=s_pos_x;
      *s_pos_y1=s_pos_y;

      for (i=0; i < blockshape_x/BLOCK_SIZE; i++)
      {
        for (j=0; j < blockshape_y/BLOCK_SIZE; j++)
        {
          all_mv[block_x+i][block_y+j][ref_frame][blocktype][0]=mv_mul**s_pos_x1;
          all_mv[block_x+i][block_y+j][ref_frame][blocktype][1]=mv_mul**s_pos_y1;
          tmp_mv[0][pic_block_y+j][pic_block_x+i+BLOCK_SIZE]=mv_mul**s_pos_x1;
          tmp_mv[1][pic_block_y+j][pic_block_x+i+BLOCK_SIZE]=mv_mul**s_pos_y1;
        }
      }
      best_inter_sad=current_inter_sad;
    }
  }

  *s_pos_x2 = 2 * *s_pos_x1;
  *s_pos_y2 = 2 * *s_pos_y1;

  return best_inter_sad;
}


/*!
 ***********************************************************************
 * \brief
 *    Eighth Pel Search
 ***********************************************************************
 */
int EighthPelSearch(int pic8_pix_x, int pic8_pix_y, int mv_mul, int blockshape_x, int blockshape_y,
       int *s_pos_x1, int *s_pos_y1, int *s_pos_x2, int *s_pos_y2, int center_h2, int center_v2,
       int predicted_x, int predicted_y, int blocktype, pel_t mo[16][16], pel_t **CurrRefPic, int ***tmp_mv,
       int *****all_mv, int block_x, int block_y, int ref_frame,
       int pic_block_x, int pic_block_y, int best_inter_sad,
       double lambda
       )
{
  int current_inter_sad;
  int lv, s_pos_x, s_pos_y, iy0, jy0, vec1_x, vec1_y, vec2_x, vec2_y, i, i22, j;
  int  m7[16][16];

  for (lv=1; lv < 9; lv++)
  {
    s_pos_x=*s_pos_x2+SpiralX[lv];
    s_pos_y=*s_pos_y2+SpiralY[lv];

    iy0=pic8_pix_x+s_pos_x;
    jy0=pic8_pix_y+s_pos_y;

    if (lambda)
      current_inter_sad = MVCostLambda (0, lambda,
      predicted_x, predicted_y, s_pos_x, s_pos_y);
    else
      current_inter_sad = MVCost (8, 8, blocktype, img->qp, predicted_x, predicted_y, s_pos_x, s_pos_y);

    for (vec1_x=0; vec1_x < blockshape_x; vec1_x += 4)
    {
      for (vec1_y=0; vec1_y < blockshape_y; vec1_y += 4)
      {
        for (i=0; i < BLOCK_SIZE; i++)
        {
          vec2_x=i+vec1_x;
          i22=iy0+vec2_x*8;
          for (j=0; j < BLOCK_SIZE; j++)
          {
            vec2_y=j+vec1_y;
            m7[i][j]=mo[vec2_y][vec2_x]-PelY_18 (CurrRefPic, jy0+vec2_y*8, i22);
          }
        }
        current_inter_sad += find_sad(input->hadamard, m7);
      }
    }
    if (current_inter_sad < best_inter_sad)
    {
      // Vectors are saved in all_mv[] to be able to assign correct vectors to each block after mode selection.
      // tmp_mv[] is a 'temporary' assignment of vectors to be used to estimate 'bit cost' in vector prediction.

      *s_pos_x1=s_pos_x;
      *s_pos_y1=s_pos_y;

      for (i=0; i < blockshape_x/BLOCK_SIZE; i++)
      {
        for (j=0; j < blockshape_y/BLOCK_SIZE; j++)
        {
          all_mv[block_x+i][block_y+j][ref_frame][blocktype][0]=*s_pos_x1;
          all_mv[block_x+i][block_y+j][ref_frame][blocktype][1]=*s_pos_y1;
          tmp_mv[0][pic_block_y+j][pic_block_x+i+BLOCK_SIZE]=*s_pos_x1;
          tmp_mv[1][pic_block_y+j][pic_block_x+i+BLOCK_SIZE]=*s_pos_y1;
        }
      }
      best_inter_sad=current_inter_sad;
    }
  }
  return best_inter_sad;
}


#define PFRAME    0
#define B_FORWARD 1
#define B_BACKWARD  2

/*!
 ************************************************************************
 * \brief
 *    In this routine motion search (integer pel+1/2 and 1/4 pel) and mode selection
 *    is performed. Since we treat all 4x4 blocks before coding/decoding the
 *    prediction may not be based on decoded pixels (except for some of the blocks).
 *    This will result in too good prediction.  To compensate for this the SAD for
 *    intra(tot_intra_sad) is given a 'handicap' depending on QP.
 *                                                                                     \par
 *    Note that a couple of speed improvements were implemented by StW, which
 *    slightly change the use of some variables.  In particular, the dimensions
 *    of the img->m7 and mo[] variables were changed in order to allow better
 *    cache access.
 *                                                                                     \par
 *    Depending on the MV-resolution 1/4-pel or 1/8-pel motion vectors are estimated
 *                                                                                     \par
 * \par Input:
 *    Best intra SAD value.
 *
 * \par Output:
 *    Reference image.
 *
 * \par Side Effects:
 *    writes refFrArr[ ][ ] to set reference frame ID for each block
 *
 *  \par Bugs and Limitations:
 *    currently the B-frame module does not support UMV in its residual coding.
 *    Hence, the old search range limitation is implemented here for all B frames.
 *    The if()s in question all start in column 1 of the file and are thus
 *    easily idetifyable.
 *
 ************************************************************************
 */

int
UnifiedMotionSearch (int tot_intra_sad, int **refFrArr, int ***tmp_mv,
         int *****img_mv,
         int *img_mb_mode, int *img_blc_size_h, int *img_blc_size_v,
         int *img_multframe_no, int BFrame, int *min_inter_sad, int *****all_mv)
{
  int predframe_no=0;
  int ref_frame,blocktype;
  int tot_inter_sad;
  int ref_inx;

  *min_inter_sad = MAX_VALUE;


  // Loop through all reference frames under consideration
  // P-Frames, B-Forward: All reference frames
  // P-Backward: use only the last reference frame, stored in mref_P
  

#ifdef _ADDITIONAL_REFERENCE_FRAME_
  for (ref_frame=0; ref_frame < img->nb_references; ref_frame++)
    if ((ref_frame <  input->no_multpred) || (ref_frame == input->add_ref_frame))
#else
  for (ref_frame=0; ref_frame < img->nb_references; ref_frame++)
#endif

  {
    ref_inx = ref_frame+1;

    //  Looping through all the chosen block sizes:
    blocktype=1;
    while (input->blc_size[blocktype][0]==0 && blocktype<=7) // skip blocksizes not chosen
      blocktype++;

    for (;blocktype <= 7;)
    {
      tot_inter_sad  = QP2QUANT[img->qp] * min(ref_frame,1) * 2; // start 'handicap '
      tot_inter_sad += SingleUnifiedMotionSearch (ref_frame, blocktype,
        refFrArr, tmp_mv, img_mv,
        BFrame, all_mv, 0.0);

        /*
        Here we update which is the best inter mode. At the end we have the best inter mode.
        predframe_no:  which reference frame to use for prediction
        img->multframe_no:  Index in the mref[] matrix used for prediction
      */

      if (tot_inter_sad <= *min_inter_sad)
      {
        if (BFrame==PFRAME)
          *img_mb_mode=blocktype;
        if (BFrame == B_FORWARD)
        {
          if (blocktype == 1)
            *img_mb_mode = blocktype;
          else
            *img_mb_mode = blocktype * 2;
        }
        if (BFrame == B_BACKWARD)
        {
          if (blocktype == 1)
            *img_mb_mode = blocktype+1;
          else
            *img_mb_mode = blocktype * 2+1;

        }
        *img_blc_size_h=input->blc_size[blocktype][0];
        *img_blc_size_v=input->blc_size[blocktype][1];
        predframe_no=ref_frame;
        *img_multframe_no=ref_inx;
        *min_inter_sad=tot_inter_sad;
      }

      while (input->blc_size[++blocktype][0]==0 && blocktype<=7); // only go through chosen blocksizes
    }
    if (BFrame == B_BACKWARD)
      ref_frame = MAX_VALUE;    // jump out of the loop
  }

  return predframe_no;
}



/*!
 ************************************************************************
 * \brief
 *    unified motion search for a single combination of reference frame and blocksize
 ************************************************************************
 */
int
SingleUnifiedMotionSearch (int    ref_frame,
                           int    blocktype,
                           int  **refFrArr,
                           int ***tmp_mv,
                           int    *****img_mv,
                           int    BFrame,
                           int    *****all_mv,
                           double lambda)
{
  int   j,i;
  int   pic_block_x,pic_block_y,block_x,ref_inx,block_y,pic_pix_y;
  int   pic4_pix_y,pic8_pix_y,pic_pix_x,pic4_pix_x,pic8_pix_x;
  int   ip0,ip1;
  int   center_h2,curr_search_range,center_v2;
  int   s_pos_x1,s_pos_y1,s_pos_x2,s_pos_y2;
  int   mb_y,mb_x;
  int   best_inter_sad, tot_inter_sad = 0;
  pel_t mo[16][16];
  int   blockshape_x,blockshape_y;
  int   mv_mul;
  pel_t **CurrRefPic;
  pel_t *CurrRefPic11;
#ifdef _FAST_FULL_ME_
  int  refindex, full_search_range;
#else
  int    center_x, center_y;
#endif



  /*  curr_search_range is the actual search range used depending on block size and reference frame.
  It may be reduced by 1/2 or 1/4 for smaller blocks and prediction from older frames due to compexity
  */
#ifdef _FULL_SEARCH_RANGE_
  if (input->full_search == 2) // no restrictions at all
  {
    curr_search_range = input->search_range;
  }
  else if (input->full_search == 1) // smaller search range for older reference frames
  {
    curr_search_range = input->search_range /  (min(ref_frame,1)+1);
  }
  else // smaller search range for older reference frames and smaller blocks
#endif
  {
    curr_search_range = input->search_range / ((min(ref_frame,1)+1) * min(2,blocktype));
  }
#ifdef _FAST_FULL_ME_
#ifdef _FULL_SEARCH_RANGE_
  if (input->full_search == 2)
    full_search_range   = input->search_range;
  else
#endif
    full_search_range   = input->search_range / (min(ref_frame,1)+1);
#endif



  // Motion vector scale-factor for 1/8-pel MV-resolution
  if(input->mv_res)
    mv_mul=2;
  else
    mv_mul=1;


  // set reference frame array
  for (j = 0;j < 4;j++)
  {
    for (i = 0;i < 4;i++)
    {
      refFrArr[img->block_y+j][img->block_x+i] = ref_frame;
    }
  }


  // find  reference image
  if (BFrame != B_BACKWARD)
  {
    ref_inx=ref_frame+1;
    CurrRefPic   = mref[ref_inx];
    CurrRefPic11 = Refbuf11[ref_inx];
  }
  else
  {
    CurrRefPic   = mref_P;
    CurrRefPic11 = Refbuf11_P;
  }


#ifdef _FAST_FULL_ME_
  //===== setup fast full integer search: get SAD's for all motion vectors and 4x4 blocks =====
  //      (this is done just once for a macroblock)
  refindex = (BFrame != B_BACKWARD ? ref_frame : img->buf_cycle);

  if (!search_setup_done[refindex])
  {
    SetupFastFullIntegerSearch (ref_frame, refFrArr, tmp_mv, img_mv, CurrRefPic11,
      full_search_range, (BFrame==B_BACKWARD), (lambda != 0));
  }
#endif

  blockshape_x=input->blc_size[blocktype][0];// input->blc_size has information of the 7 blocksizes
  blockshape_y=input->blc_size[blocktype][1];// array iz stores offset inside one MB


  //  Loop through the whole MB with all block sizes
  for (mb_y=0; mb_y < MB_BLOCK_SIZE; mb_y += blockshape_y)
  {
    block_y=mb_y/BLOCK_SIZE;
    pic_block_y=img->block_y+block_y;
    pic_pix_y=img->pix_y+mb_y;
    pic4_pix_y=pic_pix_y*4;

    for (mb_x=0; mb_x < MB_BLOCK_SIZE; mb_x += blockshape_x)
    {
      block_x=mb_x/BLOCK_SIZE;
      pic_block_x=img->block_x+block_x;
      pic_pix_x=img->pix_x+mb_x;
      pic4_pix_x=pic_pix_x*4;


      SetMotionVectorPredictor (img_mv[block_x][block_y][ref_frame][blocktype], refFrArr, tmp_mv,
        ref_frame, mb_x, mb_y, blockshape_x, blockshape_y);


      ip0=img_mv[block_x][block_y][ref_frame][blocktype][0];
      ip1=img_mv[block_x][block_y][ref_frame][blocktype][1];

      //  mo[] is the the original block to be used in the search process
      // Note: dimensions changed to speed up access
      for (i=0; i < blockshape_x; i++)
        for (j=0; j < blockshape_y; j++)
          mo[j][i]=imgY_org[pic_pix_y+j][pic_pix_x+i];

      //  Integer pel search.  center_x,center_y is the 'search center'
#ifndef _FAST_FULL_ME_
      // Note that the predictor in mv[][][][][] is in 1/4 or 1/8th pel resolution
      // The following rounding enforces a "true" integer pel search.
      center_x=ip0/(mv_mul*4);
      center_y=ip1/(mv_mul*4);

      // Limitation of center_x,center_y so that the search wlevel ow contains the (0,0) vector
      center_x=max(-(input->search_range),min(input->search_range,center_x));
      center_y=max(-(input->search_range),min(input->search_range,center_y));

      // The following was deleted to support UMV.  See Remak regarding UMV performance for B
      // in the Readme File (to be provided)
      // Search center corrected to prevent vectors outside the frame, this is not permitted in this model.
      if ((pic_pix_x + center_x > curr_search_range) &&
          (pic_pix_y + center_y > curr_search_range) &&
          (pic_pix_x + center_x < img->width  - 1 - curr_search_range - blockshape_x) &&
          (pic_pix_y + center_y < img->height - 1 - curr_search_range - blockshape_y))
      {
        PelYline_11 = FastLine16Y_11;
      }
      else
      {
        PelYline_11 = UMVLine16Y_11;
      }

      best_inter_sad = IntegerSpiralSearch (mv_mul, center_x, center_y, ip0, ip1,
                                            blockshape_x, blockshape_y, curr_search_range,
                                            blocktype, mo, CurrRefPic11, &center_h2, &center_v2,
                                            pic_pix_x, pic_pix_y,
                                            lambda);

#else // _FAST_FULL_ME_