unsigned char *s, *d;
int lx,lx2,h;
{
  unsigned char *dp,*sp,*sp2;
  int j;

  sp = s;
  sp2 = s+lx2;
  dp = d;
  for (j=0; j<h; j++)
  {
    dp[0] = (unsigned int)(sp[0]+sp2[0]+1)>>1;
    dp[1] = (unsigned int)(sp[1]+sp2[1]+1)>>1;
    dp[2] = (unsigned int)(sp[2]+sp2[2]+1)>>1;
    dp[3] = (unsigned int)(sp[3]+sp2[3]+1)>>1;
    dp[4] = (unsigned int)(sp[4]+sp2[4]+1)>>1;
    dp[5] = (unsigned int)(sp[5]+sp2[5]+1)>>1;
    dp[6] = (unsigned int)(sp[6]+sp2[6]+1)>>1;
    dp[7] = (unsigned int)(sp[7]+sp2[7]+1)>>1;
    dp[8] = (unsigned int)(sp[8]+sp2[8]+1)>>1;
    dp[9] = (unsigned int)(sp[9]+sp2[9]+1)>>1;
    dp[10] = (unsigned int)(sp[10]+sp2[10]+1)>>1;
    dp[11] = (unsigned int)(sp[11]+sp2[11]+1)>>1;
    dp[12] = (unsigned int)(sp[12]+sp2[12]+1)>>1;
    dp[13] = (unsigned int)(sp[13]+sp2[13]+1)>>1;
    dp[14] = (unsigned int)(sp[14]+sp2[14]+1)>>1;
    dp[15] = (unsigned int)(sp[15]+sp2[15]+1)>>1;
    sp+= lx2;
    sp2+= lx2;
    dp+= lx;
  }
}

static void recvc(s,d,lx,lx2,h)
unsigned char *s, *d;
int lx,lx2,h;
{
  unsigned char *dp,*sp,*sp2;
  int j;

  sp = s;
  sp2 = s+lx2;
  dp = d;

  for (j=0; j<h; j++)
  {
    dp[0] = (unsigned int)(sp[0]+sp2[0]+1)>>1;
    dp[1] = (unsigned int)(sp[1]+sp2[1]+1)>>1;
    dp[2] = (unsigned int)(sp[2]+sp2[2]+1)>>1;
    dp[3] = (unsigned int)(sp[3]+sp2[3]+1)>>1;
    dp[4] = (unsigned int)(sp[4]+sp2[4]+1)>>1;
    dp[5] = (unsigned int)(sp[5]+sp2[5]+1)>>1;
    dp[6] = (unsigned int)(sp[6]+sp2[6]+1)>>1;
    dp[7] = (unsigned int)(sp[7]+sp2[7]+1)>>1;
    sp+= lx2;
    sp2+= lx2;
    dp+= lx;
  }
}

static void recvo(s,d,lx,lx2,addflag,c,xa,xb,ya,yb)
unsigned char *s;
int *d;
int lx,lx2,addflag,c,xa,xb,ya,yb;
{
  int *dp,*om;
  unsigned char *sp,*sp2;
  int i,j;

  sp = s;
  sp2 = s+lx2;
  dp = d;
  om = &OM[c][ya][0];

  if (!addflag) 
  {
    for (j = ya; j < yb; j++)
	{
      for (i = xa; i < xb; i++) 
	  {
        dp[i] = (((unsigned int)(sp[i] + sp2[i]+1))>>1)*om[i];
      }
      sp+= lx2;
      sp2+= lx2;
      dp+= lx;
      om+= 8;
    }
  }
  else 
  {
    for (j = ya; j < yb; j++) 
	{
      for (i = xa; i < xb; i++) 
	  {
        dp[i] += (((unsigned int)(sp[i] + sp2[i]+1))>>1)*om[i];
      }
      sp+= lx2;
      sp2+= lx2;
      dp+= lx;
      om+= 8;
    }
  }
}

static void rec4(s,d,lx,lx2,h)
unsigned char *s, *d;
int lx,lx2,h;
{
  unsigned char *dp,*sp,*sp2;
  int j;
  unsigned int s1,s2,s3,s4;

  sp = s;
  sp2 = s+lx2;
  dp = d;
  for (j=0; j<h; j++)
  {
    s1=sp[0]; s3=sp2[0];
    dp[0] = (unsigned int)(s1+(s2=sp[1])+s3+(s4=sp2[1])+2)>>2;
    dp[1] = (unsigned int)(s2+(s1=sp[2])+s4+(s3=sp2[2])+2)>>2;
    dp[2] = (unsigned int)(s1+(s2=sp[3])+s3+(s4=sp2[3])+2)>>2;
    dp[3] = (unsigned int)(s2+(s1=sp[4])+s4+(s3=sp2[4])+2)>>2;
    dp[4] = (unsigned int)(s1+(s2=sp[5])+s3+(s4=sp2[5])+2)>>2;
    dp[5] = (unsigned int)(s2+(s1=sp[6])+s4+(s3=sp2[6])+2)>>2;
    dp[6] = (unsigned int)(s1+(s2=sp[7])+s3+(s4=sp2[7])+2)>>2;
    dp[7] = (unsigned int)(s2+(s1=sp[8])+s4+(s3=sp2[8])+2)>>2;
    dp[8] = (unsigned int)(s1+(s2=sp[9])+s3+(s4=sp2[9])+2)>>2;
    dp[9] = (unsigned int)(s2+(s1=sp[10])+s4+(s3=sp2[10])+2)>>2;
    dp[10] = (unsigned int)(s1+(s2=sp[11])+s3+(s4=sp2[11])+2)>>2;
    dp[11] = (unsigned int)(s2+(s1=sp[12])+s4+(s3=sp2[12])+2)>>2;
    dp[12] = (unsigned int)(s1+(s2=sp[13])+s3+(s4=sp2[13])+2)>>2;
    dp[13] = (unsigned int)(s2+(s1=sp[14])+s4+(s3=sp2[14])+2)>>2;
    dp[14] = (unsigned int)(s1+(s2=sp[15])+s3+(s4=sp2[15])+2)>>2;
    dp[15] = (unsigned int)(s2+sp[16]+s4+sp2[16]+2)>>2;
    sp+= lx2;
    sp2+= lx2;
    dp+= lx;
  }
}

static void rec4c(s,d,lx,lx2,h)
unsigned char *s, *d;
int lx,lx2,h;
{
  unsigned char *dp,*sp,*sp2;
  int j;
  unsigned int s1,s2,s3,s4;

  sp = s;
  sp2 = s+lx2;
  dp = d;
  for (j=0; j<h; j++)
  {
    s1=sp[0]; s3=sp2[0];
    dp[0] = (unsigned int)(s1+(s2=sp[1])+s3+(s4=sp2[1])+2)>>2;
    dp[1] = (unsigned int)(s2+(s1=sp[2])+s4+(s3=sp2[2])+2)>>2;
    dp[2] = (unsigned int)(s1+(s2=sp[3])+s3+(s4=sp2[3])+2)>>2;
    dp[3] = (unsigned int)(s2+(s1=sp[4])+s4+(s3=sp2[4])+2)>>2;
    dp[4] = (unsigned int)(s1+(s2=sp[5])+s3+(s4=sp2[5])+2)>>2;
    dp[5] = (unsigned int)(s2+(s1=sp[6])+s4+(s3=sp2[6])+2)>>2;
    dp[6] = (unsigned int)(s1+(s2=sp[7])+s3+(s4=sp2[7])+2)>>2;
    dp[7] = (unsigned int)(s2+sp[8]+s4+sp2[8]+2)>>2;
    sp+= lx2;
    sp2+= lx2;
    dp+= lx;
  }
}

static void rec4o(s,d,lx,lx2,addflag,c,xa,xb,ya,yb)
unsigned char *s;
int *d;
int lx,lx2,addflag,c,xa,xb,ya,yb;
{
  int *dp,*om;
  unsigned char *sp,*sp2;
  int i,j;

  sp = s;
  sp2 = s+lx2;
  dp = d;
  om = &OM[c][ya][0];

  if (!addflag) 
  {
    for (j = ya; j < yb; j++) 
	{
      for (i = xa; i < xb; i++) 
	  {
        dp[i] = (((unsigned int)(sp[i]+sp2[i]+sp[i+1]+sp2[i+1]+2))>>2)*om[i];
      }
      sp+= lx2;
      sp2+= lx2;
      dp+= lx;
      om+= 8;
    }
  }
  else 
  {
    for (j = ya; j < yb; j++)
	{
      for (i = xa; i < xb; i++)
	  {
        dp[i] += (((unsigned int)(sp[i]+sp2[i]+sp[i+1]+sp2[i+1]+2))>>2)*om[i];
      }
      sp+= lx2;
      sp2+= lx2;
      dp+= lx;
      om += 8;
    }
  }
}

static void recon_comp_obmc(src,dst,lx,lx2,comp,w,h,x,y)
unsigned char *src;
unsigned char *dst;
int lx,lx2,comp;
int w,h;
int x,y;
{
  int j,k;
  int xmb,ymb;
  int c8,t8,l8,r8;
  int ti8,li8,ri8;
  int xit=0,xib=0,xir=0,xil=0;
  int yit=0,yib=0,yir=0,yil=0;
  int vect=0,vecb=0,vecr=0,vecl=0;
  int nx[5],ny[5],xint[5],yint[5],xh[5],yh[5];
  int p[64],*pd;
  unsigned char *d,*s[5];

  xmb = (x>>4)+1;
  ymb = (y>>4)+1;

  c8  = (modemap[ymb][xmb] == MODE_INTER4V ? 1 : 0);

  t8  = (modemap[ymb-1][xmb] == MODE_INTER4V ? 1 : 0);
  ti8 = (modemap[ymb-1][xmb] == MODE_INTRA ? 1 : 0);
  ti8 = (modemap[ymb-1][xmb] == MODE_INTRA_Q ? 1 : ti8);

  l8  = (modemap[ymb][xmb-1] == MODE_INTER4V ? 1 : 0);
  li8 = (modemap[ymb][xmb-1] == MODE_INTRA ? 1 : 0);
  li8 = (modemap[ymb][xmb-1] == MODE_INTRA_Q ? 1 : li8);
  
  r8  = (modemap[ymb][xmb+1] == MODE_INTER4V ? 1 : 0);
  ri8 = (modemap[ymb][xmb+1] == MODE_INTRA ? 1 : 0);
  ri8 = (modemap[ymb][xmb+1] == MODE_INTRA_Q ? 1 : ri8);

  if (pb_frame) 
  {
    ti8 = li8 = ri8 = 0;
  }

  switch (comp+1) 
  {
  case 1:
    vect = (ti8 ? (c8 ? 1 : 0) : (t8 ? 3 : 0)); 
    yit  = (ti8 ? ymb : ymb - 1); 
    xit = xmb;

    vecb = (c8 ? 3 : 0) ; yib = ymb; xib = xmb;

    vecl = (li8 ? (c8 ? 1 : 0) : (l8 ? 2 : 0)); 
    yil = ymb; 
    xil = (li8 ? xmb : xmb-1);

    vecr = (c8 ? 2 : 0) ; yir = ymb; xir = xmb;

    /* edge handling */
    if (ymb == 1) 
	{
      yit = ymb;
      vect = (c8 ? 1 : 0);
    }
    if (xmb == 1) 
	{
      xil = xmb;
      vecl = (c8 ? 1 : 0);
    }
    break;
  
  case 2:
    vect = (ti8 ? (c8 ? 2 : 0) : (t8 ? 4 : 0)); 
    yit = (ti8 ? ymb : ymb-1); 
    xit = xmb;

    vecb = (c8 ? 4 : 0) ; yib = ymb; xib = xmb;
    vecl = (c8 ? 1 : 0) ; yil = ymb; xil = xmb;

    vecr = (ri8 ? (c8 ? 2 : 0) : (r8 ? 1 : 0)); 
    yir = ymb; 
    xir = (ri8 ? xmb : xmb+1);

    /* edge handling */
    if (ymb == 1) 
	{
      yit = ymb;
      vect = (c8 ? 2 : 0);
    }
    if (xmb == mb_width) 
	{
      xir = xmb;
      vecr = (c8 ? 2 : 0);
    }
    break;

  case 3:
    vect = (c8 ? 1 : 0) ; yit = ymb  ; xit = xmb;
    vecb = (c8 ? 3 : 0) ; yib = ymb  ; xib = xmb;
      
    vecl = (li8 ? (c8 ? 3 : 0) : (l8 ? 4 : 0)); 
    yil = ymb;  
    xil = (li8 ? xmb : xmb-1);
    
    vecr = (c8 ? 4 : 0) ; yir = ymb  ; xir = xmb;

    /* edge handling */
    if (xmb == 1) 
	{
      xil = xmb;
      vecl = (c8 ? 3 : 0);
    }
    break;

  case 4:
    vect = (c8 ? 2 : 0) ; yit = ymb  ; xit = xmb;
    vecb = (c8 ? 4 : 0) ; yib = ymb  ; xib = xmb;
    vecl = (c8 ? 3 : 0) ; yil = ymb  ; xil = xmb;

    vecr = (ri8 ? (c8 ? 4 : 0) : (r8 ? 3 : 0)); 
    yir = ymb; 
    xir = (ri8 ? xmb : xmb+1);

    /* edge handling */
    if (xmb == mb_width) 
	{
      xir = xmb;
      vecr = (c8 ? 4 : 0);
    }
    break;

  default:
    fprintf(stderr,"Illegal block number in recon_comp_obmc (recon.c)\n");
    exit(1);
    break;
  }

  nx[0] = MV[0][c8 ? comp + 1 : 0][ymb][xmb];
  ny[0] = MV[1][c8 ? comp + 1 : 0][ymb][xmb];
  
  nx[1] = MV[0][vect][yit][xit]; ny[1] = MV[1][vect][yit][xit];
  nx[2] = MV[0][vecb][yib][xib]; ny[2] = MV[1][vecb][yib][xib];
  nx[3] = MV[0][vecr][yir][xir]; ny[3] = MV[1][vecr][yir][xir];
  nx[4] = MV[0][vecl][yil][xil]; ny[4] = MV[1][vecl][yil][xil];

  for (k = 0; k < 5; k++) 
  {
    xint[k] = nx[k]>>1;
    xh[k] = nx[k] & 1;
    yint[k] = ny[k]>>1;
    yh[k] = ny[k] & 1;
    s[k] = src + lx2*(y+yint[k]) + x + xint[k];
  }
  
  d = dst + lx*y + x;

  if (!xh[0] && !yh[0])
    reco(s[0],&p[0],8,lx2,0,0,0,8,0,8);
  else if (!xh[0] && yh[0])
    recvo(s[0],&p[0],8,lx2,0,0,0,8,0,8);
  else if (xh[0] && !yh[0])
    recho(s[0],&p[0],8,lx2,0,0,0,8,0,8);
  else /* if (xh[] && yh[]) */
    rec4o(s[0],&p[0],8,lx2,0,0,0,8,0,8);

  if (!xh[1] && !yh[1])
    reco(s[1],&p[0],8,lx2,1,1,0,8,0,4);
  else if (!xh[1] && yh[1])
    recvo(s[1],&p[0],8,lx2,1,1,0,8,0,4);
  else if (xh[1] && !yh[1])
    recho(s[1],&p[0],8,lx2,1,1,0,8,0,4);
  else /* if (xh[] && yh[]) */
    rec4o(s[1],&p[0],8,lx2,1,1,0,8,0,4);

  if (!xh[2] && !yh[2])
    reco(s[2]+(lx2<<2),&p[32],8,lx2,1,2,0,8,4,8);
  else if (!xh[2] && yh[2])
    recvo(s[2]+(lx2<<2),&p[32],8,lx2,1,2,0,8,4,8);
  else if (xh[2] && !yh[2])
    recho(s[2]+(lx2<<2),&p[32],8,lx2,1,2,0,8,4,8);
  else /* if (xh[] && yh[]) */
    rec4o(s[2]+(lx2<<2),&p[32],8,lx2,1,2,0,8,4,8);

  if (!xh[3] && !yh[3])
    reco(s[3],&p[0],8,lx2,1,3,4,8,0,8);
  else if (!xh[3] && yh[3])
    recvo(s[3],&p[0],8,lx2,1,3,4,8,0,8);
  else if (xh[3] && !yh[3])
    recho(s[3],&p[0],8,lx2,1,3,4,8,0,8);
  else /* if (xh[] && yh[]) */
    rec4o(s[3],&p[0],8,lx2,1,3,4,8,0,8);

  if (!xh[4] && !yh[4])
    reco(s[4],&p[0],8,lx2,1,4,0,4,0,8);
  else if (!xh[4] && yh[4])
    recvo(s[4],&p[0],8,lx2,1,4,0,4,0,8);
  else if (xh[4] && !yh[4])
    recho(s[4],&p[0],8,lx2,1,4,0,4,0,8);
  else /* if (xh[] && yh[]) */
    rec4o(s[4],&p[0],8,lx2,1,4,0,4,0,8);

  pd = &p[0];
  for (j = 0; j < 8; j++) 
  {
    d[0] = (pd[0] + 4 )>>3;	
    d[1] = (pd[1] + 4 )>>3;	
    d[2] = (pd[2] + 4 )>>3;	
    d[3] = (pd[3] + 4 )>>3;	
    d[4] = (pd[4] + 4 )>>3;	
    d[5] = (pd[5] + 4 )>>3;	
    d[6] = (pd[6] + 4 )>>3;	
    d[7] = (pd[7] + 4 )>>3;	
    d += lx;
    pd += 8;
  }
}