/**
Code adapted from CxImage (http://www.xdp.it/cximage.htm)
*/
static BYTE* 
  Rotate90(BYTE *src, int src_width, int src_height, int bpp ) {
    int x, y, y2;

    int dst_width  = src_height;
    int dst_height = src_width;

    // get src and dst scan width
    int src_pitch  = (src_width*bpp + 7) /8 + 3 & ~3; 
    int dst_pitch  = (dst_width*bpp + 7) /8 + 3 & ~3;


    

    // allocate dst image
    BYTE *dst =  new BYTE[dst_pitch * dst_height];
    if(NULL == dst) return NULL;

    if(bpp == 1) {
      // speedy rotate for BW images

      BYTE *sbits, *dbits, *dbitsmax, bitpos, *nrow, *srcdisp;
      div_t div_r;

      BYTE *bsrc  = src; 
      BYTE *bdest = dst;
      dbitsmax = bdest + dst_height * dst_pitch - 1;

      for(y = 0; y < src_height; y++) {
        // figure out the column we are going to be copying to
        div_r = div(y, 8);
        // set bit pos of src column byte
        bitpos = (BYTE)(128 >> div_r.rem);
        srcdisp = bsrc + y * src_pitch;
        for (x = 0; x < src_pitch; x++) {
          // get source bits
          sbits = srcdisp + x;
          // get destination column
          nrow = bdest + (dst_height - 1 - (x * 8)) * dst_pitch + div_r.quot;
          for (int z = 0; z < 8; z++) {
            // get destination byte
            dbits = nrow - z * dst_pitch;
            if ((dbits < bdest) || (dbits > dbitsmax)) break;
            if (*sbits & (128 >> z)) *dbits |= bitpos;
          }
        }
      }
    }
    else if((bpp == 8) || (bpp == 24) || (bpp == 32)) {

    }

    return dst;
}