Possible Problems with Enc/Dec

• Jean-loup Gailly (jloup@chorus.fr) writes: Here are two problems with your code:
  1. If the image is a square of size 2^N+1, then

     partition_image(atx+size, aty, hsize-size,vsize, tol);

     will produce ranges of hsize 1 and break the algorithm which assumes that depth <= max_part, or equivalently that a range is never smaller than size>>max_part. The algorithm assumes that both horizontal and vertical sizes have at least max_exponent-max_part zeroes in the least significant bits of their binary form.

  2. The expression

     x_exponent = (int)floor(log((double)hsize)/log(2.0));

     is unstable if hsize is a power of two. for hsize = 64, here are some results on a 486:

     • - MSDOS, Turbo C and MSC: 6
     • - MSDOS, gcc (DJGPP): 5
     • - Unix, cc: 6
     • - Unix, gcc: 5
     As a result, enc and dec do not behave as expected when compiled with gcc on a 486.

     You can fix the problem by using hsize+0.1 instead of hsize. Similarly, you can subtract 0.1 in:

     bits_needed[i] = ceil(log((double)domain.no_h_domains[i]* (double)domain.no_v_domains[i])/log(2.0));

• Comment:

  The image size must be a multiple of the smallest range size.