From: ts <decoux@...>
Date: 2003-04-19T22:28:49+09:00
Subject: Re: Roundoff problem with Float and Marshal

>>>>> "R" == Rudi Cilibrasi <cilibrar@ofb.net> writes:

R> This is a tricky problem.  I have rewritten my patch in a way that I hope
R> is endian-independent, and furthermore allows you to load and save exactly
R> across any IEEE754 system, whether big or little endian floats.

 The test was

pigeon% cat b.rb
#!./ruby -v
io = if ARGV.size != 0
        args = true
        File.new('tt')
     else
        args = false
        File.new('tt', 'w')
     end

srand(1256457)

32.times do |cnt|
   w = rand(111111111)
   x = rand(222222223)
   y = rand(33333333334)
   z = rand(311414313)
   a = (x.to_f + y.to_f / z.to_f) * 
      Math.exp(w.to_f / (x.to_f + y.to_f / z.to_f))
   if args
      ma = Marshal.load(io)
      if a != ma 
         puts "Error #{cnt} : #{a} -- #{ma}"
      end
   else
      ma = Marshal.dump(a, io)
   end
end
pigeon% 

pigeon% ./b.rb 
ruby 1.8.0 (2003-04-18) [i686-linux]
pigeon% 

pigeon% ./b.rb 1
ruby 1.8.0 (2003-04-18) [i686-linux]
pigeon% 

nasun% ./b.rb 1
ruby 1.8.0 (2003-04-18) [sparc-solaris2.8]
Error 25 : 253341198.872703 -- 253341198.872703 -- 5.96046447753906e-08
nasun% 


 `32.times' was just to have the first error


Guy Decoux