Re: floating point discrepancy

[Hans Horn <hannes at 2horns dot com>]

On 4/6/2011 1:42 PM, Sebastian Berchtold wrote:
> Hi,
> i'm not a fortran programmer. I just try to port some code to the jvm.
>
> But i ran into significant discrepancies when dividing two numbers.
> I just give you the example im struggling with:
> 28.503076184326837 / 2.41191E-003
>
> 11817.636721240358
> 11817.636384715817
>           .
>
> the first result is consistently from java/c/wolframalpha
> while the second result is from my gfortran compiler.
> I'm using the current package on ubuntu, gcc version 4.4.5.
>
>
> This makes me serious trouble.
> So my question is how can i reach more consistent results?
> But also where can such a discrepancy come from? It's not like we're on the last digit here =)
>
> Thank you in advance,
> S.B.

Sebastian,

you are doing single precision floating point arithmetic:
28.503076184326837 / 2.41191E-003=  11817.636718750000

You need to be explicit when you want double precision:
28.503076184326837D0 / 2.41191D-003=  11817.636721240358

Kindly,
H.