Re: Double-precision floating point arithmetics behaves differently in Guile and in C

[Maxime Devos]

Panicz Maciej Godek schreef op ma 27-09-2021 om 16:04 [+0200]:
> niedz., 26 wrz 2021 o 21:11 Matt Wette <matt.wette@gmail.com>
> napisał(a):
> 
> > Your post is a little vague.  Are you comparing singular values
> > vectors
> > of A vs
> > eigenvalues, eigenvectors of A'*A ?  Also, SVD is iterative and
> > different algorithms may yield
> > different results (and some of the SVD algorithms out there are not
> > great).
> >   Are you using the exact same algorithm but noticing that (op x y)
> > in
> > Guile
> > behaves differently than (x op y) in C?
> > 
> > 
> Yes.
> I mean, I have used exactly the same algorithm (obtained by
> mechanically
> transforming the scmutils code into C), and I've been looking at its
> operation, and noticed that the results of floating point operations
> differed between Guile and C for the same input data.

If this is on x86,
https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html#index-fexcess-precision
might be relevant.

Greetings,
Maxime.