Re: [Help-gsl] Iterations number with Conjugate PR gradient minimization

[Brian Gough]

At Mon, 28 Jan 2008 12:22:32 +0100,
Mauro wrote:
> I wrote a program that uses Conjugate PR gradient algorithm to minimize a
> multivariable function.
> The program works fine, but has a strange behavior when I run it on
> Valgrind, a software for detect many memory management and threading bugs.
> The result of minimization is exactly the same that I obtain without using
> valgrind, but the number of iterations is different!
> The input data obviously not change...I thought that the cause is Valgrind
> because it modifies the memory usage...
> But if I copy initialization params array (passed by the minimization
> function that initialize PRGC method) using memcpy instead an
> element-by-element cycle, the behavior is the same as when I use Valgrind!
> What happens?Can be an unwanted compiler optimization?Or it's a bug, or it's
> normal?

Maybe rounding.  Try using gsl_ieee_env_setup and
GSL_IEEE_MODE=double-precision,mask-all.

-- 
Brian Gough