Re: [Gretl-devel] RNG (was: Can't open datafile)
On Wed, 19 Jan 2011, Riccardo (Jack) Lucchetti wrote:
On Tue, 18 Jan 2011, Allin Cottrell wrote:
> I've now modified the configure script in CVS, and I think that
> things should now be right for the case where the compiler
> supports sse2 but the CPU does not. That is, you shouldn't have to
> manually --disable-sse2 to get a working binary.
>
> However, I can't check that definitively on my own machine.
Just crossed my mind: what happens to random number generation on machines
which don't support sse2 (which I assume is the case on non-x86
architectures)? I have no time to check this right now, but I would assume
we fall back on the old Mersenne Twister. But if this is the case, we'd
have that scripts with the same seed would produce architecture-specific
results. Do we want this?
The SFMT doesn't actually require sse2. I think it will just run
slower in the absence of sse2, but it's probably worth checking
to see if you get the same sequence, for a given seed, with
and without sse2.
Allin