First,
let me thank you all for your helpful explanations.
> > Is there someone that could be able to explain me the reasons that are
> > behind the 2 conditions (q/N < 32 and maximum convolution error is < 0.5
> > for every Lucas-Lehmer iteration ) (BTW, sorry for my ignorance but what
> > does maximum convolution error mean ?)
>
> The fastest way to multiply numbers is to use FFTs, and the fastest way to
> do FFTs is with floating point math.
Is this affirmation 100% true ?
Does someone have tried to perform FFT with MMX and fixed point variables?
The only reason I can see that wouldn't make MMX a viable solution is that
Crandall algorithm necessitate 'double' (or 'float', I don't remember which one
is currently used) precision.
Any comments ?
> But then the results of the
> convolution aren't integers...they have to be rounded off. If the error
> is >.5, they'll be rounded incorrectly and the convolution fails.
>
> Hope this helps,
> jasonp
