On 17 Jun 99, at 12:15, Blosser, Jeremy wrote:
> You could go with a NTT instead of a FFT. Thus foregoing any double
> precision floats. Going with an NTT would also eliminate any precision
> problems...
This is true.
>
> Of course, NTTs right now are a bit slower than FFTs, but who knows maybe w/
> MMX instructions you could get a NTT to perform on par with an FFT...
>
> I need to look into that... hmm...
Ah. I think that might be less of a good idea than you think.
When you do your NTT, you're going to need at least twice as many
bits in the elements of the transform as there are bits in the number
you're testing (because you're going to want to square the values in
the elements, without any bits falling off the more significant end).
If you're working into millions of bits, I think this forces you to
use (at least) 64-bit elements. That scuppers any plans to use MMX
instructions.
Still, if Merced has parallel integer execution pipes which can do
all instructions (instead of some being done in one pipe & some in
another, a la Pentium), then maybe you don't need MMX to get the same
total performance as a floating-point FFT.
Regards
Brian Beesley
________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
