> At 11:36 PM 6/16/99 -0700, John R Pierce wrote:
> >*If* the IA64 has a fast pipelineable 64 bit * 64 bit -> 128 bit integer
> >multiply capability, perhaps the FPU is no longer needed?
>
> You still need floating point numbers and that's probably better handled
with
> FPU hardware.
Do you? I thought the only reason the FFT was using FP numbers was most
current generation processors have a faster and higher precision FP multiply
than fixed point. With a 64*64 bit fixed point multiply that generates a
higher precision result, you can quickly do exact fixed point multiplies of
any length.
example, multiplying 2 128 bit integers X * Y where Xh and Xl are the high
and low half of the X argument takes 4 multiples plus a few adds.
Z = XhXl * YhYl
Z = Xl * Yl
Z += (Xl * Yh) << 64
Z += (Xh * Yl) << 64
Z += (Xh * Yh) << 128
where Z is a 256 bit 'accumulator'...
-jrp
________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
