On Tuesday 04 March 2008 21:36, Daran wrote: > > My first thought is that with modern PCs working with 64 bit integers why > can't we get 32 bits precision using them anyway? But even if FP > arithmetic does give more precision, by implementing an integer transform, > the rounding could be done within FT space.
It's been discussed. Often. If you think you can get more performance from current CPUs using integer hardware, take it up with George Woltman. But basically the point seems to be that with SSE2 architecture you can get two double-precision multiplies per clock, whereas a single integer multiply takes several clocks. > > I could still be missing something. Isn't there a separate 'carry' step to > an FFT multiplication which needs to be done every time, outside FT space, > even if there was no need to round? Yes. If you stay in FT space, after you've squared your N bit inputs, you have 2*N bit products - you have to redistribute the high N bits, else the result overflows next time you square. That's independent of how you implement the FFT. There's another point that using floating point hardware lets you inspect the rounding errors and so find if the system is unreliable ... shouldn't be necessary but consumer PCs are not built with reliability in mind, most "glitch" at about the same frequency as LL tests complete (one or two per month). Regards Brian Beesley _______________________________________________ Prime mailing list [email protected] http://hogranch.com/mailman/listinfo/prime
