Quoting "Steinar H. Gunderson" <[EMAIL PROTECTED]>: > Speaking of which; what kind of FFT does Prime95 compute? (I'd normally > guess it's a Cooley-Tukey variant, but you never know with George's magical > assembler, and I have no idea about DIT/DIF and radices :-) )
If memory serves, it's a radix-4 real-valued Cooley-Tukey transform. The use of a real-valued transform is a little unusual, almost all the other Mersenne-testing programs use a complex-valued FFT that's adapted for non-complex data. The 'special sauce' in Prime95 is not the choice of algorithm, but the data layout and the method of shuffling data back and forth to keep critical information in cache as long as possible. Prime95 also makes extremely good use of prefetching, overlapping essentially all main-memory latency with useful work. This should still be present, though it's quite dated by now: www.mersenne.org/p4notes.doc jasonp ------------------------------------------------------ This message was sent using BOO.net's Webmail. http://www.boo.net/ _______________________________________________ Prime mailing list [email protected] http://hogranch.com/mailman/listinfo/prime
