George Woltman wrote:
>As recently 6 years ago, Crandall and Fagin published
>the details for doubling the speed of the test. Perhaps there are
>more theoretical or algorithmic improvements to come. (Anyone on this
>list working on that right now??)
Does 50+ bits per IEEE 64-bit float count as an algorithmic improvement?
The old:
1000 iterations of M 54011 with FFT length 3072
Res64: E4B9382BD681646B. Program: E2.6x
Clocks = 00:00:05.135
5.02u 0.09s 0:05 98% 0+5k 0+7io 0pf+0w
The new:
1000 iterations of M 54011 with FFT length 1024
Res64: E4D9382BD681646B. Program: E3.0x
Clocks = 00:00:07.760
7.38u 0.18s 0:07 96% 0+4k 0+0io 0pf+0w
Of course, cutting that 7.38s CPU time by a factor of 3-4 is a challenge-
if it weren't, it wouldn't be fun!
Cheers,
Ernst
p.s.: And no, I'm not cheating and using the nonstandard real*16 data type.
________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
