FFTW vs. gwnum?

ixfd64 · 2011-02-25, 19:58

According to Wikipedia, FFTW is the fastest free FFT library. However, it also goes without saying that gwnum (George Woltman's FFT library) is world-class code. Does anyone know how gwnum compares to FFTW? The FFTW benchmarks webpage does rank it among other programs, but there is no mention of Prime95; the closest thing is Ernst's implementation.

Nevertheless, I'd suspect gwnum to be faster because it uses assembly for its core operations, while FFTW uses C subroutines. If FFTW was faster, then George would have used it already.

After all, I don't think the FFTW folks are wrong when they claim it's the fastest free FFT library because Prime95 isn't technically free software (due to the EFF prize distribution terms). :P

henryzz · 2011-02-25, 20:03

AFAIK as long as you aren't searching for mersenne primes using GWNUM there are no restrictions.

Dubslow · 2011-09-17, 08:40

So P95 doesn't use the FFTW? I was wondering about that.
Also, how dooes the CUFFT library compare to FFTW or P95?

jasonp · 2011-09-17, 12:34

All of the FFT code in Prime95 is custom-written by George. The first version of FFTW was released around 1999. People in the GPU subforum have much more experience with CuFFTW, you should ask them for more specifics.

cheesehead · 2011-09-18, 15:49

Quote:

Originally Posted by Dubslow

So P95 doesn't use the FFTW? I was wondering about that.

Note that FFTW is written in C and is portable.

Much of Prime95's FFT code is written in assembly language for Intel or Intel-compatible CPUs. That allows it to be better optimized for speed on Intel-compatible architecture than FFTW can be, but it's not portable to other architectures.

Christenson · 2011-09-18, 20:55

Hasn't P95 also got some AMD-specific kernels, too?

jasonp · 2011-09-18, 22:24

Yes, there are SSE2 kernels designed for Opterons in 64-bit mode (and probably 32-bit mode as well, the lack of such kernels was a hot-button issue a few years ago).

Likewise, FFTW has extensive inline assembly depending on the compiler (usually assumed to be gcc)

ewmayer · 2011-09-20, 16:05

Quote:

Originally Posted by ixfd64

According to Wikipedia, FFTW is the fastest free FFT library. However, it also goes without saying that gwnum (George Woltman's FFT library) is world-class code. Does anyone know how gwnum compares to FFTW? The FFTW benchmarks webpage does rank it among other programs, but there is no mention of Prime95; the closest thing is Ernst's implementation.

If you're talking about the version of my code listed in the FFTW benchmarks, that is ~15 years old, my very early-stage Fortran-90 FFT implementation. My current FFT library is generic C with lots of SSE2 inline assembler for SSE2-supporting hardware. Not as highly-tuned as George's, probably runs at best 2/3 as fast. Haven't had time or interest to do detailed comparisons vs FFTW, but anyone who likes can download the code (I've done work on it since that snapshot, but that should suffice to give a pretty idea of relative performance), build locally (using Visual Studio or MinGW32 under Win32, GCC under Linux/MacOS, or MinGW64 under Win64) land do comparative timings.

2011-02-25, 19:58	#1
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 2×5×239 Posts	FFTW vs. gwnum? According to Wikipedia, FFTW is the fastest free FFT library. However, it also goes without saying that gwnum (George Woltman's FFT library) is world-class code. Does anyone know how gwnum compares to FFTW? The FFTW benchmarks webpage does rank it among other programs, but there is no mention of Prime95; the closest thing is Ernst's implementation. Nevertheless, I'd suspect gwnum to be faster because it uses assembly for its core operations, while FFTW uses C subroutines. If FFTW was faster, then George would have used it already. After all, I don't think the FFTW folks are wrong when they claim it's the fastest free FFT library because Prime95 isn't technically free software (due to the EFF prize distribution terms). :P Last fiddled with by ixfd64 on 2011-02-25 at 19:59

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Question about FFTW vs. gwnum?	hangxanh	Software	2	2014-10-24 15:48
llrp, a portable version of LLR using FFTW	Jean Penné	Software	16	2012-04-08 13:43
LLR V3.8.2 using gwnum 26.2 is available!	Jean Penné	Software	25	2010-11-01 15:18
GWNUM?	Unregistered	Information & Answers	3	2010-09-12 19:52
GWNUM as DLL?	Cyclamen Persicum	Software	1	2007-01-02 20:53

2011-02-25, 20:03	#2
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 2³×3×5×7² Posts	AFAIK as long as you aren't searching for mersenne primes using GWNUM there are no restrictions.

2011-09-17, 08:40	#3
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	So P95 doesn't use the FFTW? I was wondering about that. Also, how dooes the CUFFT library compare to FFTW or P95?

2011-09-17, 12:34	#4
jasonp Tribal Bullet Oct 2004 3,541 Posts	All of the FFT code in Prime95 is custom-written by George. The first version of FFTW was released around 1999. People in the GPU subforum have much more experience with CuFFTW, you should ask them for more specifics.

2011-09-18, 20:55	#6
Christenson Dec 2010 Monticello 5×359 Posts	Hasn't P95 also got some AMD-specific kernels, too?

2011-09-18, 22:24	#7
jasonp Tribal Bullet Oct 2004 DD5₁₆ Posts	Yes, there are SSE2 kernels designed for Opterons in 64-bit mode (and probably 32-bit mode as well, the lack of such kernels was a hot-button issue a few years ago). Likewise, FFTW has extensive inline assembly depending on the compiler (usually assumed to be gcc)