> Message: 1
> Date: Sat, 3 Apr 2004 07:10:06 -0800 (PST)
> From: George Duncan <email@hidden>
> Subject: need optimized 2d fft for G5
> To: email@hidden
>
> I am new to this list. I am porting a code to the G5
> and have a question about fft routines. I have been
> computing on intel boxes until now am am anxious to
> see what my new G5 can do in comparison. I have a
> fortran90 code which simulates the motion of the human
> vocal folds in phonation. The code uses a
> Navier-Stokes solver which assumes periodic bc and
> uses fft very efficiently. I want to move to the G5
> with the same code intact but with the intel MKL
> library replaced with some _optimized_ library for the
> G5 containiing, among other things, the fft routines
> [I do complex to complex and the inverse and extract
> the real part ]. I am hopeful that someone on this
> list can point me to a highly optimized fft. I am
> aware of fftw, but have not tried it on the G5. Does
> anyone know of routines which have these properties
> and are sharable.
That sounds like a really cool application. I agree with the previous
posts, check out the FFT and DSP functions in Apple's various frameworks.
Single precision implementations will use AltiVec, double will be scalar
(which is very good on a G5) and they have all been highly tuned. If you
happen to need any custom functions and can use single precision, AltiVec
has a lot of intrinsic functions that are specifically intended for DSP.
You can use the AltiVec functions to define your own libraries and get
excellent performance.
Let us know how your code performs on the G5 once you get going. And make
sure to try a couple different F90 compilers since the performance can vary
quite a bit. IBM XLF has been giving excellent results on G5.
Craig
--
Dr. Craig A. Hunter
NASA Langley Research Center
AAAC / Configuration Aerodynamics Branch
(757) 864-3020
email@hidden (NEW!!)
_______________________________________________
scitech mailing list | email@hidden
Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/scitech
Do not post admin requests to the list. They will be ignored.
