[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: IDL FFT vs C benchmark?
- Subject: Re: IDL FFT vs C benchmark?
- From: "David McClain" <dmcclain(at)azstarnet.com>
- Date: Mon, 7 Feb 2000 00:42:03 -0700
- Newsgroups: comp.lang.idl-pvwave
- Organization: Posted via Supernews, http://www.supernews.com
- References: <email@example.com>
- Xref: news.doit.wisc.edu comp.lang.idl-pvwave:18296
Actually, the IDL routines are pretty slow by comparison... More than two
years ago I investigated their performance compared to the Intel Math Kernel
Library and found that unlike the expected 2N log2 N for an NxN square image
2-D FFT, the IDL routines scaled as (N log2 N)^2 which implies a tree search
on every butterfly operation. This appaling behavior was pointed out the RSI
and they furnished the header comments from a Fortran reference that they
used for their implementation. It appears that they sacrificed speed for the
sake of arbitrary dimension FFT's. The Intel routines are strictly power of
2 but so what. I would rather have an interpolated transform to a power of 2
in size at expected speeds than to sacrifice performance due to poor scaling
of the original problem.
We now use the Intel MKL routines wrapped in a multithreaded DLL to maximize
parallel performance of FFT's. The speedup is remarkable indeed. On our old
4-processor Pentium Pro machine we reached speeds of 75 MButterflys/sec. Our
newer multiprocessors exceed that by another factor of 2-5. By comparison,
we never saw the RSI routines exceed 7-10 MButterflys/sec.
D.McClain, Sr. Scientist
Raytheon Systems Co.
Myron Brown <Myron.Brown@jhuapl.edu> wrote in message
> Hi. Has anybody done any benchmarking of IDL's
> FFT routines? They seem pretty fast, actually. I'm
> wondering how they compare to efficient C code.
> Perhaps someone has looked at fftpack or other
> efficient implementations of the FFT in C and compared
> Anyone have any idea?
> Myron Brown
> JHU Applied Physics Laboratory