[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Discuss-gnuradio] FFTW --enable-sse and gcc 3.4.4
From: |
Eric Blossom |
Subject: |
Re: [Discuss-gnuradio] FFTW --enable-sse and gcc 3.4.4 |
Date: |
Fri, 28 Oct 2005 16:34:16 -0700 |
User-agent: |
Mutt/1.5.6i |
On Fri, Oct 28, 2005 at 05:40:08PM -0400, Lee Patton wrote:
> I am happy to report that FFTW can be built with --enable-sse under
> Fedora Core 3 using gcc 3.4.4.
>
> Ignore the rest if this has never been a problem for you...
>
> I originally built FFTW with gcc 3.4.2, and ran into the same problem as
> these fine folks:
>
> http://lists.gnu.org/archive/html/discuss-gnuradio/2005-04/msg00070.html
>
> Basically, when running usrp_fft.py, a seg fault would occur in
> libfftw3f.so. Building FFTW without --enable-sse using gcc 3.4.2 got rid
> of the seg fault. This seemed to be a problem for Fedora Core 3 users
> only.
Thanks for the info.
> Anyway, I changed the -mcpu=pentiumpro to - mcpu=pentium-m, and did
> not notice any appreciable performance increases in usrp_fftw.py.
The performance of usrp_fft.py is *completely* dominated by the GUI.
> Are there standard benchmarking scripts distributed with gnuradio that I
> could use to test if changing -mcpu or enabling/disabling SSE makes a
> difference?
Yes, a few. See gnuradio-core/src/tests/benchmark*
These benchmark alternate implementations of low level primitives that
profiling has indicated are often a bottle neck. Of course, what
matters depends on your application.
FYI, "oprofile" is a wonderful profiling tool that doesn't require
modifications to your code, and works by using the h/w performance
counters. It is capable of profiling everything, include the kernel.
See http://oprofile.sf.net
Eric