FFTmt

Performs vectorized FFT on multiple processors using pthreads

Description

This mex file implements vectorized FFTs on multiple threads. When a matrix is passed to FFTmt, it is divided in "numCPU" and each fraction of the problem is executed in a separate thread

Example: if numCPU = 2 and the matrix has 256x2^14 elements, 128 FFTs are performed in each thread, allowing the OS to distribute the load on 2 processors.

A single FFT will not be accelerated by this code.
The FFTw library already provides provision for that but the is not currently used by Matlab, see www.fftw.org for details.

The same FFT library than Matlab uses is employed so that we don't even have to link against the FFTw lib.

Developed with pthreads, so UNIX only, not for Windows,

• Acknowledgements: This submission has inspired the following:; FFTmtX
• MATLAB release: MATLAB 7.2 (R2006a)

[igor scardanzan]

super idea

[Ulrich Wirleitner]

PRETTY COOL!
just tried to compile it on WINDOWS XP:
using pthread for win32 from http://sources.redhat.com/pthreads-win32/ V2.7.0
I also had to download fftw3 from http://www.fftw.org/install/windows.html (don't know how to link to the matlab internal fftw3)
I used VC 7.1 to create the fftw3-lib's (read fftw for windows installation).
Don't forget to tell mex where to find the lib's.

RESULT: it worked!
And it is faster in every case: (Factor 1.07 up to 1.46)

BUT the best is: I don't even tried it on a multicore CPU, just a Pentium-M 1.5 GHz Dothan 2MB Cache. Probably this is because of the newer FFTW library and not the thread advantage. Can sombody try it with a multicore CPU?

[abady abd]

 

[Ulrich Wirleitner]

Update for Win-XP-Version:

I could try it now on a AMD X2 DualCore CPU:
Speed up factor is about 1.6 up to 2.2. (using included test file).

Using 2D-FFT on 1MPix random images I could reach a factor of about 2.6, compiled version about 2.8 (!) compared to MATLAB R2006a internal fft. Could not try R2006b yet.

There is also a small numeric difference (or error?) between both FFT's which probably result from two different FFTW-Versions?

Hopefully FFTW's Multi-Core Support will be included in future MATLAB releases...

[jeff fessler]

I compiled it on a quad dual-core opteron box
with ncpu=8 and got only 1.8 to 2.5x speedup.
maybe the test routine doesn't have a big
enough example to illustrate the benefits of
so many cores. what i really want is to
speed up a single FFT.

[John Iversen]

Great! I've been using it on 4core intel Mac Pro for several months now, with up to ~3x speed gains. Gain increases with matrix size of course. Empirically, I've found for matrices with < 2^15 elements, the builtin matlab fft is faster.

I've also adapted a single-precision version, which has even larger speed gains. I will post this at some point, but email me now if you'd like it.

[fritz däumler]

Great... but have problems with 64bit windows compile... someone has an working pthread-win32 compile for 64 bit windows?