4
This directory contains files that allow you to call FFTW from MATLAB
5
(instead of MATLAB's own FFT functions). This is accomplished by
6
means of a "MEX" program--a MATLAB external function--that wraps
7
around the FFTW library.
9
NOTE: you must have MATLAB 5.0 or later to use these routines.
11
Once you have compiled and installed the MEX (see below), using FFTW
12
from within MATLAB is simple:
14
The forward transform:
17
The backwards transform:
20
Note that FFTW computes the unnormalized DFT, so "c" in the above code
21
is a scaled version of the original "a". (To get back the original
22
"a", you would compute: c / prod(size(c)).)
24
To get help on using FFTW in MATLAB, simply type "help fftw" at the
27
There are a few points that you should be aware of:
29
* The first call is expensive:
31
The first time you call FFTW from within MATLAB, it performs
32
expensive one-time computations. (It is figuring out a "plan"--see
33
the FFTW manual for more information on what is happening.) So, the
34
first FFT you compute is slow (it probably takes several seconds).
35
However, subsequent transforms of the same size will reuse the initial
36
computations, and will be quite fast (often 2-3 times as fast as
37
MATLAB's built-in FFT). So, you should use FFTW within MATLAB when
38
you are computing many FFTs of the same size and the initial cost is
39
unimportant. If you just need a single FFT, use MATLAB's built-in
42
To reduce the startup cost, at some slight penalty in performance,
43
replace FFTW_MEASURE in fftw.c with FFTW_ESTIMATE.
45
* Small transforms are inefficient:
47
There is a certain amount of overhead involved in calling FFTW from
48
MATLAB, and this makes small transforms relatively inefficient. So,
49
if you are doing very small transforms in MATLAB, you might be better
50
off with the built-in routines. (The exact point at which FFTW begins
51
to win will depend upon your machine. It is simple for you to use
52
MATLAB's timing routines to find out what is best in your
55
(One of the major costs is in translating the array from MATLAB's
56
representation, in which real and imaginary parts are stored
57
separately, to FFTW's representation, in which complex numbers are
58
stored as adjacent real/imaginary pairs.)
60
* FFTW computes multi-dimensional transforms:
62
The FFTW call in MATLAB computes a transform of the same
63
dimensionality as the matrix that you give it. Thus, it is analogous
64
to the "fftn" routine in MATLAB, rather than the "fft" routine.
66
* All transforms are out-of-place:
68
Although the FFTW library is capable of performing in-place
69
multi-dimensional transforms, the MATLAB routine is out-of-place.
70
This is simply a restriction of the environment--as far as we can
71
tell, we are not allowed to modify the inputs that are passed to us,
72
and must return our results in a separate array.
74
**********************************************************************
78
Installation of the FFTW MEX routines is straightforward. First, you
79
have to compile the FFTW library (see the FFTW manual). Then, you must
80
compile the file fftw.c in this directory using the MEX compilation
81
procedure on your machine. Finally, you take the MEX file that is
82
produced, along with the fftw.m file in this directory, and install
83
them wherever you typically put your MATLAB scripts.
85
The method for compiling MEX files should be described in your MATLAB
86
manual. (You will need to link with the FFTW library that you had
87
compiled earlier.) On UNIX systems, you can simply type "make", and
88
the Makefile in this directory should do the right thing.