Interface to FFTW library run-time algorithm tuning control
fftw('planner', method)
method = fftw('planner')
info = fftw('dwisdom')
info = fftw('swisdom')
fftw('dwisdom', info)
fftw('swisdom', info)
fftw
enables you to optimize the speed
of the MATLAB^{®} FFT functions fft
, ifft
, fft2
, ifft2
, fftn
,
and ifftn
. You can use fftw
to
set options for a tuning algorithm that experimentally determines
the fastest algorithm for computing an FFT of a particular size and
dimension at run time. MATLAB software records the optimal algorithm
in an internal data base and uses it to compute FFTs of the same size
throughout the current session. The tuning algorithm is part of the
FFTW library that MATLAB software uses to compute FFTs.
fftw('planner', method)
sets the method
by which the tuning algorithm searches for a good FFT algorithm when
the dimension of the FFT is not a power of 2. You can specify method
to
be one of the following. The default method is estimate
:
'estimate'
'measure'
'patient'
'exhaustive'
'hybrid'
When you call fftw('planner', method)
, the
next time you call one of the FFT functions, such as fft
, the tuning algorithm uses the specified
method to optimize the FFT computation. Because the tuning involves
trying different algorithms, the first time you call an FFT function,
it might run more slowly than if you did not call fftw
.
However, subsequent calls to any of the FFT functions, for a problem
of the same size, often run more quickly than they would without using fftw
.
Note The FFT functions only use the optimal FFT algorithm during the current MATLAB session. Reusing Optimal FFT Algorithms explains how to reuse the optimal algorithm in a future MATLAB session. |
If you set the method to 'estimate'
, the
FFTW library does not use run-time tuning to select the algorithms.
The resulting algorithms might not be optimal.
If you set the method to 'measure'
, the FFTW
library experiments with many different algorithms to compute an FFT
of a given size and chooses the fastest. Setting the method to 'patient'
or 'exhaustive'
has
a similar result, but the library experiments with even more algorithms
so that the tuning takes longer the first time you call an FFT function.
However, subsequent calls to FFT functions are faster than with 'measure'
.
If you set 'planner'
to 'hybrid'
, MATLAB software
Sets method to 'measure'
method
for FFT dimensions 8192 or smaller.
Sets method to 'estimate'
for FFT
dimensions greater than 8192.
method = fftw('planner')
returns
the current planner method.
info = fftw('dwisdom')
returns
the information in the FFTW library's internal double-precision database. info
can
be saved and then later reused in a subsequent MATLAB session
using the next syntax.
info = fftw('swisdom')
returns
the information in the FFTW library's internal single-precision database.
fftw('dwisdom', info)
loads fftw
wisdom
represented by info
into the FFTW library's internal
double-precision wisdom database. fftw('dwisdom','')
or fftw('dwisdom',[])
clears the internal wisdom database.
fftw('swisdom', info)
loads fftw
wisdom
represented by info
into the FFTW library's internal
single-precision wisdom database. fftw('swisdom','')
or fftw('swisdom',[])
clears the internal wisdom database.
Note on large powers of 2
For FFT dimensions that are powers of 2, between 2^{14} and
2^{22}, MATLAB software uses special
preloaded information in its internal database to optimize the FFT
computation. No tuning is performed when the dimension of the FFT
is a power of 2, unless you clear the database using the command |
For more information about the FFTW library, see http://www.fftw.org
.
The following example illustrates the run times for different
settings of planner
. The example first creates
some data and applies fft
to it using the default
method, estimate
.
t=0:.001:5; x = sin(2*pi*50*t)+sin(2*pi*120*t); y = x + 2*randn(size(t)); tic; Y = fft(y,1458); toc Elapsed time is 0.000521 seconds.
If you execute the commands
tic; Y = fft(y,1458); toc Elapsed time is 0.000151 seconds.
a second time, MATLAB software reports the elapsed time
as essentially 0. To measure the elapsed time more accurately, you
can execute the command Y = fft(y,1458)
1000 times
in a loop.
tic; for k=1:1000 Y = fft(y,1458); end; toc Elapsed time is 0.056532 seconds.
This tells you that it takes on order of 1/10000 of a second
to execute fft(y, 1458)
a single time.
For comparison, set planner
to patient
.
Since this planner
explores possible algorithms
more thoroughly than hybrid
, the first time you
run fft
, it takes longer to compute the results.
fftw('planner','patient') tic;Y = fft(y,1458);toc Elapsed time is 0.100637 seconds.
However, the next time you call fft
, it
runs at approximately the same speed as before you ran the method patient
.
tic;for k=1:1000 Y=fft(y,1458); end;toc Elapsed time is 0.057209 seconds.
In order to use the optimized FFT algorithm in a future MATLAB session, first save the "wisdom" using the command
info = fftw('wisdom')
You can save info
for a future session using
the command
save info
The next time you open a MATLAB session, load info
using
the command
load info
and then reload the "wisdom" into the FFTW database using the command
fftw('wisdom', info)