Revision as of 11:40, 10 February 2009

FFTW3 is a library designed to compute discrete Fourier transforms. As described in the the official FFTW site, there are various versions available, with different features and different levels of maturity. In this tutorial I deal with version 3, including the experimental MPI version. However the installation instructions seems to be valid (although not tested) also for the more popular version 2.

General Remarks

As usual we would like to install the libraries in the user space, so we will create a couple of directories for that purpose:

 mkdir $HOME/usr
 mkdir $HOME/soft

To install FFTW3, download the package from the FFTW3 download page and decompress it:

 cd ~/soft
 wget http://www.fftw.org/fftw-3.3alpha1.tar.gz
 tar -zxvf fftw-3.3alpha1.tar.gz
 cd fftw-3.3alpha1

Ubuntu only: If you want to install FFTW3 in your local Ubuntu you can skip this tutorial altogether and just run:

 sudo apt-get install libfftw3-dev libfftw3-doc

However the MPI version (e.g. for testing) will not be available. If you want to have the MPI version follow the instructions in the other sections.

Build and Install

The configure/make/install procedure works well for installation in wcr.stanford.edu. We have the option of building and using static or shared libraries. If you are going to use shared libraries read last section.

Serial version

Then configure, make and install:

 ./configure --prefix=$HOME/usr --enable-shared=yes
 make
 make install

The following files will be installed in:

 ~/usr/include/fftw3.h
 ~/usr/include/fftw3.f
 ~/usr/lib/libfftw3.a
 ~/usr/lib/libfftw3.la
 ~/usr/lib/libfftw3.so

The typical compilation options will be

 export LD_RUN_PATH=$HOME/usr/lib  #do this once *before* compiling
 cc -I$HOME/usr/include program.c -L$HOME/usr/lib -lfftw3 -o program

The official tutorial on the usage of FFTW3 (which is different from FFTW 2) is located here.

MPI version

To install the experimental MPI version of FFTW3, make sure you downloaded fftw-3.3alpha1 (and not fftw-3.2 for example). Also make sure that there is an MPI compiler available:

 $which mpicc
 /usr/bin/mpicc

If it is not available, you can choose one with the command 'mpi-selector-menu' in wcr. I tested this with the 'openmpi_gcc-1.2.2' compiler.

Do the same procedure of downloading the file and decompressing it, but add the --enable-mpi flag:

 ./configure --prefix=$HOME/usr --enable-mpi --enable-shared=yes
 make install

now the library will be installed in your home directory, besides the files mentioned above, you will find also:

 ~/usr/include/fftw3-mpi.h
 ~/usr/lib/libfftw3_mpi.a
 ~/usr/lib/libfftw3_mpi.la
 ~/usr/lib/libfftw3_mpi.so

The typical command line for compilation will be

 export LD_RUN_PATH=$HOME/usr/lib #do this *before* compiling
 mpicc -I$HOME/usr mpi_program.c -L$HOME/usr/lib -lfftw3_mpi -lfftw3 -o mpi_program

Make sure to link *first* to fftw3_mpi and *later* to fftw3. For some MPI platforms (notably openmpi-gcc) setting LD_RUN_PATH does not do the job of storing the library path inside the executables (see not in previous section), it may be necessary to use the following command:

 mpicc -Wl,rpath=$HOME/usr/lib -I$HOME/usr mpi_program.c -L$HOME/usr/lib -lfftw3_mpi -lfftw3 -o mpi_program

In any case we should always check that the executable is properly linked by doing

 ldd ./mpi_program

and checking that all shared libraries are "found".

The official tutorial for the MPI version of FFTW3 can be found here.

Compilation with Shared Libraries

In the command line compilation examples above I set the variables LD_RUN_PATH. Using LD_RUN_PATH saves us from having to set path variables before *running* the program, such as LD_LIBRARY_PATH (which is a bad practice). When LD_RUN_PATH is set before compilation the created executable will store the search path internally (but will not enforce it). I learned this trick from http://gcc.gnu.org/faq.html#rpath and it works well with gcc at least. Setting this variable before compilation can be annoying, but is better than having to set variables *each time we use* the executable. This seems to be the only good option left when using libraries installed in the home directory (does anybody know a better alternative?).

Update on LD_RUN_PATH: It seems that the trick of setting this variable does not work with mpicxx, at least with the openmpi implementation. A more general way to store the path of the libraries in the executable seems to be to use the option '-Wl,rpath=$HOME/usr/lib' (sic); as described here.

Another option to avoid this issues all together is to use the -static option when compiling to ensure that the static fftw3 library files are embedded in the executable.

Basic Usage

The following is a complete program using FFTW3, for a serial 2D transform. Sources and makefile are attached here

 /* cc -I$HOME/usr/include simple_example.c -L$HOME/usr/lib -lfftw3 -o simple_example */
 
 #include <fftw3.h>
 #include <math.h>
      
 int main(int argc, char **argv){
   const ptrdiff_t N0 = 18, N1 = 18;
   fftw_plan plan;
   fftw_complex *data;
 
   data = (fftw_complex *) fftw_malloc(sizeof(fftw_complex) * N0 * N1);
 
   /* create plan for forward DFT */
   plan = fftw_plan_dft_2d(N0, N1, data, data, FFTW_FORWARD, FFTW_ESTIMATE);
 
   /* initialize data to some function my_function(x,y) */
   int i, j;
   double pdata=0;
   for (i = 0; i < N0; ++i){
     for (j = 0; j < N1; ++j){
       data[i*N1 + j][0]=i; 
       data[i*N1 + j][1]=0;
       pdata+=data[i*N1 + j][0]*data[i*N1 + j][0]+data[i*N1 + j][1]*data[i*N1 + j][1];
     }
   }
   printf("power of original data is %f\n", pdata);
 
   /* compute transforms, in-place, as many times as desired */
   fftw_execute(plan);
 
   double normalization=sqrt((double)N0*N1);
   double ptransform = 0;
 
   /*normalize data and calculate power of transform */
   for (i = 0; i < N0; ++i){
     for (j = 0; j < N1; ++j){
       data[i*N1+j][0]/=normalization;
       data[i*N1+j][1]/=normalization;
       ptransform+=data[i*N1 + j][0]*data[i*N1 + j][0]+data[i*N1 + j][1]*data[i*N1 + j][1];
     }
   }
 
   printf("power of transform is %f\n", pdata);
  
   fftw_destroy_plan(plan);
 
   return 0;
 }

The program will output the following: