- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -
- Information contained in the HLRS Wiki is not legally binding and HLRS is not responsible for any damages that might result from its use -

FFTW

From HLRS Platforms
Revision as of 14:19, 15 October 2010 by Hpcerdei (talk | contribs)
Jump to navigationJump to search
FFTW (Fastest Fourier Transform in the West) is a free collection of fast C routines for computing the Discrete Fourier Transform in one or more dimensions. It includes complex, real, symmetric and parallel transforms and can handle arbitrary array sizes efficiently. FFTW is typically faster than other publically-available FFT implementations and is even competitive with vendor-tuned libraries.
Developer:
Platforms: NEC Nehalem Cluster
Category: Numerical Library
License:
Website: FFTW Home Page


Using FFTW on Nehalem cluster:

There are four versions of FFTW instaled on Nehalem cluster:

  • intel single and double precision
  • gnu single and double precision

Unload the loaded modules and load the modules that you need. See here how to use the module environment.

Example in case of the intel double precision version:

module load compiler/gnu/4.4.3 module load compiler/intel/11.1 module load mpi/openmpi/1.4.1-intel-11.1 (only if mpi programs are compiled) module load unsupported-modules module load numlib/fftw/3.2.2_double-openmpi-1.4.1-intel-11.1 #The command "module avail numlib" lists the available fftw modules. E.g.: #numlib/fftw/3.2.2_double-openmpi-1.4.1-gnu-4.4.3 #numlib/fftw/3.2.2_double-openmpi-1.4.1-intel-11.1 #numlib/fftw/3.2.2_single-openmpi-1.4.1-gnu-4.4.3 #numlib/fftw/3.2.2_single-openmpi-1.4.1-intel-11.1 #Here you can see the version too of the intel and gnu compiler that you need to load in the module.


Example

File: ex1_fftw.c
      /*
       * fft.c: compute FFT and IFFT from an array
       */
      #include <stdio.h>
      #include <fftw3.h>

      #define SIZE 4

      int main( int argc, char** argv )
      {
          double          input[SIZE] = { 1.0, 1.0, 1.0, 1.0 };
          fftw_complex    *data, *fft_result, *ifft_result;
          fftw_plan       plan_forward, plan_backward;
          int             i;

          data        = ( fftw_complex* ) fftw_malloc( sizeof( fftw_complex ) * SIZE );
          fft_result  = ( fftw_complex* ) fftw_malloc( sizeof( fftw_complex ) * SIZE );
          ifft_result = ( fftw_complex* ) fftw_malloc( sizeof( fftw_complex ) * SIZE );

          plan_forward  = fftw_plan_dft_1d( SIZE, data, fft_result, FFTW_FORWARD, FFTW_ESTIMATE );
          plan_backward = fftw_plan_dft_1d( SIZE, fft_result, ifft_result, FFTW_BACKWARD, FFTW_ESTIMATE );

          /* populate input data */
          for( i = 0 ; i < SIZE ; i++ ) {
              data[i][0] = input[i];
              data[i][1] = 0.0;
          }

          /* print initial data */
          for( i = 0 ; i < SIZE ; i++ ) {
              fprintf( stdout, "data[%d] = { %2.2f, %2.2f }\n",
                          i, data[i][0], data[i][1] );
          }

          fftw_execute( plan_forward );

          /* print fft result */
          for( i = 0 ; i < SIZE ; i++ ) {
              fprintf( stdout, "fft_result[%d] = { %2.2f, %2.2f }\n",
                          i, fft_result[i][0], fft_result[i][1] );
          }

          fftw_execute( plan_backward );

          /* print ifft result */
          for( i = 0 ; i < SIZE ; i++ ) {
              fprintf( stdout, "ifft_result[%d] = { %2.2f, %2.2f }\n",
                          i, ifft_result[i][0] / SIZE, ifft_result[i][1] / SIZE );
          }
          /* free memory */
          fftw_destroy_plan( plan_forward );
          fftw_destroy_plan( plan_backward );

          fftw_free( data );
          fftw_free( fft_result );
          fftw_free( ifft_result );

          return 0;
      }


Compile with:

gcc ex1_fftw.c -o ex1 -I/$FFTW_HOME/include -L/$FFTW_HOME/lib -lfftw3 (lfftw3f for single version)

The shell variable $FFTW_HOME is available after loading the required modules.

For the gnu fftw version you need the gnu version of the mpi module if mpi programs are compiled.

(e.g. mpi/openmpi/1.4.1-gnu-4.4.3)

For the single precision version all FFTW identifiers will begin with fftwf_ instead of fftw_.

See also

External links