FFTW

View FFTWlib.chpl on GitHub

Example usage of the FFTW module in Chapel. This particular file demonstrates the single-threaded version of the code. In order to initialize FFTW for multithreaded support, see the FFTW documentation.

If FFTW is in your standard include/library paths, compile this code using:

chpl testFFTW.chpl

Otherwise, use the following (where $FFTW_DIR points to your FFTW installation):

chpl testFFTW.chpl -I$FFTW_DIR/include -L$FFTW_DIR/lib

The FFTW module uses the FFTW3 API and currently just implements the basic, double-precision interface. We will assume that the reader is familiar with using FFTW; more details are at http://www.fftw.org.

The code computes a series of 1D, 2D and 3D transforms, exercising the complex<->complex and real<->complex transforms (both in- and out-of-place). The output of the code should be a series of small numbers ( <= 10^-13 ); see testFFTW.good for example values, though it is possible that your values may differ in practice.

The input data for these tests is in arr{1,2,3}d.dat. The format of these files is documented below.

use FFTW;

Config to print out the error values. Note that values may differ between platforms, so turn off for testing.

config const printErrors = true;

If we don’t print out error values, we’ll print a success/failure message. This is the epsilon value against which success should be measured.

config const epsilon = 10e-13;

Performs the tests and cleanup.

proc main() {
  testAllDims();
  cleanup();
}

A helper function that invokes the test for each rank.

proc testAllDims() {
  for param d in 1..3 {
    writeln(d, "D");
    runtest(d, "arr"+d:string+"d.dat");
  }
}

This is the main test code, parametrized by the number of dimensions, ndim. fn is the filename of the file that contains the test data.

proc runtest(param ndim : int, fn : string) {
  var dims : ndim*int(32);

We define a number of different domains below, corresponding to complex input/output arrays, real (input)->complex (output) out-of-place arrays, and real<->complex in-place arrays.

The domains are as follows:

  • D: for complex<->complex transforms. This domain is also used for the real array in a real->complex out-of-place transform.

  • cD : for the complex array in a real<->complex out-of-place transform.

  • rD : for the real array in a real<->complex in-place transform. This includes the padding needed by the in-place transform. D is a sub-domain of this, and can be used to extract the real array, without padding.

  • reD, imD : Utility domains that access the real/complex parts of the complex array in a real<->complex in-place transform.

var D : domain(ndim);
var rD,cD,reD,imD : domain(ndim,int,strideKind.any);

Read in the arrays from the file below. A and B are the arrays that will be used in the FFTW calls, while goodA and goodB store the true values.

Note that goodA is all real, allowing us to reuse the same array for the real<->complex tests.

The file format is as follows (all little-endian):

  • [ndim] int32 : dimensions of the array

  • [narr] real64 : the real components of A (imaginary components are all zero). narr = N_1*N_2*..*N_ndim (product of dimensions)

  • [narr] complex128 : the components of B (stored as real-imaginary pairs).

The arrays are all stored in row-major order, as in C.

var A,B,goodA,goodB : [D] complex(128);
{
  use IO;
  var f = open(fn,ioMode.r).reader(deserializer=new binaryDeserializer(endianness.little), locking=false);

  // Read in dimensions
  for d in dims {
    f.read(d);
  }

Set the domain D, handling the different rank cases.

  select ndim {
    when 1 do D = 0.. #dims(0);
    when 2 do D = {0.. #dims(0), 0.. #dims(1)};
    when 3 do D = {0.. #dims(0), 0.. #dims(1), 0.. #dims(2)};
  }

  // Read in the arrays
  for val in goodA {
    f.read(val.re);
    val.im = 0;
  }
  for val in goodB {
    f.read(val);
  }
  f.close();
  writeln("Data read...");
}

Now set the remaining domains.

Refer to the FFTW documentation on the storage order for the in-place transforms (Sec 2.4 and 4.3.4).

select ndim {
  when 1 {
    var ldim = dims(0)/2 + 1;
    //
    // Domains for real FFT
    //
    rD = 0.. #(2*ldim); // Padding to do in-place transforms
    cD = 0.. #ldim;
    //
    // Define domains to extract the real and imaginary parts for
    // in-place transforms
    //
    reD = rD[0..(2*ldim-1) by 2];
    imD = rD[1..(2*ldim-1) by 2];
  }
  when 2 {
    //
    // Domains for real FFT
    //
    var ldim = dims(1)/2+1;
    rD = {0.. #dims(0),0.. #(2*ldim)}; // Padding to do in-place transforms
    cD = {0.. #dims(0),0.. #ldim};
    //
    // Define domains to extract the real and imaginary parts for
    // in-place transforms
    //
    reD = rD[..,0..(2*ldim-1) by 2];
    imD = rD[..,1..(2*ldim-1) by 2];
  }
  when 3 {
    //
    // Domains for real FFT
    //
    var ldim = dims(2)/2+1;
    rD = {0.. #dims(0),0.. #dims(1),0.. #(2*ldim)}; // Padding to do in-place transforms
    cD = {0.. #dims(0),0.. #dims(1),0.. #ldim};
    //
    // Define domains to extract the real and imaginary parts for
    // in-place transforms
    //
    reD = rD[..,..,0..(2*ldim-1) by 2];
    imD = rD[..,..,1..(2*ldim-1) by 2];
  }
}

FFTW does not normalize inverse transforms, so just compute the normalization constant.

var norm = * reduce dims;

We start the FFT tests below. The structure is the same :

  • Define plans for forward and reverse transforms.

  • Execute forward transform A -> B.

  • Compare with goodB.

  • Execute reverse transform B -> A and normalize.

  • Compare with goodA.

  • Cleanup plans.

complex<->complex out-of-place transform

Unlike the basic FFTW interface, we do not have specific 1D/2D/3D planner routines. For the complex <-> complex case, the dimensions of the array are inferred automatically.

var fwd = plan_dft(A, B, FFTW_FORWARD, FFTW_ESTIMATE);
var rev = plan_dft(B, A, FFTW_BACKWARD, FFTW_ESTIMATE);

// Test forward and reverse transform
A = goodA;
execute(fwd);
printcmp(B,goodB);
execute(rev);
A /= norm;
printcmp(A,goodA);
destroy_plan(fwd);
destroy_plan(rev);

complex <-> complex in-place transform

This is the same calling sequence as above, but using the in-place versions of the routine.

fwd = plan_dft(A, FFTW_FORWARD, FFTW_ESTIMATE);
rev = plan_dft(A, FFTW_BACKWARD, FFTW_ESTIMATE);

// Test forward and reverse transform
A = goodA;
execute(fwd);
printcmp(A,goodB);
execute(rev);
A /= norm; // FFTW does an unnormalized transform
printcmp(A,goodA);
destroy_plan(fwd);
destroy_plan(rev);

real <-> complex out-of-place transform

As with FFTW, these use r2c and c2r suffixes to define the direction of the transform.

plan_dft_r2c and plan_dft_c2r are overloaded; for the out-of-place transforms, they infer the dimensions from the sizes of the arrays passed in.

var rA : [D] real(64); // No padding for an out-of-place transform
var cB : [cD] complex(128);
fwd = plan_dft_r2c(rA,cB,FFTW_ESTIMATE);
rev = plan_dft_c2r(cB,rA,FFTW_ESTIMATE);
rA[D] = goodA.re;
execute(fwd);
printcmp(cB,goodB[cD]);
execute(rev);
rA /= norm;
printcmp(rA[D],goodA.re);
destroy_plan(fwd);
destroy_plan(rev);

real <-> complex in-place transform

In this case, the first argument to the planning routines is the domain of the real array WITHOUT padding (in both r2c and c2r cases). This breaks the ambiguity of whether the leading dimension of the real array is even or odd.

This design decision was motivated by the fact that the user has likely already defined a domain to extract the unpadded real array from the fully padded array (or that such a domain is intrinsically useful).

For both the r2c and c2r transforms, a real array is passed in.

var rA2 : [rD] real(64);
fwd = plan_dft_r2c(D,rA2,FFTW_ESTIMATE);
rev = plan_dft_c2r(D,rA2,FFTW_ESTIMATE);
rA2[D] = goodA.re;
execute(fwd);
printcmp(rA2[reD],goodB[cD].re); // Check the real and complex parts separately.
printcmp(rA2[imD],goodB[cD].im);
execute(rev);
rA2 /= norm;
printcmp(rA2[D],goodA.re);
destroy_plan(fwd);
destroy_plan(rev);

This is another real <-> complex in-place transform, except we pass in a complex array instead of a real array. This can get a little ugly, so we just reverse engineer the previous case.

Note that we reuse the rA2 and cB arrays, since they’re the correct sizes.

  fwd = plan_dft_r2c(D,cB,FFTW_ESTIMATE);
  rev = plan_dft_c2r(D,cB,FFTW_ESTIMATE);

  // Zero out rA2 to ensure that anything extraneous in the padding
  // doesn't get passed in.
  rA2 = 0.0;
  rA2[D] = goodA.re;
  cB.re = rA2[reD]; // Fill the complex array
  cB.im = rA2[imD];
  execute(fwd);
  printcmp(cB.re,goodB[cD].re); // Check the real and complex parts separately.
  printcmp(cB.im,goodB[cD].im);
  execute(rev);
  cB /= norm;

  // Pull everything back out to the real array for simplicity
  rA2 = 0.0;
  rA2[reD] = cB.re;
  rA2[imD] = cB.im;
  printcmp(rA2[D],goodA.re);
  destroy_plan(fwd);
  destroy_plan(rev);
}

Utility function to print the maximum absolute deviation between values computed by this code, and “truth”. The true values are computed using Mathematica v10.

proc printcmp(x, y) {
  var err = max reduce abs(x-y);
  if (printErrors) then
    writeln(err);
  else {
    if err < epsilon then
      writeln("SUCCESS: error below threshold");
    else
      writeln("FAILURE: error (", err, ") exceeds epsilon (", epsilon, ")");
  }
}