Tutorial FFT 2D sequential

In this tutorial, we present how to use fluidfft to perform 2d Fast Fourier Transform in sequential. You will see that it is really very simple.

If you really want performance, first benchmark the different methods for an array shape representative of what you will use and in the machine where the computations will be done. We explain how to do this in the page Benchmarking.

import numpy as np
from fluidfft import get_methods, import_fft_class

print(get_methods(ndim=2, sequential=True))

{'fft2d.with_dask', 'fft2d.with_pyfftw'}

We import a class and instantiate it:

cls = import_fft_class('fft2d.with_pyfftw')

o = cls(16, 48)

Let’s have a look at the attribute of this objects. They are actually all methods:

print('\n'.join([name for name in dir(o) if not name.startswith('__')]))

_numpy_api
arrayK
arrayX
coef_norm
compute_energy_from_Fourier
compute_energy_from_K
compute_energy_from_X
compute_energy_from_spatial
create_arrayK
create_arrayX
empty_aligned
fft
fft2d
fft_as_arg
fftplan
get_is_transposed
get_k_adim_loc
get_local_size_X
get_seq_indices_first_K
get_seq_indices_first_X
get_shapeK_loc
get_shapeK_seq
get_shapeX_loc
get_shapeX_seq
get_short_name
get_x_adim_loc
ifft
ifft2d
ifft_as_arg
ifft_as_arg_destroy
ifftplan
inv_coef_norm
project_fft_on_realX
run_benchs
run_tests
shapeK
shapeK_loc
shapeK_seq
shapeX
sum_wavenumbers

Let’s run a test and benchmark the fft and ifft functions directly from C++.

print(o.run_tests())

None

t1, t2 = o.run_benchs()
print('t_fft = {} s; t_ifft = {} s'.format(t1, t2))

Internal bench (FFT2DWithPYFFTW)
time fft (FFT2DWithPYFFTW):  0.000014 s
time ifft (FFT2DWithPYFFTW): 0.000009 s
t_fft = 1.3518333435058594e-05 s; t_ifft = 8.559226989746093e-06 s

Let’s understand how the data is stored:

print(o.get_is_transposed())

False

which means that for this class, in Fourier space, the data is not transposed…

Now we can get the non dimensional wavenumber in the first and second dimensions:

k0, k1 = o.get_k_adim_loc()
print('k0:', k0)
print('k1:', k1)

k0: [ 0  1  2  3  4  5  6  7  8 -7 -6 -5 -4 -3 -2 -1]
k1: [ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
 24]

and check that the shapes of the array in one process are the same than in sequential (we are in sequential, there is only one process):

assert o.get_shapeX_loc() == o.get_shapeX_seq()
assert o.get_shapeK_loc() == o.get_shapeK_seq()

Now, let’s compute fast Fourier transforms. We first initialize arrays:

a = np.ones(o.get_shapeX_loc())
a_fft = np.empty(o.get_shapeK_loc(), dtype=np.complex128)

If we do not have the array where to put the result we can do:

a_fft = o.fft(a)

If we already have the array where to put the result we can do:

o.fft_as_arg(a, a_fft)

And finally for the inverse Fourier transform:

a = o.ifft(a_fft)

o.ifft_as_arg(a_fft, a)

Let’s mention the existence of the method ifft_as_arg_destroy, which can be slightly faster than ifft_as_arg because it avoids one copy.

o.ifft_as_arg_destroy(a_fft, a)