Threaded FFT in Enthought Python

Question

Fast Fourier Transforms (FFTs) in Numpy/SciPy are not threaded. Enthought Python is shipped with the Intel MKL numerical library, which is capable of threaded FFTs. How does one

Answer 1

The following code works for me with Enthought 7.3-1 (64-bit) on Windows 7 Ultimate 64-bit. I haven't benchmarked it but it certainly uses all cores at once rather than just one.

from ctypes import *

class Mkl_Fft:
    c_double_p = POINTER(c_double)

    def __init__(self,num_threads=8):
        self.dfti = cdll.LoadLibrary("mk2_rt.dll")
        self.dfti.MKL_Set_Num_Threads(num_threads)
        self.Create = self.dfti.DftiCreateDescriptor_d_md
        self.Commit = self.dfti.DftiCommitDescriptor
        self.ComputeForward = self.dfti.DftiComputeForward

    def fft(self,a):
        Desc_Handle = c_void_p(0)
        dims = (c_int*2)(*a.shape)
        DFTI_COMPLEX = c_int(32)
        rank = 2

        self.Create(byref(Desc_Handle), DFTI_COMPLEX, rank, dims )
        self.Commit(Desc_Handle)
        self.ComputeForward(Desc_Handle, a.ctypes.data_as(self.c_double_p) )

Usage:

import numpy as np
a = np.ones( (32,32), dtype = complex128 )
fft = Mkl_Fft()
fft.fft(a)