Sorry for the horrendous formatting of the previous message, this is straight from the "Reference in a new issue" feature from Github.

Anyway.

Coming back to the idea of computing ffts with the gil released, turns out you still need the gil to call FTW.execute().
Here is the code to the execute method.

    cpdef execute(self):
        '''execute()

        Execute the planned operation, taking the correct kind of FFT of
        the input array (i.e. :attr:`FFTW.input_array`),
        and putting the result in the output array (i.e.
        :attr:`FFTW.output_array`).
        '''
        cdef void *input_pointer = (
                <void *>np.PyArray_DATA(self._input_array))
        cdef void *output_pointer = (
                <void *>np.PyArray_DATA(self._output_array))

        cdef void *plan = self._plan
        cdef fftw_generic_execute fftw_execute = self._fftw_execute
        with nogil:
            fftw_execute(plan, input_pointer, output_pointer)

In order to call it from a nogil block, this should really be marked nogil (and probably be cdef, I'm not sure)

HOWEVER, if self._fftw_execute and self._plan were not private, but readonly, I think I could get away with it.

My question is: how do you feel about those two properties being readonly ?

I think execute can happily be nogil for calling from Cython (though we should check that np.PyArray_DATA does not do anything pythony). Is this possible? This code is pretty ancient nowadays so things might have improved to make this sort of this possible/better now.

Do you have a reference to making it readonly? Does this just make it more broadly accessible from Cython?