FFTW planned 2D real ldiv! unexpectedly overwrites input

[marius311]

The following seems unexpected:

using FFTW, LinearAlgebra
y = rand(4,4)
p = plan_rfft(y)
x = p * y
x_orig = copy(x)
ldiv!(y,p,x)
x == x_orig # unexpectedly returns false

This only happens with FFTW and not MKL, only with 2D and real FFTs, and only with ldiv! not mul!.

[stevengj]

Unfortunately, FFTW's multidimensional c2r transforms (the inverse of rfft) overwrite the input, even for out-of-place transforms. The only way around this is to make an extra copy.

[marius311]

That's fine, but shouldn't this package be making that copy for you? It kind of seems like its trying to do so here in fact, although I don't fully understand everything that's happening.

[jamesgabbard]

Bumping this issue, I spent quite some time debugging before figuring this out by experiment. I doubt that anyone expects ldiv! to modify its last argument.

Silently creating a copy seems like a poor solution. Users who are writing ldiv!(real, plan, complex) instead of real .= irfft(complex) are probably doing this for performance reasons, and won't enjoy a large and unexpected allocation. Seems better just to clearly state in the FFTW.jl documentation that any c2r transform will overwrite its input.