Support Unitful data?

[JeffFessler]

Currently fft fails for Unitful data. It is pretty easy to write a work-around that does not even add any dependency on the Unitful package, by using oneunit in Base:

using Unitful: mm
using FFTW: fft
import FFTW: fft

x = ones(8)*mm
#fft(x) # fails ERROR: LoadError: MethodError: no method matching plan_fft...

function fft(x::AbstractArray{<:Number})
    u = oneunit(eltype(x))
    return fft(x / u) * u
end

fft(x) # works as desired

Edit: perhaps the support should be added at the plan_fft level or such.
Would there be openness for a PR to add such support to FFTW?

[stevengj]

That method allocates two extra arrays, which isn't ideal. It also creates an infinite dispatch loop for any Number type that isn't supported…

[mcabbott]

It should be fairly easy to do this by reinterpreting arrays, at least for simple dense arrays of constant units. Something like PainterQubits/Unitful.jl#437 except easier since fft only acts one one array. Unitful would probably need to depend on AbstractFFTs.

[JeffFessler]

Thanks for the reinterpret tip! In the likely event that no else takes this up sooner, I'll take stab it this summer after classes end...