Large Memory allocation for `fft` applied on real valued array

[roflmaostc]

Hey,

I was wondering why the memory usage of fft is high for a real array.
I suspect that x needs to be converted to a complex array which gives extra 2MB. For the output we need another 2MB, so totally 4MB.
However, couldn't we technically use an in-place fft! for the temporary complex version of x because we can't access it afterwards?

julia> using FFTW, BenchmarkTools

julia> x = randn((256, 512));

julia> varinfo(r"x")
  name      size summary                
  –––– ––––––––– –––––––––––––––––––––––
  x    1.000 MiB 256×512 Matrix{Float64}

julia> x_c = Complex.(x);

julia> @btime fft($x);
  2.275 ms (35 allocations: 4.00 MiB)

julia> @btime fft($x_c);
  2.126 ms (33 allocations: 2.00 MiB)

So why we don't fix that similar to that one:

julia> my_fft(x::AbstractArray{<:Real, N} where N) = fft!(Complex.(x));

julia> @btime my_fft($x);
  2.111 ms (33 allocations: 2.00 MiB)

julia> my_fft(x) == fft(x)
true

I was trying to find some information on that but wasn't successful, so I'm sorry in case I missed something!

Thanks,

Felix

[stevengj]

You could, but in-place 1d FFTs can sometimes be slower anyway. If you really want to cut down on memory allocation here, you should be using an rfft.

[roflmaostc]

Yeah I agree, that's definitely better.
What's actually the reason that it is slower?