A project comparing how different FFT implementations perform
Run
julia --project -e 'import Pkg; Pkg.instantiate();'to install all dependencies and then
julia --project benchFFT.jlto run the benchmark.
On a 2.9 GHz Intel Core i7 Quad-Core, 16MB RAM MacbookPro one gets:
Performing bench tests for 2D FFTs using nx=256 and ny=256 grid-points with Float64 arithmetic.
computing ∂f/∂x using fft
1.785 ms (130 allocations: 4.51 MiB)
computing ∂f/∂x using rfft
1.677 ms (147 allocations: 2.53 MiB)
computing ∂f/∂x using fftplan
742.876 μs (19 allocations: 3.50 MiB)
computing ∂f/∂x using rfftplan
313.398 μs (11 allocations: 1.51 MiB)
computing ∂f/∂x using fftplan & mul!
423.025 μs (13 allocations: 512.45 KiB)
computing ∂f/∂x using rfftplan & mul!
229.309 μs (5 allocations: 224 bytes)
and
Performing bench tests for 2D FFTs using nx=256 and ny=256 grid-points with Float32 arithmetic.
computing ∂f/∂x using fft
1.569 ms (130 allocations: 2.26 MiB)
computing ∂f/∂x using rfft
1.579 ms (147 allocations: 1.27 MiB)
computing ∂f/∂x using fftplan
503.819 μs (19 allocations: 1.75 MiB)
computing ∂f/∂x using rfftplan
233.894 μs (11 allocations: 772.45 KiB)
computing ∂f/∂x using fftplan & mul!
340.899 μs (13 allocations: 256.45 KiB)
computing ∂f/∂x using rfftplan & mul!
196.207 μs (5 allocations: 224 bytes)