sph methods slow? #191
Description
Maybe this is totally expected, but:
- A lot slower than FFT. It has been a while, but I seem to recall (going back to 2018) that the first version was much more comparable to FFT
- Allocations that don't go away with in-place (hard to make an efficient time evolution algorithm in this case)
- Which method is the bottleneck depends on size
using FastTransforms
using FFTW
using LinearAlgebra
n,m = 1024,2047
x = randn(ComplexF64,n,m)
# FFTW
@time fft(x);
0.033259 seconds (259 allocations: 32.001 MiB)
P! = plan_fft!(x);
@time P!*x;
0.035154 seconds (232 allocations: 14.969 KiB)
# Spherical
P = plan_sph2fourier(x)
PS = plan_sph_synthesis(x)
@time PS*(P*x);
0.235185 seconds (549 allocations: 127.972 MiB, 1.57% gc time)
@time P*x; # slow
0.145820 seconds (8 allocations: 63.969 MiB, 1.85% gc time)
@time PS*x;
0.051036 seconds (545 allocations: 64.003 MiB)
@time lmul!(P,x); # slow
0.139856 seconds (4 allocations: 31.984 MiB, 0.91% gc time)
@time lmul!(PS,x);
0.047418 seconds (539 allocations: 32.018 MiB)
## small array
n,m = 60,121
x = randn(ComplexF64,n,m);
# FFTW
@time fft(x);
0.000560 seconds (242 allocations: 129.836 KiB)
P! = plan_fft!(x);
@time P!*x;
0.000411 seconds (218 allocations: 14.492 KiB)
# Spherical
P = plan_sph2fourier(x);
PS = plan_sph_synthesis(x);
@time PS*(P*x);
0.101827 seconds (19.31 k allocations: 1.626 MiB)
@time P*x;
0.000266 seconds (8 allocations: 227.406 KiB)
@time PS*x; # slow
0.031964 seconds (20.48 k allocations: 1.440 MiB)
@time lmul!(P,x);
0.000364 seconds (4 allocations: 113.594 KiB)
@time lmul!(PS,x); # slow
0.037324 seconds (20.38 k allocations: 1.326 MiB)