Loading FastTransforms.jl can make FFTs via FFTW.jl 100x slower due to threading conflicts

[ranocha]

While this is not a self-contained example, I would like to summarize an issue I had so that others may find it (and my workaround) when they search for it.

Loading FastTransforms.jl (import FastTransforms or using FastTransforms or loading a package depending on FastTransforms.jl) made some FFTs (of small real vectors with 256 to 4096 entries) via FFTW.jl roughly 100x slower on my system (when running Julia with one thread). Debugging this via the profiling tools showed that a huge amount of time was spent in __psynch_cvwait and __psynch_cvsignal. The reason appears to be

FastTransforms.jl/src/libfasttransforms.jl

Lines 15 to 20 in 14a3118

	function __init__()
	n = ceil(Int, Sys.CPU_THREADS/2)
	ft_set_num_threads(n)
	ccall((:ft_fftw_init_threads, libfasttransforms), Cint, ())
	ft_fftw_plan_with_nthreads(n)
	end

Reversing these settings by executing

FastTransforms.ft_set_num_threads(1)
ccall((:ft_fftw_init_threads, FastTransforms.libfasttransforms), Cint, ())
FastTransforms.ft_fftw_plan_with_nthreads(1)

after loading FastTransforms fixed the performance issue.