require fftw 3.3.9 and use planner_nthreads in finufft.cpp

[ahbarnett]

https://github.com/flatironinstitute/finufft/blob/cdcccbec2585ae2ab5bf2e0be27bac1c526fa6c3/src/finufft.cpp#L764

rewrite that fftw plan call using fftw_planner_nthreads. See https://www.fftw.org/release-notes.html 3.3.9 was released Dec 2020.

Julia crash may be related to this: https://github.com/ludvigak/FINUFFT.jl/issues/61

[jkrimmer]

Hi Alex,

I have implemented a first draft to use fftw_planner_nthreads at #526. Based on this PR, I have built the finufft library for Julia on Windows. Although, the library seems to work fine, the issue described in ludvigak/FINUFFT.jl/issues/61 persists.

Since FFTW.jl and FINUFFT.jl share the same libfftw(f) in julia, I have used the following commands to veriy that this PR works:

With finufft v2.2.0:

julia> @info FFTW.get_num_threads() # calls fftw_planner_nthreads internally
[ Info: 1

julia> nufft1d1(rand(1024), rand(ComplexF64, 1024), 1, 1e-6, 1024; debug=true, nthreads=4);
[finufft_makeplan] new plan: FINUFFT version 2.2.0 .................
[finufft_makeplan] 1d1: (ms,mt,mu)=(1024,1,1) (nf1,nf2,nf3)=(2048,1,1)
               ntrans=1 nthr=4 batchSize=1
[finufft_makeplan] kernel fser (ns=7):          0.000552 s
[finufft_makeplan] fwBatch 0.00GB alloc:        3.91e-06 s
[finufft_makeplan] FFTW plan (mode 64, nthr=4): 0.000163 s
[finufft_setpts] sort (didSort=1):              8.41e-05 s
[finufft_execute] start ntrans=1 (1 batches, bsize=1)...
[finufft_execute] done. tot spread:             0.000381 s
               tot FFT:                         0.00301 s
               tot deconvolve:                  1e-05 s

julia> @info FFTW.get_num_threads()
[ Info: 4

With finufft build from the PR

julia> @info FFTW.get_num_threads() # calls fftw_planner_nthreads internally
[ Info: 1

julia> nufft1d1(rand(1024), rand(ComplexF64, 1024), 1, 1e-6, 1024; debug=true, nthreads=4);
[finufft_makeplan] new plan: FINUFFT version 2.3.0-rc1 .................
[finufft_makeplan] 1d1: (ms,mt,mu)=(1024,1,1) (nf1,nf2,nf3)=(2048,1,1)
               ntrans=1 nthr=4 batchSize=1
[finufft_makeplan] kernel fser (ns=7):          0.000106 s
[finufft_makeplan] fwBatch 0.00GB alloc:        1.11e-06 s
[finufft_makeplan] FFTW plan (mode 64, nthr=4): 5.99e-05 s
[finufft_setpts] sort (didSort=1):              8.9e-05 s
[finufft_execute] start ntrans=1 (1 batches, bsize=1)...
[finufft_execute] done. tot spread:             0.000491 s
               tot FFT:                         7.11e-05 s
               tot deconvolve:                  1.31e-05 s

julia> @info FFTW.get_num_threads()
[ Info: 1

Cheers Jonas

[DiamonDinoia]

It seems more of a FFTW.jl issue https://github.com/JuliaMath/FFTW.jl/issues/306. Shall we delay this to 2.4.0 or 2.3.1, in case we do a bugfix release? More investigation is needed in any case.

[ahbarnett]

Robert has created a longer-term solution in the form of a user lock (adding 3 opts fields): https://github.com/blackwer/finufft/tree/user_fftw_lock

Propose we assess that for 2.4 and extend it to language wrappers.

[ahbarnett]

closing since we're pretty much stuck with FFTW 3.3.8.