The creation of FFTW plans is not threadsafe https://discourse.julialang.org/t/fftw-plans-for-multiple-threads/21861 . ( Or is it? ) However, if the plan is pre-allocated, that should be thread safe? I want to pre-compute a matrix factorization (that uses FFTW internally) and use it in multiple threads…

[image] adomasbaliuka: Am I right to assume the copyto! is the problem because it modifies the buffer? Reading the buffers should be fine, right? The copyto! and the mul!. The point is that you can’t have multiple threads writing to the same temporary buffer in parallel.

[image] adomasbaliuka: However, if the plan is pre-allocated, that should be thread safe? Yes. Plan execution is thread-safe. But I don’t know how ToeplitzMatrices uses FFTW.

Thanks for the response! I guess, the problem might be the copyto! in the mul! function ( ToeplitzMatrices.jl/linearalgebra.jl at 63c9ae4e0c697d745a6a454608ffe40d4b1f2118 · JuliaLinearAlgebra/ToeplitzMatrices.jl · GitHub ), which performs the multiplications. It seems to modify the factorization obje…

[image] adomasbaliuka: tmp = A.tmp It re-uses the same A.tmp buffer, which is not thread-safe.

Am I right to assume the copyto! is the problem because it modifies the buffer? Reading the buffers should be fine, right? What I tried now is to modify the mul! function by introducing an option not to reuse the buffer: function mul!( y::StridedVector, A::ToeplitzFactorization, x::StridedVect…

Thanks! It seems the copy operation I put in solves the problem, as I don’t think anything else modifies contents of the factorization. I guess the package maintainers wouldn’t want to implement thread safety for this feature but I can always use my modified version.

Thread safety of `mul!` in `ToeplitzMatrices.jl` with FFTW plan inside Factorization

General Usage

stevengj June 8, 2023, 7:07pm 2

Yes. Plan execution is thread-safe. But I don’t know how ToeplitzMatrices uses FFTW.

1 Like

Topic		Replies	Views
Is multithreaded FFTW on multiple arrays causing too much overhead? Performance fftw	15	1031	August 19, 2023
FFTW multithreading and plans General Usage fftw , multithreading	3	821	July 11, 2023
FFTW Plans for multiple threads Julia at Scale fftw , multithreading	1	2584	March 14, 2019
Is CUDA.jl and FFTW threadsafe? GPU fftw , multithreading , cuda , fft	4	543	May 22, 2023
Parallel FFT not that much faster General Usage fftw , parallel , multithreading	10	7176	December 19, 2016

Thread safety of `mul!` in `ToeplitzMatrices.jl` with FFTW plan inside Factorization

Related topics