Vector of random number generators - Parallel random number generators

[nicolasvazquez95]

Hi, I'm currently working in Montecarlo simulations in my GPU. I need to parallelize my simulations, so I need to set a random number generator for each simulation. My idea is to instantiate a CuArray with N random number generators (like MersenneTwister or Xoroshiro).

In Numba, there is a function for that: numba.cuda.random.create_xoroshiro128p_states(n, seed, subsequence_start=0, stream=0) (see https://numba.pydata.org/numba-doc/0.36.1/cuda/random.html).

In CUDA.jl, I cannot find a similar function for this purpose. Even more, I don't know how to set a single random generator.

Does anyone have faced this problem? Any support will be apreciated.

Regards!

[maleadt]

It's unclear to me at what level of parallelism you need multiple RNGs. If you want to this in a kernel, that's currently not possible; you can only use a single seed throughout the (implicitly) global RNG that you can use in a kernel. Every thread will still generate different numbers, of course, but they will be derived from a single seed.

Why do you need multiple RNGs though? Why isn't it sufficient to just call rand() in every thread and observe that every thread gets a different random number?

julia> function kernel()
       @cushow rand()
       return
       end

julia> @cuda threads=2 kernel()
rand() = 0.630449
rand() = 0.125559

From the host, you can of course maintain multiple copies of CUDA's RNG objects, but there's no reason to try and store those in a CuArray.