Generating Random Number from inside Kernel

Specific Domains GPU
2

Yes there are several possibilities. The easiest is probably to use the random number generator that is part of GPUArrays should work on the device. (It isn’t documented well, but it should get you started: https://github.com/JuliaGPU/GPUArrays.jl/blob/master/src/random.jl)

@maleadt wrote this a while back and I don’t think it is part of any repository. It is a xorshift implementation.

function xorshift(x::UInt32)::UInt32
    x = xor(x, x << 13)
    x = xor(x, x >> 17)
    x = xor(x, x << 5)
    return x
end

function gpurand(state::UInt32)::Tuple{Float32,UInt32}
    state = xorshift(state)
    res = (state >> 9) | reinterpret(UInt32, 1f0)
    val = reinterpret(Float32, res) - 1f0
    return val, state
end

function gpu_walk_kernel(f, out, N, x, a, b)
    idx = (blockIdx().x-UInt32(1)) * blockDim().x + threadIdx().x

    # random number generation
    _, state = gpurand(idx)
    function rand()
        let state=state # JuliaLang/julia#15276
            val, state = gpurand(state)
            return val
        end
    end

    ...
end

The last idea is to actually use the device version of CURAND.jl and I attempted that in GitHub - JuliaGPU/CUDAnativelib.jl