floops_cuda_demos.jl

using FLoops
using CUDA

function demo1(ex, xs = CUDA.ones(3))
    z = zero(eltype(xs))
    @floop FallbackEx() for x in xs
        @reduce(s = z + x)
    end
    return s
end

function demo2(ex, xs = CUDA.ones(3))
    @floop FallbackEx() for x in xs
        @reduce(s += x)
    end
    return s
end