test_gpu_rep.jl

using CUDA

@inline function _division(::Type{T},factor,exponent) where T
    TIME_DIVISION = (
    (:day,         24*60*60,      0),
    (:hour,           60*60,      0),
    (:minute,            60,      0),
    (:second,             1,      0),
    (:millisecond,        1,     -3),
    (:microsecond,        1,     -6),
    (:nanosecond,         1,     -9),
)
    ntuple(length(TIME_DIVISION)) do i
        (T(10)^(-exponent) * TIME_DIVISION[i][2]) ÷
            (T(10)^(-TIME_DIVISION[i][3]) * factor)
    end
end

function foo(dt)
    factor,exponent = (86400,0)
    divi = _division(Int64,factor,exponent)
    return divi[1] * dt
end

dt_cpu = [1,2,3]
dt_gpu = cu(dt_cpu)

a_d = foo.(dt_gpu) # fine

CUDA.synchronize()