c++ AMP library with c#: How to keep data in GPU memory?

Question

Let's say i have a program which generates big randomly filled integer array and allows me to check amount of the items divisible by some user-inputted number, using GPU for this purpose.
c# code

[DllImport("AMP.dll", CharSet = CharSet.Unicode, CallingConvention = CallingConvention.Cdecl)]
public static extern int RunTest(int* CPUinput, int length, int num);
static void Main(string[] args)
{
    Random rnd = new Random();
    int[] arr = new int[10000000];
    for (int i = 0; i < arr.Length; i++)
        arr[i] = rnd.Next(1, int.MaxValue);
    fixed (int* arrPtr = &arr[0])
    {
        while (true)
        {
            int num = int.Parse(Console.ReadLine());
            Console.WriteLine($"There are {RunTest(arrPtr, arr.Length, num)} numbers in array divisible by {num}");
        }
    }
}

c++ code

extern "C" { __declspec(dllexport) int RunTest(int* input, int length, int num); }
int RunTest(int* CPUinput, int length, int num)
{
    int CPUresult[1];
    CPUresult[0] = 0;
    array_view<int, 1> GPUinput(length, CPUinput);
    array_view<int, 1> GPUresult(1, CPUresult);
    parallel_for_each(GPUinput.get_extent(), [=](index<1> idx) restrict(amp) {
        if (GPUinput[idx[0]] % num == 0)
            atomic_fetch_inc(&GPUresult[0]);
    });
    GPUinput.discard_data();
    GPUresult.synchronize();
    return CPUresult[0];
}

Obviously, copying array each time i run the test is a bad idea. In fact it is a bottleneck in this case. How can i store the array in gpu memory among several library calls?

Answer 1

I've done this but it was a long time ago. You would want to create a wrapper in C++ CLI and interop with that so that your C# code has something it can maintain a reference to and thus keep the memory on the GPU allocated.

The following should get you started

C++/CLI wrapper for native C++ to use as reference in C#