I'm currently using CUDA.NET library by GASS. I need to initialize cuda arrays (actually cublas vectors, but it doesn't matters) in one CPU thread and use them in other CPU thread. But CUDA context which holding all initialized arrays and loaded functions, can be attached to only one CPU thread.

There is mechanism called context migration API to detach context from one thread and attach it to another. But i don't how to properly use it in CUDA.NET.

I tried something like this:

class Program
{
    private static float[] vector1, vector2;
    private static CUDA cuda;
    private static CUBLAS cublas;

    private static CUdeviceptr ptr;

    static void Main(string[] args)
    {
        cuda = new CUDA(false);
        cublas = new CUBLAS(cuda);
        cuda.Init();
        cuda.CreateContext(0);
        AllocateVectors();
        cuda.DetachContext();
        CUcontext context = cuda.PopCurrentContext();
        GetVectorFromDeviceAsync(context);
    }

    private static void AllocateVectors()
    {
        vector1 = new float[]{1f, 2f, 3f, 4f, 5f};
        ptr = cublas.Allocate(vector1.Length, sizeof (float));
        cublas.SetVector(vector1, ptr);

        vector2 = new float[5];
    }


    private static void GetVectorFromDevice(object objContext)
    {
        CUcontext localContext = (CUcontext) objContext;
        cuda.PushCurrentContext(localContext);
        cuda.AttachContext(localContext);

        //change vector somehow
        vector1[0] = -1;
        //copy changed vector to device
        cublas.SetVector(vector1, ptr);
        cublas.GetVector(ptr, vector2);
        CUDADriver.cuCtxPopCurrent(ref localContext);
    }

    private static void GetVectorFromDeviceAsync(CUcontext cUcontext)
    {
        Thread thread = new Thread(GetVectorFromDevice);
        thread.IsBackground = false;
        thread.Start(cUcontext);
    }
}

But execution fails on attempt to copy changed vector to device because context is not attached? Any ideas how i can get it work?