static double[] SumMatrixManagedCuda(double[][,] matrix)
{
int Z = matrix.Length;
int Y = matrix[0].GetLength(0);
int X = matrix[0].GetLength(1);
var result = new double[Y * X];
var lm = ToLinearArray(matrix);
int N = lm.Length;
matrixSumCude.SetComputeSize((uint)X, (uint)Y);
//matrixSumCude.BlockDimensions = 128;
//matrixSumCude.GridDimensions = (N + 127) / 128;
var da = cntxt.AllocateMemory(N * sizeof(double));
var db = cntxt.AllocateMemory(result.Length * sizeof(double));
cntxt.CopyToDevice(da, lm);
cntxt.CopyToDevice(db, result);
//CudaDeviceVariable<int> dA = a;
//CudaDeviceVariable<int> dB = b;
//CudaDeviceVariable<int> dC = new CudaDeviceVariable<int>(N);
// Invoke kernel
//kernel.Run(dA.DevicePointer, dC.DevicePointer, dimX, dimY, dimZ);
matrixSumCude.Run(db, da, X, Y, Z);
cntxt.CopyToHost <double>(result, db);
return(result);
}