private void HandleError(cudaError error)
{
if (error != cudaError.cudaSuccess)
{
throw new CudafyHostException(error.ToString());
}
}
public void nppsSet_32s_test()
{
int length = 1024;
int value = 75;
IntPtr ptr = Npps.nppsMalloc_32s(length);
int[] result = new int[length];
GCHandle gcHandle = GCHandle.Alloc(result, GCHandleType.Pinned);
IntPtr h_result = Marshal.UnsafeAddrOfPinnedArrayElement(result, 0);
UInt64 size = Convert.ToUInt64(sizeof(int) * result.Length);
NppStatus status = Npps.nppsSet_32s(value, ptr, length);
if (status != NppStatus.NPP_SUCCESS)
{
Assert.Fail(String.Format("Fail {0}", status.ToString()));
}
cudaError cudaStatus = CudaRuntimeApi.cudaMemcpy(h_result, ptr, size, cudaMemcpyKind.DeviceToHost);
if (cudaStatus != cudaError.cudaSuccess)
{
Assert.Fail(String.Format("Fail {0}", cudaStatus.ToString()));
}
for (int i = 0; i < result.Length; i++)
{
Assert.AreEqual(value, result[i]);
}
gcHandle.Free();
Npps.nppsFree(ptr);
}
public static void Check(cudaError error, string message)
{
if (error != cudaError.cudaSuccess)
{
message = string.Format("{0}: {1}", error.ToString(), message);
Exception exception = new CudaException(message);
exception.Data.Add("cudaError", error);
throw exception;
}
}
private void CopyFromDevice <T>(Array devArray, int devOffset, Array hostArray, int hostOffset, int count)
{
CUDevicePtrEx devPtrEx = GetDeviceMemory(devArray) as CUDevicePtrEx;
int n = hostArray.Length;
Type type = typeof(T);
int elemSize = CUDA.MSizeOf(type);
unsafe
{
GCHandle handle = GCHandle.Alloc(hostArray, GCHandleType.Pinned);
IntPtr hostPtr = new IntPtr(handle.AddrOfPinnedObject().ToInt64() + hostOffset * elemSize);
CUdeviceptr devPtr = devPtrEx.DevPtr + devOffset * elemSize;
cudaError rc = CUDARuntime.cudaMemcpy(hostPtr, devPtr, elemSize * n, cudaMemcpyKind.cudaMemcpyDeviceToHost);
handle.Free();
HandleError(rc);
}
}
public static extern string cudaGetErrorString(cudaError error);
