Add more operation and some refactoring of the code to ease usage.

user_ops/kernels/cuda_utils.h

@@ -0,0 +1,10 @@
+#ifndef LIB_CUDA_UTILS_H_
+#define LIB_CUDA_UTILS_H_
+
+template <typename T>
+__device__ T* DynamicSharedMemory() {
+  extern __shared__ __align__(sizeof(T)) unsigned char s_shm[];
+  return reinterpret_cast<T*>(s_shm);
+}
+
+#endif  // LIB_CUDA_UTILS_H_