template<typename T, unsigned long N>
class CudaVec {
public:
    CUDA_CALLABLE CudaVec(const T& val=T());
    CUDA_CALLABLE CudaVec(T array[N]);
    CUDA_CALLABLE CudaVec(const T& x, const T& y, const T& z);
    CUDA_CALLABLE CudaVec(const CudaVec<T,N>& other) = default;
    CUDA_CALLABLE CudaVec(CudaVec<T,N>&& other) = default;
    CUDA_CALLABLE virtual ~CudaVec() = default;
    CUDA_CALLABLE CudaVec<T,N>& operator=(const CudaVec<T,N>& right) = default;
    CUDA_CALLABLE CudaVec<T,N>& operator=(CudaVec<T,N>&& right) = default;

    CUDA_CALLABLE CudaVec<T,N> operator+(const CudaVec<T,N>& right) const;
    CUDA_CALLABLE CudaVec<T,N> operator-(const CudaVec<T,N>& right) const;
    CUDA_CALLABLE CudaVec<T,N> operator+(const T& right) const;
    CUDA_CALLABLE CudaVec<T,N> operator-(const T& right) const;
    CUDA_CALLABLE CudaVec<T,N> operator*(const T& right) const;
    CUDA_CALLABLE CudaVec<T,N> operator/(const T& right) const;
    CUDA_CALLABLE CudaVec<T,N> operator-() const;
    CUDA_CALLABLE CudaVec<T,N>& operator+=(const CudaVec<T,N>& right);
    CUDA_CALLABLE CudaVec<T,N>& operator-=(const CudaVec<T,N>& right);
    CUDA_CALLABLE CudaVec<T,N>& operator+=(const T& right);
    CUDA_CALLABLE CudaVec<T,N>& operator-=(const T& right);
    CUDA_CALLABLE CudaVec<T,N>& operator*=(const T& right);
    CUDA_CALLABLE CudaVec<T,N>& operator/=(const T& right);

    CUDA_CALLABLE CudaVec<T,3> cross(const CudaVec<T,3>& right) const;
    CUDA_CALLABLE CudaVec<T,N> dot(const CudaVec<T,N>& right) const;

    CUDA_CALLABLE T norm2() const;
    CUDA_CALLABLE T norm() const;
    CUDA_CALLABLE T normalize();

    CUDA_CALLABLE const T& get(unsigned long index) const;
    CUDA_CALLABLE const T& x() const;
    CUDA_CALLABLE const T& y() const;
    CUDA_CALLABLE const T& z() const;
    CUDA_CALLABLE const T& operator[](unsigned long index) const;
    CUDA_CALLABLE T& operator[](unsigned long index);

private:
    T elems[N];
};