Question

In CUDA combined __device__ and __host__ allows a function to be call from both the device and the host.

My question is: Is any example that using both will be really preferable that just defining __device__ or __host__?

Answer 1

The canonical example is using C++ classes in CUDA. In the CUDA C++ model, every member function of a class must be defined in both host and device code if that class is to be instantiated in both the device and host memory spaces.

The simplest possible case would be a trivial class:

class example
{
    public:
    float a, b;

    example(float _a, float _b) : a(_a), b(_b) {};
}

It is not possible to use this in class in CUDA, you must define the constructor in both device and host code, so:

class example
{
    public:
    float a, b;

    __device__ __host__
    example(float _a, float _b) : a(_a), b(_b) {};
}