CODEWITHSUNDEEP
c++classcudaoperator-overloading
Vitality
Vitality

Reputation: 21505

Overload of operator() for a CUDA matrix class

I have both CPU and GPU (CUDA) matrix classes and I want to overload the operator() so that I can read or write individual elements of the matrices.

For the CPU matrix class, I was able to do so by

OutType & operator()(const int i) { return data_[i]; }

(read) and

OutType operator()(const int i) const { return data_[i]; }

(write). For the GPU matrix class, I was able to overload the operator() for reading by

__host__ OutType operator()(const int i) const { OutType d; CudaSafeCall(cudaMemcpy(&d,data_+i,sizeof(OutType),cudaMemcpyDeviceToHost)); return d; }

but I was unable to do the same for writing. Could someone provide any hint to solve this issue?

The writing case for CPU returns the reference of data_[i], so the assignment job is performed by the builting C++ operator=. I cannot figure out how could I exploit the same mechanism for CUDA.

Thanks.

Upvotes: 0

Views: 1090

Answers (1)

vinograd47
vinograd47

Reputation: 6420

You can create a separate class that has overloaded assignment operator and type cast operator and emulates reference behavior:

class DeviceReferenceWrapper
{
public:
    explicit DeviceReferenceWrapper(void* ptr) : ptr_(ptr) {}

    DeviceReferenceWrapper& operator =(int val)
    {
        cudaMemcpy(ptr_, &val, sizeof(int), cudaMemcpyHostToDevice);
        return *this;
    }

    operator int() const
    {
        int val;
        cudaMemcpy(&val, ptr_, sizeof(int), cudaMemcpyDeviceToHost);
        return val;
    }

private:
    void* ptr_;
};

and use it in matrix class

class Matrix
{
    DeviceReferenceWrapper operator ()(int i)
    {
        return DeviceReferenceWrapper(data + i);
    }
};

Upvotes: 1

Related Questions