How to use template functions and CUDA

Question

So I have the following code:

File: Cuda.cu

template <typename T>
__global__ void xpy( int n, T *x, T *y, T *r )
{
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    if (i < n) r[i] = x[i] + y[i];
}

mtx_mtx_add( float *a1, float *a2, float *r, const int &numElements )
{
// snip
xpy<<<numBlocks, blockSize>>>(numElements, a1, a2, r); 
}
mtx_mtx_add( int *a1, int *a2, int *r, const int &numElements ) {:::}
mtx_mtx_add( long long *a1, long long *a2, long long *r, const int &numElements ) {:::}

File: Calling Code

extern "C" bool mtx_mtx_add( float *a1, float *a2, float *r, int &numElements );
extern "C" bool mtx_mtx_add( float *a1, float *a2, float *r, int &numElements );
extern "C" bool mtx_mtx_add( float *a1, float *a2, float *r, int &numElements );

int main()
{
... ...
mtx_mtx_add(...);
}

Now what I want is for the mtx_mtx_add function to be templated. Is this possible and if so how?

Answer 1

Programming in CUDA is basically C++. You can use all the features of the C++ language as you would use in a standard C++ program.

You can create the function template as follows:

template<typename T>
bool mtx_mtx_add(T *a1, T *a2, T *r, const int &numElements)
{
   xpy<T><<<numBlocks, blockSize>>>(numElements, a1, a2, r);
}

Then you can specialize the function template for different data-types as:

template bool mtx_mtx_add<float>(float* a1, float* a2, float* r, const int& numElements);
template bool mtx_mtx_add<int>(int* a1, int* a2, int* r, const int& numElements);