CUDA multiple files - cuda

Question

Is there any way to do like this in CUDA + C++??
class : 1
class1
{
__device__ ....//some cuda code
void ExecuteCuda1(); //this should execute the cuda code in this class
}
class : 2
class2
{
__device__ ....//some cuda code
void ExecuteCuda2(); //this should execute the cuda code in this class
}
class : 3
class3
{
cl1 = new class1();
cl1->ExecuteCuda1();
cl1 = new class2();
cl1->ExecuteCuda2();
}

Answer 1

No, your code won't work, because ExecuteCuda1 and ExecuteCuda2 are __host__ functions, meaning they will execute on the CPU. It's illegal for a __host__ function to call any function marked as __device__, even if they are member functions of a common class.
Structuring your code like this will work:
__global__ kernel1() {...}
class Class1
{
void ExecuteCuda1()
{
// launch kernel1
kernel1<<<...>>>();
}
};
__global__ kernel2() {...}
class Class2
{
void ExecuteCuda2()
{
// launch kernel2
kernel2<<<...>>>();
}
};
class Class3
{
void ExecuteCuda3()
{
Class1 cl1 = new class1();
cl1->ExecuteCuda1();
Class2 cl2 = new class2();
cl2->ExecuteCuda2();
}
};
Note that __global__ function cannot be member functions, even if they are declared static. That's why we must define the kernels outside of any class. As Dan says, the code must be defined within a single translation unit, due to the absence of a CUDA linker. You can achieve this in multiple files by using header files.

Answer 2

Yes, but device code (i.e. functions marked with device or global) must be visible in the same compilation unit where it is needed, i.e. no linking.
So as long as you had the definitions of Class1 and Class2 in headers and not just the declarations then it should work.