Test copy constructor within CUDA kernel.

copy_constructor.cu

#include <iostream>
#include <cuda_runtime.h>

class CClass {
public:
	class Inner {
	public:
		__host__ __device__ Inner() : value_( 5 ) {}
		__host__ __device__ Inner( const Inner& i ) : value_( 9 * i.value_ ) {}
		__host__ __device__ int value() const { return value_; }
	private:
		int value_;
	};
public:
	__host__ __device__ CClass() : value_( 1 ) {}
	__host__ __device__ CClass( const CClass& cc ) : value_( 3 ) {}
	__host__ __device__ int value() const { return value_; }
	__host__ __device__ Inner inner() const { return Inner(); }
private:
	int value_;
};

__global__ void test_copy_constructor_kernel( CClass cc, int* output ) {
  CClass::Inner i1;
  CClass::Inner i2 = cc.inner();
  *output = i2.value();
}

void test_copy_constructor_cpu() {
	CClass::Inner i1;
	CClass cc;
	CClass::Inner inner = cc.inner();
    std::cout << "Test copy constructor cpu - " << inner.value() << std::endl;
}

void test_copy_constructor_cuda() {
	int* cuda_out = 0;
	cudaMalloc( &cuda_out, sizeof( int ) );
	CClass cc;
	test_copy_constructor_kernel<<< dim3(4,4),dim3(4,4,4) >>>( cc, cuda_out );
	int out = 0;
	cudaMemcpy( &out, cuda_out, sizeof( int ), cudaMemcpyDeviceToHost );
	cudaFree( cuda_out );
	std::cout << "Test copy constructor cuda - " << out << std::endl;
}

int main(int, char**) {
  test_copy_constructor_cpu();
  test_copy_constructor_cuda();
  return 0;
}