// file sum.cu

#include <cuda.h>
#include <algorithm>
#include <cstdlib>
using namespace std;

// ====================================
// kernel declaration 
// ====================================
__global__ 
void sum(float *x, float *y, float *z, int size) {
	// compute Global Thread InDex
	int gtid = threadIdx.x ;
	
	if (gtid < size) {
		z[gtid] = x[gtid] + y[gtid]; // paralell part
	}
}

// ====================================
// main function
// ====================================
int main() {
	const int SIZE = 512;

	// allocate data on computer global memory
	float *x_cpu = new float[ SIZE ];
	float *y_cpu = new float[ SIZE ];
	float *z_cpu = new float[ SIZE ];

	// fill x, y
	std::fill(&x_cpu[0], &x_cpu[SIZE], 1);
	std::fill(&y_cpu[0], &y_cpu[SIZE], 2);

	// allocate data on GPU global memory
	float *x_gpu, *y_gpu, *z_gpu;

	cudaMalloc( (void**) &x_gpu, SIZE * sizeof(float) );
	cudaMalloc( (void**) &y_gpu, SIZE * sizeof(float) );
	cudaMalloc( (void**) &z_gpu, SIZE * sizeof(float) );

	// copy x and y resp. to dev_x, dev_y
	cudaMemcpy(x_gpu, x_cpu, SIZE * sizeof(float), cudaMemcpyHostToDevice);
	cudaMemcpy(y_gpu, y_cpu, SIZE * sizeof(float), cudaMemcpyHostToDevice);

	// define number of threads
	dim3 grid(1, 1, 1);
	dim3 block(SIZE, 1, 1);
	// call kernel
	sum<<< grid, block >>>(x_gpu, y_gpu, z_gpu, SIZE);

	// copy result back to computer's global memory
	cudaMemcpy(z_cpu, z_gpu, SIZE * sizeof(float), cudaMemcpyDeviceToHost);

	// free memory
	cudaFree( x_gpu );
	cudaFree( y_gpu );
	cudaFree( z_gpu );

	delete [] x_cpu;
	delete [] y_cpu;
	delete [] z_cpu;
	
	exit(EXIT_FAILURE);
}