#include "stdio.h"

__global__ void add(int a, int b, int *dev_result) {
    *dev_result = a + b;
}

int main() {
    int a, b, c;
    int *dev_result;

    printf("a = ");
    scanf("%d", &a);

    printf("b = ");
    scanf("%d", &b);

    cudaMalloc((void **) &dev_result, sizeof(int));

    add<<<8,8>>>(a, b, dev_result);

    /* Copies from GPU memory back to main memory: */
    cudaMemcpy(&c, dev_result, sizeof(int), cudaMemcpyDeviceToHost);

    printf("%d + %d is %d\n", a, b, c);

    cudaFree(dev_result);
    return 0;
}