将整数从GPU复制到CPU

#include<stdio.h>
#include <cuda_runtime.h>

__global__ void kernel(int *d_data) {
    d_data[threadIdx.x] = threadIdx.x;
}

int main(void) {
    int h_data[16];
    int *d_data;

    cudaMalloc((void **)&d_data, sizeof(int) * 16);

    kernel<<<1, 16>>>(d_data);

    cudaMemcpy(h_data, d_data, sizeof(int) * 16, cudaMemcpyDeviceToHost);

    for (int i = 0; i < 16; i++) {
        printf("h_data[%d] = %d\n", i, h_data[i]);
    }

    cudaFree(d_data);

    return 0;
}