int main(){

    clock_t start,finish;

    int *d_data,*h_data;

    h_data = (int *)calloc(MAX, sizeof(int));

    memset(h_data,,MAX*sizeof(int));

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

    start = clock();

    cudaMemcpy(d_data,h_data,MAX*sizeof(int),cudaMemcpyHostToDevice);

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

    finish = clock();

    cudaFree(d_data);

    free(h_data);

    cout<<"time is "<<finish-start<<endl;

    getchar();

return ;

}