openMP 只运行一个线程？

Question

我在 main:

int main(int argv, char **argc)
{
   // Get multi-CPU/multi-GPU data
   int num_gpus;
   cudaGetDeviceCount(&num_gpus);
   printf("### Number of host CPUs:\t%d\n", omp_get_num_procs());
   printf("### Number of CUDA devices:\t%d\n", num_gpus);
   omp_set_num_threads(num_gpus);

   #pragma omp parallel
   {    
    unsigned int cpu_thread_id = omp_get_thread_num();  

    printf("### (CPU thread %d)\n",cpu_thread_id);  

    test(cpu_thread_id, num_gpus);
   }    
}

和测试函数中有这段代码：

void test(unsigned int cpu_thread_id, int num_gpus)
{   
    printf("### Using CUDA device %d, GPU = %d\n", cpu_thread_id, num_gpus);
}

即使我得到了输出：

### Number of host CPUs:        8
### Number of CUDA devices:     4
### (CPU thread 0)
### Using CUDA device 0, GPU = 4

但我期望有更多像这样的线程：

### (CPU thread 1)
### Using CUDA device 1, GPU = 4
### (CPU thread 2)
### Using CUDA device 2, GPU = 4
### (CPU thread 3)
### Using CUDA device 3, GPU = 4

为什么其他三个线程没有运行？

提前致谢。