下面是我的代码：

//initialize the device_vector
int size = N;
thrust::device_vector<glm::vec3> value(size);
thrust::device_vector<int> key(size);
//get the device pointer of the device_vector
//so than I can write data to the device_vector in CUDA kernel
glm::vec3 * dv_value_ptr = thrust::raw_pointer_cast(&value[0]);
int* dv_key_ptr = thrust::raw_pointer_cast(&key[0]);
//run the kernel function
dim3 threads(16, 16);
dim3 blocks(iDivUp(m_width, threads.x), iDivUp(m_height, threads.y));
//the size of value and key is packed in dev_data
compute_one_i_all_j <<<blocks, threads >>>(dev_data, dv_key_ptr, dv_value_ptr);
//Finally, reduce the vector by its keys.
thrust::pair<thrust::device_vector<int>::iterator,
      thrust::device_vector<glm::vec3>::iterator> new_last;
new_last = thrust::reduce_by_key(key.begin(), key.end(), value.begin(), output_key.begin(), output_value.begin());
//get the reduced vector size
int new__size = new_last.first - output_key.begin();

完成所有这些代码后，我将output_key写入一个文件。我在文件中得到了许多重复的密钥，如下所示：

reduce_by_key()似乎不起作用。
PS. CUDA内核只写入key s和value s的一部分，因此在内核之后，key和value中的一些元素保持不变（可能为0）。