thrust::reduce_by_key() seems not work - cuda

Question

Here is my code:
//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();
After all these code, I write the output_key to a file. I get so many duplicated keys in the file, as below:
So, the reduce_by_key() seems not work .
Ps. The CUDA kernel only write part of the key and value, so after the kernel some of the elements in key and value remains unchanged (likely 0).

Answer 1

As stated in documentation:
For each group of consecutive keys in the range [keys_first, keys_last) that are equal, reduce_by_key copies the first element of the group to the keys_output. The corresponding values in the range are reduced using the plus and the result copied to values_output.
Every group of equal consecutive keys will be reduced.
So first of all you must rearrange all your keys and values so all elements with equal keys are adjacent. The simpliest method will be to use sort_by_key.
thrust::sort_by_key(key.begin(), key.end(), value.begin())
new_last = thrust::reduce_by_key(key.begin(), key.end(), value.begin(), output_key.begin(), output_value.begin());