Counting occurrences of numbers in a CUDA array

I have an array of unsigned integers stored on the GPU with CUDA (typically 1000000 elements). I would like 开发者_如何学Cto count the occurrence of every number in the array. There are only a few distinct numbers (about 10), but these numbers can span from 1 to 1000000. About 9/10th of the numbers are 0, I don't need the count of them. The result looks something like this:

58458 -> 1000 occurrences
15 -> 412 occurrences

I have an implementation using atomicAdds, but it is too slow (a lot of threads write to the same address). Does someone know of a fast/efficient method?

You can implement a histogram by first sorting the numbers, and then doing a keyed reduction.

The most straightforward method would be to use thrust::sort and then thrust::reduce_by_key. It's also often much faster than ad hoc binning based on atomics. Here's an example.

I suppose you can find help in the CUDA examples, specifically the histogram examples. They are part of the GPU computing SDK. You can find it here http://developer.nvidia.com/cuda-cc-sdk-code-samples#histogram. They even have a whitepaper explaining the algorithms.

I'm comparing two approaches suggested at the duplicate question thrust count occurence, namely,

1. Using thrust::counting_iterator and thrust::upper_bound, following the histogram Thrust example;
2. Using thrust::unique_copy and thrust::upper_bound.

Below, please find a fully worked example.

#include <time.h>       // --- time
#include <stdlib.h>     // --- srand, rand
#include <iostream>

#include <thrust\host_vector.h>
#include <thrust\device_vector.h>
#include <thrust\sort.h>
#include <thrust\iterator\zip_iterator.h>
#include <thrust\unique.h>
#include <thrust/binary_search.h>
#include <thrust\adjacent_difference.h>

#include "Utilities.cuh"
#include "TimingGPU.cuh"

//#define VERBOSE
#define NO_HISTOGRAM

/********/
/* MAIN */
/********/
int main() {

    const int N = 1048576;
    //const int N = 20;
    //const int N = 128;

    TimingGPU timerGPU;

    // --- Initialize random seed
    srand(time(NULL));

    thrust::host_vector<int> h_code(N);

    for (int k = 0; k < N; k++) {
        // --- Generate random numbers between 0 and 9
        h_code[k] = (rand() % 10);
    }

    thrust::device_vector<int> d_code(h_code);
    //thrust::device_vector<unsigned int> d_counting(N);

    thrust::sort(d_code.begin(), d_code.end());

    h_code = d_code;

    timerGPU.StartCounter();

#ifdef NO_HISTOGRAM
    // --- The number of d_cumsum bins is equal to the maximum value plus one
    int num_bins = d_code.back() + 1;

    thrust::device_vector<int> d_code_unique(num_bins);
    thrust::unique_copy(d_code.begin(), d_code.end(), d_code_unique.begin());
    thrust::device_vector<int> d_counting(num_bins);
    thrust::upper_bound(d_code.begin(), d_code.end(), d_code_unique.begin(), d_code_unique.end(), d_counting.begin());  
#else
    thrust::device_vector<int> d_cumsum;

    // --- The number of d_cumsum bins is equal to the maximum value plus one
    int num_bins = d_code.back() + 1;

    // --- Resize d_cumsum storage
    d_cumsum.resize(num_bins);

    // --- Find the end of each bin of values - Cumulative d_cumsum
    thrust::counting_iterator<int> search_begin(0);
    thrust::upper_bound(d_code.begin(), d_code.end(), search_begin, search_begin + num_bins, d_cumsum.begin());

    // --- Compute the histogram by taking differences of the cumulative d_cumsum
    //thrust::device_vector<int> d_counting(num_bins);
    //thrust::adjacent_difference(d_cumsum.begin(), d_cumsum.end(), d_counting.begin());
#endif

    printf("Timing GPU = %f\n", timerGPU.GetCounter());

#ifdef VERBOSE
    thrust::host_vector<int> h_counting(d_counting);
    printf("After\n");
    for (int k = 0; k < N; k++) printf("code = %i\n", h_code[k]);
#ifndef NO_HISTOGRAM
    thrust::host_vector<int> h_cumsum(d_cumsum);
    printf("\nCounting\n");
    for (int k = 0; k < num_bins; k++) printf("element = %i; counting = %i; cumsum = %i\n", k, h_counting[k], h_cumsum[k]);
#else
    thrust::host_vector<int> h_code_unique(d_code_unique);

    printf("\nCounting\n");
    for (int k = 0; k < N; k++) printf("element = %i; counting = %i\n", h_code_unique[k], h_counting[k]);
#endif
#endif
}

The first approach has shown to be the fastest. On an NVIDIA GTX 960 card, I have had the following timings for a number of N = 1048576 array elements:

First approach: 2.35ms
First approach without thrust::adjacent_difference: 1.52
Second approach: 4.67ms

Please, note that there is no strict need to calculate the adjacent difference explicitly, since this operation can be manually done during a kernel processing, if needed.

As others have said, you can use the sort & reduce_by_key approach to count frequencies. In my case, I needed to get mode of an array (maximum frequency/occurrence) so here is my solution:

1 - First, we create two new arrays, one containing a copy of input data and another filled with ones to later reduce it (sum):

// Input: [1 3 3 3 2 2 3]
// *(Temp) dev_keys: [1 3 3 3 2 2 3]
// *(Temp) dev_ones: [1 1 1 1 1 1 1]

// Copy input data
thrust::device_vector<int> dev_keys(myptr, myptr+size);

// Fill an array with ones
thrust::fill(dev_ones.begin(), dev_ones.end(), 1);

2 - Then, we sort the keys since the reduce_by_key function needs the array to be sorted.

// Sort keys (see below why)
thrust::sort(dev_keys.begin(), dev_keys.end());

3 - Later, we create two output vectors, for the (unique) keys and their frequencies:

thrust::device_vector<int> output_keys(N);
thrust::device_vector<int> output_freqs(N);

4 - Finally, we perform the reduction by key:

// Reduce contiguous keys: [1 3 3 3 2 2 3] => [1 3 2 1] Vs. [1 3 3 3 3 2 2] => [1 4 2] 
thrust::pair<thrust::device_vector<int>::iterator, thrust::device_vector<int>::iterator> new_end;
new_end = thrust::reduce_by_key(dev_keys.begin(), dev_keys.end(), dev_ones.begin(), output_keys.begin(), output_freqs.begin());

5 - ...and if we want, we can get the most frequent element

// Get most frequent element
// Get index of the maximum frequency
int num_keys = new_end.first  - output_keys.begin();
thrust::device_vector<int>::iterator iter = thrust::max_element(output_freqs.begin(), output_freqs.begin() + num_keys);
unsigned int index = iter - output_freqs.begin();

int most_frequent_key = output_keys[index];
int most_frequent_val = output_freqs[index];  // Frequencies