Fast, in-place radix sort.

#include "kxsort.h"
#include <cstdlib>
#include <vector>
#include <iostream>

int main(int argc, char **argv) {
	int N = 10000000;
	std::vector<int> a(N);
	for (int i = 0; i < N; ++i) {
		a[i] = rand() * (rand() % 2 ? 1 : -1);
	}

	kx::radix_sort(a.begin(), a.end());
	std::cout << (is_sorted(a.begin(), a.end()) ? "OK" : "Error") << std::endl;
	return 0;
}

Time (sec.) for sorting 100 million random integers.

To sort all C++ builtin integers in ascending order, the API is the same as std::sort.

template <class RandomIt>
void radix_sort(RandomIt s, RandomIt e);

template <class RandomIt, class RadixTraits>
inline void radix_sort(RandomIt s, RandomIt e, RadixTraits r_traits);

To sort items of other value types, the user should define the RadixTraits for them. kx::radix_sort sorts items byte by byte (i.e. the bucket size is 256 for each round). A RadixTraits struct must implement the following public interfaces,

struct SampleRadixTraits {
    static const int nBytes;
    int kth_byte(const YourType &x, int k);
    bool compare(const YourType &x, const YourType &y);
};

where nBytes specifies how many rounds the radix sort will go over, and kth_byte(const YourType &x, int k) is the function to extract the k-th (0-base) byte of x for radix sort, i.e. the byte to be used in the nBytes-1-k round.

The function compare should be compatible with kth_byte, which means it should return exactly the same results as the following function for any x and y, but the implementation may be faster:

bool another_compare(const YourType &x, const YourType &y) {
	for (int k = nBytes - 1; k >= 0; --k) {
		if (kth_byte(x, k) < kth_byte(y, k)) return true;
		if (kth_byte(x, k) > kth_byte(y, k)) return false;
	}
	return false;
}

You may want to look at some examples.

An MSD radix sort is implemented in kx::radix_sort. It partitions the input array with the most significant byte, and sort each partitions recursively. When the number of items in a partition is no more than 64, an insertion sort will be applied and no more recursion is needed. The using of insertion sort is the reason why it requires a compare function in RadixTraits.

The implement here is substantially the same as the following implementations

• ac's klib radix_sort
• Gorset's implementation
• Birkeland's PhD thesis

kx::radix_sort is a bit faster than ac's klib radix_sort probably due to the using of C++ template, which may unfold the recursions.

• It is not a stable sorting method.
• It is slower than std::sort if the input sequence is nearly sorted.