Very fast sorting of fixed length arrays using comparator networks

Here is a little class that uses the Bose-Nelson algorithm to generate a sorting network on compile time.

/**
 * A Functor class to create a sort for fixed sized arrays/containers with a
 * compile time generated Bose-Nelson sorting network.
 * \tparam NumElements  The number of elements in the array or container to sort.
 * \tparam T            The element type.
 * \tparam Compare      A comparator functor class that returns true if lhs < rhs.
 */
template <unsigned NumElements, class Compare = void> class StaticSort
{
    template <class A, class C> struct Swap
    {
        template <class T> inline void s(T &v0, T &v1)
        {
            T t = Compare()(v0, v1) ? v0 : v1; // Min
            v1 = Compare()(v0, v1) ? v1 : v0; // Max
            v0 = t;
        }

        inline Swap(A &a, const int &i0, const int &i1) { s(a[i0], a[i1]); }
    };

    template <class A> struct Swap <A, void>
    {
        template <class T> inline void s(T &v0, T &v1)
        {
            // Explicitly code out the Min and Max to nudge the compiler
            // to generate branchless code.
            T t = v0 < v1 ? v0 : v1; // Min
            v1 = v0 < v1 ? v1 : v0; // Max
            v0 = t;
        }

        inline Swap(A &a, const int &i0, const int &i1) { s(a[i0], a[i1]); }
    };

    template <class A, class C, int I, int J, int X, int Y> struct PB
    {
        inline PB(A &a)
        {
            enum { L = X >> 1, M = (X & 1 ? Y : Y + 1) >> 1, IAddL = I + L, XSubL = X - L };
            PB<A, C, I, J, L, M> p0(a);
            PB<A, C, IAddL, J + M, XSubL, Y - M> p1(a);
            PB<A, C, IAddL, J, XSubL, M> p2(a);
        }
    };

    template <class A, class C, int I, int J> struct PB <A, C, I, J, 1, 1>
    {
        inline PB(A &a) { Swap<A, C> s(a, I - 1, J - 1); }
    };

    template <class A, class C, int I, int J> struct PB <A, C, I, J, 1, 2>
    {
        inline PB(A &a) { Swap<A, C> s0(a, I - 1, J); Swap<A, C> s1(a, I - 1, J - 1); }
    };

    template <class A, class C, int I, int J> struct PB <A, C, I, J, 2, 1>
    {
        inline PB(A &a) { Swap<A, C> s0(a, I - 1, J - 1); Swap<A, C> s1(a, I, J - 1); }
    };

    template <class A, class C, int I, int M, bool Stop = false> struct PS
    {
        inline PS(A &a)
        {
            enum { L = M >> 1, IAddL = I + L, MSubL = M - L};
            PS<A, C, I, L, (L <= 1)> ps0(a);
            PS<A, C, IAddL, MSubL, (MSubL <= 1)> ps1(a);
            PB<A, C, I, IAddL, L, MSubL> pb(a);
        }
    };

    template <class A, class C, int I, int M> struct PS <A, C, I, M, true>
    {
        inline PS(A &a) {}
    };

public:
    /**
     * Sorts the array/container arr.
     * \param  arr  The array/container to be sorted.
     */
    template <class Container> inline void operator() (Container &arr) const
    {
        PS<Container, Compare, 1, NumElements, (NumElements <= 1)> ps(arr);
    };

    /**
     * Sorts the array arr.
     * \param  arr  The array to be sorted.
     */
    template <class T> inline void operator() (T *arr) const
    {
        PS<T*, Compare, 1, NumElements, (NumElements <= 1)> ps(arr);
    };
};

#include <iostream>
#include <vector>

int main(int argc, const char * argv[])
{
    enum { NumValues = 32 };

    // Arrays
    {
        int rands[NumValues];
        for (int i = 0; i < NumValues; ++i) rands[i] = rand() % 100;
        std::cout << "Before Sort: \t";
        for (int i = 0; i < NumValues; ++i) std::cout << rands[i] << " ";
        std::cout << "\n";
        StaticSort<NumValues> staticSort;
        staticSort(rands);
        std::cout << "After Sort: \t";
        for (int i = 0; i < NumValues; ++i) std::cout << rands[i] << " ";
        std::cout << "\n";
    }

    std::cout << "\n";

    // STL Vector
    {
        std::vector<int> rands(NumValues);
        for (int i = 0; i < NumValues; ++i) rands[i] = rand() % 100;
        std::cout << "Before Sort: \t";
        for (int i = 0; i < NumValues; ++i) std::cout << rands[i] << " ";
        std::cout << "\n";
        StaticSort<NumValues> staticSort;
        staticSort(rands);
        std::cout << "After Sort: \t";
        for (int i = 0; i < NumValues; ++i) std::cout << rands[i] << " ";
        std::cout << "\n";
    }

    return 0;
}

Benchmarks

The following benchmarks are compiled with clang -O3 and ran on my mid-2012 macbook air.

Time (in milliseconds) to sort 1 million arrays.
The number of milliseconds for arrays of size 2, 4, 8 are 1.943, 8.655, 20.246 respectively.

Here are the average clocks per sort for small arrays of 6 elements. The benchmark code and examples can be found at this question:
Fastest sort of fixed length 6 int array

Direct call to qsort library function   : 342.26
Naive implementation (insertion sort)   : 136.76
Insertion Sort (Daniel Stutzbach)       : 101.37
Insertion Sort Unrolled                 : 110.27
Rank Order                              : 90.88
Rank Order with registers               : 90.29
Sorting Networks (Daniel Stutzbach)     : 93.66
Sorting Networks (Paul R)               : 31.54
Sorting Networks 12 with Fast Swap      : 32.06
Sorting Networks 12 reordered Swap      : 29.74
Reordered Sorting Network w/ fast swap  : 25.28
Templated Sorting Network (this class)  : 25.01 

It performs as fast as the fastest example in the question for 6 elements.

The code used for the benchmarks can be found here.

It includes more features and further optimizations for more robust performance on real-world data.

The other answers are interesting and fairly good, but I believe that I can provide some additional elements of answer, point per point:

• Is it worth the effort? Well, if you need to sort small collections of integers and the sorting networks are tuned to take advantage of some instructions as much as possible, it might be worth the effort. The following graph presents the results of sorting a million arrays of int of size 0-14 with different sorting algorithms. As you can see, the sorting networks can provide a significant speedup if you really need it.

• No standard implementation of std::sort I know of use sorting networks; when they are not fine-tuned, they might be slower than a straight insertion sort. libc++'s std::sort has dedicated algorithms to sort 0 thru 5 values at once but they it doesn't use sorting networks either. The only sorting algorithm I know of which uses sorting networks to sort a few values is Wikisort. That said, the research paper Applying Sorting Networks to Synthesize Optimized Sorting Libraries suggests that sorting networks could be used to sort small arrays or to improve recursive sorting algorithms such as quicksort, but only if they are fine-tuned to take advantage of specific hardware instructions.

  The access aligned sort algorithm is some kind of bottom-up mergesort that apparently uses bitonic sorting networks implemented with SIMD instructions for the first pass. Apparently, the algorithm could be faster than the standard library one for some scalar types.

• I can actually provide such information for the simple reason that I developed a C++14 sorting library that happens to provide efficient sorting networks of size 0 thru 32 that implement the optimizations described in the previous section. I used it to generate the graph in the first section. I am still working on the sorting networks part of the library to provide size-optimal, depth-optimal and swaps-optimal networks. Small optimal sorting networks are found with brute force while bigger sorting networks use results from the litterature.

  Note that none of the sorting algorithms in the library directly use sorting networks, but you can adapt them so that a sorting network will be picked whenever the sorting algorithm is given a small std::array or a small fixed-size C array:

  using namespace cppsort;
  
  // Sorters are function objects that can be
  // adapted with sorter adapters from the
  // library
  using sorter = small_array_adapter<
      std_sorter,
      sorting_network_sorter
  >;
  
  // Now you can use it as a function
  sorter sort;
  
  // Instead of a size-agnostic sorting algorithm,
  // sort will use an optimal sorting network for
  // 5 inputs since the bound of the array can be
  // deduced at compile time
  int arr[] = { 2, 4, 7, 9, 3 };
  sort(arr);
  

  As mentioned above, the library provides efficient sorting networks for built-in integers, but you're probably out of luck if you need to sort small arrays of something else (e.g. my latest benchmarks show that they are not better than a straight insertion sort even for long long int).

• You could probably use template metaprogramming to generate sorting networks of any size, but no known algorithm can generate the best sorting networks, so you might as well write the best ones by hand. I don't think the ones generated by simple algorithms can actually provide usable and efficient networks anyway (Batcher's odd-even sort and pairwise sorting networks might be the only usable ones) [Another answer seems to show that generated networks could actually work].

There are known optimal or at least best length comparator networks for N<16, so there's at least a fairly good starting point. Fairly, since the optimal networks are not necessarily designed for maximum level of parallelism achievable with e.g. SSE or other vector arithmetics.

Another point is that already some optimal networks for some N are degenerate versions for a slightly larger optimal network for N+1.

From wikipedia:

The optimal depths for up to 10 inputs are known and they are respectively 0, 1, 3, 3, 5, 5, 6, 6, 7, 7.

This said, I'd pursuit for implementing networks for N={4, 6, 8 and 10}, since the depth constraint cannot be simulated by extra parallelism (I think). I also think, that the ability to work in registers of SSE (also using some min/max instructions) or even some relatively large register set in RISC architecture will provide noticeable performance advantage compared to "well known" sorting methods such as quicksort due to absence of pointer arithmetic and other overhead.

Additionally, I'd pursuit to implement the parallel network using the infamous loop unrolling trick Duff's device.

EDIT When the input values are known to be positive IEEE-754 floats or doubles, it's also worth to mention that the comparison can also be performed as integers. (float and int must have same endianness)

Let me share some thoughts.

Does anyone have any opinions on whether or not this is worth the effort?

It is impossible to give a correct answer. You have to profile your actual code to find that out. In my practice, when it comes to low-level profiling, the bottleneck was always not where I thought.

Does anyone know if this optimisation exists in any standard implementations of, for example, std::sort?

For example, Visual C++ implementation of std::sort uses insertion sort for small vectors. I'm not aware of an implementation which uses optimal sorting networks.

Perhaps it would be possible to generate a sorting network like this statically using template magic

There are algorithms for generating sorting networks, such as Bose-Nelson, Hibbard, and Batcher's algorithms. As C++ templates are Turing-complete, you can implement them using TMP. However, those algorithms are not guaranteed to give the theoretically minimal number of comparators, so you may want to hardcode the optimal network.