0

我需要在 c++ 中使用英特尔线程构建块的并行归并排序的最佳代码

4

2 回答 2

2

首先，让我说，根据我的经验， tbb::parallel_sort() 非常有效，并且比我即将发布的代码要快一些（至少对于我已经输入的数千个元素的顺序）测试）。

话虽如此，我认为以下代码正是您正在寻找的。变量应该是自我解释的，代码中的文档应该解释其余的 -

这将需要并行化：

#include<tbb/parallel_invoke.h>

如果您选择使用可能更快的 Concurrency::parallel_invoke()，请包含以下内容：

#include<ppl.h>

我推荐这些设置 -

#define MIN_ELEMENTS_FOR_RECURSION            (50)
#define MIN_ELEMENTS_FOR_PARALLEL_PROCESSING  (100)

以下是要调用的主要函数。参数是随机访问类（例如向量、双端队列等）的开始和结束的迭代器和比较函数 -

template <typename T_it, typename T_it_dereferenced>
void parallelMergeSort( T_it first, T_it last,  bool (*firstLessThanSecond)(const T_it_dereferenced& a, const T_it_dereferenced& b) )
{
    // create copy of container for extra space
    std::vector<T_it_dereferenced> copy(first, last);

    parallelMergeSortRecursive( first, last, copy.begin(), copy.end(), firstLessThanSecond );
}

这从 parallelMergeSort() 递归调用，以便对每一半进行排序 -

template <typename T_it, typename T_it_dereferenced>
void parallelMergeSortRecursive( T_it source_first, T_it source_last, T_it copy_first, T_it copy_last,
bool (*firstLessThanSecond)(const T_it_dereferenced& a, const T_it_dereferenced& b), int recursion_depth = 0 )
{
    // divide the array in two, and sort the two halves

    long num_elements = source_last - source_first;

    if ( num_elements > MIN_ELEMENTS_FOR_RECURSION )
    {
        T_it source_middle = source_first + num_elements / 2;
        T_it copy_middle = copy_first + num_elements / 2;

        if ( num_elements > MIN_ELEMENTS_FOR_PARALLEL_PROCESSING )
        {
            // Concurrency::parallel_invoke() may work faster
            tbb::parallel_invoke(
                [=] { parallelMergeSortRecursive( source_first,     source_middle,  copy_first,  copy_middle,   firstLessThanSecond, recursion_depth + 1 ); },
                [=] { parallelMergeSortRecursive( source_middle,    source_last,    copy_middle, copy_last,     firstLessThanSecond, recursion_depth + 1 ); }
            );
        }
        else // sort serially rather than in parallel
        {
            parallelMergeSortRecursive( source_first,   source_middle,  copy_first,  copy_middle,   firstLessThanSecond, recursion_depth + 1 );
            parallelMergeSortRecursive( source_middle,  source_last,    copy_middle, copy_last,     firstLessThanSecond, recursion_depth + 1 );
        }

        // merge the two sorted halves

        // we switch source <--> target with each level of recursion.
        // at even recursion depths (including zero which is the root level) we assume the source is sorted and merge into the target

        if ( recursion_depth % 2 == 0 ) 
        {
            merge( source_first, copy_first, copy_middle, copy_last, firstLessThanSecond );
        }
        else
        {
            merge( copy_first, source_first, source_middle, source_last, firstLessThanSecond );
        }
    }
    else // very few elements remain to be sorted, stop the recursion and sort in place
    {
        if ( recursion_depth % 2 == 0 )
        {
            std::stable_sort(source_first, source_last, firstLessThanSecond);
        }
        else
        {
            std::stable_sort(copy_first, copy_last, firstLessThanSecond);
        }
    }
}

这是从递归函数调用的，以合并两半 -

template <typename T_it, typename T_it_dereferenced>
void merge( T_it target_first, T_it source_first, T_it source_middle, T_it source_last,
bool (*firstLessThanSecond)(const T_it_dereferenced& a, const T_it_dereferenced& b) )
{
    // source is assumed to contain two sorted sequences (from first to middle and from middle to last)

    T_it source_it1 = source_first;
    T_it source_it2 = source_middle;
    T_it target_it = target_first;

    for ( /* intentional */ ; source_it1 < source_middle && source_it2 < source_last ; ++target_it )
    {
        //if ( source_container[i] < source_container[j] )
        if (  firstLessThanSecond(*source_it1, *source_it2)  )
        {
            *target_it = *source_it1;
            ++source_it1;
        }
        else
        {
            *target_it = *source_it2;
            ++source_it2;
        }
    }

    // insert remaining elements in non-completely-traversed-half into original container
    // only one of these two whiles will execute since one of the conditions caused the previous while to stop

    for ( /* intentional */ ; source_it1 < source_middle ; ++target_it )
    {
        *target_it = *source_it1;
        ++source_it1;
    }

    for ( /* intentional */ ; source_it2 < source_last ; ++target_it )
    {
        *target_it = *source_it2;
        ++source_it2;
    }
}
于 2013-06-20T09:43:17.360 回答
1

TBB 已经包含了一种排序方法（并行快速排序），但它的实现很差（运行时至少是线性的，与处理器数量无关）。

我的建议是从现有实现中移植并行合并排序。例如使用 OpenMP 的 gnu 并行模式排序（包含在任何最近的带有源文件的 gcc 中）。只需用#pragma omp一些 tbb 并行代码替换所有内容。

于 2011-03-29T19:33:47.793 回答