parallel_sort.h

Go to the documentation of this file.

// -*- mode: c++; tab-width: 4; indent-tabs-mode: t; c-file-style: "stroustrup"; -*-
// vi:set ts=4 sts=4 sw=4 noet cino+=(0 :
// Copyright 2011, The TPIE development team
// 
// This file is part of TPIE.
// 
// TPIE is free software: you can redistribute it and/or modify it under
// the terms of the GNU Lesser General Public License as published by the
// Free Software Foundation, either version 3 of the License, or (at your
// option) any later version.
// 
// TPIE is distributed in the hope that it will be useful, but WITHOUT ANY
// WARRANTY; without even the implied warranty of MERCHANTABILITY or
// FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
// License for more details.
// 
// You should have received a copy of the GNU Lesser General Public License
// along with TPIE.  If not, see <http://www.gnu.org/licenses/>


#ifndef __TPIE_PARALLEL_SORT_H__
#define __TPIE_PARALLEL_SORT_H__

#include <algorithm>
#include <cstdint>
#include <boost/iterator/iterator_traits.hpp>
#include <mutex>
#include <cmath>
#include <functional>
#include <tpie/progress_indicator_base.h>
#include <tpie/dummy_progress.h>
#include <tpie/internal_queue.h>
#include <tpie/job.h>
#include <tpie/config.h>

namespace tpie {

template <typename iterator_type, typename comp_type, bool Progress,
          size_t min_size=1024*1024*8/sizeof(typename boost::iterator_value<iterator_type>::type)>
class parallel_sort_impl {
private:
    typedef progress_types<Progress> P;

    struct progress_t {
        typename P::base * pi;
        std::uint64_t work_estimate;
        std::uint64_t total_work_estimate;
        std::condition_variable cond;
        std::mutex mutex;
    };

    typedef typename boost::iterator_value<iterator_type>::type value_type;

    static inline std::uint64_t sortWork(std::uint64_t n) {
        if(n == 0)
            return 0;

        return static_cast<uint64_t>(log(static_cast<double>(n)) * static_cast<double>(n) * 1.8
                / log(static_cast<double>(2)));
    }

    template <typename comp_t>
    static inline iterator_type unguarded_partition(iterator_type first, 
                                                    iterator_type last, 
                                                    comp_t & comp) {
        // Textbook partitioning.
        iterator_type pivot = first;
        while (true) {
            do --last;
            while (comp(*pivot, *last));

            do {
                if (first == last) break;
                ++first;
            } while (comp(*first, *pivot));

            if (first == last) break;

            std::iter_swap(first, last);
        }
        std::iter_swap(last, pivot);
        return last;
    }

    static inline iterator_type median(iterator_type a, iterator_type b, iterator_type c, comp_type & comp) {
        if (comp(*a, *b)) {
            if (comp(*b, *c)) return b;
            else if (comp(*a, *c)) return c;
            else return a;
        } else {
            if (comp(*a, *c)) return a;
            else if (comp(*b, *c)) return c;
            else return b;
        }
    }

    static inline iterator_type pick_pivot(iterator_type a, iterator_type b, comp_type & comp) {
        if (a == b) return a;
        assert(a < b);

        // Since (b-a) is at least min_size, which is at least 100000 in
        // realistic contexts, ((b-a)/8)*c is a good approximation of
        // (c*(b-a))/8.
        size_t step = (b-a)/8;

        return median(median(a+0, a+step, a+step*2, comp),
                      median(a+step*3, a+step*4, a+step*5, comp),
                      median(a+step*6, a+step*7, b-1, comp), comp);
    }

    static inline iterator_type partition(iterator_type a, iterator_type b, comp_type & comp) {
        iterator_type pivot = pick_pivot(a, b, comp);

        std::iter_swap(pivot, a);
        iterator_type l = unguarded_partition(a, b, comp);

        return l;
    }

#ifdef DOXYGEN
public:
#endif
    class qsort_job : public job {
    public:
        qsort_job(iterator_type a, iterator_type b, comp_type comp, qsort_job * parent, progress_t & p)
            : a(a), b(b), comp(comp), parent(parent), progress(p) {

            // Does nothing.
        }

        ~qsort_job() {
            for (size_t i = 0; i < children.size(); ++i) {
                delete children[i];
            }
            children.resize(0);
        }

        virtual void operator()() override {
            assert(a <= b);
            while (static_cast<size_t>(b - a) >= min_size) {
                iterator_type pivot = partition(a, b, comp);
                add_progress(b - a);
                //qsort_job * j = tpie_new<qsort_job>(a, pivot, comp, this);
                qsort_job * j = new qsort_job(a, pivot, comp, this, progress);
                j->enqueue(this);
                children.push_back(j);
                a = pivot+1;
            }
            std::sort(a, b, comp);
            add_progress(sortWork(b - a));
        }

    protected:
        virtual void on_done() override {
            // Unfortunately, it might not be safe to delete our children at
            // this point, as other threads might in theory wait for them to
            // .join(). It is safer to postpone deletion until our own
            // deletion.
            if (!parent) {
                std::lock_guard<std::mutex> lock(progress.mutex);
                progress.work_estimate = progress.total_work_estimate;
                progress.cond.notify_one();
            }
        }

    private:
        iterator_type a;
        iterator_type b;
        comp_type comp;
        qsort_job * parent;
        progress_t & progress;

        std::vector<qsort_job *> children;

        void add_progress(uint64_t amount) {
            std::lock_guard<std::mutex> lock(progress.mutex);
            progress.work_estimate += amount;
            progress.cond.notify_one();
        }
    };
public:
    parallel_sort_impl(typename P::base * p) {
        progress.pi = p;
    }

    void operator()(iterator_type a, iterator_type b, comp_type comp=std::less<value_type>() ) {
        progress.work_estimate = 0;
        progress.total_work_estimate = sortWork(b-a);
        if (progress.pi) progress.pi->init(progress.total_work_estimate);

        if (static_cast<size_t>(b - a) < min_size) {
            std::sort(a, b, comp);
            if (progress.pi) progress.pi->done();
            return;
        }

        qsort_job * master = new qsort_job(a, b, comp, 0, progress);
        master->enqueue();

        std::uint64_t prev_work_estimate = 0;
        std::unique_lock<std::mutex> lock(progress.mutex);
        while (progress.work_estimate < progress.total_work_estimate) {
            if (progress.pi && progress.work_estimate > prev_work_estimate) progress.pi->step(progress.work_estimate - prev_work_estimate);
            prev_work_estimate = progress.work_estimate;
            progress.cond.wait(lock);
        }
        lock.unlock();

        master->join();
        delete master;
        if (progress.pi) progress.pi->done();
    }
private:
    static const size_t max_job_count=256;
    progress_t progress;
    bool kill;
    size_t working;

    std::pair<iterator_type, iterator_type> jobs[max_job_count];
    size_t job_count;
};

template <bool Progress, typename iterator_type, typename comp_type>
void parallel_sort(iterator_type a, 
                   iterator_type b, 
                   typename tpie::progress_types<Progress>::base & pi,
                   comp_type comp=std::less<typename boost::iterator_value<iterator_type>::type>()) {
#ifdef TPIE_PARALLEL_SORT
    parallel_sort_impl<iterator_type, comp_type, Progress> s(&pi);
    s(a,b,comp);
#else
    pi.init(1);
    std::sort(a,b,comp);
    pi.done();
#endif
}

template <typename iterator_type, typename comp_type>
void parallel_sort(iterator_type a, 
                   iterator_type b, 
                   comp_type comp=std::less<typename boost::iterator_value<iterator_type>::type>()) {
#ifdef TPIE_PARALLEL_SORT
    parallel_sort_impl<iterator_type, comp_type, false> s(0);
    s(a,b,comp);
#else
    std::sort(a, b, comp);
#endif
}


}
#endif //__TPIE_PARALLEL_SORT_H__