parallel_reduce/convex_hull/convex_hull_bench.cpp

/*
    Copyright (c) 2005-2021 Intel Corporation

    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License.
*/

/*
    This file contains a few implementations, so it may look overly complicated.
    The most efficient implementation is also separated into convex_hull_sample.cpp
*/
#include "convex_hull.hpp"

typedef util::point<double> point_t;

#ifndef USETBB
#define USETBB 1
#endif
#ifndef USECONCVEC
#define USECONCVEC 1
#endif

#if !USETBB // Serial implementation of Quick Hull algorithm

typedef std::vector<point_t> pointVec_t;

void serial_initialize(pointVec_t &points);

class FindXExtremum : public std::unary_function<const point_t &, void> {
public:
    typedef enum { minX, maxX } extremumType;

    FindXExtremum(const point_t &frstPoint, extremumType exType_)
            : extrXPoint(frstPoint),
              exType(exType_) {}

    void operator()(const point_t &p) {
        if (closerToExtremum(p))
            extrXPoint = p;
    }

    operator point_t() {
        return extrXPoint;
    }

private:
    const extremumType exType;
    point_t extrXPoint;

    bool closerToExtremum(const point_t &p) const {
        switch (exType) {
            case minX: return p.x < extrXPoint.x; break;
            case maxX: return p.x > extrXPoint.x; break;
        }
        return false; // avoid warning
    }
};

template <FindXExtremum::extremumType type>
point_t extremum(const pointVec_t &points) {
    assert(!points.empty());
    return std::for_each(points.begin(), points.end(), FindXExtremum(points[0], type));
}

class SplitByCP : public std::unary_function<const point_t &, void> {
    pointVec_t &reducedSet;
    point_t p1, p2;
    point_t farPoint;
    double howFar;

public:
    SplitByCP(point_t _p1, point_t _p2, pointVec_t &_reducedSet)
            : p1(_p1),
              p2(_p2),
              reducedSet(_reducedSet),
              howFar(0),
              farPoint(p1) {}

    void operator()(const point_t &p) {
        double cp;
        if ((p != p1) && (p != p2)) {
            cp = util::cross_product(p1, p2, p);
            if (cp > 0) {
                reducedSet.push_back(p);
                if (cp > howFar) {
                    farPoint = p;
                    howFar = cp;
                }
            }
        }
    }

    operator point_t() {
        return farPoint;
    }
};

point_t divide(const pointVec_t &P, pointVec_t &P_reduced, const point_t &p1, const point_t &p2) {
    SplitByCP splitByCP(p1, p2, P_reduced);
    point_t farPoint = std::for_each(P.begin(), P.end(), splitByCP);

    if (util::verbose) {
        std::stringstream ss;
        ss << P.size() << " nodes in bucket"
           << ", "
           << "dividing by: [ " << p1 << ", " << p2 << " ], "
           << "farthest node: " << farPoint;
        util::OUTPUT.push_back(ss.str());
    }

    return farPoint;
}

void divide_and_conquer(const pointVec_t &P, pointVec_t &H, point_t p1, point_t p2) {
    assert(P.size() >= 2);
    pointVec_t P_reduced;
    pointVec_t H1, H2;
    point_t p_far = divide(P, P_reduced, p1, p2);
    if (P_reduced.size() < 2) {
        H.push_back(p1);
        H.insert(H.end(), P_reduced.begin(), P_reduced.end());
    }
    else {
        divide_and_conquer(P_reduced, H1, p1, p_far);
        divide_and_conquer(P_reduced, H2, p_far, p2);

        H.insert(H.end(), H1.begin(), H1.end());
        H.insert(H.end(), H2.begin(), H2.end());
    }
}

void quickhull(const pointVec_t &points, pointVec_t &hull) {
    if (points.size() < 2) {
        hull.insert(hull.end(), points.begin(), points.end());
        return;
    }
    point_t p_maxx = extremum<FindXExtremum::maxX>(points);
    point_t p_minx = extremum<FindXExtremum::minX>(points);

    pointVec_t H;

    divide_and_conquer(points, hull, p_maxx, p_minx);
    divide_and_conquer(points, H, p_minx, p_maxx);
    hull.insert(hull.end(), H.begin(), H.end());
}

int main(int argc, char *argv[]) {
    utility::thread_number_range single_thread([] {
        return -1;
    });
    util::ParseInputArgs(argc, argv, single_thread);

    pointVec_t points;
    pointVec_t hull;
    util::my_time_t tm_init, tm_start, tm_end;

    std::cout << "Starting serial version of QUICK HULL algorithm"
              << "\n";

    tm_init = util::gettime();
    serial_initialize(points);
    tm_start = util::gettime();
    std::cout << "Init time: " << util::time_diff(tm_init, tm_start)
              << "  Points in input: " << points.size() << "\n";
    tm_start = util::gettime();
    quickhull(points, hull);
    tm_end = util::gettime();
    std::cout << "Serial time: " << util::time_diff(tm_start, tm_end)
              << "  Points in hull: " << hull.size() << "\n";

    return 0;
}

#else // USETBB - parallel version of Quick Hull algorithm

#include "oneapi/tbb/parallel_for.h"
#include "oneapi/tbb/parallel_reduce.h"
#include "oneapi/tbb/blocked_range.h"

typedef oneapi::tbb::blocked_range<std::size_t> range_t;

#if USECONCVEC
#include "oneapi/tbb/concurrent_vector.h"

typedef oneapi::tbb::concurrent_vector<point_t> pointVec_t;

void appendVector(const point_t *src, std::size_t srcSize, pointVec_t &dest) {
    std::copy(src, src + srcSize, dest.grow_by(srcSize));
}

void appendVector(const pointVec_t &src, pointVec_t &dest) {
    std::copy(src.begin(), src.end(), dest.grow_by(src.size()));
}

void grow_vector_to_at_least(pointVec_t &vect, std::size_t size) {
    vect.grow_to_at_least(size);
}
#else // USE STD::VECTOR - include spin_mutex.h and lock vector operations
#include "oneapi/tbb/spin_mutex.h"

typedef oneapi::tbb::spin_mutex mutex_t;
typedef std::vector<point_t> pointVec_t;

void appendVector(mutex_t &insertMutex, const pointVec_t &src, pointVec_t &dest) {
    mutex_t::scoped_lock lock(insertMutex);
    dest.insert(dest.end(), src.begin(), src.end());
}

void appendVector(mutex_t &insertMutex, const point_t *src, std::size_t srcSize, pointVec_t &dest) {
    mutex_t::scoped_lock lock(insertMutex);
    dest.insert(dest.end(), src, src + srcSize);
}

void grow_vector_to_at_least(mutex_t &mutex, pointVec_t &vect, std::size_t size) {
    mutex_t::scoped_lock lock(mutex);
    if (vect.size() < size) {
        vect.resize(size);
    }
}
#endif // USECONCVEC

class FillRNDPointsVector {
    pointVec_t &points;

public:
    static const std::size_t grainSize = cfg::generateGrainSize;
#if !USECONCVEC
    static mutex_t pushBackMutex;
#endif // USECONCVEC

    explicit FillRNDPointsVector(pointVec_t &_points) : points(_points) {}

    void operator()(const range_t &range) const {
        util::rng the_rng(range.begin());
        const std::size_t i_end = range.end();
        std::size_t count = 0;
#if USECONCVEC
        points.grow_to_at_least(i_end);
#else // Locked enlarge to a not thread-safe STD::VECTOR
        grow_vector_to_at_least(pushBackMutex, points, i_end);
#endif // USECONCVEC

        for (std::size_t i = range.begin(); i != i_end; ++i) {
            points[i] = util::GenerateRNDPoint<double>(count, the_rng, util::rng::max_rand);
        }
    }
};

class FillRNDPointsVector_buf {
    pointVec_t &points;

public:
    static const std::size_t grainSize = cfg::generateGrainSize;
#if !USECONCVEC
    static mutex_t insertMutex;
#endif // USECONCVEC

    explicit FillRNDPointsVector_buf(pointVec_t &_points) : points(_points) {}

    void operator()(const range_t &range) const {
        util::rng the_rng(range.begin());
        const std::size_t i_end = range.end();
        std::size_t count = 0, j = 0;
        point_t tmp_vec[grainSize];

        for (std::size_t i = range.begin(); i != i_end; ++i) {
            tmp_vec[j++] = util::GenerateRNDPoint<double>(count, the_rng, util::rng::max_rand);
        }
#if USECONCVEC
        grow_vector_to_at_least(points, range.end());
#else // USE STD::VECTOR
        grow_vector_to_at_least(insertMutex, points, range.end());
#endif // USECONCVEC
        std::copy(tmp_vec, tmp_vec + j, points.begin() + range.begin());
    }
};

#if !USECONCVEC
mutex_t FillRNDPointsVector::pushBackMutex{};
mutex_t FillRNDPointsVector_buf::insertMutex{};
#endif

template <typename BodyType>
void initialize(pointVec_t &points) {
    //This function generate the same series of point on every call.
    //Reproducibility is needed for benchmarking to produce reliable results.
    //It is achieved through the following points:
    //      - FillRNDPointsVector_buf instance has its own local instance
    //        of random number generator, which in turn does not use any global data
    //      - oneapi::tbb::simple_partitioner produce the same set of ranges on every call to
    //        oneapi::tbb::parallel_for
    //      - local RNG instances are seeded by the starting indexes of corresponding ranges
    //      - grow_to_at_least() enables putting points into the resulting vector in deterministic order
    //        (unlike concurrent push_back or grow_by).

    // In the buffered version, a temporary storage for as much as grainSize elements
    // is allocated inside the body. Since auto_partitioner may increase effective
    // range size which would cause a crash, simple partitioner has to be used.

    oneapi::tbb::parallel_for(range_t(0, cfg::numberOfPoints, BodyType::grainSize),
                              BodyType(points),
                              oneapi::tbb::simple_partitioner());
}

class FindXExtremum {
public:
    typedef enum { minX, maxX } extremumType;

    static const std::size_t grainSize = cfg::findExtremumGrainSize;

    FindXExtremum(const pointVec_t &points_, extremumType exType_)
            : points(points_),
              exType(exType_),
              extrXPoint(points[0]) {}

    FindXExtremum(const FindXExtremum &fxex, oneapi::tbb::split)
            : points(fxex.points),
              exType(fxex.exType),
              extrXPoint(fxex.extrXPoint) {}

    void operator()(const range_t &range) {
        const std::size_t i_end = range.end();
        if (!range.empty()) {
            for (std::size_t i = range.begin(); i != i_end; ++i) {
                if (closerToExtremum(points[i])) {
                    extrXPoint = points[i];
                }
            }
        }
    }

    void join(const FindXExtremum &rhs) {
        if (closerToExtremum(rhs.extrXPoint)) {
            extrXPoint = rhs.extrXPoint;
        }
    }

    point_t extremeXPoint() {
        return extrXPoint;
    }

private:
    const pointVec_t &points;
    const extremumType exType;
    point_t extrXPoint;
    bool closerToExtremum(const point_t &p) const {
        switch (exType) {
            case minX: return p.x < extrXPoint.x; break;
            case maxX: return p.x > extrXPoint.x; break;
        }
        return false; // avoid warning
    }
};

template <FindXExtremum::extremumType type>
point_t extremum(const pointVec_t &P) {
    FindXExtremum fxBody(P, type);
    oneapi::tbb::parallel_reduce(range_t(0, P.size(), FindXExtremum::grainSize), fxBody);
    return fxBody.extremeXPoint();
}

class SplitByCP {
    const pointVec_t &initialSet;
    pointVec_t &reducedSet;
    point_t p1, p2;
    point_t farPoint;
    double howFar;

public:
    static const std::size_t grainSize = cfg::divideGrainSize;
#if !USECONCVEC
    static mutex_t pushBackMutex;
#endif // USECONCVEC

    SplitByCP(point_t _p1, point_t _p2, const pointVec_t &_initialSet, pointVec_t &_reducedSet)
            : p1(_p1),
              p2(_p2),
              initialSet(_initialSet),
              reducedSet(_reducedSet),
              howFar(0),
              farPoint(p1) {}

    SplitByCP(SplitByCP &sbcp, oneapi::tbb::split)
            : p1(sbcp.p1),
              p2(sbcp.p2),
              initialSet(sbcp.initialSet),
              reducedSet(sbcp.reducedSet),
              howFar(0),
              farPoint(p1) {}

    void operator()(const range_t &range) {
        const std::size_t i_end = range.end();
        double cp;
        for (std::size_t i = range.begin(); i != i_end; ++i) {
            if ((initialSet[i] != p1) && (initialSet[i] != p2)) {
                cp = util::cross_product(p1, p2, initialSet[i]);
                if (cp > 0) {
#if USECONCVEC
                    reducedSet.push_back(initialSet[i]);
#else // Locked push_back to a not thread-safe STD::VECTOR
                    {
                        mutex_t::scoped_lock lock(pushBackMutex);
                        reducedSet.push_back(initialSet[i]);
                    }
#endif // USECONCVEC
                    if (cp > howFar) {
                        farPoint = initialSet[i];
                        howFar = cp;
                    }
                }
            }
        }
    }

    void join(const SplitByCP &rhs) {
        if (rhs.howFar > howFar) {
            howFar = rhs.howFar;
            farPoint = rhs.farPoint;
        }
    }

    point_t farthestPoint() const {
        return farPoint;
    }
};

class SplitByCP_buf {
    const pointVec_t &initialSet;
    pointVec_t &reducedSet;
    point_t p1, p2;
    point_t farPoint;
    double howFar;

public:
    static const std::size_t grainSize = cfg::divideGrainSize;
#if !USECONCVEC
    static mutex_t insertMutex;
#endif // USECONCVEC

    SplitByCP_buf(point_t _p1, point_t _p2, const pointVec_t &_initialSet, pointVec_t &_reducedSet)
            : p1(_p1),
              p2(_p2),
              initialSet(_initialSet),
              reducedSet(_reducedSet),
              howFar(0),
              farPoint(p1) {}

    SplitByCP_buf(SplitByCP_buf &sbcp, oneapi::tbb::split)
            : p1(sbcp.p1),
              p2(sbcp.p2),
              initialSet(sbcp.initialSet),
              reducedSet(sbcp.reducedSet),
              howFar(0),
              farPoint(p1) {}

    void operator()(const range_t &range) {
        const std::size_t i_end = range.end();
        std::size_t j = 0;
        double cp;
        point_t tmp_vec[grainSize];
        for (std::size_t i = range.begin(); i != i_end; ++i) {
            if ((initialSet[i] != p1) && (initialSet[i] != p2)) {
                cp = util::cross_product(p1, p2, initialSet[i]);
                if (cp > 0) {
                    tmp_vec[j++] = initialSet[i];
                    if (cp > howFar) {
                        farPoint = initialSet[i];
                        howFar = cp;
                    }
                }
            }
        }

#if USECONCVEC
        appendVector(tmp_vec, j, reducedSet);
#else // USE STD::VECTOR
        appendVector(insertMutex, tmp_vec, j, reducedSet);
#endif // USECONCVEC
    }

    void join(const SplitByCP_buf &rhs) {
        if (rhs.howFar > howFar) {
            howFar = rhs.howFar;
            farPoint = rhs.farPoint;
        }
    }

    point_t farthestPoint() const {
        return farPoint;
    }
};

#if !USECONCVEC
mutex_t SplitByCP::pushBackMutex{};
mutex_t SplitByCP_buf::insertMutex{};
#endif

template <typename BodyType>
point_t divide(const pointVec_t &P, pointVec_t &P_reduced, const point_t &p1, const point_t &p2) {
    BodyType body(p1, p2, P, P_reduced);
    // Must use simple_partitioner (see the comment in initialize() above)
    oneapi::tbb::parallel_reduce(
        range_t(0, P.size(), BodyType::grainSize), body, oneapi::tbb::simple_partitioner());

    if (util::verbose) {
        std::stringstream ss;
        ss << P.size() << " nodes in bucket"
           << ", "
           << "dividing by: [ " << p1 << ", " << p2 << " ], "
           << "farthest node: " << body.farthestPoint();
        util::OUTPUT.push_back(ss.str());
    }

    return body.farthestPoint();
}

void divide_and_conquer(const pointVec_t &P, pointVec_t &H, point_t p1, point_t p2, bool buffered) {
    assert(P.size() >= 2);
    pointVec_t P_reduced;
    pointVec_t H1, H2;
    point_t p_far;

    if (buffered) {
        p_far = divide<SplitByCP_buf>(P, P_reduced, p1, p2);
    }
    else {
        p_far = divide<SplitByCP>(P, P_reduced, p1, p2);
    }

    if (P_reduced.size() < 2) {
        H.push_back(p1);
#if USECONCVEC
        appendVector(P_reduced, H);
#else // insert into STD::VECTOR
        H.insert(H.end(), P_reduced.begin(), P_reduced.end());
#endif
    }
    else {
        divide_and_conquer(P_reduced, H1, p1, p_far, buffered);
        divide_and_conquer(P_reduced, H2, p_far, p2, buffered);

#if USECONCVEC
        appendVector(H1, H);
        appendVector(H2, H);
#else // insert into STD::VECTOR
        H.insert(H.end(), H1.begin(), H1.end());
        H.insert(H.end(), H2.begin(), H2.end());
#endif
    }
}

void quickhull(const pointVec_t &points, pointVec_t &hull, bool buffered) {
    if (points.size() < 2) {
#if USECONCVEC
        appendVector(points, hull);
#else // STD::VECTOR
        hull.insert(hull.end(), points.begin(), points.end());
#endif // USECONCVEC
        return;
    }

    point_t p_maxx = extremum<FindXExtremum::maxX>(points);
    point_t p_minx = extremum<FindXExtremum::minX>(points);

    pointVec_t H;

    divide_and_conquer(points, hull, p_maxx, p_minx, buffered);
    divide_and_conquer(points, H, p_minx, p_maxx, buffered);
#if USECONCVEC
    appendVector(H, hull);
#else // STD::VECTOR
    hull.insert(hull.end(), H.begin(), H.end());
#endif // USECONCVEC
}

int main(int argc, char *argv[]) {
    utility::thread_number_range threads(utility::get_default_num_threads);
    util::ParseInputArgs(argc, argv, threads);

    int nthreads;
    util::my_time_t tm_init, tm_start, tm_end;

#if USECONCVEC
    std::cout << "Starting TBB unbuffered push_back version of QUICK HULL algorithm"
              << "\n";
#else
    std::cout << "Starting STL locked unbuffered push_back version of QUICK HULL algorithm"
              << "\n";
#endif // USECONCVEC

    for (nthreads = threads.first; nthreads <= threads.last; nthreads = threads.step(nthreads)) {
        pointVec_t points;
        pointVec_t hull;

        oneapi::tbb::global_control c(oneapi::tbb::global_control::max_allowed_parallelism,
                                      nthreads);
        tm_init = util::gettime();
        initialize<FillRNDPointsVector>(points);
        tm_start = util::gettime();
        std::cout << "Parallel init time on " << nthreads
                  << " threads: " << util::time_diff(tm_init, tm_start)
                  << "  Points in input: " << points.size() << "\n";

        tm_start = util::gettime();
        quickhull(points, hull, false);
        tm_end = util::gettime();
        std::cout << "Time on " << nthreads << " threads: " << util::time_diff(tm_start, tm_end)
                  << "  Points in hull: " << hull.size() << "\n";
    }

#if USECONCVEC
    std::cout << "Starting TBB buffered version of QUICK HULL algorithm"
              << "\n";
#else
    std::cout << "Starting STL locked buffered version of QUICK HULL algorithm"
              << "\n";
#endif

    for (nthreads = threads.first; nthreads <= threads.last; nthreads = threads.step(nthreads)) {
        pointVec_t points;
        pointVec_t hull;

        oneapi::tbb::global_control c(oneapi::tbb::global_control::max_allowed_parallelism,
                                      nthreads);

        tm_init = util::gettime();
        initialize<FillRNDPointsVector_buf>(points);
        tm_start = util::gettime();
        std::cout << "Init time on " << nthreads
                  << " threads: " << util::time_diff(tm_init, tm_start)
                  << "  Points in input: " << points.size() << "\n";

        tm_start = util::gettime();
        quickhull(points, hull, true);
        tm_end = util::gettime();
        std::cout << "Time on " << nthreads << " threads: " << util::time_diff(tm_start, tm_end)
                  << "  Points in hull: " << hull.size() << "\n";
    }

    return 0;
}

#endif // USETBB

void serial_initialize(pointVec_t &points) {
    points.reserve(cfg::numberOfPoints);

    unsigned int rseed = 1;
    for (std::size_t i = 0, count = 0; long(i) < cfg::numberOfPoints; ++i) {
        points.push_back(util::GenerateRNDPoint<double>(count, &std::rand, RAND_MAX));
    }
}