nanoflann.hpp

/***********************************************************************
 * Software License Agreement (BSD License)
 *
 * Copyright 2008-2009  Marius Muja (mariusm@cs.ubc.ca). All rights reserved.
 * Copyright 2008-2009  David G. Lowe (lowe@cs.ubc.ca). All rights reserved.
 * Copyright 2011-2025  Jose Luis Blanco (joseluisblancoc@gmail.com).
 *   All rights reserved.
 *
 * THE BSD LICENSE
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *************************************************************************/

 
#pragma once
 
#include <algorithm>
#include <array>
#include <atomic>
#include <cassert>
#include <cmath>  // for abs()
#include <cstdint>
#include <cstdlib>  // for abs()
#include <functional>  // std::reference_wrapper
#include <future>
#include <istream>
#include <limits>  // std::numeric_limits
#include <ostream>
#include <stack>
#include <stdexcept>
#include <unordered_set>
#include <vector>

#define NANOFLANN_VERSION 0x190
 
// Avoid conflicting declaration of min/max macros in Windows headers
#if !defined(NOMINMAX) && (defined(_WIN32) || defined(_WIN32_) || defined(WIN32) || defined(_WIN64))
#define NOMINMAX
#ifdef max
#undef max
#undef min
#endif
#endif
// Avoid conflicts with X11 headers
#ifdef None
#undef None
#endif
 
// Handle restricted pointers
#if defined(__GNUC__) || defined(__clang__)
#define NANOFLANN_RESTRICT __restrict__
#elif defined(_MSC_VER)
#define NANOFLANN_RESTRICT __restrict
#else
#define NANOFLANN_RESTRICT
#endif
 
// Memory alignment of KD-tree nodes:
#ifndef NANOFLANN_NODE_ALIGNMENT
#define NANOFLANN_NODE_ALIGNMENT 16
#endif
 
namespace nanoflann
{

template <typename T>
constexpr T pi_const()
{
    return static_cast<T>(3.14159265358979323846);
}

template <typename T, typename = int>
struct has_resize : std::false_type
{
};
 
template <typename T>
struct has_resize<T, decltype((void)std::declval<T>().resize(1), 0)> : std::true_type
{
};
 
template <typename T, typename = int>
struct has_assign : std::false_type
{
};
 
template <typename T>
struct has_assign<T, decltype((void)std::declval<T>().assign(1, 0), 0)> : std::true_type
{
};

template <typename Container>
inline typename std::enable_if<has_resize<Container>::value, void>::type resize(
    Container& c, const size_t nElements)
{
    c.resize(nElements);
}

template <typename Container>
inline typename std::enable_if<!has_resize<Container>::value, void>::type resize(
    Container& c, const size_t nElements)
{
    if (nElements != c.size()) throw std::logic_error("Attempt to resize a fixed size container.");
}

template <typename Container, typename T>
inline typename std::enable_if<has_assign<Container>::value, void>::type assign(
    Container& c, const size_t nElements, const T& value)
{
    c.assign(nElements, value);
}

template <typename Container, typename T>
inline typename std::enable_if<!has_assign<Container>::value, void>::type assign(
    Container& c, const size_t nElements, const T& value)
{
    for (size_t i = 0; i < nElements; i++) c[i] = value;
}

struct IndexDist_Sorter
{
    template <typename PairType>
    bool operator()(const PairType& p1, const PairType& p2) const
    {
        return p1.second < p2.second;
    }
};

template <typename IndexType = size_t, typename DistanceType = double>
struct ResultItem
{
    ResultItem() = default;
    ResultItem(const IndexType index, const DistanceType distance) : first(index), second(distance)
    {
    }
 
    IndexType    first;  
    DistanceType second;  
};


template <typename _DistanceType, typename _IndexType = size_t, typename _CountType = size_t>
class KNNResultSet
{
   public:
    using DistanceType = _DistanceType;
    using IndexType    = _IndexType;
    using CountType    = _CountType;
 
   private:
    IndexType*    indices;
    DistanceType* dists;
    CountType     capacity;
    CountType     count;
 
   public:
    explicit KNNResultSet(CountType capacity_)
        : indices(nullptr), dists(nullptr), capacity(capacity_), count(0)
    {
    }
 
    void init(IndexType* indices_, DistanceType* dists_)
    {
        indices = indices_;
        dists   = dists_;
        count   = 0;
    }
 
    CountType size() const noexcept { return count; }
    bool      empty() const noexcept { return count == 0; }
    bool      full() const noexcept { return count == capacity; }

    bool addPoint(DistanceType dist, IndexType index)
    {
        CountType i;
        for (i = count; i > 0; --i)
        {
#ifdef NANOFLANN_FIRST_MATCH
            if ((dists[i - 1] > dist) || ((dist == dists[i - 1]) && (indices[i - 1] > index)))
            {
#else
            if (dists[i - 1] > dist)
            {
#endif
                if (i < capacity)
                {
                    dists[i]   = dists[i - 1];
                    indices[i] = indices[i - 1];
                }
            }
            else
                break;
        }
        if (i < capacity)
        {
            dists[i]   = dist;
            indices[i] = index;
        }
        if (count < capacity) count++;
 
        // tell caller that the search shall continue
        return true;
    }

    DistanceType worstDist() const noexcept
    {
        return (count < capacity || !count) ? std::numeric_limits<DistanceType>::max()
                                            : dists[count - 1];
    }
 
    void sort()
    {
        // already sorted
    }
};

template <typename _DistanceType, typename _IndexType = size_t, typename _CountType = size_t>
class RKNNResultSet
{
   public:
    using DistanceType = _DistanceType;
    using IndexType    = _IndexType;
    using CountType    = _CountType;
 
   private:
    IndexType*    indices;
    DistanceType* dists;
    CountType     capacity;
    CountType     count;
    DistanceType  maximumSearchDistanceSquared;
 
   public:
    explicit RKNNResultSet(CountType capacity_, DistanceType maximumSearchDistanceSquared_)
        : indices(nullptr),
          dists(nullptr),
          capacity(capacity_),
          count(0),
          maximumSearchDistanceSquared(maximumSearchDistanceSquared_)
    {
    }
 
    void init(IndexType* indices_, DistanceType* dists_)
    {
        indices = indices_;
        dists   = dists_;
        count   = 0;
        if (capacity) dists[capacity - 1] = maximumSearchDistanceSquared;
    }
 
    CountType size() const noexcept { return count; }
    bool      empty() const noexcept { return count == 0; }
    bool      full() const noexcept { return count == capacity; }

    bool addPoint(DistanceType dist, IndexType index)
    {
        CountType i;
        for (i = count; i > 0; --i)
        {
#ifdef NANOFLANN_FIRST_MATCH
            if ((dists[i - 1] > dist) || ((dist == dists[i - 1]) && (indices[i - 1] > index)))
            {
#else
            if (dists[i - 1] > dist)
            {
#endif
                if (i < capacity)
                {
                    dists[i]   = dists[i - 1];
                    indices[i] = indices[i - 1];
                }
            }
            else
                break;
        }
        if (i < capacity)
        {
            dists[i]   = dist;
            indices[i] = index;
        }
        if (count < capacity) count++;
 
        // tell caller that the search shall continue
        return true;
    }

    DistanceType worstDist() const noexcept
    {
        return (count < capacity || !count) ? maximumSearchDistanceSquared : dists[count - 1];
    }
 
    void sort()
    {
        // already sorted
    }
};

template <typename _DistanceType, typename _IndexType = size_t>
class RadiusResultSet
{
   public:
    using DistanceType = _DistanceType;
    using IndexType    = _IndexType;
 
   public:
    const DistanceType radius;
 
    std::vector<ResultItem<IndexType, DistanceType>>& m_indices_dists;
 
    explicit RadiusResultSet(
        DistanceType radius_, std::vector<ResultItem<IndexType, DistanceType>>& indices_dists)
        : radius(radius_), m_indices_dists(indices_dists)
    {
        init();
    }
 
    void init() { clear(); }
    void clear() { m_indices_dists.clear(); }
 
    size_t size() const noexcept { return m_indices_dists.size(); }
    bool   empty() const noexcept { return m_indices_dists.empty(); }
    bool   full() const noexcept { return true; }

    bool addPoint(DistanceType dist, IndexType index)
    {
        if (dist < radius) m_indices_dists.emplace_back(index, dist);
        return true;
    }
 
    DistanceType worstDist() const noexcept { return radius; }

    ResultItem<IndexType, DistanceType> worst_item() const
    {
        if (m_indices_dists.empty())
            throw std::runtime_error(
                "Cannot invoke RadiusResultSet::worst_item() on "
                "an empty list of results.");
        auto it =
            std::max_element(m_indices_dists.begin(), m_indices_dists.end(), IndexDist_Sorter());
        return *it;
    }
 
    void sort() { std::sort(m_indices_dists.begin(), m_indices_dists.end(), IndexDist_Sorter()); }
};


template <typename T>
void save_value(std::ostream& stream, const T& value)
{
    stream.write(reinterpret_cast<const char*>(&value), sizeof(T));
}
 
template <typename T>
void save_value(std::ostream& stream, const std::vector<T>& value)
{
    size_t size = value.size();
    stream.write(reinterpret_cast<const char*>(&size), sizeof(size_t));
    stream.write(reinterpret_cast<const char*>(value.data()), sizeof(T) * size);
}
 
template <typename T>
void load_value(std::istream& stream, T& value)
{
    stream.read(reinterpret_cast<char*>(&value), sizeof(T));
}
 
template <typename T>
void load_value(std::istream& stream, std::vector<T>& value)
{
    size_t size;
    stream.read(reinterpret_cast<char*>(&size), sizeof(size_t));
    value.resize(size);
    stream.read(reinterpret_cast<char*>(value.data()), sizeof(T) * size);
}

 
struct Metric
{
};

template <class T, class DataSource, typename _DistanceType = T, typename IndexType = size_t>
struct L1_Adaptor
{
    using ElementType  = T;
    using DistanceType = _DistanceType;
 
    const DataSource& data_source;
 
    L1_Adaptor(const DataSource& _data_source) : data_source(_data_source) {}
 
    inline DistanceType evalMetric(
        const T* NANOFLANN_RESTRICT a, const IndexType b_idx, size_t size,
        DistanceType worst_dist = -1) const
    {
        DistanceType result  = DistanceType();
        const size_t multof4 = (size >> 2) << 2;  // largest multiple of 4, i.e. 1 << 2
        size_t       d;
 
        /* Process 4 items with each loop for efficiency. */
        if (worst_dist <= 0)
        {
            /* No checks with worst_dist. */
            for (d = 0; d < multof4; d += 4)
            {
                const DistanceType diff0 =
                    std::abs(a[d + 0] - data_source.kdtree_get_pt(b_idx, d + 0));
                const DistanceType diff1 =
                    std::abs(a[d + 1] - data_source.kdtree_get_pt(b_idx, d + 1));
                const DistanceType diff2 =
                    std::abs(a[d + 2] - data_source.kdtree_get_pt(b_idx, d + 2));
                const DistanceType diff3 =
                    std::abs(a[d + 3] - data_source.kdtree_get_pt(b_idx, d + 3));
                /* Parentheses break dependency chain: */
                result += (diff0 + diff1) + (diff2 + diff3);
            }
        }
        else
        {
            /* Check with worst_dist. */
            for (d = 0; d < multof4; d += 4)
            {
                const DistanceType diff0 =
                    std::abs(a[d + 0] - data_source.kdtree_get_pt(b_idx, d + 0));
                const DistanceType diff1 =
                    std::abs(a[d + 1] - data_source.kdtree_get_pt(b_idx, d + 1));
                const DistanceType diff2 =
                    std::abs(a[d + 2] - data_source.kdtree_get_pt(b_idx, d + 2));
                const DistanceType diff3 =
                    std::abs(a[d + 3] - data_source.kdtree_get_pt(b_idx, d + 3));
                /* Parentheses break dependency chain: */
                result += (diff0 + diff1) + (diff2 + diff3);
                if (result > worst_dist)
                {
                    return result;
                }
            }
        }
        /* Process last 0-3 components. Unrolled loop with fall-through switch.
         */
        switch (size - multof4)
        {
            case 3:
                result += std::abs(a[d + 2] - data_source.kdtree_get_pt(b_idx, d + 2));
            case 2:
                result += std::abs(a[d + 1] - data_source.kdtree_get_pt(b_idx, d + 1));
            case 1:
                result += std::abs(a[d + 0] - data_source.kdtree_get_pt(b_idx, d + 0));
            case 0:
                break;
        }
        return result;
    }
 
    template <typename U, typename V>
    inline DistanceType accum_dist(const U a, const V b, const size_t) const
    {
        return std::abs(a - b);
    }
};

template <class T, class DataSource, typename _DistanceType = T, typename IndexType = size_t>
struct L2_Adaptor
{
    using ElementType  = T;
    using DistanceType = _DistanceType;
 
    const DataSource& data_source;
 
    L2_Adaptor(const DataSource& _data_source) : data_source(_data_source) {}
 
    inline DistanceType evalMetric(
        const T* NANOFLANN_RESTRICT a, const IndexType b_idx, size_t size,
        DistanceType worst_dist = -1) const
    {
        DistanceType result  = DistanceType();
        const size_t multof4 = (size >> 2) << 2;  // largest multiple of 4, i.e. 1 << 2
        size_t       d;
 
        /* Process 4 items with each loop for efficiency. */
        if (worst_dist <= 0)
        {
            /* No checks with worst_dist. */
            for (d = 0; d < multof4; d += 4)
            {
                const DistanceType diff0 = a[d + 0] - data_source.kdtree_get_pt(b_idx, d + 0);
                const DistanceType diff1 = a[d + 1] - data_source.kdtree_get_pt(b_idx, d + 1);
                const DistanceType diff2 = a[d + 2] - data_source.kdtree_get_pt(b_idx, d + 2);
                const DistanceType diff3 = a[d + 3] - data_source.kdtree_get_pt(b_idx, d + 3);
                /* Parentheses break dependency chain: */
                result += (diff0 * diff0 + diff1 * diff1) + (diff2 * diff2 + diff3 * diff3);
            }
        }
        else
        {
            /* Check with worst_dist. */
            for (d = 0; d < multof4; d += 4)
            {
                const DistanceType diff0 = a[d + 0] - data_source.kdtree_get_pt(b_idx, d + 0);
                const DistanceType diff1 = a[d + 1] - data_source.kdtree_get_pt(b_idx, d + 1);
                const DistanceType diff2 = a[d + 2] - data_source.kdtree_get_pt(b_idx, d + 2);
                const DistanceType diff3 = a[d + 3] - data_source.kdtree_get_pt(b_idx, d + 3);
                /* Parentheses break dependency chain: */
                result += (diff0 * diff0 + diff1 * diff1) + (diff2 * diff2 + diff3 * diff3);
                if (result > worst_dist)
                {
                    return result;
                }
            }
        }
        /* Process last 0-3 components. Unrolled loop with fall-through switch.
         */
        DistanceType diff;
        switch (size - multof4)
        {
            case 3:
                diff = a[d + 2] - data_source.kdtree_get_pt(b_idx, d + 2);
                result += diff * diff;
            case 2:
                diff = a[d + 1] - data_source.kdtree_get_pt(b_idx, d + 1);
                result += diff * diff;
            case 1:
                diff = a[d + 0] - data_source.kdtree_get_pt(b_idx, d + 0);
                result += diff * diff;
            case 0:
                break;
        }
        return result;
    }
 
    template <typename U, typename V>
    inline DistanceType accum_dist(const U a, const V b, const size_t) const
    {
        auto diff = a - b;
        return diff * diff;
    }
};

template <class T, class DataSource, typename _DistanceType = T, typename IndexType = size_t>
struct L2_Simple_Adaptor
{
    using ElementType  = T;
    using DistanceType = _DistanceType;
 
    const DataSource& data_source;
 
    L2_Simple_Adaptor(const DataSource& _data_source) : data_source(_data_source) {}
 
    inline DistanceType evalMetric(const T* a, const IndexType b_idx, size_t size) const
    {
        DistanceType result = DistanceType();
        for (size_t i = 0; i < size; ++i)
        {
            const DistanceType diff = a[i] - data_source.kdtree_get_pt(b_idx, i);
            result += diff * diff;
        }
        return result;
    }
 
    template <typename U, typename V>
    inline DistanceType accum_dist(const U a, const V b, const size_t) const
    {
        auto diff = a - b;
        return diff * diff;
    }
};

template <class T, class DataSource, typename _DistanceType = T, typename IndexType = size_t>
struct SO2_Adaptor
{
    using ElementType  = T;
    using DistanceType = _DistanceType;
 
    const DataSource& data_source;
 
    SO2_Adaptor(const DataSource& _data_source) : data_source(_data_source) {}
 
    inline DistanceType evalMetric(const T* a, const IndexType b_idx, size_t size) const
    {
        return accum_dist(a[size - 1], data_source.kdtree_get_pt(b_idx, size - 1), size - 1);
    }

    template <typename U, typename V>
    inline DistanceType accum_dist(const U a, const V b, const size_t) const
    {
        DistanceType result = DistanceType();
        DistanceType PI     = pi_const<DistanceType>();
        result              = b - a;
        if (result > PI)
            result -= 2 * PI;
        else if (result < -PI)
            result += 2 * PI;
        return result;
    }
};

template <class T, class DataSource, typename _DistanceType = T, typename IndexType = size_t>
struct SO3_Adaptor
{
    using ElementType  = T;
    using DistanceType = _DistanceType;
 
    L2_Simple_Adaptor<T, DataSource, DistanceType, IndexType> distance_L2_Simple;
 
    SO3_Adaptor(const DataSource& _data_source) : distance_L2_Simple(_data_source) {}
 
    inline DistanceType evalMetric(const T* a, const IndexType b_idx, size_t size) const
    {
        return distance_L2_Simple.evalMetric(a, b_idx, size);
    }
 
    template <typename U, typename V>
    inline DistanceType accum_dist(const U a, const V b, const size_t idx) const
    {
        return distance_L2_Simple.accum_dist(a, b, idx);
    }
};

struct metric_L1 : public Metric
{
    template <class T, class DataSource, typename IndexType = size_t>
    struct traits
    {
        using distance_t = L1_Adaptor<T, DataSource, T, IndexType>;
    };
};

struct metric_L2 : public Metric
{
    template <class T, class DataSource, typename IndexType = size_t>
    struct traits
    {
        using distance_t = L2_Adaptor<T, DataSource, T, IndexType>;
    };
};

struct metric_L2_Simple : public Metric
{
    template <class T, class DataSource, typename IndexType = size_t>
    struct traits
    {
        using distance_t = L2_Simple_Adaptor<T, DataSource, T, IndexType>;
    };
};

struct metric_SO2 : public Metric
{
    template <class T, class DataSource, typename IndexType = size_t>
    struct traits
    {
        using distance_t = SO2_Adaptor<T, DataSource, T, IndexType>;
    };
};

struct metric_SO3 : public Metric
{
    template <class T, class DataSource, typename IndexType = size_t>
    struct traits
    {
        using distance_t = SO3_Adaptor<T, DataSource, T, IndexType>;
    };
};


 
enum class KDTreeSingleIndexAdaptorFlags
{
    None                  = 0,
    SkipInitialBuildIndex = 1
};
 
inline std::underlying_type<KDTreeSingleIndexAdaptorFlags>::type operator&(
    KDTreeSingleIndexAdaptorFlags lhs, KDTreeSingleIndexAdaptorFlags rhs)
{
    using underlying = typename std::underlying_type<KDTreeSingleIndexAdaptorFlags>::type;
    return static_cast<underlying>(lhs) & static_cast<underlying>(rhs);
}

struct KDTreeSingleIndexAdaptorParams
{
    KDTreeSingleIndexAdaptorParams(
        size_t                        _leaf_max_size  = 10,
        KDTreeSingleIndexAdaptorFlags _flags          = KDTreeSingleIndexAdaptorFlags::None,
        unsigned int                  _n_thread_build = 1)
        : leaf_max_size(_leaf_max_size), flags(_flags), n_thread_build(_n_thread_build)
    {
    }
 
    size_t                        leaf_max_size;
    KDTreeSingleIndexAdaptorFlags flags;
    unsigned int                  n_thread_build;
};

struct SearchParameters
{
    SearchParameters(float eps_ = 0, bool sorted_ = true) : eps(eps_), sorted(sorted_) {}
 
    float eps;  
    bool  sorted;  
};



class PooledAllocator
{
    static constexpr size_t WORDSIZE  = 16;  // WORDSIZE must >= 8
    static constexpr size_t BLOCKSIZE = 8192;
 
    /* We maintain memory alignment to word boundaries by requiring that all
        allocations be in multiples of the machine wordsize.  */
    /* Size of machine word in bytes.  Must be power of 2. */
    /* Minimum number of bytes requested at a time from the system.  Must be
     * multiple of WORDSIZE. */
 
    using Size = size_t;
 
    Size  remaining_ = 0;  
    void* base_      = nullptr;  
    void* loc_       = nullptr;  
 
    void internal_init()
    {
        remaining_   = 0;
        base_        = nullptr;
        usedMemory   = 0;
        wastedMemory = 0;
    }
 
   public:
    Size usedMemory   = 0;
    Size wastedMemory = 0;

    PooledAllocator() { internal_init(); }

    ~PooledAllocator() { free_all(); }

    void free_all()
    {
        while (base_ != nullptr)
        {
            // Get pointer to prev block
            void* prev = *(static_cast<void**>(base_));
            ::free(base_);
            base_ = prev;
        }
        internal_init();
    }

    void* malloc(const size_t req_size)
    {
        /* Round size up to a multiple of wordsize.  The following expression
            only works for WORDSIZE that is a power of 2, by masking last bits
           of incremented size to zero.
         */
        const Size size = (req_size + (WORDSIZE - 1)) & ~(WORDSIZE - 1);
 
        /* Check whether a new block must be allocated.  Note that the first
           word of a block is reserved for a pointer to the previous block.
         */
        if (size > remaining_)
        {
            wastedMemory += remaining_;
 
            /* Allocate new storage. */
            const Size blocksize = size > BLOCKSIZE ? size + WORDSIZE : BLOCKSIZE + WORDSIZE;
 
            // use the standard C malloc to allocate memory
            void* m = ::malloc(blocksize);
            if (!m)
            {
                throw std::bad_alloc();
            }
 
            /* Fill first word of new block with pointer to previous block. */
            static_cast<void**>(m)[0] = base_;
            base_                     = m;
 
            remaining_ = blocksize - WORDSIZE;
            loc_       = static_cast<char*>(m) + WORDSIZE;
        }
        void* rloc = loc_;
        loc_       = static_cast<char*>(loc_) + size;
        remaining_ -= size;
 
        usedMemory += size;
 
        return rloc;
    }

    template <typename T>
    T* allocate(const size_t count = 1)
    {
        T* mem = static_cast<T*>(this->malloc(sizeof(T) * count));
        return mem;
    }
};



template <int32_t DIM, typename T>
struct array_or_vector
{
    using type = std::array<T, DIM>;
};

template <typename T>
struct array_or_vector<-1, T>
{
    using type = std::vector<T>;
};


template <
    class Derived, typename Distance, class DatasetAdaptor, int32_t DIM = -1,
    typename index_t = uint32_t>
class KDTreeBaseClass
{
   public:
    void freeIndex(Derived& obj)
    {
        obj.pool_.free_all();
        obj.root_node_           = nullptr;
        obj.size_at_index_build_ = 0;
    }
 
    using ElementType  = typename Distance::ElementType;
    using DistanceType = typename Distance::DistanceType;
    using IndexType    = index_t;

    std::vector<IndexType> vAcc_;
 
    using Offset    = typename decltype(vAcc_)::size_type;
    using Size      = typename decltype(vAcc_)::size_type;
    using Dimension = int32_t;
 
    /*-------------------------------------------------------------------
     * Internal Data Structures
     *
     * "Node" below can be declared with alignas(N) to improve
     * cache friendliness and SIMD load/store performance.
     *
     * The optimal N depends on the underlying hardware:
     *  + Intel x86-64: 16 for SSE, 32 for AVX/AVX2 and 64 for AVX-512
     *  + NVIDIA Jetson: 16 for ARM + NEON and CUDA float4/
     *  To avoid unnecessary padding, the smallest alignment
     *  compatible with a platform's vector width should be chosen.
     * ------------------------------------------------------------------*/
    struct alignas(NANOFLANN_NODE_ALIGNMENT) Node
    {
        union
        {
            struct leaf
            {
                Offset left, right;  
            } lr;
            struct nonleaf
            {
                Dimension divfeat;  
                DistanceType divlow, divhigh;
            } sub;
        } node_type;

        Node *child1 = nullptr, *child2 = nullptr;
    };
 
    using NodePtr      = Node*;
    using NodeConstPtr = const Node*;
 
    struct Interval
    {
        ElementType low, high;
    };
 
    NodePtr root_node_ = nullptr;
 
    Size leaf_max_size_ = 0;

    Size n_thread_build_ = 1;
    Size size_ = 0;
    Size      size_at_index_build_ = 0;
    Dimension dim_                 = 0;  

    using BoundingBox = typename array_or_vector<DIM, Interval>::type;

    using distance_vector_t = typename array_or_vector<DIM, DistanceType>::type;

    BoundingBox root_bbox_;

    PooledAllocator pool_;

    Size size(const Derived& obj) const { return obj.size_; }

    Size veclen(const Derived& obj) const { return DIM > 0 ? DIM : obj.dim_; }

    ElementType dataset_get(const Derived& obj, IndexType element, Dimension component) const
    {
        return obj.dataset_.kdtree_get_pt(element, component);
    }

    Size usedMemory(const Derived& obj) const
    {
        return obj.pool_.usedMemory + obj.pool_.wastedMemory +
               obj.dataset_.kdtree_get_point_count() *
                   sizeof(IndexType);  // pool memory and vind array memory
    }

    void computeMinMax(
        const Derived& obj, Offset ind, Size count, Dimension element, ElementType& min_elem,
        ElementType& max_elem) const
    {
        min_elem = dataset_get(obj, vAcc_[ind], element);
        max_elem = min_elem;
        for (Offset i = 1; i < count; ++i)
        {
            ElementType val = dataset_get(obj, vAcc_[ind + i], element);
            if (val < min_elem) min_elem = val;
            if (val > max_elem) max_elem = val;
        }
    }

    NodePtr divideTree(Derived& obj, const Offset left, const Offset right, BoundingBox& bbox)
    {
        assert(left < obj.dataset_.kdtree_get_point_count());
 
        NodePtr    node = obj.pool_.template allocate<Node>();  // allocate memory
        const auto dims = (DIM > 0 ? DIM : obj.dim_);
 
        /* If too few exemplars remain, then make this a leaf node. */
        if ((right - left) <= static_cast<Offset>(obj.leaf_max_size_))
        {
            node->child1 = node->child2 = nullptr; /* Mark as leaf node. */
            node->node_type.lr.left     = left;
            node->node_type.lr.right    = right;
 
            // compute bounding-box of leaf points
            for (Dimension i = 0; i < dims; ++i)
            {
                bbox[i].low  = dataset_get(obj, obj.vAcc_[left], i);
                bbox[i].high = dataset_get(obj, obj.vAcc_[left], i);
            }
            for (Offset k = left + 1; k < right; ++k)
            {
                for (Dimension i = 0; i < dims; ++i)
                {
                    const auto val = dataset_get(obj, obj.vAcc_[k], i);
                    if (bbox[i].low > val) bbox[i].low = val;
                    if (bbox[i].high < val) bbox[i].high = val;
                }
            }
        }
        else
        {
            /* Determine the index, dimension and value for split plane */
            Offset       idx;
            Dimension    cutfeat;
            DistanceType cutval;
            middleSplit_(obj, left, right - left, idx, cutfeat, cutval, bbox);
 
            node->node_type.sub.divfeat = cutfeat;
 
            /* Recurse on left */
            BoundingBox left_bbox(bbox);
            left_bbox[cutfeat].high = cutval;
            node->child1            = this->divideTree(obj, left, left + idx, left_bbox);
 
            /* Recurse on right */
            BoundingBox right_bbox(bbox);
            right_bbox[cutfeat].low = cutval;
            node->child2            = this->divideTree(obj, left + idx, right, right_bbox);
 
            node->node_type.sub.divlow  = left_bbox[cutfeat].high;
            node->node_type.sub.divhigh = right_bbox[cutfeat].low;
 
            for (Dimension i = 0; i < dims; ++i)
            {
                bbox[i].low  = std::min(left_bbox[i].low, right_bbox[i].low);
                bbox[i].high = std::max(left_bbox[i].high, right_bbox[i].high);
            }
        }
 
        return node;
    }

    NodePtr divideTreeConcurrent(
        Derived& obj, const Offset left, const Offset right, BoundingBox& bbox,
        std::atomic<unsigned int>& thread_count, std::mutex& mutex)
    {
        std::unique_lock<std::mutex> lock(mutex);
        NodePtr                      node = obj.pool_.template allocate<Node>();  // allocate memory
        lock.unlock();
 
        const auto dims = (DIM > 0 ? DIM : obj.dim_);
 
        /* If too few exemplars remain, then make this a leaf node. */
        if ((right - left) <= static_cast<Offset>(obj.leaf_max_size_))
        {
            node->child1 = node->child2 = nullptr; /* Mark as leaf node. */
            node->node_type.lr.left     = left;
            node->node_type.lr.right    = right;
 
            // compute bounding-box of leaf points
            for (Dimension i = 0; i < dims; ++i)
            {
                bbox[i].low  = dataset_get(obj, obj.vAcc_[left], i);
                bbox[i].high = dataset_get(obj, obj.vAcc_[left], i);
            }
            for (Offset k = left + 1; k < right; ++k)
            {
                for (Dimension i = 0; i < dims; ++i)
                {
                    const auto val = dataset_get(obj, obj.vAcc_[k], i);
                    if (bbox[i].low > val) bbox[i].low = val;
                    if (bbox[i].high < val) bbox[i].high = val;
                }
            }
        }
        else
        {
            /* Determine the index, dimension and value for split plane */
            Offset       idx;
            Dimension    cutfeat;
            DistanceType cutval;
            middleSplit_(obj, left, right - left, idx, cutfeat, cutval, bbox);
 
            node->node_type.sub.divfeat = cutfeat;
 
            std::future<NodePtr> right_future;
 
            /* Recurse on right concurrently, if possible */
 
            BoundingBox right_bbox(bbox);
            right_bbox[cutfeat].low = cutval;
            if (++thread_count < n_thread_build_)
            {
                /* Concurrent thread for right recursion */
 
                right_future = std::async(
                    std::launch::async, &KDTreeBaseClass::divideTreeConcurrent, this, std::ref(obj),
                    left + idx, right, std::ref(right_bbox), std::ref(thread_count),
                    std::ref(mutex));
            }
            else
            {
                --thread_count;
            }
 
            /* Recurse on left in this thread */
 
            BoundingBox left_bbox(bbox);
            left_bbox[cutfeat].high = cutval;
            node->child1 =
                this->divideTreeConcurrent(obj, left, left + idx, left_bbox, thread_count, mutex);
 
            if (right_future.valid())
            {
                /* Block and wait for concurrent right from above */
 
                node->child2 = right_future.get();
                --thread_count;
            }
            else
            {
                /* Otherwise, recurse on right in this thread */
 
                node->child2 = this->divideTreeConcurrent(
                    obj, left + idx, right, right_bbox, thread_count, mutex);
            }
 
            node->node_type.sub.divlow  = left_bbox[cutfeat].high;
            node->node_type.sub.divhigh = right_bbox[cutfeat].low;
 
            for (Dimension i = 0; i < dims; ++i)
            {
                bbox[i].low  = std::min(left_bbox[i].low, right_bbox[i].low);
                bbox[i].high = std::max(left_bbox[i].high, right_bbox[i].high);
            }
        }
 
        return node;
    }
 
    void middleSplit_(
        const Derived& obj, const Offset ind, const Size count, Offset& index, Dimension& cutfeat,
        DistanceType& cutval, const BoundingBox& bbox)
    {
        const auto dims = (DIM > 0 ? DIM : obj.dim_);
        const auto EPS  = static_cast<DistanceType>(0.00001);
 
        // Pre-compute max_span once
        ElementType max_span = bbox[0].high - bbox[0].low;
        for (Dimension i = 1; i < dims; ++i)
        {
            ElementType span = bbox[i].high - bbox[i].low;
            if (span > max_span) max_span = span;
        }
 
        // Single-pass min/max computation for candidate dimensions
        cutfeat                = 0;
        ElementType max_spread = -1;
        ElementType min_elem = 0, max_elem = 0;
 
        // Only check dimensions within (1-EPS) of max_span
        std::vector<Dimension> candidates;
        candidates.reserve(dims);
        for (Dimension i = 0; i < dims; ++i)
        {
            if (bbox[i].high - bbox[i].low >= (1 - EPS) * max_span)
            {
                candidates.push_back(i);
            }
        }
 
        // Vectorized min/max for candidates
        for (Dimension dim : candidates)
        {
            ElementType local_min = dataset_get(obj, vAcc_[ind], dim);
            ElementType local_max = local_min;
 
            // Unrolled loop for better performance
            constexpr size_t UNROLL = 4;
            Offset           k      = 1;
            for (; k + UNROLL <= count; k += UNROLL)
            {
                ElementType v0 = dataset_get(obj, vAcc_[ind + k], dim);
                ElementType v1 = dataset_get(obj, vAcc_[ind + k + 1], dim);
                ElementType v2 = dataset_get(obj, vAcc_[ind + k + 2], dim);
                ElementType v3 = dataset_get(obj, vAcc_[ind + k + 3], dim);
 
                local_min = std::min({local_min, v0, v1, v2, v3});
                local_max = std::max({local_max, v0, v1, v2, v3});
            }
 
            // Handle remainder
            for (; k < count; ++k)
            {
                ElementType val = dataset_get(obj, vAcc_[ind + k], dim);
                local_min       = std::min(local_min, val);
                local_max       = std::max(local_max, val);
            }
 
            ElementType spread = local_max - local_min;
            if (spread > max_spread)
            {
                cutfeat    = dim;
                max_spread = spread;
                min_elem   = local_min;
                max_elem   = local_max;
            }
        }
 
        // Median-of-three for better balance
        DistanceType split_val = (bbox[cutfeat].low + bbox[cutfeat].high) / 2;
        if (split_val < min_elem) split_val = min_elem;
        if (split_val > max_elem) split_val = max_elem;
 
        cutval = split_val;
 
        // Optimized partitioning
        Offset lim1, lim2;
        planeSplit(obj, ind, count, cutfeat, cutval, lim1, lim2);
 
        index = (lim1 > count / 2) ? lim1 : (lim2 < count / 2) ? lim2 : count / 2;
    }

    void planeSplit(
        const Derived& obj, const Offset ind, const Size count, const Dimension cutfeat,
        const DistanceType& cutval, Offset& lim1, Offset& lim2)
    {
        // Dutch National Flag algorithm for three-way partitioning
        Offset left  = 0;
        Offset mid   = 0;
        Offset right = count - 1;
 
        while (mid <= right)
        {
            ElementType val = dataset_get(obj, vAcc_[ind + mid], cutfeat);
 
            if (val < cutval)
            {
                std::swap(vAcc_[ind + left], vAcc_[ind + mid]);
                left++;
                mid++;
            }
            else if (val > cutval)
            {
                std::swap(vAcc_[ind + mid], vAcc_[ind + right]);
                right--;
            }
            else
            {
                mid++;
            }
        }
 
        lim1 = left;
        lim2 = mid;
    }
 
    DistanceType computeInitialDistances(
        const Derived& obj, const ElementType* vec, distance_vector_t& dists) const
    {
        assert(vec);
        DistanceType dist = DistanceType();
 
        for (Dimension i = 0; i < (DIM > 0 ? DIM : obj.dim_); ++i)
        {
            if (vec[i] < obj.root_bbox_[i].low)
            {
                dists[i] = obj.distance_.accum_dist(vec[i], obj.root_bbox_[i].low, i);
                dist += dists[i];
            }
            if (vec[i] > obj.root_bbox_[i].high)
            {
                dists[i] = obj.distance_.accum_dist(vec[i], obj.root_bbox_[i].high, i);
                dist += dists[i];
            }
        }
        return dist;
    }
 
    static void save_tree(const Derived& obj, std::ostream& stream, const NodeConstPtr tree)
    {
        save_value(stream, *tree);
        if (tree->child1 != nullptr)
        {
            save_tree(obj, stream, tree->child1);
        }
        if (tree->child2 != nullptr)
        {
            save_tree(obj, stream, tree->child2);
        }
    }
 
    static void load_tree(Derived& obj, std::istream& stream, NodePtr& tree)
    {
        tree = obj.pool_.template allocate<Node>();
        load_value(stream, *tree);
        if (tree->child1 != nullptr)
        {
            load_tree(obj, stream, tree->child1);
        }
        if (tree->child2 != nullptr)
        {
            load_tree(obj, stream, tree->child2);
        }
    }

    void saveIndex(const Derived& obj, std::ostream& stream) const
    {
        save_value(stream, obj.size_);
        save_value(stream, obj.dim_);
        save_value(stream, obj.root_bbox_);
        save_value(stream, obj.leaf_max_size_);
        save_value(stream, obj.vAcc_);
        if (obj.root_node_) save_tree(obj, stream, obj.root_node_);
    }

    void loadIndex(Derived& obj, std::istream& stream)
    {
        load_value(stream, obj.size_);
        load_value(stream, obj.dim_);
        load_value(stream, obj.root_bbox_);
        load_value(stream, obj.leaf_max_size_);
        load_value(stream, obj.vAcc_);
        load_tree(obj, stream, obj.root_node_);
    }
};


template <typename Distance, class DatasetAdaptor, int32_t DIM = -1, typename index_t = uint32_t>
class KDTreeSingleIndexAdaptor
    : public KDTreeBaseClass<
          KDTreeSingleIndexAdaptor<Distance, DatasetAdaptor, DIM, index_t>, Distance,
          DatasetAdaptor, DIM, index_t>
{
   public:
    explicit KDTreeSingleIndexAdaptor(
        const KDTreeSingleIndexAdaptor<Distance, DatasetAdaptor, DIM, index_t>&) = delete;

    const DatasetAdaptor& dataset_;
 
    const KDTreeSingleIndexAdaptorParams indexParams;
 
    Distance distance_;
 
    using Base = typename nanoflann::KDTreeBaseClass<
        nanoflann::KDTreeSingleIndexAdaptor<Distance, DatasetAdaptor, DIM, index_t>, Distance,
        DatasetAdaptor, DIM, index_t>;
 
    using Offset    = typename Base::Offset;
    using Size      = typename Base::Size;
    using Dimension = typename Base::Dimension;
 
    using ElementType  = typename Base::ElementType;
    using DistanceType = typename Base::DistanceType;
    using IndexType    = typename Base::IndexType;
 
    using Node    = typename Base::Node;
    using NodePtr = Node*;
 
    using Interval = typename Base::Interval;

    using BoundingBox = typename Base::BoundingBox;

    using distance_vector_t = typename Base::distance_vector_t;

    template <class... Args>
    explicit KDTreeSingleIndexAdaptor(
        const Dimension dimensionality, const DatasetAdaptor& inputData,
        const KDTreeSingleIndexAdaptorParams& params, Args&&... args)
        : dataset_(inputData),
          indexParams(params),
          distance_(inputData, std::forward<Args>(args)...)
    {
        init(dimensionality, params);
    }
 
    explicit KDTreeSingleIndexAdaptor(
        const Dimension dimensionality, const DatasetAdaptor& inputData,
        const KDTreeSingleIndexAdaptorParams& params = {})
        : dataset_(inputData), indexParams(params), distance_(inputData)
    {
        init(dimensionality, params);
    }
 
   private:
    void init(const Dimension dimensionality, const KDTreeSingleIndexAdaptorParams& params)
    {
        Base::size_                = dataset_.kdtree_get_point_count();
        Base::size_at_index_build_ = Base::size_;
        Base::dim_                 = dimensionality;
        if (DIM > 0) Base::dim_ = DIM;
        Base::leaf_max_size_ = params.leaf_max_size;
        if (params.n_thread_build > 0)
        {
            Base::n_thread_build_ = params.n_thread_build;
        }
        else
        {
            Base::n_thread_build_ = std::max(std::thread::hardware_concurrency(), 1u);
        }
 
        if (!(params.flags & KDTreeSingleIndexAdaptorFlags::SkipInitialBuildIndex))
        {
            // Build KD-tree:
            buildIndex();
        }
    }
 
   public:
    void buildIndex()
    {
        Base::size_                = dataset_.kdtree_get_point_count();
        Base::size_at_index_build_ = Base::size_;
        init_vind();
        this->freeIndex(*this);
        Base::size_at_index_build_ = Base::size_;
        if (Base::size_ == 0) return;
        computeBoundingBox(Base::root_bbox_);
        // construct the tree
        if (Base::n_thread_build_ == 1)
        {
            Base::root_node_ = this->divideTree(*this, 0, Base::size_, Base::root_bbox_);
        }
        else
        {
#ifndef NANOFLANN_NO_THREADS
            std::atomic<unsigned int> thread_count(0u);
            std::mutex                mutex;
            Base::root_node_ = this->divideTreeConcurrent(
                *this, 0, Base::size_, Base::root_bbox_, thread_count, mutex);
#else /* NANOFLANN_NO_THREADS */
            throw std::runtime_error("Multithreading is disabled");
#endif /* NANOFLANN_NO_THREADS */
        }
    }


    template <typename RESULTSET>
    bool findNeighbors(
        RESULTSET& result, const ElementType* vec, const SearchParameters& searchParams = {}) const
    {
        assert(vec);
        if (this->size(*this) == 0) return false;
        if (!Base::root_node_)
            throw std::runtime_error(
                "[nanoflann] findNeighbors() called before building the "
                "index.");
        float epsError = 1 + searchParams.eps;
 
        // fixed or variable-sized container (depending on DIM)
        distance_vector_t dists;
        // Fill it with zeros.
        auto zero = static_cast<typename RESULTSET::DistanceType>(0);
        assign(dists, (DIM > 0 ? DIM : Base::dim_), zero);
        DistanceType dist = this->computeInitialDistances(*this, vec, dists);
        searchLevel(result, vec, Base::root_node_, dist, dists, epsError);
 
        if (searchParams.sorted) result.sort();
 
        return result.full();
    }

    template <typename RESULTSET>
    Size findWithinBox(RESULTSET& result, const BoundingBox& bbox) const
    {
        if (this->size(*this) == 0) return 0;
        if (!Base::root_node_)
            throw std::runtime_error(
                "[nanoflann] findWithinBox() called before building the "
                "index.");
 
        std::stack<NodePtr> stack;
        stack.push(Base::root_node_);
 
        while (!stack.empty())
        {
            const NodePtr node = stack.top();
            stack.pop();
 
            // If this is a leaf node, then do check and return.
            if (!node->child1)  // (if one node is nullptr, both are)
            {
                for (Offset i = node->node_type.lr.left; i < node->node_type.lr.right; ++i)
                {
                    if (contains(bbox, Base::vAcc_[i]))
                    {
                        if (!result.addPoint(0, Base::vAcc_[i]))
                        {
                            // the resultset doesn't want to receive any more
                            // points, we're done searching!
                            return result.size();
                        }
                    }
                }
            }
            else
            {
                const int  idx        = node->node_type.sub.divfeat;
                const auto low_bound  = node->node_type.sub.divlow;
                const auto high_bound = node->node_type.sub.divhigh;
 
                if (bbox[idx].low <= low_bound) stack.push(node->child1);
                if (bbox[idx].high >= high_bound) stack.push(node->child2);
            }
        }
 
        return result.size();
    }

    Size knnSearch(
        const ElementType* query_point, const Size num_closest, IndexType* out_indices,
        DistanceType* out_distances) const
    {
        nanoflann::KNNResultSet<DistanceType, IndexType> resultSet(num_closest);
        resultSet.init(out_indices, out_distances);
        findNeighbors(resultSet, query_point);
        return resultSet.size();
    }

    Size radiusSearch(
        const ElementType* query_point, const DistanceType& radius,
        std::vector<ResultItem<IndexType, DistanceType>>& IndicesDists,
        const SearchParameters&                           searchParams = {}) const
    {
        RadiusResultSet<DistanceType, IndexType> resultSet(radius, IndicesDists);
        const Size nFound = radiusSearchCustomCallback(query_point, resultSet, searchParams);
        return nFound;
    }

    template <class SEARCH_CALLBACK>
    Size radiusSearchCustomCallback(
        const ElementType* query_point, SEARCH_CALLBACK& resultSet,
        const SearchParameters& searchParams = {}) const
    {
        findNeighbors(resultSet, query_point, searchParams);
        return resultSet.size();
    }

    Size rknnSearch(
        const ElementType* query_point, const Size num_closest, IndexType* out_indices,
        DistanceType* out_distances, const DistanceType& radius) const
    {
        nanoflann::RKNNResultSet<DistanceType, IndexType> resultSet(num_closest, radius);
        resultSet.init(out_indices, out_distances);
        findNeighbors(resultSet, query_point);
        return resultSet.size();
    }

 
   public:
    void init_vind()
    {
        // Create a permutable array of indices to the input vectors.
        Base::size_ = dataset_.kdtree_get_point_count();
        if (Base::vAcc_.size() != Base::size_) Base::vAcc_.resize(Base::size_);
        for (IndexType i = 0; i < static_cast<IndexType>(Base::size_); i++) Base::vAcc_[i] = i;
    }
 
    void computeBoundingBox(BoundingBox& bbox)
    {
        const auto dims = (DIM > 0 ? DIM : Base::dim_);
        resize(bbox, dims);
        if (dataset_.kdtree_get_bbox(bbox))
        {
            // Done! It was implemented in derived class
        }
        else
        {
            const Size N = dataset_.kdtree_get_point_count();
            if (!N)
                throw std::runtime_error(
                    "[nanoflann] computeBoundingBox() called but "
                    "no data points found.");
            for (Dimension i = 0; i < dims; ++i)
            {
                bbox[i].low = bbox[i].high = this->dataset_get(*this, Base::vAcc_[0], i);
            }
            for (Offset k = 1; k < N; ++k)
            {
                for (Dimension i = 0; i < dims; ++i)
                {
                    const auto val = this->dataset_get(*this, Base::vAcc_[k], i);
                    if (val < bbox[i].low) bbox[i].low = val;
                    if (val > bbox[i].high) bbox[i].high = val;
                }
            }
        }
    }
 
    bool contains(const BoundingBox& bbox, IndexType idx) const
    {
        const auto dims = (DIM > 0 ? DIM : Base::dim_);
        for (Dimension i = 0; i < dims; ++i)
        {
            const auto point = this->dataset_.kdtree_get_pt(idx, i);
            if (point < bbox[i].low || point > bbox[i].high) return false;
        }
        return true;
    }

    template <class RESULTSET>
    bool searchLevel(
        RESULTSET& result_set, const ElementType* vec, const NodePtr node, DistanceType mindist,
        distance_vector_t& dists, const float epsError) const
    {
        // If this is a leaf node, then do check and return.
        if (!node->child1)  // (if one node is nullptr, both are)
        {
            for (Offset i = node->node_type.lr.left; i < node->node_type.lr.right; ++i)
            {
                const IndexType accessor = Base::vAcc_[i];  // reorder... : i;
                DistanceType    dist =
                    distance_.evalMetric(vec, accessor, (DIM > 0 ? DIM : Base::dim_));
                if (dist < result_set.worstDist())
                {
                    if (!result_set.addPoint(dist, Base::vAcc_[i]))
                    {
                        // the resultset doesn't want to receive any more
                        // points, we're done searching!
                        return false;
                    }
                }
            }
            return true;
        }
 
        /* Which child branch should be taken first? */
        Dimension    idx   = node->node_type.sub.divfeat;
        ElementType  val   = vec[idx];
        DistanceType diff1 = val - node->node_type.sub.divlow;
        DistanceType diff2 = val - node->node_type.sub.divhigh;
 
        NodePtr      bestChild;
        NodePtr      otherChild;
        DistanceType cut_dist;
        if ((diff1 + diff2) < 0)
        {
            bestChild  = node->child1;
            otherChild = node->child2;
            cut_dist   = distance_.accum_dist(val, node->node_type.sub.divhigh, idx);
        }
        else
        {
            bestChild  = node->child2;
            otherChild = node->child1;
            cut_dist   = distance_.accum_dist(val, node->node_type.sub.divlow, idx);
        }
 
        /* Call recursively to search next level down. */
        if (!searchLevel(result_set, vec, bestChild, mindist, dists, epsError))
        {
            // the resultset doesn't want to receive any more points, we're done
            // searching!
            return false;
        }
 
        DistanceType dst = dists[idx];
        mindist          = mindist + cut_dist - dst;
        dists[idx]       = cut_dist;
        if (mindist * epsError <= result_set.worstDist())
        {
            if (!searchLevel(result_set, vec, otherChild, mindist, dists, epsError))
            {
                // the resultset doesn't want to receive any more points, we're
                // done searching!
                return false;
            }
        }
        dists[idx] = dst;
        return true;
    }
 
   public:
    void saveIndex(std::ostream& stream) const { Base::saveIndex(*this, stream); }

    void loadIndex(std::istream& stream) { Base::loadIndex(*this, stream); }
 
};  // class KDTree

template <typename Distance, class DatasetAdaptor, int32_t DIM = -1, typename IndexType = uint32_t>
class KDTreeSingleIndexDynamicAdaptor_
    : public KDTreeBaseClass<
          KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM, IndexType>, Distance,
          DatasetAdaptor, DIM, IndexType>
{
   public:
    const DatasetAdaptor& dataset_;  
 
    KDTreeSingleIndexAdaptorParams index_params_;
 
    std::vector<int>& treeIndex_;
 
    Distance distance_;
 
    using Base = typename nanoflann::KDTreeBaseClass<
        nanoflann::KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM, IndexType>,
        Distance, DatasetAdaptor, DIM, IndexType>;
 
    using ElementType  = typename Base::ElementType;
    using DistanceType = typename Base::DistanceType;
 
    using Offset    = typename Base::Offset;
    using Size      = typename Base::Size;
    using Dimension = typename Base::Dimension;
 
    using Node    = typename Base::Node;
    using NodePtr = Node*;
 
    using Interval = typename Base::Interval;
    using BoundingBox = typename Base::BoundingBox;

    using distance_vector_t = typename Base::distance_vector_t;

    KDTreeSingleIndexDynamicAdaptor_(
        const Dimension dimensionality, const DatasetAdaptor& inputData,
        std::vector<int>&                     treeIndex,
        const KDTreeSingleIndexAdaptorParams& params = KDTreeSingleIndexAdaptorParams())
        : dataset_(inputData), index_params_(params), treeIndex_(treeIndex), distance_(inputData)
    {
        Base::size_                = 0;
        Base::size_at_index_build_ = 0;
        for (auto& v : Base::root_bbox_) v = {};
        Base::dim_ = dimensionality;
        if (DIM > 0) Base::dim_ = DIM;
        Base::leaf_max_size_ = params.leaf_max_size;
        if (params.n_thread_build > 0)
        {
            Base::n_thread_build_ = params.n_thread_build;
        }
        else
        {
            Base::n_thread_build_ = std::max(std::thread::hardware_concurrency(), 1u);
        }
    }

    KDTreeSingleIndexDynamicAdaptor_(const KDTreeSingleIndexDynamicAdaptor_& rhs) = default;

    KDTreeSingleIndexDynamicAdaptor_ operator=(const KDTreeSingleIndexDynamicAdaptor_& rhs)
    {
        KDTreeSingleIndexDynamicAdaptor_ tmp(rhs);
        std::swap(Base::vAcc_, tmp.Base::vAcc_);
        std::swap(Base::leaf_max_size_, tmp.Base::leaf_max_size_);
        std::swap(index_params_, tmp.index_params_);
        std::swap(treeIndex_, tmp.treeIndex_);
        std::swap(Base::size_, tmp.Base::size_);
        std::swap(Base::size_at_index_build_, tmp.Base::size_at_index_build_);
        std::swap(Base::root_node_, tmp.Base::root_node_);
        std::swap(Base::root_bbox_, tmp.Base::root_bbox_);
        std::swap(Base::pool_, tmp.Base::pool_);
        return *this;
    }

    void buildIndex()
    {
        Base::size_ = Base::vAcc_.size();
        this->freeIndex(*this);
        Base::size_at_index_build_ = Base::size_;
        if (Base::size_ == 0) return;
        computeBoundingBox(Base::root_bbox_);
        // construct the tree
        if (Base::n_thread_build_ == 1)
        {
            Base::root_node_ = this->divideTree(*this, 0, Base::size_, Base::root_bbox_);
        }
        else
        {
#ifndef NANOFLANN_NO_THREADS
            std::atomic<unsigned int> thread_count(0u);
            std::mutex                mutex;
            Base::root_node_ = this->divideTreeConcurrent(
                *this, 0, Base::size_, Base::root_bbox_, thread_count, mutex);
#else /* NANOFLANN_NO_THREADS */
            throw std::runtime_error("Multithreading is disabled");
#endif /* NANOFLANN_NO_THREADS */
        }
    }


    template <typename RESULTSET>
    bool findNeighbors(
        RESULTSET& result, const ElementType* vec, const SearchParameters& searchParams = {}) const
    {
        assert(vec);
        if (this->size(*this) == 0) return false;
        if (!Base::root_node_) return false;
        float epsError = 1 + searchParams.eps;
 
        // fixed or variable-sized container (depending on DIM)
        distance_vector_t dists;
        // Fill it with zeros.
        assign(
            dists, (DIM > 0 ? DIM : Base::dim_),
            static_cast<typename distance_vector_t::value_type>(0));
        DistanceType dist = this->computeInitialDistances(*this, vec, dists);
        searchLevel(result, vec, Base::root_node_, dist, dists, epsError);
 
        if (searchParams.sorted) result.sort();
 
        return result.full();
    }

    Size knnSearch(
        const ElementType* query_point, const Size num_closest, IndexType* out_indices,
        DistanceType* out_distances, const SearchParameters& searchParams = {}) const
    {
        nanoflann::KNNResultSet<DistanceType, IndexType> resultSet(num_closest);
        resultSet.init(out_indices, out_distances);
        findNeighbors(resultSet, query_point, searchParams);
        return resultSet.size();
    }

    Size radiusSearch(
        const ElementType* query_point, const DistanceType& radius,
        std::vector<ResultItem<IndexType, DistanceType>>& IndicesDists,
        const SearchParameters&                           searchParams = {}) const
    {
        RadiusResultSet<DistanceType, IndexType> resultSet(radius, IndicesDists);
        const size_t nFound = radiusSearchCustomCallback(query_point, resultSet, searchParams);
        return nFound;
    }

    template <class SEARCH_CALLBACK>
    Size radiusSearchCustomCallback(
        const ElementType* query_point, SEARCH_CALLBACK& resultSet,
        const SearchParameters& searchParams = {}) const
    {
        findNeighbors(resultSet, query_point, searchParams);
        return resultSet.size();
    }

 
   public:
    void computeBoundingBox(BoundingBox& bbox)
    {
        const auto dims = (DIM > 0 ? DIM : Base::dim_);
        resize(bbox, dims);
 
        if (dataset_.kdtree_get_bbox(bbox))
        {
            // Done! It was implemented in derived class
        }
        else
        {
            const Size N = Base::size_;
            if (!N)
                throw std::runtime_error(
                    "[nanoflann] computeBoundingBox() called but "
                    "no data points found.");
            for (Dimension i = 0; i < dims; ++i)
            {
                bbox[i].low = bbox[i].high = this->dataset_get(*this, Base::vAcc_[0], i);
            }
            for (Offset k = 1; k < N; ++k)
            {
                for (Dimension i = 0; i < dims; ++i)
                {
                    const auto val = this->dataset_get(*this, Base::vAcc_[k], i);
                    if (val < bbox[i].low) bbox[i].low = val;
                    if (val > bbox[i].high) bbox[i].high = val;
                }
            }
        }
    }

    template <class RESULTSET>
    void searchLevel(
        RESULTSET& result_set, const ElementType* vec, const NodePtr node, DistanceType mindist,
        distance_vector_t& dists, const float epsError) const
    {
        // If this is a leaf node, then do check and return.
        if (!node->child1)  // (if one node is nullptr, both are)
        {
            for (Offset i = node->node_type.lr.left; i < node->node_type.lr.right; ++i)
            {
                const IndexType index = Base::vAcc_[i];  // reorder... : i;
                if (treeIndex_[index] == -1) continue;
                DistanceType dist = distance_.evalMetric(vec, index, (DIM > 0 ? DIM : Base::dim_));
                if (dist < result_set.worstDist())
                {
                    if (!result_set.addPoint(
                            static_cast<typename RESULTSET::DistanceType>(dist),
                            static_cast<typename RESULTSET::IndexType>(Base::vAcc_[i])))
                    {
                        // the resultset doesn't want to receive any more
                        // points, we're done searching!
                        return;  // false;
                    }
                }
            }
            return;
        }
 
        /* Which child branch should be taken first? */
        Dimension    idx   = node->node_type.sub.divfeat;
        ElementType  val   = vec[idx];
        DistanceType diff1 = val - node->node_type.sub.divlow;
        DistanceType diff2 = val - node->node_type.sub.divhigh;
 
        NodePtr      bestChild;
        NodePtr      otherChild;
        DistanceType cut_dist;
        if ((diff1 + diff2) < 0)
        {
            bestChild  = node->child1;
            otherChild = node->child2;
            cut_dist   = distance_.accum_dist(val, node->node_type.sub.divhigh, idx);
        }
        else
        {
            bestChild  = node->child2;
            otherChild = node->child1;
            cut_dist   = distance_.accum_dist(val, node->node_type.sub.divlow, idx);
        }
 
        /* Call recursively to search next level down. */
        searchLevel(result_set, vec, bestChild, mindist, dists, epsError);
 
        DistanceType dst = dists[idx];
        mindist          = mindist + cut_dist - dst;
        dists[idx]       = cut_dist;
        if (mindist * epsError <= result_set.worstDist())
        {
            searchLevel(result_set, vec, otherChild, mindist, dists, epsError);
        }
        dists[idx] = dst;
    }
 
   public:
    void saveIndex(std::ostream& stream) { saveIndex(*this, stream); }

    void loadIndex(std::istream& stream) { loadIndex(*this, stream); }
};

template <typename Distance, class DatasetAdaptor, int32_t DIM = -1, typename IndexType = uint32_t>
class KDTreeSingleIndexDynamicAdaptor
{
   public:
    using ElementType  = typename Distance::ElementType;
    using DistanceType = typename Distance::DistanceType;
 
    using Offset = typename KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM>::Offset;
    using Size   = typename KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM>::Size;
    using Dimension =
        typename KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM>::Dimension;
 
   protected:
    Size leaf_max_size_;
    Size treeCount_;
    Size pointCount_;

    const DatasetAdaptor& dataset_;  

    std::vector<int>        treeIndex_;
    std::unordered_set<int> removedPoints_;
 
    KDTreeSingleIndexAdaptorParams index_params_;
 
    Dimension dim_;  
 
    using index_container_t =
        KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM, IndexType>;
    std::vector<index_container_t> index_;
 
   public:
    const std::vector<index_container_t>& getAllIndices() const { return index_; }
 
   private:
    int First0Bit(IndexType num)
    {
        int pos = 0;
        while (num & 1)
        {
            num = num >> 1;
            pos++;
        }
        return pos;
    }

    void init()
    {
        using my_kd_tree_t =
            KDTreeSingleIndexDynamicAdaptor_<Distance, DatasetAdaptor, DIM, IndexType>;
        std::vector<my_kd_tree_t> index(
            treeCount_, my_kd_tree_t(dim_ /*dim*/, dataset_, treeIndex_, index_params_));
        index_ = index;
    }
 
   public:
    Distance distance_;

    explicit KDTreeSingleIndexDynamicAdaptor(
        const int dimensionality, const DatasetAdaptor& inputData,
        const KDTreeSingleIndexAdaptorParams& params            = KDTreeSingleIndexAdaptorParams(),
        const size_t                          maximumPointCount = 1000000000U)
        : dataset_(inputData), index_params_(params), distance_(inputData)
    {
        treeCount_  = static_cast<size_t>(std::log2(maximumPointCount)) + 1;
        pointCount_ = 0U;
        dim_        = dimensionality;
        treeIndex_.clear();
        if (DIM > 0) dim_ = DIM;
        leaf_max_size_ = params.leaf_max_size;
        init();
        const size_t num_initial_points = dataset_.kdtree_get_point_count();
        if (num_initial_points > 0)
        {
            addPoints(0, num_initial_points - 1);
        }
    }

    explicit KDTreeSingleIndexDynamicAdaptor(
        const KDTreeSingleIndexDynamicAdaptor<Distance, DatasetAdaptor, DIM, IndexType>&) = delete;

    void addPoints(IndexType start, IndexType end)
    {
        const Size count    = end - start + 1;
        int        maxIndex = 0;
        treeIndex_.resize(treeIndex_.size() + count);
        for (IndexType idx = start; idx <= end; idx++)
        {
            const int pos           = First0Bit(pointCount_);
            maxIndex                = std::max(pos, maxIndex);
            treeIndex_[pointCount_] = pos;
 
            const auto it = removedPoints_.find(idx);
            if (it != removedPoints_.end())
            {
                removedPoints_.erase(it);
                treeIndex_[idx] = pos;
            }
 
            for (int i = 0; i < pos; i++)
            {
                for (int j = 0; j < static_cast<int>(index_[i].vAcc_.size()); j++)
                {
                    index_[pos].vAcc_.push_back(index_[i].vAcc_[j]);
                    if (treeIndex_[index_[i].vAcc_[j]] != -1) treeIndex_[index_[i].vAcc_[j]] = pos;
                }
                index_[i].vAcc_.clear();
            }
            index_[pos].vAcc_.push_back(idx);
            pointCount_++;
        }
 
        for (int i = 0; i <= maxIndex; ++i)
        {
            index_[i].freeIndex(index_[i]);
            if (!index_[i].vAcc_.empty()) index_[i].buildIndex();
        }
    }

    void removePoint(size_t idx)
    {
        if (idx >= pointCount_) return;
        removedPoints_.insert(idx);
        treeIndex_[idx] = -1;
    }

    template <typename RESULTSET>
    bool findNeighbors(
        RESULTSET& result, const ElementType* vec, const SearchParameters& searchParams = {}) const
    {
        for (size_t i = 0; i < treeCount_; i++)
        {
            index_[i].findNeighbors(result, &vec[0], searchParams);
        }
        return result.full();
    }
};

template <
    class MatrixType, int32_t DIM = -1, class Distance = nanoflann::metric_L2,
    bool row_major = true>
struct KDTreeEigenMatrixAdaptor
{
    using self_t    = KDTreeEigenMatrixAdaptor<MatrixType, DIM, Distance, row_major>;
    using num_t     = typename MatrixType::Scalar;
    using IndexType = typename MatrixType::Index;
    using metric_t  = typename Distance::template traits<num_t, self_t, IndexType>::distance_t;
 
    using index_t = KDTreeSingleIndexAdaptor<
        metric_t, self_t, row_major ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime,
        IndexType>;
 
    index_t* index_;  
 
    using Offset    = typename index_t::Offset;
    using Size      = typename index_t::Size;
    using Dimension = typename index_t::Dimension;

    explicit KDTreeEigenMatrixAdaptor(
        const Dimension dimensionality, const std::reference_wrapper<const MatrixType>& mat,
        const int leaf_max_size = 10, const unsigned int n_thread_build = 1)
        : m_data_matrix(mat)
    {
        const auto dims = row_major ? mat.get().cols() : mat.get().rows();
        if (static_cast<Dimension>(dims) != dimensionality)
            throw std::runtime_error(
                "Error: 'dimensionality' must match column count in data "
                "matrix");
        if (DIM > 0 && static_cast<int32_t>(dims) != DIM)
            throw std::runtime_error(
                "Data set dimensionality does not match the 'DIM' template "
                "argument");
        index_ = new index_t(
            dims, *this /* adaptor */,
            nanoflann::KDTreeSingleIndexAdaptorParams(
                leaf_max_size, nanoflann::KDTreeSingleIndexAdaptorFlags::None, n_thread_build));
    }
 
   public:
    KDTreeEigenMatrixAdaptor(const self_t&) = delete;
 
    ~KDTreeEigenMatrixAdaptor() { delete index_; }
 
    const std::reference_wrapper<const MatrixType> m_data_matrix;

    void query(
        const num_t* query_point, const Size num_closest, IndexType* out_indices,
        num_t* out_distances) const
    {
        nanoflann::KNNResultSet<num_t, IndexType> resultSet(num_closest);
        resultSet.init(out_indices, out_distances);
        index_->findNeighbors(resultSet, query_point);
    }

 
    inline const self_t& derived() const noexcept { return *this; }
    inline self_t&       derived() noexcept { return *this; }
 
    // Must return the number of data points
    inline Size kdtree_get_point_count() const
    {
        if (row_major)
            return m_data_matrix.get().rows();
        else
            return m_data_matrix.get().cols();
    }
 
    // Returns the dim'th component of the idx'th point in the class:
    inline num_t kdtree_get_pt(const IndexType idx, size_t dim) const
    {
        if (row_major)
            return m_data_matrix.get().coeff(idx, IndexType(dim));
        else
            return m_data_matrix.get().coeff(IndexType(dim), idx);
    }
 
    // Optional bounding-box computation: return false to default to a standard
    // bbox computation loop.
    //   Return true if the BBOX was already computed by the class and returned
    //   in "bb" so it can be avoided to redo it again. Look at bb.size() to
    //   find out the expected dimensionality (e.g. 2 or 3 for point clouds)
    template <class BBOX>
    inline bool kdtree_get_bbox(BBOX& /*bb*/) const
    {
        return false;
    }

 
};  // end of KDTreeEigenMatrixAdaptor

  // end of grouping
}  // namespace nanoflann
 
#undef NANOFLANN_RESTRICT