oneapi/tbb/partitioner.h

/*
    Copyright (c) 2005-2022 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.
*/

#ifndef __TBB_partitioner_H
#define __TBB_partitioner_H

#ifndef __TBB_INITIAL_CHUNKS
// initial task divisions per thread
#define __TBB_INITIAL_CHUNKS 2
#endif
#ifndef __TBB_RANGE_POOL_CAPACITY
// maximum number of elements in range pool
#define __TBB_RANGE_POOL_CAPACITY 8
#endif
#ifndef __TBB_INIT_DEPTH
// initial value for depth of range pool
#define __TBB_INIT_DEPTH 5
#endif
#ifndef __TBB_DEMAND_DEPTH_ADD
// when imbalance is found range splits this value times more
#define __TBB_DEMAND_DEPTH_ADD 1
#endif

#include "detail/_config.h"
#include "detail/_namespace_injection.h"
#include "detail/_aligned_space.h"
#include "detail/_utils.h"
#include "detail/_template_helpers.h"
#include "detail/_range_common.h"
#include "detail/_task.h"
#include "detail/_small_object_pool.h"

#include "cache_aligned_allocator.h"
#include "task_group.h" // task_group_context
#include "task_arena.h"

#include <algorithm>
#include <atomic>
#include <type_traits>

#if defined(_MSC_VER) && !defined(__INTEL_COMPILER)
    // Workaround for overzealous compiler warnings
    #pragma warning (push)
    #pragma warning (disable: 4244)
#endif

namespace tbb {
namespace detail {

namespace d1 {
class auto_partitioner;
class simple_partitioner;
class static_partitioner;
class affinity_partitioner;
class affinity_partition_type;
class affinity_partitioner_base;

inline std::size_t get_initial_auto_partitioner_divisor() {
    const std::size_t factor = 4;
    return factor * max_concurrency();
}

//! Defines entry point for affinity partitioner into oneTBB run-time library.
class affinity_partitioner_base: no_copy {
    friend class affinity_partitioner;
    friend class affinity_partition_type;
    //! Array that remembers affinities of tree positions to affinity_id.
    /** nullptr if my_size==0. */
    slot_id* my_array;
    //! Number of elements in my_array.
    std::size_t my_size;
    //! Zeros the fields.
    affinity_partitioner_base() : my_array(nullptr), my_size(0) {}
    //! Deallocates my_array.
    ~affinity_partitioner_base() { resize(0); }
    //! Resize my_array.
    /** Retains values if resulting size is the same. */
    void resize(unsigned factor) {
        // Check factor to avoid asking for number of workers while there might be no arena.
        unsigned max_threads_in_arena = max_concurrency();
        std::size_t new_size = factor ? factor * max_threads_in_arena : 0;
        if (new_size != my_size) {
            if (my_array) {
                r1::cache_aligned_deallocate(my_array);
                // Following two assignments must be done here for sake of exception safety.
                my_array = nullptr;
                my_size = 0;
            }
            if (new_size) {
                my_array = static_cast<slot_id*>(r1::cache_aligned_allocate(new_size * sizeof(slot_id)));
                std::fill_n(my_array, new_size, no_slot);
                my_size = new_size;
            }
        }
    }
};

template<typename Range, typename Body, typename Partitioner> struct start_for;
template<typename Range, typename Body, typename Partitioner> struct start_scan;
template<typename Range, typename Body, typename Partitioner> struct start_reduce;
template<typename Range, typename Body, typename Partitioner> struct start_deterministic_reduce;

struct node {
    node* my_parent{};
    std::atomic<int> m_ref_count{};

    node() = default;
    node(node* parent, int ref_count) :
        my_parent{parent}, m_ref_count{ref_count} {
        __TBB_ASSERT(ref_count > 0, "The ref count must be positive");
    }
};

struct wait_node : node {
    wait_node() : node{ nullptr, 1 } {}
    wait_context m_wait{1};
};

//! Join task node that contains shared flag for stealing feedback
struct tree_node : public node {
    small_object_allocator m_allocator;
    std::atomic<bool> m_child_stolen{false};

    tree_node(node* parent, int ref_count, small_object_allocator& alloc)
        : node{parent, ref_count}
        , m_allocator{alloc} {}

    void join(task_group_context*) {/*dummy, required only for reduction algorithms*/};

    template <typename Task>
    static void mark_task_stolen(Task &t) {
        std::atomic<bool> &flag = static_cast<tree_node*>(t.my_parent)->m_child_stolen;
#if TBB_USE_PROFILING_TOOLS
        // Threading tools respect lock prefix but report false-positive data-race via plain store
        flag.exchange(true);
#else
        flag.store(true, std::memory_order_relaxed);
#endif // TBB_USE_PROFILING_TOOLS
    }
    template <typename Task>
    static bool is_peer_stolen(Task &t) {
        return static_cast<tree_node*>(t.my_parent)->m_child_stolen.load(std::memory_order_relaxed);
    }
};

// Context used to check cancellation state during reduction join process
template<typename TreeNodeType>
void fold_tree(node* n, const execution_data& ed) {
    for (;;) {
        __TBB_ASSERT(n->m_ref_count.load(std::memory_order_relaxed) > 0, "The refcount must be positive.");
        call_itt_task_notify(releasing, n);
        if (--n->m_ref_count > 0) {
            return;
        }
        node* parent = n->my_parent;
        if (!parent) {
            break;
        };

        call_itt_task_notify(acquired, n);
        TreeNodeType* self = static_cast<TreeNodeType*>(n);
        self->join(ed.context);
        self->m_allocator.delete_object(self, ed);
        n = parent;
    }
    // Finish parallel for execution when the root (last node) is reached
    static_cast<wait_node*>(n)->m_wait.release();
}

//! Depth is a relative depth of recursive division inside a range pool. Relative depth allows
//! infinite absolute depth of the recursion for heavily unbalanced workloads with range represented
//! by a number that cannot fit into machine word.
typedef unsigned char depth_t;

//! Range pool stores ranges of type T in a circular buffer with MaxCapacity
template <typename T, depth_t MaxCapacity>
class range_vector {
    depth_t my_head;
    depth_t my_tail;
    depth_t my_size;
    depth_t my_depth[MaxCapacity]; // relative depths of stored ranges
    tbb::detail::aligned_space<T, MaxCapacity> my_pool;

public:
    //! initialize via first range in pool
    range_vector(const T& elem) : my_head(0), my_tail(0), my_size(1) {
        my_depth[0] = 0;
        new( static_cast<void *>(my_pool.begin()) ) T(elem);//TODO: std::move?
    }
    ~range_vector() {
        while( !empty() ) pop_back();
    }
    bool empty() const { return my_size == 0; }
    depth_t size() const { return my_size; }
    //! Populates range pool via ranges up to max depth or while divisible
    //! max_depth starts from 0, e.g. value 2 makes 3 ranges in the pool up to two 1/4 pieces
    void split_to_fill(depth_t max_depth) {
        while( my_size < MaxCapacity && is_divisible(max_depth) ) {
            depth_t prev = my_head;
            my_head = (my_head + 1) % MaxCapacity;
            new(my_pool.begin()+my_head) T(my_pool.begin()[prev]); // copy TODO: std::move?
            my_pool.begin()[prev].~T(); // instead of assignment
            new(my_pool.begin()+prev) T(my_pool.begin()[my_head], detail::split()); // do 'inverse' split
            my_depth[my_head] = ++my_depth[prev];
            my_size++;
        }
    }
    void pop_back() {
        __TBB_ASSERT(my_size > 0, "range_vector::pop_back() with empty size");
        my_pool.begin()[my_head].~T();
        my_size--;
        my_head = (my_head + MaxCapacity - 1) % MaxCapacity;
    }
    void pop_front() {
        __TBB_ASSERT(my_size > 0, "range_vector::pop_front() with empty size");
        my_pool.begin()[my_tail].~T();
        my_size--;
        my_tail = (my_tail + 1) % MaxCapacity;
    }
    T& back() {
        __TBB_ASSERT(my_size > 0, "range_vector::back() with empty size");
        return my_pool.begin()[my_head];
    }
    T& front() {
        __TBB_ASSERT(my_size > 0, "range_vector::front() with empty size");
        return my_pool.begin()[my_tail];
    }
    //! similarly to front(), returns depth of the first range in the pool
    depth_t front_depth() {
        __TBB_ASSERT(my_size > 0, "range_vector::front_depth() with empty size");
        return my_depth[my_tail];
    }
    depth_t back_depth() {
        __TBB_ASSERT(my_size > 0, "range_vector::back_depth() with empty size");
        return my_depth[my_head];
    }
    bool is_divisible(depth_t max_depth) {
        return back_depth() < max_depth && back().is_divisible();
    }
};

//! Provides default methods for partition objects and common algorithm blocks.
template <typename Partition>
struct partition_type_base {
    typedef detail::split split_type;
    // decision makers
    void note_affinity( slot_id ) {}
    template <typename Task>
    bool check_being_stolen(Task&, const execution_data&) { return false; } // part of old should_execute_range()
    template <typename Range> split_type get_split() { return split(); }
    Partition& self() { return *static_cast<Partition*>(this); } // CRTP helper

    template<typename StartType, typename Range>
    void work_balance(StartType &start, Range &range, const execution_data&) {
        start.run_body( range ); // static partitioner goes here
    }

    template<typename StartType, typename Range>
    void execute(StartType &start, Range &range, execution_data& ed) {
        // The algorithm in a few words ([]-denotes calls to decision methods of partitioner):
        // [If this task is stolen, adjust depth and divisions if necessary, set flag].
        // If range is divisible {
        //    Spread the work while [initial divisions left];
        //    Create trap task [if necessary];
        // }
        // If not divisible or [max depth is reached], execute, else do the range pool part
        if ( range.is_divisible() ) {
            if ( self().is_divisible() ) {
                do { // split until is divisible
                    typename Partition::split_type split_obj = self().template get_split<Range>();
                    start.offer_work( split_obj, ed );
                } while ( range.is_divisible() && self().is_divisible() );
            }
        }
        self().work_balance(start, range, ed);
    }
};

//! Provides default splitting strategy for partition objects.
template <typename Partition>
struct adaptive_mode : partition_type_base<Partition> {
    typedef Partition my_partition;
    std::size_t my_divisor;
    // For affinity_partitioner, my_divisor indicates the number of affinity array indices the task reserves.
    // A task which has only one index must produce the right split without reserved index in order to avoid
    // it to be overwritten in note_affinity() of the created (right) task.
    // I.e. a task created deeper than the affinity array can remember must not save its affinity (LIFO order)
    static const unsigned factor = 1;
    adaptive_mode() : my_divisor(get_initial_auto_partitioner_divisor() / 4 * my_partition::factor) {}
    adaptive_mode(adaptive_mode &src, split) : my_divisor(do_split(src, split())) {}
    adaptive_mode(adaptive_mode&, const proportional_split&) : my_divisor(0)
    {
        // left blank as my_divisor gets overridden in the successors' constructors
    }
    /*! Override do_split methods in order to specify splitting strategy */
    std::size_t do_split(adaptive_mode &src, split) {
        return src.my_divisor /= 2u;
    }
};

//! Helper type for checking availability of proportional_split constructor
template <typename T> using supports_proportional_splitting = typename std::is_constructible<T, T&, proportional_split&>;

//! A helper class to create a proportional_split object for a given type of Range.
/** If the Range has proportional_split constructor,
    then created object splits a provided value in an implemenation-defined proportion;
    otherwise it represents equal-size split. */
// TODO: check if this helper can be a nested class of proportional_mode.
template <typename Range, typename = void>
struct proportion_helper {
    static proportional_split get_split(std::size_t) { return proportional_split(1,1); }
};

template <typename Range>
struct proportion_helper<Range, typename std::enable_if<supports_proportional_splitting<Range>::value>::type> {
    static proportional_split get_split(std::size_t n) {
        std::size_t right = n / 2;
        std::size_t left  = n - right;
        return proportional_split(left, right);
    }
};

//! Provides proportional splitting strategy for partition objects
template <typename Partition>
struct proportional_mode : adaptive_mode<Partition> {
    typedef Partition my_partition;
    using partition_type_base<Partition>::self; // CRTP helper to get access to derived classes

    proportional_mode() : adaptive_mode<Partition>() {}
    proportional_mode(proportional_mode &src, split) : adaptive_mode<Partition>(src, split()) {}
    proportional_mode(proportional_mode &src, const proportional_split& split_obj)
        : adaptive_mode<Partition>(src, split_obj)
    {
        self().my_divisor = do_split(src, split_obj);
    }
    std::size_t do_split(proportional_mode &src, const proportional_split& split_obj) {
        std::size_t portion = split_obj.right() * my_partition::factor;
        portion = (portion + my_partition::factor/2) & (0ul - my_partition::factor);
        src.my_divisor -= portion;
        return portion;
    }
    bool is_divisible() { // part of old should_execute_range()
        return self().my_divisor > my_partition::factor;
    }
    template <typename Range>
    proportional_split get_split() {
        // Create a proportion for the number of threads expected to handle "this" subrange
        return proportion_helper<Range>::get_split( self().my_divisor / my_partition::factor );
    }
};

static std::size_t get_initial_partition_head() {
    int current_index = tbb::this_task_arena::current_thread_index();
    if (current_index == tbb::task_arena::not_initialized)
        current_index = 0;
    return size_t(current_index);
}

//! Provides default linear indexing of partitioner's sequence
template <typename Partition>
struct linear_affinity_mode : proportional_mode<Partition> {
    std::size_t my_head;
    std::size_t my_max_affinity;
    using proportional_mode<Partition>::self;
    linear_affinity_mode() : proportional_mode<Partition>(), my_head(get_initial_partition_head()),
                             my_max_affinity(self().my_divisor) {}
    linear_affinity_mode(linear_affinity_mode &src, split) : proportional_mode<Partition>(src, split())
        , my_head((src.my_head + src.my_divisor) % src.my_max_affinity), my_max_affinity(src.my_max_affinity) {}
    linear_affinity_mode(linear_affinity_mode &src, const proportional_split& split_obj) : proportional_mode<Partition>(src, split_obj)
        , my_head((src.my_head + src.my_divisor) % src.my_max_affinity), my_max_affinity(src.my_max_affinity) {}
    void spawn_task(task& t, task_group_context& ctx) {
        if (self().my_divisor) {
            spawn(t, ctx, slot_id(my_head));
        } else {
            spawn(t, ctx);
        }
    }
};

static bool is_stolen_task(const execution_data& ed) {
    return execution_slot(ed) != original_slot(ed);
}

/*! Determine work-balance phase implementing splitting & stealing actions */
template<class Mode>
struct dynamic_grainsize_mode : Mode {
    using Mode::self;
    enum {
        begin = 0,
        run,
        pass
    } my_delay;
    depth_t my_max_depth;
    static const unsigned range_pool_size = __TBB_RANGE_POOL_CAPACITY;
    dynamic_grainsize_mode(): Mode()
        , my_delay(begin)
        , my_max_depth(__TBB_INIT_DEPTH) {}
    dynamic_grainsize_mode(dynamic_grainsize_mode& p, split)
        : Mode(p, split())
        , my_delay(pass)
        , my_max_depth(p.my_max_depth) {}
    dynamic_grainsize_mode(dynamic_grainsize_mode& p, const proportional_split& split_obj)
        : Mode(p, split_obj)
        , my_delay(begin)
        , my_max_depth(p.my_max_depth) {}
    template <typename Task>
    bool check_being_stolen(Task &t, const execution_data& ed) { // part of old should_execute_range()
        if( !(self().my_divisor / Mode::my_partition::factor) ) { // if not from the top P tasks of binary tree
            self().my_divisor = 1; // TODO: replace by on-stack flag (partition_state's member)?
            if( is_stolen_task(ed) && t.my_parent->m_ref_count >= 2 ) { // runs concurrently with the left task
#if __TBB_USE_OPTIONAL_RTTI
                // RTTI is available, check whether the cast is valid
                // TODO: TBB_REVAMP_TODO __TBB_ASSERT(dynamic_cast<tree_node*>(t.m_parent), 0);
                // correctness of the cast relies on avoiding the root task for which:
                // - initial value of my_divisor != 0 (protected by separate assertion)
                // - is_stolen_task() always returns false for the root task.
#endif
                tree_node::mark_task_stolen(t);
                if( !my_max_depth ) my_max_depth++;
                my_max_depth += __TBB_DEMAND_DEPTH_ADD;
                return true;
            }
        }
        return false;
    }
    depth_t max_depth() { return my_max_depth; }
    void align_depth(depth_t base) {
        __TBB_ASSERT(base <= my_max_depth, nullptr);
        my_max_depth -= base;
    }
    template<typename StartType, typename Range>
    void work_balance(StartType &start, Range &range, execution_data& ed) {
        if( !range.is_divisible() || !self().max_depth() ) {
            start.run_body( range );
        }
        else { // do range pool
            range_vector<Range, range_pool_size> range_pool(range);
            do {
                range_pool.split_to_fill(self().max_depth()); // fill range pool
                if( self().check_for_demand( start ) ) {
                    if( range_pool.size() > 1 ) {
                        start.offer_work( range_pool.front(), range_pool.front_depth(), ed );
                        range_pool.pop_front();
                        continue;
                    }
                    if( range_pool.is_divisible(self().max_depth()) ) // was not enough depth to fork a task
                        continue; // note: next split_to_fill() should split range at least once
                }
                start.run_body( range_pool.back() );
                range_pool.pop_back();
            } while( !range_pool.empty() && !ed.context->is_group_execution_cancelled() );
        }
    }
    template <typename Task>
    bool check_for_demand(Task& t) {
        if ( pass == my_delay ) {
            if ( self().my_divisor > 1 ) // produce affinitized tasks while they have slot in array
                return true; // do not do my_max_depth++ here, but be sure range_pool is splittable once more
            else if ( self().my_divisor && my_max_depth ) { // make balancing task
                self().my_divisor = 0; // once for each task; depth will be decreased in align_depth()
                return true;
            }
            else if ( tree_node::is_peer_stolen(t) ) {
                my_max_depth += __TBB_DEMAND_DEPTH_ADD;
                return true;
            }
        } else if( begin == my_delay ) {
            my_delay = pass;
        }
        return false;
    }
};

class auto_partition_type: public dynamic_grainsize_mode<adaptive_mode<auto_partition_type> > {
public:
    auto_partition_type( const auto_partitioner& ) {
        my_divisor *= __TBB_INITIAL_CHUNKS;
    }
    auto_partition_type( auto_partition_type& src, split)
        : dynamic_grainsize_mode<adaptive_mode<auto_partition_type> >(src, split()) {}
    bool is_divisible() { // part of old should_execute_range()
        if( my_divisor > 1 ) return true;
        if( my_divisor && my_max_depth ) { // can split the task. TODO: on-stack flag instead
            // keep same fragmentation while splitting for the local task pool
            my_max_depth--;
            my_divisor = 0; // decrease max_depth once per task
            return true;
        } else return false;
    }
    template <typename Task>
    bool check_for_demand(Task& t) {
        if (tree_node::is_peer_stolen(t)) {
            my_max_depth += __TBB_DEMAND_DEPTH_ADD;
            return true;
        } else return false;
    }
    void spawn_task(task& t, task_group_context& ctx) {
        spawn(t, ctx);
    }
};

class simple_partition_type: public partition_type_base<simple_partition_type> {
public:
    simple_partition_type( const simple_partitioner& ) {}
    simple_partition_type( const simple_partition_type&, split ) {}
    //! simplified algorithm
    template<typename StartType, typename Range>
    void execute(StartType &start, Range &range, execution_data& ed) {
        split_type split_obj = split(); // start.offer_work accepts split_type as reference
        while( range.is_divisible() )
            start.offer_work( split_obj, ed );
        start.run_body( range );
    }
    void spawn_task(task& t, task_group_context& ctx) {
        spawn(t, ctx);
    }
};

class static_partition_type : public linear_affinity_mode<static_partition_type> {
public:
    typedef detail::proportional_split split_type;
    static_partition_type( const static_partitioner& ) {}
    static_partition_type( static_partition_type& p, const proportional_split& split_obj )
        : linear_affinity_mode<static_partition_type>(p, split_obj) {}
};

class affinity_partition_type : public dynamic_grainsize_mode<linear_affinity_mode<affinity_partition_type> > {
    static const unsigned factor_power = 4; // TODO: get a unified formula based on number of computing units
    slot_id* my_array;
public:
    static const unsigned factor = 1 << factor_power; // number of slots in affinity array per task
    typedef detail::proportional_split split_type;
    affinity_partition_type( affinity_partitioner_base& ap ) {
        __TBB_ASSERT( (factor&(factor-1))==0, "factor must be power of two" );
        ap.resize(factor);
        my_array = ap.my_array;
        my_max_depth = factor_power + 1;
        __TBB_ASSERT( my_max_depth < __TBB_RANGE_POOL_CAPACITY, nullptr );
    }
    affinity_partition_type(affinity_partition_type& p, split)
        : dynamic_grainsize_mode<linear_affinity_mode<affinity_partition_type> >(p, split())
        , my_array(p.my_array) {}
    affinity_partition_type(affinity_partition_type& p, const proportional_split& split_obj)
        : dynamic_grainsize_mode<linear_affinity_mode<affinity_partition_type> >(p, split_obj)
        , my_array(p.my_array) {}
    void note_affinity(slot_id id) {
        if( my_divisor )
            my_array[my_head] = id;
    }
    void spawn_task(task& t, task_group_context& ctx) {
        if (my_divisor) {
            if (!my_array[my_head]) {
                // TODO: consider new ideas with my_array for both affinity and static partitioner's, then code reuse
                spawn(t, ctx, slot_id(my_head / factor));
            } else {
                spawn(t, ctx, my_array[my_head]);
            }
        } else {
            spawn(t, ctx);
        }
    }
};

//! A simple partitioner
/** Divides the range until the range is not divisible.
    @ingroup algorithms */
class simple_partitioner {
public:
    simple_partitioner() {}
private:
    template<typename Range, typename Body, typename Partitioner> friend struct start_for;
    template<typename Range, typename Body, typename Partitioner> friend struct start_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_deterministic_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_scan;
    // new implementation just extends existing interface
    typedef simple_partition_type task_partition_type;
    // TODO: consider to make split_type public
    typedef simple_partition_type::split_type split_type;

    // for parallel_scan only
    class partition_type {
    public:
        bool should_execute_range(const execution_data& ) {return false;}
        partition_type( const simple_partitioner& ) {}
        partition_type( const partition_type&, split ) {}
    };
};

//! An auto partitioner
/** The range is initial divided into several large chunks.
    Chunks are further subdivided into smaller pieces if demand detected and they are divisible.
    @ingroup algorithms */
class auto_partitioner {
public:
    auto_partitioner() {}

private:
    template<typename Range, typename Body, typename Partitioner> friend struct start_for;
    template<typename Range, typename Body, typename Partitioner> friend struct start_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_deterministic_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_scan;
    // new implementation just extends existing interface
    typedef auto_partition_type task_partition_type;
    // TODO: consider to make split_type public
    typedef auto_partition_type::split_type split_type;

    //! Backward-compatible partition for auto and affinity partition objects.
    class partition_type {
        size_t num_chunks;
        static const size_t VICTIM_CHUNKS = 4;
        public:
        bool should_execute_range(const execution_data& ed) {
            if( num_chunks<VICTIM_CHUNKS && is_stolen_task(ed) )
                num_chunks = VICTIM_CHUNKS;
            return num_chunks==1;
        }
        partition_type( const auto_partitioner& )
            : num_chunks(get_initial_auto_partitioner_divisor()*__TBB_INITIAL_CHUNKS/4) {}
        partition_type( partition_type& pt, split ) {
            num_chunks = pt.num_chunks = (pt.num_chunks+1u) / 2u;
        }
    };
};

//! A static partitioner
class static_partitioner {
public:
    static_partitioner() {}
private:
    template<typename Range, typename Body, typename Partitioner> friend struct start_for;
    template<typename Range, typename Body, typename Partitioner> friend struct start_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_deterministic_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_scan;
    // new implementation just extends existing interface
    typedef static_partition_type task_partition_type;
    // TODO: consider to make split_type public
    typedef static_partition_type::split_type split_type;
};

//! An affinity partitioner
class affinity_partitioner : affinity_partitioner_base {
public:
    affinity_partitioner() {}

private:
    template<typename Range, typename Body, typename Partitioner> friend struct start_for;
    template<typename Range, typename Body, typename Partitioner> friend struct start_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_deterministic_reduce;
    template<typename Range, typename Body, typename Partitioner> friend struct start_scan;
    // new implementation just extends existing interface
    typedef affinity_partition_type task_partition_type;
    // TODO: consider to make split_type public
    typedef affinity_partition_type::split_type split_type;
};

} // namespace d1
} // namespace detail

inline namespace v1 {
// Partitioners
using detail::d1::auto_partitioner;
using detail::d1::simple_partitioner;
using detail::d1::static_partitioner;
using detail::d1::affinity_partitioner;
// Split types
using detail::split;
using detail::proportional_split;
} // namespace v1

} // namespace tbb

#if defined(_MSC_VER) && !defined(__INTEL_COMPILER)
    #pragma warning (pop)
#endif // warning 4244 is back

#undef __TBB_INITIAL_CHUNKS
#undef __TBB_RANGE_POOL_CAPACITY
#undef __TBB_INIT_DEPTH

#endif /* __TBB_partitioner_H */