random_shuffle.h

Go to the documentation of this file.

// -*- C++ -*-

// Copyright (C) 2007-2019 Free Software Foundation, Inc.
//
// This file is part of the GNU ISO C++ Library.  This library is free
// software; you can redistribute it and/or modify it under the terms
// of the GNU General Public License as published by the Free Software
// Foundation; either version 3, or (at your option) any later
// version.

// This library is distributed in the hope that it will be useful, but
// WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
// General Public License for more details.

// Under Section 7 of GPL version 3, you are granted additional
// permissions described in the GCC Runtime Library Exception, version
// 3.1, as published by the Free Software Foundation.

// You should have received a copy of the GNU General Public License and
// a copy of the GCC Runtime Library Exception along with this program;
// see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
// <http://www.gnu.org/licenses/>.

/** @file parallel/random_shuffle.h
 *  @brief Parallel implementation of std::random_shuffle().
 *  This file is a GNU parallel extension to the Standard C++ Library.
 */

// Written by Johannes Singler.

#ifndef _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H
#define _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H 1

#include <limits>
#include <bits/stl_numeric.h>
#include <parallel/parallel.h>
#include <parallel/random_number.h>

namespace __gnu_parallel
{
  /** @brief Type to hold the index of a bin.
    *
    *  Since many variables of this type are allocated, it should be
    *  chosen as small as possible.
    */
  typedef unsigned short _BinIndex;

  /** @brief Data known to every thread participating in
      __gnu_parallel::__parallel_random_shuffle(). */
  template<typename _RAIter>
    struct _DRandomShufflingGlobalData
    {
      typedef std::iterator_traits<_RAIter> _TraitsType;
      typedef typename _TraitsType::value_type _ValueType;
      typedef typename _TraitsType::difference_type _DifferenceType;

      /** @brief Begin iterator of the __source. */
      _RAIter& _M_source;

      /** @brief Temporary arrays for each thread. */
      _ValueType** _M_temporaries;

      /** @brief Two-dimensional array to hold the thread-bin distribution.
       *
       *  Dimensions (_M_num_threads + 1) __x (_M_num_bins + 1). */
      _DifferenceType** _M_dist;

      /** @brief Start indexes of the threads' __chunks. */
      _DifferenceType* _M_starts;

      /** @brief Number of the thread that will further process the
          corresponding bin. */
      _ThreadIndex* _M_bin_proc;

      /** @brief Number of bins to distribute to. */
      int _M_num_bins;

      /** @brief Number of bits needed to address the bins. */
      int _M_num_bits;

      /** @brief Constructor. */
      _DRandomShufflingGlobalData(_RAIter& __source)
      : _M_source(__source) { }
    };

  /** @brief Local data for a thread participating in
      __gnu_parallel::__parallel_random_shuffle().
    */
  template<typename _RAIter, typename _RandomNumberGenerator>
    struct _DRSSorterPU
    {
      /** @brief Number of threads participating in total. */
      int _M_num_threads;

      /** @brief Begin index for bins taken care of by this thread. */
      _BinIndex _M_bins_begin;

      /** @brief End index for bins taken care of by this thread. */
      _BinIndex __bins_end;

      /** @brief Random _M_seed for this thread. */
      uint32_t _M_seed;

      /** @brief Pointer to global data. */
      _DRandomShufflingGlobalData<_RAIter>* _M_sd;
    };

  /** @brief Generate a random number in @c [0,2^__logp).
    *  @param __logp Logarithm (basis 2) of the upper range __bound.
    *  @param __rng Random number generator to use.
    */
  template<typename _RandomNumberGenerator>
    inline int
    __random_number_pow2(int __logp, _RandomNumberGenerator& __rng)
    { return __rng.__genrand_bits(__logp); }

  /** @brief Random shuffle code executed by each thread.
    *  @param __pus Array of thread-local data records. */
  template<typename _RAIter, typename _RandomNumberGenerator>
    void 
    __parallel_random_shuffle_drs_pu(_DRSSorterPU<_RAIter,
                                     _RandomNumberGenerator>* __pus)
    {
      typedef std::iterator_traits<_RAIter> _TraitsType;
      typedef typename _TraitsType::value_type _ValueType;
      typedef typename _TraitsType::difference_type _DifferenceType;

      _ThreadIndex __iam = omp_get_thread_num();
      _DRSSorterPU<_RAIter, _RandomNumberGenerator>* __d = &__pus[__iam];
      _DRandomShufflingGlobalData<_RAIter>* __sd = __d->_M_sd;

      // Indexing: _M_dist[bin][processor]
      _DifferenceType __length = (__sd->_M_starts[__iam + 1]
                                  - __sd->_M_starts[__iam]);
      _BinIndex* __oracles = new _BinIndex[__length];
      _DifferenceType* __dist = new _DifferenceType[__sd->_M_num_bins + 1];
      _BinIndex* __bin_proc = new _BinIndex[__sd->_M_num_bins];
      _ValueType** __temporaries = new _ValueType*[__d->_M_num_threads];

      // Compute oracles and count appearances.
      for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b)
        __dist[__b] = 0;
      int __num_bits = __sd->_M_num_bits;

      _RandomNumber __rng(__d->_M_seed);

      // First main loop.
      for (_DifferenceType __i = 0; __i < __length; ++__i)
        {
          _BinIndex __oracle = __random_number_pow2(__num_bits, __rng);
          __oracles[__i] = __oracle;

          // To allow prefix (partial) sum.
          ++(__dist[__oracle + 1]);
        }

      for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b)
        __sd->_M_dist[__b][__iam + 1] = __dist[__b];

#     pragma omp barrier

#     pragma omp single
      {
        // Sum up bins, __sd->_M_dist[__s + 1][__d->_M_num_threads] now
        // contains the total number of items in bin __s
        for (_BinIndex __s = 0; __s < __sd->_M_num_bins; ++__s)
          __gnu_sequential::partial_sum(__sd->_M_dist[__s + 1],
                                        __sd->_M_dist[__s + 1]
                                        + __d->_M_num_threads + 1,
                                        __sd->_M_dist[__s + 1]);
      }

#     pragma omp barrier

      _SequenceIndex __offset = 0, __global_offset = 0;
      for (_BinIndex __s = 0; __s < __d->_M_bins_begin; ++__s)
        __global_offset += __sd->_M_dist[__s + 1][__d->_M_num_threads];

#     pragma omp barrier

      for (_BinIndex __s = __d->_M_bins_begin; __s < __d->__bins_end; ++__s)
        {
          for (int __t = 0; __t < __d->_M_num_threads + 1; ++__t)
            __sd->_M_dist[__s + 1][__t] += __offset;
          __offset = __sd->_M_dist[__s + 1][__d->_M_num_threads];
        }

      __sd->_M_temporaries[__iam] = static_cast<_ValueType*>
        (::operator new(sizeof(_ValueType) * __offset));

#     pragma omp barrier

      // Draw local copies to avoid false sharing.
      for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b)
        __dist[__b] = __sd->_M_dist[__b][__iam];
      for (_BinIndex __b = 0; __b < __sd->_M_num_bins; ++__b)
        __bin_proc[__b] = __sd->_M_bin_proc[__b];
      for (_ThreadIndex __t = 0; __t < __d->_M_num_threads; ++__t)
        __temporaries[__t] = __sd->_M_temporaries[__t];

      _RAIter __source = __sd->_M_source;
      _DifferenceType __start = __sd->_M_starts[__iam];

      // Distribute according to oracles, second main loop.
      for (_DifferenceType __i = 0; __i < __length; ++__i)
        {
          _BinIndex __target_bin = __oracles[__i];
          _ThreadIndex __target_p = __bin_proc[__target_bin];

          // Last column [__d->_M_num_threads] stays unchanged.
          ::new(&(__temporaries[__target_p][__dist[__target_bin + 1]++]))
              _ValueType(*(__source + __i + __start));
        }

      delete[] __oracles;
      delete[] __dist;
      delete[] __bin_proc;
      delete[] __temporaries;

#     pragma omp barrier

      // Shuffle bins internally.
      for (_BinIndex __b = __d->_M_bins_begin; __b < __d->__bins_end; ++__b)
        {
          _ValueType* __begin =
            (__sd->_M_temporaries[__iam]
             + (__b == __d->_M_bins_begin
                ? 0 : __sd->_M_dist[__b][__d->_M_num_threads])),
            *__end = (__sd->_M_temporaries[__iam]
                      + __sd->_M_dist[__b + 1][__d->_M_num_threads]);

          __sequential_random_shuffle(__begin, __end, __rng);
          std::copy(__begin, __end, __sd->_M_source + __global_offset
                    + (__b == __d->_M_bins_begin
                       ? 0 : __sd->_M_dist[__b][__d->_M_num_threads]));
        }

      for (_SequenceIndex __i = 0; __i < __offset; ++__i)
        __sd->_M_temporaries[__iam][__i].~_ValueType();
      ::operator delete(__sd->_M_temporaries[__iam]);
    }

  /** @brief Round up to the next greater power of 2.
    *  @param __x _Integer to round up */
  template<typename _Tp>
    _Tp 
    __round_up_to_pow2(_Tp __x)
    {
      if (__x <= 1)
        return 1;
      else
        return (_Tp)1 << (__rd_log2(__x - 1) + 1);
    }

  /** @brief Main parallel random shuffle step.
    *  @param __begin Begin iterator of sequence.
    *  @param __end End iterator of sequence.
    *  @param __n Length of sequence.
    *  @param __num_threads Number of threads to use.
    *  @param __rng Random number generator to use.
    */
  template<typename _RAIter, typename _RandomNumberGenerator>
    void
    __parallel_random_shuffle_drs(_RAIter __begin, _RAIter __end,
                                  typename std::iterator_traits
                                  <_RAIter>::difference_type __n,
                                  _ThreadIndex __num_threads,
                                  _RandomNumberGenerator& __rng)
    {
      typedef std::iterator_traits<_RAIter> _TraitsType;
      typedef typename _TraitsType::value_type _ValueType;
      typedef typename _TraitsType::difference_type _DifferenceType;

      _GLIBCXX_CALL(__n)

      const _Settings& __s = _Settings::get();

      if (__num_threads > __n)
        __num_threads = static_cast<_ThreadIndex>(__n);

      _BinIndex __num_bins, __num_bins_cache;

#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
      // Try the L1 cache first.

      // Must fit into L1.
      __num_bins_cache =
        std::max<_DifferenceType>(1, __n / (__s.L1_cache_size_lb
                                            / sizeof(_ValueType)));
      __num_bins_cache = __round_up_to_pow2(__num_bins_cache);

      // No more buckets than TLB entries, power of 2
      // Power of 2 and at least one element per bin, at most the TLB size.
      __num_bins = std::min<_DifferenceType>(__n, __num_bins_cache);

#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
      // 2 TLB entries needed per bin.
      __num_bins = std::min<_DifferenceType>(__s.TLB_size / 2, __num_bins);
#endif
      __num_bins = __round_up_to_pow2(__num_bins);

      if (__num_bins < __num_bins_cache)
        {
#endif
          // Now try the L2 cache
          // Must fit into L2
          __num_bins_cache = static_cast<_BinIndex>
            (std::max<_DifferenceType>(1, __n / (__s.L2_cache_size
                                                 / sizeof(_ValueType))));
          __num_bins_cache = __round_up_to_pow2(__num_bins_cache);

          // No more buckets than TLB entries, power of 2.
          __num_bins = static_cast<_BinIndex>
            (std::min(__n, static_cast<_DifferenceType>(__num_bins_cache)));
          // Power of 2 and at least one element per bin, at most the TLB size.
#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
          // 2 TLB entries needed per bin.
          __num_bins = std::min(static_cast<_DifferenceType>(__s.TLB_size / 2),
                                __num_bins);
#endif
            __num_bins = __round_up_to_pow2(__num_bins);
#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
        }
#endif

      __num_bins = __round_up_to_pow2(
                        std::max<_BinIndex>(__num_threads, __num_bins));

      if (__num_threads <= 1)
      {
        _RandomNumber __derived_rng(
                            __rng(std::numeric_limits<uint32_t>::max()));
        __sequential_random_shuffle(__begin, __end, __derived_rng);
        return;
      }

      _DRandomShufflingGlobalData<_RAIter> __sd(__begin);
      _DRSSorterPU<_RAIter, _RandomNumber >* __pus;
      _DifferenceType* __starts;

#     pragma omp parallel num_threads(__num_threads)
      {
        _ThreadIndex __num_threads = omp_get_num_threads();
#       pragma omp single
        {
          __pus = new _DRSSorterPU<_RAIter, _RandomNumber>[__num_threads];
          
          __sd._M_temporaries = new _ValueType*[__num_threads];
          __sd._M_dist = new _DifferenceType*[__num_bins + 1];
          __sd._M_bin_proc = new _ThreadIndex[__num_bins];
          for (_BinIndex __b = 0; __b < __num_bins + 1; ++__b)
            __sd._M_dist[__b] = new _DifferenceType[__num_threads + 1];
          for (_BinIndex __b = 0; __b < (__num_bins + 1); ++__b)
            {
              __sd._M_dist[0][0] = 0;
              __sd._M_dist[__b][0] = 0;
            }
          __starts = __sd._M_starts = new _DifferenceType[__num_threads + 1];
          int __bin_cursor = 0;
          __sd._M_num_bins = __num_bins;
          __sd._M_num_bits = __rd_log2(__num_bins);

          _DifferenceType __chunk_length = __n / __num_threads,
                                 __split = __n % __num_threads,
                                 __start = 0;
          _DifferenceType __bin_chunk_length = __num_bins / __num_threads,
                                 __bin_split = __num_bins % __num_threads;
          for (_ThreadIndex __i = 0; __i < __num_threads; ++__i)
            {
              __starts[__i] = __start;
              __start += (__i < __split
                          ? (__chunk_length + 1) : __chunk_length);
              int __j = __pus[__i]._M_bins_begin = __bin_cursor;

              // Range of bins for this processor.
              __bin_cursor += (__i < __bin_split
                               ? (__bin_chunk_length + 1)
                               : __bin_chunk_length);
              __pus[__i].__bins_end = __bin_cursor;
              for (; __j < __bin_cursor; ++__j)
                __sd._M_bin_proc[__j] = __i;
              __pus[__i]._M_num_threads = __num_threads;
              __pus[__i]._M_seed = __rng(std::numeric_limits<uint32_t>::max());
              __pus[__i]._M_sd = &__sd;
            }
          __starts[__num_threads] = __start;
        } //single
          // Now shuffle in parallel.
        __parallel_random_shuffle_drs_pu(__pus);
      }  // parallel

      delete[] __starts;
      delete[] __sd._M_bin_proc;
      for (int __s = 0; __s < (__num_bins + 1); ++__s)
        delete[] __sd._M_dist[__s];
      delete[] __sd._M_dist;
      delete[] __sd._M_temporaries;

      delete[] __pus;
    }

  /** @brief Sequential cache-efficient random shuffle.
   *  @param __begin Begin iterator of sequence.
   *  @param __end End iterator of sequence.
   *  @param __rng Random number generator to use.
   */
  template<typename _RAIter, typename _RandomNumberGenerator>
    void
    __sequential_random_shuffle(_RAIter __begin, _RAIter __end,
                                _RandomNumberGenerator& __rng)
    {
      typedef std::iterator_traits<_RAIter> _TraitsType;
      typedef typename _TraitsType::value_type _ValueType;
      typedef typename _TraitsType::difference_type _DifferenceType;

      _DifferenceType __n = __end - __begin;
      const _Settings& __s = _Settings::get();

      _BinIndex __num_bins, __num_bins_cache;

#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
      // Try the L1 cache first, must fit into L1.
      __num_bins_cache = std::max<_DifferenceType>
        (1, __n / (__s.L1_cache_size_lb / sizeof(_ValueType)));
      __num_bins_cache = __round_up_to_pow2(__num_bins_cache);

      // No more buckets than TLB entries, power of 2
      // Power of 2 and at least one element per bin, at most the TLB size
      __num_bins = std::min(__n, (_DifferenceType)__num_bins_cache);
#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
      // 2 TLB entries needed per bin
      __num_bins = std::min((_DifferenceType)__s.TLB_size / 2, __num_bins);
#endif
      __num_bins = __round_up_to_pow2(__num_bins);

      if (__num_bins < __num_bins_cache)
        {
#endif
          // Now try the L2 cache, must fit into L2.
          __num_bins_cache = static_cast<_BinIndex>
            (std::max<_DifferenceType>(1, __n / (__s.L2_cache_size
                                                 / sizeof(_ValueType))));
          __num_bins_cache = __round_up_to_pow2(__num_bins_cache);

          // No more buckets than TLB entries, power of 2
          // Power of 2 and at least one element per bin, at most the TLB size.
          __num_bins = static_cast<_BinIndex>
            (std::min(__n, static_cast<_DifferenceType>(__num_bins_cache)));

#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
          // 2 TLB entries needed per bin
          __num_bins = std::min<_DifferenceType>(__s.TLB_size / 2, __num_bins);
#endif
          __num_bins = __round_up_to_pow2(__num_bins);
#if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
        }
#endif

      int __num_bits = __rd_log2(__num_bins);

      if (__num_bins > 1)
        {
          _ValueType* __target =
            static_cast<_ValueType*>(::operator new(sizeof(_ValueType) * __n));
          _BinIndex* __oracles = new _BinIndex[__n];
          _DifferenceType* __dist0 = new _DifferenceType[__num_bins + 1],
                         * __dist1 = new _DifferenceType[__num_bins + 1];

          for (int __b = 0; __b < __num_bins + 1; ++__b)
            __dist0[__b] = 0;

          _RandomNumber __bitrng(__rng(0xFFFFFFFF));

          for (_DifferenceType __i = 0; __i < __n; ++__i)
            {
              _BinIndex __oracle = __random_number_pow2(__num_bits, __bitrng);
              __oracles[__i] = __oracle;

              // To allow prefix (partial) sum.
              ++(__dist0[__oracle + 1]);
            }

          // Sum up bins.
          __gnu_sequential::partial_sum(__dist0, __dist0 + __num_bins + 1,
                                        __dist0);

          for (int __b = 0; __b < __num_bins + 1; ++__b)
            __dist1[__b] = __dist0[__b];

          // Distribute according to oracles.
          for (_DifferenceType __i = 0; __i < __n; ++__i)
            ::new(&(__target[(__dist0[__oracles[__i]])++])) 
                _ValueType(*(__begin + __i));

          for (int __b = 0; __b < __num_bins; ++__b)
            __sequential_random_shuffle(__target + __dist1[__b],
                                        __target + __dist1[__b + 1], __rng);

          // Copy elements back.
          std::copy(__target, __target + __n, __begin);

          delete[] __dist0;
          delete[] __dist1;
          delete[] __oracles;
          
          for (_DifferenceType __i = 0; __i < __n; ++__i)
            __target[__i].~_ValueType();
          ::operator delete(__target);
        }
      else
        __gnu_sequential::random_shuffle(__begin, __end, __rng);
    }

  /** @brief Parallel random public call.
   *  @param __begin Begin iterator of sequence.
   *  @param __end End iterator of sequence.
   *  @param __rng Random number generator to use.
   */
  template<typename _RAIter, typename _RandomNumberGenerator>
    inline void
    __parallel_random_shuffle(_RAIter __begin, _RAIter __end,
                              _RandomNumberGenerator __rng = _RandomNumber())
    {
      typedef std::iterator_traits<_RAIter> _TraitsType;
      typedef typename _TraitsType::difference_type _DifferenceType;
      _DifferenceType __n = __end - __begin;
      __parallel_random_shuffle_drs(__begin, __end, __n,
                                    __get_max_threads(), __rng);
    }
}

#endif /* _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H */