3 // Copyright (C) 2007, 2008, 2009 Free Software Foundation, Inc.
5 // This file is part of the GNU ISO C++ Library. This library is free
6 // software; you can redistribute it and/or modify it under the terms
7 // of the GNU General Public License as published by the Free Software
8 // Foundation; either version 3, or (at your option) any later
11 // This library is distributed in the hope that it will be useful, but
12 // WITHOUT ANY WARRANTY; without even the implied warranty of
13 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 // General Public License for more details.
16 // Under Section 7 of GPL version 3, you are granted additional
17 // permissions described in the GCC Runtime Library Exception, version
18 // 3.1, as published by the Free Software Foundation.
20 // You should have received a copy of the GNU General Public License and
21 // a copy of the GCC Runtime Library Exception along with this program;
22 // see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
23 // <http://www.gnu.org/licenses/>.
25 /** @file parallel/random_shuffle.h
26 * @brief Parallel implementation of std::random_shuffle().
27 * This file is a GNU parallel extension to the Standard C++ Library.
30 // Written by Johannes Singler.
32 #ifndef _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H
33 #define _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H 1
36 #include <bits/stl_numeric.h>
37 #include <parallel/parallel.h>
38 #include <parallel/random_number.h>
40 namespace __gnu_parallel
42 /** @brief Type to hold the index of a bin.
44 * Since many variables of this type are allocated, it should be
45 * chosen as small as possible.
47 typedef unsigned short _BinIndex;
49 /** @brief Data known to every thread participating in
50 __gnu_parallel::__parallel_random_shuffle(). */
51 template<typename _RAIter>
52 struct _DRandomShufflingGlobalData
54 typedef std::iterator_traits<_RAIter> _TraitsType;
55 typedef typename _TraitsType::value_type _ValueType;
56 typedef typename _TraitsType::difference_type _DifferenceType;
58 /** @brief Begin iterator of the __source. */
61 /** @brief Temporary arrays for each thread. */
62 _ValueType** _M_temporaries;
64 /** @brief Two-dimensional array to hold the thread-bin distribution.
66 * Dimensions (_M_num_threads + 1) __x (_M_num_bins + 1). */
67 _DifferenceType** _M_dist;
69 /** @brief Start indexes of the threads' __chunks. */
70 _DifferenceType* _M_starts;
72 /** @brief Number of the thread that will further process the
74 _ThreadIndex* _M_bin_proc;
76 /** @brief Number of bins to distribute to. */
79 /** @brief Number of bits needed to address the bins. */
82 /** @brief Constructor. */
83 _DRandomShufflingGlobalData(_RAIter& __source)
84 : _M_source(__source) { }
87 /** @brief Local data for a thread participating in
88 __gnu_parallel::__parallel_random_shuffle().
90 template<typename _RAIter, typename _RandomNumberGenerator>
93 /** @brief Number of threads participating in total. */
96 /** @brief Begin index for bins taken care of by this thread. */
97 _BinIndex _M_bins_begin;
99 /** @brief End index for bins taken care of by this thread. */
100 _BinIndex __bins_end;
102 /** @brief Random _M_seed for this thread. */
105 /** @brief Pointer to global data. */
106 _DRandomShufflingGlobalData<_RAIter>* _M_sd;
109 /** @brief Generate a random number in @__c [0,2^logp).
110 * @param logp Logarithm (basis 2) of the upper range __bound.
111 * @param __rng Random number generator to use.
113 template<typename _RandomNumberGenerator>
115 __random_number_pow2(int logp, _RandomNumberGenerator& __rng)
116 { return __rng.__genrand_bits(logp); }
118 /** @brief Random shuffle code executed by each thread.
119 * @param __pus Array of thread-local data records. */
120 template<typename _RAIter, typename _RandomNumberGenerator>
122 __parallel_random_shuffle_drs_pu(_DRSSorterPU<_RAIter,
123 _RandomNumberGenerator>* __pus)
125 typedef std::iterator_traits<_RAIter> _TraitsType;
126 typedef typename _TraitsType::value_type _ValueType;
127 typedef typename _TraitsType::difference_type _DifferenceType;
129 _ThreadIndex __iam = omp_get_thread_num();
130 _DRSSorterPU<_RAIter, _RandomNumberGenerator>* d = &__pus[__iam];
131 _DRandomShufflingGlobalData<_RAIter>* _M_sd = d->_M_sd;
133 // Indexing: _M_dist[bin][processor]
134 _DifferenceType __length = _M_sd->_M_starts[__iam + 1] -
135 _M_sd->_M_starts[__iam];
136 _BinIndex* __oracles = new _BinIndex[__length];
137 _DifferenceType* _M_dist = new _DifferenceType[_M_sd->_M_num_bins + 1];
138 _BinIndex* _M_bin_proc = new _BinIndex[_M_sd->_M_num_bins];
139 _ValueType** _M_temporaries = new _ValueType*[d->_M_num_threads];
141 // Compute oracles and count appearances.
142 for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins + 1; ++__b)
144 int _M_num_bits = _M_sd->_M_num_bits;
146 _RandomNumber __rng(d->_M_seed);
149 for (_DifferenceType __i = 0; __i < __length; ++__i)
151 _BinIndex __oracle = __random_number_pow2(_M_num_bits, __rng);
152 __oracles[__i] = __oracle;
154 // To allow prefix (partial) sum.
155 ++(_M_dist[__oracle + 1]);
158 for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins + 1; ++__b)
159 _M_sd->_M_dist[__b][__iam + 1] = _M_dist[__b];
165 // Sum up bins, _M_sd->_M_dist[__s + 1][d->_M_num_threads] now contains
166 // the total number of items in bin __s
167 for (_BinIndex __s = 0; __s < _M_sd->_M_num_bins; ++__s)
168 __gnu_sequential::partial_sum(
169 _M_sd->_M_dist[__s + 1],
170 _M_sd->_M_dist[__s + 1] + d->_M_num_threads + 1,
171 _M_sd->_M_dist[__s + 1]);
176 _SequenceIndex __offset = 0, __global_offset = 0;
177 for (_BinIndex __s = 0; __s < d->_M_bins_begin; ++__s)
178 __global_offset += _M_sd->_M_dist[__s + 1][d->_M_num_threads];
182 for (_BinIndex __s = d->_M_bins_begin; __s < d->__bins_end; ++__s)
184 for (int __t = 0; __t < d->_M_num_threads + 1; ++__t)
185 _M_sd->_M_dist[__s + 1][__t] += __offset;
186 __offset = _M_sd->_M_dist[__s + 1][d->_M_num_threads];
189 _M_sd->_M_temporaries[__iam] = static_cast<_ValueType*>(
190 ::operator new(sizeof(_ValueType) * __offset));
194 // Draw local copies to avoid false sharing.
195 for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins + 1; ++__b)
196 _M_dist[__b] = _M_sd->_M_dist[__b][__iam];
197 for (_BinIndex __b = 0; __b < _M_sd->_M_num_bins; ++__b)
198 _M_bin_proc[__b] = _M_sd->_M_bin_proc[__b];
199 for (_ThreadIndex __t = 0; __t < d->_M_num_threads; ++__t)
200 _M_temporaries[__t] = _M_sd->_M_temporaries[__t];
202 _RAIter _M_source = _M_sd->_M_source;
203 _DifferenceType __start = _M_sd->_M_starts[__iam];
205 // Distribute according to oracles, second main loop.
206 for (_DifferenceType __i = 0; __i < __length; ++__i)
208 _BinIndex target_bin = __oracles[__i];
209 _ThreadIndex target_p = _M_bin_proc[target_bin];
211 // Last column [d->_M_num_threads] stays unchanged.
212 ::new(&(_M_temporaries[target_p][_M_dist[target_bin + 1]++]))
213 _ValueType(*(_M_source + __i + __start));
218 delete[] _M_bin_proc;
219 delete[] _M_temporaries;
223 // Shuffle bins internally.
224 for (_BinIndex __b = d->_M_bins_begin; __b < d->__bins_end; ++__b)
226 _ValueType* __begin =
227 _M_sd->_M_temporaries[__iam] +
228 ((__b == d->_M_bins_begin)
229 ? 0 : _M_sd->_M_dist[__b][d->_M_num_threads]),
231 _M_sd->_M_temporaries[__iam] +
232 _M_sd->_M_dist[__b + 1][d->_M_num_threads];
233 __sequential_random_shuffle(__begin, __end, __rng);
234 std::copy(__begin, __end, _M_sd->_M_source + __global_offset +
235 ((__b == d->_M_bins_begin)
236 ? 0 : _M_sd->_M_dist[__b][d->_M_num_threads]));
239 ::operator delete(_M_sd->_M_temporaries[__iam]);
242 /** @brief Round up to the next greater power of 2.
243 * @param __x _Integer to round up */
244 template<typename _Tp>
246 __round_up_to_pow2(_Tp __x)
251 return (_Tp)1 << (__rd_log2(__x - 1) + 1);
254 /** @brief Main parallel random shuffle step.
255 * @param __begin Begin iterator of sequence.
256 * @param __end End iterator of sequence.
257 * @param __n Length of sequence.
258 * @param __num_threads Number of threads to use.
259 * @param __rng Random number generator to use.
261 template<typename _RAIter, typename _RandomNumberGenerator>
263 __parallel_random_shuffle_drs(_RAIter __begin,
265 typename std::iterator_traits
266 <_RAIter>::difference_type __n,
267 _ThreadIndex __num_threads,
268 _RandomNumberGenerator& __rng)
270 typedef std::iterator_traits<_RAIter> _TraitsType;
271 typedef typename _TraitsType::value_type _ValueType;
272 typedef typename _TraitsType::difference_type _DifferenceType;
276 const _Settings& __s = _Settings::get();
278 if (__num_threads > __n)
279 __num_threads = static_cast<_ThreadIndex>(__n);
281 _BinIndex _M_num_bins, __num_bins_cache;
283 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
284 // Try the L1 cache first.
287 __num_bins_cache = std::max<_DifferenceType>(
288 1, __n / (__s.L1_cache_size_lb / sizeof(_ValueType)));
289 __num_bins_cache = __round_up_to_pow2(__num_bins_cache);
291 // No more buckets than TLB entries, power of 2
292 // Power of 2 and at least one element per bin, at most the TLB size.
293 _M_num_bins = std::min<_DifferenceType>(__n, __num_bins_cache);
295 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
296 // 2 TLB entries needed per bin.
297 _M_num_bins = std::min<_DifferenceType>(__s.TLB_size / 2, _M_num_bins);
299 _M_num_bins = __round_up_to_pow2(_M_num_bins);
301 if (_M_num_bins < __num_bins_cache)
304 // Now try the L2 cache
306 __num_bins_cache = static_cast<_BinIndex>(std::max<_DifferenceType>(
307 1, __n / (__s.L2_cache_size / sizeof(_ValueType))));
308 __num_bins_cache = __round_up_to_pow2(__num_bins_cache);
310 // No more buckets than TLB entries, power of 2.
311 _M_num_bins = static_cast<_BinIndex>(
312 std::min(__n, static_cast<_DifferenceType>(__num_bins_cache)));
313 // Power of 2 and at least one element per bin, at most the TLB size.
314 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
315 // 2 TLB entries needed per bin.
316 _M_num_bins = std::min(
317 static_cast<_DifferenceType>(__s.TLB_size / 2), _M_num_bins);
319 _M_num_bins = __round_up_to_pow2(_M_num_bins);
320 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
324 __num_threads = std::min<_BinIndex>(__num_threads, _M_num_bins);
326 if (__num_threads <= 1)
327 return __sequential_random_shuffle(__begin, __end, __rng);
329 _DRandomShufflingGlobalData<_RAIter> _M_sd(__begin);
330 _DRSSorterPU<_RAIter, _RandomNumber >* __pus;
331 _DifferenceType* _M_starts;
333 # pragma omp parallel num_threads(__num_threads)
335 _ThreadIndex __num_threads = omp_get_num_threads();
338 __pus = new _DRSSorterPU<_RAIter, _RandomNumber>
341 _M_sd._M_temporaries = new _ValueType*[__num_threads];
342 _M_sd._M_dist = new _DifferenceType*[_M_num_bins + 1];
343 _M_sd._M_bin_proc = new _ThreadIndex[_M_num_bins];
344 for (_BinIndex __b = 0; __b < _M_num_bins + 1; ++__b)
345 _M_sd._M_dist[__b] = new _DifferenceType[__num_threads + 1];
346 for (_BinIndex __b = 0; __b < (_M_num_bins + 1); ++__b)
348 _M_sd._M_dist[0][0] = 0;
349 _M_sd._M_dist[__b][0] = 0;
351 _M_starts = _M_sd._M_starts
352 = new _DifferenceType[__num_threads + 1];
354 _M_sd._M_num_bins = _M_num_bins;
355 _M_sd._M_num_bits = __rd_log2(_M_num_bins);
357 _DifferenceType __chunk_length = __n / __num_threads,
358 __split = __n % __num_threads, __start = 0;
359 _DifferenceType bin_chunk_length = _M_num_bins / __num_threads,
360 bin_split = _M_num_bins % __num_threads;
361 for (_ThreadIndex __i = 0; __i < __num_threads; ++__i)
363 _M_starts[__i] = __start;
364 __start += (__i < __split)
365 ? (__chunk_length + 1) : __chunk_length;
366 int __j = __pus[__i]._M_bins_begin = bin_cursor;
368 // Range of bins for this processor.
369 bin_cursor += (__i < bin_split) ?
370 (bin_chunk_length + 1) : bin_chunk_length;
371 __pus[__i].__bins_end = bin_cursor;
372 for (; __j < bin_cursor; ++__j)
373 _M_sd._M_bin_proc[__j] = __i;
374 __pus[__i]._M_num_threads = __num_threads;
375 __pus[__i]._M_seed = __rng(std::numeric_limits<uint32>::max());
376 __pus[__i]._M_sd = &_M_sd;
378 _M_starts[__num_threads] = __start;
380 // Now shuffle in parallel.
381 __parallel_random_shuffle_drs_pu(__pus);
385 delete[] _M_sd._M_bin_proc;
386 for (int __s = 0; __s < (_M_num_bins + 1); ++__s)
387 delete[] _M_sd._M_dist[__s];
388 delete[] _M_sd._M_dist;
389 delete[] _M_sd._M_temporaries;
394 /** @brief Sequential cache-efficient random shuffle.
395 * @param __begin Begin iterator of sequence.
396 * @param __end End iterator of sequence.
397 * @param __rng Random number generator to use.
399 template<typename _RAIter, typename _RandomNumberGenerator>
401 __sequential_random_shuffle(_RAIter __begin,
403 _RandomNumberGenerator& __rng)
405 typedef std::iterator_traits<_RAIter> _TraitsType;
406 typedef typename _TraitsType::value_type _ValueType;
407 typedef typename _TraitsType::difference_type _DifferenceType;
409 _DifferenceType __n = __end - __begin;
410 const _Settings& __s = _Settings::get();
412 _BinIndex _M_num_bins, __num_bins_cache;
414 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
415 // Try the L1 cache first, must fit into L1.
417 std::max<_DifferenceType>
418 (1, __n / (__s.L1_cache_size_lb / sizeof(_ValueType)));
419 __num_bins_cache = __round_up_to_pow2(__num_bins_cache);
421 // No more buckets than TLB entries, power of 2
422 // Power of 2 and at least one element per bin, at most the TLB size
423 _M_num_bins = std::min(__n, (_DifferenceType)__num_bins_cache);
424 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
425 // 2 TLB entries needed per bin
426 _M_num_bins = std::min((_DifferenceType)__s.TLB_size / 2, _M_num_bins);
428 _M_num_bins = __round_up_to_pow2(_M_num_bins);
430 if (_M_num_bins < __num_bins_cache)
433 // Now try the L2 cache, must fit into L2.
435 static_cast<_BinIndex>(std::max<_DifferenceType>(
436 1, __n / (__s.L2_cache_size / sizeof(_ValueType))));
437 __num_bins_cache = __round_up_to_pow2(__num_bins_cache);
439 // No more buckets than TLB entries, power of 2
440 // Power of 2 and at least one element per bin, at most the TLB size.
441 _M_num_bins = static_cast<_BinIndex>
442 (std::min(__n, static_cast<_DifferenceType>(__num_bins_cache)));
444 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
445 // 2 TLB entries needed per bin
447 std::min<_DifferenceType>(__s.TLB_size / 2, _M_num_bins);
449 _M_num_bins = __round_up_to_pow2(_M_num_bins);
450 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
454 int _M_num_bits = __rd_log2(_M_num_bins);
458 _ValueType* __target = static_cast<_ValueType*>(
459 ::operator new(sizeof(_ValueType) * __n));
460 _BinIndex* __oracles = new _BinIndex[__n];
461 _DifferenceType* __dist0 = new _DifferenceType[_M_num_bins + 1],
462 * __dist1 = new _DifferenceType[_M_num_bins + 1];
464 for (int __b = 0; __b < _M_num_bins + 1; ++__b)
467 _RandomNumber bitrng(__rng(0xFFFFFFFF));
469 for (_DifferenceType __i = 0; __i < __n; ++__i)
471 _BinIndex __oracle = __random_number_pow2(_M_num_bits, bitrng);
472 __oracles[__i] = __oracle;
474 // To allow prefix (partial) sum.
475 ++(__dist0[__oracle + 1]);
480 partial_sum(__dist0, __dist0 + _M_num_bins + 1, __dist0);
482 for (int __b = 0; __b < _M_num_bins + 1; ++__b)
483 __dist1[__b] = __dist0[__b];
485 // Distribute according to oracles.
486 for (_DifferenceType __i = 0; __i < __n; ++__i)
487 ::new(&(__target[(__dist0[__oracles[__i]])++]))
488 _ValueType(*(__begin + __i));
490 for (int __b = 0; __b < _M_num_bins; ++__b)
492 __sequential_random_shuffle(__target + __dist1[__b],
493 __target + __dist1[__b + 1],
497 // Copy elements back.
498 std::copy(__target, __target + __n, __begin);
503 ::operator delete(__target);
506 __gnu_sequential::random_shuffle(__begin, __end, __rng);
509 /** @brief Parallel random public call.
510 * @param __begin Begin iterator of sequence.
511 * @param __end End iterator of sequence.
512 * @param __rng Random number generator to use.
514 template<typename _RAIter, typename _RandomNumberGenerator>
516 __parallel_random_shuffle(_RAIter __begin,
518 _RandomNumberGenerator __rng = _RandomNumber())
520 typedef std::iterator_traits<_RAIter> _TraitsType;
521 typedef typename _TraitsType::difference_type _DifferenceType;
522 _DifferenceType __n = __end - __begin;
523 __parallel_random_shuffle_drs(
524 __begin, __end, __n, __get_max_threads(), __rng) ;
529 #endif /* _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H */