libstdc++
balanced_quicksort.h
Go to the documentation of this file.
1 // -*- C++ -*-
2 
3 // Copyright (C) 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
4 //
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
9 // version.
10 
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.
15 
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.
19 
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/>.
24 
25 /** @file parallel/balanced_quicksort.h
26  * @brief Implementation of a dynamically load-balanced parallel quicksort.
27  *
28  * It works in-place and needs only logarithmic extra memory.
29  * The algorithm is similar to the one proposed in
30  *
31  * P. Tsigas and Y. Zhang.
32  * A simple, fast parallel implementation of quicksort and
33  * its performance evaluation on SUN enterprise 10000.
34  * In 11th Euromicro Conference on Parallel, Distributed and
35  * Network-Based Processing, page 372, 2003.
36  *
37  * This file is a GNU parallel extension to the Standard C++ Library.
38  */
39 
40 // Written by Johannes Singler.
41 
42 #ifndef _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H
43 #define _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H 1
44 
46 #include <bits/stl_algo.h>
47 #include <bits/stl_function.h>
48 
49 #include <parallel/settings.h>
50 #include <parallel/partition.h>
51 #include <parallel/random_number.h>
52 #include <parallel/queue.h>
53 
54 #if _GLIBCXX_ASSERTIONS
55 #include <parallel/checkers.h>
56 #endif
57 
58 namespace __gnu_parallel
59 {
60  /** @brief Information local to one thread in the parallel quicksort run. */
61  template<typename _RAIter>
63  {
64  typedef std::iterator_traits<_RAIter> _TraitsType;
65  typedef typename _TraitsType::difference_type _DifferenceType;
66 
67  /** @brief Continuous part of the sequence, described by an
68  iterator pair. */
70 
71  /** @brief Initial piece to work on. */
73 
74  /** @brief Work-stealing queue. */
76 
77  /** @brief Number of threads involved in this algorithm. */
79 
80  /** @brief Pointer to a counter of elements left over to sort. */
81  volatile _DifferenceType* _M_elements_leftover;
82 
83  /** @brief The complete sequence to sort. */
85 
86  /** @brief Constructor.
87  * @param __queue_size size of the work-stealing queue. */
88  _QSBThreadLocal(int __queue_size) : _M_leftover_parts(__queue_size) { }
89  };
90 
91  /** @brief Balanced quicksort divide step.
92  * @param __begin Begin iterator of subsequence.
93  * @param __end End iterator of subsequence.
94  * @param __comp Comparator.
95  * @param __num_threads Number of threads that are allowed to work on
96  * this part.
97  * @pre @c (__end-__begin)>=1 */
98  template<typename _RAIter, typename _Compare>
99  typename std::iterator_traits<_RAIter>::difference_type
100  __qsb_divide(_RAIter __begin, _RAIter __end,
101  _Compare __comp, _ThreadIndex __num_threads)
102  {
103  _GLIBCXX_PARALLEL_ASSERT(__num_threads > 0);
104 
105  typedef std::iterator_traits<_RAIter> _TraitsType;
106  typedef typename _TraitsType::value_type _ValueType;
107  typedef typename _TraitsType::difference_type _DifferenceType;
108 
109  _RAIter __pivot_pos =
110  __median_of_three_iterators(__begin, __begin + (__end - __begin) / 2,
111  __end - 1, __comp);
112 
113 #if defined(_GLIBCXX_ASSERTIONS)
114  // Must be in between somewhere.
115  _DifferenceType __n = __end - __begin;
116 
117  _GLIBCXX_PARALLEL_ASSERT((!__comp(*__pivot_pos, *__begin)
118  && !__comp(*(__begin + __n / 2),
119  *__pivot_pos))
120  || (!__comp(*__pivot_pos, *__begin)
121  && !__comp(*(__end - 1), *__pivot_pos))
122  || (!__comp(*__pivot_pos, *(__begin + __n / 2))
123  && !__comp(*__begin, *__pivot_pos))
124  || (!__comp(*__pivot_pos, *(__begin + __n / 2))
125  && !__comp(*(__end - 1), *__pivot_pos))
126  || (!__comp(*__pivot_pos, *(__end - 1))
127  && !__comp(*__begin, *__pivot_pos))
128  || (!__comp(*__pivot_pos, *(__end - 1))
129  && !__comp(*(__begin + __n / 2),
130  *__pivot_pos)));
131 #endif
132 
133  // Swap pivot value to end.
134  if (__pivot_pos != (__end - 1))
135  std::iter_swap(__pivot_pos, __end - 1);
136  __pivot_pos = __end - 1;
137 
139  __pred(__comp, *__pivot_pos);
140 
141  // Divide, returning __end - __begin - 1 in the worst case.
142  _DifferenceType __split_pos = __parallel_partition(__begin, __end - 1,
143  __pred,
144  __num_threads);
145 
146  // Swap back pivot to middle.
147  std::iter_swap(__begin + __split_pos, __pivot_pos);
148  __pivot_pos = __begin + __split_pos;
149 
150 #if _GLIBCXX_ASSERTIONS
151  _RAIter __r;
152  for (__r = __begin; __r != __pivot_pos; ++__r)
153  _GLIBCXX_PARALLEL_ASSERT(__comp(*__r, *__pivot_pos));
154  for (; __r != __end; ++__r)
155  _GLIBCXX_PARALLEL_ASSERT(!__comp(*__r, *__pivot_pos));
156 #endif
157 
158  return __split_pos;
159  }
160 
161  /** @brief Quicksort conquer step.
162  * @param __tls Array of thread-local storages.
163  * @param __begin Begin iterator of subsequence.
164  * @param __end End iterator of subsequence.
165  * @param __comp Comparator.
166  * @param __iam Number of the thread processing this function.
167  * @param __num_threads
168  * Number of threads that are allowed to work on this part. */
169  template<typename _RAIter, typename _Compare>
170  void
172  _RAIter __begin, _RAIter __end,
173  _Compare __comp,
174  _ThreadIndex __iam, _ThreadIndex __num_threads,
175  bool __parent_wait)
176  {
177  typedef std::iterator_traits<_RAIter> _TraitsType;
178  typedef typename _TraitsType::value_type _ValueType;
179  typedef typename _TraitsType::difference_type _DifferenceType;
180 
181  _DifferenceType __n = __end - __begin;
182 
183  if (__num_threads <= 1 || __n <= 1)
184  {
185  __tls[__iam]->_M_initial.first = __begin;
186  __tls[__iam]->_M_initial.second = __end;
187 
188  __qsb_local_sort_with_helping(__tls, __comp, __iam, __parent_wait);
189 
190  return;
191  }
192 
193  // Divide step.
194  _DifferenceType __split_pos =
195  __qsb_divide(__begin, __end, __comp, __num_threads);
196 
197 #if _GLIBCXX_ASSERTIONS
198  _GLIBCXX_PARALLEL_ASSERT(0 <= __split_pos &&
199  __split_pos < (__end - __begin));
200 #endif
201 
203  __num_threads_leftside = std::max<_ThreadIndex>
204  (1, std::min<_ThreadIndex>(__num_threads - 1, __split_pos
205  * __num_threads / __n));
206 
207 # pragma omp atomic
208  *__tls[__iam]->_M_elements_leftover -= (_DifferenceType)1;
209 
210  // Conquer step.
211 # pragma omp parallel num_threads(2)
212  {
213  bool __wait;
214  if(omp_get_num_threads() < 2)
215  __wait = false;
216  else
217  __wait = __parent_wait;
218 
219 # pragma omp sections
220  {
221 # pragma omp section
222  {
223  __qsb_conquer(__tls, __begin, __begin + __split_pos, __comp,
224  __iam, __num_threads_leftside, __wait);
225  __wait = __parent_wait;
226  }
227  // The pivot_pos is left in place, to ensure termination.
228 # pragma omp section
229  {
230  __qsb_conquer(__tls, __begin + __split_pos + 1, __end, __comp,
231  __iam + __num_threads_leftside,
232  __num_threads - __num_threads_leftside, __wait);
233  __wait = __parent_wait;
234  }
235  }
236  }
237  }
238 
239  /**
240  * @brief Quicksort step doing load-balanced local sort.
241  * @param __tls Array of thread-local storages.
242  * @param __comp Comparator.
243  * @param __iam Number of the thread processing this function.
244  */
245  template<typename _RAIter, typename _Compare>
246  void
248  _Compare& __comp, _ThreadIndex __iam,
249  bool __wait)
250  {
251  typedef std::iterator_traits<_RAIter> _TraitsType;
252  typedef typename _TraitsType::value_type _ValueType;
253  typedef typename _TraitsType::difference_type _DifferenceType;
255 
256  _QSBThreadLocal<_RAIter>& __tl = *__tls[__iam];
257 
258  _DifferenceType
260  if (__base_case_n < 2)
261  __base_case_n = 2;
262  _ThreadIndex __num_threads = __tl._M_num_threads;
263 
264  // Every thread has its own random number generator.
265  _RandomNumber __rng(__iam + 1);
266 
267  _Piece __current = __tl._M_initial;
268 
269  _DifferenceType __elements_done = 0;
270 #if _GLIBCXX_ASSERTIONS
271  _DifferenceType __total_elements_done = 0;
272 #endif
273 
274  for (;;)
275  {
276  // Invariant: __current must be a valid (maybe empty) range.
277  _RAIter __begin = __current.first, __end = __current.second;
278  _DifferenceType __n = __end - __begin;
279 
280  if (__n > __base_case_n)
281  {
282  // Divide.
283  _RAIter __pivot_pos = __begin + __rng(__n);
284 
285  // Swap __pivot_pos value to end.
286  if (__pivot_pos != (__end - 1))
287  std::iter_swap(__pivot_pos, __end - 1);
288  __pivot_pos = __end - 1;
289 
291  <_Compare, _ValueType, _ValueType, bool>
292  __pred(__comp, *__pivot_pos);
293 
294  // Divide, leave pivot unchanged in last place.
295  _RAIter __split_pos1, __split_pos2;
296  __split_pos1 = __gnu_sequential::partition(__begin, __end - 1,
297  __pred);
298 
299  // Left side: < __pivot_pos; __right side: >= __pivot_pos.
300 #if _GLIBCXX_ASSERTIONS
301  _GLIBCXX_PARALLEL_ASSERT(__begin <= __split_pos1
302  && __split_pos1 < __end);
303 #endif
304  // Swap pivot back to middle.
305  if (__split_pos1 != __pivot_pos)
306  std::iter_swap(__split_pos1, __pivot_pos);
307  __pivot_pos = __split_pos1;
308 
309  // In case all elements are equal, __split_pos1 == 0.
310  if ((__split_pos1 + 1 - __begin) < (__n >> 7)
311  || (__end - __split_pos1) < (__n >> 7))
312  {
313  // Very unequal split, one part smaller than one 128th
314  // elements not strictly larger than the pivot.
316  <_Compare, _ValueType, _ValueType, bool>, _ValueType>
318  <_Compare, _ValueType, _ValueType, bool>
319  (__comp, *__pivot_pos));
320 
321  // Find other end of pivot-equal range.
322  __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1,
323  __end, __pred);
324  }
325  else
326  // Only skip the pivot.
327  __split_pos2 = __split_pos1 + 1;
328 
329  // Elements equal to pivot are done.
330  __elements_done += (__split_pos2 - __split_pos1);
331 #if _GLIBCXX_ASSERTIONS
332  __total_elements_done += (__split_pos2 - __split_pos1);
333 #endif
334  // Always push larger part onto stack.
335  if (((__split_pos1 + 1) - __begin) < (__end - (__split_pos2)))
336  {
337  // Right side larger.
338  if ((__split_pos2) != __end)
339  __tl._M_leftover_parts.push_front
340  (std::make_pair(__split_pos2, __end));
341 
342  //__current.first = __begin; //already set anyway
343  __current.second = __split_pos1;
344  continue;
345  }
346  else
347  {
348  // Left side larger.
349  if (__begin != __split_pos1)
350  __tl._M_leftover_parts.push_front(std::make_pair
351  (__begin, __split_pos1));
352 
353  __current.first = __split_pos2;
354  //__current.second = __end; //already set anyway
355  continue;
356  }
357  }
358  else
359  {
360  __gnu_sequential::sort(__begin, __end, __comp);
361  __elements_done += __n;
362 #if _GLIBCXX_ASSERTIONS
363  __total_elements_done += __n;
364 #endif
365 
366  // Prefer own stack, small pieces.
367  if (__tl._M_leftover_parts.pop_front(__current))
368  continue;
369 
370 # pragma omp atomic
371  *__tl._M_elements_leftover -= __elements_done;
372 
373  __elements_done = 0;
374 
375 #if _GLIBCXX_ASSERTIONS
376  double __search_start = omp_get_wtime();
377 #endif
378 
379  // Look for new work.
380  bool __successfully_stolen = false;
381  while (__wait && *__tl._M_elements_leftover > 0
382  && !__successfully_stolen
384  // Possible dead-lock.
385  && (omp_get_wtime() < (__search_start + 1.0))
386 #endif
387  )
388  {
389  _ThreadIndex __victim;
390  __victim = __rng(__num_threads);
391 
392  // Large pieces.
393  __successfully_stolen = (__victim != __iam)
394  && __tls[__victim]->_M_leftover_parts.pop_back(__current);
395  if (!__successfully_stolen)
396  __yield();
397 #if !defined(__ICC) && !defined(__ECC)
398 # pragma omp flush
399 #endif
400  }
401 
402 #if _GLIBCXX_ASSERTIONS
403  if (omp_get_wtime() >= (__search_start + 1.0))
404  {
405  sleep(1);
406  _GLIBCXX_PARALLEL_ASSERT(omp_get_wtime()
407  < (__search_start + 1.0));
408  }
409 #endif
410  if (!__successfully_stolen)
411  {
412 #if _GLIBCXX_ASSERTIONS
413  _GLIBCXX_PARALLEL_ASSERT(*__tl._M_elements_leftover == 0);
414 #endif
415  return;
416  }
417  }
418  }
419  }
420 
421  /** @brief Top-level quicksort routine.
422  * @param __begin Begin iterator of sequence.
423  * @param __end End iterator of sequence.
424  * @param __comp Comparator.
425  * @param __num_threads Number of threads that are allowed to work on
426  * this part.
427  */
428  template<typename _RAIter, typename _Compare>
429  void
430  __parallel_sort_qsb(_RAIter __begin, _RAIter __end,
431  _Compare __comp, _ThreadIndex __num_threads)
432  {
433  _GLIBCXX_CALL(__end - __begin)
434 
435  typedef std::iterator_traits<_RAIter> _TraitsType;
436  typedef typename _TraitsType::value_type _ValueType;
437  typedef typename _TraitsType::difference_type _DifferenceType;
439 
440  typedef _QSBThreadLocal<_RAIter> _TLSType;
441 
442  _DifferenceType __n = __end - __begin;
443 
444  if (__n <= 1)
445  return;
446 
447  // At least one element per processor.
448  if (__num_threads > __n)
449  __num_threads = static_cast<_ThreadIndex>(__n);
450 
451  // Initialize thread local storage
452  _TLSType** __tls = new _TLSType*[__num_threads];
453  _DifferenceType __queue_size = (__num_threads
454  * (_ThreadIndex)(__rd_log2(__n) + 1));
455  for (_ThreadIndex __t = 0; __t < __num_threads; ++__t)
456  __tls[__t] = new _QSBThreadLocal<_RAIter>(__queue_size);
457 
458  // There can never be more than ceil(__rd_log2(__n)) ranges on the
459  // stack, because
460  // 1. Only one processor pushes onto the stack
461  // 2. The largest range has at most length __n
462  // 3. Each range is larger than half of the range remaining
463  volatile _DifferenceType __elements_leftover = __n;
464  for (_ThreadIndex __i = 0; __i < __num_threads; ++__i)
465  {
466  __tls[__i]->_M_elements_leftover = &__elements_leftover;
467  __tls[__i]->_M_num_threads = __num_threads;
468  __tls[__i]->_M_global = std::make_pair(__begin, __end);
469 
470  // Just in case nothing is left to assign.
471  __tls[__i]->_M_initial = std::make_pair(__end, __end);
472  }
473 
474  // Main recursion call.
475  __qsb_conquer(__tls, __begin, __begin + __n, __comp, 0,
476  __num_threads, true);
477 
478 #if _GLIBCXX_ASSERTIONS
479  // All stack must be empty.
480  _Piece __dummy;
481  for (_ThreadIndex __i = 1; __i < __num_threads; ++__i)
482  _GLIBCXX_PARALLEL_ASSERT(
483  !__tls[__i]->_M_leftover_parts.pop_back(__dummy));
484 #endif
485 
486  for (_ThreadIndex __i = 0; __i < __num_threads; ++__i)
487  delete __tls[__i];
488  delete[] __tls;
489  }
490 } // namespace __gnu_parallel
491 
492 #endif /* _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H */