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/multiway_mergesort.h
26 * @brief Parallel multiway merge sort.
27 * This file is a GNU parallel extension to the Standard C++ Library.
30 // Written by Johannes Singler.
32 #ifndef _GLIBCXX_PARALLEL_MULTIWAY_MERGESORT_H
33 #define _GLIBCXX_PARALLEL_MULTIWAY_MERGESORT_H 1
37 #include <parallel/basic_iterator.h>
38 #include <bits/stl_algo.h>
39 #include <parallel/parallel.h>
40 #include <parallel/multiway_merge.h>
42 namespace __gnu_parallel
45 /** @brief Subsequence description. */
46 template<typename _DifferenceTp
>
49 typedef _DifferenceTp difference_type
;
51 /** @brief Begin of subsequence. */
52 difference_type begin
;
54 /** @brief End of subsequence. */
58 /** @brief Data accessed by all threads.
60 * PMWMS = parallel multiway mergesort */
61 template<typename RandomAccessIterator
>
62 struct PMWMSSortingData
64 typedef std::iterator_traits
<RandomAccessIterator
> traits_type
;
65 typedef typename
traits_type::value_type value_type
;
66 typedef typename
traits_type::difference_type difference_type
;
68 /** @brief Number of threads involved. */
69 thread_index_t num_threads
;
71 /** @brief Input begin. */
72 RandomAccessIterator source
;
74 /** @brief Start indices, per thread. */
75 difference_type
* starts
;
77 /** @brief Storage in which to sort. */
78 value_type
** temporary
;
80 /** @brief Samples. */
83 /** @brief Offsets to add to the found positions. */
84 difference_type
* offsets
;
86 /** @brief Pieces of data to merge @c [thread][sequence] */
87 std::vector
<Piece
<difference_type
> >* pieces
;
91 * @brief Select samples from a sequence.
92 * @param sd Pointer to algorithm data. Result will be placed in
94 * @param num_samples Number of samples to select.
96 template<typename RandomAccessIterator
, typename _DifferenceTp
>
98 determine_samples(PMWMSSortingData
<RandomAccessIterator
>* sd
,
99 _DifferenceTp num_samples
)
101 typedef std::iterator_traits
<RandomAccessIterator
> traits_type
;
102 typedef typename
traits_type::value_type value_type
;
103 typedef _DifferenceTp difference_type
;
105 thread_index_t iam
= omp_get_thread_num();
107 difference_type
* es
= new difference_type
[num_samples
+ 2];
109 equally_split(sd
->starts
[iam
+ 1] - sd
->starts
[iam
],
110 num_samples
+ 1, es
);
112 for (difference_type i
= 0; i
< num_samples
; ++i
)
113 ::new(&(sd
->samples
[iam
* num_samples
+ i
]))
114 value_type(sd
->source
[sd
->starts
[iam
] + es
[i
+ 1]]);
119 /** @brief Split consistently. */
120 template<bool exact
, typename RandomAccessIterator
,
121 typename Comparator
, typename SortingPlacesIterator
>
122 struct split_consistently
126 /** @brief Split by exact splitting. */
127 template<typename RandomAccessIterator
, typename Comparator
,
128 typename SortingPlacesIterator
>
129 struct split_consistently
130 <true, RandomAccessIterator
, Comparator
, SortingPlacesIterator
>
133 const thread_index_t iam
,
134 PMWMSSortingData
<RandomAccessIterator
>* sd
,
137 std::iterator_traits
<RandomAccessIterator
>::difference_type
143 std::vector
<std::pair
<SortingPlacesIterator
, SortingPlacesIterator
> >
144 seqs(sd
->num_threads
);
145 for (thread_index_t s
= 0; s
< sd
->num_threads
; s
++)
146 seqs
[s
] = std::make_pair(sd
->temporary
[s
],
148 + (sd
->starts
[s
+ 1] - sd
->starts
[s
]));
150 std::vector
<SortingPlacesIterator
> offsets(sd
->num_threads
);
152 // if not last thread
153 if (iam
< sd
->num_threads
- 1)
154 multiseq_partition(seqs
.begin(), seqs
.end(),
155 sd
->starts
[iam
+ 1], offsets
.begin(), comp
);
157 for (int seq
= 0; seq
< sd
->num_threads
; seq
++)
160 if (iam
< (sd
->num_threads
- 1))
161 sd
->pieces
[iam
][seq
].end
= offsets
[seq
] - seqs
[seq
].first
;
163 // very end of this sequence
164 sd
->pieces
[iam
][seq
].end
=
165 sd
->starts
[seq
+ 1] - sd
->starts
[seq
];
170 for (thread_index_t seq
= 0; seq
< sd
->num_threads
; seq
++)
172 // For each sequence.
174 sd
->pieces
[iam
][seq
].begin
= sd
->pieces
[iam
- 1][seq
].end
;
176 // Absolute beginning.
177 sd
->pieces
[iam
][seq
].begin
= 0;
182 /** @brief Split by sampling. */
183 template<typename RandomAccessIterator
, typename Comparator
,
184 typename SortingPlacesIterator
>
185 struct split_consistently
<false, RandomAccessIterator
, Comparator
,
186 SortingPlacesIterator
>
189 const thread_index_t iam
,
190 PMWMSSortingData
<RandomAccessIterator
>* sd
,
193 std::iterator_traits
<RandomAccessIterator
>::difference_type
197 typedef std::iterator_traits
<RandomAccessIterator
> traits_type
;
198 typedef typename
traits_type::value_type value_type
;
199 typedef typename
traits_type::difference_type difference_type
;
201 determine_samples(sd
, num_samples
);
206 __gnu_sequential::sort(sd
->samples
,
207 sd
->samples
+ (num_samples
* sd
->num_threads
),
212 for (thread_index_t s
= 0; s
< sd
->num_threads
; ++s
)
214 // For each sequence.
215 if (num_samples
* iam
> 0)
216 sd
->pieces
[iam
][s
].begin
=
217 std::lower_bound(sd
->temporary
[s
],
219 + (sd
->starts
[s
+ 1] - sd
->starts
[s
]),
220 sd
->samples
[num_samples
* iam
],
224 // Absolute beginning.
225 sd
->pieces
[iam
][s
].begin
= 0;
227 if ((num_samples
* (iam
+ 1)) < (num_samples
* sd
->num_threads
))
228 sd
->pieces
[iam
][s
].end
=
229 std::lower_bound(sd
->temporary
[s
],
231 + (sd
->starts
[s
+ 1] - sd
->starts
[s
]),
232 sd
->samples
[num_samples
* (iam
+ 1)],
237 sd
->pieces
[iam
][s
].end
= sd
->starts
[s
+ 1] - sd
->starts
[s
];
242 template<bool stable
, typename RandomAccessIterator
, typename Comparator
>
243 struct possibly_stable_sort
247 template<typename RandomAccessIterator
, typename Comparator
>
248 struct possibly_stable_sort
<true, RandomAccessIterator
, Comparator
>
250 void operator()(const RandomAccessIterator
& begin
,
251 const RandomAccessIterator
& end
, Comparator
& comp
) const
253 __gnu_sequential::stable_sort(begin
, end
, comp
);
257 template<typename RandomAccessIterator
, typename Comparator
>
258 struct possibly_stable_sort
<false, RandomAccessIterator
, Comparator
>
260 void operator()(const RandomAccessIterator begin
,
261 const RandomAccessIterator end
, Comparator
& comp
) const
263 __gnu_sequential::sort(begin
, end
, comp
);
267 template<bool stable
, typename SeqRandomAccessIterator
,
268 typename RandomAccessIterator
, typename Comparator
,
270 struct possibly_stable_multiway_merge
274 template<typename SeqRandomAccessIterator
, typename RandomAccessIterator
,
275 typename Comparator
, typename DiffType
>
276 struct possibly_stable_multiway_merge
277 <true, SeqRandomAccessIterator
, RandomAccessIterator
, Comparator
,
280 void operator()(const SeqRandomAccessIterator
& seqs_begin
,
281 const SeqRandomAccessIterator
& seqs_end
,
282 const RandomAccessIterator
& target
,
284 DiffType length_am
) const
286 stable_multiway_merge(seqs_begin
, seqs_end
, target
, length_am
, comp
,
291 template<typename SeqRandomAccessIterator
, typename RandomAccessIterator
,
292 typename Comparator
, typename DiffType
>
293 struct possibly_stable_multiway_merge
294 <false, SeqRandomAccessIterator
, RandomAccessIterator
, Comparator
,
297 void operator()(const SeqRandomAccessIterator
& seqs_begin
,
298 const SeqRandomAccessIterator
& seqs_end
,
299 const RandomAccessIterator
& target
,
301 DiffType length_am
) const
303 multiway_merge(seqs_begin
, seqs_end
, target
, length_am
, comp
,
308 /** @brief PMWMS code executed by each thread.
309 * @param sd Pointer to algorithm data.
310 * @param comp Comparator.
312 template<bool stable
, bool exact
, typename RandomAccessIterator
,
315 parallel_sort_mwms_pu(PMWMSSortingData
<RandomAccessIterator
>* sd
,
318 typedef std::iterator_traits
<RandomAccessIterator
> traits_type
;
319 typedef typename
traits_type::value_type value_type
;
320 typedef typename
traits_type::difference_type difference_type
;
322 thread_index_t iam
= omp_get_thread_num();
324 // Length of this thread's chunk, before merging.
325 difference_type length_local
= sd
->starts
[iam
+ 1] - sd
->starts
[iam
];
327 // Sort in temporary storage, leave space for sentinel.
329 typedef value_type
* SortingPlacesIterator
;
332 static_cast<value_type
*>(
333 ::operator new(sizeof(value_type
) * (length_local
+ 1)));
336 std::uninitialized_copy(sd
->source
+ sd
->starts
[iam
],
337 sd
->source
+ sd
->starts
[iam
] + length_local
,
340 possibly_stable_sort
<stable
, SortingPlacesIterator
, Comparator
>()
341 (sd
->temporary
[iam
], sd
->temporary
[iam
] + length_local
, comp
);
343 // Invariant: locally sorted subsequence in sd->temporary[iam],
344 // sd->temporary[iam] + length_local.
346 // No barrier here: Synchronization is done by the splitting routine.
348 difference_type num_samples
=
349 _Settings::get().sort_mwms_oversampling
* sd
->num_threads
- 1;
351 <exact
, RandomAccessIterator
, Comparator
, SortingPlacesIterator
>()
352 (iam
, sd
, comp
, num_samples
);
354 // Offset from target begin, length after merging.
355 difference_type offset
= 0, length_am
= 0;
356 for (thread_index_t s
= 0; s
< sd
->num_threads
; s
++)
358 length_am
+= sd
->pieces
[iam
][s
].end
- sd
->pieces
[iam
][s
].begin
;
359 offset
+= sd
->pieces
[iam
][s
].begin
;
363 std::pair
<SortingPlacesIterator
, SortingPlacesIterator
> >
365 seq_vector_type
seqs(sd
->num_threads
);
367 for (int s
= 0; s
< sd
->num_threads
; ++s
)
370 std::make_pair(sd
->temporary
[s
] + sd
->pieces
[iam
][s
].begin
,
371 sd
->temporary
[s
] + sd
->pieces
[iam
][s
].end
);
374 possibly_stable_multiway_merge
<
376 typename
seq_vector_type::iterator
,
377 RandomAccessIterator
,
378 Comparator
, difference_type
>()
379 (seqs
.begin(), seqs
.end(),
380 sd
->source
+ offset
, comp
,
385 ::operator delete(sd
->temporary
[iam
]);
388 /** @brief PMWMS main call.
389 * @param begin Begin iterator of sequence.
390 * @param end End iterator of sequence.
391 * @param comp Comparator.
392 * @param n Length of sequence.
393 * @param num_threads Number of threads to use.
395 template<bool stable
, bool exact
, typename RandomAccessIterator
,
398 parallel_sort_mwms(RandomAccessIterator begin
, RandomAccessIterator end
,
400 thread_index_t num_threads
)
402 _GLIBCXX_CALL(end
- begin
)
404 typedef std::iterator_traits
<RandomAccessIterator
> traits_type
;
405 typedef typename
traits_type::value_type value_type
;
406 typedef typename
traits_type::difference_type difference_type
;
408 difference_type n
= end
- begin
;
413 // at least one element per thread
415 num_threads
= static_cast<thread_index_t
>(n
);
418 PMWMSSortingData
<RandomAccessIterator
> sd
;
419 difference_type
* starts
;
421 # pragma omp parallel num_threads(num_threads)
423 num_threads
= omp_get_num_threads(); //no more threads than requested
427 sd
.num_threads
= num_threads
;
430 sd
.temporary
= new value_type
*[num_threads
];
434 difference_type size
=
435 (_Settings::get().sort_mwms_oversampling
* num_threads
- 1)
437 sd
.samples
= static_cast<value_type
*>(
438 ::operator new(size
* sizeof(value_type
)));
443 sd
.offsets
= new difference_type
[num_threads
- 1];
444 sd
.pieces
= new std::vector
<Piece
<difference_type
> >[num_threads
];
445 for (int s
= 0; s
< num_threads
; ++s
)
446 sd
.pieces
[s
].resize(num_threads
);
447 starts
= sd
.starts
= new difference_type
[num_threads
+ 1];
449 difference_type chunk_length
= n
/ num_threads
;
450 difference_type split
= n
% num_threads
;
451 difference_type pos
= 0;
452 for (int i
= 0; i
< num_threads
; ++i
)
455 pos
+= (i
< split
) ? (chunk_length
+ 1) : chunk_length
;
457 starts
[num_threads
] = pos
;
460 // Now sort in parallel.
461 parallel_sort_mwms_pu
<stable
, exact
>(&sd
, comp
);
465 delete[] sd
.temporary
;
468 ::operator delete(sd
.samples
);
473 } //namespace __gnu_parallel
475 #endif /* _GLIBCXX_PARALLEL_MULTIWAY_MERGESORT_H */