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
>* __sd
= __d
->_M_sd
;
133 // Indexing: _M_dist[bin][processor]
134 _DifferenceType __length
= (__sd
->_M_starts
[__iam
+ 1]
135 - __sd
->_M_starts
[__iam
]);
136 _BinIndex
* __oracles
= new _BinIndex
[__length
];
137 _DifferenceType
* __dist
= new _DifferenceType
[__sd
->_M_num_bins
+ 1];
138 _BinIndex
* __bin_proc
= new _BinIndex
[__sd
->_M_num_bins
];
139 _ValueType
** __temporaries
= new _ValueType
*[__d
->_M_num_threads
];
141 // Compute oracles and count appearances.
142 for (_BinIndex __b
= 0; __b
< __sd
->_M_num_bins
+ 1; ++__b
)
144 int __num_bits
= __sd
->_M_num_bits
;
146 _RandomNumber
__rng(__d
->_M_seed
);
149 for (_DifferenceType __i
= 0; __i
< __length
; ++__i
)
151 _BinIndex __oracle
= __random_number_pow2(__num_bits
, __rng
);
152 __oracles
[__i
] = __oracle
;
154 // To allow prefix (partial) sum.
155 ++(__dist
[__oracle
+ 1]);
158 for (_BinIndex __b
= 0; __b
< __sd
->_M_num_bins
+ 1; ++__b
)
159 __sd
->_M_dist
[__b
][__iam
+ 1] = __dist
[__b
];
165 // Sum up bins, __sd->_M_dist[__s + 1][__d->_M_num_threads] now
166 // contains the total number of items in bin __s
167 for (_BinIndex __s
= 0; __s
< __sd
->_M_num_bins
; ++__s
)
168 __gnu_sequential::partial_sum(__sd
->_M_dist
[__s
+ 1],
169 __sd
->_M_dist
[__s
+ 1]
170 + __d
->_M_num_threads
+ 1,
171 __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
+= __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 __sd
->_M_dist
[__s
+ 1][__t
] += __offset
;
186 __offset
= __sd
->_M_dist
[__s
+ 1][__d
->_M_num_threads
];
189 __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
< __sd
->_M_num_bins
+ 1; ++__b
)
196 __dist
[__b
] = __sd
->_M_dist
[__b
][__iam
];
197 for (_BinIndex __b
= 0; __b
< __sd
->_M_num_bins
; ++__b
)
198 __bin_proc
[__b
] = __sd
->_M_bin_proc
[__b
];
199 for (_ThreadIndex __t
= 0; __t
< __d
->_M_num_threads
; ++__t
)
200 __temporaries
[__t
] = __sd
->_M_temporaries
[__t
];
202 _RAIter __source
= __sd
->_M_source
;
203 _DifferenceType __start
= __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
= __bin_proc
[__target_bin
];
211 // Last column [__d->_M_num_threads] stays unchanged.
212 ::new(&(__temporaries
[__target_p
][__dist
[__target_bin
+ 1]++]))
213 _ValueType(*(__source
+ __i
+ __start
));
219 delete[] __temporaries
;
223 // Shuffle bins internally.
224 for (_BinIndex __b
= __d
->_M_bins_begin
; __b
< __d
->__bins_end
; ++__b
)
226 _ValueType
* __begin
=
227 (__sd
->_M_temporaries
[__iam
]
228 + (__b
== __d
->_M_bins_begin
229 ? 0 : __sd
->_M_dist
[__b
][__d
->_M_num_threads
])),
230 * __end
= (__sd
->_M_temporaries
[__iam
]
231 + __sd
->_M_dist
[__b
+ 1][__d
->_M_num_threads
]);
233 __sequential_random_shuffle(__begin
, __end
, __rng
);
234 std::copy(__begin
, __end
, __sd
->_M_source
+ __global_offset
235 + (__b
== __d
->_M_bins_begin
236 ? 0 : __sd
->_M_dist
[__b
][__d
->_M_num_threads
]));
239 ::operator delete(__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
, _RAIter __end
,
264 typename
std::iterator_traits
265 <_RAIter
>::difference_type __n
,
266 _ThreadIndex __num_threads
,
267 _RandomNumberGenerator
& __rng
)
269 typedef std::iterator_traits
<_RAIter
> _TraitsType
;
270 typedef typename
_TraitsType::value_type _ValueType
;
271 typedef typename
_TraitsType::difference_type _DifferenceType
;
275 const _Settings
& __s
= _Settings::get();
277 if (__num_threads
> __n
)
278 __num_threads
= static_cast<_ThreadIndex
>(__n
);
280 _BinIndex __num_bins
, __num_bins_cache
;
282 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
283 // Try the L1 cache first.
287 std::max
<_DifferenceType
>(1, __n
/ (__s
.L1_cache_size_lb
288 / 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 __num_bins
= std::min
<_DifferenceType
>(__n
, __num_bins_cache
);
295 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
296 // 2 TLB entries needed per bin.
297 __num_bins
= std::min
<_DifferenceType
>(__s
.TLB_size
/ 2, __num_bins
);
299 __num_bins
= __round_up_to_pow2(__num_bins
);
301 if (__num_bins
< __num_bins_cache
)
304 // Now try the L2 cache
306 __num_bins_cache
= static_cast<_BinIndex
>
307 (std::max
<_DifferenceType
>(1, __n
/ (__s
.L2_cache_size
308 / sizeof(_ValueType
))));
309 __num_bins_cache
= __round_up_to_pow2(__num_bins_cache
);
311 // No more buckets than TLB entries, power of 2.
312 __num_bins
= static_cast<_BinIndex
>
313 (std::min(__n
, static_cast<_DifferenceType
>(__num_bins_cache
)));
314 // Power of 2 and at least one element per bin, at most the TLB size.
315 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
316 // 2 TLB entries needed per bin.
317 __num_bins
= std::min(static_cast<_DifferenceType
>(__s
.TLB_size
/ 2),
320 __num_bins
= __round_up_to_pow2(__num_bins
);
321 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
325 __num_threads
= std::min
<_BinIndex
>(__num_threads
, __num_bins
);
327 if (__num_threads
<= 1)
328 return __sequential_random_shuffle(__begin
, __end
, __rng
);
330 _DRandomShufflingGlobalData
<_RAIter
> __sd(__begin
);
331 _DRSSorterPU
<_RAIter
, _RandomNumber
>* __pus
;
332 _DifferenceType
* __starts
;
334 # pragma omp parallel num_threads(__num_threads)
336 _ThreadIndex __num_threads
= omp_get_num_threads();
339 __pus
= new _DRSSorterPU
<_RAIter
, _RandomNumber
>[__num_threads
];
341 __sd
._M_temporaries
= new _ValueType
*[__num_threads
];
342 __sd
._M_dist
= new _DifferenceType
*[__num_bins
+ 1];
343 __sd
._M_bin_proc
= new _ThreadIndex
[__num_bins
];
344 for (_BinIndex __b
= 0; __b
< __num_bins
+ 1; ++__b
)
345 __sd
._M_dist
[__b
] = new _DifferenceType
[__num_threads
+ 1];
346 for (_BinIndex __b
= 0; __b
< (__num_bins
+ 1); ++__b
)
348 __sd
._M_dist
[0][0] = 0;
349 __sd
._M_dist
[__b
][0] = 0;
351 __starts
= __sd
._M_starts
= new _DifferenceType
[__num_threads
+ 1];
352 int __bin_cursor
= 0;
353 __sd
._M_num_bins
= __num_bins
;
354 __sd
._M_num_bits
= __rd_log2(__num_bins
);
356 _DifferenceType __chunk_length
= __n
/ __num_threads
,
357 __split
= __n
% __num_threads
,
359 _DifferenceType __bin_chunk_length
= __num_bins
/ __num_threads
,
360 __bin_split
= __num_bins
% __num_threads
;
361 for (_ThreadIndex __i
= 0; __i
< __num_threads
; ++__i
)
363 __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)
371 : __bin_chunk_length
);
372 __pus
[__i
].__bins_end
= __bin_cursor
;
373 for (; __j
< __bin_cursor
; ++__j
)
374 __sd
._M_bin_proc
[__j
] = __i
;
375 __pus
[__i
]._M_num_threads
= __num_threads
;
376 __pus
[__i
]._M_seed
= __rng(std::numeric_limits
<uint32_t>::max());
377 __pus
[__i
]._M_sd
= &__sd
;
379 __starts
[__num_threads
] = __start
;
381 // Now shuffle in parallel.
382 __parallel_random_shuffle_drs_pu(__pus
);
386 delete[] __sd
._M_bin_proc
;
387 for (int __s
= 0; __s
< (__num_bins
+ 1); ++__s
)
388 delete[] __sd
._M_dist
[__s
];
389 delete[] __sd
._M_dist
;
390 delete[] __sd
._M_temporaries
;
395 /** @brief Sequential cache-efficient random shuffle.
396 * @param __begin Begin iterator of sequence.
397 * @param __end End iterator of sequence.
398 * @param __rng Random number generator to use.
400 template<typename _RAIter
, typename _RandomNumberGenerator
>
402 __sequential_random_shuffle(_RAIter __begin
, _RAIter __end
,
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 __num_bins
, __num_bins_cache
;
414 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
415 // Try the L1 cache first, must fit into L1.
416 __num_bins_cache
= std::max
<_DifferenceType
>
417 (1, __n
/ (__s
.L1_cache_size_lb
/ sizeof(_ValueType
)));
418 __num_bins_cache
= __round_up_to_pow2(__num_bins_cache
);
420 // No more buckets than TLB entries, power of 2
421 // Power of 2 and at least one element per bin, at most the TLB size
422 __num_bins
= std::min(__n
, (_DifferenceType
)__num_bins_cache
);
423 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
424 // 2 TLB entries needed per bin
425 __num_bins
= std::min((_DifferenceType
)__s
.TLB_size
/ 2, __num_bins
);
427 __num_bins
= __round_up_to_pow2(__num_bins
);
429 if (__num_bins
< __num_bins_cache
)
432 // Now try the L2 cache, must fit into L2.
433 __num_bins_cache
= static_cast<_BinIndex
>
434 (std::max
<_DifferenceType
>(1, __n
/ (__s
.L2_cache_size
435 / sizeof(_ValueType
))));
436 __num_bins_cache
= __round_up_to_pow2(__num_bins_cache
);
438 // No more buckets than TLB entries, power of 2
439 // Power of 2 and at least one element per bin, at most the TLB size.
440 __num_bins
= static_cast<_BinIndex
>
441 (std::min(__n
, static_cast<_DifferenceType
>(__num_bins_cache
)));
443 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB
444 // 2 TLB entries needed per bin
445 __num_bins
= std::min
<_DifferenceType
>(__s
.TLB_size
/ 2, __num_bins
);
447 __num_bins
= __round_up_to_pow2(__num_bins
);
448 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1
452 int __num_bits
= __rd_log2(__num_bins
);
456 _ValueType
* __target
=
457 static_cast<_ValueType
*>(::operator new(sizeof(_ValueType
) * __n
));
458 _BinIndex
* __oracles
= new _BinIndex
[__n
];
459 _DifferenceType
* __dist0
= new _DifferenceType
[__num_bins
+ 1],
460 * __dist1
= new _DifferenceType
[__num_bins
+ 1];
462 for (int __b
= 0; __b
< __num_bins
+ 1; ++__b
)
465 _RandomNumber
__bitrng(__rng(0xFFFFFFFF));
467 for (_DifferenceType __i
= 0; __i
< __n
; ++__i
)
469 _BinIndex __oracle
= __random_number_pow2(__num_bits
, __bitrng
);
470 __oracles
[__i
] = __oracle
;
472 // To allow prefix (partial) sum.
473 ++(__dist0
[__oracle
+ 1]);
477 __gnu_sequential::partial_sum(__dist0
, __dist0
+ __num_bins
+ 1,
480 for (int __b
= 0; __b
< __num_bins
+ 1; ++__b
)
481 __dist1
[__b
] = __dist0
[__b
];
483 // Distribute according to oracles.
484 for (_DifferenceType __i
= 0; __i
< __n
; ++__i
)
485 ::new(&(__target
[(__dist0
[__oracles
[__i
]])++]))
486 _ValueType(*(__begin
+ __i
));
488 for (int __b
= 0; __b
< __num_bins
; ++__b
)
489 __sequential_random_shuffle(__target
+ __dist1
[__b
],
490 __target
+ __dist1
[__b
+ 1], __rng
);
492 // Copy elements back.
493 std::copy(__target
, __target
+ __n
, __begin
);
498 ::operator delete(__target
);
501 __gnu_sequential::random_shuffle(__begin
, __end
, __rng
);
504 /** @brief Parallel random public call.
505 * @param __begin Begin iterator of sequence.
506 * @param __end End iterator of sequence.
507 * @param __rng Random number generator to use.
509 template<typename _RAIter
, typename _RandomNumberGenerator
>
511 __parallel_random_shuffle(_RAIter __begin
, _RAIter __end
,
512 _RandomNumberGenerator __rng
= _RandomNumber())
514 typedef std::iterator_traits
<_RAIter
> _TraitsType
;
515 typedef typename
_TraitsType::difference_type _DifferenceType
;
516 _DifferenceType __n
= __end
- __begin
;
517 __parallel_random_shuffle_drs(__begin
, __end
, __n
,
518 __get_max_threads(), __rng
);
522 #endif /* _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H */