1 // -*- C++ -*- 2 3 // Copyright (C) 2007, 2008, 2009 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/random_shuffle.h 26 * @brief Parallel implementation of std::random_shuffle(). 27 * This file is a GNU parallel extension to the Standard C++ Library. 28 */ 29 30 // Written by Johannes Singler. 31 32 #ifndef _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H 33 #define _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H 1 34 35 #include <limits> 36 #include <bits/stl_numeric.h> 37 #include <parallel/parallel.h> 38 #include <parallel/random_number.h> 39 40 namespace __gnu_parallel 41 { 42 /** @brief Type to hold the index of a bin. 43 * 44 * Since many variables of this type are allocated, it should be 45 * chosen as small as possible. 46 */ 47 typedef unsigned short _BinIndex; 48 49 /** @brief Data known to every thread participating in 50 __gnu_parallel::__parallel_random_shuffle(). */ 51 template<typename _RAIter> 52 struct _DRandomShufflingGlobalData 53 { 54 typedef std::iterator_traits<_RAIter> _TraitsType; 55 typedef typename _TraitsType::value_type _ValueType; 56 typedef typename _TraitsType::difference_type _DifferenceType; 57 58 /** @brief Begin iterator of the __source. */ 59 _RAIter& _M_source; 60 61 /** @brief Temporary arrays for each thread. */ 62 _ValueType** _M_temporaries; 63 64 /** @brief Two-dimensional array to hold the thread-bin distribution. 65 * 66 * Dimensions (_M_num_threads + 1) __x (_M_num_bins + 1). */ 67 _DifferenceType** _M_dist; 68 69 /** @brief Start indexes of the threads' __chunks. */ 70 _DifferenceType* _M_starts; 71 72 /** @brief Number of the thread that will further process the 73 corresponding bin. */ 74 _ThreadIndex* _M_bin_proc; 75 76 /** @brief Number of bins to distribute to. */ 77 int _M_num_bins; 78 79 /** @brief Number of bits needed to address the bins. */ 80 int _M_num_bits; 81 82 /** @brief Constructor. */ 83 _DRandomShufflingGlobalData(_RAIter& __source) 84 : _M_source(__source) { } 85 }; 86 87 /** @brief Local data for a thread participating in 88 __gnu_parallel::__parallel_random_shuffle(). 89 */ 90 template<typename _RAIter, typename _RandomNumberGenerator> 91 struct _DRSSorterPU 92 { 93 /** @brief Number of threads participating in total. */ 94 int _M_num_threads; 95 96 /** @brief Begin index for bins taken care of by this thread. */ 97 _BinIndex _M_bins_begin; 98 99 /** @brief End index for bins taken care of by this thread. */ 100 _BinIndex __bins_end; 101 102 /** @brief Random _M_seed for this thread. */ 103 uint32_t _M_seed; 104 105 /** @brief Pointer to global data. */ 106 _DRandomShufflingGlobalData<_RAIter>* _M_sd; 107 }; 108 109 /** @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. 112 */ 113 template<typename _RandomNumberGenerator> 114 inline int 115 __random_number_pow2(int __logp, _RandomNumberGenerator& __rng) 116 { return __rng.__genrand_bits(__logp); } 117 118 /** @brief Random shuffle code executed by each thread. 119 * @param __pus Array of thread-local data records. */ 120 template<typename _RAIter, typename _RandomNumberGenerator> 121 void 122 __parallel_random_shuffle_drs_pu(_DRSSorterPU<_RAIter, 123 _RandomNumberGenerator>* __pus) 124 { 125 typedef std::iterator_traits<_RAIter> _TraitsType; 126 typedef typename _TraitsType::value_type _ValueType; 127 typedef typename _TraitsType::difference_type _DifferenceType; 128 129 _ThreadIndex __iam = omp_get_thread_num(); 130 _DRSSorterPU<_RAIter, _RandomNumberGenerator>* __d = &__pus[__iam]; 131 _DRandomShufflingGlobalData<_RAIter>* __sd = __d->_M_sd; 132 133 // 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]; 140 141 // Compute oracles and count appearances. 142 for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b) 143 __dist[__b] = 0; 144 int __num_bits = __sd->_M_num_bits; 145 146 _RandomNumber __rng(__d->_M_seed); 147 148 // First main loop. 149 for (_DifferenceType __i = 0; __i < __length; ++__i) 150 { 151 _BinIndex __oracle = __random_number_pow2(__num_bits, __rng); 152 __oracles[__i] = __oracle; 153 154 // To allow prefix (partial) sum. 155 ++(__dist[__oracle + 1]); 156 } 157 158 for (_BinIndex __b = 0; __b < __sd->_M_num_bins + 1; ++__b) 159 __sd->_M_dist[__b][__iam + 1] = __dist[__b]; 160 161 # pragma omp barrier 162 163 # pragma omp single 164 { 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]); 172 } 173 174 # pragma omp barrier 175 176 _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]; 179 180 # pragma omp barrier 181 182 for (_BinIndex __s = __d->_M_bins_begin; __s < __d->__bins_end; ++__s) 183 { 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]; 187 } 188 189 __sd->_M_temporaries[__iam] = static_cast<_ValueType*> 190 (::operator new(sizeof(_ValueType) * __offset)); 191 192 # pragma omp barrier 193 194 // 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]; 201 202 _RAIter __source = __sd->_M_source; 203 _DifferenceType __start = __sd->_M_starts[__iam]; 204 205 // Distribute according to oracles, second main loop. 206 for (_DifferenceType __i = 0; __i < __length; ++__i) 207 { 208 _BinIndex __target_bin = __oracles[__i]; 209 _ThreadIndex __target_p = __bin_proc[__target_bin]; 210 211 // Last column [__d->_M_num_threads] stays unchanged. 212 ::new(&(__temporaries[__target_p][__dist[__target_bin + 1]++])) 213 _ValueType(*(__source + __i + __start)); 214 } 215 216 delete[] __oracles; 217 delete[] __dist; 218 delete[] __bin_proc; 219 delete[] __temporaries; 220 221 # pragma omp barrier 222 223 // Shuffle bins internally. 224 for (_BinIndex __b = __d->_M_bins_begin; __b < __d->__bins_end; ++__b) 225 { 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]); 232 233 __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])); 237 } 238 239 ::operator delete(__sd->_M_temporaries[__iam]); 240 } 241 242 /** @brief Round up to the next greater power of 2. 243 * @param __x _Integer to round up */ 244 template<typename _Tp> 245 _Tp 246 __round_up_to_pow2(_Tp __x) 247 { 248 if (__x <= 1) 249 return 1; 250 else 251 return (_Tp)1 << (__rd_log2(__x - 1) + 1); 252 } 253 254 /** @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. 260 */ 261 template<typename _RAIter, typename _RandomNumberGenerator> 262 void 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) 268 { 269 typedef std::iterator_traits<_RAIter> _TraitsType; 270 typedef typename _TraitsType::value_type _ValueType; 271 typedef typename _TraitsType::difference_type _DifferenceType; 272 273 _GLIBCXX_CALL(__n) 274 275 const _Settings& __s = _Settings::get(); 276 277 if (__num_threads > __n) 278 __num_threads = static_cast<_ThreadIndex>(__n); 279 280 _BinIndex __num_bins, __num_bins_cache; 281 282 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 283 // Try the L1 cache first. 284 285 // Must fit into L1. 286 __num_bins_cache = 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); 290 291 // 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); 294 295 #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); 298 #endif 299 __num_bins = __round_up_to_pow2(__num_bins); 300 301 if (__num_bins < __num_bins_cache) 302 { 303 #endif 304 // Now try the L2 cache 305 // Must fit into L2 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); 310 311 // 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), 318 __num_bins); 319 #endif 320 __num_bins = __round_up_to_pow2(__num_bins); 321 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 322 } 323 #endif 324 325 __num_bins = __round_up_to_pow2( 326 std::max<_BinIndex>(__num_threads, __num_bins)); 327 328 if (__num_threads <= 1) 329 { 330 _RandomNumber __derived_rng( 331 __rng(std::numeric_limits<uint32_t>::max())); 332 __sequential_random_shuffle(__begin, __end, __derived_rng); 333 return; 334 } 335 336 _DRandomShufflingGlobalData<_RAIter> __sd(__begin); 337 _DRSSorterPU<_RAIter, _RandomNumber >* __pus; 338 _DifferenceType* __starts; 339 340 # pragma omp parallel num_threads(__num_threads) 341 { 342 _ThreadIndex __num_threads = omp_get_num_threads(); 343 # pragma omp single 344 { 345 __pus = new _DRSSorterPU<_RAIter, _RandomNumber>[__num_threads]; 346 347 __sd._M_temporaries = new _ValueType*[__num_threads]; 348 __sd._M_dist = new _DifferenceType*[__num_bins + 1]; 349 __sd._M_bin_proc = new _ThreadIndex[__num_bins]; 350 for (_BinIndex __b = 0; __b < __num_bins + 1; ++__b) 351 __sd._M_dist[__b] = new _DifferenceType[__num_threads + 1]; 352 for (_BinIndex __b = 0; __b < (__num_bins + 1); ++__b) 353 { 354 __sd._M_dist[0][0] = 0; 355 __sd._M_dist[__b][0] = 0; 356 } 357 __starts = __sd._M_starts = new _DifferenceType[__num_threads + 1]; 358 int __bin_cursor = 0; 359 __sd._M_num_bins = __num_bins; 360 __sd._M_num_bits = __rd_log2(__num_bins); 361 362 _DifferenceType __chunk_length = __n / __num_threads, 363 __split = __n % __num_threads, 364 __start = 0; 365 _DifferenceType __bin_chunk_length = __num_bins / __num_threads, 366 __bin_split = __num_bins % __num_threads; 367 for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) 368 { 369 __starts[__i] = __start; 370 __start += (__i < __split 371 ? (__chunk_length + 1) : __chunk_length); 372 int __j = __pus[__i]._M_bins_begin = __bin_cursor; 373 374 // Range of bins for this processor. 375 __bin_cursor += (__i < __bin_split 376 ? (__bin_chunk_length + 1) 377 : __bin_chunk_length); 378 __pus[__i].__bins_end = __bin_cursor; 379 for (; __j < __bin_cursor; ++__j) 380 __sd._M_bin_proc[__j] = __i; 381 __pus[__i]._M_num_threads = __num_threads; 382 __pus[__i]._M_seed = __rng(std::numeric_limits<uint32_t>::max()); 383 __pus[__i]._M_sd = &__sd; 384 } 385 __starts[__num_threads] = __start; 386 } //single 387 // Now shuffle in parallel. 388 __parallel_random_shuffle_drs_pu(__pus); 389 } // parallel 390 391 delete[] __starts; 392 delete[] __sd._M_bin_proc; 393 for (int __s = 0; __s < (__num_bins + 1); ++__s) 394 delete[] __sd._M_dist[__s]; 395 delete[] __sd._M_dist; 396 delete[] __sd._M_temporaries; 397 398 delete[] __pus; 399 } 400 401 /** @brief Sequential cache-efficient random shuffle. 402 * @param __begin Begin iterator of sequence. 403 * @param __end End iterator of sequence. 404 * @param __rng Random number generator to use. 405 */ 406 template<typename _RAIter, typename _RandomNumberGenerator> 407 void 408 __sequential_random_shuffle(_RAIter __begin, _RAIter __end, 409 _RandomNumberGenerator& __rng) 410 { 411 typedef std::iterator_traits<_RAIter> _TraitsType; 412 typedef typename _TraitsType::value_type _ValueType; 413 typedef typename _TraitsType::difference_type _DifferenceType; 414 415 _DifferenceType __n = __end - __begin; 416 const _Settings& __s = _Settings::get(); 417 418 _BinIndex __num_bins, __num_bins_cache; 419 420 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 421 // Try the L1 cache first, must fit into L1. 422 __num_bins_cache = std::max<_DifferenceType> 423 (1, __n / (__s.L1_cache_size_lb / sizeof(_ValueType))); 424 __num_bins_cache = __round_up_to_pow2(__num_bins_cache); 425 426 // No more buckets than TLB entries, power of 2 427 // Power of 2 and at least one element per bin, at most the TLB size 428 __num_bins = std::min(__n, (_DifferenceType)__num_bins_cache); 429 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB 430 // 2 TLB entries needed per bin 431 __num_bins = std::min((_DifferenceType)__s.TLB_size / 2, __num_bins); 432 #endif 433 __num_bins = __round_up_to_pow2(__num_bins); 434 435 if (__num_bins < __num_bins_cache) 436 { 437 #endif 438 // Now try the L2 cache, must fit into L2. 439 __num_bins_cache = static_cast<_BinIndex> 440 (std::max<_DifferenceType>(1, __n / (__s.L2_cache_size 441 / sizeof(_ValueType)))); 442 __num_bins_cache = __round_up_to_pow2(__num_bins_cache); 443 444 // No more buckets than TLB entries, power of 2 445 // Power of 2 and at least one element per bin, at most the TLB size. 446 __num_bins = static_cast<_BinIndex> 447 (std::min(__n, static_cast<_DifferenceType>(__num_bins_cache))); 448 449 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_TLB 450 // 2 TLB entries needed per bin 451 __num_bins = std::min<_DifferenceType>(__s.TLB_size / 2, __num_bins); 452 #endif 453 __num_bins = __round_up_to_pow2(__num_bins); 454 #if _GLIBCXX_RANDOM_SHUFFLE_CONSIDER_L1 455 } 456 #endif 457 458 int __num_bits = __rd_log2(__num_bins); 459 460 if (__num_bins > 1) 461 { 462 _ValueType* __target = 463 static_cast<_ValueType*>(::operator new(sizeof(_ValueType) * __n)); 464 _BinIndex* __oracles = new _BinIndex[__n]; 465 _DifferenceType* __dist0 = new _DifferenceType[__num_bins + 1], 466 * __dist1 = new _DifferenceType[__num_bins + 1]; 467 468 for (int __b = 0; __b < __num_bins + 1; ++__b) 469 __dist0[__b] = 0; 470 471 _RandomNumber __bitrng(__rng(0xFFFFFFFF)); 472 473 for (_DifferenceType __i = 0; __i < __n; ++__i) 474 { 475 _BinIndex __oracle = __random_number_pow2(__num_bits, __bitrng); 476 __oracles[__i] = __oracle; 477 478 // To allow prefix (partial) sum. 479 ++(__dist0[__oracle + 1]); 480 } 481 482 // Sum up bins. 483 __gnu_sequential::partial_sum(__dist0, __dist0 + __num_bins + 1, 484 __dist0); 485 486 for (int __b = 0; __b < __num_bins + 1; ++__b) 487 __dist1[__b] = __dist0[__b]; 488 489 // Distribute according to oracles. 490 for (_DifferenceType __i = 0; __i < __n; ++__i) 491 ::new(&(__target[(__dist0[__oracles[__i]])++])) 492 _ValueType(*(__begin + __i)); 493 494 for (int __b = 0; __b < __num_bins; ++__b) 495 __sequential_random_shuffle(__target + __dist1[__b], 496 __target + __dist1[__b + 1], __rng); 497 498 // Copy elements back. 499 std::copy(__target, __target + __n, __begin); 500 501 delete[] __dist0; 502 delete[] __dist1; 503 delete[] __oracles; 504 ::operator delete(__target); 505 } 506 else 507 __gnu_sequential::random_shuffle(__begin, __end, __rng); 508 } 509 510 /** @brief Parallel random public call. 511 * @param __begin Begin iterator of sequence. 512 * @param __end End iterator of sequence. 513 * @param __rng Random number generator to use. 514 */ 515 template<typename _RAIter, typename _RandomNumberGenerator> 516 inline void 517 __parallel_random_shuffle(_RAIter __begin, _RAIter __end, 518 _RandomNumberGenerator __rng = _RandomNumber()) 519 { 520 typedef std::iterator_traits<_RAIter> _TraitsType; 521 typedef typename _TraitsType::difference_type _DifferenceType; 522 _DifferenceType __n = __end - __begin; 523 __parallel_random_shuffle_drs(__begin, __end, __n, 524 __get_max_threads(), __rng); 525 } 526 } 527 528 #endif /* _GLIBCXX_PARALLEL_RANDOM_SHUFFLE_H */ 529