00001 // -*- C++ -*- 00002 00003 // Copyright (C) 2007, 2008, 2009 Free Software Foundation, Inc. 00004 // 00005 // This file is part of the GNU ISO C++ Library. This library is free 00006 // software; you can redistribute it and/or modify it under the terms 00007 // of the GNU General Public License as published by the Free Software 00008 // Foundation; either version 3, or (at your option) any later 00009 // version. 00010 00011 // This library is distributed in the hope that it will be useful, but 00012 // WITHOUT ANY WARRANTY; without even the implied warranty of 00013 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 00014 // General Public License for more details. 00015 00016 // Under Section 7 of GPL version 3, you are granted additional 00017 // permissions described in the GCC Runtime Library Exception, version 00018 // 3.1, as published by the Free Software Foundation. 00019 00020 // You should have received a copy of the GNU General Public License and 00021 // a copy of the GCC Runtime Library Exception along with this program; 00022 // see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 00023 // <http://www.gnu.org/licenses/>. 00024 00025 /** @file parallel/settings.h 00026 * @brief Runtime settings and tuning parameters, heuristics to decide 00027 * whether to use parallelized algorithms. 00028 * This file is a GNU parallel extension to the Standard C++ Library. 00029 * 00030 * @section parallelization_decision 00031 * The decision whether to run an algorithm in parallel. 00032 * 00033 * There are several ways the user can switch on and off the parallel 00034 * execution of an algorithm, both at compile- and run-time. 00035 * 00036 * Only sequential execution can be forced at compile-time. This 00037 * reduces code size and protects code parts that have 00038 * non-thread-safe side effects. 00039 * 00040 * Ultimately, forcing parallel execution at compile-time makes 00041 * sense. Often, the sequential algorithm implementation is used as 00042 * a subroutine, so no reduction in code size can be achieved. Also, 00043 * the machine the program is run on might have only one processor 00044 * core, so to avoid overhead, the algorithm is executed 00045 * sequentially. 00046 * 00047 * To force sequential execution of an algorithm ultimately at 00048 * compile-time, the user must add the tag 00049 * __gnu_parallel::sequential_tag() to the end of the parameter list, 00050 * e. g. 00051 * 00052 * \code 00053 * std::sort(v.begin(), v.end(), __gnu_parallel::sequential_tag()); 00054 * \endcode 00055 * 00056 * This is compatible with all overloaded algorithm variants. No 00057 * additional code will be instantiated, at all. The same holds for 00058 * most algorithm calls with iterators not providing random access. 00059 * 00060 * If the algorithm call is not forced to be executed sequentially 00061 * at compile-time, the decision is made at run-time. 00062 * The global variable __gnu_parallel::_Settings::algorithm_strategy 00063 * is checked. It is a tristate variable corresponding to: 00064 * 00065 * a. force_sequential, meaning the sequential algorithm is executed. 00066 * b. force_parallel, meaning the parallel algorithm is executed. 00067 * c. heuristic 00068 * 00069 * For heuristic, the parallel algorithm implementation is called 00070 * only if the input size is sufficiently large. For most 00071 * algorithms, the input size is the (combined) length of the input 00072 * sequence(s). The threshold can be set by the user, individually 00073 * for each algorithm. The according variables are called 00074 * __gnu_parallel::_Settings::[algorithm]_minimal_n . 00075 * 00076 * For some of the algorithms, there are even more tuning options, 00077 * e. g. the ability to choose from multiple algorithm variants. See 00078 * below for details. 00079 */ 00080 00081 // Written by Johannes Singler and Felix Putze. 00082 00083 #ifndef _GLIBCXX_PARALLEL_SETTINGS_H 00084 #define _GLIBCXX_PARALLEL_SETTINGS_H 1 00085 00086 #include <parallel/types.h> 00087 00088 /** 00089 * @brief Determine at compile(?)-time if the parallel variant of an 00090 * algorithm should be called. 00091 * @param c A condition that is convertible to bool that is overruled by 00092 * __gnu_parallel::_Settings::algorithm_strategy. Usually a decision 00093 * based on the input size. 00094 */ 00095 #define _GLIBCXX_PARALLEL_CONDITION(c) (__gnu_parallel::_Settings::get().algorithm_strategy != __gnu_parallel::force_sequential && ((__gnu_parallel::get_max_threads() > 1 && (c)) || __gnu_parallel::_Settings::get().algorithm_strategy == __gnu_parallel::force_parallel)) 00096 00097 /* 00098 inline bool 00099 parallel_condition(bool c) 00100 { 00101 bool ret = false; 00102 const _Settings& s = _Settings::get(); 00103 if (s.algorithm_strategy != force_seqential) 00104 { 00105 if (s.algorithm_strategy == force_parallel) 00106 ret = true; 00107 else 00108 ret = get_max_threads() > 1 && c; 00109 } 00110 return ret; 00111 } 00112 */ 00113 00114 namespace __gnu_parallel 00115 { 00116 /// class _Settings 00117 /// Run-time settings for the parallel mode, including all tunable parameters. 00118 struct _Settings 00119 { 00120 _AlgorithmStrategy algorithm_strategy; 00121 00122 _SortAlgorithm sort_algorithm; 00123 _PartialSumAlgorithm partial_sum_algorithm; 00124 _MultiwayMergeAlgorithm multiway_merge_algorithm; 00125 _FindAlgorithm find_algorithm; 00126 00127 _SplittingAlgorithm sort_splitting; 00128 _SplittingAlgorithm merge_splitting; 00129 _SplittingAlgorithm multiway_merge_splitting; 00130 00131 // Per-algorithm settings. 00132 00133 /// Minimal input size for accumulate. 00134 sequence_index_t accumulate_minimal_n; 00135 00136 /// Minimal input size for adjacent_difference. 00137 unsigned int adjacent_difference_minimal_n; 00138 00139 /// Minimal input size for count and count_if. 00140 sequence_index_t count_minimal_n; 00141 00142 /// Minimal input size for fill. 00143 sequence_index_t fill_minimal_n; 00144 00145 /// Block size increase factor for find. 00146 double find_increasing_factor; 00147 00148 /// Initial block size for find. 00149 sequence_index_t find_initial_block_size; 00150 00151 /// Maximal block size for find. 00152 sequence_index_t find_maximum_block_size; 00153 00154 /// Start with looking for this many elements sequentially, for find. 00155 sequence_index_t find_sequential_search_size; 00156 00157 /// Minimal input size for for_each. 00158 sequence_index_t for_each_minimal_n; 00159 00160 /// Minimal input size for generate. 00161 sequence_index_t generate_minimal_n; 00162 00163 /// Minimal input size for max_element. 00164 sequence_index_t max_element_minimal_n; 00165 00166 /// Minimal input size for merge. 00167 sequence_index_t merge_minimal_n; 00168 00169 /// Oversampling factor for merge. 00170 unsigned int merge_oversampling; 00171 00172 /// Minimal input size for min_element. 00173 sequence_index_t min_element_minimal_n; 00174 00175 /// Minimal input size for multiway_merge. 00176 sequence_index_t multiway_merge_minimal_n; 00177 00178 /// Oversampling factor for multiway_merge. 00179 int multiway_merge_minimal_k; 00180 00181 /// Oversampling factor for multiway_merge. 00182 unsigned int multiway_merge_oversampling; 00183 00184 /// Minimal input size for nth_element. 00185 sequence_index_t nth_element_minimal_n; 00186 00187 /// Chunk size for partition. 00188 sequence_index_t partition_chunk_size; 00189 00190 /// Chunk size for partition, relative to input size. If > 0.0, 00191 /// this value overrides partition_chunk_size. 00192 double partition_chunk_share; 00193 00194 /// Minimal input size for partition. 00195 sequence_index_t partition_minimal_n; 00196 00197 /// Minimal input size for partial_sort. 00198 sequence_index_t partial_sort_minimal_n; 00199 00200 /// Ratio for partial_sum. Assume "sum and write result" to be 00201 /// this factor slower than just "sum". 00202 float partial_sum_dilation; 00203 00204 /// Minimal input size for partial_sum. 00205 unsigned int partial_sum_minimal_n; 00206 00207 /// Minimal input size for random_shuffle. 00208 unsigned int random_shuffle_minimal_n; 00209 00210 /// Minimal input size for replace and replace_if. 00211 sequence_index_t replace_minimal_n; 00212 00213 /// Minimal input size for set_difference. 00214 sequence_index_t set_difference_minimal_n; 00215 00216 /// Minimal input size for set_intersection. 00217 sequence_index_t set_intersection_minimal_n; 00218 00219 /// Minimal input size for set_symmetric_difference. 00220 sequence_index_t set_symmetric_difference_minimal_n; 00221 00222 /// Minimal input size for set_union. 00223 sequence_index_t set_union_minimal_n; 00224 00225 /// Minimal input size for parallel sorting. 00226 sequence_index_t sort_minimal_n; 00227 00228 /// Oversampling factor for parallel std::sort (MWMS). 00229 unsigned int sort_mwms_oversampling; 00230 00231 /// Such many samples to take to find a good pivot (quicksort). 00232 unsigned int sort_qs_num_samples_preset; 00233 00234 /// Maximal subsequence length to switch to unbalanced base case. 00235 /// Applies to std::sort with dynamically load-balanced quicksort. 00236 sequence_index_t sort_qsb_base_case_maximal_n; 00237 00238 /// Minimal input size for parallel std::transform. 00239 sequence_index_t transform_minimal_n; 00240 00241 /// Minimal input size for unique_copy. 00242 sequence_index_t unique_copy_minimal_n; 00243 00244 sequence_index_t workstealing_chunk_size; 00245 00246 // Hardware dependent tuning parameters. 00247 00248 /// Size of the L1 cache in bytes (underestimation). 00249 unsigned long long L1_cache_size; 00250 00251 /// Size of the L2 cache in bytes (underestimation). 00252 unsigned long long L2_cache_size; 00253 00254 /// Size of the Translation Lookaside Buffer (underestimation). 00255 unsigned int TLB_size; 00256 00257 /// Overestimation of cache line size. Used to avoid false 00258 /// sharing, i. e. elements of different threads are at least this 00259 /// amount apart. 00260 unsigned int cache_line_size; 00261 00262 // Statistics. 00263 00264 /// The number of stolen ranges in load-balanced quicksort. 00265 sequence_index_t qsb_steals; 00266 00267 /// Get the global settings. 00268 static const _Settings& 00269 get() throw(); 00270 00271 /// Set the global settings. 00272 static void 00273 set(_Settings&) throw(); 00274 00275 explicit 00276 _Settings() : algorithm_strategy(heuristic), sort_algorithm(MWMS), partial_sum_algorithm(LINEAR), multiway_merge_algorithm(LOSER_TREE), find_algorithm(CONSTANT_SIZE_BLOCKS), sort_splitting(EXACT), merge_splitting(EXACT), multiway_merge_splitting(EXACT), accumulate_minimal_n(1000), adjacent_difference_minimal_n(1000), count_minimal_n(1000), fill_minimal_n(1000), find_increasing_factor(2.0), find_initial_block_size(256), find_maximum_block_size(8192), find_sequential_search_size(256), for_each_minimal_n(1000), generate_minimal_n(1000), max_element_minimal_n(1000), merge_minimal_n(1000), merge_oversampling(10), min_element_minimal_n(1000), multiway_merge_minimal_n(1000), multiway_merge_minimal_k(2), multiway_merge_oversampling(10), nth_element_minimal_n(1000), partition_chunk_size(1000), partition_chunk_share(0.0), partition_minimal_n(1000), partial_sort_minimal_n(1000), partial_sum_dilation(1.0f), partial_sum_minimal_n(1000), random_shuffle_minimal_n(1000), replace_minimal_n(1000), set_difference_minimal_n(1000), set_intersection_minimal_n(1000), set_symmetric_difference_minimal_n(1000), set_union_minimal_n(1000), sort_minimal_n(1000), sort_mwms_oversampling(10), sort_qs_num_samples_preset(100), sort_qsb_base_case_maximal_n(100), transform_minimal_n(1000), unique_copy_minimal_n(10000), workstealing_chunk_size(100), L1_cache_size(16 << 10), L2_cache_size(256 << 10), TLB_size(128), cache_line_size(64), qsb_steals(0) 00277 { } 00278 }; 00279 } 00280 00281 #endif /* _GLIBCXX_PARALLEL_SETTINGS_H */