1 //===- llvm/Support/Parallel.h - Parallel algorithms ----------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8
9 #ifndef LLVM_SUPPORT_PARALLEL_H
10 #define LLVM_SUPPORT_PARALLEL_H
11
12 #include "llvm/ADT/STLExtras.h"
13 #include "llvm/Config/llvm-config.h"
14 #include "llvm/Support/MathExtras.h"
15
16 #include <algorithm>
17 #include <condition_variable>
18 #include <functional>
19 #include <mutex>
20
21 namespace llvm {
22
23 namespace parallel {
24 struct sequential_execution_policy {};
25 struct parallel_execution_policy {};
26
27 template <typename T>
28 struct is_execution_policy
29 : public std::integral_constant<
30 bool, llvm::is_one_of<T, sequential_execution_policy,
31 parallel_execution_policy>::value> {};
32
33 constexpr sequential_execution_policy seq{};
34 constexpr parallel_execution_policy par{};
35
36 namespace detail {
37
38 #if LLVM_ENABLE_THREADS
39
40 class Latch {
41 uint32_t Count;
42 mutable std::mutex Mutex;
43 mutable std::condition_variable Cond;
44
45 public:
Count(Count)46 explicit Latch(uint32_t Count = 0) : Count(Count) {}
~Latch()47 ~Latch() { sync(); }
48
inc()49 void inc() {
50 std::lock_guard<std::mutex> lock(Mutex);
51 ++Count;
52 }
53
dec()54 void dec() {
55 std::lock_guard<std::mutex> lock(Mutex);
56 if (--Count == 0)
57 Cond.notify_all();
58 }
59
sync()60 void sync() const {
61 std::unique_lock<std::mutex> lock(Mutex);
62 Cond.wait(lock, [&] { return Count == 0; });
63 }
64 };
65
66 class TaskGroup {
67 Latch L;
68 bool Parallel;
69
70 public:
71 TaskGroup();
72 ~TaskGroup();
73
74 void spawn(std::function<void()> f);
75
sync()76 void sync() const { L.sync(); }
77 };
78
79 const ptrdiff_t MinParallelSize = 1024;
80
81 /// Inclusive median.
82 template <class RandomAccessIterator, class Comparator>
medianOf3(RandomAccessIterator Start,RandomAccessIterator End,const Comparator & Comp)83 RandomAccessIterator medianOf3(RandomAccessIterator Start,
84 RandomAccessIterator End,
85 const Comparator &Comp) {
86 RandomAccessIterator Mid = Start + (std::distance(Start, End) / 2);
87 return Comp(*Start, *(End - 1))
88 ? (Comp(*Mid, *(End - 1)) ? (Comp(*Start, *Mid) ? Mid : Start)
89 : End - 1)
90 : (Comp(*Mid, *Start) ? (Comp(*(End - 1), *Mid) ? Mid : End - 1)
91 : Start);
92 }
93
94 template <class RandomAccessIterator, class Comparator>
parallel_quick_sort(RandomAccessIterator Start,RandomAccessIterator End,const Comparator & Comp,TaskGroup & TG,size_t Depth)95 void parallel_quick_sort(RandomAccessIterator Start, RandomAccessIterator End,
96 const Comparator &Comp, TaskGroup &TG, size_t Depth) {
97 // Do a sequential sort for small inputs.
98 if (std::distance(Start, End) < detail::MinParallelSize || Depth == 0) {
99 llvm::sort(Start, End, Comp);
100 return;
101 }
102
103 // Partition.
104 auto Pivot = medianOf3(Start, End, Comp);
105 // Move Pivot to End.
106 std::swap(*(End - 1), *Pivot);
107 Pivot = std::partition(Start, End - 1, [&Comp, End](decltype(*Start) V) {
108 return Comp(V, *(End - 1));
109 });
110 // Move Pivot to middle of partition.
111 std::swap(*Pivot, *(End - 1));
112
113 // Recurse.
114 TG.spawn([=, &Comp, &TG] {
115 parallel_quick_sort(Start, Pivot, Comp, TG, Depth - 1);
116 });
117 parallel_quick_sort(Pivot + 1, End, Comp, TG, Depth - 1);
118 }
119
120 template <class RandomAccessIterator, class Comparator>
parallel_sort(RandomAccessIterator Start,RandomAccessIterator End,const Comparator & Comp)121 void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End,
122 const Comparator &Comp) {
123 TaskGroup TG;
124 parallel_quick_sort(Start, End, Comp, TG,
125 llvm::Log2_64(std::distance(Start, End)) + 1);
126 }
127
128 template <class IterTy, class FuncTy>
parallel_for_each(IterTy Begin,IterTy End,FuncTy Fn)129 void parallel_for_each(IterTy Begin, IterTy End, FuncTy Fn) {
130 // TaskGroup has a relatively high overhead, so we want to reduce
131 // the number of spawn() calls. We'll create up to 1024 tasks here.
132 // (Note that 1024 is an arbitrary number. This code probably needs
133 // improving to take the number of available cores into account.)
134 ptrdiff_t TaskSize = std::distance(Begin, End) / 1024;
135 if (TaskSize == 0)
136 TaskSize = 1;
137
138 TaskGroup TG;
139 while (TaskSize < std::distance(Begin, End)) {
140 TG.spawn([=, &Fn] { std::for_each(Begin, Begin + TaskSize, Fn); });
141 Begin += TaskSize;
142 }
143 std::for_each(Begin, End, Fn);
144 }
145
146 template <class IndexTy, class FuncTy>
parallel_for_each_n(IndexTy Begin,IndexTy End,FuncTy Fn)147 void parallel_for_each_n(IndexTy Begin, IndexTy End, FuncTy Fn) {
148 ptrdiff_t TaskSize = (End - Begin) / 1024;
149 if (TaskSize == 0)
150 TaskSize = 1;
151
152 TaskGroup TG;
153 IndexTy I = Begin;
154 for (; I + TaskSize < End; I += TaskSize) {
155 TG.spawn([=, &Fn] {
156 for (IndexTy J = I, E = I + TaskSize; J != E; ++J)
157 Fn(J);
158 });
159 }
160 for (IndexTy J = I; J < End; ++J)
161 Fn(J);
162 }
163
164 #endif
165
166 template <typename Iter>
167 using DefComparator =
168 std::less<typename std::iterator_traits<Iter>::value_type>;
169
170 } // namespace detail
171
172 // sequential algorithm implementations.
173 template <class Policy, class RandomAccessIterator,
174 class Comparator = detail::DefComparator<RandomAccessIterator>>
175 void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End,
176 const Comparator &Comp = Comparator()) {
177 static_assert(is_execution_policy<Policy>::value,
178 "Invalid execution policy!");
179 llvm::sort(Start, End, Comp);
180 }
181
182 template <class Policy, class IterTy, class FuncTy>
for_each(Policy policy,IterTy Begin,IterTy End,FuncTy Fn)183 void for_each(Policy policy, IterTy Begin, IterTy End, FuncTy Fn) {
184 static_assert(is_execution_policy<Policy>::value,
185 "Invalid execution policy!");
186 std::for_each(Begin, End, Fn);
187 }
188
189 template <class Policy, class IndexTy, class FuncTy>
for_each_n(Policy policy,IndexTy Begin,IndexTy End,FuncTy Fn)190 void for_each_n(Policy policy, IndexTy Begin, IndexTy End, FuncTy Fn) {
191 static_assert(is_execution_policy<Policy>::value,
192 "Invalid execution policy!");
193 for (IndexTy I = Begin; I != End; ++I)
194 Fn(I);
195 }
196
197 // Parallel algorithm implementations, only available when LLVM_ENABLE_THREADS
198 // is true.
199 #if LLVM_ENABLE_THREADS
200 template <class RandomAccessIterator,
201 class Comparator = detail::DefComparator<RandomAccessIterator>>
202 void sort(parallel_execution_policy policy, RandomAccessIterator Start,
203 RandomAccessIterator End, const Comparator &Comp = Comparator()) {
204 detail::parallel_sort(Start, End, Comp);
205 }
206
207 template <class IterTy, class FuncTy>
for_each(parallel_execution_policy policy,IterTy Begin,IterTy End,FuncTy Fn)208 void for_each(parallel_execution_policy policy, IterTy Begin, IterTy End,
209 FuncTy Fn) {
210 detail::parallel_for_each(Begin, End, Fn);
211 }
212
213 template <class IndexTy, class FuncTy>
for_each_n(parallel_execution_policy policy,IndexTy Begin,IndexTy End,FuncTy Fn)214 void for_each_n(parallel_execution_policy policy, IndexTy Begin, IndexTy End,
215 FuncTy Fn) {
216 detail::parallel_for_each_n(Begin, End, Fn);
217 }
218 #endif
219
220 } // namespace parallel
221 } // namespace llvm
222
223 #endif // LLVM_SUPPORT_PARALLEL_H
224