LLVM  9.0.0svn
Parallel.h
Go to the documentation of this file.
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"
15 
16 #include <algorithm>
17 #include <condition_variable>
18 #include <functional>
19 #include <mutex>
20 
21 #if defined(_MSC_VER) && LLVM_ENABLE_THREADS
22 #pragma warning(push)
23 #pragma warning(disable : 4530)
24 #include <concrt.h>
25 #include <ppl.h>
26 #pragma warning(pop)
27 #endif
28 
29 namespace llvm {
30 
31 namespace parallel {
34 
35 template <typename T>
37  : public std::integral_constant<
38  bool, llvm::is_one_of<T, sequential_execution_policy,
39  parallel_execution_policy>::value> {};
40 
43 
44 namespace detail {
45 
46 #if LLVM_ENABLE_THREADS
47 
48 class Latch {
49  uint32_t Count;
50  mutable std::mutex Mutex;
51  mutable std::condition_variable Cond;
52 
53 public:
54  explicit Latch(uint32_t Count = 0) : Count(Count) {}
55  ~Latch() { sync(); }
56 
57  void inc() {
58  std::lock_guard<std::mutex> lock(Mutex);
59  ++Count;
60  }
61 
62  void dec() {
63  std::lock_guard<std::mutex> lock(Mutex);
64  if (--Count == 0)
65  Cond.notify_all();
66  }
67 
68  void sync() const {
69  std::unique_lock<std::mutex> lock(Mutex);
70  Cond.wait(lock, [&] { return Count == 0; });
71  }
72 };
73 
74 class TaskGroup {
75  Latch L;
76 
77 public:
78  void spawn(std::function<void()> f);
79 
80  void sync() const { L.sync(); }
81 };
82 
83 #if defined(_MSC_VER)
84 template <class RandomAccessIterator, class Comparator>
85 void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End,
86  const Comparator &Comp) {
87  concurrency::parallel_sort(Start, End, Comp);
88 }
89 template <class IterTy, class FuncTy>
90 void parallel_for_each(IterTy Begin, IterTy End, FuncTy Fn) {
91  concurrency::parallel_for_each(Begin, End, Fn);
92 }
93 
94 template <class IndexTy, class FuncTy>
95 void parallel_for_each_n(IndexTy Begin, IndexTy End, FuncTy Fn) {
96  concurrency::parallel_for(Begin, End, Fn);
97 }
98 
99 #else
101 
102 /// Inclusive median.
103 template <class RandomAccessIterator, class Comparator>
104 RandomAccessIterator medianOf3(RandomAccessIterator Start,
105  RandomAccessIterator End,
106  const Comparator &Comp) {
107  RandomAccessIterator Mid = Start + (std::distance(Start, End) / 2);
108  return Comp(*Start, *(End - 1))
109  ? (Comp(*Mid, *(End - 1)) ? (Comp(*Start, *Mid) ? Mid : Start)
110  : End - 1)
111  : (Comp(*Mid, *Start) ? (Comp(*(End - 1), *Mid) ? Mid : End - 1)
112  : Start);
113 }
114 
115 template <class RandomAccessIterator, class Comparator>
116 void parallel_quick_sort(RandomAccessIterator Start, RandomAccessIterator End,
117  const Comparator &Comp, TaskGroup &TG, size_t Depth) {
118  // Do a sequential sort for small inputs.
119  if (std::distance(Start, End) < detail::MinParallelSize || Depth == 0) {
120  llvm::sort(Start, End, Comp);
121  return;
122  }
123 
124  // Partition.
125  auto Pivot = medianOf3(Start, End, Comp);
126  // Move Pivot to End.
127  std::swap(*(End - 1), *Pivot);
128  Pivot = std::partition(Start, End - 1, [&Comp, End](decltype(*Start) V) {
129  return Comp(V, *(End - 1));
130  });
131  // Move Pivot to middle of partition.
132  std::swap(*Pivot, *(End - 1));
133 
134  // Recurse.
135  TG.spawn([=, &Comp, &TG] {
136  parallel_quick_sort(Start, Pivot, Comp, TG, Depth - 1);
137  });
138  parallel_quick_sort(Pivot + 1, End, Comp, TG, Depth - 1);
139 }
140 
141 template <class RandomAccessIterator, class Comparator>
142 void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End,
143  const Comparator &Comp) {
144  TaskGroup TG;
145  parallel_quick_sort(Start, End, Comp, TG,
146  llvm::Log2_64(std::distance(Start, End)) + 1);
147 }
148 
149 template <class IterTy, class FuncTy>
150 void parallel_for_each(IterTy Begin, IterTy End, FuncTy Fn) {
151  // TaskGroup has a relatively high overhead, so we want to reduce
152  // the number of spawn() calls. We'll create up to 1024 tasks here.
153  // (Note that 1024 is an arbitrary number. This code probably needs
154  // improving to take the number of available cores into account.)
155  ptrdiff_t TaskSize = std::distance(Begin, End) / 1024;
156  if (TaskSize == 0)
157  TaskSize = 1;
158 
159  TaskGroup TG;
160  while (TaskSize < std::distance(Begin, End)) {
161  TG.spawn([=, &Fn] { std::for_each(Begin, Begin + TaskSize, Fn); });
162  Begin += TaskSize;
163  }
164  std::for_each(Begin, End, Fn);
165 }
166 
167 template <class IndexTy, class FuncTy>
168 void parallel_for_each_n(IndexTy Begin, IndexTy End, FuncTy Fn) {
169  ptrdiff_t TaskSize = (End - Begin) / 1024;
170  if (TaskSize == 0)
171  TaskSize = 1;
172 
173  TaskGroup TG;
174  IndexTy I = Begin;
175  for (; I + TaskSize < End; I += TaskSize) {
176  TG.spawn([=, &Fn] {
177  for (IndexTy J = I, E = I + TaskSize; J != E; ++J)
178  Fn(J);
179  });
180  }
181  for (IndexTy J = I; J < End; ++J)
182  Fn(J);
183 }
184 
185 #endif
186 
187 #endif
188 
189 template <typename Iter>
190 using DefComparator =
191  std::less<typename std::iterator_traits<Iter>::value_type>;
192 
193 } // namespace detail
194 
195 // sequential algorithm implementations.
196 template <class Policy, class RandomAccessIterator,
198 void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End,
199  const Comparator &Comp = Comparator()) {
201  "Invalid execution policy!");
202  llvm::sort(Start, End, Comp);
203 }
204 
205 template <class Policy, class IterTy, class FuncTy>
206 void for_each(Policy policy, IterTy Begin, IterTy End, FuncTy Fn) {
208  "Invalid execution policy!");
209  std::for_each(Begin, End, Fn);
210 }
211 
212 template <class Policy, class IndexTy, class FuncTy>
213 void for_each_n(Policy policy, IndexTy Begin, IndexTy End, FuncTy Fn) {
215  "Invalid execution policy!");
216  for (IndexTy I = Begin; I != End; ++I)
217  Fn(I);
218 }
219 
220 // Parallel algorithm implementations, only available when LLVM_ENABLE_THREADS
221 // is true.
222 #if LLVM_ENABLE_THREADS
223 template <class RandomAccessIterator,
225 void sort(parallel_execution_policy policy, RandomAccessIterator Start,
226  RandomAccessIterator End, const Comparator &Comp = Comparator()) {
227  detail::parallel_sort(Start, End, Comp);
228 }
229 
230 template <class IterTy, class FuncTy>
231 void for_each(parallel_execution_policy policy, IterTy Begin, IterTy End,
232  FuncTy Fn) {
233  detail::parallel_for_each(Begin, End, Fn);
234 }
235 
236 template <class IndexTy, class FuncTy>
237 void for_each_n(parallel_execution_policy policy, IndexTy Begin, IndexTy End,
238  FuncTy Fn) {
239  detail::parallel_for_each_n(Begin, End, Fn);
240 }
241 #endif
242 
243 } // namespace parallel
244 } // namespace llvm
245 
246 #endif // LLVM_SUPPORT_PARALLEL_H
const ptrdiff_t MinParallelSize
Definition: Parallel.h:100
void for_each_n(Policy policy, IndexTy Begin, IndexTy End, FuncTy Fn)
Definition: Parallel.h:213
This class represents lattice values for constants.
Definition: AllocatorList.h:23
void parallel_for_each_n(IndexTy Begin, IndexTy End, FuncTy Fn)
Definition: Parallel.h:168
constexpr sequential_execution_policy seq
Definition: Parallel.h:41
Latch(uint32_t Count=0)
Definition: Parallel.h:54
SmartMutex< false > Mutex
Mutex - A standard, always enforced mutex.
Definition: Mutex.h:138
auto partition(R &&Range, UnaryPredicate P) -> decltype(adl_begin(Range))
Provide wrappers to std::partition which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1274
void spawn(std::function< void()> f)
Definition: Parallel.cpp:121
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:1115
constexpr parallel_execution_policy par
Definition: Parallel.h:42
void for_each(Policy policy, IterTy Begin, IterTy End, FuncTy Fn)
Definition: Parallel.h:206
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:940
std::less< typename std::iterator_traits< Iter >::value_type > DefComparator
Definition: Parallel.h:191
void parallel_for_each(IterTy Begin, IterTy End, FuncTy Fn)
Definition: Parallel.h:150
#define I(x, y, z)
Definition: MD5.cpp:58
void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp)
Definition: Parallel.h:142
RandomAccessIterator medianOf3(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp)
Inclusive median.
Definition: Parallel.h:104
void for_each(parallel_execution_policy policy, IterTy Begin, IterTy End, FuncTy Fn)
Definition: Parallel.h:231
void parallel_quick_sort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp, TaskGroup &TG, size_t Depth)
Definition: Parallel.h:116
print Print MemDeps of function
void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:198
unsigned Log2_64(uint64_t Value)
Return the floor log base 2 of the specified value, -1 if the value is zero.
Definition: MathExtras.h:544