LLVM  15.0.0git
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"
14 #include "llvm/Support/Error.h"
16 #include "llvm/Support/Threading.h"
17 
18 #include <algorithm>
19 #include <condition_variable>
20 #include <functional>
21 #include <mutex>
22 
23 namespace llvm {
24 
25 namespace parallel {
26 
27 // Strategy for the default executor used by the parallel routines provided by
28 // this file. It defaults to using all hardware threads and should be
29 // initialized before the first use of parallel routines.
31 
32 namespace detail {
33 
34 #if LLVM_ENABLE_THREADS
35 
36 class Latch {
37  uint32_t Count;
38  mutable std::mutex Mutex;
39  mutable std::condition_variable Cond;
40 
41 public:
42  explicit Latch(uint32_t Count = 0) : Count(Count) {}
43  ~Latch() {
44  // Ensure at least that sync() was called.
45  assert(Count == 0);
46  }
47 
48  void inc() {
49  std::lock_guard<std::mutex> lock(Mutex);
50  ++Count;
51  }
52 
53  void dec() {
54  std::lock_guard<std::mutex> lock(Mutex);
55  if (--Count == 0)
56  Cond.notify_all();
57  }
58 
59  void sync() const {
60  std::unique_lock<std::mutex> lock(Mutex);
61  Cond.wait(lock, [&] { return Count == 0; });
62  }
63 };
64 
65 class TaskGroup {
66  Latch L;
67  bool Parallel;
68 
69 public:
70  TaskGroup();
71  ~TaskGroup();
72 
73  void spawn(std::function<void()> f);
74 
75  void sync() const { L.sync(); }
76 };
77 
79 
80 /// Inclusive median.
81 template <class RandomAccessIterator, class Comparator>
82 RandomAccessIterator medianOf3(RandomAccessIterator Start,
83  RandomAccessIterator End,
84  const Comparator &Comp) {
85  RandomAccessIterator Mid = Start + (std::distance(Start, End) / 2);
86  return Comp(*Start, *(End - 1))
87  ? (Comp(*Mid, *(End - 1)) ? (Comp(*Start, *Mid) ? Mid : Start)
88  : End - 1)
89  : (Comp(*Mid, *Start) ? (Comp(*(End - 1), *Mid) ? Mid : End - 1)
90  : Start);
91 }
92 
93 template <class RandomAccessIterator, class Comparator>
94 void parallel_quick_sort(RandomAccessIterator Start, RandomAccessIterator End,
95  const Comparator &Comp, TaskGroup &TG, size_t Depth) {
96  // Do a sequential sort for small inputs.
97  if (std::distance(Start, End) < detail::MinParallelSize || Depth == 0) {
98  llvm::sort(Start, End, Comp);
99  return;
100  }
101 
102  // Partition.
103  auto Pivot = medianOf3(Start, End, Comp);
104  // Move Pivot to End.
105  std::swap(*(End - 1), *Pivot);
106  Pivot = std::partition(Start, End - 1, [&Comp, End](decltype(*Start) V) {
107  return Comp(V, *(End - 1));
108  });
109  // Move Pivot to middle of partition.
110  std::swap(*Pivot, *(End - 1));
111 
112  // Recurse.
113  TG.spawn([=, &Comp, &TG] {
114  parallel_quick_sort(Start, Pivot, Comp, TG, Depth - 1);
115  });
116  parallel_quick_sort(Pivot + 1, End, Comp, TG, Depth - 1);
117 }
118 
119 template <class RandomAccessIterator, class Comparator>
120 void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End,
121  const Comparator &Comp) {
122  TaskGroup TG;
123  parallel_quick_sort(Start, End, Comp, TG,
124  llvm::Log2_64(std::distance(Start, End)) + 1);
125 }
126 
127 // TaskGroup has a relatively high overhead, so we want to reduce
128 // the number of spawn() calls. We'll create up to 1024 tasks here.
129 // (Note that 1024 is an arbitrary number. This code probably needs
130 // improving to take the number of available cores into account.)
131 enum { MaxTasksPerGroup = 1024 };
132 
133 template <class IterTy, class ResultTy, class ReduceFuncTy,
134  class TransformFuncTy>
135 ResultTy parallel_transform_reduce(IterTy Begin, IterTy End, ResultTy Init,
136  ReduceFuncTy Reduce,
137  TransformFuncTy Transform) {
138  // Limit the number of tasks to MaxTasksPerGroup to limit job scheduling
139  // overhead on large inputs.
140  size_t NumInputs = std::distance(Begin, End);
141  if (NumInputs == 0)
142  return std::move(Init);
143  size_t NumTasks = std::min(static_cast<size_t>(MaxTasksPerGroup), NumInputs);
144  std::vector<ResultTy> Results(NumTasks, Init);
145  {
146  // Each task processes either TaskSize or TaskSize+1 inputs. Any inputs
147  // remaining after dividing them equally amongst tasks are distributed as
148  // one extra input over the first tasks.
149  TaskGroup TG;
150  size_t TaskSize = NumInputs / NumTasks;
151  size_t RemainingInputs = NumInputs % NumTasks;
152  IterTy TBegin = Begin;
153  for (size_t TaskId = 0; TaskId < NumTasks; ++TaskId) {
154  IterTy TEnd = TBegin + TaskSize + (TaskId < RemainingInputs ? 1 : 0);
155  TG.spawn([=, &Transform, &Reduce, &Results] {
156  // Reduce the result of transformation eagerly within each task.
157  ResultTy R = Init;
158  for (IterTy It = TBegin; It != TEnd; ++It)
159  R = Reduce(R, Transform(*It));
160  Results[TaskId] = R;
161  });
162  TBegin = TEnd;
163  }
164  assert(TBegin == End);
165  }
166 
167  // Do a final reduction. There are at most 1024 tasks, so this only adds
168  // constant single-threaded overhead for large inputs. Hopefully most
169  // reductions are cheaper than the transformation.
170  ResultTy FinalResult = std::move(Results.front());
171  for (ResultTy &PartialResult :
172  makeMutableArrayRef(Results.data() + 1, Results.size() - 1))
173  FinalResult = Reduce(FinalResult, std::move(PartialResult));
174  return std::move(FinalResult);
175 }
176 
177 #endif
178 
179 } // namespace detail
180 } // namespace parallel
181 
182 template <class RandomAccessIterator,
183  class Comparator = std::less<
184  typename std::iterator_traits<RandomAccessIterator>::value_type>>
185 void parallelSort(RandomAccessIterator Start, RandomAccessIterator End,
186  const Comparator &Comp = Comparator()) {
187 #if LLVM_ENABLE_THREADS
188  if (parallel::strategy.ThreadsRequested != 1) {
189  parallel::detail::parallel_sort(Start, End, Comp);
190  return;
191  }
192 #endif
193  llvm::sort(Start, End, Comp);
194 }
195 
196 void parallelForEachN(size_t Begin, size_t End, function_ref<void(size_t)> Fn);
197 
198 template <class IterTy, class FuncTy>
199 void parallelForEach(IterTy Begin, IterTy End, FuncTy Fn) {
200  parallelForEachN(0, End - Begin, [&](size_t I) { Fn(Begin[I]); });
201 }
202 
203 template <class IterTy, class ResultTy, class ReduceFuncTy,
204  class TransformFuncTy>
205 ResultTy parallelTransformReduce(IterTy Begin, IterTy End, ResultTy Init,
206  ReduceFuncTy Reduce,
207  TransformFuncTy Transform) {
208 #if LLVM_ENABLE_THREADS
209  if (parallel::strategy.ThreadsRequested != 1) {
210  return parallel::detail::parallel_transform_reduce(Begin, End, Init, Reduce,
211  Transform);
212  }
213 #endif
214  for (IterTy I = Begin; I != End; ++I)
215  Init = Reduce(std::move(Init), Transform(*I));
216  return std::move(Init);
217 }
218 
219 // Range wrappers.
220 template <class RangeTy,
221  class Comparator = std::less<decltype(*std::begin(RangeTy()))>>
222 void parallelSort(RangeTy &&R, const Comparator &Comp = Comparator()) {
223  parallelSort(std::begin(R), std::end(R), Comp);
224 }
225 
226 template <class RangeTy, class FuncTy>
227 void parallelForEach(RangeTy &&R, FuncTy Fn) {
228  parallelForEach(std::begin(R), std::end(R), Fn);
229 }
230 
231 template <class RangeTy, class ResultTy, class ReduceFuncTy,
232  class TransformFuncTy>
233 ResultTy parallelTransformReduce(RangeTy &&R, ResultTy Init,
234  ReduceFuncTy Reduce,
235  TransformFuncTy Transform) {
236  return parallelTransformReduce(std::begin(R), std::end(R), Init, Reduce,
237  Transform);
238 }
239 
240 // Parallel for-each, but with error handling.
241 template <class RangeTy, class FuncTy>
242 Error parallelForEachError(RangeTy &&R, FuncTy Fn) {
243  // The transform_reduce algorithm requires that the initial value be copyable.
244  // Error objects are uncopyable. We only need to copy initial success values,
245  // so work around this mismatch via the C API. The C API represents success
246  // values with a null pointer. The joinErrors discards null values and joins
247  // multiple errors into an ErrorList.
250  [](LLVMErrorRef Lhs, LLVMErrorRef Rhs) {
251  return wrap(joinErrors(unwrap(Lhs), unwrap(Rhs)));
252  },
253  [&Fn](auto &&V) { return wrap(Fn(V)); }));
254 }
255 
256 } // namespace llvm
257 
258 #endif // LLVM_SUPPORT_PARALLEL_H
MathExtras.h
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:17
llvm::parallel::detail::TaskGroup::~TaskGroup
~TaskGroup()
Definition: Parallel.cpp:154
llvm::parallelForEachError
Error parallelForEachError(RangeTy &&R, FuncTy Fn)
Definition: Parallel.h:242
llvm::ThreadPoolStrategy
This tells how a thread pool will be used.
Definition: Threading.h:116
llvm::parallel::strategy
ThreadPoolStrategy strategy
Definition: Parallel.cpp:20
llvm::Error::success
static ErrorSuccess success()
Create a success value.
Definition: Error.h:329
Error.h
llvm::Depth
@ Depth
Definition: SIMachineScheduler.h:36
llvm::sys::path::end
const_iterator end(StringRef path)
Get end iterator over path.
Definition: Path.cpp:235
llvm::sys::path::begin
const_iterator begin(StringRef path, Style style=Style::native)
Get begin iterator over path.
Definition: Path.cpp:226
Results
Function Alias Analysis Results
Definition: AliasAnalysis.cpp:848
llvm::parallel::detail::TaskGroup::spawn
void spawn(std::function< void()> f)
Definition: Parallel.cpp:161
STLExtras.h
llvm::lltok::less
@ less
Definition: LLToken.h:32
llvm::Log2_64
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:629
ptrdiff_t
f
Itanium Name Demangler i e convert the string _Z1fv into f()". You can also use the CRTP base ManglingParser to perform some simple analysis on the mangled name
llvm::parallel::detail::parallel_transform_reduce
ResultTy parallel_transform_reduce(IterTy Begin, IterTy End, ResultTy Init, ReduceFuncTy Reduce, TransformFuncTy Transform)
Definition: Parallel.h:135
llvm::parallel::detail::MaxTasksPerGroup
@ MaxTasksPerGroup
Definition: Parallel.h:131
llvm::unwrap
Attribute unwrap(LLVMAttributeRef Attr)
Definition: Attributes.h:264
llvm::parallel::detail::Latch
Definition: Parallel.h:36
llvm::sys::SmartMutex< false >
llvm::parallel::detail::TaskGroup::TaskGroup
TaskGroup()
Definition: Parallel.cpp:153
llvm::parallelForEach
void parallelForEach(IterTy Begin, IterTy End, FuncTy Fn)
Definition: Parallel.h:199
Threading.h
llvm::parallel::detail::Latch::~Latch
~Latch()
Definition: Parallel.h:43
llvm::parallel::detail::MinParallelSize
const ptrdiff_t MinParallelSize
Definition: Parallel.h:78
move
compiles ldr LCPI1_0 ldr ldr mov lsr tst moveq r1 ldr LCPI1_1 and r0 bx lr It would be better to do something like to fold the shift into the conditional move
Definition: README.txt:546
llvm::parallel::detail::Latch::dec
void dec()
Definition: Parallel.h:53
I
#define I(x, y, z)
Definition: MD5.cpp:58
llvm::parallelTransformReduce
ResultTy parallelTransformReduce(IterTy Begin, IterTy End, ResultTy Init, ReduceFuncTy Reduce, TransformFuncTy Transform)
Definition: Parallel.h:205
llvm::joinErrors
Error joinErrors(Error E1, Error E2)
Concatenate errors.
Definition: Error.h:426
assert
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
std::swap
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:853
llvm::parallel::detail::TaskGroup::sync
void sync() const
Definition: Parallel.h:75
function
print Print MemDeps of function
Definition: MemDepPrinter.cpp:82
llvm::wrap
LLVMAttributeRef wrap(Attribute Attr)
Definition: Attributes.h:259
llvm::parallel::detail::Latch::Latch
Latch(uint32_t Count=0)
Definition: Parallel.h:42
llvm::min
Expected< ExpressionValue > min(const ExpressionValue &Lhs, const ExpressionValue &Rhs)
Definition: FileCheck.cpp:357
Cond
SmallVector< MachineOperand, 4 > Cond
Definition: BasicBlockSections.cpp:178
llvm::parallelSort
void parallelSort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:185
uint32_t
llvm::partition
auto partition(R &&Range, UnaryPredicate P)
Provide wrappers to std::partition which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1718
llvm::parallel::detail::TaskGroup
Definition: Parallel.h:65
llvm::Init
Definition: Record.h:281
llvm::parallel::detail::Latch::sync
void sync() const
Definition: Parallel.h:59
llvm::parallel::detail::parallel_sort
void parallel_sort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp)
Definition: Parallel.h:120
llvm::Error
Lightweight error class with error context and mandatory checking.
Definition: Error.h:155
llvm::sort
void sort(IteratorTy Start, IteratorTy End)
Definition: STLExtras.h:1550
llvm::parallel::detail::parallel_quick_sort
void parallel_quick_sort(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp, TaskGroup &TG, size_t Depth)
Definition: Parallel.h:94
LLVMErrorRef
struct LLVMOpaqueError * LLVMErrorRef
Opaque reference to an error instance.
Definition: Error.h:33
llvm::parallelForEachN
void parallelForEachN(size_t Begin, size_t End, function_ref< void(size_t)> Fn)
Definition: Parallel.cpp:178
llvm::makeMutableArrayRef
MutableArrayRef< T > makeMutableArrayRef(T &OneElt)
Construct a MutableArrayRef from a single element.
Definition: ArrayRef.h:533
llvm::parallel::detail::Latch::inc
void inc()
Definition: Parallel.h:48
llvm::parallel::detail::medianOf3
RandomAccessIterator medianOf3(RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp)
Inclusive median.
Definition: Parallel.h:82