aboutsummaryrefslogtreecommitdiffhomepage
path: root/unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h
blob: ed1a761b6360708993e4b12d8a014253332f3391 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2016 Dmitry Vyukov <dvyukov@google.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.

#ifndef EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H
#define EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H


namespace Eigen {

template <typename Environment>
class NonBlockingThreadPoolTempl : public Eigen::ThreadPoolInterface {
 public:
  typedef typename Environment::Task Task;
  typedef RunQueue<Task, 1024> Queue;

  NonBlockingThreadPoolTempl(int num_threads, Environment env = Environment())
      : env_(env),
        threads_(num_threads),
        queues_(num_threads),
        coprimes_(num_threads),
        waiters_(num_threads),
        blocked_(0),
        spinning_(0),
        done_(false),
        cancelled_(false),
        ec_(waiters_) {
    waiters_.resize(num_threads);

    // Calculate coprimes of num_threads.
    // Coprimes are used for a random walk over all threads in Steal
    // and NonEmptyQueueIndex. Iteration is based on the fact that if we take
    // a walk starting thread index t and calculate num_threads - 1 subsequent
    // indices as (t + coprime) % num_threads, we will cover all threads without
    // repetitions (effectively getting a presudo-random permutation of thread
    // indices).
    for (int i = 1; i <= num_threads; i++) {
      unsigned a = i;
      unsigned b = num_threads;
      // If GCD(a, b) == 1, then a and b are coprimes.
      while (b != 0) {
        unsigned tmp = a;
        a = b;
        b = tmp % b;
      }
      if (a == 1) {
        coprimes_.push_back(i);
      }
    }
    for (int i = 0; i < num_threads; i++) {
      queues_.push_back(new Queue());
    }
    for (int i = 0; i < num_threads; i++) {
      threads_.push_back(env_.CreateThread([this, i]() { WorkerLoop(i); }));
    }
  }

  ~NonBlockingThreadPoolTempl() {
    done_ = true;

    // Now if all threads block without work, they will start exiting.
    // But note that threads can continue to work arbitrary long,
    // block, submit new work, unblock and otherwise live full life.
    if (!cancelled_) {
      ec_.Notify(true);
    } else {
      // Since we were cancelled, there might be entries in the queues.
      // Empty them to prevent their destructor from asserting.
      for (size_t i = 0; i < queues_.size(); i++) {
        queues_[i]->Flush();
      }
    }

    // Join threads explicitly to avoid destruction order issues.
    for (size_t i = 0; i < threads_.size(); i++) delete threads_[i];
    for (size_t i = 0; i < threads_.size(); i++) delete queues_[i];
  }

  void Schedule(std::function<void()> fn) {
    Task t = env_.CreateTask(std::move(fn));
    PerThread* pt = GetPerThread();
    if (pt->pool == this) {
      // Worker thread of this pool, push onto the thread's queue.
      Queue* q = queues_[pt->thread_id];
      t = q->PushFront(std::move(t));
    } else {
      // A free-standing thread (or worker of another pool), push onto a random
      // queue.
      Queue* q = queues_[Rand(&pt->rand) % queues_.size()];
      t = q->PushBack(std::move(t));
    }
    // Note: below we touch this after making w available to worker threads.
    // Strictly speaking, this can lead to a racy-use-after-free. Consider that
    // Schedule is called from a thread that is neither main thread nor a worker
    // thread of this pool. Then, execution of w directly or indirectly
    // completes overall computations, which in turn leads to destruction of
    // this. We expect that such scenario is prevented by program, that is,
    // this is kept alive while any threads can potentially be in Schedule.
    if (!t.f) {
      ec_.Notify(false);
    }
    else {
      env_.ExecuteTask(t);  // Push failed, execute directly.
    }
  }

  void Cancel() {
    cancelled_ = true;
    done_ = true;

    // Let each thread know it's been cancelled.
#ifdef EIGEN_THREAD_ENV_SUPPORTS_CANCELLATION
    for (size_t i = 0; i < threads_.size(); i++) {
      threads_[i]->OnCancel();
    }
#endif

    // Wake up the threads without work to let them exit on their own.
    ec_.Notify(true);
  }

  int NumThreads() const final {
    return static_cast<int>(threads_.size());
  }

  int CurrentThreadId() const final {
    const PerThread* pt =
        const_cast<NonBlockingThreadPoolTempl*>(this)->GetPerThread();
    if (pt->pool == this) {
      return pt->thread_id;
    } else {
      return -1;
    }
  }

 private:
  typedef typename Environment::EnvThread Thread;

  struct PerThread {
    constexpr PerThread() : pool(NULL), rand(0), thread_id(-1) { }
    NonBlockingThreadPoolTempl* pool;  // Parent pool, or null for normal threads.
    uint64_t rand;  // Random generator state.
    int thread_id;  // Worker thread index in pool.
  };

  Environment env_;
  MaxSizeVector<Thread*> threads_;
  MaxSizeVector<Queue*> queues_;
  MaxSizeVector<unsigned> coprimes_;
  MaxSizeVector<EventCount::Waiter> waiters_;
  std::atomic<unsigned> blocked_;
  std::atomic<bool> spinning_;
  std::atomic<bool> done_;
  std::atomic<bool> cancelled_;
  EventCount ec_;

  // Main worker thread loop.
  void WorkerLoop(int thread_id) {
    PerThread* pt = GetPerThread();
    pt->pool = this;
    pt->rand = std::hash<std::thread::id>()(std::this_thread::get_id());
    pt->thread_id = thread_id;
    Queue* q = queues_[thread_id];
    EventCount::Waiter* waiter = &waiters_[thread_id];
    while (!cancelled_) {
      Task t = q->PopFront();
      if (!t.f) {
        t = Steal();
        if (!t.f) {
          // Leave one thread spinning. This reduces latency.
          // TODO(dvyukov): 1000 iterations is based on fair dice roll, tune it.
          // Also, the time it takes to attempt to steal work 1000 times depends
          // on the size of the thread pool. However the speed at which the user
          // of the thread pool submit tasks is independent of the size of the
          // pool. Consider a time based limit instead.
          if (!spinning_ && !spinning_.exchange(true)) {
            for (int i = 0; i < 1000 && !t.f; i++) {
              if (!cancelled_.load(std::memory_order_relaxed)) {
                t = Steal();
              } else {
                return;
              }
            }
            spinning_ = false;
          }
          if (!t.f) {
            if (!WaitForWork(waiter, &t)) {
              return;
            }
          }
        }
      }
      if (t.f) {
        env_.ExecuteTask(t);
      }
    }
  }

  // Steal tries to steal work from other worker threads in best-effort manner.
  Task Steal() {
    PerThread* pt = GetPerThread();
    const size_t size = queues_.size();
    unsigned r = Rand(&pt->rand);
    unsigned inc = coprimes_[r % coprimes_.size()];
    unsigned victim = r % size;
    for (unsigned i = 0; i < size; i++) {
      Task t = queues_[victim]->PopBack();
      if (t.f) {
        return t;
      }
      victim += inc;
      if (victim >= size) {
        victim -= size;
      }
    }
    return Task();
  }

  // WaitForWork blocks until new work is available (returns true), or if it is
  // time to exit (returns false). Can optionally return a task to execute in t
  // (in such case t.f != nullptr on return).
  bool WaitForWork(EventCount::Waiter* waiter, Task* t) {
    eigen_assert(!t->f);
    // We already did best-effort emptiness check in Steal, so prepare for
    // blocking.
    ec_.Prewait(waiter);
    // Now do a reliable emptiness check.
    int victim = NonEmptyQueueIndex();
    if (victim != -1) {
      ec_.CancelWait(waiter);
      if (cancelled_) {
        return false;
      } else {
        *t = queues_[victim]->PopBack();
        return true;
      }
    }
    // Number of blocked threads is used as termination condition.
    // If we are shutting down and all worker threads blocked without work,
    // that's we are done.
    blocked_++;
    if (done_ && blocked_ == threads_.size()) {
      ec_.CancelWait(waiter);
      // Almost done, but need to re-check queues.
      // Consider that all queues are empty and all worker threads are preempted
      // right after incrementing blocked_ above. Now a free-standing thread
      // submits work and calls destructor (which sets done_). If we don't
      // re-check queues, we will exit leaving the work unexecuted.
      if (NonEmptyQueueIndex() != -1) {
        // Note: we must not pop from queues before we decrement blocked_,
        // otherwise the following scenario is possible. Consider that instead
        // of checking for emptiness we popped the only element from queues.
        // Now other worker threads can start exiting, which is bad if the
        // work item submits other work. So we just check emptiness here,
        // which ensures that all worker threads exit at the same time.
        blocked_--;
        return true;
      }
      // Reached stable termination state.
      ec_.Notify(true);
      return false;
    }
    ec_.CommitWait(waiter);
    blocked_--;
    return true;
  }

  int NonEmptyQueueIndex() {
    PerThread* pt = GetPerThread();
    const size_t size = queues_.size();
    unsigned r = Rand(&pt->rand);
    unsigned inc = coprimes_[r % coprimes_.size()];
    unsigned victim = r % size;
    for (unsigned i = 0; i < size; i++) {
      if (!queues_[victim]->Empty()) {
        return victim;
      }
      victim += inc;
      if (victim >= size) {
        victim -= size;
      }
    }
    return -1;
  }

  static EIGEN_STRONG_INLINE PerThread* GetPerThread() {
    EIGEN_THREAD_LOCAL PerThread per_thread_;
    PerThread* pt = &per_thread_;
    return pt;
  }

  static EIGEN_STRONG_INLINE unsigned Rand(uint64_t* state) {
    uint64_t current = *state;
    // Update the internal state
    *state = current * 6364136223846793005ULL + 0xda3e39cb94b95bdbULL;
    // Generate the random output (using the PCG-XSH-RS scheme)
    return static_cast<unsigned>((current ^ (current >> 22)) >> (22 + (current >> 61)));
  }
};

typedef NonBlockingThreadPoolTempl<StlThreadEnvironment> NonBlockingThreadPool;

}  // namespace Eigen

#endif  // EIGEN_CXX11_THREADPOOL_NONBLOCKING_THREAD_POOL_H