File: thread_pool.cpp

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (167 lines) | stat: -rw-r--r-- 4,143 bytes parent folder | download
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
#include <torch/csrc/lazy/core/thread_pool.h>

#include <c10/util/Logging.h>
#include <c10/util/irange.h>
#include <torch/csrc/lazy/core/config.h>
#include <torch/csrc/lazy/core/metrics.h>

#include <condition_variable>
#include <deque>
#include <exception>
#include <mutex>

namespace torch {
namespace lazy {
namespace {

class ThreadPool {
 public:
  explicit ThreadPool(size_t num_threads) {
    threads_.reserve(num_threads);
    for (const auto i : c10::irange(num_threads)) {
      (void)i; // Suppress unused variable warning
      threads_.emplace_back([this]() { Worker(); });
    }
  }

  ~ThreadPool() {
    {
      std::lock_guard<std::mutex> lock(mutex_);
      exiting_ = true;
      cv_.notify_all();
    }
    for (auto& thread : threads_) {
      thread.join();
    }
  }

  void Schedule(std::function<void()> closure) {
    // If we have more work scheduled than waiting worker threads, just schedule
    // it on a separate thread. This prevents tricky thread-pool-size-deadlocks
    // caused by an undersized thread pool and closures that end up doing sync
    // waits on the pool threads.
    bool scheduled = false;
    {
      std::lock_guard<std::mutex> lock(mutex_);
      if (work_.size() < waiting_) {
        work_.emplace_back(std::move(closure));
        scheduled = true;
      }
    }
    if (scheduled) {
      cv_.notify_one();
    } else {
      // NOLINTNEXTLINE(bugprone-use-after-move)
      ScheduleOnThread(std::move(closure));
    }
  }

 private:
  void Worker() {
    while (true) {
      std::function<void()> closure = GetWork();
      if (closure == nullptr) {
        break;
      }
      try {
        closure();
      } catch (const std::exception& ex) {
        TORCH_LAZY_COUNTER("ThreadPoolException", 1);
        LOG(ERROR) << "Exception from running thread pool closure: "
                   << ex.what();
      }
    }
  }

  void ScheduleOnThread(std::function<void()> closure) {
    std::thread thread(std::move(closure));
    thread.detach();
  }

  std::function<void()> GetWork() {
    std::unique_lock<std::mutex> lock(mutex_);
    ++waiting_;
    cv_.wait(lock, [this] { return exiting_ || !work_.empty(); });
    --waiting_;
    if (work_.empty()) {
      return nullptr;
    }
    std::function<void()> closure(std::move(work_.front()));
    work_.pop_front();
    return closure;
  }

  std::vector<std::thread> threads_;
  std::mutex mutex_;
  std::condition_variable cv_;
  bool exiting_ = false;
  std::deque<std::function<void()>> work_;
  size_t waiting_ = 0;
};

ThreadPool* GetIoThreadPool() {
  static ThreadPool* pool =
      new ThreadPool(FLAGS_torch_lazy_io_thread_pool_size);
  return pool;
}

} // namespace

class Completion::Data {
 public:
  void Wait() {
    std::unique_lock<std::mutex> lock(mutex_);
    cv_.wait(lock, [this] { return completed_; });
    if (exptr_ != nullptr) {
      std::rethrow_exception(exptr_);
    }
  }

  static std::function<void()> GetCompleter(
      std::shared_ptr<Data> data,
      std::function<void()> closure) {
    auto closure_wrapper = [closure = std::move(closure), data]() {
      std::exception_ptr exptr;
      try {
        closure();
      } catch (...) {
        exptr = std::current_exception();
      }
      data->Complete(exptr);
    };
    return closure_wrapper;
  }

 private:
  void Complete(std::exception_ptr exptr) {
    std::lock_guard<std::mutex> lock(mutex_);
    exptr_ = std::move(exptr);
    completed_ = true;
    cv_.notify_all();
  }

  std::mutex mutex_;
  std::condition_variable cv_;
  bool completed_ = false;
  std::exception_ptr exptr_;
};

Completion::Completion(std::shared_ptr<Data> data) : data_(std::move(data)) {}

void Completion::Wait() {
  data_->Wait();
}

void ScheduleIoClosure(std::function<void()> closure) {
  GetIoThreadPool()->Schedule(std::move(closure));
}

Completion ScheduleIoClosureWithCompletion(std::function<void()> closure) {
  auto data = std::make_shared<Completion::Data>();
  GetIoThreadPool()->Schedule(
      Completion::Data::GetCompleter(data, std::move(closure)));
  return Completion(std::move(data));
}

} // namespace lazy
} // namespace torch