FM / util

/usr/local/lib64/python3.6/site-packages/torch/include/c10/util/

/usr/local/lib64/python3.6/site-packages/torch/include/c10/util

Name	Size	Mode	Actions
accumulate.h	4205	0644	edit dl rm
AlignOf.h	4835	0644	edit dl rm
Array.h	11354	0644	edit dl rm
ArrayRef.h	9058	0644	edit dl rm
Backtrace.h	364	0644	edit dl rm
BFloat16-inl.h	9222	0644	edit dl rm
BFloat16-math.h	5217	0644	edit dl rm
BFloat16.h	2372	0644	edit dl rm
Bitset.h	3414	0644	edit dl rm
C++17.h	13329	0644	edit dl rm
complex.h	17542	0644	edit dl rm
complex_math.h	10902	0644	edit dl rm
complex_utils.h	958	0644	edit dl rm
ConstexprCrc.h	6633	0644	edit dl rm
copysign.h	866	0644	edit dl rm
DeadlockDetection.h	1920	0644	edit dl rm
Deprecated.h	3579	0644	edit dl rm
either.h	6423	0644	edit dl rm
env.h	835	0644	edit dl rm
Exception.h	24723	0644	edit dl rm
ExclusivelyOwned.h	4759	0644	edit dl rm
Flags.h	10056	0644	edit dl rm
flat_hash_map.h	61655	0644	edit dl rm
FunctionRef.h	2301	0644	edit dl rm
Half-inl.h	8674	0644	edit dl rm
Half.h	18965	0644	edit dl rm
hash.h	5084	0644	edit dl rm
IdWrapper.h	2348	0644	edit dl rm
intrusive_ptr.h	35563	0644	edit dl rm
in_place.h	350	0644	edit dl rm
irange.h	2679	0644	edit dl rm
LeftRight.h	6016	0644	edit dl rm
llvmMathExtras.h	29168	0644	edit dl rm
Logging.h	11256	0644	edit dl rm
logging_is_google_glog.h	2031	0644	edit dl rm
logging_is_not_google_glog.h	8271	0644	edit dl rm
MathConstants.h	858	0644	edit dl rm
math_compat.h	7296	0644	edit dl rm
MaybeOwned.h	6689	0644	edit dl rm
Metaprogramming.h	15288	0644	edit dl rm
numa.h	696	0644	edit dl rm
Optional.h	35592	0644	edit dl rm
order_preserving_flat_hash_map.h	65482	0644	edit dl rm
overloaded.h	709	0644	edit dl rm
python_stub.h	56	0644	edit dl rm
qint8.h	472	0644	edit dl rm
qint32.h	319	0644	edit dl rm
quint4x2.h	366	0644	edit dl rm
quint8.h	320	0644	edit dl rm
Registry.h	12242	0644	edit dl rm
reverse_iterator.h	8796	0644	edit dl rm
ScopeExit.h	1345	0644	edit dl rm
signal_handler.h	3154	0644	edit dl rm
SmallBuffer.h	1243	0644	edit dl rm
SmallVector.h	34456	0644	edit dl rm
sparse_bitset.h	26511	0644	edit dl rm
StringUtil.h	4538	0644	edit dl rm
string_utils.h	3989	0644	edit dl rm
string_view.h	20189	0644	edit dl rm
tempfile.h	6029	0644	edit dl rm
ThreadLocal.h	3883	0644	edit dl rm
ThreadLocalDebugInfo.h	2603	0644	edit dl rm
thread_name.h	148	0644	edit dl rm
Type.h	607	0644	edit dl rm
TypeCast.h	6972	0644	edit dl rm
typeid.h	18793	0644	edit dl rm
TypeIndex.h	5251	0644	edit dl rm
TypeList.h	16901	0644	edit dl rm
TypeTraits.h	5368	0644	edit dl rm
Unicode.h	295	0644	edit dl rm
UniqueVoidPtr.h	4117	0644	edit dl rm
Unroll.h	667	0644	edit dl rm
variant.h	95182	0644	edit dl rm
win32-headers.h	858	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/c10/util/LeftRight.h (6016B)

#include <c10/macros/Macros.h>
#include <array>
#include <atomic>
#include <functional>
#include <mutex>
#include <thread>

namespace c10 {

namespace detail {

struct IncrementRAII final {
 public:
  explicit IncrementRAII(std::atomic<int32_t>* counter) : _counter(counter) {
    _counter->fetch_add(1);
  }

~IncrementRAII() {
    _counter->fetch_sub(1);
  }

private:
  std::atomic<int32_t>* _counter;

C10_DISABLE_COPY_AND_ASSIGN(IncrementRAII);
};

} // namespace detail

// LeftRight wait-free readers synchronization primitive
// https://hal.archives-ouvertes.fr/hal-01207881/document
//
// LeftRight is quite easy to use (it can make an arbitrary
// data structure permit wait-free reads), but it has some
// particular performance characteristics you should be aware
// of if you're deciding to use it:
//
//  - Reads still incur an atomic write (this is how LeftRight
//    keeps track of how long it needs to keep around the old
//    data structure)
//
//  - Writes get executed twice, to keep both the left and right
//    versions up to date.  So if your write is expensive or
//    nondeterministic, this is also an inappropriate structure
//
// LeftRight is used fairly rarely in PyTorch's codebase.  If you
// are still not sure if you need it or not, consult your local
// C++ expert.
//
template <class T>
class LeftRight final {
 public:
  template <class... Args>
  explicit LeftRight(const Args&... args)
      : _counters{{{0}, {0}}},
        _foregroundCounterIndex(0),
        _foregroundDataIndex(0),
        _data{{T{args...}, T{args...}}},
        _writeMutex() {}

// Copying and moving would not be threadsafe.
  // Needs more thought and careful design to make that work.
  LeftRight(const LeftRight&) = delete;
  LeftRight(LeftRight&&) noexcept = delete;
  LeftRight& operator=(const LeftRight&) = delete;
  LeftRight& operator=(LeftRight&&) noexcept = delete;

~LeftRight() {
    // wait until any potentially running writers are finished
    { std::unique_lock<std::mutex> lock(_writeMutex); }

// wait until any potentially running readers are finished
    while (_counters[0].load() != 0 || _counters[1].load() != 0) {
      std::this_thread::yield();
    }
  }

template <typename F>
  auto read(F&& readFunc) const -> typename std::result_of<F(const T&)>::type {
    detail::IncrementRAII _increment_counter(
        &_counters[_foregroundCounterIndex.load()]);

return readFunc(_data[_foregroundDataIndex.load()]);
  }

// Throwing an exception in writeFunc is ok but causes the state to be either
  // the old or the new state, depending on if the first or the second call to
  // writeFunc threw.
  template <typename F>
  auto write(F&& writeFunc) -> typename std::result_of<F(T&)>::type {
    std::unique_lock<std::mutex> lock(_writeMutex);

return _write(writeFunc);
  }

private:
  template <class F>
  auto _write(const F& writeFunc) -> typename std::result_of<F(T&)>::type {
    /*
     * Assume, A is in background and B in foreground. In simplified terms, we
     * want to do the following:
     * 1. Write to A (old background)
     * 2. Switch A/B
     * 3. Write to B (new background)
     *
     * More detailed algorithm (explanations on why this is important are below
     * in code):
     * 1. Write to A
     * 2. Switch A/B data pointers
     * 3. Wait until A counter is zero
     * 4. Switch A/B counters
     * 5. Wait until B counter is zero
     * 6. Write to B
     */

auto localDataIndex = _foregroundDataIndex.load();

// 1. Write to A
    _callWriteFuncOnBackgroundInstance(writeFunc, localDataIndex);

// 2. Switch A/B data pointers
    localDataIndex = localDataIndex ^ 1;
    _foregroundDataIndex = localDataIndex;

/*
     * 3. Wait until A counter is zero
     *
     * In the previous write run, A was foreground and B was background.
     * There was a time after switching _foregroundDataIndex (B to foreground)
     * and before switching _foregroundCounterIndex, in which new readers could
     * have read B but incremented A's counter.
     *
     * In this current run, we just switched _foregroundDataIndex (A back to
     * foreground), but before writing to the new background B, we have to make
     * sure A's counter was zero briefly, so all these old readers are gone.
     */
    auto localCounterIndex = _foregroundCounterIndex.load();
    _waitForBackgroundCounterToBeZero(localCounterIndex);

/*
     * 4. Switch A/B counters
     *
     * Now that we know all readers on B are really gone, we can switch the
     * counters and have new readers increment A's counter again, which is the
     * correct counter since they're reading A.
     */
    localCounterIndex = localCounterIndex ^ 1;
    _foregroundCounterIndex = localCounterIndex;

/*
     * 5. Wait until B counter is zero
     *
     * This waits for all the readers on B that came in while both data and
     * counter for B was in foreground, i.e. normal readers that happened
     * outside of that brief gap between switching data and counter.
     */
    _waitForBackgroundCounterToBeZero(localCounterIndex);

// 6. Write to B
    return _callWriteFuncOnBackgroundInstance(writeFunc, localDataIndex);
  }

template <class F>
  auto _callWriteFuncOnBackgroundInstance(
      const F& writeFunc,
      uint8_t localDataIndex) -> typename std::result_of<F(T&)>::type {
    try {
      return writeFunc(_data[localDataIndex ^ 1]);
    } catch (...) {
      // recover invariant by copying from the foreground instance
      _data[localDataIndex ^ 1] = _data[localDataIndex];
      // rethrow
      throw;
    }
  }

void _waitForBackgroundCounterToBeZero(uint8_t counterIndex) {
    while (_counters[counterIndex ^ 1].load() != 0) {
      std::this_thread::yield();
    }
  }

mutable std::array<std::atomic<int32_t>, 2> _counters;
  std::atomic<uint8_t> _foregroundCounterIndex;
  std::atomic<uint8_t> _foregroundDataIndex;
  std::array<T, 2> _data;
  std::mutex _writeMutex;
};

} // namespace c10

cmd: