@smake/eigen 1.1.0 → 1.1.1

package/eigen/Eigen/src/ThreadPool/Barrier.h

@@ -0,0 +1,70 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2018 Rasmus Munk Larsen <rmlarsen@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/.
+
+// Barrier is an object that allows one or more threads to wait until
+// Notify has been called a specified number of times.
+
+#ifndef EIGEN_CXX11_THREADPOOL_BARRIER_H
+#define EIGEN_CXX11_THREADPOOL_BARRIER_H
+
+// IWYU pragma: private
+#include "./InternalHeaderCheck.h"
+
+namespace Eigen {
+
+class Barrier {
+ public:
+  Barrier(unsigned int count) : state_(count << 1), notified_(false) {
+    eigen_plain_assert(((count << 1) >> 1) == count);
+  }
+  ~Barrier() { eigen_plain_assert((state_ >> 1) == 0); }
+
+  void Notify() {
+    unsigned int v = state_.fetch_sub(2, std::memory_order_acq_rel) - 2;
+    if (v != 1) {
+      // Clear the lowest bit (waiter flag) and check that the original state
+      // value was not zero. If it was zero, it means that notify was called
+      // more times than the original count.
+      eigen_plain_assert(((v + 2) & ~1) != 0);
+      return;  // either count has not dropped to 0, or waiter is not waiting
+    }
+    EIGEN_MUTEX_LOCK l(mu_);
+    eigen_plain_assert(!notified_);
+    notified_ = true;
+    cv_.notify_all();
+  }
+
+  void Wait() {
+    unsigned int v = state_.fetch_or(1, std::memory_order_acq_rel);
+    if ((v >> 1) == 0) return;
+    EIGEN_MUTEX_LOCK l(mu_);
+    while (!notified_) {
+      cv_.wait(l);
+    }
+  }
+
+ private:
+  EIGEN_MUTEX mu_;
+  EIGEN_CONDVAR cv_;
+  std::atomic<unsigned int> state_;  // low bit is waiter flag
+  bool notified_;
+};
+
+// Notification is an object that allows a user to to wait for another
+// thread to signal a notification that an event has occurred.
+//
+// Multiple threads can wait on the same Notification object,
+// but only one caller must call Notify() on the object.
+struct Notification : Barrier {
+  Notification() : Barrier(1){};
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_BARRIER_H

package/eigen/Eigen/src/ThreadPool/CoreThreadPoolDevice.h

@@ -0,0 +1,336 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2023 Charlie Schlosser <cs.schlosser@gmail.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_CORE_THREAD_POOL_DEVICE_H
+#define EIGEN_CORE_THREAD_POOL_DEVICE_H
+
+namespace Eigen {
+
+// CoreThreadPoolDevice provides an easy-to-understand Device for parallelizing Eigen Core expressions with
+// Threadpool. Expressions are recursively split evenly until the evaluation cost is less than the threshold for
+// delegating the task to a thread.
+/*
+                 a
+                / \
+               /   \
+              /     \
+             /       \
+            /         \
+           /           \
+          /             \
+         a               e
+        / \             / \
+       /   \           /   \
+      /     \         /     \
+     a       c       e       g
+    / \     / \     / \     / \
+   /   \   /   \   /   \   /   \
+  a     b c     d e     f g     h
+*/
+// Each task descends the binary tree to the left, delegates the right task to a new thread, and continues to the
+// left. This ensures that work is evenly distributed to the thread pool as quickly as possible and minimizes the number
+// of tasks created during the evaluation. Consider an expression that is divided into 8 chunks. The
+// primary task 'a' creates tasks 'e' 'c' and 'b', and executes its portion of the expression at the bottom of the
+// tree. Likewise, task 'e' creates tasks 'g' and 'f', and executes its portion of the expression.
+
+struct CoreThreadPoolDevice {
+  using Task = std::function<void()>;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoreThreadPoolDevice(ThreadPool& pool, float threadCostThreshold = 3e-5f)
+      : m_pool(pool) {
+    eigen_assert(threadCostThreshold >= 0.0f && "threadCostThreshold must be non-negative");
+    m_costFactor = threadCostThreshold;
+  }
+
+  template <int PacketSize>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int calculateLevels(Index size, float cost) const {
+    eigen_assert(cost >= 0.0f && "cost must be non-negative");
+    Index numOps = size / PacketSize;
+    int actualThreads = numOps < m_pool.NumThreads() ? static_cast<int>(numOps) : m_pool.NumThreads();
+    float totalCost = static_cast<float>(numOps) * cost;
+    float idealThreads = totalCost * m_costFactor;
+    if (idealThreads < static_cast<float>(actualThreads)) {
+      idealThreads = numext::maxi(idealThreads, 1.0f);
+      actualThreads = numext::mini(actualThreads, static_cast<int>(idealThreads));
+    }
+    int maxLevel = internal::log2_ceil(actualThreads);
+    return maxLevel;
+  }
+
+// MSVC does not like inlining parallelForImpl
+#if EIGEN_COMP_MSVC && !EIGEN_COMP_CLANG
+#define EIGEN_PARALLEL_FOR_INLINE
+#else
+#define EIGEN_PARALLEL_FOR_INLINE EIGEN_STRONG_INLINE
+#endif
+
+  template <typename UnaryFunctor, int PacketSize>
+  EIGEN_DEVICE_FUNC EIGEN_PARALLEL_FOR_INLINE void parallelForImpl(Index begin, Index end, UnaryFunctor& f,
+                                                                   Barrier& barrier, int level) {
+    while (level > 0) {
+      level--;
+      Index size = end - begin;
+      eigen_assert(size % PacketSize == 0 && "this function assumes size is a multiple of PacketSize");
+      Index mid = begin + numext::round_down(size >> 1, PacketSize);
+      Task right = [this, mid, end, &f, &barrier, level]() {
+        parallelForImpl<UnaryFunctor, PacketSize>(mid, end, f, barrier, level);
+      };
+      m_pool.Schedule(std::move(right));
+      end = mid;
+    }
+    for (Index i = begin; i < end; i += PacketSize) f(i);
+    barrier.Notify();
+  }
+
+  template <typename BinaryFunctor, int PacketSize>
+  EIGEN_DEVICE_FUNC EIGEN_PARALLEL_FOR_INLINE void parallelForImpl(Index outerBegin, Index outerEnd, Index innerBegin,
+                                                                   Index innerEnd, BinaryFunctor& f, Barrier& barrier,
+                                                                   int level) {
+    while (level > 0) {
+      level--;
+      Index outerSize = outerEnd - outerBegin;
+      if (outerSize > 1) {
+        Index outerMid = outerBegin + (outerSize >> 1);
+        Task right = [this, &f, &barrier, outerMid, outerEnd, innerBegin, innerEnd, level]() {
+          parallelForImpl<BinaryFunctor, PacketSize>(outerMid, outerEnd, innerBegin, innerEnd, f, barrier, level);
+        };
+        m_pool.Schedule(std::move(right));
+        outerEnd = outerMid;
+      } else {
+        Index innerSize = innerEnd - innerBegin;
+        eigen_assert(innerSize % PacketSize == 0 && "this function assumes innerSize is a multiple of PacketSize");
+        Index innerMid = innerBegin + numext::round_down(innerSize >> 1, PacketSize);
+        Task right = [this, &f, &barrier, outerBegin, outerEnd, innerMid, innerEnd, level]() {
+          parallelForImpl<BinaryFunctor, PacketSize>(outerBegin, outerEnd, innerMid, innerEnd, f, barrier, level);
+        };
+        m_pool.Schedule(std::move(right));
+        innerEnd = innerMid;
+      }
+    }
+    for (Index outer = outerBegin; outer < outerEnd; outer++)
+      for (Index inner = innerBegin; inner < innerEnd; inner += PacketSize) f(outer, inner);
+    barrier.Notify();
+  }
+
+#undef EIGEN_PARALLEL_FOR_INLINE
+
+  template <typename UnaryFunctor, int PacketSize>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void parallelFor(Index begin, Index end, UnaryFunctor& f, float cost) {
+    Index size = end - begin;
+    int maxLevel = calculateLevels<PacketSize>(size, cost);
+    Barrier barrier(1 << maxLevel);
+    parallelForImpl<UnaryFunctor, PacketSize>(begin, end, f, barrier, maxLevel);
+    barrier.Wait();
+  }
+
+  template <typename BinaryFunctor, int PacketSize>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void parallelFor(Index outerBegin, Index outerEnd, Index innerBegin,
+                                                         Index innerEnd, BinaryFunctor& f, float cost) {
+    Index outerSize = outerEnd - outerBegin;
+    Index innerSize = innerEnd - innerBegin;
+    Index size = outerSize * innerSize;
+    int maxLevel = calculateLevels<PacketSize>(size, cost);
+    Barrier barrier(1 << maxLevel);
+    parallelForImpl<BinaryFunctor, PacketSize>(outerBegin, outerEnd, innerBegin, innerEnd, f, barrier, maxLevel);
+    barrier.Wait();
+  }
+
+  ThreadPool& m_pool;
+  // costFactor is the cost of delegating a task to a thread
+  // the inverse is used to avoid a floating point division
+  float m_costFactor;
+};
+
+// specialization of coefficient-wise assignment loops for CoreThreadPoolDevice
+
+namespace internal {
+
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+struct Kernel;
+#endif
+
+template <typename Kernel>
+struct cost_helper {
+  using SrcEvaluatorType = typename Kernel::SrcEvaluatorType;
+  using DstEvaluatorType = typename Kernel::DstEvaluatorType;
+  using SrcXprType = typename SrcEvaluatorType::XprType;
+  using DstXprType = typename DstEvaluatorType::XprType;
+  static constexpr Index Cost = functor_cost<SrcXprType>::Cost + functor_cost<DstXprType>::Cost;
+};
+
+template <typename Kernel>
+struct dense_assignment_loop_with_device<Kernel, CoreThreadPoolDevice, DefaultTraversal, NoUnrolling> {
+  static constexpr Index XprEvaluationCost = cost_helper<Kernel>::Cost;
+  struct AssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE AssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index outer, Index inner) {
+      this->assignCoeffByOuterInner(outer, inner);
+    }
+  };
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(Kernel& kernel, CoreThreadPoolDevice& device) {
+    const Index innerSize = kernel.innerSize();
+    const Index outerSize = kernel.outerSize();
+    constexpr float cost = static_cast<float>(XprEvaluationCost);
+    AssignmentFunctor functor(kernel);
+    device.template parallelFor<AssignmentFunctor, 1>(0, outerSize, 0, innerSize, functor, cost);
+  }
+};
+
+template <typename Kernel>
+struct dense_assignment_loop_with_device<Kernel, CoreThreadPoolDevice, DefaultTraversal, InnerUnrolling> {
+  using DstXprType = typename Kernel::DstEvaluatorType::XprType;
+  static constexpr Index XprEvaluationCost = cost_helper<Kernel>::Cost, InnerSize = DstXprType::InnerSizeAtCompileTime;
+  struct AssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE AssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index outer) {
+      copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, 0, InnerSize>::run(*this, outer);
+    }
+  };
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(Kernel& kernel, CoreThreadPoolDevice& device) {
+    const Index outerSize = kernel.outerSize();
+    AssignmentFunctor functor(kernel);
+    constexpr float cost = static_cast<float>(XprEvaluationCost) * static_cast<float>(InnerSize);
+    device.template parallelFor<AssignmentFunctor, 1>(0, outerSize, functor, cost);
+  }
+};
+
+template <typename Kernel>
+struct dense_assignment_loop_with_device<Kernel, CoreThreadPoolDevice, InnerVectorizedTraversal, NoUnrolling> {
+  using PacketType = typename Kernel::PacketType;
+  static constexpr Index XprEvaluationCost = cost_helper<Kernel>::Cost, PacketSize = unpacket_traits<PacketType>::size,
+                         SrcAlignment = Kernel::AssignmentTraits::SrcAlignment,
+                         DstAlignment = Kernel::AssignmentTraits::DstAlignment;
+  struct AssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE AssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index outer, Index inner) {
+      this->template assignPacketByOuterInner<Unaligned, Unaligned, PacketType>(outer, inner);
+    }
+  };
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(Kernel& kernel, CoreThreadPoolDevice& device) {
+    const Index innerSize = kernel.innerSize();
+    const Index outerSize = kernel.outerSize();
+    const float cost = static_cast<float>(XprEvaluationCost) * static_cast<float>(innerSize);
+    AssignmentFunctor functor(kernel);
+    device.template parallelFor<AssignmentFunctor, PacketSize>(0, outerSize, 0, innerSize, functor, cost);
+  }
+};
+
+template <typename Kernel>
+struct dense_assignment_loop_with_device<Kernel, CoreThreadPoolDevice, InnerVectorizedTraversal, InnerUnrolling> {
+  using PacketType = typename Kernel::PacketType;
+  using DstXprType = typename Kernel::DstEvaluatorType::XprType;
+  static constexpr Index XprEvaluationCost = cost_helper<Kernel>::Cost, PacketSize = unpacket_traits<PacketType>::size,
+                         SrcAlignment = Kernel::AssignmentTraits::SrcAlignment,
+                         DstAlignment = Kernel::AssignmentTraits::DstAlignment,
+                         InnerSize = DstXprType::InnerSizeAtCompileTime;
+  struct AssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE AssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index outer) {
+      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, InnerSize, SrcAlignment, DstAlignment>::run(*this, outer);
+    }
+  };
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(Kernel& kernel, CoreThreadPoolDevice& device) {
+    const Index outerSize = kernel.outerSize();
+    constexpr float cost = static_cast<float>(XprEvaluationCost) * static_cast<float>(InnerSize);
+    AssignmentFunctor functor(kernel);
+    device.template parallelFor<AssignmentFunctor, PacketSize>(0, outerSize, functor, cost);
+  }
+};
+
+template <typename Kernel>
+struct dense_assignment_loop_with_device<Kernel, CoreThreadPoolDevice, SliceVectorizedTraversal, NoUnrolling> {
+  using Scalar = typename Kernel::Scalar;
+  using PacketType = typename Kernel::PacketType;
+  static constexpr Index XprEvaluationCost = cost_helper<Kernel>::Cost, PacketSize = unpacket_traits<PacketType>::size;
+  struct PacketAssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketAssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index outer, Index inner) {
+      this->template assignPacketByOuterInner<Unaligned, Unaligned, PacketType>(outer, inner);
+    }
+  };
+  struct ScalarAssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ScalarAssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index outer) {
+      const Index innerSize = this->innerSize();
+      const Index packetAccessSize = numext::round_down(innerSize, PacketSize);
+      for (Index inner = packetAccessSize; inner < innerSize; inner++) this->assignCoeffByOuterInner(outer, inner);
+    }
+  };
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(Kernel& kernel, CoreThreadPoolDevice& device) {
+    const Index outerSize = kernel.outerSize();
+    const Index innerSize = kernel.innerSize();
+    const Index packetAccessSize = numext::round_down(innerSize, PacketSize);
+    constexpr float packetCost = static_cast<float>(XprEvaluationCost);
+    const float scalarCost = static_cast<float>(XprEvaluationCost) * static_cast<float>(innerSize - packetAccessSize);
+    PacketAssignmentFunctor packetFunctor(kernel);
+    ScalarAssignmentFunctor scalarFunctor(kernel);
+    device.template parallelFor<PacketAssignmentFunctor, PacketSize>(0, outerSize, 0, packetAccessSize, packetFunctor,
+                                                                     packetCost);
+    device.template parallelFor<ScalarAssignmentFunctor, 1>(0, outerSize, scalarFunctor, scalarCost);
+  };
+};
+
+template <typename Kernel>
+struct dense_assignment_loop_with_device<Kernel, CoreThreadPoolDevice, LinearTraversal, NoUnrolling> {
+  static constexpr Index XprEvaluationCost = cost_helper<Kernel>::Cost;
+  struct AssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE AssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index index) { this->assignCoeff(index); }
+  };
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(Kernel& kernel, CoreThreadPoolDevice& device) {
+    const Index size = kernel.size();
+    constexpr float cost = static_cast<float>(XprEvaluationCost);
+    AssignmentFunctor functor(kernel);
+    device.template parallelFor<AssignmentFunctor, 1>(0, size, functor, cost);
+  }
+};
+
+template <typename Kernel>
+struct dense_assignment_loop_with_device<Kernel, CoreThreadPoolDevice, LinearVectorizedTraversal, NoUnrolling> {
+  using Scalar = typename Kernel::Scalar;
+  using PacketType = typename Kernel::PacketType;
+  static constexpr Index XprEvaluationCost = cost_helper<Kernel>::Cost,
+                         RequestedAlignment = Kernel::AssignmentTraits::LinearRequiredAlignment,
+                         PacketSize = unpacket_traits<PacketType>::size,
+                         DstIsAligned = Kernel::AssignmentTraits::DstAlignment >= RequestedAlignment,
+                         DstAlignment = packet_traits<Scalar>::AlignedOnScalar ? RequestedAlignment
+                                                                               : Kernel::AssignmentTraits::DstAlignment,
+                         SrcAlignment = Kernel::AssignmentTraits::JointAlignment;
+  struct AssignmentFunctor : public Kernel {
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE AssignmentFunctor(Kernel& kernel) : Kernel(kernel) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void operator()(Index index) {
+      this->template assignPacket<DstAlignment, SrcAlignment, PacketType>(index);
+    }
+  };
+  static constexpr bool UsePacketSegment = Kernel::AssignmentTraits::UsePacketSegment;
+  using head_loop =
+      unaligned_dense_assignment_loop<PacketType, DstAlignment, SrcAlignment, UsePacketSegment, DstIsAligned>;
+  using tail_loop = unaligned_dense_assignment_loop<PacketType, DstAlignment, SrcAlignment, UsePacketSegment, false>;
+
+  static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void run(Kernel& kernel, CoreThreadPoolDevice& device) {
+    const Index size = kernel.size();
+    const Index alignedStart =
+        DstIsAligned ? 0 : internal::first_aligned<RequestedAlignment>(kernel.dstDataPtr(), size);
+    const Index alignedEnd = alignedStart + numext::round_down(size - alignedStart, PacketSize);
+
+    head_loop::run(kernel, 0, alignedStart);
+
+    constexpr float cost = static_cast<float>(XprEvaluationCost);
+    AssignmentFunctor functor(kernel);
+    device.template parallelFor<AssignmentFunctor, PacketSize>(alignedStart, alignedEnd, functor, cost);
+
+    tail_loop::run(kernel, alignedEnd, size);
+  }
+};
+
+}  // namespace internal
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CORE_THREAD_POOL_DEVICE_H

package/eigen/Eigen/src/ThreadPool/EventCount.h

@@ -0,0 +1,241 @@
+// 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_EVENTCOUNT_H
+#define EIGEN_CXX11_THREADPOOL_EVENTCOUNT_H
+
+// IWYU pragma: private
+#include "./InternalHeaderCheck.h"
+
+namespace Eigen {
+
+// EventCount allows to wait for arbitrary predicates in non-blocking
+// algorithms. Think of condition variable, but wait predicate does not need to
+// be protected by a mutex. Usage:
+// Waiting thread does:
+//
+//   if (predicate)
+//     return act();
+//   EventCount::Waiter& w = waiters[my_index];
+//   ec.Prewait(&w);
+//   if (predicate) {
+//     ec.CancelWait(&w);
+//     return act();
+//   }
+//   ec.CommitWait(&w);
+//
+// Notifying thread does:
+//
+//   predicate = true;
+//   ec.Notify(true);
+//
+// Notify is cheap if there are no waiting threads. Prewait/CommitWait are not
+// cheap, but they are executed only if the preceding predicate check has
+// failed.
+//
+// Algorithm outline:
+// There are two main variables: predicate (managed by user) and state_.
+// Operation closely resembles Dekker mutual algorithm:
+// https://en.wikipedia.org/wiki/Dekker%27s_algorithm
+// Waiting thread sets state_ then checks predicate, Notifying thread sets
+// predicate then checks state_. Due to seq_cst fences in between these
+// operations it is guaranteed than either waiter will see predicate change
+// and won't block, or notifying thread will see state_ change and will unblock
+// the waiter, or both. But it can't happen that both threads don't see each
+// other changes, which would lead to deadlock.
+class EventCount {
+ public:
+  class Waiter;
+
+  EventCount(MaxSizeVector<Waiter>& waiters) : state_(kStackMask), waiters_(waiters) {
+    eigen_plain_assert(waiters.size() < (1 << kWaiterBits) - 1);
+  }
+
+  EventCount(const EventCount&) = delete;
+  void operator=(const EventCount&) = delete;
+
+  ~EventCount() {
+    // Ensure there are no waiters.
+    eigen_plain_assert(state_.load() == kStackMask);
+  }
+
+  // Prewait prepares for waiting.
+  // After calling Prewait, the thread must re-check the wait predicate
+  // and then call either CancelWait or CommitWait.
+  void Prewait() {
+    uint64_t state = state_.load(std::memory_order_relaxed);
+    for (;;) {
+      CheckState(state);
+      uint64_t newstate = state + kWaiterInc;
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate, std::memory_order_seq_cst)) return;
+    }
+  }
+
+  // CommitWait commits waiting after Prewait.
+  void CommitWait(Waiter* w) {
+    eigen_plain_assert((w->epoch & ~kEpochMask) == 0);
+    w->state = Waiter::kNotSignaled;
+    const uint64_t me = (w - &waiters_[0]) | w->epoch;
+    uint64_t state = state_.load(std::memory_order_seq_cst);
+    for (;;) {
+      CheckState(state, true);
+      uint64_t newstate;
+      if ((state & kSignalMask) != 0) {
+        // Consume the signal and return immediately.
+        newstate = state - kWaiterInc - kSignalInc;
+      } else {
+        // Remove this thread from pre-wait counter and add to the waiter stack.
+        newstate = ((state & kWaiterMask) - kWaiterInc) | me;
+        w->next.store(state & (kStackMask | kEpochMask), std::memory_order_relaxed);
+      }
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate, std::memory_order_acq_rel)) {
+        if ((state & kSignalMask) == 0) {
+          w->epoch += kEpochInc;
+          Park(w);
+        }
+        return;
+      }
+    }
+  }
+
+  // CancelWait cancels effects of the previous Prewait call.
+  void CancelWait() {
+    uint64_t state = state_.load(std::memory_order_relaxed);
+    for (;;) {
+      CheckState(state, true);
+      uint64_t newstate = state - kWaiterInc;
+      // We don't know if the thread was also notified or not,
+      // so we should not consume a signal unconditionally.
+      // Only if number of waiters is equal to number of signals,
+      // we know that the thread was notified and we must take away the signal.
+      if (((state & kWaiterMask) >> kWaiterShift) == ((state & kSignalMask) >> kSignalShift)) newstate -= kSignalInc;
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate, std::memory_order_acq_rel)) return;
+    }
+  }
+
+  // Notify wakes one or all waiting threads.
+  // Must be called after changing the associated wait predicate.
+  void Notify(bool notifyAll) {
+    std::atomic_thread_fence(std::memory_order_seq_cst);
+    uint64_t state = state_.load(std::memory_order_acquire);
+    for (;;) {
+      CheckState(state);
+      const uint64_t waiters = (state & kWaiterMask) >> kWaiterShift;
+      const uint64_t signals = (state & kSignalMask) >> kSignalShift;
+      // Easy case: no waiters.
+      if ((state & kStackMask) == kStackMask && waiters == signals) return;
+      uint64_t newstate;
+      if (notifyAll) {
+        // Empty wait stack and set signal to number of pre-wait threads.
+        newstate = (state & kWaiterMask) | (waiters << kSignalShift) | kStackMask;
+      } else if (signals < waiters) {
+        // There is a thread in pre-wait state, unblock it.
+        newstate = state + kSignalInc;
+      } else {
+        // Pop a waiter from list and unpark it.
+        Waiter* w = &waiters_[state & kStackMask];
+        uint64_t next = w->next.load(std::memory_order_relaxed);
+        newstate = (state & (kWaiterMask | kSignalMask)) | next;
+      }
+      CheckState(newstate);
+      if (state_.compare_exchange_weak(state, newstate, std::memory_order_acq_rel)) {
+        if (!notifyAll && (signals < waiters)) return;  // unblocked pre-wait thread
+        if ((state & kStackMask) == kStackMask) return;
+        Waiter* w = &waiters_[state & kStackMask];
+        if (!notifyAll) w->next.store(kStackMask, std::memory_order_relaxed);
+        Unpark(w);
+        return;
+      }
+    }
+  }
+
+ private:
+  // State_ layout:
+  // - low kWaiterBits is a stack of waiters committed wait
+  //   (indexes in waiters_ array are used as stack elements,
+  //   kStackMask means empty stack).
+  // - next kWaiterBits is count of waiters in prewait state.
+  // - next kWaiterBits is count of pending signals.
+  // - remaining bits are ABA counter for the stack.
+  //   (stored in Waiter node and incremented on push).
+  static const uint64_t kWaiterBits = 14;
+  static const uint64_t kStackMask = (1ull << kWaiterBits) - 1;
+  static const uint64_t kWaiterShift = kWaiterBits;
+  static const uint64_t kWaiterMask = ((1ull << kWaiterBits) - 1) << kWaiterShift;
+  static const uint64_t kWaiterInc = 1ull << kWaiterShift;
+  static const uint64_t kSignalShift = 2 * kWaiterBits;
+  static const uint64_t kSignalMask = ((1ull << kWaiterBits) - 1) << kSignalShift;
+  static const uint64_t kSignalInc = 1ull << kSignalShift;
+  static const uint64_t kEpochShift = 3 * kWaiterBits;
+  static const uint64_t kEpochBits = 64 - kEpochShift;
+  static const uint64_t kEpochMask = ((1ull << kEpochBits) - 1) << kEpochShift;
+  static const uint64_t kEpochInc = 1ull << kEpochShift;
+
+ public:
+  class Waiter {
+    friend class EventCount;
+
+    enum State {
+      kNotSignaled,
+      kWaiting,
+      kSignaled,
+    };
+
+    EIGEN_ALIGN_TO_AVOID_FALSE_SHARING std::atomic<uint64_t> next{kStackMask};
+    EIGEN_MUTEX mu;
+    EIGEN_CONDVAR cv;
+    uint64_t epoch{0};
+    unsigned state{kNotSignaled};
+  };
+
+ private:
+  static void CheckState(uint64_t state, bool waiter = false) {
+    static_assert(kEpochBits >= 20, "not enough bits to prevent ABA problem");
+    const uint64_t waiters = (state & kWaiterMask) >> kWaiterShift;
+    const uint64_t signals = (state & kSignalMask) >> kSignalShift;
+    eigen_plain_assert(waiters >= signals);
+    eigen_plain_assert(waiters < (1 << kWaiterBits) - 1);
+    eigen_plain_assert(!waiter || waiters > 0);
+    (void)waiters;
+    (void)signals;
+  }
+
+  void Park(Waiter* w) {
+    EIGEN_MUTEX_LOCK lock(w->mu);
+    while (w->state != Waiter::kSignaled) {
+      w->state = Waiter::kWaiting;
+      w->cv.wait(lock);
+    }
+  }
+
+  void Unpark(Waiter* w) {
+    for (Waiter* next; w; w = next) {
+      uint64_t wnext = w->next.load(std::memory_order_relaxed) & kStackMask;
+      next = wnext == kStackMask ? nullptr : &waiters_[internal::convert_index<size_t>(wnext)];
+      unsigned state;
+      {
+        EIGEN_MUTEX_LOCK lock(w->mu);
+        state = w->state;
+        w->state = Waiter::kSignaled;
+      }
+      // Avoid notifying if it wasn't waiting.
+      if (state == Waiter::kWaiting) w->cv.notify_one();
+    }
+  }
+
+  std::atomic<uint64_t> state_;
+  MaxSizeVector<Waiter>& waiters_;
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_CXX11_THREADPOOL_EVENTCOUNT_H