From 002cf0d1c979857e057879d8c84b92439dbcc90d Mon Sep 17 00:00:00 2001 From: Benoit Steiner Date: Tue, 22 Mar 2016 15:24:23 -0700 Subject: Use a single Barrier instead of a collection of Notifications to reduce the thread synchronization overhead --- .../CXX11/src/Tensor/TensorDeviceThreadPool.h | 75 +++++++++++++++++----- 1 file changed, 60 insertions(+), 15 deletions(-) (limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h') diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h index dcbef5b03..e4165bbf8 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h @@ -118,47 +118,82 @@ class ThreadPool : public ThreadPoolInterface { }; -// 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. -class Notification { +// Barrier is an object that allows one or more threads to wait until +// Notify has been called a specified number of times. +class Barrier { public: - Notification() : notified_(false) {} - ~Notification() {} + Barrier(unsigned int count) : state_(count << 1), notified_(false) { + eigen_assert(((count << 1) >> 1) == count); + } + ~Barrier() { + eigen_assert((state_>>1) == 0); + } void Notify() { + unsigned int v = state_.fetch_sub(2, std::memory_order_acq_rel) - 2; + if (v != 1) { + eigen_assert(((v + 2) & ~1) != 0); + return; // either count has not dropped to 0, or waiter is not waiting + } std::unique_lock l(mu_); eigen_assert(!notified_); notified_ = true; cv_.notify_all(); } - void WaitForNotification() { + void Wait() { + unsigned int v = state_.fetch_or(1, std::memory_order_acq_rel); + if ((v >> 1) == 0) return; std::unique_lock l(mu_); - cv_.wait(l, [this]() { return notified_; } ); + while (!notified_) { + cv_.wait(l); + } } private: std::mutex mu_; std::condition_variable cv_; + std::atomic 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) {}; +}; + + // Runs an arbitrary function and then calls Notify() on the passed in // Notification. -template struct FunctionWrapper +template struct FunctionWrapperWithNotification { static void run(Notification* n, Function f, Args... args) { f(args...); - n->Notify(); + if (n) { + n->Notify(); + } + } +}; + +template struct FunctionWrapperWithBarrier +{ + static void run(Barrier* b, Function f, Args... args) { + f(args...); + if (b) { + b->Notify(); + } } }; -static EIGEN_STRONG_INLINE void wait_until_ready(Notification* n) { +template +static EIGEN_STRONG_INLINE void wait_until_ready(SyncType* n) { if (n) { - n->WaitForNotification(); + n->Wait(); } } @@ -203,10 +238,20 @@ struct ThreadPoolDevice { EIGEN_STRONG_INLINE Notification* enqueue(Function&& f, Args&&... args) const { Notification* n = new Notification(); std::function func = - std::bind(&FunctionWrapper::run, n, f, args...); + std::bind(&FunctionWrapperWithNotification::run, n, f, args...); pool_->Schedule(func); return n; } + + template + EIGEN_STRONG_INLINE void enqueue_with_barrier(Barrier* b, + Function&& f, + Args&&... args) const { + std::function func = std::bind( + &FunctionWrapperWithBarrier::run, b, f, args...); + pool_->Schedule(func); + } + template EIGEN_STRONG_INLINE void enqueueNoNotification(Function&& f, Args&&... args) const { std::function func = std::bind(f, args...); -- cgit v1.2.3