[MSLITE][DEVELOP] fix bug of thread pool

mindspore/core/mindrt/src/actor/actormgr.cc

@@ -67,6 +67,10 @@ void ActorMgr::Initialize(bool use_inner_pool, size_t actor_thread_num, size_t m
       inner_pool_->DisableOccupiedActorThread();
       inner_pool_->SetKernelThreadNum(max_thread_num - actor_thread_num);
     }
+    if (inner_pool_ != nullptr) {
+      inner_pool_->SetMaxSpinCount(kDefaultSpinCount);
+      inner_pool_->SetSpinCountMaxValue();
+    }
   }
 }
 
@@ -253,7 +257,6 @@ void ActorMgr::Terminate(const AID &id) {
     std::unique_ptr<MessageBase> msg(new (std::nothrow) MessageBase("Terminate", MessageBase::Type::KTERMINATE));
     MINDRT_OOM_EXIT(msg);
     (void)actor->EnqueMessage(std::move(msg));
-    actor->SetRunningStatus(true);
 
     // Wait actor's thread to finish.
     actor->Await();

mindspore/core/mindrt/src/thread/actor_threadpool.cc

@@ -41,7 +41,7 @@ void ActorWorker::RunWithSpin() {
     } else {
       YieldAndDeactive();
     }
-    if (spin_count_ >= max_spin_count_) {
+    if (spin_count_ > max_spin_count_) {
       WaitUntilActive();
       spin_count_ = 0;
     }
@@ -58,11 +58,15 @@ bool ActorWorker::RunQueueActorTask() {
   return true;
 }
 
-bool ActorWorker::Active() {
+bool ActorWorker::ActorActive() {
   if (status_ != kThreadIdle) {
     return false;
   }
+  {
+    std::lock_guard<std::mutex> _l(mutex_);
+    active_num_++;
     status_ = kThreadBusy;
+  }
   cond_var_.notify_one();
   return true;
 }
@@ -70,6 +74,7 @@ bool ActorWorker::Active() {
 ActorThreadPool::~ActorThreadPool() {
   // wait until actor queue is empty
   bool terminate = false;
+  int count = 0;
   do {
     {
 #ifdef USE_HQUEUE
@@ -80,9 +85,12 @@ ActorThreadPool::~ActorThreadPool() {
 #endif
     }
     if (!terminate) {
+      for (auto &worker : workers_) {
+        worker->Active();
+      }
       std::this_thread::yield();
     }
-  } while (!terminate);
+  } while (!terminate && count++ < kMaxCount);
   for (auto &worker : workers_) {
     delete worker;
     worker = nullptr;
@@ -124,7 +132,7 @@ void ActorThreadPool::PushActorToQueue(ActorBase *actor) {
   // active one idle actor thread if exist
   for (size_t i = 0; i < actor_thread_num_; ++i) {
     auto worker = reinterpret_cast<ActorWorker *>(workers_[i]);
-    if (worker->Active()) {
+    if (worker->ActorActive()) {
       break;
     }
   }

mindspore/core/mindrt/src/thread/actor_threadpool.h

@@ -33,7 +33,7 @@ class ActorThreadPool;
 class ActorWorker : public Worker {
  public:
   void CreateThread(ActorThreadPool *pool);
-  bool Active();
+  bool ActorActive();
 
  private:
   void RunWithSpin();

mindspore/core/mindrt/src/thread/threadpool.cc

@@ -66,7 +66,7 @@ void Worker::Run() {
     } else {
       YieldAndDeactive();
     }
-    if (spin_count_ >= max_spin_count_) {
+    if (spin_count_ > max_spin_count_) {
       WaitUntilActive();
       spin_count_ = 0;
     }
@@ -96,7 +96,8 @@ void Worker::YieldAndDeactive() {
 
 void Worker::WaitUntilActive() {
   std::unique_lock<std::mutex> _l(mutex_);
-  cond_var_.wait(_l, [&] { return status_ == kThreadBusy || !alive_; });
+  cond_var_.wait(_l, [&] { return status_ == kThreadBusy || active_num_ > 0 || !alive_; });
+  active_num_--;
 }
 
 void Worker::set_scale(float lhs_scale, float rhs_scale) {
@@ -114,6 +115,15 @@ void Worker::Active(Task *task, int task_id) {
   cond_var_.notify_one();
 }
 
+void Worker::Active() {
+  {
+    std::lock_guard<std::mutex> _l(mutex_);
+    active_num_++;
+    status_ = kThreadBusy;
+  }
+  cond_var_.notify_one();
+}
+
 bool Worker::available() {
   int expected = kThreadIdle;
   return status_.compare_exchange_strong(expected, kThreadHeld);
@@ -262,6 +272,12 @@ void ThreadPool::ActiveWorkers(const std::vector<Worker *> &workers, Task *task,
   }
 }
 
+void ThreadPool::ActiveWorkers() const {
+  for (auto &worker : workers_) {
+    worker->Active();
+  }
+}
+
 Worker *ThreadPool::CurrentWorker() const {
   for (const auto &worker : workers_) {
     if (worker->thread_id() == std::this_thread::get_id()) {
@@ -332,6 +348,20 @@ void ThreadPool::SetSpinCountMinValue() {
   return;
 }
 
+void ThreadPool::SetMaxSpinCount(int spin_count) {
+  if (spin_count <= 0) {
+    return;
+  }
+  max_spin_count_ = spin_count;
+}
+
+void ThreadPool::SetMinSpinCount(int spin_count) {
+  if (spin_count <= 0) {
+    return;
+  }
+  min_spin_count_ = spin_count;
+}
+
 ThreadPool *ThreadPool::CreateThreadPool(size_t thread_num, const std::vector<int> &core_list) {
   ThreadPool *pool = new (std::nothrow) ThreadPool();
   if (pool == nullptr) {

mindspore/core/mindrt/src/thread/threadpool.h

@@ -30,7 +30,8 @@
 
 namespace mindspore {
 constexpr int kDefaultSpinCount = 300000;
-constexpr int kMinSpinCount = 3000;
+constexpr int kMaxCount = 30000;
+constexpr int kMinSpinCount = 1;
 constexpr int kDefaultFrequency = 1;
 constexpr float kMaxScale = 1.;
 
@@ -61,6 +62,8 @@ class Worker {
   void CreateThread();
   // assign task and then activate thread
   void Active(Task *task, int task_id);
+  // activate thread
+  void Active();
   // whether or not it is idle and marked as held
   bool available();
   // assigns task first before running
@@ -93,6 +96,7 @@ class Worker {
   cpu_set_t mask_;
 #endif
   std::atomic_int status_{kThreadBusy};
+  std::atomic_int active_num_{0};
 
   std::mutex mutex_;
   std::condition_variable cond_var_;
@@ -103,7 +107,7 @@ class Worker {
   float rhs_scale_{kMaxScale};
   int frequency_{kDefaultFrequency};
   int spin_count_{0};
-  int max_spin_count_{kDefaultSpinCount};
+  int max_spin_count_{kMinSpinCount};
 };
 
 class ThreadPool {
@@ -124,8 +128,9 @@ class ThreadPool {
   size_t GetKernelThreadNum() const { return kernel_thread_num_; }
   void SetSpinCountMaxValue();
   void SetSpinCountMinValue();
-  void SetMaxSpinCount(int spin_count) { max_spin_count_ = spin_count; }
-  void SetMinSpinCount(int spin_count) { min_spin_count_ = spin_count; }
+  void SetMaxSpinCount(int spin_count);
+  void SetMinSpinCount(int spin_count);
+  void ActiveWorkers() const;
 
  protected:
   ThreadPool() = default;

mindspore/lite/test/config/models_onnx_fp16.cfg

@@ -61,7 +61,7 @@ ml_ei_facedetection.onnx 2
 #ml_video_edit_art_generate.onnx #mul operator overflows, not suitable for fp16
 #ml_voice_detect.onnx #conv operator overflows, not suitable for fp16
 #ml_location_lane_counter.onnx has very small values during op computation (<1e-6), which causes the precision variation
-ml_location_lane_counter.onnx 7.5
+ml_location_lane_counter.onnx 8
 ml_location_lane_counter0.onnx 1.0
 #The encoder an decoder model are used in ml_asr scene, both have value overflow. Not suitable for fp16.
 #But added for guarding process.