Workqueue threadnames (#40035)

* add align for WorkQueue

* add spinlock

* merge develop

* merge

* Add EventsWaiter

* Revert "Add EventsWaiter"

This reverts commit e206173aa9be7401b83a53581627bfaf557c8fb2.

* Set thread name for WorkQueue

* Add thread names

* fix ut
Co-authored-by: Nliutiexing <liutiexing@google.com>

paddle/fluid/framework/new_executor/workqueue/nonblocking_threadpool.h

@@ -409,7 +409,7 @@ class ThreadPoolTempl {
       return false;
     }
     platform::RecordEvent("SleepWaitForWork",
-                          platform::TracerEventType::UserDefined, 2);
+                          platform::TracerEventType::UserDefined, 10);
     ec_.CommitWait(waiter);
     blocked_--;
     return true;

paddle/fluid/platform/init.cc

@@ -28,6 +28,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/device/device_wrapper.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/init.h"
+#include "paddle/fluid/platform/os_info.h"
 #include "paddle/fluid/platform/place.h"
 
 #ifdef PADDLE_WITH_XPU
@@ -161,6 +162,8 @@ void LoadCustomDevice(const std::string &library_dir) {
 #endif
 
 void InitDevices() {
+  // set name at the entry point of Paddle
+  platform::SetCurrentThreadName("MainThread");
 // CUPTI attribute should be set before any CUDA context is created (see CUPTI
 // documentation about CUpti_ActivityAttribute).
 #ifdef PADDLE_WITH_CUDA

paddle/fluid/platform/os_info_test.cc

@@ -30,8 +30,7 @@ TEST(ThreadInfo, TestThreadNameUtils) {
   using paddle::platform::GetCurrentThreadName;
   using paddle::platform::SetCurrentThreadName;
   using paddle::platform::GetAllThreadNames;
-  EXPECT_EQ("unset", GetCurrentThreadName());
-  EXPECT_TRUE(SetCurrentThreadName("MainThread"));
+  SetCurrentThreadName("MainThread");
   EXPECT_FALSE(SetCurrentThreadName("MainThread"));
   auto names = GetAllThreadNames();
   EXPECT_TRUE(names.find(GetCurrentThreadStdId()) != names.end());

paddle/fluid/platform/profiler/host_event_recorder.h

@@ -189,7 +189,10 @@ struct ThreadEventSection {
 
 class ThreadEventRecorder {
  public:
-  ThreadEventRecorder() { thread_id_ = GetCurrentThreadSysId(); }
+  ThreadEventRecorder() {
+    thread_id_ = GetCurrentThreadSysId();
+    thread_name_ = GetCurrentThreadName();
+  }
 
   DISABLE_COPY_AND_ASSIGN(ThreadEventRecorder);
 
@@ -202,7 +205,7 @@ class ThreadEventRecorder {
 
   ThreadEventSection GatherEvents() {
     ThreadEventSection thr_sec;
-    thr_sec.thread_name = GetCurrentThreadName();
+    thr_sec.thread_name = thread_name_;
     thr_sec.thread_id = thread_id_;
     thr_sec.events = std::move(base_evt_cntr_.Reduce());
     return thr_sec;
@@ -210,6 +213,7 @@ class ThreadEventRecorder {
 
  private:
   uint64_t thread_id_;
+  std::string thread_name_;
   EventContainer<CommonEvent> base_evt_cntr_;
 };