Update profiler (#39779)

* add align for WorkQueue

* add spinlock

* merge develop

* merge

* Add EventsWaiter

* Revert "Add EventsWaiter"

This reverts commit e206173aa9be7401b83a53581627bfaf557c8fb2.

* add log for Executor

* update the profiler
Co-authored-by: Nliutiexing <liutiexing@google.com>

paddle/fluid/platform/profiler/cupti_data_process.cc

@@ -14,6 +14,7 @@
 
 #include "paddle/fluid/platform/profiler/cupti_data_process.h"
 #include <cstdio>
+#include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/os_info.h"
 
 namespace paddle {
@@ -26,7 +27,7 @@ void AddKernelRecord(const CUpti_ActivityKernel4* kernel, uint64_t start_ns,
     return;
   }
   DeviceTraceEvent event;
-  event.name = kernel->name;
+  event.name = demangle(kernel->name);
   event.type = TracerEventType::Kernel;
   event.start_ns = kernel->start;
   event.end_ns = kernel->end;

paddle/fluid/platform/profiler/event_tracing.h

@@ -21,26 +21,55 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
+// Default tracing level.
+// It is Recommended to set the level explicitly.
 static constexpr uint32_t kDefaultTraceLevel = 4;
-// CPU event tracing. A trace marks something that happens but has no duration
+
+// Host event tracing. A trace marks something that happens but has no duration
 // associated with it. For example, thread starts working.
 // Chrome Trace Viewer Format: Instant Event
 struct RecordInstantEvent {
+  /**
+   * @param name: It is the caller's reponsibility to manage the underlying
+   * storage. RecordInstantEvent stores the pointer.
+   * @param type: Classification which is used to instruct the profiling
+   * data statistics.
+   * @param level: Used to filter events, works like glog VLOG(level).
+   * RecordEvent will works if HostTraceLevel >= level.
+   */
   explicit RecordInstantEvent(const char* name, TracerEventType type,
                               uint32_t level = kDefaultTraceLevel);
 };
 
-// CPU event tracing. A trace starts when an object of this clas is created and
+// Host event tracing. A trace starts when an object of this clas is created and
 // stops when the object is destroyed.
 // Chrome Trace Viewer Format: Duration Event/Complte Event
 class RecordEvent {
  public:
+  /**
+   * @param name: If your string argument has a longer lifetime (e.g.: string
+   * literal, static variables, etc) than the event, use 'const char* name'.
+   * Do your best to avoid using 'std::string' as the argument type. It will
+   * cause deep-copy to harm performance.
+   * @param type: Classification which is used to instruct the profiling
+   * data statistics.
+   * @param level: Used to filter events, works like glog VLOG(level).
+   * RecordEvent will works if HostTraceLevel >= level.
+   */
   explicit RecordEvent(
       const std::string& name,
       const TracerEventType type = TracerEventType::UserDefined,
       uint32_t level = kDefaultTraceLevel,
       const EventRole role = EventRole::kOrdinary);
 
+  /**
+   * @param name: It is the caller's reponsibility to manage the underlying
+   * storage. RecordEvent stores the pointer.
+   * @param type: Classification which is used to instruct the profiling
+   * data statistics.
+   * @param level: Used to filter events, works like glog VLOG(level).
+   * RecordEvent will works if HostTraceLevel >= level.
+   */
   explicit RecordEvent(const char* name, const TracerEventType type =
                                              TracerEventType::UserDefined,
                        uint32_t level = kDefaultTraceLevel,

paddle/fluid/platform/profiler/host_event_recorder.h

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