Update host tracer (#39975)

* add align for WorkQueue

* add spinlock

* merge develop

* merge

* Add EventsWaiter

* Revert "Add EventsWaiter"

This reverts commit e206173aa9be7401b83a53581627bfaf557c8fb2.

* update HostTracer

* fix

* update

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

paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc

@@ -231,6 +231,8 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
     OpHandleBase *op,
     const std::shared_ptr<BlockingQueue<size_t>> &complete_q) {
   ++remaining_;
+  platform::RecordEvent("WorkQueue::AddTask",
+                        platform::TracerEventType::UserDefined, 10 /*level*/);
   this->pool_->enqueue([=] {
     std::deque<OpHandleBase *> op_queue;
     op_queue.push_front(op);

paddle/fluid/platform/os_info.cc

@@ -95,8 +95,6 @@ std::unordered_map<uint64_t, ThreadId> GetAllThreadIds() {
   return res;
 }
 
-static constexpr const char* kDefaultThreadName = "unset";
-
 std::string GetCurrentThreadName() {
   const auto& thread_name =
       internal::ThreadDataRegistry<std::string>::GetInstance()
@@ -112,7 +110,7 @@ std::unordered_map<uint64_t, std::string> GetAllThreadNames() {
 bool SetCurrentThreadName(const std::string& name) {
   auto& instance = internal::ThreadDataRegistry<std::string>::GetInstance();
   const auto& cur_name = instance.GetCurrentThreadData();
-  if (!cur_name.empty() || cur_name == kDefaultThreadName) {
+  if (!cur_name.empty() || name.empty() || name == kDefaultThreadName) {
     return false;
   }
   instance.SetCurrentThreadData(name);

paddle/fluid/platform/os_info.h

@@ -57,7 +57,8 @@ ThreadId GetCurrentThreadId();
 // create/destory when using it.
 std::unordered_map<uint64_t, ThreadId> GetAllThreadIds();
 
-// Returns 'unset' if SetCurrentThreadName is never called.
+static constexpr const char* kDefaultThreadName = "unset";
+// Returns kDefaultThreadName if SetCurrentThreadName is never called.
 std::string GetCurrentThreadName();
 
 // Return the map from StdTid to ThreadName

paddle/fluid/platform/profiler/host_tracer.cc

@@ -26,6 +26,9 @@ void ProcessHostEvents(const HostEventSection& host_events,
                        TraceEventCollector* collector) {
   for (const auto& thr_sec : host_events.thr_sections) {
     uint64_t tid = thr_sec.thread_id;
+    if (thr_sec.thread_name != kDefaultThreadName) {
+      collector->AddThreadName(tid, thr_sec.thread_name);
+    }
     for (const auto& evt : thr_sec.events) {
       HostTraceEvent event;
       event.name = evt.name;
@@ -41,12 +44,18 @@ void ProcessHostEvents(const HostEventSection& host_events,
 
 }  // namespace
 
+void HostTracer::PrepareTracing() {
+  // warm up
+  HostTraceLevel::GetInstance().SetLevel(options_.trace_level);
+  state_ = TracerState::READY;
+}
+
 void HostTracer::StartTracing() {
   PADDLE_ENFORCE_EQ(
       state_ == TracerState::READY || state_ == TracerState::STOPED, true,
       platform::errors::PreconditionNotMet("TracerState must be READY"));
   HostEventRecorder::GetInstance().GatherEvents();
-  HostTraceLevel::GetInstance().SetLevel(trace_level_);
+  HostTraceLevel::GetInstance().SetLevel(options_.trace_level);
   state_ = TracerState::STARTED;
 }
 

paddle/fluid/platform/profiler/host_tracer.h

@@ -45,9 +45,9 @@ struct HostTracerOptions {
 
 class HostTracer : public TracerBase {
  public:
-  explicit HostTracer(const HostTracerOptions& options) {
-    trace_level_ = options.trace_level;
-  }
+  explicit HostTracer(const HostTracerOptions& options) : options_(options) {}
+
+  void PrepareTracing() override;
 
   void StartTracing() override;
 
@@ -56,7 +56,7 @@ class HostTracer : public TracerBase {
   void CollectTraceData(TraceEventCollector* collector) override;
 
  private:
-  uint32_t trace_level_;
+  HostTracerOptions options_;
 };
 
 }  // namespace platform

paddle/fluid/platform/profiler/trace_event_collector.h

@@ -15,6 +15,8 @@ limitations under the License. */
 #pragma once
 
 #include <list>
+#include <string>
+#include <unordered_map>
 #include "paddle/fluid/platform/profiler/trace_event.h"
 
 namespace paddle {
@@ -32,6 +34,10 @@ class TraceEventCollector {
     device_events_.push_back(event);
   }
 
+  void AddThreadName(uint64_t tid, const std::string& name) {
+    thread_names_[tid] = name;
+  }
+
   const std::list<HostTraceEvent>& HostEvents() const { return host_events_; }
 
   const std::list<RuntimeTraceEvent>& RuntimeEvents() const {
@@ -42,7 +48,12 @@ class TraceEventCollector {
     return device_events_;
   }
 
+  const std::unordered_map<uint64_t, std::string>& ThreadNames() const {
+    return thread_names_;
+  }
+
  private:
+  std::unordered_map<uint64_t, std::string> thread_names_;
   std::list<HostTraceEvent> host_events_;
   std::list<RuntimeTraceEvent> runtime_events_;
   std::list<DeviceTraceEvent> device_events_;