Code cleanup in the profiler code. (#9782)

paddle/fluid/platform/profiler.h

@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 #include <forward_list>
 #include <list>
-#include <mutex>
+#include <string>
 #include <vector>
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/profiler.pb.h"
@@ -23,16 +23,16 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
-enum EventKind { kMark, kPushRange, kPopRange };
+enum EventType { kMark, kPushRange, kPopRange };
 
 class Event {
  public:
   // The DeviceContext is used to get the cuda stream.
   // If CPU profiling mode, can pass nullptr.
-  Event(EventKind kind, std::string name, uint32_t thread_id,
+  Event(EventType type, std::string name, uint32_t thread_id,
         const DeviceContext* dev_ctx);
 
-  std::string kind() const;
+  const EventType& type() const;
   std::string name() const { return name_; }
   uint32_t thread_id() const { return thread_id_; }
   bool has_cuda() const { return has_cuda_; }
@@ -46,7 +46,7 @@ class Event {
   double CudaElapsedMs(const Event& e) const;
 
  private:
-  EventKind kind_;
+  EventType type_;
   std::string name_;
   uint32_t thread_id_;
   int64_t cpu_ns_;
@@ -57,39 +57,6 @@ class Event {
 #endif
 };
 
-struct EventList {
-  constexpr static size_t kMB = 1024 * 1024;
-  constexpr static size_t kEventBlockSize = 16 * kMB;
-  constexpr static size_t kEventSize = sizeof(Event);
-  constexpr static size_t kEventAlign = alignof(Event);
-  constexpr static size_t kNumBlock =
-      kEventBlockSize /
-      ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign);
-
-  template <typename... Args>
-  void Record(Args&&... args) {
-    if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) {
-      event_blocks.emplace_front();
-      event_blocks.front().reserve(kNumBlock);
-    }
-    event_blocks.front().emplace_back(std::forward<Args>(args)...);
-  }
-
-  std::vector<Event> Reduce() {
-    std::vector<Event> result;
-    for (auto& block : event_blocks) {
-      result.insert(result.begin(), std::make_move_iterator(block.begin()),
-                    std::make_move_iterator(block.end()));
-    }
-    event_blocks.clear();
-    return result;
-  }
-
-  void Clear() { event_blocks.clear(); }
-
-  std::forward_list<std::vector<Event>> event_blocks;
-};
-
 enum ProfilerState {
   kDisabled,  // disabled state
   kCPU,       // CPU profiling state
@@ -136,16 +103,6 @@ struct RecordThread {
 // event_lists, event_lists[i][j] represents the j-th Event of i-th thread.
 std::vector<std::vector<Event>> GetAllEvents();
 
-// The information of each event given in the profiling report
-struct EventItem {
-  std::string name;
-  int calls;
-  double total_time;
-  double min_time;
-  double max_time;
-  double ave_time;
-};
-
 // Candidate keys to sort the profiling report
 enum EventSortingKey { kDefault, kCalls, kTotal, kMin, kMax, kAve };
 
@@ -158,14 +115,5 @@ void ResetProfiler();
 void DisableProfiler(EventSortingKey sorted_key,
                      const std::string& profile_path);
 
-// Parse the event list and output the profiling report
-void ParseEvents(std::vector<std::vector<Event>>&,
-                 EventSortingKey sorted_by = EventSortingKey::kDefault);
-
-// Print results
-void PrintProfiler(std::vector<std::vector<EventItem>>& events_table,
-                   std::string& sorted_domain, const size_t name_width,
-                   const size_t data_width);
-
 }  // namespace platform
 }  // namespace paddle

paddle/fluid/platform/profiler_test.cc

@@ -13,22 +13,23 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/platform/profiler.h"
+#include <string>
 #ifdef PADDLE_WITH_CUDA
-#include "cuda_runtime.h"
+#include <cuda_runtime.h>
 #endif
 #include "gtest/gtest.h"
 
 TEST(Event, CpuElapsedTime) {
   using paddle::platform::Event;
-  using paddle::platform::EventKind;
+  using paddle::platform::EventType;
 
-  Event start_event(EventKind::kPushRange, "test", 0, nullptr);
+  Event start_event(EventType::kPushRange, "test", 0, nullptr);
   EXPECT_TRUE(start_event.has_cuda() == false);
   int counter = 0;
   while (counter != 1000) {
     counter++;
   }
-  Event stop_event(EventKind::kPopRange, "test", 0, nullptr);
+  Event stop_event(EventType::kPopRange, "test", 0, nullptr);
   EXPECT_GT(start_event.CpuElapsedMs(stop_event), 0);
 }
 
@@ -38,16 +39,16 @@ TEST(Event, CudaElapsedTime) {
   using paddle::platform::CUDADeviceContext;
   using paddle::platform::CUDAPlace;
   using paddle::platform::Event;
-  using paddle::platform::EventKind;
+  using paddle::platform::EventType;
 
   DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(0));
-  Event start_event(EventKind::kPushRange, "test", 0, dev_ctx);
+  Event start_event(EventType::kPushRange, "test", 0, dev_ctx);
   EXPECT_TRUE(start_event.has_cuda() == true);
   int counter = 0;
   while (counter != 1000) {
     counter++;
   }
-  Event stop_event(EventKind::kPopRange, "test", 0, dev_ctx);
+  Event stop_event(EventType::kPopRange, "test", 0, dev_ctx);
   EXPECT_GT(start_event.CudaElapsedMs(stop_event), 0);
 }
 #endif
@@ -55,7 +56,7 @@ TEST(Event, CudaElapsedTime) {
 TEST(RecordEvent, RecordEvent) {
   using paddle::platform::DeviceContext;
   using paddle::platform::Event;
-  using paddle::platform::EventKind;
+  using paddle::platform::EventType;
   using paddle::platform::RecordEvent;
   using paddle::platform::ProfilerState;
   using paddle::platform::EventSortingKey;