fix record event for operator type in new dygraph (#44582)

* fix new dygraph record event for op

* update unit test

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -1386,7 +1386,7 @@ static std::string GenerateGradNodeCreationContent(
       "%s"
       "  {\n"
       "    paddle::platform::RecordEvent node_creation_record_event(\"%s\", "
-      "paddle::platform::TracerEventType::Operator, 1);\n"
+      "paddle::platform::TracerEventType::OperatorInner, 1);\n"
       "%s"
       "    if(require_any_grad) {\n"
       "      VLOG(6) << \" Construct Grad for %s \"; \n"

paddle/fluid/eager/auto_code_generator/final_state_generator/python_c_gen.py

@@ -111,7 +111,7 @@ PARSE_PYTHON_C_ARGS_TEMPLATE = \
 
 
 RECORD_EVENT_TEMPLATE = \
-"paddle::platform::RecordEvent {}(\"{} {}\", paddle::platform::TracerEventType::Operator, 1);"
+"paddle::platform::RecordEvent {}(\"{} {}\", paddle::platform::TracerEventType::UserDefined, 1);"
 
 
 RETURN_INPLACE_PYOBJECT_TEMPLATE = \

paddle/fluid/eager/backward.cc

@@ -851,7 +851,7 @@ void Backward(
     bool retain_graph) {
   VLOG(3) << "Run in Backward";
   paddle::platform::RecordEvent backward_record_event(
-      "backward", paddle::platform::TracerEventType::Operator, 1);
+      "backward", paddle::platform::TracerEventType::UserDefined, 1);
   RunBackward(tensors, grad_tensors, retain_graph);
   phi::autotune::AutoTuneStatus::Instance().Update();
 }

paddle/fluid/platform/profiler/chrometracing_logger.cc

@@ -603,7 +603,7 @@ void ChromeTracingLogger::StartLog() {
         std::string(
             R"JSON(
     {
-       "id": %d, "name": "%s", "totalGlobalMem": %u,
+       "id": %d, "name": "%s", "totalGlobalMem": %llu,
       "computeMajor": %d, "computeMinor": %d,
       "maxThreadsPerBlock": %d, "maxThreadsPerMultiprocessor": %d,
       "regsPerBlock": %d, "regsPerMultiprocessor": %d, "warpSize": %d,
@@ -633,7 +633,7 @@ void ChromeTracingLogger::StartLog() {
         std::string(
             R"JSON(
     {
-       "id": %d, "name": "%s", "totalGlobalMem": %u,
+       "id": %d, "name": "%s", "totalGlobalMem": %llu,
       "computeMajor": %d, "computeMinor": %d,
       "maxThreadsPerBlock": %d, "maxThreadsPerMultiprocessor": %d,
       "regsPerBlock": %d, "regsPerMultiprocessor": %d, "warpSize": %d,

python/paddle/fluid/tests/unittests/test_profiler_statistic.py

@@ -82,9 +82,9 @@ class TestProfilerStatistic(unittest.TestCase):
                                       profiler.TracerEventType.Forward, 50, 110,
                                       1000, 1001)
 
-        userdefined_node = HostPythonNode('Communication Time',
-                                          profiler.TracerEventType.UserDefined,
-                                          100, 110, 1000, 1001)
+        userdefined_node = HostPythonNode(
+            'Communication Time', profiler.TracerEventType.PythonUserDefined,
+            100, 110, 1000, 1001)
 
         communication_node = HostPythonNode(
             'Communication', profiler.TracerEventType.Communication, 105, 110,
@@ -209,7 +209,7 @@ class TestProfilerStatistic(unittest.TestCase):
                 0, profiler.TracerEventType.Memcpy), 60)
         self.assertEqual(
             time_range_summary.get_cpu_range_sum(
-                profiler.TracerEventType.UserDefined), 25)
+                profiler.TracerEventType.UserDefined), 15)
         self.assertEqual(
             time_range_summary.get_cpu_range_sum(
                 profiler.TracerEventType.Communication), 5)
@@ -277,9 +277,9 @@ class TestProfilerStatistic(unittest.TestCase):
                                       profiler.TracerEventType.Forward, 50, 110,
                                       1000, 1001)
 
-        userdefined_node = HostPythonNode('Communication Time',
-                                          profiler.TracerEventType.UserDefined,
-                                          100, 110, 1000, 1001)
+        userdefined_node = HostPythonNode(
+            'Communication Time', profiler.TracerEventType.PythonUserDefined,
+            100, 110, 1000, 1001)
         allreduce_launchkernel0 = HostPythonNode(
             'cudalaunchkernel', profiler.TracerEventType.CudaRuntime, 102, 104,
             1000, 1001)
@@ -451,7 +451,7 @@ class TestProfilerStatistic(unittest.TestCase):
                 0, profiler.TracerEventType.Memcpy), 60)
         self.assertEqual(
             time_range_summary.get_cpu_range_sum(
-                profiler.TracerEventType.UserDefined), 25)
+                profiler.TracerEventType.UserDefined), 15)
         self.assertEqual(
             time_range_summary.get_cpu_range_sum(
                 profiler.TracerEventType.Communication), 5)
@@ -518,9 +518,9 @@ class TestProfilerStatistic(unittest.TestCase):
         optimization_node = HostPythonNode(
             'Optimization', profiler.TracerEventType.Optimization, 220, 300,
             1000, 1001)
-        userdefined_node = HostPythonNode('Communication Time',
-                                          profiler.TracerEventType.UserDefined,
-                                          60, 70, 1000, 1001)
+        userdefined_node = HostPythonNode(
+            'Communication Time', profiler.TracerEventType.PythonUserDefined,
+            60, 70, 1000, 1001)
 
         conv2d_node = HostPythonNode('conv2d',
                                      profiler.TracerEventType.Operator, 25, 25,

python/paddle/profiler/profiler_statistic.py

@@ -514,7 +514,8 @@ class EventSummary:
                         or 'memset' in host_statistic_node.name.lower():
                         self.add_memory_manipulation_item(host_statistic_node)
                     else:
-                        self.add_userdefined_item(host_statistic_node)
+                        if host_statistic_node.type == TracerEventType.PythonUserDefined:
+                            self.add_userdefined_item(host_statistic_node)
             self.add_kernel_item(host_statistic_nodes[0])
 
         for threadid, root_statistic_node in node_statistic_trees.items():