Fix operator type record in profiler [cherry-pick PR44582] (#44654)

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

* fix new dygraph record event for op

* update unit test

* fix file mode

Author: rainyfly

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -1,11 +1,11 @@
 # Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
-# 
+#
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
-# 
+#
 #     http://www.apache.org/licenses/LICENSE-2.0
-# 
+#
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
@@ -71,7 +71,7 @@ def FindParsingFunctionFromAttributeType(atype):
 
 
 RECORD_EVENT_TEMPLATE = \
-"    paddle::platform::RecordEvent {}(\"{} {}\", paddle::platform::TracerEventType::Operator, 1);"
+"paddle::platform::RecordEvent {}(\"{} {}\", paddle::platform::TracerEventType::UserDefined, 1);"
 
 
 RETURN_INPLACE_PYOBJECT_TEMPLATE = \
@@ -253,6 +253,7 @@ def FindParsingFunctionFromAttributeType(atype):
 ## Generator Classes ##
 #######################
 class PythonCSingleFunctionGenerator(FunctionGeneratorBase):
+
     def __init__(self, forward_api_contents, namespace):
         # Members from Parent:
         #self.namespace
@@ -265,7 +266,7 @@ def __init__(self, forward_api_contents, namespace):
         #self.forward_outputs_position_map
         #self.optional_inputs
         #self.no_need_buffers
-        #self.intermediate_outputs   
+        #self.intermediate_outputs
         #self.inplace_map
         FunctionGeneratorBase.__init__(self, forward_api_contents, namespace)
 
@@ -327,16 +328,16 @@ def GeneratePythonCFunction(self):
         set_device_str = FUNCTION_SET_DEVICE_TEMPLATE.format(expected_place_str)
 
         # Generate Dygraph Function Call Logic
-        num_args = len(forward_inputs_position_map.keys()) + len(
-            orig_forward_attrs_list)
+        num_args = len(
+            forward_inputs_position_map.keys()) + len(orig_forward_attrs_list)
         dygraph_function_call_list = ["" for i in range(num_args)]
         for name, (_, pos) in forward_inputs_position_map.items():
             dygraph_function_call_list[pos] = f"{name}"
         for name, _, _, pos in orig_forward_attrs_list:
             dygraph_function_call_list[pos] = f"{name}"
         dygraph_function_call_str = ",".join(dygraph_function_call_list)
 
-        # Generate Python-C Function Definitions 
+        # Generate Python-C Function Definitions
         if is_forward_only:
             fwd_function_name = FUNCTION_NAME_TEMPLATE.format(
                 "paddle::experimental::", namespace, forward_api_name)
@@ -441,8 +442,9 @@ def run(self):
 
 
 class PythonCYamlGenerator(YamlGeneratorBase):
+
     def __init__(self, path):
-        # Parent members: 
+        # Parent members:
         # self.namespace
         # self.api_yaml_path
         # self.forward_api_list
@@ -457,8 +459,8 @@ def GeneratePythonCFunctions(self):
         forward_api_list = self.forward_api_list
 
         for forward_api_content in forward_api_list:
-            f_generator = PythonCSingleFunctionGenerator(forward_api_content,
-                                                         namespace)
+            f_generator = PythonCSingleFunctionGenerator(
+                forward_api_content, namespace)
             status = f_generator.run()
 
             if status == True:

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

@@ -30,10 +30,10 @@
 namespace egr {
 
 /*
-* GeneralGrad is Helpper class to implement custom grad operation between
-* outputs and inputs.
-*
-* **/
+ * GeneralGrad is Helpper class to implement custom grad operation between
+ * outputs and inputs.
+ *
+ * **/
 class GeneralGrad {
  public:
   static GeneralGrad& Instance() { return *general_grad_; }
@@ -64,7 +64,8 @@ class GeneralGrad {
                                 paddle::platform::errors::Fatal(
                                     "There is no grad op for %s:[%d] or it's"
                                     "stop_gradient=True.",
-                                    msg, i));
+                                    msg,
+                                    i));
         if (is_no_grad_vars) {
           (no_grad_var_nodes_inputmeta_map)[target_node] = auto_grad_meta;
         } else {  // normal input
@@ -248,7 +249,8 @@ class GeneralGrad {
 
   std::vector<paddle::experimental::Tensor> GetResults(
       const std::vector<paddle::experimental::Tensor>& inputs,
-      bool allow_unused, bool create_graph) {
+      bool allow_unused,
+      bool create_graph) {
     VLOG(6) << "Running in GetResults";
     if (inputs.empty()) return {};
 
@@ -276,7 +278,8 @@ class GeneralGrad {
         tensor_auto_grad_meta->SetStopGradient(!create_graph);
         results.emplace_back(iter->second);
       } else {
-        PADDLE_ENFORCE_EQ(allow_unused, true,
+        PADDLE_ENFORCE_EQ(allow_unused,
+                          true,
                           paddle::platform::errors::InvalidArgument(
                               "The %d-th input does not appear in the backward "
                               "graph. Please check the input tensor or set "
@@ -493,7 +496,8 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
 void EnforceGradNodeHasInput(GradNodeBase* node) {
   VLOG(6) << "Running in EnforceGradNodeHasInput";
   PADDLE_ENFORCE_NE(
-      node->IsTensorWrappersCleared(), true,
+      node->IsTensorWrappersCleared(),
+      true,
       paddle::platform::errors::Fatal(
           "The TensorWrappers of %s do not exist. This may be because:\n"
           "You calculate backward twice for the same subgraph without "
@@ -509,10 +513,13 @@ void DuplicateCheck(const std::vector<paddle::experimental::Tensor>& inputs,
   for (auto in : inputs) {
     AutogradMeta* auto_grad_meta = EagerUtils::unsafe_autograd_meta(in);
     PADDLE_ENFORCE_EQ(
-        visisted_ins.count(auto_grad_meta), 0,
+        visisted_ins.count(auto_grad_meta),
+        0,
         paddle::platform::errors::AlreadyExists(
-            "%s contain duplicate tensor %s, please check %s carefully.", msg,
-            in.name(), msg));
+            "%s contain duplicate tensor %s, please check %s carefully.",
+            msg,
+            in.name(),
+            msg));
     visisted_ins.insert(auto_grad_meta);
   }
 }
@@ -522,7 +529,8 @@ GeneralGrad* GeneralGrad::general_grad_ = new GeneralGrad();
 std::vector<paddle::experimental::Tensor> RunBackward(
     const std::vector<paddle::experimental::Tensor>& tensors,  // output
     const std::vector<paddle::experimental::Tensor>& grad_tensors,
-    bool retain_graph, bool create_graph = false,
+    bool retain_graph,
+    bool create_graph = false,
     const std::vector<paddle::experimental::Tensor>& inputs = {},
     bool allow_unused = false,
     const std::vector<paddle::experimental::Tensor>& no_grad_vars = {}) {
@@ -631,8 +639,8 @@ std::vector<paddle::experimental::Tensor> RunBackward(
 
   if (is_general_grad) {
     // Prepare several vital preprocess for GeneralGrad
-    GeneralGrad::Instance().PreparedForGeneralGrad(inputs, no_grad_vars, &queue,
-                                                   node_input_buffers_dict);
+    GeneralGrad::Instance().PreparedForGeneralGrad(
+        inputs, no_grad_vars, &queue, node_input_buffers_dict);
   }
 
   VLOG(6) << " startup_ops' size is :" << queue.size();
@@ -651,7 +659,8 @@ std::vector<paddle::experimental::Tensor> RunBackward(
 
     paddle::platform::RecordEvent node_record_event(
         std::string((*node).name()) + " grad_node",
-        paddle::platform::TracerEventType::Operator, 1);
+        paddle::platform::TracerEventType::Operator,
+        1);
 
     if (queue.size() > 1 && node_in_degree_map[node] != 0) {
       queue.pop();
@@ -716,7 +725,8 @@ std::vector<paddle::experimental::Tensor> RunBackward(
                        "Number of edges should be either empty ( for leaf node "
                        ") or the same as number of output grad tensors, but we "
                        "got edges size is: %d, grad_output size is: %d",
-                       edges.size(), grad_output_tensors.size()));
+                       edges.size(),
+                       grad_output_tensors.size()));
 
     for (size_t i = 0; i < edges.size(); i++) {
       for (size_t j = 0; j < edges[i].size(); j++) {
@@ -739,7 +749,8 @@ std::vector<paddle::experimental::Tensor> RunBackward(
         }
 
         PADDLE_ENFORCE_LT(
-            j, grad_output_tensors[i].size(),
+            j,
+            grad_output_tensors[i].size(),
             paddle::platform::errors::Fatal(
                 "Rank of grad_output_tensors should be less than "
                 "grad_output_tensors[i].size(), which is: %d. This error may "
@@ -771,9 +782,10 @@ std::vector<paddle::experimental::Tensor> RunBackward(
         VLOG(6) << "Sum grad inputs for edge slot: " << edge_rank.first
                 << ", rank: " << edge_rank.second;
 
-        node_input_buffers_dict[next_node]->add(
-            edge_rank.first, edge_rank.second, grad_output_tensor,
-            create_graph);
+        node_input_buffers_dict[next_node]->add(edge_rank.first,
+                                                edge_rank.second,
+                                                grad_output_tensor,
+                                                create_graph);
 
         // Update queue
         node_in_degree_map[next_node]--;
@@ -810,7 +822,7 @@ void Backward(
     bool retain_graph) {
   VLOG(6) << "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();
 }
@@ -819,14 +831,22 @@ std::vector<paddle::experimental::Tensor> Grad(
     const std::vector<paddle::experimental::Tensor>& tensors,  // outputs
     const std::vector<paddle::experimental::Tensor>& inputs,
     const std::vector<paddle::experimental::Tensor>& grad_tensors,
-    bool retain_graph, bool create_graph, bool only_inputs, bool allow_unused,
+    bool retain_graph,
+    bool create_graph,
+    bool only_inputs,
+    bool allow_unused,
     const std::vector<paddle::experimental::Tensor>& no_grad_vars) {
   VLOG(6) << "Run in Grad";
 
   DuplicateCheck(inputs, true /* is_input */);
   DuplicateCheck(tensors, false /* is_input */);
 
-  return RunBackward(tensors, grad_tensors, retain_graph, create_graph, inputs,
-                     allow_unused, no_grad_vars);
+  return RunBackward(tensors,
+                     grad_tensors,
+                     retain_graph,
+                     create_graph,
+                     inputs,
+                     allow_unused,
+                     no_grad_vars);
 }
 }  // namespace egr

paddle/fluid/eager/backward.cc

@@ -588,7 +588,7 @@ 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,
@@ -618,7 +618,7 @@ 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,