Added more profile signposts to dygraph (#40201)

* Added more signposts to dygraph profiling

* Fixed minor issues

* Refactored signpost names

* Fixed typo

* Removed debug codes

* Fixed typo

* Adjusted signpost names

* Fixed issues from branch merge

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -1553,9 +1553,23 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
   core_ops_returns_info[op_type] = return_contents;
 
   // [Generation] ComputeRequireGrad -> GradNodeCreation
+
   if (!bwd_info.GenerateForwardOnly()) {
     std::string grad_node_creation_body_str =
         GenerateGradNodeCreationContent(fwd_info, bwd_info);
+
+    // Add event record
+    std::string event_name = op_type + " node_creation";
+    const char* NODE_CREATION_TEMPLATE =
+        "{\n"
+        "   paddle::platform::RecordEvent node_creation_record_event(\"%s\", "
+        "paddle::platform::TracerEventType::Operator, 1);\n"
+        "   %s\n"
+        "}";
+
+    grad_node_creation_body_str = paddle::string::Sprintf(
+        NODE_CREATION_TEMPLATE, event_name, grad_node_creation_body_str);
+
     generated_function_body += grad_node_creation_body_str;
     generated_function_body += "\n";
 
@@ -1614,10 +1628,20 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
     if ((*iter) == ',') dygraph_function_args_str.erase(iter);
   }
 
-  const char* FWD_FUNCTION_TEMPLATE = "%s %s(%s) {\n\n%s\n}\n\n";
+  const char* DYGRAPH_FUNCTION_EVENT_RECORD_FUNCTION_TEMPLATE =
+      "paddle::platform::RecordEvent dygraph_entrance_record_event(\"%s\", "
+      "paddle::platform::TracerEventType::Operator, 1);";
+  std::string event_name = op_type + " dygraph";
+  std::string fwd_record_event_str = paddle::string::Sprintf(
+      DYGRAPH_FUNCTION_EVENT_RECORD_FUNCTION_TEMPLATE, event_name);
+  const char* FWD_FUNCTION_TEMPLATE =
+      "%s %s(%s) {\n\n"
+      " %s\n"
+      " %s\n"
+      "}\n\n";
   std::string fwd_function_str = paddle::string::Sprintf(
       FWD_FUNCTION_TEMPLATE, function_proto_return_type_str, function_name,
-      dygraph_function_args_str, generated_function_body);
+      dygraph_function_args_str, fwd_record_event_str, generated_function_body);
 
   // [Generation] Generate forward functions header
   const char* FWD_HEADER_TEMPLATE = "%s %s(%s);\n";
@@ -2240,8 +2264,9 @@ static void GenerateForwardDygraphFile(const std::string& forward_cc_path,
       "\"paddle/fluid/eager/api/generated/fluid_generated/"
       "dygraph_forward_api.h\"\n"
       "#include "
-      "\"paddle/fluid/eager/api/generated/fluid_generated/nodes/nodes.h\"\n\n"
-      "#include \"paddle/fluid/eager/api/utils/global_utils.h\"\n";
+      "\"paddle/fluid/eager/api/generated/fluid_generated/nodes/nodes.h\"\n"
+      "#include \"paddle/fluid/eager/api/utils/global_utils.h\"\n"
+      "#include \"paddle/fluid/platform/profiler/event_tracing.h\"\n\n";
   std::string forward_cc_include_str =
       paddle::string::Sprintf(FORWARD_INCLUDE_TEMPLATE);
   std::ofstream forward_cc_stream(forward_cc_path, std::ios::out);

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

@@ -923,8 +923,20 @@ def GenerateForwardDefinition(fwd_api_name, bwd_api_name,
         backward_fwd_input_map, backward_grad_input_map,
         backward_grad_output_map, backward_attrs_list, optional_inputs)
 
+    node_event_name = fwd_api_name + " node_creation"
+    NODE_CREATION_TEMPLATE = """{{\n
+           paddle::platform::RecordEvent node_creation_record_event(\"{}\", paddle::platform::TracerEventType::Operator, 1);\n
+           {}\n
+        }}"""
+    node_creation_str = NODE_CREATION_TEMPLATE.format(node_event_name,
+                                                      node_creation_str)
+
+    dygraph_event_str = f"paddle::platform::RecordEvent dygraph_entrance_record_event(\"{fwd_api_name} dygraph\", paddle::platform::TracerEventType::Operator, 1);"
+
     FORWARD_FUNCTION_TEMPLATE = """
 {} {}({}) {{
+    {}
+
     // Forward API Call
     {}
     
@@ -938,7 +950,7 @@ def GenerateForwardDefinition(fwd_api_name, bwd_api_name,
     forward_function_name = GetForwardFunctionName(fwd_api_name)
     forward_function_str = FORWARD_FUNCTION_TEMPLATE.format(
         returns_type_str, forward_function_name, inputs_args_definition_str,
-        forward_call_str, node_creation_str, returns_str)
+        dygraph_event_str, forward_call_str, node_creation_str, returns_str)
     forward_function_declaration_str = f"{returns_type_str} {forward_function_name}({inputs_args_declaration_str});"
 
     return forward_function_str, forward_function_declaration_str
@@ -1065,6 +1077,8 @@ def GenerateForwardCCFile(filepath, forward_definition_str):
 
 #include "paddle/phi/api/include/sparse_api.h"
 #include "paddle/fluid/eager/api/utils/global_utils.h"
+#include "paddle/fluid/platform/profiler/event_tracing.h"
+
 """
 
     file_contents += GenerateCoreOpInfoDefinition()

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

@@ -94,9 +94,13 @@ def GeneratePythonCFunction(fwd_api_name, forward_inputs_position_map,
         dygraph_function_call_list[pos] = f"{name}"
     dygraph_function_call_str = ",".join(dygraph_function_call_list)
 
+    pythonc_event_str = f"paddle::platform::RecordEvent pythonc_record_event(\"{fwd_api_name} pybind_imperative_func\", paddle::platform::TracerEventType::Operator, 1);"
+
     PYTHON_C_FUNCTION_TEMPLATE = """
 static PyObject * eager_final_state_api_{}(PyObject *self, PyObject *args, PyObject *kwargs)
 {{
+  {}
+
   PyThreadState *tstate = nullptr;
   try
   {{
@@ -136,8 +140,8 @@ def GeneratePythonCFunction(fwd_api_name, forward_inputs_position_map,
         fwd_function_name = namespace_str + GetForwardFunctionName(fwd_api_name)
 
     python_c_function_str = PYTHON_C_FUNCTION_TEMPLATE.format(
-        fwd_api_name, fwd_api_name, get_eager_tensor_str, parse_attributes_str,
-        fwd_function_name, dygraph_function_call_str)
+        fwd_api_name, pythonc_event_str, fwd_api_name, get_eager_tensor_str,
+        parse_attributes_str, fwd_function_name, dygraph_function_call_str)
 
     python_c_function_reg_str = f"{{\"final_state_{fwd_api_name}\", (PyCFunction)(void(*)(void)) {namespace_str}eager_final_state_api_{fwd_api_name}, METH_VARARGS | METH_KEYWORDS, \"C++ interface function for {fwd_api_name} in dygraph.\"}}\n"
 
@@ -231,6 +235,7 @@ def GeneratePythonCWrappers(python_c_function_str, python_c_function_reg_str):
 #include  "paddle/fluid/pybind/op_function_common.h"
 #include  "paddle/fluid/eager/api/generated/eager_generated/forwards/dygraph_functions.h"
 #include  "paddle/fluid/pybind/exception.h"
+#include  "paddle/fluid/platform/profiler/event_tracing.h"
 #include  <Python.h>
 
 namespace paddle {{

paddle/fluid/eager/backward.cc

@@ -19,6 +19,8 @@
 #include "paddle/fluid/eager/grad_node_info.h"
 #include "paddle/fluid/eager/grad_tensor_holder.h"
 #include "paddle/fluid/eager/utils.h"
+#include "paddle/fluid/platform/profiler.h"
+#include "paddle/fluid/platform/profiler/event_tracing.h"
 
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/errors.h"
@@ -77,6 +79,9 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
 void RunBackward(const std::vector<paddle::experimental::Tensor>& tensors,
                  const std::vector<paddle::experimental::Tensor>& grad_tensors,
                  bool retain_graph) {
+  paddle::platform::RecordEvent backward_record_event(
+      "backward", paddle::platform::TracerEventType::Operator, 1);
+
   VLOG(6) << "Start Backward";
   // *Gradient Hook should happen at node-level
   // *Inplace version check should perform at node-level
@@ -160,6 +165,10 @@ void RunBackward(const std::vector<paddle::experimental::Tensor>& tensors,
   while (!queue.empty()) {
     GradNodeBase* node = queue.front();
 
+    paddle::platform::RecordEvent node_record_event(
+        std::string(typeid(*node).name()) + " grad_node",
+        paddle::platform::TracerEventType::Operator, 1);
+
     if (queue.size() > 1 && node_in_degree_map[node] != 0) {
       queue.pop();
       continue;

paddle/fluid/imperative/basic_engine.cc

@@ -389,6 +389,9 @@ static void PerformBackwardInplace(const std::string& op_type,
 }
 
 void BasicEngine::Execute() {
+  platform::RecordEvent backward_record_event(
+      "backward", platform::TracerEventType::Operator, 1);
+
   if (init_nodes_.empty()) {
     return;
   }
@@ -412,7 +415,7 @@ void BasicEngine::Execute() {
 
     for (auto& cur_op : *shared_cur_node) {
       platform::RecordEvent op_type_record_event(
-          cur_op.Type(), platform::TracerEventType::Operator, 1);
+          cur_op.Type() + " grad_node", platform::TracerEventType::Operator, 1);
 
       ++op_num;
 

paddle/fluid/imperative/tracer.cc

@@ -177,7 +177,7 @@ void Tracer::TraceOp(const std::string& type, const NameVarMap<VarType>& ins,
                      paddle::framework::AttributeMap* passed_default_attrs_,
                      bool use_default_attr_map) {
   platform::RecordEvent op_type_record_event(
-      type, platform::TracerEventType::Operator, 1);
+      type + " trace_op", platform::TracerEventType::Operator, 1);
   platform::ScopedFlushDenormal flush;
   VLOG(1) << "Trace Op: " << type;
   if (FLAGS_use_mkldnn) {
@@ -297,19 +297,24 @@ void Tracer::TraceOp(const std::string& type, const NameVarMap<VarType>& ins,
     program_desc_tracer_->InsertOp(type, new_ins, outs, attrs);
   }
 
-  if (ComputeRequiredGrad(new_ins, outs, trace_backward)) {
-    PADDLE_ENFORCE_EQ(
-        passed_default_attrs_, nullptr,
-        paddle::platform::errors::PermissionDenied(
-            "We expect passed_default_attrs_ is nullptr while "
-            "use_default_attr_map is true, however we got not null "
-            "passed_default_attrs_. Please check your usage of trace_op. "));
-    CreateGradOpNode(*op, new_ins, outs, attrs, default_attrs, place,
-                     inplace_map);
-  } else {
-    VLOG(3) << "No Grad to track for Op: " << type;
+  {
+    platform::RecordEvent node_creation_record_event(
+        type + " node_creation", platform::TracerEventType::Operator, 1);
+
+    if (ComputeRequiredGrad(new_ins, outs, trace_backward)) {
+      PADDLE_ENFORCE_EQ(
+          passed_default_attrs_, nullptr,
+          paddle::platform::errors::PermissionDenied(
+              "We expect passed_default_attrs_ is nullptr while "
+              "use_default_attr_map is true, however we got not null "
+              "passed_default_attrs_. Please check your usage of trace_op. "));
+      CreateGradOpNode(*op, new_ins, outs, attrs, default_attrs, place,
+                       inplace_map);
+    } else {
+      VLOG(3) << "No Grad to track for Op: " << type;
+    }
+    VLOG(6) << "Finish Trace Op: " << type;
   }
-  VLOG(6) << "Finish Trace Op: " << type;
 }
 
 template void Tracer::TraceOp<VarBase>(

python/paddle/utils/code_gen/api_base.py

@@ -696,6 +696,7 @@ def gen_dense_tensor_kernel_code(self, code_indent, inplace_flag=False):
             code_indent)
         outputs_args, kernel_output_names, output_create = self.gene_output(
             self.outputs['types'], 'SetKernelOutput', code_indent, inplace_flag)
+        api_func_name = self.get_api_func_name() + ('_' if inplace_flag else '')
         return f"""
 {code_indent}  auto kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
 {code_indent}      "{self.kernel['func'][0]}", {{kernel_backend, kernel_layout, kernel_data_type}});
@@ -709,7 +710,10 @@ def gen_dense_tensor_kernel_code(self, code_indent, inplace_flag=False):
 
 {code_indent}  using kernel_signature = {kernel_signature};
 {code_indent}  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-{code_indent}  (*kernel_fn)({kernel_args}, {outputs_args});
+{code_indent}  {{
+{code_indent}    paddle::platform::RecordEvent kernel_record_event(\"{api_func_name} compute\", paddle::platform::TracerEventType::Operator, 1);
+{code_indent}    (*kernel_fn)({kernel_args}, {outputs_args});
+{code_indent}  }}
 
 {code_indent}  return {self.gene_return_code()};"""
 
@@ -719,6 +723,7 @@ def gen_selected_rows_kernel_code(self, code_indent, inplace_flag=False):
         outputs_args, kernel_output_names, output_create = self.gene_output(
             self.outputs['types'], 'SetSelectedRowsKernelOutput', code_indent,
             inplace_flag)
+        api_func_name = self.get_api_func_name() + ('_' if inplace_flag else '')
         return f"""
 {code_indent}  auto kernel = phi::KernelFactory::Instance().SelectKernelOrThrowError(
 {code_indent}      "{self.kernel['func'][1]}", {{kernel_backend, kernel_layout, kernel_data_type}});
@@ -732,7 +737,10 @@ def gen_selected_rows_kernel_code(self, code_indent, inplace_flag=False):
 
 {code_indent}  using kernel_signature = {kernel_signature};
 {code_indent}  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
-{code_indent}  (*kernel_fn)({kernel_args}, {outputs_args});
+{code_indent}  {{
+{code_indent}    paddle::platform::RecordEvent kernel_record_event(\"{api_func_name} compute\", paddle::platform::TracerEventType::Operator, 1);
+{code_indent}    (*kernel_fn)({kernel_args}, {outputs_args});
+{code_indent}  }}
 
 {code_indent}  return {self.gene_return_code()};"""
 

python/paddle/utils/code_gen/api_gen.py

@@ -147,6 +147,9 @@ def source_include(header_file_path):
 #include "paddle/phi/infermeta/multiary.h"
 #include "paddle/phi/infermeta/nullary.h"
 #include "paddle/phi/infermeta/unary.h"
+#include "paddle/phi/kernels/declarations.h"
+
+#include "paddle/fluid/platform/profiler/event_tracing.h"
 """
 
 

python/paddle/utils/code_gen/backward_api_gen.py

@@ -154,6 +154,8 @@ def source_include(header_file_path):
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/api/include/api.h"
 #include "paddle/phi/infermeta/backward.h"
+
+#include "paddle/fluid/platform/profiler/event_tracing.h"
 """