Add gperf tools

CMakeLists.txt

@@ -81,6 +81,12 @@ option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 option(WITH_FAST_MATH   "Make use of fast math library, might affect the precision to some extent" ON)
 
+if (WITH_PROFILER)
+    find_package(Gperftools REQUIRED)
+    include_directories(${GPERFTOOLS_INCLUDE_DIR})
+    add_definitions(-DWITH_GPERFTOOLS)
+endif()
+
 # PY_VERSION
 if(NOT PY_VERSION)
   set(PY_VERSION 2.7)

cmake/generic.cmake

@@ -110,6 +110,14 @@ function(find_fluid_modules TARGET_NAME)
   endif()
 endfunction(find_fluid_modules)
 
+
+function(common_link TARGET_NAME)
+  if (WITH_PROFILER)
+    target_link_libraries(${TARGET_NAME} gperftools::profiler)
+  endif()
+endfunction()
+
+
 # find all third_party modules is used for paddle static library
 # for reduce the dependency when building the inference libs.
 set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY)
@@ -274,6 +282,7 @@ function(cc_library TARGET_NAME)
       endif()
       target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
       add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
+      common_link(${TARGET_NAME})
     endif()
 
     # cpplint code style
@@ -340,6 +349,7 @@ function(cc_binary TARGET_NAME)
   if(cc_binary_DEPS)
     target_link_libraries(${TARGET_NAME} ${cc_binary_DEPS})
     add_dependencies(${TARGET_NAME} ${cc_binary_DEPS})
+    common_link(${TARGET_NAME})
   endif()
 endfunction(cc_binary)
 
@@ -362,6 +372,7 @@ function(cc_test TARGET_NAME)
       target_link_libraries(${TARGET_NAME} ${win32_deps})
     endif(WIN32)
     add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    common_link(${TARGET_NAME})
     add_test(NAME ${TARGET_NAME}
              COMMAND ${TARGET_NAME} ${cc_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
@@ -420,6 +431,7 @@ function(nv_binary TARGET_NAME)
     if(nv_binary_DEPS)
       target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS})
       add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
+      common_link(${TARGET_NAME})
     endif()
   endif()
 endfunction(nv_binary)
@@ -433,6 +445,7 @@ function(nv_test TARGET_NAME)
     cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
     target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
     add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
+    common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})
     if (nv_test_SERIAL)
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
@@ -499,6 +512,7 @@ function(hip_binary TARGET_NAME)
     if(hip_binary_DEPS)
       target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS})
       add_dependencies(${TARGET_NAME} ${hip_binary_DEPS})
+      common_link(${TARGET_NAME})
     endif()
   endif()
 endfunction(hip_binary)
@@ -518,6 +532,7 @@ function(hip_test TARGET_NAME)
     set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
     target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
     add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
+    common_link(${TARGET_NAME})
     add_test(${TARGET_NAME} ${TARGET_NAME})
   endif()
 endfunction(hip_test)
@@ -560,6 +575,7 @@ function(go_library TARGET_NAME)
   endif()
   if(go_library_DEPS)
     add_dependencies(${TARGET_NAME} ${go_library_DEPS})
+    common_link(${TARGET_NAME})
   endif(go_library_DEPS)
 
   # The "source file" of the library is `${dummyfile}` which never

paddle/fluid/framework/parallel_executor.cc

@@ -30,13 +30,36 @@ limitations under the License. */
 #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
 #include "paddle/fluid/platform/profiler.h"
 
+#ifdef WITH_GPERFTOOLS
+#include "gperftools/profiler.h"
+#endif
+DEFINE_string(pe_profile_fname, "",
+              "Profiler filename for PE, which generated by gperftools."
+              "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable.");
+
 namespace paddle {
 namespace framework {
 
+static std::once_flag gProfileOnce;
+#ifdef WITH_GPERFTOOLS
+static bool gProfileStarted = false;
+#endif
 class ParallelExecutorPrivate {
  public:
   explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
-      : places_(places) {}
+      : places_(places) {
+    if (!FLAGS_pe_profile_fname.empty()) {
+      std::call_once(gProfileOnce, [] {
+#ifdef WITH_GPERFTOOLS
+        ProfilerStart(FLAGS_pe_profile_fname.c_str());
+        gProfileStarted = true;
+#else
+        LOG(WARNING) << "Paddle is not compiled with gperftools. "
+                        "FLAGS_pe_profile_fname will be ignored";
+#endif
+      });
+    }
+  }
 
   ~ParallelExecutorPrivate() {
     if (own_local_scope_) {
@@ -270,6 +293,12 @@ void ParallelExecutor::BCastParamsToDevices(
 
 void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
                            const std::string &fetched_var_name) {
+#ifdef WITH_GPERFTOOLS
+  if (gProfileStarted) {
+    ProfilerFlush();
+  }
+#endif
+
   platform::RecordBlock b(0);
 #ifdef PADDLE_WITH_CUDA
   if (!gcs_.empty()) {

python/paddle/fluid/__init__.py

@@ -125,7 +125,8 @@ def __bootstrap__():
         'use_ngraph', 'initial_cpu_memory_in_mb', 'init_allocated_mem',
         'free_idle_memory', 'paddle_num_threads', "dist_threadpool_size",
         'eager_delete_tensor_gb', 'allocator_strategy',
-        'reader_queue_speed_test_mode', 'print_sub_graph_dir'
+        'reader_queue_speed_test_mode', 'print_sub_graph_dir',
+        'pe_profile_fname'
     ]
     if 'Darwin' not in sysstr:
         read_env_flags.append('use_pinned_memory')