diff --git a/cmake/generic.cmake b/cmake/generic.cmake
index 2c65554e48ba946faec84eb826fbf93b315f519f..fe834604bec8ef9d28f1778a3362a613c7cb2be0 100644
--- a/cmake/generic.cmake
+++ b/cmake/generic.cmake
@@ -382,8 +382,7 @@ function(cc_test_run TARGET_NAME)
     set(multiValueArgs COMMAND ARGS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
-	    COMMAND ${cc_test_COMMAND}
-	    ARGS ${cc_test_ARGS}
+	    COMMAND ${cc_test_COMMAND} ${cc_test_ARGS}
             WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index 88723e24184cce240d4bccd1b4c15a1ea8f14ed6..e0244a876f2cfc6b77db105d94d3262bed4f4014 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -63,10 +63,6 @@ if(WITH_TESTING)
   endif()
 endif()
 
-if(NOT ON_INFER)
-  return()
-endif()
-
 set(SHARED_INFERENCE_SRCS
     io.cc
     ${CMAKE_CURRENT_SOURCE_DIR}/../framework/data_feed.cc
diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt
index d375a4443e8c8b71830093b9ab1941e749c99d5b..3433e3d759aef9c1bae643697d727f74aa571f58 100755
--- a/paddle/fluid/inference/api/CMakeLists.txt
+++ b/paddle/fluid/inference/api/CMakeLists.txt
@@ -45,10 +45,21 @@ cc_library(analysis_predictor SRCS analysis_predictor.cc ${mkldnn_quantizer_src}
 cc_test(test_paddle_inference_api SRCS api_tester.cc DEPS paddle_inference_api)
 
 if(WITH_TESTING)
-  inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps}
-                      ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book)
+  if (NOT APPLE AND NOT WIN32)
+    inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS paddle_fluid_shared
+                        ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book)
+  else()
+    inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps}
+                        ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book)
+  endif()
   set_tests_properties(test_api_impl PROPERTIES DEPENDS test_image_classification)
   set_tests_properties(test_api_impl PROPERTIES LABELS "RUN_TYPE=DIST")
 endif()
-cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor benchmark ${inference_deps}
-        ARGS --dirname=${WORD2VEC_MODEL_DIR})
+
+if (NOT APPLE AND NOT WIN32)
+  cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS paddle_fluid_shared
+          ARGS --dirname=${WORD2VEC_MODEL_DIR})
+else()
+  cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor benchmark ${inference_deps}
+          ARGS --dirname=${WORD2VEC_MODEL_DIR})
+endif()
diff --git a/paddle/fluid/inference/tests/api/CMakeLists.txt b/paddle/fluid/inference/tests/api/CMakeLists.txt
index 8274ca86dc4225946848778cbac689674b789843..d23ef93b922a048722e802543cf2b24324acaaaf 100644
--- a/paddle/fluid/inference/tests/api/CMakeLists.txt
+++ b/paddle/fluid/inference/tests/api/CMakeLists.txt
@@ -1,7 +1,11 @@
-set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor benchmark)
+if (NOT APPLE AND NOT WIN32)
+    set(INFERENCE_EXTRA_DEPS paddle_fluid_shared)
+else()
+    set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor benchmark)
+endif()
 
 if(WITH_GPU AND TENSORRT_FOUND)
-    set(INFERENCE_EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps} ir_pass_manager analysis_predictor)
+    set(INFERENCE_EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps})
 endif()
 
 function(download_data install_dir data_file)
@@ -33,13 +37,13 @@ endfunction()
 
 function(inference_analysis_api_test target install_dir filename)
     inference_analysis_test(${target} SRCS ${filename}
-        EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} benchmark
+        EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
         ARGS --infer_model=${install_dir}/model --infer_data=${install_dir}/data.txt --refer_result=${install_dir}/result.txt)
 endfunction()
 
 function(inference_analysis_api_test_build TARGET_NAME filename)
 	inference_analysis_test_build(${TARGET_NAME} SRCS ${filename}
-        EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} benchmark)
+        EXTRA_DEPS ${INFERENCE_EXTRA_DEPS})
 endfunction()
 
 function(inference_analysis_api_int8_test_run TARGET_NAME test_binary model_dir data_path)
@@ -49,7 +53,7 @@ function(inference_analysis_api_int8_test_run TARGET_NAME test_binary model_dir
              --infer_data=${data_path}
              --warmup_batch_size=${WARMUP_BATCH_SIZE}
              --batch_size=50
-             --paddle_num_threads=${CPU_NUM_THREADS_ON_CI}
+             --cpu_num_threads=${CPU_NUM_THREADS_ON_CI}
 	     --iterations=2)
 endfunction()
 
@@ -65,7 +69,7 @@ function(inference_analysis_api_object_dection_int8_test_run TARGET_NAME test_bi
              --infer_data=${data_path}
              --warmup_batch_size=10
              --batch_size=300
-             --paddle_num_threads=${CPU_NUM_THREADS_ON_CI}
+             --cpu_num_threads=${CPU_NUM_THREADS_ON_CI}
 	     --iterations=1)
 endfunction()
 
@@ -88,7 +92,7 @@ function(inference_analysis_api_qat_test_run TARGET_NAME test_binary fp32_model_
              --int8_model=${int8_model_dir}
              --infer_data=${data_path}
              --batch_size=50
-             --paddle_num_threads=${CPU_NUM_THREADS_ON_CI}
+             --cpu_num_threads=${CPU_NUM_THREADS_ON_CI}
              --with_accuracy_layer=false
              --iterations=2)
 endfunction()
@@ -167,7 +171,7 @@ set(ERNIE_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/Ernie_Large")
 download_model_and_data(${ERNIE_INSTALL_DIR} "Ernie_large_model.tar.gz" "Ernie_large_data.txt.tar.gz" "Ernie_large_result.txt.tar.gz")
 download_result(${ERNIE_INSTALL_DIR} "Ernie_large_result.txt.tar.gz")
 inference_analysis_test(test_analyzer_ernie_large SRCS analyzer_ernie_tester.cc
-    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} benchmark
+    EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
     ARGS --infer_model=${ERNIE_INSTALL_DIR}/model --infer_data=${ERNIE_INSTALL_DIR}/data.txt --refer_result=${ERNIE_INSTALL_DIR}/result.txt --ernie_large=true)
 
 # text_classification
@@ -186,7 +190,7 @@ download_model_and_data(${TRANSFORMER_INSTALL_DIR} "temp%2Ftransformer_model.tar
 inference_analysis_test(test_analyzer_transformer SRCS analyzer_transformer_tester.cc 
   EXTRA_DEPS ${INFERENCE_EXTRA_DEPS}
   ARGS --infer_model=${TRANSFORMER_INSTALL_DIR}/model --infer_data=${TRANSFORMER_INSTALL_DIR}/data.txt --batch_size=8 
-       --paddle_num_threads=${CPU_NUM_THREADS_ON_CI})
+       --cpu_num_threads=${CPU_NUM_THREADS_ON_CI})
 
 # ocr
 set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr")
diff --git a/paddle/fluid/inference/tests/api/analyzer_detect_tester.cc b/paddle/fluid/inference/tests/api/analyzer_detect_tester.cc
new file mode 100644
index 0000000000000000000000000000000000000000..5333f0052d74250204daa09117602bf8bfe2aae1
--- /dev/null
+++ b/paddle/fluid/inference/tests/api/analyzer_detect_tester.cc
@@ -0,0 +1,129 @@
+/* Copyright (c) 2020 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.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <fstream>
+#include <iostream>
+#include "paddle/fluid/inference/tests/api/tester_helper.h"
+
+DEFINE_string(infer_shape, "", "data shape file");
+DEFINE_int32(sample, 20, "number of sample");
+
+namespace paddle {
+namespace inference {
+namespace analysis {
+
+struct Record {
+  std::vector<float> data;
+  std::vector<int32_t> shape;
+};
+
+Record ProcessALine(const std::string &line, const std::string &shape_line) {
+  VLOG(3) << "process a line";
+  std::vector<std::string> columns;
+
+  Record record;
+  std::vector<std::string> data_strs;
+  split(line, ' ', &data_strs);
+  for (auto &d : data_strs) {
+    record.data.push_back(std::stof(d));
+  }
+
+  std::vector<std::string> shape_strs;
+  split(shape_line, ' ', &shape_strs);
+  for (auto &s : shape_strs) {
+    record.shape.push_back(std::stoi(s));
+  }
+  return record;
+}
+
+void SetConfig(AnalysisConfig *cfg) {
+  cfg->SetModel(FLAGS_infer_model + "/model", FLAGS_infer_model + "/params");
+  cfg->DisableGpu();
+  cfg->SwitchIrDebug();
+  cfg->SwitchSpecifyInputNames(false);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
+}
+
+void SetInput(std::vector<std::vector<PaddleTensor>> *inputs,
+              const std::string &line, const std::string &shape_line) {
+  auto record = ProcessALine(line, shape_line);
+
+  PaddleTensor input;
+  input.shape = record.shape;
+  input.dtype = PaddleDType::FLOAT32;
+  size_t input_size = record.data.size() * sizeof(float);
+  input.data.Resize(input_size);
+  memcpy(input.data.data(), record.data.data(), input_size);
+  std::vector<PaddleTensor> input_slots;
+  input_slots.assign({input});
+  (*inputs).emplace_back(input_slots);
+}
+
+void profile(int cache_capacity = 1) {
+  AnalysisConfig cfg;
+  SetConfig(&cfg);
+  cfg.EnableMKLDNN();
+  cfg.SetMkldnnCacheCapacity(cache_capacity);
+
+  std::vector<std::vector<PaddleTensor>> outputs;
+  std::vector<std::vector<PaddleTensor>> input_slots_all;
+
+  Timer run_timer;
+  double elapsed_time = 0;
+
+  int num_times = FLAGS_repeat;
+  int sample = FLAGS_sample;
+  auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
+  outputs.resize(sample);
+
+  std::vector<std::thread> threads;
+
+  std::ifstream file(FLAGS_infer_data);
+  std::ifstream infer_file(FLAGS_infer_shape);
+  std::string line;
+  std::string shape_line;
+
+  for (int i = 0; i < sample; i++) {
+    threads.emplace_back([&, i]() {
+      std::getline(file, line);
+      std::getline(infer_file, shape_line);
+      SetInput(&input_slots_all, line, shape_line);
+
+      run_timer.tic();
+      predictor->Run(input_slots_all[0], &outputs[0], FLAGS_batch_size);
+      elapsed_time += run_timer.toc();
+    });
+    threads[0].join();
+    threads.clear();
+    std::vector<std::vector<PaddleTensor>>().swap(input_slots_all);
+  }
+  file.close();
+  infer_file.close();
+
+  auto batch_latency = elapsed_time / (sample * num_times);
+  PrintTime(FLAGS_batch_size, num_times, FLAGS_num_threads, 0, batch_latency,
+            sample, VarType::FP32);
+}
+
+#ifdef PADDLE_WITH_MKLDNN
+TEST(Analyzer_detect, profile_mkldnn) {
+  profile(5 /* cache_capacity */);
+  profile(10 /* cache_capacity */);
+}
+#endif
+
+}  // namespace analysis
+}  // namespace inference
+}  // namespace paddle
diff --git a/paddle/fluid/inference/tests/api/analyzer_ernie_tester.cc b/paddle/fluid/inference/tests/api/analyzer_ernie_tester.cc
index 199eee02d75c1ec0998fd2b657bd567d01d17435..87c8d783160218db2f791816ffbc5a161a66b6a1 100644
--- a/paddle/fluid/inference/tests/api/analyzer_ernie_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_ernie_tester.cc
@@ -143,7 +143,7 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false,
   }
   cfg->SwitchSpecifyInputNames();
   cfg->SwitchIrOptim();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
 }
 
 void profile(bool use_mkldnn = false, bool use_gpu = false) {
diff --git a/paddle/fluid/inference/tests/api/analyzer_image_classification_tester.cc b/paddle/fluid/inference/tests/api/analyzer_image_classification_tester.cc
index 07934f96dc55ef1e80b54472c74975ff62b6add3..1faffacebcfdb173b96815a6ad223f06ea69c07f 100644
--- a/paddle/fluid/inference/tests/api/analyzer_image_classification_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_image_classification_tester.cc
@@ -27,7 +27,7 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchIrOptim();
   cfg->SwitchSpecifyInputNames();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
 }
 
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
@@ -40,7 +40,7 @@ void SetOptimConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchIrOptim();
   cfg->SwitchSpecifyInputNames();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
 }
 
 // Easy for profiling independently.
diff --git a/paddle/fluid/inference/tests/api/analyzer_int8_image_classification_tester.cc b/paddle/fluid/inference/tests/api/analyzer_int8_image_classification_tester.cc
index eb2c7935026a2705adb5fd9e21d639536fdf0fa7..5f2c879fe0a0c755d192a6be34ac6a1173412b06 100644
--- a/paddle/fluid/inference/tests/api/analyzer_int8_image_classification_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_int8_image_classification_tester.cc
@@ -26,7 +26,7 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchIrOptim();
   cfg->SwitchSpecifyInputNames();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
   cfg->EnableMKLDNN();
 }
 
diff --git a/paddle/fluid/inference/tests/api/analyzer_int8_object_detection_tester.cc b/paddle/fluid/inference/tests/api/analyzer_int8_object_detection_tester.cc
index 8dcc611bb64557dd49fc3b7599aad5886ce4bfb8..7f06a3b9023ba3e907c9731d576f014a3e451113 100644
--- a/paddle/fluid/inference/tests/api/analyzer_int8_object_detection_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_int8_object_detection_tester.cc
@@ -27,7 +27,7 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchIrOptim(true);
   cfg->SwitchSpecifyInputNames(false);
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
   cfg->EnableMKLDNN();
 }
 
diff --git a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc
index 11a49ed2914ae22c2ddb4cfe384900adfce4f21d..06a8e01b10c6eb70fe2cbac19725d96281863c29 100644
--- a/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc
@@ -107,7 +107,7 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
   cfg->SwitchIrOptim();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
   if (FLAGS_zero_copy) {
     cfg->SwitchUseFeedFetchOps(false);
   }
diff --git a/paddle/fluid/inference/tests/api/analyzer_qat_image_classification_tester.cc b/paddle/fluid/inference/tests/api/analyzer_qat_image_classification_tester.cc
index fd3210c3384e3f3acbdcf22b6afdacfacc68c9ba..7b2b1c31cc5a7ee84fefc5abc37c342155151d94 100644
--- a/paddle/fluid/inference/tests/api/analyzer_qat_image_classification_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_qat_image_classification_tester.cc
@@ -26,7 +26,7 @@ void SetConfig(AnalysisConfig *cfg, std::string model_path) {
   cfg->DisableGpu();
   cfg->SwitchIrOptim(false);
   cfg->SwitchSpecifyInputNames();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
   cfg->EnableMKLDNN();
 }
 
diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc
index b88607717036515bbb6b836e89cbf8d7bac9bc51..56f706ae56bda8b06eba5dd9e080552aa9785c6e 100644
--- a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc
@@ -143,7 +143,7 @@ void SetConfig(AnalysisConfig *cfg, bool use_mkldnn = false) {
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
   cfg->SwitchIrDebug();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
   if (FLAGS_zero_copy) {
     cfg->SwitchUseFeedFetchOps(false);
   }
diff --git a/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc b/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc
index f2195966add8c4c159d26682c9578c95301a345f..9726109bf89ac0d5e1048f6cae0483248696f3e2 100644
--- a/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_transformer_tester.cc
@@ -165,7 +165,7 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchSpecifyInputNames();
   cfg->SwitchIrOptim();
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
 }
 
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h
index c8e5c826a1333f6443c8bbad2ae00d890b823c0a..1aaac6b09d820dece6118dde5655297a938287bc 100644
--- a/paddle/fluid/inference/tests/api/tester_helper.h
+++ b/paddle/fluid/inference/tests/api/tester_helper.h
@@ -66,8 +66,8 @@ DEFINE_bool(warmup, false,
             "Use warmup to calculate elapsed_time more accurately. "
             "To reduce CI time, it sets false in default.");
 
-DECLARE_bool(profile);
-DECLARE_int32(paddle_num_threads);
+DEFINE_bool(enable_profile, false, "Turn on profiler for fluid");
+DEFINE_int32(cpu_num_threads, 1, "Number of threads for each paddle instance.");
 
 namespace paddle {
 namespace inference {
@@ -355,7 +355,7 @@ void PredictionWarmUp(PaddlePredictor *predictor,
     predictor->ZeroCopyRun();
   }
   PrintTime(batch_size, 1, num_threads, tid, warmup_timer.toc(), 1, data_type);
-  if (FLAGS_profile) {
+  if (FLAGS_enable_profile) {
     paddle::platform::ResetProfiler();
   }
 }
diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc
index c6c84f8b9f0e56165ea07a75c4cf22c6dad3fcd4..3ce3f6d1e1eda58b78d58ece6f0cbd4cbd30fde9 100644
--- a/paddle/fluid/platform/init.cc
+++ b/paddle/fluid/platform/init.cc
@@ -38,6 +38,16 @@ DEFINE_int32(multiple_of_cupti_buffer_size, 1,
              "Multiple of the CUPTI device buffer size. If the timestamps have "
              "been dropped when you are profiling, try increasing this value.");
 
+namespace paddle {
+namespace platform {
+
+void ParseCommandLineFlags(int argc, char **argv, bool remove) {
+  google::ParseCommandLineFlags(&argc, &argv, remove);
+}
+
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
diff --git a/paddle/fluid/platform/init.h b/paddle/fluid/platform/init.h
index d189f0022bf6360bd5a55116c786e5375836e645..e087bb3f37c802ba654fd1e2706bf7258a8b4829 100644
--- a/paddle/fluid/platform/init.h
+++ b/paddle/fluid/platform/init.h
@@ -19,6 +19,14 @@ limitations under the License. */
 #include "gflags/gflags.h"
 #include "glog/logging.h"
 
+namespace paddle {
+namespace platform {
+
+void ParseCommandLineFlags(int argc, char **argv, bool remove);
+
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
diff --git a/paddle/testing/paddle_gtest_main.cc b/paddle/testing/paddle_gtest_main.cc
index d5acff56a9aa9136b84e216f6f8b0f28b528dbc5..c19bd56fbbf4de343abea198674eaf8b545321f5 100644
--- a/paddle/testing/paddle_gtest_main.cc
+++ b/paddle/testing/paddle_gtest_main.cc
@@ -23,10 +23,41 @@ limitations under the License. */
 int main(int argc, char** argv) {
   paddle::memory::allocation::UseAllocatorStrategyGFlag();
   testing::InitGoogleTest(&argc, argv);
-  std::vector<char*> new_argv;
-  std::string gflags_env;
+  // Because the dynamic library libpaddle_fluid.so clips the symbol table, the
+  // external program cannot recognize the flag inside the so, and the flag
+  // defined by the external program cannot be accessed inside the so.
+  // Therefore, the ParseCommandLine function needs to be called separately
+  // inside and outside.
+  std::vector<char*> external_argv;
+  std::vector<char*> internal_argv;
+
+  // ParseNewCommandLineFlags in gflags.cc starts processing
+  // commandline strings from idx 1.
+  // The reason is, it assumes that the first one (idx 0) is
+  // the filename of executable file.
+  external_argv.push_back(argv[0]);
+  internal_argv.push_back(argv[0]);
+
+  std::vector<google::CommandLineFlagInfo> all_flags;
+  std::vector<std::string> external_flags_name;
+  google::GetAllFlags(&all_flags);
+  for (size_t i = 0; i < all_flags.size(); ++i) {
+    external_flags_name.push_back(all_flags[i].name);
+  }
+
   for (int i = 0; i < argc; ++i) {
-    new_argv.push_back(argv[i]);
+    bool flag = true;
+    std::string tmp(argv[i]);
+    for (size_t j = 0; j < external_flags_name.size(); ++j) {
+      if (tmp.find(external_flags_name[j]) != std::string::npos) {
+        external_argv.push_back(argv[i]);
+        flag = false;
+        break;
+      }
+    }
+    if (flag) {
+      internal_argv.push_back(argv[i]);
+    }
   }
 
   std::vector<std::string> envs;
@@ -70,7 +101,7 @@ int main(int argc, char** argv) {
     }
     env_string = env_string.substr(0, env_string.length() - 1);
     env_str = strdup(env_string.c_str());
-    new_argv.push_back(env_str);
+    internal_argv.push_back(env_str);
     VLOG(1) << "gtest env_string:" << env_string;
   }
 
@@ -82,13 +113,17 @@ int main(int argc, char** argv) {
     }
     undefok_string = undefok_string.substr(0, undefok_string.length() - 1);
     undefok_str = strdup(undefok_string.c_str());
-    new_argv.push_back(undefok_str);
+    internal_argv.push_back(undefok_str);
     VLOG(1) << "gtest undefok_string:" << undefok_string;
   }
 
-  int new_argc = static_cast<int>(new_argv.size());
-  char** new_argv_address = new_argv.data();
-  google::ParseCommandLineFlags(&new_argc, &new_argv_address, false);
+  int new_argc = static_cast<int>(external_argv.size());
+  char** external_argv_address = external_argv.data();
+  google::ParseCommandLineFlags(&new_argc, &external_argv_address, false);
+
+  int internal_argc = internal_argv.size();
+  char** arr = internal_argv.data();
+  paddle::platform::ParseCommandLineFlags(internal_argc, arr, true);
   paddle::framework::InitDevices(true);
 
   int ret = RUN_ALL_TESTS();