[Inference] [unittest] Inference unit tests rely on dynamic libraries (#24743)

cmake/generic.cmake

@@ -380,8 +380,7 @@ 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)

paddle/fluid/inference/CMakeLists.txt

@@ -63,10 +63,6 @@ if(WITH_TESTING AND WITH_INFERENCE_API_TEST)
     add_subdirectory(tests/api)
 endif()
 
-if(NOT ON_INFER)
-  return()
-endif()
-
 set(SHARED_INFERENCE_SRCS
     io.cc
     ${CMAKE_CURRENT_SOURCE_DIR}/../framework/data_feed.cc

paddle/fluid/inference/api/CMakeLists.txt

@@ -49,10 +49,21 @@ 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)
+  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 ${inference_deps}
+
+if (NOT APPLE AND NOT WIN32)
+  cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS paddle_fluid_shared
           ARGS --dirname=${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()

paddle/fluid/inference/tests/api/CMakeLists.txt

-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")

paddle/fluid/inference/tests/api/analyzer_detect_tester.cc

@@ -53,7 +53,7 @@ void SetConfig(AnalysisConfig *cfg) {
   cfg->DisableGpu();
   cfg->SwitchIrDebug();
   cfg->SwitchSpecifyInputNames(false);
-  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
+  cfg->SetCpuMathLibraryNumThreads(FLAGS_cpu_num_threads);
 }
 
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs,

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) {

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.

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();
 }
 

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();
 }
 

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);
   }

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();
 }
 

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);
   }

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) {

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 {
@@ -373,7 +373,7 @@ 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();
   }
 }

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 {
 

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 {
 

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();