code optimization

PaddlePaddle · Mar 3, 2022 · 25e8d04 · 25e8d04
1 parent 4e9a2ee
commit 25e8d04
Show file tree

Hide file tree

Showing 9 changed files with 29 additions and 35 deletions.
diff --git a/cmake/external/onnxruntime.cmake b/cmake/external/onnxruntime.cmake
@@ -31,6 +31,8 @@ SET(ONNXRUNTIME_SOURCE_DIR     ${THIRD_PARTY_PATH}/onnxruntime/src/${ONNXRUNTIME
 SET(ONNXRUNTIME_INSTALL_DIR    ${THIRD_PARTY_PATH}/install/onnxruntime)
 SET(ONNXRUNTIME_INC_DIR        "${ONNXRUNTIME_INSTALL_DIR}/include" CACHE PATH "onnxruntime include directory." FORCE)
 SET(ONNXRUNTIME_LIB_DIR        "${ONNXRUNTIME_INSTALL_DIR}/lib" CACHE PATH "onnxruntime lib directory." FORCE)
+SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${ONNXRUNTIME_LIB_DIR}")
+
 
 if (WIN32)
   SET(ONNXRUNTIME_URL             "https://github.com/microsoft/onnxruntime/releases/download/v1.10.0/onnxruntime-win-x64-1.10.0.zip")
@@ -44,16 +46,16 @@ endif()
 INCLUDE_DIRECTORIES(${ONNXRUNTIME_INC_DIR}) # For ONNXRUNTIME code to include internal headers.
 if (WIN32)
   SET(ONNXRUNTIME_SOURCE_LIB "${ONNXRUNTIME_SOURCE_DIR}/lib/onnxruntime.dll" CACHE FILEPATH "ONNXRUNTIME source library." FORCE)
-  SET(ONNXRUNTIME_SHARE_LIB "${ONNXRUNTIME_INSTALL_DIR}/lib/onnxruntime.dll" CACHE FILEPATH "ONNXRUNTIME share library." FORCE)
+  SET(ONNXRUNTIME_SHARED_LIB "${ONNXRUNTIME_INSTALL_DIR}/lib/onnxruntime.dll" CACHE FILEPATH "ONNXRUNTIME shared library." FORCE)
   SET(ONNXRUNTIME_LIB "${ONNXRUNTIME_INSTALL_DIR}/lib/onnxruntime.lib" CACHE FILEPATH "ONNXRUNTIME static library." FORCE)
 elseif (APPLE)
   SET(ONNXRUNTIME_SOURCE_LIB "${ONNXRUNTIME_SOURCE_DIR}/lib/libonnxruntime.1.10.0.dylib" CACHE FILEPATH "ONNXRUNTIME source library." FORCE)
   SET(ONNXRUNTIME_LIB "${ONNXRUNTIME_INSTALL_DIR}/lib/libonnxruntime.1.10.0.dylib" CACHE FILEPATH "ONNXRUNTIME static library." FORCE)
-  SET(ONNXRUNTIME_SHARE_LIB ${ONNXRUNTIME_LIB} CACHE FILEPATH "ONNXRUNTIME share library." FORCE)
+  SET(ONNXRUNTIME_SHARED_LIB ${ONNXRUNTIME_LIB} CACHE FILEPATH "ONNXRUNTIME shared library." FORCE)
 else ()
   SET(ONNXRUNTIME_SOURCE_LIB "${ONNXRUNTIME_SOURCE_DIR}/lib/libonnxruntime.so.1.10.0" CACHE FILEPATH "ONNXRUNTIME source library." FORCE)
   SET(ONNXRUNTIME_LIB "${ONNXRUNTIME_INSTALL_DIR}/lib/libonnxruntime.so.1.10.0" CACHE FILEPATH "ONNXRUNTIME static library." FORCE)
-  SET(ONNXRUNTIME_SHARE_LIB ${ONNXRUNTIME_LIB} CACHE FILEPATH "ONNXRUNTIME share library." FORCE)
+  SET(ONNXRUNTIME_SHARED_LIB ${ONNXRUNTIME_LIB} CACHE FILEPATH "ONNXRUNTIME shared library." FORCE)
 endif ()
 
 
@@ -67,7 +69,7 @@ if (WIN32)
       CONFIGURE_COMMAND     ""
       BUILD_COMMAND         ""
       UPDATE_COMMAND        ""
-      INSTALL_COMMAND       ${CMAKE_COMMAND} -E copy ${ONNXRUNTIME_SOURCE_LIB} ${ONNXRUNTIME_SHARE_LIB} &&
+      INSTALL_COMMAND       ${CMAKE_COMMAND} -E copy ${ONNXRUNTIME_SOURCE_LIB} ${ONNXRUNTIME_SHARED_LIB} &&
                             ${CMAKE_COMMAND} -E copy ${ONNXRUNTIME_SOURCE_DIR}/lib/onnxruntime.lib ${ONNXRUNTIME_LIB} &&
                             ${CMAKE_COMMAND} -E copy_directory ${ONNXRUNTIME_SOURCE_DIR}/include ${ONNXRUNTIME_INC_DIR}
       BUILD_BYPRODUCTS      ${ONNXRUNTIME_LIB}

diff --git a/cmake/external/paddle2onnx.cmake b/cmake/external/paddle2onnx.cmake
@@ -29,19 +29,20 @@ SET(PADDLE2ONNX_INSTALL_DIR    ${THIRD_PARTY_PATH}/install/paddle2onnx)
 SET(PADDLE2ONNX_INC_DIR        "${PADDLE2ONNX_INSTALL_DIR}/include" CACHE PATH "paddle2onnx include directory." FORCE)
 SET(PADDLE2ONNX_REPOSITORY     ${GIT_URL}/PaddlePaddle/Paddle2ONNX.git)
 SET(PADDLE2ONNX_TAG            cpp)
-
 SET(LIBDIR "lib")
+SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${PADDLE2ONNX_INSTALL_DIR}/${LIBDIR}")
 
 INCLUDE_DIRECTORIES(${PADDLE2ONNX_INC_DIR}) # For PADDLE2ONNX code to include internal headers.
 if(WIN32)
     SET(PADDLE2ONNX_LIB "${PADDLE2ONNX_INSTALL_DIR}/${LIBDIR}/paddle2onnx.lib" CACHE FILEPATH "paddle2onnx static library." FORCE)
-    SET(PADDLE2ONNX_SHARE_LIB "${PADDLE2ONNX_INSTALL_DIR}/${LIBDIR}/paddle2onnx.dll" CACHE FILEPATH "paddle2onnx share library." FORCE)
+    SET(PADDLE2ONNX_SHARED_LIB "${PADDLE2ONNX_INSTALL_DIR}/${LIBDIR}/paddle2onnx.dll" CACHE FILEPATH "paddle2onnx shared library." FORCE)
 elseif(APPLE)
     SET(PADDLE2ONNX_LIB "${PADDLE2ONNX_INSTALL_DIR}/${LIBDIR}/libpaddle2onnx.dylib" CACHE FILEPATH "PADDLE2ONNX library." FORCE)
 else()
     SET(PADDLE2ONNX_LIB "${PADDLE2ONNX_INSTALL_DIR}/${LIBDIR}/libpaddle2onnx.so" CACHE FILEPATH "PADDLE2ONNX library." FORCE)
 endif(WIN32)
 
+
 # The protoc path is required to compile onnx.
 string(REPLACE "/" ";" PROTOC_BIN_PATH ${PROTOBUF_PROTOC_EXECUTABLE})
 list(POP_BACK PROTOC_BIN_PATH)

diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake
@@ -123,7 +123,7 @@ function(copy_part_of_thrid_party TARGET DST)
         set(dst_dir "${DST}/third_party/install/paddle2onnx")
         if(WIN32)
             copy(${TARGET}
-                SRCS ${PADDLE2ONNX_INC_DIR}/paddle2onnx ${PADDLE2ONNX_SHARE_LIB} ${PADDLE2ONNX_LIB}
+                SRCS ${PADDLE2ONNX_INC_DIR}/paddle2onnx ${PADDLE2ONNX_SHARED_LIB} ${PADDLE2ONNX_LIB}
                 DSTS ${dst_dir}/include ${dst_dir}/lib ${dst_dir}/lib)
         else()
             copy(${TARGET}

diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -36,7 +36,6 @@
 #include "paddle/fluid/inference/analysis/helper.h"
 #include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
 #include "paddle/fluid/inference/api/helper.h"
-#include "paddle/fluid/inference/api/onnxruntime_predictor.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_inference_pass.h"
 #include "paddle/fluid/inference/utils/io_utils.h"
@@ -57,6 +56,10 @@
 #include "paddle/fluid/inference/api/mkldnn_quantizer.h"
 #endif
 
+#ifdef PADDLE_WITH_ONNXRUNTIME
+#include "paddle/fluid/inference/api/onnxruntime_predictor.h"
+#endif
+
 #if PADDLE_WITH_TENSORRT
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/inference/tensorrt/helper.h"
@@ -1495,8 +1498,8 @@ Predictor::Predictor(const Config &config) {
   if (config.use_onnxruntime()) {
 #ifdef PADDLE_WITH_ONNXRUNTIME
     if (config.use_gpu()) {
-      LOG(WARNING) << "ONNXRuntime not support gpu for now， fall back to "
-                      "using Paddle Inference.";
+      LOG(WARNING) << "The current ONNXRuntime backend doesn't support GPU,"
+                      "and it falls back to use Paddle Inference.";
     } else if (!paddle::CheckConvertToONNX(config)) {
       LOG(WARNING)
           << "Paddle2ONNX do't support convert the Model， fall back to using "

diff --git a/paddle/fluid/inference/api/onnxruntime_predictor.cc b/paddle/fluid/inference/api/onnxruntime_predictor.cc
@@ -66,10 +66,6 @@ framework::proto::VarType::Type ConvertONNXType(
   }
 }
 
-/*
-bool CheckConvertToONNX(const AnalysisConfig &config) { return true; }
-*/
-
 bool CheckConvertToONNX(const AnalysisConfig &config) {
   if (!config.model_dir().empty()) {
     LOG(ERROR) << "Paddle2ONNX not support model_dir config";
@@ -104,11 +100,14 @@ bool ONNXRuntimePredictor::Init() {
                       config_.model_from_memory());
 
   Ort::SessionOptions session_options;
-  // Turn optimization off first, and then turn it on when it's stabl
+  // Turn optimization off first, and then turn it on when it's stable
   // session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
   // session_options.SetExecutionMode(ExecutionMode::ORT_SEQUENTIAL);
   // session_options.EnableCpuMemArena();
   // session_options.EnableMemPattern();
+  // session_options.SetInterOpNumThreads(config_.cpu_math_library_num_threads());
+  session_options.SetIntraOpNumThreads(config_.cpu_math_library_num_threads());
+  VLOG(2) << "ONNXRuntime threads " << config_.cpu_math_library_num_threads();
   if (config_.profile_enabled()) {
     LOG(WARNING) << "ONNXRuntime Profiler is activated, which might affect the "
                     "performance";
@@ -122,10 +121,6 @@ bool ONNXRuntimePredictor::Init() {
                "will be "
                "generated.";
   }
-  // session_options.SetInterOpNumThreads(config_.cpu_math_library_num_threads());
-  session_options.SetIntraOpNumThreads(config_.cpu_math_library_num_threads());
-  VLOG(2) << "ONNXRuntime threads " << config_.cpu_math_library_num_threads();
-  // session_ = {env_, config_.prog_file().c_str(), session_options};
   session_ = {env_, onnx_proto.data(), onnx_proto.size(), session_options};
 
   auto memory_info =

diff --git a/paddle/fluid/inference/api/onnxruntime_predictor.h b/paddle/fluid/inference/api/onnxruntime_predictor.h
@@ -28,11 +28,9 @@
 #include "paddle/fluid/platform/device/gpu/gpu_types.h"
 #include "paddle/fluid/string/printf.h"
 
-#ifdef PADDLE_WITH_ONNXRUNTIME
 #include "onnxruntime_c_api.h"    // NOLINT
 #include "onnxruntime_cxx_api.h"  // NOLINT
 #include "paddle2onnx/converter.h"
-#endif
 
 ///
 /// \file onnxruntime_predictor.h
@@ -51,9 +49,7 @@ bool CheckConvertToONNX(const AnalysisConfig &config);
 struct ONNXDesc {
   std::string name;
   std::vector<int64_t> shape;
-#ifdef PADDLE_WITH_ONNXRUNTIME
   ONNXTensorElementDataType dtype;
-#endif
 };
 
 ///
@@ -95,9 +91,7 @@ class ONNXRuntimePredictor : public PaddlePredictor {
   explicit ONNXRuntimePredictor(const AnalysisConfig &config)
       : config_(config) {
     predictor_id_ = inference::GetUniqueId();
-#ifdef PADDLE_WITH_ONNXRUNTIME
     env_ = Ort::Env(ORT_LOGGING_LEVEL_VERBOSE, "onnx");
-#endif
   }
   ///
   /// \brief Destroy the ONNXRuntime Predictor object
@@ -181,7 +175,6 @@ class ONNXRuntimePredictor : public PaddlePredictor {
   std::shared_ptr<framework::Scope> scope_;
 
  private:
-#ifdef PADDLE_WITH_ONNXRUNTIME
   ///
   /// \brief get the Ort Value(input Tensor).
   ///
@@ -203,16 +196,13 @@ class ONNXRuntimePredictor : public PaddlePredictor {
   /// \return get a Ort::Value
   ///
   void AsTensor(const Ort::Value &value, const ONNXDesc &desc);
-#endif
 
  private:
   AnalysisConfig config_;
 
-#ifdef PADDLE_WITH_ONNXRUNTIME
   // ONNXRuntime
   Ort::Env env_;
   Ort::Session session_{nullptr};
-#endif
 
   platform::Place place_;
   framework::Scope *sub_scope_{nullptr};

diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt
@@ -246,13 +246,13 @@ if(WITH_PYTHON)
     endif()
     if(WITH_ONNXRUNTIME)
       ADD_CUSTOM_COMMAND(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/paddle2onnx.dll
-        COMMAND ${CMAKE_COMMAND} -E copy ${PADDLE2ONNX_SHARE_LIB} ${CMAKE_CURRENT_BINARY_DIR}
+        COMMAND ${CMAKE_COMMAND} -E copy ${PADDLE2ONNX_SHARED_LIB} ${CMAKE_CURRENT_BINARY_DIR}
         DEPENDS paddle2onnx)
       list(APPEND OP_IMPL_DEPS ${CMAKE_CURRENT_BINARY_DIR}/paddle2onnx.dll)
       list(APPEND EAGER_OP_IMPL_DEPS ${CMAKE_CURRENT_BINARY_DIR}/paddle2onnx.dll)
 
       ADD_CUSTOM_COMMAND(OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/onnxruntime.dll
-        COMMAND ${CMAKE_COMMAND} -E copy ${ONNXRUNTIME_SHARE_LIB} ${CMAKE_CURRENT_BINARY_DIR} 
+        COMMAND ${CMAKE_COMMAND} -E copy ${ONNXRUNTIME_SHARED_LIB} ${CMAKE_CURRENT_BINARY_DIR} 
         DEPENDS onnxruntime)
       list(APPEND OP_IMPL_DEPS ${CMAKE_CURRENT_BINARY_DIR}/onnxruntime.dll)
       list(APPEND EAGER_OP_IMPL_DEPS ${CMAKE_CURRENT_BINARY_DIR}/onnxruntime.dll)

diff --git a/paddle/fluid/pybind/inference_api.cc b/paddle/fluid/pybind/inference_api.cc
@@ -28,12 +28,15 @@
 #include <vector>
 #include "paddle/fluid/inference/api/analysis_predictor.h"
 #include "paddle/fluid/inference/api/helper.h"
-#include "paddle/fluid/inference/api/onnxruntime_predictor.h"
 #include "paddle/fluid/inference/api/paddle_infer_contrib.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 #include "paddle/fluid/inference/api/paddle_pass_builder.h"
 #include "paddle/fluid/inference/utils/io_utils.h"
 
+#ifdef PADDLE_WITH_ONNXRUNTIME
+#include "paddle/fluid/inference/api/onnxruntime_predictor.h"
+#endif
+
 namespace py = pybind11;
 
 namespace pybind11 {
@@ -559,7 +562,7 @@ void BindAnalysisConfig(py::module *m) {
       .def("disable_gpu", &AnalysisConfig::DisableGpu)
       .def("enable_onnxruntime", &AnalysisConfig::EnableONNXRuntime)
       .def("disable_onnxruntime", &AnalysisConfig::DisableONNXRuntime)
-      .def("onnxruntime_enable", &AnalysisConfig::use_onnxruntime)
+      .def("onnxruntime_enabled", &AnalysisConfig::use_onnxruntime)
       .def("use_gpu", &AnalysisConfig::use_gpu)
       .def("use_xpu", &AnalysisConfig::use_xpu)
       .def("use_npu", &AnalysisConfig::use_npu)

diff --git a/python/setup.py.in b/python/setup.py.in
@@ -504,9 +504,9 @@ if '${WITH_MKLDNN}' == 'ON':
         package_data['paddle.libs']+=['mkldnn.dll']
 
 if '${WITH_ONNXRUNTIME}' == 'ON':
-    shutil.copy('${ONNXRUNTIME_SHARE_LIB}', libs_path)
+    shutil.copy('${ONNXRUNTIME_SHARED_LIB}', libs_path)
     if os.name == 'nt':
-        shutil.copy('${PADDLE2ONNX_SHARE_LIB}', libs_path)
+        shutil.copy('${PADDLE2ONNX_SHARED_LIB}', libs_path)
         package_data['paddle.libs']+=['paddle2onnx.dll', 'onnxruntime.dll']
     else:
         shutil.copy('${PADDLE2ONNX_LIB}', libs_path)