[Backend] Add fp16 support for ONNXRuntime-GPU (#1239)

* add fp16 support for ort-gpu

* add enable_ort_fp16 option

* fix

* fix bugs

* use fp16 for test

* update code

* update code

* fix windows bug

cmake/paddle2onnx.cmake

@@ -43,7 +43,7 @@ else()
 endif(WIN32)
 
 set(PADDLE2ONNX_URL_BASE "https://bj.bcebos.com/fastdeploy/third_libs/")
-set(PADDLE2ONNX_VERSION "1.0.5")
+set(PADDLE2ONNX_VERSION "1.0.6")
 if(WIN32)
   set(PADDLE2ONNX_FILE "paddle2onnx-win-x64-${PADDLE2ONNX_VERSION}.zip")
   if(NOT CMAKE_CL_64)

fastdeploy/runtime/backends/ort/option.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include "fastdeploy/core/fd_type.h"
+#include "fastdeploy/runtime/enum_variables.h"
 #include <iostream>
 #include <memory>
 #include <string>
@@ -25,18 +26,27 @@ namespace fastdeploy {
 /*! @brief Option object to configure ONNX Runtime backend
  */
 struct OrtBackendOption {
-  /// Level of graph optimization, -1: mean default(Enable all the optimization strategy)/0: disable all the optimization strategy/1: enable basic strategy/2:enable extend strategy/99: enable all
+  /// Level of graph optimization,
+  ///         /-1: mean default(Enable all the optimization strategy)
+  ///         /0: disable all the optimization strategy/1: enable basic strategy
+  ///         /2:enable extend strategy/99: enable all
   int graph_optimization_level = -1;
   /// Number of threads to execute the operator, -1: default
   int intra_op_num_threads = -1;
-  /// Number of threads to execute the graph, -1: default. This parameter only will bring effects while the `OrtBackendOption::execution_mode` set to 1.
+  /// Number of threads to execute the graph,
+  ///         -1: default. This parameter only will bring effects
+  ///         while the `OrtBackendOption::execution_mode` set to 1.
   int inter_op_num_threads = -1;
-  /// Execution mode for the graph, -1: default(Sequential mode)/0: Sequential mode, execute the operators in graph one by one. /1: Parallel mode, execute the operators in graph parallelly.
+  /// Execution mode for the graph, -1: default(Sequential mode)
+  ///         /0: Sequential mode, execute the operators in graph one by one.
+  ///         /1: Parallel mode, execute the operators in graph parallelly.
   int execution_mode = -1;
-  // Inference device, OrtBackend supports CPU/GPU
+  /// Inference device, OrtBackend supports CPU/GPU
   Device device = Device::CPU;
-  // Inference device id
+  /// Inference device id
   int device_id = 0;
   void* external_stream_ = nullptr;
+  /// Use fp16 to infer
+  bool enable_fp16 = false;
 };
 }  // namespace fastdeploy

fastdeploy/runtime/backends/ort/option_pybind.cc

@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "fastdeploy/pybind/main.h"
 #include "fastdeploy/runtime/backends/ort/option.h"
+#include "fastdeploy/pybind/main.h"
 
 namespace fastdeploy {
 
@@ -28,7 +28,8 @@ void BindOrtOption(pybind11::module& m) {
                      &OrtBackendOption::inter_op_num_threads)
       .def_readwrite("execution_mode", &OrtBackendOption::execution_mode)
       .def_readwrite("device", &OrtBackendOption::device)
-      .def_readwrite("device_id", &OrtBackendOption::device_id);
+      .def_readwrite("device_id", &OrtBackendOption::device_id)
+      .def_readwrite("enable_fp16", &OrtBackendOption::enable_fp16);
 }
 
 }  // namespace fastdeploy

fastdeploy/runtime/backends/ort/ort_backend.cc

@@ -181,22 +181,21 @@ bool OrtBackend::InitFromPaddle(const std::string& model_buffer,
   strcpy(ops[1].op_name, "pool2d");
   strcpy(ops[1].export_op_name, "AdaptivePool2d");
 
-  if (!paddle2onnx::Export(model_buffer.c_str(), model_buffer.size(),
-                           params_buffer.c_str(), params_buffer.size(),
-                           &model_content_ptr, &model_content_size, 11, true,
-                           verbose, true, true, true, ops.data(), 2,
-                           "onnxruntime", nullptr, 0, "", &save_external)) {
+  if (!paddle2onnx::Export(
+          model_buffer.c_str(), model_buffer.size(), params_buffer.c_str(),
+          params_buffer.size(), &model_content_ptr, &model_content_size, 11,
+          true, verbose, true, true, true, ops.data(), 2, "onnxruntime",
+          nullptr, 0, "", &save_external, false)) {
     FDERROR << "Error occured while export PaddlePaddle to ONNX format."
             << std::endl;
     return false;
   }
-
   std::string onnx_model_proto(model_content_ptr,
                                model_content_ptr + model_content_size);
   delete[] model_content_ptr;
   model_content_ptr = nullptr;
   if (save_external) {
-    std::string model_file_name = "model.onnx";
+    model_file_name = "model.onnx";
     std::fstream f(model_file_name, std::ios::out);
     FDASSERT(f.is_open(), "Can not open file: %s to save model.",
              model_file_name.c_str());
@@ -219,6 +218,22 @@ bool OrtBackend::InitFromOnnx(const std::string& model_file,
             << std::endl;
     return false;
   }
+  std::string onnx_model_buffer;
+  if (option.enable_fp16) {
+    if (option.device == Device::CPU) {
+      FDWARNING << "Turning on FP16 on CPU may result in slower inference."
+                << std::endl;
+    }
+    char* model_content_ptr;
+    int model_content_size = 0;
+    paddle2onnx::ConvertFP32ToFP16(model_file.c_str(), model_file.size(),
+                                   &model_content_ptr, &model_content_size);
+    std::string onnx_model_proto(model_content_ptr,
+                                 model_content_ptr + model_content_size);
+    onnx_model_buffer = onnx_model_proto;
+  } else {
+    onnx_model_buffer = model_file;
+  }
 
   if (!BuildOption(option)) {
     FDERROR << "Create Ort option fail." << std::endl;
@@ -226,7 +241,19 @@ bool OrtBackend::InitFromOnnx(const std::string& model_file,
   }
 
   InitCustomOperators();
-  session_ = {env_, model_file.data(), model_file.size(), session_options_};
+  if (model_file_name.size()) {
+#ifdef WIN32
+    std::wstring widestr =
+        std::wstring(model_file_name.begin(), model_file_name.end());
+    session_ = {env_, widestr.c_str(), session_options_};
+#else
+    session_ = {env_, model_file_name.c_str(), session_options_};
+#endif
+  } else {
+    session_ = {env_, onnx_model_buffer.data(), onnx_model_buffer.size(),
+                session_options_};
+  }
+
   binding_ = std::make_shared<Ort::IoBinding>(session_);
 
   Ort::MemoryInfo memory_info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);

fastdeploy/runtime/backends/ort/ort_backend.h

@@ -74,6 +74,10 @@ class OrtBackend : public BaseBackend {
   std::shared_ptr<Ort::IoBinding> binding_;
   std::vector<OrtValueInfo> inputs_desc_;
   std::vector<OrtValueInfo> outputs_desc_;
+
+  // the ONNX model file name,
+  // when ONNX is bigger than 2G, we will set this name
+  std::string model_file_name;
 #ifndef NON_64_PLATFORM
   Ort::CustomOpDomain custom_op_domain_ = Ort::CustomOpDomain("Paddle");
 #endif

fastdeploy/runtime/runtime_option.h

@@ -14,7 +14,6 @@
 
 /*! \file runtime_option.h
     \brief A brief file description.
-
     More details
  */
 
@@ -160,7 +159,8 @@ struct FASTDEPLOY_DECL RuntimeOption {
 
   //  \brief Set the profile mode as 'true'.
   //
-  // \param[in] inclue_h2d_d2h Whether to include time of H2D_D2H for time of runtime.
+  // \param[in] inclue_h2d_d2h Whether to
+  //            include time of H2D_D2H for time of runtime.
   // \param[in] repeat Repeat times for runtime inference.
   // \param[in] warmup Warmup times for runtime inference.
   //
@@ -178,12 +178,14 @@ struct FASTDEPLOY_DECL RuntimeOption {
     benchmark_option.enable_profile = false;
   }
 
-  // \brief Enable to check if current backend set by user can be found at valid_xxx_backend.
+  // \brief Enable to check if current backend set by
+  //        user can be found at valid_xxx_backend.
   //
   void EnableValidBackendCheck() {
     enable_valid_backend_check = true;
   }
-  // \brief Disable to check if current backend set by user can be found at valid_xxx_backend.
+  // \brief Disable to check if current backend set by
+  //        user can be found at valid_xxx_backend.
   //
   void DisableValidBackendCheck() {
     enable_valid_backend_check = false;

fastdeploy/vision/facedet/contrib/centerface/preprocessor.h

@@ -51,7 +51,6 @@ class FASTDEPLOY_DECL CenterFacePreprocessor{
 
   // target size, tuple of (width, height), default size = {640, 640}
   std::vector<int> size_;
-
 };
 
 }  // namespace facedet