[Backend] Support Intel GPU with OpenVINO (PaddlePaddle#472)

* Update ov_backend.cc * Update ov_backend.cc * support set openvino device
felixhjh · Nov 25, 2022 · ff02ff0 · ff02ff0
1 parent 4c4397a
commit ff02ff0
Show file tree

Hide file tree

Showing 7 changed files with 24 additions and 3 deletions.
diff --git a/fastdeploy/backends/openvino/ov_backend.cc b/fastdeploy/backends/openvino/ov_backend.cc
@@ -158,7 +158,8 @@ bool OpenVINOBackend::InitFromPaddle(const std::string& model_file,
   } else if (option_.ov_num_streams > 0) {
     properties["NUM_STREAMS"] = option_.ov_num_streams;
   }
-  compiled_model_ = core_.compile_model(model, "CPU", properties);
+  FDINFO << "Compile OpenVINO model on device_name:" << option.device << "." << std::endl;
+  compiled_model_ = core_.compile_model(model, option.device, properties);
 
   request_ = compiled_model_.create_infer_request();
   initialized_ = true;
@@ -255,7 +256,8 @@ bool OpenVINOBackend::InitFromOnnx(const std::string& model_file,
   } else if (option_.ov_num_streams > 0) {
     properties["NUM_STREAMS"] = option_.ov_num_streams;
   }
-  compiled_model_ = core_.compile_model(model, "CPU", properties);
+  FDINFO << "Compile OpenVINO model on device_name:" << option.device << "." << std::endl;
+  compiled_model_ = core_.compile_model(model, option.device, properties);
 
   request_ = compiled_model_.create_infer_request();
 

diff --git a/fastdeploy/backends/openvino/ov_backend.h b/fastdeploy/backends/openvino/ov_backend.h
@@ -26,6 +26,7 @@
 namespace fastdeploy {
 
 struct OpenVINOBackendOption {
+  std::string device = "CPU";
   int cpu_thread_num = -1;
   int ov_num_streams = 1;
   std::map<std::string, std::vector<int64_t>> shape_infos;

diff --git a/fastdeploy/pybind/runtime.cc b/fastdeploy/pybind/runtime.cc
@@ -33,6 +33,7 @@ void BindRuntime(pybind11::module& m) {
       .def("use_openvino_backend", &RuntimeOption::UseOpenVINOBackend)
       .def("use_lite_backend", &RuntimeOption::UseLiteBackend)
       .def("set_paddle_mkldnn", &RuntimeOption::SetPaddleMKLDNN)
+      .def("set_openvino_device", &RuntimeOption::SetOpenVINODevice)
       .def("enable_paddle_log_info", &RuntimeOption::EnablePaddleLogInfo)
       .def("disable_paddle_log_info", &RuntimeOption::DisablePaddleLogInfo)
       .def("set_paddle_mkldnn_cache_size",

diff --git a/fastdeploy/runtime.cc b/fastdeploy/runtime.cc
@@ -332,6 +332,10 @@ void RuntimeOption::SetPaddleMKLDNNCacheSize(int size) {
   pd_mkldnn_cache_size = size;
 }
 
+void RuntimeOption::SetOpenVINODevice(const std::string& name) {
+  openvino_device = name;
+}
+
 void RuntimeOption::EnableLiteFP16() {
   lite_enable_fp16 = true;
 }
@@ -641,6 +645,7 @@ void Runtime::CreateOpenVINOBackend() {
 #ifdef ENABLE_OPENVINO_BACKEND
   auto ov_option = OpenVINOBackendOption();
   ov_option.cpu_thread_num = option.cpu_thread_num;
+  ov_option.device = option.openvino_device;
   ov_option.ov_num_streams = option.ov_num_streams;
   FDASSERT(option.model_format == ModelFormat::PADDLE ||
                option.model_format == ModelFormat::ONNX,

diff --git a/fastdeploy/runtime.h b/fastdeploy/runtime.h
@@ -168,6 +168,11 @@ struct FASTDEPLOY_DECL RuntimeOption {
    */
   void SetPaddleMKLDNNCacheSize(int size);
 
+  /**
+   * @brief Set device name for OpenVINO, default 'CPU', can also be 'AUTO', 'GPU', 'GPU.1'....
+   */
+   void SetOpenVINODevice(const std::string& name = "CPU");
+
   /**
    * @brief Set optimzed model dir for Paddle Lite backend.
    */
@@ -344,6 +349,9 @@ struct FASTDEPLOY_DECL RuntimeOption {
   size_t trt_max_batch_size = 32;
   size_t trt_max_workspace_size = 1 << 30;
 
+  // ======Only for OpenVINO Backend======
+  std::string openvino_device = "CPU";
+
   // ======Only for Poros Backend=======
   bool is_dynamic = false;
   bool long_to_int = true;

diff --git a/fastdeploy/vision/common/processors/transform.h b/fastdeploy/vision/common/processors/transform.h
@@ -37,7 +37,6 @@ namespace fastdeploy {
 namespace vision {
 
 void FuseTransforms(std::vector<std::shared_ptr<Processor>>* processors);
-
 // Fuse Normalize + Cast(Float) to Normalize
 void FuseNormalizeCast(std::vector<std::shared_ptr<Processor>>* processors);
 // Fuse Normalize + HWC2CHW to NormalizeAndPermute

diff --git a/python/fastdeploy/runtime.py b/python/fastdeploy/runtime.py
@@ -269,6 +269,11 @@ def set_paddle_mkldnn(self, use_mkldnn=True):
         """
         return self._option.set_paddle_mkldnn(use_mkldnn)
 
+    def set_openvino_device(self, name="CPU"):
+        """Set device name for OpenVINO, default 'CPU', can also be 'AUTO', 'GPU', 'GPU.1'....
+        """
+        return self._option.set_openvino_device(name)
+
     def enable_paddle_log_info(self):
         """Enable print out the debug log information while using Paddle Inference backend, the log information is disabled by default.
         """