[Backend] optimize cpu usage on openvino backend (#1531)

* optimiza cpu usage on openvino backend

conver the hint varible to string

* update with affinity mode set-up

---------

Co-authored-by: Ethan <yicheng@intel.com>
Co-authored-by: Jason <jiangjiajun@baidu.com>

cmake/openvino.cmake

@@ -18,6 +18,12 @@ if (OPENVINO_DIRECTORY)
   message(STATUS "Use the openvino lib specified by user. The OpenVINO path: ${OPENVINO_DIRECTORY}")
   STRING(REGEX REPLACE "\\\\" "/" OPENVINO_DIRECTORY ${OPENVINO_DIRECTORY})
   get_openvino_libs(${OPENVINO_DIRECTORY}/runtime)
+  include_directories(${OPENVINO_DIRECTORY}/runtime/include ${OPENVINO_DIRECTORY}/runtime/include/ie)
+  set(OPENVINO_LIB_DIR
+    "${OPENVINO_DIRECTORY}/runtime/lib/intel64"
+    CACHE PATH "openvino lib directory." FORCE)
+  set(CMAKE_BUILD_RPATH "${CMAKE_BUILD_RPATH}" "${OPENVINO_LIB_DIR}")
+
 else()
   set(OPENVINO_PROJECT "extern_openvino")
 

fastdeploy/runtime/backends/openvino/option.h

@@ -30,7 +30,13 @@ struct OpenVINOBackendOption {
   int cpu_thread_num = -1;
 
   /// Number of streams while use OpenVINO
-  int num_streams = 0;
+  int num_streams = 1;
+
+  /// Affinity mode
+  std::string affinity  = "YES";
+
+  /// Performance hint mode
+  std::string hint = "UNDEFINED";
 
   /**
    * @brief Set device name for OpenVINO, default 'CPU', can also be 'AUTO', 'GPU', 'GPU.1'....
@@ -56,6 +62,39 @@ struct OpenVINOBackendOption {
     }
   }
 
+  /**
+   * @brief Set Affinity mode
+   */
+  void SetAffinity (const std::string& _affinity) {
+    FDASSERT(_affinity == "YES" || _affinity == "NO" || _affinity == "NUMA" ||
+                 _affinity == "HYBRID_AWARE",
+             "The affinity mode should be one of the list "
+             "['YES', 'NO', 'NUMA', "
+             "'HYBRID_AWARE'] ");
+    affinity = _affinity;
+  }
+
+  /**
+   * @brief Set the Performance Hint
+   */
+  void SetPerformanceHint (const std::string& _hint) {
+    FDASSERT(_hint == "LATENCY" || _hint == "THROUGHPUT" ||
+                 _hint == "CUMULATIVE_THROUGHPUT" || _hint == "UNDEFINED",
+             "The performance hint should be one of the list "
+             "['LATENCY', 'THROUGHPUT', 'CUMULATIVE_THROUGHPUT', "
+             "'UNDEFINED'] ");
+    hint = _hint;
+  }
+
+  /**
+   * @brief Set the number of streams
+   */
+  void SetStreamNum (int _num_streams) {
+    FDASSERT(_num_streams > 0, "The stream_num must be greater than 0.");
+    num_streams = _num_streams;
+  }
+  
+
   std::map<std::string, std::vector<int64_t>> shape_infos;
   std::set<std::string> cpu_operators{"MulticlassNms"};
 };

fastdeploy/runtime/backends/openvino/option_pybind.cc

@@ -22,9 +22,14 @@ void BindOpenVINOOption(pybind11::module& m) {
       .def(pybind11::init())
       .def_readwrite("cpu_thread_num", &OpenVINOBackendOption::cpu_thread_num)
       .def_readwrite("num_streams", &OpenVINOBackendOption::num_streams)
+      .def_readwrite("affinity", &OpenVINOBackendOption::affinity)
+      .def_readwrite("hint", &OpenVINOBackendOption::hint)
       .def("set_device", &OpenVINOBackendOption::SetDevice)
       .def("set_shape_info", &OpenVINOBackendOption::SetShapeInfo)
-      .def("set_cpu_operators", &OpenVINOBackendOption::SetCpuOperators);
+      .def("set_cpu_operators", &OpenVINOBackendOption::SetCpuOperators)
+      .def("set_affinity", &OpenVINOBackendOption::SetAffinity)
+      .def("set_performance_hint", &OpenVINOBackendOption::SetPerformanceHint)
+      .def("set_stream_num", &OpenVINOBackendOption::SetStreamNum);
 }
 
 }  // namespace fastdeploy

fastdeploy/runtime/backends/openvino/ov_backend.cc

@@ -203,10 +203,10 @@ bool OpenVINOBackend::InitFromPaddle(const std::string& model_file,
   }
 
   ov::AnyMap properties;
+  if (option_.hint == "UNDEFINED") {
     if (option_.device == "CPU" && option_.cpu_thread_num > 0) {
       properties["INFERENCE_NUM_THREADS"] = option_.cpu_thread_num;
     }
-  if (option_.device == "CPU") {
     if (option_.num_streams == -1) {
       properties["NUM_STREAMS"] = ov::streams::AUTO;
     } else if (option_.num_streams == -2) {
@@ -214,17 +214,32 @@ bool OpenVINOBackend::InitFromPaddle(const std::string& model_file,
     } else if (option_.num_streams > 0) {
       properties["NUM_STREAMS"] = option_.num_streams;
     }
-  } else {
-    if (option_.num_streams != 0) {
-      FDWARNING << "NUM_STREAMS only available on device CPU, currently the "
-                   "device is set as "
-                << option_.device << ", the NUM_STREAMS will be ignored."
-                << std::endl;
+
+    FDINFO << "number of streams:" << option_.num_streams << "." << std::endl;
+    if (option_.affinity == "YES") {
+      properties["AFFINITY"] = "CORE";
+    } else if (option_.affinity == "NO") {
+      properties["AFFINITY"] = "NONE";
+    } else if (option_.affinity == "NUMA") {
+      properties["AFFINITY"] = "NUMA";
+    } else if (option_.affinity == "HYBRID_AWARE") {
+      properties["AFFINITY"] = "HYBRID_AWARE";
     }
+    FDINFO << "affinity:" << option_.affinity << "." << std::endl;
+  } else if (option_.hint == "LATENCY") {
+    properties.emplace(
+        ov::hint::performance_mode(ov::hint::PerformanceMode::LATENCY));
+  } else if (option_.hint == "THROUGHPUT") {
+    properties.emplace(
+        ov::hint::performance_mode(ov::hint::PerformanceMode::THROUGHPUT));
+  } else if (option_.hint == "CUMULATIVE_THROUGHPUT") {
+    properties.emplace(ov::hint::performance_mode(
+        ov::hint::PerformanceMode::CUMULATIVE_THROUGHPUT));
   }
 
   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();
@@ -332,10 +347,10 @@ bool OpenVINOBackend::InitFromOnnx(const std::string& model_file,
   }
 
   ov::AnyMap properties;
+  if (option_.hint == "UNDEFINED") {
     if (option_.device == "CPU" && option_.cpu_thread_num > 0) {
       properties["INFERENCE_NUM_THREADS"] = option_.cpu_thread_num;
     }
-  if (option_.device == "CPU") {
     if (option_.num_streams == -1) {
       properties["NUM_STREAMS"] = ov::streams::AUTO;
     } else if (option_.num_streams == -2) {
@@ -343,13 +358,27 @@ bool OpenVINOBackend::InitFromOnnx(const std::string& model_file,
     } else if (option_.num_streams > 0) {
       properties["NUM_STREAMS"] = option_.num_streams;
     }
-  } else {
-    if (option_.num_streams != 0) {
-      FDWARNING << "NUM_STREAMS only available on device CPU, currently the "
-                   "device is set as "
-                << option_.device << ", the NUM_STREAMS will be ignored."
-                << std::endl;
+
+    FDINFO << "number of streams:" << option_.num_streams << "." << std::endl;
+    if (option_.affinity == "YES") {
+      properties["AFFINITY"] = "CORE";
+    } else if (option_.affinity == "NO") {
+      properties["AFFINITY"] = "NONE";
+    } else if (option_.affinity == "NUMA") {
+      properties["AFFINITY"] = "NUMA";
+    } else if (option_.affinity == "HYBRID_AWARE") {
+      properties["AFFINITY"] = "HYBRID_AWARE";
     }
+    FDINFO << "affinity:" << option_.affinity << "." << std::endl;
+  } else if (option_.hint == "LATENCY") {
+    properties.emplace(
+        ov::hint::performance_mode(ov::hint::PerformanceMode::LATENCY));
+  } else if (option_.hint == "THROUGHPUT") {
+    properties.emplace(
+        ov::hint::performance_mode(ov::hint::PerformanceMode::THROUGHPUT));
+  } else if (option_.hint == "CUMULATIVE_THROUGHPUT") {
+    properties.emplace(ov::hint::performance_mode(
+        ov::hint::PerformanceMode::CUMULATIVE_THROUGHPUT));
   }
 
   FDINFO << "Compile OpenVINO model on device_name:" << option.device << "."
@@ -384,7 +413,8 @@ bool OpenVINOBackend::Infer(std::vector<FDTensor>& inputs,
   }
 
   RUNTIME_PROFILE_LOOP_BEGIN(1)
-  request_.infer();
+  request_.start_async();
+  request_.wait();
   RUNTIME_PROFILE_LOOP_END
 
   outputs->resize(output_infos_.size());