[lite] Add threads and power_mode option support (#298)

* [cmake] support Android arm64-v8a & armeabi-v7a native c++ sdk

* [cmake] fixed patchelf download on mac and android

* [lite] Add threads and power_mode option support

* [pybind] update runtime pybind for lite power mode

* [python] Add set_lite_power_mode api to runtime

fastdeploy/backends/lite/lite_backend.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "fastdeploy/backends/lite/lite_backend.h"
+
 #include <cstring>
 
 namespace fastdeploy {
@@ -40,8 +41,16 @@ FDDataType LiteDataTypeToFD(const paddle::lite_api::PrecisionType& dtype) {
 
 void LiteBackend::BuildOption(const LiteBackendOption& option) {
   std::vector<paddle::lite_api::Place> valid_places;
-  valid_places.push_back(paddle::lite_api::Place{TARGET(kARM), PRECISION(kFloat)});
+  valid_places.push_back(
+      paddle::lite_api::Place{TARGET(kARM), PRECISION(kFloat)});
   config_.set_valid_places(valid_places);
+  if (option.threads > 0) {
+    config_.set_threads(option.threads);
+  }
+  if (option.power_mode > 0) {
+    config_.set_power_mode(
+        static_cast<paddle::lite_api::PowerMode>(option.power_mode));
+  }
 }
 
 bool LiteBackend::InitFromPaddle(const std::string& model_file,
@@ -56,7 +65,9 @@ bool LiteBackend::InitFromPaddle(const std::string& model_file,
   config_.set_model_file(model_file);
   config_.set_param_file(params_file);
   BuildOption(option);
-  predictor_ = paddle::lite_api::CreatePaddlePredictor<paddle::lite_api::CxxConfig>(config_);
+  predictor_ =
+      paddle::lite_api::CreatePaddlePredictor<paddle::lite_api::CxxConfig>(
+          config_);
 
   inputs_desc_.clear();
   outputs_desc_.clear();
@@ -103,9 +114,7 @@ TensorInfo LiteBackend::GetOutputInfo(int index) {
   return outputs_desc_[index];
 }
 
-std::vector<TensorInfo> LiteBackend::GetOutputInfos() {
-  return outputs_desc_;
-}
+std::vector<TensorInfo> LiteBackend::GetOutputInfos() { return outputs_desc_; }
 
 bool LiteBackend::Infer(std::vector<FDTensor>& inputs,
                         std::vector<FDTensor>* outputs) {
@@ -119,12 +128,15 @@ bool LiteBackend::Infer(std::vector<FDTensor>& inputs,
   for (size_t i = 0; i < inputs.size(); ++i) {
     auto iter = inputs_order_.find(inputs[i].name);
     if (iter == inputs_order_.end()) {
-      FDERROR << "Cannot find input with name:" << inputs[i].name << " in loaded model." << std::endl;
+      FDERROR << "Cannot find input with name:" << inputs[i].name
+              << " in loaded model." << std::endl;
       return false;
     }
     auto tensor = predictor_->GetInput(iter->second);
     tensor->Resize(inputs[i].shape);
-    tensor->ShareExternalMemory(const_cast<void*>(inputs[i].CpuData()), inputs[i].Nbytes(), paddle::lite_api::TargetType::kARM);
+    tensor->ShareExternalMemory(const_cast<void*>(inputs[i].CpuData()),
+                                inputs[i].Nbytes(),
+                                paddle::lite_api::TargetType::kARM);
   }
 
   predictor_->Run();
@@ -132,8 +144,10 @@ bool LiteBackend::Infer(std::vector<FDTensor>& inputs,
   outputs->resize(outputs_desc_.size());
   for (size_t i = 0; i < outputs_desc_.size(); ++i) {
     auto tensor = predictor_->GetOutput(i);
-    (*outputs)[i].Resize(tensor->shape(), outputs_desc_[i].dtype, outputs_desc_[i].name);
-    memcpy((*outputs)[i].MutableData(), tensor->data<void>(), (*outputs)[i].Nbytes());
+    (*outputs)[i].Resize(tensor->shape(), outputs_desc_[i].dtype,
+                         outputs_desc_[i].name);
+    memcpy((*outputs)[i].MutableData(), tensor->data<void>(),
+           (*outputs)[i].Nbytes());
   }
   return true;
 }

fastdeploy/backends/lite/lite_backend.h

@@ -25,6 +25,18 @@
 namespace fastdeploy {
 
 struct LiteBackendOption {
+  // cpu num threads
+  int threads = 1;
+  // lite power mode
+  // 0: LITE_POWER_HIGH
+  // 1: LITE_POWER_LOW
+  // 2: LITE_POWER_FULL
+  // 3: LITE_POWER_NO_BIND
+  // 4: LITE_POWER_RAND_HIGH
+  // 5: LITE_POWER_RAND_LOW
+  int power_mode = 0;
+  // TODO(qiuyanjun): support more options for lite backend.
+  // Such as fp16, different device target (kARM/kXPU/kNPU/...)
 };
 
 // Convert data type from paddle lite to fastdeploy

fastdeploy/pybind/runtime.cc

@@ -34,6 +34,7 @@ void BindRuntime(pybind11::module& m) {
       .def("disable_paddle_log_info", &RuntimeOption::DisablePaddleLogInfo)
       .def("set_paddle_mkldnn_cache_size",
            &RuntimeOption::SetPaddleMKLDNNCacheSize)
+      .def("set_lite_power_mode", &RuntimeOption::SetLitePowerMode)
       .def("set_trt_input_shape", &RuntimeOption::SetTrtInputShape)
       .def("enable_trt_fp16", &RuntimeOption::EnableTrtFP16)
       .def("disable_trt_fp16", &RuntimeOption::DisableTrtFP16)

fastdeploy/runtime.cc

@@ -148,7 +148,9 @@ void RuntimeOption::SetModelPath(const std::string& model_path,
     model_file = model_path;
     model_format = ModelFormat::ONNX;
   } else {
-    FDASSERT(false, "The model format only can be ModelFormat::PADDLE/ModelFormat::ONNX.");
+    FDASSERT(
+        false,
+        "The model format only can be ModelFormat::PADDLE/ModelFormat::ONNX.");
   }
 }
 
@@ -228,6 +230,11 @@ void RuntimeOption::SetPaddleMKLDNNCacheSize(int size) {
   pd_mkldnn_cache_size = size;
 }
 
+void RuntimeOption::SetLitePowerMode(int mode) {
+  FDASSERT(mode > -1, "Parameter mode must greater than -1.");
+  lite_power_mode = mode;
+}
+
 void RuntimeOption::SetTrtInputShape(const std::string& input_name,
                                      const std::vector<int32_t>& min_shape,
                                      const std::vector<int32_t>& opt_shape,
@@ -465,6 +472,8 @@ void Runtime::CreateTrtBackend() {
 void Runtime::CreateLiteBackend() {
 #ifdef ENABLE_LITE_BACKEND
   auto lite_option = LiteBackendOption();
+  lite_option.threads = option.cpu_thread_num;
+  lite_option.power_mode = option.lite_power_mode;
   FDASSERT(option.model_format == ModelFormat::PADDLE,
            "LiteBackend only support model format of ModelFormat::PADDLE");
   backend_ = utils::make_unique<LiteBackend>();

fastdeploy/runtime.h

@@ -84,6 +84,9 @@ struct FASTDEPLOY_DECL RuntimeOption {
   // set size of cached shape while enable mkldnn with paddle inference backend
   void SetPaddleMKLDNNCacheSize(int size);
 
+  // set the power mode of paddle lite backend.
+  void SetLitePowerMode(int mode);
+
   // set tensorrt shape while the inputs of model contain dynamic shape
   // min_shape: the minimum shape
   // opt_shape: the most common shape while inference, default be empty
@@ -126,6 +129,12 @@ struct FASTDEPLOY_DECL RuntimeOption {
   int pd_mkldnn_cache_size = 1;
   std::vector<std::string> pd_delete_pass_names;
 
+  // ======Only for Paddle-Lite Backend=====
+  // 0: LITE_POWER_HIGH 1: LITE_POWER_LOW 2: LITE_POWER_FULL
+  // 3: LITE_POWER_NO_BIND 4: LITE_POWER_RAND_HIGH
+  // 5: LITE_POWER_RAND_LOW
+  int lite_power_mode = 0;
+
   // ======Only for Trt Backend=======
   std::map<std::string, std::vector<int32_t>> trt_max_shape;
   std::map<std::string, std::vector<int32_t>> trt_min_shape;

python/fastdeploy/runtime.py

@@ -54,7 +54,9 @@ class RuntimeOption:
     def __init__(self):
         self._option = C.RuntimeOption()
 
-    def set_model_path(self, model_path, params_path="",
+    def set_model_path(self,
+                       model_path,
+                       params_path="",
                        model_format=C.ModelFormat.PADDLE):
         return self._option.set_model_path(model_path, params_path,
                                            model_format)
@@ -98,6 +100,9 @@ class RuntimeOption:
     def set_paddle_mkldnn_cache_size(self, cache_size):
         return self._option.set_paddle_mkldnn_cache_size(cache_size)
 
+    def set_lite_power_mode(self, mode):
+        return self._option.set_lite_power_mode(mode)
+
     def set_trt_input_shape(self,
                             tensor_name,
                             min_shape,