Merge branch 'develop' into matting

fastdeploy/runtime/backends/rknpu2/option.h

@@ -34,4 +34,19 @@ typedef enum _rknpu2_core_mask {
   RKNN_NPU_CORE_UNDEFINED,
 } CoreMask;
 }  // namespace rknpu2
+
+struct RKNPU2BackendOption {
+  rknpu2::CpuName cpu_name = rknpu2::CpuName::RK3588;
+
+  // The specification of NPU core setting.It has the following choices :
+  // RKNN_NPU_CORE_AUTO : Referring to automatic mode, meaning that it will
+  // select the idle core inside the NPU.
+  // RKNN_NPU_CORE_0 : Running on the NPU0 core
+  // RKNN_NPU_CORE_1: Runing on the NPU1 core
+  // RKNN_NPU_CORE_2: Runing on the NPU2 core
+  // RKNN_NPU_CORE_0_1: Running on both NPU0 and NPU1 core simultaneously.
+  // RKNN_NPU_CORE_0_1_2: Running on both NPU0, NPU1 and NPU2 simultaneously.
+  rknpu2::CoreMask core_mask = rknpu2::CoreMask::RKNN_NPU_CORE_AUTO;
+};
+
 }  // namespace fastdeploy

fastdeploy/runtime/backends/rknpu2/rknpu2_backend.cc

@@ -76,17 +76,24 @@ void RKNPU2Backend::BuildOption(const RKNPU2BackendOption& option) {
 }
 
 /***************************************************************
- *  @name       InitFromRKNN
+ *  @name       Init
  *  @brief      Initialize RKNN model
  *  @param      model_file: Binary data for the RKNN model or the path of RKNN
  *model. params_file: None option: config
  *  @return     bool
  *  @note       None
  ***************************************************************/
-bool RKNPU2Backend::InitFromRKNN(const std::string& model_file,
-                                 const RKNPU2BackendOption& option) {
+bool RKNPU2Backend::Init(const RuntimeOption& runtime_option) {
+  if (!(Supported(runtime_option.model_format, Backend::RKNPU2) && Supported(runtime_option.device, Backend::RKNPU2))) {
+    return false;
+  }
+  if (runtime_option.model_from_memory_) {
+    FDERROR << "RKNPU2 backend doesn't support load model from memory, please load model from disk." << std::endl;
+    return false;
+  }
+
   // LoadModel
-  if (!this->LoadModel((char*)model_file.data())) {
+  if (!this->LoadModel((char*)runtime_option.model_file.data())) {
     FDERROR << "load model failed" << std::endl;
     return false;
   }
@@ -98,7 +105,7 @@ bool RKNPU2Backend::InitFromRKNN(const std::string& model_file,
   }
 
   // BuildOption
-  this->BuildOption(option);
+  this->BuildOption(runtime_option.rknpu2_option);
 
   // SetCoreMask if RK3588
   if (this->option_.cpu_name == rknpu2::CpuName::RK3588) {
@@ -124,7 +131,7 @@ bool RKNPU2Backend::InitFromRKNN(const std::string& model_file,
  *  @return     bool
  *  @note       Only support RK3588
  ***************************************************************/
-bool RKNPU2Backend::SetCoreMask(const rknpu2::CoreMask& core_mask) {
+bool RKNPU2Backend::SetCoreMask(const rknpu2::CoreMask& core_mask) const {
   int ret = rknn_set_core_mask(ctx, static_cast<rknn_core_mask>(core_mask));
   if (ret != RKNN_SUCC) {
     FDERROR << "rknn_set_core_mask fail! ret=" << ret << std::endl;

fastdeploy/runtime/backends/rknpu2/rknpu2_backend.h

@@ -24,40 +24,13 @@
 #include <vector>
 
 namespace fastdeploy {
-struct RKNPU2BackendOption {
-  rknpu2::CpuName cpu_name = rknpu2::CpuName::RK356X;
-
-  // The specification of NPU core setting.It has the following choices :
-  // RKNN_NPU_CORE_AUTO : Referring to automatic mode, meaning that it will
-  // select the idle core inside the NPU.
-  // RKNN_NPU_CORE_0 : Running on the NPU0 core
-  // RKNN_NPU_CORE_1: Runing on the NPU1 core
-  // RKNN_NPU_CORE_2: Runing on the NPU2 core
-  // RKNN_NPU_CORE_0_1: Running on both NPU0 and NPU1 core simultaneously.
-  // RKNN_NPU_CORE_0_1_2: Running on both NPU0, NPU1 and NPU2 simultaneously.
-  rknpu2::CoreMask core_mask = rknpu2::CoreMask::RKNN_NPU_CORE_AUTO;
-};
-
 class RKNPU2Backend : public BaseBackend {
  public:
   RKNPU2Backend() = default;
 
   virtual ~RKNPU2Backend();
 
-  // RKNN API
-  bool LoadModel(void* model);
-
-  bool GetSDKAndDeviceVersion();
-
-  bool SetCoreMask(const rknpu2::CoreMask& core_mask);
-
-  bool GetModelInputOutputInfos();
-
-  // BaseBackend API
-  void BuildOption(const RKNPU2BackendOption& option);
-
-  bool InitFromRKNN(const std::string& model_file,
-                    const RKNPU2BackendOption& option = RKNPU2BackendOption());
+  bool Init(const RuntimeOption& runtime_option);
 
   int NumInputs() const override {
     return static_cast<int>(inputs_desc_.size());
@@ -75,6 +48,18 @@ class RKNPU2Backend : public BaseBackend {
              bool copy_to_fd = true) override;
 
  private:
+  // BaseBackend API
+  void BuildOption(const RKNPU2BackendOption& option);
+  
+  // RKNN API
+  bool LoadModel(void* model);
+
+  bool GetSDKAndDeviceVersion();
+
+  bool SetCoreMask(const rknpu2::CoreMask& core_mask) const;
+
+  bool GetModelInputOutputInfos();
+
   // The object of rknn context.
   rknn_context ctx{};
   // The structure rknn_sdk_version is used to indicate the version

fastdeploy/runtime/enum_variables.h

@@ -96,6 +96,37 @@ static std::map<Device, std::vector<Backend>>
   {Device::SOPHGOTPUD, {Backend::SOPHGOTPU}}
 };
 
+inline bool Supported(ModelFormat format, Backend backend) {
+  auto iter = s_default_backends_by_format.find(format);
+  if (iter == s_default_backends_by_format.end()) {
+    FDERROR << "Didn't find format is registered in s_default_backends_by_format." << std::endl;
+    return false;
+  }
+  for (size_t i = 0; i < iter->second.size(); ++i) {
+    if (iter->second[i] == backend) {
+      return true;
+    }
+  }
+  std::string msg = Str(iter->second); 
+  FDERROR << backend << " only supports " << msg << ", but now it's " << format << "." << std::endl;
+  return false;
+}
+
+inline bool Supported(Device device, Backend backend) {
+  auto iter = s_default_backends_by_device.find(device);
+  if (iter == s_default_backends_by_device.end()) {
+    FDERROR << "Didn't find device is registered in s_default_backends_by_device." << std::endl;
+    return false;
+  }
+  for (size_t i = 0; i < iter->second.size(); ++i) {
+    if (iter->second[i] == backend) {
+      return true;
+    }
+  }
+  std::string msg = Str(iter->second); 
+  FDERROR << backend << " only supports " << msg << ", but now it's " << device << "." << std::endl;
+  return false;
+}
 
 FASTDEPLOY_DECL std::ostream& operator<<(std::ostream& o, const Backend& b);
 FASTDEPLOY_DECL std::ostream& operator<<(std::ostream& o, const Device& d);

fastdeploy/runtime/runtime.cc

@@ -169,6 +169,7 @@ bool Runtime::Init(const RuntimeOption& _option) {
             << std::endl;
     return false;
   }
+  backend_->benchmark_option_ = option.benchmark_option;
   return true;
 }
 
@@ -282,7 +283,6 @@ void Runtime::CreatePaddleBackend() {
   option.paddle_infer_option.trt_option.gpu_id = option.device_id;
   backend_ = utils::make_unique<PaddleBackend>();
   auto casted_backend = dynamic_cast<PaddleBackend*>(backend_.get());
-  casted_backend->benchmark_option_ = option.benchmark_option;
 
   if (option.model_from_memory_) {
     FDASSERT(
@@ -313,7 +313,6 @@ void Runtime::CreatePaddleBackend() {
 void Runtime::CreateOpenVINOBackend() {
 #ifdef ENABLE_OPENVINO_BACKEND
   backend_ = utils::make_unique<OpenVINOBackend>();
-  backend_->benchmark_option_ = option.benchmark_option;
   FDASSERT(backend_->Init(option), "Failed to initialize OpenVINOBackend.");
 #else
   FDASSERT(false,
@@ -327,7 +326,6 @@ void Runtime::CreateOpenVINOBackend() {
 void Runtime::CreateOrtBackend() {
 #ifdef ENABLE_ORT_BACKEND
   backend_ = utils::make_unique<OrtBackend>();
-  backend_->benchmark_option_ = option.benchmark_option;
 
   FDASSERT(backend_->Init(option), "Failed to initialize Backend::ORT.");
 #else
@@ -348,7 +346,6 @@ void Runtime::CreateTrtBackend() {
   option.trt_option.enable_pinned_memory = option.enable_pinned_memory;
   option.trt_option.external_stream_ = option.external_stream_;
   backend_ = utils::make_unique<TrtBackend>();
-  backend_->benchmark_option_ = option.benchmark_option;
   FDASSERT(backend_->Init(option), "Failed to initialize TensorRT backend.");
 #else
   FDASSERT(false,
@@ -362,7 +359,6 @@ void Runtime::CreateTrtBackend() {
 void Runtime::CreateLiteBackend() {
 #ifdef ENABLE_LITE_BACKEND
   backend_ = utils::make_unique<LiteBackend>();
-  backend_->benchmark_option_ = option.benchmark_option;
 
   FDASSERT(backend_->Init(option),
            "Load model from nb file failed while initializing LiteBackend.");
@@ -376,20 +372,9 @@ void Runtime::CreateLiteBackend() {
 }
 
 void Runtime::CreateRKNPU2Backend() {
-  FDASSERT(option.model_from_memory_ == false,
-           "RKNPU2Backend don't support to load model from memory");
-  FDASSERT(option.device == Device::RKNPU,
-           "Backend::RKNPU2 only supports Device::RKNPU2");
-  FDASSERT(option.model_format == ModelFormat::RKNN,
-           "RKNPU2Backend only support model format of ModelFormat::RKNN");
 #ifdef ENABLE_RKNPU2_BACKEND
-  auto rknpu2_option = RKNPU2BackendOption();
-  rknpu2_option.cpu_name = option.rknpu2_cpu_name_;
-  rknpu2_option.core_mask = option.rknpu2_core_mask_;
   backend_ = utils::make_unique<RKNPU2Backend>();
-  auto casted_backend = dynamic_cast<RKNPU2Backend*>(backend_.get());
-  FDASSERT(casted_backend->InitFromRKNN(option.model_file, rknpu2_option),
-           "Load model from nb file failed while initializing LiteBackend.");
+  FDASSERT(backend_->Init(option), "Failed to initialize RKNPU2 backend.");
 #else
   FDASSERT(false,
            "RKNPU2Backend is not available, please compiled with "

fastdeploy/runtime/runtime_option.cc

@@ -60,8 +60,8 @@ void RuntimeOption::UseCpu() { device = Device::CPU; }
 
 void RuntimeOption::UseRKNPU2(fastdeploy::rknpu2::CpuName rknpu2_name,
                               fastdeploy::rknpu2::CoreMask rknpu2_core) {
-  rknpu2_cpu_name_ = rknpu2_name;
-  rknpu2_core_mask_ = rknpu2_core;
+  rknpu2_option.cpu_name = rknpu2_name;
+  rknpu2_option.core_mask = rknpu2_core;
   device = Device::RKNPU;
 }
 

fastdeploy/runtime/runtime_option.h

@@ -151,6 +151,8 @@ struct FASTDEPLOY_DECL RuntimeOption {
   OpenVINOBackendOption openvino_option;
   /// Option to configure Paddle Lite backend
   LiteBackendOption paddle_lite_option;
+  /// Option to configure RKNPU2 backend
+  RKNPU2BackendOption rknpu2_option;
 
   /** \brief Set the profile mode as 'true'.
    *
@@ -199,12 +201,6 @@ struct FASTDEPLOY_DECL RuntimeOption {
 
   bool enable_pinned_memory = false;
 
-  // ======Only for RKNPU2 Backend=======
-  fastdeploy::rknpu2::CpuName rknpu2_cpu_name_ =
-      fastdeploy::rknpu2::CpuName::RK3588;
-  fastdeploy::rknpu2::CoreMask rknpu2_core_mask_ =
-      fastdeploy::rknpu2::CoreMask::RKNN_NPU_CORE_AUTO;
-
   // *** The belowing api are deprecated, will be removed in v1.2.0
   // *** Do not use it anymore
   void SetPaddleMKLDNN(bool pd_mkldnn = true);

fastdeploy/utils/utils.h

@@ -207,7 +207,7 @@ template <typename T>
 std::string Str(const std::vector<T>& shape) {
   std::ostringstream oss;
   oss << "[ " << shape[0];
-  for (int i = 1; i < shape.size(); ++i) {
+  for (size_t i = 1; i < shape.size(); ++i) {
     oss << " ," << shape[i];
   }
   oss << " ]";