Merge branch 'develop' of github.com:yunyaoXYY/FastDeploy into huawei

.github/workflows/build.yml

@@ -1,25 +1,7 @@
 name: Build
-on: [push, pull_request]
+on: [pull_request]
 
 jobs:
-  macOS-latest-cpp:
-    runs-on: macOS-latest
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v1
-
-      - name: Get CMake
-        uses: lukka/get-cmake@latest
-
-      - name: Build FastDeploy
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DENABLE_ORT_BACKEND=ON -DENABLE_PADDLE_BACKEND=OFF -DENABLE_OPENVINO_BACKEND=OFF -DCMAKE_INSTALL_PREFIX=${PWD}/compiled_fastdeploy_sdk -DENABLE_VISION=ON -DENABLE_TEXT=ON
-          make -j12
-          make install
-          ls -l
   macOS-latest-py:
     runs-on: macos-latest
 

README_CN.md

@@ -289,7 +289,7 @@ int main(int argc, char* argv[]) {
 | Classification         | [PaddleClas/SqueeezeNetV1.1](./examples/vision/classification/paddleclas)                    | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ✅                           | ✅       |✅       | ✅       |
 | Classification         | [PaddleClas/Inceptionv3](./examples/vision/classification/paddleclas)                        | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ✅       |
 | Classification         | [PaddleClas/PP-HGNet](./examples/vision/classification/paddleclas)                           | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ✅                           | ✅       |✅       | ✅       |
-| Detection              | [PaddleDetection/PP-YOLOE+](./examples/vision/detection/paddledetection)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ✅       |
+| Detection              | 🔥🔥[PaddleDetection/PP-YOLOE+](./examples/vision/detection/paddledetection)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ✅       |
 | Detection              | [🔥PaddleDetection/YOLOv8](./examples/vision/detection/paddledetection)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ❔      |
 | Detection              | [🔥ultralytics/YOLOv8](./examples/vision/detection/yolov8)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ❔               | ✅                           | ❔                           | ❔      |❔       | ❔      |
 | Detection              | [PaddleDetection/PicoDet](./examples/vision/detection/paddledetection)                       | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       | ❔       | ✅       |

README_EN.md

@@ -275,7 +275,7 @@ Notes: ✅: already supported; ❔: to be supported in the future;  N/A: Not Ava
 | Classification         | [PaddleClas/SqueeezeNetV1.1](./examples/vision/classification/paddleclas)                    | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ✅                           | ✅       |✅       | ✅       |
 | Classification         | [PaddleClas/Inceptionv3](./examples/vision/classification/paddleclas)                        | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ✅       |
 | Classification         | [PaddleClas/PP-HGNet](./examples/vision/classification/paddleclas)                           | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ✅                           | ✅       |✅       | ✅       |
-| Detection              | [PaddleDetection/PP-YOLOE+](./examples/vision/detection/paddledetection)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ✅       |
+| Detection              | 🔥🔥[PaddleDetection/PP-YOLOE+](./examples/vision/detection/paddledetection)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ✅       |
 | Detection              | [🔥PaddleDetection/YOLOv8](./examples/vision/detection/paddledetection)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       |✅       | ❔      |
 | Detection              | [🔥ultralytics/YOLOv8](./examples/vision/detection/yolov8)                      | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ❔               | ✅                           | ❔                           | ❔      |❔       | ❔      |
 | Detection              | [PaddleDetection/PicoDet](./examples/vision/detection/paddledetection)                       | ✅                                                | ✅                        | ✅                        | ✅                        | ✅                       | ✅                     | ✅                          | ✅                | ✅                           | ❔                           | ✅       | ❔       | ✅       |

cmake/ascend.cmake

@@ -26,7 +26,7 @@ if(CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "aarch64")
   else ()
     message(STATUS "Build FastDeploy Ascend Python library on aarch64 platform.")
     if(NOT PADDLELITE_URL)
-      set(PADDLELITE_URL "https://bj.bcebos.com/fastdeploy/third_libs/lite-linux_arm64_huawei_ascend_npu_python_0118.tgz")
+      set(PADDLELITE_URL "https://bj.bcebos.com/fastdeploy/third_libs/lite-linux_arm64_huawei_ascend_npu_python_0202.tgz")
     endif()
   endif()	
 endif()

docs/cn/faq/rknpu2/build.md

@@ -47,8 +47,7 @@ cmake ..  -DCMAKE_C_COMPILER=/home/zbc/opt/gcc-linaro-6.3.1-2017.05-x86_64_aarch
           -DENABLE_ORT_BACKEND=OFF \
 	      -DENABLE_RKNPU2_BACKEND=ON \
 	      -DENABLE_VISION=ON \
-	      -DRKNN2_TARGET_SOC=RK3588 \
+	      -DRKNN2_TARGET_SOC=RK356X \
-	      -DENABLE_FLYCV=ON \
           -DCMAKE_INSTALL_PREFIX=${PWD}/fastdeploy-0.0.0
 make -j8
 make install

examples/vision/detection/rkyolo/cpp/CMakeLists.txt

@@ -4,34 +4,9 @@ project(rknpu2_test)
 set(CMAKE_CXX_STANDARD 14)
 
 # 指定下载解压后的fastdeploy库路径
-set(FASTDEPLOY_INSTALL_DIR "thirdpartys/fastdeploy-0.0.3")
+option(FASTDEPLOY_INSTALL_DIR "Path of downloaded fastdeploy sdk.")
-
+include(${FASTDEPLOY_INSTALL_DIR}/FastDeploy.cmake)
-include(${FASTDEPLOY_INSTALL_DIR}/FastDeployConfig.cmake)
+# 添加FastDeploy依赖头文件
-include_directories(${FastDeploy_INCLUDE_DIRS})
+include_directories(${FASTDEPLOY_INCS})
-
 add_executable(infer_rkyolo infer_rkyolo.cc)
-target_link_libraries(infer_rkyolo ${FastDeploy_LIBS})
+target_link_libraries(infer_rkyolo ${FASTDEPLOY_LIBS})
-
-
-
-set(CMAKE_INSTALL_PREFIX ${CMAKE_SOURCE_DIR}/build/install)
-
-install(TARGETS infer_rkyolo DESTINATION ./)
-
-install(DIRECTORY model DESTINATION ./)
-install(DIRECTORY images DESTINATION ./)
-
-file(GLOB FASTDEPLOY_LIBS ${FASTDEPLOY_INSTALL_DIR}/lib/*)
-message("${FASTDEPLOY_LIBS}")
-install(PROGRAMS ${FASTDEPLOY_LIBS} DESTINATION lib)
-
-file(GLOB ONNXRUNTIME_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/onnxruntime/lib/*)
-install(PROGRAMS ${ONNXRUNTIME_LIBS} DESTINATION lib)
-
-install(DIRECTORY ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/opencv/lib DESTINATION ./)
-
-file(GLOB PADDLETOONNX_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/paddle2onnx/lib/*)
-install(PROGRAMS ${PADDLETOONNX_LIBS} DESTINATION lib)
-
-file(GLOB RKNPU2_LIBS ${FASTDEPLOY_INSTALL_DIR}/third_libs/install/rknpu2_runtime/${RKNN2_TARGET_SOC}/lib/*)
-install(PROGRAMS ${RKNPU2_LIBS} DESTINATION lib)

examples/vision/detection/rkyolo/cpp/README_CN.md

@@ -10,58 +10,12 @@
 
 以上步骤请参考[RK2代NPU部署库编译](../../../../../docs/cn/build_and_install/rknpu2.md)实现
 
-## 生成基本目录文件
-
-该例程由以下几个部分组成
-```text
-.
-├── CMakeLists.txt
-├── build  # 编译文件夹
-├── image  # 存放图片的文件夹
-├── infer_rkyolo.cc
-├── model  # 存放模型文件的文件夹
-└── thirdpartys  # 存放sdk的文件夹
-```
-
-首先需要先生成目录结构
-```bash
-mkdir build
-mkdir images
-mkdir model
-mkdir thirdpartys
-```
-
-## 编译
-
-### 编译并拷贝SDK到thirdpartys文件夹
-
-请参考[RK2代NPU部署库编译](../../../../../docs/cn/build_and_install/rknpu2.md)仓库编译SDK，编译完成后，将在build目录下生成
-fastdeploy-0.0.3目录，请移动它至thirdpartys目录下.
-
-### 拷贝模型文件，以及配置文件至model文件夹
-在Paddle动态图模型 -> Paddle静态图模型 -> ONNX模型的过程中，将生成ONNX文件以及对应的yaml配置文件，请将配置文件存放到model文件夹内。
-转换为RKNN后的模型文件也需要拷贝至model。
-
-### 准备测试图片至image文件夹
 ```bash
 wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg
-cp 000000014439.jpg ./images
-```
-
-### 编译example
-
-```bash
 cd build
-cmake ..
+cmake .. -DFASTDEPLOY_INSTALL_DIR=${PWD}/fastdeploy-linux-x64-x.x.x
 make -j8
-make install
+./infer_rkyolo /path/to/model 000000014439.jpg
-```
-
-## 运行例程
-
-```bash
-cd ./build/install
-./infer_picodet model/ images/000000014439.jpg
 ```
 
 

fastdeploy/runtime/backends/lite/configure_hardware.cc

@@ -0,0 +1,159 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/runtime/backends/lite/lite_backend.h"
+// https://github.com/PaddlePaddle/Paddle-Lite/issues/8290
+// When compiling the FastDeploy dynamic library, namely,
+// WITH_STATIC_LIB=OFF, and depending on the Paddle Lite
+// static library, you need to include the fake registration
+// codes of Paddle Lite. When you compile the FastDeploy static
+// library and depends on the Paddle Lite static library,
+// WITH_STATIC_LIB=ON, you do not need to include the fake
+// registration codes for Paddle Lite, but wait until you
+// use the FastDeploy static library.
+#if (defined(WITH_LITE_STATIC) && (!defined(WITH_STATIC_LIB)))
+#warning You are compiling the FastDeploy dynamic library with \
+Paddle Lite static lib We will automatically add some registration \
+codes for ops, kernels and passes for Paddle Lite.
+#include "paddle_use_kernels.h"  // NOLINT
+#include "paddle_use_ops.h"      // NOLINT
+#include "paddle_use_passes.h"   // NOLINT
+#endif
+
+#include <cstring>
+
+namespace fastdeploy {
+
+#if defined(__arm__) || defined(__aarch64__)
+#define FD_LITE_HOST TARGET(kARM)
+#elif defined(__x86_64__)
+#define FD_LITE_HOST TARGET(kX86)
+#endif
+
+std::vector<paddle::lite_api::Place> GetPlacesForCpu(
+    const LiteBackendOption& option) {
+  std::vector<paddle::lite_api::Place> valid_places;
+  valid_places.push_back(
+      paddle::lite_api::Place{FD_LITE_HOST, PRECISION(kInt8)});
+  if (option.enable_fp16) {
+    paddle::lite_api::MobileConfig check_fp16_config;
+    if (check_fp16_config.check_fp16_valid()) {
+      valid_places.push_back(
+          paddle::lite_api::Place{FD_LITE_HOST, PRECISION(kFP16)});
+    } else {
+      FDWARNING << "Current CPU doesn't support float16 precision, will "
+                   "fallback to float32."
+                << std::endl;
+    }
+  }
+  valid_places.push_back(
+      paddle::lite_api::Place{FD_LITE_HOST, PRECISION(kFloat)});
+  return valid_places;
+}
+
+void LiteBackend::ConfigureCpu(const LiteBackendOption& option) {
+  config_.set_valid_places(GetPlacesForCpu(option));
+}
+
+void LiteBackend::ConfigureKunlunXin(const LiteBackendOption& option) {
+  std::vector<paddle::lite_api::Place> valid_places;
+  valid_places.push_back(
+      paddle::lite_api::Place{TARGET(kXPU), PRECISION(kInt8)});
+  if (option.enable_fp16) {
+    valid_places.push_back(
+        paddle::lite_api::Place{TARGET(kXPU), PRECISION(kFP16)});
+  }
+  valid_places.push_back(
+      paddle::lite_api::Place{TARGET(kXPU), PRECISION(kFloat)});
+
+  config_.set_xpu_dev_per_thread(option.device_id);
+  config_.set_xpu_workspace_l3_size_per_thread(
+      option.kunlunxin_l3_workspace_size);
+  config_.set_xpu_l3_cache_method(option.kunlunxin_l3_workspace_size,
+                                  option.kunlunxin_locked);
+  config_.set_xpu_conv_autotune(option.kunlunxin_autotune,
+                                option.kunlunxin_autotune_file);
+  config_.set_xpu_multi_encoder_method(option.kunlunxin_precision,
+                                       option.kunlunxin_adaptive_seqlen);
+  if (option.kunlunxin_enable_multi_stream) {
+    config_.enable_xpu_multi_stream();
+  }
+  auto cpu_places = GetPlacesForCpu(option);
+  valid_places.insert(valid_places.end(), cpu_places.begin(), cpu_places.end());
+  config_.set_valid_places(valid_places);
+}
+
+void LiteBackend::ConfigureTimvx(const LiteBackendOption& option) {
+  config_.set_nnadapter_device_names({"verisilicon_timvx"});
+  std::vector<paddle::lite_api::Place> valid_places;
+  valid_places.push_back(
+      paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)});
+  valid_places.push_back(
+      paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kFloat)});
+  auto cpu_places = GetPlacesForCpu(option);
+  valid_places.insert(valid_places.end(), cpu_places.begin(), cpu_places.end());
+  config_.set_valid_places(valid_places);
+  ConfigureNNAdapter(option);
+}
+
+void LiteBackend::ConfigureAscend(const LiteBackendOption& option) {
+  config_.set_nnadapter_device_names({"huawei_ascend_npu"});
+  std::vector<paddle::lite_api::Place> valid_places;
+  valid_places.push_back(
+      paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)});
+  valid_places.push_back(
+      paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kFloat)});
+  auto cpu_places = GetPlacesForCpu(option);
+  valid_places.insert(valid_places.end(), cpu_places.begin(), cpu_places.end());
+  config_.set_valid_places(valid_places);
+  ConfigureNNAdapter(option);
+}
+
+void LiteBackend::ConfigureNNAdapter(const LiteBackendOption& option) {
+  if (!option.nnadapter_subgraph_partition_config_path.empty()) {
+    std::vector<char> nnadapter_subgraph_partition_config_buffer;
+    if (ReadFile(option.nnadapter_subgraph_partition_config_path,
+                 &nnadapter_subgraph_partition_config_buffer, false)) {
+      if (!nnadapter_subgraph_partition_config_buffer.empty()) {
+        std::string nnadapter_subgraph_partition_config_string(
+            nnadapter_subgraph_partition_config_buffer.data(),
+            nnadapter_subgraph_partition_config_buffer.size());
+        config_.set_nnadapter_subgraph_partition_config_buffer(
+            nnadapter_subgraph_partition_config_string);
+      }
+    }
+  }
+
+  if (!option.nnadapter_context_properties.empty()) {
+    config_.set_nnadapter_context_properties(
+        option.nnadapter_context_properties);
+  }
+
+  if (!option.nnadapter_model_cache_dir.empty()) {
+    config_.set_nnadapter_model_cache_dir(option.nnadapter_model_cache_dir);
+  }
+
+  if (!option.nnadapter_mixed_precision_quantization_config_path.empty()) {
+    config_.set_nnadapter_mixed_precision_quantization_config_path(
+        option.nnadapter_mixed_precision_quantization_config_path);
+  }
+
+  if (!option.nnadapter_subgraph_partition_config_path.empty()) {
+    config_.set_nnadapter_subgraph_partition_config_path(
+        option.nnadapter_subgraph_partition_config_path);
+  }
+
+  config_.set_nnadapter_dynamic_shape_info(option.nnadapter_dynamic_shape_info);
+}
+}  // namespace fastdeploy

fastdeploy/runtime/backends/lite/lite_backend.cc

@@ -26,8 +26,8 @@
 #warning You are compiling the FastDeploy dynamic library with \
 Paddle Lite static lib We will automatically add some registration \
 codes for ops, kernels and passes for Paddle Lite.
-#include "paddle_use_ops.h"       // NOLINT
 #include "paddle_use_kernels.h"  // NOLINT
+#include "paddle_use_ops.h"      // NOLINT
 #include "paddle_use_passes.h"   // NOLINT
 #endif
 
@@ -35,143 +35,20 @@ codes for ops, kernels and passes for Paddle Lite.
 
 namespace fastdeploy {
 
-// Convert data type from paddle lite to fastdeploy
-FDDataType LiteDataTypeToFD(const paddle::lite_api::PrecisionType& dtype) {
-  if (dtype == paddle::lite_api::PrecisionType::kFloat) {
-    return FDDataType::FP32;
-  } else if (dtype == paddle::lite_api::PrecisionType::kInt8) {
-    return FDDataType::INT8;
-  } else if (dtype == paddle::lite_api::PrecisionType::kInt32) {
-    return FDDataType::INT32;
-  } else if (dtype == paddle::lite_api::PrecisionType::kInt64) {
-    return FDDataType::INT64;
-  } else if (dtype == paddle::lite_api::PrecisionType::kInt16) {
-    return FDDataType::INT16;
-  } else if (dtype == paddle::lite_api::PrecisionType::kUInt8) {
-    return FDDataType::UINT8;
-  } else if (dtype == paddle::lite_api::PrecisionType::kFP64) {
-    return FDDataType::FP64;
-  }
-  FDASSERT(false, "Unexpected data type of %d.", dtype);
-  return FDDataType::FP32;
-}
-
 void LiteBackend::BuildOption(const LiteBackendOption& option) {
   option_ = option;
-  std::vector<paddle::lite_api::Place> valid_places;
-  if (option_.enable_int8) {
-    if (option_.enable_kunlunxin) {
-      valid_places.push_back(
-          paddle::lite_api::Place{TARGET(kXPU), PRECISION(kInt8)});
-    } else {
-      valid_places.push_back(
-          paddle::lite_api::Place{TARGET(kARM), PRECISION(kInt8)});
-    }
-    FDINFO << "Lite::Backend enable_int8 option is ON ! Lite::Backend will "
-           << "inference with int8 precision!" << std::endl;
-  }
-  if (option_.enable_fp16) {
-    if (option_.enable_kunlunxin) {
-      valid_places.push_back(
-          paddle::lite_api::Place{TARGET(kXPU), PRECISION(kFP16)});
-    } else {
-      paddle::lite_api::MobileConfig check_fp16_config;
-      // Determine whether the device supports the FP16
-      // instruction set (or whether it is an arm device
-      // of the armv8.2 architecture)
-      supported_fp16_ = check_fp16_config.check_fp16_valid();
-      if (supported_fp16_) {
-        valid_places.push_back(
-            paddle::lite_api::Place{TARGET(kARM), PRECISION(kFP16)});
-        FDINFO << "The device supports FP16, Lite::Backend will inference with "
-                  "FP16 precision."
-               << std::endl;
-      } else {
-        FDWARNING << "The device doesn't support FP16, will fallback to FP32.";
-      }
-    }
-  }
-  if (!option_.nnadapter_subgraph_partition_config_path.empty()) {
-    std::vector<char> nnadapter_subgraph_partition_config_buffer;
-    if (ReadFile(option_.nnadapter_subgraph_partition_config_path,
-                 &nnadapter_subgraph_partition_config_buffer, false)) {
-      if (!nnadapter_subgraph_partition_config_buffer.empty()) {
-        std::string nnadapter_subgraph_partition_config_string(
-            nnadapter_subgraph_partition_config_buffer.data(),
-            nnadapter_subgraph_partition_config_buffer.size());
-        config_.set_nnadapter_subgraph_partition_config_buffer(
-            nnadapter_subgraph_partition_config_string);
-      }
-    }
-  }
-  if (option_.enable_timvx) {
-    config_.set_nnadapter_device_names({"verisilicon_timvx"});
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)});
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kFloat)});
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kARM), PRECISION(kInt8)});
-  }
 
-  if (option_.enable_ascend) {
+  if (option_.device == Device::CPU) {
-    if (option_.nnadapter_device_names.empty()) {
+    ConfigureCpu(option_);
-      config_.set_nnadapter_device_names({"huawei_ascend_npu"});
+  } else if (option_.device == Device::TIMVX) {
-    } else {
+    ConfigureTimvx(option_);
-      config_.set_nnadapter_device_names(option_.nnadapter_device_names);
+  } else if (option_.device == Device::KUNLUNXIN) {
+    ConfigureKunlunXin(option_);
+  } else if (option_.device == Device::ASCEND) {
+    ConfigureAscend(option_);
   }
-
+  if (option_.cpu_threads > 0) {
-    if (!option_.nnadapter_context_properties.empty()) {
+    config_.set_threads(option_.cpu_threads);
-      config_.set_nnadapter_context_properties(
-          option_.nnadapter_context_properties);
-    }
-
-    if (!option_.nnadapter_model_cache_dir.empty()) {
-      config_.set_nnadapter_model_cache_dir(option_.nnadapter_model_cache_dir);
-    }
-
-    if (!option_.nnadapter_mixed_precision_quantization_config_path.empty()) {
-      config_.set_nnadapter_mixed_precision_quantization_config_path(
-          option_.nnadapter_mixed_precision_quantization_config_path);
-    }
-
-    if (!option_.nnadapter_subgraph_partition_config_path.empty()) {
-      config_.set_nnadapter_subgraph_partition_config_path(
-          option_.nnadapter_subgraph_partition_config_path);
-    }
-
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kInt8)});
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kNNAdapter), PRECISION(kFloat)});
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kARM), PRECISION(kInt8)});
-  }
-
-  if (option_.enable_kunlunxin) {
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kXPU), PRECISION(kFloat)});
-    valid_places.push_back(
-        paddle::lite_api::Place{TARGET(kX86), PRECISION(kFloat)});
-    config_.set_xpu_dev_per_thread(option_.device_id);
-    config_.set_xpu_workspace_l3_size_per_thread(
-        option_.kunlunxin_l3_workspace_size);
-    config_.set_xpu_l3_cache_method(option_.kunlunxin_l3_workspace_size,
-                                    option_.kunlunxin_locked);
-    config_.set_xpu_conv_autotune(option_.kunlunxin_autotune,
-                                  option_.kunlunxin_autotune_file);
-    config_.set_xpu_multi_encoder_method(option_.kunlunxin_precision,
-                                         option_.kunlunxin_adaptive_seqlen);
-    if (option_.kunlunxin_enable_multi_stream) {
-      config_.enable_xpu_multi_stream();
-    }
-  } else {
-    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(
@@ -179,29 +56,6 @@ void LiteBackend::BuildOption(const LiteBackendOption& option) {
   }
 }
 
-bool LiteBackend::ReadFile(const std::string& filename,
-                           std::vector<char>* contents, const bool binary) {
-  FILE* fp = fopen(filename.c_str(), binary ? "rb" : "r");
-  if (!fp) {
-    FDERROR << "Cannot open file " << filename << "." << std::endl;
-    return false;
-  }
-  fseek(fp, 0, SEEK_END);
-  size_t size = ftell(fp);
-  fseek(fp, 0, SEEK_SET);
-  contents->clear();
-  contents->resize(size);
-  size_t offset = 0;
-  char* ptr = reinterpret_cast<char*>(&(contents->at(0)));
-  while (offset < size) {
-    size_t already_read = fread(ptr, 1, size - offset, fp);
-    offset += already_read;
-    ptr += already_read;
-  }
-  fclose(fp);
-  return true;
-}
-
 bool LiteBackend::InitFromPaddle(const std::string& model_file,
                                  const std::string& params_file,
                                  const LiteBackendOption& option) {
@@ -246,7 +100,7 @@ bool LiteBackend::InitFromPaddle(const std::string& model_file,
     auto shape = tensor->shape();
     info.shape.assign(shape.begin(), shape.end());
     info.name = output_names[i];
-    if (!option_.enable_kunlunxin) {
+    if (!option_.device == Device::KUNLUNXIN) {
       info.dtype = LiteDataTypeToFD(tensor->precision());
     }
     outputs_desc_.emplace_back(info);
@@ -337,4 +191,49 @@ bool LiteBackend::Infer(std::vector<FDTensor>& inputs,
   return true;
 }
 
+bool ReadFile(const std::string& filename, std::vector<char>* contents,
+              bool binary) {
+  FILE* fp = fopen(filename.c_str(), binary ? "rb" : "r");
+  if (!fp) {
+    FDERROR << "Cannot open file " << filename << "." << std::endl;
+    return false;
+  }
+  fseek(fp, 0, SEEK_END);
+  size_t size = ftell(fp);
+  fseek(fp, 0, SEEK_SET);
+  contents->clear();
+  contents->resize(size);
+  size_t offset = 0;
+  char* ptr = reinterpret_cast<char*>(&(contents->at(0)));
+  while (offset < size) {
+    size_t already_read = fread(ptr, 1, size - offset, fp);
+    offset += already_read;
+    ptr += already_read;
+  }
+  fclose(fp);
+  return true;
+}
+
+// Convert data type from paddle lite to fastdeploy
+FDDataType LiteDataTypeToFD(const paddle::lite_api::PrecisionType& dtype) {
+  if (dtype == paddle::lite_api::PrecisionType::kFloat) {
+    return FDDataType::FP32;
+  } else if (dtype == paddle::lite_api::PrecisionType::kInt8) {
+    return FDDataType::INT8;
+  } else if (dtype == paddle::lite_api::PrecisionType::kInt32) {
+    return FDDataType::INT32;
+  } else if (dtype == paddle::lite_api::PrecisionType::kInt64) {
+    return FDDataType::INT64;
+  } else if (dtype == paddle::lite_api::PrecisionType::kInt16) {
+    return FDDataType::INT16;
+  } else if (dtype == paddle::lite_api::PrecisionType::kUInt8) {
+    return FDDataType::UINT8;
+  } else if (dtype == paddle::lite_api::PrecisionType::kFP64) {
+    return FDDataType::FP64;
+  }
+  FDASSERT(false, "Unexpected data type of %s.",
+           paddle::lite_api::PrecisionToStr(dtype).c_str());
+  return FDDataType::FP32;
+}
+
 }  // namespace fastdeploy

fastdeploy/runtime/backends/lite/lite_backend.h

@@ -19,13 +19,12 @@
 #include <string>
 #include <vector>
 
-#include "fastdeploy/runtime/backends/backend.h"
-#include "fastdeploy/runtime/backends/lite/option.h"
 #include "paddle_api.h"  // NOLINT
 
+#include "fastdeploy/runtime/backends/backend.h"
+#include "fastdeploy/runtime/backends/lite/option.h"
+
 namespace fastdeploy {
-// Convert data type from paddle lite to fastdeploy
-FDDataType LiteDataTypeToFD(const paddle::lite_api::PrecisionType& dtype);
 
 class LiteBackend : public BaseBackend {
  public:
@@ -51,15 +50,26 @@ class LiteBackend : public BaseBackend {
   std::vector<TensorInfo> GetOutputInfos() override;
 
  private:
+  void ConfigureCpu(const LiteBackendOption& option);
+  void ConfigureTimvx(const LiteBackendOption& option);
+  void ConfigureAscend(const LiteBackendOption& option);
+  void ConfigureKunlunXin(const LiteBackendOption& option);
+  void ConfigureNNAdapter(const LiteBackendOption& option);
+
   paddle::lite_api::CxxConfig config_;
   std::shared_ptr<paddle::lite_api::PaddlePredictor> predictor_;
   std::vector<TensorInfo> inputs_desc_;
   std::vector<TensorInfo> outputs_desc_;
   std::map<std::string, int> inputs_order_;
   LiteBackendOption option_;
-  bool supported_fp16_ = false;
+};
+
+// Convert data type from paddle lite to fastdeploy
+FDDataType LiteDataTypeToFD(const paddle::lite_api::PrecisionType& dtype);
+
+// Helper function to read file
 bool ReadFile(const std::string& filename,
              std::vector<char>* contents,
-               const bool binary = true);
+             bool binary = true);
-};
+
 }  // namespace fastdeploy

fastdeploy/runtime/backends/lite/option.h

@@ -51,24 +51,19 @@ enum LitePowerMode {
 };
 
 struct LiteBackendOption {
-  // cpu num threads
+  /// Paddle Lite power mode for mobile device.
-  int threads = 1;
+  LitePowerMode power_mode = LITE_POWER_NO_BIND;
-  // lite power mode
+  /// Number of threads while use CPU
-  // 0: LITE_POWER_HIGH
+  int cpu_threads = 1;
-  // 1: LITE_POWER_LOW
+  /// Enable use half precision
-  // 2: LITE_POWER_FULL
-  // 3: LITE_POWER_NO_BIND
-  // 4: LITE_POWER_RAND_HIGH
-  // 5: LITE_POWER_RAND_LOW
-  int power_mode = 3;
-  // enable fp16
   bool enable_fp16 = false;
-  // enable int8
+  /// Enable use int8 precision for quantized model
   bool enable_int8 = false;
+
+  Device device = Device::CPU;
+
   // optimized model dir for CxxConfig
   std::string optimized_model_dir = "";
-  // TODO(qiuyanjun): support more options for lite backend.
-  // Such as fp16, different device target (kARM/kXPU/kNPU/...)
   std::string nnadapter_subgraph_partition_config_path = "";
   std::string nnadapter_subgraph_partition_config_buffer = "";
   std::string nnadapter_context_properties = "";
@@ -77,9 +72,6 @@ struct LiteBackendOption {
   std::map<std::string, std::vector<std::vector<int64_t>>>
     nnadapter_dynamic_shape_info = {{"", {{0}}}};
   std::vector<std::string> nnadapter_device_names = {};
-  bool enable_timvx = false;
-  bool enable_ascend = false;
-  bool enable_kunlunxin = false;
   int device_id = 0;
   int kunlunxin_l3_workspace_size = 0xfffc00;
   bool kunlunxin_locked = false;

fastdeploy/runtime/backends/paddle/paddle_backend.cc

@@ -92,23 +92,19 @@ void PaddleBackend::BuildOption(const PaddleBackendOption& option) {
 bool PaddleBackend::InitFromPaddle(const std::string& model_buffer,
                                    const std::string& params_buffer,
                                    const PaddleBackendOption& option) {
-  // bool PaddleBackend::InitFromPaddle(const std::string& contents) {
   if (initialized_) {
     FDERROR << "PaddleBackend is already initlized, cannot initialize again."
             << std::endl;
     return false;
   }
+  config_.SetModelBuffer(model_buffer.c_str(), model_buffer.size(),
+                         params_buffer.c_str(), params_buffer.size());
+  config_.EnableMemoryOptim();
+  BuildOption(option);
   
   // The input/output information get from predictor is not right, use
   // PaddleReader instead now
-  std::string contents;
+  auto reader = paddle2onnx::PaddleReader(model_buffer.c_str(), model_buffer.size());
-
-  config_.SetModelBuffer(model_buffer.c_str(), model_buffer.size(),
-                         params_buffer.c_str(), params_buffer.size());
-  contents = model_buffer;
-  config_.EnableMemoryOptim();
-  BuildOption(option);
-  auto reader = paddle2onnx::PaddleReader(contents.c_str(), contents.size());
   // If it's a quantized model, and use cpu with mkldnn, automaticaly switch to
   // int8 mode
   if (reader.is_quantize_model) {

fastdeploy/runtime/runtime_option.cc

@@ -58,7 +58,7 @@ void RuntimeOption::UseRKNPU2(fastdeploy::rknpu2::CpuName rknpu2_name,
 
 void RuntimeOption::UseTimVX() {
   device = Device::TIMVX;
-  paddle_lite_option.enable_timvx = true;
+  paddle_lite_option.device = device;
 }
 
 void RuntimeOption::UseKunlunXin(int kunlunxin_id, int l3_workspace_size,
@@ -68,7 +68,7 @@ void RuntimeOption::UseKunlunXin(int kunlunxin_id, int l3_workspace_size,
                                  bool adaptive_seqlen,
                                  bool enable_multi_stream) {
   device = Device::KUNLUNXIN;
-  paddle_lite_option.enable_kunlunxin = true;
+  paddle_lite_option.device = device;
   paddle_lite_option.device_id = kunlunxin_id;
   paddle_lite_option.kunlunxin_l3_workspace_size = l3_workspace_size;
   paddle_lite_option.kunlunxin_locked = locked;
@@ -81,7 +81,7 @@ void RuntimeOption::UseKunlunXin(int kunlunxin_id, int l3_workspace_size,
 
 void RuntimeOption::UseAscend() {
   device = Device::ASCEND;
-  paddle_lite_option.enable_ascend = true;
+  paddle_lite_option.device = device;
 }
 
 void RuntimeOption::UseSophgo() {
@@ -96,7 +96,7 @@ void RuntimeOption::SetExternalStream(void* external_stream) {
 void RuntimeOption::SetCpuThreadNum(int thread_num) {
   FDASSERT(thread_num > 0, "The thread_num must be greater than 0.");
   cpu_thread_num = thread_num;
-  paddle_lite_option.threads = thread_num;
+  paddle_lite_option.cpu_threads = thread_num;
   ort_option.intra_op_num_threads = thread_num;
 }
 

fastdeploy/vision/detection/contrib/rknpu2/postprocessor.cc

@@ -141,7 +141,6 @@ int RKYOLOPostprocessor::ProcessFP16(float* input, int* anchor, int grid_h,
           } else {
             limit_score = box_conf_f32 * class_prob_f32;
           }
-          // printf("limit score: %f", limit_score);
           if (limit_score > conf_threshold_) {
             float box_x, box_y, box_w, box_h;
             if (anchor_per_branch_ == 1) {

fastdeploy/vision/detection/contrib/rknpu2/postprocessor.h

@@ -55,28 +55,32 @@ class FASTDEPLOY_DECL RKYOLOPostprocessor {
   /// Get nms_threshold, default 0.45
   float GetNMSThreshold() const { return nms_threshold_; }
 
-  // Set height and weight
+  /// Set height and weight
   void SetHeightAndWeight(int& height, int& width) {
     height_ = height;
     width_ = width;
   }
 
-  // Set pad_hw_values
+  /// Set pad_hw_values
   void SetPadHWValues(std::vector<std::vector<int>> pad_hw_values) {
     pad_hw_values_ = pad_hw_values;
   }
 
-  // Set scale
+  /// Set scale
-  void SetScale(std::vector<float> scale) {
+  void SetScale(std::vector<float> scale) { scale_ = scale; }
-    scale_ = scale;
-  }
 
-  // Set Anchor
+  /// Set Anchor
   void SetAnchor(std::vector<int> anchors, int anchor_per_branch) {
     anchors_ = anchors;
     anchor_per_branch_ = anchor_per_branch;
   }
 
+  /// Set the number of class
+  void SetClassNum(int num) {
+    obj_class_num_ = num;
+    prob_box_size_ = obj_class_num_ + 5;
+  }
+
  private:
   std::vector<int> anchors_ = {10, 13, 16,  30,  33, 23,  30,  61,  62,
                                45, 59, 119, 116, 90, 156, 198, 373, 326};
@@ -85,12 +89,9 @@ class FASTDEPLOY_DECL RKYOLOPostprocessor {
   int width_ = 0;
   int anchor_per_branch_ = 0;
 
-  int ProcessFP16(float *input, int *anchor, int grid_h,
+  int ProcessFP16(float* input, int* anchor, int grid_h, int grid_w, int stride,
-              int grid_w, int stride,
+                  std::vector<float>& boxes, std::vector<float>& boxScores,
-              std::vector<float> &boxes,
+                  std::vector<int>& classId, float threshold);
-              std::vector<float> &boxScores,
-              std::vector<int> &classId,
-              float threshold);
   // Model
   int QuickSortIndiceInverse(std::vector<float>& input, int left, int right,
                              std::vector<int>& indices);