Ncnn demo sdk (#164)

* fix different dims and data sync and fp16 macro * fix pad is negative bug, but still need debugging info * split detection_output to dets and labels * fix clang++ compile bug * for fp16 compile macro of cmake * fix pad_val dict input * fix yolox missing normalization and get int tensor * about score_threshold * remove -lstdc++fs for android * move -pthread for android * fix link libraries for CLANG++; * fix clang-format * for mobileyolov3 conf_thre * fix lint * fix bug * support ncnn vulkan net * remove unused debugging info; * Change INFO to DEBUG * support vulkan precision mapping * fix pad * optimize getdetslabels * remove EVAL_MODE * ncnn_net support 4-dim data. * remove FP16 compile options * remove use_gpu_vulkan compile options * fix pad.cpp * fix yapf * fix clang-format * rm redundant lines * fix pad according to mmdet * add android build docs * fix lint * use cpp style string comparision * fix use after free bug * Add missing -DBUILD_SHARED_LIBS=OFF for en * Add missing -DBUILD_SHARED_LIBS=OFF for ZH-CN * reset img2tensor_impl * reset img2tensor_impl * add blank line * sync android.md docs * fix some dirs * fix docs * update docs * fix code
2022-03-25 23:28:16 +08:00 · 2022-03-25 23:28:16 +08:00 · 6948e3d04b
parent e01056cc0a
commit 6948e3d04b
11 changed files with 540 additions and 32 deletions
--- a/configs/_base_/backends/ncnn.py
+++ b/configs/_base_/backends/ncnn.py
@ -1 +1 @@
-backend_config = dict(type='ncnn')
+backend_config = dict(type='ncnn', precision='FP32', use_vulkan=False)
--- a/csrc/codebase/mmdet/object_detection.cpp
+++ b/csrc/codebase/mmdet/object_detection.cpp
@ -12,19 +12,66 @@ namespace mmdeploy::mmdet {

 ResizeBBox::ResizeBBox(const Value& cfg) : MMDetection(cfg) {
  if (cfg.contains("params")) {
-    score_thr_ = cfg["params"].value("score_thr", 0.f);
+    if (cfg["params"].contains("conf_thr")) {
+      // for mobilev2yolov3
+      score_thr_ = cfg["params"].value("conf_thr", 0.f);
+    } else {
+      score_thr_ = cfg["params"].value("score_thr", 0.f);
+    }
    min_bbox_size_ = cfg["params"].value("min_bbox_size", 0.f);
  }
 }
+std::vector<Tensor> ResizeBBox::GetDetsLabels(const Value& prep_res, const Value& infer_res) {
+  std::vector<Tensor> results;
+  if (infer_res.contains("dets") && infer_res.contains("labels")) {
+    results.push_back(infer_res["dets"].get<Tensor>());
+    results.push_back(infer_res["labels"].get<Tensor>());
+    return results;
+  } else if (infer_res.contains("detection_output") && (!infer_res.contains("dets")) &&
+             (!infer_res.contains("labels"))) {
+    int img_width = prep_res["img_metas"]["img_shape"][2].get<int>();
+    int img_height = prep_res["img_metas"]["img_shape"][1].get<int>();
+    auto detection_output = infer_res["detection_output"].get<Tensor>();
+    auto* detection_output_ptr = detection_output.data<float>();
+    // detection_output: (1, num_det, 6)
+    TensorDesc labeldesc = detection_output.desc();
+    int batch_size = detection_output.shape()[0];
+    int num_det = detection_output.shape()[1];
+    labeldesc.shape = {batch_size, num_det};
+    Tensor labels(labeldesc);
+    TensorDesc detdesc = detection_output.desc();
+    detdesc.shape = {batch_size, num_det, 5};
+    Tensor dets(detdesc);
+    auto* dets_ptr = dets.data<float>();
+    auto* labels_ptr = labels.data<float>();
+
+    for (int i = 0; i < batch_size * num_det; ++i) {
+      *labels_ptr++ = detection_output_ptr[0] - 1;
+      dets_ptr[4] = detection_output_ptr[1];
+      dets_ptr[0] = detection_output_ptr[2] * img_width;
+      dets_ptr[1] = detection_output_ptr[3] * img_height;
+      dets_ptr[2] = detection_output_ptr[4] * img_width;
+      dets_ptr[3] = detection_output_ptr[5] * img_height;
+      dets_ptr += 5;
+      detection_output_ptr += 6;
+    }
+    results.push_back(dets);
+    results.push_back(labels);
+    return results;
+  } else {
+    MMDEPLOY_ERROR("No support for another key of detection results!");
+    return results;
+  }
+}
 Result<Value> ResizeBBox::operator()(const Value& prep_res, const Value& infer_res) {
  MMDEPLOY_DEBUG("prep_res: {}\ninfer_res: {}", prep_res, infer_res);
  try {
-    auto dets = infer_res["dets"].get<Tensor>();
-    auto labels = infer_res["labels"].get<Tensor>();
-
+    Tensor dets, labels;
+    vector<Tensor> outputs = GetDetsLabels(prep_res, infer_res);
+    dets = outputs[0];
+    labels = outputs[1];
    MMDEPLOY_DEBUG("dets.shape: {}", dets.shape());
    MMDEPLOY_DEBUG("labels.shape: {}", labels.shape());
-
    // `dets` is supposed to have 3 dims. They are 'batch', 'bboxes_number'
    // and 'channels' respectively
    if (!(dets.shape().size() == 3 && dets.data_type() == DataType::kFLOAT)) {
@ -72,7 +119,6 @@ Result<DetectorOutput> ResizeBBox::GetBBoxes(const Value& prep_res, const Tensor
  DetectorOutput objs;
  auto* dets_ptr = dets.data<float>();
  auto* labels_ptr = labels.data<T>();
-
  vector<float> scale_factor;
  if (prep_res.contains("scale_factor")) {
    from_value(prep_res["scale_factor"], scale_factor);
--- a/csrc/codebase/mmdet/object_detection.h
+++ b/csrc/codebase/mmdet/object_detection.h
@ -26,6 +26,8 @@ class ResizeBBox : public MMDetection {
                                        const float* scale_factor, float x_offset, float y_offset,
                                        int ori_width, int ori_height);

+  std::vector<Tensor> GetDetsLabels(const Value& prep_res, const Value& infer_res);
+
 protected:
  constexpr static Device kHost{0, 0};
  float score_thr_{0.f};
--- a/csrc/core/CMakeLists.txt
+++ b/csrc/core/CMakeLists.txt
@ -48,7 +48,7 @@ target_include_directories(${PROJECT_NAME}
        $<INSTALL_INTERFACE:include/cpp>
        )
 target_link_libraries(${PROJECT_NAME} PUBLIC ${SPDLOG_LIB})
-if (NOT MSVC)
+if (NOT (MSVC OR ANDROID))
    target_link_libraries(${PROJECT_NAME} PUBLIC stdc++fs)
 endif ()
 add_library(mmdeploy::core ALIAS ${PROJECT_NAME})
--- a/csrc/core/utils/stacktrace.cpp
+++ b/csrc/core/utils/stacktrace.cpp
@ -36,7 +36,7 @@ std::string Stacktrace::to_string() const {
 }  // namespace mmdeploy

 #else
-
+#include <string>
 namespace mmdeploy {

 struct Stacktrace::Impl {};
--- a/csrc/model/CMakeLists.txt
+++ b/csrc/model/CMakeLists.txt
@ -12,7 +12,7 @@ endif ()
 foreach (MODEL_NAME ${MODEL_NAMES})
    set(TARGET_MODEL_NAME mmdeploy_${MODEL_NAME})
    mmdeploy_add_module(${TARGET_MODEL_NAME} ${MODEL_NAME}_impl.cpp)
-    if (NOT MSVC)
+    if (NOT (MSVC OR ANDROID))
        target_link_libraries(${TARGET_MODEL_NAME} PUBLIC stdc++fs)
    endif ()
    if (${MODEL_NAME} STREQUAL "zip_model")
--- a/csrc/net/ncnn/ncnn_net.cpp
+++ b/csrc/net/ncnn/ncnn_net.cpp
@ -22,7 +22,9 @@ Result<void> NCNNNet::Init(const Value& args) {
  auto& context = args["context"];
  device_ = context["device"].get<Device>();
  stream_ = context["stream"].get<Stream>();
-
+  if (args.contains("use_vulkan")) {
+    net_.opt.use_vulkan_compute = args["use_vulkan"].get<bool>();
+  }
  if (!device_.is_host()) {
    return Status(eNotSupported);
  }
@ -30,10 +32,29 @@ Result<void> NCNNNet::Init(const Value& args) {
  auto name = args["name"].get<std::string>();
  auto model = context["model"].get<Model>();
  OUTCOME_TRY(auto config, model.GetModelConfig(name));
-
+  auto precision = config.precision;
+  if (precision == "FP16") {
+    net_.opt.use_fp16_packed = true;
+    net_.opt.use_fp16_storage = true;
+    net_.opt.use_fp16_arithmetic = true;
+  } else if (precision == "INT8") {
+    // in android platform, ncnn will automatically start FP16 accelerate.
+    // In INT8 case, we set fp16 as false explicitly.
+    net_.opt.use_int8_packed = true;
+    net_.opt.use_int8_storage = true;
+    net_.opt.use_int8_arithmetic = true;
+    net_.opt.use_fp16_packed = false;
+    net_.opt.use_fp16_storage = false;
+    net_.opt.use_fp16_arithmetic = false;
+  } else {
+    // in android platform, ncnn will automatically start FP16 accelerate.
+    // In FP32 case, we set fp16 as false explicitly.
+    net_.opt.use_fp16_packed = false;
+    net_.opt.use_fp16_storage = false;
+    net_.opt.use_fp16_arithmetic = false;
+  }
  OUTCOME_TRY(params_, model.ReadFile(config.net));
  OUTCOME_TRY(weights_, model.ReadFile(config.weights));
-
  register_mmdeploy_custom_layers(net_);

  OUTCOME_TRY(ncnn_status(net_.load_param_mem(params_.c_str())));
@ -57,7 +78,6 @@ Result<void> NCNNNet::Init(const Value& args) {
        x,
    });
  }
-
  return success();
 }

@ -91,10 +111,24 @@ Result<void> NCNNNet::Forward() {
    OUTCOME_TRY(ncnn_status(extractor.extract(output_indices_[i], outputs[i])));
    auto& tensor = output_tensors_[i];
    auto shape = outputs[i].shape();
-    tensor.Reshape({1, shape.w, shape.h, shape.c});
+    if (outputs[i].dims == 1) {
+      tensor.Reshape({1, shape.w});
+    } else if (outputs[i].dims == 2) {
+      tensor.Reshape({1, shape.h, shape.w});
+    } else if (outputs[i].dims == 3) {
+      tensor.Reshape({1, shape.d, shape.h, shape.w});
+    } else {
+      // for dim==4 case and blank image.
+      tensor.Reshape({1, shape.c, shape.d, shape.h, shape.w});
+    }
+    // tensor.Reshape({1, shape.c, shape.h, shape.w});
    // ncnn Mat may be padded, flatten to avoid that
-    auto flattened = outputs[i].reshape(shape.w * shape.h * shape.c);
-    OUTCOME_TRY(tensor.CopyFrom(flattened.data, stream_));
+    auto flattened = outputs[i].reshape(shape.c * shape.h * shape.w);
+    // if ((shape.c * shape.h * shape.w) > 0)
+    if (outputs[i].dims > 0) {
+      OUTCOME_TRY(tensor.CopyFrom(flattened.data, stream_));
+    }
+    OUTCOME_TRY(stream_.Wait());
  }
  return success();
 }
--- a/csrc/preprocess/transform/pad.cpp
+++ b/csrc/preprocess/transform/pad.cpp
@ -22,7 +22,17 @@ PadImpl::PadImpl(const Value& args) : TransformImpl(args) {
  }

  arg_.size_divisor = args.value("size_divisor", 1);
-  arg_.pad_val = args.value("pad_val", 0.0f);
+  if (args.contains("pad_val")) {
+    if (args["pad_val"].is_number()) {
+      arg_.pad_val = args["pad_val"].get<float>();
+    } else if (args["pad_val"].contains("img")) {
+      arg_.pad_val = args["pad_val"]["img"][0].get<float>();
+    } else {
+      throw std::invalid_argument("args must be number or img dict");
+    }
+  } else {
+    arg_.pad_val = 0.0f;
+  }
  arg_.pad_to_square = args.value("pad_to_square", false);
  arg_.padding_mode = args.value("padding_mode", std::string("constant"));
 }
@ -31,43 +41,39 @@ Result<Value> PadImpl::Process(const Value& input) {
  MMDEPLOY_DEBUG("input: {}", to_json(input).dump(2));
  Value output = input;
  auto img_fields = GetImageFields(input);
-
  for (auto& key : img_fields) {
    Tensor output_tensor;
    auto tensor = input[key].get<Tensor>();
-
    assert(tensor.desc().shape.size() == 4);
    assert(tensor.desc().shape[0] == 1);
    assert(tensor.desc().shape[3] == 3 || tensor.desc().shape[3] == 1);

    int height = tensor.desc().shape[1];
    int width = tensor.desc().shape[2];
-
    if (arg_.pad_to_square) {
      int max_size = std::max(tensor.desc().shape[1], tensor.desc().shape[2]);
      std::array padding{0, 0, max_size - width, max_size - height};
-
      OUTCOME_TRY(output_tensor, PadImage(tensor, padding));
-
      output["pad_fixed_size"].push_back(max_size);
      output["pad_fixed_size"].push_back(max_size);
+    } else if (arg_.size[0] != 0 && arg_.size[1] != 0) {
+      output_tensor = tensor;
+      std::array padding{0, 0, arg_.size[1] - width, arg_.size[0] - height};
+      OUTCOME_TRY(output_tensor, PadImage(tensor, padding));
+      output["pad_fixed_size"].push_back(arg_.size[0]);
+      output["pad_fixed_size"].push_back(arg_.size[1]);
    } else if (arg_.size_divisor != 1) {
      auto pad_h = (height + arg_.size_divisor - 1) / arg_.size_divisor * arg_.size_divisor;
      auto pad_w = (width + arg_.size_divisor - 1) / arg_.size_divisor * arg_.size_divisor;
      std::array padding{0, 0, pad_w - width, pad_h - height};
-
      OUTCOME_TRY(output_tensor, PadImage(tensor, padding));
-
      output["pad_size_divisor"] = arg_.size_divisor;
      output["pad_fixed_size"].push_back(pad_h);
      output["pad_fixed_size"].push_back(pad_w);
    } else {
-      std::array padding{0, 0, arg_.size[1] - width, arg_.size[0] - height};
-
-      OUTCOME_TRY(output_tensor, PadImage(tensor, padding));
-
-      output["pad_fixed_size"].push_back(arg_.size[0]);
-      output["pad_fixed_size"].push_back(arg_.size[1]);
+      output_tensor = tensor;
+      output["pad_fixed_size"].push_back(height);
+      output["pad_fixed_size"].push_back(width);
    }
    output[key] = output_tensor;
    for (auto& v : output_tensor.desc().shape) {
--- a/docs/en/build/android.md
+++ b/docs/en/build/android.md
@ -0,0 +1,210 @@
+# Build for Android
+
+- [Build for Android](#build-for-android)
+  - [Build From Source](#build-from-source)
+    - [Install Toolchains](#install-toolchains)
+    - [Install Dependencies](#install-dependencies)
+      - [Install Dependencies for SDK](#install-dependencies-for-sdk)
+    - [Build MMDeploy](#build-mmdeploy)
+      - [Build Options Spec](#build-options-spec)
+      - [Build SDK](#build-sdk)
+      - [Build Demo](#build-demo)
+
+---
+MMDeploy provides cross compile for android platform.
+
+Model converter is executed on linux platform, and SDK is executed on android platform.
+
+Here are two steps for android build.
+
+1. Build model converter on linux, please refer to [How to build linux](./linux.md)
+
+2. Build SDK using android toolchain on linux.
+
+This doc is only for how to build SDK using android toolchain on linux.
+
+## Build From Source
+
+### Install Toolchains
+
+- cmake
+
+    **Make sure cmake version >= 3.14.0**. If not, you can follow instructions below to install cmake 3.20.0. For more versions of cmake, please refer to [cmake website](https://cmake.org/install).
+
+    ```bash
+    sudo apt-get install -y libssl-dev
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.0/cmake-3.20.0.tar.gz
+    tar -zxvf cmake-3.20.0.tar.gz
+    cd cmake-3.20.0
+    ./bootstrap
+    make
+    sudo make install
+    ```
+
+- ANDROID NDK 19+
+
+    **Make sure android ndk version >= 19.0**. If not, you can follow instructions below to install android ndk r23b. For more versions of android ndk, please refer to [android ndk website](https://developer.android.com/ndk/downloads).
+
+    ```bash
+    wget https://dl.google.com/android/repository/android-ndk-r23b-linux.zip
+    unzip android-ndk-r23b-linux.zip
+    cd android-ndk-r23b
+    export NDK_PATH=${PWD}
+    ```
+
+### Install Dependencies
+
+#### Install Dependencies for SDK
+
+You can skip this chapter if only interested in model converter.
+<table>
+<thead>
+  <tr>
+    <th>NAME </th>
+    <th>INSTALLATION </th>
+  </tr>
+</thead>
+<tbody>
+  <tr>
+    <td>spdlog </td>
+    <td>
+<pre><code>
+git clone -b v1.9.2 https://github.com/gabime/spdlog.git
+cd spdlog
+export SPDLOG_DIR=${PWD}
+mkdir -p build
+cd build
+cmake -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DCMAKE_INSTALL_PREFIX=${SPDLOG_DIR} -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-30 ..
+make install
+</code></pre>
+   </td>
+  </tr>
+  <tr>
+    <td>OpenCV<br>(>=3.0) </td>
+    <td>
+<pre><code>
+export OPENCV_VERSION=4.5.4
+wget https://github.com/opencv/opencv/releases/download/${OPENCV_VERSION}/opencv-${OPENCV_VERSION}-android-sdk.zip
+unzip opencv-${OPENCV_VERSION}-android-sdk.zip
+export OPENCV_ANDROID_SDK_DIR=${PWD}/OpenCV-android-sdk
+</code></pre>
+    </td>
+
+  </tr>
+  <tr>
+    <td>ncnn </td>
+    <td>A high-performance neural network inference computing framework supporting for android.</br>
+  <b> Now, MMDeploy supports v20211208 and has to use <code>git clone</code> to download it.</b><br>
+<pre><code>
+git clone -b 20211208 https://github.com/Tencent/ncnn.git
+cd ncnn
+git submodule update --init
+export NCNN_DIR=${PWD}
+mkdir -p build
+cd build
+cmake -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-30 -DNCNN_VULKAN=ON -DNCNN_DISABLE_EXCEPTION=OFF -DNCNN_DISABLE_RTTI=OFF ..
+make install
+</code></pre>
+   </td>
+  </tr>
+</tbody>
+</table>
+
+### Build MMDeploy
+#### Build Options Spec
+<table>
+<thead>
+  <tr>
+    <th>NAME</th>
+    <th>VALUE</th>
+    <th>DEFAULT</th>
+    <th>REMARK</th>
+  </tr>
+</thead>
+<tbody>
+  <tr>
+    <td>MMDEPLOY_BUILD_SDK</td>
+    <td>{ON, OFF}</td>
+    <td>OFF</td>
+    <td>Switch to build MMDeploy SDK</td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_BUILD_SDK_PYTHON_API</td>
+    <td>{ON, OFF}</td>
+    <td>OFF</td>
+    <td>switch to build MMDeploy SDK python package</td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_BUILD_TEST</td>
+    <td>{ON, OFF}</td>
+    <td>OFF</td>
+    <td>Switch to build MMDeploy SDK unittest cases</td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_TARGET_DEVICES</td>
+    <td>{"cpu"}</td>
+    <td>cpu</td>
+    <td>Enable target device. <br>If you want use ncnn vulkan accelerate, you still fill <code>{"cpu"}</code> here. Because, vulkan accelerate is only for ncnn net. The other part of inference is still using cpu.</td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_TARGET_BACKENDS</td>
+    <td>{"ncnn"}</td>
+    <td>N/A</td>
+    <td>Enabling inference engine. <br><b>By default, no target inference engine is set, since it highly depends on the use case.</b><br> Only ncnn backend is supported for android platform now.<br>
+    After specifying the inference engine, it's package path has to be passed to cmake as follows, <br>
+    1. <b>ncnn</b>: ncnn. <code>ncnn_DIR</code> is needed.
+<pre><code>-Dncnn_DIR=${NCNN_DIR}/build/install/lib/cmake/ncnn</code></pre>
+   </td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_CODEBASES</td>
+    <td>{"mmcls", "mmdet", "mmseg", "mmedit", "mmocr", "all"}</td>
+    <td>N/A</td>
+    <td>Enable codebase's postprocess modules. It MUST be set by a semicolon separated list of codebase names. The currently supported codebases are 'mmcls', 'mmdet', 'mmedit', 'mmseg', 'mmocr'. Instead of listing them one by one, you can also pass <code>all</code> to enable them all, i.e., <code>-DMMDEPLOY_CODEBASES=all</code></td>
+  </tr>
+  <tr>
+    <td>BUILD_SHARED_LIBS</td>
+    <td>{ON, OFF}</td>
+    <td>ON</td>
+    <td>Switch to build shared library or static library of MMDeploy SDK. Now you should build static library for android. Bug will be fixed soon.</td>
+  </tr>
+</tbody>
+</table>
+
+#### Build SDK
+MMDeploy provides a recipe as shown below for building SDK with ncnn as inference engine for android.
+
+- cpu + ncnn
+  ```Bash
+  cd ${MMDEPLOY_DIR}
+  mkdir -p build && cd build
+  cmake .. \
+      -DMMDEPLOY_BUILD_SDK=ON \
+      -DOpenCV_DIR=${OPENCV_ANDROID_SDK_DIR}/sdk/native/jni/abi-arm64-v8a \
+      -Dspdlog_DIR=${SPDLOG_DIR}/lib/cmake/spdlog \
+      -Dncnn_DIR=${NCNN_DIR}/build/install/lib/cmake/ncnn \
+      -DMMDEPLOY_TARGET_BACKENDS=ncnn \
+      -DMMDEPLOY_CODEBASES=all \
+      -DBUILD_SHARED_LIBS=OFF \
+      -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=android-30 \
+      -DANDROID_CPP_FEATURES="rtti exceptions"
+
+  make -j$(nproc) && make install
+  ```
+
+#### Build Demo
+
+```Bash
+cd ${MMDEPLOY_DIR}/build/install/example
+mkdir -p build && cd build
+cmake .. \
+      -DOpenCV_DIR=${OPENCV_ANDROID_SDK_DIR}/sdk/native/jni/abi-arm64-v8a \
+      -Dspdlog_DIR=${SPDLOG_DIR}/lib/cmake/spdlog \
+      -DMMDeploy_DIR=${MMDEPLOY_DIR}/build/install/lib/cmake/MMDeploy \
+      -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=android-30
+cmake --build . -- -j$(nproc)
+```
--- a/docs/zh_cn/build/android.md
+++ b/docs/zh_cn/build/android.md
@ -0,0 +1,205 @@
+# Android 下构建方式
+
+- [Android 下构建方式](#Android 下构建方式)
+  - [源码安装](#源码安装)
+    - [安装构建和编译工具链](#安装构建和编译工具链)
+    - [安装依赖包](#安装依赖包)
+      - [安装 MMDeploy SDK 依赖](#安装-MMDeploy-SDK-依赖)
+    - [编译 MMDeploy](#编译-mmdeploy)
+      - [编译选项说明](#编译选项说明)
+      - [编译 SDK](#编译-sdk)
+      - [编译 Demo](#编译-demo)
+
+---
+MMDeploy 为 android 平台提供交叉编译的构建方式.
+
+MMDeploy converter 部分在 linux 平台上执行,SDK 部分在 android 平台上执行.
+
+MMDeploy 的交叉编译分为两步:
+
+1. 在 linux 平台上构建 MMDeploy converter. 请根据 [How to build linux](./linux.md) 进行构建.
+
+2. 使用 android 工具链构建 MMDeploy SDK.
+
+本文档仅提供在 linux 平台上使用 android 工具链进行交叉编译构建 MMDeploy SDK 的方法.
+
+## 源码安装
+
+### 安装构建和编译工具链
+
+- cmake
+
+    **保证 cmake的版本 >= 3.14.0**. 如果不是,可以参考以下命令安装 3.20.0 版本. 如要获取其他版本,请参考 [这里](https://cmake.org/install)
+
+    ```bash
+    sudo apt-get install -y libssl-dev
+    wget https://github.com/Kitware/CMake/releases/download/v3.20.0/cmake-3.20.0.tar.gz
+    tar -zxvf cmake-3.20.0.tar.gz
+    cd cmake-3.20.0
+    ./bootstrap
+    make
+    sudo make install
+    ```
+
+- ANDROID NDK 19+
+
+    **保证 android ndk 的版本 >= 19.0**. 如果不是,可以参考以下命令安装 r23b 版本. 如要获取其他版本,请参考 [这里](https://developer.android.com/ndk/downloads)
+
+    ```bash
+    wget https://dl.google.com/android/repository/android-ndk-r23b-linux.zip
+    unzip android-ndk-r23b-linux.zip
+    cd android-ndk-r23b
+    export NDK_PATH=${PWD}
+    ```
+
+### 安装依赖包
+
+#### 安装 MMDeploy SDK 依赖
+
+如果您只对模型转换感兴趣,那么可以跳过本章节.
+<table>
+<thead>
+  <tr>
+    <th>名称 </th>
+    <th>安装方式 </th>
+  </tr>
+</thead>
+<tbody>
+  <tr>
+    <td>spdlog </td>
+    <td>
+<pre><code>
+git clone -b v1.9.2 https://github.com/gabime/spdlog.git
+cd spdlog
+export SPDLOG_DIR=${PWD}
+mkdir -p build
+cd build
+cmake -DCMAKE_POSITION_INDEPENDENT_CODE=ON -DCMAKE_INSTALL_PREFIX=${SPDLOG_DIR} -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-30 ..
+make install
+</code></pre>
+   </td>
+  </tr>
+  <tr>
+    <td>OpenCV<br>(>=3.0) </td>
+    <td>
+<pre><code>
+export OPENCV_VERSION=4.5.4
+wget https://github.com/opencv/opencv/releases/download/${OPENCV_VERSION}/opencv-${OPENCV_VERSION}-android-sdk.zip
+unzip opencv-${OPENCV_VERSION}-android-sdk.zip
+export OPENCV_ANDROID_SDK_DIR=${PWD}/OpenCV-android-sdk
+</code></pre>
+    </td>
+
+  </tr>
+  <tr>
+    <td>ncnn </td>
+    <td>ncnn 是支持 android 平台的高效神经网络推理计算框架</br>
+  <b> 目前, MMDeploy 支持 ncnn 的 20211208 版本, 且必须使用<code>git clone</code> 下载源码的方式安装</b><br>
+<pre><code>
+git clone -b 20211208 https://github.com/Tencent/ncnn.git
+cd ncnn
+git submodule update --init
+export NCNN_DIR=${PWD}
+mkdir -p build
+cd build
+cmake -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-30 -DNCNN_VULKAN=ON -DNCNN_DISABLE_EXCEPTION=OFF -DNCNN_DISABLE_RTTI=OFF ..
+make install
+</code></pre>
+   </td>
+  </tr>
+</tbody>
+</table>
+
+### 编译 MMDeploy
+#### 编译选项说明
+<table>
+<thead>
+  <tr>
+    <th>编译选项</th>
+    <th>取值范围</th>
+    <th>缺省值</th>
+    <th>说明</th>
+  </tr>
+</thead>
+<tbody>
+  <tr>
+    <td>MMDEPLOY_BUILD_SDK</td>
+    <td>{ON, OFF}</td>
+    <td>OFF</td>
+    <td>MMDeploy SDK 编译开关</td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_BUILD_TEST</td>
+    <td>{ON, OFF}</td>
+    <td>OFF</td>
+    <td>MMDeploy SDK 的单元测试程序编译开关</td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_TARGET_DEVICES</td>
+    <td>{"cpu"}</td>
+    <td>cpu</td>
+    <td>设置目标设备. <br>如果您想使用 ncnn 的 vulkan 加速功能, 您仍旧需要在这里填写<code>{"cpu"}</code>参数. 这是因为 vulkan 加速仅用于 ncnn 网络计算内部,推理过程的其他部分仍旧使用 cpu 计算.</td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_TARGET_BACKENDS</td>
+    <td>{"ncnn"}</td>
+    <td>N/A</td>
+    <td>设置推理后端. <br><b>默认情况下,SDK不设置任何后端, 因为它与应用场景高度相关.</b><br> Android 端目前只支持ncnn一个后端 <br>
+    构建时,几乎每个后端,都需传入一些路径变量,用来查找依赖包. <br>
+    1. <b>ncnn</b>: 表示 ncnn. 需要设置<code>ncnn_DIR</code>.
+<pre><code>-Dncnn_DIR=${NCNN_DIR}/build/install/lib/cmake/ncnn</code></pre>
+   </td>
+  </tr>
+  <tr>
+    <td>MMDEPLOY_CODEBASES</td>
+    <td>{"mmcls", "mmdet", "mmseg", "mmedit", "mmocr", "all"}</td>
+    <td>N/A</td>
+    <td>用来设置 SDK 后处理组件,加载 OpenMMLab 算法仓库的后处理功能. 已支持的算法仓库有'mmcls','mmdet','mmedit','mmseg'和'mmocr'. 如果选择多个codebase,中间使用分号隔开. 比如, 'mmcls', 'mmdet', 'mmedit', 'mmseg', 'mmocr'. 也可以通过 <code>all</code> 的方式, 加载所有codebase, 即 <code>-DMMDEPLOY_CODEBASES=all.</code></td>
+  </tr>
+  <tr>
+    <td>BUILD_SHARED_LIBS</td>
+    <td>{ON, OFF}</td>
+    <td>ON</td>
+    <td>MMDeploy SDK 的动态库的编译开关.设置 OFF 时, 编译静态库. 目前 android 端 SDK 仅支持静态库加载, 后续会进行对动态库加载的支持.</td>
+  </tr>
+</tbody>
+</table>
+
+#### 编译 SDK
+
+下文展示构建SDK的样例，用 ncnn 作为推理引擎。
+
+- cpu + ncnn
+  ```Bash
+  cd ${MMDEPLOY_DIR}
+  mkdir -p build && cd build
+  cmake .. \
+      -DMMDEPLOY_BUILD_SDK=ON \
+      -DOpenCV_DIR=${OPENCV_ANDROID_SDK_DIR}/sdk/native/jni/abi-arm64-v8a \
+      -Dspdlog_DIR=${SPDLOG_DIR}/lib/cmake/spdlog \
+      -Dncnn_DIR=${NCNN_DIR}/build/install/lib/cmake/ncnn \
+      -DMMDEPLOY_TARGET_BACKENDS=ncnn \
+      -DMMDEPLOY_CODEBASES=all \
+      -DBUILD_SHARED_LIBS=OFF \
+      -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=android-30 \
+      -DANDROID_CPP_FEATURES="rtti exceptions"
+
+  make -j$(nproc) && make install
+  ```
+
+#### 编译 Demo
+
+```Bash
+cd ${MMDEPLOY_DIR}/build/install/example
+mkdir -p build && cd build
+cmake .. \
+      -DOpenCV_DIR=${OPENCV_ANDROID_SDK_DIR}/sdk/native/jni/abi-arm64-v8a \
+      -Dspdlog_DIR=${SPDLOG_DIR}/lib/cmake/spdlog \
+      -DMMDeploy_DIR=${MMDEPLOY_DIR}/build/install/lib/cmake/MMDeploy \
+      -DCMAKE_TOOLCHAIN_FILE=${NDK_PATH}/build/cmake/android.toolchain.cmake \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=android-30
+cmake --build . -- -j$(nproc)
+```
--- a/mmdeploy/utils/export_info.py
+++ b/mmdeploy/utils/export_info.py
@ -112,6 +112,8 @@ def get_models(deploy_cfg: Union[str, mmcv.Config],
    elif backend == Backend.NCNN:
        net = onnx_name.replace('.onnx', '.param')
        weights = onnx_name.replace('.onnx', '.bin')
+        if 'precision' in deploy_cfg['backend_config']:
+            precision = deploy_cfg['backend_config']['precision']
    elif backend == Backend.ONNXRUNTIME:
        pass
    else:
@ -153,13 +155,16 @@ def get_inference_info(deploy_cfg: mmcv.Config, model_cfg: mmcv.Config,
    input_names = onnx_config.get('input_names', None)
    input_name = input_names[0] if input_names else 'input'
    input_map = dict(img=input_name)
-    return dict(
+    return_dict = dict(
        name=name,
        type=type,
        module=module,
        input=input,
        output=output,
        input_map=input_map)
+    if 'use_vulkan' in deploy_cfg['backend_config']:
+        return_dict['use_vulkan'] = deploy_cfg['backend_config']['use_vulkan']
+    return return_dict


 def get_preprocess(deploy_cfg: mmcv.Config, model_cfg: mmcv.Config):