merge conflicts

e5f779de · hanbuhe · 8d3d2e58 · e5f779de · e5f779de · e5f779de
24 changed file
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -22,9 +22,6 @@ if (WITH_PADDLE_MOBILE)
    return()
 endif(WITH_PADDLE_MOBILE)

-# set(CMAKE_BUILD_TYPE DEBUG)
-
-
 set(PADDLE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
 set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
 set(CMAKE_CXX_STANDARD 11)

--- a/docs/demo_guides/ios_app_demo.md
+++ b/docs/demo_guides/ios_app_demo.md
@@ -90,7 +90,7 @@ ios-detection_demo/detection_demo/ViewController.mm

 ## 代码讲解 （如何使用Paddle-Lite C++ API 执行预测）

-IOS 示例基于C++ API 开发，调用Paddle-Lite C++ API包括以下五步。更详细的API 描述参考： [Paddle-Lite C++ API](https://paddle-lite.readthedocs.io/zh/latest/api_reference/java_api_doc.html)。
+IOS 示例基于C++ API 开发，调用Paddle-Lite C++ API包括以下五步。更详细的API 描述参考： [Paddle-Lite C++ API](https://paddle-lite.readthedocs.io/zh/latest/api_reference/cxx_api_doc.html)。

 ```c++
 #include <iostream>

--- a/docs/demo_guides/linux_arm_demo.md
+++ b/docs/demo_guides/linux_arm_demo.md
+# Linux(ARM) Demo
+
+## 多种应用场景
+
+我们提供Paddle-Lite示例工程[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)，其中包含[Android](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo)、[iOS](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-ios-demo)和[Armlinux](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-armlinux-demo)平台的示例工程。Linux(ARM) demo涵盖[图像分类](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/image_classification_demo)、[目标检测](https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo/object_detection_demo)2个应用场景。
+
+### 1. 图像分类
+
+Paddle-Lite提供的图像分类demo ，在移动端上提供了实时的物体识别能力，可以应用到生产线自动分拣或质检、识别医疗图像、辅助医生肉眼诊断等场景。在移动端预测的效果图如下：
+
+<p align="center"><img width="250" height="250"  src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/tabby_cat.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="250" height="250"  src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/tabby_cat2.jpg"/></p>
+
+### 2. 物体检测
+
+Paddle-Lite提供的物体检测demo ，在移动端上提供了检测多个物体的位置、名称、位置及数量的能力。可以应用到视频监控（是否有违规物体或行为）、工业质检（微小瑕疵的数量和位置）、医疗诊断（细胞计数、中药识别）等场景。在移动端预测的效果图如下：
+
+<p align="center"><img width="250" height="250"  src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/dog.jpg"/>&#8194;&#8194;&#8194;&#8194;&#8194;<img width="250" height="250"  src="https://paddlelite-data.bj.bcebos.com/doc_images/Android_iOS_demo/demo/dog2.jpg"/></p>
+
+## Linux(ARM) demo部署方法
+
+下面我们以**目标检测（object_detection_demo)**为例讲解如何部署iOS工程。
+
+**目的**：将基于Paddle-Lite的预测库部署到Linux(ARM)设备，实现物体检测的目标。
+
+**需要的环境**：Linux(ARM)设备、下载到本地的[Paddle-Lite-Demo](https://github.com/PaddlePaddle/Paddle-Lite-Demo)工程
+
+**部署步骤**：
+
+1、 目标检测的Linux(ARM)示例位于 `Paddle-Lite-Demo\PaddleLite-armlinux-demo\object_detection_demo`
+
+2、终端中执行 `download_models_and_libs.sh` 脚本自动下载模型和Paddle-Lite预测库
+
+```shell
+cd PaddleLite-armlinux-demo          # 1. 终端中进入 Paddle-Lite-Demo\PaddleLite-armlinux-demo
+sh download_models_and_libs.sh       # 2. 执行脚本下载依赖项 （需要联网）
+```
+
+下载完成后会出现提示： `Download successful!`
+
+3、终端中执行 `download_models_and_libs.sh` 脚本自动下载模型和Paddle-Lite预测库
+```shell
+cd object_detection_demo    # 1. 终端中进入
+sh run.sh                   # 2. 执行脚本编译并执行物体检测demo，输出预测数据和运行时间
+```
+demo结果如下:
+<img width="836" alt="image" src="https://user-images.githubusercontent.com/50474132/82852558-da228580-9f35-11ea-837c-e4d71066da57.png">
+
+## 使用C++接口预测
+Linux(ARM) demo 示例基于C++ API 开发，调用Paddle-Lite C++ API包括以下五步。更详细的API 描述参考： [Paddle-Lite C++ API](https://paddle-lite.readthedocs.io/zh/latest/api_reference/cxx_api_doc.html)。
+
+```c++
+#include <iostream>
+// 引入C++ API
+#include "paddle_lite/paddle_api.h"
+#include "paddle_lite/paddle_use_ops.h"
+#include "paddle_lite/paddle_use_kernels.h"
+
+// 1. 设置MobileConfig
+MobileConfig config;
+config.set_model_from_file(<modelPath>); // 设置NaiveBuffer格式模型路径
+config.set_power_mode(LITE_POWER_NO_BIND); // 设置CPU运行模式
+config.set_threads(4); // 设置工作线程数
+
+// 2. 创建PaddlePredictor
+std::shared_ptr<PaddlePredictor> predictor = CreatePaddlePredictor<MobileConfig>(config);
+
+// 3. 设置输入数据
+std::unique_ptr<Tensor> input_tensor(std::move(predictor->GetInput(0)));
+input_tensor->Resize({1, 3, 224, 224});
+auto* data = input_tensor->mutable_data<float>();
+for (int i = 0; i < ShapeProduction(input_tensor->shape()); ++i) {
+  data[i] = 1;
+}
+
+// 4. 执行预测
+predictor->run();
+
+// 5. 获取输出数据
+std::unique_ptr<const Tensor> output_tensor(std::move(predictor->GetOutput(0)));
+std::cout << "Output shape " << output_tensor->shape()[1] << std::endl;
+for (int i = 0; i < ShapeProduction(output_tensor->shape()); i += 100) {
+  std::cout << "Output[" << i << "]: " << output_tensor->data<float>()[i]
+            << std::endl;
+}
+```
+
+## 使用Python接口预测
+
+1. Python预测库编译参考[编译Linux](../user_guides/Compile/Linux)，建议在开发版上编译。
+2. [Paddle-Lite Python API](https://paddle-lite.readthedocs.io/zh/latest/api_reference/python_api_doc.html)。
+3. 代码参考，[Python预测](python_demo)
--- a/lite/api/CMakeLists.txt
+++ b/lite/api/CMakeLists.txt
--- a/lite/api/paddle_use_passes.h
+++ b/lite/api/paddle_use_passes.h
@@ -44,6 +44,7 @@ USE_MIR_PASS(type_layout_cast_pass);
 USE_MIR_PASS(type_layout_cast_preprocess_pass);
 USE_MIR_PASS(memory_optimize_pass);
 USE_MIR_PASS(kernel_place_correct_pass)
+
 USE_MIR_PASS(multi_stream_analysis_pass);
 USE_MIR_PASS(elementwise_mul_constant_eliminate_pass)
 USE_MIR_PASS(npu_subgraph_pass);

--- a/lite/backends/arm/math/pooling.h
+++ b/lite/backends/arm/math/pooling.h
@@ -76,30 +76,55 @@ void pooling1x1s2p0_max(const float* din,
                        int pad_bottom,
                        int pad_right);

-void pooling2x2s2_max(const float* din,
-                      float* dout,
-                      int num,
-                      int chout,
-                      int hout,
-                      int wout,
-                      int chin,
-                      int hin,
-                      int win,
-                      int pad_bottom,
-                      int pad_right);
-
-void pooling2x2s2_avg(const float* din,
-                      float* dout,
-                      int num,
-                      int chout,
-                      int hout,
-                      int wout,
-                      int chin,
-                      int hin,
-                      int win,
-                      bool exclusive,
-                      int pad_bottom,
-                      int pad_right);
+void pooling2x2s2p0_max(const float* din,
+                        float* dout,
+                        int num,
+                        int chout,
+                        int hout,
+                        int wout,
+                        int chin,
+                        int hin,
+                        int win,
+                        int pad_bottom,
+                        int pad_right);
+
+void pooling2x2s2p0_avg(const float* din,
+                        float* dout,
+                        int num,
+                        int chout,
+                        int hout,
+                        int wout,
+                        int chin,
+                        int hin,
+                        int win,
+                        bool exclusive,
+                        int pad_bottom,
+                        int pad_right);
+
+void pooling2x2s2p1_max(const float* din,
+                        float* dout,
+                        int num,
+                        int chout,
+                        int hout,
+                        int wout,
+                        int chin,
+                        int hin,
+                        int win,
+                        int pad_bottom,
+                        int pad_right);
+
+void pooling2x2s2p1_avg(const float* din,
+                        float* dout,
+                        int num,
+                        int chout,
+                        int hout,
+                        int wout,
+                        int chin,
+                        int hin,
+                        int win,
+                        bool exclusive,
+                        int pad_bottom,
+                        int pad_right);

 void pooling3x3s1p1_max(const float* din,
                        float* dout,

--- a/lite/backends/fpga/KD/llapi/zynqmp_api.cpp
+++ b/lite/backends/fpga/KD/llapi/zynqmp_api.cpp
@@ -204,7 +204,8 @@ int get_device_info(const struct DeviceInfo &args) {
 int perform_bypass(const struct BypassArgs &args) {
  int ret = -1;
  int size = args.image.channels * args.image.width * args.image.height;
-  int max_size = 1 << 20;
+
+  int max_size = 1 << 21;

  float times = 1.0 * size / max_size;
  int count = static_cast<int>(times);

--- a/lite/backends/fpga/KD/pes/prior_box_pe.cpp
+++ b/lite/backends/fpga/KD/pes/prior_box_pe.cpp
@@ -241,13 +241,10 @@ void PriorBoxPE::compute_prior_box() {
  }

  boxes.flush();
-  // boxes.syncToCPU();
+  boxes.syncToCPU();
  variances.flush();
  output_boxes->copyFrom(&boxes);
  output_variances->copyFrom(&variances);
-
-  output_boxes->invalidate();
-  output_variances->invalidate();
 }

 void PriorBoxPE::apply() {}
@@ -256,9 +253,8 @@ bool PriorBoxPE::dispatch() {
  if (cachedBoxes_ == nullptr) {
    cachedBoxes_ = new Tensor();
    cachedVariances_ = new Tensor();
-    cachedBoxes_->mutableData<float16>(FP16, param_.outputBoxes->shape());
-    cachedVariances_->mutableData<float16>(FP16,
-                                           param_.outputVariances->shape());
+    cachedBoxes_->mutableData<float>(FP32, param_.outputBoxes->shape());
+    cachedVariances_->mutableData<float>(FP32, param_.outputVariances->shape());
    cachedBoxes_->setDataLocation(CPU);
    cachedVariances_->setDataLocation(CPU);
    compute_prior_box();

--- a/lite/backends/fpga/KD/pes/scale_pe.hpp
+++ b/lite/backends/fpga/KD/pes/scale_pe.hpp
@@ -36,7 +36,6 @@ class ScalePE : public PE {
  }

  inline int lcm(int a, int b) { return a * b / gcd(a, b); }
-
  bool init() {
    Tensor* output = param_.output;
    output->setAligned(true);

--- a/lite/core/mir/static_kernel_pick_pass.h
+++ b/lite/core/mir/static_kernel_pick_pass.h
@@ -164,11 +164,6 @@ class StaticKernelPickPass : public mir::StmtPass {
    // might have different data layout.
    // TODO(Superjomn) reconsider the idea of taking the data layout as a kernel
    // specification.
-
-    if (kernel.place().target == TARGET(kFPGA)) {
-      final_score = 1000;
-    }
-
    return final_score;
  }


--- a/lite/demo/python/mobilenetv1_full_api.py
+++ b/lite/demo/python/mobilenetv1_full_api.py
@@ -20,9 +20,6 @@ from __future__ import division
 from __future__ import print_function

 import argparse
-import sys
-sys.path.append('../../python/lib')
-
 from paddlelite.lite import *

 # Command arguments
@@ -42,8 +39,8 @@ def RunModel(args):
        config.set_param_file(args.param_file)
    else:
        config.set_model_dir(args.model_dir)
-    # For x86, you can set places = [Place(TargetType.X86, PrecisionType.FP32)]
-    places = [Place(TargetType.ARM, PrecisionType.FP32)]
+    # For arm platform (armlinux), you can set places = [Place(TargetType.ARM, PrecisionType.FP32)]
+    places = [Place(TargetType.X86, PrecisionType.FP32)]
    config.set_valid_places(places)

    # 2. Create paddle predictor

--- a/lite/demo/python/mobilenetv1_light_api.py
+++ b/lite/demo/python/mobilenetv1_light_api.py
@@ -20,8 +20,6 @@ from __future__ import division
 from __future__ import print_function

 import argparse
-import sys
-sys.path.append('../../python/lib')

 from paddlelite.lite import *

@@ -33,7 +31,7 @@ parser.add_argument(
 def RunModel(args):
    # 1. Set config information
    config = MobileConfig()
-    config.set_model_dir(args.model_dir)
+    config.set_model_from_file(args.model_dir)

    # 2. Create paddle predictor
    predictor = create_paddle_predictor(config)

--- a/lite/kernels/arm/cast_compute.cc
+++ b/lite/kernels/arm/cast_compute.cc
@@ -62,10 +62,6 @@ void CastCompute::Run() {
    int32_t* out_data = param.Out->mutable_data<int32_t>();
    std::transform(
        x_data_begin, x_data_end, out_data, TransOp<int64_t, int32_t>);
-  } else if (param.in_dtype == 3 && param.out_dtype == 5) {
-    const auto* x_data = param.X->data<float>();
-    auto* o_data = param.Out->mutable_data<float>();
-    memcpy(o_data, x_data, sizeof(float) * param.X->numel());
  } else {
    LOG(FATAL) << "other has not been implemented";
  }

--- a/lite/kernels/arm/lookup_table_compute.cc
+++ b/lite/kernels/arm/lookup_table_compute.cc
@@ -34,7 +34,7 @@ void LookupTableCompute::Run() {

  auto table_dim = w->dims();
  int64_t ids_numel = ids->numel();
-  auto ids_data = ids->data<float>();
+  auto ids_data = ids->data<int64_t>();

  int64_t row_number = table_dim[0];
  int64_t row_width = table_dim[1];
@@ -73,6 +73,7 @@ REGISTER_LITE_KERNEL(lookup_table,
    .BindInput("Ids", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kInt64))})
    .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kARM))})
    .Finalize();
+
 REGISTER_LITE_KERNEL(lookup_table_v2,
                     kARM,
                     kAny,

--- a/lite/kernels/arm/pool_compute.cc
+++ b/lite/kernels/arm/pool_compute.cc
@@ -58,6 +58,7 @@ void PoolCompute::Run() {
  bool global_pooling = (paddings[0] == 0) && (ksize[0] == in_dims[2]) &&
                        (ksize[1] == in_dims[3]) && kps_equal && pads_equal;
  global_pooling = param.global_pooling || global_pooling;
+
  if (global_pooling) {
    for (size_t i = 0; i < ksize.size(); ++i) {
      paddings[2 * i] = 0;
@@ -107,35 +108,65 @@ void PoolCompute::Run() {
    } else if (ksize[0] == 2 && strides[0] == 2 && paddings[0] == 0 &&
               kps_equal) {
      if (pooling_type == "max") {
-        lite::arm::math::pooling2x2s2_max(din,
-                                          dout,
-                                          out_dims[0],
-                                          out_dims[1],
-                                          out_dims[2],
-                                          out_dims[3],
-                                          in_dims[1],
-                                          in_dims[2],
-                                          in_dims[3],
-                                          paddings[1],
-                                          paddings[3]);
+        lite::arm::math::pooling2x2s2p0_max(din,
+                                            dout,
+                                            out_dims[0],
+                                            out_dims[1],
+                                            out_dims[2],
+                                            out_dims[3],
+                                            in_dims[1],
+                                            in_dims[2],
+                                            in_dims[3],
+                                            paddings[1],
+                                            paddings[3]);
        return;
      } else if (pooling_type == "avg") {
-        lite::arm::math::pooling2x2s2_avg(din,
-                                          dout,
-                                          out_dims[0],
-                                          out_dims[1],
-                                          out_dims[2],
-                                          out_dims[3],
-                                          in_dims[1],
-                                          in_dims[2],
-                                          in_dims[3],
-                                          exclusive,
-                                          paddings[1],
-                                          paddings[3]);
+        lite::arm::math::pooling2x2s2p0_avg(din,
+                                            dout,
+                                            out_dims[0],
+                                            out_dims[1],
+                                            out_dims[2],
+                                            out_dims[3],
+                                            in_dims[1],
+                                            in_dims[2],
+                                            in_dims[3],
+                                            exclusive,
+                                            paddings[1],
+                                            paddings[3]);
        return;
      }
-    } else if (ksize[0] == 3 && strides[0] == 1 && paddings[0] == 1 &&
+    } else if (ksize[0] == 2 && strides[0] == 2 && paddings[0] == 1 &&
               kps_equal) {
+      if (pooling_type == "max") {
+        lite::arm::math::pooling2x2s2p1_max(din,
+                                            dout,
+                                            out_dims[0],
+                                            out_dims[1],
+                                            out_dims[2],
+                                            out_dims[3],
+                                            in_dims[1],
+                                            in_dims[2],
+                                            in_dims[3],
+                                            paddings[1],
+                                            paddings[3]);
+        return;
+      } else if (pooling_type == "avg") {
+        lite::arm::math::pooling2x2s2p1_avg(din,
+                                            dout,
+                                            out_dims[0],
+                                            out_dims[1],
+                                            out_dims[2],
+                                            out_dims[3],
+                                            in_dims[1],
+                                            in_dims[2],
+                                            in_dims[3],
+                                            exclusive,
+                                            paddings[1],
+                                            paddings[3]);
+        return;
+      }
+    } else if (ksize[0] == 3 && strides[0] == 1 && paddings[0] == 1 &&
+               pads_equal && kps_equal) {
      if (pooling_type == "max") {
        lite::arm::math::pooling3x3s1p1_max(din,
                                            dout,
@@ -165,7 +196,7 @@ void PoolCompute::Run() {
        return;
      }
    } else if (ksize[0] == 3 && strides[0] == 1 && paddings[0] == 0 &&
-               kps_equal) {
+               pads_equal && kps_equal) {
      if (pooling_type == "max") {
        lite::arm::math::pooling3x3s1p0_max(din,
                                            dout,
@@ -195,7 +226,7 @@ void PoolCompute::Run() {
        return;
      }
    } else if (ksize[0] == 3 && strides[0] == 2 && paddings[0] == 0 &&
-               kps_equal) {
+               pads_equal && kps_equal) {
      if (pooling_type == "max") {
        lite::arm::math::pooling3x3s2p0_max(din,
                                            dout,
@@ -225,7 +256,7 @@ void PoolCompute::Run() {
        return;
      }
    } else if (ksize[0] == 3 && strides[0] == 2 && paddings[0] == 1 &&
-               kps_equal) {
+               pads_equal && kps_equal) {
      if (pooling_type == "max") {
        lite::arm::math::pooling3x3s2p1_max(din,
                                            dout,

--- a/lite/kernels/fpga/prior_box_compute.cc
+++ b/lite/kernels/fpga/prior_box_compute.cc
@@ -132,26 +132,3 @@ REGISTER_LITE_KERNEL(prior_box,
    .BindOutput("Variances", {LiteType::GetTensorTy(TARGET(kARM))})
    .Finalize();

-// REGISTER_LITE_KERNEL(prior_box,
-//                      kFPGA,
-//                      kFP16,
-//                      kNHWC,
-//                      paddle::lite::kernels::fpga::PriorBoxCompute,
-//                      def)
-//     .BindInput("Input",
-//                {LiteType::GetTensorTy(TARGET(kFPGA),
-//                                       PRECISION(kFP16),
-//                                       DATALAYOUT(kNHWC))})
-//     .BindInput("Image",
-//                {LiteType::GetTensorTy(TARGET(kFPGA),
-//                                       PRECISION(kFP16),
-//                                       DATALAYOUT(kNHWC))})
-//     .BindOutput("Boxes",
-//                {LiteType::GetTensorTy(TARGET(kFPGA),
-//                                       PRECISION(kFP16),
-//                                       DATALAYOUT(kNHWC))})
-//     .BindOutput("Variances",
-//                {LiteType::GetTensorTy(TARGET(kFPGA),
-//                                       PRECISION(kFP16),
-//                                       DATALAYOUT(kNHWC))})
-//     .Finalize();
--- a/lite/kernels/host/multiclass_nms_compute.cc
+++ b/lite/kernels/host/multiclass_nms_compute.cc
@@ -428,14 +428,8 @@ REGISTER_LITE_KERNEL(multiclass_nms,
                     kNCHW,
                     paddle::lite::kernels::host::MulticlassNmsCompute,
                     def)
-    .BindInput("BBoxes",
-               {LiteType::GetTensorTy(TARGET(kHost),
-                                      PRECISION(kAny),
-                                      DATALAYOUT(kAny))})
-    .BindInput("Scores",
-               {LiteType::GetTensorTy(TARGET(kHost),
-                                      PRECISION(kAny),
-                                      DATALAYOUT(kAny))})
+    .BindInput("BBoxes", {LiteType::GetTensorTy(TARGET(kHost))})
+    .BindInput("Scores", {LiteType::GetTensorTy(TARGET(kHost))})
    .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kHost))})
    .Finalize();


--- a/lite/kernels/host/reshape_compute.cc
+++ b/lite/kernels/host/reshape_compute.cc
@@ -46,21 +46,17 @@ REGISTER_LITE_KERNEL(reshape,
                     paddle::lite::kernels::host::ReshapeCompute,
                     def)
    .BindInput("X",
-               {LiteType::GetTensorTy(TARGET(kHost),
-                                      PRECISION(kAny),
-                                      DATALAYOUT(kAny))})
+               {LiteType::GetTensorTy(
+                   TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny), -1)})
    .BindInput("ShapeTensor",
-               {LiteType::GetTensorTy(TARGET(kHost),
-                                      PRECISION(kAny),
-                                      DATALAYOUT(kAny))})
+               {LiteType::GetTensorTy(
+                   TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny), -1)})
    .BindInput("Shape",
-               {LiteType::GetTensorTy(TARGET(kHost),
-                                      PRECISION(kAny),
-                                      DATALAYOUT(kAny))})
+               {LiteType::GetTensorTy(
+                   TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny), -1)})
    .BindOutput("Out",
-                {LiteType::GetTensorTy(TARGET(kHost),
-                                       PRECISION(kAny),
-                                       DATALAYOUT(kAny))})
+                {LiteType::GetTensorTy(
+                    TARGET(kHost), PRECISION(kAny), DATALAYOUT(kAny), -1)})
    .Finalize();

 REGISTER_LITE_KERNEL(reshape2,

--- a/lite/tools/ci_build.sh
+++ b/lite/tools/ci_build.sh
@@ -289,6 +289,17 @@ function build_test_server {
    test_model_optimize_tool_compile
 }

+# Build the code and run lite server tests. This is executed in the CI system.
+function build_test_coverage {
+    mkdir -p ./build
+    cd ./build
+    export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:$PWD/third_party/install/mklml/lib"
+    cmake_x86_for_CI
+    build
+
+    test_server
+}
+
 # The CUDA version of CI is cuda_10.1.243_418.87.00_linux.
 # The cuDNN version is cudnn-10.1-linux-x64-v7.5.0.56.
 function build_test_cuda_server {
@@ -1136,7 +1147,7 @@ function main {
                shift
                ;;
            build_check_coverage)
-                build_test_server
+                build_test_coverage
                check_coverage
                shift
                ;;

--- a/lite/utils/cv/image_resize.cc
+++ b/lite/utils/cv/image_resize.cc
@@ -678,15 +678,9 @@ void resize(const uint8_t* src,
  } else if (srcFormat == NV12 || srcFormat == NV21) {
    nv21_resize(src, dst, srcw, srch, dstw, dsth);
    return;
-    num = 1;
-    int hout = static_cast<int>(0.5 * dsth);
-    dsth += hout;
  } else if (srcFormat == BGR || srcFormat == RGB) {
    bgr_resize(src, dst, srcw, srch, dstw, dsth);
    return;
-    w_in = srcw * 3;
-    w_out = dstw * 3;
-    num = 3;
  } else if (srcFormat == BGRA || srcFormat == RGBA) {
    w_in = srcw * 4;
    w_out = dstw * 4;

--- a/mobile/src/fpga/KD/pes/conv_pe.hpp
+++ b/mobile/src/fpga/KD/pes/conv_pe.hpp
@@ -29,6 +29,7 @@ namespace zynqmp {
 class ConvPE : public PE {
 public:
  bool init() {
+    std::cout << "Conv init" << std::endl;
    return true;
  }


--- a/tools/coverage/gcda_clean.py
+++ b/tools/coverage/gcda_clean.py
@@ -10,8 +10,7 @@ import sys

 from github import Github

-token = os.getenv('GITHUB_API_TOKEN', 'e51cb020919a6eef689257966e8fb6477981788a')
-
+token = os.getenv('GITHUB_API_TOKEN')

 def get_pull(pull_id):
    """

--- a/tools/coverage/paddle_lite_coverage.sh
+++ b/tools/coverage/paddle_lite_coverage.sh
-et -xe
+#!/bin/bash
+# The git version of CI is 2.7.4. This script is not compatible with git version 1.7.1.
+set -xe

 PADDLE_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}")/../../" && pwd )"

@@ -26,7 +28,6 @@ function gen_full_html_report() {
        '/Paddle-Lite/lite/kernels/*' \
        '/Paddle-Lite/lite/model_parser/*' \
        '/Paddle-Lite/lite/opreators/*' \
-        '/Paddle-Lite/lite/tests/*' \
        '/Paddle-Lite/lite/tools/*' \
        '/Paddle-Lite/lite/utils/*' \
        -o coverage-full.tmp \
@@ -35,7 +36,21 @@ function gen_full_html_report() {
    mv -f coverage-full.tmp coverage-full.info

    lcov --remove coverage-full.info \
-        '/Paddle-Lite/lite/demo*' \
+        '/Paddle-Lite/lite/tests/*' \
+        '/Paddle-Lite/lite/demo/*' \
+        '/Paddle-Lite/lite/fluid/*_test*' \
+        '/Paddle-Lite/lite/model_parser/*_test*' \
+        '/Paddle-Lite/lite/kernels/*/*test*' \
+        '/Paddle-Lite/lite/kernels/*/bridges/*test*' \
+        '/Paddle-Lite/lite/utils/*_test*' \
+        '/Paddle-Lite/lite/api/*test*' \
+        '/Paddle-Lite/lite/core/*_test*' \
+        '/Paddle-Lite/lite/core/*/*test*' \
+        '/Paddle-Lite/lite/core/mir/*/*_test*' \
+        '/Paddle-Lite/lite/core/mir/*_test*' \
+        '/Paddle-Lite/lite/backends/x86/*/*test*' \
+        '/Paddle-Lite/lite/backends/opencl/*test*' \
+        '/Paddle-Lite/lite/operators/*test*' \
        -o coverage-full.tmp \
        --rc lcov_branch_coverage=0

@@ -48,7 +63,7 @@ gen_full_html_report || true
 function gen_diff_html_report() {
    if [ "${GIT_PR_ID}" != "" ]; then
        COVERAGE_DIFF_PATTERN="`python ${PADDLE_ROOT}/tools/coverage/pull_request.py files ${GIT_PR_ID}`"
-
+        sleep 5
        python ${PADDLE_ROOT}/tools/coverage/pull_request.py diff ${GIT_PR_ID} > git-diff.out
    fi

@@ -57,6 +72,7 @@ function gen_diff_html_report() {
        -o coverage-diff.info \
        --rc lcov_branch_coverage=0
    
+    sleep 5 
    python ${PADDLE_ROOT}/tools/coverage/coverage_diff.py coverage-diff.info git-diff.out > coverage-diff.tmp

    mv -f coverage-diff.tmp coverage-diff.info

--- a/tools/coverage/pull_request.py
+++ b/tools/coverage/pull_request.py
@@ -11,8 +11,7 @@ import os

 from github import Github

-token = os.getenv('GITHUB_API_TOKEN', 'e51cb020919a6eef689257966e8fb6477981788a')
-
+token = os.getenv('GITHUB_API_TOKEN')

 def get_pull(pull_id):
    """