Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle-Lite …

…into fix_fill_op test=huawei_ascend_npu
PaddlePaddle · Oct 22, 2021 · e55d24f · e55d24f
2 parents 3bbbe1e + 66a7c61
commit e55d24f
Show file tree

Hide file tree

Showing 23 changed files with 328 additions and 158 deletions.
diff --git a/cmake/postproject.cmake b/cmake/postproject.cmake
@@ -115,8 +115,17 @@ endif()
 if (LITE_ON_TINY_PUBLISH)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -ffast-math -Ofast -Os -fomit-frame-pointer")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden -ffunction-sections")
+    # 1. strip useless symbols from third-party libs
+    # exclude-libs is not supported on macOs system
     if(NOT ARMMACOS)
-    check_linker_flag(-Wl,--gc-sections)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,--exclude-libs,ALL")
+      check_linker_flag(-Wl,--gc-sections)
+    endif()
+    # 2. strip rtti lib to reduce lib size
+    #     2.1 replace typeid by fastTypeId
+    #     2.2 replace dynamic_cast by static_cast
+    if(NOT LITE_WITH_NNADAPTER)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fno-rtti")
     endif()
 endif()
 

diff --git a/docs/api_reference/cxx_api_doc.md b/docs/api_reference/cxx_api_doc.md
@@ -3,7 +3,7 @@
 
 ## CreatePaddlePredictor
 
- \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_api.h)&gt;
+ \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_api.h)&gt;
 
 ```c++
 template <typename ConfigT>
@@ -32,7 +32,7 @@ std::shared_ptr<PaddlePredictor> predictor = CreatePaddlePredictor<MobileConfig>
 
 ## CxxConfig
 
- \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_api.h)&gt;
+ \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_api.h)&gt;
 
 ```c++
 class CxxConfig;
@@ -244,7 +244,7 @@ int x86_math_num_threads() const;
 
 ## MobileConfig
 
- \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_api.h)&gt;
+ \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_api.h)&gt;
 
 ```c++
 class MobileConfig;
@@ -474,7 +474,7 @@ int threads() const
 
 ## PaddlePredictor
 
- \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_api.h)&gt;
+ \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_api.h)&gt;
 
 ```c++
 class PaddlePredictor
@@ -650,7 +650,7 @@ virtual std::string GetVersion() const = 0;
 
 ## TargetType
 
- \#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_place.h)&gt;
+ \#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_place.h)&gt;
 
 ```c++
 class TargetType;
@@ -661,7 +661,7 @@ class TargetType;
 
 ## PrecisionType
 
-\#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_place.h)&gt;
+\#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_place.h)&gt;
 
 ```c++
 class PrecisionType;
@@ -672,7 +672,7 @@ class PrecisionType;
 
 ## DataLayoutType
 
-\#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_place.h)&gt;
+\#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_place.h)&gt;
 
 ```c++
 class DataLayoutType;
@@ -683,7 +683,7 @@ class DataLayoutType;
 
 ## Place
 
- \#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_place.h)&gt;
+ \#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_place.h)&gt;
 
 `Place`是`TargetType`、`PrecisionType`和`DataLayoutType`的集合，说明运行时的设备类型、数据精度和数据格式。
 
@@ -694,7 +694,7 @@ Place{TargetType(ARM), PrecisionType(FP32), DataLayoutType(NCHW)}
 
 ## PowerMode
 
-\#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_place.h)&gt;
+\#include &lt;[paddle\_place.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_place.h)&gt;
 
 ```c++
 enum PowerMode;
@@ -728,7 +728,7 @@ PowerMode详细说明如下：
 
 
 ## Tensor
- \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/v2.9.1/lite/api/paddle_api.h)&gt;
+ \#include &lt;[paddle\_api.h](https://github.com/PaddlePaddle/Paddle-Lite/tree/develop/lite/api/paddle_api.h)&gt;
 
 ```c++
 struct Tensor

diff --git a/docs/develop_guides/add_hardware.md b/docs/develop_guides/add_hardware.md
diff --git a/docs/source_compile/compile_options.md b/docs/source_compile/compile_options.md
@@ -5,9 +5,9 @@
 | :-- |  :-- |--: |
 | LITE_WITH_LOG |  是否输出日志信息 | ON |
 | LITE_WITH_EXCEPTION | 是否在错误发生时抛出异常 | OFF |
-| LITE_WITH_TRAIN |  打开[模型训练功能](../demo_guides/cpp_train_demo.md)，支持移动端模型训练 | OFF |
-| LITE_BUILD_EXTRA |  编译[全量预测库](library.md)，包含更多算子和模型支持 | OFF |
-| LITE_BUILD_TAILOR | 编译时[根据模型裁剪预测库](library_tailoring.rst)，缩小预测库大小 | OFF |
+| LITE_WITH_TRAIN |  打开[模型训练功能](https://paddle-lite.readthedocs.io/zh/develop/demo_guides/cpp_train_demo.html)，支持移动端模型训练 | OFF |
+| LITE_BUILD_EXTRA |  编译[全量预测库](https://paddle-lite.readthedocs.io/zh/develop/source_compile/library.html)，包含更多算子和模型支持 | OFF |
+| LITE_BUILD_TAILOR | 编译时[根据模型裁剪预测库](https://paddle-lite.readthedocs.io/zh/develop/source_compile/library_tailoring.html)，缩小预测库大小 | OFF |
 | WITH_SYSTEM_BLAS |  编译时强制使用 reference BLAS |  OFF |
 
 ### 轻量级编译选项
@@ -25,26 +25,26 @@
 
 | 编译选项 |  说明  | 默认值 |
 | :-- |  :-- | --: |
-| LITE_WITH_PROFILE |  编译[性能 Profiler 工具](../user_guides/profiler.md)，用于 kernel 耗时统计 | OFF |
-| LITE_WITH_PRECISION_PROFILE |  编译[精度 Profiler 工具](../user_guides/profiler.md)，用于 kernel 精度分析 | OFF |
+| LITE_WITH_PROFILE |  编译[性能 Profiler 工具](https://paddle-lite.readthedocs.io/zh/develop/user_guides/profiler.html)，用于 kernel 耗时统计 | OFF |
+| LITE_WITH_PRECISION_PROFILE |  编译[精度 Profiler 工具](https://paddle-lite.readthedocs.io/zh/develop/user_guides/profiler.html)，用于 kernel 精度分析 | OFF |
 | WITH_TESTING |  编译 Lite 单测模块 | OFF |
 
 ## 部分平台相关编译选项
 
 | 编译选项 |  说明  | 适用平台 | 默认值 |
 | :-- |  :-- | --: | --: |
 | LITE_WITH_ARM |  编译支持 Andriod 或 ARMLinux 平台预测库 | Andriod / ARMLinux | OFF |
-| LITE_WITH_JAVA |  编译支持 [Java API](../api_reference/java_api_doc.md) 的预测库 | Andriod / ARMLinux | OFF |
+| LITE_WITH_JAVA |  编译支持 [Java API](https://paddle-lite.readthedocs.io/zh/develop/api_reference/java_api_doc.html) 的预测库 | Andriod / ARMLinux | OFF |
 | LITE_WITH_ARM_CLANG | 使用 clang 编译 ARM 平台预测库 | Andriod / ARMLinux |OFF |
 | WITH_ARM_DOTPROD |  编译 ARM 点积指令优化的预测库 | Andriod / ARMLinux |ON |
-| LITE_WITH_CV |  编译 [CV 图像加速库](../api_reference/cv.md) | Andirod / ARMLinux |OFF |
+| LITE_WITH_CV |  编译 [CV 图像加速库](https://paddle-lite.readthedocs.io/zh/develop/api_reference/cv.html) | Andirod / ARMLinux |OFF |
 | ANDROID_API_LEVEL | 设置安卓 API LEVEL | Android | Default，即 ARMv7 下为16，ARMv8 下为21 |
 | LITE_WITH_OPENMP |  编译时打开 OpenMP | ARMLinux / X86 | ON |
-| LITE_WITH_X86 |  编译[ X86 平台](../demo_guides/x86.md)预测库 | X86 | ON |
+| LITE_WITH_X86 |  编译[ X86 平台](https://paddle-lite.readthedocs.io/zh/develop/demo_guides/x86.html)预测库 | X86 | ON |
 | WITH_AVX |  编译有 AVX 指令优化的预测库 | X86 |ON IF ${AVX_FOUND} |
 | WITH_MKL | 编译有 Intel MKL 支持的预测库 | X86 |ON IF ${AVX_FOUND} |
-| LITE_ON_MODEL_OPTIMIZE_TOOL |  编译[模型优化工具 opt](../user_guides/model_optimize_tool.md) | X86 |OFF|
-| LITE_WITH_PYTHON |  编译支持 [Python API](../api_reference/python_api_doc.md) 的预测库 | X86 / CUDA |OFF |
-| LITE_WITH_OPENCL |  编译 [OpenCL 平台](../demo_guides/opencl.md)预测库 | OpenCL | OFF |
-| LITE_WITH_XPU |  编译[百度 XPU 平台](../demo_guides/baidu_xpu.md)预测库 | XPU |OFF |
+| LITE_ON_MODEL_OPTIMIZE_TOOL |  编译[模型优化工具 opt](https://paddle-lite.readthedocs.io/zh/develop/user_guides/model_optimize_tool.html) | X86 |OFF|
+| LITE_WITH_PYTHON |  编译支持 [Python API](https://paddle-lite.readthedocs.io/zh/develop/api_reference/python_api_doc.html) 的预测库 | X86 / CUDA |OFF |
+| LITE_WITH_OPENCL |  编译 [OpenCL 平台](https://paddle-lite.readthedocs.io/zh/develop/demo_guides/opencl.html)预测库 | OpenCL | OFF |
+| LITE_WITH_XPU |  编译[百度 XPU 平台](https://paddle-lite.readthedocs.io/zh/develop/demo_guides/baidu_xpu.html)预测库 | XPU |OFF |
 | LITE_WITH_XTCL | 通过 XTCL 方式支持百度 XPU，默认 Kernel 方式 | XPU |OFF IF LITE_WITH_XPU |
diff --git a/lite/backends/opencl/cl_kernel/image/elementwise_broadcast_kernel.cl b/lite/backends/opencl/cl_kernel/image/elementwise_broadcast_kernel.cl
@@ -63,19 +63,19 @@ __kernel void broadcast_elementwise_common(
     in0 = READ_IMG_TYPE(
         CL_DTYPE_CHAR, input_x, SAMPLER, (int2)(cur_index.x * 4, cur_index.y));
 
-    if (cur_index.x * 4 + 1 < input_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 1 < bias_width) {
       in1 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
                           (int2)(cur_index.x * 4 + 1, cur_index.y));
     }
-    if (cur_index.x * 4 + 2 < input_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 2 < bias_width) {
       in2 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
                           (int2)(cur_index.x * 4 + 2, cur_index.y));
     }
-    if (cur_index.x * 4 + 3 < input_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 3 < bias_width) {
       in3 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
@@ -100,19 +100,19 @@ __kernel void broadcast_elementwise_common(
     in0 = READ_IMG_TYPE(
         CL_DTYPE_CHAR, input_x, SAMPLER, (int2)(cur_index.y, cur_index.x * 4));
 
-    if (cur_index.x * 4 + 1 < input_nhwc4.z * input_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 1 < bias_width) {
       in1 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
                           (int2)(cur_index.y, cur_index.x * 4 + 1));
     }
-    if (cur_index.x * 4 + 2 < input_nhwc4.z * input_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 2 < bias_width) {
       in2 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
                           (int2)(cur_index.y, cur_index.x * 4 + 2));
     }
-    if (cur_index.x * 4 + 3 < input_nhwc4.z * input_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 3 < bias_width) {
       in3 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
@@ -140,19 +140,19 @@ __kernel void broadcast_elementwise_common(
                         SAMPLER,
                         (int2)(tmp_c4 * input_nhwc4.y + tmp_w, tmp_h));
 
-    if (cur_index.x + 1 < input_nhwc4.x * input_nhwc4.y) {
+    if (tmp_h + 1 < bias_width) {
       in1 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
                           (int2)(tmp_c4 * input_nhwc4.y + tmp_w, tmp_h + 1));
     }
-    if (cur_index.x + 2 < input_nhwc4.x * input_nhwc4.y) {
+    if (tmp_h + 2 < bias_width) {
       in2 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
                           (int2)(tmp_c4 * input_nhwc4.y + tmp_w, tmp_h + 2));
     }
-    if (cur_index.x + 3 < input_nhwc4.x * input_nhwc4.y) {
+    if (tmp_h + 3 < bias_width) {
       in3 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_x,
                           SAMPLER,
@@ -235,19 +235,19 @@ __kernel void broadcast_elementwise_common(
     in0 = READ_IMG_TYPE(
         CL_DTYPE_CHAR, input_y, SAMPLER, (int2)(cur_index.y, cur_index.x * 4));
 
-    if (cur_index.x * 4 + 1 < bias_nhwc4.z * bias_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 1 < bias_width) {
       in1 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_y,
                           SAMPLER,
                           (int2)(cur_index.y, cur_index.x * 4 + 1));
     }
-    if (cur_index.x * 4 + 2 < bias_nhwc4.z * bias_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 2 < bias_width) {
       in2 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_y,
                           SAMPLER,
                           (int2)(cur_index.y, cur_index.x * 4 + 2));
     }
-    if (cur_index.x * 4 + 3 < bias_nhwc4.z * bias_nhwc4.w * 4) {
+    if (cur_index.x * 4 + 3 < bias_width) {
       in3 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_y,
                           SAMPLER,
@@ -275,19 +275,19 @@ __kernel void broadcast_elementwise_common(
                         SAMPLER,
                         (int2)(tmp_c4 * bias_nhwc4.y + tmp_w, tmp_h));
 
-    if (cur_index.x + 1 < bias_nhwc4.x * bias_nhwc4.y) {
+    if (tmp_h + 1 < bias_width) {
       in1 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_y,
                           SAMPLER,
                           (int2)(tmp_c4 * bias_nhwc4.y + tmp_w, tmp_h + 1));
     }
-    if (cur_index.x + 2 < bias_nhwc4.x * bias_nhwc4.y) {
+    if (tmp_h + 2 < bias_width) {
       in2 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_y,
                           SAMPLER,
                           (int2)(tmp_c4 * bias_nhwc4.y + tmp_w, tmp_h + 2));
     }
-    if (cur_index.x + 3 < bias_nhwc4.x * bias_nhwc4.y) {
+    if (tmp_h + 3 < bias_width) {
       in3 = READ_IMG_TYPE(CL_DTYPE_CHAR,
                           input_y,
                           SAMPLER,

diff --git a/lite/core/optimizer/mir/memory_optimize_pass.cc b/lite/core/optimizer/mir/memory_optimize_pass.cc
@@ -132,9 +132,9 @@ void MemoryOptimizePass::CollectLifeCycleByDevice(
                             {"flatten", {{"X"}, {"Out"}}},
                             {"flatten2", {{"X"}, {"Out"}}},
                             {"squeeze", {{"X"}, {"Out"}}},
-                            {"squeeze2", {{"X"}, {"Out"}}},
+                            {"squeeze2", {{"X"}, {{"Out"}, {"XShape"}}}},
                             {"unsqueeze", {{"X"}, {"Out"}}},
-                            {"unsqueeze2", {{"X"}, {"Out"}}}};
+                            {"unsqueeze2", {{"X"}, {{"Out"}, {"XShape"}}}}};
     auto inplace_op_node = inplace_op_nodes.find(op_type);
     if (inplace_op_node != inplace_op_nodes.end()) {
       bool inplace = false;

diff --git a/lite/core/optimizer/mir/pass_manager.h b/lite/core/optimizer/mir/pass_manager.h
@@ -73,7 +73,7 @@ class PassManager {
   template <typename PassTy>
   PassTy* LookUp(const std::string& key) {
     auto it = pass_map_.find(key);
-    if (it != pass_map_.end()) return dynamic_cast<PassTy*>(it->second);
+    if (it != pass_map_.end()) return static_cast<PassTy*>(it->second);
     return nullptr;
   }
 

diff --git a/lite/kernels/npu/bridges/graph.h b/lite/kernels/npu/bridges/graph.h
@@ -79,9 +79,9 @@ class Graph {
                             PrecisionType precision = PRECISION(kFloat),
                             DataLayoutType layout = DATALAYOUT(kNCHW)) {
     Node::Role role = Node::Role::kVar;
-    if (typeid(T) == typeid(ge::op::Const)) {
+    if (FastTypeId<T>() == FastTypeId<ge::op::Const>()) {
       role = Node::Role::kConst;
-    } else if (typeid(T) == typeid(ge::op::Data)) {
+    } else if (FastTypeId<T>() == FastTypeId<ge::op::Data>()) {
       role = Node::Role::kData;
     }
     auto node = std::make_shared<Node>(precision, layout, role);

diff --git a/lite/kernels/opencl/elementwise_image_compute.cc b/lite/kernels/opencl/elementwise_image_compute.cc
@@ -261,7 +261,7 @@ class ElementwiseImageCompute : public KernelLite<TARGET(kOpenCL),
 
     int inputx_broadcast_c_flag = (x_nchw_[1] == 1) ? 1 : 0;
     int inputy_broadcast_c_flag = (y_nchw_[1] == 1) ? 1 : 0;
-    int bias_width = y_nchw_[1];
+    int bias_width = out_nchw_[1];
 
     if (y_dims_ == x_dims_) {
       cl_int status = kernel_.setArg(0, *x_img);

diff --git a/lite/kernels/opencl/elementwise_image_compute_test.cc b/lite/kernels/opencl/elementwise_image_compute_test.cc
@@ -528,6 +528,10 @@ void test_elementwise_all_dim_data_gpu() {
   c = randint(1, 40);
   h = randint(1, 40);
   w = randint(1, 40);
+  n = 2;
+  c = 3;
+  h = 4;
+  w = 5;
   std::vector<bool> xy_swap_flags{false, true};
   for (auto xy_swap_flag : xy_swap_flags) {
     RunElementwiseCommonSize<float>({n, c, h, w},

diff --git a/lite/operators/conv_op.h b/lite/operators/conv_op.h
@@ -175,7 +175,7 @@ class ConvOpLite : public OpLite {
       padding_algorithm_ = op_desc.GetAttr<std::string>("padding_algorithm");
     }
     // For Int8
-    const OpInfo* op_info = dynamic_cast<const OpInfo*>(&op_desc);
+    const OpInfo* op_info = static_cast<const OpInfo*>(&op_desc);
     if (op_info != nullptr && op_info->HasAttr("enable_int8")) {
       param_.enable_int8 = op_info->GetAttr<bool>("enable_int8");
       auto input_scale_name = "Input0_scale";

diff --git a/lite/operators/conv_transpose_op.cc b/lite/operators/conv_transpose_op.cc
@@ -151,7 +151,7 @@ bool ConvTransposeOpLite::AttachImpl(const cpp::OpDesc& op_desc,
   param_.dilations = std::make_shared<std::vector<int>>(dilations);
 
   // For Int8
-  const OpInfo* op_info = dynamic_cast<const OpInfo*>(&op_desc);
+  const OpInfo* op_info = static_cast<const OpInfo*>(&op_desc);
   if (op_info != nullptr && op_info->HasAttr("enable_int8")) {
     param_.enable_int8 = op_info->GetAttr<bool>("enable_int8");
     auto input_scale_name = "Input0_scale";

diff --git a/lite/operators/fc_op.cc b/lite/operators/fc_op.cc
@@ -115,7 +115,7 @@ bool FcOpLite::AttachImpl(const cpp::OpDesc& op_desc, lite::Scope* scope) {
   }
 
   // For Int8
-  const OpInfo* op_info = dynamic_cast<const OpInfo*>(&op_desc);
+  const OpInfo* op_info = static_cast<const OpInfo*>(&op_desc);
   if (op_info != nullptr && op_info->HasAttr("enable_int8")) {
     param_.enable_int8 = op_info->GetAttr<bool>("enable_int8");
     auto input_scale_name = "Input0_scale";

diff --git a/lite/operators/gru_op.cc b/lite/operators/gru_op.cc
@@ -102,7 +102,7 @@ bool GRUOpLite::AttachImpl(const cpp::OpDesc& op_desc, lite::Scope* scope) {
   }
 
   // For int8
-  const OpInfo* op_info = dynamic_cast<const OpInfo*>(&op_desc);
+  const OpInfo* op_info = static_cast<const OpInfo*>(&op_desc);
   if (op_info != nullptr && op_info->HasAttr("enable_int8") &&
       op_info->GetAttr<bool>("enable_int8")) {
     param_.enable_int8 = true;

diff --git a/lite/operators/lstm_op.cc b/lite/operators/lstm_op.cc
@@ -117,7 +117,7 @@ bool LstmOp::AttachImpl(const cpp::OpDesc &opdesc, lite::Scope *scope) {
       GetActivationType(opdesc.GetAttr<std::string>("candidate_activation"));
 
   // For int8
-  const OpInfo *op_info = dynamic_cast<const OpInfo *>(&opdesc);
+  const OpInfo *op_info = static_cast<const OpInfo *>(&opdesc);
   if (op_info != nullptr && op_info->HasAttr("enable_int8") &&
       op_info->GetAttr<bool>("enable_int8")) {
     param_.enable_int8 = true;

diff --git a/lite/operators/matmul_op.cc b/lite/operators/matmul_op.cc
@@ -148,7 +148,7 @@ bool MatMulOpLite::AttachImpl(const cpp::OpDesc &op_desc, lite::Scope *scope) {
   param_.transpose_Y = op_desc.GetAttr<bool>("transpose_Y");
   param_.alpha = op_desc.GetAttr<float>("alpha");
 
-  const OpInfo *op_info = dynamic_cast<const OpInfo *>(&op_desc);
+  const OpInfo *op_info = static_cast<const OpInfo *>(&op_desc);
   if (op_info != nullptr && op_info->HasAttr("enable_int8")) {
     param_.enable_int8 = op_info->GetAttr<bool>("enable_int8");
     auto input_scale_name = "X0_scale";

diff --git a/lite/operators/sparse_conv_op.h b/lite/operators/sparse_conv_op.h
@@ -122,7 +122,7 @@ class SparseConvOp : public OpLite {
     }
 
     // For Int8
-    const OpInfo* op_info = dynamic_cast<const OpInfo*>(&op_desc);
+    const OpInfo* op_info = static_cast<const OpInfo*>(&op_desc);
     if (op_info != nullptr && op_info->HasAttr("enable_int8")) {
       param_.enable_int8 = op_info->GetAttr<bool>("enable_int8");
       auto input_scale_name = "Input0_scale";

diff --git a/lite/operators/subgraph_op.cc b/lite/operators/subgraph_op.cc
@@ -40,7 +40,7 @@ bool SubgraphOp::AttachImpl(const cpp::OpDesc& op_desc, lite::Scope* scope) {
   param_.output_data_names =
       op_desc.GetAttr<std::vector<std::string>>("output_data_names");
   // Get the quantization parameters of input and output data variables
-  auto op_info = dynamic_cast<const OpInfo*>(&op_desc);
+  auto op_info = static_cast<const OpInfo*>(&op_desc);
   param_.input_data_scales.clear();
   param_.output_data_scales.clear();
   for (auto& input_data_name : param_.input_data_names) {