Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into...

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/pybind_for_protobuf_desc

cmake/generic.cmake

@@ -106,22 +106,22 @@ function(merge_static_libs TARGET_NAME)
   endforeach()
   list(REMOVE_DUPLICATES libs_deps)
 
-  if(APPLE) # Use OSX's libtool to merge archives
   # To produce a library we need at least one source file.
   # It is created by add_custom_command below and will helps
   # also help to track dependencies.
-    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c)
+  set(target_SRCS ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c)
 
+  if(APPLE) # Use OSX's libtool to merge archives
     # Make the generated dummy source file depended on all static input
     # libs. If input lib changes,the source file is touched
     # which causes the desired effect (relink).
-    add_custom_command(OUTPUT ${dummyfile}
-      COMMAND ${CMAKE_COMMAND} -E touch ${dummyfile}
+    add_custom_command(OUTPUT ${target_SRCS}
+      COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
       DEPENDS ${libs})
 
     # Generate dummy staic lib
-    file(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
-    add_library(${TARGET_NAME} STATIC ${dummyfile})
+    file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";")
+    add_library(${TARGET_NAME} STATIC ${target_SRCS})
     target_link_libraries(${TARGET_NAME} ${libs_deps})
 
     foreach(lib ${libs})
@@ -130,11 +130,14 @@ function(merge_static_libs TARGET_NAME)
     endforeach()
     add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
       COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
-      COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles})
+      COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}
+      )
   else() # general UNIX: use "ar" to extract objects and re-add to a common lib
+    set(target_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.dir)
+
     foreach(lib ${libs})
-      set(objlistfile ${lib}.objlist) # list of objects in the input library
-      set(objdir ${lib}.objdir)
+      set(objlistfile ${target_DIR}/${lib}.objlist) # list of objects in the input library
+      set(objdir ${target_DIR}/${lib}.objdir)
 
       add_custom_command(OUTPUT ${objdir}
         COMMAND ${CMAKE_COMMAND} -E make_directory ${objdir}
@@ -142,31 +145,32 @@ function(merge_static_libs TARGET_NAME)
 
       add_custom_command(OUTPUT ${objlistfile}
         COMMAND ${CMAKE_AR} -x "$<TARGET_FILE:${lib}>"
-        COMMAND ${CMAKE_AR} -t "$<TARGET_FILE:${lib}>" > ../${objlistfile}
+        COMMAND ${CMAKE_AR} -t "$<TARGET_FILE:${lib}>" > ${objlistfile}
         DEPENDS ${lib} ${objdir}
         WORKING_DIRECTORY ${objdir})
 
-      # Empty dummy source file that goes into merged library		
-      set(mergebase ${lib}.mergebase.c)		
-      add_custom_command(OUTPUT ${mergebase}		
-        COMMAND ${CMAKE_COMMAND} -E touch ${mergebase}		
-        DEPENDS ${objlistfile})		
-
-      list(APPEND mergebases "${mergebase}")
+      list(APPEND target_OBJS "${objlistfile}")
     endforeach()
 
-    add_library(${TARGET_NAME} STATIC ${mergebases})
+    # Make the generated dummy source file depended on all static input
+    # libs. If input lib changes,the source file is touched
+    # which causes the desired effect (relink).
+    add_custom_command(OUTPUT ${target_SRCS}
+      COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
+      DEPENDS ${libs} ${target_OBJS})
+
+    # Generate dummy staic lib
+    file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";")
+    add_library(${TARGET_NAME} STATIC ${target_SRCS})
     target_link_libraries(${TARGET_NAME} ${libs_deps})
 
     # Get the file name of the generated library
-    set(outlibfile "$<TARGET_FILE:${TARGET_NAME}>")
+    set(target_LIBNAME "$<TARGET_FILE:${TARGET_NAME}>")
 
-    foreach(lib ${libs})
     add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
-        COMMAND ${CMAKE_AR} cr ${outlibfile} *.o
-        COMMAND ${CMAKE_RANLIB} ${outlibfile}
-        WORKING_DIRECTORY ${lib}.objdir)
-    endforeach()
+        COMMAND ${CMAKE_AR} crs ${target_LIBNAME} `find ${target_DIR} -name '*.o'`
+        COMMAND ${CMAKE_RANLIB} ${target_LIBNAME}
+        WORKING_DIRECTORY ${target_DIR})
   endif()
 endfunction(merge_static_libs)
 
@@ -196,7 +200,7 @@ function(cc_library TARGET_NAME)
     add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS})
 
   else(cc_library_SRCS)
-    if (cc_library_DEPS)
+    if(cc_library_DEPS)
       merge_static_libs(${TARGET_NAME} ${cc_library_DEPS})
     else()
       message(FATAL "Please specify source file or library in cc_library.")

doc/faq/index_cn.rst

@@ -158,17 +158,23 @@ PaddlePaddle的参数使用名字 :code:`name` 作为参数的ID，相同名字
 
 这里 :code:`hidden_a` 和 :code:`hidden_b` 使用了同样的parameter和bias。并且softmax层的两个输入也使用了同样的参数 :code:`softmax_param`。
 
-7. \*-cp27mu-linux_x86_64.whl is not a supported wheel on this platform.
+7. paddlepaddle\*.whl is not a supported wheel on this platform.
 ------------------------------------------------------------------------
 
-出现这个问题的主要原因是，系统编译wheel包的时候，使用的 :code:`wheel` 包是最新的，
-而系统中的 :code:`pip` 包比较老。具体的解决方法是，更新 :code:`pip` 包并重新编译PaddlePaddle。
+出现这个问题的主要原因是，没有找到和当前系统匹配的paddlepaddle安装包。最新的paddlepaddle python安装包支持Linux x86_64和MacOS 10.12操作系统，并安装了python 2.7和pip 9.0.1。
+
 更新 :code:`pip` 包的方法是\:
 
 ..  code-block:: bash
 
     pip install --upgrade pip
 
+如果还不行，可以执行 :code:`python -c "import pip; print(pip.pep425tags.get_supported())"` 获取当前系统支持的python包的后缀，
+并对比是否和正在安装的后缀一致。
+
+如果系统支持的是 :code:`linux_x86_64` 而安装包是 :code:`manylinux1_x86_64` ，需要升级pip版本到最新；
+如果系统支持 :code:`manylinux1_x86_64` 而安装包（本地）是 :code:`linux_x86_64` ，可以重命名这个whl包为 :code:`manylinux1_x86_64` 再安装。
+
 8.  python相关的单元测试都过不了
 --------------------------------
 
@@ -310,7 +316,7 @@ Paddle二进制在运行时捕获了浮点数异常，只要出现浮点数异
 * 模型一直不收敛，发散到了一个数值特别大的地方。
 * 训练数据有问题，导致参数收敛到了一些奇异的情况。或者输入数据尺度过大，有些特征的取值达到数百万，这时进行矩阵乘法运算就可能导致浮点数溢出。
 
-主要的解决办法是减小学习律或者对数据进行归一化处理。
+主要的解决办法是减小学习率或者对数据进行归一化处理。
 
 15. 编译安装后执行 import paddle.v2 as paddle 报ImportError: No module named v2
 ------------------------------------------------------------------------
@@ -373,3 +379,15 @@ PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数
 
     parameters = paddle.parameters.create(my_cost)
     parameters.set('emb', load_parameter(emb_param_file, 30000, 256))
+
+18. 集群多节点训练，日志中保存均为网络通信类错误
+------------------------------
+
+集群多节点训练，日志报错为网络通信类错误，比如 :code:`Connection reset by peer` 等。
+此类报错通常是由于某一个节点的错误导致这个节点的训练进程退出，从而引发其他节点无法连接导致，可以参考下面的步骤排查：
+
+* 从 :code:`train.log` ， :code:`server.log` 找到最早报错的地方，查看是否是其他错误引发的报错（比如FPE，内存不足，磁盘空间不足等）。
+
+* 如果发现最早的报错就是网络通信的问题，很有可能是非独占方式执行导致的端口冲突，可以联系OP，看当前MPI集群是否支持resource=full参数提交，如果支持增加此参数提交，并更换job 端口。
+
+* 如果当前MPI集群并不支持任务独占模式，可以联系OP是否可以更换集群或升级当前集群。
\ No newline at end of file

paddle/capi/CMakeLists.txt

@@ -62,6 +62,7 @@ if(ANDROID)
           LIBRARY DESTINATION lib/${ANDROID_ABI})
   execute_process(
     COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
     OUTPUT_VARIABLE GIT_COMMITS_LIST
     RESULT_VARIABLE GIT_COMMITS_LIST_RESULT
     ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
@@ -81,8 +82,7 @@ if(ANDROID)
       )"
   )
 else(ANDROID)
-  install(TARGETS paddle_capi_whole
-          ARCHIVE DESTINATION lib)
+  install(TARGETS paddle_capi_whole ARCHIVE DESTINATION lib)
   if(NOT IOS)
     install(TARGETS paddle_capi_shared DESTINATION lib)
   endif()

paddle/framework/attribute.cc

@@ -31,6 +31,10 @@ ProgramDesc& GetProgramDesc() {
   return *g_program_desc;
 }
 
+template <>
+AttrType AttrTypeID<bool>() {
+  return BOOLEAN;
+}
 template <>
 AttrType AttrTypeID<int>() {
   return INT;
@@ -44,6 +48,10 @@ AttrType AttrTypeID<std::string>() {
   return STRING;
 }
 template <>
+AttrType AttrTypeID<std::vector<bool>>() {
+  return BOOLEANS;
+}
+template <>
 AttrType AttrTypeID<std::vector<int>>() {
   return INTS;
 }
@@ -66,6 +74,9 @@ AttrType AttrTypeID<BlockDesc>() {
 
 Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
   switch (attr_desc.type()) {
+    case framework::AttrType::BOOLEAN: {
+      return attr_desc.b();
+    }
     case framework::AttrType::INT: {
       return attr_desc.i();
     }
@@ -75,6 +86,13 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
     case framework::AttrType::STRING: {
       return attr_desc.s();
     }
+    case framework::AttrType::BOOLEANS: {
+      std::vector<bool> val(attr_desc.bools_size());
+      for (int i = 0; i < attr_desc.bools_size(); ++i) {
+        val[i] = attr_desc.bools(i);
+      }
+      return val;
+    }
     case framework::AttrType::INTS: {
       std::vector<int> val(attr_desc.ints_size());
       for (int i = 0; i < attr_desc.ints_size(); ++i) {

paddle/framework/attribute.h

@@ -27,8 +27,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
-                       std::vector<float>, std::vector<std::string>,
+typedef boost::variant<boost::blank, bool, int, float, std::string,
+                       std::vector<bool>, std::vector<int>, std::vector<float>,
+                       std::vector<std::string>,
                        std::vector<std::pair<int, int>>, BlockDesc*>
     Attribute;
 

paddle/framework/backward.cc

@@ -166,9 +166,8 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
 
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
-        net->AppendOp(OpRegistry::CreateOp("fill_zeros_like",
-                                           {{"Src", {prefix}}},
-                                           {{"Dst", {grad_input}}}, {}));
+        net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}},
+                                           {{"Y", {grad_input}}}, {}));
       }
       return false;
     });

paddle/framework/backward_test.cc

@@ -127,8 +127,8 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker {
  public:
   FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("Src", "x");
-    AddOutput("Dst", "out");
+    AddInput("X", "x");
+    AddOutput("Y", "out");
     AddComment("");
   }
 };
@@ -325,10 +325,10 @@ TEST(Backward, op_part_of_output_are_not_need) {
 
   auto &fill_zero = *net->ops_[0];
   ASSERT_EQ("fill_zeros_like", fill_zero.Type());
-  ASSERT_EQ(1UL, fill_zero.Inputs("Src").size());
-  ASSERT_EQ("Z", fill_zero.Input("Src"));
-  ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size());
-  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst"));
+  ASSERT_EQ(1UL, fill_zero.Inputs("X").size());
+  ASSERT_EQ("Z", fill_zero.Input("X"));
+  ASSERT_EQ(1UL, fill_zero.Outputs("Y").size());
+  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Y"));
 
   auto &d_many_out = *net->ops_[1];
   ASSERT_EQ("many_output_op_grad", d_many_out.Type());

paddle/framework/framework.proto

@@ -23,7 +23,9 @@ enum AttrType {
   FLOATS = 4;
   STRINGS = 5;
   INT_PAIRS = 6;
-  BLOCK = 7;
+  BOOLEAN = 7;
+  BOOLEANS = 8;
+  BLOCK = 9;
 }
 
 message IntPair {
@@ -45,7 +47,9 @@ message OpDesc {
     repeated float floats = 7;
     repeated string strings = 8;
     repeated IntPair int_pairs = 9;
-    optional int32 block_idx = 10;
+    optional bool b = 10;
+    repeated bool bools = 11;
+    optional int32 block_idx = 12;
   };
 
   message Var {

paddle/framework/operator.cc

@@ -207,23 +207,22 @@ const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
 }
 
 template <>
-Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
-  auto* var = OutputVar(name);
-  return var == nullptr ? nullptr : const_cast<Tensor*>(GetTensorFromVar(var));
+Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const {
+  auto var = OutputVar(name);
+  return var == nullptr ? nullptr : var->GetMutable<LoDTensor>();
 }
 
 template <>
-std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
     const std::string& name) const {
   auto names = op().Outputs(name);
   std::vector<Tensor*> res;
   res.reserve(names.size());
   std::transform(names.begin(), names.end(), std::back_inserter(res),
                  [&](const std::string& sub_name) {
-                   auto var = scope().FindVar(sub_name);
-                   return var == nullptr
-                              ? nullptr
-                              : const_cast<Tensor*>(GetTensorFromVar(var));
+                   auto var = scope_.FindVar(sub_name);
+                   return var == nullptr ? nullptr
+                                         : var->GetMutable<LoDTensor>();
                  });
   return res;
 }

paddle/framework/operator.h

@@ -212,9 +212,9 @@ class InferShapeContext {
     return res;
   }
 
-  std::vector<const Variable*> MultiOutputVar(const std::string& name) const {
+  std::vector<Variable*> MultiOutputVar(const std::string& name) const {
     auto names = op_.Outputs(name);
-    std::vector<const Variable*> res;
+    std::vector<Variable*> res;
     res.reserve(names.size());
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [this](const std::string& name) {
@@ -271,6 +271,20 @@ class InferShapeContext {
     return &var->Get<Tensor>();
   }
 
+  void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
+                size_t j = 0) const {
+    PADDLE_ENFORCE_LT(i, InputSize(in));
+    PADDLE_ENFORCE_LT(j, OutputSize(out));
+    auto* in_var = MultiInputVar(in)[i];
+    auto* out_var = MultiOutputVar(out)[j];
+    if (!in_var->IsType<LoDTensor>()) return;
+    PADDLE_ENFORCE(out_var->IsType<LoDTensor>(),
+                   "The %d-th output of Output(%s) must be LoDTensor.", j, out);
+    auto in_tensor = in_var->Get<LoDTensor>();
+    auto* out_tensor = out_var->GetMutable<LoDTensor>();
+    out_tensor->set_lod(in_tensor.lod());
+  }
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;
@@ -283,6 +297,13 @@ template <>
 const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
     const std::string& name) const;
 
+template <>
+Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const;
+
+template <>
+std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
+    const std::string& name) const;
+
 template <typename T>
 struct EigenDeviceConverter;
 
@@ -315,38 +336,10 @@ class ExecutionContext : public InferShapeContext {
     return device_context_;
   }
 
-  // redefine Output function,
-  // use Variable::Get instead of Variable::GetMutable
-  template <typename T>
-  T* Output(const std::string& name) const {
-    auto var = OutputVar(name);
-    return var == nullptr ? nullptr : const_cast<T*>(&var->Get<T>());
-  }
-
-  // redefine MultiOutput function.
-  // use Variable::Get instead of Variable::GetMutable
-  template <typename T>
-  std::vector<T*> MultiOutput(const std::string& name) const {
-    auto names = op().Outputs(name);
-    std::vector<T*> res;
-    res.reserve(names.size());
-    std::transform(
-        names.begin(), names.end(), std::back_inserter(res),
-        [&](const std::string& sub_name) { return Output<T>(sub_name); });
-    return res;
-  }
-
  private:
   const platform::DeviceContext& device_context_;
 };
 
-template <>
-Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
-
-template <>
-std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
-    const std::string& name) const;
-
 class OpKernel {
  public:
   /**

paddle/framework/tensor.h

@@ -165,12 +165,6 @@ class Tensor {
   /*! points to dimensions of memory block. */
   DDim dims_;
 
-  /**
-   * A cache of the number of elements in a tensor.
-   * Would be 0 for an uninitialized tensor.
-   */
-  int64_t numel_;
-
   /**
    * @brief   A PlaceHolder may be shared by more than one tensor.
    *

paddle/framework/tensor_impl.h

@@ -147,13 +147,12 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
 
 inline Tensor& Tensor::Resize(const DDim& dims) {
   dims_ = dims;
-  numel_ = product(dims_);
   return *this;
 }
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
-inline int64_t Tensor::numel() const { return numel_; }
+inline int64_t Tensor::numel() const { return product(dims_); }
 
 template <typename T>
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {

paddle/operators/accuracy_op.cc

@@ -39,7 +39,8 @@ class AccuracyOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0],
                       "inference size must be the same as label size");
 
-    ctx.Output<framework::LoDTensor>("Accuracy")->Resize({1});
+    ctx.Output<framework::Tensor>("Accuracy")->Resize({1});
+    ctx.ShareLoD("Inference", /*->*/ "Accuracy");
   }
 };
 
@@ -54,11 +55,15 @@ class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
     // TODO(typhoonzero): AddInput("Weight", ...
     AddOutput("Accuracy", "The accuracy of current batch");
 
-    AddComment(
-        R"DOC(Accuracy. It will print accuracy rate for classification.
+    AddComment(R"DOC(
+Accuracy. It will print accuracy rate for classification.
 The accuracy is:
 ..  math::
-accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
+accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})
+
+Both the input `Inference` and `Label` can carry the LoD (Level of Details)
+information, or not. But the output only shares the LoD with input `Inference`.
+)DOC");
   }
 };
 

paddle/operators/activation_op.cc

@@ -23,8 +23,9 @@ class ActivationOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::LoDTensor>("Y")->Resize(
+    ctx.Output<framework::Tensor>("Y")->Resize(
         ctx.Input<framework::Tensor>("X")->dims());
+    ctx.ShareLoD("X", /*->*/ "Y");
   }
 };
 
@@ -34,7 +35,7 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
+    ctx.Output<framework::Tensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<framework::Tensor>("Y")->dims());
   }
 };

paddle/operators/add_op.cc

@@ -33,7 +33,7 @@ class AddOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
                       ctx.Input<Tensor>("Y")->dims(),
                       "Two input of Add Op's dimension must be same.");
-    ctx.Output<framework::LoDTensor>("Out")->Resize(
+    ctx.Output<framework::Tensor>("Out")->Resize(
         ctx.Input<Tensor>("X")->dims());
   }
 };

paddle/operators/clip_op.cc

@@ -17,8 +17,6 @@
 namespace paddle {
 namespace operators {
 
-using framework::LoDTensor;
-
 class ClipOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -29,11 +27,12 @@ class ClipOp : public framework::OperatorWithKernel {
                             "Input(X) of ClipOp should not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
                             "Output(Out) of ClipOp should not be null.");
-    auto x_dims = ctx.Input<LoDTensor>("X")->dims();
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto max = Attr<float>("max");
     auto min = Attr<float>("min");
     PADDLE_ENFORCE_LT(min, max, "max should be greater than min.");
-    ctx.Output<LoDTensor>("Out")->Resize(x_dims);
+    ctx.Output<Tensor>("Out")->Resize(x_dims);
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -66,8 +65,8 @@ class ClipOpGrad : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) should not be null");
-    auto x_dims = ctx.Input<LoDTensor>("X")->dims();
-    auto *x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
     if (x_grad != nullptr) {
       x_grad->Resize(x_dims);
     }

paddle/operators/concat_op.cc

@@ -29,7 +29,7 @@ class ConcatOp : public framework::OperatorWithKernel {
                             "Output(Out) of ConcatOp should not be null.");
 
     auto ins = ctx.MultiInput<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    auto *out = ctx.Output<framework::Tensor>("Out");
     size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
     size_t n = ins.size();
 

paddle/operators/conv2d_op.cc

@@ -37,7 +37,7 @@ class Conv2DOp : public framework::OperatorWithKernel {
 
     auto in = ctx.Input<Tensor>("Input");
     auto filter = ctx.Input<Tensor>("Filter");
-    auto out = ctx.Output<framework::LoDTensor>("Output");
+    auto out = ctx.Output<framework::Tensor>("Output");
     std::vector<int> strides = Attr<std::vector<int>>("strides");
     std::vector<int> paddings = Attr<std::vector<int>>("paddings");
     int groups = Attr<int>("groups");
@@ -111,10 +111,9 @@ class Conv2DOpGrad : public framework::OperatorWithKernel {
   void InferShape(const framework::InferShapeContext &ctx) const override {
     auto in = ctx.Input<Tensor>("Input");
     auto filter = ctx.Input<Tensor>("Filter");
-    auto d_in =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Input"));
+    auto d_in = ctx.Output<framework::Tensor>(framework::GradVarName("Input"));
     auto d_filter =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Filter"));
+        ctx.Output<framework::Tensor>(framework::GradVarName("Filter"));
     if (d_in) d_in->Resize(in->dims());
     if (d_filter) d_filter->Resize(filter->dims());
   }

paddle/operators/cos_sim_op.cc

@@ -54,9 +54,10 @@ class CosSimOp : public framework::OperatorWithKernel {
                    " just 1 (which will be broadcasted to match Input(X)).");
 
     // resize tensor
-    ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
-    ctx.Output<framework::LoDTensor>("XNorm")->Resize({x_dims[0], 1});
-    ctx.Output<framework::LoDTensor>("YNorm")->Resize({y_dims[0], 1});
+    ctx.Output<framework::Tensor>("Out")->Resize({x_dims[0], 1});
+    ctx.Output<framework::Tensor>("XNorm")->Resize({x_dims[0], 1});
+    ctx.Output<framework::Tensor>("YNorm")->Resize({y_dims[0], 1});
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -81,10 +82,13 @@ Cosine Similarity Operator.
 
 The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
 
-Input(X) and Input(Y) must have the same shape, except that the 1st dimension
-of Input(Y) could be just 1 (different from Input(X)), which will be
-broadcasted to match the shape of Input(X) before computing their cosine
+The input `X` and `Y` must have the same shape, except that the 1st dimension
+of input `Y` could be just 1 (different from input `X`), which will be
+broadcasted to match the shape of input `X` before computing their cosine
 similarity.
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };
@@ -139,10 +143,8 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
                       "Shape of Input(Out@Grad) must be [X.Dim(0), 1].");
 
     // resize tensor
-    auto *x_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
-    auto *y_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+    auto *x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto *y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);
   }

paddle/operators/crop_op.cc

@@ -19,7 +19,6 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
-using framework::LoDTensor;
 
 class CropOp : public framework::OperatorWithKernel {
  public:
@@ -31,9 +30,9 @@ class CropOp : public framework::OperatorWithKernel {
                             "Input(X) of CropOp should not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
                             "Output(Out) of CropOp should not be null.");
-    auto x_dim = ctx.Input<LoDTensor>("X")->dims();
-    auto *y = ctx.Input<LoDTensor>("Y");
-    auto *out = ctx.Output<LoDTensor>("Out");
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto *y = ctx.Input<Tensor>("Y");
+    auto *out = ctx.Output<Tensor>("Out");
     if (y == nullptr) {
       auto shape = Attr<std::vector<int>>("shape");
       PADDLE_ENFORCE_EQ(
@@ -121,8 +120,8 @@ class CropOpGrad : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) should not be null");
-    auto x_dims = ctx.Input<LoDTensor>("X")->dims();
-    auto *x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
     if (x_grad != nullptr) {
       x_grad->Resize(x_dims);
     }

paddle/operators/cross_entropy_op.cc

@@ -17,8 +17,6 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-using framework::LoDTensor;
-
 class CrossEntropyOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -35,23 +33,21 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2.");
     PADDLE_ENFORCE_EQ(label->dims().size(), 2,
                       "Input(Label)'s rank must be 2.");
-    // TODO(xinghai-sun): remove this check after swtiching to bool
-    PADDLE_ENFORCE(ctx.Attr<int>("soft_label") == 0 ||
-                   ctx.Attr<int>("soft_label") == 1);
     PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
                       "The 1st dimension of Input(X) and Input(Label) must "
                       "be equal.");
-    if (ctx.Attr<int>("soft_label") == 1) {
+    if (ctx.Attr<bool>("soft_label")) {
       PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
-                        "If Attr(soft_label) == 1, The 2nd dimension of "
+                        "If Attr(soft_label) == true, The 2nd dimension of "
                         "Input(X) and Input(Label) must be equal.");
     } else {
       PADDLE_ENFORCE_EQ(label->dims()[1], 1,
-                        "If Attr(soft_label) == 0, The 2nd dimension of "
+                        "If Attr(soft_label) == false, The 2nd dimension of "
                         "Input(Label) must be 1.");
     }
 
-    ctx.Output<LoDTensor>("Y")->Resize({x->dims()[0], 1});
+    ctx.Output<Tensor>("Y")->Resize({x->dims()[0], 1});
+    ctx.ShareLoD("X", /*->*/ "Y");
   }
 };
 
@@ -74,9 +70,6 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(dy->dims().size(), 2, "Input(Y@Grad)'s rank must be 2.");
     PADDLE_ENFORCE_EQ(label->dims().size(), 2,
                       "Input(Label)'s rank must be 2.");
-    // TODO(xinghai-sun): remove this check after swtiching to bool
-    PADDLE_ENFORCE(ctx.Attr<int>("soft_label") == 0 ||
-                   ctx.Attr<int>("soft_label") == 1);
     PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
                       "The 1st dimension of Input(X) and Input(Label) must "
                       "be equal.");
@@ -85,17 +78,17 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
                       "be equal.");
     PADDLE_ENFORCE_EQ(dy->dims()[1], 1,
                       "The 2nd dimension of Input(Y@Grad) must be 1.");
-    if (ctx.Attr<int>("soft_label") == 1) {
+    if (ctx.Attr<bool>("soft_label")) {
       PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
-                        "If Attr(soft_label) == 1, The 2nd dimension of "
+                        "If Attr(soft_label) == true, The 2nd dimension of "
                         "Input(X) and Input(Label) must be equal.");
     } else {
       PADDLE_ENFORCE_EQ(label->dims()[1], 1,
-                        "If Attr(soft_label) == 0, The 2nd dimension of "
+                        "If Attr(soft_label) == false, The 2nd dimension of "
                         "Input(Label) must be 1.");
     }
 
-    auto dx = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     dx->Resize(x->dims());
   }
 };
@@ -108,7 +101,8 @@ class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "The first input of CrossEntropyOp");
     AddInput("Label", "The second input of CrossEntropyOp");
     AddOutput("Y", "The output of CrossEntropyOp");
-    AddAttr<int>("soft_label", "Is soft label. Default zero.").SetDefault(0);
+    AddAttr<bool>("soft_label", "Is soft label. Default zero.")
+        .SetDefault(false);
 
     AddComment(R"DOC(
 CrossEntropy Operator.
@@ -116,12 +110,12 @@ CrossEntropy Operator.
 It supports both standard cross-entropy and soft-label cross-entropy loss
 computation.
 1) One-hot cross-entropy:
-    soft_label = 0, Label[i, 0] indicates the class index for sample i:
+    soft_label = False, Label[i, 0] indicates the class index for sample i:
 
                 Y[i] = -log(X[i, Label[i]])
 
 2) Soft-label cross-entropy:
-    soft_label = 1, Label[i, j] indicates the soft label of class j
+    soft_label = True, Label[i, j] indicates the soft label of class j
     for sample i:
 
                 Y[i] = \sum_j{-Label[i, j] * log(X[i, j])}
@@ -133,6 +127,9 @@ computation.
      As a special case of 2), when each row of Input(Label) has only one
      non-zero element (equals 1), soft-label cross-entropy degenerates to a
      one-hot cross-entropy with one-hot label representation.
+
+Both the input `X` and `Label` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };

paddle/operators/cross_entropy_op.cu

@@ -102,7 +102,7 @@ class CrossEntropyOpCUDAKernel : public framework::OpKernel {
     int grid = (n + block - 1) / block;
     // TODO(qingqing) launch kernel on specified stream
     // base on ExecutionContext.
-    if (ctx.Attr<int>("soft_label") == 1) {
+    if (ctx.Attr<bool>("soft_label")) {
       auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
       SoftCrossEntropyKernel<T><<<grid, block>>>(y_data, x_data, label_data, n,
                                                  d);
@@ -137,7 +137,7 @@ class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel {
     grid = (n + block - 1) / block;
     // TODO(qingqing): launch kernel on specified stream
     // base on ExecutionContext.
-    if (ctx.Attr<int>("soft_label") == 1) {
+    if (ctx.Attr<bool>("soft_label")) {
       auto* label_data = label->data<T>();
       SoftCrossEntropyGradientKernel<T><<<grid, block>>>(
           dx_data, dy_data, x_data, label_data, n, d);

paddle/operators/cross_entropy_op.h

@@ -51,7 +51,7 @@ class CrossEntropyOpKernel : public framework::OpKernel {
     int batch_size = x->dims()[0];
     int class_num = x->dims()[1];
 
-    if (ctx.Attr<int>("soft_label") == 1) {
+    if (ctx.Attr<bool>("soft_label")) {
       auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
       int index = 0;
       for (int i = 0; i < batch_size; ++i) {
@@ -92,7 +92,7 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel {
     int class_num = x->dims()[1];
 
     // TODO(qingqing): make zero setting an common function.
-    if (ctx.Attr<int>("soft_label") == 1) {
+    if (ctx.Attr<bool>("soft_label")) {
       auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
       int index = 0;
       for (int i = 0; i < batch_size; ++i) {

paddle/operators/dropout_op.cc

@@ -18,7 +18,6 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
-using framework::LoDTensor;
 
 class DropoutOp : public framework::OperatorWithKernel {
  public:
@@ -29,15 +28,13 @@ class DropoutOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
     PADDLE_ENFORCE_GE(ctx.Attr<float>("dropout_prob"), 0);
     PADDLE_ENFORCE_LE(ctx.Attr<float>("dropout_prob"), 1);
-    // TODO(xinghai-sun): remove this check after swtiching to bool
-    PADDLE_ENFORCE(ctx.Attr<int>("is_training") == 0 ||
-                   ctx.Attr<int>("is_training") == 1);
 
     auto dims = ctx.Input<Tensor>("X")->dims();
-    ctx.Output<LoDTensor>("Out")->Resize(dims);
-    if (ctx.Attr<int>("is_training") == 1) {
-      ctx.Output<LoDTensor>("Mask")->Resize(dims);
+    ctx.Output<Tensor>("Out")->Resize(dims);
+    if (ctx.Attr<bool>("is_training")) {
+      ctx.Output<Tensor>("Mask")->Resize(dims);
     }
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -49,8 +46,7 @@ class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddAttr<AttrType>("dropout_prob", "Probability of setting units to zero.")
         .SetDefault(.5f);
-    // TODO(xinghai-sun): use bool for is_training after bool is supported.
-    AddAttr<int>("is_training", "Whether in training phase.").SetDefault(1);
+    AddAttr<bool>("is_training", "Whether in training phase.").SetDefault(true);
     AddAttr<int>("seed", "Dropout random seed.").SetDefault(0);
     AddInput("X", "The input of dropout op.");
     AddOutput("Out", "The output of dropout op.");
@@ -59,7 +55,7 @@ class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 Dropout Operator.
 
-"Dropout" refers to randomly dropping out units in a nerual network. It is a
+'Dropout' refers to randomly dropping out units in a nerual network. It is a
 regularization technique for reducing overfitting by preventing neuron
 co-adaption during training. The dropout operator randomly set (according to
 the given dropout probability) the outputs of some units to zero, while others
@@ -75,7 +71,7 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.Attr<int>("is_training"), 1,
+    PADDLE_ENFORCE(ctx.Attr<bool>("is_training"),
                    "GradOp is only callable when is_training is true");
 
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
@@ -85,9 +81,6 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
 
     PADDLE_ENFORCE_GE(ctx.Attr<AttrType>("dropout_prob"), 0);
     PADDLE_ENFORCE_LE(ctx.Attr<AttrType>("dropout_prob"), 1);
-    // TODO(xinghai-sun): remove this check after swtiching to bool
-    PADDLE_ENFORCE(ctx.Attr<int>("is_training") == 0 ||
-                   ctx.Attr<int>("is_training") == 1);
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
     PADDLE_ENFORCE_EQ(x_dims, out_dims,
@@ -96,7 +89,7 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(x_dims, mask_dims,
                       "Dimensions of Input(X) and Mask must be the same.");
 
-    auto *x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
     x_grad->Resize(x_dims);
   }
 };

paddle/operators/dropout_op.cu

@@ -59,7 +59,7 @@ class GPUDropoutKernel : public framework::OpKernel {
     auto Y = EigenMatrix<T>::Reshape(*y, 1);
 
     auto place = context.GetEigenDevice<Place>();
-    if (context.Attr<int>("is_training") == 1) {
+    if (context.Attr<bool>("is_training")) {
       auto* mask = context.Output<Tensor>("Mask");
       auto* mask_data = mask->mutable_data<T>(context.GetPlace());
       int size = framework::product(mask->dims());

paddle/operators/dropout_op.h

@@ -35,7 +35,7 @@ class CPUDropoutKernel : public framework::OpKernel {
     auto* y_data = y->mutable_data<T>(context.GetPlace());
     AttrType dropout_prob = context.Attr<AttrType>("dropout_prob");
 
-    if (context.Attr<int>("is_training") == 1) {
+    if (context.Attr<bool>("is_training")) {
       auto* mask = context.Output<Tensor>("Mask");
       auto* mask_data = mask->mutable_data<T>(context.GetPlace());
       int seed = context.Attr<int>("seed");
@@ -65,7 +65,7 @@ template <typename Place, typename T>
 class DropoutGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    PADDLE_ENFORCE_EQ(context.Attr<int>("is_training"), 1,
+    PADDLE_ENFORCE(context.Attr<bool>("is_training"),
                    "GradOp is only callable when is_training is true");
 
     auto* grad_x = context.Output<Tensor>(framework::GradVarName("X"));

paddle/operators/elementwise_add_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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 "paddle/operators/elementwise_add_op.h"
+
+namespace paddle {
+namespace operators {
+class ElementwiseAddOpMaker : public ElementwiseOpMaker {
+ public:
+  ElementwiseAddOpMaker(framework::OpProto* proto,
+                        framework::OpAttrChecker* op_checker)
+      : ElementwiseOpMaker(proto, op_checker) {
+    SetComment("add", "Out = X + Y");
+    AddComment(comment_);
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(elementwise_add, ops::ElementwiseOp, ops::ElementwiseAddOpMaker,
+            elementwise_add_grad, ops::ElementwiseOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_add,
+    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_add_grad,
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/elementwise_add_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/elementwise_add_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    elementwise_add,
+    ops::ElementwiseAddKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    elementwise_add_grad,
+    ops::ElementwiseAddGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/elementwise_add_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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 "paddle/operators/elementwise_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ElementwiseAddKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseCompute<EigenAddFunctor, Place, T>(ctx);
+  }
+};
+
+template <typename T>
+struct ElementwiseAddGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = dz_e;
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseAddOneGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = dz_e.sum();
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseAddBroadCastGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = dz_e.reshape(Eigen::DSizes<int, 2>(pre, n))
+                           .sum(Eigen::array<int, 1>{{0}});
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseAddBroadCast2GradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N, typename Post>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
+                  Post post) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = dz_e.reshape(Eigen::DSizes<int, 3>(pre, n, post))
+                           .sum(Eigen::array<int, 2>{{0, 2}});
+    }
+  }
+};
+
+template <typename Place, typename T>
+class ElementwiseAddGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseGradCompute<Place, T, ElementwiseAddGradFunctor<T>,
+                           ElementwiseAddOneGradFunctor<T>,
+                           ElementwiseAddBroadCastGradFunctor<T>,
+                           ElementwiseAddBroadCast2GradFunctor<T>>(ctx);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/elementwise_div_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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 "paddle/operators/elementwise_div_op.h"
+
+namespace paddle {
+namespace operators {
+class ElementwiseDivOpMaker : public ElementwiseOpMaker {
+ public:
+  ElementwiseDivOpMaker(framework::OpProto* proto,
+                        framework::OpAttrChecker* op_checker)
+      : ElementwiseOpMaker(proto, op_checker) {
+    SetComment("Div", "Out = X / Y");
+    AddComment(comment_);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(elementwise_div, ops::ElementwiseOp, ops::ElementwiseDivOpMaker,
+            elementwise_div_grad, ops::ElementwiseOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_div,
+    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_div_grad,
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/elementwise_div_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/elementwise_div_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    elementwise_div,
+    ops::ElementwiseDivKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    elementwise_div_grad,
+    ops::ElementwiseDivGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/elementwise_div_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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 "paddle/operators/elementwise_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ElementwiseDivKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseCompute<EigenDivFunctor, Place, T>(ctx);
+  }
+};
+
+template <typename T>
+struct ElementwiseDivGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto z_e = framework::EigenVector<T>::Flatten(*z);
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e / y_e;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = -1.0 * dz_e * z_e / y_e;
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseDivBroadCastGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+
+    auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e / y_e_bcast;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
+                           .reshape(Eigen::DSizes<int, 2>(pre, n))
+                           .sum(Eigen::array<int, 1>{{0}});
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseDivBroadCast2GradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N, typename Post>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
+                  Post post) {
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+
+    auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e / y_e_bcast;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = (-1.0 * (x_e * dz_e) / (y_e_bcast * y_e_bcast))
+                           .reshape(Eigen::DSizes<int, 3>(pre, n, post))
+                           .sum(Eigen::array<int, 2>{{0, 2}});
+    }
+  }
+};
+
+template <typename Place, typename T>
+class ElementwiseDivGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseGradCompute<Place, T, ElementwiseDivGradFunctor<T>,
+                           ElementwiseDivGradFunctor<T>,
+                           ElementwiseDivBroadCastGradFunctor<T>,
+                           ElementwiseDivBroadCast2GradFunctor<T>>(ctx);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/elementwise_mul_op.cc

@@ -17,101 +17,25 @@
 namespace paddle {
 namespace operators {
 
-using Tensor = framework::Tensor;
-
-class ElementWiseMulOp : public framework::OperatorWithKernel {
+class ElementwiseMulOpMaker : public ElementwiseOpMaker {
  public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of ElementWiseMulOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Input(Y) of ElementWiseMulOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.OutputVar("Out"),
-        "Output(Out) of ElementWiseMulOp should not be null.");
-
-    auto x_dim = ctx.Input<Tensor>("X")->dims();
-    auto y_dim = ctx.Input<Tensor>("Y")->dims();
-    PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
-                      "Rank of first input must >= rank of second input.")
-    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dim);
+  ElementwiseMulOpMaker(framework::OpProto* proto,
+                        framework::OpAttrChecker* op_checker)
+      : ElementwiseOpMaker(proto, op_checker) {
+    SetComment("Mul", "Out = X ⊙ Y");
+    AddComment(comment_);
   }
 };
 
-class ElementWiseMulOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  ElementWiseMulOpMaker(framework::OpProto *proto,
-                        framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The first input of elementwise mul op");
-    AddInput("Y", "The second input of elementwise mul op");
-    AddAttr<int>("axis",
-                 R"DOC(
-When shape(Y) does not equal shape(X),Y will be broadcasted 
-to match the shape of X and axis should be dimension index Y in X
-        )DOC")
-        .SetDefault(-1)
-        .EqualGreaterThan(-1);
-
-    AddOutput("Out", "The output of elementwise mul op");
-    AddComment(R"DOC(
-Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
-1. The shape of Y should be same with X or
-2. Y's shape is a subset of X. 
-   Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
-   example:
-      shape(X) = (2, 3, 4, 5), shape(Y) = (,)
-      shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
-      shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
-      shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
-      shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
-)DOC");
-  }
-};
-
-class ElementWiseMulOpGrad : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Input(Out@GRAD) should not be null");
-
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto y_dims = ctx.Input<Tensor>("Y")->dims();
-    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    auto *x_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
-    auto *y_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
-
-    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
-                      "Rank of first input must >= rank of second input.")
-
-    if (x_grad) {
-      x_grad->Resize(x_dims);
-    }
-
-    if (y_grad) {
-      y_grad->Resize(y_dims);
-    }
-  }
-};
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(elementwise_mul, ops::ElementWiseMulOp, ops::ElementWiseMulOpMaker,
-            elementwise_mul_grad, ops::ElementWiseMulOpGrad);
+REGISTER_OP(elementwise_mul, ops::ElementwiseOp, ops::ElementwiseMulOpMaker,
+            elementwise_mul_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul,
-    ops::ElementWiseMulKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul_grad,
-    ops::ElementWiseMulGradKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/elementwise_mul_op.cu

@@ -19,7 +19,7 @@ namespace ops = paddle::operators;
 
 REGISTER_OP_GPU_KERNEL(
     elementwise_mul,
-    ops::ElementWiseMulKernel<paddle::platform::GPUPlace, float>);
+    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, float>);
 REGISTER_OP_GPU_KERNEL(
     elementwise_mul_grad,
-    ops::ElementWiseMulGradKernel<paddle::platform::GPUPlace, float>);
+    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/elementwise_mul_op.h

@@ -13,169 +13,102 @@
    limitations under the License. */
 
 #pragma once
-#include "paddle/framework/eigen.h"
-#include "paddle/framework/op_registry.h"
+#include "paddle/operators/elementwise_op.h"
 
 namespace paddle {
 namespace operators {
-/*
- * Out = X ⊙ Y
- * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
- *    pre=2, n=3*4, post=5
- * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)
- *    pre=2*3, n=4*5, post=1
- */
-
-inline void get_mid_dims(const framework::DDim& x_dims,
-                         const framework::DDim& y_dims, const int axis,
-                         int& pre, int& n, int& post) {
-  pre = 1;
-  n = 1;
-  post = 1;
-  for (int i = 0; i < axis; ++i) {
-    pre *= x_dims[i];
-  }
-
-  for (int i = 0; i < y_dims.size(); ++i) {
-    PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
-                      "Broadcast dimension mismatch.");
-    n *= y_dims[i];
-  }
-
-  for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
-    post *= x_dims[i];
-  }
-}
 
 template <typename Place, typename T>
-class ElementWiseMulKernel : public framework::OpKernel {
+class ElementwiseMulKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    using Tensor = framework::Tensor;
-
-    auto* x = ctx.Input<Tensor>("X");
-    auto* y = ctx.Input<Tensor>("Y");
-    auto* z = ctx.Output<Tensor>("Out");
-    z->mutable_data<T>(ctx.GetPlace());
-
-    auto x_e = framework::EigenVector<T>::Flatten(*x);
-    auto y_e = framework::EigenVector<T>::Flatten(*y);
-    auto z_e = framework::EigenVector<T>::Flatten(*z);
-
-    auto x_dims = x->dims();
-    auto y_dims = y->dims();
-    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
-                      "Rank of first input must >= rank of second input.")
-
-    if (x_dims == y_dims || product(y_dims) == 1) {
-      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_e;
-      return;
-    }
-
-    int axis = ctx.Attr<int>("axis");
-    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
-    PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
-                   "Axis should be in range [0, x_dims)");
-
-    int pre, n, post;
-    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
-    if (post == 1) {
-      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
-                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))
-                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
-      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
-      return;
-    } else {
-      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
-                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
-                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
-      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
-      return;
-    }
+    ElementwiseCompute<EigenMulFunctor, Place, T>(ctx);
   }
 };
 
-template <typename Place, typename T>
-class ElementWiseMulGradKernel : public framework::OpKernel {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    using Tensor = framework::Tensor;
-
-    auto* x = ctx.Input<Tensor>("X");
-    auto* y = ctx.Input<Tensor>("Y");
-    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
-
+template <typename T>
+struct ElementwiseMulGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
     auto x_e = framework::EigenVector<T>::Flatten(*x);
     auto y_e = framework::EigenVector<T>::Flatten(*y);
-    auto dout_e = framework::EigenVector<T>::Flatten(*dout);
-
-    auto x_dims = x->dims();
-    auto y_dims = y->dims();
-
-    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    if (dx) {
-      dx->mutable_data<T>(ctx.GetPlace());
-    }
-    if (dy) {
-      dy->mutable_data<T>(ctx.GetPlace());
-    }
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
 
-    if (x_dims == y_dims || product(y_dims) == 1) {
     if (dx) {
       auto dx_e = framework::EigenVector<T>::Flatten(*dx);
-        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e;
+      dx_e.device(d) = dz_e * y_e;
     }
 
     if (dy) {
       auto dy_e = framework::EigenVector<T>::Flatten(*dy);
-        dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;
+      dy_e.device(d) = x_e * dz_e;
     }
-      return;
   }
+};
 
-    int axis = ctx.Attr<int>("axis");
-    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
-
-    int pre, n, post;
-    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+template <typename T>
+struct ElementwiseMulBroadCastGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
 
-    // TODO(gongweibao): wrap reshape to a function.
-    if (post == 1) {
     auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
                          .broadcast(Eigen::DSizes<int, 2>(pre, 1))
                          .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+
     if (dx) {
       auto dx_e = framework::EigenVector<T>::Flatten(*dx);
-        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      dx_e.device(d) = dz_e * y_e_bcast;
     }
 
     if (dy) {
       auto dy_e = framework::EigenVector<T>::Flatten(*dy);
-        dy_e.device(ctx.GetEigenDevice<Place>()) =
-            (x_e * dout_e)
+      dy_e.device(d) = (x_e * dz_e)
                            .reshape(Eigen::DSizes<int, 2>(pre, n))
                            .sum(Eigen::array<int, 1>{{0}});
     }
-      return;
-    } else {
+  }
+};
+
+template <typename T>
+struct ElementwiseMulBroadCast2GradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N, typename Post>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
+                  Post post) {
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+
     auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
                          .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
                          .reshape(Eigen::DSizes<int, 1>(x_e.size()));
     if (dx) {
       auto dx_e = framework::EigenVector<T>::Flatten(*dx);
-        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      dx_e.device(d) = dz_e * y_e_bcast;
     }
 
     if (dy) {
       auto dy_e = framework::EigenVector<T>::Flatten(*dy);
-        dy_e.device(ctx.GetEigenDevice<Place>()) =
-            (x_e * dout_e)
+      dy_e.device(d) = (x_e * dz_e)
                            .reshape(Eigen::DSizes<int, 3>(pre, n, post))
                            .sum(Eigen::array<int, 2>{{0, 2}});
     }
-      return;
   }
+};
+
+template <typename Place, typename T>
+class ElementwiseMulGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseGradCompute<Place, T, ElementwiseMulGradFunctor<T>,
+                           ElementwiseMulGradFunctor<T>,
+                           ElementwiseMulBroadCastGradFunctor<T>,
+                           ElementwiseMulBroadCast2GradFunctor<T>>(ctx);
   }
 };
 

paddle/operators/elementwise_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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. */
+
+#pragma once
+#include <iostream>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+/*
+ * Out = X ⊙ Y
+ * If Y's shape does not match X' shape, they will be reshaped.
+ * For example:
+ * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+ *    pre=2, n=3*4, post=5
+ *    x.shape(2, 12, 5) * y.shape(1,12,1).broadcast(2,12,5)
+ * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)
+ *    pre=2*3, n=4*5, post=1
+ *    x.shape(2, 3, 20) * y.shape(1,1,20).broadcast(2,3,20)
+ */
+inline void get_mid_dims(const framework::DDim& x_dims,
+                         const framework::DDim& y_dims, const int axis,
+                         int& pre, int& n, int& post) {
+  pre = 1;
+  n = 1;
+  post = 1;
+  for (int i = 0; i < axis; ++i) {
+    pre *= x_dims[i];
+  }
+
+  for (int i = 0; i < y_dims.size(); ++i) {
+    PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
+                      "Broadcast dimension mismatch.");
+    n *= y_dims[i];
+  }
+
+  for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
+    post *= x_dims[i];
+  }
+}
+
+#define EIGEN_FUNCTOR(name, eigen_op)                                          \
+  struct Eigen##name##Functor {                                                \
+    template <typename Place, typename T>                                      \
+    inline void Run(const framework::Tensor* x, const framework::Tensor* y,    \
+                    framework::Tensor* z,                                      \
+                    const framework::ExecutionContext& ctx) {                  \
+      auto x_e = framework::EigenVector<T>::Flatten(*x);                       \
+      auto y_e = framework::EigenVector<T>::Flatten(*y);                       \
+      auto z_e = framework::EigenVector<T>::Flatten(*z);                       \
+      z_e.device(ctx.GetEigenDevice<Place>()) = eigen_op(x_e, y_e);            \
+    }                                                                          \
+    template <typename Place, typename T>                                      \
+    inline void RunBroadCast(const framework::Tensor* x,                       \
+                             const framework::Tensor* y, framework::Tensor* z, \
+                             const framework::ExecutionContext& ctx, int pre,  \
+                             int n) {                                          \
+      auto x_e = framework::EigenVector<T>::Flatten(*x);                       \
+      auto y_e = framework::EigenVector<T>::Flatten(*y);                       \
+      auto z_e = framework::EigenVector<T>::Flatten(*z);                       \
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))                  \
+                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))             \
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));          \
+      z_e.device(ctx.GetEigenDevice<Place>()) = eigen_op(x_e, y_bcast);        \
+    }                                                                          \
+    template <typename Place, typename T>                                      \
+    inline void RunBroadCast2(const framework::Tensor* x,                      \
+                              const framework::Tensor* y,                      \
+                              framework::Tensor* z,                            \
+                              const framework::ExecutionContext& ctx, int pre, \
+                              int n, int post) {                               \
+      auto x_e = framework::EigenVector<T>::Flatten(*x);                       \
+      auto y_e = framework::EigenVector<T>::Flatten(*y);                       \
+      auto z_e = framework::EigenVector<T>::Flatten(*z);                       \
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))               \
+                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))       \
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));          \
+      z_e.device(ctx.GetEigenDevice<Place>()) = eigen_op(x_e, y_bcast);        \
+    }                                                                          \
+  }
+
+template <class functor, typename Place, typename T>
+void ElementwiseCompute(const framework::ExecutionContext& ctx) {
+  using Tensor = framework::Tensor;
+
+  auto* x = ctx.Input<Tensor>("X");
+  auto* y = ctx.Input<Tensor>("Y");
+  auto* z = ctx.Output<Tensor>("Out");
+  z->mutable_data<T>(ctx.GetPlace());
+
+  auto x_dims = x->dims();
+  auto y_dims = y->dims();
+  PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                    "Rank of first input must >= rank of second input.")
+
+  if (x_dims == y_dims || product(y_dims) == 1) {
+    functor f;
+    f.template Run<Place, T>(x, y, z, ctx);
+    return;
+  }
+
+  int axis = ctx.Attr<int>("axis");
+  axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+  PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
+                 "Axis should be in range [0, x_dims)");
+
+  int pre, n, post;
+  get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+  if (post == 1) {
+    functor f;
+    f.template RunBroadCast<Place, T>(x, y, z, ctx, pre, n);
+    return;
+  } else {
+    functor f;
+    f.template RunBroadCast2<Place, T>(x, y, z, ctx, pre, n, post);
+    return;
+  }
+}
+
+#define EIGEN_ADD(x, y) ((x) + (y))
+EIGEN_FUNCTOR(Add, EIGEN_ADD);
+
+#define EIGEN_SUB(x, y) ((x) - (y))
+EIGEN_FUNCTOR(Sub, EIGEN_SUB);
+
+#define EIGEN_MUL(x, y) ((x) * (y))
+EIGEN_FUNCTOR(Mul, EIGEN_MUL);
+
+#define EIGEN_DIV(x, y) ((x) / (y))
+EIGEN_FUNCTOR(Div, EIGEN_DIV);
+
+template <typename Place, typename T, typename functor, typename functor1,
+          typename broadcastfunctor, typename broadcast2functor>
+void ElementwiseGradCompute(const framework::ExecutionContext& ctx) {
+  using Tensor = framework::Tensor;
+
+  auto* x = ctx.Input<Tensor>("X");
+  auto* y = ctx.Input<Tensor>("Y");
+  auto* out = ctx.Input<Tensor>("Out");
+  auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+  auto place = ctx.GetEigenDevice<Place>();
+
+  auto x_dims = x->dims();
+  auto y_dims = y->dims();
+
+  auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+  auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+  if (dx) {
+    dx->mutable_data<T>(ctx.GetPlace());
+  }
+  if (dy) {
+    dy->mutable_data<T>(ctx.GetPlace());
+  }
+
+  if (x_dims == y_dims) {
+    functor f;
+    f(place, x, y, out, dx, dy, dout);
+    return;
+  }
+
+  if (product(y_dims) == 1) {
+    functor1 f;
+    f(place, x, y, out, dx, dy, dout);
+    return;
+  }
+
+  int axis = ctx.Attr<int>("axis");
+  axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+
+  int pre, n, post;
+  get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+
+  if (post == 1) {
+    broadcastfunctor f;
+    f(place, x, y, out, dx, dy, dout, pre, n);
+    return;
+  } else {
+    broadcast2functor f;
+    f(place, x, y, out, dx, dy, dout, pre, n, post);
+    return;
+  }
+}
+
+class ElementwiseOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  using Tensor = framework::Tensor;
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of elementwise op should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of elementwise op should not be null");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of elementwise op should not be null.");
+
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto y_dim = ctx.Input<Tensor>("Y")->dims();
+    PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
+                      "Rank of first input must >= rank of second input.")
+    ctx.Output<framework::Tensor>("Out")->Resize(x_dim);
+    ctx.ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ElementwiseOpMaker(framework::OpProto* proto,
+                     framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", R"DOC(
+The first input of elementwise op, it's a tensor of any dimensions.
+)DOC");
+    AddInput("Y", R"DOC(
+The sencond input of elementwise op, it's a tensor and it's dimensions
+must be small or equal to X's dimensions.
+)DOC");
+    AddAttr<int>("axis",
+                 R"DOC(
+When the shape(Y) does not equal the shape(X),Y will be broadcasted 
+to match the shape of X and axis should be dimension index Y in X
+        )DOC")
+        .SetDefault(-1)
+        .EqualGreaterThan(-1);
+
+    AddOutput("Out", "The output of elementwise op");
+    comment_ = R"DOC(
+Limited elementwise {name} operator.The equation is: Out = {equation}.
+1. The shape of Y should be same with X or
+2. Y's shape is a subset of X. 
+   Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
+
+   example:
+      shape(X) = (2, 3, 4, 5), shape(Y) = (,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+      shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+
+Both the input X and Y can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input X.
+)DOC";
+    AddComment(comment_);
+  }
+
+ protected:
+  std::string comment_;
+
+  void Replace(std::string& src, std::string from, std::string to) {
+    std::size_t len_from = std::strlen(from.c_str());
+    std::size_t len_to = std::strlen(to.c_str());
+    for (std::size_t pos = src.find(from); pos != std::string::npos;
+         pos = src.find(from, pos + len_to)) {
+      src.replace(pos, len_from, to);
+    }
+  }
+
+  void SetComment(std::string name, std::string equation) {
+    Replace(comment_, "{name}", name);
+    Replace(comment_, "{equation}", equation);
+  }
+};
+
+class ElementwiseOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  using Tensor = framework::Tensor;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.")
+
+    if (x_grad) {
+      x_grad->Resize(x_dims);
+    }
+
+    if (y_grad) {
+      y_grad->Resize(y_dims);
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle

paddle/operators/elementwise_sub_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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 "paddle/operators/elementwise_sub_op.h"
+
+namespace paddle {
+namespace operators {
+class ElementwiseSubOpMaker : public ElementwiseOpMaker {
+ public:
+  ElementwiseSubOpMaker(framework::OpProto* proto,
+                        framework::OpAttrChecker* op_checker)
+      : ElementwiseOpMaker(proto, op_checker) {
+    SetComment("Sub", "Out = X - Y");
+    AddComment(comment_);
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(elementwise_sub, ops::ElementwiseOp, ops::ElementwiseSubOpMaker,
+            elementwise_sub_grad, ops::ElementwiseOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_sub,
+    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_sub_grad,
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/elementwise_sub_op.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/elementwise_sub_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    elementwise_sub,
+    ops::ElementwiseSubKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    elementwise_sub_grad,
+    ops::ElementwiseSubGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/elementwise_sub_op.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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 "paddle/operators/elementwise_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ElementwiseSubKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseCompute<EigenSubFunctor, Place, T>(ctx);
+  }
+};
+
+template <typename T>
+struct ElementwiseSubGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = (-1.0) * dz_e;
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseSubOneGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = (-1.0) * dz_e.sum();
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseSubBroadCastGradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = (-1.0) *
+                       dz_e.reshape(Eigen::DSizes<int, 2>(pre, n))
+                           .sum(Eigen::array<int, 1>{{0}});
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseSubBroadCast2GradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N, typename Post>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
+                  Post post) {
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = dz_e;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = (-1.0) *
+                       dz_e.reshape(Eigen::DSizes<int, 3>(pre, n, post))
+                           .sum(Eigen::array<int, 2>{{0, 2}});
+    }
+  }
+};
+
+template <typename Place, typename T>
+class ElementwiseSubGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseGradCompute<Place, T, ElementwiseSubGradFunctor<T>,
+                           ElementwiseSubOneGradFunctor<T>,
+                           ElementwiseSubBroadCastGradFunctor<T>,
+                           ElementwiseSubBroadCast2GradFunctor<T>>(ctx);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/fc_op.cc

@@ -186,6 +186,9 @@ W_i is a 2-D matrix of size (K x N), where N means the number of neurons
 in the fully connected layer. B is a 1-D vector of size N.
 Thus, the output Out is a 2-D matrix of size (M x N).
 Activation type can be set to `identity` (default), `sigmoid` or `softmax`.
+
+All the inputs can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with first input (`X[0]`).
 )DOC");
   }
 };

paddle/operators/fill_zeros_like_op.cc

@@ -23,15 +23,14 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.InputVar("Src"),
-        "Input(Src) of FillZerosLikeOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.OutputVar("Dst"),
-        "Output(Dst) of FillZerosLikeOp should not be null.");
-
-    ctx.Output<framework::LoDTensor>("Dst")->Resize(
-        ctx.Input<framework::Tensor>("Src")->dims());
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of FillZerosLikeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of FillZerosLikeOp should not be null.");
+
+    ctx.Output<framework::Tensor>("Y")->Resize(
+        ctx.Input<framework::Tensor>("X")->dims());
+    ctx.ShareLoD("X", /*->*/ "Y");
   }
 };
 
@@ -40,8 +39,8 @@ class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker {
   FillZerosLikeOpMaker(framework::OpProto *proto,
                        framework::OpAttrChecker *op_checker)
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("Src", "The input of fill-zeros-like op.");
-    AddOutput("Dst", "The varibale will be filled up with zeros.");
+    AddInput("X", "The input of fill-zeros-like op.");
+    AddOutput("Y", "The varibale will be filled up with zeros.");
     AddComment(R"DOC(
 Fill up a vriable with zeros.
 

paddle/operators/fill_zeros_like_op.h

@@ -23,7 +23,7 @@ template <typename Place, typename T>
 class FillZerosLikeKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* output = context.Output<framework::Tensor>("Dst");
+    auto* output = context.Output<framework::Tensor>("Y");
     output->mutable_data<T>(context.GetPlace());
     auto t = framework::EigenVector<T>::Flatten(*output);
     t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));

paddle/operators/gather_op.cc

@@ -35,7 +35,7 @@ class GatherOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
     framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
     output_dims[0] = batch_size;
-    ctx.Output<framework::LoDTensor>("Out")->Resize(output_dims);
+    ctx.Output<framework::Tensor>("Out")->Resize(output_dims);
   }
 };
 
@@ -45,7 +45,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto X_grad = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto X_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
     auto X = ctx.Input<Tensor>("X");
 
     X_grad->Resize(X->dims());

paddle/operators/gaussian_random_op.cc

@@ -48,7 +48,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
         ctx.OutputVar("Out"),
         "Output(Out) of GaussianRandomOp should not be null.");
 
-    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
+    auto* tensor = ctx.Output<framework::Tensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());

paddle/operators/lookup_table_op.cc

@@ -32,9 +32,10 @@ class LookupTableOp : public framework::OperatorWithKernel {
 
     auto table_t = ctx.Input<Tensor>("W");
     auto ids_t = ctx.Input<Tensor>("Ids");
-    auto output_t = ctx.Output<framework::LoDTensor>("Out");
+    auto output_t = ctx.Output<framework::Tensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
+    ctx.ShareLoD("Ids", /*->*/ "Out");
   }
 };
 
@@ -50,9 +51,13 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
              "An input with type int32 or int64"
              "contains the ids to be looked up in W.");
     AddOutput("Out", "The lookup results, which have the same type with W.");
-    AddComment(
-        "This operator is used to perform lookups on the parameter W,"
-        "then concatenated into a dense tensor.");
+    AddComment(R"DOC(
+This operator is used to perform lookups on the parameter W,
+then concatenated into a dense tensor.
+
+The input `Ids` can carry the LoD (Level of Details) information,
+or not. And the output only shares the LoD with input `Ids`.
+)DOC");
   }
 };
 
@@ -64,7 +69,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
   void InferShape(const framework::InferShapeContext &context) const override {
     auto table = context.Input<Tensor>("W");
     auto d_table =
-        context.Output<framework::LoDTensor>(framework::GradVarName("W"));
+        context.Output<framework::Tensor>(framework::GradVarName("W"));
     d_table->Resize(table->dims());
   }
 };

paddle/operators/mean_op.cc

@@ -27,7 +27,7 @@ class MeanOp : public framework::OperatorWithKernel {
                             "Input(X) of MeanOp should not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
                             "Output(Out) of MeanOp should not be null.");
-    ctx.Output<framework::LoDTensor>("Out")->Resize({1});
+    ctx.Output<framework::Tensor>("Out")->Resize({1});
   }
 };
 
@@ -37,7 +37,8 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
     AddOutput("Out", "The output of mean op").NotInGradient();
-    AddComment("Mean Operator");
+    AddComment(R"DOC( Mean Operator
+)DOC");
   }
 };
 
@@ -47,7 +48,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
+    ctx.Output<framework::Tensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };

paddle/operators/minus_op.cc

@@ -40,7 +40,8 @@ class MinusOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         left_tensor->numel(), right_tensor->numel(),
         "Minus operator must take two tensor with same num of elements");
-    ctx.Output<framework::LoDTensor>("Out")->Resize(left_tensor->dims());
+    ctx.Output<framework::Tensor>("Out")->Resize(left_tensor->dims());
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -54,7 +55,12 @@ class MinusOpMaker : public framework::OpProtoAndCheckerMaker {
 
     AddComment(R"DOC(Minus Operator
 
-Equation: Out = X - Y
+Equation:
+
+    Out = X - Y
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };

paddle/operators/modified_huber_loss_op.cc

@@ -34,8 +34,8 @@ class ModifiedHuberLossOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(x->dims().size(), 2, "The tensor rank of X must be 2.");
     PADDLE_ENFORCE_EQ(x->dims()[1], 1, "The 2nd dimension of X must be 1.");
 
-    context.Output<framework::LoDTensor>("IntermediateVal")->Resize(x->dims());
-    context.Output<framework::LoDTensor>("Out")->Resize({x->dims()[0], 1});
+    context.Output<framework::Tensor>("IntermediateVal")->Resize(x->dims());
+    context.Output<framework::Tensor>("Out")->Resize({x->dims()[0], 1});
   }
 };
 
@@ -81,7 +81,7 @@ class ModifiedHuberLossGradOp : public framework::OperatorWithKernel {
     auto* intermediate_val = context.Input<Tensor>("IntermediateVal");
     auto* out_grad = context.Input<Tensor>(framework::GradVarName("Out"));
     auto* x_grad =
-        context.Output<framework::LoDTensor>(framework::GradVarName("X"));
+        context.Output<framework::Tensor>(framework::GradVarName("X"));
 
     PADDLE_ENFORCE_NOT_NULL(x, "X must be initialized.");
     PADDLE_ENFORCE_NOT_NULL(y, "Y must be initialized.");

paddle/operators/modified_huber_loss_op.h

@@ -52,8 +52,8 @@ class ModifiedHuberLossKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* in0 = context.Input<Tensor>("X");
     auto* in1 = context.Input<Tensor>("Y");
-    auto* out0 = context.Output<framework::LoDTensor>("IntermediateVal");
-    auto* out1 = context.Output<framework::LoDTensor>("Out");
+    auto* out0 = context.Output<framework::Tensor>("IntermediateVal");
+    auto* out1 = context.Output<framework::Tensor>("Out");
 
     out0->mutable_data<T>(context.GetPlace());
     out1->mutable_data<T>(context.GetPlace());
@@ -77,11 +77,9 @@ class ModifiedHuberLossGradCPUKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* in0 = context.Input<Tensor>("Y");
-    auto* in1 = context.Input<framework::LoDTensor>("IntermediateVal");
-    auto* in2 =
-        context.Input<framework::LoDTensor>(framework::GradVarName("Out"));
-    auto* out0 =
-        context.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto* in1 = context.Input<framework::Tensor>("IntermediateVal");
+    auto* in2 = context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* out0 = context.Output<framework::Tensor>(framework::GradVarName("X"));
 
     if (out0) {
       const T* y_ptr = in0->data<T>();

paddle/operators/mul_op.cc

@@ -18,7 +18,6 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
-using framework::LoDTensor;
 
 class MulOp : public framework::OperatorWithKernel {
  public:
@@ -53,8 +52,9 @@ class MulOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         x_mat_dims[1], y_mat_dims[0],
         "First matrix's width must be equal with second matrix's height.");
-    ctx.Output<framework::LoDTensor>("Out")->Resize(
+    ctx.Output<framework::Tensor>("Out")->Resize(
         {x_mat_dims[0], y_mat_dims[1]});
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -83,9 +83,14 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker {
         .SetDefault(1)
         .EqualGreaterThan(1);
     AddComment(R"DOC(
-Two Element Mul Operator.
+Mul operator is used to perform matrix multiplication for input X and Y.
 
-The equation is: Out = X * Y
+The equation is:
+
+    Out = X * Y
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };
@@ -103,10 +108,8 @@ class MulOpGrad : public framework::OperatorWithKernel {
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    auto *x_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
-    auto *y_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+    auto *x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto *y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
 
     auto x_mat_dims =
         framework::flatten_to_2d(x_dims, Attr<int>("x_num_col_dims"));

paddle/operators/pad_op.cc

@@ -39,8 +39,13 @@ class PadOp : public framework::OperatorWithKernel {
     for (int i = 0; i < x_dim.size(); ++i) {
       out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
     }
-    ctx.Output<framework::LoDTensor>("Out")->Resize(
+    ctx.Output<framework::Tensor>("Out")->Resize(
         framework::make_ddim(out_dims));
+    if (out_dims[0] == x_dim[0]) {
+      // Only pass LoD when the first dimension is equal between
+      // output and input.
+      ctx.ShareLoD("X", /*->*/ "Out");
+    }
   }
 };
 
@@ -101,7 +106,7 @@ class PadOpGrad : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) should not be null");
     auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto *x_g = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *x_g = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
     if (x_g != nullptr) {
       x_g->Resize(x_dims);
     }

paddle/operators/prelu_op.cc

@@ -36,8 +36,9 @@ class PReluOp : public framework::OperatorWithKernel {
 
     PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
                             "Output(Out) should not be null");
-    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    auto *out = ctx.Output<framework::Tensor>("Out");
     out->Resize(in->dims());
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -55,6 +56,8 @@ The equation is:
   f(x) = alpha * x , for x < 0
   f(x) = x         , for x >= 0
 
+The input `X` can carry the LoD (Level of Details) information,
+or not. And the output shares the LoD with input `X`.
 )DOC");
   }
 };
@@ -69,11 +72,11 @@ class PReluGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) should not be null");
-    auto *dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
     auto *x = ctx.Input<framework::Tensor>("X");
 
     auto *dalpha =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Alpha"));
+        ctx.Output<framework::Tensor>(framework::GradVarName("Alpha"));
     auto *alpha = ctx.Input<framework::Tensor>("Alpha");
 
     dx->Resize(x->dims());

paddle/operators/rank_loss_op.cc

@@ -40,7 +40,7 @@ class RankLossOp : public framework::OperatorWithKernel {
                    "All inputs must have the same size");
     PADDLE_ENFORCE((label_dims.size() == 2) && (label_dims[1] == 1),
                    "All inputs must be row vector with size batch_size x 1.");
-    ctx.Output<framework::LoDTensor>("Out")->Resize(label_dims);
+    ctx.Output<framework::Tensor>("Out")->Resize(label_dims);
   }
 };
 
@@ -102,9 +102,9 @@ class RankLossGradOp : public framework::OperatorWithKernel {
                             "Input(Out@GRAD) shouldn't be null.");
     auto dims = ctx.Input<framework::Tensor>("Left")->dims();
     auto *left_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Left"));
+        ctx.Output<framework::Tensor>(framework::GradVarName("Left"));
     auto *right_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Right"));
+        ctx.Output<framework::Tensor>(framework::GradVarName("Right"));
     if (left_grad) {
       left_grad->Resize(dims);
     }

paddle/operators/rank_loss_op.h

@@ -24,7 +24,7 @@ template <typename Place, typename T>
 class RankLossKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
-    auto* out_t = ctx.Output<framework::LoDTensor>("Out");
+    auto* out_t = ctx.Output<framework::Tensor>("Out");
     auto* label_t = ctx.Input<framework::Tensor>("Label");
     auto* left_t = ctx.Input<framework::Tensor>("Left");
     auto* right_t = ctx.Input<framework::Tensor>("Right");
@@ -46,9 +46,9 @@ class RankLossGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
     auto* d_left_t =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Left"));
+        ctx.Output<framework::Tensor>(framework::GradVarName("Left"));
     auto* d_right_t =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Right"));
+        ctx.Output<framework::Tensor>(framework::GradVarName("Right"));
 
     auto* d_out_t = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
     auto* label_t = ctx.Input<framework::Tensor>("Label");

paddle/operators/reshape_op.cc

@@ -50,7 +50,12 @@ class ReshapeOp : public framework::OperatorWithKernel {
     std::transform(shape.begin(), shape.end(), shape_int64.begin(),
                    [](int a) { return static_cast<int64_t>(a); });
     auto out_dims = framework::make_ddim(shape_int64);
-    ctx.Output<framework::LoDTensor>("Out")->Resize(out_dims);
+    ctx.Output<framework::Tensor>("Out")->Resize(out_dims);
+    if (shape[0] == in->dims()[0]) {
+      // Only pass LoD when the first dimension is equal between
+      // output and input.
+      ctx.ShareLoD("X", /*->*/ "Out");
+    }
   }
 };
 
@@ -94,7 +99,7 @@ class ReshapeGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) shouldn't be null.");
     auto dims = ctx.Input<framework::Tensor>("X")->dims();
-    auto *d_in = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *d_in = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
     d_in->Resize(dims);
   }
 };

paddle/operators/rowwise_add_op.cc

@@ -44,7 +44,8 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
     PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1");
-    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dims);
+    ctx.Output<framework::Tensor>("Out")->Resize(x_dims);
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -83,8 +84,8 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
-    auto *dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
-    auto *db = ctx.Output<framework::LoDTensor>(framework::GradVarName("b"));
+    auto *dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto *db = ctx.Output<framework::Tensor>(framework::GradVarName("b"));
     if (dx) dx->Resize(x_dims);
     if (db) db->Resize(b_dims);
   }

paddle/operators/scale_op.cc

@@ -33,8 +33,9 @@ class ScaleOp : public framework::OperatorWithKernel {
                             "Output(Out) of ScaleOp should not be null.");
 
     auto *in = ctx.Input<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    auto *out = ctx.Output<framework::Tensor>("Out");
     out->Resize(in->dims());
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 

paddle/operators/scatter_op.cc

@@ -44,7 +44,7 @@ class ScatterOp : public framework::OperatorWithKernel {
     framework::DDim data_dim(ctx.Input<Tensor>("Updates")->dims());
     for (int i = 1; i < data_dim.size(); ++i)
       PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input<Tensor>("Updates")->dims()[i]);
-    ctx.Output<framework::LoDTensor>("Out")->Resize(
+    ctx.Output<framework::Tensor>("Out")->Resize(
         ctx.Input<Tensor>("Ref")->dims());
   }
 };
@@ -56,10 +56,9 @@ class ScatterGradOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     auto *dUpdates =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Updates"));
+        ctx.Output<framework::Tensor>(framework::GradVarName("Updates"));
     auto *Updates = ctx.Input<Tensor>("Updates");
-    auto *dRef =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Ref"));
+    auto *dRef = ctx.Output<framework::Tensor>(framework::GradVarName("Ref"));
     auto *Ref = ctx.Input<Tensor>("Ref");
 
     dRef->Resize(Ref->dims());

paddle/operators/sgd_op.cc

@@ -33,7 +33,7 @@ class SGDOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("param")->dims(),
                       ctx.Input<Tensor>("grad")->dims(),
                       "Two input of SGD Op's dimension must be same.");
-    ctx.Output<framework::LoDTensor>("param_out")
+    ctx.Output<framework::Tensor>("param_out")
         ->Resize(ctx.Input<Tensor>("param")->dims());
   }
 };

paddle/operators/smooth_l1_loss_op.cc

@@ -44,8 +44,8 @@ class SmoothL1LossOp : public framework::OperatorWithKernel {
                         "The shape of OutsideWeight must be same as X.");
     }
 
-    auto* diff = ctx.Output<framework::LoDTensor>("Diff");
-    auto* out = ctx.Output<framework::LoDTensor>("Out");
+    auto* diff = ctx.Output<framework::Tensor>("Diff");
+    auto* out = ctx.Output<framework::Tensor>("Out");
     diff->Resize(x->dims());
     // loss is a two-rank tensor
     out->Resize({x->dims()[0], 1});
@@ -103,10 +103,8 @@ class SmoothL1LossGradOp : public framework::OperatorWithKernel {
     auto in_dims = ctx.Input<framework::Tensor>("X")->dims();
     auto out_dims =
         ctx.Input<framework::Tensor>(framework::GradVarName("Out"))->dims();
-    auto* x_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
-    auto* y_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
 
     PADDLE_ENFORCE_GE(out_dims.size(), 2,
                       "The tensor rank of Input(Out@Grad) should be 2.");

paddle/operators/softmax_op.cc

@@ -30,8 +30,7 @@ class SoftmaxOp : public framework::OperatorWithKernel {
 
     PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be a matrix.");
-    ctx.Output<framework::LoDTensor>("Y")->Resize(
-        ctx.Input<Tensor>("X")->dims());
+    ctx.Output<framework::Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -77,7 +76,7 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
                       ctx.Input<Tensor>(framework::GradVarName("Y"))->dims(),
                       "Input(Y) and its gradients should have a same shape.");
 
-    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
+    ctx.Output<framework::Tensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };

paddle/operators/split_op.cc

@@ -27,7 +27,7 @@ class SplitOp : public framework::OperatorWithKernel {
   void InferShape(const framework::InferShapeContext &ctx) const override {
     // infershape
     auto *in = ctx.Input<framework::Tensor>("X");
-    auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
+    auto outs = ctx.MultiOutput<framework::Tensor>("Out");
     size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
     size_t num = static_cast<size_t>(ctx.Attr<int>("num"));
     std::vector<int> sections =

paddle/operators/squared_l2_distance_op.cc

@@ -54,9 +54,10 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
                    "First dimension of target must be equal to input "
                    "or to 1.");
 
-    ctx.Output<framework::LoDTensor>("sub_result")
+    ctx.Output<framework::Tensor>("sub_result")
         ->Resize({x_dims[0], x->numel() / x_dims[0]});
-    ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
+    ctx.Output<framework::Tensor>("Out")->Resize({x_dims[0], 1});
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -79,6 +80,9 @@ class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker {
     input or to 1. If the first dimension of target is 1, SquaredL2DistanceOp
     will broadcast target's first dimension to input's first dimension.
     You can decide whether calculate the gradient of input and target.
+
+    Both the input X and Y can carry the LoD (Level of Details) information,
+    or not. But the output only shares the LoD with input X.
     )DOC");
   }
 };
@@ -100,10 +104,8 @@ class SquaredL2DistanceGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(out_dims[1], 1,
                       "Second dimension of output gradient "
                       "must be 1.");
-    auto* x_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
-    auto* y_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);
   }

paddle/operators/sum_op.cc

@@ -28,7 +28,7 @@ class SumOp : public framework::OperatorWithKernel {
                             "Output(Out) of SumOp should not be null.");
 
     auto ins = ctx.MultiInput<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    auto *out = ctx.Output<framework::Tensor>("Out");
     int N = ins.size();
 
     auto in_dim = ins[0]->dims();
@@ -39,6 +39,7 @@ class SumOp : public framework::OperatorWithKernel {
       PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape");
     }
     out->Resize(in_dim);
+    ctx.ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -49,8 +50,11 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "the input tensors of sum operator.").AsDuplicable();
     AddOutput("Out", "the output tensor of sum operator.");
     AddComment(R"DOC(
-            Sum the input tensors.
-        )DOC");
+Sum the input tensors.
+
+All the inputs can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with the first input.
+)DOC");
   }
 };
 
@@ -61,7 +65,7 @@ class SumGradOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     auto outputs =
-        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
+        ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
     auto dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
     for (auto output : outputs) {
       output->Resize(dims);

paddle/operators/top_k_op.cc

@@ -40,8 +40,8 @@ class TopkOp : public framework::OperatorWithKernel {
 
     framework::DDim dims = input->dims();
     dims[dims.size() - 1] = k;
-    ctx.Output<framework::LoDTensor>("Out")->Resize(dims);
-    ctx.Output<framework::LoDTensor>("Indices")->Resize(dims);
+    ctx.Output<framework::Tensor>("Out")->Resize(dims);
+    ctx.Output<framework::Tensor>("Indices")->Resize(dims);
   }
 };
 

paddle/operators/transpose_op.cc

@@ -51,7 +51,7 @@ class TransposeOp : public framework::OperatorWithKernel {
     for (size_t i = 0; i < axis_size; i++) {
       out_dims[i] = x_dims[axis[i]];
     }
-    ctx.Output<framework::LoDTensor>("Out")->Resize(out_dims);
+    ctx.Output<framework::Tensor>("Out")->Resize(out_dims);
   }
 };
 
@@ -99,8 +99,7 @@ class TransposeOpGrad : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) should not be null");
     auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto *x_grad =
-        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
 
     if (x_grad) x_grad->Resize(x_dims);
   }

paddle/operators/uniform_random_op.cc

@@ -54,7 +54,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 
     PADDLE_ENFORCE(Attr<float>("min") < Attr<float>("max"),
                    "uniform_random's min must less then max");
-    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
+    auto* tensor = ctx.Output<framework::Tensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());

python/paddle/v2/framework/op.py

@@ -89,12 +89,16 @@ class OpDescCreationMethod(object):
                     new_attr.f = user_defined_attr
                 elif attr.type == framework_pb2.STRING:
                     new_attr.s = user_defined_attr
+                elif attr.type == framework_pb2.BOOLEAN:
+                    new_attr.b = user_defined_attr
                 elif attr.type == framework_pb2.INTS:
                     new_attr.ints.extend(user_defined_attr)
                 elif attr.type == framework_pb2.FLOATS:
                     new_attr.floats.extend(user_defined_attr)
                 elif attr.type == framework_pb2.STRINGS:
                     new_attr.strings.extend(user_defined_attr)
+                elif attr.type == framework_pb2.BOOLEANS:
+                    new_attr.bools.extend(user_defined_attr)
                 elif attr.type == framework_pb2.INT_PAIRS:
                     for p in user_defined_attr:
                         pair = new_attr.int_pairs.add()

python/paddle/v2/framework/tests/test_cos_sim_op.py

@@ -24,15 +24,15 @@ class TestCosSimOp(OpTest):
         self.check_output()
 
     def test_check_grad_normal(self):
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05)
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.06)
 
     def test_check_grad_ingore_x(self):
         self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set("X"))
+            ['Y'], 'Out', max_relative_error=0.06, no_grad_set=set("X"))
 
     def test_check_grad_ingore_y(self):
         self.check_grad(
-            ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
+            ['X'], 'Out', max_relative_error=0.06, no_grad_set=set('Y'))
 
 
 class TestCosSimOp2(TestCosSimOp):

python/paddle/v2/framework/tests/test_cross_entropy_op.py

@@ -19,7 +19,7 @@ class TestCrossEntropyOp1(OpTest):
             dtype="float32")
         self.inputs = {"X": X, "Label": label}
         self.outputs = {"Y": cross_entropy}
-        self.attrs = {'soft_label': 0}
+        self.attrs = {'soft_label': False}
 
     def test_check_output(self):
         self.check_output()
@@ -45,7 +45,7 @@ class TestCrossEntropyOp2(OpTest):
             axis=1, keepdims=True).astype("float32")
         self.inputs = {'X': X, 'Label': label}
         self.outputs = {'Y': cross_entropy}
-        self.attrs = {'soft_label': 1}
+        self.attrs = {'soft_label': True}
 
     def test_check_output(self):
         self.check_output()
@@ -76,7 +76,7 @@ class TestCrossEntropyOp3(OpTest):
             axis=1, keepdims=True).astype("float32")
         self.inputs = {'X': X, 'Label': label}
         self.outputs = {'Y': cross_entropy}
-        self.attrs = {'soft_label': 1}
+        self.attrs = {'soft_label': True}
 
     def test_check_output(self):
         self.check_output()

python/paddle/v2/framework/tests/test_dropout_op.py

@@ -7,7 +7,7 @@ class TestDropoutOp(OpTest):
     def setUp(self):
         self.op_type = "dropout"
         self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
-        self.attrs = {'dropout_prob': 0.0, 'is_training': 1}
+        self.attrs = {'dropout_prob': 0.0, 'is_training': True}
         self.outputs = {'Out': self.inputs['X'], 'Mask': np.ones((32, 64))}
 
     def test_check_output(self):
@@ -21,7 +21,7 @@ class TestDropoutOp2(TestDropoutOp):
     def setUp(self):
         self.op_type = "dropout"
         self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
-        self.attrs = {'dropout_prob': 1.0, 'is_training': 1}
+        self.attrs = {'dropout_prob': 1.0, 'is_training': True}
         self.outputs = {'Out': np.zeros((32, 64)), 'Mask': np.zeros((32, 64))}
 
 
@@ -29,7 +29,7 @@ class TestDropoutOp3(TestDropoutOp):
     def setUp(self):
         self.op_type = "dropout"
         self.inputs = {'X': np.random.random((32, 64, 2)).astype("float32")}
-        self.attrs = {'dropout_prob': 0.0, 'is_training': 1}
+        self.attrs = {'dropout_prob': 0.0, 'is_training': True}
         self.outputs = {'Out': self.inputs['X'], 'Mask': np.ones((32, 64, 2))}
 
 
@@ -37,7 +37,7 @@ class TestDropoutOp4(OpTest):
     def setUp(self):
         self.op_type = "dropout"
         self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
-        self.attrs = {'dropout_prob': 0.35, 'is_training': 0}
+        self.attrs = {'dropout_prob': 0.35, 'is_training': False}
         self.outputs = {'Out': self.inputs['X'] * self.attrs['dropout_prob']}
 
     def test_check_output(self):
@@ -48,7 +48,7 @@ class TestDropoutOp5(OpTest):
     def setUp(self):
         self.op_type = "dropout"
         self.inputs = {'X': np.random.random((32, 64, 3)).astype("float32")}
-        self.attrs = {'dropout_prob': 0.75, 'is_training': 0}
+        self.attrs = {'dropout_prob': 0.75, 'is_training': False}
         self.outputs = {'Out': self.inputs['X'] * self.attrs['dropout_prob']}
 
     def test_check_output(self):

python/paddle/v2/framework/tests/test_elementwise_add_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestElementwiseOp(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_add"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+        }
+        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['Y'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
+
+
+class TestElementwiseAddOp_Vector(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_add"
+        self.inputs = {
+            'X': np.random.random((32, )).astype("float32"),
+            'Y': np.random.random((32, )).astype("float32")
+        }
+        self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['Y'])}
+
+
+class TestElementwiseAddOp_broadcast_0(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_add"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(2).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 0}
+        self.outputs = {
+            'Out': self.inputs['X'] + self.inputs['Y'].reshape(2, 1, 1)
+        }
+
+
+class TestElementwiseAddOp_broadcast_1(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_add"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(3).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 1)
+        }
+
+
+class TestElementwiseAddOp_broadcast_2(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_add"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(4).astype(np.float32)
+        }
+
+        self.outputs = {
+            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 1, 4)
+        }
+
+
+class TestElementwiseAddOp_broadcast_3(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_add"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
+            'Y': np.random.rand(3, 4).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] + self.inputs['Y'].reshape(1, 3, 4, 1)
+        }
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_elementwise_div_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class ElementwiseDivOp(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        """ Warning
+        CPU gradient check error!
+        'X': np.random.random((32,84)).astype("float32"),
+        'Y': np.random.random((32,84)).astype("float32")
+        """
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+        }
+        self.outputs = {'Out': np.divide(self.inputs['X'], self.inputs['Y'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.05)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
+
+
+class TestElementwiseDivOp_Vector(ElementwiseDivOp):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [32]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [32]).astype("float32")
+        }
+        self.outputs = {'Out': np.divide(self.inputs['X'], self.inputs['Y'])}
+
+
+class TestElementwiseDivOp_broadcast_0(ElementwiseDivOp):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [2]).astype("float32")
+        }
+
+        self.attrs = {'axis': 0}
+        self.outputs = {
+            'Out':
+            np.divide(self.inputs['X'], self.inputs['Y'].reshape(2, 1, 1))
+        }
+
+
+class TestElementwiseDivOp_broadcast_1(ElementwiseDivOp):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [3]).astype("float32")
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out':
+            np.divide(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 1))
+        }
+
+
+class TestElementwiseDivOp_broadcast_2(ElementwiseDivOp):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [4]).astype("float32")
+        }
+
+        self.outputs = {
+            'Out':
+            np.divide(self.inputs['X'], self.inputs['Y'].reshape(1, 1, 4))
+        }
+
+
+class TestElementwiseDivOp_broadcast_3(ElementwiseDivOp):
+    def setUp(self):
+        self.op_type = "elementwise_div"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [2, 3, 4, 5]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [3, 4]).astype("float32")
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out':
+            np.divide(self.inputs['X'], self.inputs['Y'].reshape(1, 3, 4, 1))
+        }
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_elementwise_mul_op.py

@@ -3,14 +3,9 @@ import numpy as np
 from op_test import OpTest
 
 
-class TestElementwiseMulOp_Matrix(OpTest):
+class ElementwiseMulOp(OpTest):
     def setUp(self):
         self.op_type = "elementwise_mul"
-        """ Warning
-        CPU gradient check error!
-        'X': np.random.random((32,84)).astype("float32"),
-        'Y': np.random.random((32,84)).astype("float32")
-        """
         self.inputs = {
             'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
             'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
@@ -32,7 +27,7 @@ class TestElementwiseMulOp_Matrix(OpTest):
             ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
 
 
-class TestElementwiseMulOp_Vector(OpTest):
+class TestElementwiseMulOp_Vector(ElementwiseMulOp):
     def setUp(self):
         self.op_type = "elementwise_mul"
         self.inputs = {
@@ -41,22 +36,8 @@ class TestElementwiseMulOp_Vector(OpTest):
         }
         self.outputs = {'Out': np.multiply(self.inputs['X'], self.inputs['Y'])}
 
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad_normal(self):
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
-
 
-class TestElementwiseMulOp_broadcast_0(OpTest):
+class TestElementwiseMulOp_broadcast_0(ElementwiseMulOp):
     def setUp(self):
         self.op_type = "elementwise_mul"
         self.inputs = {
@@ -69,22 +50,8 @@ class TestElementwiseMulOp_broadcast_0(OpTest):
             'Out': self.inputs['X'] * self.inputs['Y'].reshape(2, 1, 1)
         }
 
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad_normal(self):
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
-
 
-class TestElementwiseMulOp_broadcast_1(OpTest):
+class TestElementwiseMulOp_broadcast_1(ElementwiseMulOp):
     def setUp(self):
         self.op_type = "elementwise_mul"
         self.inputs = {
@@ -97,22 +64,8 @@ class TestElementwiseMulOp_broadcast_1(OpTest):
             'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 3, 1)
         }
 
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad_normal(self):
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
-
 
-class TestElementwiseMulOp_broadcast_2(OpTest):
+class TestElementwiseMulOp_broadcast_2(ElementwiseMulOp):
     def setUp(self):
         self.op_type = "elementwise_mul"
         self.inputs = {
@@ -124,22 +77,8 @@ class TestElementwiseMulOp_broadcast_2(OpTest):
             'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 1, 4)
         }
 
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad_normal(self):
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
-
 
-class TestElementwiseMulOp_broadcast_3(OpTest):
+class TestElementwiseMulOp_broadcast_3(ElementwiseMulOp):
     def setUp(self):
         self.op_type = "elementwise_mul"
         self.inputs = {

python/paddle/v2/framework/tests/test_elementwise_sub_op.py

+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestElementwiseOp(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_sub"
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+        }
+        self.outputs = {'Out': self.inputs['X'] - self.inputs['Y']}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.005)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.005, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.005, no_grad_set=set('Y'))
+
+
+class TestElementwiseSubOp_Vector(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_sub"
+        self.inputs = {
+            'X': np.random.random((32, )).astype("float32"),
+            'Y': np.random.random((32, )).astype("float32")
+        }
+        self.outputs = {'Out': self.inputs['X'] - self.inputs['Y']}
+
+
+class TestElementwiseSubOp_broadcast_0(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_sub"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(2).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 0}
+        self.outputs = {
+            'Out': self.inputs['X'] - self.inputs['Y'].reshape(2, 1, 1)
+        }
+
+
+class TestElementwiseSubOp_broadcast_1(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_sub"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(3).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] - self.inputs['Y'].reshape(1, 3, 1)
+        }
+
+
+class TestElementwiseSubOp_broadcast_2(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_sub"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(4).astype(np.float32)
+        }
+
+        self.outputs = {
+            'Out': self.inputs['X'] - self.inputs['Y'].reshape(1, 1, 4)
+        }
+
+
+class TestElementwiseSubOp_broadcast_3(TestElementwiseOp):
+    def setUp(self):
+        self.op_type = "elementwise_sub"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
+            'Y': np.random.rand(3, 4).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] - self.inputs['Y'].reshape(1, 3, 4, 1)
+        }
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_fill_zeros_like_op.py

@@ -6,8 +6,8 @@ from op_test import OpTest
 class TestFillZerosLikeOp(OpTest):
     def setUp(self):
         self.op_type = "fill_zeros_like"
-        self.inputs = {'Src': np.random.random((219, 232)).astype("float32")}
-        self.outputs = {'Dst': np.zeros_like(self.inputs["Src"])}
+        self.inputs = {'X': np.random.random((219, 232)).astype("float32")}
+        self.outputs = {'Y': np.zeros_like(self.inputs["X"])}
 
     def test_check_output(self):
         self.check_output()