Merge remote-tracking branch 'upstream/develop' into remove-flag

Dockerfile

@@ -71,20 +71,6 @@ RUN pip install -r /root/requirements.txt
 RUN apt-get install -y libssl-dev libffi-dev
 RUN pip install certifi urllib3[secure]
 
-# TODO(qijun) The template library Eigen doesn't work well with GCC 5 
-# coming with the default Docker image, so we switch to use GCC 4.8 
-# by default. And I will check Eigen library later.
-
-RUN ln -sf gcc-4.8 /usr/bin/gcc && \
-    ln -sf gcc-ar-4.8 /usr/bin/gcc-ar && \
-    ln -sf gcc-nm-4.8 /usr/bin/gcc-nm && \
-    ln -sf gcc-ranlib-4.8 /usr/bin/gcc-ranlib && \
-    ln -sf gcc-4.8 /usr/bin/x86_64-linux-gnu-gcc && \
-    ln -sf gcc-ar-4.8 /usr/bin/x86_64-linux-gnu-gcc-ar && \
-    ln -sf gcc-nm-4.8 /usr/bin/x86_64-linux-gnu-gcc-nm && \
-    ln -sf gcc-ranlib-4.8 /usr/bin/x86_64-linux-gnu-gcc-ranlib && \
-    ln -sf g++-4.8 /usr/bin/g++ && \
-    ln -sf g++-4.8 /usr/bin/x86_64-linux-gnu-g++ 
 
 # Install woboq_codebrowser to /woboq
 RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \

cmake/flags.cmake

@@ -9,13 +9,6 @@ function(CheckCompilerCXX11Flag)
         if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
             message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
         endif()
-        if(NOT ANDROID)
-            # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
-            # Use Debug mode instead for now.
-            if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9)
-                set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
-            endif()
-        endif()
     elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
         # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"
         # Apple Clang is a different compiler than upstream Clang which havs different version numbers.
@@ -160,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
 
 # Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
 # So, don't set these flags here.
-LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread)
+LIST(APPEND CUDA_NVCC_FLAGS -std=c++11)
 LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math)
 
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")

doc/design/mkldnn/README.MD

@@ -101,6 +101,7 @@ if use_mkldnn
 5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
 6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
 7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
+8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
 
 ## References
 

doc/getstarted/build_and_install/build_from_source_en.md

@@ -68,7 +68,7 @@ As a simple example, consider the following:
 
 1. **BLAS Dependencies(optional)**
   
-    CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically.
+    CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically.
     To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
 
     ```bash
@@ -131,9 +131,9 @@ As a simple example, consider the following:
     To build GPU version, you will need the following installed:
 
         1. a CUDA-capable GPU
-        2. A supported version of Linux with a gcc compiler and toolchain
+        2. A supported version of Linux with a GCC compiler and toolchain
         3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
 
     The CUDA development environment relies on tight integration with the host development environment,
     including the host compiler and C runtime libraries, and is therefore only supported on
@@ -172,6 +172,7 @@ export PATH=<path to install>/bin:$PATH
 # install PaddlePaddle Python modules.
 sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
 ```
+
 ## <span id="centos">Build on Centos 7</span>
 
 ### Install Dependencies
@@ -192,9 +193,9 @@ sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
     To build GPU version, you will need the following installed:
 
         1. a CUDA-capable GPU
-        2. A supported version of Linux with a gcc compiler and toolchain
+        2. A supported version of Linux with a GCC compiler and toolchain
         3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
 
     The CUDA development environment relies on tight integration with the host development environment,
     including the host compiler and C runtime libraries, and is therefore only supported on
@@ -222,7 +223,7 @@ mkdir build && cd build
 ``` 
 
 Finally, you can build and install PaddlePaddle:
-
+  
 ```bash
 # you can add build option here, such as:    
 cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>

paddle/capi/gradient_machine.cpp

@@ -146,3 +146,19 @@ paddle_error paddle_gradient_machine_randomize_param(
   m->machine->randParameters();
   return kPD_NO_ERROR;
 }
+
+paddle_error paddle_gradient_machine_get_layer_output(
+    paddle_gradient_machine machine,
+    const char* layerName,
+    paddle_arguments args) {
+  auto m = cast(machine);
+  auto out = paddle::capi::cast<paddle::capi::CArguments>(args);
+  if (m == nullptr || layerName == nullptr || out == nullptr ||
+      m->machine == nullptr) {
+    return kPD_NULLPTR;
+  }
+
+  auto layerOutput = m->machine->getLayerOutput(layerName);
+  out->args.push_back(layerOutput);
+  return kPD_NO_ERROR;
+}

paddle/capi/gradient_machine.h

@@ -39,7 +39,11 @@ PD_API paddle_error paddle_gradient_machine_create_for_inference(
 /**
  * @brief Create a gradient machine used for model inference, using config with
  *        parameters which is generated by `paddle merge_model`.
- * @param [out] machine that used for model inference.
+ *        Example:
+ *          paddle merge_model \
+ *                 --model_dir="pass-00000" \
+ *                 --model_file="merged_model.paddle"
+ * @param [out] machine that used for model inference
  * @param [in] mergedModel
  * @param [in] size
  * @return paddle_error
@@ -97,6 +101,18 @@ paddle_gradient_machine_randomize_param(paddle_gradient_machine machine);
 PD_API paddle_error
 paddle_gradient_machine_destroy(paddle_gradient_machine machine);
 
+/**
+ * @brief Get the output of the layer named `layerName`.
+ * @param [in] gradient machine that have run a inference
+ * @param [in] layerName name of specified layer
+ * @param [out] args output of the specified layer
+ * @return paddle_error
+ */
+PD_API paddle_error
+paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine,
+                                         const char* layerName,
+                                         paddle_arguments args);
+
 #ifdef __cplusplus
 }
 #endif

paddle/framework/op_registry.cc

@@ -17,5 +17,49 @@ limitations under the License. */
 #include <vector>
 
 namespace paddle {
-namespace framework {}  // namespace framework
+namespace framework {
+
+std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
+                                                   const VarNameMap& inputs,
+                                                   const VarNameMap& outputs,
+                                                   AttributeMap attrs) {
+  auto it = op_info_map().find(type);
+  PADDLE_ENFORCE(it != op_info_map().end(),
+                 "Operator '%s' has not been registered.", type);
+  it->second.checker_->Check(attrs);
+  auto op = it->second.creator_(type, inputs, outputs, attrs);
+  return std::shared_ptr<OperatorBase>(op);
+}
+
+std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+  VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
+  VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
+  AttributeMap attrs;
+  for (auto& attr : op_desc.attrs()) {
+    attrs[attr.name()] = GetAttrValue(attr);
+  }
+
+  return CreateOp(op_desc.type(), inputs, outputs, attrs);
+}
+
+OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
+    const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
+  VarNameMap ret_val;
+  for (auto& var : op_desc_vars) {
+    auto& var_names = ret_val[var.parameter()];
+    auto& var_names_in_proto = var.arguments();
+    var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
+    std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
+              std::back_inserter(var_names));
+  }
+  return ret_val;
+}
+
+std::shared_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
+  PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
+  std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
+  return grad_op;
+}
+
+}  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry.h

@@ -29,103 +29,6 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-// this class not only make proto but also init attribute checkers.
-class OpProtoAndCheckerMaker {
- public:
-  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : proto_(proto), op_checker_(op_checker) {}
-
-  ~OpProtoAndCheckerMaker() {
-    PADDLE_ENFORCE(validated_, "should call Validate after build");
-  }
-
-  void Validate() {
-    validated_ = true;
-    CheckNoDuplicatedInOutAttrs();
-  }
-
- protected:
-  struct VariableBuilder {
-    OpProto::Var* var_;
-
-    VariableBuilder& AsDuplicable() {
-      var_->set_duplicable(true);
-      return *this;
-    }
-
-    VariableBuilder& AsIntermediate() {
-      var_->set_intermediate(true);
-      return *this;
-    }
-
-    // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
-    // means that input/output is not needed when calculate gradient. It does
-    // not mean no gradient when backward. It should be changed soon.
-    VariableBuilder& AsNoGradient() {
-      var_->set_no_gradient(true);
-      return *this;
-    }
-  };
-
-  VariableBuilder AddInput(const std::string& name,
-                           const std::string& comment) {
-    auto* input = proto_->add_inputs();
-    input->set_name(name);
-    input->set_comment(comment);
-    return VariableBuilder{input};
-  }
-
-  VariableBuilder AddOutput(const std::string& name,
-                            const std::string& comment) {
-    auto* output = proto_->add_outputs();
-    output->set_name(name);
-    output->set_comment(comment);
-    return VariableBuilder{output};
-  }
-
-  template <typename T>
-  TypedAttrChecker<T>& AddAttr(const std::string& name,
-                               const std::string& comment,
-                               bool generated = false) {
-    auto* attr = proto_->add_attrs();
-    attr->set_name(name);
-    attr->set_comment(comment);
-    attr->set_generated(generated);
-    attr->set_type(AttrTypeID<T>());
-    return op_checker_->AddAttrChecker<T>(name);
-  }
-
-  void AddComment(const std::string& comment) { proto_->set_comment(comment); }
-
- private:
-  void CheckNoDuplicatedInOutAttrs() {
-    std::unordered_set<std::string> names;
-    auto checker = [&](const std::string& name) {
-      PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
-      names.insert(name);
-    };
-    for (auto& attr : proto_->attrs()) {
-      checker(attr.name());
-    }
-    for (auto& input : proto_->inputs()) {
-      checker(input.name());
-    }
-    for (auto& output : proto_->outputs()) {
-      checker(output.name());
-    }
-  }
-
-  OpProto* proto_;
-  OpAttrChecker* op_checker_;
-  bool validated_{false};
-};
-
-class NOPMaker : public OpProtoAndCheckerMaker {
- public:
-  NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {}
-};
-
 class OpRegistry {
   using VarNameMap = OperatorBase::VarNameMap;
   using OpCreator = std::function<OperatorBase*(
@@ -177,45 +80,14 @@ class OpRegistry {
   static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
                                                 const VarNameMap& inputs,
                                                 const VarNameMap& outputs,
-                                                AttributeMap attrs) {
-    auto it = op_info_map().find(type);
-    PADDLE_ENFORCE(it != op_info_map().end(),
-                   "Operator '%s' has not been registered.", type);
-    it->second.checker_->Check(attrs);
-    auto op = it->second.creator_(type, inputs, outputs, attrs);
-    return std::shared_ptr<OperatorBase>(op);
-  }
-
-  static VarNameMap ConvertOpDescVarsToVarNameMap(
-      const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
-    VarNameMap ret_val;
-    for (auto& var : op_desc_vars) {
-      auto& var_names = ret_val[var.parameter()];
-      auto& var_names_in_proto = var.arguments();
-      var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
-      std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
-                std::back_inserter(var_names));
-    }
-    return ret_val;
-  }
+                                                AttributeMap attrs);
 
-  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
-    VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
-    VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
-    AttributeMap attrs;
-    for (auto& attr : op_desc.attrs()) {
-      attrs[attr.name()] = GetAttrValue(attr);
-    }
+  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
-    return CreateOp(op_desc.type(), inputs, outputs, attrs);
-  }
+  static VarNameMap ConvertOpDescVarsToVarNameMap(
+      const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars);
 
-  static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
-    PADDLE_ENFORCE(!op.IsNetOp(),
-                   "Use framework::Backward to get backward ops");
-    std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
-    return grad_op;
-  }
+  static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
 
   static std::unordered_map<std::string, const OpInfo>& op_info_map() {
     static std::unordered_map<std::string, const OpInfo> op_info_map_;
@@ -272,8 +144,18 @@ class OpKernelRegistrar : public Registrar {
                     grad_op_class)                                            \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
       __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
-  static ::paddle::framework::OpRegistrar<op_class, op_maker_class,           \
-                                          grad_op_class>                      \
+  class _OpClass_##op_type##_ : public op_class {                             \
+   public:                                                                    \
+    DEFINE_OP_CLONE_METHOD(_OpClass_##op_type##_);                            \
+    DEFINE_OP_CONSTRUCTOR(_OpClass_##op_type##_, op_class);                   \
+  };                                                                          \
+  class _OpGradClass_##op_type##_ : public grad_op_class {                    \
+   public:                                                                    \
+    DEFINE_OP_CLONE_METHOD(_OpGradClass_##op_type##_);                        \
+    DEFINE_OP_CONSTRUCTOR(_OpGradClass_##op_type##_, grad_op_class);          \
+  };                                                                          \
+  static ::paddle::framework::OpRegistrar<                                    \
+      _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_>       \
       __op_registrar_##op_type##__(#op_type, #grad_op_type);                  \
   int TouchOpRegistrar_##op_type() {                                          \
     __op_registrar_##op_type##__.Touch();                                     \
@@ -304,7 +186,8 @@ class OpKernelRegistrar : public Registrar {
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
 /**
- * Macro to mark what Operator and Kernel we will use and tell the compiler to
+ * Macro to mark what Operator and Kernel
+ * we will use and tell the compiler to
  * link them into target.
  */
 #define USE_OP_ITSELF(op_type)                                    \
@@ -324,7 +207,8 @@ class OpKernelRegistrar : public Registrar {
       __attribute__((unused)) =                                  \
           TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE()
 
-// TODO(fengjiayi): The following macros seems ugly, do we have better method?
+// TODO(fengjiayi): The following macros
+// seems ugly, do we have better method?
 
 #ifdef PADDLE_ONLY_CPU
 #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)

paddle/framework/operator.cc

@@ -164,5 +164,43 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   return ret_val;
 }
 
+void OpProtoAndCheckerMaker::Validate() {
+  validated_ = true;
+  CheckNoDuplicatedInOutAttrs();
+}
+
+OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput(
+    const std::string& name, const std::string& comment) {
+  auto* input = proto_->add_inputs();
+  input->set_name(name);
+  input->set_comment(comment);
+  return OpProtoAndCheckerMaker::VariableBuilder{input};
+}
+
+OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput(
+    const std::string& name, const std::string& comment) {
+  auto* output = proto_->add_outputs();
+  output->set_name(name);
+  output->set_comment(comment);
+  return OpProtoAndCheckerMaker::VariableBuilder{output};
+}
+
+void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() {
+  std::unordered_set<std::string> names;
+  auto checker = [&](const std::string& name) {
+    PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
+    names.insert(name);
+  };
+  for (auto& attr : proto_->attrs()) {
+    checker(attr.name());
+  }
+  for (auto& input : proto_->inputs()) {
+    checker(input.name());
+  }
+  for (auto& output : proto_->outputs()) {
+    checker(output.name());
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator.h

@@ -67,10 +67,6 @@ class OperatorBase {
   OperatorBase(const std::string& type, const VarNameMap& inputs,
                const VarNameMap& outputs, const AttributeMap& attrs);
 
-  OperatorBase(const OperatorBase& o) = delete;
-  OperatorBase& operator=(const OperatorBase& o) = delete;
-  OperatorBase(OperatorBase&& o) = delete;
-
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -116,10 +112,14 @@ class OperatorBase {
   void SetType(const std::string& type) { type_ = type; }
   const AttributeMap& Attrs() const { return attrs_; }
 
+  // Return a new operator instance, which is as same as this.
+  // Use unique_ptr to prevent caller forget to delete this pointer.
+  virtual std::unique_ptr<OperatorBase> Clone() const = 0;
+
  protected:
   std::string type_;
   // NOTE: in case of OpGrad, inputs_ contains:
-  // I (Inputs)
+  // I (Inputs)opear
   // O (Outputs)
   // OG (Output Gradients)
   VarNameMap inputs_;
@@ -130,12 +130,100 @@ class OperatorBase {
   AttributeMap attrs_;
 };
 
+// Macro for define a clone method.
+// If you are writing an kernel operator, `Clone` will be defined when you
+// register it. i.e. `Clone` method is not needed to define by yourself.
+#define DEFINE_OP_CLONE_METHOD(CLS)                       \
+  std::unique_ptr<OperatorBase> Clone() const final {     \
+    return std::unique_ptr<OperatorBase>(new CLS(*this)); \
+  }
+
+// Macro for define a default constructor for Operator.
+// You can also use
+//   using PARENT_CLASS::PARENT_CLASS;
+// to use parent's constructor.
+#define DEFINE_OP_CONSTRUCTOR(CLS, PARENT_CLS)                                 \
+  CLS(const std::string& type, const VarNameMap& inputs,                       \
+      const VarNameMap& outputs, const paddle::framework::AttributeMap& attrs) \
+      : PARENT_CLS(type, inputs, outputs, attrs) {}
+
 class NOP : public OperatorBase {
  public:
   using OperatorBase::OperatorBase;
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
+  std::unique_ptr<OperatorBase> Clone() const override {
+    return std::unique_ptr<OperatorBase>(new NOP(*this));
+  }
+};
+
+// this class not only make proto but also init attribute checkers.
+class OpProtoAndCheckerMaker {
+ public:
+  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : proto_(proto), op_checker_(op_checker) {}
+
+  ~OpProtoAndCheckerMaker() {
+    PADDLE_ENFORCE(validated_, "should call Validate after build");
+  }
+
+  void Validate();
+
+ protected:
+  struct VariableBuilder {
+    OpProto::Var* var_;
+
+    VariableBuilder& AsDuplicable() {
+      var_->set_duplicable(true);
+      return *this;
+    }
+
+    VariableBuilder& AsIntermediate() {
+      var_->set_intermediate(true);
+      return *this;
+    }
+
+    // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
+    // means that input/output is not needed when calculate gradient. It does
+    // not mean no gradient when backward. It should be changed soon.
+    VariableBuilder& AsNoGradient() {
+      var_->set_no_gradient(true);
+      return *this;
+    }
+  };
+
+  VariableBuilder AddInput(const std::string& name, const std::string& comment);
+
+  VariableBuilder AddOutput(const std::string& name,
+                            const std::string& comment);
+
+  template <typename T>
+  TypedAttrChecker<T>& AddAttr(const std::string& name,
+                               const std::string& comment,
+                               bool generated = false) {
+    auto* attr = proto_->add_attrs();
+    attr->set_name(name);
+    attr->set_comment(comment);
+    attr->set_generated(generated);
+    attr->set_type(AttrTypeID<T>());
+    return op_checker_->AddAttrChecker<T>(name);
+  }
+
+  void AddComment(const std::string& comment) { proto_->set_comment(comment); }
+
+ private:
+  void CheckNoDuplicatedInOutAttrs();
+
+  OpProto* proto_;
+  OpAttrChecker* op_checker_;
+  bool validated_{false};
+};
+
+class NOPMaker : public OpProtoAndCheckerMaker {
+ public:
+  NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {}
 };
 
 class InferShapeContext {

paddle/framework/operator_test.cc

@@ -245,3 +245,21 @@ TEST(OpKernel, multi_inputs) {
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   op->Run(scope, cpu_device_context);
 }
+
+class OperatorClone : public paddle::framework::OperatorBase {
+ public:
+  DEFINE_OP_CLONE_METHOD(OperatorClone);
+  OperatorClone(const std::string& type, const VarNameMap& inputs,
+                const VarNameMap& outputs,
+                const paddle::framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void InferShape(const paddle::framework::Scope& scope) const override {}
+  void Run(const paddle::framework::Scope& scope,
+           const paddle::platform::DeviceContext& dev_ctx) const override {}
+};
+
+TEST(Operator, Clone) {
+  OperatorClone a("ABC", {}, {}, {});
+  auto b = a.Clone();
+  ASSERT_EQ(a.Type(), b->Type());
+}
\ No newline at end of file

paddle/memory/CMakeLists.txt

 add_subdirectory(detail)
 
 cc_library(memory SRCS memory.cc)
-cc_library(memcpy SRCS memcpy.cc DEPS device_context)
+cc_library(memcpy SRCS memcpy.cc)
 
 cc_library(paddle_memory
     DEPS

paddle/memory/memcpy.cc

@@ -16,8 +16,6 @@ limitations under the License. */
 
 #include <cstring>  // for memcpy
 
-#include "paddle/platform/device_context.h"
-
 namespace paddle {
 namespace memory {
 

paddle/operators/net_op.cc

@@ -85,7 +85,14 @@ NetOp::NetOp(const std::string& type,
              const framework::OperatorBase::VarNameMap& inputs,
              const framework::OperatorBase::VarNameMap& outputs,
              const framework::AttributeMap& attrs)
-    : OperatorBase(type, inputs, outputs, attrs) {}
+    : framework::OperatorBase(type, inputs, outputs, attrs) {}
+
+std::unique_ptr<framework::OperatorBase> NetOp::Clone() const {
+  PADDLE_ENFORCE(
+      add_op_done_,
+      "Must clone a sealed NetOp, invoke Net::CompleteAddOp before clone");
+  return std::unique_ptr<OperatorBase>(new NetOp(*this));
+}
 
 }  // namespace operators
 }  // namespace paddle

paddle/operators/net_op.h

@@ -41,6 +41,18 @@ class NetOp : public framework::OperatorBase {
   NetOp(const std::string& type, const VarNameMap& inputs,
         const VarNameMap& outputs, const framework::AttributeMap& attrs);
 
+  NetOp(const NetOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    this->ops_.reserve(o.ops_.size());
+    std::transform(o.ops_.begin(), o.ops_.end(), std::back_inserter(this->ops_),
+                   [](const std::shared_ptr<OperatorBase>& op)
+                       -> std::shared_ptr<OperatorBase> {
+                         return std::shared_ptr<OperatorBase>(op->Clone());
+                       });
+    this->CompleteAddOp();
+  }
+
   /**
    * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
@@ -98,6 +110,8 @@ class NetOp : public framework::OperatorBase {
   bool IsNetOp() const override;
   std::vector<std::string> OutputVars(bool has_intermediate) const override;
 
+  std::unique_ptr<framework::OperatorBase> Clone() const override;
+
   std::vector<std::shared_ptr<OperatorBase>> ops_;
 
  private:

paddle/operators/net_op_test.cc

@@ -13,6 +13,7 @@ static int run_cnt = 0;
 class TestOp : public framework::OperatorBase {
  public:
   using framework::OperatorBase::OperatorBase;
+  DEFINE_OP_CLONE_METHOD(TestOp);
   void InferShape(const Scope& scope) const override { ++infer_shape_cnt; }
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {
@@ -70,5 +71,21 @@ TEST(NetOp, insert_op) {
   ASSERT_EQ(3UL, net.ops_.size());
 }
 
+TEST(NetOp, Clone) {
+  NetOp net;
+  net.AddOp(
+      std::shared_ptr<framework::NOP>(new framework::NOP{"empty", {}, {}, {}}));
+  net.AddOp(std::shared_ptr<framework::NOP>(
+      new framework::NOP{"empty2", {}, {}, {}}));
+  net.CompleteAddOp(true);
+  auto new_net_op = net.Clone();
+  ASSERT_NE(new_net_op, nullptr);
+  ASSERT_TRUE(new_net_op->IsNetOp());
+  auto* new_net = static_cast<NetOp*>(new_net_op.get());
+  ASSERT_EQ(2, new_net->ops_.size());
+  ASSERT_EQ(new_net->ops_[0]->Type(), "empty");
+  ASSERT_EQ(new_net->ops_[1]->Type(), "empty2");
+}
+
 }  // namespace operators
 }  // namespace paddle

paddle/operators/recurrent_op.h

@@ -110,13 +110,20 @@ class RecurrentGradientAlgorithm {
   std::shared_ptr<NetOp>* stepnet_;
 };
 
-class RecurrentOp final : public framework::OperatorBase {
+class RecurrentOp : public framework::OperatorBase {
  public:
   RecurrentOp(const std::string& type, const VarNameMap& inputs,
               const VarNameMap& outputs, const framework::AttributeMap& attrs);
+
+  RecurrentOp(const RecurrentOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement copy ctor well.
+    PADDLE_THROW("Not implemented");
+  }
   /**
-     * InferShape must be called before Run.
-     */
+   * InferShape must be called before Run.
+   */
   void InferShape(const framework::Scope& scope) const override {
     alg_.InferShape(scope);
   }
@@ -137,12 +144,19 @@ class RecurrentOp final : public framework::OperatorBase {
   std::shared_ptr<NetOp> stepnet_;
 };
 
-class RecurrentGradientOp final : public framework::OperatorBase {
+class RecurrentGradientOp : public framework::OperatorBase {
  public:
   RecurrentGradientOp(const std::string& type, const VarNameMap& inputs,
                       const VarNameMap& outputs,
                       const framework::AttributeMap& attrs);
 
+  RecurrentGradientOp(const RecurrentGradientOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement Copy ctor.
+    PADDLE_THROW("Not Implemented");
+  }
+
   /**
    * InferShape must be called before Run.
    */

paddle/platform/CMakeLists.txt

@@ -16,5 +16,8 @@ ELSE()
     set(GPU_CTX_DEPS)
 ENDIF()
 
-cc_library(device_context SRCS device_context.cc DEPS place eigen3 ${GPU_CTX_DEPS})
+# memcpy deoends on device_context, here add deps individually for
+# avoiding cycle dependencies
+cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
+    system_allocator memory_block meta_data meta_cache place eigen3 ${GPU_CTX_DEPS})
 nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)

paddle/platform/device_context.cc

@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/platform/device_context.h"
+#include "paddle/memory/memory.h"
 
 namespace paddle {
 namespace platform {
@@ -36,6 +37,59 @@ Place CPUDeviceContext::GetPlace() const { return CPUPlace(); }
 
 #ifndef PADDLE_ONLY_CPU
 
+class EigenCudaStreamDevice : public Eigen::StreamInterface {
+ public:
+  EigenCudaStreamDevice() : scratch_(nullptr), semaphore_(nullptr) {
+    Eigen::initializeDeviceProp();
+  }
+  ~EigenCudaStreamDevice() override {}
+
+  void Reinitialize(const cudaStream_t* cuda_stream, GPUPlace place) {
+    stream_ = cuda_stream;
+    place_ = place;
+    device_prop_ = &Eigen::m_deviceProperties[place.device];
+  }
+
+  const cudaStream_t& stream() const override { return *stream_; }
+
+  const cudaDeviceProp& deviceProperties() const override {
+    return *device_prop_;
+  }
+
+  void* allocate(size_t num_bytes) const override {
+    return paddle::memory::Alloc(place_, num_bytes);
+  }
+
+  void deallocate(void* buffer) const override {
+    paddle::memory::Free(place_, buffer);
+  }
+
+  void* scratchpad() const override {
+    if (scratch_ == NULL) {
+      scratch_ = allocate(Eigen::kCudaScratchSize + sizeof(unsigned int));
+    }
+    return scratch_;
+  }
+
+  unsigned int* semaphore() const override {
+    if (semaphore_ == NULL) {
+      char* scratch =
+          static_cast<char*>(scratchpad()) + Eigen::kCudaScratchSize;
+      semaphore_ = reinterpret_cast<unsigned int*>(scratch);
+      PADDLE_ENFORCE(
+          cudaMemsetAsync(semaphore_, 0, sizeof(unsigned int), *stream_));
+    }
+    return semaphore_;
+  }
+
+ private:
+  GPUPlace place_;
+  const cudaStream_t* stream_;         // not owned;
+  const cudaDeviceProp* device_prop_;  // not owned;
+  mutable void* scratch_;
+  mutable unsigned int* semaphore_;
+};
+
 template <>
 Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() const {
   return reinterpret_cast<const CUDADeviceContext*>(this)->eigen_device();
@@ -43,19 +97,9 @@ Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() const {
 
 CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) {
   SetDeviceId(place_.device);
-  // TODO(qijun) Pass a created cuda stream to Eigen::CudaStreamDevice directly
-  // here will cause segment fault. We must implement a class derived from
-  // Eigen::StreamInterface, and reinitialize it with a cuda stream and a gpu id
-  // later. Please refer to the implementation of class EigenCudaStreamDevice
-  // in TensorFlow.
-  //
-  // We find that CUDA 7 introduces a new option, the per-thread default stream,
-  // that has two effects. Please refer to https://devblogs.nvidia.com/
-  // parallelforall/gpu-pro-tip-cuda-7-streams-simplify-concurrency/
-  //
-  // So, we decide to use default stream and add –default-stream per-thread nvcc
-  // flag. Than, two threads with two CUDADeviceContexts will run parallelly.
-  eigen_stream_.reset(new Eigen::CudaStreamDevice());
+  PADDLE_ENFORCE(cudaStreamCreate(&stream_));
+  eigen_stream_.reset(new EigenCudaStreamDevice());
+  eigen_stream_->Reinitialize(&stream_, place);
   eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
 }
 
@@ -75,12 +119,13 @@ CUDADeviceContext::~CUDADeviceContext() {
   }
   eigen_stream_.reset();
   eigen_device_.reset();
+  PADDLE_ENFORCE(cudaStreamDestroy(stream_));
 }
 
 Place CUDADeviceContext::GetPlace() const { return place_; }
 
 void CUDADeviceContext::Wait() const {
-  PADDLE_ENFORCE(cudaStreamSynchronize(0));
+  PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
 }
 
 Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
@@ -91,6 +136,7 @@ cublasHandle_t CUDADeviceContext::cublas_handle() {
   if (!cublas_handle_) {
     SetDeviceId(place_.device);
     PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
+    PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
   }
   return cublas_handle_;
 }
@@ -99,10 +145,13 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() {
   if (!cudnn_handle_) {
     SetDeviceId(place_.device);
     PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
+    PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
   }
   return cudnn_handle_;
 }
 
+cudaStream_t CUDADeviceContext::stream() { return stream_; }
+
 curandGenerator_t CUDADeviceContext::curand_generator() {
   if (!curand_generator_) {
     SetDeviceId(place_.device);
@@ -110,6 +159,8 @@ curandGenerator_t CUDADeviceContext::curand_generator() {
                                                   CURAND_RNG_PSEUDO_DEFAULT));
     PADDLE_ENFORCE(
         dynload::curandSetPseudoRandomGeneratorSeed(curand_generator_, seed_));
+
+    PADDLE_ENFORCE(dynload::curandSetStream(curand_generator_, stream_));
   }
   return curand_generator_;
 }

paddle/platform/device_context.h

@@ -52,6 +52,7 @@ class CPUDeviceContext : public DeviceContext {
 };
 
 #ifndef PADDLE_ONLY_CPU
+class EigenCudaStreamDevice;
 
 class CUDADeviceContext : public DeviceContext {
  public:
@@ -76,6 +77,9 @@ class CUDADeviceContext : public DeviceContext {
 
   /*! \brief  Return curand handle in the device context. */
   curandGenerator_t curand_generator();
+
+  /*! \brief  Return cuda stream in the device context. */
+  cudaStream_t      stream();
   // clang-format on
 
  private:
@@ -83,15 +87,16 @@ class CUDADeviceContext : public DeviceContext {
 
  private:
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
-  std::unique_ptr<Eigen::CudaStreamDevice> eigen_stream_;
+  std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
 
  private:
   uint64_t seed_;
 
   // clang-format off
-  cudnnHandle_t     cudnn_handle_     = nullptr;
-  cublasHandle_t    cublas_handle_    = nullptr;
-  curandGenerator_t curand_generator_ = nullptr;
+  cudaStream_t       stream_{nullptr};
+  cudnnHandle_t      cudnn_handle_{nullptr};
+  cublasHandle_t     cublas_handle_{nullptr};
+  curandGenerator_t  curand_generator_{nullptr};
   // clang-format on
 };
 

paddle/platform/device_context_test.cc

@@ -45,6 +45,7 @@ TEST(Device, CUDADeviceContext) {
     ASSERT_NE(nullptr, cublas_handle);
     curandGenerator_t curand_handle = device_context->curand_generator();
     ASSERT_NE(nullptr, curand_handle);
+    ASSERT_NE(nullptr, device_context->stream());
     delete device_context;
   }
 }

paddle/platform/enforce.h

@@ -86,7 +86,7 @@ struct EnforceNotMet : public std::exception {
                                   2 + sizeof(void*) * 2, call_stack[i],
                                   demangled, addr_offset);
         } else {
-          sout << string::Sprintf("%-3d %*0p %s\n", i, 2 + sizeof(void*) * 2,
+          sout << string::Sprintf("%-3d %*0p\n", i, 2 + sizeof(void*) * 2,
                                   call_stack[i]);
         }
       }

paddle/scripts/docker/build.sh

@@ -120,25 +120,6 @@ EOF
     /woboq/indexgenerator/codebrowser_indexgenerator $WOBOQ_OUT
 fi
 
-# generate deb package for current build
-# FIXME(typhoonzero): should we remove paddle/scripts/deb ?
-if [[ ${WITH_DEB:-ON} == "ON" ]]; then
-    cat <<EOF
-========================================
-Generating .deb package ...
-========================================
-EOF
-    set +e
-    cpack -D CPACK_GENERATOR='DEB' -j `nproc` ..
-    err_code=$?
-    if [ ${err_code} -ne 0 ]; then
-        # cat error logs if cpack failed.
-        cat /paddle/build/_CPack_Packages/Linux/DEB/PreinstallOutput.log
-        exit ${err_code}
-    fi
-    set -e
-fi
-
 cat <<EOF
 ========================================
 Generate /paddle/build/Dockerfile ...
@@ -158,14 +139,13 @@ EOF
 fi
 
 cat >> /paddle/build/Dockerfile <<EOF
-# Use different deb file when building different type of images
-ADD *.deb /
+ADD python/dist/*.whl /
 # run paddle version to install python packages first
 RUN apt-get update &&\
     apt-get install -y wget python-pip && pip install -U pip && \
-    dpkg -i /*.deb ; apt-get install -f -y && \
+    pip install /*.whl; apt-get install -f -y && \
     apt-get clean -y && \
-    rm -f /*.deb && \
+    rm -f /*.whl && \
     paddle version
 ${DOCKERFILE_CUDNN_DSO}
 ${DOCKERFILE_GPU_ENV}

paddle/scripts/submit_local.sh.in

@@ -56,8 +56,7 @@ if [ -z "${PADDLE_NO_STAT+x}" ]; then
     fi
 fi
 
-
-MYDIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
+PADDLE_BIN_PATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
 
 if [ ! -z "${DEBUGGER}" ]; then
     echo "Using debug command ${DEBUGGER}"
@@ -93,34 +92,16 @@ else:
   sys.exit(0)
 EOF
 
-if [ $? -eq 1 ]; then  # Older version installed, or not installed at all
-    echo "First time run paddle, need to install some python dependencies."
-    # setuptools normalizes package version, so we need to use normalized
-    # package version for paddle python package
-    PYTHON_PADDLE_VERSION=$(python -c 'import packaging.version
-import setuptools
-print str(packaging.version.Version("@PADDLE_VERSION@"))
-' 2>/dev/null)
-    BASEDIR=$(dirname "$0")
-    pip install ${BASEDIR}/../opt/paddle/share/wheels/*-${PYTHON_PADDLE_VERSION}-*.whl
-    if [ $? -ne 0 ]; then
-	echo "pip install wheels failed. "
-	echo "Please use 'sudo paddle' at the first time you use PaddlePaddle"
-	echo "PaddlePaddle will install some python dependencies automatically."
-	exit 1
-    fi
-    echo "Python dependencies are installed."
-fi
 
 case "$1" in
     "train")
-        ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_trainer ${@:2}
+        ${DEBUGGER} $PADDLE_BIN_PATH/paddle_trainer ${@:2}
         ;;
     "merge_model")
-        ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_merge_model ${@:2}
+        ${DEBUGGER} $PADDLE_BIN_PATH/paddle_merge_model ${@:2}
         ;;
     "pserver")
-        ${DEBUGGER} $MYDIR/../opt/paddle/bin/paddle_pserver_main ${@:2}
+        ${DEBUGGER} $PADDLE_BIN_PATH/paddle_pserver_main ${@:2}
         ;;
     "dump_config")
         python -m paddle.utils.dump_config ${@:2}
@@ -129,7 +110,7 @@ case "$1" in
         python -m paddle.utils.make_model_diagram ${@:2}
         ;;
     "usage")
-        $MYDIR/../opt/paddle/bin/paddle_usage ${@:2}
+        $PADDLE_BIN_PATH/paddle_usage ${@:2}
         ;;
     "version")
         version

python/paddle/v2/trainer.py

@@ -27,16 +27,24 @@ class SGD(object):
     SGD Trainer combines data reader, network topolopy and update_equation together
     to train/test a neural network.
 
-    :param update_equation: The optimizer object.
-    :type update_equation: paddle.v2.optimizer.Optimizer
     :param cost: Target cost that neural network should be optimized.
     :type cost: paddle.v2.config_base.Layer
     :param parameters: The parameters dictionary.
     :type parameters: paddle.v2.parameters.Parameters
+    :param update_equation: The optimizer object.
+    :type update_equation: paddle.v2.optimizer.Optimizer
     :param extra_layers: Some layers in the neural network graph are not
                          in the path of cost layer.
-    :param pserver_spec: pserver location, eg: localhost:3000
     :type extra_layers: paddle.v2.config_base.Layer
+    :param is_local: Whether trainning locally
+    :type is_local: bool
+    :param pserver_spec: comma string for pserver location,
+                         eg:127.10.0.10:3000,127.10.0.11:3000,
+                         and this parameter is only used for fault
+                         tolerant mode cluster training.
+    :type pserver_spec: string
+    :param use_etcd: Whether using etcd pserver.
+    :param use_etcd: bool
     """
 
     def __init__(self,

python/setup.py.in

@@ -24,14 +24,17 @@ if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
     setup_requires+=["opencv-python"]
 
 # the prefix is sys.prefix which should always be usr
-paddle_bin_dir = 'local/opt/paddle/bin'
+paddle_bin_dir = 'opt/paddle/bin'
 paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage',
                '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer',
                '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model',
-               '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main']
+               '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main',
+               '${PADDLE_BINARY_DIR}/paddle/scripts/paddle']
 
-paddle_rt_lib_dir = 'local/lib'
-paddle_rt_libs = [] if '${MKL_SHARED_LIBS}'== '' else '${MKL_SHARED_LIBS}'.split(';')
+paddle_rt_lib_dir = 'lib'
+paddle_rt_libs = ['${WARPCTC_LIBRARIES}']
+if '${MKL_SHARED_LIBS}'!= '':
+  paddle_rt_libs += '${MKL_SHARED_LIBS}'.split(';')
 
 setup(name='paddlepaddle',
       version='${PADDLE_VERSION}',
@@ -50,8 +53,7 @@ setup(name='paddlepaddle',
           'paddle.v2.framework.proto': '${PADDLE_BINARY_DIR}/paddle/framework',
           'py_paddle': '${PADDLE_SOURCE_DIR}/paddle/py_paddle'
       },
-      scripts=['${PADDLE_BINARY_DIR}/paddle/scripts/paddle'],
+      scripts=paddle_bins,
       distclass=BinaryDistribution,
-      data_files=[(paddle_bin_dir, paddle_bins),
-                  (paddle_rt_lib_dir, paddle_rt_libs)]
+      data_files=[(paddle_rt_lib_dir, paddle_rt_libs)]
 )