diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake
index bd401faa6eb8a583bce542db68852f8571681daf..8a594a825abdca6a0f989b94fa42f97d6df5e10a 100644
--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -52,6 +52,7 @@ ExternalProject_Add(
 
 ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})
-ADD_DEPENDENCIES(glog extern_glog)
+ADD_DEPENDENCIES(glog extern_glog gflags)
+LINK_LIBRARIES(glog gflags)
 
 LIST(APPEND external_project_dependencies glog)
diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt
index 0b5e9a2599dbec6d40eb70609d61ebea2b7e1486..4b06966fba2bc9f92756be0cb8110bbcd5272423 100644
--- a/paddle/CMakeLists.txt
+++ b/paddle/CMakeLists.txt
@@ -14,6 +14,7 @@ if(Boost_FOUND)
   add_subdirectory(memory)
   add_subdirectory(platform)
   add_subdirectory(framework)
+  add_subdirectory(operators)
   add_subdirectory(pybind)
 endif()
 
diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index de31952e79262ed02193ab5ce06c9e2105c232e6..8415ce67e90397b31864a90cda54b81c19b3b34e 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -11,8 +11,8 @@ proto_library(op_proto SRCS op_proto.proto DEPS attr_type)
 cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
 cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
-cc_library(operator SRCS operator.cc DEPS op_desc protobuf device_context)
-cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry place)
+cc_library(operator SRCS operator.cc DEPS op_desc device_context)
+cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto op_desc)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry operator)
 py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)
diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index e46da822c6c2e99b682212dc724126ac47150a24..61dfcb704964cd730a8fc9ab6ad394cd47cb4666 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -1,6 +1,7 @@
 #pragma once
 
 #include <algorithm>
+#include <type_traits>
 #include "paddle/framework/attr_checker.h"
 #include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/op_proto.pb.h"
@@ -81,8 +82,6 @@ class OpProtoAndCheckerMaker {
     return op_checker_->AddAttrChecker<T>(name);
   }
 
-  void AddType(const std::string& op_type) { proto_->set_type(op_type); }
-
   void AddComment(const std::string& comment) {
     *(proto_->mutable_comment()) = comment;
   }
@@ -101,8 +100,11 @@ class OpRegistry {
     OpProto& op_proto = protos()[op_type];
     OpAttrChecker& op_checker = op_checkers()[op_type];
     ProtoMakerType(&op_proto, &op_checker);
-    PADDLE_ENFORCE(op_proto.IsInitialized(),
-                   "Fail to initialize %s's OpProto !", op_type);
+    *op_proto.mutable_type() = op_type;
+    PADDLE_ENFORCE(
+        op_proto.IsInitialized(),
+        "Fail to initialize %s's OpProto, because %s is not initialized",
+        op_type, op_proto.InitializationErrorString());
   }
 
   static OperatorBase* CreateOp(const OpDesc& op_desc) {
@@ -143,18 +145,73 @@ class OpRegistry {
 template <typename OpType, typename ProtoMakerType>
 class OpRegisterHelper {
  public:
-  OpRegisterHelper(std::string op_type) {
+  OpRegisterHelper(const char* op_type) {
     OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
   }
 };
 
-#define REGISTER_OP(type, op_class, op_maker_class)                         \
-  class op_class##Register {                                                \
-   private:                                                                 \
-    const static OpRegisterHelper<op_class, op_maker_class> reg;            \
-  };                                                                        \
-  const OpRegisterHelper<op_class, op_maker_class> op_class##Register::reg( \
-      #type)
+#define STATIC_ASSERT_GLOBAL_NAMESPACE(uniq_name, msg)                        \
+  struct __test_global_namespace_##uniq_name##__ {};                          \
+  static_assert(std::is_same<::__test_global_namespace_##uniq_name##__,       \
+                             __test_global_namespace_##uniq_name##__>::value, \
+                msg)
+
+#define REGISTER_OP(__op_type, __op_class, __op_maker_class)                 \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(__reg_op__##__op_type,                      \
+                                 "REGISTER_OP must be in global namespace"); \
+  static ::paddle::framework::OpRegisterHelper<__op_class, __op_maker_class> \
+      __op_register_##__op_type##__(#__op_type);                             \
+  int __op_register_##__op_type##_handle__() { return 0; }
+
+#define REGISTER_OP_KERNEL(type, GPU_OR_CPU, PlaceType, KernelType)       \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
+      __reg_op_kernel_##type##_##GPU_OR_CPU##__,                          \
+      "REGISTER_OP_KERNEL must be in global namespace");                  \
+  struct __op_kernel_register__##type##__ {                               \
+    __op_kernel_register__##type##__() {                                  \
+      ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
+      key.place_ = PlaceType();                                           \
+      ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
+          .reset(new KernelType());                                       \
+    }                                                                     \
+  };                                                                      \
+  static __op_kernel_register__##type##__ __reg_kernel_##type##__;        \
+  int __op_kernel_register_##type##_handle_##GPU_OR_CPU##__() { return 0; }
+
+#define REGISTER_OP_GPU_KERNEL(type, KernelType) \
+  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, KernelType)
+
+#define REGISTER_OP_CPU_KERNEL(type, KernelType) \
+  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, KernelType)
+
+#define USE_OP_WITHOUT_KERNEL(op_type)                      \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                           \
+      __use_op_without_kernel_##op_type,                    \
+      "USE_OP_WITHOUT_KERNEL must be in global namespace"); \
+  extern int __op_register_##op_type##_handle__();          \
+  static int __use_op_ptr_##op_type##_without_kernel__      \
+      __attribute__((unused)) = __op_register_##op_type##_handle__()
+
+#define USE_OP_KERNEL(op_type, DEVICE_TYPE)                               \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
+      __use_op_kernel_##op_type##_##DEVICE_TYPE##__,                      \
+      "USE_OP_KERNEL must be in global namespace");                       \
+  extern int __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__(); \
+  static int __use_op_ptr_##op_type##_##DEVICE_TYPE##_kernel__            \
+      __attribute__((unused)) =                                           \
+          __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
+
+#ifdef PADDLE_ONLY_CPU
+#define USE_OP(op_type)           \
+  USE_OP_WITHOUT_KERNEL(op_type); \
+  USE_OP_KERNEL(op_type, CPU);
+
+#else
+#define USE_OP(op_type)           \
+  USE_OP_WITHOUT_KERNEL(op_type); \
+  USE_OP_KERNEL(op_type, CPU);    \
+  USE_OP_KERNEL(op_type, GPU)
+#endif
 
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc
index f5162fb870a91e566a0d2b1419050fe0799b199b..9bcc0407addca555e0b47b8178f6304396ce81fc 100644
--- a/paddle/framework/op_registry_test.cc
+++ b/paddle/framework/op_registry_test.cc
@@ -1,8 +1,6 @@
 #include "paddle/framework/op_registry.h"
 #include <gtest/gtest.h>
 
-using namespace paddle::framework;
-
 namespace paddle {
 namespace framework {
 class CosineOp : public OperatorBase {
@@ -21,13 +19,10 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .LargerThan(0.0);
-    AddType("cos");
     AddComment("This is cos op");
   }
 };
 
-REGISTER_OP(cos_sim, CosineOp, CosineOpProtoAndCheckerMaker);
-
 class MyTestOp : public OperatorBase {
  public:
   void InferShape(const std::shared_ptr<Scope>& scope) const override {}
@@ -48,15 +43,17 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     };
     AddAttr<int>("test_attr", "a simple test attribute")
         .AddCustomChecker(my_checker);
-    AddType("my_test_op");
     AddComment("This is my_test op");
   }
 };
-
-REGISTER_OP(my_test_op, MyTestOp, MyTestOpProtoAndCheckerMaker);
 }  // namespace framework
 }  // namespace paddle
 
+REGISTER_OP(cos_sim, paddle::framework::CosineOp,
+            paddle::framework::CosineOpProtoAndCheckerMaker);
+REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
+            paddle::framework::MyTestOpProtoAndCheckerMaker);
+
 TEST(OpRegistry, CreateOp) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("cos_sim");
@@ -71,7 +68,7 @@ TEST(OpRegistry, CreateOp) {
 
   paddle::framework::OperatorBase* op =
       paddle::framework::OpRegistry::CreateOp(op_desc);
-  auto scope = std::make_shared<Scope>();
+  auto scope = std::make_shared<paddle::framework::Scope>();
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
   float scale_get = op->GetAttr<float>("scale");
@@ -114,7 +111,7 @@ TEST(OpRegistry, DefaultValue) {
 
   paddle::framework::OperatorBase* op =
       paddle::framework::OpRegistry::CreateOp(op_desc);
-  auto scope = std::make_shared<Scope>();
+  auto scope = std::make_shared<paddle::framework::Scope>();
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
   ASSERT_EQ(op->GetAttr<float>("scale"), 1.0);
@@ -169,13 +166,8 @@ TEST(OpRegistry, CustomChecker) {
   paddle::framework::OperatorBase* op =
       paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::platform::CPUDeviceContext dev_ctx;
-  auto scope = std::make_shared<Scope>();
+  auto scope = std::make_shared<paddle::framework::Scope>();
   op->Run(scope, dev_ctx);
   int test_attr = op->GetAttr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
 }
-
-int main(int argc, char** argv) {
-  testing::InitGoogleTest(&argc, argv);
-  return RUN_ALL_TESTS();
-}
\ No newline at end of file
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index 4336115670031fed3672e799340241c9d6f45d5f..d3c55e0ceb6737772be38b536a107900ee895b12 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <paddle/framework/attr_checker.h>
 #include <paddle/framework/op_desc.pb.h>
 #include <paddle/framework/scope.h>
+#include <paddle/framework/tensor.h>
 #include <paddle/platform/device_context.h>
 #include <paddle/platform/place.h>
 #include <paddle/utils/Error.h>
@@ -103,6 +104,19 @@ class OpKernel {
   virtual ~OpKernel() {}
 };
 
+template <typename T>
+struct VarToTensor {};
+
+template <>
+struct VarToTensor<Tensor*> {
+  Tensor* operator()(Variable* var) { return var->GetMutable<Tensor>(); }
+};
+
+template <>
+struct VarToTensor<const Tensor*> {
+  const Tensor* operator()(Variable* var) { return &var->Get<Tensor>(); }
+};
+
 class OperatorWithKernel : public OperatorBase {
  public:
   struct OpKernelKey {
@@ -136,19 +150,36 @@ class OperatorWithKernel : public OperatorBase {
   AllOpKernels() {
     static std::unordered_map<std::string, OpKernelMap> g_all_op_kernels;
     return g_all_op_kernels;
+  }
+  void InferShape(const std::shared_ptr<Scope>& scope) const final {
+    std::vector<const Tensor*> ins;
+    VarNamesToTensors(scope, inputs_, &ins);
+    std::vector<Tensor*> outs;
+    VarNamesToTensors(scope, outputs_, &outs);
+    InferShape(ins, outs);
   };
+
+ private:
+  template <typename T>
+  void VarNamesToTensors(const std::shared_ptr<Scope>& scope,
+                         const std::vector<std::string>& var_names,
+                         std::vector<T>* container) const {
+    container->reserve(var_names.size());
+    VarToTensor<T> convert;
+    for (auto& name : var_names) {
+      auto var = scope->GetVariable(name);
+      if (var != nullptr) {
+        container->push_back(convert(var));
+      } else {
+        container->push_back(nullptr);
+      }
+    }
+  }
+
+ protected:
+  virtual void InferShape(const std::vector<const Tensor*>& inputs,
+                          const std::vector<Tensor*>& outputs) const = 0;
 };
 
 }  // namespace framework
 }  // namespace paddle
-
-#define REGISTER_OP_KERNEL(type, PlaceType, KernelType)                   \
-  struct __op_kernel_register__##type##__ {                               \
-    __op_kernel_register__##type##__() {                                  \
-      ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
-      key.place_ = PlaceType();                                           \
-      ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
-          .reset(new KernelType());                                       \
-    }                                                                     \
-  };                                                                      \
-  static __op_kernel_register__##type##__ __reg_kernel_##type##__
diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc
index 01b87bb50e7a10134fc43fa6aa2716e13c4dbe33..204b601a4aad52ee57b81235851c9806b607799f 100644
--- a/paddle/framework/operator_test.cc
+++ b/paddle/framework/operator_test.cc
@@ -36,44 +36,6 @@ class OperatorTest : public OperatorBase {
   float x = 0;
 };
 
-class OperatorTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
- public:
-  OperatorTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("input", "input of test op");
-    AddOutput("output", "output of test op");
-    AddAttr<float>("scale", "scale of cosine op")
-        .SetDefault(1.0)
-        .LargerThan(0.0);
-    AddType("test_operator");
-    AddComment("This is test op");
-  }
-};
-
-REGISTER_OP(test_operator, OperatorTest, OperatorTestProtoAndCheckerMaker);
-
-TEST(OperatorBase, all) {
-  OpDesc op_desc;
-  op_desc.set_type("test_operator");
-  *op_desc.mutable_inputs()->Add() = "IN1";
-  *op_desc.mutable_outputs()->Add() = "OUT1";
-  auto attr = op_desc.mutable_attrs()->Add();
-  attr->set_name("scale");
-  attr->set_type(paddle::framework::AttrType::FLOAT);
-  float scale = 3.14;
-  attr->set_f(scale);
-
-  platform::CPUDeviceContext device_context;
-  auto scope = std::make_shared<Scope>();
-
-  OperatorBase* op = paddle::framework::OpRegistry::CreateOp(op_desc);
-  ASSERT_EQ(op->GetAttr<float>("scale"), scale);
-  scope->CreateVariable("OUT1");
-  op->Run(scope, device_context);
-  std::cout << op->DebugString() << std::endl;
-  delete op;
-}
-
 class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
   OpKernelTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
@@ -83,14 +45,14 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .LargerThan(0.0);
-    AddType("test_operator");
     AddComment("This is test op");
   }
 };
 
 class OpWithKernelTest : public OperatorWithKernel {
- public:
-  void InferShape(const std::shared_ptr<Scope>& scope) const override {}
+ protected:
+  void InferShape(const std::vector<const Tensor*>& inputs,
+                  const std::vector<Tensor*>& outputs) const override {}
 };
 
 class CPUKernelTest : public OpKernel {
@@ -103,10 +65,16 @@ class CPUKernelTest : public OpKernel {
   }
 };
 
-REGISTER_OP(op_with_kernel, OpWithKernelTest, OpKernelTestProtoAndCheckerMaker);
-REGISTER_OP_KERNEL(op_with_kernel, platform::CPUPlace, CPUKernelTest);
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest,
+            paddle::framework::OpKernelTestProtoAndCheckerMaker);
+REGISTER_OP_CPU_KERNEL(op_with_kernel, paddle::framework::CPUKernelTest);
 
 TEST(OpKernel, all) {
+  using namespace paddle::framework;
+
   OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
   *op_desc.mutable_inputs()->Add() = "IN1";
@@ -116,7 +84,7 @@ TEST(OpKernel, all) {
   attr->set_type(paddle::framework::AttrType::FLOAT);
   attr->set_f(3.14);
 
-  platform::CPUDeviceContext cpu_device_context;
+  paddle::platform::CPUDeviceContext cpu_device_context;
   auto scope = std::make_shared<Scope>();
 
   OperatorBase* op = paddle::framework::OpRegistry::CreateOp(op_desc);
@@ -124,5 +92,3 @@ TEST(OpKernel, all) {
 
   delete op;
 }
-}  // namespace framework
-}  // namespace paddle
\ No newline at end of file
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..40bb326512c118178184120d4bc26dc127689ff3
--- /dev/null
+++ b/paddle/operators/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(WITH_GPU)
+    nv_library(add_op SRCS add_op.cc add_op.cu DEPS operator op_registry glog ddim)
+else()
+    cc_library(add_op SRCS add_op.cc DEPS operator op_registry glog ddim)
+endif()
+cc_test(add_op_test SRCS add_op_test.cc DEPS add_op)
diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..2766f0bf258ed863a4297c1e4a2be4673cbf3044
--- /dev/null
+++ b/paddle/operators/add_op.cc
@@ -0,0 +1,44 @@
+#include <paddle/framework/op_registry.h>
+#include <paddle/framework/tensor.h>
+#include <paddle/operators/add_op.h>
+
+namespace paddle {
+namespace operators {
+
+class AddOp : public framework::OperatorWithKernel {
+protected:
+  void InferShape(
+      const std::vector<const framework::Tensor *> &inputs,
+      const std::vector<framework::Tensor *> &outputs) const override {
+    PADDLE_ENFORCE(inputs.size() == 2, "Input size of AddOp must be two");
+    PADDLE_ENFORCE(outputs.size() == 1, "Output size of AddOp must be one");
+    PADDLE_ENFORCE(
+        inputs[0] != nullptr && inputs[1] != nullptr && outputs[0] != nullptr,
+        "Inputs/Outputs of AddOp must all be set");
+    PADDLE_ENFORCE(inputs[0]->dims() == inputs[1]->dims(),
+                   "Two input of Add Op's dimension must be same.");
+    // Need set dims in Tensor
+    // outputs[0]->set_dims(inputs[0]->dims())
+  }
+};
+
+class AddOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  AddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of add op");
+    AddInput("Y", "The second input of add op");
+    AddOutput("Out", "The output of add op");
+    AddComment(R"DOC(
+Two Element Add Operator.
+
+The equation is: Out = X + Y
+)DOC");
+  }
+};
+}  // namespace op
+}  // namespace paddle
+
+REGISTER_OP(add_two, paddle::operators::AddOp, paddle::operators::AddOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    add_two, ::paddle::operators::AddKernel<::paddle::platform::CPUPlace>);
\ No newline at end of file
diff --git a/paddle/operators/add_op.cu b/paddle/operators/add_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..5979345fffd68d71ba09dc96874d8ff9471bdbcc
--- /dev/null
+++ b/paddle/operators/add_op.cu
@@ -0,0 +1,5 @@
+#include <paddle/operators/add_op.h>
+#include <paddle/framework/op_registry.h>
+
+REGISTER_OP_GPU_KERNEL(add_two,
+                       paddle::operators::AddKernel<paddle::platform::GPUPlace>);
\ No newline at end of file
diff --git a/paddle/operators/add_op.h b/paddle/operators/add_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..17d459dbc86af73fa86934a5ccce9da509c59c6b
--- /dev/null
+++ b/paddle/operators/add_op.h
@@ -0,0 +1,17 @@
+#pragma once
+#include <glog/logging.h>
+#include <paddle/framework/operator.h>
+
+namespace paddle {
+namespace operators {
+
+template <typename Place>
+class AddKernel : public framework::OpKernel {
+public:
+  void Compute(const KernelContext &context) const override {
+    LOG(INFO) << "Add kernel in " << typeid(Place).name();
+  }
+};
+
+}  // namespace op
+}  // namespace paddle
diff --git a/paddle/operators/add_op_test.cc b/paddle/operators/add_op_test.cc
new file mode 100644
index 0000000000000000000000000000000000000000..f554ac1bef3255f136ad4407a7a1096bdc2b1db5
--- /dev/null
+++ b/paddle/operators/add_op_test.cc
@@ -0,0 +1,9 @@
+#include <gtest/gtest.h>
+#define private public
+#include <paddle/framework/op_registry.h>
+USE_OP(add_two);
+TEST(AddOp, GetOpProto) {
+  auto& protos = paddle::framework::OpRegistry::protos();
+  auto it = protos.find("add_two");
+  ASSERT_NE(it, protos.end());
+}
\ No newline at end of file