diff --git a/cmake/operators.cmake b/cmake/operators.cmake
index 89726bf9859e71ee04c2f9380554090845fd44e5..2ced43f9e6c60da642f7a6252f889d9c9ab9748f 100644
--- a/cmake/operators.cmake
+++ b/cmake/operators.cmake
@@ -166,6 +166,8 @@ function(op_library TARGET)
       # Append first implemented MKLDNN activation operator
       if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
+      elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
+        file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
       else()
         file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
       endif()
diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index c701a2ad63048f69e8443fee7434928698aa20cc..e4c471d86b7bff1bfb3b697ab24219144b4667f5 100644
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -118,8 +118,9 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
 cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
 
 cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
+cc_library(op_kernel_type SRCS op_kernel_type.cc DEPS device_context place)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
-    shape_inference data_transform lod_tensor profiler transfer_scope_cache)
+    shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type)
 
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
 
@@ -191,7 +192,7 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
 
-cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
+cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto op_kernel_type)
 cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
 
 cc_test(tuple_test SRCS tuple_test.cc )
diff --git a/paddle/fluid/framework/op_kernel_type.cc b/paddle/fluid/framework/op_kernel_type.cc
new file mode 100644
index 0000000000000000000000000000000000000000..6d4801e4a0eed7083e671e1d49b8628dfb280cf9
--- /dev/null
+++ b/paddle/fluid/framework/op_kernel_type.cc
@@ -0,0 +1,54 @@
+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+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/fluid/framework/op_kernel_type.h"
+
+namespace paddle {
+namespace framework {
+
+size_t OpKernelType::Hash::operator()(const OpKernelType& key) const {
+  int cur_loc = 0;
+
+  int place = key.place_.which();
+  cur_loc += OpKernelType::kPlaceBits;
+
+  int data_type = static_cast<int>(key.data_type_) << cur_loc;
+  cur_loc += OpKernelType::kPrimaryDTypeBits;
+
+  int data_layout = static_cast<int>(key.data_layout_) << cur_loc;
+  cur_loc += OpKernelType::kLayoutBits;
+
+  int library_type = static_cast<int>(key.library_type_) << cur_loc;
+  cur_loc += OpKernelType::kLibBits;
+
+  int customized_value = key.customized_type_value_;
+  PADDLE_ENFORCE(customized_value < (1 << OpKernelType::kCustomizeBits));
+  customized_value = customized_value << cur_loc;
+  cur_loc += OpKernelType::kCustomizeBits;
+  PADDLE_ENFORCE(cur_loc < 64);
+
+  std::hash<int> hasher;
+  return hasher(place + data_type + data_layout + library_type +
+                customized_value);
+}
+
+bool OpKernelType::operator==(const OpKernelType& o) const {
+  return platform::places_are_same_class(place_, o.place_) &&
+         data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
+         library_type_ == o.library_type_ &&
+         customized_type_value_ == o.customized_type_value_;
+}
+
+}  // namespace framework
+}  // namespace paddle
diff --git a/paddle/fluid/framework/op_kernel_type.h b/paddle/fluid/framework/op_kernel_type.h
index ac0330218973123771367ed5ba9477c90143a043..9edc1a3e150027b5a3dbd8483dc8b58d1d4ab918 100644
--- a/paddle/fluid/framework/op_kernel_type.h
+++ b/paddle/fluid/framework/op_kernel_type.h
@@ -24,54 +24,55 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-struct OpKernelType {
-  struct Hash {
-    size_t operator()(const OpKernelType& key) const {
-      int place = key.place_.which();
-      int data_type = static_cast<int>(key.data_type_) << LEFT_SHIFT;
-      int data_layout = static_cast<int>(key.data_layout_) << (LEFT_SHIFT * 2);
-      int library_type = static_cast<int>(key.library_type_)
-                         << (LEFT_SHIFT * 3);
-
-      std::hash<int> hasher;
-      return hasher(place + data_type + data_layout + library_type);
-    }
-  };
+class OpKernelType {
+ public:
+  constexpr static int kDefaultCustomizedTypeValue = 0;
 
-  // place, data_type, library_type kinds less than 2^8
-  constexpr static int LEFT_SHIFT = 8;
-
-  proto::VarType::Type data_type_;
-  DataLayout data_layout_;
-  platform::Place place_;
-  LibraryType library_type_;
+  // In total should be smaller than 64.
+  constexpr static int kPlaceBits = 4;
+  constexpr static int kPrimaryDTypeBits = 8;
+  constexpr static int kLayoutBits = 4;
+  constexpr static int kLibBits = 4;
+  constexpr static int kCustomizeBits = 4;
 
   OpKernelType(proto::VarType::Type data_type, platform::Place place,
                DataLayout data_layout = DataLayout::kAnyLayout,
-               LibraryType library_type = LibraryType::kPlain)
+               LibraryType library_type = LibraryType::kPlain,
+               int customized_type_value = kDefaultCustomizedTypeValue)
       : data_type_(data_type),
         data_layout_(data_layout),
         place_(place),
-        library_type_(library_type) {}
+        library_type_(library_type),
+        customized_type_value_(customized_type_value) {}
 
   OpKernelType(proto::VarType::Type data_type,
                const platform::DeviceContext& dev_ctx,
                DataLayout data_layout = DataLayout::kAnyLayout,
-               LibraryType library_type = LibraryType::kPlain)
+               LibraryType library_type = LibraryType::kPlain,
+               int customized_type_value = kDefaultCustomizedTypeValue)
       : data_type_(data_type),
         data_layout_(data_layout),
         place_(dev_ctx.GetPlace()),
-        library_type_(library_type) {}
+        library_type_(library_type),
+        customized_type_value_(customized_type_value) {}
+
+  virtual ~OpKernelType() {}
+
+  struct Hash {
+    size_t operator()(const OpKernelType& key) const;
+  };
 
   size_t hash_key() const { return Hash()(*this); }
 
-  bool operator==(const OpKernelType& o) const {
-    return platform::places_are_same_class(place_, o.place_) &&
-           data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
-           library_type_ == o.library_type_;
-  }
+  bool operator==(const OpKernelType& o) const;
 
   bool operator!=(const OpKernelType& o) const { return !(*this == o); }
+
+  proto::VarType::Type data_type_;
+  DataLayout data_layout_;
+  platform::Place place_;
+  LibraryType library_type_;
+  int customized_type_value_;
 };
 
 inline std::ostream& operator<<(std::ostream& os,
diff --git a/paddle/fluid/framework/op_registry.h b/paddle/fluid/framework/op_registry.h
index 0e6e74293c30d5f8caa58fe6bfa63657d2669b46..36673e48c2047bca54f604b082dfec123f1e2c82 100644
--- a/paddle/fluid/framework/op_registry.h
+++ b/paddle/fluid/framework/op_registry.h
@@ -35,6 +35,7 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
+
 class Registrar {
  public:
   // In our design, various kinds of classes, e.g., operators and kernels,
@@ -78,7 +79,7 @@ struct OpKernelRegistrarFunctor;
 
 template <typename PlaceType, typename T, typename Func>
 inline void RegisterKernelClass(const char* op_type, const char* library_type,
-                                Func func) {
+                                int customized_type_value, Func func) {
   std::string library(library_type);
   std::string data_layout = "ANYLAYOUT";
   if (library == "MKLDNN") {
@@ -86,7 +87,7 @@ inline void RegisterKernelClass(const char* op_type, const char* library_type,
   }
   OpKernelType key(ToDataType(std::type_index(typeid(T))), PlaceType(),
                    StringToDataLayout(data_layout),
-                   StringToLibraryType(library_type));
+                   StringToLibraryType(library_type), customized_type_value);
   OperatorWithKernel::AllOpKernels()[op_type][key] = func;
 }
 
@@ -95,22 +96,26 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
   using KERNEL_TYPE =
       typename std::tuple_element<I, std::tuple<KernelTypes...>>::type;
 
-  void operator()(const char* op_type, const char* library_type) const {
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {
     using T = typename KERNEL_TYPE::ELEMENT_TYPE;
     RegisterKernelClass<PlaceType, T>(
-        op_type, library_type, [](const framework::ExecutionContext& ctx) {
+        op_type, library_type, customized_type_value,
+
+        [](const framework::ExecutionContext& ctx) {
           KERNEL_TYPE().Compute(ctx);
         });
     constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
     OpKernelRegistrarFunctor<PlaceType, I + 1 == size, I + 1, KernelTypes...>
         func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
 template <typename PlaceType, size_t I, typename... KernelType>
 struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
-  void operator()(const char* op_type, const char* library_type) const {}
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {}
 };
 
 // User can register many kernel in one place. The data type could be
@@ -118,9 +123,10 @@ struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
 template <typename PlaceType, typename... KernelType>
 class OpKernelRegistrar : public Registrar {
  public:
-  explicit OpKernelRegistrar(const char* op_type, const char* library_type) {
+  explicit OpKernelRegistrar(const char* op_type, const char* library_type,
+                             int customized_type_value) {
     OpKernelRegistrarFunctor<PlaceType, false, 0, KernelType...> func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
@@ -130,17 +136,19 @@ struct OpKernelRegistrarFunctorEx;
 template <typename PlaceType, typename... DataTypeAndKernelType>
 class OpKernelRegistrarEx : public Registrar {
  public:
-  explicit OpKernelRegistrarEx(const char* op_type, const char* library_type) {
+  explicit OpKernelRegistrarEx(const char* op_type, const char* library_type,
+                               int customized_type_value) {
     OpKernelRegistrarFunctorEx<PlaceType, false, 0, DataTypeAndKernelType...>
         func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
 template <typename PlaceType, size_t I, typename... DataTypeAndKernelType>
 struct OpKernelRegistrarFunctorEx<PlaceType, true, I,
                                   DataTypeAndKernelType...> {
-  void operator()(const char* op_type, const char* library_type) const {}
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {}
 };
 
 template <typename PlaceType, size_t I, typename... DataTypeAndKernelType>
@@ -153,18 +161,21 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
       typename std::tuple_element<I,
                                   std::tuple<DataTypeAndKernelType...>>::type;
 
-  void operator()(const char* op_type, const char* library_type) const {
-    RegisterKernelClass<PlaceType, T>(op_type, library_type, Functor());
+  void operator()(const char* op_type, const char* library_type,
+                  int customized_type_value) const {
+    RegisterKernelClass<PlaceType, T>(op_type, library_type,
+                                      customized_type_value, Functor());
 
     constexpr auto size =
         std::tuple_size<std::tuple<DataTypeAndKernelType...>>::value;
     OpKernelRegistrarFunctorEx<PlaceType, I + 2 >= size, I + 2,
                                DataTypeAndKernelType...>
         func;
-    func(op_type, library_type);
+    func(op_type, library_type, customized_type_value);
   }
 };
 
+// clang-format off
 /**
  * check if MACRO is used in GLOBAL NAMESPACE.
  */
@@ -199,42 +210,64 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
 /**
  * Macro to register OperatorKernel.
  */
-#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...)        \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                          \
-      __reg_op_kernel_##op_type##_##library_type##__,                      \
-      "REGISTER_OP_KERNEL must be called in global namespace");            \
-  static ::paddle::framework::OpKernelRegistrar<place_class, __VA_ARGS__>  \
-      __op_kernel_registrar_##op_type##_##library_type##__(#op_type,       \
-                                                           #library_type); \
-  int TouchOpKernelRegistrar_##op_type##_##library_type() {                \
-    __op_kernel_registrar_##op_type##_##library_type##__.Touch();          \
-    return 0;                                                              \
+#define REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(op_type, library_type,             \
+                                            place_class, customized_name,      \
+                                            customized_type_value, ...)        \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                              \
+      __reg_op_kernel_##op_type##_##library_type##_##customized_name##__,      \
+                                 "REGISTER_OP_KERNEL must be called in "       \
+                                 "global namespace");                          \
+  static ::paddle::framework::OpKernelRegistrar<place_class,                   \
+                                                __VA_ARGS__>                   \
+      __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__(\
+          #op_type, #library_type, customized_type_value);                     \
+  int TouchOpKernelRegistrar_##op_type##_##library_type##_##customized_name() {\
+    __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__   \
+        .Touch();                                                              \
+    return 0;                                                                  \
   }
 
+#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...)   \
+  REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(                                \
+      op_type, library_type, place_class, DEFAULT_TYPE,               \
+      ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
+      __VA_ARGS__)
+
 #define REGISTER_OP_CUDA_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__)
 
 #define REGISTER_OP_CPU_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
-#define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class, ...)      \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                           \
-      __reg_op_kernel_##op_type##_##library_type##__,                       \
-      "REGISTER_OP_KERNEL_EX must be called in global namespace");          \
-  static ::paddle::framework::OpKernelRegistrarEx<place_class, __VA_ARGS__> \
-      __op_kernel_registrar_##op_type##_##library_type##__(#op_type,        \
-                                                           #library_type);  \
-  int TouchOpKernelRegistrar_##op_type##_##library_type() {                 \
-    __op_kernel_registrar_##op_type##_##library_type##__.Touch();           \
-    return 0;                                                               \
+#define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class,  \
+                              customized_name,                     \
+                              customized_type_value,               \
+                              ...)                                 \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                  \
+      __reg_op_kernel_##op_type##_##library_type##_##customized_name##__, \
+                                 "REGISTER_OP_KERNEL_EX must be called in "  \
+                                 "global namespace");  \
+  static ::paddle::framework::OpKernelRegistrarEx<place_class,  \
+                                                  __VA_ARGS__>  \
+      __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__(\
+          #op_type, #library_type, customized_type_value);  \
+  int TouchOpKernelRegistrar_##op_type##_##library_type##_##customized_name() {\
+    __op_kernel_registrar_##op_type##_##library_type##_##customized_name##__   \
+        .Touch();                                                              \
+    return 0;                                                                  \
   }
 
 #define REGISTER_OP_CUDA_KERNEL_FUNCTOR(op_type, ...)                 \
-  REGISTER_OP_KERNEL_EX(op_type, CUDA, ::paddle::platform::CUDAPlace, \
-                        __VA_ARGS__)
+  REGISTER_OP_KERNEL_EX(                                              \
+      op_type, CUDA, ::paddle::platform::CUDAPlace, DEFAULT_TYPE,     \
+      ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
+      __VA_ARGS__)
 
-#define REGISTER_OP_CPU_KERNEL_FUNCTOR(op_type, ...) \
-  REGISTER_OP_KERNEL_EX(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
+#define REGISTER_OP_CPU_KERNEL_FUNCTOR(op_type, ...)                  \
+  REGISTER_OP_KERNEL_EX(                                              \
+      op_type, CPU, ::paddle::platform::CPUPlace, DEFAULT_TYPE,       \
+      ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
+      __VA_ARGS__)
 
 /**
  * Macro to mark what Operator and Kernel
@@ -248,13 +281,19 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
   extern int TouchOpRegistrar_##op_type();                 \
   UNUSED static int use_op_itself_##op_type##_ = TouchOpRegistrar_##op_type()
 
-#define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE)               \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                 \
-      __use_op_kernel_##op_type##_##LIBRARY_TYPE##__,             \
-      "USE_OP_DEVICE_KERNEL must be in global namespace");        \
-  extern int TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE(); \
-  UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_ = \
-      TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE()
+#define USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(op_type,                     \
+                                              LIBRARY_TYPE,                \
+                                              customized_name)             \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                          \
+      __use_op_kernel_##op_type##_##LIBRARY_TYPE##_##customized_name##__,  \
+      "USE_OP_DEVICE_KERNEL must be in global namespace");                 \
+  extern int                                                               \
+      TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE##_##customized_name(); \
+  UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_##DEFAULT_TYPE##_ = /* NOLINT */ \
+      TouchOpKernelRegistrar_##op_type##_##LIBRARY_TYPE##_##customized_name()
+
+#define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE) \
+  USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(op_type, LIBRARY_TYPE, DEFAULT_TYPE)
 
 // TODO(fengjiayi): The following macros
 // seems ugly, do we have better method?
@@ -280,6 +319,7 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
 #define USE_OP(op_type)   \
   USE_OP_ITSELF(op_type); \
   USE_OP_KERNEL(op_type)
+// clang-format off
 
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/fluid/framework/operator_test.cc b/paddle/fluid/framework/operator_test.cc
index ac9dd8245ad4e0e8842f219b23d3866b03fdaedb..ab14732e4d6eab9dd15364da02b436c10ed68a19 100644
--- a/paddle/fluid/framework/operator_test.cc
+++ b/paddle/fluid/framework/operator_test.cc
@@ -50,6 +50,8 @@ class OpWithoutKernelCheckerMaker : public OpProtoAndCheckerMaker {
     AddInput("input", "input of test op");
     AddOutput("output", "output of test op");
     AddAttr<float>("scale", "scale of cosine op");
+    AddAttr<int>("kernel_sub_type", "kernels with different implementations.")
+        .SetDefault(0);
     AddComment("This is test op");
   }
 };
@@ -95,6 +97,8 @@ TEST(OperatorBase, all) {
 namespace paddle {
 namespace framework {
 
+static int special_type_value = 1;
+
 class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
   void Make() {
@@ -103,11 +107,14 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .GreaterThan(0.0);
+    AddAttr<int>("kernel_sub_type", "kernels with different implementations.")
+        .SetDefault(0);
     AddComment("This is test op");
   }
 };
 
 static int cpu_kernel_run_num = 0;
+static int cpu_kernel2_run_num = 0;
 
 class OpWithKernelTest : public OperatorWithKernel {
  public:
@@ -117,7 +124,10 @@ class OpWithKernelTest : public OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override {}
   OpKernelType GetExpectedKernelType(
       const ExecutionContext& ctx) const override {
-    return OpKernelType(proto::VarType::FP32, ctx.GetPlace());
+    int sub_type = ctx.Attr<int>("kernel_sub_type");
+    return OpKernelType(proto::VarType::FP32, ctx.GetPlace(),
+                        framework::DataLayout::kAnyLayout,
+                        framework::LibraryType::kPlain, sub_type);
   }
 };
 
@@ -132,6 +142,17 @@ class CPUKernelTest : public OpKernel<float> {
   }
 };
 
+template <typename T1, typename T2>
+class CPUKernel2Test : public OpKernel<float> {
+ public:
+  void Compute(const ExecutionContext& ctx) const {
+    std::cout << ctx.op().DebugString() << std::endl;
+    cpu_kernel2_run_num++;
+    ASSERT_EQ(ctx.op().Input("x"), "IN1");
+    ASSERT_EQ(ctx.op().Output("y"), "OUT1");
+  }
+};
+
 class OpKernelTestMultiInputsProtoAndCheckerMaker
     : public OpProtoAndCheckerMaker {
  public:
@@ -142,6 +163,8 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .GreaterThan(0.0);
+    AddAttr<int>("kernel_sub_type", "kernels with different implementations.")
+        .SetDefault(0);
     AddComment("This is test op");
   }
 };
@@ -189,9 +212,15 @@ class CPUKernalMultiInputsTest : public OpKernel<float> {
 REGISTER_OP_WITHOUT_GRADIENT(
     op_with_kernel, paddle::framework::OpWithKernelTest,
     paddle::framework::OpKernelTestProtoAndCheckerMaker);
+
 REGISTER_OP_CPU_KERNEL(op_with_kernel,
                        paddle::framework::CPUKernelTest<float, float>);
 
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(
+    op_with_kernel, CPU, paddle::platform::CPUPlace, MY_SPECIAL_NAME,
+    paddle::framework::special_type_value,
+    paddle::framework::CPUKernel2Test<float, float>);
+
 // test with single input
 TEST(OpKernel, all) {
   paddle::framework::InitDevices(true);
@@ -211,7 +240,19 @@ TEST(OpKernel, all) {
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0);
   op->Run(scope, cpu_place);
+  // kerne_sub_type = 0, hence cpu_kernel is called, cpu_kernel2 is not called.
+  ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
+  ASSERT_EQ(paddle::framework::cpu_kernel2_run_num, 0);
+
+  attr = op_desc.mutable_attrs()->Add();
+  attr->set_name("kernel_sub_type");
+  attr->set_type(paddle::framework::proto::AttrType::INT);
+  attr->set_i(1);
+  auto op2 = paddle::framework::OpRegistry::CreateOp(op_desc);
+  op2->Run(scope, cpu_place);
+  // kerne_sub_type = 1, hence cpu_kernel2 is called, cpu_kernel is not called.
   ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
+  ASSERT_EQ(paddle::framework::cpu_kernel2_run_num, 1);
 }
 
 REGISTER_OP_WITHOUT_GRADIENT(
diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc
index 05e268bf6a8d9a2562a4c278d317f75dac28e52c..ce45dd58419ab20cccf00544288b79d869515578 100644
--- a/paddle/fluid/operators/conv_mkldnn_op.cc
+++ b/paddle/fluid/operators/conv_mkldnn_op.cc
@@ -491,8 +491,12 @@ class ConvMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_KERNEL(conv2d, MKLDNN, ::paddle::platform::CPUPlace,
-                   ops::ConvMKLDNNOpKernel<float>);
-
-REGISTER_OP_KERNEL(conv2d_grad, MKLDNN, ::paddle::platform::CPUPlace,
-                   ops::ConvMKLDNNGradOpKernel<float>);
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN,
+                                    ::paddle::platform::CPUPlace, FP32,
+                                    ops::kConvMKLDNNFP32,
+                                    ops::ConvMKLDNNOpKernel<float>);
+
+REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(conv2d_grad, MKLDNN,
+                                    ::paddle::platform::CPUPlace, FP32,
+                                    ops::kConvMKLDNNFP32,
+                                    ops::ConvMKLDNNGradOpKernel<float>);
diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc
index 342525be49e28f1785e25d4daad38c3c81b4774f..7455b9492f054b32ee7fb1fc90b1a344367ceb81 100644
--- a/paddle/fluid/operators/conv_op.cc
+++ b/paddle/fluid/operators/conv_op.cc
@@ -74,6 +74,8 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
 
 framework::OpKernelType ConvOp::GetExpectedKernelType(
     const framework::ExecutionContext& ctx) const {
+  int customized_type_value =
+      framework::OpKernelType::kDefaultCustomizedTypeValue;
   framework::LibraryType library{framework::LibraryType::kPlain};
   // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
   std::string data_format = ctx.Attr<std::string>("data_format");
@@ -89,6 +91,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
       platform::CanMKLDNNBeUsed(ctx)) {
     library = framework::LibraryType::kMKLDNN;
     layout = framework::DataLayout::kMKLDNN;
+    customized_type_value = kConvMKLDNNFP32;
   }
 #endif
 
@@ -105,7 +108,7 @@ framework::OpKernelType ConvOp::GetExpectedKernelType(
   }
 
   return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
-                                 library);
+                                 library, customized_type_value);
 }
 
 void Conv2DOpMaker::Make() {
@@ -342,6 +345,8 @@ void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
 
 framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
     const framework::ExecutionContext& ctx) const {
+  int customized_type_value =
+      framework::OpKernelType::kDefaultCustomizedTypeValue;
   framework::LibraryType library_{framework::LibraryType::kPlain};
   // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
   std::string data_format = ctx.Attr<std::string>("data_format");
@@ -357,12 +362,13 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
       platform::CanMKLDNNBeUsed(ctx)) {
     library_ = framework::LibraryType::kMKLDNN;
     layout_ = framework::DataLayout::kMKLDNN;
+    customized_type_value = kConvMKLDNNFP32;
   }
 #endif
 
   return framework::OpKernelType(
       framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
-      layout_, library_);
+      layout_, library_, customized_type_value);
 }
 
 }  // namespace operators
diff --git a/paddle/fluid/operators/conv_op.h b/paddle/fluid/operators/conv_op.h
index e69814001e4da5d10e51ee57c1dbe291338b8b49..249f308c13ff5636fbaa6747b28cab7886b7e736 100644
--- a/paddle/fluid/operators/conv_op.h
+++ b/paddle/fluid/operators/conv_op.h
@@ -27,6 +27,8 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+constexpr int kConvMKLDNNFP32 = 1;
+constexpr int kConvMKLDNNINT8 = 2;
 
 // Base convolution operator definations for other conv
 // like operators to reuse the implementation.