diff --git a/paddle/fluid/framework/op_registry.h b/paddle/fluid/framework/op_registry.h
index a1f07f9f25..179df3b981 100644
--- a/paddle/fluid/framework/op_registry.h
+++ b/paddle/fluid/framework/op_registry.h
@@ -178,9 +178,8 @@ struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelTypes...> {
     RegisterKernelClass<PlaceType, T>(
         op_type, library_type, customized_type_value,
 
-        [op_type](const framework::ExecutionContext& ctx) {
+        [](const framework::ExecutionContext& ctx) {
           KERNEL_TYPE().Compute(ctx);
-          CheckKernelLaunch<PlaceType>(op_type);
         });
     constexpr auto size = std::tuple_size<std::tuple<KernelTypes...>>::value;
     OpKernelRegistrarFunctor<PlaceType, I + 1 == size, I + 1, KernelTypes...>
@@ -240,13 +239,8 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
 
   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,
-
-        [op_type](const framework::ExecutionContext& ctx) {
-          Functor()(ctx);
-          CheckKernelLaunch<PlaceType>(op_type);
-        });
+    RegisterKernelClass<PlaceType, T>(op_type, library_type,
+                                      customized_type_value, Functor());
 
     constexpr auto size =
         std::tuple_size<std::tuple<DataTypeAndKernelType...>>::value;
@@ -275,7 +269,7 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
     VarTypeInference
     InferShapeBase
 */
-#define REGISTER_OPERATOR(op_type, op_class, ...)                        \
+#define REGISTER_OPERATOR__(op_type, op_class, ...)                        \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                        \
       __reg_op__##op_type,                                               \
       "REGISTER_OPERATOR must be called in global namespace");           \
@@ -286,15 +280,22 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
     return 0;                                                            \
   }
 
-#define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, ...) \
+#define REGISTER_OPERATOR(op_type, op_class, ...)
+
+#define REGISTER_OP_WITHOUT_GRADIENT__(op_type, op_class, ...) \
   REGISTER_OPERATOR(op_type, op_class, __VA_ARGS__, \
         paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,   \
         paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>)
 
+#define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, ...)
 /**
  * Macro to register OperatorKernel.
  */
 #define REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(op_type, library_type,             \
+                                            place_class, customized_name,      \
+                                            customized_type_value, ...)
+
+#define REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE__(op_type, library_type,             \
                                             place_class, customized_name,      \
                                             customized_type_value, ...)        \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                              \
@@ -311,18 +312,22 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
     return 0;                                                                  \
   }
 
-#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...)   \
-  REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(                                \
+#define REGISTER_OP_KERNEL(op_type, library_type, place_class, ...)
+
+#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__)
 
-#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+#define REGISTER_OP_CUDA_KERNEL__(op_type, ...) \
+  REGISTER_OP_KERNEL__(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__)
+
 #define REGISTER_OP_CUDA_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::CUDAPlace, __VA_ARGS__)
-#else
-#define REGISTER_OP_CUDA_KERNEL(op_type, ...)
-#endif
+
+#define REGISTER_OP_CPU_KERNEL__(op_type, ...) \
+  REGISTER_OP_KERNEL__(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
 #define REGISTER_OP_CPU_KERNEL(op_type, ...) \
   REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
@@ -340,6 +345,11 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
   REGISTER_OP_KERNEL(op_type, MLU, ::paddle::platform::MLUPlace, __VA_ARGS__)
 
 #define REGISTER_OP_KERNEL_EX(op_type, library_type, place_class,  \
+                              customized_name,                     \
+                              customized_type_value,               \
+			      ...)
+
+#define REGISTER_OP_KERNEL_EX__(op_type, library_type, place_class,  \
                               customized_name,                     \
                               customized_type_value,               \
                               ...)                                 \
@@ -357,8 +367,10 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
     return 0;                                                                  \
   }
 
-#define REGISTER_OP_CUDA_KERNEL_FUNCTOR(op_type, ...)                 \
-  REGISTER_OP_KERNEL_EX(                                              \
+#define REGISTER_OP_CUDA_KERNEL_FUNCTOR(op_type, ...)
+
+#define REGISTER_OP_CUDA_KERNEL_FUNCTOR__(op_type, ...)                 \
+  REGISTER_OP_KERNEL_EX__(                                              \
       op_type, CUDA, ::paddle::platform::CUDAPlace, DEFAULT_TYPE,     \
       ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
       __VA_ARGS__)
@@ -375,12 +387,6 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
       ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
       __VA_ARGS__)
 
-#define REGISTER_OP_NPU_KERNEL_FUNCTOR(op_type, ...)                  \
-  REGISTER_OP_KERNEL_EX(                                              \
-      op_type, NPU, ::paddle::platform::NPUPlace, DEFAULT_TYPE,       \
-      ::paddle::framework::OpKernelType::kDefaultCustomizedTypeValue, \
-      __VA_ARGS__)
-
 #define REGISTER_OP_MLU_KERNEL_FUNCTOR(op_type, ...)                  \
   REGISTER_OP_KERNEL_EX(                                              \
       op_type, MLU, ::paddle::platform::MLUPlace, DEFAULT_TYPE,       \
@@ -392,7 +398,9 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
  * we will use and tell the compiler to
  * link them into target.
  */
-#define USE_OP_ITSELF(op_type)                             \
+#define USE_OP_ITSELF(op_type)
+
+#define USE_OP_ITSELF__(op_type)                             \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                          \
       __use_op_itself_##op_type,                           \
       "USE_OP_ITSELF must be called in global namespace"); \
@@ -400,6 +408,10 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
   UNUSED static int use_op_itself_##op_type##_ = TouchOpRegistrar_##op_type()
 
 #define USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(op_type,                     \
+                                              LIBRARY_TYPE,                \
+                                              customized_name)
+
+#define USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE__(op_type,                     \
                                               LIBRARY_TYPE,                \
                                               customized_name)             \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                          \
@@ -410,33 +422,58 @@ struct OpKernelRegistrarFunctorEx<PlaceType, false, I,
   UNUSED static int use_op_kernel_##op_type##_##LIBRARY_TYPE##_##customized_name##_ = /* 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)
+#define USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(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)
+
+#define USE_OP_DEVICE_KERNEL(op_type, LIBRARY_TYPE)
 
 // TODO(fengjiayi): The following macros
 // seems ugly, do we have better method?
 
-#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
+#ifndef PADDLE_WITH_CUDA
 #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)
+#define USE_OP_KERNEL__(op_type) USE_OP_DEVICE_KERNEL__(op_type, CPU)
 #else
 #define USE_OP_KERNEL(op_type)        \
   USE_OP_DEVICE_KERNEL(op_type, CPU); \
   USE_OP_DEVICE_KERNEL(op_type, CUDA)
+
+#define USE_OP_KERNEL__(op_type)        \
+  USE_OP_DEVICE_KERNEL__(op_type, CPU); \
+  USE_OP_DEVICE_KERNEL__(op_type, CUDA)
 #endif
 
 #define USE_NO_KERNEL_OP(op_type) USE_OP_ITSELF(op_type);
 
+#define USE_NO_KERNEL_OP__(op_type) USE_OP_ITSELF__(op_type);
+
 #define USE_CPU_ONLY_OP(op_type) \
   USE_OP_ITSELF(op_type);        \
   USE_OP_DEVICE_KERNEL(op_type, CPU);
 
+#define USE_CPU_ONLY_OP__(op_type) \
+  USE_OP_ITSELF__(op_type);        \
+  USE_OP_DEVICE_KERNEL__(op_type, CPU);
+
 #define USE_CUDA_ONLY_OP(op_type) \
   USE_OP_ITSELF(op_type);         \
   USE_OP_DEVICE_KERNEL(op_type, CUDA)
 
+#define USE_CUDA_ONLY_OP__(op_type) \
+  USE_OP_ITSELF__(op_type);         \
+  USE_OP_DEVICE_KERNEL__(op_type, CUDA)
+
 #define USE_OP(op_type)   \
   USE_OP_ITSELF(op_type); \
   USE_OP_KERNEL(op_type)
+
+#define USE_OP__(op_type)   \
+  USE_OP_ITSELF__(op_type); \
+  USE_OP_KERNEL__(op_type)
 // clang-format on
 
 }  // namespace framework