diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index 0db67e4c6785250c0f4c03b1c698ecffcd52372b..804f901dfa278384eb281a2650ae7aa724639fb9 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -100,14 +100,38 @@ class OpRegistrar : public Registrar {
   }
 };
 
-template <typename PlaceType, typename KernelType>
+template <typename PlaceType, bool at_end, size_t I, typename... KernelType>
+struct OpKernelRegistrarFunctor;
+
+template <typename PlaceType, size_t I, typename... KernelType>
+struct OpKernelRegistrarFunctor<PlaceType, false, I, KernelType...> {
+  using KT = typename std::tuple_element<I, std::tuple<KernelType...>>::type;
+
+  void operator()(const char* op_type) const {
+    using T = typename KT::ELEMENT_TYPE;
+    OperatorWithKernel::OpKernelKey key(ToDataType(std::type_index(typeid(T))),
+                                        PlaceType());
+    OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KT);
+
+    constexpr auto size = std::tuple_size<std::tuple<KernelType...>>::value;
+    OpKernelRegistrarFunctor<PlaceType, I + 1 == size, I + 1, KernelType...>
+        func;
+    func(op_type);
+  }
+};
+
+template <typename PlaceType, size_t I, typename... KernelType>
+struct OpKernelRegistrarFunctor<PlaceType, true, I, KernelType...> {
+  void operator()(const char* op_type) const {}
+};
+
+// User can register many kernel in one place. The data type could be different.
+template <typename PlaceType, typename... KernelType>
 class OpKernelRegistrar : public Registrar {
  public:
   explicit OpKernelRegistrar(const char* op_type) {
-    using T = typename KernelType::ELEMENT_TYPE;
-    OperatorWithKernel::OpKernelKey key(ToDataType(std::type_index(typeid(T))),
-                                        PlaceType());
-    OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KernelType);
+    OpKernelRegistrarFunctor<PlaceType, false, 0, KernelType...> func;
+    func(op_type);
   }
 };
 
diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc
index bda5dfe03e974740fe4a07191ae6b68ebfcd5d3a..da7765aa6a7a81c9e0b4f462022cad54c16aec47 100644
--- a/paddle/operators/elementwise_mul_op.cc
+++ b/paddle/operators/elementwise_mul_op.cc
@@ -36,7 +36,9 @@ REGISTER_OP(elementwise_mul, ops::ElementwiseOp, ops::ElementwiseMulOpMaker,
             elementwise_mul_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul,
-    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, double>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul_grad,
-    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, double>);
diff --git a/paddle/operators/elementwise_mul_op.cu b/paddle/operators/elementwise_mul_op.cu
index da08a75596c4d3b89dc8892bd4405464fec96389..056f081d3e6ac349978ff00689700c035bed8e39 100644
--- a/paddle/operators/elementwise_mul_op.cu
+++ b/paddle/operators/elementwise_mul_op.cu
@@ -19,7 +19,9 @@ namespace ops = paddle::operators;
 
 REGISTER_OP_GPU_KERNEL(
     elementwise_mul,
-    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, float>);
+    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, float>,
+    ops::ElementwiseMulKernel<paddle::platform::GPUPlace, double>);
 REGISTER_OP_GPU_KERNEL(
     elementwise_mul_grad,
-    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, float>);
+    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, float>,
+    ops::ElementwiseMulGradKernel<paddle::platform::GPUPlace, double>);