diff --git a/mace/kernels/addn.h b/mace/kernels/addn.h
index c8bb601620f1965b126bc39d2ef2259d26e91e68..61e906bf575f408f7425b6d80414f75008ddec57 100644
--- a/mace/kernels/addn.h
+++ b/mace/kernels/addn.h
@@ -77,12 +77,6 @@ struct AddNFunctor {
   }
 };
 
-template <>
-void AddNFunctor<DeviceType::NEON, float>::operator()(
-    const std::vector<const Tensor *> &input_tensors,
-    Tensor *output_tensor,
-    StatsFuture *future);
-
 template <typename T>
 struct AddNFunctor<DeviceType::OPENCL, T> {
   void operator()(const std::vector<const Tensor *> &input_tensors,
diff --git a/mace/ops/addn.cc b/mace/ops/addn.cc
index 8e253cf1dbbd45791e72b65352af60d389e44b87..e30275bb5ac95a7ad15a227190490c5ccd4e495a 100644
--- a/mace/ops/addn.cc
+++ b/mace/ops/addn.cc
@@ -25,6 +25,12 @@ void Register_AddN(OperatorRegistry *op_registry) {
                                      .TypeConstraint<half>("T")
                                      .Build(),
                     AddNOp<DeviceType::OPENCL, half>);
+
+  REGISTER_OPERATOR(op_registry, OpKeyBuilder("AddN")
+                                     .Device(DeviceType::NEON)
+                                     .TypeConstraint<float>("T")
+                                     .Build(),
+                    AddNOp<DeviceType::NEON, float>);
 }
 
 }  // namespace ops