diff --git a/paddle/pten/core/utils/array.h b/paddle/pten/core/utils/array.h
index cd43dc7b420f5e015cac2a79ca0e64afe86b29c3..2d6bfbe213baff95b37a32ef14dcd36acef1d726 100644
--- a/paddle/pten/core/utils/array.h
+++ b/paddle/pten/core/utils/array.h
@@ -104,28 +104,22 @@ class Array<T, 0> {
   HOSTDEVICE inline T *GetMutable() { return nullptr; }
 
   HOSTDEVICE inline T &operator[](size_t) {
-#if defined(__HIPCC__)
-    // HIP will have compile error, if use "obj()"
+#if defined(__HIPCC__) || defined(__CUDA_ARCH__)
+    // HIP and CUDA will have compile error, if use "obj()"
     // function declared in block scope cannot have 'static' storage class
     static T obj{};
     return obj;
-#elif defined(__CUDA_ARCH__)
-    static T obj();
-    return obj;
 #else
     PADDLE_THROW(pten::errors::Unavailable("Array<T, 0> has no element."));
 #endif
   }
 
   HOSTDEVICE inline const T &operator[](size_t) const {
-#if defined(__HIPCC__)
-    // HIP will have compile error, if use "obj()"
+#if defined(__HIPCC__) || defined(__CUDA_ARCH__)
+    // HIP and CUDA will have compile error, if use "obj()"
     // function declared in block scope cannot have 'static' storage class
     static const T obj{};
     return obj;
-#elif defined(__CUDA_ARCH__)
-    static const T obj();
-    return obj;
 #else
     PADDLE_THROW(pten::errors::Unavailable("Array<T, 0> has no element."));
 #endif
diff --git a/paddle/pten/kernels/funcs/elementwise_base.h b/paddle/pten/kernels/funcs/elementwise_base.h
index 34f2ab4a62ab945bf57ceaa21d6237845409ad4e..21abf82c7b2283da1a82f18c22d31335ee4d3335 100644
--- a/paddle/pten/kernels/funcs/elementwise_base.h
+++ b/paddle/pten/kernels/funcs/elementwise_base.h
@@ -31,8 +31,6 @@ namespace kps = pten::kps;
 
 #endif
 
-#define BASE_SIZE 1  // To avoid running errors when Arity == 0 in args[Arity]
-
 namespace pten {
 
 enum ElementwiseType { kUnary = 1, kBinary = 2, kTernary = 3, kAny = -1 };
@@ -482,7 +480,7 @@ struct ElementwisePrimitiveCaller<InT, OutT, VecSize, Functor, 0, false> {
   __device__ inline void operator()(Functor func,
                                     InT (*args)[VecSize],
                                     OutT *result) {
-    kps::ElementwiseFillConst<InT, OutT, VecSize, 1, 1, Functor>(result, func);
+    kps::ElementwiseConstant<InT, OutT, VecSize, 1, 1, Functor>(result, func);
   }
 };
 
@@ -560,13 +558,12 @@ template <typename InT,
           bool IsBoundary>
 __device__ void VectorizedElementwiseKernelImpl(
 
-    const pten::framework::Array<const _ptr_ InT *__restrict__,
-                                 Arity + BASE_SIZE> &in,
+    const pten::framework::Array<const _ptr_ InT *__restrict__, Arity> &in,
     pten::framework::Array<_ptr_ OutT *, NumOuts> outs,
     int num,
     int data_offset,
     Functor func) {
-  InT args[Arity + BASE_SIZE][VecSize];
+  InT args[Arity > 1 ? Arity : 1][VecSize];
   ConditionalT<OutT, NumOuts> result[VecSize];
 
 #pragma unroll
@@ -596,8 +593,7 @@ template <typename InT,
           int NumOuts,
           int VecSize>
 __global__ void VectorizedElementwiseKernel(
-    pten::framework::Array<const _ptr_ InT *__restrict__, Arity + BASE_SIZE>
-        ins,
+    pten::framework::Array<const _ptr_ InT *__restrict__, Arity> ins,
     pten::framework::Array<_ptr_ OutT *, NumOuts> outs,
     int size,
     int main_offset,
@@ -637,9 +633,9 @@ void ElementwiseCudaKernel(const KPDevice &ctx,
                            const std::vector<const DenseTensor *> &ins,
                            std::vector<DenseTensor *> *outs,
                            Functor func) {
-  auto numel = (*outs)[0]->numel();
-  pten::framework::Array<const _ptr_ InT *__restrict__, Arity + BASE_SIZE>
-      ins_data;
+  auto numel =
+      (*outs)[0]->numel();  // To avoid running errors when ins.size()== 0
+  pten::framework::Array<const _ptr_ InT *__restrict__, Arity> ins_data;
   pten::framework::Array<_ptr_ OutT *, NumOuts> outs_data;
 
   for (int i = 0; i < Arity; ++i) {
diff --git a/paddle/pten/kernels/gpu/full_kernel.cu b/paddle/pten/kernels/gpu/full_kernel.cu
index 6464dc97d5b7efdb68b50ec8592b4c68e7aa8247..5dbae41e00c950dc6e8caa9da7a4fec13b65574b 100644
--- a/paddle/pten/kernels/gpu/full_kernel.cu
+++ b/paddle/pten/kernels/gpu/full_kernel.cu
@@ -62,10 +62,9 @@ void FullLikeKernel(const ContextT& dev_ctx,
   auto value = val.to<float>();
   using CommonType = typename std::common_type<
       float,
-      typename std::conditional<
-          std::is_same<T, paddle::platform::float16>::value,
-          float,
-          T>::type>::type;
+      typename std::conditional<std::is_same<T, pten::dtype::float16>::value,
+                                float,
+                                T>::type>::type;
 
   auto common_type_value = static_cast<CommonType>(value);
 
@@ -75,7 +74,7 @@ void FullLikeKernel(const ContextT& dev_ctx,
           (common_type_value <=
            static_cast<CommonType>(std::numeric_limits<T>::max())),
       true,
-      paddle::platform::errors::InvalidArgument(
+      pten::errors::InvalidArgument(
           "The filled value is out of range for target type, "
           "current kernel type is %s, the range should between %f "
           "and %f, but now value is %f.",
diff --git a/paddle/pten/kernels/primitive/compute_primitives.h b/paddle/pten/kernels/primitive/compute_primitives.h
index f854cf95aac29708f0dca4810e32a23fedc47e60..449c81b915e6d6de2eecd033fbcd374c5c426292 100644
--- a/paddle/pten/kernels/primitive/compute_primitives.h
+++ b/paddle/pten/kernels/primitive/compute_primitives.h
@@ -420,8 +420,7 @@ template <typename InT,
           int NY,
           int BlockSize,
           class OpFunc>
-__device__ __forceinline__ void ElementwiseFillConst(OutT* out,
-                                                     OpFunc compute) {
+__device__ __forceinline__ void ElementwiseConstant(OutT* out, OpFunc compute) {
 #pragma unroll
   for (int idx = 0; idx < NX * NY; idx++) {
     out[idx] = static_cast<OutT>(compute());
diff --git a/paddle/pten/kernels/primitive/compute_primitives_xpu2.h b/paddle/pten/kernels/primitive/compute_primitives_xpu2.h
index d7282c089fc9cc332abc132941188c7804e68f80..e083280aa2ecc47fb18382e0e77739a09d88720b 100644
--- a/paddle/pten/kernels/primitive/compute_primitives_xpu2.h
+++ b/paddle/pten/kernels/primitive/compute_primitives_xpu2.h
@@ -348,5 +348,18 @@ __device__ __forceinline__ void Reduce(T* out,
   }
 }
 
+template <typename InT,
+          typename OutT,
+          int NX,
+          int NY,
+          int BlockSize,
+          class OpFunc>
+__device__ __forceinline__ void ElementwiseConstant(OutT* out, OpFunc compute) {
+#pragma unroll
+  for (int idx = 0; idx < NX * NY; idx++) {
+    out[idx] = static_cast<OutT>(compute());
+  }
+}
+
 }  // namespace kps
 }  // namespace pten