diff --git a/mindspore/ccsrc/backend/kernel_compiler/gpu/math/addn_gpu_kernel.h b/mindspore/ccsrc/backend/kernel_compiler/gpu/math/addn_gpu_kernel.h
index b692727e3d..218d0c607a 100644
--- a/mindspore/ccsrc/backend/kernel_compiler/gpu/math/addn_gpu_kernel.h
+++ b/mindspore/ccsrc/backend/kernel_compiler/gpu/math/addn_gpu_kernel.h
@@ -115,20 +115,20 @@ class AddNGpuFwdKernel : public GpuKernel {
     for (size_t i = input_shape.size(); i < 4; i++) {
       (void)input_shape.insert(input_shape.begin(), 1);
     }
-    int dimA[4];
+    std::vector<int> dimA;
     for (size_t i = 0; i < input_shape.size(); i++) {
-      dimA[i] = SizeToInt(input_shape[i]);
+      dimA.push_back(SizeToInt(input_shape[i]));
     }
     auto input_format = AnfAlgo::GetInputFormat(kernel_node, 0);
     if (input_format == kOpFormat_NHWC) {
       CHECK_CUDNN_RET_WITH_EXCEPT(kernel_node_,
                                   cudnnSetTensorNdDescriptorEx(input_descriptor_, CUDNN_TENSOR_NHWC, cudnn_data_type_,
-                                                               SizeToInt(input_shape.size()), dimA),
+                                                               SizeToInt(input_shape.size()), dimA.data()),
                                   "cudnnSetTensorNdDescriptor failed");
     } else {
       CHECK_CUDNN_RET_WITH_EXCEPT(kernel_node_,
                                   cudnnSetTensorNdDescriptorEx(input_descriptor_, CUDNN_TENSOR_NCHW, cudnn_data_type_,
-                                                               SizeToInt(input_shape.size()), dimA),
+                                                               SizeToInt(input_shape.size()), dimA.data()),
                                   "cudnnSetTensorNdDescriptor failed");
     }
     InitSizeLists();