diff --git a/mace/kernels/neon/conv_2d_neon.cc b/mace/kernels/neon/conv_2d_neon.cc
index 9891eae992993d0230d8d9b933920b51292f104e..06d3b3e75487b5c478343d3457ad55912febb153 100644
--- a/mace/kernels/neon/conv_2d_neon.cc
+++ b/mace/kernels/neon/conv_2d_neon.cc
@@ -31,9 +31,9 @@ static inline void ConstructInputWithPadding(const float* input,
 
   // Skip the padded top rows
   output_ptr += padded_top * output_width;
-  for (; batch > 0; --batch) {
-    for (; channels > 0; --channels) {
-      for(; height > 0; --height) {
+  for (int i = 0; i < batch; ++i) {
+    for (int j = 0; j < channels; ++j) {
+      for (int k = 0; k < height; ++k) {
         memcpy(output_ptr + padded_left, input, width * sizeof(float));
         input += width;
         output_ptr += output_width;
diff --git a/mace/ops/conv_2d_test.cc b/mace/ops/conv_2d_test.cc
index 32a4e50db53c9e0d6c0e80a09505e01de0fc5f19..40d6e3a68f2f2756527a42ea1eaf1f4d5b858e58 100644
--- a/mace/ops/conv_2d_test.cc
+++ b/mace/ops/conv_2d_test.cc
@@ -199,11 +199,11 @@ TEST_F(Conv2dOpTest, ConvNxNS12) {
     srand(time(NULL));
 
     // generate random input
-    index_t batch = 1 + rand() % 5;
-    index_t input_channels = 3 + rand() % 50;
-    index_t height = 10 + rand() % 100;
-    index_t width = 10 + rand() % 100;
-    index_t output_channels = 3 + rand() % 50;
+    index_t batch = 1 + rand() % 10;
+    index_t input_channels = 1 + rand() % 50;
+    index_t height = 7 + rand() % 100;
+    index_t width = 7 + rand() % 100;
+    index_t output_channels = 1 + rand() % 50;
     // Construct graph
     auto& net = test_net();
     OpDefBuilder("Conv2d", "Conv2dTest")
@@ -236,11 +236,11 @@ TEST_F(Conv2dOpTest, ConvNxNS12) {
     // Run NEON
     net.RunOp(DeviceType::NEON);
 
-    ExpectTensorNear<float>(expected, *net.GetOutput("Output"), 1e-5);
+    ExpectTensorNear<float>(expected, *net.GetOutput("Output"), 1e-3);
 
   };
 
-  for (int kernel_size : {1}) { // TODO(liu1i10) 3x3
+  for (int kernel_size : {1, 3}) {
     for (int stride : {1, 2}) {
       func(kernel_size, kernel_size, stride, stride, VALID);
       func(kernel_size, kernel_size, stride, stride, SAME);