Finish conv2d 3x3 stride 2 neon kernel.

mace/ops/conv_2d_benchmark.cc

@@ -76,6 +76,10 @@ BM_CONV_2D(1, 64, 32, 32, 3, 3, 1, VALID, 128, float);
 BM_CONV_2D(1, 64, 33, 31, 3, 3, 1, VALID, 128, float);
 BM_CONV_2D(1, 64, 32, 32, 3, 3, 1, SAME, 128, float);
 BM_CONV_2D(1, 64, 33, 31, 3, 3, 1, SAME, 128, float);
+BM_CONV_2D(1, 64, 32, 32, 3, 3, 2, VALID, 128, float);
+BM_CONV_2D(1, 64, 33, 31, 3, 3, 2, VALID, 128, float);
+BM_CONV_2D(1, 64, 32, 32, 3, 3, 2, SAME, 128, float);
+BM_CONV_2D(1, 64, 33, 31, 3, 3, 2, SAME, 128, float);
 BM_CONV_2D(1, 64, 32, 32, 5, 5, 1, VALID, 128, float);
 BM_CONV_2D(1, 64, 32, 31, 5, 5, 1, VALID, 128, float);
 BM_CONV_2D(1, 64, 32, 32, 5, 5, 1, SAME, 128, float);

mace/ops/conv_2d_test.cc

@@ -174,8 +174,8 @@ TEST_F(Conv2dOpTest, ConvNxNS12) {
     // generate random input
     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 height = 11 + rand() % 100;
+    index_t width = 11 + rand() % 100;
     index_t output_channels = 1 + rand() % 50;
     // Construct graph
     auto& net = test_net();