Add conv 1x1 unit test.

mace/kernels/opencl/cl/conv_2d_3x3.cl

@@ -33,7 +33,7 @@ void kernel conv_2d_3x3(global const float *input,
   const float *input_base = input + in_offset + in_pixel_begin;
   float *output_base = output + out_offset + out_pixel_begin;
 
-  uint pixels = out_pixel_end - out_pixel_begin;
+  const int pixels = out_pixel_end - out_pixel_begin;
 
   for (int i = out_chan_begin; i < out_chan_end; ++i) {
     float *output_ptr = output_base + i * out_pixel;

mace/kernels/opencl/cl/depthwise_conv_3x3.cl

@@ -35,7 +35,7 @@ void kernel depthwise_conv_3x3(global const float *input, /* n, c, h, w */
   const float *input_base = input + in_offset + in_pixel_begin;
   float *output_base = output + out_offset + out_pixel_begin;
 
-  uint pixels = out_pixel_end - out_pixel_begin;
+  const int pixels = out_pixel_end - out_pixel_begin;
 
   for (int i = out_chan_begin; i < out_chan_end; ++i) {
     float bias_value = bias[i];

mace/ops/conv_2d_test.cc

@@ -296,7 +296,7 @@ static void TestUnalignedConvNxNS12() {
     ExpectTensorNear<float>(expected, *net.GetOutput("Output"), 0.001);
   };
 
-  for (int kernel_size : {3}) {
+  for (int kernel_size : {1, 3, 5}) {
     for (int stride : {1, 2}) {
       func(kernel_size, kernel_size, stride, stride, VALID);
       func(kernel_size, kernel_size, stride, stride, SAME);