use gemm to s1p0 in conv

src/operators/kernel/central-arm-func/dwconv_bn_relu_arm_func.h

@@ -118,7 +118,8 @@ void DWConvBNReluCompute(const FusionDWConvBNReluParam<CPU> &param) {
   if (param.Groups() == param.Input()->dims()[1] &&
       param.Input()->dims()[1] == param.Output()->dims()[1] &&
       param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
-      param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1) {
+      param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1 &&
+      param.paddings_[0] == 1) {
     math::DepthwiseConvAddBNRelu3x3s1p1(param.Input(), param.Filter(),
                                         param.Output(), param.NewScale(),
                                         param.NewBias(), true);