neon optimize for depthwise convolution 5x5 :P

src/layer/arm/convolutiondepthwise_arm.cpp

@@ -23,6 +23,7 @@
 namespace ncnn {
 
 #include "convolutiondepthwise_3x3.h"
+#include "convolutiondepthwise_5x5.h"
 
 #include "convolutiondepthwise_3x3_int8.h"
 
@@ -65,6 +66,13 @@ int ConvolutionDepthWise_arm::load_model(const ModelBin& mb)
                 return 0;
             }
         }
+        if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1)
+        {
+            if ((stride_w == 1 && stride_h == 1) || (stride_w == 2 && stride_h == 2))
+            {
+                return 0;
+            }
+        }
     }
 
     const int channels_g = channels / group;
@@ -243,6 +251,19 @@ int ConvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, con
                     return 0;
                 }
             }
+            if (kernel_w == 5 && kernel_h == 5 && dilation_w == 1 && dilation_h == 1)
+            {
+                if (stride_w == 1 && stride_h == 1)
+                {
+                    convdw5x5s1_neon(bottom_blob_bordered, top_blob, weight_data, bias_data, opt);
+                    return 0;
+                }
+                else if (stride_w == 2 && stride_h == 2)
+                {
+                    convdw5x5s2_neon(bottom_blob_bordered, top_blob, weight_data, bias_data, opt);
+                    return 0;
+                }
+            }
         }
 
         #pragma omp parallel for num_threads(opt.num_threads)