conv7x7s2 pack1to4 bf16s neon kernel

src/layer/arm/convolution_arm.cpp

@@ -51,6 +51,7 @@ namespace ncnn {
 #include "convolution_3x3_pack4_bf16s.h"
 #include "convolution_3x3_pack1to4_bf16s.h"
 #include "convolution_3x3_pack4to1_bf16s.h"
+#include "convolution_7x7_pack1to4_bf16s.h"
 #endif // __ARM_NEON
 
 
@@ -1375,6 +1376,15 @@ int Convolution_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const
                 activation->forward_inplace(top_blob, opt);
             }
         }
+        else if (kernel_w == 7 && kernel_h == 7 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
+        {
+            conv7x7s2_pack1to4_bf16s_neon(bottom_blob_bordered, top_blob, weight_data_pack1to4_bf16, bias_data, opt);
+
+            if (activation)
+            {
+                activation->forward_inplace(top_blob, opt);
+            }
+        }
         else
         {
             // num_output

src/layer/arm/packing_arm.cpp

@@ -31,7 +31,13 @@ Packing_arm::Packing_arm()
 
 int Packing_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
 {
-    if (opt.use_bf16_storage)
+    size_t elemsize = bottom_blob.elemsize;
+    int elempack = bottom_blob.elempack;
+
+    bool elemtype_is_bf16 = (elemsize == 2u && elempack == 1) || (elemsize == 8u && elempack == 4);
+    bool elemtype_is_fp32 = (elemsize == 4u && elempack == 1) || (elemsize == 16u && elempack == 4);
+
+    if (opt.use_bf16_storage && elemtype_is_bf16)
         return forward_bf16s(bottom_blob, top_blob, opt);
 
     if (use_padding)
@@ -39,8 +45,11 @@ int Packing_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
         return Packing::forward(bottom_blob, top_blob, opt);
     }
 
-    size_t elemsize = bottom_blob.elemsize;
-    int elempack = bottom_blob.elempack;
+    if (!elemtype_is_fp32)
+    {
+        // non-fp32 type
+        return Packing::forward(bottom_blob, top_blob, opt);
+    }
 
     if (elempack == out_elempack)
     {