feat(dnn/arm): add armv7 nchw_nchw44 3x3s2 asm kernel

GitOrigin-RevId: 50ce91e41d200d355a5b142f33d5e01f07544650

dnn/src/arm_common/handle.h

@@ -10,6 +10,9 @@
  */
 #pragma once
 #include "src/fallback/handle.h"
+#if MGB_ENABLE_CPUINFO
+#include "cpuinfo.h"
+#endif
 
 namespace megdnn {
 namespace arm_common {
@@ -20,6 +23,9 @@ class HandleImpl: public fallback::HandleImpl {
                 HandleType type = HandleType::ARM_COMMON):
             fallback::HandleImpl::HandleImpl(computing_handle, type)
         {
+            #if MGB_ENABLE_CPUINFO
+            cpuinfo_initialize();
+            #endif
         }
 
         template <typename Opr>

dnn/test/arm_common/conv_bias.cpp

@@ -243,6 +243,7 @@ static void benchmark_convbias(Handle* handle, std::string int_name,
 
     if (is_fp32) {
         run(1, 1, 4, 112, 112, 2, 2, true);
+        run(1, 3, 24, 224, 224, 3, 2, true);
         run(1, 3, 32, 224, 224, 3, 2, true);
         run(1, 3, 64, 224, 224, 7, 2, true);