[ARM]fix conv3x3_dw v7 clang bug (#3129)

* fix v7 clang bug. test=develop

* fix format, test=develop

* fix .a build bug, test=develop

lite/CMakeLists.txt

@@ -187,7 +187,6 @@ if (LITE_WITH_LIGHT_WEIGHT_FRAMEWORK AND LITE_WITH_ARM)
                 add_dependencies(publish_inference tiny_publish_cxx_lib)
                 if(NOT "${CMAKE_BUILD_TYPE}" STREQUAL "Debug")
                     add_custom_command(TARGET tiny_publish_cxx_lib POST_BUILD
-                                COMMAND ${CMAKE_STRIP} "-s" ${INFER_LITE_PUBLISH_ROOT}/cxx/lib/libpaddle_api_light_bundled.a
                                 COMMAND ${CMAKE_STRIP} "-s" ${INFER_LITE_PUBLISH_ROOT}/cxx/lib/libpaddle_light_api_shared.so)
                 endif()
             endif()

lite/kernels/arm/conv_compute.cc

@@ -58,12 +58,13 @@ void ConvCompute<PRECISION(kFloat), PRECISION(kFloat)>::PrepareForRun() {
 
   bool flag_dw_3x3 = (kw == 3) && (kh == 3) && (stride == 1 || stride == 2);
   bool flag_dw_5x5 = (kw == 5) && (kh == 5) && (stride == 1 || stride == 2);
-  bool flag_dw = flag_dw_3x3 || flag_dw_5x5;
 
 #ifdef LITE_WITH_ARM_CLANG  // clang
   flag_dw_3x3 =
       (stride == 1 && (paddings[0] > 1 || paddings[2] > 1)) ? false : true;
 #endif
+  bool flag_dw = flag_dw_3x3 || flag_dw_5x5;
+
   /// select conv impl
   if (param.groups == ic && ic == oc && ks_equal && no_dilation && flag_dw) {
     impl_ = new DepthwiseConv<PRECISION(kFloat), PRECISION(kFloat)>;