fix kernel chosee. test=develop

lite/backends/arm/math/CMakeLists.txt

@@ -80,6 +80,7 @@ if (NOT HAS_ARM_MATH_LIB_DIR)
       conv3x3s2_depthwise_int8.cc
       conv5x5s1_depthwise_int8.cc
       conv5x5s1_depthwise_fp32.cc
+      conv5x5s1_depthwise_fp32_c4.cc
       conv5x5s2_depthwise_int8.cc
       conv5x5s2_depthwise_fp32.cc
       conv3x3_winograd_fp32_c4.cc

lite/backends/arm/math/conv_impl.cc

@@ -735,6 +735,7 @@ void conv_depthwise_5x5_fp32(const void* din,
   bool flag_relu = param.fuse_relu;
   bool flag_bias = param.bias != nullptr;
   ctx->ExtendWorkspace((w_in + w_out) * sizeof(float));
+  bool flag_act = param.act_param.has_active;
   if (stride == 2) {
     conv_depthwise_5x5s2_fp32(reinterpret_cast<const float*>(din),
                               reinterpret_cast<float*>(dout),
@@ -751,7 +752,7 @@ void conv_depthwise_5x5_fp32(const void* din,
                               act_param,
                               ctx);
   } else if (stride == 1) {
-    if (h_in < 5 || w_in < 5) {
+    if (h_in < 5 || w_in < 5 || flag_act) {
       conv_depthwise_5x5s1_fp32(reinterpret_cast<float*>(dout),
                                 reinterpret_cast<const float*>(din),
                                 reinterpret_cast<const float*>(weights),

lite/kernels/arm/conv_depthwise.cc

@@ -58,7 +58,8 @@ void DepthwiseConv<PRECISION(kFloat), PRECISION(kFloat)>::PrepareForRun() {
     auto strides = param.strides;
     auto hin = param.x->dims()[2];
     auto win = param.x->dims()[3];
-    if (win >= kw && hin >= kw && (strides[0] == 1 && strides[1] == 1)) {
+    bool flag_act = param.act_param.has_active;
+    if (win >= kw && hin >= kw && (strides[0] == 1 && strides[1] == 1) && !flag_act) {
       flag_trans_weights_ = false;
       impl_ = lite::arm::math::conv_depthwise_5x5_fp32;
 #ifdef LITE_WITH_PROFILE