diff --git a/src/operators/kernel/arm/conv_kernel.cpp b/src/operators/kernel/arm/conv_kernel.cpp
index f7f55b790db9cf8151b792129cfa858075dbe9a0..de19127e68361bd51f25a15c4c7ab69639707433 100644
--- a/src/operators/kernel/arm/conv_kernel.cpp
+++ b/src/operators/kernel/arm/conv_kernel.cpp
@@ -57,6 +57,8 @@ bool ConvKernel<CPU, float>::Init(ConvParam<CPU> *param) {
                param->Strides()[0] == 2 && param->Paddings()[0] == 1 &&
                param->Paddings()[0] == param->Paddings()[1]) {
       param->ExecMode() = ConvParam<CPU>::EXEC_DEPTHWISE3x3S2P1_FLOAT;
+    } else if (depth3x3) {
+      param->ExecMode() = ConvParam<CPU>::EXEC_DEPTHWISE3x3_FLOAT;
 #ifndef __aarch64__
     } else if (depth5x5 && param->Strides()[0] == param->Strides()[1] &&
                param->Strides()[0] == 1) {
@@ -106,6 +108,10 @@ void ConvKernel<CPU, float>::Compute(const ConvParam<CPU> &param) {
       math::DepthwiseConv3x3s2p0(param.Input(), param.Filter(), param.Output(),
                                  nullptr, false, false);
       break;
+    case ConvParam<CPU>::EXEC_DEPTHWISE3x3_FLOAT:
+      math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
+                             param.Filter(), nullptr, param.Output(), false);
+      break;
 #ifndef __aarch64__
     case ConvParam<CPU>::EXEC_DEPTHWISE5x5_FLOAT:
       DepthwiseConv5x5<float, float>(param);
diff --git a/src/operators/kernel/central-arm-func/conv_arm_func.h b/src/operators/kernel/central-arm-func/conv_arm_func.h
index 86a3c7a9694e8d686f41911ea4af784a33c2cd0a..93be71f554d86b554e6b8ba07b2341b675052785 100644
--- a/src/operators/kernel/central-arm-func/conv_arm_func.h
+++ b/src/operators/kernel/central-arm-func/conv_arm_func.h
@@ -90,8 +90,8 @@ inline void GemmConv(const ConvParam<CPU> &param) {
       Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
 
       if (!is_expand) {
-        col.ShareDataWith(in_slice);
-        col_matrix.ShareDataWith(col);
+        // col_matrix.ShareDataWith(in_slice);
+        col_matrix = in_slice;
         col_matrix.Resize(col_matrix_shape);
       } else if (data_dim == 2U) {
         // im2col