diff --git a/src/operators/fusion_conv_add_bn_relu_op.h b/src/operators/fusion_conv_add_bn_relu_op.h
index 753ce395980c9186c948bc1ae4c89c3d2c417fdc..389c76cc83a532fe706d911903a8412bb8bfb4ca 100644
--- a/src/operators/fusion_conv_add_bn_relu_op.h
+++ b/src/operators/fusion_conv_add_bn_relu_op.h
@@ -79,11 +79,11 @@ class FusionConvAddBNReluOp
 
 #ifdef PADDLE_MOBILE_CPU
 
-//#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
-// static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
-//    new FusionConvAddBNReluMatcher());
-//#define FUSION_CONV_ADD_BN_RELU_REGISTER
-//#endif
+#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
+static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
+    new FusionConvAddBNReluMatcher());
+#define FUSION_CONV_ADD_BN_RELU_REGISTER
+#endif
 
 #endif
 
diff --git a/src/operators/kernel/central-arm-func/conv_add_arm_func.h b/src/operators/kernel/central-arm-func/conv_add_arm_func.h
index 71ce3b61a928e252923822e934c65abdc9962a43..9dfe7a6fa8b3979b82e013cd2e7a7e00b21e6a26 100644
--- a/src/operators/kernel/central-arm-func/conv_add_arm_func.h
+++ b/src/operators/kernel/central-arm-func/conv_add_arm_func.h
@@ -152,9 +152,15 @@ void ConvAddCompute(const FusionConvAddParam &param) {
   } else if (param.Groups() == param.Input()->dims()[1] &&
              param.Input()->dims()[1] == param.Output()->dims()[1] &&
              param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
-             param.Filter()->dims()[2] == 3) {
-    math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
-                           param.Filter(), param.Bias(), param.Output(), true);
+             param.Filter()->dims()[2] == 3 && param.Strides()[0] == 2) {
+    //        math::DepthwiseConv3x3(param.Input(), param.Strides(),
+    //        param.Paddings(),
+    //                               param.Filter(), param.Bias(),
+    //                               param.Output(), false);
+
+    math::DepthwiseConv3x3s2p1v2(param.Input(), param.Filter(), param.Output(),
+                                 *param.Bias(), true);
+
   } else {
     ConvAddBasic(param);
   }
diff --git a/src/operators/kernel/central-arm-func/conv_add_bn_relu_func.h b/src/operators/kernel/central-arm-func/conv_add_bn_relu_func.h
index f184a59a3abbe463f75778c4db216a306c315e8d..fb49a33c67face81a2615516bffd6aa151868fe3 100644
--- a/src/operators/kernel/central-arm-func/conv_add_bn_relu_func.h
+++ b/src/operators/kernel/central-arm-func/conv_add_bn_relu_func.h
@@ -26,8 +26,6 @@ void ConvAddBNReluBasic(const FusionConvAddBNReluParam &param) {
   Tensor bias = *param.Bias();
   Tensor new_bias = *param.NewBias();
   Tensor new_scale = *param.NewScale();
-  auto new_bias_ptr = new_bias.data<float>();
-  auto new_scale_ptr = new_scale.data<float>();
   int axis = param.Axis();
   Tensor *output = param.Output();
   math::expand_bias(bias, axis, output->dims());
@@ -106,20 +104,10 @@ void ConvAddBNReluBasic(const FusionConvAddBNReluParam &param) {
       // gemm
       Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
       Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
-      math::matmul<float>(filter_slice, false, col_matrix, false,
-                          static_cast<float>(1), &out_slice,
-                          static_cast<float>(0));
-    }
-  }
-  /// todo : use neon in special case instead of 2for(300ms)
-  auto output_ptr = output->data<float>();
-  for (int c = 0; c < output_matrix_shape[0]; c++) {
-    int start = c * output_matrix_shape[1];
-    for (int j = 0; j < output_matrix_shape[1]; j++) {
-      output_ptr[start + j] =
-          output_ptr[start + j] * new_scale_ptr[c] + new_bias_ptr[c];
-      output_ptr[start + j] =
-          output_ptr[start + j] < 0 ? 0 : output_ptr[start + j];
+
+      math::matmulWithBn<float>(
+          filter_slice, false, col_matrix, false, static_cast<float>(1),
+          &out_slice, static_cast<float>(0), true, &new_scale, &new_bias);
     }
   }
 }
@@ -138,9 +126,12 @@ void ConvAddBNReluCompute(const FusionConvAddBNReluParam &param) {
              param.Input()->dims()[1] == param.Output()->dims()[1] &&
              param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
              param.Filter()->dims()[2] == 3 && param.Strides()[0] == 2) {
-    math::DepthwiseConvAddBNRelu3x3s2p1(param.Input(), param.Filter(),
-                                        param.Output(), param.NewScale(),
-                                        param.NewBias(), 1);
+    //    math::DepthwiseConvAddBNRelu3x3s2p1(param.Input(), param.Filter(),
+    //                                        param.Output(), param.NewScale(),
+    //                                        param.NewBias(), 1);
+    math::DepthwiseConvAddBNRelu3x3s2p1v2(param.Input(), param.Filter(),
+                                          param.Output(), param.NewScale(),
+                                          param.NewBias(), true);
   } else {
     ConvAddBNReluBasic(param);
   }
diff --git a/src/operators/kernel/central-arm-func/depthwise_conv_arm_func.h b/src/operators/kernel/central-arm-func/depthwise_conv_arm_func.h
index 885f2051f645546c2585caa72aa9c80f8d352e6c..60b09df597f218eefc7f95ba3f342ae0c51c7000 100644
--- a/src/operators/kernel/central-arm-func/depthwise_conv_arm_func.h
+++ b/src/operators/kernel/central-arm-func/depthwise_conv_arm_func.h
@@ -37,8 +37,12 @@ void DepthwiseConvCompute(const ConvParam &param) {
              param.Input()->dims()[1] == param.Output()->dims()[1] &&
              param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
              param.Filter()->dims()[2] == 3 && param.Strides()[0] == 2) {
-    math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
-                           param.Filter(), &Bias, param.Output(), false);
+    //    math::DepthwiseConv3x3(param.Input(), param.Strides(),
+    //    param.Paddings(),
+    //                           param.Filter(), &Bias, param.Output(), false);
+    math::DepthwiseConv3x3s2p1v2(param.Input(), param.Filter(), param.Output(),
+                                 Bias, false);
+
   } else {
     ConvBasic(param);
   }
diff --git a/src/operators/math/depthwise_conv_3x3.cpp b/src/operators/math/depthwise_conv_3x3.cpp
index d1faa1fd641ffd771b5970165d51794a42655670..c8a6473567e6572d506aa5339f8e647c5c25fd5d 100644
--- a/src/operators/math/depthwise_conv_3x3.cpp
+++ b/src/operators/math/depthwise_conv_3x3.cpp
@@ -1010,6 +1010,442 @@ void DepthwiseConvAddBNRelu3x3s2p1(const Tensor *input, const Tensor *filter,
     output_data += output_batch_stride;
   }
 }
+
+void DepthwiseConv3x3s2p1v2(const Tensor *input, const Tensor *filter,
+                            Tensor *output, Tensor bias, bool if_bias) {
+  const float *input_data = input->data<float>();
+  const float *filter_data = filter->data<float>();
+  float *output_data = output->data<float>();
+  const float *bias_data = bias.data<float>();
+
+  const int in_h = static_cast<int>(input->dims()[2]);
+  const int in_w = static_cast<int>(input->dims()[3]);
+  const int out_h = static_cast<int>(output->dims()[2]);
+  const int out_w = static_cast<int>(output->dims()[3]);
+  const int out_l = out_h;
+  const int in_l = in_h;
+  const int inhxw = in_h * in_w;
+  const int outhxw = out_h * out_w;
+  const int if_pad = in_l - 1 == (out_l - 1) * 2 ? 1 : 0;
+  const int batch_size = static_cast<int>(input->dims()[0]);
+  const int c = static_cast<int>(input->dims()[1]);
+  const float *input_row_ptr;
+  float *output_row_ptr;
+
+  const int w_times = (out_w - 2) / 3;
+
+  float32x4_t vbias = vdupq_n_f32(0.0);
+
+  float32x4x2_t input_buff_mid{}, input_buff_bottom[w_times + 1];
+  float32x4_t elewise_res0, elewise_res1, elewise_res2, res3;
+  int out2in_mid;
+  float32x4_t zero = vdupq_n_f32(0.0);
+  for (int b = batch_size; b > 0; --b) {
+    const float *filter_data_tmp = filter_data;
+    for (int j = 0; j < c; ++j) {
+      auto output_data_tmp = output_data + j * out_h * out_w;
+      auto input_data_tmp = input_data + j * in_h * in_w;
+      auto input_const = input_data_tmp;
+
+      if (if_bias) {
+        vbias = vdupq_n_f32(bias_data[j]);
+      }
+
+      float w00 = filter_data_tmp[0];
+      float w01 = filter_data_tmp[1];
+      float w02 = filter_data_tmp[2];
+      float w10 = filter_data_tmp[3];
+      float w11 = filter_data_tmp[4];
+      float w12 = filter_data_tmp[5];
+      float w20 = filter_data_tmp[6];
+      float w21 = filter_data_tmp[7];
+      float w22 = filter_data_tmp[8];
+
+      int h_mid = 0;
+
+      for (; h_mid < out_h - 1; h_mid++) {
+        input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
+        output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
+
+        for (int w4 = 0; w4 < w_times + 1; w4++) {
+          if (h_mid == 0) {
+            elewise_res1 = zero;
+            elewise_res0 = zero;
+            elewise_res2 = zero;
+          } else {
+            elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
+            elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
+            elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
+          }
+          input_buff_mid = vld2q_f32(input_row_ptr);
+          input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
+
+          elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1], w11);
+          elewise_res0 = vmlaq_n_f32(elewise_res0, input_buff_mid.val[0], w10);
+          elewise_res2 = vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
+
+          elewise_res1 =
+              vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1], w21);
+          elewise_res0 =
+              vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0], w20);
+          elewise_res2 =
+              vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0], w22);
+
+          res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
+                           vaddq_f32(elewise_res0, elewise_res1));
+          res3 = vaddq_f32(res3, vbias);
+          vst1q_f32(output_row_ptr, res3);
+
+          input_row_ptr += 6;
+          output_row_ptr += 3;
+        }
+      }
+      clock();
+
+      input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
+      output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
+
+      for (int w4 = 0; w4 < w_times + 1; w4++) {
+        elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
+        elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
+        elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
+
+        input_buff_mid = vld2q_f32(input_row_ptr);
+        input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
+
+        elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1], w11);
+        elewise_res0 = vmlaq_n_f32(elewise_res0, input_buff_mid.val[0], w10);
+        elewise_res2 = vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
+
+        if (!if_pad) {
+          elewise_res1 =
+              vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1], w21);
+          elewise_res0 =
+              vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0], w20);
+          elewise_res2 =
+              vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0], w22);
+        }
+        res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
+                         vaddq_f32(elewise_res0, elewise_res1));
+        res3 = vaddq_f32(res3, vbias);
+
+        if ((w4 != w_times)) {
+          vst1q_f32(output_row_ptr, res3);
+        } else {
+          if (out_l - 2 - w_times * 3 == 1) {
+            vst1q_lane_f32(output_row_ptr, res3, 0);
+          } else if (out_l - 2 - w_times * 3 == 2) {
+            vst1q_lane_f32(output_row_ptr, res3, 0);
+            vst1q_lane_f32(output_row_ptr + 1, res3, 1);
+          }
+        }
+        input_row_ptr += 6;
+        output_row_ptr += 3;
+      }
+
+      output_data_tmp[0] = input_const[0] * w11 + input_const[1] * w12 +
+                           input_const[in_l] * w21 +
+                           input_const[in_l + 1] * w22;
+
+      out2in_mid = (out_l - 1) * 2;
+      output_data_tmp[out_l - 1] =
+          w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
+          w20 * input_const[out2in_mid + in_w - 1] +
+          w21 * input_const[out2in_mid + in_w] +
+          (1 - if_pad) * (w12 * input_const[out2in_mid + 1] +
+                          w22 * input_const[out2in_mid + in_w + 1]);
+
+      out2in_mid = (out_l - 1) * 2 * in_w;
+
+      output_data_tmp[out_l * (out_l - 1)] =
+          w01 * input_const[out2in_mid - in_w] +
+          w02 * input_const[out2in_mid - in_w + 1] +
+          w11 * input_const[out2in_mid] + w12 * input_const[out2in_mid + 1] +
+          (1 - if_pad) * (w21 * input_const[out2in_mid + in_w] +
+                          w22 * input_const[out2in_mid + in_w + 1]);
+      out2in_mid = (out_l - 1) * 2 * in_w + (out_l - 1) * 2;
+
+      output_data_tmp[out_l * out_l - 1] =
+          w00 * input_const[out2in_mid - in_w - 1] +
+          w01 * input_const[out2in_mid - in_w] +
+          w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
+          (1 - if_pad) * (w20 * input_const[out2in_mid + in_w - 1] +
+                          w21 * input_const[out2in_mid + in_w] +
+                          w02 * input_const[out2in_mid - in_w + 1] +
+                          w12 * input_const[out2in_mid + 1] +
+                          w22 * input_const[out2in_mid + in_w + 1]);
+      if (if_bias) {
+        output_data_tmp[0] += bias_data[j];
+        output_data_tmp[out_l - 1] += bias_data[j];
+        output_data_tmp[out_l * (out_l - 1)] += bias_data[j];
+        output_data_tmp[out_l * out_l - 1] += bias_data[j];
+      }
+      for (int i = 1; i < out_h - 1; i++) {
+        out2in_mid = i * 2 * in_w;
+        output_data_tmp[i * out_l] = w01 * input_const[out2in_mid - in_w] +
+                                     w02 * input_const[out2in_mid - in_w + 1] +
+                                     w11 * input_const[out2in_mid] +
+                                     w12 * input_const[out2in_mid + 1] +
+                                     w21 * input_const[out2in_mid + in_w] +
+                                     w22 * input_const[out2in_mid + in_w + 1];
+
+        out2in_mid = i * 2 * in_w + (out_l - 1) * 2;
+        output_data_tmp[i * out_l + out_l - 1] =
+            w00 * input_const[out2in_mid - in_w - 1] +
+            w01 * input_const[out2in_mid - in_w] +
+            w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
+            w20 * input_const[out2in_mid + in_w - 1] +
+            w21 * input_const[out2in_mid + in_w] +
+            (1 - if_pad) * (w02 * input_const[out2in_mid - in_w + 1] +
+                            w12 * input_const[out2in_mid + 1] +
+                            w22 * input_const[out2in_mid + in_w + 1]);
+        if (if_bias) {
+          output_data_tmp[i * out_l] += bias_data[j];
+          output_data_tmp[i * out_l + out_l - 1] += bias_data[j];
+        }
+      }
+      filter_data_tmp += 9;
+    }
+    input_data += inhxw * c;
+    output_data += outhxw * c;
+  }
+}
+
+void DepthwiseConvAddBNRelu3x3s2p1v2(const Tensor *input, const Tensor *filter,
+                                     Tensor *output, const Tensor *new_scale,
+                                     const Tensor *new_bias, bool if_relu) {
+  const float *input_data = input->data<float>();
+  const float *filter_data = filter->data<float>();
+  float *output_data = output->data<float>();
+  const float *newscale_data = new_scale->data<float>();
+  const float *newbias_data = new_bias->data<float>();
+
+  float32x4_t vnewbias = vdupq_n_f32(0.0);
+  float32x4_t vnewscale = vdupq_n_f32(1.0);
+
+  const int in_h = static_cast<int>(input->dims()[2]);
+  const int in_w = static_cast<int>(input->dims()[3]);
+  const int out_h = static_cast<int>(output->dims()[2]);
+  const int out_w = static_cast<int>(output->dims()[3]);
+  const int out_l = out_h;
+  const int in_l = in_h;
+  const int inhxw = in_h * in_w;
+  const int outhxw = out_h * out_w;
+  const int if_pad = in_l - 1 == (out_l - 1) * 2 ? 1 : 0;
+  const int batch_size = static_cast<int>(input->dims()[0]);
+  const int c = static_cast<int>(input->dims()[1]);
+  const float *input_row_ptr;
+  float *output_row_ptr;
+
+  const int w_times = (out_w - 2) / 3;
+
+  float32x4x2_t input_buff_mid{}, input_buff_bottom[w_times + 1];
+  float32x4_t elewise_res0, elewise_res1, elewise_res2, res3;
+  int out2in_mid;
+  float32x4_t zero = vdupq_n_f32(0.0);
+  for (int b = batch_size; b > 0; --b) {
+    const float *filter_data_tmp = filter_data;
+    for (int j = 0; j < c; ++j) {
+      auto output_data_tmp = output_data + j * out_h * out_w;
+      auto input_data_tmp = input_data + j * in_h * in_w;
+      auto input_const = input_data_tmp;
+
+      vnewbias = vdupq_n_f32(newbias_data[j]);
+      vnewscale = vdupq_n_f32(newscale_data[j]);
+
+      float w00 = filter_data_tmp[0];
+      float w01 = filter_data_tmp[1];
+      float w02 = filter_data_tmp[2];
+      float w10 = filter_data_tmp[3];
+      float w11 = filter_data_tmp[4];
+      float w12 = filter_data_tmp[5];
+      float w20 = filter_data_tmp[6];
+      float w21 = filter_data_tmp[7];
+      float w22 = filter_data_tmp[8];
+
+      int h_mid = 0;
+
+      for (; h_mid < out_h - 1; h_mid++) {
+        input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
+        output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
+
+        for (int w4 = 0; w4 < w_times + 1; w4++) {
+          if (h_mid == 0) {
+            elewise_res1 = zero;
+            elewise_res0 = zero;
+            elewise_res2 = zero;
+          } else {
+            elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
+            elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
+            elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
+          }
+          input_buff_mid = vld2q_f32(input_row_ptr);
+          input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
+
+          elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1], w11);
+          elewise_res0 = vmlaq_n_f32(elewise_res0, input_buff_mid.val[0], w10);
+          elewise_res2 = vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
+
+          elewise_res1 =
+              vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1], w21);
+          elewise_res0 =
+              vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0], w20);
+          elewise_res2 =
+              vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0], w22);
+
+          res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
+                           vaddq_f32(elewise_res0, elewise_res1));
+          res3 = vmlaq_f32(vnewbias, vnewscale, res3);
+
+          if (if_relu) {
+            res3 = vmaxq_f32(res3, zero);
+          }
+          vst1q_f32(output_row_ptr, res3);
+
+          input_row_ptr += 6;
+          output_row_ptr += 3;
+        }
+      }
+      clock();
+
+      input_row_ptr = input_data_tmp + 1 + h_mid * 2 * in_w;
+      output_row_ptr = output_data_tmp + 1 + h_mid * out_w;
+
+      for (int w4 = 0; w4 < w_times + 1; w4++) {
+        elewise_res1 = vmulq_n_f32(input_buff_bottom[w4].val[1], w01);
+        elewise_res0 = vmulq_n_f32(input_buff_bottom[w4].val[0], w00);
+        elewise_res2 = vmulq_n_f32(input_buff_bottom[w4].val[0], w02);
+
+        input_buff_mid = vld2q_f32(input_row_ptr);
+        input_buff_bottom[w4] = vld2q_f32(input_row_ptr + in_w);
+
+        elewise_res1 = vmlaq_n_f32(elewise_res1, input_buff_mid.val[1], w11);
+        elewise_res0 = vmlaq_n_f32(elewise_res0, input_buff_mid.val[0], w10);
+        elewise_res2 = vmlaq_n_f32(elewise_res2, input_buff_mid.val[0], w12);
+
+        if (!if_pad) {
+          elewise_res1 =
+              vmlaq_n_f32(elewise_res1, input_buff_bottom[w4].val[1], w21);
+          elewise_res0 =
+              vmlaq_n_f32(elewise_res0, input_buff_bottom[w4].val[0], w20);
+          elewise_res2 =
+              vmlaq_n_f32(elewise_res2, input_buff_bottom[w4].val[0], w22);
+        }
+        res3 = vaddq_f32(vextq_f32(elewise_res2, zero, 1),
+                         vaddq_f32(elewise_res0, elewise_res1));
+        res3 = vmlaq_f32(vnewbias, vnewscale, res3);
+
+        if (if_relu) {
+          res3 = vmaxq_f32(res3, zero);
+        }
+        if ((w4 != w_times)) {
+          vst1q_f32(output_row_ptr, res3);
+        } else {
+          if (out_l - 2 - w_times * 3 == 1) {
+            vst1q_lane_f32(output_row_ptr, res3, 0);
+          } else if (out_l - 2 - w_times * 3 == 2) {
+            vst1q_lane_f32(output_row_ptr, res3, 0);
+            vst1q_lane_f32(output_row_ptr + 1, res3, 1);
+          }
+        }
+        input_row_ptr += 6;
+        output_row_ptr += 3;
+      }
+
+      output_data_tmp[0] = input_const[0] * w11 + input_const[1] * w12 +
+                           input_const[in_l] * w21 +
+                           input_const[in_l + 1] * w22;
+
+      out2in_mid = (out_l - 1) * 2;
+      output_data_tmp[out_l - 1] =
+          w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
+          w20 * input_const[out2in_mid + in_w - 1] +
+          w21 * input_const[out2in_mid + in_w] +
+          (1 - if_pad) * (w12 * input_const[out2in_mid + 1] +
+                          w22 * input_const[out2in_mid + in_w + 1]);
+
+      out2in_mid = (out_l - 1) * 2 * in_w;
+
+      output_data_tmp[out_l * (out_l - 1)] =
+          w01 * input_const[out2in_mid - in_w] +
+          w02 * input_const[out2in_mid - in_w + 1] +
+          w11 * input_const[out2in_mid] + w12 * input_const[out2in_mid + 1] +
+          (1 - if_pad) * (w21 * input_const[out2in_mid + in_w] +
+                          w22 * input_const[out2in_mid + in_w + 1]);
+      out2in_mid = (out_l - 1) * 2 * in_w + (out_l - 1) * 2;
+
+      output_data_tmp[out_l * out_l - 1] =
+          w00 * input_const[out2in_mid - in_w - 1] +
+          w01 * input_const[out2in_mid - in_w] +
+          w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
+          (1 - if_pad) * (w20 * input_const[out2in_mid + in_w - 1] +
+                          w21 * input_const[out2in_mid + in_w] +
+                          w02 * input_const[out2in_mid - in_w + 1] +
+                          w12 * input_const[out2in_mid + 1] +
+                          w22 * input_const[out2in_mid + in_w + 1]);
+      output_data_tmp[0] =
+          output_data_tmp[0] * newscale_data[j] + newbias_data[j];
+      output_data_tmp[out_l - 1] =
+          output_data_tmp[out_l - 1] * newscale_data[j] + newbias_data[j];
+      output_data_tmp[out_l * (out_l - 1)] =
+          output_data_tmp[out_l * (out_l - 1)] * newscale_data[j] +
+          newbias_data[j];
+      output_data_tmp[out_l * out_l - 1] =
+          output_data_tmp[out_l * out_l - 1] * newscale_data[j] +
+          newbias_data[j];
+      if (if_relu) {
+        output_data_tmp[0] = output_data_tmp[0] < 0 ? 0 : output_data_tmp[0];
+        output_data_tmp[out_l - 1] =
+            output_data_tmp[out_l - 1] < 0 ? 0 : output_data_tmp[out_l - 1];
+        output_data_tmp[out_l * (out_l - 1)] =
+            output_data_tmp[out_l * (out_l - 1)] < 0
+                ? 0
+                : output_data_tmp[out_l * (out_l - 1)];
+        output_data_tmp[out_l * out_l - 1] =
+            output_data_tmp[out_l * out_l - 1] < 0
+                ? 0
+                : output_data_tmp[out_l * out_l - 1];
+      }
+      for (int i = 1; i < out_h - 1; i++) {
+        out2in_mid = i * 2 * in_w;
+        output_data_tmp[i * out_l] = w01 * input_const[out2in_mid - in_w] +
+                                     w02 * input_const[out2in_mid - in_w + 1] +
+                                     w11 * input_const[out2in_mid] +
+                                     w12 * input_const[out2in_mid + 1] +
+                                     w21 * input_const[out2in_mid + in_w] +
+                                     w22 * input_const[out2in_mid + in_w + 1];
+
+        out2in_mid = i * 2 * in_w + (out_l - 1) * 2;
+        output_data_tmp[i * out_l + out_l - 1] =
+            w00 * input_const[out2in_mid - in_w - 1] +
+            w01 * input_const[out2in_mid - in_w] +
+            w10 * input_const[out2in_mid - 1] + w11 * input_const[out2in_mid] +
+            w20 * input_const[out2in_mid + in_w - 1] +
+            w21 * input_const[out2in_mid + in_w] +
+            (1 - if_pad) * (w02 * input_const[out2in_mid - in_w + 1] +
+                            w12 * input_const[out2in_mid + 1] +
+                            w22 * input_const[out2in_mid + in_w + 1]);
+        output_data_tmp[i * out_l] =
+            output_data_tmp[i * out_l] * newscale_data[j] + newbias_data[j];
+        output_data_tmp[i * out_l + out_l - 1] =
+            output_data_tmp[i * out_l + out_l - 1] * newscale_data[j] +
+            newbias_data[j];
+        if (if_relu) {
+          output_data_tmp[i * out_l] =
+              output_data_tmp[i * out_l] < 0 ? 0 : output_data_tmp[i * out_l];
+          output_data_tmp[i * out_l + out_l - 1] =
+              output_data_tmp[i * out_l + out_l - 1] < 0
+                  ? 0
+                  : output_data_tmp[i * out_l + out_l - 1];
+        }
+      }
+      filter_data_tmp += 9;
+    }
+    input_data += inhxw * c;
+    output_data += outhxw * c;
+  }
+}
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile
diff --git a/src/operators/math/depthwise_conv_3x3.h b/src/operators/math/depthwise_conv_3x3.h
index b5103a53ad7d988b8d18a57cb0f0d8d4bb9fee0f..60e979648f871e640924a3373c625c311c3dd067 100644
--- a/src/operators/math/depthwise_conv_3x3.h
+++ b/src/operators/math/depthwise_conv_3x3.h
@@ -38,6 +38,11 @@ void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
 void DepthwiseConvAddBNRelu3x3s2p1(const Tensor *input, const Tensor *filter,
                                    Tensor *output, const Tensor *new_scale,
                                    const Tensor *new_bias, bool if_relu);
+void DepthwiseConv3x3s2p1v2(const Tensor *input, const Tensor *filter,
+                            Tensor *output, Tensor bias, bool if_bias);
+void DepthwiseConvAddBNRelu3x3s2p1v2(const Tensor *input, const Tensor *filter,
+                                     Tensor *output, const Tensor *new_scale,
+                                     const Tensor *new_bias, bool if_relu);
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle_mobile
diff --git a/test/common/test_gemm.cpp b/test/common/test_gemm.cpp
index aaf3c183f3e125f09695fad8a41cfb5360e9da13..8cb778c458034aecf6cea89fcf0d3e2a3d8118ba 100644
--- a/test/common/test_gemm.cpp
+++ b/test/common/test_gemm.cpp
@@ -52,8 +52,9 @@ int main() {
   }
 
   auto time1 = time();
-  paddle_mobile::operators::math::sgemm(m, n, k, 0.9, a, lda, b, ldb, 0.3, c,
-                                        ldc);
+  //  paddle_mobile::operators::math::Sgemm(m, n, k, 0.9, a, lda, b, ldb, 0.3,
+  //  c,
+  //                                        ldc);
   auto time2 = time();
   DLOG << "gemm cost :" << time_diff(time1, time2) << "ms\n";
   for (int i = 0; i < m * n; ++i) {