fuse conv add batch relu when using faster depthwise conv (#1749)

src/operators/kernel/arm/convolution/conv_add_bn_relu_kernel.cpp

@@ -61,16 +61,61 @@ bool ConvAddBNReluKernel<CPU, float>::Init(
   param->SetNewBias(new_bias);
 
   InitBaseConvKernel(param);
+
+  // try to use faster depthwise conv
+  switch (param->ExecMode()) {
+    case ConvParam<CPU>::EXEC_DEPTHWISE3x3S1_FLOAT:
+    case ConvParam<CPU>::EXEC_DEPTHWISE3x3S2_FLOAT:
+      const std::vector<int> &paddings = param->Paddings();
+      const std::vector<int> &strides = param->Strides();
+      if (paddings.size() == 2 && paddings[0] == paddings[1] &&
+          strides.size() == 2 && strides[0] == strides[1]) {
+        int pad = paddings[0];
+        int stride = strides[0];
+        const int hin = param->Input()->dims()[2];
+        if (pad == 0 && hin > 2) {
+          could_use_faster_depthwise_conv_ = true;
+        } else if (pad == 1) {
+          could_use_faster_depthwise_conv_ = true;
+        }
+      }
+      break;
+  }
+
+  if (could_use_faster_depthwise_conv_) {
+    auto filter_data = param->Filter()->data<float>();
+    auto filter_dim = param->Filter()->dims();
+    int len = 1;
+    for (int i = 0; i < filter_dim.size(); i++) {
+      len *= filter_dim[i];
+    }
+    int batch = filter_dim[0];
+    int step = len / batch;
+    for (int i = 0; i < batch; i++) {
+      for (int k = 0; k < step; k++) {
+        filter_data[i * step + k] =
+            filter_data[i * step + k] * new_scale_ptr[i];
+      }
+    }
+  }
+
   return true;
 }
 
 template <>
 void ConvAddBNReluKernel<CPU, float>::Compute(
     const FusionConvAddBNReluParam<CPU> &param) {
+  bool fusion_has_been_computed = false;
   switch (param.ExecMode()) {
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S1_FLOAT:
     case ConvParam<CPU>::EXEC_DEPTHWISE3x3S2_FLOAT:
+      if (could_use_faster_depthwise_conv_) {
+        FasterDepthwiseConv3x3_bias_relu(param, param.NewBias()->data<float>(),
+                                         true);
+        fusion_has_been_computed = true;
+      } else {
         DepthwiseConv3x3<float, float>(param);
+      }
       break;
     case ConvParam<CPU>::EXEC_DEPTHWISE5x5_FLOAT:
       DepthwiseConv5x5<float, float>(param);
@@ -89,8 +134,10 @@ void ConvAddBNReluKernel<CPU, float>::Compute(
       PADDLE_MOBILE_THROW_EXCEPTION("Invalid convolution execute mode %d",
                                     param.ExecMode());
   }
+  if (!fusion_has_been_computed) {
     math::ScaleAddChannelWise<RELU>(param.Output(), param.NewScale(),
                                     param.NewBias(), param.Output());
+  }
 }
 
 template class ConvAddBNReluKernel<CPU, float>;

src/operators/kernel/central-arm-func/conv_arm_func.cpp

@@ -212,8 +212,8 @@ void DepthwiseConv3x3(const ConvParam<CPU> &param) {
   }
 }
 
-template <>
-void DepthwiseConv3x3<float, float>(const ConvParam<CPU> &param) {
+void FasterDepthwiseConv3x3_bias_relu(const ConvParam<CPU> &param,
+                                      const float *bias, bool flag_relu) {
   const Tensor *input = param.Input();
   const Tensor *filter = param.Filter();
   const std::vector<int> &paddings = param.Paddings();
@@ -222,14 +222,11 @@ void DepthwiseConv3x3<float, float>(const ConvParam<CPU> &param) {
   Tensor *output = param.Output();
   output->mutable_data<float>();
 
-  if (paddings.size() == 2 && paddings[0] == paddings[1] &&
-      strides.size() == 2 && strides[0] == strides[1]) {
   int pad = paddings[0];
   int stride = strides[0];
   const float *din = input->data<float>();
   float *dout = output->mutable_data<float>();
   const float *weights = filter->data<float>();
-    const float *bias = nullptr;
   const int num = input->dims()[0];
   const int chin = input->dims()[1];
   const int hin = input->dims()[2];
@@ -237,37 +234,15 @@ void DepthwiseConv3x3<float, float>(const ConvParam<CPU> &param) {
   const int chout = output->dims()[1];
   const int hout = output->dims()[2];
   const int wout = output->dims()[3];
-    bool flag_relu = false;
   bool flag_bias = bias != nullptr;
   if (pad == 0 && hin > 2) {
     math::depthwise::conv_depthwise_3x3p0(din, dout, num, chout, hout, wout,
-                                            chin, hin, win, weights, bias,
-                                            stride, flag_bias, flag_relu);
+                                          chin, hin, win, weights, bias, stride,
+                                          flag_bias, flag_relu);
   } else if (pad == 1) {
     math::depthwise::conv_depthwise_3x3p1(din, dout, num, chout, hout, wout,
-                                            chin, hin, win, weights, bias,
-                                            stride, flag_bias, flag_relu);
-    } else {
-      GemmConv<float, float>(param);
-    }
-  } else {
-    if (strides[0] == 1) {
-      for (int i = 0; i < batch_size; i++) {
-        Tensor in_batch = input->Slice(i, i + 1);
-        Tensor out_batch = output->Slice(i, i + 1);
-        math::DepthwiseConv3x3S1<float, float>(in_batch, *filter, paddings,
-                                               &out_batch);
-      }
-    } else if (strides[0] == 2) {
-      for (int i = 0; i < batch_size; i++) {
-        Tensor in_batch = input->Slice(i, i + 1);
-        Tensor out_batch = output->Slice(i, i + 1);
-        math::DepthwiseConv3x3S2<float, float>(in_batch, *filter, paddings,
-                                               &out_batch);
-      }
-    } else {
-      GemmConv<float, float>(param);
-    }
+                                          chin, hin, win, weights, bias, stride,
+                                          flag_bias, flag_relu);
   }
 }
 

src/operators/kernel/central-arm-func/conv_arm_func.h

@@ -44,6 +44,9 @@ void DepthwiseConv5x5(const ConvParam<CPU> &param);
 template <typename Itype, typename Otype>
 void SlidingwindowConv3x3(const ConvParam<CPU> &param);
 
+void FasterDepthwiseConv3x3_bias_relu(const ConvParam<CPU> &param,
+                                      const float *bias, bool flag_relu);
+
 }  // namespace operators
 }  // namespace paddle_mobile
 

src/operators/kernel/conv_add_bn_relu_kernel.h

@@ -36,6 +36,9 @@ class ConvAddBNReluKernel
  public:
   void Compute(const FusionConvAddBNReluParam<DeviceType> &param);
   bool Init(FusionConvAddBNReluParam<DeviceType> *param);
+
+ private:
+  bool could_use_faster_depthwise_conv_ = false;
 };
 
 }  // namespace operators