diff --git a/src/operators/feed_op.h b/src/operators/feed_op.h
index e45ad38fd68cb9b4616b7e363be117e2039c93a9..8056e78aed3b67daf7dead7987e80529b85f0d60 100644
--- a/src/operators/feed_op.h
+++ b/src/operators/feed_op.h
@@ -30,18 +30,45 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
       : framework::OperatorBase<DeviceType>(type, inputs, outputs, attrs,
                                             scope),
         param_(inputs, outputs, attrs, scope.get()) {}
-  void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
 
-  void Init() {}
+ protected:
+  FeedParam param_;
+
+#ifdef PADDLE_MOBILE_FPGA
+  void RunImpl() const { fpga::PerformBypass(param_.FpgaArgs()); }
+
+  void Init() {
+    const Tensor *input = param_.InputX();
+    auto input_ptr = input->data<float>();
+    Tensor *output = param_.Out();
+    auto output_ptr = output->mutable_data<half>();
+    fpga::BypassArgs args;
+    args.convert_type = fpga::DATA_FP32_TO_FP16;
+    args.layout_type = fpga::LAYOUT_CHW_TO_HWC;
+    args.image.address = (void *)input_ptr;
+    args.image.channels = input->dims()[1];
+    args.image.height = input->dims()[2];
+    args.image.width = input->dims()[3];
+    args.output.address = output_ptr;
+    param_.SetFpgaArgs(args);
+  }
 
   void InferShape() const {
     auto out_dims = param_.Out()->dims();
     out_dims[0] = param_.BatchSize();
     param_.Out()->Resize(out_dims);
+    param_.Out()->ShareDataWith(*param_.InputX());  // TODO How to handle fp16
   }
+#else
+  void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
 
- protected:
-  FeedParam param_;
+  void Init() {}
+
+  void InferShape() const {
+    auto out_dims = param_.Out()->dims();
+    out_dims[0] = param_.BatchSize();
+    param_.Out()->Resize(out_dims);
+#endif
 };
 
 }  // namespace operators
diff --git a/src/operators/kernel/fpga/conv_add_bn_kernel.cpp b/src/operators/kernel/fpga/conv_add_bn_kernel.cpp
index 6f9da6bc1dde924e2c499bb2478d29a8d4a9e5d9..6719db3a80cb3c3a2ee603096b2659fa5489497d 100644
--- a/src/operators/kernel/fpga/conv_add_bn_kernel.cpp
+++ b/src/operators/kernel/fpga/conv_add_bn_kernel.cpp
@@ -24,13 +24,13 @@ template <>
 bool ConvAddBNKernel<FPGA, float>::Init(FusionConvAddBNParam *param) {
   bool relu_enabled = false;
   const Tensor *input = param->Input();
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   const Tensor *filter = param->Filter();
   auto filter_ptr = filter->data<float>();
   Tensor *out = param->Output();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
   auto bn_mean_ptr = param->InputMean()->data<float>();
   auto bn_var_ptr = param->InputVariance()->data<float>();
   auto bn_scale_ptr = param->InputScale()->data<float>();
diff --git a/src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp b/src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp
index 66a593df84c12f87371a9bde9f0aef514b392584..2f80ec9742cae356c3b9a2dcd0c79027da37e7b4 100644
--- a/src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp
+++ b/src/operators/kernel/fpga/conv_add_bn_relu_kernel.cpp
@@ -24,13 +24,13 @@ template <>
 bool ConvAddBNReluKernel<FPGA, float>::Init(FusionConvAddBNReluParam *param) {
   bool relu_enabled = true;
   const Tensor *input = param->Input();
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   const Tensor *filter = param->Filter();
   auto filter_ptr = filter->data<float>();
   Tensor *out = param->Output();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
   auto bn_mean_ptr = param->InputMean()->data<float>();
   auto bn_var_ptr = param->InputVariance()->data<float>();
   auto bn_scale_ptr = param->InputScale()->data<float>();
diff --git a/src/operators/kernel/fpga/conv_add_relu_kernel.cpp b/src/operators/kernel/fpga/conv_add_relu_kernel.cpp
index 9692bcef872f956e2cdbe82545b3ab4173bf1348..a20f4e4837c6ceadf7f445b0b82d951ba3c1721b 100644
--- a/src/operators/kernel/fpga/conv_add_relu_kernel.cpp
+++ b/src/operators/kernel/fpga/conv_add_relu_kernel.cpp
@@ -24,13 +24,13 @@ template <>
 bool ConvAddReluKernel<FPGA, float>::Init(FusionConvAddReluParam *param) {
   bool relu_enabled = true;
   const Tensor *input = param->Input();
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   const Tensor *bias = param->Bias();
   auto bias_ptr = bias->data<float>();
   const Tensor *filter = param->Filter();
   auto filter_ptr = filter->data<float>();
   Tensor *out = param->Output();
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
 
   PADDLE_MOBILE_ENFORCE(input->dims()[1] == bias->dims()[0],
                         "Image channel should be equal to bias number");
diff --git a/src/operators/op_param.h b/src/operators/op_param.h
index 0821ab8c32a6ba232a673ddd100a4e7fe6475571..844d2f1068b7499bef35a7fd346317c1c3f6ad65 100644
--- a/src/operators/op_param.h
+++ b/src/operators/op_param.h
@@ -665,6 +665,16 @@ class FeedParam : public OpParam {
   Tensor *input_x_;
   Tensor *out_;
   int batch_size;
+
+#ifdef PADDLE_MOBILE_FPGA
+
+ private:
+  fpga::BypassArgs fpga_bypass_args;
+
+ public:
+  const fpga::BypassArgs &FpgaArgs() const { return fpga_bypass_args; }
+  void SetFpgaArgs(const fpga::BypassArgs &args) { fpga_bypass_args = args; }
+#endif
 };
 
 class FetchParam : public OpParam {