reconstruct code to support RFCN for FPGA track

src/fpga/V1/api.cpp

@@ -28,11 +28,13 @@ void format_image(framework::Tensor *image_tensor) {
   auto dims = image_tensor->dims();
   auto channel = dims[1], height = dims[2], width = dims[3];
   auto data_ptr = image_tensor->data<float>();
-  size_t memory_size = channel * height * width * sizeof(float);
-  auto new_data = (float *)fpga_malloc(memory_size);  // NOLINT
-  fpga_copy(new_data, data_ptr, memory_size);
-  image::format_image(&new_data, channel, height, width);
-  image_tensor->reset_data_ptr(new_data);
+  auto external_ptr = reinterpret_cast<float *>(image_tensor->external_data);
+  float *p_data = external_ptr == nullptr ? data_ptr : external_ptr;
+  float *old_p = p_data;
+  image::format_image(&p_data, channel, height, width);
+  if (old_p != p_data) {
+    image_tensor->reset_data_ptr(p_data);
+  }
 }
 
 void format_fp16_ofm(framework::Tensor *ofm_tensor) {
@@ -50,6 +52,7 @@ void format_fp16_ofm(framework::Tensor *ofm_tensor) {
   auto p = fpga_malloc(memory_size);
   memset(p, 0, memory_size);
   ofm_tensor->reset_data_ptr(p);
+  ofm_tensor->set_type(typeid(half));
 }
 
 void format_fp16_ofm(framework::Tensor *ofm_tensor, framework::DDim dims) {
@@ -67,6 +70,7 @@ void format_fp16_ofm(framework::Tensor *ofm_tensor, framework::DDim dims) {
   auto p = fpga_malloc(memory_size);
   memset(p, 0, memory_size);
   ofm_tensor->reset_data_ptr(p);
+  ofm_tensor->set_type(typeid(half));
 }
 void format_fp32_ofm(framework::Tensor *ofm_tensor) {
   auto dims = ofm_tensor->dims();
@@ -83,6 +87,7 @@ void format_fp32_ofm(framework::Tensor *ofm_tensor) {
   auto p = fpga_malloc(memory_size);
   memset(p, 0, memory_size);
   ofm_tensor->reset_data_ptr(p);
+  ofm_tensor->set_type(typeid(float));
 }
 
 float filter_find_max(framework::Tensor *filter_tensor) {
@@ -139,6 +144,7 @@ void format_filter(framework::Tensor *filter_tensor, float max_value,
   filter::format_filter(&new_data, num, channel, height, width, group_num,
                         max_value);
   filter_tensor->reset_data_ptr(new_data);
+  filter_tensor->set_type(typeid(int8_t));
 }
 void format_dwconv_filter(framework::Tensor *filter_tensor, float *scale_ptr) {
   auto dims = filter_tensor->dims();
@@ -149,6 +155,7 @@ void format_dwconv_filter(framework::Tensor *filter_tensor, float *scale_ptr) {
   fpga_copy(new_data, data_ptr, memory_size);
   filter::format_dwconv_filter(&new_data, num, height, width, scale_ptr);
   filter_tensor->reset_data_ptr(new_data);
+  filter_tensor->set_type(typeid(int8_t));
 }
 
 void format_DWDconv_filter(framework::Tensor *filter_tensor, float *scale_ptr,
@@ -173,6 +180,7 @@ void format_DWDconv_filter(framework::Tensor *filter_tensor, float *scale_ptr,
   //      framework::make_ddim({num, 1, height, width});
   //  filter_tensor->Resize(dims_new);
   filter_tensor->reset_data_ptr(new_data);
+  filter_tensor->set_type(typeid(int8_t));
 }
 
 void format_fc_filter(framework::Tensor *filter_tensor, float max_value) {
@@ -187,6 +195,7 @@ void format_fc_filter(framework::Tensor *filter_tensor, float max_value) {
   filter::format_fc_filter(&new_data, num, channel, height, width, 1,
                            max_value);
   filter_tensor->reset_data_ptr(new_data);
+  filter_tensor->set_type(typeid(int8_t));
 }
 void format_deconv_filter(framework::Tensor *filter_tensor, float max_value,
                           int group_num, int stride) {
@@ -213,6 +222,7 @@ void format_deconv_filter(framework::Tensor *filter_tensor, float max_value,
       framework::make_ddim({num, channel, height, width});
   filter_tensor->Resize(dims_new);
   filter_tensor->reset_data_ptr(new_data);
+  filter_tensor->set_type(typeid(int8_t));
 }
 
 void format_bias_scale_array(float **bias_scale_array,
@@ -236,6 +246,7 @@ void format_concat_output(framework::Tensor *out, int height, int width,
   auto ddim = framework::make_ddim({1, sum_channel, height, width});
   out->Resize(ddim);
   out->reset_data_ptr(data_ptr);
+  out->set_type(typeid(half));
 }
 void format_conv_data(framework::Tensor *filter_tensor,
                       framework::Tensor *ofm_tensor, float **bs_ptr,
@@ -447,9 +458,9 @@ void fill_split_arg(struct SplitConvArgs *arg, framework::Tensor *input,
                     int16_t leaky_relu_negative_slope, int group_num,
                     int stride_h, int stride_w, int padding_h, int padding_w,
                     float *bs_ptr) {
-  auto input_ptr = input->data<float>();
-  auto filter_ptr = filter->data<float>();
-  auto out_ptr = out->data<float>();
+  auto input_ptr = input->data<half>();
+  auto filter_ptr = filter->data<int8_t>();
+  auto out_ptr = out->data<half>();
   auto deleter = [](void *p) { fpga_free(p); };
 
   arg->group_num = (uint32_t)group_num;
@@ -571,8 +582,8 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
                      int16_t leaky_relu_negative_slope, int group_num,
                      int stride_h, int stride_w, int padding_h, int padding_w,
                      float *bs_ptr) {
-  auto input_ptr = input->data<float>();
-  auto filter_ptr = filter->data<float>();
+  auto input_ptr = input->data<half>();
+  auto filter_ptr = filter->data<int8_t>();
   auto deleter = [](void *p) { fpga_free(p); };
 
   arg->group_num = (uint32_t)group_num;
@@ -603,7 +614,7 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
   framework::DDim dims_out_new = framework::make_ddim(
       {1, arg->filter_num, sub_output_height * sub_conv_num, real_out_width});
   fpga::format_fp16_ofm(out, dims_out_new);
-  auto out_ptr = out->data<float>();
+  auto out_ptr = out->data<half>();
   arg->output.address =
       (half *)out_ptr +  // NOLINT
       omit_size * sizeof(half) *
@@ -793,7 +804,7 @@ void fill_deconv_arg(struct DeconvArgs *arg, framework::Tensor *input,
                     arg->split_conv_args[i]->conv_arg[j].output.scale_address),
                 deleter));
       }
-      arg->split_conv_args[i]->concat_arg.images_in[j] = static_cast<int16_t *>(
+      arg->split_conv_args[i]->concat_arg.images_in[j] = static_cast<half *>(
           arg->split_conv_args[i]->conv_arg[j].output.address);
       arg->split_conv_args[i]->concat_arg.scales_in[j] =
           arg->split_conv_args[i]->conv_arg[j].output.scale_address;
@@ -818,9 +829,9 @@ void fill_dwconv_arg(struct DWconvArgs *arg, framework::Tensor *input,
                      int16_t leaky_relu_negative_slope, int stride_h,
                      int stride_w, int padding_h, int padding_w,
                      float *bias_ptr) {
-  auto filter_ptr = filter->data<float>();
-  auto input_ptr = input->data<float>();
-  auto output_ptr = out->mutable_data<float>();
+  auto filter_ptr = filter->data<uint8_t>();
+  auto input_ptr = input->data<half>();
+  auto output_ptr = out->mutable_data<half>();
   arg->sub_conv_num = 1;
   // arg->relu_enabled = relu_enabled;
   arg->output.activation.activation_type = activation_enable;
@@ -848,9 +859,8 @@ void fill_DWDeconv_arg(struct DWDeconvArgs *arg, framework::Tensor *input,
                        int16_t leaky_relu_negative_slope, int stride_h,
                        int stride_w, int padding_h, int padding_w,
                        float *bias_ptr) {
-  auto filter_ptr = filter->data<float>();
-  auto input_ptr = input->data<float>();
-  auto output_ptr = out->mutable_data<float>();
+  auto filter_ptr = filter->data<int8_t>();
+  auto input_ptr = input->data<half>();
 
   auto deleter = [](void *p) { fpga_free(p); };
 
@@ -885,7 +895,7 @@ void fill_DWDeconv_arg(struct DWDeconvArgs *arg, framework::Tensor *input,
   framework::DDim dims_out_new = framework::make_ddim(
       {1, arg->filter_num, real_out_height, real_out_width});
   fpga::format_fp16_ofm(out, dims_out_new);
-  auto out_ptr = out->data<float>();
+  auto out_ptr = out->data<half>();
 
   /*====For Addition
   arg->output.address =

src/fpga/V1/image.cpp

@@ -22,7 +22,6 @@ namespace fpga {
 namespace image {
 
 void convert_to_hwc(float **data_in, int channel, int height, int width) {
-  float *tmp = *data_in;
   float *data_tmp =
       (float *)fpga_malloc(channel * height * width * sizeof(float));  // NOLINT
   int64_t amount_per_row = width * channel;
@@ -35,33 +34,35 @@ void convert_to_hwc(float **data_in, int channel, int height, int width) {
     }
   }
   *data_in = data_tmp;
-  fpga_free(tmp);
 }
 
 void align_element_conv(float **data_in, int height, int cw) {
   int h = 0;
   int align_cw = align_to_x(cw, IMAGE_ALIGNMENT);
-  if (align_cw != cw) {
-    float *tmp = *data_in;
-    float *data_tmp =
-        (float *)fpga_malloc(height * align_cw * sizeof(float));  // NOLINT
 
-    memset(data_tmp, 0, height * align_cw * sizeof(float));
+  float *data_tmp =
+      (float *)fpga_malloc(height * align_cw * sizeof(float));  // NOLINT
 
-    for (h = 0; h < height; h++) {
-      memcpy((void *)(data_tmp + h * align_cw),  // NOLINT
-             (void *)(*data_in + h * cw),        // NOLINT
-             cw * sizeof(float));
-    }
+  memset(data_tmp, 0, height * align_cw * sizeof(float));
 
-    *data_in = data_tmp;
-    fpga_free(tmp);
+  for (h = 0; h < height; h++) {
+    memcpy((void *)(data_tmp + h * align_cw),  // NOLINT
+           (void *)(*data_in + h * cw),        // NOLINT
+           cw * sizeof(float));
   }
+
+  *data_in = data_tmp;
 }
 
 void format_image(float **data_in, int channel, int height, int width) {
   convert_to_hwc(data_in, channel, height, width);
-  align_element_conv(data_in, height, channel * width);
+  int cw = channel * width;
+  int align_cw = align_to_x(cw, IMAGE_ALIGNMENT);
+  if (align_cw != cw) {
+    float *hwc_temp = *data_in;
+    align_element_conv(data_in, height, channel * width);
+    fpga_free(hwc_temp);
+  }
   fpga_flush(*data_in, align_to_x(channel * width, IMAGE_ALIGNMENT) * height *
                            sizeof(float));
 }

src/fpga/common/fpga_common.cpp

@@ -164,7 +164,7 @@ void fpga_free(void *ptr) {
     //    DLOG << "Address: " << ptr << ", " << size << " bytes freed. Total "
     //         << counter << " bytes";
   } else {
-    DLOG << "Invalid pointer";
+    DLOG << "Address: " << ptr << "  Invalid pointer";
   }
 }
 void fpga_copy(void *dest, const void *src, size_t num) {

src/fpga/common/fpga_common.h

@@ -19,17 +19,16 @@ limitations under the License. */
 #include <memory>
 #include <vector>
 
-namespace paddle_mobile {
-namespace fpga {
-
 #ifdef PADDLE_MOBILE_FPGA_V1
-#define IMAGE_ALIGNMENT 16           // Aligned to 16
-#define FILTER_NUM_ALIGNMENT 32      // Filter number aligned to 32
-#define FILTER_ELEMENT_ALIGNMENT 16  // Filter element number aligned to 16
-#define BS_NUM_ALIGNMENT 8
-#define BIAS_NUM_ALIGNMENT 16
+#define IMAGE_ALIGNMENT (16)           // Aligned to 16
+#define FILTER_NUM_ALIGNMENT (32)      // Filter number aligned to 32
+#define FILTER_ELEMENT_ALIGNMENT (16)  // Filter element number aligned to 16
+#define BS_NUM_ALIGNMENT (8)
+#define BIAS_NUM_ALIGNMENT (16)
 #endif
 
+namespace paddle_mobile {
+namespace fpga {
 enum DataType {
   DATA_TYPE_FP32 = 1,
   DATA_TYPE_FP16 = 0,
@@ -49,7 +48,7 @@ enum ActivationType {
 };
 
 struct ActivationArgs {
-  enum ActivationType activation_type;
+  enum ActivationType activation_type = NONE;
   int16_t leaky_relu_negative_slope;
 };
 

src/framework/executor.cpp

@@ -84,6 +84,11 @@ Executor<Device, T>::Executor(const Program<Device> &program,
     InitMemory();
   }
 
+#ifdef PADDLE_MOBILE_FPGA
+  program_.scope->EraseVars({"feed", "fetch"});
+  program_.scope->print_vars();
+#endif
+
   int count = 0;
   for (int block_id = 0; block_id < ops_of_block_.size(); ++block_id) {
     for (auto &op_handler : ops_of_block_[block_id]) {
@@ -92,14 +97,6 @@ Executor<Device, T>::Executor(const Program<Device> &program,
       ops_list_.push_back(op_handler);
     }
   }
-#ifdef PADDLE_MOBILE_FPGA
-  TalorFeedOp();
-  DLOG << "TalorFeed finished";
-  TalorFetchdOp();
-  DLOG << "TalorFetch finished";
-  program_.scope->print_vars();
-
-#endif
 }
 
 template <typename T>
@@ -451,49 +448,6 @@ std::shared_ptr<LoDTensor> Executor<Device, T>::GetOutput(
 }
 
 #ifdef PADDLE_MOBILE_FPGA
-template <typename Device, typename T>
-void Executor<Device, T>::TalorFeedOp() {
-  auto &ops = ops_of_block_[0];
-  int num = 0;
-  program_.scope->EraseVars(std::vector<string>{string("feed")});
-  for (auto op : ops) {
-    if (op->Type() == "feed") {
-      auto new_name = string("feed") + std::to_string(num++);
-      auto var = program_.scope->Var(new_name);
-      auto tensor = var->template GetMutable<LoDTensor>();
-      auto output_map = op->Outputs();
-      std::vector<std::string> out_keys = op->GetOutKeys();
-      PADDLE_MOBILE_ENFORCE(!out_keys.empty(), "this op contains no output");
-      auto output_tensor =
-          GetVarValue<LoDTensor>(out_keys[0], output_map, *(program_.scope));
-      tensor->Resize(output_tensor->dims());
-      tensor->init(typeid(float));
-      op->ChangeNameMap("X", std::vector<string>{new_name});
-    }
-  }
-}
-template <typename Device, typename T>
-void Executor<Device, T>::TalorFetchdOp() {
-  auto &ops = ops_of_block_[0];
-  int num = 0;
-  program_.scope->EraseVars(std::vector<string>{string("fetch")});
-  for (auto op : ops) {
-    if (op->Type() == "fetch") {
-      auto new_name = string("fetch") + std::to_string(num++);
-      auto var = program_.scope->Var(new_name);
-      auto tensor = var->template GetMutable<LoDTensor>();
-      auto input_map = op->Inputs();
-      std::vector<std::string> in_keys = op->GetInputKeys();
-      PADDLE_MOBILE_ENFORCE(!in_keys.empty(), "this op contains no input");
-      auto input_tensor =
-          GetVarValue<LoDTensor>(in_keys[0], input_map, *(program_.scope));
-      tensor->Resize(input_tensor->dims());
-      tensor->init(typeid(float));
-      op->ChangeNameMap("Out", std::vector<string>{new_name});
-    }
-  }
-}
-
 template <typename Device, typename T>
 void Executor<Device, T>::InjectVariable(const Tensor &t,
                                          std::string var_name) {
@@ -509,18 +463,29 @@ void Executor<Device, T>::FeedData(const Tensor &t) {
 }
 
 template <typename Device, typename T>
-void Executor<Device, T>::FeedData(const std::vector<Tensor> &v) {
+void Executor<Device, T>::FeedData(const std::vector<void *> &v) {
   auto input_size = v.size();
-  PADDLE_MOBILE_ENFORCE(input_size > 0, "Empty input");
-  int counter = 0;
   auto vars = program_.scope->VarContain("feed");
-  for (auto var : vars) {
-    Tensor *feed_tensor = var->template GetMutable<LoDTensor>();
-    feed_tensor->Resize(v[counter].dims());
-    feed_tensor->ShareDataWith(v[counter]);
-    if (++counter > v.size()) {
-      return;
-    }
+  PADDLE_MOBILE_ENFORCE(input_size == vars.size(),
+                        "input data number not correct");
+  for (int i = 0; i < input_size; i++) {
+    auto var = program_.scope->Var("feed", i);
+    auto feed_tensor = var->template GetMutable<LoDTensor>();
+    feed_tensor->external_data = v[i];
+  }
+}
+
+template <typename Device, typename T>
+void Executor<Device, T>::GetResults(std::vector<void *> *v) {
+  auto output_size = v->size();
+  PADDLE_MOBILE_ENFORCE(output_size > 0, "Empty output");
+  auto vars = program_.scope->VarContain("fetch");
+  PADDLE_MOBILE_ENFORCE(output_size == vars.size(),
+                        "output data number not correct");
+  for (int i = 0; i < output_size; i++) {
+    auto var = program_.scope->Var("fetch", i);
+    auto fetch_tensor = var->template GetMutable<LoDTensor>();
+    (*v)[i] = fetch_tensor->template data<float>();
   }
 }
 

src/framework/executor.h

@@ -50,11 +50,10 @@ class Executor {
   std::shared_ptr<LoDTensor> GetOutput(const std::string &var_name);
 
 #ifdef PADDLE_MOBILE_FPGA
-  void TalorFeedOp();
-  void TalorFetchdOp();
   void InjectVariable(const Tensor &t, std::string var_name);
   void FeedData(const Tensor &t);
-  void FeedData(const std::vector<Tensor> &v);
+  void FeedData(const std::vector<void *> &v);
+  void GetResults(std::vector<void *> *v);
   std::shared_ptr<Tensor> FetchResult(int id = -1);
   void Predict_From_To(int start = 0, int end = -1);
   void Predict_From(int start);

src/framework/operator.cpp

@@ -50,6 +50,9 @@ OperatorBase<Dtype>::OperatorBase(const std::string &type,
       attrs_(attrs),
       scope_(scope) {
   CheckAllInputOutputSet();
+#ifdef PADDLE_MOBILE_FPGA
+  InsertTensors();
+#endif
 }
 
 template <typename Dtype>
@@ -133,15 +136,19 @@ void OperatorBase<GPU_CL>::Run() {
 
 #ifdef PADDLE_MOBILE_FPGA
 template <typename Dtype>
-void OperatorBase<Dtype>::ChangeNameMap(string key, std::vector<string> value) {
-  auto it = inputs_.find(key);
-  if (it != inputs_.end()) {
-    inputs_[key] = value;
-    return;
-  }
-  it = outputs_.find(key);
-  if (it != outputs_.end()) {
-    inputs_[key] = value;
+void OperatorBase<Dtype>::InsertTensors() {
+  static int feed_num = 0;
+  static int fetch_num = 0;
+  if (type_ == "feed") {
+    auto new_name = string("feed") + std::to_string(feed_num++);
+    auto var = scope_->Var(new_name);
+    var->template GetMutable<framework::LoDTensor>();
+    inputs_.at("X") = {string(new_name)};
+  } else if (type_ == "fetch") {
+    auto new_name = string("fetch") + std::to_string(fetch_num++);
+    auto var = scope_->Var(new_name);
+    var->template GetMutable<framework::LoDTensor>();
+    outputs_.at("Out") = {string(new_name)};
   }
 }
 #endif

src/framework/operator.h

@@ -79,6 +79,7 @@ class OperatorBase {
     }
   }
 #ifdef PADDLE_MOBILE_FPGA
+  void InsertTensors();
   void ChangeNameMap(string key, std::vector<string> value);
 #endif
  protected:
@@ -95,6 +96,7 @@ class OperatorBase {
 template <typename Dtype, typename ParamType, typename KernelType>
 class OperatorWithKernel : public OperatorBase<Dtype> {
  public:
+#ifndef PADDLE_MOBILE_FPGA1
   OperatorWithKernel(const std::string &type, const VariableNameMap &inputs,
                      const VariableNameMap &outputs, const AttributeMap &attrs,
                      std::shared_ptr<Scope> scope)
@@ -104,7 +106,25 @@ class OperatorWithKernel : public OperatorBase<Dtype> {
     kernel_.InitCLHelper(scope->GetCLScpoe());
 #endif
   }
-
+#else
+  OperatorWithKernel(const std::string &type, const VariableNameMap inputs,
+                     const VariableNameMap &outputs, const AttributeMap &attrs,
+                     std::shared_ptr<Scope> scope)
+      : OperatorBase<Dtype>(type, inputs, outputs, attrs, scope) {
+    static int feed_num = 0;
+    static int fetch_num = 0;
+    if (type == "feed") {
+      auto new_name = string("feed") + std::to_string(feed_num++);
+      auto var = scope->Var(new_name);
+      (const_cast<VariableNameMap &>(inputs)).at("X") = {string(new_name)};
+    } else if (type == "fetch") {
+      auto new_name = string("fetch") + std::to_string(fetch_num++);
+      auto var = scope->Var(new_name);
+      (const_cast<VariableNameMap &>(outputs)).at("Out") = {string(new_name)};
+    }
+    param_ = ParamType(inputs, outputs, attrs, *scope);
+  }
+#endif
   virtual void RunImpl() { this->kernel_.Compute(this->param_); }
 
   virtual void InferShape() const = 0;

src/framework/tensor.h

@@ -202,6 +202,10 @@ class Tensor : public TensorBase {
   inline void reset_data_ptr(void *p) {
     ((PlaceholderImpl *)(holder_.get()))->ptr_.reset((uint8_t *)p);  // NOLINT
   }
+  inline void set_type(std::type_index type) { holder_->set_type(type); }
+  inline void *get_data() {
+    return (void *)(((PlaceholderImpl *)(holder_.get()))->ptr_.get());
+  }  // NOLINT
 
   inline void *init(std::type_index type) {
     if (holder_ != nullptr) {
@@ -217,7 +221,8 @@ class Tensor : public TensorBase {
         reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
   }
 
-  float scale[2];  // scale[0]= MAX/127.0, scale[1]= 127.0/MAX
+  float scale[2];                 // scale[0]= MAX/127.0, scale[1]= 127.0/MAX
+  void *external_data = nullptr;  // only used for Feed
 #endif
 };
 

src/io/api_paddle_mobile.cc

@@ -177,6 +177,23 @@ bool PaddleMobilePredictor<Device, T>::Run(
 
   return true;
 }
+template <typename Device, typename T>
+void PaddleMobilePredictor<Device, T>::FeedData(
+    const std::vector<void *> &inputs) {
+  paddle_mobile_->FeedData(inputs);
+}
+
+template <typename Device, typename T>
+void PaddleMobilePredictor<Device, T>::GetResults(
+    std::vector<void *> *outputs) {
+  paddle_mobile_->GetResults(outputs);
+}
+
+template <typename Device, typename T>
+void PaddleMobilePredictor<Device, T>::Predict_From_To(int start, int end) {
+  paddle_mobile_->Predict_From_To(start, end);
+}
+
 #endif
 template <typename Device, typename T>
 PaddleMobilePredictor<Device, T>::~PaddleMobilePredictor() {

src/io/api_paddle_mobile.h

@@ -35,6 +35,9 @@ class PaddleMobilePredictor : public PaddlePredictor {
   bool Run(const std::vector<PaddleTensor>& inputs,
            std::vector<PaddleTensor>* output_data, std::vector<int>* index_data,
            int batch_size = -1) override;
+  void FeedData(const std::vector<void*>& inputs) override;
+  void GetResults(std::vector<void*>* outputs) override;
+  void Predict_From_To(int start = 0, int end = -1) override;
 #endif
   ~PaddleMobilePredictor() override;
 

src/io/paddle_inference_api.h

@@ -119,6 +119,9 @@ class PaddlePredictor {
   virtual bool Run(const std::vector<PaddleTensor>& inputs,
                    std::vector<PaddleTensor>* output_data,
                    std::vector<int>* index_data, int batch_size = -1) = 0;
+  virtual void FeedData(const std::vector<void*>& inputs) = 0;
+  virtual void GetResults(std::vector<void*>* outputs) = 0;
+  virtual void Predict_From_To(int start = 0, int end = -1) = 0;
 #endif
 
  protected:

src/io/paddle_mobile.cpp

@@ -228,10 +228,14 @@ void PaddleMobile<Device, T>::FeedData(const framework::Tensor &t) {
   executor_->FeedData(t);
 }
 template <typename Device, typename T>
-void PaddleMobile<Device, T>::FeedData(
-    const std::vector<framework::Tensor> &v) {
+void PaddleMobile<Device, T>::FeedData(const std::vector<void *> &v) {
   executor_->FeedData(v);
 };
+template <typename Device, typename T>
+void PaddleMobile<Device, T>::GetResults(std::vector<void *> *v) {
+  executor_->GetResults(v);
+}
+
 template <typename Device, typename T>
 std::shared_ptr<framework::Tensor> PaddleMobile<Device, T>::FetchResult(
     int id) {

src/io/paddle_mobile.h

@@ -90,7 +90,8 @@ class PaddleMobile {
 #ifdef PADDLE_MOBILE_FPGA
   void InjectVariable(const framework::Tensor &t, std::string var_name);
   void FeedData(const framework::Tensor &t);
-  void FeedData(const std::vector<framework::Tensor> &v);
+  void FeedData(const std::vector<void *> &v);
+  void GetResults(std::vector<void *> *v);
   std::shared_ptr<framework::Tensor> FetchResult(int id = -1);
   void Predict_From_To(int start = 0, int end = -1);
   void Predict_From(int start);

src/operators/kernel/detection_kernel.h

@@ -103,6 +103,10 @@ class ProposalParam : public OpParam {
   float nms_thresh_;
   float min_size_;
   float eta_;
+#ifdef PADDLE_MOBILE_FPGA
+  std::shared_ptr<Tensor> float_score, float_bbox;
+  fpga::BypassArgs score_arg, bbox_arg;
+#endif
 };
 
 DECLARE_KERNEL(Proposal, ProposalParam);
@@ -133,6 +137,10 @@ class PSRoiPoolParam : public OpParam {
   int pooled_height_;
   int pooled_width_;
   float spatial_scale_;
+#ifdef PADDLE_MOBILE_FPGA
+  std::shared_ptr<Tensor> float_input, float_output;
+  fpga::BypassArgs input_arg, output_arg;
+#endif
 };
 
 DECLARE_KERNEL(PSRoiPool, PSRoiPoolParam);

src/operators/kernel/fpga/V1/anchor_generator_kernel.cpp

@@ -23,15 +23,46 @@ namespace operators {
 template <>
 bool AnchorGeneratorKernel<FPGA, float>::Init(
     AnchorGeneratorParam<FPGA> *param) {
-  // TODO zhangyang
+  auto input = param->input_;
+  auto anchors = param->output_anchors_;
+  auto anchor_ptr = anchors->mutable_data<float>();
+  auto stride = param->stride_;
+  auto feature_width = input->dims()[3], feature_height = input->dims()[2];
+  auto stride_width = stride[0], stride_height = stride[1];
+
+  int anchors_offset[] = {-2,  -2,   18,   18,  -10, -9,   26,   25,   -23,
+                          -20, 39,   36,   -43, -34, 59,   49,   -63,  -54,
+                          79,  69,   -96,  -77, 112, 93,   -137, -118, 153,
+                          134, -204, -188, 220, 204, -281, -395, 296,  441};
+  int num_anchors = sizeof(anchors_offset) / (sizeof(int) * 4);
+
+  //  DLOG << "feature_height: " << feature_height;
+  //  DLOG << "feature_width: " << feature_width;
+  //  DLOG << "num_anchors: " << num_anchors;
+  //  DLOG << "stride_width: " << stride_width;
+  //  DLOG << "stride_height: " << stride_height;
+
+  for (int h_idx = 0; h_idx < feature_height; ++h_idx) {
+    for (int w_idx = 0; w_idx < feature_width; ++w_idx) {
+      int offset = h_idx * w_idx * num_anchors * 4;
+      for (int idx = 0; idx < num_anchors; idx++) {
+        anchor_ptr[offset + 0] =
+            anchors_offset[idx * 4 + 0] + w_idx * stride_width;
+        anchor_ptr[offset + 1] =
+            anchors_offset[idx * 4 + 1] + h_idx * stride_height;
+        anchor_ptr[offset + 2] =
+            anchors_offset[idx * 4 + 2] + w_idx * stride_width;
+        anchor_ptr[offset + 3] =
+            anchors_offset[idx * 4 + 3] + h_idx * stride_height;
+      }
+    }
+  }
   return true;
 }
 
 template <>
 void AnchorGeneratorKernel<FPGA, float>::Compute(
-    const AnchorGeneratorParam<FPGA> &param) {
-  // TODO(hjchen2)
-}
+    const AnchorGeneratorParam<FPGA> &param) {}
 
 }  // namespace operators
 }  // namespace paddle_mobile

src/operators/kernel/fpga/V1/concat_kernel.cpp

@@ -38,7 +38,7 @@ bool ConcatKernel<FPGA, float>::Init(ConcatParam<FPGA> *param) {
     PADDLE_MOBILE_ENFORCE(
         input->dims()[2] == height && input->dims()[3] == width,
         "Image height & width should be unified");
-    images_in[i] = (half *)input->data<float>();      // NOLINT
+    images_in[i] = input->data<half>();
     channel_num[i] = (uint32_t)inputs[i]->dims()[1];  // NOLINT
     scales_in[i] = input->scale;
   }
@@ -48,7 +48,7 @@ bool ConcatKernel<FPGA, float>::Init(ConcatParam<FPGA> *param) {
   concatArgs.image_num = image_num;
   concatArgs.images_in = images_in;
   concatArgs.scales_in = scales_in;
-  concatArgs.image_out = (half *)out->data<float>();  // NOLINT
+  concatArgs.image_out = out->data<half>();
   concatArgs.scale_out = out->scale;
   concatArgs.channel_num = channel_num;
   concatArgs.height = height;

src/operators/kernel/fpga/V1/elementwise_add_kernel.cpp

@@ -27,10 +27,10 @@ bool ElementwiseAddKernel<FPGA, float>::Init(ElementwiseAddParam<FPGA> *param) {
   auto *input_x = const_cast<LoDTensor *>(param->InputX());
   auto *input_y = const_cast<LoDTensor *>(param->InputY());
   auto *out = param->Out();
-  auto input_x_ptr = input_x->data<float>();
-  auto input_y_ptr = input_y->data<float>();
+  auto input_x_ptr = input_x->data<half>();
+  auto input_y_ptr = input_y->data<half>();
   fpga::format_fp16_ofm(out);
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
 
   fpga::EWAddArgs ewaddArgs = {0};
   // ewaddArgs.relu_enabled = relu_enabled;

src/operators/kernel/fpga/V1/elementwise_add_relu_kernel.cpp

@@ -28,10 +28,10 @@ bool ElementwiseAddReluKernel<FPGA, float>::Init(
   auto *input_x = const_cast<LoDTensor *>(param->InputX());
   auto *input_y = const_cast<LoDTensor *>(param->InputY());
   auto *out = param->Out();
-  auto input_x_ptr = input_x->data<float>();
-  auto input_y_ptr = input_y->data<float>();
+  auto input_x_ptr = input_x->data<half>();
+  auto input_y_ptr = input_y->data<half>();
   fpga::format_fp16_ofm(out);
-  auto out_ptr = out->mutable_data<float>();
+  auto out_ptr = out->mutable_data<half>();
 
   fpga::EWAddArgs ewaddArgs = {0};
   // ewaddArgs.relu_enabled = relu_enabled;

src/operators/kernel/fpga/V1/feed_kernel.cpp

@@ -19,19 +19,35 @@ namespace operators {
 
 template <>
 bool FeedKernel<FPGA, float>::Init(FeedParam<FPGA> *param) {
-  Tensor *output = param->Out();
+  auto output = param->Out();
+  auto input = const_cast<LoDTensor *>(param->InputX());
+  input->init(typeid(float));
+  input->Resize(output->dims());
+
+  if (output->dims().size() != 4) {
+    auto input_ptr = input->mutable_data<float>();
+    size_t size = output->numel() * sizeof(float);
+    auto p = fpga::fpga_malloc(size);
+    memcpy(p, input_ptr, size);
+    output->reset_data_ptr(p);
+    return true;
+  }
   fpga::format_fp16_ofm(output);
   return true;
 }
 
 template <>
 void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
-  auto input =
-      reinterpret_cast<Tensor *>(const_cast<LoDTensor *>(param.InputX()));
+  auto output = param.Out();
+  auto input = const_cast<LoDTensor *>(param.InputX());
+
+  if (input->dims().size() != 4) {
+    return;
+  }
+
   fpga::format_image(input);
   auto input_ptr = input->data<float>();
-  Tensor *output = param.Out();
-  auto output_ptr = output->data<float>();
+  auto output_ptr = output->data<half>();
 
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP32};
 
@@ -39,7 +55,7 @@ void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
   args.output_data_type = fpga::DATA_TYPE_FP16;
   args.input_layout_type = fpga::LAYOUT_CHW;
   args.output_layout_type = fpga::LAYOUT_HWC;
-  args.image.address = reinterpret_cast<void *>(input_ptr);
+  args.image.address = input_ptr;
   args.image.channels = (uint32_t)input->dims()[1];
   args.image.height = (uint32_t)input->dims()[2];
   args.image.width = (uint32_t)input->dims()[3];
@@ -48,6 +64,8 @@ void FeedKernel<FPGA, float>::Compute(const FeedParam<FPGA> &param) {
   args.output.address = output_ptr;
   args.output.scale_address = output->scale;
   fpga::PerformBypass(args);
+
+  input->external_data = nullptr;
 }
 template class FeedKernel<FPGA, float>;
 

src/operators/kernel/fpga/V1/fetch_kernel.cpp

@@ -19,20 +19,15 @@ namespace operators {
 
 template <>
 bool FetchKernel<FPGA, float>::Init(FetchParam<FPGA> *param) {
-  Tensor *output = param->Out();
-  // fpga::format_fp16_ofm(output);
-  return true;
-}
-
-template <>
-void FetchKernel<FPGA, float>::Compute(const FetchParam<FPGA> &param) {
-  param.Out()->ShareDataWith(*(param.InputX()));
-  /*auto input =
-          reinterpret_cast<Tensor *>(const_cast<Tensor *>(param.InputX()));
-  fpga::format_image(input);
-  auto input_ptr = input->data<float>();
-  Tensor *output = param.Out();
-  auto output_ptr = output->data<float>();
+  auto input = const_cast<Tensor *>(param->InputX());
+  auto output = param->Out();
+  if (input->type() == typeid(float)) {
+    output->ShareDataWith(*input);
+    return true;
+  }
+  output->init(typeid(float));
+  output->Resize(input->dims());
+  fpga::format_fp32_ofm(output);
 
   fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
 
@@ -40,13 +35,28 @@ void FetchKernel<FPGA, float>::Compute(const FetchParam<FPGA> &param) {
   args.output_data_type = fpga::DATA_TYPE_FP32;
   args.input_layout_type = fpga::LAYOUT_CHW;
   args.output_layout_type = fpga::LAYOUT_HWC;
-  args.image.address = reinterpret_cast<void *>(input_ptr);
-  args.image.channels = (uint32_t)input->dims()[1];
-  args.image.height = (input->dims().size() == 4) ? (uint32_t)input->dims()[2] :
-  1; args.image.width = (input->dims().size() == 4) ? (uint32_t)input->dims()[3]
-  : 1; args.image.pad_height = 0; args.image.pad_width = 0; args.output.address
-  = output_ptr; args.output.scale_address = output->scale;
-  fpga::PerformBypass(args);*/
+  args.image.address = input->data<half>();
+  args.image.channels = (uint32_t)product(input->dims());
+  args.image.height = 1;
+  args.image.width = 1;
+  args.image.pad_height = 0;
+  args.image.pad_width = 0;
+  args.output.address = output->data<float>();
+  args.output.scale_address = output->scale;
+  param->fpga_bypass_args = args;
+
+  return true;
+}
+
+template <>
+void FetchKernel<FPGA, float>::Compute(const FetchParam<FPGA> &param) {
+  auto input = param.InputX();
+  if (input->type() == typeid(float)) {
+    return;
+  }
+  fpga::PerformBypass(param.fpga_bypass_args);
+
+  // TODO: DEalign: get rid of extra 0
 }
 
 template class FetchKernel<FPGA, float>;

src/operators/kernel/fpga/V1/pool_kernel.cpp

@@ -22,10 +22,10 @@ namespace operators {
 template <>
 bool PoolKernel<FPGA, float>::Init(PoolParam<FPGA> *param) {
   auto *input = const_cast<Tensor *>(param->Input());
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   Tensor *output = param->Output();
   fpga::format_fp16_ofm(output);
-  auto output_ptr = output->mutable_data<float>();
+  auto output_ptr = output->mutable_data<half>();
   vector<int> ksize = param->Ksize();
   vector<int> strides = param->Strides();
   vector<int> paddings = param->Paddings();

src/operators/kernel/fpga/V1/proposal_kernel.cpp

@@ -14,20 +14,422 @@ limitations under the License. */
 
 #ifdef PROPOSAL_OP
 
+#include <algorithm>
 #include <vector>
 #include "operators/kernel/detection_kernel.h"
 
 namespace paddle_mobile {
 namespace operators {
+static const double kBBoxClipDefault = std::log(1000.0 / 16.0);
 
 template <>
 bool ProposalKernel<FPGA, float>::Init(ProposalParam<FPGA> *param) {
+  int post_nms_top_n = param->post_nms_topn_;
+  int64_t batch = param->scores_->dims()[0];
+  auto total = post_nms_top_n * batch;
+  param->rpn_rois_->mutable_data<float>({total, 4});
+  param->rpn_probs_->mutable_data<float>({total, 1});
+
+  //  DLOG << *param->rpn_rois_;
+  //  DLOG << *param->rpn_probs_;
+
+  param->float_bbox = std::make_shared<Tensor>();
+  param->float_bbox->Resize(param->bbox_deltas_->dims());
+  param->float_bbox->init(typeid(float));
+  fpga::format_fp32_ofm(param->float_bbox.get());
+  param->float_score = std::make_shared<Tensor>();
+  param->float_score->Resize(param->scores_->dims());
+  param->float_score->init(typeid(float));
+  fpga::format_fp32_ofm(param->float_score.get());
+
+  auto input = param->bbox_deltas_;
+  fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
+  args.input_layout_type = fpga::LAYOUT_HWC;
+  args.output_layout_type = fpga::LAYOUT_HWC;
+  args.input_data_type = fpga::DATA_TYPE_FP16;
+  args.output_data_type = fpga::DATA_TYPE_FP32;
+  args.image.address = input->data<half>();
+  args.image.height = (uint32_t)input->dims()[2];
+  args.image.width = (uint32_t)input->dims()[3];
+  args.image.channels = (uint32_t)input->dims()[1];
+  args.output.address = param->float_bbox->mutable_data<float>();
+  args.output.scale_address = param->float_bbox->scale;
+  param->bbox_arg = args;
+
+  input = param->scores_;
+  args.image.address = input->data<half>();
+  args.image.height = (uint32_t)input->dims()[2];
+  args.image.width = (uint32_t)input->dims()[3];
+  args.image.channels = (uint32_t)input->dims()[1];
+  args.output.address = param->float_score->mutable_data<float>();
+  args.output.scale_address = param->float_score->scale;
+  param->score_arg = args;
+
   return true;
 }
 
+void AppendProposals(Tensor *dst, int64_t offset, const Tensor &src) {
+  auto *out_data = dst->data<void>();
+  auto *to_add_data = src.data<void>();
+  size_t size_of_t = framework::SizeOfType(src.type());
+  offset *= size_of_t;
+  std::memcpy(
+      reinterpret_cast<void *>(reinterpret_cast<uintptr_t>(out_data) + offset),
+      to_add_data, src.numel() * size_of_t);
+}
+
+template <class T>
+static inline void BoxCoder(Tensor *all_anchors, Tensor *bbox_deltas,
+                            Tensor *variances, Tensor *proposals) {
+  T *proposals_data = proposals->mutable_data<T>();
+
+  int64_t row = all_anchors->dims()[0];
+  int64_t len = all_anchors->dims()[1];
+
+  auto *bbox_deltas_data = bbox_deltas->data<T>();
+  auto *anchor_data = all_anchors->data<T>();
+  const T *variances_data = nullptr;
+  if (variances) {
+    variances_data = variances->data<T>();
+  }
+
+  for (int64_t i = 0; i < row; ++i) {
+    T anchor_width = anchor_data[i * len + 2] - anchor_data[i * len] + 1.0;
+    T anchor_height = anchor_data[i * len + 3] - anchor_data[i * len + 1] + 1.0;
+
+    T anchor_center_x = anchor_data[i * len] + 0.5 * anchor_width;
+    T anchor_center_y = anchor_data[i * len + 1] + 0.5 * anchor_height;
+
+    T bbox_center_x = 0, bbox_center_y = 0;
+    T bbox_width = 0, bbox_height = 0;
+
+    if (variances) {
+      bbox_center_x =
+          variances_data[i * len] * bbox_deltas_data[i * len] * anchor_width +
+          anchor_center_x;
+      bbox_center_y = variances_data[i * len + 1] *
+                          bbox_deltas_data[i * len + 1] * anchor_height +
+                      anchor_center_y;
+      bbox_width = std::exp(std::min<T>(variances_data[i * len + 2] *
+                                            bbox_deltas_data[i * len + 2],
+                                        kBBoxClipDefault)) *
+                   anchor_width;
+      bbox_height = std::exp(std::min<T>(variances_data[i * len + 3] *
+                                             bbox_deltas_data[i * len + 3],
+                                         kBBoxClipDefault)) *
+                    anchor_height;
+    } else {
+      bbox_center_x =
+          bbox_deltas_data[i * len] * anchor_width + anchor_center_x;
+      bbox_center_y =
+          bbox_deltas_data[i * len + 1] * anchor_height + anchor_center_y;
+      bbox_width = std::exp(std::min<T>(bbox_deltas_data[i * len + 2],
+                                        kBBoxClipDefault)) *
+                   anchor_width;
+      bbox_height = std::exp(std::min<T>(bbox_deltas_data[i * len + 3],
+                                         kBBoxClipDefault)) *
+                    anchor_height;
+    }
+
+    proposals_data[i * len] = bbox_center_x - bbox_width / 2;
+    proposals_data[i * len + 1] = bbox_center_y - bbox_height / 2;
+    proposals_data[i * len + 2] = bbox_center_x + bbox_width / 2 - 1;
+    proposals_data[i * len + 3] = bbox_center_y + bbox_height / 2 - 1;
+  }
+  // return proposals;
+}
+
+template <class T>
+static inline void ClipTiledBoxes(const Tensor &im_info, Tensor *boxes) {
+  T *boxes_data = boxes->mutable_data<T>();
+  const T *im_info_data = im_info.data<T>();
+  T zero(0);
+  for (int64_t i = 0; i < boxes->numel(); ++i) {
+    if (i % 4 == 0) {
+      boxes_data[i] =
+          std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero);
+    } else if (i % 4 == 1) {
+      boxes_data[i] =
+          std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero);
+    } else if (i % 4 == 2) {
+      boxes_data[i] =
+          std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero);
+    } else {
+      boxes_data[i] =
+          std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero);
+    }
+  }
+}
+
+template <class T>
+static inline void FilterBoxes(Tensor *boxes, float min_size,
+                               const Tensor &im_info, Tensor *keep) {
+  const T *im_info_data = im_info.data<T>();
+  T *boxes_data = boxes->mutable_data<T>();
+  T im_scale = im_info_data[2];
+  keep->Resize({boxes->dims()[0]});
+  min_size = std::max(min_size, 1.0f);
+  int *keep_data = keep->mutable_data<int>();
+
+  int keep_len = 0;
+  for (int i = 0; i < boxes->dims()[0]; ++i) {
+    T ws = boxes_data[4 * i + 2] - boxes_data[4 * i] + 1;
+    T hs = boxes_data[4 * i + 3] - boxes_data[4 * i + 1] + 1;
+    T ws_origin_scale =
+        (boxes_data[4 * i + 2] - boxes_data[4 * i]) / im_scale + 1;
+    T hs_origin_scale =
+        (boxes_data[4 * i + 3] - boxes_data[4 * i + 1]) / im_scale + 1;
+    T x_ctr = boxes_data[4 * i] + ws / 2;
+    T y_ctr = boxes_data[4 * i + 1] + hs / 2;
+    if (ws_origin_scale >= min_size && hs_origin_scale >= min_size &&
+        x_ctr <= im_info_data[1] && y_ctr <= im_info_data[0]) {
+      keep_data[keep_len++] = i;
+    }
+  }
+  keep->Resize({keep_len});
+}
+
+template <class T>
+static inline std::vector<std::pair<T, int>> GetSortedScoreIndex(
+    const std::vector<T> &scores) {
+  std::vector<std::pair<T, int>> sorted_indices;
+  sorted_indices.reserve(scores.size());
+  for (size_t i = 0; i < scores.size(); ++i) {
+    sorted_indices.emplace_back(scores[i], i);
+  }
+  // Sort the score pair according to the scores in descending order
+  std::stable_sort(sorted_indices.begin(), sorted_indices.end(),
+                   [](const std::pair<T, int> &a, const std::pair<T, int> &b) {
+                     return a.first < b.first;
+                   });
+  return sorted_indices;
+}
+
+template <class T>
+static inline T BBoxArea(const T *box, bool normalized) {
+  if (box[2] < box[0] || box[3] < box[1]) {
+    // If coordinate values are is invalid
+    // (e.g. xmax < xmin or ymax < ymin), return 0.
+    return static_cast<T>(0.);
+  } else {
+    const T w = box[2] - box[0];
+    const T h = box[3] - box[1];
+    if (normalized) {
+      return w * h;
+    } else {
+      // If coordinate values are not within range [0, 1].
+      return (w + 1) * (h + 1);
+    }
+  }
+}
+
+template <typename T>
+static inline Tensor VectorToTensor(const std::vector<T> &selected_indices,
+                                    int selected_num) {
+  Tensor keep_nms;
+  keep_nms.Resize({selected_num});
+  auto *keep_data = keep_nms.mutable_data<T>();
+  for (int i = 0; i < selected_num; ++i) {
+    keep_data[i] = selected_indices[i];
+  }
+  return keep_nms;
+}
+
+template <class T>
+static inline T JaccardOverlap(const T *box1, const T *box2, bool normalized) {
+  if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] ||
+      box2[3] < box1[1]) {
+    return static_cast<T>(0.);
+  } else {
+    const T inter_xmin = std::max(box1[0], box2[0]);
+    const T inter_ymin = std::max(box1[1], box2[1]);
+    const T inter_xmax = std::min(box1[2], box2[2]);
+    const T inter_ymax = std::min(box1[3], box2[3]);
+    const T inter_w = std::max(T(0), inter_xmax - inter_xmin + 1);
+    const T inter_h = std::max(T(0), inter_ymax - inter_ymin + 1);
+    const T inter_area = inter_w * inter_h;
+    const T bbox1_area = BBoxArea<T>(box1, normalized);
+    const T bbox2_area = BBoxArea<T>(box2, normalized);
+    return inter_area / (bbox1_area + bbox2_area - inter_area);
+  }
+}
+
+template <class T>
+static inline Tensor NMS(Tensor *bbox, Tensor *scores, T nms_threshold,
+                         float eta) {
+  int64_t num_boxes = bbox->dims()[0];
+  // 4: [xmin ymin xmax ymax]
+  int64_t box_size = bbox->dims()[1];
+
+  std::vector<T> scores_data(num_boxes);
+  std::copy_n(scores->data<T>(), num_boxes, scores_data.begin());
+  std::vector<std::pair<T, int>> sorted_indices =
+      GetSortedScoreIndex<T>(scores_data);
+
+  std::vector<int> selected_indices;
+  int selected_num = 0;
+  T adaptive_threshold = nms_threshold;
+  const T *bbox_data = bbox->data<T>();
+  while (sorted_indices.size() != 0) {
+    int idx = sorted_indices.back().second;
+    bool flag = true;
+    for (int kept_idx : selected_indices) {
+      if (flag) {
+        T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
+                                      bbox_data + kept_idx * box_size, false);
+        flag = (overlap <= adaptive_threshold);
+      } else {
+        break;
+      }
+    }
+    if (flag) {
+      selected_indices.push_back(idx);
+      ++selected_num;
+    }
+    sorted_indices.erase(sorted_indices.end() - 1);
+    if (flag && eta < 1 && adaptive_threshold > 0.5) {
+      adaptive_threshold *= eta;
+    }
+  }
+  return VectorToTensor(selected_indices, selected_num);
+}
+
+template <typename T>
+std::pair<Tensor, Tensor> ProposalForOneImage(
+    const Tensor &im_info_slice, const Tensor &anchors, const Tensor &variances,
+    const Tensor &bbox_deltas_slice,  // [M, 4]
+    const Tensor &scores_slice,       // [N, 1]
+    int pre_nms_top_n, int post_nms_top_n, float nms_thresh, float min_size,
+    float eta) {
+  auto *scores_data = scores_slice.data<T>();
+
+  // Sort index
+  Tensor index_t;
+  index_t.Resize({scores_slice.numel()});
+  int *index = index_t.mutable_data<int>();
+  for (int i = 0; i < scores_slice.numel(); ++i) {
+    index[i] = i;
+  }
+  auto compare = [scores_data](const int64_t &i, const int64_t &j) {
+    return scores_data[i] > scores_data[j];
+  };
+
+  if (pre_nms_top_n <= 0 || pre_nms_top_n >= scores_slice.numel()) {
+    std::sort(index, index + scores_slice.numel(), compare);
+  } else {
+    std::nth_element(index, index + pre_nms_top_n, index + scores_slice.numel(),
+                     compare);
+    index_t.Resize({pre_nms_top_n});
+  }
+
+  Tensor scores_sel, bbox_sel, anchor_sel, var_sel;
+  scores_sel.mutable_data<T>({index_t.numel(), 1});
+  bbox_sel.mutable_data<T>({index_t.numel(), 4});
+  anchor_sel.mutable_data<T>({index_t.numel(), 4});
+  var_sel.mutable_data<T>({index_t.numel(), 4});
+
+  Tensor proposals;
+  proposals.mutable_data<T>({index_t.numel(), 4});
+  BoxCoder<T>(&anchor_sel, &bbox_sel, &var_sel, &proposals);
+
+  ClipTiledBoxes<T>(im_info_slice, &proposals);
+
+  Tensor keep;
+  FilterBoxes<T>(&proposals, min_size, im_info_slice, &keep);
+
+  Tensor scores_filter;
+  bbox_sel.mutable_data<T>({keep.numel(), 4});
+  scores_filter.mutable_data<T>({keep.numel(), 1});
+
+  if (nms_thresh <= 0) {
+    return std::make_pair(bbox_sel, scores_filter);
+  }
+
+  Tensor keep_nms = NMS<T>(&bbox_sel, &scores_filter, nms_thresh, eta);
+
+  if (post_nms_top_n > 0 && post_nms_top_n < keep_nms.numel()) {
+    keep_nms.Resize({post_nms_top_n});
+  }
+
+  proposals.mutable_data<T>({keep_nms.numel(), 4});
+  scores_sel.mutable_data<T>({keep_nms.numel(), 1});
+
+  return std::make_pair(proposals, scores_sel);
+}
+
 template <>
 void ProposalKernel<FPGA, float>::Compute(const ProposalParam<FPGA> &param) {
-  // TODO(hjchen2)
+  auto score_tensor = param.float_score.get();
+  fpga::PerformBypass(param.score_arg);
+  fpga::fpga_invalidate(score_tensor->data<float>(),
+                        score_tensor->numel() * sizeof(float));
+
+  auto bbox_tensor = param.float_bbox.get();
+  fpga::PerformBypass(param.bbox_arg);
+  fpga::fpga_invalidate(bbox_tensor->data<float>(),
+                        bbox_tensor->numel() * sizeof(float));
+
+  auto *scores = param.float_score.get();
+  auto *bbox_deltas = param.float_bbox.get();
+  auto *im_info = param.im_info_;
+  auto anchors = *param.anchors_;
+  auto variances = *param.variances_;
+
+  auto *rpn_rois = param.rpn_rois_;
+  auto *rpn_roi_probs = param.rpn_probs_;
+
+  int pre_nms_top_n = param.pre_nms_topn_;
+  int post_nms_top_n = param.post_nms_topn_;
+  float nms_thresh = param.nms_thresh_;
+  float min_size = param.min_size_;
+  float eta = param.eta_;
+
+  auto &scores_dim = scores->dims();
+  int64_t num = scores_dim[0];
+  int64_t c_score = scores_dim[1];
+  int64_t h_score = scores_dim[2];
+  int64_t w_score = scores_dim[3];
+
+  auto &bbox_dim = bbox_deltas->dims();
+  int64_t c_bbox = bbox_dim[1];
+  int64_t h_bbox = bbox_dim[2];
+  int64_t w_bbox = bbox_dim[3];
+
+  //
+  Tensor bbox_deltas_swap, scores_swap;
+  bbox_deltas_swap.mutable_data<float>({num, h_bbox, w_bbox, c_bbox});
+  scores_swap.mutable_data<float>({num, h_score, w_score, c_score});
+
+  framework::LoD lod;
+  lod.resize(1);
+  auto &lod0 = lod[0];
+  lod0.push_back(0);
+  anchors.Resize({anchors.numel() / 4, 4});
+
+  int64_t num_proposals = 0;
+  for (int64_t i = 0; i < num; ++i) {
+    Tensor im_info_slice = im_info->Slice(i, i + 1);
+    Tensor bbox_deltas_slice = bbox_deltas_swap.Slice(i, i + 1);
+    Tensor scores_slice = scores_swap.Slice(i, i + 1);
+
+    bbox_deltas_slice.Resize({h_bbox * w_bbox * c_bbox / 4, 4});
+    scores_slice.Resize({h_score * w_score * c_score, 1});
+
+    std::pair<Tensor, Tensor> tensor_pair = ProposalForOneImage<float>(
+        im_info_slice, anchors, variances, bbox_deltas_slice, scores_slice,
+        pre_nms_top_n, post_nms_top_n, nms_thresh, min_size, eta);
+    Tensor &proposals = tensor_pair.first;
+    Tensor &scores = tensor_pair.second;
+
+    AppendProposals(rpn_rois, 4 * num_proposals, proposals);
+    AppendProposals(rpn_roi_probs, num_proposals, scores);
+    num_proposals += proposals.dims()[0];
+    lod0.push_back(num_proposals);
+  }
+  rpn_rois->set_lod(lod);
+  rpn_roi_probs->set_lod(lod);
+  rpn_rois->Resize({num_proposals, 4});
+  rpn_roi_probs->Resize({num_proposals, 1});
 }
 
 }  // namespace operators

src/operators/kernel/fpga/V1/psroi_pool_kernel.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #ifdef PSROI_POOL_OP
 
+#include <cmath>
 #include <vector>
 #include "operators/kernel/detection_kernel.h"
 
@@ -21,13 +22,180 @@ namespace paddle_mobile {
 namespace operators {
 
 template <>
-bool PSRoiPoolKernel<FPGA, float>::Init(PSRoiPoolParam<FPGA> *param) {
+bool PSRoiPoolKernel<FPGA, float>::Init(PSRoiPoolParam<FPGA>* param) {
+  auto dims = param->input_x_->dims();
+  PADDLE_MOBILE_ENFORCE(dims[1] * dims[3] % IMAGE_ALIGNMENT == 0,
+                        "data not aligned");
+
+  param->float_input = std::make_shared<Tensor>();
+  param->float_input->mutable_data<float>(param->input_x_->dims());
+  param->float_output = std::make_shared<Tensor>();
+  param->float_output->mutable_data<float>(param->output_->dims());
+
+  auto input = param->input_x_;
+  fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
+  args.input_layout_type = fpga::LAYOUT_HWC;
+  args.output_layout_type = fpga::LAYOUT_HWC;
+  args.input_data_type = fpga::DATA_TYPE_FP16;
+  args.output_data_type = fpga::DATA_TYPE_FP32;
+  args.image.address = input->data<half>();
+  args.image.height = (uint32_t)input->dims()[2];
+  args.image.width = (uint32_t)input->dims()[3];
+  args.image.channels = (uint32_t)input->dims()[1];
+  args.output.address = param->float_input->mutable_data<float>();
+  args.output.scale_address = param->float_input->scale;
+  param->input_arg = args;
+
+  fpga::format_fp16_ofm(param->output_);
+
+  input = param->float_output.get();
+  args.input_data_type = fpga::DATA_TYPE_FP32;
+  args.output_data_type = fpga::DATA_TYPE_FP16;
+  args.image.address = input->data<float>();
+  args.image.height = (uint32_t)input->dims()[2];
+  args.image.width = (uint32_t)input->dims()[3];
+  args.image.channels = (uint32_t)input->dims()[1];
+  args.output.address = param->output_->mutable_data<half>();
+  args.output.scale_address = param->output_->scale;
+  param->input_arg = args;
+
   return true;
 }
 
 template <>
-void PSRoiPoolKernel<FPGA, float>::Compute(const PSRoiPoolParam<FPGA> &param) {
-  // TODO(hjchen2)
+void PSRoiPoolKernel<FPGA, float>::Compute(const PSRoiPoolParam<FPGA>& param) {
+  auto input_tensor = param.float_input.get();
+  fpga::PerformBypass(param.input_arg);
+  fpga::fpga_invalidate(input_tensor->data<float>(),
+                        input_tensor->numel() * sizeof(float));
+
+  auto* in = input_tensor;
+  auto* rois = param.input_rois_;
+  auto* out = param.float_output.get();
+
+  auto pooled_height = param.pooled_height_;
+  auto pooled_width = param.pooled_width_;
+  auto spatial_scale = param.spatial_scale_;
+  auto output_channels = param.output_channels_;
+
+  auto in_dims = in->dims();
+  int batch_size = in_dims[0];
+  int input_channels = in_dims[1];
+  int height = in_dims[2];
+  int width = in_dims[3];
+  int rois_num = rois->dims()[0];
+
+  // TODO   auto in_stride = framework::stride(in_dims);
+  // TODO   auto out_stride = framework::stride(out->dims());
+  auto in_stride =
+      framework::stride({batch_size, height, width, input_channels});
+  auto out_stride = framework::stride(
+      {out->dims()[0], out->dims()[2], out->dims()[3], out->dims()[1]});
+
+  const float* input_data = in->data<float>();
+  framework::Tensor rois_batch_id_list;
+  rois_batch_id_list.Resize({rois_num});
+  auto rois_batch_id_data = rois_batch_id_list.mutable_data<int>();
+  return;
+
+  PADDLE_MOBILE_ENFORCE(rois->NumLevels() > 0, "ROIS should not be empty");
+
+  auto rois_lod = rois->lod().back();
+  int rois_batch_size = rois_lod.size() - 1;
+  PADDLE_MOBILE_ENFORCE(
+      rois_batch_size == batch_size,
+      "the rois_batch_size and input(X) batch_size should be the same.");
+  int rois_num_with_lod = rois_lod[rois_batch_size];
+  PADDLE_MOBILE_ENFORCE(rois_num_with_lod == rois_num,
+                        "the rois_num from input and lod must be the same");
+
+  PADDLE_MOBILE_ENFORCE(
+      input_channels == output_channels * pooled_height * pooled_width,
+      "the channels of input X should equal the product of "
+      "output_channels x pooled_height x pooled_width");
+
+  // calculate batch id index for each roi according to LoD
+  for (int n = 0; n < rois_batch_size; ++n) {
+    for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) {
+      rois_batch_id_data[i] = n;
+    }
+  }
+  auto output_data = out->mutable_data<float>();
+  auto input_rois = rois->data<float>();
+
+  // calculate psroipooling, parallel processing can be implemented per ROI
+  for (int n = 0; n < rois_num; ++n) {
+    // set roi batch id
+    int roi_batch_id = rois_batch_id_data[n];
+
+    // [start, end) interval for spatial sampling
+    auto offset_input_rois = input_rois + n * 4;
+    auto roi_start_w =
+        static_cast<float>(round(offset_input_rois[0])) * spatial_scale;
+    auto roi_start_h =
+        static_cast<float>(round(offset_input_rois[1])) * spatial_scale;
+    auto roi_end_w =
+        static_cast<float>(round(offset_input_rois[2]) + 1.) * spatial_scale;
+    auto roi_end_h =
+        static_cast<float>(round(offset_input_rois[3]) + 1.) * spatial_scale;
+
+    // Force too small rois to be 1 x 1
+    auto roi_height = std::max(roi_end_h - roi_start_h, 0.1f);  // avoid 0
+    auto roi_width = std::max(roi_end_w - roi_start_w, 0.1f);
+
+    // Compute bin size w and h at input feature map
+    auto bin_size_h = roi_height / static_cast<float>(pooled_height);
+    auto bin_size_w = roi_width / static_cast<float>(pooled_width);
+    DLOG << 3;
+
+    // calculate each pixel of the output feature map.
+    int out_roi_offset = n * out_stride[0];
+    for (int c = 0; c < output_channels; ++c) {
+      // per category
+      // int out_plane_offset = out_roi_offset + c * out_stride[1];
+      int out_plane_offset = out_roi_offset + c;
+      for (int ph = 0; ph < pooled_height; ++ph) {
+        // TODO         int out_row_offset = out_plane_offset + ph *
+        // out_stride[2];
+        int out_row_offset = out_plane_offset + ph * out_stride[1];
+        for (int pw = 0; pw < pooled_width; ++pw) {
+          // calculate w and h at input feature map
+          int hstart = floor(static_cast<float>(ph) * bin_size_h + roi_start_h);
+          int wstart = floor(static_cast<float>(pw) * bin_size_w + roi_start_w);
+          int hend =
+              ceil(static_cast<float>(ph + 1) * bin_size_h + roi_start_h);
+          int wend =
+              ceil(static_cast<float>(pw + 1) * bin_size_w + roi_start_w);
+          //  Add roi offsets and clip to input boundaries
+          hstart = std::min(std::max(hstart, 0), height);
+          wstart = std::min(std::max(wstart, 0), width);
+          hend = std::min(std::max(hend, 0), height);
+          wend = std::min(std::max(wend, 0), width);
+
+          // TODO           int output_index = out_row_offset + pw;
+          int output_index = out_row_offset + pw * output_channels;
+          int input_channel = (c * pooled_height + ph) * pooled_width + pw;
+          // TODO          int input_plane_offset =
+          // TODO           roi_batch_id * in_stride[0] + input_channel *
+          // in_stride[1];
+          int input_plane_offset = roi_batch_id * in_stride[0] + input_channel;
+          auto offset_input_data = input_data + input_plane_offset;
+          float out_sum = 0.;
+          bool is_empty = (hend <= hstart) || (wend <= wstart);
+          for (int ih = hstart; ih < hend; ++ih) {
+            for (int iw = wstart; iw < wend; ++iw) {
+              int input_index = ih * in_stride[1] + iw * input_channel;
+              out_sum += offset_input_data[input_index];
+            }
+          }
+          float bin_area = (hend - hstart) * (wend - wstart);
+          output_data[output_index] = is_empty ? 0. : out_sum / bin_area;
+        }
+      }
+    }
+  }
+  fpga::format_image(out);
+  fpga::PerformBypass(param.output_arg);
 }
 
 }  // namespace operators

src/operators/kernel/fpga/V1/reshape2_kernel.cpp

@@ -15,18 +15,61 @@ limitations under the License. */
 #ifdef RESHAPE2_OP
 
 #include "operators/kernel/reshape2_kernel.h"
+#include "framework/ddim.h"
 
 namespace paddle_mobile {
 namespace operators {
 
 template <>
 bool Reshape2Kernel<FPGA, float>::Init(Reshape2Param<FPGA> *param) {
+  auto input = const_cast<LoDTensor *>(param->InputX());
+  auto output = param->Out();
+  auto shape = param->Shape();
+  output->ShareDataWith(*input);
+
+  auto num_in = framework::product(input->dims());
+  auto num_shape = framework::product(framework::make_ddim(shape));
+  PADDLE_MOBILE_ENFORCE(num_shape != 0, "0 index is not supported");
+
+  for (int i = 0; i < shape.size(); i++) {
+    if (shape[i] == -1) {
+      shape[i] = static_cast<int>(-num_in / num_shape);
+      break;
+    }
+  }
+  output->Resize(framework::make_ddim(shape));
+  DLOG << "input: " << input;
+  DLOG << "output: " << output;
+
   return true;
 }
 
 template <>
 void Reshape2Kernel<FPGA, float>::Compute(const Reshape2Param<FPGA> &param) {
-  return;
+  auto input = const_cast<LoDTensor *>(param.InputX());
+  auto output = param.Out();
+  auto shape = param.Shape();
+
+  if (output->type() != typeid(half)) {
+    DLOG << "wrong type";
+  }
+
+  auto num_in = framework::product(input->dims());
+  auto num_shape = framework::product(framework::make_ddim(shape));
+  PADDLE_MOBILE_ENFORCE(num_shape != 0, "0 index is not supported");
+
+  for (int i = 0; i < shape.size(); i++) {
+    if (shape[i] == -1) {
+      shape[i] = static_cast<int>(-num_in / num_shape);
+      break;
+    }
+  }
+  output->Resize(framework::make_ddim(shape));
+  if (output->type() != typeid(half)) {
+    DLOG << "wrong type";
+    DLOG << output;
+  }
+  //
 }
 
 }  // namespace operators

src/operators/kernel/fpga/V1/sigmoid_kernel.cpp

@@ -25,7 +25,7 @@ bool SigmoidKernel<FPGA, float>::Init(SigmoidParam<FPGA> *param) {
       paddle_mobile::fpga::SIGMOID;
   int16_t leaky_relu_negative_slope = 0;
   auto input = const_cast<Tensor *>(param->InputX());
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   auto out = param->Out();
   fpga::format_fp16_ofm(out);
 
@@ -38,7 +38,7 @@ bool SigmoidKernel<FPGA, float>::Init(SigmoidParam<FPGA> *param) {
   args.image.width =
       (input->dims().size() == 4) ? (uint32_t)input->dims()[3] : 1;
   args.image.channels = (uint32_t)input->dims()[1];
-  args.output.address = out->data<float>();
+  args.output.address = out->data<half>();
   args.output.scale_address = out->scale;
   args.output.activation.activation_type = activation_enable;
   args.output.activation.leaky_relu_negative_slope = leaky_relu_negative_slope;

src/operators/kernel/fpga/V1/slice_kernel.cpp

@@ -21,10 +21,37 @@ namespace operators {
 
 template <>
 bool SliceKernel<FPGA, float>::Init(SliceParam<FPGA>* param) {
+  auto output = param->output_;
+  fpga::format_fp16_ofm(output);
+  DLOG << "input: " << param->input_;
+  DLOG << "output: " << param->output_;
+  if (param->input_->type() != typeid(half)) {
+    DLOG << "wrong type";
+  }
   return true;
 }
 template <>
-void SliceKernel<FPGA, float>::Compute(const SliceParam<FPGA>& param) {}
+void SliceKernel<FPGA, float>::Compute(const SliceParam<FPGA>& param) {
+  // Only support slicing in channel dimension
+
+  auto input = param.input_;
+  DLOG << input;
+  int HW = input->dims()[2] * input->dims()[3];
+  int channel = input->dims()[1];
+  auto input_ptr = input->data<half>();
+  auto output_ptr = param.output_->data<half>();
+
+  int start = param.starts_[0], end = param.ends_[0];
+  start = start < 0 ? start + channel : start;
+  end = end < 0 ? end + channel : end;
+  start = start > channel ? channel : start;
+  end = end > channel ? channel : end;
+  int len = end - start;
+
+  for (int i = 0; i < HW; i++) {
+    memcpy(output_ptr + len * i, input_ptr + i * channel + start, len);
+  }
+}
 }  // namespace operators
 }  // namespace paddle_mobile
 #endif

src/operators/kernel/fpga/V1/softmax_kernel.cpp

@@ -23,49 +23,72 @@ namespace operators {
 template <>
 bool SoftmaxKernel<FPGA, float>::Init(SoftmaxParam<FPGA> *param) {
   auto input = const_cast<LoDTensor *>(param->InputX());
-  auto input_ptr = input->data<float>();
+  auto input_ptr = input->data<half>();
   auto out = param->Out();
-  fpga::format_fp32_ofm(out);
+
   auto float_input = new Tensor;
-  if (input->dims().size() == 2) {
-    float_input->mutable_data<float>({1, input->dims()[1]});
-  } else if (input->dims().size() == 4) {
-    float_input->mutable_data<float>(
-        {1, input->dims()[2], input->dims()[3], input->dims()[1]});
-  } else {
-    DLOG << "wrong dimension of softmax input";
+
+  PADDLE_MOBILE_ENFORCE(input->dims().size() == 4,
+                        "Softmax should have 4-order input");
+  auto dims = framework::vectorize(input->dims());
+  auto channel = dims[3];
+  if (channel == 1) {  // This input is generated by FC op, dims = [N C 1 1]
+    PADDLE_MOBILE_ENFORCE(dims[2] == 1, "Softmax input must come from FC op");
+    dims[3] = dims[1];
+    dims[1] = 1;
+  }
+  input->Resize(framework::make_ddim(dims));
+  float_input->Resize(framework::make_ddim(dims));
+
+  if (channel != 2) {  // Use CPU
+    float_input->init(typeid(float));
+    fpga::format_fp32_ofm(float_input);
+    fpga::format_fp32_ofm(out);
+
+    fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
+    args.input_layout_type = fpga::LAYOUT_HWC;
+    args.output_layout_type = fpga::LAYOUT_CHW;
+    args.input_data_type = fpga::DATA_TYPE_FP16;
+    args.output_data_type = fpga::DATA_TYPE_FP32;
+    args.image.address = input_ptr;
+    args.image.height = (uint32_t)dims[1];
+    args.image.width = (uint32_t)dims[2];
+    args.image.channels = (uint32_t)dims[3];
+    args.output.address = float_input->data<float>();
+    args.output.scale_address = float_input->scale;
+    param->SetFloatInput(float_input);
+    param->SetFpgaArgs(args);
+  } else {  // Use FPGA
+    fpga::format_fp16_ofm(out);
+    fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
+    args.input_layout_type = fpga::LAYOUT_HWC;
+    args.output_layout_type = fpga::LAYOUT_CHW;
+    args.input_data_type = fpga::DATA_TYPE_FP16;
+    args.output_data_type = fpga::DATA_TYPE_FP16;
+    args.image.address = input_ptr;
+    args.image.height = (uint32_t)input->dims()[1];
+    args.image.width = (uint32_t)input->dims()[2];
+    args.image.channels = (uint32_t)input->dims()[3];
+    args.output.address = out->data<half>();
+    args.output.scale_address = out->scale;
+    args.output.activation.activation_type = fpga::SOFTMAX;
+    param->SetFpgaArgs(args);
   }
-  fpga::format_fp32_ofm(float_input);
-  fpga::BypassArgs args = {fpga::DATA_TYPE_FP16};
-  args.input_layout_type = fpga::LAYOUT_HWC;
-  args.output_layout_type = fpga::LAYOUT_CHW;
-  args.input_data_type = fpga::DATA_TYPE_FP16;
-  args.output_data_type = fpga::DATA_TYPE_FP32;
-  args.image.address = input_ptr;
-  args.image.height =
-      (input->dims().size() == 4) ? (uint32_t)input->dims()[2] : 1;
-  args.image.width =
-      (input->dims().size() == 4) ? (uint32_t)input->dims()[3] : 1;
-  args.image.channels = (uint32_t)input->dims()[1];
-  args.output.address = float_input->data<float>();
-  args.output.scale_address = float_input->scale;
-  param->SetFloatInput(float_input);
-  param->SetFpgaArgs(args);
 
   return true;
 }
 
 template <>
 void SoftmaxKernel<FPGA, float>::Compute(const SoftmaxParam<FPGA> &param) {
-  Tensor *in_x = param.FloatInput();
-  Tensor *out = param.Out();
-
   fpga::PerformBypass(param.FpgaArgs());
-  fpga::fpga_invalidate((void *)in_x->data<float>(),  // NOLINT
-                        in_x->numel() * sizeof(float));
-  // TODO: In general case, 0 should be squeezed before softmax input  // NOLINT
-  math::SoftmaxFuntor<CPU, float>()(in_x, out);
-  fpga::fpga_flush(out->data<float>(), out->memory_size());
+
+  if (param.FpgaArgs().output.activation.activation_type != fpga::SOFTMAX) {
+    Tensor *out = param.Out();
+    Tensor *in_x = param.FloatInput();
+    fpga::fpga_invalidate(in_x->data<float>(), in_x->numel() * sizeof(float));
+    math::SoftmaxFuntor<CPU, float>()(in_x, out);
+    fpga::fpga_flush(out->data<float>(), out->memory_size());
+  }
 }
 
 }  // namespace operators

src/operators/kernel/fpga/V1/split_kernel.cpp

@@ -34,16 +34,18 @@ bool SplitKernel<FPGA, float>::Init(SplitParam<FPGA> *param) {
       fpga::fpga_malloc(image_num * sizeof(float *)));
   auto out_channels = reinterpret_cast<uint32_t *>(
       fpga::fpga_malloc(image_num * sizeof(uint32_t)));
+  DLOG << "input: " << in;
   for (int i = 0; i < image_num; i++) {
     fpga::format_fp16_ofm(outs[i]);
-    images_out[i] = outs[i]->mutable_data<float>();
+    DLOG << "output: " << outs[i];
+    images_out[i] = outs[i]->mutable_data<half>();
     scales_out[i] = outs[i]->scale;
     out_channels[i] = (uint32_t)sections[i];
   }
 
   fpga::SplitArgs arg = {0};
   arg.image_num = image_num;
-  arg.image_in = (half *)in->data<float>();
+  arg.image_in = in->data<half>();
   arg.scale_in = in->scale;
   arg.images_out = images_out;
   arg.scales_out = scales_out;

src/operators/kernel/fpga/V1/tanh_kernel.cpp

@@ -22,8 +22,10 @@ namespace operators {
 template <>
 bool TanhKernel<FPGA, float>::Init(TanhParam<FPGA> *param) {
   auto input = const_cast<Tensor *>(param->InputX());
-  auto input_ptr = input->data<float>();
+  DLOG << "input: " << input;
+  auto input_ptr = input->data<half>();
   auto float_input = new Tensor;
+
   float_input->mutable_data<float>(
       {1, input->dims()[1], input->dims()[2], input->dims()[3]});
   fpga::format_fp32_ofm(float_input);

src/operators/kernel/fpga/V1/transpose2_kernel.cpp

@@ -20,7 +20,21 @@ namespace operators {
 
 template <>
 bool Transpose2Kernel<FPGA, float>::Init(Transpose2Param<FPGA> *param) {
-  param->Out()->ShareDataWith(*param->InputX());
+  auto input = param->InputX();
+  auto output = param->Out();
+  auto axis = param->Axis();
+  auto dim = input->dims();
+  output->ShareDataWith(*input);
+
+  auto dim_v = vectorize(dim);
+
+  for (int i = 0; i < axis.size(); i++) {
+    dim_v[i] = dim[axis[i]];
+  }
+  output->Resize(framework::make_ddim(dim_v));
+
+  DLOG << "input: " << input;
+  DLOG << "output: " << output;
   return true;
 }
 

src/operators/op_param.h

@@ -1172,6 +1172,12 @@ class FeedParam : public OpParam {
  public:
   FeedParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
             const AttributeMap &attrs, const Scope &scope) {
+#ifdef PADDLE_MOBILE_FPGA
+    static int feed_num = 0;
+    auto new_name = std::string("feed") + std::to_string(feed_num++);
+    const_cast<VariableNameMap &>(inputs).at("X") = {string(new_name)};
+#endif
+
     input_x_ = InputXFrom<LoDTensor>(inputs, scope);
     out_ = OutFrom<GType>(outputs, scope);
     auto var = scope.FindVar("batch_size");
@@ -1195,6 +1201,11 @@ class FetchParam : public OpParam {
  public:
   FetchParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
              const AttributeMap &attrs, const Scope &scope) {
+#ifdef PADDLE_MOBILE_FPGA
+    static int fetch_num = 0;
+    auto new_name = std::string("fetch") + std::to_string(fetch_num++);
+    const_cast<VariableNameMap &>(outputs).at("Out") = {string(new_name)};
+#endif
     input_x_ = InputXFrom<GType>(inputs, scope);
     out_ = OutFrom(outputs, scope);
   }
@@ -1210,18 +1221,9 @@ class FetchParam : public OpParam {
   RType *input_x_;
   Tensor *out_;
 #ifdef PADDLE_MOBILE_FPGA
-
- private:
-  std::shared_ptr<RType> float_input_x_;
+ public:
   fpga::BypassArgs fpga_bypass_args;
 
- public:
-  RType *FloatInput() const {
-    return float_input_x_ == nullptr ? input_x_ : float_input_x_.get();
-  }
-  void SetFloatInput(Tensor *input) { float_input_x_.reset(input); }
-  const fpga::BypassArgs &FpgaArgs() const { return fpga_bypass_args; }
-  void SetFpgaArgs(const fpga::BypassArgs &args) { fpga_bypass_args = args; }
 #endif
 };
 

test/fpga/test_resnet50.cpp

@@ -51,8 +51,8 @@ void convert_to_chw(int16_t **data_in, int channel, int height, int width,
   }
 }
 
-void dump(std::string filename, const Tensor input_tensor) {
-  auto dataptr = input_tensor.data<float>();
+void dump(std::string filename, Tensor input_tensor) {
+  auto dataptr = reinterpret_cast<half *>(input_tensor.get_data());
   std::ofstream out(filename.c_str());
   float result = 0;
   for (int i = 0; i < input_tensor.numel(); ++i) {
@@ -61,12 +61,11 @@ void dump(std::string filename, const Tensor input_tensor) {
   }
   out.close();
 }
-void dump_stride(std::string filename, const Tensor input_tensor,
-                 const int dumpnum) {
+void dump_stride(std::string filename, Tensor input_tensor, const int dumpnum) {
   int c = (input_tensor.dims())[1];
   int h = (input_tensor.dims())[2];
   int w = (input_tensor.dims())[3];
-  auto data_ptr = input_tensor.data<float>();
+  auto data_ptr = input_tensor.get_data();
   int16_t *data_tmp = (int16_t *)malloc(c * h * w * sizeof(int16_t));
   int16_t *data_ptr_16 = (int16_t *)data_ptr;
   convert_to_chw(&data_ptr_16, c, h, w, data_tmp);
@@ -98,9 +97,9 @@ int main() {
     for (int i = 0; i < 73; i++) {
       auto tensor_ptr = paddle_mobile.FetchResult(i);
       std::string saveName = "resnet50_result_" + std::to_string(i);
-      paddle_mobile::fpga::fpga_invalidate((*tensor_ptr).data<float>(),
+      paddle_mobile::fpga::fpga_invalidate((*tensor_ptr).get_data(),
                                            tensor_ptr->numel() * sizeof(half));
-      dump_stride(saveName, (*tensor_ptr), 20);
+      // dump_stride(saveName, (*tensor_ptr), 20);
       // dump(saveName, (*tensor_ptr));
     }
 

test/fpga/test_rfcn.cpp

@@ -23,29 +23,38 @@ limitations under the License. */
 #include "fpga/V2/api.h"
 #endif
 
-// static const char *g_densebox_combine = "../models/densebox";
-static const char *g_densebox_combine = "../models/rfcn";
+void readStream(std::string filename, uint8_t *buf) {
+  std::ifstream in;
+  in.open(filename, std::ios::in);
+  if (!in.is_open()) {
+    std::cout << "open File Failed." << std::endl;
+    return;
+  }
+  int i = 0;
+  while (!in.eof()) {
+    in >> buf[i];
+    i++;
+  }
+  in.close();
+}
+
+static const char *g_rfcn_combine = "../models/rfcn";
+const std::string g_image_src_float = "../models/rfcn/data.bin";
 int main() {
   paddle_mobile::fpga::open_device();
   paddle_mobile::PaddleMobile<paddle_mobile::FPGA> paddle_mobile;
-  // paddle_mobile.SetThreadNum(4);
-  if (paddle_mobile.Load(std::string(g_densebox_combine) + "/model",
-                         std::string(g_densebox_combine) + "/params", true,
-                         false, 1, true)) {
-    // std::vector<float> input;
-    // std::vector<int64_t> dims{1, 3, 512, 1024};
-    // GetInput<float>(g_test_image_1x3x224x224_banana, &input, dims);
-
-    // auto vec_result = paddle_mobile.Predict(input, dims);
-    return 0;
-
-    Tensor input_tensor;
-    SetupTensor<float>(&input_tensor, {1, 3, 512, 1024}, static_cast<float>(0),
-                       static_cast<float>(1));
-    // readStream(g_image_src_float,
-    //           input_tensor.mutable_data<float>({1, 3, 224, 224}));
-    paddle_mobile.FeedData(input_tensor);
+
+  if (paddle_mobile.Load(std::string(g_rfcn_combine) + "/model",
+                         std::string(g_rfcn_combine) + "/params", true, false,
+                         1, true)) {
+    float img_info[3] = {768, 1536, 768.0f / 960.0f};
+    auto img = fpga::fpga_malloc(768 * 1536 * 3 * sizeof(float));
+    readStream(g_image_src_float, reinterpret_cast<uint8_t *>(img));
+    std::vector<void *> v(3, nullptr);
+    paddle_mobile.FeedData({img_info, img});
     paddle_mobile.Predict_To(-1);
+    paddle_mobile.GetResults(&v);
+    DLOG << "Computation done";
   }
 
   return 0;