fix bugs

src/fpga/V1/pe.cpp

@@ -290,14 +290,11 @@ int ComputeBasicConv(const struct ConvArgs &args) {
   reg_writeq(args.driver.deconv_param, 0xd18);
   reg_writeq(args.driver.fpga_bias_scale_len / 4, 0xd20);
   reg_writeq(args.driver.cmd, REG_CONV_CMD);
-  DLOG << "before reg poll";
   if (0 != fpga_regpoll(REG_INTERRUPT, INTERRUPT_CONV, PE_IRQ_TIMEOUT)) {
     g_fpgainfo.pe_data->pes[PE_IDX_CONV]->status = ERROR;
     ret = -EIO;
     DLOG << "Conv Wait Irq Timeout!";
   }
-  DLOG << "after reg poll";
-
   output_scale = reg_readq(REG_SCALE_PARAMETER);
   output_scale = (output_scale << 32) | (output_scale >> 32);
   fpga_copy(args.output.scale_address, &output_scale, sizeof(float) * 2);

src/framework/executor.cpp

@@ -459,7 +459,7 @@ void Executor<Device, T>::InjectVariable(const Tensor &t,
 
 template <typename Device, typename T>
 void Executor<Device, T>::FeedData(const Tensor &t) {
-  InjectVariable(t, "feed");
+  InjectVariable(t, "feed0");
 }
 
 template <typename Device, typename T>

src/framework/operator.h

@@ -80,7 +80,6 @@ class OperatorBase {
   }
 #ifdef PADDLE_MOBILE_FPGA
   void InsertTensors();
-  void ChangeNameMap(string key, std::vector<string> value);
 #endif
  protected:
   std::shared_ptr<Scope> scope_;
@@ -96,7 +95,6 @@ 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)
@@ -106,25 +104,6 @@ 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/scope.cpp

@@ -126,8 +126,6 @@ std::vector<Variable *> Scope::VarContain(const std::string substring) {
   return v;
 }
 
-void Scope::InsertVar(const std::string str, Variable *var) {}
-
 void Scope::print_vars() {
   DLOG << "====================start to print variables=================";
   for (auto pair : vars_) {

src/framework/scope.h

@@ -86,7 +86,6 @@ class Scope {
 #ifdef PADDLE_MOBILE_FPGA
   Variable *Var(const std::string &name, const int id);
   std::vector<Variable *> VarContain(const std::string substring);
-  void InsertVar(const std::string str, Variable *var);
   void print_vars();
 #endif
 

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

@@ -22,7 +22,6 @@ bool FetchKernel<FPGA, float>::Init(FetchParam<FPGA> *param) {
   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));
@@ -52,9 +51,13 @@ template <>
 void FetchKernel<FPGA, float>::Compute(const FetchParam<FPGA> &param) {
   auto input = param.InputX();
   if (input->type() == typeid(float)) {
+    auto output = param.Out();
+    output->ShareDataWith(*input);
     return;
   }
   fpga::PerformBypass(param.fpga_bypass_args);
+  fpga::fpga_invalidate(param.fpga_bypass_args.output.address,
+                        param.fpga_bypass_args.image.channels * sizeof(float));
 
   // TODO: DEalign: get rid of extra 0
 }

test/fpga/test_resnet50.cpp

@@ -61,15 +61,16 @@ void dump(std::string filename, Tensor input_tensor) {
   }
   out.close();
 }
-void dump_stride(std::string filename, Tensor input_tensor, const int dumpnum) {
+void dump_stride_half(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.get_data();
-  int16_t *data_tmp = (int16_t *)malloc(c * h * w * sizeof(int16_t));
-  int16_t *data_ptr_16 = (int16_t *)data_ptr;
+  auto *data_tmp =
+      reinterpret_cast<half *>(malloc(c * h * w * sizeof(int16_t)));
+  auto *data_ptr_16 = reinterpret_cast<half *>(data_ptr);
   convert_to_chw(&data_ptr_16, c, h, w, data_tmp);
-  // const int16_t *dataptr = input_tensor.data<int16_t>();
   std::ofstream out(filename.c_str());
   float result = 0;
   int stride = input_tensor.numel() / dumpnum;
@@ -81,6 +82,20 @@ void dump_stride(std::string filename, Tensor input_tensor, const int dumpnum) {
   out.close();
   free(data_tmp);
 }
+
+void dump_stride_float(std::string filename, Tensor input_tensor,
+                       const int dumpnum) {
+  auto data_ptr = reinterpret_cast<float *>(input_tensor.get_data());
+  std::ofstream out(filename.c_str());
+  float result = 0;
+  int stride = input_tensor.numel() / dumpnum;
+  stride = stride > 0 ? stride : 1;
+  for (int i = 0; i < input_tensor.numel(); i += stride) {
+    result = data_ptr[i];
+    out << result << std::endl;
+  }
+  out.close();
+}
 static const char *g_resnet50 = "../models/resnet50";
 const std::string g_image_src_float = "../images/image_src_float";
 int main() {
@@ -99,22 +114,19 @@ int main() {
       std::string saveName = "resnet50_result_" + std::to_string(i);
       paddle_mobile::fpga::fpga_invalidate((*tensor_ptr).get_data(),
                                            tensor_ptr->numel() * sizeof(half));
-      // dump_stride(saveName, (*tensor_ptr), 20);
+      dump_stride_half(saveName, (*tensor_ptr), 20);
       // dump(saveName, (*tensor_ptr));
     }
 
-    std::shared_ptr<Tensor> output_tensor = paddle_mobile.FetchResult(73);
-    //(*output_tensor).dump<float>("resnet50_result_73");
-    output_tensor = paddle_mobile.FetchResult(74);
-    //(*output_tensor).dump<float>("resnet50_result_74");
-    // std::shared_ptr<Tensor> output_tensor = paddle_mobile.FetchResult(74);
-
-    // output_tensor = paddle_mobile.FetchResult(74);
+    auto tensor_ptr = paddle_mobile.FetchResult(73);
+    dump_stride_float("resnet50_result_73", (*tensor_ptr), 20);
+    tensor_ptr = paddle_mobile.FetchResult(74);
+    dump_stride_float("resnet50_result_74", (*tensor_ptr), 9999);
 
     float max = 0;
-    auto data_ptr = output_tensor->data<float>();
+    auto data_ptr = tensor_ptr->data<float>();
     int maximumIdx = 0;
-    for (int i = 0; i < (*output_tensor).numel(); i++) {
+    for (int i = 0; i < (*tensor_ptr).numel(); i++) {
       if (data_ptr[i] > max) {
         maximumIdx = i;
         max = data_ptr[i];