fix bugs for FPGA track

src/framework/executor.cpp

@@ -465,11 +465,12 @@ void Executor<Device, T>::FeedData(const Tensor &t) {
 template <typename Device, typename T>
 void Executor<Device, T>::FeedData(const std::vector<void *> &v) {
   auto input_size = v.size();
-  auto vars = program_.scope->VarContain("feed");
+  int index = 0;
+  auto vars = program_.scope->VarContain("feed", &index);
   PADDLE_MOBILE_ENFORCE(input_size == vars.size(),
                         "input data number not correct");
   for (int i = 0; i < input_size; i++) {
-    auto var = vars[i];
+    auto var = program_.scope->Var("feed", i + index);
     auto feed_tensor = var->template GetMutable<LoDTensor>();
     feed_tensor->external_data = v[i];
   }
@@ -478,11 +479,12 @@ void Executor<Device, T>::FeedData(const std::vector<void *> &v) {
 template <typename Device, typename T>
 void Executor<Device, T>::FeedTensorData(const vector<framework::Tensor> &v) {
   auto input_size = v.size();
-  auto vars = program_.scope->VarContain("feed");
+  int index = 0;
+  auto vars = program_.scope->VarContain("feed", &index);
   PADDLE_MOBILE_ENFORCE(input_size == vars.size(),
                         "input data number not correct");
   for (int i = 0; i < input_size; i++) {
-    auto var = vars[i];
+    auto var = program_.scope->Var("feed", i + index);
     auto feed_tensor = var->template GetMutable<LoDTensor>();
     feed_tensor->ShareDataWith(v[i]);
   }
@@ -492,12 +494,13 @@ 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");
+  int index = 0;
+  auto vars = program_.scope->VarContain("fetch", &index);
   PADDLE_MOBILE_ENFORCE(output_size == vars.size(),
                         "output data number not correct");
 
   for (int i = 0; i < output_size; i++) {
-    auto var = vars[i];
+    auto var = program_.scope->Var("fetch", i + index);
     auto fetch_tensor = var->template GetMutable<LoDTensor>();
     (*v)[i] = fetch_tensor->template data<float>();
   }
@@ -506,11 +509,11 @@ void Executor<Device, T>::GetResults(std::vector<void *> *v) {
 template <typename Device, typename T>
 void Executor<Device, T>::GetTensorResults(
     std::vector<framework::Tensor *> *v) {
-  auto vars = program_.scope->VarContain("fetch");
+  int index = 0;
+  auto vars = program_.scope->VarContain("fetch", &index);
   auto output_size = vars.size();
-
   for (int i = 0; i < output_size; i++) {
-    auto var = vars[i];
+    auto var = program_.scope->Var("fetch", i + index);
     auto fetch_tensor = var->template GetMutable<LoDTensor>();
     v->push_back(fetch_tensor);
   }

src/framework/scope.cpp

@@ -116,18 +116,26 @@ Variable *Scope::Var(const std::string &name, const int id) {
   return Var(name + std::to_string(id));
 }
 
-std::vector<Variable *> Scope::VarContain(const std::string substring) {
+std::vector<Variable *> Scope::VarContain(const std::string substring,
+                                          int *min) {
   std::vector<Variable *> v;
+
+  int temp = 9999;
+  auto len0 = substring.length();
   for (auto pair : vars_) {
     if (pair.first.find(substring) == 0) {
       v.push_back(pair.second);
+      auto len1 = pair.first.length();
+      int index = std::stoi(pair.first.substr(len0, len1));
+      if (index < temp) {
+        temp = index;
+      }
     }
   }
+  *min = temp;
   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

@@ -85,8 +85,7 @@ 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);
+  std::vector<Variable *> VarContain(const std::string substring, int *min);
   void print_vars();
 #endif
 

src/io/api_paddle_mobile.cc

@@ -171,18 +171,6 @@ void PaddleMobilePredictor<Device, T>::GetPaddleTensor(const std::string &name,
   ConvertTensors(*t, output);
 };
 
-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);

src/io/api_paddle_mobile.h

@@ -32,8 +32,6 @@ class PaddleMobilePredictor : public PaddlePredictor {
            std::vector<PaddleTensor>* output_data,
            int batch_size = -1) override;
 #ifdef PADDLE_MOBILE_FPGA
-  void FeedData(const std::vector<void*>& inputs) override;
-  void GetResults(std::vector<void*>* outputs) override;
   void Predict_From_To(int start, int end) override;
   void FeedPaddleTensors(const std::vector<PaddleTensor>& inputs) override;
   void FetchPaddleTensors(std::vector<PaddleTensor>* outputs) override;

src/io/paddle_inference_api.h

@@ -113,6 +113,7 @@ class PaddlePredictor {
   // `inputs`. `inputs` should be available until Run returns. Caller should be
   // responsible for the output tensor's buffer, either allocated or passed from
   // outside.
+
   virtual bool Run(const std::vector<PaddleTensor>& inputs,
                    std::vector<PaddleTensor>* output_data,
                    int batch_size = -1) = 0;
@@ -126,8 +127,6 @@ class PaddlePredictor {
     std::string param_file;
   };
 #ifdef PADDLE_MOBILE_FPGA
-  virtual void FeedData(const std::vector<void*>& inputs) = 0;
-  virtual void GetResults(std::vector<void*>* outputs) = 0;
   virtual void Predict_From_To(int start, int end) = 0;
   virtual void FeedPaddleTensors(const std::vector<PaddleTensor>& inputs) = 0;
   virtual void FetchPaddleTensors(std::vector<PaddleTensor>* outputs) = 0;

test/fpga/test_rfcn_api.cpp

@@ -55,8 +55,21 @@ PaddleMobileConfig GetConfig() {
   return config;
 }
 
+PaddleMobileConfig GetConfig1() {
+  PaddleMobileConfig config;
+  config.precision = PaddleMobileConfig::FP32;
+  config.device = PaddleMobileConfig::kFPGA;
+  config.model_dir = "../models/resnet50";
+  config.thread_num = 1;
+  config.batch_size = 1;
+  config.optimize = true;
+  config.quantification = false;
+  return config;
+}
+
 int main() {
   open_device();
+
   PaddleMobileConfig config = GetConfig();
   auto predictor =
       CreatePaddlePredictor<PaddleMobileConfig,
@@ -64,46 +77,22 @@ int main() {
 
   std::cout << "Finishing loading model" << std::endl;
 
-  float img_info[3] = {768, 1536, 768.0f / 960.0f};
-  int img_length = 768 * 1536 * 3;
+  float img_info[3] = {432, 1280, 1.0f};
+  int img_length = 432 * 1280 * 3;
   auto img = reinterpret_cast<float *>(fpga_malloc(img_length * sizeof(float)));
   readStream(g_image, reinterpret_cast<char *>(img));
 
   std::cout << "Finishing initializing data" << std::endl;
-  /*
-    predictor->FeedData({img_info, img});
-    predictor->Predict_From_To(0, -1);
-    std::cout << " Finishing predicting " << std::endl;
-      std::vector<void *> v(3, nullptr);
-      predictor->GetResults(&v);
-    int post_nms = 300;
-    for (int num = 0; num < post_nms; num ++){
-      for (int i = 0; i < 8; i ++){
-        std:: cout << ((float*)(v[0]))[num * 8 + i] << std::endl;
-      }
-    }
-    for (int num = 0; num < post_nms; num ++){
-      for (int i = 0; i < 8; i ++){
-        std:: cout << ((float*)(v[1]))[num * 8 + i] << std::endl;
-      }
-    }
-    for (int num = 0; num < post_nms; num ++){
-      for (int i = 0; i < 4; i ++){
-        std:: cout << ((float*)(v[2]))[num * 4 + i] << std::endl;
-      }
-    }
-  */
-
   struct PaddleTensor t_img_info, t_img;
-  t_img_info.dtype = FLOAT32;
+  t_img.dtypeid = typeid(float);
   t_img_info.layout = LAYOUT_HWC;
   t_img_info.shape = std::vector<int>({1, 3});
   t_img_info.name = "Image information";
   t_img_info.data.Reset(img_info, 3 * sizeof(float));
 
-  t_img.dtype = FLOAT32;
+  t_img.dtypeid = typeid(float);
   t_img.layout = LAYOUT_HWC;
-  t_img.shape = std::vector<int>({1, 768, 1536, 3});
+  t_img.shape = std::vector<int>({1, 432, 1280, 3});
   t_img.name = "Image information";
   t_img.data.Reset(img, img_length * sizeof(float));
   predictor->FeedPaddleTensors({t_img_info, t_img});
@@ -116,6 +105,9 @@ int main() {
   std::vector<PaddleTensor> v;        // No need to initialize v
   predictor->FetchPaddleTensors(&v);  // Old data in v will be cleared
   std::cout << "Output number is " << v.size() << std::endl;
+  std::cout << "out[0] length " << v[0].data.length() << std::endl;
+  std::cout << "out[1] length " << v[1].data.length() << std::endl;
+  std::cout << "out[2] length " << v[2].data.length() << std::endl;
 
   auto post_nms = v[0].data.length() / sizeof(float) / 8;
   for (int num = 0; num < post_nms; num++) {
@@ -138,6 +130,8 @@ int main() {
   }
   std::cout << "Finish getting vector values" << std::endl;
 
+  ////////////////////////////////////////////////////
+
   PaddleTensor tensor;
   predictor->GetPaddleTensor("fetch2", &tensor);
   for (int i = 0; i < post_nms; i++) {
@@ -145,5 +139,36 @@ int main() {
     std::cout << p[+i] << std::endl;
   }
 
+  //////////////////////////////////////////////////////
+
+  PaddleMobileConfig config1 = GetConfig1();
+  auto predictor1 =
+      CreatePaddlePredictor<PaddleMobileConfig,
+                            PaddleEngineKind::kPaddleMobile>(config1);
+
+  std::cout << "Finishing loading model" << std::endl;
+
+  int img_length1 = 224 * 224 * 3;
+  auto img1 =
+      reinterpret_cast<float *>(fpga_malloc(img_length1 * sizeof(float)));
+
+  std::cout << "Finishing initializing data" << std::endl;
+
+  struct PaddleTensor t_img1;
+
+  t_img1.dtypeid = typeid(float);
+  t_img1.layout = LAYOUT_HWC;
+  t_img1.shape = std::vector<int>({1, 224, 224, 3});
+  t_img1.name = "Image information";
+  t_img1.data.Reset(img1, img_length1 * sizeof(float));
+  predictor1->FeedPaddleTensors({t_img1});
+  predictor1->Predict_From_To(0, -1);
+  std::cout << "Finishing predicting " << std::endl;
+
+  std::vector<PaddleTensor> v1;         // No need to initialize v
+  predictor1->FetchPaddleTensors(&v1);  // Old data in v will be cleared
+  std::cout << "Output number is " << v1.size() << std::endl;
+  std::cout << "out[0] length " << v1[0].data.length() << std::endl;
+
   return 0;
 }