Merge pull request #1480 from zhangyang0701/develop

add new interface function for FPGA track close #1479

src/framework/executor.cpp

@@ -505,18 +505,23 @@ 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 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");
+  auto output_size = vars.size();
+
   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;
+    v->push_back(fetch_tensor);
   }
 }
 
+template <typename Device, typename T>
+framework::Tensor *Executor<Device, T>::GetTensorByName(
+    const std::string &name) {
+  auto var = program_.scope->Var(name);
+  return var->template GetMutable<LoDTensor>();
+};
+
 template <typename Device, typename T>
 std::shared_ptr<Tensor> Executor<Device, T>::FetchResult(int id) {
   auto &ops = ops_of_block_[0];

src/framework/executor.h

@@ -57,6 +57,7 @@ class Executor {
 
   void GetResults(std::vector<void *> *v);
   void GetTensorResults(std::vector<framework::Tensor *> *v);
+  framework::Tensor *GetTensorByName(const std::string &name);
 
   std::shared_ptr<Tensor> FetchResult(int id = -1);
   void Predict_From_To(int start = 0, int end = -1);

src/io/api_paddle_mobile.cc

@@ -151,15 +151,26 @@ void PaddleMobilePredictor<Device, T>::FeedPaddleTensors(
 template <typename Device, typename T>
 void PaddleMobilePredictor<Device, T>::FetchPaddleTensors(
     std::vector<PaddleTensor> *outputs) {
-  auto num = outputs->size();
-  PADDLE_MOBILE_ENFORCE(num > 0, "0 output pointers is not permitted");
-  std::vector<framework::Tensor *> tensors(num, nullptr);
+  //  auto num = outputs->size();
+  //  PADDLE_MOBILE_ENFORCE(num > 0, "0 output pointers is not permitted");
+  //  std::vector<framework::Tensor *> tensors(num, nullptr);
+  outputs->clear();
+  std::vector<framework::Tensor *> tensors;
   paddle_mobile_->GetTensorResults(&tensors);
+  auto num = tensors.size();
+  outputs->resize(num, PaddleTensor());
   for (int i = 0; i < num; i++) {
     ConvertTensors(*tensors[i], &(*outputs)[i]);
   }
 }
 
+template <typename Device, typename T>
+void PaddleMobilePredictor<Device, T>::GetPaddleTensor(const std::string &name,
+                                                       PaddleTensor *output) {
+  framework::Tensor *t = paddle_mobile_->GetTensorByName(name);
+  ConvertTensors(*t, output);
+};
+
 template <typename Device, typename T>
 void PaddleMobilePredictor<Device, T>::FeedData(
     const std::vector<void *> &inputs) {

src/io/api_paddle_mobile.h

@@ -37,6 +37,8 @@ class PaddleMobilePredictor : public PaddlePredictor {
   void Predict_From_To(int start, int end) override;
   void FeedPaddleTensors(const std::vector<PaddleTensor>& inputs) override;
   void FetchPaddleTensors(std::vector<PaddleTensor>* outputs) override;
+  void GetPaddleTensor(const std::string& name, PaddleTensor* output) override;
+
 #endif
 
   ~PaddleMobilePredictor() override;

src/io/paddle_inference_api.h

@@ -27,8 +27,6 @@ limitations under the License. */
 #include <typeindex>
 #include <vector>
 
-// #define PADDLE_MOBILE_FPGA
-
 namespace paddle_mobile {
 
 #ifdef PADDLE_MOBILE_FPGA
@@ -133,6 +131,8 @@ class PaddlePredictor {
   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;
+  virtual void GetPaddleTensor(const std::string& name,
+                               PaddleTensor* output) = 0;
 #endif
 
  protected:

src/io/paddle_mobile.cpp

@@ -249,6 +249,12 @@ void PaddleMobile<Device, T>::GetTensorResults(
   executor_->GetTensorResults(v);
 }
 
+template <typename Device, typename T>
+framework::Tensor *PaddleMobile<Device, T>::GetTensorByName(
+    const std::string &name) {
+  return executor_->GetTensorByName(name);
+};
+
 template <typename Device, typename T>
 std::shared_ptr<framework::Tensor> PaddleMobile<Device, T>::FetchResult(
     int id) {

src/io/paddle_mobile.h

@@ -95,6 +95,7 @@ class PaddleMobile {
 
   void GetResults(std::vector<void *> *v);
   void GetTensorResults(std::vector<framework::Tensor *> *v);
+  framework::Tensor *GetTensorByName(const std::string &name);
 
   std::shared_ptr<framework::Tensor> FetchResult(int id = -1);
   void Predict_From_To(int start = 0, int end = -1);

test/fpga/test_rfcn.cpp

@@ -133,39 +133,16 @@ int main() {
     readStream(g_image_src_float, reinterpret_cast<char *>(img));
 
     std::vector<void *> v(3, nullptr);
-    paddle_mobile.FeedData({img_info, img});
+    paddle_mobile.FeedData(std::vector<void *>({img_info, img}));
     paddle_mobile.Predict_To(-1);
 
-    for (int i = 55; i < 69; i++) {
+    for (int i = 65; i < 69; i++) {
       auto tensor_ptr = paddle_mobile.FetchResult(i);
       std::string saveName = "rfcn_" + std::to_string(i);
-      // if(i != 58)
       paddle_mobile::fpga::fpga_invalidate((*tensor_ptr).get_data(),
                                            tensor_ptr->numel() * sizeof(float));
-      //                                   tensor_ptr->numel() * sizeof(float));
-      if ((i == 48) || (i == 47)) {
-        dump_stride(saveName, (*tensor_ptr), 20,
-                    false);  // 20);//tensor_ptr->numel());
-      } else if (i == 55) {
-        dump_stride(saveName, (*tensor_ptr), tensor_ptr->numel(),
-                    true);  // 20);//tensor_ptr->numel());
-      } else {
-        dump_stride(saveName, (*tensor_ptr), tensor_ptr->numel(),
-                    true);  // 20);//tensor_ptr->numel());
-      }
-      /*    float result = 0;
-          std::string str = "softmax_input_data";
-          float* data =
-         static_cast<float*>(fpga::fpga_malloc(tensor_ptr->numel() *
-         sizeof(float))); str = "softmax_output_data"; auto output_ptr =
-         static_cast<half*>((*tensor_ptr).get_data()); for (int idx = 0; idx <
-         tensor_ptr->numel(); ++idx)
-          {
-              data[idx] = fpga::fp16_2_fp32(output_ptr[idx]);
+      dump_stride(saveName, (*tensor_ptr), tensor_ptr->numel(), true);
     }
-          fpga::savefile<float>(str,data, tensor_ptr->numel(), result );   */
-    }
-
     //   paddle_mobile.GetResults(&v);
     DLOG << "Computation done";
     fpga::fpga_free(img);

test/fpga/test_rfcn_api.cpp

@@ -12,6 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#ifndef PADDLE_MOBILE_FPGA
+#define PADDLE_MOBILE_FPGA
+#endif
 #include <fstream>
 #include <iostream>
 #include "../../src/io/paddle_inference_api.h"
@@ -59,14 +62,14 @@ int main() {
       CreatePaddlePredictor<PaddleMobileConfig,
                             PaddleEngineKind::kPaddleMobile>(config);
 
-  std::cout << "after loading model" << std::endl;
+  std::cout << "Finishing loading model" << std::endl;
 
   float img_info[3] = {768, 1536, 768.0f / 960.0f};
   int img_length = 768 * 1536 * 3;
   auto img = reinterpret_cast<float *>(fpga_malloc(img_length * sizeof(float)));
   readStream(g_image, reinterpret_cast<char *>(img));
 
-  std::cout << "after initializing data" << std::endl;
+  std::cout << "Finishing initializing data" << std::endl;
   /*
     predictor->FeedData({img_info, img});
     predictor->Predict_From_To(0, -1);
@@ -110,8 +113,10 @@ int main() {
   predictor->Predict_From_To(0, -1);
   std::cout << "Finishing predicting " << std::endl;
 
-  std::vector<PaddleTensor> v(3, PaddleTensor());
-  predictor->FetchPaddleTensors(&v);
+  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;
+
   auto post_nms = v[0].data.length() / sizeof(float) / 8;
   for (int num = 0; num < post_nms; num++) {
     for (int i = 0; i < 8; i++) {
@@ -131,5 +136,14 @@ int main() {
       std::cout << p[num * 4 + i] << std::endl;
     }
   }
+  std::cout << "Finish getting vector values" << std::endl;
+
+  PaddleTensor tensor;
+  predictor->GetPaddleTensor("fetch2", &tensor);
+  for (int i = 0; i < post_nms; i++) {
+    auto p = reinterpret_cast<float *>(tensor.data.data());
+    std::cout << p[+i] << std::endl;
+  }
+
   return 0;
 }