add compare between zerocopy and analysis

paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc

@@ -134,7 +134,7 @@ TEST(Analyzer_Pyramid_DNN, profile) {
   TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
                  input_slots_all, &outputs, FLAGS_num_threads);
 
-  if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
+  if (FLAGS_num_threads == 1 && !FLAGS_test_all_data && !FLAGS_zero_copy) {
     PADDLE_ENFORCE_EQ(outputs.size(), 1UL);
     size_t size = GetSize(outputs[0]);
     PADDLE_ENFORCE_GT(size, 0);
@@ -167,6 +167,12 @@ TEST(Analyzer_Pyramid_DNN, compare) {
   SetInput(&input_slots_all);
   CompareNativeAndAnalysis(
       reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
+
+  // Compare AnalysisConfig and AnalysisConfig + ZeroCopy
+  std::vector<std::string> outputs_name;
+  outputs_name.emplace_back("cos_sim_2.tmp_0");
+  CompareAnalysisAndZeroCopy(reinterpret_cast<PaddlePredictor::Config *>(&cfg),
+                             input_slots_all, outputs_name);
 }
 
 // Compare Deterministic result

paddle/fluid/inference/tests/api/tester_helper.h

@@ -69,6 +69,7 @@ void PrintConfig(const PaddlePredictor::Config *config, bool use_analysis) {
   LOG(INFO) << analysis_config->ToNativeConfig();
 }
 
+// Compare result between two PaddleTensor
 void CompareResult(const std::vector<PaddleTensor> &outputs,
                    const std::vector<PaddleTensor> &ref_outputs) {
   EXPECT_GT(outputs.size(), 0UL);
@@ -102,6 +103,41 @@ void CompareResult(const std::vector<PaddleTensor> &outputs,
   }
 }
 
+// Compare result between a PaddleTensor and a ZeroCopyTensor
+void CompareResult(const std::vector<PaddleTensor> &outputs,
+                   const std::vector<ZeroCopyTensor> &ref_outputs) {
+  EXPECT_GT(outputs.size(), 0UL);
+  EXPECT_EQ(outputs.size(), ref_outputs.size());
+  for (size_t i = 0; i < outputs.size(); i++) {
+    auto &out = outputs[i];
+    auto &ref_out = ref_outputs[i];
+    size_t size = VecReduceToInt(out.shape);
+    EXPECT_GT(size, 0UL);
+    int ref_size = 0;  // this is the number of elements not memory size
+    PaddlePlace place;
+    switch (out.dtype) {
+      case PaddleDType::INT64: {
+        int64_t *pdata = static_cast<int64_t *>(out.data.data());
+        int64_t *pdata_ref = ref_out.data<int64_t>(&place, &ref_size);
+        EXPECT_EQ(size, ref_size);
+        for (size_t j = 0; j < size; ++j) {
+          EXPECT_EQ(pdata_ref[j], pdata[j]);
+        }
+        break;
+      }
+      case PaddleDType::FLOAT32: {
+        float *pdata = static_cast<float *>(out.data.data());
+        float *pdata_ref = ref_out.data<float>(&place, &ref_size);
+        EXPECT_EQ(size, ref_size);
+        for (size_t j = 0; j < size; ++j) {
+          CHECK_LE(std::abs(pdata_ref[j] - pdata[j]), FLAGS_accuracy);
+        }
+        break;
+      }
+    }
+  }
+}
+
 std::unique_ptr<PaddlePredictor> CreateTestPredictor(
     const PaddlePredictor::Config *config, bool use_analysis = true) {
   const auto *analysis_config =
@@ -377,6 +413,30 @@ void CompareNativeAndAnalysis(
   CompareResult(analysis_outputs, native_outputs);
 }
 
+void CompareAnalysisAndZeroCopy(
+    PaddlePredictor::Config *config,
+    const std::vector<std::vector<PaddleTensor>> &inputs,
+    const std::vector<std::string> &outputs_name) {
+  int batch_size = FLAGS_batch_size;
+  // analysis
+  std::vector<PaddleTensor> analysis_outputs;
+  auto predictor = CreateTestPredictor(config, true);
+  predictor->Run(inputs[0], &analysis_outputs, batch_size);
+  // analysis + zero_copy
+  std::vector<ZeroCopyTensor> zerocopy_outputs;
+  reinterpret_cast<AnalysisConfig *>(config)->SwitchUseFeedFetchOps(false);
+  predictor = CreateTestPredictor(config, true);
+  ConvertPaddleTensorToZeroCopyTensor(predictor.get(), inputs[0]);
+  predictor->ZeroCopyRun();
+  for (size_t i = 0; i < outputs_name.size(); i++) {
+    ZeroCopyTensor zerocopy_output =
+        *predictor->GetOutputTensor(outputs_name[i]).get();
+    zerocopy_outputs.emplace_back(zerocopy_output);
+  }
+  // compare
+  CompareResult(analysis_outputs, zerocopy_outputs);
+}
+
 template <typename T>
 std::string LoDTensorSummary(const framework::LoDTensor &tensor) {
   std::stringstream ss;