Merge pull request #14825 from NHZlX/add_benchmark_for_trt

an sample of recording benchmark to file for trt

paddle/fluid/inference/tests/api/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
+set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor benchmark)
 
 if(WITH_GPU AND TENSORRT_FOUND)
     set(INFERENCE_EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} analysis ${analysis_deps} ir_pass_manager analysis_predictor)

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

@@ -30,8 +30,10 @@
 #include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/inference/tests/api/config_printer.h"
 #include "paddle/fluid/inference/tests/test_helper.h"
+#include "paddle/fluid/inference/utils/benchmark.h"
 #include "paddle/fluid/platform/profiler.h"
 
+DEFINE_string(model_name, "", "model name");
 DEFINE_string(infer_model, "", "model path");
 DEFINE_string(infer_data, "", "data file");
 DEFINE_int32(batch_size, 1, "batch size.");
@@ -40,6 +42,8 @@ DEFINE_bool(test_all_data, false, "Test the all dataset in data file.");
 DEFINE_int32(num_threads, 1, "Running the inference program in multi-threads.");
 DEFINE_bool(use_analysis, true,
             "Running the inference program in analysis mode.");
+DEFINE_bool(record_benchmark, false,
+            "Record benchmark after profiling the model");
 
 DECLARE_bool(profile);
 DECLARE_int32(paddle_num_threads);
@@ -192,8 +196,16 @@ void TestOneThreadPrediction(
         predictor->Run(inputs[j], outputs, batch_size);
       }
     }
-    PrintTime(batch_size, num_times, 1, 0, run_timer.toc() / num_times,
-              inputs.size());
+
+    double latency = run_timer.toc() / num_times;
+    PrintTime(batch_size, num_times, 1, 0, latency, inputs.size());
+    if (FLAGS_record_benchmark) {
+      Benchmark benchmark;
+      benchmark.SetName(FLAGS_model_name);
+      benchmark.SetBatchSize(batch_size);
+      benchmark.SetLatency(latency);
+      benchmark.PersistToFile("benchmark_record.txt");
+    }
   }
 }
 

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

@@ -135,6 +135,9 @@ TEST(TensorRT_resnext50, compare) {
 
 TEST(TensorRT_resnext50, profile) {
   std::string model_dir = FLAGS_infer_model + "/resnext50";
+  // Set FLAGS_record_benchmark to true to record benchmark to file.
+  // FLAGS_record_benchmark=true;
+  FLAGS_model_name = "resnext50";
   profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
 }
 

paddle/fluid/inference/utils/benchmark.cc

@@ -30,7 +30,7 @@ std::string Benchmark::SerializeToString() const {
   ss << '\n';
 
   ss << name_ << "\t";
-  ss << batch_size_ << "\t";
+  ss << batch_size_ << "\t\t";
   ss << num_threads_ << "\t";
   ss << latency_ << "\t";
   ss << 1000.0 / latency_;