optimize Paddle-TRT performance (#46684)

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1244,6 +1244,16 @@ void AnalysisPredictor::PrepareArgument() {
 // NOTE All the members in AnalysisConfig should be copied to Argument.
 void AnalysisPredictor::OptimizeInferenceProgram() {
   PrepareArgument();
+
+#ifdef PADDLE_WITH_TENSORRT
+  if (config_.tensorrt_engine_enabled()) {
+    inference::tensorrt::TensorRTEngine::predictor_id_per_thread =
+        predictor_id_;
+    VLOG(3) << "thread_local var predictor_id in TensorRTEngine is set to: "
+            << inference::tensorrt::TensorRTEngine::predictor_id_per_thread;
+  }
+#endif
+
   Analyzer().Run(&argument_);
 
   PADDLE_ENFORCE_EQ(

paddle/fluid/inference/tensorrt/engine.cc

@@ -646,9 +646,8 @@ void TensorRTEngine::GetEngineInfo() {
   LOG(INFO) << "====== engine info ======";
   std::unique_ptr<nvinfer1::IEngineInspector> infer_inspector(
       infer_engine_->createEngineInspector());
-  auto infer_context = infer_ptr<nvinfer1::IExecutionContext>(
-      infer_engine_->createExecutionContextWithoutDeviceMemory());
-  infer_inspector->setExecutionContext(infer_context.get());
+  auto infer_context = context();
+  infer_inspector->setExecutionContext(infer_context);
   LOG(INFO) << infer_inspector->getEngineInformation(
       nvinfer1::LayerInformationFormat::kONELINE);
   LOG(INFO) << "====== engine info end ======";