🐛 [Bug] current BN implementation results in slower performance #3628
Description
Bug Description
current BN implementation results in slower performance. We should perform BN constant folding manually and using scaleND layer.
To Reproduce
In ResNet 18, we notice that the performance gap between Torch-TensorRT and ONNX TensorRT is around 15%. In Torch-TensorRT engine graph, we did not find the CONV layer and BN layer are fused together. This could be the main cause of the performance gap.