Created by: grygielski

Merging https://github.com/PaddlePaddle/Paddle/pull/22695 introduced small overhead in convolutions with residual connection. It resulted in performance drop on all Resnet models:

Performance change [FPS]

CLX platform:

resnet50_int8     47.97 -> 46.60
resnet50_fp32     13.28 -> 12.01
resnet101_fp32    7.29 -> 6.65
resnet101_int8    26.78 -> 25.01

SKX platform:

resnet50_int8     22.87 -> 22.46
resnet50_fp32     13.65 -> 12.91
resnet101_fp32    7.37 -> 6.99
resnet101_int8    12.20 -> 12.07

It's inevitable result of change from moving Tensor's pointer to copying it's data. I just wanted to let you know and ask what do you think about it?