Created by: zhiqiu
Test for inplace addto strategy
on ResNet-50 amp training.
Related PR: https://github.com/PaddlePaddle/Paddle/pull/27112
Result
The performance of ResNet50 with amp enabled and batch_size=128 on V100 single card.
- before: 958.9 img/s
- after: 1016.5 img/s, 6% imporved.