Optimize perf of softmax_with_cross_entropy_bwd

[ZzSean]

PR types

Performance optimization

PR changes

OPs

Describe

Delete D2D memcpy to optimize perf of softmax_with_cross_entropy_bwd with hard label

case	pytorch	paddle 优化前	diff	paddle 优化后	diff	加速比
label(Variable)-dtype:int64,shape:[2,1024,1024,1] logits(Variable)-dtype:float32,shape:[2,1024,1024,19] axis(int):3 ignore_index(int):255 soft_label(bool):False	0.90763	0.98579	差于 (8.61%)	0.54260	优于 (67.27%)	1.82
label(Variable)-dtype:int64,shape:[8,1024,1] logits(Variable)-dtype:float32,shape:[8,1024,50257] axis(int):2 ignore_index(int):-100 soft_label(bool):False	8.12188	8.89369	差于 (9.50%)	4.74061	优于 (71.33%)	1.88
label(Variable)-dtype:int64,shape:[8,1024,1] logits(Variable)-dtype:float32,shape:[8,1024,50304] axis(int):2 ignore_index(int):-100 soft_label(bool):False	7.90362	8.88303	差于 (12.39%)	4.75493	优于 (66.22%)	1.87
label(Variable)-dtype:int64,shape:[8,1024,1] logits(Variable)-dtype:float16,shape:[8,1024,50304] axis(int):2 ignore_index(int):-100 soft_label(bool):False	3.98370	6.45549	差于 (62.05%)	4.50929	差于 (13.19%)	1.43

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.