Optimize perf of softmax_with_cross_entropy

[ZzSean]

PR types

Performance optimization

PR changes

OPs

Describe

Optimize perf of softmax_with_cross_entropy

Replace cudnn with CUDA kernel, and make the process of cross entrop into softmax

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	0.71130	0.71989	打平 (1.21%)	0.71510	打平 (0.53%)	1.01
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	8.20172	9.96242	差于 (21.47%)	8.32200	打平 (1.47%)	1.20
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	8.01143	9.27125	差于 (15.73%)	8.07147	打平 (0.75%)	1.15
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	3.99666	9.66784	差于 (1.42x)	4.20327	差于 (5.17%)	2.30

[paddle-bot-old]

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

[xingfeng01]

else 的情况会有问题吗？好像有些情况没有给 loss 赋值

[ZzSean]

label_value的有效范围是[0, size)，而loss_id的遍历范围也是[0, size)，因此不会出现loss不被赋值的情况

[xingfeng01]

建议给函数加点注释，说明一下函数功能

[ZzSean]

好的，下个PR会加一些注释。