【PaddlePaddle Hackathon 4 No.36】为 Paddle 优化 tile op 在 GPU 上的计算性能

[zeroRains]

PR types

Performance optimization

PR changes

OPs

Describe

目前Paddle中的Tile算子在GPU和CPU的计算逻辑相同，没有编写对应的Cuda代码，存在一定优化空间
设计文档：https://github.com/PaddlePaddle/community/blob/master/rfcs/OPs-Perf/20230319_tile_op_optimization.md

• 开发环境

  1. 设备：Tesla V100
  2. 环境：CUDA11.2，cuDNN 8

• 优化方法

  • 使用phi::funcs::BroadcastKernel与kps::IdentityFunctor<T>()的组合方式，加速tile执行中的复制操作
    完成优化后，Paddle与优化前的Paddle的性能对比效果:

Case No.	device	repeat_times	input_shape	input_type	Paddle Perf(ms)	old Paddle Perf(ms)	diff
1	Tesla V100	[1,10,128,128]	[16L,100L,2L,2L]	float32	5.1831	10.1888	faster than 96.58%
2	Tesla V100	[1,10,128,128]	[16L,100L,2L,2L]	float16	3.5461	16.7348	faster than 372%
3	Tesla V100	[4,1,807]	[32L, 807L, 1L]	float32	0.3885	0.7381	faster than 89.99%
4	Tesla V100	[4,1,807]	[32L, 807L, 1L]	float16	0.2465	0.9850	faster than 300%

完成优化后，Paddle与Pytorch的性能对比效果如下:

Case No.	device	repeat_times	input_shape	input_type	Paddle Perf(ms)	Pytorch Perf(ms)	diff
1	Tesla V100	[1,10,128,128]	[16L,100L,2L,2L]	float32	5.1831	8.0796	faster than 55.88%
2	Tesla V100	[1,10,128,128]	[16L,100L,2L,2L]	float16	3.5461	7.7898	faster than 120%
3	Tesla V100	[4,1,807]	[32L, 807L, 1L]	float32	0.3885	0.5342	faster than 37.50%
4	Tesla V100	[4,1,807]	[32L, 807L, 1L]	float16	0.2465	0.3768	faster than 52.86%

针对四种不同case, 优化后的性能有不同程度的提升。
感谢 @AndPuQing @Asthestarsfalll 在我Debug时提供帮助。

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[paddle-bot]

很抱歉，经过我们的反复讨论，你的PR暂未达到合入标准，请阅读飞桨原生算子开发规范，你可以重新提交新的PR，我们先将此PR关闭，感谢你的贡献。
Sorry to inform you that through our discussion, your PR fails to meet the merging standard (Reference: Paddle Custom Operator Design Doc). You can also submit an new one. Thank you.

[zeroRains]

CI已过，麻烦老师reviewe一下 @JamesLim-sy

[JamesLim-sy]

LGTM