【PFCC算子性能优化】为Paddle优化adaptive_pooling_op性能

[OuyangChao]

PR types

Performance optimization

PR changes

OPs

Describe

• document: PFCC PR 244
• device: GeForce RTX 3080 (Compute Capability: 8.6)
• data_type: float32
• data_format: NCHW
• diff: (before - after) / before

Case No.	input_shape	output_shape	pooling_type	before kernel_time(ms)	after kernel_time(ms)	diff
0	[128,64,112,112]	[56,56]	AVG	5.7967	0.7395	87.24%
1	[128,512,7,7]	[1,1]	AVG	0.0341	0.0341	0.00%
2	[128,2048,7,7]	[1,1]	AVG	0.1239	0.1242	-0.24%
3	[4,2048,64,128]	[32,32]	AVG	2.1829	0.4311	80.25%
4	[128,64,224,224]	[112,112]	AVG	19.1376	3.2681	82.92%
5	[128,128,112,112]	[56,56]	AVG	11.0257	1.4704	86.66%
6	[128,256,56,56]	[28,28]	AVG	6.7391	0.7834	88.38%
7	[128,512,28,28]	[14,14]	AVG	3.5089	0.4116	88.27%
8	[128,512,14,14]	[7,7]	AVG	0.9216	0.1138	87.65%
9	[128,64,112,112]	[56,56]	MAX	5.9602	0.9065	84.79%
10	[128,512,7,7]	[1,1]	MAX	0.0895	0.0739	17.43%
11	[128,2048,7,7]	[1,1]	MAX	0.3027	0.2711	10.44%
12	[4,2048,64,128]	[32,32]	MAX	2.2829	0.4811	78.93%
13	[128,64,224,224]	[112,112]	MAX	21.6426	4.2093	80.55%
14	[128,128,112,112]	[56,56]	MAX	11.8736	1.8121	84.74%
15	[128,256,56,56]	[28,28]	MAX	6.8098	0.891	86.92%
16	[128,512,28,28]	[14,14]	MAX	3.5734	0.4537	87.30%
17	[128,512,14,14]	[7,7]	MAX	0.9262	0.1214	86.89%

TODO

• adaptive pooling forward optimization
• adaptive pooling backward optimization

[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.

[JamesLim-sy]

之前优化的时候没想到这里可以有性能提升，Good one!

[JamesLim-sy]

divmods 这参数在计算中没有采用，可以删除，或者想采用的话可以替换 Line1986-1987中的计算div 和 mod计算.

[JamesLim-sy]

divmods 参数使用的建议同下

[JamesLim-sy]

这些建议留在反向的PR中修改吧

[JamesLim-sy]

LGTM

[OuyangChao]

这些建议留在反向的PR中修改吧

好的