implementation of broadcast sub backward by reduce

[Zjq9409]

PR types

Performance optimization

PR changes

OPs

Describe

使用reduce实现broadcast sub 反向，相比于原始性能数据如下：

Case	pytorch	paddle(优化前)	优化前相比pytorch	paddle(优化后)	优化后相比pytorch	加速比
[50, 128, 1000], [128, 1000]	0.30086	0.18073	优于 (39.93%)	0.17129	优于(43.07%)	1.06
[50, 128, 1000], [1, 128, 1000]	0.30359	0.17959	优于 (40.84%)	0.17206	优于(43.32%)	1.04
[16, 2048, 7, 7], [16, 2048]	0.09000	0.07593	优于 (15.63%)	0.06041	优于(32.88%)	1.26
[16, 2048, 16, 16], [16, 2048, 16, 16]	0.38284	0.25730	优于 (32.79%)	0.25788	优于(32.64%)	1.00
[6, 1, 80, 46080], [1]	0.20880	1.85418	差于 (7.88x)	0.11859	优于(43.2%)	15.64
[512, 896, 4, 12], [512, 896, 4, 1]	1.11503	2.82007	差于 (1.53x)	0.64902	优于(41.79%)	4.35
[512, 896, 4, 12], [512, 896, 4, 1] fp16	0.71971	2.73426	差于 (2.80x)	0.43191	优于(39.99%)	6.33
[32, 12, 128, 128], [32, 1, 1, 128] fp16	0.18400	0.45639	差于 (1.48x)	0.09958	优于(45.88%)	4.58
[32, 1, 1, 128], [1, 12, 128, 1] fp16	0.19077	0.31292	差于 (64.03%)	0.10816	优于(43.3%)	2.89

[paddle-bot-old]

✅ This PR's description meets the template requirements!
Please wait for other CI results.

[paddle-bot-old]

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