【Hackathon 4th No.54】为 Paddle allclose、isclose 算子实现 float16 数据类型支持

[AndPuQing]

PR types

New features

PR changes

OPs

Describe

allclose		before		after
Case No.	input_shape	fp32(ms)	fp16(ms)	fp32(ms)	fp16(ms)
0	[16,10,100]	3.9588	3.923	4.0632	4.5353
1	[16, 256, 10, 10]	11.5456	9.1728	9.0451	6.6540
2	[100,100,100]	18.3031	17.8495	11.3204	10.9800

isclose		before		after
Case No.	input_shape	fp32(ms)	fp16(ms)	fp32(ms)	fp16(ms)
0	[1700971,1]	12.2054	11.5510	11.6839	12.3301
1	[17009,100]	13.7947	13.3367	12.0218	12.5707

优化思路

• 将输入 tensor 的元素加载到共享内存中，然后在共享内存上进行比较。这种方法将减少全局内存访问次数，从而提高访问效率。(但是对于小数量可能反而会增加线程同步的开销，导致性能下降)
• 输入 tensor 只读。添加 __restrict__ 关键字，编译器就可以更好地对代码进行优化

[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.

[luotao1]

close due to the following PR is merged:

• No.54：为 Paddle allclose、isclose 算子实现 float16 数据类型支持 #51168