Optimization of elementwise CUDA kernel

[JamesLim-sy]

PR types

Performance optimization

PR changes

OPs

Describe

Design Notion:

1. Owing to that the addition is just reading 2 elements from respective space and write back the addition result back to another space, the serial operations above are quite independent, therefore the restrict words should benefit the performance optimization.
2. Using Grid-loop stride writing forms to slightly optimized performance.
3. The GPU forward performance comparison chart is below.

Test Case	paddle baseline/us	paddle current /us	pytorch /us	perf diff (With respect to pytorch)	perf diff
x.shape=[32, 128, 768] y.shape=[768]	33.857	33.693	32.228	slower 5.05% -> slower 4.55%	perf increase ↑
x.shape=[16, 2048, 7, 7] y.shape=[16, 2048, 1, 1]	18.821	18.594	18.267	slower 3.03% -> slower 1.79%	perf increase ↑
x.shape=[16, 1, 513, 513] y.shape=[1]	44.577	44.519	42.565	slower 4.73% -> slower 4.83%	perf decrease ↓

[paddle-bot-old]

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

[paddle-bot-old]

Sorry to inform you that 8e19ebe's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.