interpolate (grad) op support fp16 on gpu

[yuanlehome]

PR types

Others

PR changes

Others

Describe

interpolate (grad) kernel support fp16 on gpu
include bilinear_interp_v2 nearest_interp_v2 trilinear_interp_v2 linear_interp_v2 bicubic_interp_v2 (grad)

performance

interp_bicubic

Config _id in Json	FP32 Perf(ms)	(grad)FP32 Perf(ms)	FP16 Perf(ms)	(grad)FP16 Perf(ms)
0	4.2525	4.8649	4.3282	15.479
1	4.2448	6.2143	4.1762	56.991
2	6.3533	7.2487	6.4865	32.324

interp_bilinear

Config _id in Json	FP32 Perf(ms)	(grad)FP32 Perf(ms)	FP16 Perf(ms)	(grad)FP16 Perf(ms)
0	16.681	38.686	14.645	98.237
1	26.222	30.251	25.933	74.770
2	0.22262	7.8993	0.2272	82.860
3	0.30256	2.4894	0.26096	147.38
4	3.1332	13.898	1.6822	209.75
5	5.4776	8.1423	4.7927	105.74
6	0.18054	20.567	0.10122	139.88

interp_nearest

Config _id in Json	FP32 Perf(ms)	(grad)FP32 Perf(ms)	FP16 Perf(ms)	(grad)FP16 Perf(ms)
0	0.16115	0.16214	0.15923	0.54301
1	0.17987	0.26403	0.25824	0.75984

interp_trilinear

Config _id in Json	FP32 Perf(ms)	(grad)FP32 Perf(ms)	FP16 Perf(ms)	(grad)FP16 Perf(ms)
0	5.7762	5.7804	5.7707	9.5857
1	0.73469	0.73664	0.73149	8.9316
2	1.9602	2.0339	1.9574	86.897

interp_linear

Config _id in Json	FP32 Perf(us)	(grad)FP32 Perf(us)	FP16 Perf(us)	(grad)FP16 Perf(us)
0	-	-	-	-
1	30.144	30.336	28.064	52.511

inference 前向使用。由于反向性能下降明显，不适用于训练，已将其添加进amp训练黑名单PR链接

[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.

[CLAassistant]

All committers have signed the CLA.

[zhangting2020]

LGTM for fp16 tests

[Aurelius84]

LGTM for dtye registeration