benchmark: benchmark against fp16 implementation

[VoVAllen]

No description provided.

[cutecutecat]

Conceptions

non-fp16 instance: i4i.xlarge without avx512_fp16 instruction set
fp16 instance: i7ie.xlarge with avx512_fp16 instruction set

disk: Restart docker image and clean page cache by sudo echo 3 > sudo /proc/sys/vm/drop_caches
mem: Run a second time to ensure all results have been cached

Top10 bench for fp16-cast laion-5m

nprob	epsilon	QPS on non-fp16	recall on non-fp16	QPS on fp16	recall on fp16
20	1.9	343.72	0.9454	375.12	0.9358
20	1.5	367.93	0.9434	459.34	0.9339
20	1.0	432.48	0.9318	503.48	0.9233
50	1.9	199.44	0.9728	273.95	0.9616
50	1.5	225.60	0.9708	301.19	0.9600
50	1.0	260.27	0.9579	321.94	0.9483
50	0.8	273.05	0.9455	341.62	0.9369
100	1.0	164.62	0.9667	211.23	0.9507
100	0.8	176.30	0.9540	213.27	0.9451
200	0.8	99.91	0.9581	147.84	0.9490

Top100 bench for fp16-cast laion-5m

nprob	epsilon	QPS disk/mem on non-fp16	recall on non-fp16	QPS disk/mem on fp16	recall on fp16
20	1.9	113.98 / 252.11	0.9173	147.84 / 283.80	0.9144
50	1.9	83.90 / 168.04	0.9625	105.84 / 188.92	0.9582
50	1.5	94.01 / 198.50	0.9601	154.10 / 224.01	0.9560
50	1.0	99.75 / 232.28	0.9474	156.61 / 266.03	0.9445
100	1.0	85.07 / 150.88	0.9631	138.59 / 177.93	0.9597
100	0.8	93.32 / 154.57	0.9512	157.31 / 193.96	0.9486
200	0.8	64.91 / 92.95	0.9585	107.63 / 117.77	0.9556