Implement GatherV2

[cognaiger9]

• Add GatherV2 operation with non-batched and batched backward kernels.
• Add driver and gtest for kernels.

Average improvement over ROCm

type	bwd
float16	4.65
float	5.59
bfloat16	5.16

Detail Benchmark

float16

op_name	dim	batch dim	dtype	param size	indices size	direction	ROCm	MIOpen	MIOpen vs ROCm
GatherV2	0	0	float16	[2 2 4 6 8]	[16 16]	bwd	570977	244854	2.33
GatherV2	1	0	float16	[2 2 4 6 8]	[16 16]	bwd	452338	252338	1.79
GatherV2	2	0	float16	[2 2 4 6 8]	[16 16]	bwd	341809	105227	3.25
GatherV2	3	0	float16	[2 2 4 6 8]	[16 16]	bwd	387715	81778	4.74
GatherV2	4	0	float16	[2 2 4 6 8]	[16 16]	bwd	417489	84355	4.95
GatherV2	0	0	float16	[2 4 8 32 64]	[8 8]	bwd	411457	146649	2.81
GatherV2	1	0	float16	[2 4 8 32 64]	[8 8]	bwd	346176	112445	3.08
GatherV2	2	0	float16	[2 4 8 32 64]	[8 8]	bwd	499682	114063	4.38
GatherV2	0	0	float16	[2 4 8 32 64]	[16 64]	bwd	3129323	1710280	1.83
GatherV2	1	0	float16	[2 4 8 32 64]	[16 64]	bwd	1854078	1284180	1.44
GatherV2	4	0	float16	[2 4 8 32 64]	[16 64]	bwd	2486584	382152	6.51
GatherV2	0	0	float16	[4 16 32 64 64]	[8 16]	bwd	9583599	5384230	1.78
GatherV2	1	0	float16	[4 16 32 64 64]	[8 16]	bwd	2583113	1443680	1.79
GatherV2	2	0	float16	[4 16 32 64 64]	[8 16]	bwd	1550325	1038899	1.49
GatherV2	3	0	float16	[4 16 32 64 64]	[8 16]	bwd	1307337	633816	2.06
GatherV2	0	0	float16	[16 16 32 64 128]	[16 32]	bwd	51681692	43223100	1.20
GatherV2	1	0	float16	[16 16 32 64 128]	[16 32]	bwd	50925790	45607700	1.12
GatherV2	1	1	float16	[2 4 8 32 64]	[2 8]	bwd	287308	45692	6.29
GatherV2	2	2	float16	[2 4 8 32 64]	[2 4]	bwd	261031	32002	8.16
GatherV2	4	2	float16	[2 4 8 32 64]	[2 4]	bwd	321297	28944	11.10
GatherV2	1	1	float16	[4 16 32 64 64]	[4 16]	bwd	2953109	247983	11.91
GatherV2	2	2	float16	[4 16 32 64 64]	[4 16]	bwd	488315	66778	7.31
GatherV2	4	2	float16	[4 16 32 64 64]	[4 16]	bwd	645411	101625	6.35

float32

op_name	dim	batch dim	dtype	param size	indices size	direction	ROCm	MIOpen	MIOpen vs ROCm
GatherV2	0	0	float32	[2 2 4 6 8]	[16 16]	bwd	412338	74560	5.53
GatherV2	1	0	float32	[2 2 4 6 8]	[16 16]	bwd	362641	81280	4.46
GatherV2	2	0	float32	[2 2 4 6 8]	[16 16]	bwd	355501	43715	8.13
GatherV2	3	0	float32	[2 2 4 6 8]	[16 16]	bwd	352977	43217	8.17
GatherV2	4	0	float32	[2 2 4 6 8]	[16 16]	bwd	441666	49848	8.86
GatherV2	0	0	float32	[2 4 8 32 64]	[8 8]	bwd	336881	79929	4.21
GatherV2	1	0	float32	[2 4 8 32 64]	[8 8]	bwd	388658	58702	6.62
GatherV2	2	0	float32	[2 4 8 32 64]	[8 8]	bwd	361425	56569	6.39
GatherV2	0	0	float32	[2 4 8 32 64]	[16 64]	bwd	2090839	905531	2.31
GatherV2	1	0	float32	[2 4 8 32 64]	[16 64]	bwd	1279412	551184	2.32
GatherV2	2	0	float32	[2 4 8 32 64]	[16 64]	bwd	1098788	657868	1.67
GatherV2	3	0	float32	[2 4 8 32 64]	[16 64]	bwd	696770	275432	2.53
GatherV2	0	0	float32	[4 16 32 64 64]	[8 16]	bwd	7004803	4010500	1.75
GatherV2	1	0	float32	[4 16 32 64 64]	[8 16]	bwd	2017621	1158890	1.74
GatherV2	2	0	float32	[4 16 32 64 64]	[8 16]	bwd	1344719	624252	2.15
GatherV2	3	0	float32	[4 16 32 64 64]	[8 16]	bwd	1471460	357056	4.12
GatherV2	0	0	float32	[16 16 32 64 128]	[16 32]	bwd	37657097	32223999	1.17
GatherV2	3	0	float32	[16 16 32 64 128]	[16 32]	bwd	15822371	11934400	1.33
GatherV2	1	1	float32	[2 2 4 6 8]	[2 4]	bwd	302031	30544	9.89
GatherV2	2	2	float32	[2 2 4 6 8]	[2 2]	bwd	266457	34598	7.70
GatherV2	4	2	float32	[2 2 4 6 8]	[2 2]	bwd	278411	32678	8.52
GatherV2	1	1	float32	[2 4 8 32 64]	[2 8]	bwd	269914	28997	9.31
GatherV2	2	2	float32	[2 4 8 32 64]	[2 4]	bwd	301791	40998	7.36
GatherV2	4	2	float32	[2 4 8 32 64]	[2 4]	bwd	245494	30153	8.14
GatherV2	1	1	float32	[4 16 32 64 64]	[4 16]	bwd	2256154	222577	10.14
GatherV2	2	2	float32	[4 16 32 64 64]	[4 16]	bwd	562390	66049	8.51
GatherV2	4	2	float32	[4 16 32 64 64]	[4 16]	bwd	912272	114960	7.94

bfloat16

op_name	dim	batch dim	dtype	param size	indices size	direction	ROCm	MIOpen	MIOpen vs ROCm
GatherV2	0	0	bfloat16	[2 2 4 6 8]	[16 16]	bwd	594514	236658	2.51
GatherV2	1	0	bfloat16	[2 2 4 6 8]	[16 16]	bwd	472834	248161	1.91
GatherV2	2	0	bfloat16	[2 2 4 6 8]	[16 16]	bwd	366961	111147	3.30
GatherV2	3	0	bfloat16	[2 2 4 6 8]	[16 16]	bwd	420721	87947	4.78
GatherV2	4	0	bfloat16	[2 2 4 6 8]	[16 16]	bwd	493618	82115	6.01
GatherV2	0	0	bfloat16	[2 4 8 32 64]	[8 8]	bwd	416145	130294	3.19
GatherV2	1	0	bfloat16	[2 4 8 32 64]	[8 8]	bwd	445201	117511	3.79
GatherV2	2	0	bfloat16	[2 4 8 32 64]	[8 8]	bwd	430690	107502	4.01
GatherV2	3	0	bfloat16	[2 4 8 32 64]	[8 8]	bwd	358578	38897	9.22
GatherV2	0	0	bfloat16	[2 4 8 32 64]	[16 64]	bwd	3190058	1470390	2.17
GatherV2	1	0	bfloat16	[2 4 8 32 64]	[16 64]	bwd	1907398	1150150	1.66
GatherV2	4	0	bfloat16	[2 4 8 32 64]	[16 64]	bwd	2475236	365103	6.78
GatherV2	0	0	bfloat16	[4 16 32 64 64]	[8 16]	bwd	9823571	5387530	1.82
GatherV2	1	0	bfloat16	[4 16 32 64 64]	[8 16]	bwd	2671833	1452310	1.84
GatherV2	2	0	bfloat16	[4 16 32 64 64]	[8 16]	bwd	1596746	1028559	1.55
GatherV2	3	0	bfloat16	[4 16 32 64 64]	[8 16]	bwd	1274308	631826	2.02
GatherV2	0	0	bfloat16	[16 16 32 64 128]	[16 32]	bwd	52973827	43275600	1.22
GatherV2	1	0	bfloat16	[16 16 32 64 128]	[16 32]	bwd	52040938	45934000	1.13
GatherV2	1	1	bfloat16	[2 2 4 6 8]	[2 4]	bwd	305423	30651	9.96
GatherV2	2	2	bfloat16	[2 2 4 6 8]	[2 2]	bwd	264632	27735	9.54
GatherV2	4	2	bfloat16	[2 2 4 6 8]	[2 2]	bwd	265897	27095	9.81
GatherV2	1	1	bfloat16	[2 4 8 32 64]	[2 8]	bwd	298350	30971	9.63
GatherV2	4	2	bfloat16	[2 4 8 32 64]	[2 4]	bwd	263545	28339	9.30
GatherV2	1	1	bfloat16	[4 16 32 64 64]	[4 16]	bwd	3029180	251201	12.06
GatherV2	2	2	bfloat16	[4 16 32 64 64]	[4 16]	bwd	510302	66333	7.69
GatherV2	4	2	bfloat16	[4 16 32 64 64]	[4 16]	bwd	739456	101572	7.28