Implement GatherND

[cognaiger9]

• Detail of operation (Tensorflow)
• Add GatherND operation with backward kernel.
• Add driver and gtest for kernels.

Average improvement over ROCm

type	bwd
float16	4.14
float	3.74
bfloat16	5.27

Detail Benchmark

float16

op_name	dtype	param grad size	indices size	contiguous	direction	ROCm	MIOpen	MIOpen vs ROCm
GatherND	float16	[1 512 85742]	[512 2]	contiguous	bwd	3610262	1305290	2.77
GatherND	float16	[1 25088 32317]	[25088 2]	contiguous	bwd	87586133	22057100	3.97
GatherND	float16	[1 512 4096]	[512 2]	contiguous	bwd	1393790	1239260	1.12
GatherND	float16	[1 1024 4096]	[1024 2]	contiguous	bwd	2649580	2244160	1.18
GatherND	float16	[1 4096 9192]	[4096 2]	contiguous	bwd	12533864	4564960	2.75
GatherND	float16	[1 9192 18384]	[9192 2]	contiguous	bwd	24809847	7643800	3.25
GatherND	float16	[1 18384 18384]	[18384 2]	contiguous	bwd	55369070	11581500	4.78
GatherND	float16	[2 2 8]	[96 1]	contiguous	bwd	178046	57423	3.10
GatherND	float16	[10 20 3]	[288 2]	contiguous	bwd	222762	53814	4.14
GatherND	float16	[16 128 256]	[16 2]	contiguous	bwd	244585	28889	8.47
GatherND	float16	[256 256 72]	[16 3]	contiguous	bwd	286021	44445	6.44
GatherND	float16	[32 1600 64]	[96 3]	contiguous	bwd	304995	39165	7.79

float32

op_name	dtype	param grad size	indices size	contiguous	direction	ROCm	MIOpen	MIOpen vs ROCm
GatherND	float32	[1 512 85742]	[512 2]	contiguous	bwd	2564567	894720	2.87
GatherND	float32	[1 25088 32317]	[25088 2]	contiguous	bwd	60985836	18362300	3.32
GatherND	float32	[1 512 1024]	[512 2]	contiguous	bwd	849864	398033	2.14
GatherND	float32	[1 512 4096]	[512 2]	contiguous	bwd	1041142	428007	2.43
GatherND	float32	[1 1024 4096]	[1024 2]	contiguous	bwd	1900522	721967	2.63
GatherND	float32	[1 4096 9192]	[4096 2]	contiguous	bwd	7941847	2002540	3.97
GatherND	float32	[1 9192 18384]	[9192 2]	contiguous	bwd	16906965	6008330	2.81
GatherND	float32	[1 18384 18384]	[18384 2]	contiguous	bwd	36964247	12001700	3.08
GatherND	float32	[2 2 8]	[96 1]	contiguous	bwd	178383	38329	4.65
GatherND	float32	[16 16 3]	[864 1]	contiguous	bwd	376348	168642	2.23
GatherND	float32	[10 20 3]	[288 2]	contiguous	bwd	210139	41102	5.11
GatherND	float32	[16 128 256]	[16 2]	contiguous	bwd	215755	29387	7.34
GatherND	float32	[256 256 72]	[16 3]	contiguous	bwd	237562	44978	5.28
GatherND	float32	[32 1600 64]	[96 3]	contiguous	bwd	181503	39253	4.62

bfloat16

op_name	dtype	param grad size	indices size	contiguous	direction	ROCm	MIOpen	MIOpen vs ROCm
GatherND	bfloat16	[1 512 85742]	[512 2]	contiguous	bwd	3703774	1218990	3.04
GatherND	bfloat16	[1 25088 32317]	[25088 2]	contiguous	bwd	88624446	19149400	4.63
GatherND	bfloat16	[1 512 4096]	[512 2]	contiguous	bwd	1435258	758537	1.89
GatherND	bfloat16	[1 1024 4096]	[1024 2]	contiguous	bwd	2698582	806822	3.34
GatherND	bfloat16	[1 4096 9192]	[4096 2]	contiguous	bwd	12357887	1230440	10.04
GatherND	bfloat16	[1 9192 18384]	[9192 2]	contiguous	bwd	26294178	4329720	6.07
GatherND	bfloat16	[1 18384 18384]	[18384 2]	contiguous	bwd	55355429	8296139	6.67
GatherND	bfloat16	[2 2 8]	[96 1]	contiguous	bwd	176559	54667	3.23
GatherND	bfloat16	[10 20 3]	[288 2]	contiguous	bwd	228746	54063	4.23
GatherND	bfloat16	[16 128 256]	[16 2]	contiguous	bwd	236586	31502	7.51
GatherND	bfloat16	[256 256 72]	[16 3]	contiguous	bwd	266646	46951	5.68
GatherND	bfloat16	[32 1600 64]	[96 3]	contiguous	bwd	269126	38951	6.91