Implement Gather

[cognaiger9]

• Add Gather operation with forward kernel.
• Add driver and gtest for kernel.

Average improvement over ROCm

type	bwd
float16	1.39
float	1.38
bfloat16	1.38

Detail Benchmark

float16

op_name	dtype	input size	indices size	contiguous	dim	direction	ROCm	MIOpen	MIOpen vs ROCm
Gather	float16	[512 85742]	[512 1]	contiguous	1	fwd	20144	14009	1.44
Gather	float16	[4096 9192]	[4096 1]	contiguous	1	fwd	15248	11396	1.34
Gather	float16	[4096 9192]	[4096 1]	noncontiguous	1	fwd	14576	10880	1.34
Gather	float16	[9192 18384]	[9192 1]	contiguous	1	fwd	19248	11662	1.65
Gather	float16	[9192 18384]	[9192 1]	noncontiguous	1	fwd	16224	10773	1.51
Gather	float16	[2048 2048]	[2048 1]	noncontiguous	1	fwd	12417	10472	1.19
Gather	float16	[512 1024]	[512 2]	contiguous	1	fwd	12658	11200	1.13
Gather	float16	[512 4096]	[512 4]	noncontiguous	1	fwd	14578	10969	1.33
Gather	float16	[1024 1024]	[1024 16]	contiguous	1	fwd	13314	9120	1.46
Gather	float16	[1024 1024]	[1024 16]	noncontiguous	1	fwd	15314	10632	1.44
Gather	float16	[4096 9192]	[4096 4]	contiguous	1	fwd	16178	11432	1.42
Gather	float16	[4096 9192]	[4096 4]	noncontiguous	1	fwd	15922	11094	1.44

float32

op_name	dtype	input size	indices size	contiguous	dim	direction	ROCm	MIOpen	MIOpen vs ROCm
Gather	float32	[4096 9192]	[4096 1]	contiguous	1	fwd	18160	10933	1.66
Gather	float32	[4096 9192]	[4096 1]	noncontiguous	1	fwd	15200	11467	1.33
Gather	float32	[9192 18384]	[9192 1]	contiguous	1	fwd	23376	11644	2.01
Gather	float32	[9192 18384]	[9192 1]	noncontiguous	1	fwd	14688	10844	1.35
Gather	float32	[256 512]	[256 16]	contiguous	0	fwd	12097	9672	1.25
Gather	float32	[2048 2048]	[2048 1]	noncontiguous	1	fwd	12897	10454	1.23
Gather	float32	[512 1024]	[512 2]	contiguous	1	fwd	14065	11058	1.27
Gather	float32	[512 1024]	[512 2]	noncontiguous	1	fwd	12850	10792	1.19
Gather	float32	[512 4096]	[512 4]	noncontiguous	1	fwd	12193	10151	1.20
Gather	float32	[1024 1024]	[1024 16]	contiguous	1	fwd	13234	10027	1.32
Gather	float32	[1024 1024]	[1024 16]	noncontiguous	1	fwd	15282	10720	1.43
Gather	float32	[4096 9192]	[4096 4]	contiguous	1	fwd	15171	11378	1.33
Gather	float32	[4096 9192]	[4096 4]	noncontiguous	1	fwd	16402	11094	1.48

bfloat16

op_name	dtype	input size	indices size	contiguous	dim	direction	ROCm	MIOpen	MIOpen vs ROCm
Gather	bfloat16	[1024 4096]	[1024 1]	noncontiguous	1	fwd	13152	10827	1.21
Gather	bfloat16	[4096 9192]	[4096 1]	contiguous	1	fwd	16112	11342	1.42
Gather	bfloat16	[4096 9192]	[4096 1]	noncontiguous	1	fwd	15152	11413	1.33
Gather	bfloat16	[9192 18384]	[9192 1]	contiguous	1	fwd	19184	11751	1.63
Gather	bfloat16	[9192 18384]	[9192 1]	noncontiguous	1	fwd	15808	11022	1.43
Gather	bfloat16	[18384 18384]	[18384 1]	contiguous	1	fwd	21216	15893	1.33
Gather	bfloat16	[18384 18384]	[18384 1]	noncontiguous	1	fwd	16000	11146	1.44
Gather	bfloat16	[512 4096]	[512 4]	noncontiguous	1	fwd	15506	10987	1.41
Gather	bfloat16	[1024 1024]	[1024 16]	contiguous	1	fwd	11634	9423	1.23
Gather	bfloat16	[1024 1024]	[1024 16]	noncontiguous	1	fwd	15201	10596	1.43
Gather	bfloat16	[4096 9192]	[4096 4]	contiguous	1	fwd	15154	11396	1.33
Gather	bfloat16	[4096 9192]	[4096 4]	noncontiguous	1	fwd	15634	11076	1.41