Implement CosineSimilarity

[cognaiger9]

• Add CosineSimilarity operation with forward and backward kernels.
• Add driver and gtest for kernels.
• MIOpen performs better if:
  • Number of output elements exceeds 20000

Average improvement over ROCm

type	fwd	bwd
float16	1.86	2.72
float	1.69	1.82
bfloat16	2.04	1.96

Detail Benchmark

float16

op_name	dtype	input1	input2	contiguous	direction	ROCm	MIOpen	MIOpen vs ROCm
Cosine Similarities	float16	[100 512 200]	[100 512 200]	contiguous	fwd	426253	347020	1.23
Cosine Similarities	float16	[100 512 200]	[100 512 200]	contiguous	bwd	1462341	947513	1.54
Cosine Similarities	float16	[100 1024 200]	[100 1024 200]	contiguous	fwd	798794	688528	1.16
Cosine Similarities	float16	[100 1024 200]	[100 1024 200]	contiguous	bwd	2784315	1897090	1.47
Cosine Similarities	float16	[48 512 512]	[48 512 512]	contiguous	fwd	515276	352034	1.46
Cosine Similarities	float16	[48 512 512]	[48 512 512]	contiguous	bwd	1786899	966217	1.85
Cosine Similarities	float16	[48 512 512]	[48 512 512]	noncontiguous	fwd	2821500	1338300	2.11
Cosine Similarities	float16	[48 512 512]	[48 512 512]	noncontiguous	bwd	6846559	4324440	1.58
Cosine Similarities	float16	[48 1024 512]	[48 1024 512]	noncontiguous	fwd	5630840	3726720	1.51
Cosine Similarities	float16	[48 1024 512]	[48 1024 512]	noncontiguous	bwd	14110539	10911400	1.29
Cosine Similarities	float16	[48 2048 512]	[48 2048 512]	contiguous	fwd	2693597	1422960	1.89
Cosine Similarities	float16	[48 2048 512]	[48 2048 512]	contiguous	bwd	7302252	3825210	1.91
Cosine Similarities	float16	[48 2048 512]	[48 2048 512]	noncontiguous	fwd	10481655	8927840	1.17
Cosine Similarities	float16	[48 2048 512]	[48 2048 512]	noncontiguous	bwd	33497829	27936900	1.20
Cosine Similarities	float16	[8192 512 8]	[8192 512 8]	contiguous	fwd	1855731	760218	2.44
Cosine Similarities	float16	[8192 512 8]	[8192 512 8]	contiguous	bwd	4881614	3379450	1.44
Cosine Similarities	float16	[8192 512 8]	[8192 512 8]	noncontiguous	fwd	3529784	890530	3.96
Cosine Similarities	float16	[8192 512 8]	[8192 512 8]	noncontiguous	bwd	10447912	3381180	3.09
Cosine Similarities	float16	[8192 512 16]	[8192 512 16]	contiguous	fwd	2646878	1552040	1.71
Cosine Similarities	float16	[8192 512 16]	[8192 512 16]	contiguous	bwd	8911075	4909360	1.82

float32

op_name	dtype	input1	input2	contiguous	direction	ROCm	MIOpen	MIOpen vs ROCm
Cosine Similarities	float32	[100 512 200]	[100 512 200]	contiguous	fwd	577659	370984	1.56
Cosine Similarities	float32	[100 512 200]	[100 512 200]	contiguous	bwd	1849378	1052470	1.76
Cosine Similarities	float32	[100 1024 200]	[100 1024 200]	contiguous	fwd	1136999	746626	1.52
Cosine Similarities	float32	[100 1024 200]	[100 1024 200]	contiguous	bwd	3618709	2147220	1.69
Cosine Similarities	float32	[100 2048 200]	[100 2048 200]	contiguous	fwd	2632380	1524860	1.73
Cosine Similarities	float32	[100 2048 200]	[100 2048 200]	contiguous	bwd	8504242	4356880	1.95
Cosine Similarities	float32	[48 512 512]	[48 512 512]	contiguous	fwd	837514	383944	2.18
Cosine Similarities	float32	[48 512 512]	[48 512 512]	contiguous	bwd	2348079	1042270	2.25
Cosine Similarities	float32	[48 512 512]	[48 512 512]	noncontiguous	fwd	3506695	1811310	1.94
Cosine Similarities	float32	[48 512 512]	[48 512 512]	noncontiguous	bwd	9166318	5483190	1.67
Cosine Similarities	float32	[48 1024 512]	[48 1024 512]	contiguous	fwd	1160312	738841	1.57
Cosine Similarities	float32	[48 1024 512]	[48 1024 512]	contiguous	bwd	4295953	2045450	2.10
Cosine Similarities	float32	[48 1024 512]	[48 1024 512]	noncontiguous	fwd	6700240	5319740	1.26
Cosine Similarities	float32	[48 1024 512]	[48 1024 512]	noncontiguous	bwd	19835299	14577400	1.36
Cosine Similarities	float32	[48 2048 512]	[48 2048 512]	contiguous	fwd	2119473	1474480	1.44
Cosine Similarities	float32	[48 2048 512]	[48 2048 512]	contiguous	bwd	8352629	4055320	2.06
Cosine Similarities	float32	[48 2048 512]	[48 2048 512]	noncontiguous	fwd	12393130	9287640	1.33
Cosine Similarities	float32	[48 2048 512]	[48 2048 512]	noncontiguous	bwd	45345268	31751600	1.43
Cosine Similarities	float32	[8192 512 8]	[8192 512 8]	contiguous	fwd	2172193	903117	2.41
Cosine Similarities	float32	[8192 512 8]	[8192 512 8]	contiguous	bwd	6099446	3222970	1.89

bfloat16

op_name	dtype	input1	input2	contiguous	direction	ROCm	MIOpen	MIOpen vs ROCm
Cosine Similarities	bfloat16	[100 1024 200]	[100 1024 200]	contiguous	fwd	846298	693346	1.22
Cosine Similarities	bfloat16	[100 1024 200]	[100 1024 200]	contiguous	bwd	2977722	1905840	1.56
Cosine Similarities	bfloat16	[48 512 512]	[48 512 512]	contiguous	fwd	520316	354896	1.47
Cosine Similarities	bfloat16	[48 512 512]	[48 512 512]	contiguous	bwd	1880802	1017110	1.85
Cosine Similarities	bfloat16	[48 512 512]	[48 512 512]	noncontiguous	fwd	2847387	1336790	2.13
Cosine Similarities	bfloat16	[48 512 512]	[48 512 512]	noncontiguous	bwd	6875854	4354790	1.58
Cosine Similarities	bfloat16	[48 1024 512]	[48 1024 512]	contiguous	fwd	850234	740263	1.15
Cosine Similarities	bfloat16	[48 1024 512]	[48 1024 512]	contiguous	bwd	3515447	2027460	1.73
Cosine Similarities	bfloat16	[48 1024 512]	[48 1024 512]	noncontiguous	fwd	6123076	3721010	1.65
Cosine Similarities	bfloat16	[48 1024 512]	[48 1024 512]	noncontiguous	bwd	14947030	13634400	1.10
Cosine Similarities	bfloat16	[8192 512 8]	[8192 512 8]	contiguous	fwd	1918707	745036	2.58
Cosine Similarities	bfloat16	[8192 512 8]	[8192 512 8]	contiguous	bwd	5203020	3372960	1.54
Cosine Similarities	bfloat16	[8192 512 8]	[8192 512 8]	noncontiguous	fwd	3599208	862476	4.17
Cosine Similarities	bfloat16	[8192 512 8]	[8192 512 8]	noncontiguous	bwd	11091124	3497940	3.17
Cosine Similarities	bfloat16	[8192 512 16]	[8192 512 16]	contiguous	fwd	2800925	1561570	1.79
Cosine Similarities	bfloat16	[8192 512 16]	[8192 512 16]	contiguous	bwd	9585856	4725360	2.03
Cosine Similarities	bfloat16	[8192 512 32]	[8192 512 32]	contiguous	fwd	4520643	2975230	1.52
Cosine Similarities	bfloat16	[8192 512 32]	[8192 512 32]	contiguous	bwd	18413546	8472450	2.17
Cosine Similarities	bfloat16	[4096 512 8]	[4096 512 8]	contiguous	fwd	1258264	469138	2.68
Cosine Similarities	bfloat16	[4096 512 8]	[4096 512 8]	contiguous	bwd	3423051	1197620	2.86