Implement GLU

[cognaiger9]

• Add GLU operation with contiguous forward and contiguous backward kernels.
• Add driver and gtest for kernels.
• MIOpen performs better if:
  • Input and output tensors are contiguous
  • Split dimension is 0
  • Number of elements in input tensor are less than 400,000 (forward case) or less than 800,000 (backward case)

Average improvement over ROCm

type	fwd	bwd
float16	1.3	1.56
float	1.11	1.53
bfloat16	1.27	1.54

Detail Benchmark

float16

op_name	dtype	size	dim	direction	ROCm	MIOpen	Improvement	MIOpen vs ROCm
GLU	float16	[2 320 4 4 4]	0	fwd	6816	4889	1,39	1,33
GLU	float16	[2 320 4 4 4]	0	bwd	7968	5066	1,57	1,55
GLU	float16	[32 64 3 3 3]	0	fwd	6448	4889	1,32	1,26
GLU	float16	[32 64 3 3 3]	0	bwd	7776	4836	1,61	1,62
GLU	float16	[64 3 11 11]	0	fwd	6384	4747	1,34	1,32
GLU	float16	[64 3 11 11]	0	bwd	7904	4853	1,63	1,58
GLU	float16	[256 256 1 1]	0	fwd	6176	5031	1,23	1,26
GLU	float16	[256 256 1 1]	0	bwd	8048	5066	1,59	1,54
GLU	float16	[128 64 7 7]	0	fwd	7488	5777	1,30	1,16
GLU	float16	[128 64 7 7]	0	bwd	9344	6382	1,46	1,42
GLU	float16	[64 64 7 7]	0	fwd	6880	5600	1,23	1,31
GLU	float16	[64 64 7 7]	0	bwd	8608	5778	1,49	1,48
GLU	float16	[64 32 7 7]	0	fwd	6192	4995	1,24	1,30
GLU	float16	[64 32 7 7]	0	bwd	7904	5262	1,50	1,56
GLU	float16	[32 32 7 7]	0	fwd	6368	4907	1,30	1,26
GLU	float16	[32 32 7 7]	0	bwd	7936	4995	1,59	1,59

float32

op_name	dtype	size	dim	direction	ROCm	MIOpen	Improvement	MIOpen vs ROCm
GLU	float32	[2 320 4 4 4]	0	fwd	5936	5013	1,18	1,33
GLU	float32	[2 320 4 4 4]	0	bwd	7744	5013	1,54	1,55
GLU	float32	[32 64 3 3 3]	0	fwd	5408	4942	1,09	1,26
GLU	float32	[32 64 3 3 3]	0	bwd	7968	5262	1,51	1,62
GLU	float32	[64 3 11 11]	0	fwd	5376	4960	1,08	1,32
GLU	float32	[64 3 11 11]	0	bwd	7904	4889	1,62	1,58
GLU	float32	[256 256 1 1]	0	fwd	5680	5191	1,09	1,26
GLU	float32	[256 256 1 1]	0	bwd	8064	5386	1,50	1,54
GLU	float32	[128 64 7 7]	0	fwd	7056	6524	1,08	1,16
GLU	float32	[128 64 7 7]	0	bwd	10064	7182	1,40	1,42
GLU	float32	[64 64 7 7]	0	fwd	6128	5635	1,09	1,31
GLU	float32	[64 64 7 7]	0	bwd	8944	5902	1,52	1,48
GLU	float32	[64 32 7 7]	0	fwd	5856	5155	1,14	1,30
GLU	float32	[64 32 7 7]	0	bwd	8320	5351	1,55	1,56
GLU	float32	[32 32 7 7]	0	fwd	5472	4942	1,11	1,26
GLU	float32	[32 32 7 7]	0	bwd	8112	5031	1,61	1,59

bfloat16

op_name	dtype	size	dim	direction	ROCm	MIOpen	Improvement	MIOpen vs ROCm
GLU	bfloat16	[2 320 4 4 4]	0	fwd	6928	5226	1,33	1,33
GLU	bfloat16	[2 320 4 4 4]	0	bwd	7776	5013	1,55	1,55
GLU	bfloat16	[32 64 3 3 3]	0	fwd	6320	5031	1,26	1,26
GLU	bfloat16	[32 64 3 3 3]	0	bwd	7904	4889	1,62	1,62
GLU	bfloat16	[64 3 11 11]	0	fwd	6416	4871	1,32	1,32
GLU	bfloat16	[64 3 11 11]	0	bwd	7792	4942	1,58	1,58
GLU	bfloat16	[256 256 1 1]	0	fwd	6240	4960	1,26	1,26
GLU	bfloat16	[256 256 1 1]	0	bwd	8016	5191	1,54	1,54
GLU	bfloat16	[128 64 7 7]	0	fwd	7392	6365	1,16	1,16
GLU	bfloat16	[128 64 7 7]	0	bwd	9344	6560	1,42	1,42
GLU	bfloat16	[64 64 7 7]	0	fwd	7024	5369	1,31	1,31
GLU	bfloat16	[64 64 7 7]	0	bwd	8576	5778	1,48	1,48
GLU	bfloat16	[64 32 7 7]	0	fwd	6560	5031	1,30	1,30
GLU	bfloat16	[64 32 7 7]	0	bwd	8096	5191	1,56	1,56
GLU	bfloat16	[32 32 7 7]	0	fwd	6384	5049	1,26	1,26
GLU	bfloat16	[32 32 7 7]	0	bwd	7968	5013	1,59	1,59