[MV-378] Impl GLU opeartion (forward only, contiguous)

[cognaiger9]

Checklist

• Port kernel from OpenCL to HIP
• CPU implementation of splitting input tensor and performing operation
• Implement driver (float, fp16, bfp16)
• Implement gtest and pass
• Benchmark performance compared to MoDNN

float16

op_name	dtype	size	dim	direction	ROCm pytorch	MIOpen HIP	Improvement
GLU	float16	[8 120 1]	0	fwd	6752	4071	1.66
GLU	float16	[8 1023 1]	0	fwd	7632	4888	1.56
GLU	float16	[8 1024 768]	0	fwd	28078	36194	0.78
GLU	float16	[16 1024 768]	0	fwd	51755	68229	0.76
GLU	float16	[48 8 512 512]	0	fwd	386416	537105	0.72

float32

op_name	dtype	size	dim	direction	ROCm pytorch	MIOpen HIP	Improvement
GLU	float32	[8 120 1]	0	fwd	5743	3911	1.47
GLU	float32	[8 1023 1]	0	fwd	6224	5048	1.23
GLU	float32	[8 1024 768]	0	fwd	32669	35127	0.93
GLU	float32	[16 1024 768]	0	fwd	61835	66362	0.93
GLU	float32	[48 8 512 512]	0	fwd	455882	509159	0.9

bfloat16

op_name	dtype	size	dim	direction	ROCm pytorch	MIOpen HIP	Improvement
GLU	bfloat16	[8 120 1]	0	fwd	6496	3928	1.65
GLU	bfloat16	[8 1023 1]	0	fwd	7967	4871	1.64
GLU	bfloat16	[8 1024 768]	0	fwd	29694	36105	0.82
GLU	bfloat16	[16 1024 768]	0	fwd	54795	69082	0.79
GLU	bfloat16	[48 8 512 512]	0	fwd	412382	530351	0.78

• Average over all cases:

type	average
float16	1.09
float32	1.09
bfloat16	1.14

[kyeonghwanryu]

Good work. You can change your ticket to done.
IsApplicable condition needs to be narrowed down to guarantee the better performance. Think about it while you do the next work.