Impl SmoothL1Loss

[long10024070]

• Added SmoothL1Loss forward and backward.
• Added driver test and gtest for both direction of SmoothL1Loss.
• New API is guarded by MIOPEN_BETA_API macro.
• Compared to ROCm pytorch:

float16

op_name	dtype	size	contiguous	reduction	model	beta	direction	ROCm pytorch	MIOpen HIP	Improvement
SmoothL1Loss	float16	[7 4]	true	sum	ssd/ssdlite	1	fwd	14520	8173	1.78
SmoothL1Loss	float16	[27 4]	true	sum	ssd/ssdlite	1	fwd	14145	8263	1.71
SmoothL1Loss	float16	[41 4]	true	sum	ssd/ssdlite	1	fwd	14088	8497	1.66
SmoothL1Loss	float16	[62 4]	true	sum	ssd/ssdlite	1	fwd	14189	8565	1.66
SmoothL1Loss	float16	[3 4]	true	sum	ssd/ssdlite	1	fwd	13449	8140	1.65
SmoothL1Loss	float16	[20 4]	false	sum	ssd/ssdlite	1	fwd	15286	8276	1.85
SmoothL1Loss	float16	[3 4]	false	sum	ssd/ssdlite	1	fwd	13516	8134	1.66
SmoothL1Loss	float16	[34 4]	false	sum	ssd/ssdlite	1	fwd	13654	8541	1.60
SmoothL1Loss	float16	[18 4]	false	sum	ssd/ssdlite	1	fwd	13049	8199	1.59
SmoothL1Loss	float16	[22 4]	false	sum	ssd/ssdlite	1	fwd	13062	8218	1.59
SmoothL1Loss	float16	[155 4]	false	sum	ssdlite	1	bwd	12649	7866	1.61
SmoothL1Loss	float16	[163 4]	false	sum	ssd/ssdlite	1	bwd	11184	8011	1.40
SmoothL1Loss	float16	[129 4]	false	sum	ssd/ssdlite	1	bwd	10881	7839	1.39
SmoothL1Loss	float16	[98 4]	false	sum	ssdlite	1	bwd	10078	7762	1.30
SmoothL1Loss	float16	[108 4]	false	sum	ssd/ssdlite	1	bwd	10073	7789	1.29

float32

op_name	dtype	size	contiguous	reduction	model	beta	direction	ROCm pytorch	MIOpen HIP	Improvement
SmoothL1Loss	float32	[20 4]	true	sum	ssd/ssdlite	1	fwd	17193	8389	2.05
SmoothL1Loss	float32	[7 4]	true	sum	ssd/ssdlite	1	fwd	15565	8129	1.91
SmoothL1Loss	float32	[3 4]	true	sum	ssd/ssdlite	1	fwd	13710	8102	1.69
SmoothL1Loss	float32	[47 4]	true	sum	ssd/ssdlite	1	fwd	14861	8785	1.69
SmoothL1Loss	float32	[34 4]	true	sum	ssd/ssdlite	1	fwd	14504	8668	1.67
SmoothL1Loss	float32	[3 4]	false	sum	ssd/ssdlite	1	fwd	13745	8154	1.69
SmoothL1Loss	float32	[34 4]	false	sum	ssd/ssdlite	1	fwd	13998	8670	1.61
SmoothL1Loss	float32	[22 4]	false	sum	ssd/ssdlite	1	fwd	13561	8424	1.61
SmoothL1Loss	float32	[30 4]	false	sum	ssd/ssdlite	1	fwd	13558	8423	1.61
SmoothL1Loss	float32	[20 4]	false	sum	ssd/ssdlite	1	fwd	13561	8435	1.61
SmoothL1Loss	float32	[104 4]	false	sum	ssd/ssdlite	1	bwd	12889	8029	1.61
SmoothL1Loss	float32	[129 4]	false	sum	ssd/ssdlite	1	bwd	12197	8120	1.50
SmoothL1Loss	float32	[131 4]	false	sum	ssd/ssdlite	1	bwd	11667	8111	1.44
SmoothL1Loss	float32	[137 4]	false	sum	ssdlite	1	bwd	11638	8132	1.43
SmoothL1Loss	float32	[155 4]	false	sum	ssdlite	1	bwd	12251	8569	1.43

For the bfloat16 datatype, ROCm pytorch SmoothL1Loss operator doesn't support this datatype.

• Average over all cases:

type	average
float16	1.48
float32	1.63
bfloat16	-

[iq136boy]

The following clang-format issue need to be fixed:
[2024-11-19T07:52:05.153Z] 453,454c453
[2024-11-19T07:52:05.153Z] < if(out_dev->FromGPU(GetStream(), out.data()) != 0)
[2024-11-19T07:52:05.153Z] < {
[2024-11-19T07:52:05.153Z] ---
[2024-11-19T07:52:05.153Z] > if(out_dev->FromGPU(GetStream(), out.data()) != 0) {