Implement RoIAlign

[anhskrttt]

• Add RoIAlign operation [ref] for both forward and backward.
• Add driver and gtest.
• Performance condition:
  • For RoIAlignForward, MIOpen performs better if input and output tensors are contiguous.
    • ROCm pytorch's RoIAlign doesn't support for bfloat16 yet, so this is an always winning case.
  • For RoIAlignBackward, MIOpen performs better if:
    • Input tensor is float16 type
    • (Input tensors are contiguous) or (Input tensors are noncontiguous but input_numel size is smaller)

Average improvement over ROCm

type	fwd	bwd
float	6.24	0.97
float16	6.32	1.21

Detail Benchmark

fp32_fwd

input_shape	rois_shape	output_shape	is_contiguous	spatial_scale	sampling_ratio	ROCm	MIOpen	improvement
[16 256 160 160]	[6 5]	[7 7]	TRUE	0.0625	-1	838074	11022	76.04
[6 256 200 320]	[6 5]	[7 7]	TRUE	0.0625	-1	738074	13866	53.23
[16 256 160 160]	[6 5]	[7 7]	TRUE	0.25	2	847994	16337	51.91
[6 256 200 320]	[6 5]	[7 7]	TRUE	0.25	-1	755995	33812	22.36
[16 256 160 160]	[6 5]	[28 28]	TRUE	0.0625	-1	875194	66168	13.23
[6 256 200 320]	[6 5]	[28 28]	TRUE	0.25	-1	780314	76088	10.26
[16 256 160 160]	[6 5]	[28 28]	TRUE	0.25	2	905273	112017	8.08
[6 256 200 320]	[6 5]	[28 28]	TRUE	0.0625	2	797754	111875	7.13
[16 256 160 160]	[512 5]	[7 7]	TRUE	0.0625	-1	1028473	347748	2.96
[6 256 200 320]	[512 5]	[7 7]	TRUE	0.0625	2	1098552	575622	1.91
[16 256 160 160]	[512 5]	[7 7]	TRUE	1	2	1881266	1121320	1.68
[1 1 64 64]	[512 5]	[7 7]	TRUE	0.25	-1	16960	12817	1.32
[6 1 800 1060]	[512 5]	[7 7]	TRUE	0.25	2	20480	14684	1.39
[6 256 200 320]	[512 5]	[7 7]	TRUE	0.25	-1	1757267	1468540	1.20
[1 3 96 96]	[6 5]	[28 28]	TRUE	0.25	-1	12480	10062	1.24
[6 256 200 320]	[2149 5]	[7 7]	TRUE	0.25	2	2980138	2481650	1.20
[1 1 64 64]	[512 5]	[7 7]	TRUE	0.0625	2	17760	14897	1.19

fp16_fwd

input_shape	rois_shape	output_shape	is_contiguous	spatial_scale	sampling_ratio	ROCm	MIOpen	improvement
[16 256 160 160]	[6 5]	[7 7]	contiguous	0.0625	-1	773275	10631	72.74
[6 256 200 320]	[6 5]	[7 7]	contiguous	0.0625	-1	687035	12906	53.23
[16 256 160 160]	[6 5]	[7 7]	contiguous	0.0625	2	775515	14950	51.87
[16 256 160 160]	[6 5]	[7 7]	contiguous	0.25	2	784474	15430	50.84
[6 256 200 320]	[6 5]	[7 7]	contiguous	0.0625	2	691834	14577	47.46
[16 256 160 160]	[6 5]	[7 7]	contiguous	1	2	777275	18133	42.87
[6 256 200 320]	[6 5]	[7 7]	contiguous	0.25	2	692795	16248	42.64
[16 256 160 160]	[6 5]	[7 7]	contiguous	0.25	-1	779514	21262	36.66
[6 256 200 320]	[6 5]	[7 7]	contiguous	1	2	704475	21741	32.40
[6 256 200 320]	[6 5]	[7 7]	contiguous	0.25	-1	699675	32070	21.82
[16 256 160 160]	[6 5]	[28 28]	contiguous	0.0625	-1	805754	65155	12.37
[16 256 160 160]	[6 5]	[28 28]	contiguous	0.25	-1	806394	65386	12.33

fp16_bwd

input_size	rois_size	output_size	contiguous	spatial_scale	sampling_ratio	ROCm	MIOpen	Improvement
[3 1 800 1060]	[512 5]	[28 28]	TRUE	0.25	2	3776773	1101540	3.43
[1 1 800 1201]	[512 5]	[28 28]	TRUE	0.25	2	4223650	1260380	3.35
[2 1 800 1060]	[512 5]	[28 28]	TRUE	0.25	2	2998538	1017500	2.95
[1 1 800 1201]	[6 5]	[28 28]	TRUE	0.0625	-1	1359510	475799	2.86
[3 1 800 1060]	[512 5]	[28 28]	TRUE	1	2	1239991	446023	2.78
[1 1 64 64]	[6 5]	[28 28]	FALSE	0.0625	-1	5180762	1944090	2.66
[1 1 64 64]	[6 5]	[28 28]	TRUE	0.0625	-1	4979164	1887840	2.64
[1 1 64 64]	[6 5]	[28 28]	FALSE	0.25	2	4175810	1677920	2.49
[1 1 800 1201]	[512 5]	[28 28]	TRUE	1	2	1037592	455659	2.28
[6 1 800 1060]	[2149 5]	[28 28]	TRUE	0.25	2	7006829	3170180	2.21
[6 256 200 320]	[512 5]	[28 28]	TRUE	1	2	113196296	68001300	1.66
[6 1 800 1060]	[512 5]	[28 28]	TRUE	0.25	2	1637108	993130	1.65
[1 3 96 96]	[512 5]	[28 28]	FALSE	0.25	-1	11071280	6859840	1.61
[2 1 800 1060]	[512 5]	[28 28]	TRUE	1	2	689115	427890	1.61
[1 1 64 64]	[2149 5]	[28 28]	TRUE	0.25	-1	38307082	23917300	1.60
[1 3 96 96]	[2149 5]	[28 28]	TRUE	0.0625	-1	39848670	25024400	1.59
[1 1 64 64]	[2149 5]	[28 28]	FALSE	0.0625	2	14476410	10476100	1.38
[1 3 96 96]	[512 5]	[7 7]	FALSE	1	2	1018780	743758	1.37
[6 1 800 1060]	[2149 5]	[28 28]	TRUE	0.0625	-1	4765470	3529800	1.35
[1 3 96 96]	[512 5]	[28 28]	TRUE	0.0625	-1	5826490	4406120	1.32
[1 1 64 64]	[512 5]	[28 28]	FALSE	0.25	-1	3054872	2344600	1.30

[amd-jnovotny]

@anhskrttt Do we need a changelog entry for this addition?

[anhskrttt]

Do we need a changelog entry for this addition?

@amd-jnovotny I actually couldn't find any MIOpen documentation specifying which types of additions require a changelog entry...? I noticed that similar ops, such as PReLU and GLU, have been included in the changelog here. Based on that, I think this op could also be added.

cc @long10024070 Do you see any issues with adding this op to the changelog?

[long10024070]

@anhskrttt Do we need a changelog entry for this addition?

@amd-jnovotny I think editing changelog should be totally handled from AMD side. It is not our decision.

[amd-jnovotny]

@long10024070 thanks for the feedback. @BrianHarrisonAMD Should we add an entry to the MIOpen changelog? Maybe as part of a separate PR? If you can add something, I can review it.

[BradPepersAMD]

MIOpen is moving to the new monorepo setup and all older unmerged PR's are being closed. Please re-open this as part of the new repo if these changes are still needed.