Implement Hinge Embedding Loss

[littlecutebird]

• Add HingeEmbeddingLoss API wrapped by MIOPEN_BETA_API, solver, kernel, driver and gtest for fp32, fp16 and bfp16

• Performance speedup compared to ROCm/pytorch:

type	Forward
float32	2.24

(Note that fp16 is ignored because using this op in fp16 will cause overflow, underflow output. Bfp16 is ignored because pytorch rocm doesn't support HingeEmbeddingLoss in bfp16)

fp32

Size	Cont	Reduction	Margin	Forward	ROCm	MIOpen	Improvement
[10 100 100 100]	TRUE	none	0.4	fwd	700242	226738	3.088331025
[10 100 100 100]	TRUE	none	0.4	bwd	463201	301511	1.536265675
[10 100 100 100]	TRUE	sum	0.4	fwd	756242	297334	2.543409095
[10 100 100 100]	TRUE	sum	0.4	bwd	458242	231413	1.98019126
[10 100 100 100]	TRUE	mean	0.4	fwd	756322	311503	2.427976617
[10 100 100 100]	TRUE	mean	0.4	bwd	519442	260587	1.993353467
[32 32 16 64 32]	TRUE	none	0.4	fwd	2262166	750561	3.013966886
[32 32 16 64 32]	TRUE	none	0.4	bwd	1457844	999503	1.458568909
[32 32 16 64 32]	TRUE	sum	0.4	fwd	2373046	948178	2.502743156
[32 32 16 64 32]	TRUE	sum	0.4	bwd	1435444	764250	1.878238796
[32 32 16 64 32]	TRUE	mean	0.4	fwd	2371846	995254	2.383156461
[32 32 16 64 32]	TRUE	mean	0.4	bwd	1637844	861903	1.900264879
[128 256 8 256]	TRUE	none	0.4	fwd	4502972	1497850	3.00629035
[128 256 8 256]	TRUE	none	0.4	bwd	2887608	1995250	1.447241198
[128 256 8 256]	TRUE	sum	0.4	fwd	4708332	1874560	2.511699812
[128 256 8 256]	TRUE	sum	0.4	bwd	2826648	1526630	1.851560627
[128 256 8 256]	TRUE	mean	0.4	fwd	4702252	1969170	2.387936034
[128 256 8 256]	TRUE	mean	0.4	bwd	3228249	1722420	1.874251925
[20 50 100 100]	TRUE	none	0.4	fwd	698082	226542	3.081468337
[20 50 100 100]	TRUE	none	0.4	bwd	464722	302063	1.538493626
[20 50 100 100]	TRUE	sum	0.4	fwd	753842	296534	2.542177288
[20 50 100 100]	TRUE	sum	0.4	bwd	455762	231058	1.972500411
[20 50 100 100]	TRUE	mean	0.4	fwd	752082	311609	2.413543896
[20 50 100 100]	TRUE	mean	0.4	bwd	516721	260427	1.98412991
[2048 3096]	TRUE	none	0.4	fwd	452961	145440	3.114418317
[2048 3096]	TRUE	none	0.4	bwd	308241	193476	1.593174347
[2048 3096]	TRUE	sum	0.4	fwd	494082	197494	2.501757015
[2048 3096]	TRUE	sum	0.4	bwd	307200	149120	2.060085837
[2048 3096]	TRUE	mean	0.4	fwd	496162	209476	2.368586377
[2048 3096]	TRUE	mean	0.4	bwd	350241	167111	2.095858441
[2048 4096]	TRUE	none	0.4	fwd	583281	190596	3.060300321
[2048 4096]	TRUE	none	0.4	bwd	391281	254329	1.538483618
[2048 4096]	TRUE	sum	0.4	fwd	634882	252676	2.512632779
[2048 4096]	TRUE	sum	0.4	bwd	385361	194507	1.981219185
[2048 4096]	TRUE	mean	0.4	fwd	632002	265120	2.383833736
[2048 4096]	TRUE	mean	0.4	bwd	438081	219751	1.99353359
[256 512 512]	TRUE	none	0.4	fwd	4498412	1498030	3.002885122
[256 512 512]	TRUE	none	0.4	bwd	2886328	1998040	1.444579688
[256 512 512]	TRUE	sum	0.4	fwd	4703132	1876320	2.50657244
[256 512 512]	TRUE	sum	0.4	bwd	2828648	1525570	1.854158118
[256 512 512]	TRUE	mean	0.4	fwd	4709052	1970400	2.389896468
[256 512 512]	TRUE	mean	0.4	bwd	3223449	1721140	1.872856944
[25 1000 1000]	TRUE	none	0.4	fwd	1691124	560765	3.015744563
[25 1000 1000]	TRUE	none	0.4	bwd	1089603	747379	1.457898871
[25 1000 1000]	TRUE	sum	0.4	fwd	1779364	713956	2.492260027
[25 1000 1000]	TRUE	sum	0.4	bwd	1071043	570561	1.877175271
[25 1000 1000]	TRUE	mean	0.4	fwd	1776005	750561	2.366236722
[25 1000 1000]	TRUE	mean	0.4	bwd	1225283	644676	1.900618295
[2525 3333]	TRUE	none	0.4	fwd	586642	191591	3.061949674
[2525 3333]	TRUE	none	0.4	bwd	395041	255467	1.546348452
[2525 3333]	TRUE	sum	0.4	fwd	639682	254417	2.514305255
[2525 3333]	TRUE	sum	0.4	bwd	389281	195840	1.987750204
[2525 3333]	TRUE	mean	0.4	fwd	637602	268391	2.375645979
[2525 3333]	TRUE	mean	0.4	bwd	443681	220925	2.008287881
[555 111 111]	TRUE	none	0.4	fwd	505681	156551	3.230135866
[555 111 111]	TRUE	none	0.4	bwd	334881	208018	1.609865492
[555 111 111]	TRUE	sum	0.4	fwd	550482	210880	2.610404021
[555 111 111]	TRUE	sum	0.4	bwd	329521	159769	2.062483961
[555 111 111]	TRUE	mean	0.4	fwd	550002	221654	2.481353822
[555 111 111]	TRUE	mean	0.4	bwd	377441	179627	2.101248699

• Fix bug in SoftMarginLoss backward in case of reduction != None: the output gradient tensor shape is set to be similar as input tensor shape. It should be set to be a scalar