Implement LogCumSumExp

[anhskrttt]

• Added LogCumSumExp [ref] forward and backward operations and kernels.
• This implementation works when value of the cumulative operated dimension is less than the highest number of threads inside a kernel block. Otherwise, it returns invalid results.
• Added driver test and gtest for LogCumSumExp operations.
• New API is guarded by MIOPEN_BETA_API macro.

Average improvement over ROCm

type	fwd	bwd
float	13.72	15.37
float16	-	-
bfloat16	13.11	15.38

Detail Benchmark

float32 Forward

op_name	dtype	size	dim	contiguous	direction	ROCm pytorch	MIOpen HIP	Improvement
LogCumSumExp	float32	[512 64 112 112]	-1	contiguous	fwd	145809128	7904310	18.44678764
LogCumSumExp	float32	[512 64 56 56]	-1	contiguous	fwd	69732172	2374840	29.36289266
LogCumSumExp	float32	[512 128 56 56]	-1	contiguous	fwd	139499742	4745900	29.39373817
LogCumSumExp	float32	[512 128 28 28]	-1	contiguous	fwd	69528870	2347620	29.61674803
LogCumSumExp	float32	[512 256 28 28]	-1	contiguous	fwd	139033407	4694170	29.61831527
LogCumSumExp	float32	[512 256 14 14]	-1	contiguous	fwd	69378332	2329770	29.77904772
LogCumSumExp	float32	[512 512 14 14]	-1	contiguous	fwd	138751082	4658060	29.78731103
LogCumSumExp	float32	[512 512 7 7]	-1	contiguous	fwd	69328528	2316700	29.92555273
LogCumSumExp	float32	[512 1024 7 7]	-1	contiguous	fwd	138644482	4633250	29.92380769
LogCumSumExp	float32	[512 1024 100]	-1	contiguous	fwd	20109653	1117820	17.9900637
LogCumSumExp	float32	[1024 1024 7 7]	-1	contiguous	fwd	277302592	11619000	23.8663045
LogCumSumExp	float32	[1024 1024 100]	-1	contiguous	fwd	40241550	2233520	18.01709857
LogCumSumExp	float32	[64 112 112 512]	-1	contiguous	fwd	43455625	7936570	5.475365933
LogCumSumExp	float32	[64 56 56 512]	-1	contiguous	fwd	5013870	1983640	2.527610857
LogCumSumExp	float32	[128 56 56 512]	-1	contiguous	fwd	21722237	3965290	5.478095423
LogCumSumExp	float32	[128 28 28 512]	-1	contiguous	fwd	2557599	993785	2.573593886
LogCumSumExp	float32	[256 28 28 512]	-1	contiguous	fwd	5011870	1983890	2.526284219
LogCumSumExp	float32	[256 14 14 512]	-1	contiguous	fwd	1330935	498963	2.667402192
LogCumSumExp	float32	[512 14 14 512]	-1	contiguous	fwd	2558607	994034	2.573963265
LogCumSumExp	float32	[512 7 7 512]	-1	contiguous	fwd	682939	251962	2.710484121
LogCumSumExp	float32	[1024 7 7 512]	-1	contiguous	fwd	1330327	499088	2.665515901
LogCumSumExp	float32	[1024 100 512]	-1	contiguous	fwd	2564831	1014140	2.529069951
LogCumSumExp	float32	[1024 7 7 1024]	-1	contiguous	fwd	2653214	1187580	2.234134964
LogCumSumExp	float32	[1024 100 1024]	-1	contiguous	fwd	5115869	2420190	2.113829493

float32 Backward

op_name	dtype	size	dim	contiguous	direction	ROCm pytorch	MIOpen HIP	Improvement
LogCumSumExp	float32	[512 64 112 112]	-1	contiguous	bwd	341245668	15222100	22.41777862
LogCumSumExp	float32	[512 64 56 56]	-1	contiguous	bwd	154188858	3600300	42.82666944
LogCumSumExp	float32	[512 128 56 56]	-1	contiguous	bwd	309668527	7199660	43.01154874
LogCumSumExp	float32	[512 128 28 28]	-1	contiguous	bwd	146470549	3496350	41.89241609
LogCumSumExp	float32	[512 256 28 28]	-1	contiguous	bwd	292874323	6990730	41.89466951
LogCumSumExp	float32	[512 256 14 14]	-1	contiguous	bwd	142612281	3445940	41.38559609
LogCumSumExp	float32	[512 512 14 14]	-1	contiguous	bwd	286393459	6889560	41.5691944
LogCumSumExp	float32	[512 512 7 7]	-1	contiguous	bwd	140653822	3396510	41.41127864
LogCumSumExp	float32	[512 1024 7 7]	-1	contiguous	bwd	281166630	6790880	41.40356331
LogCumSumExp	float32	[512 1024 100]	-1	contiguous	bwd	47806229	2152690	22.20766994
LogCumSumExp	float32	[1024 1024 7 7]	-1	contiguous	bwd	562217927	13587700	41.37697528
LogCumSumExp	float32	[1024 1024 100]	-1	contiguous	bwd	95480131	4303720	22.1854886
LogCumSumExp	float32	[64 112 112 512]	-1	contiguous	bwd	143977019	15467800	9.308176922
LogCumSumExp	float32	[64 56 56 512]	-1	contiguous	bwd	24262716	3865060	6.277448733
LogCumSumExp	float32	[128 56 56 512]	-1	contiguous	bwd	71758202	7731350	9.281458219
LogCumSumExp	float32	[128 28 28 512]	-1	contiguous	bwd	12271517	1936850	6.335811756
LogCumSumExp	float32	[256 28 28 512]	-1	contiguous	bwd	24321708	3866090	6.291035129
LogCumSumExp	float32	[256 14 14 512]	-1	contiguous	bwd	6260454	971563	6.443693307
LogCumSumExp	float32	[512 14 14 512]	-1	contiguous	bwd	12272334	1937170	6.335186896
LogCumSumExp	float32	[512 7 7 512]	-1	contiguous	bwd	3209674	489470	6.557447852
LogCumSumExp	float32	[1024 7 7 512]	-1	contiguous	bwd	6262886	971367	6.447497187
LogCumSumExp	float32	[1024 100 512]	-1	contiguous	bwd	12416076	1975960	6.283566469
LogCumSumExp	float32	[1024 7 7 1024]	-1	contiguous	bwd	12463612	2356880	5.288182682
LogCumSumExp	float32	[1024 100 1024]	-1	contiguous	bwd	24801400	4803570	5.163118264

float16 Forward

This operation doesn't work with float16 tensors using ROCm/Pytorch.

float16 Backward

This operation doesn't work with float16 tensors using ROCm/Pytorch.

bfloat16 Forward

op_name	dtype	size	dim	contiguous	direction	ROCm pytorch	MIOpen HIP	Improvement
LogCumSumExp	bfloat16	[512 64 112 112]	-1	contiguous	fwd	155711190	7766570	20.04890061
LogCumSumExp	bfloat16	[512 64 56 56]	-1	contiguous	fwd	78634221	2369570	33.18501711
LogCumSumExp	bfloat16	[512 128 56 56]	-1	contiguous	fwd	154616908	4737550	32.63646991
LogCumSumExp	bfloat16	[512 128 28 28]	-1	contiguous	fwd	77031384	2345150	32.84710317
LogCumSumExp	bfloat16	[512 256 28 28]	-1	contiguous	fwd	155450902	4690190	33.14383895
LogCumSumExp	bfloat16	[512 256 14 14]	-1	contiguous	fwd	76914969	2329770	33.0139752
LogCumSumExp	bfloat16	[512 512 14 14]	-1	contiguous	fwd	153749491	4657050	33.01435265
LogCumSumExp	bfloat16	[512 512 7 7]	-1	contiguous	fwd	76870091	2317380	33.1711204
LogCumSumExp	bfloat16	[512 1024 7 7]	-1	contiguous	fwd	153749893	5881890	26.13953899
LogCumSumExp	bfloat16	[512 1024 100]	-1	contiguous	fwd	22237226	1104490	20.1334788
LogCumSumExp	bfloat16	[1024 1024 7 7]	-1	contiguous	fwd	307385054	9265820	33.17408001
LogCumSumExp	bfloat16	[1024 1024 100]	-1	contiguous	fwd	44433574	2205060	20.15073241
LogCumSumExp	bfloat16	[64 112 112 512]	-1	contiguous	fwd	45994588	8012800	5.740139277
LogCumSumExp	bfloat16	[64 56 56 512]	-1	contiguous	fwd	5199885	2002470	2.596735532
LogCumSumExp	bfloat16	[128 56 56 512]	-1	contiguous	fwd	23011078	4003970	5.747065537
LogCumSumExp	bfloat16	[128 28 28 512]	-1	contiguous	fwd	2649870	1003100	2.64168079
LogCumSumExp	bfloat16	[256 28 28 512]	-1	contiguous	fwd	5200894	2002679	2.596968361
LogCumSumExp	bfloat16	[256 14 14 512]	-1	contiguous	fwd	1375703	503568	2.731911083
LogCumSumExp	bfloat16	[512 14 14 512]	-1	contiguous	fwd	2650126	1002850	2.642594605
LogCumSumExp	bfloat16	[512 7 7 512]	-1	contiguous	fwd	1908003	254218	7.505381208
LogCumSumExp	bfloat16	[1024 7 7 512]	-1	contiguous	fwd	1375319	503923	2.729224505
LogCumSumExp	bfloat16	[1024 100 512]	-1	contiguous	fwd	2657150	1023539	2.596041773
LogCumSumExp	bfloat16	[1024 7 7 1024]	-1	contiguous	fwd	2742302	1202620	2.28027307
LogCumSumExp	bfloat16	[1024 100 1024]	-1	contiguous	fwd	5305517	2450040	2.165481788

bfloat16 Backward

op_name	dtype	size	dim	contiguous	direction	ROCm pytorch	MIOpen HIP	Improvement
LogCumSumExp	bfloat16	[512 64 112 112]	-1	contiguous	bwd	351717497	15334900	22.93575419
LogCumSumExp	bfloat16	[512 64 56 56]	-1	contiguous	bwd	164634038	3604780	45.67103624
LogCumSumExp	bfloat16	[512 128 56 56]	-1	contiguous	bwd	329388221	7209500	45.68808114
LogCumSumExp	bfloat16	[512 128 28 28]	-1	contiguous	bwd	159209435	3536600	45.01765396
LogCumSumExp	bfloat16	[512 256 28 28]	-1	contiguous	bwd	318429575	8015910	39.72469439
LogCumSumExp	bfloat16	[512 256 14 14]	-1	contiguous	bwd	156468733	3479140	44.97339371
LogCumSumExp	bfloat16	[512 512 14 14]	-1	contiguous	bwd	312781554	6956140	44.96481583
LogCumSumExp	bfloat16	[512 512 7 7]	-1	contiguous	bwd	155144215	3436940	45.1402163
LogCumSumExp	bfloat16	[512 1024 7 7]	-1	contiguous	bwd	310310835	6874340	45.14045494
LogCumSumExp	bfloat16	[512 1024 100]	-1	contiguous	bwd	49671856	2181040	22.7743902
LogCumSumExp	bfloat16	[1024 1024 7 7]	-1	contiguous	bwd	620280020	13746200	45.12374474
LogCumSumExp	bfloat16	[1024 1024 100]	-1	contiguous	bwd	99066678	4358260	22.7307866
LogCumSumExp	bfloat16	[64 112 112 512]	-1	contiguous	bwd	131557996	15735600	8.36053255
LogCumSumExp	bfloat16	[64 56 56 512]	-1	contiguous	bwd	20352407	3931210	5.177135538
LogCumSumExp	bfloat16	[128 56 56 512]	-1	contiguous	bwd	65836374	7861480	8.37455212
LogCumSumExp	bfloat16	[128 28 28 512]	-1	contiguous	bwd	10347722	1968140	5.257614804
LogCumSumExp	bfloat16	[256 28 28 512]	-1	contiguous	bwd	20364119	3931120	5.180233369
LogCumSumExp	bfloat16	[256 14 14 512]	-1	contiguous	bwd	5309644	987474	5.376996255
LogCumSumExp	bfloat16	[512 14 14 512]	-1	contiguous	bwd	10341083	1967850	5.25501588
LogCumSumExp	bfloat16	[512 7 7 512]	-1	contiguous	bwd	2735166	497826	5.494220872
LogCumSumExp	bfloat16	[1024 7 7 512]	-1	contiguous	bwd	5307532	987901	5.372534292
LogCumSumExp	bfloat16	[1024 100 512]	-1	contiguous	bwd	10427898	2008300	5.192400538
LogCumSumExp	bfloat16	[1024 7 7 1024]	-1	contiguous	bwd	10520905	2365660	4.447344504
LogCumSumExp	bfloat16	[1024 100 1024]	-1	contiguous	bwd	20781652	4823750	4.308194247