rocm-systems/tests/workloads/no_roof_SPI/mi200/timestamps.csv

1	Index	KernelName	gpu-id	queue-id	queue-index	pid	tid	grd	wgr	lds	scr	vgpr	sgpr	fbar	sig	obj	DispatchNs	BeginNs	EndNs	CompleteNs
2	0	__amd_rocclr_fillBufferAligned.kd	0	0	0	158945	158945	33554432	256	0	0	4	32	4160	0x0	0x7f9532004280	16419337834810	16419337860341	16419338099862	16419338190962
3	1	void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]	0	0	2	158945	158945	32768	256	0	0	12	24	13888	0x0	0x7f9532023f80	16419342856499	16419342871250	16419342884850	16419342902458
4	2	void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]	0	0	5	158945	158945	4194304	256	0	0	12	24	14336	0x7f9535013380	0x7f9532023fc0	16419342906708	16419342956210	16419343048851	16419343050865
5	3	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	8	158945	158945	4194304	256	0	0	16	24	15040	0x7f9535013200	0x7f9532024000	16419343072585	16419343082611	16419343262772	16419343264920
6	4	void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]	0	0	11	158945	158945	4194304	256	0	0	16	24	15488	0x7f9535013080	0x7f9532024040	16419343278430	16419343287252	16419343466293	16419343468516
7	5	void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]	0	0	14	158945	158945	4194304	256	0	0	12	24	15936	0x7f953286df00	0x7f9532024080	16419343482226	16419343490933	16419343583734	16419343585953
8	6	void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]	0	0	17	158945	158945	4194304	256	0	0	12	24	16384	0x7f953286dd80	0x7f95320240c0	16419343601053	16419343609334	16419343701335	16419343703191
9	7	void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]	0	0	20	158945	158945	4194304	256	0	0	12	24	17088	0x7f953286dc00	0x7f9532024100	16419343729500	16419343740375	16419343832535	16419343834298
10	8	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	23	158945	158945	4194304	256	0	0	16	24	17792	0x7f953286da80	0x7f9532024140	16419343854828	16419343864376	16419344043257	16419344044953
11	9	void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]	0	0	26	158945	158945	4194304	256	0	0	16	24	18496	0x7f953286d900	0x7f9532024180	16419344063513	16419344087897	16419344267098	16419344269388
12	10	void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]	0	0	29	158945	158945	4194304	256	0	0	16	24	19200	0x7f953286d780	0x7f95320241c0	16419344287828	16419344301338	16419344395419	16419344397406
13	11	void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]	0	0	32	158945	158945	4194304	256	0	0	16	24	19904	0x7f953286d600	0x7f9532024200	16419344413635	16419344422299	16419344515899	16419344517783
14	12	void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]	0	0	35	158945	158945	4194304	256	0	0	12	24	20608	0x7f9535013480	0x7f9532024240	16419344539002	16419344548060	16419344641340	16419344643230
15	13	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	38	158945	158945	4194304	256	0	0	16	24	21312	0x7f9535013300	0x7f9532024280	16419344659290	16419344668060	16419344847421	16419344849256
16	14	void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]	0	0	41	158945	158945	4194304	256	0	0	16	24	22016	0x7f9535013180	0x7f95320242c0	16419344869275	16419344886782	16419345066303	16419345068381
17	15	void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]	0	0	44	158945	158945	4194304	256	0	0	16	24	22720	0x7f9535013000	0x7f9532024300	16419345083360	16419345091583	16419345185023	16419345186908
18	16	void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]	0	0	47	158945	158945	4194304	256	0	0	12	24	23424	0x7f953286de80	0x7f9532024340	16419345207048	16419345222624	16419345315424	16419345317375
19	17	void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]	0	0	50	158945	158945	4194304	256	0	0	12	24	24128	0x7f953286dd00	0x7f9532024380	16419345337965	16419345348384	16419345440865	16419345442793
20	18	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	53	158945	158945	4194304	256	0	0	16	24	24832	0x7f953286db80	0x7f95320243c0	16419345460802	16419345469505	16419345649026	16419345650918
21	19	void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]	0	0	56	158945	158945	4194304	256	0	0	16	24	25536	0x7f953286da00	0x7f9532024400	16419345674487	16419345684546	16419345865188	16419345866963
22	20	void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]	0	0	59	158945	158945	4194304	256	0	0	12	24	26240	0x7f953286d880	0x7f9532024440	16419345882813	16419345890948	16419345984708	16419345986611
23	21	void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]	0	0	62	158945	158945	4194304	256	0	0	12	24	26944	0x7f953286d700	0x7f9532024480	16419346007220	16419346017348	16419346110309	16419346112318
24	22	void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]	0	0	65	158945	158945	4194304	256	0	0	12	24	27648	0x7f9535013580	0x7f95320244c0	16419346133977	16419346144549	16419346236390	16419346238345
25	23	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	68	158945	158945	4194304	256	0	0	16	24	28608	0x7f9535013400	0x7f9532024500	16419346259125	16419346269830	16419346448551	16419346450400
26	24	void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]	0	0	71	158945	158945	4194304	256	0	0	16	24	29312	0x7f9535013280	0x7f9532024540	16419346464220	16419346479591	16419346659112	16419346661166
27	25	void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]	0	0	74	158945	158945	4194304	256	0	0	12	24	30016	0x7f9535013100	0x7f9532024580	16419346675045	16419346683272	16419346775113	16419346776973
28	26	void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]	0	0	77	158945	158945	4194304	256	0	0	8	24	30976	0x7f953286df80	0x7f95320245c0	16419346793773	16419346802313	16419346894634	16419346896541
29	27	void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]	0	0	80	158945	158945	4194304	256	0	0	12	24	31680	0x7f953286de00	0x7f9532024600	16419346916740	16419346927434	16419347020394	16419347022348
30	28	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	83	158945	158945	4194304	256	0	0	16	24	32640	0x7f953286dc80	0x7f9532024640	16419347039607	16419347048555	16419347228236	16419347230193
31	29	void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]	0	0	86	158945	158945	4194304	256	0	0	12	24	33600	0x7f953286db00	0x7f9532024680	16419347249893	16419347260716	16419347441517	16419347442819
32	30	void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]	0	0	89	158945	158945	4194304	256	0	0	12	24	34560	0x7f953286d980	0x7f95320246c0	16419347456768	16419347464717	16419347558318	16419347559796
33	31	void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]	0	0	92	158945	158945	4194304	256	0	0	8	24	35520	0x7f953286d800	0x7f9532024700	16419347573986	16419347581998	16419347676398	16419347677903
34	32	void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]	0	0	95	158945	158945	4194304	256	0	0	12	24	36224	0x7f953286d680	0x7f9532024740	16419347699233	16419347708559	16419347802639	16419347804001
35	33	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	98	158945	158945	4194304	256	0	0	12	24	37184	0x7f9535013500	0x7f9532024780	16419347819600	16419347827279	16419348007600	16419348009046
36	34	void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]	0	0	101	158945	158945	4194304	256	0	0	12	24	38144	0x7f9535013380	0x7f95320247c0	16419348022836	16419348031601	16419348211922	16419348213412
37	35	void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]	0	0	104	158945	158945	4194304	256	0	0	12	24	39104	0x7f9535013200	0x7f9532024800	16419348227281	16419348235762	16419348329362	16419348330739
38	36	void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]	0	0	107	158945	158945	4194304	256	0	0	8	24	40320	0x7f9535013080	0x7f9532024840	16419348344959	16419348352882	16419348447763	16419348449206
39	37	void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]	0	0	110	158945	158945	4194304	256	0	0	12	24	41280	0x7f953286df00	0x7f9532024880	16419348469756	16419348480243	16419348574004	16419348575394
40	38	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	113	158945	158945	4194304	256	0	0	12	24	42240	0x7f953286dd80	0x7f95320248c0	16419348590683	16419348598804	16419348776885	16419348778279
41	39	void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]	0	0	116	158945	158945	4194304	256	0	0	12	24	43200	0x7f953286dc00	0x7f9532024900	16419348792239	16419348800245	16419348981686	16419348983005
42	40	void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]	0	0	119	158945	158945	4194304	256	0	0	8	24	44160	0x7f953286da80	0x7f9532024940	16419348997914	16419349006646	16419349101047	16419349102352
43	41	void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]	0	0	122	158945	158945	4194304	256	0	0	12	24	45376	0x7f953286d900	0x7f9532024980	16419349116792	16419349124567	16419349220248	16419349221599
44	42	void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]	0	0	125	158945	158945	4194304	256	0	0	12	24	46336	0x7f953286d780	0x7f95320249c0	16419349241209	16419349252408	16419349345048	16419349346447
45	43	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	128	158945	158945	4194304	256	0	0	12	24	47552	0x7f953286d600	0x7f9532024a00	16419349361656	16419349369529	16419349548250	16419349549672
46	44	void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]	0	0	131	158945	158945	4194304	256	0	0	12	24	48512	0x7f9535013480	0x7f9532024a40	16419349564592	16419349572570	16419349751771	16419349753208
47	45	void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]	0	0	134	158945	158945	4194304	256	0	0	8	24	49472	0x7f9535013300	0x7f9532024a80	16419349766967	16419349774971	16419349867932	16419349869275
48	46	void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]	0	0	137	158945	158945	4194304	256	0	0	12	24	50688	0x7f9535013180	0x7f9532024ac0	16419349884795	16419349893212	16419349986652	16419349988033
49	47	void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]	0	0	140	158945	158945	4194304	256	0	0	12	24	51648	0x7f9535013000	0x7f9532024b00	16419350006462	16419350014492	16419350106333	16419350107740
50	48	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	143	158945	158945	4194304	256	0	0	12	24	52864	0x7f953286de80	0x7f9532024b40	16419350124450	16419350132253	16419350312894	16419350314105
51	49	void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]	0	0	146	158945	158945	4194304	256	0	0	12	24	54080	0x7f953286dd00	0x7f9532024b80	16419350328185	16419350335774	16419350516575	16419350518011
52	50	void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]	0	0	149	158945	158945	4194304	256	0	0	8	24	55296	0x7f953286db80	0x7f9532024bc0	16419350531531	16419350539456	16419350634336	16419350635658
53	51	void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]	0	0	152	158945	158945	4194304	256	0	0	12	24	56768	0x7f953286da00	0x7f9532024c00	16419350649858	16419350657856	16419350752737	16419350754186
54	52	void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]	0	0	155	158945	158945	4194304	256	0	0	12	24	57728	0x7f953286d880	0x7f9532024c40	16419350774225	16419350784897	16419350878178	16419350879673
55	53	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	158	158945	158945	4194304	256	0	0	12	24	58944	0x7f953286d700	0x7f9532024c80	16419350893893	16419350902018	16419351082019	16419351083268
56	54	void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]	0	0	161	158945	158945	4194304	256	0	0	12	24	60160	0x7f9535013580	0x7f9532024cc0	16419351097418	16419351105859	16419351286020	16419351287504
57	55	void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]	0	0	164	158945	158945	4194304	256	0	0	8	24	61376	0x7f9535013400	0x7f9532024d00	16419351300824	16419351308740	16419351402501	16419351403861
58	56	void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]	0	0	167	158945	158945	4194304	256	0	0	12	24	62848	0x7f9535013280	0x7f9532024d40	16419351418211	16419351426341	16419351525221	16419351526609
59	57	void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]	0	0	170	158945	158945	4194304	256	0	0	12	24	63808	0x7f9535013100	0x7f9532024d80	16419351550108	16419351560742	16419351654502	16419351656026
60	58	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	173	158945	158945	4194304	256	0	0	12	24	65024	0x7f953286df80	0x7f9532024dc0	16419351672835	16419351681542	16419351860263	16419351861661
61	59	void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]	0	0	176	158945	158945	4194304	256	0	0	12	24	66240	0x7f953286de00	0x7f9532024e00	16419351876241	16419351884584	16419352063945	16419352065317
62	60	void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]	0	0	179	158945	158945	4194304	256	0	0	8	24	67456	0x7f953286dc80	0x7f9532024e40	16419352079437	16419352087785	16419352182025	16419352183354
63	61	void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]	0	0	182	158945	158945	4194304	256	0	0	12	24	68928	0x7f953286db00	0x7f9532024e80	16419352198724	16419352206666	16419352311626	16419352313021
64	62	void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]	0	0	185	158945	158945	4194304	256	0	0	12	24	70144	0x7f953286d980	0x7f9532024ec0	16419352332191	16419352344106	16419352436587	16419352437959
65	63	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	188	158945	158945	4194304	256	0	0	12	24	71616	0x7f953286d800	0x7f9532024f00	16419352452178	16419352460587	16419352642188	16419352643454
66	64	void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]	0	0	191	158945	158945	4194304	256	0	0	12	24	72832	0x7f953286d680	0x7f9532024f40	16419352657494	16419352666508	16419352847309	16419352848610
67	65	void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]	0	0	194	158945	158945	4194304	256	0	0	8	24	74048	0x7f9535013500	0x7f9532024f80	16419352863559	16419352872270	16419352965870	16419352967217
68	66	void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]	0	0	197	158945	158945	4194304	256	0	0	12	24	75776	0x7f9535013380	0x7f9532024fc0	16419352980887	16419352988590	16419353095311	16419353096554
69	67	void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]	0	0	200	158945	158945	4194304	256	0	0	12	24	76992	0x7f9535013200	0x7f9532025000	16419353115964	16419353127151	16419353220432	16419353221881
70	68	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	203	158945	158945	4194304	256	0	0	12	24	78464	0x7f9535013080	0x7f9532025040	16419353236101	16419353244112	16419353423473	16419353424897
71	69	void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]	0	0	206	158945	158945	4194304	256	0	0	12	24	79936	0x7f953286df00	0x7f9532025080	16419353439207	16419353447313	16419353626834	16419353628092
72	70	void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]	0	0	209	158945	158945	4194304	256	0	0	8	24	81408	0x7f953286dd80	0x7f95320250c0	16419353641462	16419353650034	16419353744915	16419353746280
73	71	void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]	0	0	212	158945	158945	4194304	256	0	0	12	24	83136	0x7f953286dc00	0x7f9532025100	16419353760070	16419353768115	16419353883796	16419353885057
74	72	void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]	0	0	215	158945	158945	4194304	256	0	0	12	24	84352	0x7f953286da80	0x7f9532025140	16419353918446	16419353927956	16419354021876	16419354023174
75	73	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	218	158945	158945	4194304	256	0	0	12	24	85824	0x7f953286d900	0x7f9532025180	16419354037153	16419354045077	16419354225878	16419354227289
76	74	void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]	0	0	221	158945	158945	4194304	256	0	0	12	24	87296	0x7f953286d780	0x7f95320251c0	16419354242139	16419354250358	16419354431959	16419354433285
77	75	void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]	0	0	224	158945	158945	4194304	256	0	0	12	24	88768	0x7f953286d600	0x7f9532025200	16419354446804	16419354454839	16419354550040	16419354551402
78	76	void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]	0	0	227	158945	158945	4194304	256	0	0	12	24	90752	0x7f9535013480	0x7f9532025240	16419354566052	16419354575160	16419354697240	16419354698679
79	77	void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]	0	0	230	158945	158945	4194304	256	0	0	12	24	91968	0x7f9535013300	0x7f9532025280	16419354718068	16419354730201	16419354824761	16419354826076
80	78	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	233	158945	158945	4194304	256	0	0	12	24	93440	0x7f9535013180	0x7f95320252c0	16419354841006	16419354849561	16419355029242	16419355030641
81	79	void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]	0	0	236	158945	158945	4194304	256	0	0	12	24	94912	0x7f9535013000	0x7f9532025300	16419355045141	16419355053243	16419355234204	16419355235587
82	80	void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]	0	0	239	158945	158945	4194304	256	0	0	12	24	96384	0x7f953286de80	0x7f9532025340	16419355250747	16419355259164	16419355354844	16419355356194
83	81	void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]	0	0	242	158945	158945	4194304	256	0	0	12	24	98368	0x7f953286dd00	0x7f9532025380	16419355372294	16419355379965	16419355506845	16419355508331
84	82	void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]	0	0	245	158945	158945	4194304	256	0	0	12	24	99584	0x7f953286db80	0x7f95320253c0	16419355528930	16419355539646	16419355633246	16419355634598
85	83	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	248	158945	158945	4194304	256	0	0	12	24	101312	0x7f953286da00	0x7f9532025400	16419355651168	16419355659166	16419355837087	16419355838474
86	84	void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]	0	0	251	158945	158945	4194304	256	0	0	12	24	102784	0x7f953286d880	0x7f9532025440	16419355854603	16419355863007	16419356042209	16419356043719
87	85	void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]	0	0	254	158945	158945	4194304	256	0	0	12	24	104256	0x7f953286d700	0x7f9532025480	16419356058199	16419356065729	16419356159169	16419356160496
88	86	void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]	0	0	257	158945	158945	4194304	256	0	0	12	24	106240	0x7f9535013580	0x7f95320254c0	16419356174306	16419356182369	16419356313250	16419356314663
89	87	void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]	0	0	260	158945	158945	4194304	256	0	0	12	24	107712	0x7f9535013400	0x7f9532025500	16419356335312	16419356345730	16419356439331	16419356440670
90	88	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	263	158945	158945	4194304	256	0	0	16	24	109440	0x7f9535013280	0x7f9532025540	16419356455630	16419356464291	16419356642372	16419356643796
91	89	void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]	0	0	266	158945	158945	4194304	256	0	0	16	24	111168	0x7f9535013100	0x7f9532025580	16419356657595	16419356665412	16419356843493	16419356844931
92	90	void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]	0	0	269	158945	158945	4194304	256	0	0	8	24	112896	0x7f953286df80	0x7f95320255c0	16419356859051	16419356867013	16419356962534	16419356963839
93	91	void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]	0	0	272	158945	158945	4194304	256	0	0	12	24	115136	0x7f953286de00	0x7f9532025600	16419356977128	16419356985254	16419357123975	16419357125375
94	92	void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]	0	0	275	158945	158945	4194304	256	0	0	12	24	116608	0x7f953286dc80	0x7f9532025640	16419357144965	16419357156455	16419357251496	16419357252902
95	93	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	278	158945	158945	4194304	256	0	0	16	24	118336	0x7f953286db00	0x7f9532025680	16419357267332	16419357276136	16419357456137	16419357457388
96	94	void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]	0	0	281	158945	158945	4194304	256	0	0	16	24	120064	0x7f953286d980	0x7f95320256c0	16419357471487	16419357480137	16419357659658	16419357661073
97	95	void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]	0	0	284	158945	158945	4194304	256	0	0	12	24	121792	0x7f953286d800	0x7f9532025700	16419357675013	16419357683018	16419357781099	16419357782370
98	96	void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]	0	0	287	158945	158945	4194304	256	0	0	12	24	124032	0x7f953286d680	0x7f9532025740	16419357798020	16419357806219	16419357950700	16419357952067
99	97	void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]	0	0	290	158945	158945	4194304	256	0	0	12	24	125504	0x7f9535013500	0x7f9532025780	16419357971716	16419357982700	16419358077741	16419358081174
100	98	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	293	158945	158945	4194304	256	0	0	12	24	127488	0x7f9535013380	0x7f95320257c0	16419358096483	16419358104301	16419358284462	16419358285829
101	99	void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]	0	0	296	158945	158945	4194304	256	0	0	12	24	129216	0x7f9535013200	0x7f9532025800	16419358300859	16419358308622	16419358488943	16419358490485
102	100	void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]	0	0	299	158945	158945	4194304	256	0	0	12	24	130944	0x7f9535013080	0x7f9532025840	16419358504354	16419358512463	16419358613264	16419358614712
103	101	void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]	0	0	302	158945	158945	4194304	256	0	0	12	24	133440	0x7f953286df00	0x7f9532025880	16419358628622	16419358636464	16419358791985	16419358793428
104	102	void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]	0	0	305	158945	158945	4194304	256	0	0	12	24	134912	0x7f953286dd80	0x7f95320258c0	16419358812958	16419358823985	16419358919026	16419358920555
105	103	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	308	158945	158945	4194304	256	0	0	12	24	136896	0x7f953286dc00	0x7f9532025900	16419358935105	16419358942866	16419359121587	16419359123021
106	104	void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]	0	0	311	158945	158945	4194304	256	0	0	12	24	138880	0x7f953286da80	0x7f9532025940	16419359138291	16419359145907	16419359325908	16419359327306
107	105	void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]	0	0	314	158945	158945	4194304	256	0	0	12	24	140864	0x7f953286d900	0x7f9532025980	16419359341076	16419359348788	16419359455189	16419359456674
108	106	void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]	0	0	317	158945	158945	4194304	256	0	0	12	24	143616	0x7f953286d780	0x7f95320259c0	16419359471113	16419359479189	16419359649270	16419359650779
109	107	void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]	0	0	320	158945	158945	4194304	256	0	0	12	24	145344	0x7f953286d600	0x7f9532025a00	16419359670489	16419359681270	16419359777111	16419359778446
110	108	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	323	158945	158945	4194304	256	0	0	12	24	147584	0x7f9535013480	0x7f9532025a40	16419359792816	16419359800791	16419359984312	16419359985872
111	109	void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]	0	0	326	158945	158945	4194304	256	0	0	12	24	149568	0x7f9535013300	0x7f9532025a80	16419359999502	16419360007192	16419360188313	16419360189807
112	110	void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]	0	0	329	158945	158945	4194304	256	0	0	12	24	151552	0x7f9535013180	0x7f9532025ac0	16419360203827	16419360211834	16419360323674	16419360325204
113	111	void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]	0	0	332	158945	158945	4194304	256	0	0	12	24	154304	0x7f9535013000	0x7f9532025b00	16419360339044	16419360347194	16419360527995	16419360529270
114	112	void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]	0	0	335	158945	158945	4194304	256	0	0	12	24	156288	0x7f953286de80	0x7f9532025b40	16419360550669	16419360560156	16419360661276	16419360662527
115	113	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	338	158945	158945	4194304	256	0	0	12	24	158784	0x7f953286dd00	0x7f9532025b80	16419360676517	16419360684636	16419360871517	16419360872902
116	114	void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]	0	0	341	158945	158945	4194304	256	0	0	12	24	161024	0x7f953286db80	0x7f9532025bc0	16419360886462	16419360894558	16419361079679	16419361081048
117	115	void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]	0	0	344	158945	158945	4194304	256	0	0	12	24	163264	0x7f953286da00	0x7f9532025c00	16419361095697	16419361103679	16419361226400	16419361227765
118	116	void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]	0	0	347	158945	158945	4194304	256	0	0	12	24	166528	0x7f953286d880	0x7f9532025c40	16419361241174	16419361249600	16419361456001	16419361477429
119	117	void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]	0	0	350	158945	158945	4194304	256	0	0	12	24	168512	0x7f953286d700	0x7f9532025c80	16419361491859	16419361502561	16419361614402	16419361615816
120	118	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	353	158945	158945	4194304	256	0	0	12	24	171264	0x7f9535013580	0x7f9532025cc0	16419361631126	16419361639842	16419361823043	16419361824391
121	119	void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]	0	0	356	158945	158945	4194304	256	0	0	12	24	173760	0x7f9535013400	0x7f9532025d00	16419361837851	16419361845763	16419362034884	16419362036177
122	120	void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]	0	0	359	158945	158945	4194304	256	0	0	12	24	176256	0x7f9535013280	0x7f9532025d40	16419362051046	16419362059365	16419362193925	16419362195363
123	121	void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]	0	0	362	158945	158945	4194304	256	0	0	12	24	179776	0x7f9535013100	0x7f9532025d80	16419362209553	16419362217286	16419362449447	16419362495287
124	122	void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]	0	0	365	158945	158945	4194304	256	0	0	12	24	182272	0x7f953286df80	0x7f9532025dc0	16419362509836	16419362520647	16419362647528	16419362648933
125	123	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	368	158945	158945	4194304	256	0	0	12	24	185536	0x7f953286de00	0x7f9532025e00	16419362663203	16419362671208	16419362860649	16419362862289
126	124	void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]	0	0	371	158945	158945	4194304	256	0	0	12	24	188544	0x7f953286dc80	0x7f9532025e40	16419362876058	16419362884010	16419363078731	16419363080224
127	125	void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]	0	0	374	158945	158945	4194304	256	0	0	12	24	191552	0x7f953286db00	0x7f9532025e80	16419363094693	16419363102411	16419363263052	16419363264630
128	126	void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]	0	0	377	158945	158945	4194304	256	0	0	12	24	195840	0x7f953286d980	0x7f9532025ec0	16419363278419	16419363286572	16419363570414	16419363616342
129	127	void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]	0	0	380	158945	158945	4194304	256	0	0	12	24	198592	0x7f953286d800	0x7f9532025f00	16419363632332	16419363643214	16419363792975	16419363794308
130	128	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	383	158945	158945	4194304	256	0	0	12	24	202368	0x7f953286d680	0x7f9532025f40	16419363809908	16419363817615	16419364025936	16419364071652
131	129	void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]	0	0	386	158945	158945	4194304	256	0	0	12	24	205888	0x7f9535013500	0x7f9532025f80	16419364080202	16419364090097	16419364299538	16419364345306
132	130	void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]	0	0	389	158945	158945	4194304	256	0	0	12	24	209408	0x7f9535013380	0x7f9532025fc0	16419364352616	16419364363058	16419364548980	16419364550271
133	131	void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]	0	0	392	158945	158945	4194304	256	0	0	12	24	214464	0x7f9535013200	0x7f9532026000	16419364566341	16419364574100	16419364908822	16419364949213
134	132	void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]	0	0	395	158945	158945	4194304	256	0	0	12	24	217728	0x7f9535013080	0x7f9532026040	16419364963922	16419364974582	16419365141303	16419365142818
135	133	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	398	158945	158945	4194304	256	0	0	12	24	222016	0x7f953286df00	0x7f9532026080	16419365157028	16419365164983	16419365394105	16419365434122
136	134	void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]	0	0	401	158945	158945	4194304	256	0	0	12	24	226048	0x7f953286dd80	0x7f95320260c0	16419365443672	16419365454265	16419365684346	16419365701026
137	135	void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]	0	0	404	158945	158945	4194304	256	0	0	12	24	230080	0x7f953286dc00	0x7f9532026100	16419365711086	16419365721627	16419365933148	16419365949391
138	136	void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]	0	0	407	158945	158945	4194304	256	0	0	12	24	235904	0x7f953286da80	0x7f9532026140	16419365959010	16419365967868	16419366353310	16419366395701
139	137	void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]	0	0	410	158945	158945	4194304	256	0	0	12	24	239424	0x7f953286d900	0x7f9532026180	16419366413480	16419366423711	16419366611392	16419366612836
140	138	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	413	158945	158945	4194304	256	0	0	12	24	244224	0x7f953286d780	0x7f95320261c0	16419366627766	16419366635552	16419366886274	16419366929729
141	139	void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]	0	0	416	158945	158945	4194304	256	0	0	12	24	248768	0x7f953286d600	0x7f9532026200	16419366937449	16419366948034	16419367196035	16419367239992
142	140	void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]	0	0	419	158945	158945	4194304	256	0	0	12	24	253312	0x7f9535013480	0x7f9532026240	16419367249702	16419367260036	16419367497157	16419367539856
143	141	void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]	0	0	422	158945	158945	4194304	256	0	0	8	24	255552	0x7f9535013300	0x7f9532026280	16419367547715	16419367558118	16419368012040	16419368054984
144	142	void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]	0	0	425	158945	158945	4194304	256	0	0	12	24	259840	0x7f9535013180	0x7f95320262c0	16419368070374	16419368081161	16419368306922	16419368331558
145	143	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	428	158945	158945	4194304	256	0	0	12	24	265664	0x7f9535013000	0x7f9532026300	16419368342398	16419368353002	16419368651084	16419368673271
146	144	void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]	0	0	431	158945	158945	4194304	256	0	0	12	24	271232	0x7f953286de80	0x7f9532026340	16419368684990	16419368694284	16419368992526	16419369014333
147	145	void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]	0	0	434	158945	158945	4194304	256	0	0	8	24	273216	0x7f953286dd00	0x7f9532026380	16419369022473	16419369032686	16419369329488	16419369345556
148	146	void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]	0	0	437	158945	158945	4194304	256	0	0	8	24	275712	0x7f953286db80	0x7f95320263c0	16419369354426	16419369364848	16419369922932	16419369939233
149	147	void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]	0	0	440	158945	158945	4194304	256	0	0	12	24	280768	0x7f953286da00	0x7f9532026400	16419369955242	16419369965652	16419370229654	16419370252796
150	148	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	443	158945	158945	4194304	256	0	0	12	24	287616	0x7f953286d880	0x7f9532026440	16419370263306	16419370273654	16419370623736	16419370645557
151	149	void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]	0	0	446	158945	158945	4194304	256	0	0	12	24	294208	0x7f953286d700	0x7f9532026480	16419370654137	16419370664376	16419371015258	16419371036789
152	150	void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]	0	0	449	158945	158945	4194304	256	0	0	8	24	296960	0x7f9535013580	0x7f95320264c0	16419371046438	16419371056699	16419371403581	16419371425130
153	151	void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]	0	0	452	158945	158945	4194304	256	0	0	8	24	299968	0x7f9535013400	0x7f9532026500	16419371434780	16419371444701	16419372106145	16419372124934
154	152	void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]	0	0	455	158945	158945	4194304	256	0	0	8	24	302720	0x7f9535013280	0x7f9532026540	16419372141584	16419372151265	16419372598468	16419372622323
155	153	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	458	158945	158945	4194304	256	0	0	12	24	305472	0x7f9535013100	0x7f9532026580	16419372631213	16419372640708	16419373105191	16419373123752
156	154	void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]	0	0	461	158945	158945	4194304	256	0	0	12	24	308224	0x7f953286df80	0x7f95320265c0	16419373131892	16419373142151	16419373604554	16419373625161
157	155	void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]	0	0	464	158945	158945	4194304	256	0	0	8	24	310976	0x7f953286de00	0x7f9532026600	16419373636731	16419373652394	16419374104077	16419374130280
158	156	void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]	0	0	467	158945	158945	4194304	256	0	0	8	24	313216	0x7f953286dc80	0x7f9532026640	16419374139320	16419374149517	16419375019762	16419375067319
159	157	void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]	0	0	470	158945	158945	4194304	256	0	0	8	24	315968	0x7f953286db00	0x7f9532026680	16419375084169	16419375094643	16419375948088	16419375993399
160	158	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	473	158945	158945	4194304	256	0	0	12	24	318720	0x7f953286d980	0x7f95320266c0	16419376004469	16419376019448	16419376905534	16419376950598
161	159	void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]	0	0	476	158945	158945	4194304	256	0	0	12	24	321472	0x7f953286d800	0x7f9532026700	16419376962008	16419376977374	16419377857699	16419377902717
162	160	void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]	0	0	479	158945	158945	4194304	256	0	0	8	24	324224	0x7f953286d680	0x7f9532026740	16419377912327	16419377929700	16419378790345	16419378838326
163	161	void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]	0	0	482	158945	158945	4194304	256	0	0	8	24	326464	0x7f9535013500	0x7f9532026780	16419378848836	16419378858185	16419380558356	16419380631497
164	162	void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]	0	0	485	158945	158945	4194304	256	0	0	8	24	329216	0x7f9535013380	0x7f95320267c0	16419380648556	16419380659156	16419382325246	16419382395668
165	163	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0	0	488	158945	158945	4194304	256	0	0	12	24	331968	0x7f9535013200	0x7f9532026800	16419382405348	16419382415487	16419384142217	16419384209488
166	164	void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]	0	0	491	158945	158945	4194304	256	0	0	12	24	334720	0x7f9535013080	0x7f9532026840	16419384219768	16419384228778	16419385946708	16419386013678
167	165	void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]	0	0	494	158945	158945	4194304	256	0	0	8	24	337472	0x7f953286df00	0x7f9532026880	16419386023368	16419386032628	16419387707678	16419387777980
168	166	void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]	0	0	497	158945	158945	4194304	256	0	0	8	24	0	0x7f953286dd80	0x7f95320268c0	16419387786719	16419387797119	16419391155699	16419391227223