rocm-systems/tests/workloads/no_roof_SPI/mi100/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_fillBuffer.kd	0	0	0	867387	867392	33554432	256	0	0	8	32	6464	0x0	0x7fe62c404180	12075254841546349	12075254841595086	12075254841918603	12075254842023395
3	1	void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]	0	0	2	867387	867392	32768	256	0	0	24	24	12480	0x0	0x7fe62c435100	12075254857277052	12075254857383717	12075254857390437	12075254857396494
4	2	void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]	0	0	4	867387	867392	4194304	256	0	0	24	24	12928	0x7fe737fe9900	0x7fe62c435140	12075254857442800	12075254857456036	12075254857593475	12075254857597277
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	6	867387	867392	4194304	256	0	0	36	24	13632	0x7fe737fe9800	0x7fe62c435180	12075254857642000	12075254857655395	12075254857898913	12075254857994085
6	4	void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]	0	0	8	867387	867392	4194304	256	0	0	28	24	14080	0x7fe737fe9700	0x7fe62c4351c0	12075254858026435	12075254858038911	12075254858281149	12075254858378269
7	5	void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]	0	0	10	867387	867392	4194304	256	0	0	24	24	14528	0x7fe737fe9600	0x7fe62c435200	12075254858403986	12075254858415228	12075254858540987	12075254858544187
8	6	void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]	0	0	12	867387	867392	4194304	256	0	0	24	24	14976	0x7fe737fe9500	0x7fe62c435240	12075254858569404	12075254858580827	12075254858705306	12075254858708693
9	7	void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]	0	0	14	867387	867392	4194304	256	0	0	24	24	15424	0x7fe737fe9400	0x7fe62c435280	12075254858760439	12075254858775065	12075254858899064	12075254858902202
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	16	867387	867392	4194304	256	0	0	28	24	16128	0x7fe737fe9300	0x7fe62c4352c0	12075254858928100	12075254858940184	12075254859191702	12075254859289522
11	9	void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]	0	0	18	867387	867392	4194304	256	0	0	28	24	16832	0x7fe737fe9a00	0x7fe62c435300	12075254859310822	12075254859322741	12075254859566579	12075254859661774
12	10	void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]	0	0	20	867387	867392	4194304	256	0	0	24	24	17536	0x7fe737fe9900	0x7fe62c435340	12075254859684726	12075254859696338	12075254859821137	12075254859824797
13	11	void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]	0	0	22	867387	867392	4194304	256	0	0	24	24	18240	0x7fe737fe9800	0x7fe62c435380	12075254859849072	12075254859861297	12075254859986416	12075254859989793
14	12	void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]	0	0	24	867387	867392	4194304	256	0	0	24	24	18944	0x7fe737fe9700	0x7fe62c4353c0	12075254860033935	12075254860045935	12075254860170894	12075254860174487
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	26	867387	867392	4194304	256	0	0	28	24	19904	0x7fe737fe9600	0x7fe62c435400	12075254860201196	12075254860212334	12075254860464812	12075254860535958
16	14	void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]	0	0	28	867387	867392	4194304	256	0	0	28	24	20608	0x7fe737fe9500	0x7fe62c435440	12075254860560263	12075254860571851	12075254860816009	12075254860882973
17	15	void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]	0	0	30	867387	867392	4194304	256	0	0	24	24	21312	0x7fe737fe9400	0x7fe62c435480	12075254860907779	12075254860919368	12075254861043847	12075254861047309
18	16	void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]	0	0	32	867387	867392	4194304	256	0	0	24	24	22016	0x7fe737fe9300	0x7fe62c4354c0	12075254861073778	12075254861084967	12075254861209766	12075254861212916
19	17	void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]	0	0	34	867387	867392	4194304	256	0	0	24	24	22720	0x7fe737fe9a00	0x7fe62c435500	12075254861251167	12075254861262725	12075254861387364	12075254861390576
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	36	867387	867392	4194304	256	0	0	28	24	23680	0x7fe737fe9900	0x7fe62c435540	12075254861416966	12075254861430404	12075254861684162	12075254861725068
21	19	void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]	0	0	38	867387	867392	4194304	256	0	0	28	24	24384	0x7fe737fe9800	0x7fe62c435580	12075254861746748	12075254861759841	12075254862040159	12075254862105806
22	20	void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]	0	0	40	867387	867392	4194304	256	0	0	24	24	25088	0x7fe737fe9700	0x7fe62c4355c0	12075254862131323	12075254862145918	12075254862270397	12075254862273888
23	21	void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]	0	0	42	867387	867392	4194304	256	0	0	24	24	25792	0x7fe737fe9600	0x7fe62c435600	12075254862299746	12075254862311357	12075254862436956	12075254862440067
24	22	void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]	0	0	44	867387	867392	4194304	256	0	0	24	24	26496	0x7fe737fe9500	0x7fe62c435640	12075254862476775	12075254862489275	12075254862613754	12075254862617166
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	46	867387	867392	4194304	256	0	0	28	24	27712	0x7fe737fe9400	0x7fe62c435680	12075254862644176	12075254862657914	12075254862905432	12075254862945977
26	24	void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]	0	0	48	867387	867392	4194304	256	0	0	28	24	28416	0x7fe737fe9300	0x7fe62c4356c0	12075254862968429	12075254862979991	12075254863224629	12075254863292381
27	25	void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]	0	0	50	867387	867392	4194304	256	0	0	24	24	29120	0x7fe737fe9a00	0x7fe62c435700	12075254863314782	12075254863326548	12075254863450547	12075254863453971
28	26	void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]	0	0	52	867387	867392	4194304	256	0	0	24	24	30080	0x7fe737fe9900	0x7fe62c435740	12075254863479849	12075254863491027	12075254863615826	12075254863618968
29	27	void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]	0	0	54	867387	867392	4194304	256	0	0	24	24	30784	0x7fe737fe9800	0x7fe62c435780	12075254863654674	12075254863666386	12075254863791025	12075254863794133
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	56	867387	867392	4194304	256	0	0	28	24	32000	0x7fe737fe9700	0x7fe62c4357c0	12075254863818969	12075254863832144	12075254864082542	12075254864149043
31	29	void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]	0	0	58	867387	867392	4194304	256	0	0	28	24	32960	0x7fe737fe9600	0x7fe62c435800	12075254864173037	12075254864185102	12075254864428620	12075254864474548
32	30	void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]	0	0	60	867387	867392	4194304	256	0	0	24	24	33920	0x7fe737fe9500	0x7fe62c435840	12075254864496028	12075254864507819	12075254864632778	12075254864636048
33	31	void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]	0	0	62	867387	867392	4194304	256	0	0	24	24	34880	0x7fe737fe9400	0x7fe62c435880	12075254864660894	12075254864672298	12075254864797577	12075254864800543
34	32	void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]	0	0	64	867387	867392	4194304	256	0	0	24	24	35840	0x7fe737fe9300	0x7fe62c4358c0	12075254864836901	12075254864849736	12075254864974855	12075254864977833
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	66	867387	867392	4194304	256	0	0	28	24	37312	0x7fe737fe9a00	0x7fe62c435900	12075254865007789	12075254865020935	12075254865304932	12075254865351587
36	34	void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]	0	0	68	867387	867392	4194304	256	0	0	28	24	38272	0x7fe737fe9900	0x7fe62c435940	12075254865372737	12075254865384612	12075254865629890	12075254865676231
37	35	void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]	0	0	70	867387	867392	4194304	256	0	0	24	24	39232	0x7fe737fe9800	0x7fe62c435980	12075254865699234	12075254865711169	12075254865835808	12075254865839103
38	36	void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]	0	0	72	867387	867392	4194304	256	0	0	24	24	40192	0x7fe737fe9700	0x7fe62c4359c0	12075254865865893	12075254865877568	12075254866003647	12075254866006424
39	37	void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]	0	0	74	867387	867392	4194304	256	0	0	24	24	41152	0x7fe737fe9600	0x7fe62c435a00	12075254866043624	12075254866058846	12075254866183485	12075254866186850
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	76	867387	867392	4194304	256	0	0	28	24	42624	0x7fe737fe9500	0x7fe62c435a40	12075254866212728	12075254866225565	12075254866477723	12075254866553732
41	39	void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]	0	0	78	867387	867392	4194304	256	0	0	28	24	43584	0x7fe737fe9400	0x7fe62c435a80	12075254866574039	12075254866588602	12075254866838360	12075254866922106
42	40	void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]	0	0	80	867387	867392	4194304	256	0	0	24	24	44544	0x7fe737fe9300	0x7fe62c435ac0	12075254866944207	12075254866957399	12075254867082678	12075254867086301
43	41	void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]	0	0	82	867387	867392	4194304	256	0	0	24	24	45760	0x7fe737fe9a00	0x7fe62c435b00	12075254867113051	12075254867124598	12075254867250197	12075254867253712
44	42	void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]	0	0	84	867387	867392	4194304	256	0	0	24	24	46720	0x7fe737fe9900	0x7fe62c435b40	12075254867289639	12075254867306036	12075254867431315	12075254867434549
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	86	867387	867392	4194304	256	0	0	28	24	48448	0x7fe737fe9800	0x7fe62c435b80	12075254867460587	12075254867473875	12075254867737713	12075254867762458
46	44	void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]	0	0	88	867387	867392	4194304	256	0	0	28	24	49408	0x7fe737fe9700	0x7fe62c435bc0	12075254867783768	12075254867794992	12075254868039310	12075254868068326
47	45	void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]	0	0	90	867387	867392	4194304	256	0	0	24	24	50368	0x7fe737fe9600	0x7fe62c435c00	12075254868087051	12075254868098350	12075254868223149	12075254868225779
48	46	void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]	0	0	92	867387	867392	4194304	256	0	0	24	24	51584	0x7fe737fe9500	0x7fe62c435c40	12075254868251887	12075254868264748	12075254868391147	12075254868394643
49	47	void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]	0	0	94	867387	867392	4194304	256	0	0	24	24	52544	0x7fe737fe9400	0x7fe62c435c80	12075254868430549	12075254868442347	12075254868567146	12075254868570399
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	96	867387	867392	4194304	256	0	0	28	24	54272	0x7fe737fe9300	0x7fe62c435cc0	12075254868596047	12075254868609066	12075254868866343	12075254868890144
51	49	void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]	0	0	98	867387	867392	4194304	256	0	0	28	24	55488	0x7fe737fe9a00	0x7fe62c435d00	12075254868911463	12075254868924103	12075254869168581	12075254869196152
52	50	void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]	0	0	100	867387	867392	4194304	256	0	0	24	24	56704	0x7fe737fe9900	0x7fe62c435d40	12075254869216720	12075254869230341	12075254869355300	12075254869358584
53	51	void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]	0	0	102	867387	867392	4194304	256	0	0	24	24	58176	0x7fe737fe9800	0x7fe62c435d80	12075254869384963	12075254869397539	12075254869525698	12075254869528951
54	52	void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]	0	0	104	867387	867392	4194304	256	0	0	24	24	59392	0x7fe737fe9700	0x7fe62c435dc0	12075254869565198	12075254869576738	12075254869702017	12075254869704898
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	106	867387	867392	4194304	256	0	0	28	24	61376	0x7fe737fe9600	0x7fe62c435e00	12075254869731307	12075254869744576	12075254870004574	12075254870032547
56	54	void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]	0	0	108	867387	867392	4194304	256	0	0	28	24	62592	0x7fe737fe9500	0x7fe62c435e40	12075254870053235	12075254870066334	12075254870312092	12075254870335369
57	55	void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]	0	0	110	867387	867392	4194304	256	0	0	24	24	63808	0x7fe737fe9400	0x7fe62c435e80	12075254870356469	12075254870369051	12075254870494010	12075254870497370
58	56	void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]	0	0	112	867387	867392	4194304	256	0	0	24	24	65280	0x7fe737fe9300	0x7fe62c435ec0	12075254870522407	12075254870533850	12075254870662809	12075254870666114
59	57	void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]	0	0	114	867387	867392	4194304	256	0	0	24	24	66496	0x7fe737fe9a00	0x7fe62c435f00	12075254870702391	12075254870715768	12075254870840887	12075254870844185
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	116	867387	867392	4194304	256	0	0	28	24	68480	0x7fe737fe9900	0x7fe62c435f40	12075254870868831	12075254870881687	12075254871136565	12075254871164220
61	59	void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]	0	0	118	867387	867392	4194304	256	0	0	28	24	69696	0x7fe737fe9800	0x7fe62c435f80	12075254871187102	12075254871199765	12075254871448243	12075254871476600
62	60	void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]	0	0	120	867387	867392	4194304	256	0	0	24	24	70912	0x7fe737fe9700	0x7fe62c435fc0	12075254871497439	12075254871510002	12075254871635441	12075254871638441
63	61	void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]	0	0	122	867387	867392	4194304	256	0	0	24	24	72384	0x7fe737fe9600	0x7fe62c436000	12075254871663968	12075254871675601	12075254871811600	12075254871814688
64	62	void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]	0	0	124	867387	867392	4194304	256	0	0	24	24	73600	0x7fe737fe9500	0x7fe62c436040	12075254871850735	12075254871861999	12075254871986638	12075254871989724
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	126	867387	867392	4194304	256	0	0	28	24	75840	0x7fe737fe9400	0x7fe62c436080	12075254872020341	12075254872032078	12075254872327755	12075254872357287
66	64	void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]	0	0	128	867387	867392	4194304	256	0	0	28	24	77056	0x7fe737fe9300	0x7fe62c4360c0	12075254872378586	12075254872390635	12075254872634153	12075254872657725
67	65	void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]	0	0	130	867387	867392	4194304	256	0	0	24	24	78272	0x7fe737fe9a00	0x7fe62c436100	12075254872678343	12075254872689992	12075254872814311	12075254872817422
68	66	void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]	0	0	132	867387	867392	4194304	256	0	0	24	24	80000	0x7fe737fe9900	0x7fe62c436140	12075254872845484	12075254872856871	12075254872998630	12075254873005742
69	67	void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]	0	0	134	867387	867392	4194304	256	0	0	24	24	81216	0x7fe737fe9800	0x7fe62c436180	12075254873038493	12075254873050150	12075254873175588	12075254873178934
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	136	867387	867392	4194304	256	0	0	28	24	83456	0x7fe737fe9700	0x7fe62c4361c0	12075254873204692	12075254873217668	12075254873473186	12075254873499570
71	69	void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]	0	0	138	867387	867392	4194304	256	0	0	28	24	84928	0x7fe737fe9600	0x7fe62c436200	12075254873519727	12075254873532546	12075254873779744	12075254873805799
72	70	void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]	0	0	140	867387	867392	4194304	256	0	0	24	24	86400	0x7fe737fe9500	0x7fe62c436240	12075254873825565	12075254873838463	12075254873963742	12075254873967048
73	71	void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]	0	0	142	867387	867392	4194304	256	0	0	24	24	88128	0x7fe737fe9400	0x7fe62c436280	12075254873992896	12075254874005502	12075254874155901	12075254874159255
74	72	void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]	0	0	144	867387	867392	4194304	256	0	0	24	24	89600	0x7fe737fe9300	0x7fe62c4362c0	12075254874210080	12075254874223900	12075254874349499	12075254874352845
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	146	867387	867392	4194304	256	0	0	28	24	92096	0x7fe737fe9a00	0x7fe62c436300	12075254874377651	12075254874390619	12075254874644857	12075254874667420
76	74	void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]	0	0	148	867387	867392	4194304	256	0	0	28	24	93568	0x7fe737fe9900	0x7fe62c436340	12075254874689270	12075254874702456	12075254874951254	12075254874973929
77	75	void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]	0	0	150	867387	867392	4194304	256	0	0	24	24	95040	0x7fe737fe9800	0x7fe62c436380	12075254874993866	12075254875006614	12075254875132053	12075254875135329
78	76	void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]	0	0	152	867387	867392	4194304	256	0	0	24	24	96768	0x7fe737fe9700	0x7fe62c4363c0	12075254875161127	12075254875172852	12075254875329971	12075254875333267
79	77	void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]	0	0	154	867387	867392	4194304	256	0	0	24	24	98240	0x7fe737fe9600	0x7fe62c436400	12075254875367391	12075254875383571	12075254875508690	12075254875512059
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	156	867387	867392	4194304	256	0	0	28	24	100736	0x7fe737fe9500	0x7fe62c436440	12075254875538178	12075254875551409	12075254875807567	12075254875826895
81	79	void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]	0	0	158	867387	867392	4194304	256	0	0	28	24	102208	0x7fe737fe9400	0x7fe62c436480	12075254875847002	12075254875860047	12075254876109485	12075254876136821
82	80	void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]	0	0	160	867387	867392	4194304	256	0	0	24	24	103680	0x7fe737fe9300	0x7fe62c4364c0	12075254876156828	12075254876169324	12075254876294923	12075254876298200
83	81	void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]	0	0	162	867387	867392	4194304	256	0	0	24	24	105664	0x7fe737fe9a00	0x7fe62c436500	12075254876325952	12075254876338603	12075254876504201	12075254876507680
84	82	void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]	0	0	164	867387	867392	4194304	256	0	0	24	24	107136	0x7fe737fe9900	0x7fe62c436540	12075254876542114	12075254876555241	12075254876681480	12075254876684709
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	166	867387	867392	4194304	256	0	0	28	24	109888	0x7fe737fe9800	0x7fe62c436580	12075254876709725	12075254876722760	12075254876980198	12075254876999003
86	84	void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]	0	0	168	867387	867392	4194304	256	0	0	28	24	111360	0x7fe737fe9700	0x7fe62c4365c0	12075254877027636	12075254877040837	12075254877288355	12075254877311373
87	85	void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]	0	0	170	867387	867392	4194304	256	0	0	24	24	112832	0x7fe737fe9600	0x7fe62c436600	12075254877331571	12075254877344195	12075254877470434	12075254877473645
88	86	void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]	0	0	172	867387	867392	4194304	256	0	0	24	24	114816	0x7fe737fe9500	0x7fe62c436640	12075254877497900	12075254877510913	12075254877682912	12075254877686019
89	87	void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]	0	0	174	867387	867392	4194304	256	0	0	24	24	116288	0x7fe737fe9400	0x7fe62c436680	12075254877721746	12075254877734431	12075254877859870	12075254877863189
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	176	867387	867392	4194304	256	0	0	28	24	119040	0x7fe737fe9300	0x7fe62c4366c0	12075254877888335	12075254877902110	12075254878153788	12075254878184346
91	89	void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]	0	0	178	867387	867392	4194304	256	0	0	28	24	120768	0x7fe737fe9a00	0x7fe62c436700	12075254878205184	12075254878217788	12075254878464826	12075254878490805
92	90	void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]	0	0	180	867387	867392	4194304	256	0	0	24	24	122496	0x7fe737fe9900	0x7fe62c436740	12075254878511944	12075254878525145	12075254878650904	12075254878654259
93	91	void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]	0	0	182	867387	867392	4194304	256	0	0	24	24	124736	0x7fe737fe9800	0x7fe62c436780	12075254878679165	12075254878690424	12075254878871702	12075254878874829
94	92	void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]	0	0	184	867387	867392	4194304	256	0	0	24	24	126464	0x7fe737fe9700	0x7fe62c4367c0	12075254878909623	12075254878922422	12075254879047861	12075254879051327
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	186	867387	867392	4194304	256	0	0	20	24	127936	0x7fe737fe9600	0x7fe62c436800	12075254879077385	12075254879090420	12075254879353938	12075254879379657
96	94	void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]	0	0	188	867387	867392	4194304	256	0	0	28	24	129664	0x7fe737fe9500	0x7fe62c436840	12075254879399474	12075254879411858	12075254879656176	12075254879682540
97	95	void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]	0	0	190	867387	867392	4194304	256	0	0	24	24	131392	0x7fe737fe9400	0x7fe62c436880	12075254879703188	12075254879716015	12075254879841774	12075254879844952
98	96	void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]	0	0	192	867387	867392	4194304	256	0	0	24	24	133632	0x7fe737fe9300	0x7fe62c4368c0	12075254879874326	12075254879886734	12075254880077932	12075254880081301
99	97	void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]	0	0	194	867387	867392	4194304	256	0	0	24	24	135360	0x7fe737fe9a00	0x7fe62c436900	12075254880116877	12075254880129772	12075254880257771	12075254880260995
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	196	867387	867392	4194304	256	0	0	20	24	136832	0x7fe737fe9900	0x7fe62c436940	12075254880286192	12075254880299851	12075254880582088	12075254880601087
101	99	void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]	0	0	198	867387	867392	4194304	256	0	0	28	24	138560	0x7fe737fe9800	0x7fe62c436980	12075254880625602	12075254880638248	12075254880890886	12075254880913297
102	100	void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]	0	0	200	867387	867392	4194304	256	0	0	24	24	140288	0x7fe737fe9700	0x7fe62c4369c0	12075254880932713	12075254880943845	12075254881070244	12075254881073695
103	101	void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]	0	0	202	867387	867392	4194304	256	0	0	24	24	142784	0x7fe737fe9600	0x7fe62c436a00	12075254881100385	12075254881111524	12075254881320322	12075254881346773
104	102	void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]	0	0	204	867387	867392	4194304	256	0	0	24	24	144768	0x7fe737fe9500	0x7fe62c436a40	12075254881376327	12075254881389122	12075254881517601	12075254881520786
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	206	867387	867392	4194304	256	0	0	20	24	145728	0x7fe737fe9400	0x7fe62c436a80	12075254881545742	12075254881558880	12075254881895518	12075254881918525
106	104	void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]	0	0	208	867387	867392	4194304	256	0	0	28	24	147712	0x7fe737fe9300	0x7fe62c436ac0	12075254881940927	12075254881953757	12075254882202715	12075254882234342
107	105	void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]	0	0	210	867387	867392	4194304	256	0	0	24	24	149696	0x7fe737fe9a00	0x7fe62c436b00	12075254882255251	12075254882268635	12075254882395834	12075254882399008
108	106	void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]	0	0	212	867387	867392	4194304	256	0	0	24	24	152192	0x7fe737fe9900	0x7fe62c436b40	12075254882425066	12075254882437593	12075254882664791	12075254882691211
109	107	void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]	0	0	214	867387	867392	4194304	256	0	0	24	24	154176	0x7fe737fe9800	0x7fe62c436b80	12075254882720095	12075254882734231	12075254882861750	12075254882865104
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	216	867387	867392	4194304	256	0	0	20	24	155648	0x7fe737fe9700	0x7fe62c436bc0	12075254882890201	12075254882902389	12075254883230227	12075254883262803
111	109	void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]	0	0	218	867387	867392	4194304	256	0	0	28	24	157632	0x7fe737fe9600	0x7fe62c436c00	12075254883284013	12075254883296626	12075254883547984	12075254883574653
112	110	void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]	0	0	220	867387	867392	4194304	256	0	0	24	24	159616	0x7fe737fe9500	0x7fe62c436c40	12075254883595501	12075254883608464	12075254883736463	12075254883739740
113	111	void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]	0	0	222	867387	867392	4194304	256	0	0	24	24	162368	0x7fe737fe9400	0x7fe62c436c80	12075254883765938	12075254883778542	12075254884022860	12075254884050878
114	112	void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]	0	0	224	867387	867392	4194304	256	0	0	24	24	164608	0x7fe737fe9300	0x7fe62c436cc0	12075254884082226	12075254884096300	12075254884224779	12075254884228057
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	226	867387	867392	4194304	256	0	0	20	24	166336	0x7fe737fe9a00	0x7fe62c436d00	12075254884252322	12075254884263658	12075254884638375	12075254884660701
116	114	void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]	0	0	228	867387	867392	4194304	256	0	0	28	24	168576	0x7fe737fe9900	0x7fe62c436d40	12075254884679636	12075254884691015	12075254884944133	12075254884970597
117	115	void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]	0	0	230	867387	867392	4194304	256	0	0	24	24	170816	0x7fe737fe9800	0x7fe62c436d80	12075254884991165	12075254885003812	12075254885134211	12075254885137617
118	116	void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]	0	0	232	867387	867392	4194304	256	0	0	24	24	174080	0x7fe737fe9700	0x7fe62c436dc0	12075254885162934	12075254885174531	12075254885453089	12075254885479292
119	117	void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]	0	0	234	867387	867392	4194304	256	0	0	24	24	176576	0x7fe737fe9600	0x7fe62c436e00	12075254885509599	12075254885522528	12075254885657407	12075254885660689
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	236	867387	867392	4194304	256	0	0	20	24	177792	0x7fe737fe9500	0x7fe62c436e40	12075254885684914	12075254885697727	12075254886120283	12075254886148506
121	119	void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]	0	0	238	867387	867392	4194304	256	0	0	28	24	180288	0x7fe737fe9400	0x7fe62c436e80	12075254886172430	12075254886185403	12075254886451801	12075254886477778
122	120	void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]	0	0	240	867387	867392	4194304	256	0	0	24	24	182784	0x7fe737fe9300	0x7fe62c436ec0	12075254886497735	12075254886510360	12075254886645719	12075254886649146
123	121	void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]	0	0	242	867387	867392	4194304	256	0	0	24	24	186304	0x7fe737fe9a00	0x7fe62c436f00	12075254886673572	12075254886686519	12075254886999636	12075254887026858
124	122	void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]	0	0	244	867387	867392	4194304	256	0	0	24	24	189312	0x7fe737fe9900	0x7fe62c436f40	12075254887058537	12075254887072276	12075254887223954	12075254887227251
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	246	867387	867392	4194304	256	0	0	20	24	190784	0x7fe737fe9800	0x7fe62c436f80	12075254887252548	12075254887265554	12075254887780110	12075254887806066
126	124	void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]	0	0	248	867387	867392	4194304	256	0	0	28	24	193792	0x7fe737fe9700	0x7fe62c436fc0	12075254887827105	12075254887839789	12075254888137387	12075254888170984
127	125	void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]	0	0	250	867387	867392	4194304	256	0	0	24	24	196800	0x7fe737fe9600	0x7fe62c437000	12075254888192394	12075254888205866	12075254888361065	12075254888364584
128	126	void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]	0	0	252	867387	867392	4194304	256	0	0	24	24	201088	0x7fe737fe9500	0x7fe62c437040	12075254888390212	12075254888402825	12075254888786982	12075254888812987
129	127	void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]	0	0	254	867387	867392	4194304	256	0	0	24	24	204608	0x7fe737fe9400	0x7fe62c437080	12075254888842893	12075254888856421	12075254889032260	12075254889035802
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	256	867387	867392	4194304	256	0	0	20	24	206080	0x7fe737fe9300	0x7fe62c4370c0	12075254889061760	12075254889073379	12075254889681374	12075254889707209
131	129	void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]	0	0	258	867387	867392	4194304	256	0	0	28	24	209600	0x7fe737fe9a00	0x7fe62c437100	12075254889729601	12075254889740734	12075254890067131	12075254890094489
132	130	void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]	0	0	260	867387	867392	4194304	256	0	0	24	24	213120	0x7fe737fe9900	0x7fe62c437140	12075254890115618	12075254890128411	12075254890306169	12075254890309569
133	131	void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]	0	0	262	867387	867392	4194304	256	0	0	24	24	218176	0x7fe737fe9800	0x7fe62c437180	12075254890334755	12075254890347449	12075254890802965	12075254890828984
134	132	void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]	0	0	264	867387	867392	4194304	256	0	0	24	24	222208	0x7fe737fe9700	0x7fe62c4371c0	12075254890859731	12075254890871445	12075254891071923	12075254891111168
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	266	867387	867392	4194304	256	0	0	20	24	223936	0x7fe737fe9600	0x7fe62c437200	12075254891130915	12075254891142963	12075254891844237	12075254891871852
136	134	void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]	0	0	268	867387	867392	4194304	256	0	0	28	24	227968	0x7fe737fe9500	0x7fe62c437240	12075254891895356	12075254891907916	12075254892280713	12075254892312851
137	135	void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]	0	0	270	867387	867392	4194304	256	0	0	24	24	232000	0x7fe737fe9400	0x7fe62c437280	12075254892334922	12075254892347593	12075254892548071	12075254892573626
138	136	void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]	0	0	272	867387	867392	4194304	256	0	0	24	24	237824	0x7fe737fe9300	0x7fe62c4372c0	12075254892592822	12075254892604071	12075254893129186	12075254893156549
139	137	void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]	0	0	274	867387	867392	4194304	256	0	0	24	24	242368	0x7fe737fe9a00	0x7fe62c437300	12075254893187256	12075254893199746	12075254893423104	12075254893448833
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	276	867387	867392	4194304	256	0	0	20	24	243584	0x7fe737fe9900	0x7fe62c437340	12075254893470022	12075254893481344	12075254894278297	12075254894305004
141	139	void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]	0	0	278	867387	867392	4194304	256	0	0	28	24	248128	0x7fe737fe9800	0x7fe62c437380	12075254894327576	12075254894340217	12075254894760053	12075254894785907
142	140	void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]	0	0	280	867387	867392	4194304	256	0	0	24	24	252672	0x7fe737fe9700	0x7fe62c4373c0	12075254894807217	12075254894820213	12075254895045171	12075254895078381
143	141	void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]	0	0	282	867387	867392	4194304	256	0	0	24	24	259264	0x7fe737fe9600	0x7fe62c437400	12075254895101354	12075254895114130	12075254895709805	12075254895736544
144	142	void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]	0	0	284	867387	867392	4194304	256	0	0	24	24	264832	0x7fe737fe9500	0x7fe62c437440	12075254895766540	12075254895779885	12075254896047883	12075254896075203
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	286	867387	867392	4194304	256	0	0	20	24	266304	0x7fe737fe9400	0x7fe62c437480	12075254896095862	12075254896107402	12075254897090754	12075254897118021
146	144	void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]	0	0	288	867387	867392	4194304	256	0	0	28	24	271872	0x7fe737fe9300	0x7fe62c4374c0	12075254897137708	12075254897148994	12075254897680029	12075254897706264
147	145	void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]	0	0	290	867387	867392	4194304	256	0	0	24	24	277440	0x7fe737fe9a00	0x7fe62c437500	12075254897728185	12075254897739229	12075254898009947	12075254898037530
148	146	void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]	0	0	292	867387	867392	4194304	256	0	0	12	24	279936	0x7fe737fe9900	0x7fe62c437540	12075254898057928	12075254898071226	12075254898807860	12075254898833990
149	147	void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]	0	0	294	867387	867392	4194304	256	0	0	24	24	286528	0x7fe737fe9800	0x7fe62c437580	12075254898865338	12075254898878260	12075254899193777	12075254899225818
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	296	867387	867392	4194304	256	0	0	20	24	288000	0x7fe737fe9700	0x7fe62c4375c0	12075254899247909	12075254899261457	12075254900430727	12075254900457928
151	149	void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]	0	0	298	867387	867392	4194304	256	0	0	32	24	294592	0x7fe737fe9600	0x7fe62c437600	12075254900480239	12075254900492967	12075254901105282	12075254901133584
152	150	void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]	0	0	300	867387	867392	4194304	256	0	0	24	24	301184	0x7fe737fe9500	0x7fe62c437640	12075254901154112	12075254901166721	12075254901484639	12075254901511716
153	151	void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]	0	0	302	867387	867392	4194304	256	0	0	12	24	304960	0x7fe737fe9400	0x7fe62c437680	12075254901533787	12075254901544798	12075254902421591	12075254902450912
154	152	void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]	0	0	304	867387	867392	4194304	256	0	0	24	24	313600	0x7fe737fe9300	0x7fe62c4376c0	12075254902482530	12075254902494390	12075254902904467	12075254902932406
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	306	867387	867392	4194304	256	0	0	20	24	314816	0x7fe737fe9a00	0x7fe62c437700	12075254902954497	12075254902966227	12075254904515014	12075254904589055
156	154	void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]	0	0	308	867387	867392	4194304	256	0	0	20	24	317568	0x7fe737fe9900	0x7fe62c437740	12075254904608591	12075254904622073	12075254905415188	12075254905489528
157	155	void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]	0	0	310	867387	867392	4194304	256	0	0	12	24	320320	0x7fe737fe9800	0x7fe62c437780	12075254905510257	12075254905521906	12075254905935664	12075254906007531
158	156	void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]	0	0	312	867387	867392	4194304	256	0	0	12	24	324096	0x7fe737fe9700	0x7fe62c4377c0	12075254906032037	12075254906043983	12075254907202535	12075254907272552
159	157	void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]	0	0	314	867387	867392	4194304	256	0	0	12	24	328896	0x7fe737fe9600	0x7fe62c437800	12075254907303409	12075254907315174	12075254908100289	12075254908170190
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	316	867387	867392	4194304	256	0	0	20	24	330112	0x7fe737fe9500	0x7fe62c437840	12075254908190177	12075254908202689	12075254911254028	12075254911324231
161	159	void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]	0	0	318	867387	867392	4194304	256	0	0	20	24	332864	0x7fe737fe9400	0x7fe62c437880	12075254911344720	12075254911356108	12075254912897857	12075254912968197
162	160	void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]	0	0	320	867387	867392	4194304	256	0	0	12	24	335616	0x7fe737fe9300	0x7fe62c4378c0	12075254912988765	12075254913002977	12075254913791291	12075254913864693
163	161	void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]	0	0	322	867387	867392	4194304	256	0	0	12	24	339392	0x7fe737fe9a00	0x7fe62c437900	12075254913885601	12075254913898651	12075254916180396	12075254916250266
164	162	void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]	0	0	324	867387	867392	4194304	256	0	0	12	24	344192	0x7fe737fe9900	0x7fe62c437940	12075254916282075	12075254916295915	12075254917831585	12075254917901965
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	326	867387	867392	4194304	256	0	0	20	24	345408	0x7fe737fe9800	0x7fe62c437980	12075254917922764	12075254917936544	12075254923992184	12075254924066582
166	164	void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]	0	0	328	867387	867392	4194304	256	0	0	20	24	348160	0x7fe737fe9700	0x7fe62c4379c0	12075254924091418	12075254924104663	12075254927146403	12075254927218709
167	165	void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]	0	0	330	867387	867392	4194304	256	0	0	12	24	350912	0x7fe737fe9600	0x7fe62c437a00	12075254927239528	12075254927251522	12075254928790552	12075254928863376
168	166	void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]	0	0	332	867387	867392	4194304	256	0	0	12	24	0	0x7fe737fe9500	0x7fe62c437a40	12075254928882802	12075254928895191	12075254933422201	12075254933498349