rocm-systems/tests/workloads/no_roof_mixbench2/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	767988	767993	33554432	256	0	0	8	32	6464	0x0	0x7f5046004180	12073517721390637	12073517721430385	12073517721754222	12073517721862724
3	1	void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]	0	0	2	767988	767993	32768	256	0	0	24	24	12480	0x0	0x7f5046035100	12073517737447236	12073517737554518	12073517737561398	12073517737566849
4	2	void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]	0	0	4	767988	767993	4194304	256	0	0	24	24	12928	0x7f5151c37900	0x7f5046035140	12073517737624716	12073517737638997	12073517737768436	12073517737771760
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	767988	767993	4194304	256	0	0	36	24	13632	0x7f5151c37800	0x7f5046035180	12073517737826181	12073517737841555	12073517738087473	12073517738188765
6	4	void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]	0	0	8	767988	767993	4194304	256	0	0	28	24	14080	0x7f5151c37700	0x7f50460351c0	12073517738227196	12073517738239632	12073517738488109	12073517738586664
7	5	void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]	0	0	10	767988	767993	4194304	256	0	0	24	24	14528	0x7f5151c37600	0x7f5046035200	12073517738622320	12073517738634188	12073517738759467	12073517738762741
8	6	void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]	0	0	12	767988	767993	4194304	256	0	0	24	24	14976	0x7f5151c37500	0x7f5046035240	12073517738799590	12073517738811786	12073517738937065	12073517738939991
9	7	void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]	0	0	14	767988	767993	4194304	256	0	0	24	24	15424	0x7f5151c37400	0x7f5046035280	12073517739007496	12073517739021064	12073517739145703	12073517739148268
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	767988	767993	4194304	256	0	0	28	24	16128	0x7f5151c37300	0x7f50460352c0	12073517739186769	12073517739199143	12073517739442501	12073517739540747
11	9	void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]	0	0	18	767988	767993	4194304	256	0	0	28	24	16832	0x7f5151c37a00	0x7f5046035300	12073517739570142	12073517739581859	12073517739827617	12073517739925232
12	10	void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]	0	0	20	767988	767993	4194304	256	0	0	24	24	17536	0x7f5151c37900	0x7f5046035340	12073517739956410	12073517739968576	12073517740093215	12073517740097071
13	11	void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]	0	0	22	767988	767993	4194304	256	0	0	24	24	18240	0x7f5151c37800	0x7f5046035380	12073517740133399	12073517740146814	12073517740272253	12073517740275252
14	12	void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]	0	0	24	767988	767993	4194304	256	0	0	24	24	18944	0x7f5151c37700	0x7f50460353c0	12073517740322410	12073517740336412	12073517740460571	12073517740463452
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	767988	767993	4194304	256	0	0	28	24	19904	0x7f5151c37600	0x7f5046035400	12073517740499579	12073517740514651	12073517740763929	12073517740862303
16	14	void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]	0	0	28	767988	767993	4194304	256	0	0	28	24	20608	0x7f5151c37500	0x7f5046035440	12073517740892219	12073517740904567	12073517741150485	12073517741222543
17	15	void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]	0	0	30	767988	767993	4194304	256	0	0	24	24	21312	0x7f5151c37400	0x7f5046035480	12073517741257968	12073517741270804	12073517741395443	12073517741398079
18	16	void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]	0	0	32	767988	767993	4194304	256	0	0	24	24	22016	0x7f5151c37300	0x7f50460354c0	12073517741434837	12073517741446962	12073517741572081	12073517741574998
19	17	void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]	0	0	34	767988	767993	4194304	256	0	0	24	24	22720	0x7f5151c37a00	0x7f5046035500	12073517741620662	12073517741633201	12073517741758159	12073517741761033
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	767988	767993	4194304	256	0	0	28	24	23680	0x7f5151c37900	0x7f5046035540	12073517741796539	12073517741809679	12073517742055597	12073517742124739
21	19	void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]	0	0	38	767988	767993	4194304	256	0	0	28	24	24384	0x7f5151c37800	0x7f5046035580	12073517742159033	12073517742171596	12073517742421353	12073517742490208
22	20	void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]	0	0	40	767988	767993	4194304	256	0	0	24	24	25088	0x7f5151c37700	0x7f50460355c0	12073517742519863	12073517742533192	12073517742657351	12073517742659864
23	21	void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]	0	0	42	767988	767993	4194304	256	0	0	24	24	25792	0x7f5151c37600	0x7f5046035600	12073517742693266	12073517742708231	12073517742833190	12073517742836161
24	22	void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]	0	0	44	767988	767993	4194304	256	0	0	24	24	26496	0x7f5151c37500	0x7f5046035640	12073517742879482	12073517742892709	12073517743016708	12073517743019692
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	767988	767993	4194304	256	0	0	28	24	27712	0x7f5151c37400	0x7f5046035680	12073517743056741	12073517743073988	12073517743320865	12073517743393868
26	24	void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]	0	0	48	767988	767993	4194304	256	0	0	28	24	28416	0x7f5151c37300	0x7f50460356c0	12073517743424394	12073517743436224	12073517743680542	12073517743778593
27	25	void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]	0	0	50	767988	767993	4194304	256	0	0	24	24	29120	0x7f5151c37a00	0x7f5046035700	12073517743806464	12073517743818781	12073517743943260	12073517743946084
28	26	void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]	0	0	52	767988	767993	4194304	256	0	0	24	24	30080	0x7f5151c37900	0x7f5046035740	12073517743979175	12073517743991259	12073517744116218	12073517744119176
29	27	void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]	0	0	54	767988	767993	4194304	256	0	0	24	24	30784	0x7f5151c37800	0x7f5046035780	12073517744163538	12073517744175418	12073517744299896	12073517744302807
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	767988	767993	4194304	256	0	0	28	24	32000	0x7f5151c37700	0x7f50460357c0	12073517744336590	12073517744349976	12073517744635093	12073517744707789
31	29	void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]	0	0	58	767988	767993	4194304	256	0	0	28	24	32960	0x7f5151c37600	0x7f5046035800	12073517744735461	12073517744748692	12073517744993010	12073517745071014
32	30	void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]	0	0	60	767988	767993	4194304	256	0	0	24	24	33920	0x7f5151c37500	0x7f5046035840	12073517745101110	12073517745113329	12073517745238288	12073517745241080
33	31	void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]	0	0	62	767988	767993	4194304	256	0	0	24	24	34880	0x7f5151c37400	0x7f5046035880	12073517745274623	12073517745286127	12073517745411886	12073517745414723
34	32	void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]	0	0	64	767988	767993	4194304	256	0	0	24	24	35840	0x7f5151c37300	0x7f50460358c0	12073517745455900	12073517745468366	12073517745593325	12073517745596220
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	767988	767993	4194304	256	0	0	28	24	37312	0x7f5151c37a00	0x7f5046035900	12073517745629472	12073517745641804	12073517745890442	12073517745943977
36	34	void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]	0	0	68	767988	767993	4194304	256	0	0	28	24	38272	0x7f5151c37900	0x7f5046035940	12073517745971017	12073517745983081	12073517746231239	12073517746312361
37	35	void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]	0	0	70	767988	767993	4194304	256	0	0	24	24	39232	0x7f5151c37800	0x7f5046035980	12073517746342457	12073517746354278	12073517746479877	12073517746482798
38	36	void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]	0	0	72	767988	767993	4194304	256	0	0	24	24	40192	0x7f5151c37700	0x7f50460359c0	12073517746519426	12073517746532996	12073517746659715	12073517746662762
39	37	void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]	0	0	74	767988	767993	4194304	256	0	0	24	24	41152	0x7f5151c37600	0x7f5046035a00	12073517746705552	12073517746719234	12073517746843073	12073517746846003
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	767988	767993	4194304	256	0	0	28	24	42624	0x7f5151c37500	0x7f5046035a40	12073517746878734	12073517746892833	12073517747141631	12073517747171187
41	39	void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]	0	0	78	767988	767993	4194304	256	0	0	28	24	43584	0x7f5151c37400	0x7f5046035a80	12073517747200842	12073517747213950	12073517747463708	12073517747515056
42	40	void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]	0	0	80	767988	767993	4194304	256	0	0	24	24	44544	0x7f5151c37300	0x7f5046035ac0	12073517747544501	12073517747558427	12073517747683226	12073517747685754
43	41	void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]	0	0	82	767988	767993	4194304	256	0	0	24	24	45760	0x7f5151c37a00	0x7f5046035b00	12073517747719496	12073517747737145	12073517747864184	12073517747867561
44	42	void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]	0	0	84	767988	767993	4194304	256	0	0	24	24	46720	0x7f5151c37900	0x7f5046035b40	12073517747910050	12073517747923543	12073517748048342	12073517748051694
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	767988	767993	4194304	256	0	0	28	24	48448	0x7f5151c37800	0x7f5046035b80	12073517748085877	12073517748100982	12073517748363379	12073517748393058
46	44	void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]	0	0	88	767988	767993	4194304	256	0	0	28	24	49408	0x7f5151c37700	0x7f5046035bc0	12073517748426610	12073517748440819	12073517748686736	12073517748715127
47	45	void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]	0	0	90	767988	767993	4194304	256	0	0	24	24	50368	0x7f5151c37600	0x7f5046035c00	12073517748741205	12073517748755536	12073517748880175	12073517748883099
48	46	void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]	0	0	92	767988	767993	4194304	256	0	0	24	24	51584	0x7f5151c37500	0x7f5046035c40	12073517748915208	12073517748928654	12073517749055213	12073517749058545
49	47	void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]	0	0	94	767988	767993	4194304	256	0	0	24	24	52544	0x7f5151c37400	0x7f5046035c80	12073517749104029	12073517749117453	12073517749242571	12073517749245162
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	767988	767993	4194304	256	0	0	28	24	54272	0x7f5151c37300	0x7f5046035cc0	12073517749279085	12073517749294411	12073517749547689	12073517749569164
51	49	void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]	0	0	98	767988	767993	4194304	256	0	0	28	24	55488	0x7f5151c37a00	0x7f5046035d00	12073517749599170	12073517749611848	12073517749856966	12073517749881614
52	50	void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]	0	0	100	767988	767993	4194304	256	0	0	24	24	56704	0x7f5151c37900	0x7f5046035d40	12073517749909176	12073517749923205	12073517750048164	12073517750051119
53	51	void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]	0	0	102	767988	767993	4194304	256	0	0	24	24	58176	0x7f5151c37800	0x7f5046035d80	12073517750087467	12073517750101124	12073517750229922	12073517750232677
54	52	void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]	0	0	104	767988	767993	4194304	256	0	0	24	24	59392	0x7f5151c37700	0x7f5046035dc0	12073517750277190	12073517750290562	12073517750415201	12073517750417901
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	767988	767993	4194304	256	0	0	28	24	61376	0x7f5151c37600	0x7f5046035e00	12073517750452145	12073517750465120	12073517750726558	12073517750746932
56	54	void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]	0	0	108	767988	767993	4194304	256	0	0	28	24	62592	0x7f5151c37500	0x7f5046035e40	12073517750776307	12073517750790237	12073517751037115	12073517751066877
57	55	void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]	0	0	110	767988	767993	4194304	256	0	0	24	24	63808	0x7f5151c37400	0x7f5046035e80	12073517751095150	12073517751108314	12073517751233433	12073517751236623
58	56	void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]	0	0	112	767988	767993	4194304	256	0	0	24	24	65280	0x7f5151c37300	0x7f5046035ec0	12073517751270636	12073517751284633	12073517751413272	12073517751415906
59	57	void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]	0	0	114	767988	767993	4194304	256	0	0	24	24	66496	0x7f5151c37a00	0x7f5046035f00	12073517751460028	12073517751474391	12073517751599670	12073517751602573
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	767988	767993	4194304	256	0	0	28	24	68480	0x7f5151c37900	0x7f5046035f40	12073517751635574	12073517751649590	12073517751903667	12073517751928118
61	59	void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]	0	0	118	767988	767993	4194304	256	0	0	28	24	69696	0x7f5151c37800	0x7f5046035f80	12073517751959887	12073517751974227	12073517752221104	12073517752249244
62	60	void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]	0	0	120	767988	767993	4194304	256	0	0	24	24	70912	0x7f5151c37700	0x7f5046035fc0	12073517752277918	12073517752291184	12073517752416943	12073517752419651
63	61	void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]	0	0	122	767988	767993	4194304	256	0	0	24	24	72384	0x7f5151c37600	0x7f5046036000	12073517752453794	12073517752467182	12073517752604141	12073517752606899
64	62	void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]	0	0	124	767988	767993	4194304	256	0	0	24	24	73600	0x7f5151c37500	0x7f5046036040	12073517752650961	12073517752665420	12073517752790539	12073517752793165
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	767988	767993	4194304	256	0	0	28	24	75840	0x7f5151c37400	0x7f5046036080	12073517752826156	12073517752838059	12073517753094056	12073517753123008
66	64	void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]	0	0	128	767988	767993	4194304	256	0	0	28	24	77056	0x7f5151c37300	0x7f50460360c0	12073517753153034	12073517753167336	12073517753412133	12073517753440458
67	65	void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]	0	0	130	767988	767993	4194304	256	0	0	24	24	78272	0x7f5151c37a00	0x7f5046036100	12073517753469081	12073517753483013	12073517753607492	12073517753610244
68	66	void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]	0	0	132	767988	767993	4194304	256	0	0	24	24	80000	0x7f5151c37900	0x7f5046036140	12073517753645199	12073517753658211	12073517753800610	12073517753803633
69	67	void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]	0	0	134	767988	767993	4194304	256	0	0	24	24	81216	0x7f5151c37800	0x7f5046036180	12073517753846042	12073517753857409	12073517753983008	12073517753985862
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	767988	767993	4194304	256	0	0	28	24	83456	0x7f5151c37700	0x7f50460361c0	12073517754027930	12073517754041248	12073517754292605	12073517754319762
71	69	void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]	0	0	138	767988	767993	4194304	256	0	0	28	24	84928	0x7f5151c37600	0x7f5046036200	12073517754349638	12073517754363645	12073517754612603	12073517754641680
72	70	void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]	0	0	140	767988	767993	4194304	256	0	0	24	24	86400	0x7f5151c37500	0x7f5046036240	12073517754672087	12073517754685882	12073517754811961	12073517754814822
73	71	void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]	0	0	142	767988	767993	4194304	256	0	0	24	24	88128	0x7f5151c37400	0x7f5046036280	12073517754847363	12073517754861400	12073517755010359	12073517755013471
74	72	void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]	0	0	144	767988	767993	4194304	256	0	0	24	24	89600	0x7f5151c37300	0x7f50460362c0	12073517755077841	12073517755090678	12073517755215957	12073517755218953
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	767988	767993	4194304	256	0	0	28	24	92096	0x7f5151c37a00	0x7f5046036300	12073517755252105	12073517755266357	12073517755523634	12073517755547704
76	74	void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]	0	0	148	767988	767993	4194304	256	0	0	28	24	93568	0x7f5151c37900	0x7f5046036340	12073517755581497	12073517755595474	12073517755845391	12073517755868851
77	75	void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]	0	0	150	767988	767993	4194304	256	0	0	24	24	95040	0x7f5151c37800	0x7f5046036380	12073517755897294	12073517755910671	12073517756036270	12073517756039408
78	76	void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]	0	0	152	767988	767993	4194304	256	0	0	24	24	96768	0x7f5151c37700	0x7f50460363c0	12073517756074864	12073517756088429	12073517756245388	12073517756248156
79	77	void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]	0	0	154	767988	767993	4194304	256	0	0	24	24	98240	0x7f5151c37600	0x7f5046036400	12073517756293670	12073517756307467	12073517756432586	12073517756435314
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	767988	767993	4194304	256	0	0	28	24	100736	0x7f5151c37500	0x7f5046036440	12073517756471070	12073517756484266	12073517756746343	12073517756771559
81	79	void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]	0	0	158	767988	767993	4194304	256	0	0	28	24	102208	0x7f5151c37400	0x7f5046036480	12073517756800643	12073517756814183	12073517757059140	12073517757086043
82	80	void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]	0	0	160	767988	767993	4194304	256	0	0	24	24	103680	0x7f5151c37300	0x7f50460364c0	12073517757114786	12073517757128260	12073517757253859	12073517757256580
83	81	void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]	0	0	162	767988	767993	4194304	256	0	0	24	24	105664	0x7f5151c37a00	0x7f5046036500	12073517757291866	12073517757305378	12073517757470497	12073517757473213
84	82	void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]	0	0	164	767988	767993	4194304	256	0	0	24	24	107136	0x7f5151c37900	0x7f5046036540	12073517757515701	12073517757529216	12073517757654815	12073517757657645
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	767988	767993	4194304	256	0	0	28	24	109888	0x7f5151c37800	0x7f5046036580	12073517757689344	12073517757702654	12073517757956252	12073517757980105
86	84	void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]	0	0	168	767988	767993	4194304	256	0	0	28	24	111360	0x7f5151c37700	0x7f50460365c0	12073517758013156	12073517758026492	12073517758274009	12073517758303666
87	85	void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]	0	0	170	767988	767993	4194304	256	0	0	24	24	112832	0x7f5151c37600	0x7f5046036600	12073517758332810	12073517758346329	12073517758472407	12073517758475175
88	86	void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]	0	0	172	767988	767993	4194304	256	0	0	24	24	114816	0x7f5151c37500	0x7f5046036640	12073517758508547	12073517758522167	12073517758732085	12073517758755766
89	87	void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]	0	0	174	767988	767993	4194304	256	0	0	24	24	116288	0x7f5151c37400	0x7f5046036680	12073517758791532	12073517758806484	12073517758932883	12073517758935951
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	767988	767993	4194304	256	0	0	28	24	119040	0x7f5151c37300	0x7f50460366c0	12073517758969413	12073517758982963	12073517759235601	12073517759263990
91	89	void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]	0	0	178	767988	767993	4194304	256	0	0	28	24	120768	0x7f5151c37a00	0x7f5046036700	12073517759297863	12073517759311120	12073517759556878	12073517759581110
92	90	void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]	0	0	180	767988	767993	4194304	256	0	0	24	24	122496	0x7f5151c37900	0x7f5046036740	12073517759608030	12073517759622317	12073517759747916	12073517759750725
93	91	void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]	0	0	182	767988	767993	4194304	256	0	0	24	24	124736	0x7f5151c37800	0x7f5046036780	12073517759783285	12073517759796395	12073517759978954	12073517759981854
94	92	void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]	0	0	184	767988	767993	4194304	256	0	0	24	24	126464	0x7f5151c37700	0x7f50460367c0	12073517760032068	12073517760044873	12073517760171432	12073517760174252
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	767988	767993	4194304	256	0	0	20	24	127936	0x7f5151c37600	0x7f5046036800	12073517760207173	12073517760220552	12073517760483749	12073517760508102
96	94	void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]	0	0	188	767988	767993	4194304	256	0	0	28	24	129664	0x7f5151c37500	0x7f5046036840	12073517760537717	12073517760551269	12073517760800706	12073517760828728
97	95	void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]	0	0	190	767988	767993	4194304	256	0	0	24	24	131392	0x7f5151c37400	0x7f5046036880	12073517760859706	12073517760873506	12073517760998944	12073517761006048
98	96	void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]	0	0	192	767988	767993	4194304	256	0	0	24	24	133632	0x7f5151c37300	0x7f50460368c0	12073517761038258	12073517761051744	12073517761243422	12073517761246234
99	97	void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]	0	0	194	767988	767993	4194304	256	0	0	24	24	135360	0x7f5151c37a00	0x7f5046036900	12073517761288813	12073517761301502	12073517761428861	12073517761431979
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	767988	767993	4194304	256	0	0	20	24	136832	0x7f5151c37900	0x7f5046036940	12073517761464209	12073517761478780	12073517761761018	12073517761785837
101	99	void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]	0	0	198	767988	767993	4194304	256	0	0	28	24	138560	0x7f5151c37800	0x7f5046036980	12073517761816374	12073517761829817	12073517762081495	12073517762109569
102	100	void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]	0	0	200	767988	767993	4194304	256	0	0	24	24	140288	0x7f5151c37700	0x7f50460369c0	12073517762139494	12073517762153654	12073517762280213	12073517762283051
103	101	void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]	0	0	202	767988	767993	4194304	256	0	0	24	24	142784	0x7f5151c37600	0x7f5046036a00	12073517762319058	12073517762332852	12073517762541650	12073517762569123
104	102	void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]	0	0	204	767988	767993	4194304	256	0	0	24	24	144768	0x7f5151c37500	0x7f5046036a40	12073517762605921	12073517762620050	12073517762747409	12073517762750299
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	767988	767993	4194304	256	0	0	20	24	145728	0x7f5151c37400	0x7f5046036a80	12073517762783802	12073517762797968	12073517763103885	12073517763135185
106	104	void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]	0	0	208	767988	767993	4194304	256	0	0	28	24	147712	0x7f5151c37300	0x7f5046036ac0	12073517763169959	12073517763183085	12073517763430922	12073517763455169
107	105	void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]	0	0	210	767988	767993	4194304	256	0	0	24	24	149696	0x7f5151c37a00	0x7f5046036b00	12073517763484754	12073517763498602	12073517763625161	12073517763628351
108	106	void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]	0	0	212	767988	767993	4194304	256	0	0	24	24	152192	0x7f5151c37900	0x7f5046036b40	12073517763659739	12073517763672360	12073517763899718	12073517763923339
109	107	void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]	0	0	214	767988	767993	4194304	256	0	0	24	24	154176	0x7f5151c37800	0x7f5046036b80	12073517763961821	12073517763976677	12073517764104196	12073517764107051
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	767988	767993	4194304	256	0	0	20	24	155648	0x7f5151c37700	0x7f5046036bc0	12073517764140523	12073517764153636	12073517764482273	12073517764505281
111	109	void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]	0	0	218	767988	767993	4194304	256	0	0	28	24	157632	0x7f5151c37600	0x7f5046036c00	12073517764535537	12073517764548672	12073517764799230	12073517764822711
112	110	void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]	0	0	220	767988	767993	4194304	256	0	0	24	24	159616	0x7f5151c37500	0x7f5046036c40	12073517764851865	12073517764866429	12073517764995548	12073517764998447
113	111	void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]	0	0	222	767988	767993	4194304	256	0	0	24	24	162368	0x7f5151c37400	0x7f5046036c80	12073517765037410	12073517765051228	12073517765296025	12073517765322159
114	112	void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]	0	0	224	767988	767993	4194304	256	0	0	24	24	164608	0x7f5151c37300	0x7f5046036cc0	12073517765360540	12073517765375865	12073517765505463	12073517765508235
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	767988	767993	4194304	256	0	0	20	24	166336	0x7f5151c37a00	0x7f5046036d00	12073517765540314	12073517765553943	12073517765929140	12073517765956438
116	114	void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]	0	0	228	767988	767993	4194304	256	0	0	28	24	168576	0x7f5151c37900	0x7f5046036d40	12073517765985522	12073517765998419	12073517766251857	12073517766281021
117	115	void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]	0	0	230	767988	767993	4194304	256	0	0	24	24	170816	0x7f5151c37800	0x7f5046036d80	12073517766311728	12073517766325936	12073517766455215	12073517766458200
118	116	void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]	0	0	232	767988	767993	4194304	256	0	0	24	24	174080	0x7f5151c37700	0x7f5046036dc0	12073517766491953	12073517766505614	12073517766784332	12073517766807499
119	117	void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]	0	0	234	767988	767993	4194304	256	0	0	24	24	176576	0x7f5151c37600	0x7f5046036e00	12073517766846191	12073517766859371	12073517766994730	12073517766997633
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	767988	767993	4194304	256	0	0	20	24	177792	0x7f5151c37500	0x7f5046036e40	12073517767035102	12073517767049449	12073517767472166	12073517767496880
121	119	void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]	0	0	238	767988	767993	4194304	256	0	0	28	24	180288	0x7f5151c37400	0x7f5046036e80	12073517767529581	12073517767542405	12073517767802883	12073517767830651
122	120	void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]	0	0	240	767988	767993	4194304	256	0	0	24	24	182784	0x7f5151c37300	0x7f5046036ec0	12073517767859945	12073517767873922	12073517768009441	12073517768012629
123	121	void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]	0	0	242	767988	767993	4194304	256	0	0	24	24	186304	0x7f5151c37a00	0x7f5046036f00	12073517768045891	12073517768058560	12073517768372157	12073517768399838
124	122	void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]	0	0	244	767988	767993	4194304	256	0	0	24	24	189312	0x7f5151c37900	0x7f5046036f40	12073517768438129	12073517768450077	12073517768601595	12073517768604509
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	767988	767993	4194304	256	0	0	20	24	190784	0x7f5151c37800	0x7f5046036f80	12073517768636578	12073517768648635	12073517769164630	12073517769194185
126	124	void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]	0	0	248	767988	767993	4194304	256	0	0	28	24	193792	0x7f5151c37700	0x7f5046036fc0	12073517769228879	12073517769243349	12073517769542867	12073517769570935
127	125	void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]	0	0	250	767988	767993	4194304	256	0	0	24	24	196800	0x7f5151c37600	0x7f5046037000	12073517769599458	12073517769613106	12073517769766705	12073517769769634
128	126	void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]	0	0	252	767988	767993	4194304	256	0	0	24	24	201088	0x7f5151c37500	0x7f5046037040	12073517769803587	12073517769816944	12073517770201581	12073517770234167
129	127	void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]	0	0	254	767988	767993	4194304	256	0	0	24	24	204608	0x7f5151c37400	0x7f5046037080	12073517770276967	12073517770289100	12073517770503658	12073517770531119
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	767988	767993	4194304	256	0	0	20	24	206080	0x7f5151c37300	0x7f50460370c0	12073517770563159	12073517770577897	12073517771199492	12073517771228525
131	129	void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]	0	0	258	767988	767993	4194304	256	0	0	28	24	209600	0x7f5151c37a00	0x7f5046037100	12073517771260474	12073517771274211	12073517771600608	12073517771627577
132	130	void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]	0	0	260	767988	767993	4194304	256	0	0	24	24	213120	0x7f5151c37900	0x7f5046037140	12073517771659927	12073517771673567	12073517771851646	12073517771854458
133	131	void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]	0	0	262	767988	767993	4194304	256	0	0	24	24	218176	0x7f5151c37800	0x7f5046037180	12073517771891457	12073517771904605	12073517772359001	12073517772391787
134	132	void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]	0	0	264	767988	767993	4194304	256	0	0	24	24	222208	0x7f5151c37700	0x7f50460371c0	12073517772435177	12073517772449240	12073517772648758	12073517772675795
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	767988	767993	4194304	256	0	0	20	24	223936	0x7f5151c37600	0x7f5046037200	12073517772706081	12073517772719958	12073517773422671	12073517773451667
136	134	void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]	0	0	268	767988	767993	4194304	256	0	0	28	24	227968	0x7f5151c37500	0x7f5046037240	12073517773485509	12073517773499311	12073517773872107	12073517773898928
137	135	void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]	0	0	270	767988	767993	4194304	256	0	0	24	24	232000	0x7f5151c37400	0x7f5046037280	12073517773931298	12073517773944587	12073517774146665	12073517774175241
138	136	void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]	0	0	272	767988	767993	4194304	256	0	0	24	24	237824	0x7f5151c37300	0x7f50460372c0	12073517774204706	12073517774217864	12073517774743939	12073517774771479
139	137	void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]	0	0	274	767988	767993	4194304	256	0	0	24	24	242368	0x7f5151c37a00	0x7f5046037300	12073517774810352	12073517774824259	12073517775046497	12073517775076967
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	767988	767993	4194304	256	0	0	20	24	243584	0x7f5151c37900	0x7f5046037340	12073517775104328	12073517775116096	12073517775914649	12073517775942045
141	139	void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]	0	0	278	767988	767993	4194304	256	0	0	28	24	248128	0x7f5151c37800	0x7f5046037380	12073517775971560	12073517775985208	12073517776406004	12073517776437716
142	140	void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]	0	0	280	767988	767993	4194304	256	0	0	24	24	252672	0x7f5151c37700	0x7f50460373c0	12073517776466249	12073517776480084	12073517776705042	12073517776732153
143	141	void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]	0	0	282	767988	767993	4194304	256	0	0	24	24	259264	0x7f5151c37600	0x7f5046037400	12073517776762289	12073517776774801	12073517777383435	12073517777411586
144	142	void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]	0	0	284	767988	767993	4194304	256	0	0	24	24	264832	0x7f5151c37500	0x7f5046037440	12073517777449536	12073517777462315	12073517777731272	12073517777758971
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	767988	767993	4194304	256	0	0	20	24	266304	0x7f5151c37400	0x7f5046037480	12073517777787164	12073517777801832	12073517778786463	12073517778814703
146	144	void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]	0	0	288	767988	767993	4194304	256	0	0	28	24	271872	0x7f5151c37300	0x7f50460374c0	12073517778844659	12073517778858142	12073517779372217	12073517779403257
147	145	void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]	0	0	290	767988	767993	4194304	256	0	0	24	24	277440	0x7f5151c37a00	0x7f5046037500	12073517779433774	12073517779447577	12073517779718294	12073517779745493
148	146	void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]	0	0	292	767988	767993	4194304	256	0	0	12	24	279936	0x7f5151c37900	0x7f5046037540	12073517779776050	12073517779789174	12073517780526287	12073517780554126
149	147	void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]	0	0	294	767988	767993	4194304	256	0	0	24	24	286528	0x7f5151c37800	0x7f5046037580	12073517780594851	12073517780606286	12073517780921483	12073517780945653
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	767988	767993	4194304	256	0	0	20	24	288000	0x7f5151c37700	0x7f50460375c0	12073517780975138	12073517780988683	12073517782159872	12073517782192009
151	149	void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]	0	0	298	767988	767993	4194304	256	0	0	32	24	294592	0x7f5151c37600	0x7f5046037600	12073517782232034	12073517782245951	12073517782853946	12073517782883364
152	150	void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]	0	0	300	767988	767993	4194304	256	0	0	24	24	301184	0x7f5151c37500	0x7f5046037640	12073517782919602	12073517782932825	12073517783251542	12073517783284410
153	151	void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]	0	0	302	767988	767993	4194304	256	0	0	12	24	304960	0x7f5151c37400	0x7f5046037680	12073517783315928	12073517783328821	12073517784205613	12073517784236469
154	152	void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]	0	0	304	767988	767993	4194304	256	0	0	24	24	313600	0x7f5151c37300	0x7f50460376c0	12073517784277505	12073517784292533	12073517784704850	12073517784737961
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	767988	767993	4194304	256	0	0	20	24	314816	0x7f5151c37a00	0x7f5046037700	12073517784768568	12073517784783088	12073517786331876	12073517786405049
156	154	void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]	0	0	308	767988	767993	4194304	256	0	0	20	24	317568	0x7f5151c37900	0x7f5046037740	12073517786438311	12073517786452675	12073517787246589	12073517787320129
157	155	void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]	0	0	310	767988	767993	4194304	256	0	0	12	24	320320	0x7f5151c37800	0x7f5046037780	12073517787356307	12073517787370108	12073517787784024	12073517787857107
158	156	void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]	0	0	312	767988	767993	4194304	256	0	0	12	24	324096	0x7f5151c37700	0x7f50460377c0	12073517787888295	12073517787902104	12073517789060654	12073517789136234
159	157	void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]	0	0	314	767988	767993	4194304	256	0	0	12	24	328896	0x7f5151c37600	0x7f5046037800	12073517789175648	12073517789189133	12073517789974247	12073517790047859
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	767988	767993	4194304	256	0	0	20	24	330112	0x7f5151c37500	0x7f5046037840	12073517790080990	12073517790094406	12073517793145742	12073517793219853
161	159	void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]	0	0	318	767988	767993	4194304	256	0	0	20	24	332864	0x7f5151c37400	0x7f5046037880	12073517793252735	12073517793266861	12073517794808129	12073517794881181
162	160	void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]	0	0	320	767988	767993	4194304	256	0	0	12	24	335616	0x7f5151c37300	0x7f50460378c0	12073517794911788	12073517794926208	12073517795715322	12073517795790240
163	161	void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]	0	0	322	767988	767993	4194304	256	0	0	12	24	339392	0x7f5151c37a00	0x7f5046037900	12073517795820607	12073517795834841	12073517798116263	12073517798190260
164	162	void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]	0	0	324	767988	767993	4194304	256	0	0	12	24	344192	0x7f5151c37900	0x7f5046037940	12073517798229233	12073517798243942	12073517799779450	12073517799852920
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	767988	767993	4194304	256	0	0	20	24	345408	0x7f5151c37800	0x7f5046037980	12073517799888556	12073517799902489	12073517805955881	12073517806031342
166	164	void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]	0	0	328	767988	767993	4194304	256	0	0	20	24	348160	0x7f5151c37700	0x7f50460379c0	12073517806067910	12073517806082280	12073517809125136	12073517809198748
167	165	void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]	0	0	330	767988	767993	4194304	256	0	0	12	24	350912	0x7f5151c37600	0x7f5046037a00	12073517809236018	12073517809249135	12073517810788803	12073517810862300
168	166	void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]	0	0	332	767988	767993	4194304	256	0	0	12	24	0	0x7f5151c37500	0x7f5046037a40	12073517810895712	12073517810909442	12073517815436447	12073517815512001