rocm-systems/tests/workloads/no_roof_TA/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	859514	859519	33554432	256	0	0	8	32	6464	0x0	0x7f624e604180	12075053930622059	12075053930670524	12075053930994841	12075053931103644
3	1	void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]	0	0	2	859514	859519	32768	256	0	0	24	24	12480	0x0	0x7f624e635100	12075053946363703	12075053946463595	12075053946470315	12075053946476022
4	2	void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]	0	0	4	859514	859519	4194304	256	0	0	24	24	12928	0x7f635a251900	0x7f624e635140	12075053946523830	12075053946537675	12075053946667114	12075053946670623
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	859514	859519	4194304	256	0	0	36	24	13632	0x7f635a251800	0x7f624e635180	12075053946715106	12075053946727593	12075053946972551	12075053947071137
6	4	void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]	0	0	8	859514	859519	4194304	256	0	0	28	24	14080	0x7f635a251700	0x7f624e6351c0	12075053947097937	12075053947109990	12075053947359268	12075053947457435
7	5	void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]	0	0	10	859514	859519	4194304	256	0	0	24	24	14528	0x7f635a251600	0x7f624e635200	12075053947480999	12075053947492867	12075053947618466	12075053947621821
8	6	void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]	0	0	12	859514	859519	4194304	256	0	0	24	24	14976	0x7f635a251500	0x7f624e635240	12075053947648901	12075053947660225	12075053947785344	12075053947788711
9	7	void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]	0	0	14	859514	859519	4194304	256	0	0	24	24	15424	0x7f635a251400	0x7f624e635280	12075053947839776	12075053947852704	12075053947976703	12075053947979796
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	859514	859519	4194304	256	0	0	28	24	16128	0x7f635a251300	0x7f624e6352c0	12075053948013038	12075053948025183	12075053948269661	12075053948371534
11	9	void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]	0	0	18	859514	859519	4194304	256	0	0	28	24	16832	0x7f635a251a00	0x7f624e635300	12075053948393174	12075053948404859	12075053948648057	12075053948745018
12	10	void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]	0	0	20	859514	859519	4194304	256	0	0	24	24	17536	0x7f635a251900	0x7f624e635340	12075053948767831	12075053948780216	12075053948905335	12075053948908893
13	11	void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]	0	0	22	859514	859519	4194304	256	0	0	24	24	18240	0x7f635a251800	0x7f624e635380	12075053948935232	12075053948946935	12075053949071734	12075053949075482
14	12	void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]	0	0	24	859514	859519	4194304	256	0	0	24	24	18944	0x7f635a251700	0x7f624e6353c0	12075053949115777	12075053949127414	12075053949251733	12075053949254976
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	859514	859519	4194304	256	0	0	28	24	19904	0x7f635a251600	0x7f624e635400	12075053949281495	12075053949293652	12075053949544530	12075053949618572
16	14	void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]	0	0	28	859514	859519	4194304	256	0	0	28	24	20608	0x7f635a251500	0x7f624e635440	12075053949644590	12075053949656049	12075053949938127	12075053950012113
17	15	void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]	0	0	30	859514	859519	4194304	256	0	0	24	24	21312	0x7f635a251400	0x7f624e635480	12075053950039494	12075053950052206	12075053950177165	12075053950180786
18	16	void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]	0	0	32	859514	859519	4194304	256	0	0	24	24	22016	0x7f635a251300	0x7f624e6354c0	12075053950206674	12075053950220205	12075053950345644	12075053950348829
19	17	void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]	0	0	34	859514	859519	4194304	256	0	0	24	24	22720	0x7f635a251a00	0x7f624e635500	12075053950387160	12075053950402123	12075053950526282	12075053950529625
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	859514	859519	4194304	256	0	0	28	24	23680	0x7f635a251900	0x7f624e635540	12075053950556154	12075053950569162	12075053950819080	12075053950861952
21	19	void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]	0	0	38	859514	859519	4194304	256	0	0	28	24	24384	0x7f635a251800	0x7f624e635580	12075053950887870	12075053950901319	12075053951147397	12075053951245475
22	20	void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]	0	0	40	859514	859519	4194304	256	0	0	24	24	25088	0x7f635a251700	0x7f624e6355c0	12075053951269720	12075053951281316	12075053951406595	12075053951409910
23	21	void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]	0	0	42	859514	859519	4194304	256	0	0	24	24	25792	0x7f635a251600	0x7f624e635600	12075053951436580	12075053951448035	12075053951573314	12075053951576640
24	22	void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]	0	0	44	859514	859519	4194304	256	0	0	24	24	26496	0x7f635a251500	0x7f624e635640	12075053951613559	12075053951625953	12075053951749952	12075053951753098
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	859514	859519	4194304	256	0	0	28	24	27712	0x7f635a251400	0x7f624e635680	12075053951779878	12075053951793152	12075053952040670	12075053952139196
26	24	void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]	0	0	48	859514	859519	4194304	256	0	0	28	24	28416	0x7f635a251300	0x7f624e6356c0	12075053952161106	12075053952171869	12075053952416187	12075053952513571
27	25	void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]	0	0	50	859514	859519	4194304	256	0	0	24	24	29120	0x7f635a251a00	0x7f624e635700	12075053952534841	12075053952545786	12075053952669785	12075053952673228
28	26	void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]	0	0	52	859514	859519	4194304	256	0	0	24	24	30080	0x7f635a251900	0x7f624e635740	12075053952699146	12075053952710425	12075053952835544	12075053952839207
29	27	void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]	0	0	54	859514	859519	4194304	256	0	0	24	24	30784	0x7f635a251800	0x7f624e635780	12075053952875204	12075053952886743	12075053953011062	12075053953014322
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	859514	859519	4194304	256	0	0	28	24	32000	0x7f635a251700	0x7f624e6357c0	12075053953041342	12075053953053462	12075053953297940	12075053953397644
31	29	void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]	0	0	58	859514	859519	4194304	256	0	0	28	24	32960	0x7f635a251600	0x7f624e635800	12075053953419796	12075053953431219	12075053953679217	12075053953776879
32	30	void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]	0	0	60	859514	859519	4194304	256	0	0	24	24	33920	0x7f635a251500	0x7f624e635840	12075053953798239	12075053953809616	12075053953933935	12075053953937307
33	31	void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]	0	0	62	859514	859519	4194304	256	0	0	24	24	34880	0x7f635a251400	0x7f624e635880	12075053953963646	12075053953975054	12075053954100173	12075053954103847
34	32	void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]	0	0	64	859514	859519	4194304	256	0	0	24	24	35840	0x7f635a251300	0x7f624e6358c0	12075053954140515	12075053954155533	12075053954280172	12075053954283691
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	859514	859519	4194304	256	0	0	28	24	37312	0x7f635a251a00	0x7f624e635900	12075053954309789	12075053954322892	12075053954567690	12075053954621268
36	34	void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]	0	0	68	859514	859519	4194304	256	0	0	28	24	38272	0x7f635a251900	0x7f624e635940	12075053954641766	12075053954661289	12075053954906727	12075053954985165
37	35	void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]	0	0	70	859514	859519	4194304	256	0	0	24	24	39232	0x7f635a251800	0x7f624e635980	12075053955010922	12075053955022246	12075053955147365	12075053955150702
38	36	void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]	0	0	72	859514	859519	4194304	256	0	0	24	24	40192	0x7f635a251700	0x7f624e6359c0	12075053955179456	12075053955191205	12075053955316803	12075053955320067
39	37	void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]	0	0	74	859514	859519	4194304	256	0	0	24	24	41152	0x7f635a251600	0x7f624e635a00	12075053955355894	12075053955370083	12075053955495042	12075053955498559
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	859514	859519	4194304	256	0	0	28	24	42624	0x7f635a251500	0x7f624e635a40	12075053955525739	12075053955537442	12075053955784160	12075053955813784
41	39	void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]	0	0	78	859514	859519	4194304	256	0	0	28	24	43584	0x7f635a251400	0x7f624e635a80	12075053955835234	12075053955846559	12075053956092637	12075053956144539
42	40	void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]	0	0	80	859514	859519	4194304	256	0	0	24	24	44544	0x7f635a251300	0x7f624e635ac0	12075053956165789	12075053956177597	12075053956302876	12075053956305789
43	41	void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]	0	0	82	859514	859519	4194304	256	0	0	24	24	45760	0x7f635a251a00	0x7f624e635b00	12075053956333160	12075053956344475	12075053956470234	12075053956473721
44	42	void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]	0	0	84	859514	859519	4194304	256	0	0	24	24	46720	0x7f635a251900	0x7f624e635b40	12075053956514697	12075053956527034	12075053956651993	12075053956654958
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	859514	859519	4194304	256	0	0	28	24	48448	0x7f635a251800	0x7f624e635b80	12075053956682359	12075053956693752	12075053956947990	12075053956970674
46	44	void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]	0	0	88	859514	859519	4194304	256	0	0	28	24	49408	0x7f635a251700	0x7f624e635bc0	12075053956994689	12075053957006230	12075053957293747	12075053957349007
47	45	void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]	0	0	90	859514	859519	4194304	256	0	0	24	24	50368	0x7f635a251600	0x7f624e635c00	12075053957368814	12075053957380307	12075053957505106	12075053957508955
48	46	void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]	0	0	92	859514	859519	4194304	256	0	0	24	24	51584	0x7f635a251500	0x7f624e635c40	12075053957535304	12075053957549265	12075053957675984	12075053957679131
49	47	void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]	0	0	94	859514	859519	4194304	256	0	0	24	24	52544	0x7f635a251400	0x7f624e635c80	12075053957715889	12075053957728784	12075053957854063	12075053957857132
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	859514	859519	4194304	256	0	0	28	24	54272	0x7f635a251300	0x7f624e635cc0	12075053957884793	12075053957897903	12075053958159180	12075053958189239
51	49	void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]	0	0	98	859514	859519	4194304	256	0	0	28	24	55488	0x7f635a251a00	0x7f624e635d00	12075053958212151	12075053958224620	12075053958471338	12075053958496179
52	50	void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]	0	0	100	859514	859519	4194304	256	0	0	24	24	56704	0x7f635a251900	0x7f624e635d40	12075053958517148	12075053958529737	12075053958654856	12075053958657960
53	51	void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]	0	0	102	859514	859519	4194304	256	0	0	24	24	58176	0x7f635a251800	0x7f624e635d80	12075053958685742	12075053958699176	12075053958827815	12075053958830921
54	52	void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]	0	0	104	859514	859519	4194304	256	0	0	24	24	59392	0x7f635a251700	0x7f624e635dc0	12075053958868461	12075053958880935	12075053959006214	12075053959009533
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	859514	859519	4194304	256	0	0	28	24	61376	0x7f635a251600	0x7f624e635e00	12075053959036774	12075053959050053	12075053959312611	12075053959337813
56	54	void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]	0	0	108	859514	859519	4194304	256	0	0	28	24	62592	0x7f635a251500	0x7f624e635e40	12075053959359784	12075053959371011	12075053959619649	12075053959648371
57	55	void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]	0	0	110	859514	859519	4194304	256	0	0	24	24	63808	0x7f635a251400	0x7f624e635e80	12075053959670131	12075053959682688	12075053959807967	12075053959811063
58	56	void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]	0	0	112	859514	859519	4194304	256	0	0	24	24	65280	0x7f635a251300	0x7f624e635ec0	12075053959838524	12075053959851487	12075053959981406	12075053959984605
59	57	void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]	0	0	114	859514	859519	4194304	256	0	0	24	24	66496	0x7f635a251a00	0x7f624e635f00	12075053960026844	12075053960038045	12075053960163164	12075053960166403
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	859514	859519	4194304	256	0	0	28	24	68480	0x7f635a251900	0x7f624e635f40	12075053960193413	12075053960204604	12075053960472602	12075053960497749
61	59	void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]	0	0	118	859514	859519	4194304	256	0	0	28	24	69696	0x7f635a251800	0x7f624e635f80	12075053960523246	12075053960535321	12075053960779319	12075053960799259
62	60	void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]	0	0	120	859514	859519	4194304	256	0	0	24	24	70912	0x7f635a251700	0x7f624e635fc0	12075053960822172	12075053960834039	12075053960959478	12075053960962543
63	61	void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]	0	0	122	859514	859519	4194304	256	0	0	24	24	72384	0x7f635a251600	0x7f624e636000	12075053960988631	12075053961001877	12075053961138036	12075053961141405
64	62	void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]	0	0	124	859514	859519	4194304	256	0	0	24	24	73600	0x7f635a251500	0x7f624e636040	12075053961179075	12075053961190676	12075053961315475	12075053961318675
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	859514	859519	4194304	256	0	0	28	24	75840	0x7f635a251400	0x7f624e636080	12075053961345174	12075053961356275	12075053961608433	12075053961632127
66	64	void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]	0	0	128	859514	859519	4194304	256	0	0	28	24	77056	0x7f635a251300	0x7f624e6360c0	12075053961653066	12075053961665072	12075053961911310	12075053961939128
67	65	void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]	0	0	130	859514	859519	4194304	256	0	0	24	24	78272	0x7f635a251a00	0x7f624e636100	12075053961960567	12075053961971630	12075053962096589	12075053962099816
68	66	void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]	0	0	132	859514	859519	4194304	256	0	0	24	24	80000	0x7f635a251900	0x7f624e636140	12075053962129551	12075053962140748	12075053962283467	12075053962286593
69	67	void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]	0	0	134	859514	859519	4194304	256	0	0	24	24	81216	0x7f635a251800	0x7f624e636180	12075053962323702	12075053962335147	12075053962460746	12075053962463793
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	859514	859519	4194304	256	0	0	28	24	83456	0x7f635a251700	0x7f624e6361c0	12075053962490031	12075053962501545	12075053962755303	12075053962783487
71	69	void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]	0	0	138	859514	859519	4194304	256	0	0	28	24	84928	0x7f635a251600	0x7f624e636200	12075053962804496	12075053962817543	12075053963099940	12075053963128017
72	70	void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]	0	0	140	859514	859519	4194304	256	0	0	24	24	86400	0x7f635a251500	0x7f624e636240	12075053963149687	12075053963162020	12075053963287619	12075053963291230
73	71	void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]	0	0	142	859514	859519	4194304	256	0	0	24	24	88128	0x7f635a251400	0x7f624e636280	12075053963318060	12075053963330179	12075053963478657	12075053963481684
74	72	void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]	0	0	144	859514	859519	4194304	256	0	0	24	24	89600	0x7f635a251300	0x7f624e6362c0	12075053963532549	12075053963543937	12075053963669536	12075053963672759
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	859514	859519	4194304	256	0	0	28	24	92096	0x7f635a251a00	0x7f624e636300	12075053963699810	12075053963712096	12075053963968893	12075053963992714
76	74	void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]	0	0	148	859514	859519	4194304	256	0	0	28	24	93568	0x7f635a251900	0x7f624e636340	12075053964021057	12075053964034173	12075053964279291	12075053964306657
77	75	void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]	0	0	150	859514	859519	4194304	256	0	0	24	24	95040	0x7f635a251800	0x7f624e636380	12075053964328879	12075053964341850	12075053964467289	12075053964470482
78	76	void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]	0	0	152	859514	859519	4194304	256	0	0	24	24	96768	0x7f635a251700	0x7f624e6363c0	12075053964496881	12075053964509689	12075053964666968	12075053964670083
79	77	void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]	0	0	154	859514	859519	4194304	256	0	0	24	24	98240	0x7f635a251600	0x7f624e636400	12075053964706781	12075053964717847	12075053964842646	12075053964845669
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	859514	859519	4194304	256	0	0	28	24	100736	0x7f635a251500	0x7f624e636440	12075053964872839	12075053964884406	12075053965143764	12075053965172687
81	79	void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]	0	0	158	859514	859519	4194304	256	0	0	28	24	102208	0x7f635a251400	0x7f624e636480	12075053965194527	12075053965205683	12075053965451761	12075053965476782
82	80	void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]	0	0	160	859514	859519	4194304	256	0	0	24	24	103680	0x7f635a251300	0x7f624e6364c0	12075053965498462	12075053965511441	12075053965637360	12075053965640506
83	81	void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]	0	0	162	859514	859519	4194304	256	0	0	24	24	105664	0x7f635a251a00	0x7f624e636500	12075053965669520	12075053965682800	12075053965847598	12075053965850757
84	82	void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]	0	0	164	859514	859519	4194304	256	0	0	24	24	107136	0x7f635a251900	0x7f624e636540	12075053965888126	12075053965899278	12075053966024237	12075053966027615
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	859514	859519	4194304	256	0	0	28	24	109888	0x7f635a251800	0x7f624e636580	12075053966055217	12075053966066476	12075053966326954	12075053966354693
86	84	void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]	0	0	168	859514	859519	4194304	256	0	0	28	24	111360	0x7f635a251700	0x7f624e6365c0	12075053966376674	12075053966387594	12075053966636072	12075053966664289
87	85	void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]	0	0	170	859514	859519	4194304	256	0	0	24	24	112832	0x7f635a251600	0x7f624e636600	12075053966685749	12075053966698631	12075053966824230	12075053966827372
88	86	void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]	0	0	172	859514	859519	4194304	256	0	0	24	24	114816	0x7f635a251500	0x7f624e636640	12075053966854031	12075053966867750	12075053967040069	12075053967043373
89	87	void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]	0	0	174	859514	859519	4194304	256	0	0	24	24	116288	0x7f635a251400	0x7f624e636680	12075053967081624	12075053967094308	12075053967220387	12075053967223798
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	859514	859519	4194304	256	0	0	28	24	119040	0x7f635a251300	0x7f624e6366c0	12075053967250538	12075053967261827	12075053967514465	12075053967538684
91	89	void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]	0	0	178	859514	859519	4194304	256	0	0	28	24	120768	0x7f635a251a00	0x7f624e636700	12075053967561616	12075053967573184	12075053967822142	12075053967845894
92	90	void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]	0	0	180	859514	859519	4194304	256	0	0	24	24	122496	0x7f635a251900	0x7f624e636740	12075053967867445	12075053967878782	12075053968004701	12075053968008086
93	91	void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]	0	0	182	859514	859519	4194304	256	0	0	24	24	124736	0x7f635a251800	0x7f624e636780	12075053968035196	12075053968046300	12075053968227579	12075053968230860
94	92	void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]	0	0	184	859514	859519	4194304	256	0	0	24	24	126464	0x7f635a251700	0x7f624e6367c0	12075053968268279	12075053968279739	12075053968406138	12075053968409251
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	859514	859519	4194304	256	0	0	20	24	127936	0x7f635a251600	0x7f624e636800	12075053968435921	12075053968446937	12075053968708535	12075053968732853
96	94	void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]	0	0	188	859514	859519	4194304	256	0	0	28	24	129664	0x7f635a251500	0x7f624e636840	12075053968753722	12075053968765335	12075053969012053	12075053969037248
97	95	void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]	0	0	190	859514	859519	4194304	256	0	0	24	24	131392	0x7f635a251400	0x7f624e636880	12075053969059420	12075053969070292	12075053969196371	12075053969199650
98	96	void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]	0	0	192	859514	859519	4194304	256	0	0	24	24	133632	0x7f635a251300	0x7f624e6368c0	12075053969229546	12075053969241171	12075053969431889	12075053969435208
99	97	void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]	0	0	194	859514	859519	4194304	256	0	0	24	24	135360	0x7f635a251a00	0x7f624e636900	12075053969472387	12075053969485169	12075053969611088	12075053969614631
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	859514	859519	4194304	256	0	0	20	24	136832	0x7f635a251900	0x7f624e636940	12075053969641742	12075053969653327	12075053969935565	12075053969959643
101	99	void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]	0	0	198	859514	859519	4194304	256	0	0	28	24	138560	0x7f635a251800	0x7f624e636980	12075053969981463	12075053969992685	12075053970238603	12075053970270791
102	100	void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]	0	0	200	859514	859519	4194304	256	0	0	24	24	140288	0x7f635a251700	0x7f624e6369c0	12075053970291409	12075053970302762	12075053970429961	12075053970433423
103	101	void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]	0	0	202	859514	859519	4194304	256	0	0	24	24	142784	0x7f635a251600	0x7f624e636a00	12075053970460513	12075053970472521	12075053970680999	12075053970708785
104	102	void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]	0	0	204	859514	859519	4194304	256	0	0	24	24	144768	0x7f635a251500	0x7f624e636a40	12075053970740544	12075053970751719	12075053970878758	12075053970881937
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	859514	859519	4194304	256	0	0	20	24	145728	0x7f635a251400	0x7f624e636a80	12075053970908686	12075053970922437	12075053971228355	12075053971257665
106	104	void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]	0	0	208	859514	859519	4194304	256	0	0	28	24	147712	0x7f635a251300	0x7f624e636ac0	12075053971282471	12075053971296034	12075053971545792	12075053971569294
107	105	void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]	0	0	210	859514	859519	4194304	256	0	0	24	24	149696	0x7f635a251a00	0x7f624e636b00	12075053971590383	12075053971603392	12075053971730271	12075053971733449
108	106	void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]	0	0	212	859514	859519	4194304	256	0	0	24	24	152192	0x7f635a251900	0x7f624e636b40	12075053971760549	12075053971771870	12075053971999228	12075053972027305
109	107	void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]	0	0	214	859514	859519	4194304	256	0	0	24	24	154176	0x7f635a251800	0x7f624e636b80	12075053972059755	12075053972072668	12075053972200827	12075053972204274
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	859514	859519	4194304	256	0	0	20	24	155648	0x7f635a251700	0x7f624e636bc0	12075053972230533	12075053972241787	12075053972569944	12075053972595421
111	109	void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]	0	0	218	859514	859519	4194304	256	0	0	28	24	157632	0x7f635a251600	0x7f624e636c00	12075053972617552	12075053972628663	12075053972881301	12075053972909545
112	110	void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]	0	0	220	859514	859519	4194304	256	0	0	24	24	159616	0x7f635a251500	0x7f624e636c40	12075053972931085	12075053972943861	12075053973071540	12075053973075042
113	111	void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]	0	0	222	859514	859519	4194304	256	0	0	24	24	162368	0x7f635a251400	0x7f624e636c80	12075053973104226	12075053973115539	12075053973358897	12075053973382894
114	112	void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]	0	0	224	859514	859519	4194304	256	0	0	24	24	164608	0x7f635a251300	0x7f624e636cc0	12075053973414743	12075053973426577	12075053973556496	12075053973559773
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	859514	859519	4194304	256	0	0	20	24	166336	0x7f635a251a00	0x7f624e636d00	12075053973585942	12075053973596976	12075053973971373	12075053973999330
116	114	void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]	0	0	228	859514	859519	4194304	256	0	0	28	24	168576	0x7f635a251900	0x7f624e636d40	12075053974025198	12075053974038092	12075053974298250	12075053974329513
117	115	void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]	0	0	230	859514	859519	4194304	256	0	0	24	24	170816	0x7f635a251800	0x7f624e636d80	12075053974351053	12075053974363849	12075053974492168	12075053974495442
118	116	void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]	0	0	232	859514	859519	4194304	256	0	0	24	24	174080	0x7f635a251700	0x7f624e636dc0	12075053974521741	12075053974533128	12075053974811206	12075053974834702
119	117	void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]	0	0	234	859514	859519	4194304	256	0	0	24	24	176576	0x7f635a251600	0x7f624e636e00	12075053974866612	12075053974877925	12075053975012484	12075053975015869
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	859514	859519	4194304	256	0	0	20	24	177792	0x7f635a251500	0x7f624e636e40	12075053975044081	12075053975056164	12075053975477440	12075053975500730
121	119	void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]	0	0	238	859514	859519	4194304	256	0	0	28	24	180288	0x7f635a251400	0x7f624e636e80	12075053975524514	12075053975536000	12075053975801598	12075053975824832
122	120	void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]	0	0	240	859514	859519	4194304	256	0	0	24	24	182784	0x7f635a251300	0x7f624e636ec0	12075053975847054	12075053975858237	12075053975992796	12075053975996090
123	121	void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]	0	0	242	859514	859519	4194304	256	0	0	24	24	186304	0x7f635a251a00	0x7f624e636f00	12075053976026347	12075053976037596	12075053976350873	12075053976378040
124	122	void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]	0	0	244	859514	859519	4194304	256	0	0	24	24	189312	0x7f635a251900	0x7f624e636f40	12075053976410601	12075053976423673	12075053976574871	12075053976578172
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	859514	859519	4194304	256	0	0	20	24	190784	0x7f635a251800	0x7f624e636f80	12075053976606465	12075053976617591	12075053977133267	12075053977161486
126	124	void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]	0	0	248	859514	859519	4194304	256	0	0	28	24	193792	0x7f635a251700	0x7f624e636fc0	12075053977185321	12075053977196626	12075053977493904	12075053977520784
127	125	void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]	0	0	250	859514	859519	4194304	256	0	0	24	24	196800	0x7f635a251600	0x7f624e637000	12075053977543947	12075053977554863	12075053977708622	12075053977711919
128	126	void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]	0	0	252	859514	859519	4194304	256	0	0	24	24	201088	0x7f635a251500	0x7f624e637040	12075053977741173	12075053977754222	12075053978138699	12075053978166865
129	127	void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]	0	0	254	859514	859519	4194304	256	0	0	24	24	204608	0x7f635a251400	0x7f624e637080	12075053978200257	12075053978213898	12075053978389737	12075053978393105
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	859514	859519	4194304	256	0	0	20	24	206080	0x7f635a251300	0x7f624e6370c0	12075053978419935	12075053978431016	12075053979039811	12075053979067017
131	129	void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]	0	0	258	859514	859519	4194304	256	0	0	28	24	209600	0x7f635a251a00	0x7f624e637100	12075053979090822	12075053979103331	12075053979432128	12075053979458034
132	130	void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]	0	0	260	859514	859519	4194304	256	0	0	24	24	213120	0x7f635a251900	0x7f624e637140	12075053979480826	12075053979491648	12075053979669886	12075053979673144
133	131	void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]	0	0	262	859514	859519	4194304	256	0	0	24	24	218176	0x7f635a251800	0x7f624e637180	12075053979699533	12075053979710526	12075053980165082	12075053980192238
134	132	void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]	0	0	264	859514	859519	4194304	256	0	0	24	24	222208	0x7f635a251700	0x7f624e6371c0	12075053980225911	12075053980237082	12075053980435480	12075053980461379
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	859514	859519	4194304	256	0	0	20	24	223936	0x7f635a251600	0x7f624e637200	12075053980481857	12075053980494520	12075053981195474	12075053981223475
136	134	void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]	0	0	268	859514	859519	4194304	256	0	0	28	24	227968	0x7f635a251500	0x7f624e637240	12075053981247119	12075053981258674	12075053981632750	12075053981660157
137	135	void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]	0	0	270	859514	859519	4194304	256	0	0	24	24	232000	0x7f635a251400	0x7f624e637280	12075053981683059	12075053981695470	12075053981897068	12075053981922885
138	136	void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]	0	0	272	859514	859519	4194304	256	0	0	24	24	237824	0x7f635a251300	0x7f624e6372c0	12075053981944285	12075053981955628	12075053982480104	12075053982513883
139	137	void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]	0	0	274	859514	859519	4194304	256	0	0	24	24	242368	0x7f635a251a00	0x7f624e637300	12075053982547045	12075053982561223	12075053982784101	12075053982806497
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	859514	859519	4194304	256	0	0	20	24	243584	0x7f635a251900	0x7f624e637340	12075053982829970	12075053982841381	12075053983638494	12075053983666465
141	139	void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]	0	0	278	859514	859519	4194304	256	0	0	28	24	248128	0x7f635a251800	0x7f624e637380	12075053983689077	12075053983700094	12075053984117210	12075053984144383
142	140	void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]	0	0	280	859514	859519	4194304	256	0	0	24	24	252672	0x7f635a251700	0x7f624e6373c0	12075053984167566	12075053984180570	12075053984406008	12075053984432168
143	141	void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]	0	0	282	859514	859519	4194304	256	0	0	24	24	259264	0x7f635a251600	0x7f624e637400	12075053984456643	12075053984467768	12075053985063443	12075053985091183
144	142	void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]	0	0	284	859514	859519	4194304	256	0	0	24	24	264832	0x7f635a251500	0x7f624e637440	12075053985124194	12075053985135122	12075053985403120	12075053985429652
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	859514	859519	4194304	256	0	0	20	24	266304	0x7f635a251400	0x7f624e637480	12075053985452644	12075053985463759	12075053986447112	12075053986474954
146	144	void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]	0	0	288	859514	859519	4194304	256	0	0	28	24	271872	0x7f635a251300	0x7f624e6374c0	12075053986496404	12075053986507271	12075053987020707	12075053987048921
147	145	void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]	0	0	290	859514	859519	4194304	256	0	0	24	24	277440	0x7f635a251a00	0x7f624e637500	12075053987070822	12075053987084226	12075053987355584	12075053987381880
148	146	void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]	0	0	292	859514	859519	4194304	256	0	0	12	24	279936	0x7f635a251900	0x7f624e637540	12075053987404832	12075053987416064	12075053988152538	12075053988180293
149	147	void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]	0	0	294	859514	859519	4194304	256	0	0	24	24	286528	0x7f635a251800	0x7f624e637580	12075053988213545	12075053988224697	12075053988539255	12075053988566561
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	859514	859519	4194304	256	0	0	20	24	288000	0x7f635a251700	0x7f624e6375c0	12075053988589083	12075053988600214	12075053989768525	12075053989795846
151	149	void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]	0	0	298	859514	859519	4194304	256	0	0	32	24	294592	0x7f635a251600	0x7f624e637600	12075053989821654	12075053989833004	12075053990438919	12075053990472663
152	150	void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]	0	0	300	859514	859519	4194304	256	0	0	24	24	301184	0x7f635a251500	0x7f624e637640	12075053990496508	12075053990507879	12075053990825796	12075053990852529
153	151	void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]	0	0	302	859514	859519	4194304	256	0	0	12	24	304960	0x7f635a251400	0x7f624e637680	12075053990875121	12075053990886276	12075053991762268	12075053991791023
154	152	void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]	0	0	304	859514	859519	4194304	256	0	0	24	24	313600	0x7f635a251300	0x7f624e6376c0	12075053991827852	12075053991840828	12075053992273944	12075053992302724
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	859514	859519	4194304	256	0	0	20	24	314816	0x7f635a251a00	0x7f624e637700	12075053992326038	12075053992338986	12075053993885855	12075053993960645
156	154	void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]	0	0	308	859514	859519	4194304	256	0	0	20	24	317568	0x7f635a251900	0x7f624e637740	12075053993983017	12075053993996733	12075053994788408	12075053994859005
157	155	void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]	0	0	310	859514	859519	4194304	256	0	0	12	24	320320	0x7f635a251800	0x7f624e637780	12075053994881096	12075053994894807	12075053995307764	12075053995381656
158	156	void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]	0	0	312	859514	859519	4194304	256	0	0	12	24	324096	0x7f635a251700	0x7f624e6377c0	12075053995408867	12075053995420083	12075053996577995	12075053996648841
159	157	void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]	0	0	314	859514	859519	4194304	256	0	0	12	24	328896	0x7f635a251600	0x7f624e637800	12075053996680881	12075053996692074	12075053997476709	12075053997547070
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	859514	859519	4194304	256	0	0	20	24	330112	0x7f635a251500	0x7f624e637840	12075053997569352	12075053997581508	12075054000631406	12075054000702164
161	159	void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]	0	0	318	859514	859519	4194304	256	0	0	20	24	332864	0x7f635a251400	0x7f624e637880	12075054000724325	12075054000737006	12075054002277315	12075054002348052
162	160	void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]	0	0	320	859514	859519	4194304	256	0	0	12	24	335616	0x7f635a251300	0x7f624e6378c0	12075054002371756	12075054002385154	12075054003173788	12075054003244678
163	161	void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]	0	0	322	859514	859519	4194304	256	0	0	12	24	339392	0x7f635a251a00	0x7f624e637900	12075054003268132	12075054003279388	12075054005560171	12075054005631444
164	162	void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]	0	0	324	859514	859519	4194304	256	0	0	12	24	344192	0x7f635a251900	0x7f624e637940	12075054005663794	12075054005676011	12075054007211040	12075054007282031
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	859514	859519	4194304	256	0	0	20	24	345408	0x7f635a251800	0x7f624e637980	12075054007305555	12075054007316959	12075054013370676	12075054013441869
166	164	void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]	0	0	328	859514	859519	4194304	256	0	0	20	24	348160	0x7f635a251700	0x7f624e6379c0	12075054013466805	12075054013480275	12075054016520254	12075054016591131
167	165	void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]	0	0	330	859514	859519	4194304	256	0	0	12	24	350912	0x7f635a251600	0x7f624e637a00	12075054016615366	12075054016627773	12075054018166322	12075054018238112
168	166	void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]	0	0	332	859514	859519	4194304	256	0	0	12	24	0	0x7f635a251500	0x7f624e637a40	12075054018260864	12075054018273041	12075054022797969	12075054022869539