rocm-systems/tests/workloads/D_int_inv2/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	825649	825654	33554432	256	0	0	8	32	6464	0x0	0x7fc02b204180	12074387702272847	12074387702319716	12074387702643873	12074387702729034
3	1	void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]	0	0	2	825649	825654	32768	256	0	0	24	24	12480	0x0	0x7fc02b235100	12074387717454209	12074387717568261	12074387717574821	12074387717580784
4	2	void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]	0	0	4	825649	825654	4194304	256	0	0	24	24	12928	0x7fc15b1a3900	0x7fc02b235140	12074387717628603	12074387717642980	12074387717777059	12074387717780556
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	825649	825654	4194304	256	0	0	36	24	13632	0x7fc15b1a3800	0x7fc02b235180	12074387717824367	12074387717837539	12074387718080096	12074387718180449
6	4	void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]	0	0	8	825649	825654	4194304	256	0	0	28	24	14080	0x7fc15b1a3700	0x7fc02b2351c0	12074387718204514	12074387718216575	12074387718461853	12074387718559353
7	5	void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]	0	0	10	825649	825654	4194304	256	0	0	24	24	14528	0x7fc15b1a3600	0x7fc02b235200	12074387718587405	12074387718599292	12074387718724571	12074387718727926
8	6	void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]	0	0	12	825649	825654	4194304	256	0	0	24	24	14976	0x7fc15b1a3500	0x7fc02b235240	12074387718759986	12074387718771290	12074387718895929	12074387718899335
9	7	void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]	0	0	14	825649	825654	4194304	256	0	0	24	24	15424	0x7fc15b1a3400	0x7fc02b235280	12074387718953225	12074387718968729	12074387719092728	12074387719096120
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	825649	825654	4194304	256	0	0	28	24	16128	0x7fc15b1a3300	0x7fc02b2352c0	12074387719125355	12074387719138167	12074387719383125	12074387719480645
11	9	void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]	0	0	18	825649	825654	4194304	256	0	0	28	24	16832	0x7fc15b1a3a00	0x7fc02b235300	12074387719503417	12074387719514804	12074387719761202	12074387719857696
12	10	void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]	0	0	20	825649	825654	4194304	256	0	0	24	24	17536	0x7fc15b1a3900	0x7fc02b235340	12074387719880037	12074387719892561	12074387720017359	12074387720021260
13	11	void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]	0	0	22	825649	825654	4194304	256	0	0	24	24	18240	0x7fc15b1a3800	0x7fc02b235380	12074387720049672	12074387720060879	12074387720185998	12074387720189532
14	12	void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]	0	0	24	825649	825654	4194304	256	0	0	24	24	18944	0x7fc15b1a3700	0x7fc02b2353c0	12074387720228835	12074387720241197	12074387720365996	12074387720369357
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	825649	825654	4194304	256	0	0	28	24	19904	0x7fc15b1a3600	0x7fc02b235400	12074387720395846	12074387720407436	12074387720652234	12074387720725428
16	14	void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]	0	0	28	825649	825654	4194304	256	0	0	28	24	20608	0x7fc15b1a3500	0x7fc02b235440	12074387720748231	12074387720760073	12074387721003911	12074387721073315
17	15	void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]	0	0	30	825649	825654	4194304	256	0	0	24	24	21312	0x7fc15b1a3400	0x7fc02b235480	12074387721094414	12074387721106310	12074387721230949	12074387721234364
18	16	void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]	0	0	32	825649	825654	4194304	256	0	0	24	24	22016	0x7fc15b1a3300	0x7fc02b2354c0	12074387721261314	12074387721272388	12074387721397667	12074387721401164
19	17	void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]	0	0	34	825649	825654	4194304	256	0	0	24	24	22720	0x7fc15b1a3a00	0x7fc02b235500	12074387721438634	12074387721450627	12074387721575106	12074387721578163
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	825649	825654	4194304	256	0	0	28	24	23680	0x7fc15b1a3900	0x7fc02b235540	12074387721605674	12074387721617985	12074387721869983	12074387721937411
21	19	void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]	0	0	38	825649	825654	4194304	256	0	0	28	24	24384	0x7fc15b1a3800	0x7fc02b235580	12074387721959842	12074387721971582	12074387722217660	12074387722287501
22	20	void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]	0	0	40	825649	825654	4194304	256	0	0	24	24	25088	0x7fc15b1a3700	0x7fc02b2355c0	12074387722313570	12074387722325019	12074387722449978	12074387722453259
23	21	void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]	0	0	42	825649	825654	4194304	256	0	0	24	24	25792	0x7fc15b1a3600	0x7fc02b235600	12074387722480961	12074387722492218	12074387722617017	12074387722620270
24	22	void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]	0	0	44	825649	825654	4194304	256	0	0	24	24	26496	0x7fc15b1a3500	0x7fc02b235640	12074387722657409	12074387722669976	12074387722794455	12074387722797759
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	825649	825654	4194304	256	0	0	28	24	27712	0x7fc15b1a3400	0x7fc02b235680	12074387722825110	12074387722836535	12074387723085012	12074387723183146
26	24	void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]	0	0	48	825649	825654	4194304	256	0	0	28	24	28416	0x7fc15b1a3300	0x7fc02b2356c0	12074387723205397	12074387723217331	12074387723466129	12074387723562641
27	25	void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]	0	0	50	825649	825654	4194304	256	0	0	24	24	29120	0x7fc15b1a3a00	0x7fc02b235700	12074387723584762	12074387723595888	12074387723720367	12074387723723720
28	26	void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]	0	0	52	825649	825654	4194304	256	0	0	24	24	30080	0x7fc15b1a3900	0x7fc02b235740	12074387723751782	12074387723763887	12074387723888685	12074387723892163
29	27	void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]	0	0	54	825649	825654	4194304	256	0	0	24	24	30784	0x7fc15b1a3800	0x7fc02b235780	12074387723929573	12074387723942765	12074387724067564	12074387724071136
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	825649	825654	4194304	256	0	0	28	24	32000	0x7fc15b1a3700	0x7fc02b2357c0	12074387724099479	12074387724111243	12074387724359721	12074387724414384
31	29	void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]	0	0	58	825649	825654	4194304	256	0	0	28	24	32960	0x7fc15b1a3600	0x7fc02b235800	12074387724436054	12074387724447081	12074387724691878	12074387724769454
32	30	void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]	0	0	60	825649	825654	4194304	256	0	0	24	24	33920	0x7fc15b1a3500	0x7fc02b235840	12074387724790913	12074387724802117	12074387724926756	12074387724930222
33	31	void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]	0	0	62	825649	825654	4194304	256	0	0	24	24	34880	0x7fc15b1a3400	0x7fc02b235880	12074387724956772	12074387724968036	12074387725093155	12074387725096662
34	32	void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]	0	0	64	825649	825654	4194304	256	0	0	24	24	35840	0x7fc15b1a3300	0x7fc02b2358c0	12074387725135123	12074387725146914	12074387725271233	12074387725274612
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	825649	825654	4194304	256	0	0	28	24	37312	0x7fc15b1a3a00	0x7fc02b235900	12074387725301412	12074387725312513	12074387725558591	12074387725588896
36	34	void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]	0	0	68	825649	825654	4194304	256	0	0	28	24	38272	0x7fc15b1a3900	0x7fc02b235940	12074387725609966	12074387725620990	12074387725902748	12074387725950358
37	35	void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]	0	0	70	825649	825654	4194304	256	0	0	24	24	39232	0x7fc15b1a3800	0x7fc02b235980	12074387725973341	12074387725986107	12074387726111546	12074387726115335
38	36	void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]	0	0	72	825649	825654	4194304	256	0	0	24	24	40192	0x7fc15b1a3700	0x7fc02b2359c0	12074387726142485	12074387726153945	12074387726279864	12074387726283157
39	37	void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]	0	0	74	825649	825654	4194304	256	0	0	24	24	41152	0x7fc15b1a3600	0x7fc02b235a00	12074387726318943	12074387726333464	12074387726457303	12074387726460456
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	825649	825654	4194304	256	0	0	28	24	42624	0x7fc15b1a3500	0x7fc02b235a40	12074387726487506	12074387726499382	12074387726749780	12074387726780040
41	39	void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]	0	0	78	825649	825654	4194304	256	0	0	28	24	43584	0x7fc15b1a3400	0x7fc02b235a80	12074387726803984	12074387726816180	12074387727062258	12074387727113129
42	40	void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]	0	0	80	825649	825654	4194304	256	0	0	24	24	44544	0x7fc15b1a3300	0x7fc02b235ac0	12074387727139047	12074387727152337	12074387727276976	12074387727280069
43	41	void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]	0	0	82	825649	825654	4194304	256	0	0	24	24	45760	0x7fc15b1a3a00	0x7fc02b235b00	12074387727306989	12074387727318255	12074387727444974	12074387727448372
44	42	void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]	0	0	84	825649	825654	4194304	256	0	0	24	24	46720	0x7fc15b1a3900	0x7fc02b235b40	12074387727485301	12074387727496814	12074387727621773	12074387727625110
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	825649	825654	4194304	256	0	0	28	24	48448	0x7fc15b1a3800	0x7fc02b235b80	12074387727651760	12074387727663532	12074387727918090	12074387727964521
46	44	void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]	0	0	88	825649	825654	4194304	256	0	0	28	24	49408	0x7fc15b1a3700	0x7fc02b235bc0	12074387727988185	12074387727999369	12074387728244007	12074387728297249
47	45	void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]	0	0	90	825649	825654	4194304	256	0	0	24	24	50368	0x7fc15b1a3600	0x7fc02b235c00	12074387728318228	12074387728330086	12074387728454085	12074387728457327
48	46	void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]	0	0	92	825649	825654	4194304	256	0	0	24	24	51584	0x7fc15b1a3500	0x7fc02b235c40	12074387728485309	12074387728496165	12074387728623044	12074387728626341
49	47	void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]	0	0	94	825649	825654	4194304	256	0	0	24	24	52544	0x7fc15b1a3400	0x7fc02b235c80	12074387728663420	12074387728674883	12074387728800322	12074387728803500
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	825649	825654	4194304	256	0	0	28	24	54272	0x7fc15b1a3300	0x7fc02b235cc0	12074387728829919	12074387728842722	12074387729093120	12074387729146057
51	49	void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]	0	0	98	825649	825654	4194304	256	0	0	28	24	55488	0x7fc15b1a3a00	0x7fc02b235d00	12074387729167427	12074387729178719	12074387729425437	12074387729476771
52	50	void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]	0	0	100	825649	825654	4194304	256	0	0	24	24	56704	0x7fc15b1a3900	0x7fc02b235d40	12074387729498712	12074387729510076	12074387729634875	12074387729638111
53	51	void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]	0	0	102	825649	825654	4194304	256	0	0	24	24	58176	0x7fc15b1a3800	0x7fc02b235d80	12074387729664871	12074387729676475	12074387729805113	12074387729808348
54	52	void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]	0	0	104	825649	825654	4194304	256	0	0	24	24	59392	0x7fc15b1a3700	0x7fc02b235dc0	12074387729844194	12074387729855833	12074387729980632	12074387729983724
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	825649	825654	4194304	256	0	0	28	24	61376	0x7fc15b1a3600	0x7fc02b235e00	12074387730019270	12074387730031031	12074387730281909	12074387730306974
56	54	void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]	0	0	108	825649	825654	4194304	256	0	0	28	24	62592	0x7fc15b1a3500	0x7fc02b235e40	12074387730336669	12074387730348309	12074387730595826	12074387730616570
57	55	void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]	0	0	110	825649	825654	4194304	256	0	0	24	24	63808	0x7fc15b1a3400	0x7fc02b235e80	12074387730640143	12074387730651346	12074387730775665	12074387730778851
58	56	void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]	0	0	112	825649	825654	4194304	256	0	0	24	24	65280	0x7fc15b1a3300	0x7fc02b235ec0	12074387730805871	12074387730817264	12074387730946543	12074387730949759
59	57	void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]	0	0	114	825649	825654	4194304	256	0	0	24	24	66496	0x7fc15b1a3a00	0x7fc02b235f00	12074387730986878	12074387730998223	12074387731123822	12074387731127018
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	825649	825654	4194304	256	0	0	28	24	68480	0x7fc15b1a3900	0x7fc02b235f40	12074387731155040	12074387731167181	12074387731419499	12074387731446191
61	59	void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]	0	0	118	825649	825654	4194304	256	0	0	28	24	69696	0x7fc15b1a3800	0x7fc02b235f80	12074387731468964	12074387731482059	12074387731730696	12074387731754434
62	60	void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]	0	0	120	825649	825654	4194304	256	0	0	24	24	70912	0x7fc15b1a3700	0x7fc02b235fc0	12074387731777657	12074387731790536	12074387731915335	12074387731918709
63	61	void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]	0	0	122	825649	825654	4194304	256	0	0	24	24	72384	0x7fc15b1a3600	0x7fc02b236000	12074387731945099	12074387731956294	12074387732092293	12074387732095859
64	62	void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]	0	0	124	825649	825654	4194304	256	0	0	24	24	73600	0x7fc15b1a3500	0x7fc02b236040	12074387732133659	12074387732144933	12074387732269252	12074387732272477
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	825649	825654	4194304	256	0	0	28	24	75840	0x7fc15b1a3400	0x7fc02b236080	12074387732299357	12074387732311011	12074387732569089	12074387732593493
66	64	void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]	0	0	128	825649	825654	4194304	256	0	0	28	24	77056	0x7fc15b1a3300	0x7fc02b2360c0	12074387732614723	12074387732626209	12074387732874846	12074387732898470
67	65	void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]	0	0	130	825649	825654	4194304	256	0	0	24	24	78272	0x7fc15b1a3a00	0x7fc02b236100	12074387732919089	12074387732931486	12074387733056605	12074387733060241
68	66	void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]	0	0	132	825649	825654	4194304	256	0	0	24	24	80000	0x7fc15b1a3900	0x7fc02b236140	12074387733087572	12074387733098524	12074387733240603	12074387733243892
69	67	void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]	0	0	134	825649	825654	4194304	256	0	0	24	24	81216	0x7fc15b1a3800	0x7fc02b236180	12074387733280190	12074387733291323	12074387733416282	12074387733419569
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	825649	825654	4194304	256	0	0	28	24	83456	0x7fc15b1a3700	0x7fc02b2361c0	12074387733446348	12074387733458201	12074387733710519	12074387733731458
71	69	void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]	0	0	138	825649	825654	4194304	256	0	0	28	24	84928	0x7fc15b1a3600	0x7fc02b236200	12074387733753619	12074387733765239	12074387734012756	12074387734038489
72	70	void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]	0	0	140	825649	825654	4194304	256	0	0	24	24	86400	0x7fc15b1a3500	0x7fc02b236240	12074387734063185	12074387734074996	12074387734200115	12074387734203425
73	71	void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]	0	0	142	825649	825654	4194304	256	0	0	24	24	88128	0x7fc15b1a3400	0x7fc02b236280	12074387734228672	12074387734240114	12074387734390513	12074387734393819
74	72	void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]	0	0	144	825649	825654	4194304	256	0	0	24	24	89600	0x7fc15b1a3300	0x7fc02b2362c0	12074387734447860	12074387734459792	12074387734585231	12074387734588621
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	825649	825654	4194304	256	0	0	28	24	92096	0x7fc15b1a3a00	0x7fc02b236300	12074387734615631	12074387734626991	12074387734880269	12074387734900441
76	74	void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]	0	0	148	825649	825654	4194304	256	0	0	28	24	93568	0x7fc15b1a3900	0x7fc02b236340	12074387734925597	12074387734936908	12074387735188266	12074387735215556
77	75	void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]	0	0	150	825649	825654	4194304	256	0	0	24	24	95040	0x7fc15b1a3800	0x7fc02b236380	12074387735237878	12074387735249226	12074387735375944	12074387735379481
78	76	void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]	0	0	152	825649	825654	4194304	256	0	0	24	24	96768	0x7fc15b1a3700	0x7fc02b2363c0	12074387735405810	12074387735416904	12074387735573063	12074387735576467
79	77	void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]	0	0	154	825649	825654	4194304	256	0	0	24	24	98240	0x7fc15b1a3600	0x7fc02b236400	12074387735613075	12074387735624742	12074387735749381	12074387735752684
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	825649	825654	4194304	256	0	0	28	24	100736	0x7fc15b1a3500	0x7fc02b236440	12074387735779264	12074387735790181	12074387736046819	12074387736073651
81	79	void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]	0	0	158	825649	825654	4194304	256	0	0	28	24	102208	0x7fc15b1a3400	0x7fc02b236480	12074387736095652	12074387736107938	12074387736353376	12074387736377506
82	80	void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]	0	0	160	825649	825654	4194304	256	0	0	24	24	103680	0x7fc15b1a3300	0x7fc02b2364c0	12074387736398735	12074387736409695	12074387736535294	12074387736538565
83	81	void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]	0	0	162	825649	825654	4194304	256	0	0	24	24	105664	0x7fc15b1a3a00	0x7fc02b236500	12074387736564012	12074387736575774	12074387736741532	12074387736744728
84	82	void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]	0	0	164	825649	825654	4194304	256	0	0	24	24	107136	0x7fc15b1a3900	0x7fc02b236540	12074387736780795	12074387736791932	12074387736917691	12074387736921096
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	825649	825654	4194304	256	0	0	28	24	109888	0x7fc15b1a3800	0x7fc02b236580	12074387736947495	12074387736958810	12074387737217368	12074387737242072
86	84	void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]	0	0	168	825649	825654	4194304	256	0	0	28	24	111360	0x7fc15b1a3700	0x7fc02b2365c0	12074387737263272	12074387737274488	12074387737519446	12074387737542721
87	85	void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]	0	0	170	825649	825654	4194304	256	0	0	24	24	112832	0x7fc15b1a3600	0x7fc02b236600	12074387737566074	12074387737578165	12074387737703764	12074387737707006
88	86	void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]	0	0	172	825649	825654	4194304	256	0	0	24	24	114816	0x7fc15b1a3500	0x7fc02b236640	12074387737733145	12074387737744244	12074387737916722	12074387737919932
89	87	void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]	0	0	174	825649	825654	4194304	256	0	0	24	24	116288	0x7fc15b1a3400	0x7fc02b236680	12074387737955358	12074387737966802	12074387738092561	12074387738096089
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	825649	825654	4194304	256	0	0	28	24	119040	0x7fc15b1a3300	0x7fc02b2366c0	12074387738123049	12074387738134800	12074387738391758	12074387738415333
91	89	void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]	0	0	178	825649	825654	4194304	256	0	0	28	24	120768	0x7fc15b1a3a00	0x7fc02b236700	12074387738435580	12074387738446637	12074387738691755	12074387738711273
92	90	void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]	0	0	180	825649	825654	4194304	256	0	0	24	24	122496	0x7fc15b1a3900	0x7fc02b236740	12074387738738994	12074387738750795	12074387738877034	12074387738880357
93	91	void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]	0	0	182	825649	825654	4194304	256	0	0	24	24	124736	0x7fc15b1a3800	0x7fc02b236780	12074387738906575	12074387738917833	12074387739099912	12074387739103461
94	92	void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]	0	0	184	825649	825654	4194304	256	0	0	24	24	126464	0x7fc15b1a3700	0x7fc02b2367c0	12074387739141442	12074387739152711	12074387739279270	12074387739283005
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	825649	825654	4194304	256	0	0	20	24	127936	0x7fc15b1a3600	0x7fc02b236800	12074387739310276	12074387739322150	12074387739585027	12074387739608490
96	94	void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]	0	0	188	825649	825654	4194304	256	0	0	28	24	129664	0x7fc15b1a3500	0x7fc02b236840	12074387739629830	12074387739640707	12074387739889665	12074387739913627
97	95	void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]	0	0	190	825649	825654	4194304	256	0	0	24	24	131392	0x7fc15b1a3400	0x7fc02b236880	12074387739935137	12074387739945824	12074387740071743	12074387740075438
98	96	void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]	0	0	192	825649	825654	4194304	256	0	0	24	24	133632	0x7fc15b1a3300	0x7fc02b2368c0	12074387740102308	12074387740113823	12074387740304861	12074387740308160
99	97	void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]	0	0	194	825649	825654	4194304	256	0	0	24	24	135360	0x7fc15b1a3a00	0x7fc02b236900	12074387740344227	12074387740355421	12074387740482939	12074387740486231
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	825649	825654	4194304	256	0	0	20	24	136832	0x7fc15b1a3900	0x7fc02b236940	12074387740512710	12074387740525659	12074387740807577	12074387740831683
101	99	void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]	0	0	198	825649	825654	4194304	256	0	0	28	24	138560	0x7fc15b1a3800	0x7fc02b236980	12074387740853664	12074387740866456	12074387741116854	12074387741142220
102	100	void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]	0	0	200	825649	825654	4194304	256	0	0	24	24	140288	0x7fc15b1a3700	0x7fc02b2369c0	12074387741165153	12074387741177813	12074387741304052	12074387741307417
103	101	void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]	0	0	202	825649	825654	4194304	256	0	0	24	24	142784	0x7fc15b1a3600	0x7fc02b236a00	12074387741334808	12074387741346772	12074387741554770	12074387741577810
104	102	void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]	0	0	204	825649	825654	4194304	256	0	0	24	24	144768	0x7fc15b1a3500	0x7fc02b236a40	12074387741608707	12074387741620049	12074387741748208	12074387741751693
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	825649	825654	4194304	256	0	0	20	24	145728	0x7fc15b1a3400	0x7fc02b236a80	12074387741777872	12074387741789488	12074387742095725	12074387742120939
106	104	void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]	0	0	208	825649	825654	4194304	256	0	0	28	24	147712	0x7fc15b1a3300	0x7fc02b236ac0	12074387742142319	12074387742153805	12074387742402923	12074387742426427
107	105	void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]	0	0	210	825649	825654	4194304	256	0	0	24	24	149696	0x7fc15b1a3a00	0x7fc02b236b00	12074387742449099	12074387742460522	12074387742587241	12074387742590832
108	106	void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]	0	0	212	825649	825654	4194304	256	0	0	24	24	152192	0x7fc15b1a3900	0x7fc02b236b40	12074387742617171	12074387742628521	12074387742856839	12074387742879859
109	107	void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]	0	0	214	825649	825654	4194304	256	0	0	24	24	154176	0x7fc15b1a3800	0x7fc02b236b80	12074387742912871	12074387742924358	12074387743050597	12074387743054233
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	825649	825654	4194304	256	0	0	20	24	155648	0x7fc15b1a3700	0x7fc02b236bc0	12074387743082436	12074387743094437	12074387743451873	12074387743477700
111	109	void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]	0	0	218	825649	825654	4194304	256	0	0	28	24	157632	0x7fc15b1a3600	0x7fc02b236c00	12074387743500483	12074387743511553	12074387743767711	12074387743790782
112	110	void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]	0	0	220	825649	825654	4194304	256	0	0	24	24	159616	0x7fc15b1a3500	0x7fc02b236c40	12074387743811761	12074387743823230	12074387743953149	12074387743956490
113	111	void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]	0	0	222	825649	825654	4194304	256	0	0	24	24	162368	0x7fc15b1a3400	0x7fc02b236c80	12074387743982378	12074387743993469	12074387744236186	12074387744261287
114	112	void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]	0	0	224	825649	825654	4194304	256	0	0	24	24	164608	0x7fc15b1a3300	0x7fc02b236cc0	12074387744290751	12074387744303066	12074387744431065	12074387744434689
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	825649	825654	4194304	256	0	0	20	24	166336	0x7fc15b1a3a00	0x7fc02b236d00	12074387744461288	12074387744472664	12074387744847701	12074387744870789
116	114	void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]	0	0	228	825649	825654	4194304	256	0	0	28	24	168576	0x7fc15b1a3900	0x7fc02b236d40	12074387744892069	12074387744904501	12074387745156338	12074387745185434
117	115	void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]	0	0	230	825649	825654	4194304	256	0	0	24	24	170816	0x7fc15b1a3800	0x7fc02b236d80	12074387745208797	12074387745221618	12074387745351537	12074387745354949
118	116	void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]	0	0	232	825649	825654	4194304	256	0	0	24	24	174080	0x7fc15b1a3700	0x7fc02b236dc0	12074387745381799	12074387745392656	12074387745670734	12074387745698077
119	117	void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]	0	0	234	825649	825654	4194304	256	0	0	24	24	176576	0x7fc15b1a3600	0x7fc02b236e00	12074387745727992	12074387745740013	12074387745875052	12074387745878502
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	825649	825654	4194304	256	0	0	20	24	177792	0x7fc15b1a3500	0x7fc02b236e40	12074387745904590	12074387745916172	12074387746338408	12074387746363032
121	119	void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]	0	0	238	825649	825654	4194304	256	0	0	28	24	180288	0x7fc15b1a3400	0x7fc02b236e80	12074387746386416	12074387746397607	12074387746657285	12074387746681214
122	120	void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]	0	0	240	825649	825654	4194304	256	0	0	24	24	182784	0x7fc15b1a3300	0x7fc02b236ec0	12074387746702654	12074387746713605	12074387746849444	12074387746852833
123	121	void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]	0	0	242	825649	825654	4194304	256	0	0	24	24	186304	0x7fc15b1a3a00	0x7fc02b236f00	12074387746879883	12074387746890563	12074387747204160	12074387747238970
124	122	void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]	0	0	244	825649	825654	4194304	256	0	0	24	24	189312	0x7fc15b1a3900	0x7fc02b236f40	12074387747271992	12074387747283200	12074387747434238	12074387747437710
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	825649	825654	4194304	256	0	0	20	24	190784	0x7fc15b1a3800	0x7fc02b236f80	12074387747463618	12074387747474718	12074387747990873	12074387748019371
126	124	void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]	0	0	248	825649	825654	4194304	256	0	0	28	24	193792	0x7fc15b1a3700	0x7fc02b236fc0	12074387748040800	12074387748051673	12074387748350230	12074387748377105
127	125	void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]	0	0	250	825649	825654	4194304	256	0	0	24	24	196800	0x7fc15b1a3600	0x7fc02b237000	12074387748397082	12074387748410870	12074387748565428	12074387748568842
128	126	void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]	0	0	252	825649	825654	4194304	256	0	0	24	24	201088	0x7fc15b1a3500	0x7fc02b237040	12074387748594840	12074387748605908	12074387748990545	12074387749018567
129	127	void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]	0	0	254	825649	825654	4194304	256	0	0	24	24	204608	0x7fc15b1a3400	0x7fc02b237080	12074387749051929	12074387749063824	12074387749239822	12074387749243245
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	825649	825654	4194304	256	0	0	20	24	206080	0x7fc15b1a3300	0x7fc02b2370c0	12074387749269774	12074387749280942	12074387749903017	12074387749929570
131	129	void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]	0	0	258	825649	825654	4194304	256	0	0	28	24	209600	0x7fc15b1a3a00	0x7fc02b237100	12074387749950379	12074387749961576	12074387750289893	12074387750321779
132	130	void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]	0	0	260	825649	825654	4194304	256	0	0	24	24	213120	0x7fc15b1a3900	0x7fc02b237140	12074387750345774	12074387750357253	12074387750535331	12074387750538893
133	131	void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]	0	0	262	825649	825654	4194304	256	0	0	24	24	218176	0x7fc15b1a3800	0x7fc02b237180	12074387750565763	12074387750577091	12074387751031647	12074387751059721
134	132	void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]	0	0	264	825649	825654	4194304	256	0	0	24	24	222208	0x7fc15b1a3700	0x7fc02b2371c0	12074387751092421	12074387751103486	12074387751302204	12074387751328370
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	825649	825654	4194304	256	0	0	20	24	223936	0x7fc15b1a3600	0x7fc02b237200	12074387751348658	12074387751360124	12074387752062998	12074387752090847
136	134	void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]	0	0	268	825649	825654	4194304	256	0	0	28	24	227968	0x7fc15b1a3500	0x7fc02b237240	12074387752115583	12074387752126517	12074387752499794	12074387752526246
137	135	void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]	0	0	270	825649	825654	4194304	256	0	0	24	24	232000	0x7fc15b1a3400	0x7fc02b237280	12074387752547465	12074387752558833	12074387752760112	12074387752787411
138	136	void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]	0	0	272	825649	825654	4194304	256	0	0	24	24	237824	0x7fc15b1a3300	0x7fc02b2372c0	12074387752808981	12074387752821711	12074387753347466	12074387753380413
139	137	void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]	0	0	274	825649	825654	4194304	256	0	0	24	24	242368	0x7fc15b1a3a00	0x7fc02b237300	12074387753412152	12074387753423946	12074387753648104	12074387753670983
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	825649	825654	4194304	256	0	0	20	24	243584	0x7fc15b1a3900	0x7fc02b237340	12074387753693605	12074387753704423	12074387754502976	12074387754530531
141	139	void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]	0	0	278	825649	825654	4194304	256	0	0	28	24	248128	0x7fc15b1a3800	0x7fc02b237380	12074387754553193	12074387754564416	12074387754984732	12074387755013057
142	140	void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]	0	0	280	825649	825654	4194304	256	0	0	24	24	252672	0x7fc15b1a3700	0x7fc02b2373c0	12074387755037373	12074387755049051	12074387755273689	12074387755300772
143	141	void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]	0	0	282	825649	825654	4194304	256	0	0	24	24	259264	0x7fc15b1a3600	0x7fc02b237400	12074387755323114	12074387755334009	12074387755930164	12074387755957332
144	142	void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]	0	0	284	825649	825654	4194304	256	0	0	24	24	264832	0x7fc15b1a3500	0x7fc02b237440	12074387755988710	12074387755999603	12074387756268401	12074387756297584
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	825649	825654	4194304	256	0	0	20	24	266304	0x7fc15b1a3400	0x7fc02b237480	12074387756320447	12074387756332560	12074387757317031	12074387757346213
146	144	void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]	0	0	288	825649	825654	4194304	256	0	0	28	24	271872	0x7fc15b1a3300	0x7fc02b2374c0	12074387757366030	12074387757376871	12074387757891906	12074387757918757
147	145	void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]	0	0	290	825649	825654	4194304	256	0	0	24	24	277440	0x7fc15b1a3a00	0x7fc02b237500	12074387757943704	12074387757955106	12074387758225983	12074387758257988
148	146	void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]	0	0	292	825649	825654	4194304	256	0	0	12	24	279936	0x7fc15b1a3900	0x7fc02b237540	12074387758281321	12074387758294143	12074387759036056	12074387759064627
149	147	void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]	0	0	294	825649	825654	4194304	256	0	0	24	24	286528	0x7fc15b1a3800	0x7fc02b237580	12074387759094682	12074387759106136	12074387759422133	12074387759449051
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	825649	825654	4194304	256	0	0	20	24	288000	0x7fc15b1a3700	0x7fc02b2375c0	12074387759471583	12074387759482932	12074387760654922	12074387760682964
151	149	void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]	0	0	298	825649	825654	4194304	256	0	0	32	24	294592	0x7fc15b1a3600	0x7fc02b237600	12074387760703101	12074387760713962	12074387761322276	12074387761352238
152	150	void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]	0	0	300	825649	825654	4194304	256	0	0	24	24	301184	0x7fc15b1a3500	0x7fc02b237640	12074387761375601	12074387761386596	12074387761705793	12074387761733436
153	151	void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]	0	0	302	825649	825654	4194304	256	0	0	12	24	304960	0x7fc15b1a3400	0x7fc02b237680	12074387761755848	12074387761767232	12074387762644185	12074387762673443
154	152	void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]	0	0	304	825649	825654	4194304	256	0	0	24	24	313600	0x7fc15b1a3300	0x7fc02b2376c0	12074387762707637	12074387762718904	12074387763127380	12074387763157332
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	825649	825654	4194304	256	0	0	20	24	314816	0x7fc15b1a3a00	0x7fc02b237700	12074387763180335	12074387763192820	12074387764741766	12074387764814692
156	154	void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]	0	0	308	825649	825654	4194304	256	0	0	20	24	317568	0x7fc15b1a3900	0x7fc02b237740	12074387764836353	12074387764848984	12074387765642898	12074387765721978
157	155	void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]	0	0	310	825649	825654	4194304	256	0	0	12	24	320320	0x7fc15b1a3800	0x7fc02b237780	12074387765742096	12074387765753616	12074387766167053	12074387766238659
158	156	void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]	0	0	312	825649	825654	4194304	256	0	0	12	24	324096	0x7fc15b1a3700	0x7fc02b2377c0	12074387766260559	12074387766272652	12074387767430564	12074387767502297
159	157	void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]	0	0	314	825649	825654	4194304	256	0	0	12	24	328896	0x7fc15b1a3600	0x7fc02b237800	12074387767534747	12074387767546403	12074387768331677	12074387768403091
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	825649	825654	4194304	256	0	0	20	24	330112	0x7fc15b1a3500	0x7fc02b237840	12074387768426675	12074387768439357	12074387771489734	12074387771561400
161	159	void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]	0	0	318	825649	825654	4194304	256	0	0	20	24	332864	0x7fc15b1a3400	0x7fc02b237880	12074387771585435	12074387771598374	12074387773138682	12074387773210545
162	160	void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]	0	0	320	825649	825654	4194304	256	0	0	12	24	335616	0x7fc15b1a3300	0x7fc02b2378c0	12074387773236072	12074387773249241	12074387774038356	12074387774110277
163	161	void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]	0	0	322	825649	825654	4194304	256	0	0	12	24	339392	0x7fc15b1a3a00	0x7fc02b237900	12074387774134712	12074387774147475	12074387776429378	12074387776501771
164	162	void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]	0	0	324	825649	825654	4194304	256	0	0	12	24	344192	0x7fc15b1a3900	0x7fc02b237940	12074387776536065	12074387776549697	12074387778084886	12074387778157238
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	825649	825654	4194304	256	0	0	20	24	345408	0x7fc15b1a3800	0x7fc02b237980	12074387778180210	12074387778194005	12074387784248681	12074387784320872
166	164	void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]	0	0	328	825649	825654	4194304	256	0	0	20	24	348160	0x7fc15b1a3700	0x7fc02b2379c0	12074387784343925	12074387784356680	12074387787397938	12074387787470605
167	165	void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]	0	0	330	825649	825654	4194304	256	0	0	12	24	350912	0x7fc15b1a3600	0x7fc02b237a00	12074387787492546	12074387787505617	12074387789044966	12074387789118748
168	166	void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]	0	0	332	825649	825654	4194304	256	0	0	12	24	0	0x7fc15b1a3500	0x7fc02b237a40	12074387789139907	12074387789153605	12074387793679812	12074387793752740