rocm-systems/tests/workloads_old/no_roof_CPF/mi100/timestamps.csv

1	Dispatch_ID	Kernel_Name	GPU_ID	queue-id	queue-index	pid	tid	grd	Workgroup_Size	LDS_Per_Workgroup	scr	vgpr	SGPR	fbar	sig	obj	DispatchNs	Start_Timestamp	End_Timestamp	CompleteNs
2	0	__amd_rocclr_fillBuffer.kd	0	0	0	870469	870474	33554432	256	0	0	8	32	6464	0x0	0x7fda00804180	12075337326665086	12075337326712913	12075337327035950	12075337327141892
3	1	void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]	0	0	2	870469	870474	32768	256	0	0	24	24	12480	0x0	0x7fda00835100	12075337341805278	12075337341910373	12075337341917253	12075337341922797
4	2	void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]	0	0	4	870469	870474	4194304	256	0	0	24	24	12928	0x7fdb0c52f900	0x7fda00835140	12075337341964654	12075337341978052	12075337342114851	12075337342118360
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	870469	870474	4194304	256	0	0	36	24	13632	0x7fdb0c52f800	0x7fda00835180	12075337342162262	12075337342175970	12075337342423008	12075337342520888
6	4	void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]	0	0	8	870469	870474	4194304	256	0	0	28	24	14080	0x7fdb0c52f700	0x7fda008351c0	12075337342547037	12075337342558527	12075337342806524	12075337342902477
7	5	void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]	0	0	10	870469	870474	4194304	256	0	0	24	24	14528	0x7fdb0c52f600	0x7fda00835200	12075337342928105	12075337342939963	12075337343064922	12075337343068646
8	6	void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]	0	0	12	870469	870474	4194304	256	0	0	24	24	14976	0x7fdb0c52f500	0x7fda00835240	12075337343096417	12075337343108922	12075337343233560	12075337343236828
9	7	void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]	0	0	14	870469	870474	4194304	256	0	0	24	24	15424	0x7fdb0c52f400	0x7fda00835280	12075337343290187	12075337343305400	12075337343429879	12075337343433073
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	870469	870474	4194304	256	0	0	28	24	16128	0x7fdb0c52f300	0x7fda008352c0	12075337343460444	12075337343474358	12075337343717876	12075337343814712
11	9	void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]	0	0	18	870469	870474	4194304	256	0	0	28	24	16832	0x7fdb0c52fa00	0x7fda00835300	12075337343835841	12075337343847155	12075337344090673	12075337344190410
12	10	void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]	0	0	20	870469	870474	4194304	256	0	0	24	24	17536	0x7fdb0c52f900	0x7fda00835340	12075337344213894	12075337344225391	12075337344350350	12075337344353844
13	11	void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]	0	0	22	870469	870474	4194304	256	0	0	24	24	18240	0x7fdb0c52f800	0x7fda00835380	12075337344379912	12075337344391150	12075337344515949	12075337344519351
14	12	void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]	0	0	24	870469	870474	4194304	256	0	0	24	24	18944	0x7fdb0c52f700	0x7fda008353c0	12075337344555398	12075337344567308	12075337344691627	12075337344694837
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	870469	870474	4194304	256	0	0	28	24	19904	0x7fdb0c52f600	0x7fda00835400	12075337344721407	12075337344732267	12075337344984424	12075337345060948
16	14	void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]	0	0	28	870469	870474	4194304	256	0	0	28	24	20608	0x7fdb0c52f500	0x7fda00835440	12075337345081766	12075337345095143	12075337345344261	12075337345411018
17	15	void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]	0	0	30	870469	870474	4194304	256	0	0	24	24	21312	0x7fdb0c52f400	0x7fda00835480	12075337345432137	12075337345443620	12075337345568579	12075337345571867
18	16	void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]	0	0	32	870469	870474	4194304	256	0	0	24	24	22016	0x7fdb0c52f300	0x7fda008354c0	12075337345597304	12075337345610499	12075337345736417	12075337345739539
19	17	void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]	0	0	34	870469	870474	4194304	256	0	0	24	24	22720	0x7fdb0c52fa00	0x7fda00835500	12075337345777078	12075337345787937	12075337345912576	12075337345915576
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	870469	870474	4194304	256	0	0	28	24	23680	0x7fdb0c52f900	0x7fda00835540	12075337345942315	12075337345953375	12075337346200253	12075337346270024
21	19	void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]	0	0	38	870469	870474	4194304	256	0	0	28	24	24384	0x7fdb0c52f800	0x7fda00835580	12075337346291544	12075337346303132	12075337346546330	12075337346611900
22	20	void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]	0	0	40	870469	870474	4194304	256	0	0	24	24	25088	0x7fdb0c52f700	0x7fda008355c0	12075337346631767	12075337346642809	12075337346767608	12075337346770574
23	21	void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]	0	0	42	870469	870474	4194304	256	0	0	24	24	25792	0x7fdb0c52f600	0x7fda00835600	12075337346796953	12075337346808408	12075337346933366	12075337346936523
24	22	void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]	0	0	44	870469	870474	4194304	256	0	0	24	24	26496	0x7fdb0c52f500	0x7fda00835640	12075337346972269	12075337346983446	12075337347107285	12075337347110446
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	870469	870474	4194304	256	0	0	28	24	27712	0x7fdb0c52f400	0x7fda00835680	12075337347137166	12075337347148724	12075337347402962	12075337347443304
26	24	void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]	0	0	48	870469	870474	4194304	256	0	0	28	24	28416	0x7fdb0c52f300	0x7fda008356c0	12075337347465115	12075337347476721	12075337347721839	12075337347786733
27	25	void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]	0	0	50	870469	870474	4194304	256	0	0	24	24	29120	0x7fdb0c52fa00	0x7fda00835700	12075337347806329	12075337347817998	12075337347942317	12075337347945217
28	26	void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]	0	0	52	870469	870474	4194304	256	0	0	24	24	30080	0x7fdb0c52f900	0x7fda00835740	12075337347970193	12075337347981677	12075337348106956	12075337348110384
29	27	void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]	0	0	54	870469	870474	4194304	256	0	0	24	24	30784	0x7fdb0c52f800	0x7fda00835780	12075337348145890	12075337348160235	12075337348284554	12075337348287513
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	870469	870474	4194304	256	0	0	28	24	32000	0x7fdb0c52f700	0x7fda008357c0	12075337348313091	12075337348324714	12075337348570631	12075337348622596
31	29	void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]	0	0	58	870469	870474	4194304	256	0	0	28	24	32960	0x7fdb0c52f600	0x7fda00835800	12075337348642933	12075337348653831	12075337348901348	12075337348978196
32	30	void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]	0	0	60	870469	870474	4194304	256	0	0	24	24	33920	0x7fdb0c52f500	0x7fda00835840	12075337348997743	12075337349013347	12075337349137666	12075337349141169
33	31	void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]	0	0	62	870469	870474	4194304	256	0	0	24	24	34880	0x7fdb0c52f400	0x7fda00835880	12075337349167338	12075337349178786	12075337349304545	12075337349307939
34	32	void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]	0	0	64	870469	870474	4194304	256	0	0	24	24	35840	0x7fdb0c52f300	0x7fda008358c0	12075337349342814	12075337349357344	12075337349482143	12075337349485379
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	870469	870474	4194304	256	0	0	28	24	37312	0x7fdb0c52fa00	0x7fda00835900	12075337349509975	12075337349521663	12075337349769820	12075337349822275
36	34	void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]	0	0	68	870469	870474	4194304	256	0	0	28	24	38272	0x7fdb0c52f900	0x7fda00835940	12075337349842793	12075337349860699	12075337350104537	12075337350164040
37	35	void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]	0	0	70	870469	870474	4194304	256	0	0	24	24	39232	0x7fdb0c52f800	0x7fda00835980	12075337350184618	12075337350195736	12075337350320215	12075337350323987
38	36	void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]	0	0	72	870469	870474	4194304	256	0	0	24	24	40192	0x7fdb0c52f700	0x7fda008359c0	12075337350349264	12075337350360695	12075337350486614	12075337350490026
39	37	void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]	0	0	74	870469	870474	4194304	256	0	0	24	24	41152	0x7fdb0c52f600	0x7fda00835a00	12075337350524730	12075337350535733	12075337350659252	12075337350662296
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	870469	870474	4194304	256	0	0	28	24	42624	0x7fdb0c52f500	0x7fda00835a40	12075337350688374	12075337350700372	12075337350950769	12075337351013939
41	39	void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]	0	0	78	870469	870474	4194304	256	0	0	28	24	43584	0x7fdb0c52f400	0x7fda00835a80	12075337351037453	12075337351048689	12075337351292206	12075337351338823
42	40	void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]	0	0	80	870469	870474	4194304	256	0	0	24	24	44544	0x7fdb0c52f300	0x7fda00835ac0	12075337351358449	12075337351370126	12075337351495084	12075337351498530
43	41	void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]	0	0	82	870469	870474	4194304	256	0	0	24	24	45760	0x7fdb0c52fa00	0x7fda00835b00	12075337351524708	12075337351537644	12075337351663563	12075337351666672
44	42	void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]	0	0	84	870469	870474	4194304	256	0	0	24	24	46720	0x7fdb0c52f900	0x7fda00835b40	12075337351702018	12075337351713002	12075337351837481	12075337351840635
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	870469	870474	4194304	256	0	0	28	24	48448	0x7fdb0c52f800	0x7fda00835b80	12075337351867505	12075337351878921	12075337352130759	12075337352179305
46	44	void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]	0	0	88	870469	870474	4194304	256	0	0	28	24	49408	0x7fdb0c52f700	0x7fda00835bc0	12075337352202207	12075337352213478	12075337352460036	12075337352484482
47	45	void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]	0	0	90	870469	870474	4194304	256	0	0	24	24	50368	0x7fdb0c52f600	0x7fda00835c00	12075337352504329	12075337352515395	12075337352639874	12075337352642896
48	46	void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]	0	0	92	870469	870474	4194304	256	0	0	24	24	51584	0x7fdb0c52f500	0x7fda00835c40	12075337352669005	12075337352680674	12075337352807392	12075337352810477
49	47	void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]	0	0	94	870469	870474	4194304	256	0	0	24	24	52544	0x7fdb0c52f400	0x7fda00835c80	12075337352846975	12075337352858272	12075337352983551	12075337352986494
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	870469	870474	4194304	256	0	0	28	24	54272	0x7fdb0c52f300	0x7fda00835cc0	12075337353020207	12075337353031390	12075337353283868	12075337353310346
51	49	void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]	0	0	98	870469	870474	4194304	256	0	0	28	24	55488	0x7fdb0c52fa00	0x7fda00835d00	12075337353330283	12075337353341947	12075337353590585	12075337353609442
52	50	void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]	0	0	100	870469	870474	4194304	256	0	0	24	24	56704	0x7fdb0c52f900	0x7fda00835d40	12075337353629399	12075337353640505	12075337353765304	12075337353768297
53	51	void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]	0	0	102	870469	870474	4194304	256	0	0	24	24	58176	0x7fdb0c52f800	0x7fda00835d80	12075337353794536	12075337353806583	12075337353936342	12075337353939265
54	52	void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]	0	0	104	870469	870474	4194304	256	0	0	24	24	59392	0x7fdb0c52f700	0x7fda00835dc0	12075337353973969	12075337353985462	12075337354110420	12075337354113699
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	870469	870474	4194304	256	0	0	28	24	61376	0x7fdb0c52f600	0x7fda00835e00	12075337354140439	12075337354152020	12075337354410738	12075337354435667
56	54	void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]	0	0	108	870469	870474	4194304	256	0	0	28	24	62592	0x7fdb0c52f500	0x7fda00835e40	12075337354456205	12075337354467217	12075337354714895	12075337354737448
57	55	void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]	0	0	110	870469	870474	4194304	256	0	0	24	24	63808	0x7fdb0c52f400	0x7fda00835e80	12075337354757686	12075337354768814	12075337354893613	12075337354896574
58	56	void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]	0	0	112	870469	870474	4194304	256	0	0	24	24	65280	0x7fdb0c52f300	0x7fda00835ec0	12075337354922842	12075337354934253	12075337355062572	12075337355065768
59	57	void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]	0	0	114	870469	870474	4194304	256	0	0	24	24	66496	0x7fdb0c52fa00	0x7fda00835f00	12075337355101484	12075337355113451	12075337355238410	12075337355241334
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	870469	870474	4194304	256	0	0	28	24	68480	0x7fdb0c52f900	0x7fda00835f40	12075337355266150	12075337355277610	12075337355535207	12075337355558193
61	59	void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]	0	0	118	870469	870474	4194304	256	0	0	28	24	69696	0x7fdb0c52f800	0x7fda00835f80	12075337355580034	12075337355593287	12075337355840644	12075337355863470
62	60	void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]	0	0	120	870469	870474	4194304	256	0	0	24	24	70912	0x7fdb0c52f700	0x7fda00835fc0	12075337355882205	12075337355894564	12075337356019363	12075337356022556
63	61	void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]	0	0	122	870469	870474	4194304	256	0	0	24	24	72384	0x7fdb0c52f600	0x7fda00836000	12075337356048414	12075337356061122	12075337356197281	12075337356200346
64	62	void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]	0	0	124	870469	870474	4194304	256	0	0	24	24	73600	0x7fdb0c52f500	0x7fda00836040	12075337356235582	12075337356246721	12075337356371840	12075337356374850
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	870469	870474	4194304	256	0	0	28	24	75840	0x7fdb0c52f400	0x7fda00836080	12075337356400658	12075337356412639	12075337356670717	12075337356690166
66	64	void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]	0	0	128	870469	870474	4194304	256	0	0	28	24	77056	0x7fdb0c52f300	0x7fda008360c0	12075337356709572	12075337356722716	12075337356968154	12075337356991286
67	65	void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]	0	0	130	870469	870474	4194304	256	0	0	24	24	78272	0x7fdb0c52fa00	0x7fda00836100	12075337357013547	12075337357025594	12075337357150552	12075337357153658
68	66	void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]	0	0	132	870469	870474	4194304	256	0	0	24	24	80000	0x7fdb0c52f900	0x7fda00836140	12075337357181950	12075337357193112	12075337357335191	12075337357338261
69	67	void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]	0	0	134	870469	870474	4194304	256	0	0	24	24	81216	0x7fdb0c52f800	0x7fda00836180	12075337357374087	12075337357385270	12075337357509909	12075337357513065
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	870469	870474	4194304	256	0	0	28	24	83456	0x7fdb0c52f700	0x7fda008361c0	12075337357537942	12075337357549269	12075337357807986	12075337357828061
71	69	void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]	0	0	138	870469	870474	4194304	256	0	0	28	24	84928	0x7fdb0c52f600	0x7fda00836200	12075337357848739	12075337357861106	12075337358108464	12075337358136544
72	70	void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]	0	0	140	870469	870474	4194304	256	0	0	24	24	86400	0x7fdb0c52f500	0x7fda00836240	12075337358156912	12075337358169263	12075337358294222	12075337358297653
73	71	void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]	0	0	142	870469	870474	4194304	256	0	0	24	24	88128	0x7fdb0c52f400	0x7fda00836280	12075337358322920	12075337358334222	12075337358483340	12075337358486524
74	72	void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]	0	0	144	870469	870474	4194304	256	0	0	24	24	89600	0x7fdb0c52f300	0x7fda008362c0	12075337358540715	12075337358552939	12075337358678058	12075337358681386
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	870469	870474	4194304	256	0	0	28	24	92096	0x7fdb0c52fa00	0x7fda00836300	12075337358706373	12075337358717258	12075337358974056	12075337358997784
76	74	void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]	0	0	148	870469	870474	4194304	256	0	0	28	24	93568	0x7fdb0c52f900	0x7fda00836340	12075337359026588	12075337359039335	12075337359284933	12075337359309153
77	75	void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]	0	0	150	870469	870474	4194304	256	0	0	24	24	95040	0x7fdb0c52f800	0x7fda00836380	12075337359329110	12075337359341412	12075337359467491	12075337359470713
78	76	void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]	0	0	152	870469	870474	4194304	256	0	0	24	24	96768	0x7fdb0c52f700	0x7fda008363c0	12075337359496731	12075337359507971	12075337359665249	12075337359668320
79	77	void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]	0	0	154	870469	870474	4194304	256	0	0	24	24	98240	0x7fdb0c52f600	0x7fda00836400	12075337359704016	12075337359715169	12075337359839808	12075337359842935
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	870469	870474	4194304	256	0	0	28	24	100736	0x7fdb0c52f500	0x7fda00836440	12075337359868542	12075337359879647	12075337360136125	12075337360164101
81	79	void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]	0	0	158	870469	870474	4194304	256	0	0	28	24	102208	0x7fdb0c52f400	0x7fda00836480	12075337360183467	12075337360194364	12075337360438362	12075337360462736
82	80	void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]	0	0	160	870469	870474	4194304	256	0	0	24	24	103680	0x7fdb0c52f300	0x7fda008364c0	12075337360483134	12075337360494042	12075337360620280	12075337360623445
83	81	void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]	0	0	162	870469	870474	4194304	256	0	0	24	24	105664	0x7fdb0c52fa00	0x7fda00836500	12075337360650675	12075337360662040	12075337360827159	12075337360830369
84	82	void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]	0	0	164	870469	870474	4194304	256	0	0	24	24	107136	0x7fdb0c52f900	0x7fda00836540	12075337360865214	12075337360876118	12075337361001237	12075337361005094
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	870469	870474	4194304	256	0	0	28	24	109888	0x7fdb0c52f800	0x7fda00836580	12075337361030481	12075337361042037	12075337361298194	12075337361320881
86	84	void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]	0	0	168	870469	870474	4194304	256	0	0	28	24	111360	0x7fdb0c52f700	0x7fda008365c0	12075337361340848	12075337361352914	12075337361637871	12075337361656825
87	85	void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]	0	0	170	870469	870474	4194304	256	0	0	24	24	112832	0x7fdb0c52f600	0x7fda00836600	12075337361675910	12075337361687311	12075337361813069	12075337361816331
88	86	void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]	0	0	172	870469	870474	4194304	256	0	0	24	24	114816	0x7fdb0c52f500	0x7fda00836640	12075337361841658	12075337361853229	12075337362025708	12075337362028896
89	87	void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]	0	0	174	870469	870474	4194304	256	0	0	24	24	116288	0x7fdb0c52f400	0x7fda00836680	12075337362064793	12075337362076747	12075337362202986	12075337362206126
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	870469	870474	4194304	256	0	0	28	24	119040	0x7fdb0c52f300	0x7fda008366c0	12075337362231613	12075337362242506	12075337362495943	12075337362519017
91	89	void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]	0	0	178	870469	870474	4194304	256	0	0	28	24	120768	0x7fdb0c52fa00	0x7fda00836700	12075337362538744	12075337362550023	12075337362794500	12075337362817111
92	90	void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]	0	0	180	870469	870474	4194304	256	0	0	24	24	122496	0x7fdb0c52f900	0x7fda00836740	12075337362837399	12075337362848420	12075337362973699	12075337362976928
93	91	void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]	0	0	182	870469	870474	4194304	256	0	0	24	24	124736	0x7fdb0c52f800	0x7fda00836780	12075337363006292	12075337363017858	12075337363200577	12075337363203729
94	92	void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]	0	0	184	870469	870474	4194304	256	0	0	24	24	126464	0x7fdb0c52f700	0x7fda008367c0	12075337363240798	12075337363251296	12075337363377055	12075337363380177
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	870469	870474	4194304	256	0	0	20	24	127936	0x7fdb0c52f600	0x7fda00836800	12075337363406286	12075337363417535	12075337363682012	12075337363704710
96	94	void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]	0	0	188	870469	870474	4194304	256	0	0	28	24	129664	0x7fdb0c52f500	0x7fda00836840	12075337363724146	12075337363734812	12075337363981050	12075337364008926
97	95	void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]	0	0	190	870469	870474	4194304	256	0	0	24	24	131392	0x7fdb0c52f400	0x7fda00836880	12075337364028953	12075337364039929	12075337364166648	12075337364169734
98	96	void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]	0	0	192	870469	870474	4194304	256	0	0	24	24	133632	0x7fdb0c52f300	0x7fda008368c0	12075337364196434	12075337364207767	12075337364399126	12075337364421522
99	97	void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]	0	0	194	870469	870474	4194304	256	0	0	24	24	135360	0x7fdb0c52fa00	0x7fda00836900	12075337364450536	12075337364462005	12075337364588564	12075337364591879
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	870469	870474	4194304	256	0	0	20	24	136832	0x7fdb0c52f900	0x7fda00836940	12075337364617516	12075337364628724	12075337364910801	12075337364933554
101	99	void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]	0	0	198	870469	870474	4194304	256	0	0	28	24	138560	0x7fdb0c52f800	0x7fda00836980	12075337364953971	12075337364966800	12075337365214478	12075337365237749
102	100	void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]	0	0	200	870469	870474	4194304	256	0	0	24	24	140288	0x7fdb0c52f700	0x7fda008369c0	12075337365256734	12075337365268078	12075337365393517	12075337365396935
103	101	void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]	0	0	202	870469	870474	4194304	256	0	0	24	24	142784	0x7fdb0c52f600	0x7fda00836a00	12075337365422402	12075337365434796	12075337365645354	12075337365667778
104	102	void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]	0	0	204	870469	870474	4194304	256	0	0	24	24	144768	0x7fdb0c52f500	0x7fda00836a40	12075337365697643	12075337365709834	12075337365837992	12075337365841340
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	870469	870474	4194304	256	0	0	20	24	145728	0x7fdb0c52f400	0x7fda00836a80	12075337365867850	12075337365880552	12075337366186789	12075337366216207
106	104	void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]	0	0	208	870469	870474	4194304	256	0	0	28	24	147712	0x7fdb0c52f300	0x7fda00836ac0	12075337366239099	12075337366252069	12075337366502466	12075337366526423
107	105	void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]	0	0	210	870469	870474	4194304	256	0	0	24	24	149696	0x7fdb0c52fa00	0x7fda00836b00	12075337366547953	12075337366560706	12075337366687265	12075337366690528
108	106	void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]	0	0	212	870469	870474	4194304	256	0	0	24	24	152192	0x7fdb0c52f900	0x7fda00836b40	12075337366715976	12075337366727104	12075337366954942	12075337366977301
109	107	void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]	0	0	214	870469	870474	4194304	256	0	0	24	24	154176	0x7fdb0c52f800	0x7fda00836b80	12075337367012547	12075337367023422	12075337367150140	12075337367153479
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	870469	870474	4194304	256	0	0	20	24	155648	0x7fdb0c52f700	0x7fda00836bc0	12075337367179637	12075337367190780	12075337367519417	12075337367542181
111	109	void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]	0	0	218	870469	870474	4194304	256	0	0	28	24	157632	0x7fdb0c52f600	0x7fda00836c00	12075337367562739	12075337367573656	12075337367821334	12075337367844212
112	110	void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]	0	0	220	870469	870474	4194304	256	0	0	24	24	159616	0x7fdb0c52f500	0x7fda00836c40	12075337367863217	12075337367874294	12075337368002453	12075337368005952
113	111	void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]	0	0	222	870469	870474	4194304	256	0	0	24	24	162368	0x7fdb0c52f400	0x7fda00836c80	12075337368032882	12075337368044052	12075337368321490	12075337368344051
114	112	void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]	0	0	224	870469	870474	4194304	256	0	0	24	24	164608	0x7fdb0c52f300	0x7fda00836cc0	12075337368373425	12075337368385489	12075337368514928	12075337368518725
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	870469	870474	4194304	256	0	0	20	24	166336	0x7fdb0c52fa00	0x7fda00836d00	12075337368543692	12075337368556687	12075337368932524	12075337368955076
116	114	void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]	0	0	228	870469	870474	4194304	256	0	0	28	24	168576	0x7fdb0c52f900	0x7fda00836d40	12075337368973771	12075337368985964	12075337369243081	12075337369266395
117	115	void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]	0	0	230	870469	870474	4194304	256	0	0	24	24	170816	0x7fdb0c52f800	0x7fda00836d80	12075337369285109	12075337369297481	12075337369426599	12075337369429999
118	116	void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]	0	0	232	870469	870474	4194304	256	0	0	24	24	174080	0x7fdb0c52f700	0x7fda00836dc0	12075337369454875	12075337369467399	12075337369745477	12075337369768057
119	117	void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]	0	0	234	870469	870474	4194304	256	0	0	24	24	176576	0x7fdb0c52f600	0x7fda00836e00	12075337369797652	12075337369810116	12075337369945955	12075337369949274
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	870469	870474	4194304	256	0	0	20	24	177792	0x7fdb0c52f500	0x7fda00836e40	12075337369974250	12075337369986274	12075337370408990	12075337370438392
121	119	void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]	0	0	238	870469	870474	4194304	256	0	0	28	24	180288	0x7fdb0c52f400	0x7fda00836e80	12075337370460163	12075337370472990	12075337370740187	12075337370762324
122	120	void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]	0	0	240	870469	870474	4194304	256	0	0	24	24	182784	0x7fdb0c52f300	0x7fda00836ec0	12075337370781570	12075337370793627	12075337370929466	12075337370932721
123	121	void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]	0	0	242	870469	870474	4194304	256	0	0	24	24	186304	0x7fdb0c52fa00	0x7fda00836f00	12075337370957497	12075337370968505	12075337371282742	12075337371306145
124	122	void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]	0	0	244	870469	870474	4194304	256	0	0	24	24	189312	0x7fdb0c52f900	0x7fda00836f40	12075337371340879	12075337371353782	12075337371505300	12075337371508751
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	870469	870474	4194304	256	0	0	20	24	190784	0x7fdb0c52f800	0x7fda00836f80	12075337371534529	12075337371545780	12075337372062255	12075337372090312
126	124	void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]	0	0	248	870469	870474	4194304	256	0	0	28	24	193792	0x7fdb0c52f700	0x7fda00836fc0	12075337372110800	12075337372121775	12075337372417292	12075337372443638
127	125	void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]	0	0	250	870469	870474	4194304	256	0	0	24	24	196800	0x7fdb0c52f600	0x7fda00837000	12075337372462734	12075337372475371	12075337372628810	12075337372632078
128	126	void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]	0	0	252	870469	870474	4194304	256	0	0	24	24	201088	0x7fdb0c52f500	0x7fda00837040	12075337372658538	12075337372669610	12075337373053606	12075337373081774
129	127	void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]	0	0	254	870469	870474	4194304	256	0	0	24	24	204608	0x7fdb0c52f400	0x7fda00837080	12075337373112221	12075337373123205	12075337373299364	12075337373302705
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	870469	870474	4194304	256	0	0	20	24	206080	0x7fdb0c52f300	0x7fda008370c0	12075337373327781	12075337373339043	12075337373948958	12075337373975715
131	129	void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]	0	0	258	870469	870474	4194304	256	0	0	28	24	209600	0x7fdb0c52fa00	0x7fda00837100	12075337373995612	12075337374006557	12075337374333114	12075337374364227
132	130	void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]	0	0	260	870469	870474	4194304	256	0	0	24	24	213120	0x7fdb0c52f900	0x7fda00837140	12075337374384895	12075337374395834	12075337374573432	12075337374576782
133	131	void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]	0	0	262	870469	870474	4194304	256	0	0	24	24	218176	0x7fdb0c52f800	0x7fda00837180	12075337374602440	12075337374615032	12075337375091347	12075337375119771
134	132	void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]	0	0	264	870469	870474	4194304	256	0	0	24	24	222208	0x7fdb0c52f700	0x7fda008371c0	12075337375149877	12075337375162067	12075337375360305	12075337375386587
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	870469	870474	4194304	256	0	0	20	24	223936	0x7fdb0c52f600	0x7fda00837200	12075337375405061	12075337375416624	12075337376118378	12075337376146108
136	134	void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]	0	0	268	870469	870474	4194304	256	0	0	28	24	227968	0x7fdb0c52f500	0x7fda00837240	12075337376168830	12075337376180137	12075337376552614	12075337376578612
137	135	void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]	0	0	270	870469	870474	4194304	256	0	0	24	24	232000	0x7fdb0c52f400	0x7fda00837280	12075337376599290	12075337376610213	12075337376811332	12075337376837272
138	136	void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]	0	0	272	870469	870474	4194304	256	0	0	24	24	237824	0x7fdb0c52f300	0x7fda008372c0	12075337376857169	12075337376869571	12075337377394846	12075337377427329
139	137	void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]	0	0	274	870469	870474	4194304	256	0	0	24	24	242368	0x7fdb0c52fa00	0x7fda00837300	12075337377459589	12075337377475645	12075337377699643	12075337377722056
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	870469	870474	4194304	256	0	0	20	24	243584	0x7fdb0c52f900	0x7fda00837340	12075337377744969	12075337377757563	12075337378555635	12075337378583206
141	139	void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]	0	0	278	870469	870474	4194304	256	0	0	28	24	248128	0x7fdb0c52f800	0x7fda00837380	12075337378604075	12075337378616115	12075337379038191	12075337379065973
142	140	void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]	0	0	280	870469	870474	4194304	256	0	0	24	24	252672	0x7fdb0c52f700	0x7fda008373c0	12075337379085770	12075337379098670	12075337379324428	12075337379350111
143	141	void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]	0	0	282	870469	870474	4194304	256	0	0	24	24	259264	0x7fdb0c52f600	0x7fda00837400	12075337379371952	12075337379383948	12075337379980582	12075337380008515
144	142	void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]	0	0	284	870469	870474	4194304	256	0	0	24	24	264832	0x7fdb0c52f500	0x7fda00837440	12075337380038801	12075337380051142	12075337380319939	12075337380346683
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	870469	870474	4194304	256	0	0	20	24	266304	0x7fdb0c52f400	0x7fda00837480	12075337380367772	12075337380378819	12075337381363130	12075337381390583
146	144	void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]	0	0	288	870469	870474	4194304	256	0	0	28	24	271872	0x7fdb0c52f300	0x7fda008374c0	12075337381409388	12075337381420089	12075337381950484	12075337381976752
147	145	void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]	0	0	290	870469	870474	4194304	256	0	0	24	24	277440	0x7fdb0c52fa00	0x7fda00837500	12075337381998232	12075337382012724	12075337382284241	12075337382315331
148	146	void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]	0	0	292	870469	870474	4194304	256	0	0	12	24	279936	0x7fdb0c52f900	0x7fda00837540	12075337382335418	12075337382347121	12075337383084074	12075337383111971
149	147	void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]	0	0	294	870469	870474	4194304	256	0	0	24	24	286528	0x7fdb0c52f800	0x7fda00837580	12075337383143951	12075337383155273	12075337383469670	12075337383492288
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	870469	870474	4194304	256	0	0	20	24	288000	0x7fdb0c52f700	0x7fda008375c0	12075337383516183	12075337383527270	12075337384697979	12075337384725349
151	149	void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]	0	0	298	870469	870474	4194304	256	0	0	32	24	294592	0x7fdb0c52f600	0x7fda00837600	12075337384748362	12075337384760538	12075337385370613	12075337385398530
152	150	void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]	0	0	300	870469	870474	4194304	256	0	0	24	24	301184	0x7fdb0c52f500	0x7fda00837640	12075337385420220	12075337385432692	12075337385751249	12075337385774629
153	151	void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]	0	0	302	870469	870474	4194304	256	0	0	12	24	304960	0x7fdb0c52f400	0x7fda00837680	12075337385797962	12075337385810449	12075337386687721	12075337386716880
154	152	void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]	0	0	304	870469	870474	4194304	256	0	0	24	24	313600	0x7fdb0c52f300	0x7fda008376c0	12075337386754109	12075337386765160	12075337387175716	12075337387202011
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	870469	870474	4194304	256	0	0	20	24	314816	0x7fdb0c52fa00	0x7fda00837700	12075337387225846	12075337387236996	12075337388785941	12075337388855794
156	154	void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]	0	0	308	870469	870474	4194304	256	0	0	20	24	317568	0x7fdb0c52f900	0x7fda00837740	12075337388877865	12075337388889251	12075337389683164	12075337389754875
157	155	void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]	0	0	310	870469	870474	4194304	256	0	0	12	24	320320	0x7fdb0c52f800	0x7fda00837780	12075337389775072	12075337389787642	12075337390200919	12075337390270262
158	156	void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]	0	0	312	870469	870474	4194304	256	0	0	12	24	324096	0x7fdb0c52f700	0x7fda008377c0	12075337390298505	12075337390310998	12075337391469229	12075337391535383
159	157	void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]	0	0	314	870469	870474	4194304	256	0	0	12	24	328896	0x7fdb0c52f600	0x7fda00837800	12075337391570739	12075337391584108	12075337392369221	12075337392435035
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	870469	870474	4194304	256	0	0	20	24	330112	0x7fdb0c52f500	0x7fda00837840	12075337392461244	12075337392474340	12075337395525356	12075337395596499
161	159	void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]	0	0	318	870469	870474	4194304	256	0	0	20	24	332864	0x7fdb0c52f400	0x7fda00837880	12075337395617478	12075337395630475	12075337397171742	12075337397242879
162	160	void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]	0	0	320	870469	870474	4194304	256	0	0	12	24	335616	0x7fdb0c52f300	0x7fda008378c0	12075337397263707	12075337397276541	12075337398065495	12075337398135898
163	161	void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]	0	0	322	870469	870474	4194304	256	0	0	12	24	339392	0x7fdb0c52fa00	0x7fda00837900	12075337398157668	12075337398170774	12075337400451876	12075337400522212
164	162	void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]	0	0	324	870469	870474	4194304	256	0	0	12	24	344192	0x7fdb0c52f900	0x7fda00837940	12075337400553450	12075337400566275	12075337402101782	12075337402172268
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	870469	870474	4194304	256	0	0	20	24	345408	0x7fdb0c52f800	0x7fda00837980	12075337402194609	12075337402208341	12075337408263012	12075337408333256
166	164	void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]	0	0	328	870469	870474	4194304	256	0	0	20	24	348160	0x7fdb0c52f700	0x7fda008379c0	12075337408356559	12075337408369251	12075337411410667	12075337411481315
167	165	void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]	0	0	330	870469	870474	4194304	256	0	0	12	24	350912	0x7fdb0c52f600	0x7fda00837a00	12075337411503036	12075337411516106	12075337413054813	12075337413127033
168	166	void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]	0	0	332	870469	870474	4194304	256	0	0	12	24	0	0x7fdb0c52f500	0x7fda00837a40	12075337413146820	12075337413159772	12075337417685176	12075337417742580