rocm-systems/tests/workloads/D_val_int2/mi200/timestamps.csv

Index,KernelName,gpu-id,queue-id,queue-index,pid,tid,grd,wgr,lds,scr,vgpr,sgpr,fbar,sig,obj,DispatchNs,BeginNs,EndNs,CompleteNs
0,"__amd_rocclr_fillBufferAligned.kd",0,0,0,92848,92848,33554432,256,0,0,4,32,4160,0x0,0x7f8e1d204280,15168985964210,15168985990090,15168986231051,15168986352141
1,"void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]",0,0,2,92848,92848,32768,256,0,0,12,24,13888,0x0,0x7f8e1d223f80,15168991007582,15168991022750,15168991036030,15168991053271
2,"void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]",0,0,5,92848,92848,4194304,256,0,0,12,24,14336,0x7f8e20219380,0x7f8e1d223fc0,15168991057700,15168991106270,15168991198591,15168991200567
3,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,8,92848,92848,4194304,256,0,0,16,24,15040,0x7f8e20219200,0x7f8e1d224000,15168991221467,15168991230911,15168991410912,15168991412952
4,"void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]",0,0,11,92848,92848,4194304,256,0,0,16,24,15488,0x7f8e20219080,0x7f8e1d224040,15168991429042,15168991437472,15168991617473,15168991619277
5,"void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]",0,0,14,92848,92848,4194304,256,0,0,12,24,15936,0x7f8e1da7df00,0x7f8e1d224080,15168991634207,15168991642753,15168991736513,15168991738414
6,"void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]",0,0,17,92848,92848,4194304,256,0,0,12,24,16384,0x7f8e1da7dd80,0x7f8e1d2240c0,15168991754834,15168991763233,15168991855234,15168991857092
7,"void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]",0,0,20,92848,92848,4194304,256,0,0,12,24,17088,0x7f8e1da7dc00,0x7f8e1d224100,15168991883971,15168991894274,15168991986754,15168991988598
8,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,23,92848,92848,4194304,256,0,0,16,24,17792,0x7f8e1da7da80,0x7f8e1d224140,15168992007068,15168992016034,15168992196195,15168992198103
9,"void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]",0,0,26,92848,92848,4194304,256,0,0,16,24,18496,0x7f8e1da7d900,0x7f8e1d224180,15168992216993,15168992237795,15168992416996,15168992418838
10,"void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]",0,0,29,92848,92848,4194304,256,0,0,16,24,19200,0x7f8e1da7d780,0x7f8e1d2241c0,15168992435878,15168992450596,15168992544196,15168992546365
11,"void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]",0,0,32,92848,92848,4194304,256,0,0,16,24,19904,0x7f8e1da7d600,0x7f8e1d224200,15168992564635,15168992574277,15168992667237,15168992669162
12,"void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]",0,0,35,92848,92848,4194304,256,0,0,12,24,20608,0x7f8e20219480,0x7f8e1d224240,15168992692242,15168992710117,15168992802917,15168992805129
13,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,38,92848,92848,4194304,256,0,0,16,24,21312,0x7f8e20219300,0x7f8e1d224280,15168992825359,15168992835558,15168993015238,15168993017064
14,"void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]",0,0,41,92848,92848,4194304,256,0,0,16,24,22016,0x7f8e20219180,0x7f8e1d2242c0,15168993035914,15168993045159,15168993225959,15168993227649
15,"void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]",0,0,44,92848,92848,4194304,256,0,0,16,24,22720,0x7f8e20219000,0x7f8e1d224300,15168993240859,15168993250119,15168993343400,15168993345296
16,"void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]",0,0,47,92848,92848,4194304,256,0,0,12,24,23424,0x7f8e1da7de80,0x7f8e1d224340,15168993366646,15168993387240,15168993481640,15168993483773
17,"void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]",0,0,50,92848,92848,4194304,256,0,0,12,24,24128,0x7f8e1da7dd00,0x7f8e1d224380,15168993506432,15168993516520,15168993609801,15168993611600
18,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,53,92848,92848,4194304,256,0,0,16,24,24832,0x7f8e1da7db80,0x7f8e1d2243c0,15168993630390,15168993639081,15168993818762,15168993820605
19,"void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]",0,0,56,92848,92848,4194304,256,0,0,16,24,25536,0x7f8e1da7da00,0x7f8e1d224400,15168993838045,15168993847242,15168994028043,15168994029800
20,"void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]",0,0,59,92848,92848,4194304,256,0,0,12,24,26240,0x7f8e1da7d880,0x7f8e1d224440,15168994047210,15168994056203,15168994149803,15168994151867
21,"void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]",0,0,62,92848,92848,4194304,256,0,0,12,24,26944,0x7f8e1da7d700,0x7f8e1d224480,15168994171977,15168994189163,15168994282764,15168994284914
22,"void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]",0,0,65,92848,92848,4194304,256,0,0,12,24,27648,0x7f8e20219580,0x7f8e1d2244c0,15168994306234,15168994316684,15168994409004,15168994410831
23,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,68,92848,92848,4194304,256,0,0,16,24,28608,0x7f8e20219400,0x7f8e1d224500,15168994430821,15168994440844,15168994620205,15168994622516
24,"void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]",0,0,71,92848,92848,4194304,256,0,0,16,24,29312,0x7f8e20219280,0x7f8e1d224540,15168994637106,15168994646445,15168994826126,15168994827951
25,"void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]",0,0,74,92848,92848,4194304,256,0,0,12,24,30016,0x7f8e20219100,0x7f8e1d224580,15168994845841,15168994854926,15168994946926,15168994948788
26,"void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]",0,0,77,92848,92848,4194304,256,0,0,8,24,30976,0x7f8e1da7df80,0x7f8e1d2245c0,15168994967378,15168994992527,15168995085007,15168995087035
27,"void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]",0,0,80,92848,92848,4194304,256,0,0,12,24,31680,0x7f8e1da7de00,0x7f8e1d224600,15168995108085,15168995118447,15168995210608,15168995212532
28,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,83,92848,92848,4194304,256,0,0,16,24,32640,0x7f8e1da7dc80,0x7f8e1d224640,15168995233422,15168995243568,15168995423568,15168995425327
29,"void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]",0,0,86,92848,92848,4194304,256,0,0,12,24,33600,0x7f8e1da7db00,0x7f8e1d224680,15168995445547,15168995455089,15168995637969,15168995640182
30,"void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]",0,0,89,92848,92848,4194304,256,0,0,12,24,34560,0x7f8e1da7d980,0x7f8e1d2246c0,15168995660712,15168995677170,15168995771250,15168995773389
31,"void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]",0,0,92,92848,92848,4194304,256,0,0,8,24,35520,0x7f8e1da7d800,0x7f8e1d224700,15168995791209,15168995801330,15168995895410,15168995897386
32,"void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]",0,0,95,92848,92848,4194304,256,0,0,12,24,36224,0x7f8e1da7d680,0x7f8e1d224740,15168995920056,15168995929171,15168996022451,15168996024293
33,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,98,92848,92848,4194304,256,0,0,12,24,37184,0x7f8e20219500,0x7f8e1d224780,15168996042183,15168996050451,15168996230452,15168996231708
34,"void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]",0,0,101,92848,92848,4194304,256,0,0,12,24,38144,0x7f8e20219380,0x7f8e1d2247c0,15168996246168,15168996255092,15168996435573,15168996436903
35,"void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]",0,0,104,92848,92848,4194304,256,0,0,12,24,39104,0x7f8e20219200,0x7f8e1d224800,15168996452023,15168996460053,15168996553333,15168996554581
36,"void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]",0,0,107,92848,92848,4194304,256,0,0,8,24,40320,0x7f8e20219080,0x7f8e1d224840,15168996569310,15168996577653,15168996671574,15168996672868
37,"void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]",0,0,110,92848,92848,4194304,256,0,0,12,24,41280,0x7f8e1da7df00,0x7f8e1d224880,15168996694127,15168996704694,15168996798934,15168996800225
38,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,113,92848,92848,4194304,256,0,0,12,24,42240,0x7f8e1da7dd80,0x7f8e1d2248c0,15168996817135,15168996825974,15168997004535,15168997005880
39,"void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]",0,0,116,92848,92848,4194304,256,0,0,12,24,43200,0x7f8e1da7dc00,0x7f8e1d224900,15168997020850,15168997035415,15168997215896,15168997217255
40,"void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]",0,0,119,92848,92848,4194304,256,0,0,8,24,44160,0x7f8e1da7da80,0x7f8e1d224940,15168997232855,15168997242456,15168997336536,15168997337792
41,"void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]",0,0,122,92848,92848,4194304,256,0,0,12,24,45376,0x7f8e1da7d900,0x7f8e1d224980,15168997354592,15168997362457,15168997458137,15168997459489
42,"void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]",0,0,125,92848,92848,4194304,256,0,0,12,24,46336,0x7f8e1da7d780,0x7f8e1d2249c0,15168997479779,15168997490617,15168997583417,15168997584776
43,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,128,92848,92848,4194304,256,0,0,12,24,47552,0x7f8e1da7d600,0x7f8e1d224a00,15168997599806,15168997608058,15168997787258,15168997788552
44,"void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]",0,0,131,92848,92848,4194304,256,0,0,12,24,48512,0x7f8e20219480,0x7f8e1d224a40,15168997802771,15168997811898,15168997992699,15168997993927
45,"void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]",0,0,134,92848,92848,4194304,256,0,0,8,24,49472,0x7f8e20219300,0x7f8e1d224a80,15168998007847,15168998017019,15168998109820,15168998111524
46,"void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]",0,0,137,92848,92848,4194304,256,0,0,12,24,50688,0x7f8e20219180,0x7f8e1d224ac0,15168998127214,15168998136380,15168998229820,15168998231381
47,"void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]",0,0,140,92848,92848,4194304,256,0,0,12,24,51648,0x7f8e20219000,0x7f8e1d224b00,15168998251531,15168998262140,15168998353981,15168998355338
48,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,143,92848,92848,4194304,256,0,0,12,24,52864,0x7f8e1da7de80,0x7f8e1d224b40,15168998370548,15168998378621,15168998557662,15168998559084
49,"void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]",0,0,146,92848,92848,4194304,256,0,0,12,24,54080,0x7f8e1da7dd00,0x7f8e1d224b80,15168998573133,15168998581022,15168998761822,15168998763029
50,"void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]",0,0,149,92848,92848,4194304,256,0,0,8,24,55296,0x7f8e1da7db80,0x7f8e1d224bc0,15168998777778,15168998785663,15168998879743,15168998881346
51,"void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]",0,0,152,92848,92848,4194304,256,0,0,12,24,56768,0x7f8e1da7da00,0x7f8e1d224c00,15168998895886,15168998904223,15168998999583,15168999000883
52,"void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]",0,0,155,92848,92848,4194304,256,0,0,12,24,57728,0x7f8e1da7d880,0x7f8e1d224c40,15168999020613,15168999031424,15168999124704,15168999126120
53,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,158,92848,92848,4194304,256,0,0,12,24,58944,0x7f8e1da7d700,0x7f8e1d224c80,15168999141260,15168999149184,15168999328545,15168999329895
54,"void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]",0,0,161,92848,92848,4194304,256,0,0,12,24,60160,0x7f8e20219580,0x7f8e1d224cc0,15168999343585,15168999352065,15168999533186,15168999534461
55,"void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]",0,0,164,92848,92848,4194304,256,0,0,8,24,61376,0x7f8e20219400,0x7f8e1d224d00,15168999548950,15168999557186,15168999651586,15168999652858
56,"void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]",0,0,167,92848,92848,4194304,256,0,0,12,24,62848,0x7f8e20219280,0x7f8e1d224d40,15168999672217,15168999679746,15168999780227,15168999781635
57,"void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]",0,0,170,92848,92848,4194304,256,0,0,12,24,63808,0x7f8e20219100,0x7f8e1d224d80,15168999803764,15168999814467,15168999907907,15168999909252
58,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,173,92848,92848,4194304,256,0,0,12,24,65024,0x7f8e1da7df80,0x7f8e1d224dc0,15168999929581,15168999940067,15169000120228,15169000121567
59,"void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]",0,0,176,92848,92848,4194304,256,0,0,12,24,66240,0x7f8e1da7de00,0x7f8e1d224e00,15169000136796,15169000145988,15169000325989,15169000327222
60,"void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]",0,0,179,92848,92848,4194304,256,0,0,8,24,67456,0x7f8e1da7dc80,0x7f8e1d224e40,15169000342042,15169000351109,15169000445189,15169000446529
61,"void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]",0,0,182,92848,92848,4194304,256,0,0,12,24,68928,0x7f8e1da7db00,0x7f8e1d224e80,15169000461519,15169000469990,15169000574150,15169000575456
62,"void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]",0,0,185,92848,92848,4194304,256,0,0,12,24,70144,0x7f8e1da7d980,0x7f8e1d224ec0,15169000595376,15169000606470,15169000698791,15169000700083
63,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,188,92848,92848,4194304,256,0,0,12,24,71616,0x7f8e1da7d800,0x7f8e1d224f00,15169000714863,15169000723591,15169000902311,15169000903728
64,"void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]",0,0,191,92848,92848,4194304,256,0,0,12,24,72832,0x7f8e1da7d680,0x7f8e1d224f40,15169000921458,15169000930952,15169001111432,15169001112734
65,"void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]",0,0,194,92848,92848,4194304,256,0,0,8,24,74048,0x7f8e20219500,0x7f8e1d224f80,15169001127143,15169001136712,15169001229993,15169001231191
66,"void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]",0,0,197,92848,92848,4194304,256,0,0,12,24,75776,0x7f8e20219380,0x7f8e1d224fc0,15169001246980,15169001255593,15169001362313,15169001363668
67,"void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]",0,0,200,92848,92848,4194304,256,0,0,12,24,76992,0x7f8e20219200,0x7f8e1d225000,15169001385087,15169001395913,15169001490794,15169001492075
68,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,203,92848,92848,4194304,256,0,0,12,24,78464,0x7f8e20219080,0x7f8e1d225040,15169001506804,15169001515594,15169001696875,15169001698070
69,"void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]",0,0,206,92848,92848,4194304,256,0,0,12,24,79936,0x7f8e1da7df00,0x7f8e1d225080,15169001712659,15169001721835,15169001900876,15169001902175
70,"void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]",0,0,209,92848,92848,4194304,256,0,0,8,24,81408,0x7f8e1da7dd80,0x7f8e1d2250c0,15169001916015,15169001925516,15169002021356,15169002022652
71,"void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]",0,0,212,92848,92848,4194304,256,0,0,12,24,83136,0x7f8e1da7dc00,0x7f8e1d225100,15169002036662,15169002045196,15169002160397,15169002161749
72,"void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]",0,0,215,92848,92848,4194304,256,0,0,12,24,84352,0x7f8e1da7da80,0x7f8e1d225140,15169002196048,15169002205677,15169002299117,15169002300466
73,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,218,92848,92848,4194304,256,0,0,12,24,85824,0x7f8e1da7d900,0x7f8e1d225180,15169002314735,15169002322637,15169002502478,15169002503671
74,"void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]",0,0,221,92848,92848,4194304,256,0,0,12,24,87296,0x7f8e1da7d780,0x7f8e1d2251c0,15169002518410,15169002526318,15169002705999,15169002707376
75,"void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]",0,0,224,92848,92848,4194304,256,0,0,12,24,88768,0x7f8e1da7d600,0x7f8e1d225200,15169002722206,15169002730159,15169002825359,15169002826693
76,"void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]",0,0,227,92848,92848,4194304,256,0,0,12,24,90752,0x7f8e20219480,0x7f8e1d225240,15169002840723,15169002849520,15169002970800,15169002972140
77,"void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]",0,0,230,92848,92848,4194304,256,0,0,12,24,91968,0x7f8e20219300,0x7f8e1d225280,15169002991809,15169003002640,15169003096561,15169003097997
78,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,233,92848,92848,4194304,256,0,0,12,24,93440,0x7f8e20219180,0x7f8e1d2252c0,15169003112486,15169003120881,15169003301041,15169003302502
79,"void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]",0,0,236,92848,92848,4194304,256,0,0,12,24,94912,0x7f8e20219000,0x7f8e1d225300,15169003316532,15169003324722,15169003503922,15169003505427
80,"void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]",0,0,239,92848,92848,4194304,256,0,0,12,24,96384,0x7f8e1da7de80,0x7f8e1d225340,15169003520667,15169003528562,15169003623763,15169003625294
81,"void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]",0,0,242,92848,92848,4194304,256,0,0,12,24,98368,0x7f8e1da7dd00,0x7f8e1d225380,15169003641214,15169003649043,15169003775283,15169003776501
82,"void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]",0,0,245,92848,92848,4194304,256,0,0,12,24,99584,0x7f8e1da7db80,0x7f8e1d2253c0,15169003799890,15169003810484,15169003904564,15169003906048
83,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,248,92848,92848,4194304,256,0,0,12,24,101312,0x7f8e1da7da00,0x7f8e1d225400,15169003922957,15169003930964,15169004110805,15169004112143
84,"void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]",0,0,251,92848,92848,4194304,256,0,0,12,24,102784,0x7f8e1da7d880,0x7f8e1d225440,15169004126572,15169004135765,15169004317366,15169004318678
85,"void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]",0,0,254,92848,92848,4194304,256,0,0,12,24,104256,0x7f8e1da7d700,0x7f8e1d225480,15169004333597,15169004342806,15169004437046,15169004438405
86,"void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]",0,0,257,92848,92848,4194304,256,0,0,12,24,106240,0x7f8e20219580,0x7f8e1d2254c0,15169004451945,15169004460246,15169004590647,15169004591881
87,"void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]",0,0,260,92848,92848,4194304,256,0,0,12,24,107712,0x7f8e20219400,0x7f8e1d225500,15169004612351,15169004622807,15169004715767,15169004717078
88,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,263,92848,92848,4194304,256,0,0,16,24,109440,0x7f8e20219280,0x7f8e1d225540,15169004733218,15169004742167,15169004920408,15169004921774
89,"void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]",0,0,266,92848,92848,4194304,256,0,0,16,24,111168,0x7f8e20219100,0x7f8e1d225580,15169004936183,15169004945368,15169005125689,15169005126899
90,"void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]",0,0,269,92848,92848,4194304,256,0,0,8,24,112896,0x7f8e1da7df80,0x7f8e1d2255c0,15169005141308,15169005150009,15169005246170,15169005247556
91,"void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]",0,0,272,92848,92848,4194304,256,0,0,12,24,115136,0x7f8e1da7de00,0x7f8e1d225600,15169005262055,15169005270490,15169005408730,15169005409902
92,"void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]",0,0,275,92848,92848,4194304,256,0,0,12,24,116608,0x7f8e1da7dc80,0x7f8e1d225640,15169005431291,15169005442010,15169005537051,15169005538329
93,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,278,92848,92848,4194304,256,0,0,16,24,118336,0x7f8e1da7db00,0x7f8e1d225680,15169005555199,15169005562971,15169005744732,15169005746144
94,"void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]",0,0,281,92848,92848,4194304,256,0,0,16,24,120064,0x7f8e1da7d980,0x7f8e1d2256c0,15169005760794,15169005770492,15169005950173,15169005951399
95,"void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]",0,0,284,92848,92848,4194304,256,0,0,12,24,121792,0x7f8e1da7d800,0x7f8e1d225700,15169005964979,15169005974013,15169006071293,15169006072596
96,"void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]",0,0,287,92848,92848,4194304,256,0,0,12,24,124032,0x7f8e1da7d680,0x7f8e1d225740,15169006096606,15169006106013,15169006250494,15169006252202
97,"void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]",0,0,290,92848,92848,4194304,256,0,0,12,24,125504,0x7f8e20219500,0x7f8e1d225780,15169006274292,15169006284734,15169006380094,15169006381439
98,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,293,92848,92848,4194304,256,0,0,12,24,127488,0x7f8e20219380,0x7f8e1d2257c0,15169006396519,15169006404574,15169006586655,15169006588124
99,"void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]",0,0,296,92848,92848,4194304,256,0,0,12,24,129216,0x7f8e20219200,0x7f8e1d225800,15169006602704,15169006610815,15169006791936,15169006793309
100,"void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]",0,0,299,92848,92848,4194304,256,0,0,12,24,130944,0x7f8e20219080,0x7f8e1d225840,15169006807319,15169006816096,15169006916097,15169006917426
101,"void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]",0,0,302,92848,92848,4194304,256,0,0,12,24,133440,0x7f8e1da7df00,0x7f8e1d225880,15169006931506,15169006939617,15169007095137,15169007096542
102,"void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]",0,0,305,92848,92848,4194304,256,0,0,12,24,134912,0x7f8e1da7dd80,0x7f8e1d2258c0,15169007116722,15169007127297,15169007223138,15169007224599
103,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,308,92848,92848,4194304,256,0,0,12,24,136896,0x7f8e1da7dc00,0x7f8e1d225900,15169007239859,15169007247938,15169007427139,15169007428634
104,"void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]",0,0,311,92848,92848,4194304,256,0,0,12,24,138880,0x7f8e1da7da80,0x7f8e1d225940,15169007443974,15169007451939,15169007632580,15169007634120
105,"void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]",0,0,314,92848,92848,4194304,256,0,0,12,24,140864,0x7f8e1da7d900,0x7f8e1d225980,15169007648059,15169007656260,15169007762980,15169007764447
106,"void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]",0,0,317,92848,92848,4194304,256,0,0,12,24,143616,0x7f8e1da7d780,0x7f8e1d2259c0,15169007781216,15169007789700,15169007961381,15169007962762
107,"void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]",0,0,320,92848,92848,4194304,256,0,0,12,24,145344,0x7f8e1da7d600,0x7f8e1d225a00,15169007983841,15169007994341,15169008089861,15169008091239
108,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,323,92848,92848,4194304,256,0,0,12,24,147584,0x7f8e20219480,0x7f8e1d225a40,15169008106108,15169008114022,15169008294822,15169008296114
109,"void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]",0,0,326,92848,92848,4194304,256,0,0,12,24,149568,0x7f8e20219300,0x7f8e1d225a80,15169008310804,15169008319462,15169008499463,15169008500789
110,"void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]",0,0,329,92848,92848,4194304,256,0,0,12,24,151552,0x7f8e20219180,0x7f8e1d225ac0,15169008515009,15169008523303,15169008633704,15169008634896
111,"void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]",0,0,332,92848,92848,4194304,256,0,0,12,24,154304,0x7f8e20219000,0x7f8e1d225b00,15169008649276,15169008657704,15169008839145,15169008840491
112,"void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]",0,0,335,92848,92848,4194304,256,0,0,12,24,156288,0x7f8e1da7de80,0x7f8e1d225b40,15169008862131,15169008871945,15169008973065,15169008974428
113,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,338,92848,92848,4194304,256,0,0,12,24,158784,0x7f8e1da7dd00,0x7f8e1d225b80,15169008990398,15169008998985,15169009185226,15169009186533
114,"void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]",0,0,341,92848,92848,4194304,256,0,0,12,24,161024,0x7f8e1da7db80,0x7f8e1d225bc0,15169009200073,15169009208586,15169009393547,15169009394968
115,"void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]",0,0,344,92848,92848,4194304,256,0,0,12,24,163264,0x7f8e1da7da00,0x7f8e1d225c00,15169009409188,15169009417547,15169009539148,15169009540325
116,"void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]",0,0,347,92848,92848,4194304,256,0,0,12,24,166528,0x7f8e1da7d880,0x7f8e1d225c40,15169009554274,15169009562988,15169009769548,15169009791099
117,"void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]",0,0,350,92848,92848,4194304,256,0,0,12,24,168512,0x7f8e1da7d700,0x7f8e1d225c80,15169009806588,15169009816909,15169009926989,15169009928366
118,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,353,92848,92848,4194304,256,0,0,12,24,171264,0x7f8e20219580,0x7f8e1d225cc0,15169009943735,15169009952269,15169010134350,15169010135761
119,"void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]",0,0,356,92848,92848,4194304,256,0,0,12,24,173760,0x7f8e20219400,0x7f8e1d225d00,15169010150290,15169010158510,15169010345711,15169010347246
120,"void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]",0,0,359,92848,92848,4194304,256,0,0,12,24,176256,0x7f8e20219280,0x7f8e1d225d40,15169010361975,15169010370191,15169010504752,15169010506202
121,"void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]",0,0,362,92848,92848,4194304,256,0,0,12,24,179776,0x7f8e20219100,0x7f8e1d225d80,15169010521272,15169010529392,15169010761713,15169010806235
122,"void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]",0,0,365,92848,92848,4194304,256,0,0,12,24,182272,0x7f8e1da7df80,0x7f8e1d225dc0,15169010821025,15169010831953,15169010958513,15169010959901
123,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,368,92848,92848,4194304,256,0,0,12,24,185536,0x7f8e1da7de00,0x7f8e1d225e00,15169010974841,15169010983154,15169011176274,15169011177636
124,"void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]",0,0,371,92848,92848,4194304,256,0,0,12,24,188544,0x7f8e1da7dc80,0x7f8e1d225e40,15169011192686,15169011201554,15169011393555,15169011394901
125,"void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]",0,0,374,92848,92848,4194304,256,0,0,12,24,191552,0x7f8e1da7db00,0x7f8e1d225e80,15169011408241,15169011416435,15169011576756,15169011578197
126,"void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]",0,0,377,92848,92848,4194304,256,0,0,12,24,195840,0x7f8e1da7d980,0x7f8e1d225ec0,15169011592336,15169011600756,15169011883957,15169011927688
127,"void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]",0,0,380,92848,92848,4194304,256,0,0,12,24,198592,0x7f8e1da7d800,0x7f8e1d225f00,15169011942488,15169011953398,15169012102838,15169012104214
128,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,383,92848,92848,4194304,256,0,0,12,24,202368,0x7f8e1da7d680,0x7f8e1d225f40,15169012121444,15169012129398,15169012338999,15169012366558
129,"void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]",0,0,386,92848,92848,4194304,256,0,0,12,24,205888,0x7f8e20219500,0x7f8e1d225f80,15169012374708,15169012384439,15169012591800,15169012611172
130,"void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]",0,0,389,92848,92848,4194304,256,0,0,12,24,209408,0x7f8e20219380,0x7f8e1d225fc0,15169012619412,15169012630040,15169012815801,15169012817117
131,"void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]",0,0,392,92848,92848,4194304,256,0,0,12,24,214464,0x7f8e20219200,0x7f8e1d226000,15169012832057,15169012840921,15169013175323,15169013190898
132,"void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]",0,0,395,92848,92848,4194304,256,0,0,12,24,217728,0x7f8e20219080,0x7f8e1d226040,15169013207028,15169013217403,15169013384924,15169013386234
133,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,398,92848,92848,4194304,256,0,0,12,24,222016,0x7f8e1da7df00,0x7f8e1d226080,15169013402633,15169013410524,15169013637725,15169013653247
134,"void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]",0,0,401,92848,92848,4194304,256,0,0,12,24,226048,0x7f8e1da7dd80,0x7f8e1d2260c0,15169013661187,15169013671485,15169013900926,15169013916431
135,"void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]",0,0,404,92848,92848,4194304,256,0,0,12,24,230080,0x7f8e1da7dc00,0x7f8e1d226100,15169013926701,15169013937086,15169014148287,15169014170005
136,"void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]",0,0,407,92848,92848,4194304,256,0,0,12,24,235904,0x7f8e1da7da80,0x7f8e1d226140,15169014183785,15169014193407,15169014578849,15169014594545
137,"void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]",0,0,410,92848,92848,4194304,256,0,0,12,24,239424,0x7f8e1da7d900,0x7f8e1d226180,15169014610525,15169014621089,15169014808130,15169014809300
138,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,413,92848,92848,4194304,256,0,0,12,24,244224,0x7f8e1da7d780,0x7f8e1d2261c0,15169014823880,15169014831970,15169015082531,15169015101223
139,"void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]",0,0,416,92848,92848,4194304,256,0,0,12,24,248768,0x7f8e1da7d600,0x7f8e1d226200,15169015110793,15169015120771,15169015368132,15169015386737
140,"void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]",0,0,419,92848,92848,4194304,256,0,0,12,24,253312,0x7f8e20219480,0x7f8e1d226240,15169015397226,15169015407652,15169015644293,15169015662850
141,"void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]",0,0,422,92848,92848,4194304,256,0,0,8,24,255552,0x7f8e20219300,0x7f8e1d226280,15169015671080,15169015681253,15169016134855,15169016153619
142,"void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]",0,0,425,92848,92848,4194304,256,0,0,12,24,259840,0x7f8e20219180,0x7f8e1d2262c0,15169016169718,15169016180615,15169016406696,15169016425752
143,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,428,92848,92848,4194304,256,0,0,12,24,265664,0x7f8e20219000,0x7f8e1d226300,15169016436502,15169016447016,15169016745098,15169016760564
144,"void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]",0,0,431,92848,92848,4194304,256,0,0,12,24,271232,0x7f8e1da7de80,0x7f8e1d226340,15169016770164,15169016779658,15169017078379,15169017096326
145,"void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]",0,0,434,92848,92848,4194304,256,0,0,8,24,273216,0x7f8e1da7dd00,0x7f8e1d226380,15169017106016,15169017115979,15169017412780,15169017428879
146,"void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]",0,0,437,92848,92848,4194304,256,0,0,8,24,275712,0x7f8e1da7db80,0x7f8e1d2263c0,15169017437938,15169017447981,15169018005423,15169018020655
147,"void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]",0,0,440,92848,92848,4194304,256,0,0,12,24,280768,0x7f8e1da7da00,0x7f8e1d226400,15169018036424,15169018046703,15169018310544,15169018336307
148,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,443,92848,92848,4194304,256,0,0,12,24,287616,0x7f8e1da7d880,0x7f8e1d226440,15169018345617,15169018355664,15169018706386,15169018729658
149,"void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]",0,0,446,92848,92848,4194304,256,0,0,12,24,294208,0x7f8e1da7d700,0x7f8e1d226480,15169018739228,15169018748786,15169019098708,15169019120659
150,"void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]",0,0,449,92848,92848,4194304,256,0,0,8,24,296960,0x7f8e20219580,0x7f8e1d2264c0,15169019131669,15169019142388,15169019489429,15169019512870
151,"void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]",0,0,452,92848,92848,4194304,256,0,0,8,24,299968,0x7f8e20219400,0x7f8e1d226500,15169019521929,15169019531829,15169020192952,15169020211913
152,"void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]",0,0,455,92848,92848,4194304,256,0,0,8,24,302720,0x7f8e20219280,0x7f8e1d226540,15169020229473,15169020238872,15169020685754,15169020703502
153,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,458,92848,92848,4194304,256,0,0,12,24,305472,0x7f8e20219100,0x7f8e1d226580,15169020714761,15169020725114,15169021190396,15169021210450
154,"void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]",0,0,461,92848,92848,4194304,256,0,0,12,24,308224,0x7f8e1da7df80,0x7f8e1d2265c0,15169021221659,15169021231836,15169021694718,15169021711828
155,"void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]",0,0,464,92848,92848,4194304,256,0,0,8,24,310976,0x7f8e1da7de00,0x7f8e1d226600,15169021720688,15169021731199,15169022182720,15169022208026
156,"void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]",0,0,467,92848,92848,4194304,256,0,0,8,24,313216,0x7f8e1da7dc80,0x7f8e1d226640,15169022218606,15169022229281,15169023099204,15169023121795
157,"void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]",0,0,470,92848,92848,4194304,256,0,0,8,24,315968,0x7f8e1da7db00,0x7f8e1d226680,15169023138014,15169023148324,15169024002248,15169024025713
158,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,473,92848,92848,4194304,256,0,0,12,24,318720,0x7f8e1da7d980,0x7f8e1d2266c0,15169024035583,15169024045608,15169024930572,15169024973241
159,"void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]",0,0,476,92848,92848,4194304,256,0,0,12,24,321472,0x7f8e1da7d800,0x7f8e1d226700,15169024983091,15169024993292,15169025872976,15169025915198
160,"void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]",0,0,479,92848,92848,4194304,256,0,0,8,24,324224,0x7f8e1da7d680,0x7f8e1d226740,15169025925948,15169025935536,15169026795540,15169026842597
161,"void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]",0,0,482,92848,92848,4194304,256,0,0,8,24,326464,0x7f8e20219500,0x7f8e1d226780,15169026851696,15169026871060,15169028569147,15169028614065
162,"void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]",0,0,485,92848,92848,4194304,256,0,0,8,24,329216,0x7f8e20219380,0x7f8e1d2267c0,15169028630945,15169028641467,15169030307394,15169030375994
163,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,488,92848,92848,4194304,256,0,0,12,24,331968,0x7f8e20219200,0x7f8e1d226800,15169030385553,15169030395875,15169032120042,15169032163531
164,"void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]",0,0,491,92848,92848,4194304,256,0,0,12,24,334720,0x7f8e20219080,0x7f8e1d226840,15169032172881,15169032183242,15169033898929,15169033943429
165,"void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]",0,0,494,92848,92848,4194304,256,0,0,8,24,337472,0x7f8e1da7df00,0x7f8e1d226880,15169033952999,15169033963410,15169035637337,15169035705858
166,"void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]",0,0,497,92848,92848,4194304,256,0,0,8,24,0,0x7f8e1da7dd80,0x7f8e1d2268c0,15169035715457,15169035726137,15169039082631,15169039157896