rocm-systems/tests/workloads/D_str_inv1/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",0,0,0,111153,111153,33554432,256,0,0,4,32,4160,0x0,0x7ff5c8604280,15444196302369,15444196327800,15444196567481,15444196654711
1,"void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]",0,0,2,111153,111153,32768,256,0,0,12,24,13888,0x0,0x7ff5c8623f80,15444201320695,15444201335821,15444201348941,15444201366234
2,"void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]",0,0,5,111153,111153,4194304,256,0,0,12,24,14336,0x7ff5d6ded380,0x7ff5c8623fc0,15444201370974,15444201416301,15444201508942,15444201511020
3,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,8,111153,111153,4194304,256,0,0,16,24,15040,0x7ff5d6ded200,0x7ff5c8624000,15444201532650,15444201542702,15444201721263,15444201723176
4,"void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]",0,0,11,111153,111153,4194304,256,0,0,16,24,15488,0x7ff5d6ded080,0x7ff5c8624040,15444201738065,15444201746223,15444201924143,15444201926061
5,"void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]",0,0,14,111153,111153,4194304,256,0,0,12,24,15936,0x7ff5d474df00,0x7ff5c8624080,15444201940241,15444201948144,15444202040944,15444202042938
6,"void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]",0,0,17,111153,111153,4194304,256,0,0,12,24,16384,0x7ff5d474dd80,0x7ff5c86240c0,15444202057818,15444202066064,15444202157904,15444202159916
7,"void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]",0,0,20,111153,111153,4194304,256,0,0,12,24,17088,0x7ff5d474dc00,0x7ff5c8624100,15444202188345,15444202198385,15444202290385,15444202292313
8,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,23,111153,111153,4194304,256,0,0,16,24,17792,0x7ff5d474da80,0x7ff5c8624140,15444202318692,15444202328465,15444202507186,15444202509098
9,"void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]",0,0,26,111153,111153,4194304,256,0,0,16,24,18496,0x7ff5d474d900,0x7ff5c8624180,15444202527987,15444202547826,15444202726067,15444202727883
10,"void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]",0,0,29,111153,111153,4194304,256,0,0,16,24,19200,0x7ff5d474d780,0x7ff5c86241c0,15444202742712,15444202758387,15444202851668,15444202853590
11,"void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]",0,0,32,111153,111153,4194304,256,0,0,16,24,19904,0x7ff5d474d600,0x7ff5c8624200,15444202873879,15444202888468,15444202981588,15444202983397
12,"void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]",0,0,35,111153,111153,4194304,256,0,0,12,24,20608,0x7ff5d6ded480,0x7ff5c8624240,15444203006476,15444203027028,15444203120469,15444203122553
13,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,38,111153,111153,4194304,256,0,0,16,24,21312,0x7ff5d6ded300,0x7ff5c8624280,15444203142153,15444203151189,15444203329910,15444203331779
14,"void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]",0,0,41,111153,111153,4194304,256,0,0,16,24,22016,0x7ff5d6ded180,0x7ff5c86242c0,15444203350868,15444203360310,15444203540951,15444203542744
15,"void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]",0,0,44,111153,111153,4194304,256,0,0,16,24,22720,0x7ff5d6ded000,0x7ff5c8624300,15444203556923,15444203565911,15444203659191,15444203661101
16,"void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]",0,0,47,111153,111153,4194304,256,0,0,12,24,23424,0x7ff5d474de80,0x7ff5c8624340,15444203677431,15444203687351,15444203779992,15444203781848
17,"void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]",0,0,50,111153,111153,4194304,256,0,0,12,24,24128,0x7ff5d474dd00,0x7ff5c8624380,15444203803648,15444203814072,15444203907032,15444203908825
18,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,53,111153,111153,4194304,256,0,0,16,24,24832,0x7ff5d474db80,0x7ff5c86243c0,15444203927205,15444203935672,15444204114713,15444204116691
19,"void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]",0,0,56,111153,111153,4194304,256,0,0,16,24,25536,0x7ff5d474da00,0x7ff5c8624400,15444204138470,15444204148473,15444204326874,15444204328566
20,"void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]",0,0,59,111153,111153,4194304,256,0,0,12,24,26240,0x7ff5d474d880,0x7ff5c8624440,15444204350755,15444204361274,15444204454715,15444204456493
21,"void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]",0,0,62,111153,111153,4194304,256,0,0,12,24,26944,0x7ff5d474d700,0x7ff5c8624480,15444204479472,15444204489915,15444204583195,15444204585030
22,"void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]",0,0,65,111153,111153,4194304,256,0,0,12,24,27648,0x7ff5d6ded580,0x7ff5c86244c0,15444204605379,15444204615675,15444204707836,15444204709597
23,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,68,111153,111153,4194304,256,0,0,16,24,28608,0x7ff5d6ded400,0x7ff5c8624500,15444204731716,15444204741916,15444204921277,15444204923092
24,"void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]",0,0,71,111153,111153,4194304,256,0,0,16,24,29312,0x7ff5d6ded280,0x7ff5c8624540,15444204938102,15444204947517,15444205128797,15444205130557
25,"void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]",0,0,74,111153,111153,4194304,256,0,0,12,24,30016,0x7ff5d6ded100,0x7ff5c8624580,15444205147887,15444205175998,15444205268478,15444205270304
26,"void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]",0,0,77,111153,111153,4194304,256,0,0,8,24,30976,0x7ff5d474df80,0x7ff5c86245c0,15444205284484,15444205293598,15444205385599,15444205387351
27,"void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]",0,0,80,111153,111153,4194304,256,0,0,12,24,31680,0x7ff5d474de00,0x7ff5c8624600,15444205407941,15444205418719,15444205511679,15444205513398
28,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,83,111153,111153,4194304,256,0,0,16,24,32640,0x7ff5d474dc80,0x7ff5c8624640,15444205537258,15444205547679,15444205728480,15444205730213
29,"void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]",0,0,86,111153,111153,4194304,256,0,0,12,24,33600,0x7ff5d474db00,0x7ff5c8624680,15444205745903,15444205754720,15444205935841,15444205937629
30,"void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]",0,0,89,111153,111153,4194304,256,0,0,12,24,34560,0x7ff5d474d980,0x7ff5c86246c0,15444205954598,15444205963201,15444206056641,15444206058396
31,"void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]",0,0,92,111153,111153,4194304,256,0,0,8,24,35520,0x7ff5d474d800,0x7ff5c8624700,15444206075055,15444206099522,15444206193602,15444206195763
32,"void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]",0,0,95,111153,111153,4194304,256,0,0,12,24,36224,0x7ff5d474d680,0x7ff5c8624740,15444206217662,15444206227042,15444206321283,15444206323100
33,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,98,111153,111153,4194304,256,0,0,12,24,37184,0x7ff5d6ded500,0x7ff5c8624780,15444206339329,15444206347843,15444206527524,15444206528845
34,"void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]",0,0,101,111153,111153,4194304,256,0,0,12,24,38144,0x7ff5d6ded380,0x7ff5c86247c0,15444206544165,15444206552644,15444206732004,15444206733450
35,"void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]",0,0,104,111153,111153,4194304,256,0,0,12,24,39104,0x7ff5d6ded200,0x7ff5c8624800,15444206748380,15444206756005,15444206849605,15444206850988
36,"void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]",0,0,107,111153,111153,4194304,256,0,0,8,24,40320,0x7ff5d6ded080,0x7ff5c8624840,15444206865077,15444206872965,15444206966885,15444206968355
37,"void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]",0,0,110,111153,111153,4194304,256,0,0,12,24,41280,0x7ff5d474df00,0x7ff5c8624880,15444206989305,15444206999846,15444207093286,15444207094732
38,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,113,111153,111153,4194304,256,0,0,12,24,42240,0x7ff5d474dd80,0x7ff5c86248c0,15444207111872,15444207119526,15444207299207,15444207300428
39,"void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]",0,0,116,111153,111153,4194304,256,0,0,12,24,43200,0x7ff5d474dc00,0x7ff5c8624900,15444207316147,15444207324647,15444207504008,15444207505343
40,"void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]",0,0,119,111153,111153,4194304,256,0,0,8,24,44160,0x7ff5d474da80,0x7ff5c8624940,15444207520632,15444207529288,15444207623208,15444207624540
41,"void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]",0,0,122,111153,111153,4194304,256,0,0,12,24,45376,0x7ff5d474d900,0x7ff5c8624980,15444207638990,15444207647528,15444207742089,15444207743397
42,"void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]",0,0,125,111153,111153,4194304,256,0,0,12,24,46336,0x7ff5d474d780,0x7ff5c86249c0,15444207762917,15444207774729,15444207868169,15444207869465
43,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,128,111153,111153,4194304,256,0,0,12,24,47552,0x7ff5d474d600,0x7ff5c8624a00,15444207883814,15444207892009,15444208071210,15444208072400
44,"void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]",0,0,131,111153,111153,4194304,256,0,0,12,24,48512,0x7ff5d6ded480,0x7ff5c8624a40,15444208086360,15444208094090,15444208276171,15444208277375
45,"void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]",0,0,134,111153,111153,4194304,256,0,0,8,24,49472,0x7ff5d6ded300,0x7ff5c8624a80,15444208291755,15444208299851,15444208392492,15444208393933
46,"void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]",0,0,137,111153,111153,4194304,256,0,0,12,24,50688,0x7ff5d6ded180,0x7ff5c8624ac0,15444208409132,15444208417132,15444208510732,15444208512340
47,"void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]",0,0,140,111153,111153,4194304,256,0,0,12,24,51648,0x7ff5d6ded000,0x7ff5c8624b00,15444208532759,15444208543372,15444208635213,15444208636727
48,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,143,111153,111153,4194304,256,0,0,12,24,52864,0x7ff5d474de80,0x7ff5c8624b40,15444208651787,15444208659373,15444208837774,15444208839312
49,"void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]",0,0,146,111153,111153,4194304,256,0,0,12,24,54080,0x7ff5d474dd00,0x7ff5c8624b80,15444208853952,15444208861454,15444209041935,15444209043258
50,"void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]",0,0,149,111153,111153,4194304,256,0,0,8,24,55296,0x7ff5d474db80,0x7ff5c8624bc0,15444209056237,15444209064175,15444209158095,15444209159545
51,"void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]",0,0,152,111153,111153,4194304,256,0,0,12,24,56768,0x7ff5d474da00,0x7ff5c8624c00,15444209174205,15444209182415,15444209277296,15444209278742
52,"void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]",0,0,155,111153,111153,4194304,256,0,0,12,24,57728,0x7ff5d474d880,0x7ff5c8624c40,15444209299282,15444209309776,15444209402576,15444209404229
53,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,158,111153,111153,4194304,256,0,0,12,24,58944,0x7ff5d474d700,0x7ff5c8624c80,15444209419779,15444209427696,15444209606417,15444209607685
54,"void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]",0,0,161,111153,111153,4194304,256,0,0,12,24,60160,0x7ff5d6ded580,0x7ff5c8624cc0,15444209623364,15444209632017,15444209811858,15444209813030
55,"void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]",0,0,164,111153,111153,4194304,256,0,0,8,24,61376,0x7ff5d6ded400,0x7ff5c8624d00,15444209828170,15444209836498,15444209930258,15444209931607
56,"void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]",0,0,167,111153,111153,4194304,256,0,0,12,24,62848,0x7ff5d6ded280,0x7ff5c8624d40,15444209947197,15444209955698,15444210056019,15444210057385
57,"void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]",0,0,170,111153,111153,4194304,256,0,0,12,24,63808,0x7ff5d6ded100,0x7ff5c8624d80,15444210106473,15444210116819,15444210211220,15444210212521
58,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,173,111153,111153,4194304,256,0,0,12,24,65024,0x7ff5d474df80,0x7ff5c8624dc0,15444210229491,15444210238100,15444210417621,15444210418896
59,"void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]",0,0,176,111153,111153,4194304,256,0,0,12,24,66240,0x7ff5d474de00,0x7ff5c8624e00,15444210433776,15444210441781,15444210621621,15444210622922
60,"void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]",0,0,179,111153,111153,4194304,256,0,0,8,24,67456,0x7ff5d474dc80,0x7ff5c8624e40,15444210636191,15444210644181,15444210738902,15444210740169
61,"void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]",0,0,182,111153,111153,4194304,256,0,0,12,24,68928,0x7ff5d474db00,0x7ff5c8624e80,15444210755369,15444210763862,15444210867062,15444210868366
62,"void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]",0,0,185,111153,111153,4194304,256,0,0,12,24,70144,0x7ff5d474d980,0x7ff5c8624ec0,15444210888346,15444210899223,15444210992023,15444210993353
63,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,188,111153,111153,4194304,256,0,0,12,24,71616,0x7ff5d474d800,0x7ff5c8624f00,15444211008783,15444211016503,15444211197144,15444211198458
64,"void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]",0,0,191,111153,111153,4194304,256,0,0,12,24,72832,0x7ff5d474d680,0x7ff5c8624f40,15444211213018,15444211221784,15444211401465,15444211402824
65,"void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]",0,0,194,111153,111153,4194304,256,0,0,8,24,74048,0x7ff5d6ded500,0x7ff5c8624f80,15444211417713,15444211426425,15444211520185,15444211521431
66,"void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]",0,0,197,111153,111153,4194304,256,0,0,12,24,75776,0x7ff5d6ded380,0x7ff5c8624fc0,15444211534451,15444211542265,15444211649306,15444211650558
67,"void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]",0,0,200,111153,111153,4194304,256,0,0,12,24,76992,0x7ff5d6ded200,0x7ff5c8625000,15444211669938,15444211681306,15444211775866,15444211777165
68,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,203,111153,111153,4194304,256,0,0,12,24,78464,0x7ff5d6ded080,0x7ff5c8625040,15444211791135,15444211799227,15444211980347,15444211981691
69,"void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]",0,0,206,111153,111153,4194304,256,0,0,12,24,79936,0x7ff5d474df00,0x7ff5c8625080,15444211996230,15444212004027,15444212183388,15444212184796
70,"void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]",0,0,209,111153,111153,4194304,256,0,0,8,24,81408,0x7ff5d474dd80,0x7ff5c86250c0,15444212200126,15444212209148,15444212304029,15444212305313
71,"void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]",0,0,212,111153,111153,4194304,256,0,0,12,24,83136,0x7ff5d474dc00,0x7ff5c8625100,15444212318643,15444212326589,15444212442109,15444212443400
72,"void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]",0,0,215,111153,111153,4194304,256,0,0,12,24,84352,0x7ff5d474da80,0x7ff5c8625140,15444212474489,15444212484190,15444212577950,15444212579337
73,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,218,111153,111153,4194304,256,0,0,12,24,85824,0x7ff5d474d900,0x7ff5c8625180,15444212595127,15444212602910,15444212783711,15444212785162
74,"void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]",0,0,221,111153,111153,4194304,256,0,0,12,24,87296,0x7ff5d474d780,0x7ff5c86251c0,15444212800162,15444212807871,15444212986592,15444212988058
75,"void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]",0,0,224,111153,111153,4194304,256,0,0,12,24,88768,0x7ff5d474d600,0x7ff5c8625200,15444213003237,15444213010912,15444213105952,15444213107105
76,"void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]",0,0,227,111153,111153,4194304,256,0,0,12,24,90752,0x7ff5d6ded480,0x7ff5c8625240,15444213123944,15444213131872,15444213253153,15444213254422
77,"void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]",0,0,230,111153,111153,4194304,256,0,0,12,24,91968,0x7ff5d6ded300,0x7ff5c8625280,15444213275781,15444213286273,15444213380513,15444213381849
78,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,233,111153,111153,4194304,256,0,0,12,24,93440,0x7ff5d6ded180,0x7ff5c86252c0,15444213397398,15444213405794,15444213585154,15444213586464
79,"void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]",0,0,236,111153,111153,4194304,256,0,0,12,24,94912,0x7ff5d6ded000,0x7ff5c8625300,15444213601084,15444213608834,15444213788355,15444213789759
80,"void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]",0,0,239,111153,111153,4194304,256,0,0,12,24,96384,0x7ff5d474de80,0x7ff5c8625340,15444213804689,15444213812195,15444213907076,15444213908437
81,"void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]",0,0,242,111153,111153,4194304,256,0,0,12,24,98368,0x7ff5d474dd00,0x7ff5c8625380,15444213923006,15444213931556,15444214057956,15444214059253
82,"void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]",0,0,245,111153,111153,4194304,256,0,0,12,24,99584,0x7ff5d474db80,0x7ff5c86253c0,15444214090622,15444214100997,15444214195077,15444214196830
83,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,248,111153,111153,4194304,256,0,0,12,24,101312,0x7ff5d474da00,0x7ff5c8625400,15444214210150,15444214218117,15444214396998,15444214398275
84,"void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]",0,0,251,111153,111153,4194304,256,0,0,12,24,102784,0x7ff5d474d880,0x7ff5c8625440,15444214412425,15444214420358,15444214599559,15444214600821
85,"void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]",0,0,254,111153,111153,4194304,256,0,0,12,24,104256,0x7ff5d474d700,0x7ff5c8625480,15444214615010,15444214623559,15444214717159,15444214718538
86,"void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]",0,0,257,111153,111153,4194304,256,0,0,12,24,106240,0x7ff5d6ded580,0x7ff5c86254c0,15444214732418,15444214740999,15444214871240,15444214872495
87,"void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]",0,0,260,111153,111153,4194304,256,0,0,12,24,107712,0x7ff5d6ded400,0x7ff5c8625500,15444214892894,15444214903400,15444214996200,15444214997542
88,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,263,111153,111153,4194304,256,0,0,16,24,109440,0x7ff5d6ded280,0x7ff5c8625540,15444215012821,15444215021641,15444215201481,15444215202817
89,"void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]",0,0,266,111153,111153,4194304,256,0,0,16,24,111168,0x7ff5d6ded100,0x7ff5c8625580,15444215215967,15444215224041,15444215402922,15444215404102
90,"void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]",0,0,269,111153,111153,4194304,256,0,0,8,24,112896,0x7ff5d474df80,0x7ff5c86255c0,15444215418142,15444215426762,15444215521963,15444215523360
91,"void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]",0,0,272,111153,111153,4194304,256,0,0,12,24,115136,0x7ff5d474de00,0x7ff5c8625600,15444215536899,15444215545003,15444215682763,15444215684066
92,"void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]",0,0,275,111153,111153,4194304,256,0,0,12,24,116608,0x7ff5d474dc80,0x7ff5c8625640,15444215704666,15444215715404,15444215810284,15444215811633
93,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,278,111153,111153,4194304,256,0,0,16,24,118336,0x7ff5d474db00,0x7ff5c8625680,15444215826443,15444215835084,15444216014445,15444216015688
94,"void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]",0,0,281,111153,111153,4194304,256,0,0,16,24,120064,0x7ff5d474d980,0x7ff5c86256c0,15444216029978,15444216039085,15444216219086,15444216220354
95,"void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]",0,0,284,111153,111153,4194304,256,0,0,12,24,121792,0x7ff5d474d800,0x7ff5c8625700,15444216234643,15444216243406,15444216340046,15444216341371
96,"void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]",0,0,287,111153,111153,4194304,256,0,0,12,24,124032,0x7ff5d474d680,0x7ff5c8625740,15444216356521,15444216364846,15444216510127,15444216511507
97,"void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]",0,0,290,111153,111153,4194304,256,0,0,12,24,125504,0x7ff5d6ded500,0x7ff5c8625780,15444216532607,15444216543567,15444216639248,15444216640614
98,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,293,111153,111153,4194304,256,0,0,12,24,127488,0x7ff5d6ded380,0x7ff5c86257c0,15444216654924,15444216662768,15444216844369,15444216845769
99,"void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]",0,0,296,111153,111153,4194304,256,0,0,12,24,129216,0x7ff5d6ded200,0x7ff5c8625800,15444216861279,15444216870129,15444217052369,15444217053745
100,"void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]",0,0,299,111153,111153,4194304,256,0,0,12,24,130944,0x7ff5d6ded080,0x7ff5c8625840,15444217067594,15444217075890,15444217175410,15444217176732
101,"void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]",0,0,302,111153,111153,4194304,256,0,0,12,24,133440,0x7ff5d474df00,0x7ff5c8625880,15444217191942,15444217200530,15444217356211,15444217357578
102,"void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]",0,0,305,111153,111153,4194304,256,0,0,12,24,134912,0x7ff5d474dd80,0x7ff5c86258c0,15444217378547,15444217389331,15444217485011,15444217486225
103,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,308,111153,111153,4194304,256,0,0,12,24,136896,0x7ff5d474dc00,0x7ff5c8625900,15444217502334,15444217510131,15444217690772,15444217692140
104,"void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]",0,0,311,111153,111153,4194304,256,0,0,12,24,138880,0x7ff5d474da80,0x7ff5c8625940,15444217708380,15444217716852,15444217899733,15444217901085
105,"void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]",0,0,314,111153,111153,4194304,256,0,0,12,24,140864,0x7ff5d474d900,0x7ff5c8625980,15444217915325,15444217923573,15444218031414,15444218032722
106,"void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]",0,0,317,111153,111153,4194304,256,0,0,12,24,143616,0x7ff5d474d780,0x7ff5c86259c0,15444218046062,15444218054774,15444218225815,15444218227318
107,"void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]",0,0,320,111153,111153,4194304,256,0,0,12,24,145344,0x7ff5d474d600,0x7ff5c8625a00,15444218247907,15444218257975,15444218353655,15444218355145
108,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,323,111153,111153,4194304,256,0,0,12,24,147584,0x7ff5d6ded480,0x7ff5c8625a40,15444218369945,15444218377975,15444218557656,15444218558940
109,"void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]",0,0,326,111153,111153,4194304,256,0,0,12,24,149568,0x7ff5d6ded300,0x7ff5c8625a80,15444218572170,15444218580216,15444218760537,15444218761846
110,"void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]",0,0,329,111153,111153,4194304,256,0,0,12,24,151552,0x7ff5d6ded180,0x7ff5c8625ac0,15444218775535,15444218783417,15444218893497,15444218894883
111,"void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]",0,0,332,111153,111153,4194304,256,0,0,12,24,154304,0x7ff5d6ded000,0x7ff5c8625b00,15444218909542,15444218917498,15444219098938,15444219100128
112,"void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]",0,0,335,111153,111153,4194304,256,0,0,12,24,156288,0x7ff5d474de80,0x7ff5c8625b40,15444219121757,15444219131418,15444219233019,15444219234355
113,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,338,111153,111153,4194304,256,0,0,12,24,158784,0x7ff5d474dd00,0x7ff5c8625b80,15444219250655,15444219258459,15444219443580,15444219444860
114,"void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]",0,0,341,111153,111153,4194304,256,0,0,12,24,161024,0x7ff5d474db80,0x7ff5c8625bc0,15444219459270,15444219467260,15444219650461,15444219651755
115,"void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]",0,0,344,111153,111153,4194304,256,0,0,12,24,163264,0x7ff5d474da00,0x7ff5c8625c00,15444219665745,15444219674461,15444219796861,15444219798352
116,"void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]",0,0,347,111153,111153,4194304,256,0,0,12,24,166528,0x7ff5d474d880,0x7ff5c8625c40,15444219812042,15444219820061,15444220026462,15444220047276
117,"void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]",0,0,350,111153,111153,4194304,256,0,0,12,24,168512,0x7ff5d474d700,0x7ff5c8625c80,15444220061066,15444220071743,15444220182783,15444220183953
118,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,353,111153,111153,4194304,256,0,0,12,24,171264,0x7ff5d6ded580,0x7ff5c8625cc0,15444220199883,15444220208863,15444220390944,15444220392248
119,"void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]",0,0,356,111153,111153,4194304,256,0,0,12,24,173760,0x7ff5d6ded400,0x7ff5c8625d00,15444220407358,15444220415424,15444220602465,15444220603794
120,"void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]",0,0,359,111153,111153,4194304,256,0,0,12,24,176256,0x7ff5d6ded280,0x7ff5c8625d40,15444220620583,15444220629185,15444220763746,15444220764970
121,"void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]",0,0,362,111153,111153,4194304,256,0,0,12,24,179776,0x7ff5d6ded100,0x7ff5c8625d80,15444220779770,15444220787906,15444221019907,15444221039584
122,"void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]",0,0,365,111153,111153,4194304,256,0,0,12,24,182272,0x7ff5d474df80,0x7ff5c8625dc0,15444221054233,15444221064867,15444221191107,15444221192450
123,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,368,111153,111153,4194304,256,0,0,12,24,185536,0x7ff5d474de00,0x7ff5c8625e00,15444221207410,15444221215588,15444221407588,15444221408935
124,"void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]",0,0,371,111153,111153,4194304,256,0,0,12,24,188544,0x7ff5d474dc80,0x7ff5c8625e40,15444221422685,15444221431269,15444221625349,15444221626670
125,"void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]",0,0,374,111153,111153,4194304,256,0,0,12,24,191552,0x7ff5d474db00,0x7ff5c8625e80,15444221640830,15444221649509,15444221809190,15444221810466
126,"void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]",0,0,377,111153,111153,4194304,256,0,0,12,24,195840,0x7ff5d474d980,0x7ff5c8625ec0,15444221824976,15444221833350,15444222116552,15444222144558
127,"void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]",0,0,380,111153,111153,4194304,256,0,0,12,24,198592,0x7ff5d474d800,0x7ff5c8625f00,15444222163378,15444222173992,15444222323912,15444222325244
128,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,383,111153,111153,4194304,256,0,0,12,24,202368,0x7ff5d474d680,0x7ff5c8625f40,15444222340234,15444222348073,15444222554633,15444222575049
129,"void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]",0,0,386,111153,111153,4194304,256,0,0,12,24,205888,0x7ff5d6ded500,0x7ff5c8625f80,15444222583328,15444222593674,15444222802954,15444222845912
130,"void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]",0,0,389,111153,111153,4194304,256,0,0,12,24,209408,0x7ff5d6ded380,0x7ff5c8625fc0,15444222853572,15444222864075,15444223049516,15444223050788
131,"void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]",0,0,392,111153,111153,4194304,256,0,0,12,24,214464,0x7ff5d6ded200,0x7ff5c8626000,15444223065277,15444223072876,15444223406957,15444223446778
132,"void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]",0,0,395,111153,111153,4194304,256,0,0,12,24,217728,0x7ff5d6ded080,0x7ff5c8626040,15444223462688,15444223473197,15444223641518,15444223642904
133,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,398,111153,111153,4194304,256,0,0,12,24,222016,0x7ff5d474df00,0x7ff5c8626080,15444223656664,15444223664398,15444223893199,15444223933567
134,"void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]",0,0,401,111153,111153,4194304,256,0,0,12,24,226048,0x7ff5d474dd80,0x7ff5c86260c0,15444223941347,15444223952080,15444224181361,15444224220921
135,"void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]",0,0,404,111153,111153,4194304,256,0,0,12,24,230080,0x7ff5d474dc00,0x7ff5c8626100,15444224228080,15444224238641,15444224449682,15444224465645
136,"void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]",0,0,407,111153,111153,4194304,256,0,0,12,24,235904,0x7ff5d474da80,0x7ff5c8626140,15444224474505,15444224483602,15444224869524,15444224884685
137,"void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]",0,0,410,111153,111153,4194304,256,0,0,12,24,239424,0x7ff5d474d900,0x7ff5c8626180,15444224898945,15444224909684,15444225096884,15444225098200
138,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,413,111153,111153,4194304,256,0,0,12,24,244224,0x7ff5d474d780,0x7ff5c86261c0,15444225113220,15444225121045,15444225371606,15444225390144
139,"void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]",0,0,416,111153,111153,4194304,256,0,0,12,24,248768,0x7ff5d474d600,0x7ff5c8626200,15444225400084,15444225410646,15444225658167,15444225674247
140,"void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]",0,0,419,111153,111153,4194304,256,0,0,12,24,253312,0x7ff5d6ded480,0x7ff5c8626240,15444225682257,15444225692567,15444225929048,15444225944571
141,"void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]",0,0,422,111153,111153,4194304,256,0,0,8,24,255552,0x7ff5d6ded300,0x7ff5c8626280,15444225952851,15444225963128,15444226416890,15444226458339
142,"void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]",0,0,425,111153,111153,4194304,256,0,0,12,24,259840,0x7ff5d6ded180,0x7ff5c86262c0,15444226474599,15444226485371,15444226710492,15444226751703
143,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,428,111153,111153,4194304,256,0,0,12,24,265664,0x7ff5d6ded000,0x7ff5c8626300,15444226759732,15444226770012,15444227067933,15444227107914
144,"void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]",0,0,431,111153,111153,4194304,256,0,0,12,24,271232,0x7ff5d474de80,0x7ff5c8626340,15444227118084,15444227127933,15444227426495,15444227469296
145,"void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]",0,0,434,111153,111153,4194304,256,0,0,8,24,273216,0x7ff5d474dd00,0x7ff5c8626380,15444227477716,15444227487935,15444227783776,15444227826718
146,"void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]",0,0,437,111153,111153,4194304,256,0,0,8,24,275712,0x7ff5d474db80,0x7ff5c86263c0,15444227836748,15444227847296,15444228404259,15444228443894
147,"void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]",0,0,440,111153,111153,4194304,256,0,0,12,24,280768,0x7ff5d474da00,0x7ff5c8626400,15444228458594,15444228469539,15444228733380,15444228778876
148,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,443,111153,111153,4194304,256,0,0,12,24,287616,0x7ff5d474d880,0x7ff5c8626440,15444228787156,15444228797381,15444229147622,15444229190827
149,"void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]",0,0,446,111153,111153,4194304,256,0,0,12,24,294208,0x7ff5d474d700,0x7ff5c8626480,15444229199257,15444229209702,15444229559944,15444229579798
150,"void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]",0,0,449,111153,111153,4194304,256,0,0,8,24,296960,0x7ff5d6ded580,0x7ff5c86264c0,15444229588348,15444229598664,15444229945386,15444229959269
151,"void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]",0,0,452,111153,111153,4194304,256,0,0,8,24,299968,0x7ff5d6ded400,0x7ff5c8626500,15444229969069,15444229979306,15444230639949,15444230688593
152,"void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]",0,0,455,111153,111153,4194304,256,0,0,8,24,302720,0x7ff5d6ded280,0x7ff5c8626540,15444230708022,15444230717389,15444231165391,15444231211741
153,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,458,111153,111153,4194304,256,0,0,12,24,305472,0x7ff5d6ded100,0x7ff5c8626580,15444231222630,15444231231471,15444231694993,15444231739529
154,"void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]",0,0,461,111153,111153,4194304,256,0,0,12,24,308224,0x7ff5d474df80,0x7ff5c86265c0,15444231750438,15444231766834,15444232228276,15444232272896
155,"void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]",0,0,464,111153,111153,4194304,256,0,0,8,24,310976,0x7ff5d474de00,0x7ff5c8626600,15444232281126,15444232291476,15444232743958,15444232788435
156,"void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]",0,0,467,111153,111153,4194304,256,0,0,8,24,313216,0x7ff5d474dc80,0x7ff5c8626640,15444232797454,15444232807478,15444233677402,15444233722423
157,"void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]",0,0,470,111153,111153,4194304,256,0,0,8,24,315968,0x7ff5d474db00,0x7ff5c8626680,15444233738643,15444233749562,15444234602846,15444234653282
158,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,473,111153,111153,4194304,256,0,0,12,24,318720,0x7ff5d474d980,0x7ff5c86266c0,15444234662972,15444234672766,15444235557250,15444235603950
159,"void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]",0,0,476,111153,111153,4194304,256,0,0,12,24,321472,0x7ff5d474d800,0x7ff5c8626700,15444235613530,15444235623650,15444236502374,15444236549139
160,"void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]",0,0,479,111153,111153,4194304,256,0,0,8,24,324224,0x7ff5d474d680,0x7ff5c8626740,15444236558749,15444236575175,15444237434858,15444237480018
161,"void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]",0,0,482,111153,111153,4194304,256,0,0,8,24,326464,0x7ff5d6ded500,0x7ff5c8626780,15444237489747,15444237499019,15444239197266,15444239268077
162,"void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]",0,0,485,111153,111153,4194304,256,0,0,8,24,329216,0x7ff5d6ded380,0x7ff5c86267c0,15444239285486,15444239295826,15444240961274,15444241029666
163,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,488,111153,111153,4194304,256,0,0,12,24,331968,0x7ff5d6ded200,0x7ff5c8626800,15444241037986,15444241048634,15444242773122,15444242842485
164,"void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]",0,0,491,111153,111153,4194304,256,0,0,12,24,334720,0x7ff5d6ded080,0x7ff5c8626840,15444242851965,15444242861922,15444244577289,15444244645603
165,"void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]",0,0,494,111153,111153,4194304,256,0,0,8,24,337472,0x7ff5d474df00,0x7ff5c8626880,15444244655103,15444244664810,15444246337617,15444246414593
166,"void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]",0,0,497,111153,111153,4194304,256,0,0,8,24,0,0x7ff5d474dd80,0x7ff5c86268c0,15444246422873,15444246466738,15444249824512,15444249893563