rocm-systems/tests/workloads/L2/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",0,0,0,177088,177088,33554432,256,0,0,4,32,4160,0x0,0x7f0d5c804280,16811968799934,16811968824945,16811969064464,16811969154652
1,"void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]",0,0,2,177088,177088,32768,256,0,0,12,24,13888,0x0,0x7f0d5c823f80,16811973880645,16811973896601,16811973909720,16811973927494
2,"void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]",0,0,5,177088,177088,4194304,256,0,0,12,24,14336,0x7f0d5f807380,0x7f0d5c823fc0,16811973932533,16811973977400,16811974069880,16811974072089
3,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,8,177088,177088,4194304,256,0,0,16,24,15040,0x7f0d5f807200,0x7f0d5c824000,16811974106908,16811974116760,16811974295159,16811974297432
4,"void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]",0,0,11,177088,177088,4194304,256,0,0,16,24,15488,0x7f0d5f807080,0x7f0d5c824040,16811974311692,16811974319959,16811974497878,16811974499986
5,"void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]",0,0,14,177088,177088,4194304,256,0,0,12,24,15936,0x7f0d5d167f00,0x7f0d5c824080,16811974512235,16811974520438,16811974613877,16811974615872
6,"void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]",0,0,17,177088,177088,4194304,256,0,0,12,24,16384,0x7f0d5d167d80,0x7f0d5c8240c0,16811974629832,16811974637717,16811974729557,16811974731569
7,"void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]",0,0,20,177088,177088,4194304,256,0,0,12,24,17088,0x7f0d5d167c00,0x7f0d5c824100,16811974757498,16811974768437,16811974860276,16811974862234
8,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,23,177088,177088,4194304,256,0,0,16,24,17792,0x7f0d5d167a80,0x7f0d5c824140,16811974879114,16811974888276,16811975067635,16811975069618
9,"void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]",0,0,26,177088,177088,4194304,256,0,0,16,24,18496,0x7f0d5d167900,0x7f0d5c824180,16811975086167,16811975095155,16811975274834,16811975276772
10,"void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]",0,0,29,177088,177088,4194304,256,0,0,16,24,19200,0x7f0d5d167780,0x7f0d5c8241c0,16811975291121,16811975299474,16811975392594,16811975394718
11,"void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]",0,0,32,177088,177088,4194304,256,0,0,16,24,19904,0x7f0d5d167600,0x7f0d5c824200,16811975411777,16811975421073,16811975514513,16811975516604
12,"void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]",0,0,35,177088,177088,4194304,256,0,0,12,24,20608,0x7f0d5f807480,0x7f0d5c824240,16811975535123,16811975550673,16811975643312,16811975645400
13,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,38,177088,177088,4194304,256,0,0,16,24,21312,0x7f0d5f807300,0x7f0d5c824280,16811975662729,16811975687152,16811975865711,16811975867813
14,"void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]",0,0,41,177088,177088,4194304,256,0,0,16,24,22016,0x7f0d5f807180,0x7f0d5c8242c0,16811975884113,16811975893711,16811976073070,16811976075327
15,"void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]",0,0,44,177088,177088,4194304,256,0,0,16,24,22720,0x7f0d5f807000,0x7f0d5c824300,16811976088376,16811976097390,16811976190990,16811976192933
16,"void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]",0,0,47,177088,177088,4194304,256,0,0,12,24,23424,0x7f0d5d167e80,0x7f0d5c824340,16811976206863,16811976217230,16811976310189,16811976312149
17,"void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]",0,0,50,177088,177088,4194304,256,0,0,12,24,24128,0x7f0d5d167d00,0x7f0d5c824380,16811976332599,16811976342989,16811976436749,16811976438715
18,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,53,177088,177088,4194304,256,0,0,16,24,24832,0x7f0d5d167b80,0x7f0d5c8243c0,16811976455685,16811976472429,16811976652108,16811976654129
19,"void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]",0,0,56,177088,177088,4194304,256,0,0,16,24,25536,0x7f0d5d167a00,0x7f0d5c824400,16811976667038,16811976676588,16811976856747,16811976858722
20,"void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]",0,0,59,177088,177088,4194304,256,0,0,12,24,26240,0x7f0d5d167880,0x7f0d5c824440,16811976874272,16811976887147,16811976980586,16811976982648
21,"void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]",0,0,62,177088,177088,4194304,256,0,0,12,24,26944,0x7f0d5d167700,0x7f0d5c824480,16811976997998,16811977007626,16811977101546,16811977103595
22,"void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]",0,0,65,177088,177088,4194304,256,0,0,12,24,27648,0x7f0d5f807580,0x7f0d5c8244c0,16811977123924,16811977134666,16811977226825,16811977228921
23,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,68,177088,177088,4194304,256,0,0,16,24,28608,0x7f0d5f807400,0x7f0d5c824500,16811977243380,16811977252265,16811977431784,16811977433774
24,"void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]",0,0,71,177088,177088,4194304,256,0,0,16,24,29312,0x7f0d5f807280,0x7f0d5c824540,16811977447544,16811977461864,16811977643143,16811977645188
25,"void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]",0,0,74,177088,177088,4194304,256,0,0,12,24,30016,0x7f0d5f807100,0x7f0d5c824580,16811977661307,16811977690823,16811977783143,16811977785273
26,"void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]",0,0,77,177088,177088,4194304,256,0,0,8,24,30976,0x7f0d5d167f80,0x7f0d5c8245c0,16811977799063,16811977808102,16811977900102,16811977902100
27,"void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]",0,0,80,177088,177088,4194304,256,0,0,12,24,31680,0x7f0d5d167e00,0x7f0d5c824600,16811977921359,16811977932262,16811978026021,16811978028016
28,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,83,177088,177088,4194304,256,0,0,16,24,32640,0x7f0d5d167c80,0x7f0d5c824640,16811978049585,16811978060261,16811978240740,16811978242919
29,"void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]",0,0,86,177088,177088,4194304,256,0,0,12,24,33600,0x7f0d5d167b00,0x7f0d5c824680,16811978255979,16811978265540,16811978445859,16811978447903
30,"void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]",0,0,89,177088,177088,4194304,256,0,0,12,24,34560,0x7f0d5d167980,0x7f0d5c8246c0,16811978462442,16811978471459,16811978564579,16811978566779
31,"void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]",0,0,92,177088,177088,4194304,256,0,0,8,24,35520,0x7f0d5d167800,0x7f0d5c824700,16811978584468,16811978598819,16811978692578,16811978694275
32,"void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]",0,0,95,177088,177088,4194304,256,0,0,12,24,36224,0x7f0d5d167680,0x7f0d5c824740,16811978715274,16811978724898,16811978817538,16811978819071
33,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,98,177088,177088,4194304,256,0,0,12,24,37184,0x7f0d5f807500,0x7f0d5c824780,16811978833521,16811978841538,16811979022337,16811979023995
34,"void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]",0,0,101,177088,177088,4194304,256,0,0,12,24,38144,0x7f0d5f807380,0x7f0d5c8247c0,16811979037184,16811979046017,16811979224576,16811979226238
35,"void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]",0,0,104,177088,177088,4194304,256,0,0,12,24,39104,0x7f0d5f807200,0x7f0d5c824800,16811979240718,16811979249056,16811979343295,16811979344905
36,"void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]",0,0,107,177088,177088,4194304,256,0,0,8,24,40320,0x7f0d5f807080,0x7f0d5c824840,16811979358894,16811979367135,16811979461695,16811979463131
37,"void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]",0,0,110,177088,177088,4194304,256,0,0,12,24,41280,0x7f0d5d167f00,0x7f0d5c824880,16811979482270,16811979494175,16811979587134,16811979588697
38,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,113,177088,177088,4194304,256,0,0,12,24,42240,0x7f0d5d167d80,0x7f0d5c8248c0,16811979603047,16811979611774,16811979790333,16811979791851
39,"void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]",0,0,116,177088,177088,4194304,256,0,0,12,24,43200,0x7f0d5d167c00,0x7f0d5c824900,16811979805470,16811979813533,16811979995292,16811979996964
40,"void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]",0,0,119,177088,177088,4194304,256,0,0,8,24,44160,0x7f0d5d167a80,0x7f0d5c824940,16811980014514,16811980022652,16811980115772,16811980117291
41,"void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]",0,0,122,177088,177088,4194304,256,0,0,12,24,45376,0x7f0d5d167900,0x7f0d5c824980,16811980132500,16811980140572,16811980234971,16811980236457
42,"void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]",0,0,125,177088,177088,4194304,256,0,0,12,24,46336,0x7f0d5d167780,0x7f0d5c8249c0,16811980256146,16811980266811,16811980360731,16811980362253
43,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,128,177088,177088,4194304,256,0,0,12,24,47552,0x7f0d5d167600,0x7f0d5c824a00,16811980376452,16811980384891,16811980563770,16811980565347
44,"void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]",0,0,131,177088,177088,4194304,256,0,0,12,24,48512,0x7f0d5f807480,0x7f0d5c824a40,16811980579156,16811980587290,16811980766809,16811980768320
45,"void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]",0,0,134,177088,177088,4194304,256,0,0,8,24,49472,0x7f0d5f807300,0x7f0d5c824a80,16811980782060,16811980790329,16811980882648,16811980884127
46,"void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]",0,0,137,177088,177088,4194304,256,0,0,12,24,50688,0x7f0d5f807180,0x7f0d5c824ac0,16811980900146,16811980908568,16811981001528,16811981003003
47,"void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]",0,0,140,177088,177088,4194304,256,0,0,12,24,51648,0x7f0d5f807000,0x7f0d5c824b00,16811981024872,16811981035287,16811981127127,16811981128619
48,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,143,177088,177088,4194304,256,0,0,12,24,52864,0x7f0d5d167e80,0x7f0d5c824b40,16811981144769,16811981152887,16811981332246,16811981333783
49,"void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]",0,0,146,177088,177088,4194304,256,0,0,12,24,54080,0x7f0d5d167d00,0x7f0d5c824b80,16811981348632,16811981358806,16811981540725,16811981542196
50,"void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]",0,0,149,177088,177088,4194304,256,0,0,8,24,55296,0x7f0d5d167b80,0x7f0d5c824bc0,16811981556276,16811981564725,16811981658805,16811981660402
51,"void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]",0,0,152,177088,177088,4194304,256,0,0,12,24,56768,0x7f0d5d167a00,0x7f0d5c824c00,16811981673572,16811981681844,16811981777044,16811981778489
52,"void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]",0,0,155,177088,177088,4194304,256,0,0,12,24,57728,0x7f0d5d167880,0x7f0d5c824c40,16811981798608,16811981809844,16811981903123,16811981904675
53,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,158,177088,177088,4194304,256,0,0,12,24,58944,0x7f0d5d167700,0x7f0d5c824c80,16811981919094,16811981927123,16811982106803,16811982108758
54,"void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]",0,0,161,177088,177088,4194304,256,0,0,12,24,60160,0x7f0d5f807580,0x7f0d5c824cc0,16811982124798,16811982132722,16811982312242,16811982313742
55,"void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]",0,0,164,177088,177088,4194304,256,0,0,8,24,61376,0x7f0d5f807400,0x7f0d5c824d00,16811982327982,16811982335921,16811982430001,16811982431508
56,"void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]",0,0,167,177088,177088,4194304,256,0,0,12,24,62848,0x7f0d5f807280,0x7f0d5c824d40,16811982446708,16811982454641,16811982553680,16811982555215
57,"void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]",0,0,170,177088,177088,4194304,256,0,0,12,24,63808,0x7f0d5f807100,0x7f0d5c824d80,16811982574804,16811982585680,16811982678800,16811982680351
58,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,173,177088,177088,4194304,256,0,0,12,24,65024,0x7f0d5d167f80,0x7f0d5c824dc0,16811982694380,16811982703120,16811982882959,16811982884484
59,"void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]",0,0,176,177088,177088,4194304,256,0,0,12,24,66240,0x7f0d5d167e00,0x7f0d5c824e00,16811982898954,16811982907759,16811983088238,16811983089808
60,"void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]",0,0,179,177088,177088,4194304,256,0,0,8,24,67456,0x7f0d5d167c80,0x7f0d5c824e40,16811983103248,16811983111438,16811983205997,16811983207474
61,"void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]",0,0,182,177088,177088,4194304,256,0,0,12,24,68928,0x7f0d5d167b00,0x7f0d5c824e80,16811983221334,16811983229837,16811983333837,16811983335380
62,"void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]",0,0,185,177088,177088,4194304,256,0,0,12,24,70144,0x7f0d5d167980,0x7f0d5c824ec0,16811983355190,16811983365837,16811983458476,16811983460027
63,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,188,177088,177088,4194304,256,0,0,12,24,71616,0x7f0d5d167800,0x7f0d5c824f00,16811983474436,16811983482316,16811983660875,16811983662420
64,"void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]",0,0,191,177088,177088,4194304,256,0,0,12,24,72832,0x7f0d5d167680,0x7f0d5c824f40,16811983676910,16811983685515,16811983865674,16811983867234
65,"void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]",0,0,194,177088,177088,4194304,256,0,0,8,24,74048,0x7f0d5f807500,0x7f0d5c824f80,16811983882103,16811983890634,16811983984074,16811983985530
66,"void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]",0,0,197,177088,177088,4194304,256,0,0,12,24,75776,0x7f0d5f807380,0x7f0d5c824fc0,16811983998820,16811984006794,16811984113833,16811984115376
67,"void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]",0,0,200,177088,177088,4194304,256,0,0,12,24,76992,0x7f0d5f807200,0x7f0d5c825000,16811984135536,16811984146313,16811984239913,16811984241492
68,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,203,177088,177088,4194304,256,0,0,12,24,78464,0x7f0d5f807080,0x7f0d5c825040,16811984256332,16811984264553,16811984445032,16811984446546
69,"void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]",0,0,206,177088,177088,4194304,256,0,0,12,24,79936,0x7f0d5d167f00,0x7f0d5c825080,16811984461275,16811984478152,16811984658791,16811984660459
70,"void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]",0,0,209,177088,177088,4194304,256,0,0,8,24,81408,0x7f0d5d167d80,0x7f0d5c8250c0,16811984676709,16811984685511,16811984780550,16811984782116
71,"void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]",0,0,212,177088,177088,4194304,256,0,0,12,24,83136,0x7f0d5d167c00,0x7f0d5c825100,16811984796135,16811984804550,16811984918470,16811984919931
72,"void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]",0,0,215,177088,177088,4194304,256,0,0,12,24,84352,0x7f0d5d167a80,0x7f0d5c825140,16811984953790,16811984963429,16811985057509,16811985059157
73,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,218,177088,177088,4194304,256,0,0,12,24,85824,0x7f0d5d167900,0x7f0d5c825180,16811985074206,16811985082469,16811985262628,16811985264171
74,"void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]",0,0,221,177088,177088,4194304,256,0,0,12,24,87296,0x7f0d5d167780,0x7f0d5c8251c0,16811985278010,16811985285988,16811985465347,16811985466874
75,"void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]",0,0,224,177088,177088,4194304,256,0,0,12,24,88768,0x7f0d5d167600,0x7f0d5c825200,16811985481014,16811985489187,16811985583906,16811985585441
76,"void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]",0,0,227,177088,177088,4194304,256,0,0,12,24,90752,0x7f0d5f807480,0x7f0d5c825240,16811985600000,16811985608546,16811985729506,16811985730936
77,"void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]",0,0,230,177088,177088,4194304,256,0,0,12,24,91968,0x7f0d5f807300,0x7f0d5c825280,16811985751595,16811985762466,16811985855745,16811985857362
78,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,233,177088,177088,4194304,256,0,0,12,24,93440,0x7f0d5f807180,0x7f0d5c8252c0,16811985873622,16811985881825,16811986060544,16811986062076
79,"void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]",0,0,236,177088,177088,4194304,256,0,0,12,24,94912,0x7f0d5f807000,0x7f0d5c825300,16811986087125,16811986096064,16811986275263,16811986276769
80,"void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]",0,0,239,177088,177088,4194304,256,0,0,12,24,96384,0x7f0d5d167e80,0x7f0d5c825340,16811986291559,16811986299423,16811986394463,16811986395895
81,"void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]",0,0,242,177088,177088,4194304,256,0,0,12,24,98368,0x7f0d5d167d00,0x7f0d5c825380,16811986411535,16811986419423,16811986545502,16811986547251
82,"void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]",0,0,245,177088,177088,4194304,256,0,0,12,24,99584,0x7f0d5d167b80,0x7f0d5c8253c0,16811986566830,16811986577502,16811986671901,16811986673327
83,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,248,177088,177088,4194304,256,0,0,12,24,101312,0x7f0d5d167a00,0x7f0d5c825400,16811986688356,16811986696381,16811986876540,16811986878260
84,"void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]",0,0,251,177088,177088,4194304,256,0,0,12,24,102784,0x7f0d5d167880,0x7f0d5c825440,16811986891830,16811986899900,16811987080220,16811987081854
85,"void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]",0,0,254,177088,177088,4194304,256,0,0,12,24,104256,0x7f0d5d167700,0x7f0d5c825480,16811987095724,16811987103739,16811987197499,16811987199250
86,"void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]",0,0,257,177088,177088,4194304,256,0,0,12,24,106240,0x7f0d5f807580,0x7f0d5c8254c0,16811987212890,16811987220859,16811987351578,16811987353126
87,"void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]",0,0,260,177088,177088,4194304,256,0,0,12,24,107712,0x7f0d5f807400,0x7f0d5c825500,16811987372655,16811987383578,16811987476538,16811987478252
88,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,263,177088,177088,4194304,256,0,0,16,24,109440,0x7f0d5f807280,0x7f0d5c825540,16811987493701,16811987501658,16811987681177,16811987682925
89,"void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]",0,0,266,177088,177088,4194304,256,0,0,16,24,111168,0x7f0d5f807100,0x7f0d5c825580,16811987696125,16811987703897,16811987883256,16811987884899
90,"void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]",0,0,269,177088,177088,4194304,256,0,0,8,24,112896,0x7f0d5d167f80,0x7f0d5c8255c0,16811987898539,16811987906616,16811988001975,16811988003495
91,"void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]",0,0,272,177088,177088,4194304,256,0,0,12,24,115136,0x7f0d5d167e00,0x7f0d5c825600,16811988017655,16811988025655,16811988164214,16811988165670
92,"void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]",0,0,275,177088,177088,4194304,256,0,0,12,24,116608,0x7f0d5d167c80,0x7f0d5c825640,16811988185080,16811988197494,16811988291734,16811988293316
93,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,278,177088,177088,4194304,256,0,0,16,24,118336,0x7f0d5d167b00,0x7f0d5c825680,16811988307436,16811988315414,16811988494453,16811988496040
94,"void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]",0,0,281,177088,177088,4194304,256,0,0,16,24,120064,0x7f0d5d167980,0x7f0d5c8256c0,16811988509560,16811988518293,16811988698932,16811988700664
95,"void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]",0,0,284,177088,177088,4194304,256,0,0,12,24,121792,0x7f0d5d167800,0x7f0d5c825700,16811988714283,16811988722452,16811988818291,16811988819830
96,"void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]",0,0,287,177088,177088,4194304,256,0,0,12,24,124032,0x7f0d5d167680,0x7f0d5c825740,16811988834499,16811988842931,16811988988371,16811988989905
97,"void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]",0,0,290,177088,177088,4194304,256,0,0,12,24,125504,0x7f0d5f807500,0x7f0d5c825780,16811989010024,16811989020531,16811989115250,16811989116821
98,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,293,177088,177088,4194304,256,0,0,12,24,127488,0x7f0d5f807380,0x7f0d5c8257c0,16811989131120,16811989139250,16811989319889,16811989321354
99,"void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]",0,0,296,177088,177088,4194304,256,0,0,12,24,129216,0x7f0d5f807200,0x7f0d5c825800,16811989337114,16811989345009,16811989525968,16811989527528
100,"void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]",0,0,299,177088,177088,4194304,256,0,0,12,24,130944,0x7f0d5f807080,0x7f0d5c825840,16811989541397,16811989550128,16811989651568,16811989652964
101,"void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]",0,0,302,177088,177088,4194304,256,0,0,12,24,133440,0x7f0d5d167f00,0x7f0d5c825880,16811989666173,16811989674447,16811989829487,16811989830948
102,"void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]",0,0,305,177088,177088,4194304,256,0,0,12,24,134912,0x7f0d5d167d80,0x7f0d5c8258c0,16811989851418,16811989861807,16811989956686,16811989958234
103,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,308,177088,177088,4194304,256,0,0,12,24,136896,0x7f0d5d167c00,0x7f0d5c825900,16811989972984,16811989980846,16811990160685,16811990162388
104,"void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]",0,0,311,177088,177088,4194304,256,0,0,12,24,138880,0x7f0d5d167a80,0x7f0d5c825940,16811990176988,16811990184845,16811990365484,16811990366962
105,"void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]",0,0,314,177088,177088,4194304,256,0,0,12,24,140864,0x7f0d5d167900,0x7f0d5c825980,16811990380131,16811990388524,16811990495724,16811990497428
106,"void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]",0,0,317,177088,177088,4194304,256,0,0,12,24,143616,0x7f0d5d167780,0x7f0d5c8259c0,16811990510567,16811990518924,16811990689003,16811990690671
107,"void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]",0,0,320,177088,177088,4194304,256,0,0,12,24,145344,0x7f0d5d167600,0x7f0d5c825a00,16811990710101,16811990721323,16811990817002,16811990818917
108,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,323,177088,177088,4194304,256,0,0,12,24,147584,0x7f0d5f807480,0x7f0d5c825a40,16811990833377,16811990841162,16811991022441,16811991024141
109,"void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]",0,0,326,177088,177088,4194304,256,0,0,12,24,149568,0x7f0d5f807300,0x7f0d5c825a80,16811991037181,16811991045161,16811991225640,16811991227185
110,"void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]",0,0,329,177088,177088,4194304,256,0,0,12,24,151552,0x7f0d5f807180,0x7f0d5c825ac0,16811991240594,16811991248520,16811991358600,16811991360171
111,"void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]",0,0,332,177088,177088,4194304,256,0,0,12,24,154304,0x7f0d5f807000,0x7f0d5c825b00,16811991373870,16811991382440,16811991563399,16811991564974
112,"void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]",0,0,335,177088,177088,4194304,256,0,0,12,24,156288,0x7f0d5d167e80,0x7f0d5c825b40,16811991585514,16811991595559,16811991696358,16811991698020
113,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,338,177088,177088,4194304,256,0,0,12,24,158784,0x7f0d5d167d00,0x7f0d5c825b80,16811991713380,16811991721478,16811991907397,16811991908934
114,"void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]",0,0,341,177088,177088,4194304,256,0,0,12,24,161024,0x7f0d5d167b80,0x7f0d5c825bc0,16811991922293,16811991930597,16811992112996,16811992114627
115,"void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]",0,0,344,177088,177088,4194304,256,0,0,12,24,163264,0x7f0d5d167a00,0x7f0d5c825c00,16811992128807,16811992136836,16811992258756,16811992260223
116,"void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]",0,0,347,177088,177088,4194304,256,0,0,12,24,166528,0x7f0d5d167880,0x7f0d5c825c40,16811992274632,16811992282755,16811992488834,16811992514365
117,"void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]",0,0,350,177088,177088,4194304,256,0,0,12,24,168512,0x7f0d5d167700,0x7f0d5c825c80,16811992529964,16811992540834,16811992652034,16811992653550
118,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,353,177088,177088,4194304,256,0,0,12,24,171264,0x7f0d5f807580,0x7f0d5c825cc0,16811992668860,16811992677474,16811992860993,16811992862654
119,"void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]",0,0,356,177088,177088,4194304,256,0,0,12,24,173760,0x7f0d5f807400,0x7f0d5c825d00,16811992876533,16811992885153,16811993071392,16811993072977
120,"void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]",0,0,359,177088,177088,4194304,256,0,0,12,24,176256,0x7f0d5f807280,0x7f0d5c825d40,16811993087487,16811993095552,16811993229791,16811993231252
121,"void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]",0,0,362,177088,177088,4194304,256,0,0,12,24,179776,0x7f0d5f807100,0x7f0d5c825d80,16811993246112,16811993254111,16811993486270,16811993505854
122,"void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]",0,0,365,177088,177088,4194304,256,0,0,12,24,182272,0x7f0d5d167f80,0x7f0d5c825dc0,16811993521503,16811993532190,16811993658749,16811993660289
123,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,368,177088,177088,4194304,256,0,0,12,24,185536,0x7f0d5d167e00,0x7f0d5c825e00,16811993675058,16811993682909,16811993874268,16811993875852
124,"void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]",0,0,371,177088,177088,4194304,256,0,0,12,24,188544,0x7f0d5d167c80,0x7f0d5c825e40,16811993890242,16811993899388,16811994092827,16811994094405
125,"void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]",0,0,374,177088,177088,4194304,256,0,0,12,24,191552,0x7f0d5d167b00,0x7f0d5c825e80,16811994118495,16811994128987,16811994288826,16811994290239
126,"void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]",0,0,377,177088,177088,4194304,256,0,0,12,24,195840,0x7f0d5d167980,0x7f0d5c825ec0,16811994305389,16811994313626,16811994596665,16811994616889
127,"void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]",0,0,380,177088,177088,4194304,256,0,0,12,24,198592,0x7f0d5d167800,0x7f0d5c825f00,16811994634998,16811994645624,16811994793624,16811994795153
128,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,383,177088,177088,4194304,256,0,0,12,24,202368,0x7f0d5d167680,0x7f0d5c825f40,16811994811033,16811994818584,16811995026423,16811995070795
129,"void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]",0,0,386,177088,177088,4194304,256,0,0,12,24,205888,0x7f0d5f807500,0x7f0d5c825f80,16811995078905,16811995088822,16811995294741,16811995335157
130,"void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]",0,0,389,177088,177088,4194304,256,0,0,12,24,209408,0x7f0d5f807380,0x7f0d5c825fc0,16811995342886,16811995353941,16811995539220,16811995540770
131,"void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]",0,0,392,177088,177088,4194304,256,0,0,12,24,214464,0x7f0d5f807200,0x7f0d5c826000,16811995555290,16811995563220,16811995897299,16811995916899
132,"void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]",0,0,395,177088,177088,4194304,256,0,0,12,24,217728,0x7f0d5f807080,0x7f0d5c826040,16811995934568,16811995945298,16811996112818,16811996114262
133,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,398,177088,177088,4194304,256,0,0,12,24,222016,0x7f0d5d167f00,0x7f0d5c826080,16811996129972,16811996137938,16811996367057,16811996386734
134,"void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]",0,0,401,177088,177088,4194304,256,0,0,12,24,226048,0x7f0d5d167d80,0x7f0d5c8260c0,16811996394304,16811996404656,16811996634575,16811996650746
135,"void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]",0,0,404,177088,177088,4194304,256,0,0,12,24,230080,0x7f0d5d167c00,0x7f0d5c826100,16811996658495,16811996668655,16811996879534,16811996895818
136,"void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]",0,0,407,177088,177088,4194304,256,0,0,12,24,235904,0x7f0d5d167a80,0x7f0d5c826140,16811996905008,16811996914414,16811997300012,16811997315725
137,"void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]",0,0,410,177088,177088,4194304,256,0,0,12,24,239424,0x7f0d5d167900,0x7f0d5c826180,16811997333744,16811997344172,16811997531211,16811997532828
138,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,413,177088,177088,4194304,256,0,0,12,24,244224,0x7f0d5d167780,0x7f0d5c8261c0,16811997547898,16811997555851,16811997806410,16811997822519
139,"void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]",0,0,416,177088,177088,4194304,256,0,0,12,24,248768,0x7f0d5d167600,0x7f0d5c826200,16811997830419,16811997840810,16811998088649,16811998113210
140,"void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]",0,0,419,177088,177088,4194304,256,0,0,12,24,253312,0x7f0d5f807480,0x7f0d5c826240,16811998121320,16811998131528,16811998368487,16811998387301
141,"void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]",0,0,422,177088,177088,4194304,256,0,0,8,24,255552,0x7f0d5f807300,0x7f0d5c826280,16811998395071,16811998405607,16811998859205,16811998878466
142,"void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]",0,0,425,177088,177088,4194304,256,0,0,12,24,259840,0x7f0d5f807180,0x7f0d5c8262c0,16811998894506,16811998905125,16811999130084,16811999152578
143,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,428,177088,177088,4194304,256,0,0,12,24,265664,0x7f0d5f807000,0x7f0d5c826300,16811999160307,16811999170724,16811999468962,16811999491097
144,"void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]",0,0,431,177088,177088,4194304,256,0,0,12,24,271232,0x7f0d5d167e80,0x7f0d5c826340,16811999501287,16811999510722,16811999809281,16811999831346
145,"void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]",0,0,434,177088,177088,4194304,256,0,0,8,24,273216,0x7f0d5d167d00,0x7f0d5c826380,16811999839206,16811999849600,16812000145919,16812000165186
146,"void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]",0,0,437,177088,177088,4194304,256,0,0,8,24,275712,0x7f0d5d167b80,0x7f0d5c8263c0,16812000173036,16812000183359,16812000741756,16812000766447
147,"void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]",0,0,440,177088,177088,4194304,256,0,0,12,24,280768,0x7f0d5d167a00,0x7f0d5c826400,16812000782937,16812000793436,16812001057435,16812001079658
148,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,443,177088,177088,4194304,256,0,0,12,24,287616,0x7f0d5d167880,0x7f0d5c826440,16812001088337,16812001098715,16812001448633,16812001470745
149,"void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]",0,0,446,177088,177088,4194304,256,0,0,12,24,294208,0x7f0d5d167700,0x7f0d5c826480,16812001479495,16812001489753,16812001840151,16812001858513
150,"void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]",0,0,449,177088,177088,4194304,256,0,0,8,24,296960,0x7f0d5f807580,0x7f0d5c8264c0,16812001867633,16812001877751,16812002224949,16812002249331
151,"void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]",0,0,452,177088,177088,4194304,256,0,0,8,24,299968,0x7f0d5f807400,0x7f0d5c826500,16812002258041,16812002267989,16812002929266,16812002949389
152,"void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]",0,0,455,177088,177088,4194304,256,0,0,8,24,302720,0x7f0d5f807280,0x7f0d5c826540,16812002968719,16812002978386,16812003426224,16812003448544
153,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,458,177088,177088,4194304,256,0,0,12,24,305472,0x7f0d5f807100,0x7f0d5c826580,16812003456874,16812003467024,16812003931821,16812003953918
154,"void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]",0,0,461,177088,177088,4194304,256,0,0,12,24,308224,0x7f0d5d167f80,0x7f0d5c8265c0,16812003962708,16812003973101,16812004434379,16812004455873
155,"void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]",0,0,464,177088,177088,4194304,256,0,0,8,24,310976,0x7f0d5d167e00,0x7f0d5c826600,16812004464952,16812004475019,16812004927337,16812004952527
156,"void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]",0,0,467,177088,177088,4194304,256,0,0,8,24,313216,0x7f0d5d167c80,0x7f0d5c826640,16812004961007,16812004971337,16812005839813,16812005862999
157,"void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]",0,0,470,177088,177088,4194304,256,0,0,8,24,315968,0x7f0d5d167b00,0x7f0d5c826680,16812005879118,16812005889732,16812006742688,16812006768691
158,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,473,177088,177088,4194304,256,0,0,12,24,318720,0x7f0d5d167980,0x7f0d5c8266c0,16812006777790,16812006787648,16812007672284,16812007696052
159,"void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]",0,0,476,177088,177088,4194304,256,0,0,12,24,321472,0x7f0d5d167800,0x7f0d5c826700,16812007704642,16812007714844,16812008594840,16812008617393
160,"void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]",0,0,479,177088,177088,4194304,256,0,0,8,24,324224,0x7f0d5d167680,0x7f0d5c826740,16812008626353,16812008649080,16812009507956,16812009551694
161,"void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]",0,0,482,177088,177088,4194304,256,0,0,8,24,326464,0x7f0d5f807500,0x7f0d5c826780,16812009560334,16812009570835,16812011270028,16812011343238
162,"void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]",0,0,485,177088,177088,4194304,256,0,0,8,24,329216,0x7f0d5f807380,0x7f0d5c8267c0,16812011359358,16812011370187,16812013035619,16812013106743
163,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,488,177088,177088,4194304,256,0,0,12,24,331968,0x7f0d5f807200,0x7f0d5c826800,16812013115453,16812013125859,16812014850971,16812014922486
164,"void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]",0,0,491,177088,177088,4194304,256,0,0,12,24,334720,0x7f0d5f807080,0x7f0d5c826840,16812014931996,16812014942811,16812016657523,16812016725400
165,"void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]",0,0,494,177088,177088,4194304,256,0,0,8,24,337472,0x7f0d5d167f00,0x7f0d5c826880,16812016733420,16812016747602,16812018421834,16812018493005
166,"void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]",0,0,497,177088,177088,4194304,256,0,0,8,24,0,0x7f0d5d167d80,0x7f0d5c8268c0,16812018501865,16812018511914,16812021869338,16812021943638