rocm-systems/tests/workloads/TCP/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",0,0,0,151766,151766,33554432,256,0,0,4,32,4160,0x0,0x7fde9d204280,16227562875467,16227562900659,16227563139060,16227563229520
1,"void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]",0,0,2,151766,151766,32768,256,0,0,12,24,13888,0x0,0x7fde9d223f80,16227567920646,16227567936047,16227567949487,16227567968405
2,"void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]",0,0,5,151766,151766,4194304,256,0,0,12,24,14336,0x7fdea01c0380,0x7fde9d223fc0,16227567970275,16227568019567,16227568112208,16227568114092
3,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,8,151766,151766,4194304,256,0,0,16,24,15040,0x7fdea01c0200,0x7fde9d224000,16227568137061,16227568146928,16227568325489,16227568327447
4,"void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]",0,0,11,151766,151766,4194304,256,0,0,16,24,15488,0x7fdea01c0080,0x7fde9d224040,16227568342057,16227568350929,16227568530610,16227568532762
5,"void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]",0,0,14,151766,151766,4194304,256,0,0,12,24,15936,0x7fde9db91f00,0x7fde9d224080,16227568546912,16227568555250,16227568649011,16227568651010
6,"void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]",0,0,17,151766,151766,4194304,256,0,0,12,24,16384,0x7fde9db91d80,0x7fde9d2240c0,16227568665919,16227568674291,16227568766291,16227568768197
7,"void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]",0,0,20,151766,151766,4194304,256,0,0,12,24,17088,0x7fde9db91c00,0x7fde9d224100,16227568795687,16227568806292,16227568898772,16227568900584
8,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,23,151766,151766,4194304,256,0,0,16,24,17792,0x7fde9db91a80,0x7fde9d224140,16227568920864,16227568930772,16227569110773,16227569112570
9,"void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]",0,0,26,151766,151766,4194304,256,0,0,16,24,18496,0x7fde9db91900,0x7fde9d224180,16227569131079,16227569139573,16227569321014,16227569322845
10,"void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]",0,0,29,151766,151766,4194304,256,0,0,16,24,19200,0x7fde9db91780,0x7fde9d2241c0,16227569340985,16227569349655,16227569443095,16227569445082
11,"void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]",0,0,32,151766,151766,4194304,256,0,0,16,24,19904,0x7fde9db91600,0x7fde9d224200,16227569464942,16227569483255,16227569576376,16227569578279
12,"void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]",0,0,35,151766,151766,4194304,256,0,0,12,24,20608,0x7fdea01c0480,0x7fde9d224240,16227569600749,16227569620856,16227569714457,16227569716386
13,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,38,151766,151766,4194304,256,0,0,16,24,21312,0x7fdea01c0300,0x7fde9d224280,16227569736816,16227569754297,16227569933178,16227569935041
14,"void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]",0,0,41,151766,151766,4194304,256,0,0,16,24,22016,0x7fdea01c0180,0x7fde9d2242c0,16227569954251,16227569971578,16227570151099,16227570152867
15,"void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]",0,0,44,151766,151766,4194304,256,0,0,16,24,22720,0x7fdea01c0000,0x7fde9d224300,16227570169056,16227570179259,16227570273660,16227570275444
16,"void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]",0,0,47,151766,151766,4194304,256,0,0,12,24,23424,0x7fde9db91e80,0x7fde9d224340,16227570301303,16227570311580,16227570405021,16227570406931
17,"void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]",0,0,50,151766,151766,4194304,256,0,0,12,24,24128,0x7fde9db91d00,0x7fde9d224380,16227570428001,16227570438141,16227570531421,16227570533308
18,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,53,151766,151766,4194304,256,0,0,16,24,24832,0x7fde9db91b80,0x7fde9d2243c0,16227570552278,16227570563262,16227570741823,16227570743744
19,"void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]",0,0,56,151766,151766,4194304,256,0,0,16,24,25536,0x7fde9db91a00,0x7fde9d224400,16227570763743,16227570773983,16227570954624,16227570956549
20,"void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]",0,0,59,151766,151766,4194304,256,0,0,12,24,26240,0x7fde9db91880,0x7fde9d224440,16227570974518,16227570983744,16227571077345,16227571079276
21,"void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]",0,0,62,151766,151766,4194304,256,0,0,12,24,26944,0x7fde9db91700,0x7fde9d224480,16227571097596,16227571107425,16227571200865,16227571202853
22,"void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]",0,0,65,151766,151766,4194304,256,0,0,12,24,27648,0x7fdea01c0580,0x7fde9d2244c0,16227571225223,16227571235585,16227571327746,16227571329551
23,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,68,151766,151766,4194304,256,0,0,16,24,28608,0x7fdea01c0400,0x7fde9d224500,16227571347940,16227571374466,16227571553987,16227571555936
24,"void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]",0,0,71,151766,151766,4194304,256,0,0,16,24,29312,0x7fdea01c0280,0x7fde9d224540,16227571571285,16227571580707,16227571762308,16227571763671
25,"void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]",0,0,74,151766,151766,4194304,256,0,0,12,24,30016,0x7fdea01c0100,0x7fde9d224580,16227571780251,16227571788229,16227571879909,16227571881438
26,"void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]",0,0,77,151766,151766,4194304,256,0,0,8,24,30976,0x7fde9db91f80,0x7fde9d2245c0,16227571897798,16227571906469,16227571998470,16227571999816
27,"void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]",0,0,80,151766,151766,4194304,256,0,0,12,24,31680,0x7fde9db91e00,0x7fde9d224600,16227572022465,16227572032870,16227572125831,16227572127253
28,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,83,151766,151766,4194304,256,0,0,16,24,32640,0x7fde9db91c80,0x7fde9d224640,16227572143763,16227572152071,16227572332072,16227572333508
29,"void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]",0,0,86,151766,151766,4194304,256,0,0,12,24,33600,0x7fde9db91b00,0x7fde9d224680,16227572349158,16227572357352,16227572536393,16227572537734
30,"void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]",0,0,89,151766,151766,4194304,256,0,0,12,24,34560,0x7fde9db91980,0x7fde9d2246c0,16227572554713,16227572563273,16227572656874,16227572658111
31,"void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]",0,0,92,151766,151766,4194304,256,0,0,8,24,35520,0x7fde9db91800,0x7fde9d224700,16227572673641,16227572681834,16227572775434,16227572776758
32,"void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]",0,0,95,151766,151766,4194304,256,0,0,12,24,36224,0x7fde9db91680,0x7fde9d224740,16227572799638,16227572809354,16227572902315,16227572903736
33,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,98,151766,151766,4194304,256,0,0,12,24,37184,0x7fdea01c0500,0x7fde9d224780,16227572918465,16227572926475,16227573105676,16227573107061
34,"void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]",0,0,101,151766,151766,4194304,256,0,0,12,24,38144,0x7fdea01c0380,0x7fde9d2247c0,16227573122361,16227573132556,16227573313197,16227573314547
35,"void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]",0,0,104,151766,151766,4194304,256,0,0,12,24,39104,0x7fdea01c0200,0x7fde9d224800,16227573329456,16227573338958,16227573432398,16227573433944
36,"void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]",0,0,107,151766,151766,4194304,256,0,0,8,24,40320,0x7fdea01c0080,0x7fde9d224840,16227573448394,16227573457678,16227573551919,16227573553251
37,"void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]",0,0,110,151766,151766,4194304,256,0,0,12,24,41280,0x7fde9db91f00,0x7fde9d224880,16227573574201,16227573585039,16227573677839,16227573679388
38,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,113,151766,151766,4194304,256,0,0,12,24,42240,0x7fde9db91d80,0x7fde9d2248c0,16227573694538,16227573702800,16227573882001,16227573883394
39,"void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]",0,0,116,151766,151766,4194304,256,0,0,12,24,43200,0x7fde9db91c00,0x7fde9d224900,16227573899404,16227573908881,16227574088722,16227574090229
40,"void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]",0,0,119,151766,151766,4194304,256,0,0,8,24,44160,0x7fde9db91a80,0x7fde9d224940,16227574108029,16227574116562,16227574209683,16227574210977
41,"void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]",0,0,122,151766,151766,4194304,256,0,0,12,24,45376,0x7fde9db91900,0x7fde9d224980,16227574226496,16227574234323,16227574328563,16227574330014
42,"void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]",0,0,125,151766,151766,4194304,256,0,0,12,24,46336,0x7fde9db91780,0x7fde9d2249c0,16227574351094,16227574361203,16227574454964,16227574456331
43,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,128,151766,151766,4194304,256,0,0,12,24,47552,0x7fde9db91600,0x7fde9d224a00,16227574473281,16227574481524,16227574660245,16227574661557
44,"void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]",0,0,131,151766,151766,4194304,256,0,0,12,24,48512,0x7fdea01c0480,0x7fde9d224a40,16227574678386,16227574686165,16227574865206,16227574866572
45,"void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]",0,0,134,151766,151766,4194304,256,0,0,8,24,49472,0x7fdea01c0300,0x7fde9d224a80,16227574882212,16227574890006,16227574982167,16227574983490
46,"void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]",0,0,137,151766,151766,4194304,256,0,0,12,24,50688,0x7fdea01c0180,0x7fde9d224ac0,16227575000179,16227575009687,16227575102968,16227575104317
47,"void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]",0,0,140,151766,151766,4194304,256,0,0,12,24,51648,0x7fdea01c0000,0x7fde9d224b00,16227575126737,16227575137528,16227575229528,16227575230974
48,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,143,151766,151766,4194304,256,0,0,12,24,52864,0x7fde9db91e80,0x7fde9d224b40,16227575248014,16227575255929,16227575433850,16227575435230
49,"void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]",0,0,146,151766,151766,4194304,256,0,0,12,24,54080,0x7fde9db91d00,0x7fde9d224b80,16227575450529,16227575459610,16227575638331,16227575639645
50,"void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]",0,0,149,151766,151766,4194304,256,0,0,8,24,55296,0x7fde9db91b80,0x7fde9d224bc0,16227575654615,16227575663451,16227575757371,16227575758783
51,"void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]",0,0,152,151766,151766,4194304,256,0,0,12,24,56768,0x7fde9db91a00,0x7fde9d224c00,16227575775022,16227575783932,16227575879132,16227575880410
52,"void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]",0,0,155,151766,151766,4194304,256,0,0,12,24,57728,0x7fde9db91880,0x7fde9d224c40,16227575900330,16227575911292,16227576006013,16227576007327
53,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,158,151766,151766,4194304,256,0,0,12,24,58944,0x7fde9db91700,0x7fde9d224c80,16227576023857,16227576031613,16227576211294,16227576212583
54,"void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]",0,0,161,151766,151766,4194304,256,0,0,12,24,60160,0x7fdea01c0580,0x7fde9d224cc0,16227576229002,16227576237054,16227576417375,16227576418718
55,"void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]",0,0,164,151766,151766,4194304,256,0,0,8,24,61376,0x7fdea01c0400,0x7fde9d224d00,16227576433588,16227576441855,16227576536736,16227576538085
56,"void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]",0,0,167,151766,151766,4194304,256,0,0,12,24,62848,0x7fdea01c0280,0x7fde9d224d40,16227576554725,16227576563616,16227576663137,16227576664563
57,"void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]",0,0,170,151766,151766,4194304,256,0,0,12,24,63808,0x7fdea01c0100,0x7fde9d224d80,16227576685402,16227576696097,16227576789217,16227576790440
58,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,173,151766,151766,4194304,256,0,0,12,24,65024,0x7fde9db91f80,0x7fde9d224dc0,16227576809030,16227576817857,16227576996098,16227576997475
59,"void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]",0,0,176,151766,151766,4194304,256,0,0,12,24,66240,0x7fde9db91e00,0x7fde9d224e00,16227577013545,16227577021699,16227577203140,16227577204631
60,"void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]",0,0,179,151766,151766,4194304,256,0,0,8,24,67456,0x7fde9db91c80,0x7fde9d224e40,16227577220100,16227577227940,16227577322660,16227577323918
61,"void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]",0,0,182,151766,151766,4194304,256,0,0,12,24,68928,0x7fde9db91b00,0x7fde9d224e80,16227577337948,16227577346981,16227577450981,16227577452365
62,"void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]",0,0,185,151766,151766,4194304,256,0,0,12,24,70144,0x7fde9db91980,0x7fde9d224ec0,16227577473235,16227577483941,16227577576582,16227577577993
63,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,188,151766,151766,4194304,256,0,0,12,24,71616,0x7fde9db91800,0x7fde9d224f00,16227577595232,16227577603622,16227577784583,16227577785848
64,"void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]",0,0,191,151766,151766,4194304,256,0,0,12,24,72832,0x7fde9db91680,0x7fde9d224f40,16227577802908,16227577811303,16227577992104,16227577993443
65,"void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]",0,0,194,151766,151766,4194304,256,0,0,8,24,74048,0x7fdea01c0500,0x7fde9d224f80,16227578009483,16227578017544,16227578111465,16227578112751
66,"void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]",0,0,197,151766,151766,4194304,256,0,0,12,24,75776,0x7fdea01c0380,0x7fde9d224fc0,16227578128260,16227578136905,16227578244106,16227578245518
67,"void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]",0,0,200,151766,151766,4194304,256,0,0,12,24,76992,0x7fdea01c0200,0x7fde9d225000,16227578265097,16227578276106,16227578370346,16227578371895
68,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,203,151766,151766,4194304,256,0,0,12,24,78464,0x7fdea01c0080,0x7fde9d225040,16227578387445,16227578395467,16227578574668,16227578576081
69,"void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]",0,0,206,151766,151766,4194304,256,0,0,12,24,79936,0x7fde9db91f00,0x7fde9d225080,16227578591450,16227578599788,16227578779789,16227578781186
70,"void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]",0,0,209,151766,151766,4194304,256,0,0,8,24,81408,0x7fde9db91d80,0x7fde9d2250c0,16227578797216,16227578805549,16227578900589,16227578901993
71,"void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]",0,0,212,151766,151766,4194304,256,0,0,12,24,83136,0x7fde9db91c00,0x7fde9d225100,16227578917883,16227578926350,16227579041870,16227579043200
72,"void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]",0,0,215,151766,151766,4194304,256,0,0,12,24,84352,0x7fde9db91a80,0x7fde9d225140,16227579079479,16227579089230,16227579183471,16227579184857
73,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,218,151766,151766,4194304,256,0,0,12,24,85824,0x7fde9db91900,0x7fde9d225180,16227579200727,16227579209231,16227579388752,16227579390093
74,"void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]",0,0,221,151766,151766,4194304,256,0,0,12,24,87296,0x7fde9db91780,0x7fde9d2251c0,16227579406162,16227579414832,16227579593073,16227579594528
75,"void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]",0,0,224,151766,151766,4194304,256,0,0,12,24,88768,0x7fde9db91600,0x7fde9d225200,16227579610888,16227579619474,16227579714034,16227579715445
76,"void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]",0,0,227,151766,151766,4194304,256,0,0,12,24,90752,0x7fdea01c0480,0x7fde9d225240,16227579730315,16227579739954,16227579860595,16227579861962
77,"void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]",0,0,230,151766,151766,4194304,256,0,0,12,24,91968,0x7fdea01c0300,0x7fde9d225280,16227579882882,16227579893555,16227579987796,16227579989129
78,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,233,151766,151766,4194304,256,0,0,12,24,93440,0x7fdea01c0180,0x7fde9d2252c0,16227580004139,16227580012276,16227580192597,16227580193945
79,"void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]",0,0,236,151766,151766,4194304,256,0,0,12,24,94912,0x7fdea01c0000,0x7fde9d225300,16227580209754,16227580217877,16227580399478,16227580400920
80,"void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]",0,0,239,151766,151766,4194304,256,0,0,12,24,96384,0x7fde9db91e80,0x7fde9d225340,16227580416270,16227580424438,16227580519959,16227580521257
81,"void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]",0,0,242,151766,151766,4194304,256,0,0,12,24,98368,0x7fde9db91d00,0x7fde9d225380,16227580538037,16227580546199,16227580672280,16227580673634
82,"void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]",0,0,245,151766,151766,4194304,256,0,0,12,24,99584,0x7fde9db91b80,0x7fde9d2253c0,16227580694104,16227580704280,16227580798520,16227580799791
83,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,248,151766,151766,4194304,256,0,0,12,24,101312,0x7fde9db91a00,0x7fde9d225400,16227580816531,16227580824280,16227581004441,16227581005907
84,"void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]",0,0,251,151766,151766,4194304,256,0,0,12,24,102784,0x7fde9db91880,0x7fde9d225440,16227581022136,16227581030202,16227581211483,16227581212842
85,"void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]",0,0,254,151766,151766,4194304,256,0,0,12,24,104256,0x7fde9db91700,0x7fde9d225480,16227581226122,16227581234363,16227581328443,16227581329870
86,"void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]",0,0,257,151766,151766,4194304,256,0,0,12,24,106240,0x7fdea01c0580,0x7fde9d2254c0,16227581344229,16227581352443,16227581482844,16227581484396
87,"void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]",0,0,260,151766,151766,4194304,256,0,0,12,24,107712,0x7fdea01c0400,0x7fde9d225500,16227581504656,16227581515484,16227581608925,16227581610323
88,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,263,151766,151766,4194304,256,0,0,16,24,109440,0x7fdea01c0280,0x7fde9d225540,16227581628063,16227581636285,16227581816606,16227581817929
89,"void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]",0,0,266,151766,151766,4194304,256,0,0,16,24,111168,0x7fdea01c0100,0x7fde9d225580,16227581834168,16227581842046,16227582020607,16227582021974
90,"void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]",0,0,269,151766,151766,4194304,256,0,0,8,24,112896,0x7fde9db91f80,0x7fde9d2255c0,16227582036574,16227582044767,16227582139648,16227582141002
91,"void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]",0,0,272,151766,151766,4194304,256,0,0,12,24,115136,0x7fde9db91e00,0x7fde9d225600,16227582155921,16227582164288,16227582303969,16227582305528
92,"void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]",0,0,275,151766,151766,4194304,256,0,0,12,24,116608,0x7fde9db91c80,0x7fde9d225640,16227582325507,16227582336289,16227582430210,16227582431795
93,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,278,151766,151766,4194304,256,0,0,16,24,118336,0x7fde9db91b00,0x7fde9d225680,16227582445945,16227582454050,16227582633411,16227582634671
94,"void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]",0,0,281,151766,151766,4194304,256,0,0,16,24,120064,0x7fde9db91980,0x7fde9d2256c0,16227582648810,16227582657731,16227582838052,16227582839486
95,"void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]",0,0,284,151766,151766,4194304,256,0,0,12,24,121792,0x7fde9db91800,0x7fde9d225700,16227582854386,16227582862372,16227582959493,16227582960903
96,"void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]",0,0,287,151766,151766,4194304,256,0,0,12,24,124032,0x7fde9db91680,0x7fde9d225740,16227582978283,16227582986533,16227583131494,16227583132790
97,"void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]",0,0,290,151766,151766,4194304,256,0,0,12,24,125504,0x7fdea01c0500,0x7fde9d225780,16227583154079,16227583164774,16227583259654,16227583261027
98,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,293,151766,151766,4194304,256,0,0,12,24,127488,0x7fdea01c0380,0x7fde9d2257c0,16227583276986,16227583284935,16227583465736,16227583467142
99,"void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]",0,0,296,151766,151766,4194304,256,0,0,12,24,129216,0x7fdea01c0200,0x7fde9d225800,16227583481452,16227583491016,16227583671817,16227583673208
100,"void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]",0,0,299,151766,151766,4194304,256,0,0,12,24,130944,0x7fdea01c0080,0x7fde9d225840,16227583687727,16227583698057,16227583797738,16227583799145
101,"void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]",0,0,302,151766,151766,4194304,256,0,0,12,24,133440,0x7fde9db91f00,0x7fde9d225880,16227583815584,16227583823658,16227583979019,16227583980421
102,"void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]",0,0,305,151766,151766,4194304,256,0,0,12,24,134912,0x7fde9db91d80,0x7fde9d2258c0,16227584000800,16227584011499,16227584106379,16227584107698
103,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,308,151766,151766,4194304,256,0,0,12,24,136896,0x7fde9db91c00,0x7fde9d225900,16227584123848,16227584131979,16227584311340,16227584312673
104,"void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]",0,0,311,151766,151766,4194304,256,0,0,12,24,138880,0x7fde9db91a80,0x7fde9d225940,16227584329343,16227584337901,16227584518862,16227584520269
105,"void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]",0,0,314,151766,151766,4194304,256,0,0,12,24,140864,0x7fde9db91900,0x7fde9d225980,16227584534738,16227584543662,16227584649582,16227584651026
106,"void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]",0,0,317,151766,151766,4194304,256,0,0,12,24,143616,0x7fde9db91780,0x7fde9d2259c0,16227584667415,16227584675823,16227584846064,16227584847412
107,"void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]",0,0,320,151766,151766,4194304,256,0,0,12,24,145344,0x7fde9db91600,0x7fde9d225a00,16227584868981,16227584879824,16227584975184,16227584976599
108,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,323,151766,151766,4194304,256,0,0,12,24,147584,0x7fdea01c0480,0x7fde9d225a40,16227584993238,16227585001904,16227585182705,16227585184114
109,"void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]",0,0,326,151766,151766,4194304,256,0,0,12,24,149568,0x7fdea01c0300,0x7fde9d225a80,16227585198384,16227585207506,16227585390387,16227585391689
110,"void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]",0,0,329,151766,151766,4194304,256,0,0,12,24,151552,0x7fdea01c0180,0x7fde9d225ac0,16227585407579,16227585416787,16227585526547,16227585527876
111,"void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]",0,0,332,151766,151766,4194304,256,0,0,12,24,154304,0x7fdea01c0000,0x7fde9d225b00,16227585542006,16227585551508,16227585732789,16227585734192
112,"void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]",0,0,335,151766,151766,4194304,256,0,0,12,24,156288,0x7fde9db91e80,0x7fde9d225b40,16227585755451,16227585764949,16227585865909,16227585867239
113,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,338,151766,151766,4194304,256,0,0,12,24,158784,0x7fde9db91d00,0x7fde9d225b80,16227585883189,16227585891190,16227586075671,16227586082044
114,"void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]",0,0,341,151766,151766,4194304,256,0,0,12,24,161024,0x7fde9db91b80,0x7fde9d225bc0,16227586097224,16227586107671,16227586289912,16227586291290
115,"void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]",0,0,344,151766,151766,4194304,256,0,0,12,24,163264,0x7fde9db91a00,0x7fde9d225c00,16227586305609,16227586314712,16227586436633,16227586438226
116,"void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]",0,0,347,151766,151766,4194304,256,0,0,12,24,166528,0x7fde9db91880,0x7fde9d225c40,16227586455146,16227586463193,16227586669754,16227586690611
117,"void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]",0,0,350,151766,151766,4194304,256,0,0,12,24,168512,0x7fde9db91700,0x7fde9d225c80,16227586707340,16227586717914,16227586829275,16227586830518
118,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,353,151766,151766,4194304,256,0,0,12,24,171264,0x7fdea01c0580,0x7fde9d225cc0,16227586846307,16227586855355,16227587038076,16227587039523
119,"void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]",0,0,356,151766,151766,4194304,256,0,0,12,24,173760,0x7fdea01c0400,0x7fde9d225d00,16227587054673,16227587062556,16227587247357,16227587248738
120,"void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]",0,0,359,151766,151766,4194304,256,0,0,12,24,176256,0x7fdea01c0280,0x7fde9d225d40,16227587265918,16227587273757,16227587407838,16227587409215
121,"void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]",0,0,362,151766,151766,4194304,256,0,0,12,24,179776,0x7fdea01c0100,0x7fde9d225d80,16227587425195,16227587434718,16227587667040,16227587686839
122,"void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]",0,0,365,151766,151766,4194304,256,0,0,12,24,182272,0x7fde9db91f80,0x7fde9d225dc0,16227587703598,16227587714400,16227587840961,16227587842415
123,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,368,151766,151766,4194304,256,0,0,12,24,185536,0x7fde9db91e00,0x7fde9d225e00,16227587858575,16227587866721,16227588056162,16227588057461
124,"void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]",0,0,371,151766,151766,4194304,256,0,0,12,24,188544,0x7fde9db91c80,0x7fde9d225e40,16227588072560,16227588082242,16227588274083,16227588275456
125,"void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]",0,0,374,151766,151766,4194304,256,0,0,12,24,191552,0x7fde9db91b00,0x7fde9d225e80,16227588291226,16227588299203,16227588459364,16227588460732
126,"void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]",0,0,377,151766,151766,4194304,256,0,0,12,24,195840,0x7fde9db91980,0x7fde9d225ec0,16227588477221,16227588485444,16227588769126,16227588788845
127,"void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]",0,0,380,151766,151766,4194304,256,0,0,12,24,198592,0x7fde9db91800,0x7fde9d225f00,16227588805124,16227588815526,16227588965127,16227588966491
128,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,383,151766,151766,4194304,256,0,0,12,24,202368,0x7fde9db91680,0x7fde9d225f40,16227588982530,16227588990567,16227589198408,16227589217415
129,"void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]",0,0,386,151766,151766,4194304,256,0,0,12,24,205888,0x7fdea01c0500,0x7fde9d225f80,16227589228975,16227589239849,16227589448490,16227589466650
130,"void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]",0,0,389,151766,151766,4194304,256,0,0,12,24,209408,0x7fdea01c0380,0x7fde9d225fc0,16227589475159,16227589485290,16227589670731,16227589672135
131,"void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]",0,0,392,151766,151766,4194304,256,0,0,12,24,214464,0x7fdea01c0200,0x7fde9d226000,16227589688275,16227589696971,16227590031853,16227590047597
132,"void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]",0,0,395,151766,151766,4194304,256,0,0,12,24,217728,0x7fdea01c0080,0x7fde9d226040,16227590064686,16227590075054,16227590243534,16227590244882
133,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,398,151766,151766,4194304,256,0,0,12,24,222016,0x7fde9db91f00,0x7fde9d226080,16227590262262,16227590270095,16227590497936,16227590513857
134,"void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]",0,0,401,151766,151766,4194304,256,0,0,12,24,226048,0x7fde9db91d80,0x7fde9d2260c0,16227590522186,16227590532496,16227590761777,16227590777591
135,"void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]",0,0,404,151766,151766,4194304,256,0,0,12,24,230080,0x7fde9db91c00,0x7fde9d226100,16227590789090,16227590799378,16227591010739,16227591026165
136,"void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]",0,0,407,151766,151766,4194304,256,0,0,12,24,235904,0x7fde9db91a80,0x7fde9d226140,16227591038785,16227591048339,16227591433621,16227591449216
137,"void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]",0,0,410,151766,151766,4194304,256,0,0,12,24,239424,0x7fde9db91900,0x7fde9d226180,16227591465415,16227591475862,16227591662903,16227591664301
138,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,413,151766,151766,4194304,256,0,0,12,24,244224,0x7fde9db91780,0x7fde9d2261c0,16227591680531,16227591688503,16227591939704,16227591955405
139,"void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]",0,0,416,151766,151766,4194304,256,0,0,12,24,248768,0x7fde9db91600,0x7fde9d226200,16227591964134,16227591974104,16227592220986,16227592237568
140,"void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]",0,0,419,151766,151766,4194304,256,0,0,12,24,253312,0x7fdea01c0480,0x7fde9d226240,16227592248468,16227592259066,16227592495867,16227592510842
141,"void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]",0,0,422,151766,151766,4194304,256,0,0,8,24,255552,0x7fdea01c0300,0x7fde9d226280,16227592519442,16227592530108,16227592983550,16227592999151
142,"void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]",0,0,425,151766,151766,4194304,256,0,0,12,24,259840,0x7fdea01c0180,0x7fde9d2262c0,16227593015471,16227593026270,16227593251872,16227593267805
143,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,428,151766,151766,4194304,256,0,0,12,24,265664,0x7fdea01c0000,0x7fde9d226300,16227593278925,16227593289152,16227593587714,16227593602968
144,"void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]",0,0,431,151766,151766,4194304,256,0,0,12,24,271232,0x7fde9db91e80,0x7fde9d226340,16227593615268,16227593625154,16227593924036,16227593941601
145,"void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]",0,0,434,151766,151766,4194304,256,0,0,8,24,273216,0x7fde9db91d00,0x7fde9d226380,16227593953910,16227593964516,16227594261158,16227594279713
146,"void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]",0,0,437,151766,151766,4194304,256,0,0,8,24,275712,0x7fde9db91b80,0x7fde9d2263c0,16227594291123,16227594301478,16227594860521,16227594903779
147,"void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]",0,0,440,151766,151766,4194304,256,0,0,12,24,280768,0x7fde9db91a00,0x7fde9d226400,16227594920569,16227594931241,16227595195083,16227595238142
148,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,443,151766,151766,4194304,256,0,0,12,24,287616,0x7fde9db91880,0x7fde9d226440,16227595247222,16227595257803,16227595608205,16227595651213
149,"void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]",0,0,446,151766,151766,4194304,256,0,0,12,24,294208,0x7fde9db91700,0x7fde9d226480,16227595663863,16227595681326,16227596031248,16227596074204
150,"void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]",0,0,449,151766,151766,4194304,256,0,0,8,24,296960,0x7fdea01c0580,0x7fde9d2264c0,16227596083223,16227596093328,16227596440530,16227596462705
151,"void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]",0,0,452,151766,151766,4194304,256,0,0,8,24,299968,0x7fdea01c0400,0x7fde9d226500,16227596471345,16227596480850,16227597143254,16227597162120
152,"void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]",0,0,455,151766,151766,4194304,256,0,0,8,24,302720,0x7fdea01c0280,0x7fde9d226540,16227597181389,16227597191094,16227597638457,16227597654409
153,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,458,151766,151766,4194304,256,0,0,12,24,305472,0x7fdea01c0100,0x7fde9d226580,16227597666789,16227597677017,16227598141660,16227598187987
154,"void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]",0,0,461,151766,151766,4194304,256,0,0,12,24,308224,0x7fde9db91f80,0x7fde9d2265c0,16227598197867,16227598208540,16227598669343,16227598689116
155,"void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]",0,0,464,151766,151766,4194304,256,0,0,8,24,310976,0x7fde9db91e00,0x7fde9d226600,16227598697856,16227598708223,16227599160226,16227599203515
156,"void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]",0,0,467,151766,151766,4194304,256,0,0,8,24,313216,0x7fde9db91c80,0x7fde9d226640,16227599214344,16227599224546,16227600093351,16227600138954
157,"void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]",0,0,470,151766,151766,4194304,256,0,0,8,24,315968,0x7fde9db91b00,0x7fde9d226680,16227600155663,16227600166631,16227601021196,16227601064633
158,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,473,151766,151766,4194304,256,0,0,12,24,318720,0x7fde9db91980,0x7fde9d2266c0,16227601076303,16227601086797,16227601971122,16227601987493
159,"void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]",0,0,476,151766,151766,4194304,256,0,0,12,24,321472,0x7fde9db91800,0x7fde9d226700,16227601999653,16227602010162,16227602890807,16227602934392
160,"void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]",0,0,479,151766,151766,4194304,256,0,0,8,24,324224,0x7fde9db91680,0x7fde9d226740,16227602944532,16227602954807,16227603814972,16227603858571
161,"void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]",0,0,482,151766,151766,4194304,256,0,0,8,24,326464,0x7fdea01c0500,0x7fde9d226780,16227603867511,16227603877693,16227605577222,16227605648162
162,"void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]",0,0,485,151766,151766,4194304,256,0,0,8,24,329216,0x7fdea01c0380,0x7fde9d2267c0,16227605665631,16227605676103,16227607341073,16227607409063
163,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,488,151766,151766,4194304,256,0,0,12,24,331968,0x7fdea01c0200,0x7fde9d226800,16227607418723,16227607429393,16227609154203,16227609222143
164,"void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]",0,0,491,151766,151766,4194304,256,0,0,12,24,334720,0x7fdea01c0080,0x7fde9d226840,16227609234053,16227609251644,16227610968613,16227611036953
165,"void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]",0,0,494,151766,151766,4194304,256,0,0,8,24,337472,0x7fde9db91f00,0x7fde9d226880,16227611050283,16227611065894,16227612740144,16227612807654
166,"void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]",0,0,497,151766,151766,4194304,256,0,0,8,24,0,0x7fde9db91d80,0x7fde9d2268c0,16227612816853,16227612828144,16227616185763,16227616241408