rocm-systems/tests/workloads/D_val_int/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",0,0,0,86746,86746,33554432,256,0,0,4,32,4160,0x0,0x7f936c204280,15073258600671,15073258626891,15073258868332,15073258979812
1,"void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]",0,0,2,86746,86746,32768,256,0,0,12,24,13888,0x0,0x7f936c223f80,15073263643990,15073263659546,15073263673146,15073263689999
2,"void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]",0,0,5,86746,86746,4194304,256,0,0,12,24,14336,0x7f936f231380,0x7f936c223fc0,15073263695949,15073263741946,15073263834267,15073263836336
3,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,8,86746,86746,4194304,256,0,0,16,24,15040,0x7f936f231200,0x7f936c224000,15073263858055,15073263867867,15073264046747,15073264049101
4,"void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]",0,0,11,86746,86746,4194304,256,0,0,16,24,15488,0x7f936f231080,0x7f936c224040,15073264064450,15073264073147,15073264252028,15073264254006
5,"void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]",0,0,14,86746,86746,4194304,256,0,0,12,24,15936,0x7f936cacdf00,0x7f936c224080,15073264268245,15073264276828,15073264370428,15073264372413
6,"void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]",0,0,17,86746,86746,4194304,256,0,0,12,24,16384,0x7f936cacdd80,0x7f936c2240c0,15073264388593,15073264396668,15073264488189,15073264490240
7,"void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]",0,0,20,86746,86746,4194304,256,0,0,12,24,17088,0x7f936cacdc00,0x7f936c224100,15073264517920,15073264528509,15073264620989,15073264622917
8,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,23,86746,86746,4194304,256,0,0,16,24,17792,0x7f936cacda80,0x7f936c224140,15073264641697,15073264665469,15073264845470,15073264847912
9,"void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]",0,0,26,86746,86746,4194304,256,0,0,16,24,18496,0x7f936cacd900,0x7f936c224180,15073264867201,15073264877630,15073265057630,15073265059707
10,"void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]",0,0,29,86746,86746,4194304,256,0,0,16,24,19200,0x7f936cacd780,0x7f936c2241c0,15073265078576,15073265087711,15073265181631,15073265183904
11,"void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]",0,0,32,86746,86746,4194304,256,0,0,16,24,19904,0x7f936cacd600,0x7f936c224200,15073265200223,15073265209471,15073265303391,15073265305381
12,"void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]",0,0,35,86746,86746,4194304,256,0,0,12,24,20608,0x7f936f231480,0x7f936c224240,15073265328370,15073265347071,15073265440672,15073265442847
13,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,38,86746,86746,4194304,256,0,0,16,24,21312,0x7f936f231300,0x7f936c224280,15073265461947,15073265472832,15073265652192,15073265654242
14,"void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]",0,0,41,86746,86746,4194304,256,0,0,16,24,22016,0x7f936f231180,0x7f936c2242c0,15073265673422,15073265684032,15073265864193,15073265866437
15,"void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]",0,0,44,86746,86746,4194304,256,0,0,16,24,22720,0x7f936f231000,0x7f936c224300,15073265885617,15073265906113,15073265999233,15073266001644
16,"void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]",0,0,47,86746,86746,4194304,256,0,0,12,24,23424,0x7f936cacde80,0x7f936c224340,15073266016494,15073266026593,15073266119554,15073266121661
17,"void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]",0,0,50,86746,86746,4194304,256,0,0,12,24,24128,0x7f936cacdd00,0x7f936c224380,15073266144111,15073266154274,15073266247554,15073266249598
18,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,53,86746,86746,4194304,256,0,0,16,24,24832,0x7f936cacdb80,0x7f936c2243c0,15073266270038,15073266293314,15073266473155,15073266475203
19,"void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]",0,0,56,86746,86746,4194304,256,0,0,16,24,25536,0x7f936cacda00,0x7f936c224400,15073266493062,15073266503875,15073266682595,15073266684638
20,"void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]",0,0,59,86746,86746,4194304,256,0,0,12,24,26240,0x7f936cacd880,0x7f936c224440,15073266701657,15073266729316,15073266822596,15073266824774
21,"void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]",0,0,62,86746,86746,4194304,256,0,0,12,24,26944,0x7f936cacd700,0x7f936c224480,15073266840084,15073266849316,15073266943396,15073266945432
22,"void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]",0,0,65,86746,86746,4194304,256,0,0,12,24,27648,0x7f936f231580,0x7f936c2244c0,15073266966481,15073266976676,15073267068997,15073267070979
23,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,68,86746,86746,4194304,256,0,0,16,24,28608,0x7f936f231400,0x7f936c224500,15073267086918,15073267095237,15073267273477,15073267275284
24,"void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]",0,0,71,86746,86746,4194304,256,0,0,16,24,29312,0x7f936f231280,0x7f936c224540,15073267292583,15073267300837,15073267480678,15073267482159
25,"void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]",0,0,74,86746,86746,4194304,256,0,0,12,24,30016,0x7f936f231100,0x7f936c224580,15073267497018,15073267505318,15073267597798,15073267599346
26,"void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]",0,0,77,86746,86746,4194304,256,0,0,8,24,30976,0x7f936cacdf80,0x7f936c2245c0,15073267614785,15073267623078,15073267714759,15073267716323
27,"void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]",0,0,80,86746,86746,4194304,256,0,0,12,24,31680,0x7f936cacde00,0x7f936c224600,15073267737922,15073267748519,15073267842759,15073267844280
28,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,83,86746,86746,4194304,256,0,0,16,24,32640,0x7f936cacdc80,0x7f936c224640,15073267860650,15073267869159,15073268048840,15073268050405
29,"void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]",0,0,86,86746,86746,4194304,256,0,0,12,24,33600,0x7f936cacdb00,0x7f936c224680,15073268065125,15073268073160,15073268254760,15073268256240
30,"void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]",0,0,89,86746,86746,4194304,256,0,0,12,24,34560,0x7f936cacd980,0x7f936c2246c0,15073268271100,15073268279240,15073268372681,15073268374187
31,"void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]",0,0,92,86746,86746,4194304,256,0,0,8,24,35520,0x7f936cacd800,0x7f936c224700,15073268388277,15073268396681,15073268490441,15073268491954
32,"void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]",0,0,95,86746,86746,4194304,256,0,0,12,24,36224,0x7f936cacd680,0x7f936c224740,15073268513574,15073268522921,15073268617001,15073268618421
33,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,98,86746,86746,4194304,256,0,0,12,24,37184,0x7f936f231500,0x7f936c224780,15073268632411,15073268641002,15073268820522,15073268822016
34,"void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]",0,0,101,86746,86746,4194304,256,0,0,12,24,38144,0x7f936f231380,0x7f936c2247c0,15073268836096,15073268845962,15073269025643,15073269027242
35,"void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]",0,0,104,86746,86746,4194304,256,0,0,12,24,39104,0x7f936f231200,0x7f936c224800,15073269044571,15073269052523,15073269146763,15073269148319
36,"void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]",0,0,107,86746,86746,4194304,256,0,0,8,24,40320,0x7f936f231080,0x7f936c224840,15073269164698,15073269173963,15073269268524,15073269270026
37,"void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]",0,0,110,86746,86746,4194304,256,0,0,12,24,41280,0x7f936cacdf00,0x7f936c224880,15073269292185,15073269302764,15073269395724,15073269397173
38,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,113,86746,86746,4194304,256,0,0,12,24,42240,0x7f936cacdd80,0x7f936c2248c0,15073269413242,15073269421644,15073269601325,15073269602818
39,"void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]",0,0,116,86746,86746,4194304,256,0,0,12,24,43200,0x7f936cacdc00,0x7f936c224900,15073269619467,15073269627885,15073269809805,15073269811283
40,"void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]",0,0,119,86746,86746,4194304,256,0,0,8,24,44160,0x7f936cacda80,0x7f936c224940,15073269826482,15073269835405,15073269929006,15073269930480
41,"void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]",0,0,122,86746,86746,4194304,256,0,0,12,24,45376,0x7f936cacd900,0x7f936c224980,15073269945049,15073269953486,15073270048366,15073270049847
42,"void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]",0,0,125,86746,86746,4194304,256,0,0,12,24,46336,0x7f936cacd780,0x7f936c2249c0,15073270070696,15073270081166,15073270174606,15073270176264
43,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,128,86746,86746,4194304,256,0,0,12,24,47552,0x7f936cacd600,0x7f936c224a00,15073270191204,15073270199726,15073270379247,15073270380729
44,"void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]",0,0,131,86746,86746,4194304,256,0,0,12,24,48512,0x7f936f231480,0x7f936c224a40,15073270395769,15073270403727,15073270583408,15073270585004
45,"void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]",0,0,134,86746,86746,4194304,256,0,0,8,24,49472,0x7f936f231300,0x7f936c224a80,15073270598654,15073270606608,15073270699248,15073270700921
46,"void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]",0,0,137,86746,86746,4194304,256,0,0,12,24,50688,0x7f936f231180,0x7f936c224ac0,15073270715091,15073270723568,15073270817008,15073270818629
47,"void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]",0,0,140,86746,86746,4194304,256,0,0,12,24,51648,0x7f936f231000,0x7f936c224b00,15073270839368,15073270850128,15073270941969,15073270943616
48,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,143,86746,86746,4194304,256,0,0,12,24,52864,0x7f936cacde80,0x7f936c224b40,15073270960395,15073270968369,15073271147249,15073271148871
49,"void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]",0,0,146,86746,86746,4194304,256,0,0,12,24,54080,0x7f936cacdd00,0x7f936c224b80,15073271163370,15073271171889,15073271352370,15073271353956
50,"void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]",0,0,149,86746,86746,4194304,256,0,0,8,24,55296,0x7f936cacdb80,0x7f936c224bc0,15073271368056,15073271376370,15073271469970,15073271471583
51,"void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]",0,0,152,86746,86746,4194304,256,0,0,12,24,56768,0x7f936cacda00,0x7f936c224c00,15073271486143,15073271494290,15073271589811,15073271591480
52,"void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]",0,0,155,86746,86746,4194304,256,0,0,12,24,57728,0x7f936cacd880,0x7f936c224c40,15073271611810,15073271622611,15073271715731,15073271717267
53,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,158,86746,86746,4194304,256,0,0,12,24,58944,0x7f936cacd700,0x7f936c224c80,15073271731747,15073271739731,15073271919412,15073271921072
54,"void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]",0,0,161,86746,86746,4194304,256,0,0,12,24,60160,0x7f936f231580,0x7f936c224cc0,15073271936102,15073271944372,15073272125012,15073272126747
55,"void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]",0,0,164,86746,86746,4194304,256,0,0,8,24,61376,0x7f936f231400,0x7f936c224d00,15073272140577,15073272148692,15073272243573,15073272245195
56,"void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]",0,0,167,86746,86746,4194304,256,0,0,12,24,62848,0x7f936f231280,0x7f936c224d40,15073272260414,15073272268853,15073272368373,15073272369872
57,"void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]",0,0,170,86746,86746,4194304,256,0,0,12,24,63808,0x7f936f231100,0x7f936c224d80,15073272394931,15073272405333,15073272499894,15073272501429
58,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,173,86746,86746,4194304,256,0,0,12,24,65024,0x7f936cacdf80,0x7f936c224dc0,15073272519208,15073272528694,15073272707254,15073272708874
59,"void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]",0,0,176,86746,86746,4194304,256,0,0,12,24,66240,0x7f936cacde00,0x7f936c224e00,15073272725763,15073272733654,15073272912055,15073272913519
60,"void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]",0,0,179,86746,86746,4194304,256,0,0,8,24,67456,0x7f936cacdc80,0x7f936c224e40,15073272928938,15073272937335,15073273031415,15073273033036
61,"void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]",0,0,182,86746,86746,4194304,256,0,0,12,24,68928,0x7f936cacdb00,0x7f936c224e80,15073273047825,15073273056215,15073273160056,15073273161623
62,"void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]",0,0,185,86746,86746,4194304,256,0,0,12,24,70144,0x7f936cacd980,0x7f936c224ec0,15073273183102,15073273193336,15073273285336,15073273286940
63,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,188,86746,86746,4194304,256,0,0,12,24,71616,0x7f936cacd800,0x7f936c224f00,15073273301169,15073273309816,15073273489657,15073273491265
64,"void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]",0,0,191,86746,86746,4194304,256,0,0,12,24,72832,0x7f936cacd680,0x7f936c224f40,15073273506825,15073273515417,15073273694297,15073273695840
65,"void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]",0,0,194,86746,86746,4194304,256,0,0,8,24,74048,0x7f936f231500,0x7f936c224f80,15073273710200,15073273718457,15073273812058,15073273813657
66,"void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]",0,0,197,86746,86746,4194304,256,0,0,12,24,75776,0x7f936f231380,0x7f936c224fc0,15073273830487,15073273838458,15073273945338,15073273946944
67,"void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]",0,0,200,86746,86746,4194304,256,0,0,12,24,76992,0x7f936f231200,0x7f936c225000,15073273966574,15073273977498,15073274071578,15073274073121
68,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,203,86746,86746,4194304,256,0,0,12,24,78464,0x7f936f231080,0x7f936c225040,15073274115740,15073274126299,15073274305659,15073274307315
69,"void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]",0,0,206,86746,86746,4194304,256,0,0,12,24,79936,0x7f936cacdf00,0x7f936c225080,15073274322105,15073274330459,15073274511100,15073274512800
70,"void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]",0,0,209,86746,86746,4194304,256,0,0,8,24,81408,0x7f936cacdd80,0x7f936c2250c0,15073274526790,15073274535420,15073274630300,15073274631978
71,"void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]",0,0,212,86746,86746,4194304,256,0,0,12,24,83136,0x7f936cacdc00,0x7f936c225100,15073274647137,15073274655740,15073274771421,15073274773014
72,"void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]",0,0,215,86746,86746,4194304,256,0,0,12,24,84352,0x7f936cacda80,0x7f936c225140,15073274809913,15073274819421,15073274913341,15073274914911
73,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,218,86746,86746,4194304,256,0,0,12,24,85824,0x7f936cacd900,0x7f936c225180,15073274930570,15073274938621,15073275118302,15073275119836
74,"void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]",0,0,221,86746,86746,4194304,256,0,0,12,24,87296,0x7f936cacd780,0x7f936c2251c0,15073275136526,15073275144382,15073275323742,15073275325241
75,"void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]",0,0,224,86746,86746,4194304,256,0,0,12,24,88768,0x7f936cacd600,0x7f936c225200,15073275339761,15073275348222,15073275442783,15073275444508
76,"void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]",0,0,227,86746,86746,4194304,256,0,0,12,24,90752,0x7f936f231480,0x7f936c225240,15073275459218,15073275467743,15073275587903,15073275589545
77,"void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]",0,0,230,86746,86746,4194304,256,0,0,12,24,91968,0x7f936f231300,0x7f936c225280,15073275610294,15073275620383,15073275713984,15073275715432
78,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,233,86746,86746,4194304,256,0,0,12,24,93440,0x7f936f231180,0x7f936c2252c0,15073275731051,15073275739264,15073275918784,15073275920487
79,"void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]",0,0,236,86746,86746,4194304,256,0,0,12,24,94912,0x7f936f231000,0x7f936c225300,15073275934886,15073275943104,15073276122625,15073276124272
80,"void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]",0,0,239,86746,86746,4194304,256,0,0,12,24,96384,0x7f936cacde80,0x7f936c225340,15073276139781,15073276147905,15073276243105,15073276244969
81,"void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]",0,0,242,86746,86746,4194304,256,0,0,12,24,98368,0x7f936cacdd00,0x7f936c225380,15073276263928,15073276273825,15073276401346,15073276402905
82,"void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]",0,0,245,86746,86746,4194304,256,0,0,12,24,99584,0x7f936cacdb80,0x7f936c2253c0,15073276424545,15073276434786,15073276529666,15073276531402
83,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,248,86746,86746,4194304,256,0,0,12,24,101312,0x7f936cacda00,0x7f936c225400,15073276546532,15073276554626,15073276733507,15073276735017
84,"void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]",0,0,251,86746,86746,4194304,256,0,0,12,24,102784,0x7f936cacd880,0x7f936c225440,15073276749057,15073276757827,15073276937987,15073276939532
85,"void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]",0,0,254,86746,86746,4194304,256,0,0,12,24,104256,0x7f936cacd700,0x7f936c225480,15073276953992,15073276962787,15073277057188,15073277058739
86,"void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]",0,0,257,86746,86746,4194304,256,0,0,12,24,106240,0x7f936f231580,0x7f936c2254c0,15073277073709,15073277081828,15073277212868,15073277214466
87,"void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]",0,0,260,86746,86746,4194304,256,0,0,12,24,107712,0x7f936f231400,0x7f936c225500,15073277234805,15073277245188,15073277338629,15073277340083
88,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,263,86746,86746,4194304,256,0,0,16,24,109440,0x7f936f231280,0x7f936c225540,15073277356792,15073277365349,15073277545829,15073277547328
89,"void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]",0,0,266,86746,86746,4194304,256,0,0,16,24,111168,0x7f936f231100,0x7f936c225580,15073277562177,15073277570789,15073277750470,15073277752043
90,"void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]",0,0,269,86746,86746,4194304,256,0,0,8,24,112896,0x7f936cacdf80,0x7f936c2255c0,15073277768682,15073277776550,15073277872390,15073277873880
91,"void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]",0,0,272,86746,86746,4194304,256,0,0,12,24,115136,0x7f936cacde00,0x7f936c225600,15073277888899,15073277897030,15073278036551,15073278038156
92,"void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]",0,0,275,86746,86746,4194304,256,0,0,12,24,116608,0x7f936cacdc80,0x7f936c225640,15073278058955,15073278069511,15073278164871,15073278166433
93,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,278,86746,86746,4194304,256,0,0,16,24,118336,0x7f936cacdb00,0x7f936c225680,15073278182852,15073278191271,15073278370792,15073278372558
94,"void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]",0,0,281,86746,86746,4194304,256,0,0,16,24,120064,0x7f936cacd980,0x7f936c2256c0,15073278386518,15073278395432,15073278575432,15073278577143
95,"void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]",0,0,284,86746,86746,4194304,256,0,0,12,24,121792,0x7f936cacd800,0x7f936c225700,15073278591833,15073278600073,15073278697513,15073278699100
96,"void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]",0,0,287,86746,86746,4194304,256,0,0,12,24,124032,0x7f936cacd680,0x7f936c225740,15073278714650,15073278722953,15073278869353,15073278870866
97,"void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]",0,0,290,86746,86746,4194304,256,0,0,12,24,125504,0x7f936f231500,0x7f936c225780,15073278890686,15073278901033,15073278996074,15073278997733
98,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,293,86746,86746,4194304,256,0,0,12,24,127488,0x7f936f231380,0x7f936c2257c0,15073279012353,15073279020554,15073279200714,15073279202358
99,"void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]",0,0,296,86746,86746,4194304,256,0,0,12,24,129216,0x7f936f231200,0x7f936c225800,15073279217008,15073279225195,15073279405355,15073279407063
100,"void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]",0,0,299,86746,86746,4194304,256,0,0,12,24,130944,0x7f936f231080,0x7f936c225840,15073279421423,15073279430635,15073279530635,15073279532100
101,"void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]",0,0,302,86746,86746,4194304,256,0,0,12,24,133440,0x7f936cacdf00,0x7f936c225880,15073279547820,15073279556076,15073279711756,15073279713316
102,"void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]",0,0,305,86746,86746,4194304,256,0,0,12,24,134912,0x7f936cacdd80,0x7f936c2258c0,15073279735405,15073279746316,15073279841516,15073279843173
103,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,308,86746,86746,4194304,256,0,0,12,24,136896,0x7f936cacdc00,0x7f936c225900,15073279858842,15073279866796,15073280045357,15073280046818
104,"void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]",0,0,311,86746,86746,4194304,256,0,0,12,24,138880,0x7f936cacda80,0x7f936c225940,15073280063798,15073280071917,15073280251918,15073280253443
105,"void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]",0,0,314,86746,86746,4194304,256,0,0,12,24,140864,0x7f936cacd900,0x7f936c225980,15073280267713,15073280275918,15073280382958,15073280384520
106,"void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]",0,0,317,86746,86746,4194304,256,0,0,12,24,143616,0x7f936cacd780,0x7f936c2259c0,15073280398210,15073280407598,15073280577519,15073280579045
107,"void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]",0,0,320,86746,86746,4194304,256,0,0,12,24,145344,0x7f936cacd600,0x7f936c225a00,15073280599705,15073280610159,15073280706479,15073280707942
108,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,323,86746,86746,4194304,256,0,0,12,24,147584,0x7f936f231480,0x7f936c225a40,15073280725032,15073280733359,15073280913520,15073280915117
109,"void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]",0,0,326,86746,86746,4194304,256,0,0,12,24,149568,0x7f936f231300,0x7f936c225a80,15073280929297,15073280937840,15073281119120,15073281120722
110,"void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]",0,0,329,86746,86746,4194304,256,0,0,12,24,151552,0x7f936f231180,0x7f936c225ac0,15073281134992,15073281143600,15073281254001,15073281255499
111,"void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]",0,0,332,86746,86746,4194304,256,0,0,12,24,154304,0x7f936f231000,0x7f936c225b00,15073281269269,15073281277361,15073281458801,15073281460284
112,"void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]",0,0,335,86746,86746,4194304,256,0,0,12,24,156288,0x7f936cacde80,0x7f936c225b40,15073281481664,15073281491122,15073281592242,15073281593861
113,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,338,86746,86746,4194304,256,0,0,12,24,158784,0x7f936cacdd00,0x7f936c225b80,15073281609611,15073281617842,15073281801683,15073281803276
114,"void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]",0,0,341,86746,86746,4194304,256,0,0,12,24,161024,0x7f936cacdb80,0x7f936c225bc0,15073281820046,15073281828403,15073282016083,15073282017651
115,"void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]",0,0,344,86746,86746,4194304,256,0,0,12,24,163264,0x7f936cacda00,0x7f936c225c00,15073282032941,15073282041203,15073282164084,15073282165937
116,"void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]",0,0,347,86746,86746,4194304,256,0,0,12,24,166528,0x7f936cacd880,0x7f936c225c40,15073282180827,15073282189364,15073282396244,15073282443021
117,"void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]",0,0,350,86746,86746,4194304,256,0,0,12,24,168512,0x7f936cacd700,0x7f936c225c80,15073282458730,15073282469685,15073282581205,15073282582637
118,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,353,86746,86746,4194304,256,0,0,12,24,171264,0x7f936f231580,0x7f936c225cc0,15073282599517,15073282608885,15073282792086,15073282793692
119,"void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]",0,0,356,86746,86746,4194304,256,0,0,12,24,173760,0x7f936f231400,0x7f936c225d00,15073282807692,15073282815926,15073283002006,15073283003657
120,"void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]",0,0,359,86746,86746,4194304,256,0,0,12,24,176256,0x7f936f231280,0x7f936c225d40,15073283019557,15073283027606,15073283162327,15073283164013
121,"void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]",0,0,362,86746,86746,4194304,256,0,0,12,24,179776,0x7f936f231100,0x7f936c225d80,15073283178383,15073283186647,15073283418808,15073283467276
122,"void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]",0,0,365,86746,86746,4194304,256,0,0,12,24,182272,0x7f936cacdf80,0x7f936c225dc0,15073283483706,15073283494328,15073283621048,15073283622552
123,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,368,86746,86746,4194304,256,0,0,12,24,185536,0x7f936cacde00,0x7f936c225e00,15073283637732,15073283645688,15073283836249,15073283837837
124,"void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]",0,0,371,86746,86746,4194304,256,0,0,12,24,188544,0x7f936cacdc80,0x7f936c225e40,15073283853077,15073283862169,15073284054010,15073284055672
125,"void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]",0,0,374,86746,86746,4194304,256,0,0,12,24,191552,0x7f936cacdb00,0x7f936c225e80,15073284072542,15073284081370,15073284241690,15073284243188
126,"void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]",0,0,377,86746,86746,4194304,256,0,0,12,24,195840,0x7f936cacd980,0x7f936c225ec0,15073284258377,15073284266970,15073284550491,15073284575610
127,"void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]",0,0,380,86746,86746,4194304,256,0,0,12,24,198592,0x7f936cacd800,0x7f936c225f00,15073284601289,15073284611771,15073284762012,15073284763545
128,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,383,86746,86746,4194304,256,0,0,12,24,202368,0x7f936cacd680,0x7f936c225f40,15073284779475,15073284787612,15073284998812,15073285018089
129,"void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]",0,0,386,86746,86746,4194304,256,0,0,12,24,205888,0x7f936f231500,0x7f936c225f80,15073285027239,15073285037213,15073285247773,15073285266773
130,"void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]",0,0,389,86746,86746,4194304,256,0,0,12,24,209408,0x7f936f231380,0x7f936c225fc0,15073285275693,15073285285533,15073285471134,15073285472678
131,"void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]",0,0,392,86746,86746,4194304,256,0,0,12,24,214464,0x7f936f231200,0x7f936c226000,15073285490008,15073285499294,15073285833695,15073285850429
132,"void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]",0,0,395,86746,86746,4194304,256,0,0,12,24,217728,0x7f936f231080,0x7f936c226040,15073285866969,15073285877215,15073286045536,15073286046934
133,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,398,86746,86746,4194304,256,0,0,12,24,222016,0x7f936cacdf00,0x7f936c226080,15073286062754,15073286071296,15073286299137,15073286317258
134,"void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]",0,0,401,86746,86746,4194304,256,0,0,12,24,226048,0x7f936cacdd80,0x7f936c2260c0,15073286328528,15073286338817,15073286568737,15073286585221
135,"void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]",0,0,404,86746,86746,4194304,256,0,0,12,24,230080,0x7f936cacdc00,0x7f936c226100,15073286596291,15073286606177,15073286817538,15073286833225
136,"void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]",0,0,407,86746,86746,4194304,256,0,0,12,24,235904,0x7f936cacda80,0x7f936c226140,15073286846375,15073286855618,15073287241539,15073287257035
137,"void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]",0,0,410,86746,86746,4194304,256,0,0,12,24,239424,0x7f936cacd900,0x7f936c226180,15073287276285,15073287286980,15073287473700,15073287475210
138,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,413,86746,86746,4194304,256,0,0,12,24,244224,0x7f936cacd780,0x7f936c2261c0,15073287493180,15073287501060,15073287751461,15073287767733
139,"void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]",0,0,416,86746,86746,4194304,256,0,0,12,24,248768,0x7f936cacd600,0x7f936c226200,15073287777103,15073287787141,15073288034662,15073288050016
140,"void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]",0,0,419,86746,86746,4194304,256,0,0,12,24,253312,0x7f936f231480,0x7f936c226240,15073288060066,15073288070342,15073288307303,15073288323080
141,"void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]",0,0,422,86746,86746,4194304,256,0,0,8,24,255552,0x7f936f231300,0x7f936c226280,15073288332290,15073288342343,15073288795944,15073288811988
142,"void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]",0,0,425,86746,86746,4194304,256,0,0,12,24,259840,0x7f936f231180,0x7f936c2262c0,15073288828908,15073288839304,15073289065705,15073289081482
143,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,428,86746,86746,4194304,256,0,0,12,24,265664,0x7f936f231000,0x7f936c226300,15073289092332,15073289102185,15073289400746,15073289416604
144,"void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]",0,0,431,86746,86746,4194304,256,0,0,12,24,271232,0x7f936cacde80,0x7f936c226340,15073289428804,15073289438346,15073289736747,15073289755996
145,"void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]",0,0,434,86746,86746,4194304,256,0,0,8,24,273216,0x7f936cacdd00,0x7f936c226380,15073289769635,15073289785067,15073290082188,15073290106147
146,"void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]",0,0,437,86746,86746,4194304,256,0,0,8,24,275712,0x7f936cacdb80,0x7f936c2263c0,15073290117977,15073290128108,15073290686830,15073290705843
147,"void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]",0,0,440,86746,86746,4194304,256,0,0,12,24,280768,0x7f936cacda00,0x7f936c226400,15073290722893,15073290733390,15073290997551,15073291042865
148,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,443,86746,86746,4194304,256,0,0,12,24,287616,0x7f936cacd880,0x7f936c226440,15073291056125,15073291066031,15073291417232,15073291460535
149,"void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]",0,0,446,86746,86746,4194304,256,0,0,12,24,294208,0x7f936cacd700,0x7f936c226480,15073291472735,15073291487313,15073291837714,15073291885575
150,"void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]",0,0,449,86746,86746,4194304,256,0,0,8,24,296960,0x7f936f231580,0x7f936c2264c0,15073291896915,15073291906354,15073292254035,15073292297445
151,"void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]",0,0,452,86746,86746,4194304,256,0,0,8,24,299968,0x7f936f231400,0x7f936c226500,15073292306105,15073292315795,15073292977397,15073293021548
152,"void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]",0,0,455,86746,86746,4194304,256,0,0,8,24,302720,0x7f936f231280,0x7f936c226540,15073293039228,15073293048758,15073293496439,15073293540326
153,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,458,86746,86746,4194304,256,0,0,12,24,305472,0x7f936f231100,0x7f936c226580,15073293550185,15073293559799,15073294024441,15073294043934
154,"void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]",0,0,461,86746,86746,4194304,256,0,0,12,24,308224,0x7f936cacdf80,0x7f936c2265c0,15073294054623,15073294065241,15073294526202,15073294572101
155,"void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]",0,0,464,86746,86746,4194304,256,0,0,8,24,310976,0x7f936cacde00,0x7f936c226600,15073294580711,15073294590842,15073295043324,15073295089578
156,"void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]",0,0,467,86746,86746,4194304,256,0,0,8,24,313216,0x7f936cacdc80,0x7f936c226640,15073295098558,15073295108764,15073295978367,15073296022186
157,"void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]",0,0,470,86746,86746,4194304,256,0,0,8,24,315968,0x7f936cacdb00,0x7f936c226680,15073296039206,15073296049727,15073296904610,15073296947944
158,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,473,86746,86746,4194304,256,0,0,12,24,318720,0x7f936cacd980,0x7f936c2266c0,15073296958904,15073296969410,15073297854212,15073297897461
159,"void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]",0,0,476,86746,86746,4194304,256,0,0,12,24,321472,0x7f936cacd800,0x7f936c226700,15073297910671,15073297920933,15073298801895,15073298849028
160,"void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]",0,0,479,86746,86746,4194304,256,0,0,8,24,324224,0x7f936cacd680,0x7f936c226740,15073298859428,15073298870536,15073299731018,15073299777386
161,"void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]",0,0,482,86746,86746,4194304,256,0,0,8,24,326464,0x7f936f231500,0x7f936c226780,15073299786216,15073299795819,15073301494064,15073301566004
162,"void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]",0,0,485,86746,86746,4194304,256,0,0,8,24,329216,0x7f936f231380,0x7f936c2267c0,15073301582553,15073301593104,15073303258389,15073303328751
163,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,488,86746,86746,4194304,256,0,0,12,24,331968,0x7f936f231200,0x7f936c226800,15073303340431,15073303375350,15073305099995,15073305169617
164,"void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]",0,0,491,86746,86746,4194304,256,0,0,12,24,334720,0x7f936f231080,0x7f936c226840,15073305182257,15073305197275,15073306913761,15073306983884
165,"void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]",0,0,494,86746,86746,4194304,256,0,0,8,24,337472,0x7f936cacdf00,0x7f936c226880,15073306994993,15073307004481,15073308678886,15073308746452
166,"void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]",0,0,497,86746,86746,4194304,256,0,0,8,24,0,0x7f936cacdd80,0x7f936c2268c0,15073308756891,15073308767046,15073312125137,15073312203149