rocm-systems/tests/workloads/D_str_inv1/mi100/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_fillBuffer.kd",0,0,0,830681,830686,33554432,256,0,0,8,32,6464,0x0,0x7fbf15604180,12074463093222399,12074463093264885,12074463093587922,12074463093698343
1,"void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]",0,0,2,830681,830686,32768,256,0,0,24,24,12480,0x0,0x7fbf15635100,12074463108256507,12074463108356859,12074463108363579,12074463108368686
2,"void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]",0,0,4,830681,830686,4194304,256,0,0,24,24,12928,0x7fc021203900,0x7fbf15635140,12074463108410934,12074463108424379,12074463108559737,12074463108562777
3,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,6,830681,830686,4194304,256,0,0,36,24,13632,0x7fc021203800,0x7fbf15635180,12074463108604284,12074463108621657,12074463108868054,12074463108967409
4,"void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]",0,0,8,830681,830686,4194304,256,0,0,28,24,14080,0x7fc021203700,0x7fbf156351c0,12074463108990111,12074463109002293,12074463109247571,12074463109349388
5,"void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]",0,0,10,830681,830686,4194304,256,0,0,24,24,14528,0x7fc021203600,0x7fbf15635200,12074463109376328,12074463109387889,12074463109513808,12074463109516268
6,"void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]",0,0,12,830681,830686,4194304,256,0,0,24,24,14976,0x7fc021203500,0x7fbf15635240,12074463109547346,12074463109559248,12074463109685006,12074463109687837
7,"void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]",0,0,14,830681,830686,4194304,256,0,0,24,24,15424,0x7fc021203400,0x7fbf15635280,12074463109734985,12074463109746446,12074463109870605,12074463109873362
8,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,16,830681,830686,4194304,256,0,0,28,24,16128,0x7fc021203300,0x7fbf156352c0,12074463109899841,12074463109911244,12074463110193642,12074463110295897
9,"void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]",0,0,18,830681,830686,4194304,256,0,0,28,24,16832,0x7fc021203a00,0x7fbf15635300,12074463110316305,12074463110332680,12074463110576678,12074463110677797
10,"void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]",0,0,20,830681,830686,4194304,256,0,0,24,24,17536,0x7fc021203900,0x7fbf15635340,12074463110698626,12074463110709957,12074463110834595,12074463110837574
11,"void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]",0,0,22,830681,830686,4194304,256,0,0,24,24,18240,0x7fc021203800,0x7fbf15635380,12074463110863542,12074463110874915,12074463111000034,12074463111007710
12,"void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]",0,0,24,830681,830686,4194304,256,0,0,24,24,18944,0x7fc021203700,0x7fbf156353c0,12074463111040752,12074463111053473,12074463111177792,12074463111180441
13,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,26,830681,830686,4194304,256,0,0,28,24,19904,0x7fc021203600,0x7fbf15635400,12074463111205838,12074463111217312,12074463111462589,12074463111537925
14,"void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]",0,0,28,830681,830686,4194304,256,0,0,28,24,20608,0x7fc021203500,0x7fbf15635440,12074463111562752,12074463111573948,12074463111819226,12074463111891112
15,"void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]",0,0,30,830681,830686,4194304,256,0,0,24,24,21312,0x7fc021203400,0x7fbf15635480,12074463111916218,12074463111929625,12074463112054424,12074463112057261
16,"void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]",0,0,32,830681,830686,4194304,256,0,0,24,24,22016,0x7fc021203300,0x7fbf156354c0,12074463112083830,12074463112094903,12074463112220502,12074463112223069
17,"void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]",0,0,34,830681,830686,4194304,256,0,0,24,24,22720,0x7fc021203a00,0x7fbf15635500,12074463112257533,12074463112268502,12074463112393621,12074463112396321
18,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,36,830681,830686,4194304,256,0,0,28,24,23680,0x7fc021203900,0x7fbf15635540,12074463112421137,12074463112433940,12074463112719057,12074463112763653
19,"void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]",0,0,38,830681,830686,4194304,256,0,0,28,24,24384,0x7fc021203800,0x7fbf15635580,12074463112783420,12074463112796337,12074463113040974,12074463113113734
20,"void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]",0,0,40,830681,830686,4194304,256,0,0,24,24,25088,0x7fc021203700,0x7fbf156355c0,12074463113139702,12074463113153453,12074463113277932,12074463113280814
21,"void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]",0,0,42,830681,830686,4194304,256,0,0,24,24,25792,0x7fc021203600,0x7fbf15635600,12074463113306091,12074463113317612,12074463113443050,12074463113445681
22,"void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]",0,0,44,830681,830686,4194304,256,0,0,24,24,26496,0x7fc021203500,0x7fbf15635640,12074463113480485,12074463113491530,12074463113615529,12074463113618432
23,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,46,830681,830686,4194304,256,0,0,28,24,27712,0x7fc021203400,0x7fbf15635680,12074463113643769,12074463113655048,12074463113904646,12074463113978571
24,"void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]",0,0,48,830681,830686,4194304,256,0,0,28,24,28416,0x7fc021203300,0x7fbf156356c0,12074463113999260,12074463114021285,12074463114267203,12074463114367734
25,"void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]",0,0,50,830681,830686,4194304,256,0,0,24,24,29120,0x7fc021203a00,0x7fbf15635700,12074463114387792,12074463114398721,12074463114522880,12074463114525558
26,"void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]",0,0,52,830681,830686,4194304,256,0,0,24,24,30080,0x7fc021203900,0x7fbf15635740,12074463114550514,12074463114561920,12074463114687359,12074463114690194
27,"void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]",0,0,54,830681,830686,4194304,256,0,0,24,24,30784,0x7fc021203800,0x7fbf15635780,12074463114725179,12074463114741118,12074463114865917,12074463114868325
28,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,56,830681,830686,4194304,256,0,0,28,24,32000,0x7fc021203700,0x7fbf156357c0,12074463114893431,12074463114904797,12074463115149594,12074463115230968
29,"void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]",0,0,58,830681,830686,4194304,256,0,0,28,24,32960,0x7fc021203600,0x7fbf15635800,12074463115251096,12074463115261593,12074463115505751,12074463115585257
30,"void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]",0,0,60,830681,830686,4194304,256,0,0,24,24,33920,0x7fc021203500,0x7fbf15635840,12074463115605414,12074463115621430,12074463115745748,12074463115748420
31,"void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]",0,0,62,830681,830686,4194304,256,0,0,24,24,34880,0x7fc021203400,0x7fbf15635880,12074463115774348,12074463115785748,12074463115911347,12074463115914429
32,"void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]",0,0,64,830681,830686,4194304,256,0,0,24,24,35840,0x7fc021203300,0x7fbf156358c0,12074463115949263,12074463115965586,12074463116090065,12074463116092900
33,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,66,830681,830686,4194304,256,0,0,28,24,37312,0x7fc021203a00,0x7fbf15635900,12074463116117316,12074463116129105,12074463116374382,12074463116456155
34,"void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]",0,0,68,830681,830686,4194304,256,0,0,28,24,38272,0x7fc021203900,0x7fbf15635940,12074463116476653,12074463116495821,12074463116777579,12074463116828928
35,"void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]",0,0,70,830681,830686,4194304,256,0,0,24,24,39232,0x7fc021203800,0x7fbf15635980,12074463116848825,12074463116860778,12074463116985737,12074463116988575
36,"void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]",0,0,72,830681,830686,4194304,256,0,0,24,24,40192,0x7fc021203700,0x7fbf156359c0,12074463117022878,12074463117034216,12074463117160135,12074463117163109
37,"void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]",0,0,74,830681,830686,4194304,256,0,0,24,24,41152,0x7fc021203600,0x7fbf15635a00,12074463117197814,12074463117210214,12074463117334533,12074463117337233
38,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,76,830681,830686,4194304,256,0,0,28,24,42624,0x7fc021203500,0x7fbf15635a40,12074463117361548,12074463117372933,12074463117620931,12074463117675541
39,"void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]",0,0,78,830681,830686,4194304,256,0,0,28,24,43584,0x7fc021203400,0x7fbf15635a80,12074463117696621,12074463117707810,12074463117951327,12074463117977412
40,"void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]",0,0,80,830681,830686,4194304,256,0,0,24,24,44544,0x7fc021203300,0x7fbf15635ac0,12074463117999403,12074463118019487,12074463118144285,12074463118146967
41,"void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]",0,0,82,830681,830686,4194304,256,0,0,24,24,45760,0x7fc021203a00,0x7fbf15635b00,12074463118172765,12074463118183965,12074463118309724,12074463118312104
42,"void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]",0,0,84,830681,830686,4194304,256,0,0,24,24,46720,0x7fc021203900,0x7fbf15635b40,12074463118346819,12074463118358363,12074463118483322,12074463118485627
43,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,86,830681,830686,4194304,256,0,0,28,24,48448,0x7fc021203800,0x7fbf15635b80,12074463118511545,12074463118523162,12074463118781879,12074463118807896
44,"void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]",0,0,88,830681,830686,4194304,256,0,0,28,24,49408,0x7fc021203700,0x7fbf15635bc0,12074463118828234,12074463118841399,12074463119085556,12074463119112953
45,"void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]",0,0,90,830681,830686,4194304,256,0,0,24,24,50368,0x7fc021203600,0x7fbf15635c00,12074463119135034,12074463119146196,12074463119270835,12074463119273441
46,"void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]",0,0,92,830681,830686,4194304,256,0,0,24,24,51584,0x7fc021203500,0x7fbf15635c40,12074463119299549,12074463119310674,12074463119436913,12074463119439269
47,"void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]",0,0,94,830681,830686,4194304,256,0,0,24,24,52544,0x7fc021203400,0x7fbf15635c80,12074463119475727,12074463119487153,12074463119611151,12074463119613443
48,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,96,830681,830686,4194304,256,0,0,28,24,54272,0x7fc021203300,0x7fbf15635cc0,12074463119638860,12074463119650511,12074463119907149,12074463119932425
49,"void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]",0,0,98,830681,830686,4194304,256,0,0,28,24,55488,0x7fc021203a00,0x7fbf15635d00,12074463119952543,12074463119965228,12074463120210186,12074463120237262
50,"void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]",0,0,100,830681,830686,4194304,256,0,0,24,24,56704,0x7fc021203900,0x7fbf15635d40,12074463120257520,12074463120269065,12074463120394344,12074463120396628
51,"void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]",0,0,102,830681,830686,4194304,256,0,0,24,24,58176,0x7fc021203800,0x7fbf15635d80,12074463120421044,12074463120432424,12074463120561382,12074463120563729
52,"void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]",0,0,104,830681,830686,4194304,256,0,0,24,24,59392,0x7fc021203700,0x7fbf15635dc0,12074463120598433,12074463120610022,12074463120735141,12074463120737732
53,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,106,830681,830686,4194304,256,0,0,28,24,61376,0x7fc021203600,0x7fbf15635e00,12074463120763239,12074463120774820,12074463121033378,12074463121061093
54,"void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]",0,0,108,830681,830686,4194304,256,0,0,28,24,62592,0x7fc021203500,0x7fbf15635e40,12074463121081601,12074463121092737,12074463121336415,12074463121364587
55,"void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]",0,0,110,830681,830686,4194304,256,0,0,24,24,63808,0x7fc021203400,0x7fbf15635e80,12074463121387279,12074463121398334,12074463121523133,12074463121525496
56,"void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]",0,0,112,830681,830686,4194304,256,0,0,24,24,65280,0x7fc021203300,0x7fbf15635ec0,12074463121550172,12074463121561373,12074463121689372,12074463121691745
57,"void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]",0,0,114,830681,830686,4194304,256,0,0,24,24,66496,0x7fc021203a00,0x7fbf15635f00,12074463121727461,12074463121739771,12074463121866010,12074463121868634
58,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,116,830681,830686,4194304,256,0,0,28,24,68480,0x7fc021203900,0x7fbf15635f40,12074463121894251,12074463121905369,12074463122156887,12074463122184421
59,"void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]",0,0,118,830681,830686,4194304,256,0,0,28,24,69696,0x7fc021203800,0x7fbf15635f80,12074463122205570,12074463122218806,12074463122463284,12074463122493826
60,"void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]",0,0,120,830681,830686,4194304,256,0,0,24,24,70912,0x7fc021203700,0x7fbf15635fc0,12074463122514895,12074463122526004,12074463122650322,12074463122652761
61,"void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]",0,0,122,830681,830686,4194304,256,0,0,24,24,72384,0x7fc021203600,0x7fbf15636000,12074463122677066,12074463122688402,12074463122825361,12074463122827786
62,"void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]",0,0,124,830681,830686,4194304,256,0,0,24,24,73600,0x7fc021203500,0x7fbf15636040,12074463122862841,12074463122874160,12074463122999119,12074463123006288
63,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,126,830681,830686,4194304,256,0,0,28,24,75840,0x7fc021203400,0x7fbf15636080,12074463123027427,12074463123039279,12074463123301516,12074463123328938
64,"void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]",0,0,128,830681,830686,4194304,256,0,0,28,24,77056,0x7fc021203300,0x7fbf156360c0,12074463123348995,12074463123360236,12074463123606473,12074463123635828
65,"void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]",0,0,130,830681,830686,4194304,256,0,0,24,24,78272,0x7fc021203a00,0x7fbf15636100,12074463123657909,12074463123669033,12074463123793511,12074463123795715
66,"void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]",0,0,132,830681,830686,4194304,256,0,0,24,24,80000,0x7fc021203900,0x7fbf15636140,12074463123820110,12074463123831111,12074463123973830,12074463123976160
67,"void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]",0,0,134,830681,830686,4194304,256,0,0,24,24,81216,0x7fc021203800,0x7fbf15636180,12074463124018429,12074463124029829,12074463124155108,12074463124157728
68,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,136,830681,830686,4194304,256,0,0,28,24,83456,0x7fc021203700,0x7fbf156361c0,12074463124182033,12074463124192868,12074463124449825,12074463124475037
69,"void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]",0,0,138,830681,830686,4194304,256,0,0,28,24,84928,0x7fc021203600,0x7fbf15636200,12074463124495666,12074463124506785,12074463124750622,12074463124774103
70,"void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]",0,0,140,830681,830686,4194304,256,0,0,24,24,86400,0x7fc021203500,0x7fbf15636240,12074463124796785,12074463124807902,12074463124932541,12074463124934912
71,"void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]",0,0,142,830681,830686,4194304,256,0,0,24,24,88128,0x7fc021203400,0x7fbf15636280,12074463124959127,12074463124970140,12074463125121179,12074463125123503
72,"void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]",0,0,144,830681,830686,4194304,256,0,0,24,24,89600,0x7fc021203300,0x7fbf156362c0,12074463125178635,12074463125191258,12074463125316857,12074463125319387
73,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,146,830681,830686,4194304,256,0,0,28,24,92096,0x7fc021203a00,0x7fbf15636300,12074463125344273,12074463125355896,12074463125608694,12074463125638329
74,"void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]",0,0,148,830681,830686,4194304,256,0,0,28,24,93568,0x7fc021203900,0x7fbf15636340,12074463125658397,12074463125669493,12074463125916051,12074463125939529
75,"void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]",0,0,150,830681,830686,4194304,256,0,0,24,24,95040,0x7fc021203800,0x7fbf15636380,12074463125961149,12074463125972371,12074463126098289,12074463126100689
76,"void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]",0,0,152,830681,830686,4194304,256,0,0,24,24,96768,0x7fc021203700,0x7fbf156363c0,12074463126125925,12074463126137169,12074463126294767,12074463126297304
77,"void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]",0,0,154,830681,830686,4194304,256,0,0,24,24,98240,0x7fc021203600,0x7fbf15636400,12074463126331678,12074463126342767,12074463126467886,12074463126470225
78,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,156,830681,830686,4194304,256,0,0,28,24,100736,0x7fc021203500,0x7fbf15636440,12074463126494851,12074463126506925,12074463126771723,12074463126796572
79,"void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]",0,0,158,830681,830686,4194304,256,0,0,28,24,102208,0x7fc021203400,0x7fbf15636480,12074463126816519,12074463126827562,12074463127072520,12074463127098373
80,"void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]",0,0,160,830681,830686,4194304,256,0,0,24,24,103680,0x7fc021203300,0x7fbf156364c0,12074463127121576,12074463127132679,12074463127258118,12074463127260484
81,"void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]",0,0,162,830681,830686,4194304,256,0,0,24,24,105664,0x7fc021203a00,0x7fbf15636500,12074463127285631,12074463127296518,12074463127462436,12074463127464713
82,"void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]",0,0,164,830681,830686,4194304,256,0,0,24,24,107136,0x7fc021203900,0x7fbf15636540,12074463127500099,12074463127511396,12074463127636995,12074463127639398
83,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,166,830681,830686,4194304,256,0,0,28,24,109888,0x7fc021203800,0x7fbf15636580,12074463127664244,12074463127675394,12074463127934112,12074463127953923
84,"void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]",0,0,168,830681,830686,4194304,256,0,0,28,24,111360,0x7fc021203700,0x7fbf156365c0,12074463127975292,12074463127986751,12074463128234109,12074463128260061
85,"void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]",0,0,170,830681,830686,4194304,256,0,0,24,24,112832,0x7fc021203600,0x7fbf15636600,12074463128281792,12074463128293148,12074463128418587,12074463128421011
86,"void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]",0,0,172,830681,830686,4194304,256,0,0,24,24,114816,0x7fc021203500,0x7fbf15636640,12074463128446578,12074463128457787,12074463128630265,12074463128632664
87,"void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]",0,0,174,830681,830686,4194304,256,0,0,24,24,116288,0x7fc021203400,0x7fbf15636680,12074463128667779,12074463128679545,12074463128805623,12074463128808050
88,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,176,830681,830686,4194304,256,0,0,28,24,119040,0x7fc021203300,0x7fbf156366c0,12074463128834088,12074463128845463,12074463129098101,12074463129124167
89,"void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]",0,0,178,830681,830686,4194304,256,0,0,28,24,120768,0x7fc021203a00,0x7fbf15636700,12074463129145156,12074463129156500,12074463129401618,12074463129425217
90,"void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]",0,0,180,830681,830686,4194304,256,0,0,24,24,122496,0x7fc021203900,0x7fbf15636740,12074463129449171,12074463129460017,12074463129585456,12074463129588019
91,"void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]",0,0,182,830681,830686,4194304,256,0,0,24,24,124736,0x7fc021203800,0x7fbf15636780,12074463129613577,12074463129624176,12074463129807054,12074463129809441
92,"void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]",0,0,184,830681,830686,4194304,256,0,0,24,24,126464,0x7fc021203700,0x7fbf156367c0,12074463129846359,12074463129857613,12074463129983052,12074463129985458
93,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,186,830681,830686,4194304,256,0,0,20,24,127936,0x7fc021203600,0x7fbf15636800,12074463130014772,12074463130025932,12074463130323369,12074463130351518
94,"void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]",0,0,188,830681,830686,4194304,256,0,0,28,24,129664,0x7fc021203500,0x7fbf15636840,12074463130372908,12074463130384648,12074463130630086,12074463130654060
95,"void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]",0,0,190,830681,830686,4194304,256,0,0,24,24,131392,0x7fc021203400,0x7fbf15636880,12074463130676001,12074463130687045,12074463130813124,12074463130815490
96,"void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]",0,0,192,830681,830686,4194304,256,0,0,24,24,133632,0x7fc021203300,0x7fbf156368c0,12074463130840667,12074463130851524,12074463131043362,12074463131046309
97,"void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]",0,0,194,830681,830686,4194304,256,0,0,24,24,135360,0x7fc021203a00,0x7fbf15636900,12074463131081966,12074463131092962,12074463131220320,12074463131223028
98,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,196,830681,830686,4194304,256,0,0,20,24,136832,0x7fc021203900,0x7fbf15636940,12074463131248966,12074463131260320,12074463131543197,12074463131568119
99,"void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]",0,0,198,830681,830686,4194304,256,0,0,28,24,138560,0x7fc021203800,0x7fbf15636980,12074463131589248,12074463131600157,12074463131847194,12074463131867636
100,"void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]",0,0,200,830681,830686,4194304,256,0,0,24,24,140288,0x7fc021203700,0x7fbf156369c0,12074463131894115,12074463131905594,12074463132032793,12074463132035718
101,"void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]",0,0,202,830681,830686,4194304,256,0,0,24,24,142784,0x7fc021203600,0x7fbf15636a00,12074463132061586,12074463132072472,12074463132282070,12074463132306501
102,"void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]",0,0,204,830681,830686,4194304,256,0,0,24,24,144768,0x7fc021203500,0x7fbf15636a40,12074463132335746,12074463132346710,12074463132474388,12074463132477118
103,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,206,830681,830686,4194304,256,0,0,20,24,145728,0x7fc021203400,0x7fbf15636a80,12074463132502445,12074463132513588,12074463132819665,12074463132844230
104,"void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]",0,0,208,830681,830686,4194304,256,0,0,28,24,147712,0x7fc021203300,0x7fbf15636ac0,12074463132864218,12074463132875184,12074463133123022,12074463133155880
105,"void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]",0,0,210,830681,830686,4194304,256,0,0,24,24,149696,0x7fc021203a00,0x7fbf15636b00,12074463133175897,12074463133187822,12074463133317100,12074463133319814
106,"void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]",0,0,212,830681,830686,4194304,256,0,0,24,24,152192,0x7fc021203900,0x7fbf15636b40,12074463133346594,12074463133357900,12074463133586538,12074463133614392
107,"void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]",0,0,214,830681,830686,4194304,256,0,0,24,24,154176,0x7fc021203800,0x7fbf15636b80,12074463133645990,12074463133656937,12074463133785096,12074463133787864
108,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,216,830681,830686,4194304,256,0,0,20,24,155648,0x7fc021203700,0x7fbf15636bc0,12074463133813231,12074463133824455,12074463134182212,12074463134216691
109,"void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]",0,0,218,830681,830686,4194304,256,0,0,28,24,157632,0x7fc021203600,0x7fbf15636c00,12074463134235857,12074463134247171,12074463134496929,12074463134526246
110,"void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]",0,0,220,830681,830686,4194304,256,0,0,24,24,159616,0x7fc021203500,0x7fbf15636c40,12074463134548478,12074463134560288,12074463134689247,12074463134691814
111,"void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]",0,0,222,830681,830686,4194304,256,0,0,24,24,162368,0x7fc021203400,0x7fbf15636c80,12074463134715849,12074463134726687,12074463134972284,12074463134995919
112,"void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]",0,0,224,830681,830686,4194304,256,0,0,24,24,164608,0x7fc021203300,0x7fbf15636cc0,12074463135036635,12074463135047484,12074463135176442,12074463135179510
113,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,226,830681,830686,4194304,256,0,0,20,24,166336,0x7fc021203a00,0x7fbf15636d00,12074463135206150,12074463135217722,12074463135593878,12074463135618807
114,"void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]",0,0,228,830681,830686,4194304,256,0,0,28,24,168576,0x7fc021203900,0x7fbf15636d40,12074463135641188,12074463135652598,12074463135908435,12074463135932279
115,"void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]",0,0,230,830681,830686,4194304,256,0,0,24,24,170816,0x7fc021203800,0x7fbf15636d80,12074463135957296,12074463135968275,12074463136099154,12074463136101914
116,"void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]",0,0,232,830681,830686,4194304,256,0,0,24,24,174080,0x7fc021203700,0x7fbf15636dc0,12074463136128023,12074463136138833,12074463136417871,12074463136442006
117,"void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]",0,0,234,830681,830686,4194304,256,0,0,24,24,176576,0x7fc021203600,0x7fbf15636e00,12074463136473535,12074463136484750,12074463136620429,12074463136623253
118,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,236,830681,830686,4194304,256,0,0,20,24,177792,0x7fc021203500,0x7fbf15636e40,12074463136648160,12074463136659148,12074463137082984,12074463137113424
119,"void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]",0,0,238,830681,830686,4194304,256,0,0,28,24,180288,0x7fc021203400,0x7fbf15636e80,12074463137134854,12074463137147464,12074463137402021,12074463137434781
120,"void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]",0,0,240,830681,830686,4194304,256,0,0,24,24,182784,0x7fc021203300,0x7fbf15636ec0,12074463137459187,12074463137470661,12074463137606499,12074463137609526
121,"void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]",0,0,242,830681,830686,4194304,256,0,0,24,24,186304,0x7fc021203a00,0x7fbf15636f00,12074463137634372,12074463137645059,12074463137959456,12074463137988060
122,"void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]",0,0,244,830681,830686,4194304,256,0,0,24,24,189312,0x7fc021203900,0x7fbf15636f40,12074463138031580,12074463138043935,12074463138194974,12074463138197799
123,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,246,830681,830686,4194304,256,0,0,20,24,190784,0x7fc021203800,0x7fbf15636f80,12074463138222335,12074463138233373,12074463138750968,12074463138778008
124,"void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]",0,0,248,830681,830686,4194304,256,0,0,28,24,193792,0x7fc021203700,0x7fbf15636fc0,12074463138805178,12074463138816248,12074463139114805,12074463139149308
125,"void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]",0,0,250,830681,830686,4194304,256,0,0,24,24,196800,0x7fc021203600,0x7fbf15637000,12074463139175166,12074463139188564,12074463139342163,12074463139345312
126,"void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]",0,0,252,830681,830686,4194304,256,0,0,24,24,201088,0x7fc021203500,0x7fbf15637040,12074463139370569,12074463139381362,12074463139766159,12074463139792162
127,"void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]",0,0,254,830681,830686,4194304,256,0,0,24,24,204608,0x7fc021203400,0x7fbf15637080,12074463139827628,12074463139838798,12074463140015116,12074463140017661
128,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,256,830681,830686,4194304,256,0,0,20,24,206080,0x7fc021203300,0x7fbf156370c0,12074463140044311,12074463140055916,12074463140667590,12074463140698116
129,"void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]",0,0,258,830681,830686,4194304,256,0,0,28,24,209600,0x7fc021203a00,0x7fbf15637100,12074463140718504,12074463140729509,12074463141056386,12074463141086377
130,"void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]",0,0,260,830681,830686,4194304,256,0,0,24,24,213120,0x7fc021203900,0x7fbf15637140,12074463141111103,12074463141122626,12074463141301184,12074463141304012
131,"void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]",0,0,262,830681,830686,4194304,256,0,0,24,24,218176,0x7fc021203800,0x7fbf15637180,12074463141327806,12074463141338143,12074463141794459,12074463141822295
132,"void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]",0,0,264,830681,830686,4194304,256,0,0,24,24,222208,0x7fc021203700,0x7fbf156371c0,12074463141852491,12074463141863418,12074463142063257,12074463142093659
133,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,266,830681,830686,4194304,256,0,0,20,24,223936,0x7fc021203600,0x7fbf15637200,12074463142113186,12074463142124216,12074463142828689,12074463142859342
134,"void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]",0,0,268,830681,830686,4194304,256,0,0,28,24,227968,0x7fc021203500,0x7fbf15637240,12074463142880802,12074463142891889,12074463143263245,12074463143297547
135,"void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]",0,0,270,830681,830686,4194304,256,0,0,24,24,232000,0x7fc021203400,0x7fbf15637280,12074463143317223,12074463143328524,12074463143531402,12074463143559193
136,"void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]",0,0,272,830681,830686,4194304,256,0,0,24,24,237824,0x7fc021203300,0x7fbf156372c0,12074463143579450,12074463143590122,12074463144116037,12074463144146504
137,"void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]",0,0,274,830681,830686,4194304,256,0,0,24,24,242368,0x7fc021203a00,0x7fbf15637300,12074463144179165,12074463144190596,12074463144414914,12074463144442715
138,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,276,830681,830686,4194304,256,0,0,20,24,243584,0x7fc021203900,0x7fbf15637340,12074463144465427,12074463144476193,12074463145275546,12074463145307171
139,"void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]",0,0,278,830681,830686,4194304,256,0,0,28,24,248128,0x7fc021203800,0x7fbf15637380,12074463145328300,12074463145339865,12074463145759221,12074463145787484
140,"void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]",0,0,280,830681,830686,4194304,256,0,0,24,24,252672,0x7fc021203700,0x7fbf156373c0,12074463145809966,12074463145820661,12074463146046738,12074463146079126
141,"void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]",0,0,282,830681,830686,4194304,256,0,0,24,24,259264,0x7fc021203600,0x7fbf15637400,12074463146102229,12074463146113458,12074463146710092,12074463146734694
142,"void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]",0,0,284,830681,830686,4194304,256,0,0,24,24,264832,0x7fc021203500,0x7fbf15637440,12074463146764800,12074463146775371,12074463147044969,12074463147076870
143,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,286,830681,830686,4194304,256,0,0,20,24,266304,0x7fc021203400,0x7fbf15637480,12074463147098791,12074463147110248,12074463148095839,12074463148129135
144,"void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]",0,0,288,830681,830686,4194304,256,0,0,28,24,271872,0x7fc021203300,0x7fbf156374c0,12074463148148782,12074463148159678,12074463148674393,12074463148703192
145,"void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]",0,0,290,830681,830686,4194304,256,0,0,24,24,277440,0x7fc021203a00,0x7fbf15637500,12074463148727477,12074463148738233,12074463149042710,12074463149074292
146,"void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]",0,0,292,830681,830686,4194304,256,0,0,12,24,279936,0x7fc021203900,0x7fbf15637540,12074463149096162,12074463149107029,12074463149844462,12074463149873206
147,"void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]",0,0,294,830681,830686,4194304,256,0,0,24,24,286528,0x7fc021203800,0x7fbf15637580,12074463149904184,12074463149915981,12074463150233258,12074463150265606
148,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,296,830681,830686,4194304,256,0,0,20,24,288000,0x7fc021203700,0x7fbf156375c0,12074463150288478,12074463150299658,12074463151473886,12074463151506151
149,"void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]",0,0,298,830681,830686,4194304,256,0,0,32,24,294592,0x7fc021203600,0x7fbf15637600,12074463151525577,12074463151536766,12074463152147000,12074463152179282
150,"void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]",0,0,300,830681,830686,4194304,256,0,0,24,24,301184,0x7fc021203500,0x7fbf15637640,12074463152202185,12074463152213399,12074463152533396,12074463152562675
151,"void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]",0,0,302,830681,830686,4194304,256,0,0,12,24,304960,0x7fc021203400,0x7fbf15637680,12074463152585277,12074463152595956,12074463153474187,12074463153507520
152,"void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]",0,0,304,830681,830686,4194304,256,0,0,24,24,313600,0x7fc021203300,0x7fbf156376c0,12074463153539430,12074463153550667,12074463153961063,12074463153991410
153,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,306,830681,830686,4194304,256,0,0,20,24,314816,0x7fc021203a00,0x7fbf15637700,12074463154020343,12074463154032742,12074463155585367,12074463155661744
154,"void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]",0,0,308,830681,830686,4194304,256,0,0,20,24,317568,0x7fc021203900,0x7fbf15637740,12074463155682703,12074463155695993,12074463156493266,12074463156572526
155,"void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]",0,0,310,830681,830686,4194304,256,0,0,12,24,320320,0x7fc021203800,0x7fbf15637780,12074463156592604,12074463156604624,12074463157019661,12074463157095037
156,"void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]",0,0,312,830681,830686,4194304,256,0,0,12,24,324096,0x7fc021203700,0x7fbf156377c0,12074463157118070,12074463157129260,12074463158289250,12074463158364386
157,"void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]",0,0,314,830681,830686,4194304,256,0,0,12,24,328896,0x7fc021203600,0x7fbf15637800,12074463158398349,12074463158410209,12074463159196283,12074463159271362
158,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,316,830681,830686,4194304,256,0,0,20,24,330112,0x7fc021203500,0x7fbf15637840,12074463159292872,12074463159304282,12074463162363936,12074463162439369
159,"void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]",0,0,318,830681,830686,4194304,256,0,0,20,24,332864,0x7fc021203400,0x7fbf15637880,12074463162460428,12074463162471775,12074463164018323,12074463164093633
160,"void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]",0,0,320,830681,830686,4194304,256,0,0,12,24,335616,0x7fc021203300,0x7fbf156378c0,12074463164117367,12074463164129042,12074463164919275,12074463164993495
161,"void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]",0,0,322,830681,830686,4194304,256,0,0,12,24,339392,0x7fc021203a00,0x7fbf15637900,12074463165021688,12074463165034794,12074463167317975,12074463167393065
162,"void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]",0,0,324,830681,830686,4194304,256,0,0,12,24,344192,0x7fc021203900,0x7fbf15637940,12074463167423702,12074463167436534,12074463168973321,12074463169049012
163,"void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]",0,0,326,830681,830686,4194304,256,0,0,20,24,345408,0x7fc021203800,0x7fbf15637980,12074463169070101,12074463169083401,12074463175147830,12074463175223497
164,"void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]",0,0,328,830681,830686,4194304,256,0,0,20,24,348160,0x7fc021203700,0x7fbf156379c0,12074463175244997,12074463175258069,12074463178307484,12074463178382828
165,"void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]",0,0,330,830681,830686,4194304,256,0,0,12,24,350912,0x7fc021203600,0x7fbf15637a00,12074463178403146,12074463178416443,12074463179956910,12074463180032784
166,"void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]",0,0,332,830681,830686,4194304,256,0,0,12,24,0,0x7fc021203500,0x7fbf15637a40,12074463180053172,12074463180064429,12074463184595432,12074463184673057