{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 23687,
  "matchedSampleCount": 22887,
  "profileSpanCycles": 246186,
  "profileSpanMs": 0.164124,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 61,
      "tileCount": 44,
      "phaseCount": 3,
      "phaseCycles": 178242,
      "phaseMs": 0.118828,
      "workCycles": 4470390,
      "averageActiveTiles": 25.080452418621874,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 73285.08196721312,
      "p50Cycles": 78480,
      "p95Cycles": 79926,
      "maximumCycles": 79992,
      "firstOffset": 6750,
      "lastOffset": 245136
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 8771,
      "tileCount": 1472,
      "phaseCount": 6,
      "phaseCycles": 108060,
      "phaseMs": 0.07204,
      "workCycles": 116323020,
      "averageActiveTiles": 1076.4669627984454,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 13262.230076388098,
      "p50Cycles": 13458,
      "p95Cycles": 24330,
      "maximumCycles": 34968,
      "firstOffset": 0,
      "lastOffset": 245718
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k96_c256_r104",
      "dimensions": {},
      "sampleCount": 1224,
      "tileCount": 1224,
      "phaseCount": 3,
      "phaseCycles": 58380,
      "phaseMs": 0.03892,
      "workCycles": 67755744,
      "averageActiveTiles": 1160.5985611510791,
      "usefulCycles": 48342528.0,
      "estimatedWorkCycles": 67755744,
      "usefulUtilization": 0.713482358041851,
      "workLaneOccupancy": 0.9889705882352942,
      "deviceUsefulUtilization": 0.5625452433084588,
      "meanCycles": 55356.0,
      "p50Cycles": 55356,
      "p95Cycles": 55356,
      "maximumCycles": 55356,
      "firstOffset": 26532,
      "lastOffset": 84912
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k96_c256_r105",
      "dimensions": {},
      "sampleCount": 204,
      "tileCount": 204,
      "phaseCount": 3,
      "phaseCycles": 56454,
      "phaseMs": 0.037636,
      "workCycles": 11341584,
      "averageActiveTiles": 200.89956424699756,
      "usefulCycles": 8134560.0,
      "estimatedWorkCycles": 11341584,
      "usefulUtilization": 0.7172331483856224,
      "workLaneOccupancy": 0.9889705882352942,
      "deviceUsefulUtilization": 0.09788846941180276,
      "meanCycles": 55596.0,
      "p50Cycles": 55596,
      "p95Cycles": 55596,
      "maximumCycles": 55596,
      "firstOffset": 28698,
      "lastOffset": 85152
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k160_c128_r121",
      "dimensions": {},
      "sampleCount": 729,
      "tileCount": 729,
      "phaseCount": 13,
      "phaseCycles": 54204,
      "phaseMs": 0.036136,
      "workCycles": 37677636,
      "averageActiveTiles": 695.1080363072836,
      "usefulCycles": 28122336.0,
      "estimatedWorkCycles": 37677636,
      "usefulUtilization": 0.7463933246767392,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.352461955318555,
      "meanCycles": 51684.0,
      "p50Cycles": 51684,
      "p95Cycles": 51684,
      "maximumCycles": 51684,
      "firstOffset": 171066,
      "lastOffset": 225270
    },
    {
      "name": "gemm_f16_init_rows_interleaved_k160_c128_r122",
      "dimensions": {},
      "sampleCount": 729,
      "tileCount": 729,
      "phaseCount": 10,
      "phaseCycles": 53268,
      "phaseMs": 0.035512,
      "workCycles": 37747620,
      "averageActiveTiles": 708.6359540437036,
      "usefulCycles": 28354752.0,
      "estimatedWorkCycles": 37747620,
      "usefulUtilization": 0.7511666165972848,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.36161934239008003,
      "meanCycles": 51780.0,
      "p50Cycles": 51780,
      "p95Cycles": 51780,
      "maximumCycles": 51780,
      "firstOffset": 172014,
      "lastOffset": 225282
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2886,
      "tileCount": 1458,
      "phaseCount": 18,
      "phaseCycles": 29826,
      "phaseMs": 0.019884,
      "workCycles": 29163348,
      "averageActiveTiles": 977.7827398913699,
      "usefulCycles": 14087196.0,
      "estimatedWorkCycles": 29163348,
      "usefulUtilization": 0.48304453933066943,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.32086454698118655,
      "meanCycles": 10105.110187110187,
      "p50Cycles": 7698,
      "p95Cycles": 12708,
      "maximumCycles": 12708,
      "firstOffset": 98106,
      "lastOffset": 244902
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r121_p121_c144_p160",
      "dimensions": {},
      "sampleCount": 27,
      "tileCount": 27,
      "phaseCount": 6,
      "phaseCycles": 19890,
      "phaseMs": 0.01326,
      "workCycles": 468990,
      "averageActiveTiles": 23.579185520361992,
      "usefulCycles": 235224.0,
      "estimatedWorkCycles": 468990,
      "usefulUtilization": 0.5015544041450777,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.00803413338579579,
      "meanCycles": 17370.0,
      "p50Cycles": 17370,
      "p95Cycles": 17370,
      "maximumCycles": 17370,
      "firstOffset": 153696,
      "lastOffset": 173586
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r122_p122_c144_p160",
      "dimensions": {},
      "sampleCount": 27,
      "tileCount": 27,
      "phaseCount": 4,
      "phaseCycles": 17922,
      "phaseMs": 0.011948,
      "workCycles": 468990,
      "averageActiveTiles": 26.168396384332105,
      "usefulCycles": 237168.0,
      "estimatedWorkCycles": 468990,
      "usefulUtilization": 0.505699481865285,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.008990043813044933,
      "meanCycles": 17370.0,
      "p50Cycles": 17370,
      "p95Cycles": 17370,
      "maximumCycles": 17370,
      "firstOffset": 155580,
      "lastOffset": 173502
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r122_p122_c160_p160",
      "dimensions": {},
      "sampleCount": 702,
      "tileCount": 702,
      "phaseCount": 6,
      "phaseCycles": 16596,
      "phaseMs": 0.011064,
      "workCycles": 11309220,
      "averageActiveTiles": 681.4425162689805,
      "usefulCycles": 6851520.0,
      "estimatedWorkCycles": 11309220,
      "usefulUtilization": 0.6058348851644941,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.28046307648778646,
      "meanCycles": 16110.0,
      "p50Cycles": 16110,
      "p95Cycles": 16110,
      "maximumCycles": 16110,
      "firstOffset": 155904,
      "lastOffset": 172500
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r121_p121_c160_p160",
      "dimensions": {},
      "sampleCount": 702,
      "tileCount": 702,
      "phaseCount": 8,
      "phaseCycles": 16554,
      "phaseMs": 0.011036,
      "workCycles": 11309220,
      "averageActiveTiles": 683.1714389271475,
      "usefulCycles": 6795360.0,
      "estimatedWorkCycles": 11309220,
      "usefulUtilization": 0.6008690254500311,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.27886994342625715,
      "meanCycles": 16110.0,
      "p50Cycles": 16110,
      "p95Cycles": 16110,
      "maximumCycles": 16110,
      "firstOffset": 155880,
      "lastOffset": 172434
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 6,
      "phaseCycles": 13662,
      "phaseMs": 0.009108,
      "workCycles": 16898664,
      "averageActiveTiles": 1236.9099692577954,
      "usefulCycles": 9412848.0,
      "estimatedWorkCycles": 16898664,
      "usefulUtilization": 0.5570172884673014,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4680572263275477,
      "meanCycles": 11480.070652173914,
      "p50Cycles": 11370,
      "p95Cycles": 11874,
      "maximumCycles": 12354,
      "firstOffset": 119574,
      "lastOffset": 133236
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r104_p104_c96_p96",
      "dimensions": {},
      "sampleCount": 1224,
      "tileCount": 1224,
      "phaseCount": 3,
      "phaseCycles": 12114,
      "phaseMs": 0.008076,
      "workCycles": 11126160,
      "averageActiveTiles": 918.4546805349182,
      "usefulCycles": 6110208.0,
      "estimatedWorkCycles": 11126160,
      "usefulUtilization": 0.5491749174917492,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3426577944311648,
      "meanCycles": 9090.0,
      "p50Cycles": 9090,
      "p95Cycles": 9090,
      "maximumCycles": 9090,
      "firstOffset": 17442,
      "lastOffset": 29556
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 2688,
      "tileCount": 1450,
      "phaseCount": 16,
      "phaseCycles": 11418,
      "phaseMs": 0.007612,
      "workCycles": 3211524,
      "averageActiveTiles": 281.26852338413033,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1194.763392857143,
      "p50Cycles": 846,
      "p95Cycles": 1698,
      "maximumCycles": 1698,
      "firstOffset": 97596,
      "lastOffset": 246162
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r105_p105_c96_p96",
      "dimensions": {},
      "sampleCount": 204,
      "tileCount": 204,
      "phaseCount": 3,
      "phaseCycles": 9948,
      "phaseMs": 0.006632,
      "workCycles": 1854360,
      "averageActiveTiles": 186.4053075995175,
      "usefulCycles": 1028160.0,
      "estimatedWorkCycles": 1854360,
      "usefulUtilization": 0.5544554455445545,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07021293334032622,
      "meanCycles": 9090.0,
      "p50Cycles": 9090,
      "p95Cycles": 9090,
      "maximumCycles": 9090,
      "firstOffset": 19608,
      "lastOffset": 29556
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 441,
      "tileCount": 63,
      "phaseCount": 15,
      "phaseCycles": 5814,
      "phaseMs": 0.003876,
      "workCycles": 135630,
      "averageActiveTiles": 23.328173374613,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 129600,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 307.55102040816325,
      "p50Cycles": 300,
      "p95Cycles": 300,
      "maximumCycles": 1356,
      "firstOffset": 0,
      "lastOffset": 135018
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 761,
      "tileCount": 751,
      "phaseCount": 13,
      "phaseCycles": 5616,
      "phaseMs": 0.003744,
      "workCycles": 244836,
      "averageActiveTiles": 43.59615384615385,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 321.7293035479632,
      "p50Cycles": 330,
      "p95Cycles": 330,
      "maximumCycles": 330,
      "firstOffset": 0,
      "lastOffset": 135336
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 35,
      "tileCount": 33,
      "phaseCount": 11,
      "phaseCycles": 3156,
      "phaseMs": 0.002104,
      "workCycles": 7602,
      "averageActiveTiles": 2.408745247148289,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 217.2,
      "p50Cycles": 198,
      "p95Cycles": 288,
      "maximumCycles": 336,
      "firstOffset": 552,
      "lastOffset": 246186
    }
  ],
  "samples": []
}
