{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 26537,
  "matchedSampleCount": 25587,
  "profileSpanCycles": 222180,
  "profileSpanMs": 0.14812,
  "atOffset": null,
  "groups": [
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 10241,
      "tileCount": 1472,
      "phaseCount": 7,
      "phaseCycles": 128904,
      "phaseMs": 0.085936,
      "workCycles": 137619402,
      "averageActiveTiles": 1067.6115714019736,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 13438.082413826774,
      "p50Cycles": 9078,
      "p95Cycles": 44820,
      "maximumCycles": 76632,
      "firstOffset": 0,
      "lastOffset": 221556
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 63,
      "tileCount": 32,
      "phaseCount": 4,
      "phaseCycles": 113592,
      "phaseMs": 0.075728,
      "workCycles": 2088774,
      "averageActiveTiles": 18.388390027466723,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 33155.142857142855,
      "p50Cycles": 38232,
      "p95Cycles": 38526,
      "maximumCycles": 38550,
      "firstOffset": 10428,
      "lastOffset": 221046
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2898,
      "tileCount": 1458,
      "phaseCount": 15,
      "phaseCycles": 38766,
      "phaseMs": 0.025844,
      "workCycles": 43389834,
      "averageActiveTiles": 1119.2754991487386,
      "usefulCycles": 32912892.0,
      "estimatedWorkCycles": 43389834,
      "usefulUtilization": 0.758539246773795,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5767760829189182,
      "meanCycles": 14972.337474120082,
      "p50Cycles": 6966,
      "p95Cycles": 24798,
      "maximumCycles": 24798,
      "firstOffset": 91236,
      "lastOffset": 220830
    },
    {
      "name": "gemm_f8_init_rows_k32_c1088_r73",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 3,
      "phaseCycles": 35616,
      "phaseMs": 0.023744,
      "workCycles": 45357408,
      "averageActiveTiles": 1273.512129380054,
      "usefulCycles": 25449552.0,
      "estimatedWorkCycles": 45357408,
      "usefulUtilization": 0.5610892050974341,
      "workLaneOccupancy": 0.9889705882352942,
      "deviceUsefulUtilization": 0.4854306442634478,
      "meanCycles": 34998.0,
      "p50Cycles": 34998,
      "p95Cycles": 34998,
      "maximumCycles": 34998,
      "firstOffset": 13092,
      "lastOffset": 48708
    },
    {
      "name": "gemm_f8_init_rows_k32_c1088_r72",
      "dimensions": {},
      "sampleCount": 144,
      "tileCount": 144,
      "phaseCount": 3,
      "phaseCycles": 35454,
      "phaseMs": 0.023636,
      "workCycles": 5022432,
      "averageActiveTiles": 141.66051785412083,
      "usefulCycles": 2788992.0,
      "estimatedWorkCycles": 5022432,
      "usefulUtilization": 0.555307070359539,
      "workLaneOccupancy": 0.9889705882352942,
      "deviceUsefulUtilization": 0.05344095594781726,
      "meanCycles": 34878.0,
      "p50Cycles": 34878,
      "p95Cycles": 34878,
      "maximumCycles": 34878,
      "firstOffset": 12954,
      "lastOffset": 48408
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k160_c192_r81",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 10,
      "phaseCycles": 29382,
      "phaseMs": 0.019588,
      "workCycles": 42585264,
      "averageActiveTiles": 1449.3657341229325,
      "usefulCycles": 28238544.0,
      "estimatedWorkCycles": 42585264,
      "usefulUtilization": 0.6631059983566147,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.6529097229004448,
      "meanCycles": 29208.0,
      "p50Cycles": 29208,
      "p95Cycles": 29208,
      "maximumCycles": 29208,
      "firstOffset": 174312,
      "lastOffset": 203694
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 7,
      "phaseCycles": 13392,
      "phaseMs": 0.008928,
      "workCycles": 16898664,
      "averageActiveTiles": 1261.847670250896,
      "usefulCycles": 9412848.0,
      "estimatedWorkCycles": 16898664,
      "usefulUtilization": 0.5570172884673014,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4774938639551192,
      "meanCycles": 11480.070652173914,
      "p50Cycles": 11370,
      "p95Cycles": 11874,
      "maximumCycles": 12354,
      "firstOffset": 125100,
      "lastOffset": 138492
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r81_p81_c160_p160",
      "dimensions": {},
      "sampleCount": 1404,
      "tileCount": 1404,
      "phaseCount": 10,
      "phaseCycles": 13014,
      "phaseMs": 0.008676,
      "workCycles": 15188472,
      "averageActiveTiles": 1167.0871369294605,
      "usefulCycles": 9097920.0,
      "estimatedWorkCycles": 15188472,
      "usefulUtilization": 0.5990016638935108,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.47492332671838355,
      "meanCycles": 10818.0,
      "p50Cycles": 10818,
      "p95Cycles": 10818,
      "maximumCycles": 10818,
      "firstOffset": 145038,
      "lastOffset": 158052
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r81_p81_c144_p160",
      "dimensions": {},
      "sampleCount": 54,
      "tileCount": 54,
      "phaseCount": 8,
      "phaseCycles": 12462,
      "phaseMs": 0.008308,
      "workCycles": 629532,
      "averageActiveTiles": 50.516129032258064,
      "usefulCycles": 314928.0,
      "estimatedWorkCycles": 629532,
      "usefulUtilization": 0.5002573340195574,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.0171678424148542,
      "meanCycles": 11658.0,
      "p50Cycles": 11658,
      "p95Cycles": 11658,
      "maximumCycles": 11658,
      "firstOffset": 145992,
      "lastOffset": 158454
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 10,
      "phaseCycles": 10428,
      "phaseMs": 0.006952,
      "workCycles": 11394756,
      "averageActiveTiles": 1092.7077100115075,
      "usefulCycles": 2353212.0,
      "estimatedWorkCycles": 11394756,
      "usefulUtilization": 0.20651710313059796,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.15330355355981387,
      "meanCycles": 7815.333333333333,
      "p50Cycles": 7812,
      "p95Cycles": 7830,
      "maximumCycles": 7830,
      "firstOffset": 155856,
      "lastOffset": 166284
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 1084,
      "tileCount": 902,
      "phaseCount": 10,
      "phaseCycles": 8112,
      "phaseMs": 0.005408,
      "workCycles": 753420,
      "averageActiveTiles": 92.87721893491124,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 695.0369003690037,
      "p50Cycles": 330,
      "p95Cycles": 1578,
      "maximumCycles": 1578,
      "firstOffset": 91164,
      "lastOffset": 138822
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 2261,
      "tileCount": 1458,
      "phaseCount": 16,
      "phaseCycles": 3804,
      "phaseMs": 0.002536,
      "workCycles": 1274202,
      "averageActiveTiles": 334.9637223974763,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 563.5568332596197,
      "p50Cycles": 654,
      "p95Cycles": 654,
      "maximumCycles": 654,
      "firstOffset": 125400,
      "lastOffset": 222180
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 293,
      "tileCount": 261,
      "phaseCount": 15,
      "phaseCycles": 3378,
      "phaseMs": 0.002252,
      "workCycles": 66816,
      "averageActiveTiles": 19.7797513321492,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 228.04095563139933,
      "p50Cycles": 240,
      "p95Cycles": 288,
      "maximumCycles": 336,
      "firstOffset": 0,
      "lastOffset": 222156
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r73_p73_c32_p32",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 3,
      "phaseCycles": 3324,
      "phaseMs": 0.002216,
      "workCycles": 3506976,
      "averageActiveTiles": 1055.0469314079423,
      "usefulCycles": 756864.0,
      "estimatedWorkCycles": 3506976,
      "usefulUtilization": 0.21581670362158167,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.15468529273269502,
      "meanCycles": 2706.0,
      "p50Cycles": 2706,
      "p95Cycles": 2706,
      "maximumCycles": 2706,
      "firstOffset": 10386,
      "lastOffset": 13710
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r72_p72_c32_p32",
      "dimensions": {},
      "sampleCount": 144,
      "tileCount": 144,
      "phaseCount": 3,
      "phaseCycles": 3102,
      "phaseMs": 0.002068,
      "workCycles": 363744,
      "averageActiveTiles": 117.26112185686654,
      "usefulCycles": 82944.0,
      "estimatedWorkCycles": 363744,
      "usefulUtilization": 0.22802850356294538,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.018164998738541754,
      "meanCycles": 2526.0,
      "p50Cycles": 2526,
      "p95Cycles": 2526,
      "maximumCycles": 2526,
      "firstOffset": 10428,
      "lastOffset": 13530
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 21,
      "tileCount": 21,
      "phaseCount": 1,
      "phaseCycles": 378,
      "phaseMs": 0.000252,
      "workCycles": 4290,
      "averageActiveTiles": 11.34920634920635,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 204.28571428571428,
      "p50Cycles": 186,
      "p95Cycles": 372,
      "maximumCycles": 378,
      "firstOffset": 0,
      "lastOffset": 378
    }
  ],
  "samples": []
}
