{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 26676,
  "matchedSampleCount": 25901,
  "profileSpanCycles": 200514,
  "profileSpanMs": 0.133676,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 75,
      "tileCount": 44,
      "phaseCount": 4,
      "phaseCycles": 152958,
      "phaseMs": 0.101972,
      "workCycles": 4277844,
      "averageActiveTiles": 27.967442042913742,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 57037.92,
      "p50Cycles": 77970,
      "p95Cycles": 78366,
      "maximumCycles": 78420,
      "firstOffset": 4932,
      "lastOffset": 199380
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 10229,
      "tileCount": 1472,
      "phaseCount": 7,
      "phaseCycles": 81042,
      "phaseMs": 0.054028,
      "workCycles": 84249396,
      "averageActiveTiles": 1039.5769600947656,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 8236.327695766937,
      "p50Cycles": 8460,
      "p95Cycles": 18180,
      "maximumCycles": 20856,
      "firstOffset": 0,
      "lastOffset": 199890
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k192_c256_r52",
      "dimensions": {},
      "sampleCount": 1326,
      "tileCount": 1326,
      "phaseCount": 3,
      "phaseCycles": 38148,
      "phaseMs": 0.025432,
      "workCycles": 46988136,
      "averageActiveTiles": 1231.7326203208556,
      "usefulCycles": 26185536.0,
      "estimatedWorkCycles": 46988136,
      "usefulUtilization": 0.557279735463437,
      "workLaneOccupancy": 0.9889705882352942,
      "deviceUsefulUtilization": 0.46631768261826934,
      "meanCycles": 35436.0,
      "p50Cycles": 35436,
      "p95Cycles": 35436,
      "maximumCycles": 35436,
      "firstOffset": 44934,
      "lastOffset": 83082
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k192_c256_r53",
      "dimensions": {},
      "sampleCount": 102,
      "tileCount": 102,
      "phaseCount": 3,
      "phaseCycles": 35856,
      "phaseMs": 0.023904,
      "workCycles": 3615084,
      "averageActiveTiles": 100.82228915662651,
      "usefulCycles": 2053008.0,
      "estimatedWorkCycles": 3615084,
      "usefulUtilization": 0.5679004969179139,
      "workLaneOccupancy": 0.9889705882352942,
      "deviceUsefulUtilization": 0.038897437576392524,
      "meanCycles": 35442.0,
      "p50Cycles": 35442,
      "p95Cycles": 35442,
      "maximumCycles": 35442,
      "firstOffset": 47256,
      "lastOffset": 83112
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r52_p52_c192_p192",
      "dimensions": {},
      "sampleCount": 1326,
      "tileCount": 1326,
      "phaseCount": 3,
      "phaseCycles": 31992,
      "phaseMs": 0.021328,
      "workCycles": 38825280,
      "averageActiveTiles": 1213.5933983495875,
      "usefulCycles": 3309696.0,
      "estimatedWorkCycles": 38825280,
      "usefulUtilization": 0.08524590163934426,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0702811572458332,
      "meanCycles": 29280.0,
      "p50Cycles": 29280,
      "p95Cycles": 29280,
      "maximumCycles": 29280,
      "firstOffset": 15654,
      "lastOffset": 47646
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r53_p53_c192_p192",
      "dimensions": {},
      "sampleCount": 102,
      "tileCount": 102,
      "phaseCount": 3,
      "phaseCycles": 29694,
      "phaseMs": 0.019796,
      "workCycles": 2986560,
      "averageActiveTiles": 100.57789452414629,
      "usefulCycles": 259488.0,
      "estimatedWorkCycles": 2986560,
      "usefulUtilization": 0.08688524590163935,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.005936640691575812,
      "meanCycles": 29280.0,
      "p50Cycles": 29280,
      "p95Cycles": 29280,
      "maximumCycles": 29280,
      "firstOffset": 17976,
      "lastOffset": 47670
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k160_c192_r81",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 9,
      "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": 152646,
      "lastOffset": 182028
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2886,
      "tileCount": 1458,
      "phaseCount": 13,
      "phaseCycles": 21366,
      "phaseMs": 0.014244,
      "workCycles": 23215356,
      "averageActiveTiles": 1086.5560235888795,
      "usefulCycles": 9380772.0,
      "estimatedWorkCycles": 23215356,
      "usefulUtilization": 0.4040761640700233,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.29826860737701916,
      "meanCycles": 8044.128898128898,
      "p50Cycles": 6966,
      "p95Cycles": 9252,
      "maximumCycles": 9252,
      "firstOffset": 89742,
      "lastOffset": 199164
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 6,
      "phaseCycles": 13074,
      "phaseMs": 0.008716,
      "workCycles": 16898664,
      "averageActiveTiles": 1292.5396971087655,
      "usefulCycles": 9412848.0,
      "estimatedWorkCycles": 16898664,
      "usefulUtilization": 0.5570172884673014,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4891079873096953,
      "meanCycles": 11480.070652173914,
      "p50Cycles": 11370,
      "p95Cycles": 11874,
      "maximumCycles": 12354,
      "firstOffset": 104100,
      "lastOffset": 117174
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r81_p81_c144_p160",
      "dimensions": {},
      "sampleCount": 54,
      "tileCount": 54,
      "phaseCount": 7,
      "phaseCycles": 12354,
      "phaseMs": 0.008236,
      "workCycles": 629532,
      "averageActiveTiles": 50.95774647887324,
      "usefulCycles": 314928.0,
      "estimatedWorkCycles": 629532,
      "usefulUtilization": 0.5002573340195574,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.01731792554427012,
      "meanCycles": 11658.0,
      "p50Cycles": 11658,
      "p95Cycles": 11658,
      "maximumCycles": 11658,
      "firstOffset": 124434,
      "lastOffset": 136788
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r81_p81_c160_p160",
      "dimensions": {},
      "sampleCount": 1404,
      "tileCount": 1404,
      "phaseCount": 9,
      "phaseCycles": 12036,
      "phaseMs": 0.008024,
      "workCycles": 15188472,
      "averageActiveTiles": 1261.9202392821535,
      "usefulCycles": 9097920.0,
      "estimatedWorkCycles": 15188472,
      "usefulUtilization": 0.5990016638935108,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5135138064068664,
      "meanCycles": 10818.0,
      "p50Cycles": 10818,
      "p95Cycles": 10818,
      "maximumCycles": 10818,
      "firstOffset": 124356,
      "lastOffset": 136392
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 9,
      "phaseCycles": 9444,
      "phaseMs": 0.006296,
      "workCycles": 11392182,
      "averageActiveTiles": 1206.2878017789074,
      "usefulCycles": 2353212.0,
      "estimatedWorkCycles": 11392182,
      "usefulUtilization": 0.20656376451850927,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.16927673194851114,
      "meanCycles": 7813.567901234568,
      "p50Cycles": 7812,
      "p95Cycles": 7830,
      "maximumCycles": 7830,
      "firstOffset": 135174,
      "lastOffset": 144618
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3067,
      "tileCount": 1458,
      "phaseCount": 18,
      "phaseCycles": 8052,
      "phaseMs": 0.005368,
      "workCycles": 3108186,
      "averageActiveTiles": 386.01415797317435,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1013.4287577437235,
      "p50Cycles": 654,
      "p95Cycles": 1698,
      "maximumCycles": 1698,
      "firstOffset": 89076,
      "lastOffset": 200514
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 166,
      "tileCount": 149,
      "phaseCount": 13,
      "phaseCycles": 4998,
      "phaseMs": 0.003332,
      "workCycles": 35892,
      "averageActiveTiles": 7.181272509003602,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 216.21686746987953,
      "p50Cycles": 198,
      "p95Cycles": 288,
      "maximumCycles": 1206,
      "firstOffset": 0,
      "lastOffset": 200490
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 756,
      "tileCount": 736,
      "phaseCount": 9,
      "phaseCycles": 3624,
      "phaseMs": 0.002416,
      "workCycles": 243360,
      "averageActiveTiles": 67.1523178807947,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 321.9047619047619,
      "p50Cycles": 330,
      "p95Cycles": 330,
      "maximumCycles": 330,
      "firstOffset": 0,
      "lastOffset": 117462
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 20,
      "tileCount": 20,
      "phaseCount": 1,
      "phaseCycles": 1374,
      "phaseMs": 0.000916,
      "workCycles": 12426,
      "averageActiveTiles": 9.043668122270743,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 621.3,
      "p50Cycles": 606,
      "p95Cycles": 1332,
      "maximumCycles": 1374,
      "firstOffset": 0,
      "lastOffset": 1374
    }
  ],
  "samples": []
}
