{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 28825,
  "matchedSampleCount": 27717,
  "profileSpanCycles": 608454,
  "profileSpanMs": 0.405636,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 138,
      "tileCount": 32,
      "phaseCount": 6,
      "phaseCycles": 597906,
      "phaseMs": 0.398604,
      "workCycles": 16190400,
      "averageActiveTiles": 27.078503978886314,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 117321.73913043478,
      "p50Cycles": 74856,
      "p95Cycles": 245724,
      "maximumCycles": 246528,
      "firstOffset": 7746,
      "lastOffset": 605652
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 10166,
      "tileCount": 1472,
      "phaseCount": 7,
      "phaseCycles": 236316,
      "phaseMs": 0.157544,
      "workCycles": 262488048,
      "averageActiveTiles": 1110.7502158127254,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 25820.189651780445,
      "p50Cycles": 16368,
      "p95Cycles": 52704,
      "maximumCycles": 57636,
      "firstOffset": 0,
      "lastOffset": 607320
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k384_c432_r61",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 5,
      "phaseCycles": 132714,
      "phaseMs": 0.088476,
      "workCycles": 140583600,
      "averageActiveTiles": 1059.2974365929745,
      "usefulCycles": 85064256.0,
      "estimatedWorkCycles": 140583600,
      "usefulUtilization": 0.6050795114081585,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.4354342223978205,
      "meanCycles": 130170.0,
      "p50Cycles": 130170,
      "p95Cycles": 130170,
      "maximumCycles": 130170,
      "firstOffset": 121302,
      "lastOffset": 254016
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k384_c432_r60",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 5,
      "phaseCycles": 132324,
      "phaseMs": 0.088216,
      "workCycles": 46742400,
      "averageActiveTiles": 353.2420422599075,
      "usefulCycles": 27889920.0,
      "estimatedWorkCycles": 46742400,
      "usefulUtilization": 0.5966728280961183,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.14318609263428503,
      "meanCycles": 129840.0,
      "p50Cycles": 129840,
      "p95Cycles": 129840,
      "maximumCycles": 129840,
      "firstOffset": 114978,
      "lastOffset": 247302
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k224_c384_r122",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 15,
      "phaseCycles": 110628,
      "phaseMs": 0.073752,
      "workCycles": 78153120,
      "averageActiveTiles": 706.4497233973316,
      "usefulCycles": 56709504.0,
      "estimatedWorkCycles": 78153120,
      "usefulUtilization": 0.7256204742689735,
      "workLaneOccupancy": 0.9607142857142857,
      "deviceUsefulUtilization": 0.3482434669420903,
      "meanCycles": 108546.0,
      "p50Cycles": 108546,
      "p95Cycles": 108546,
      "maximumCycles": 108546,
      "firstOffset": 438390,
      "lastOffset": 549018
    },
    {
      "name": "gemm_f8_init_rows_interleaved_k224_c384_r121",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 15,
      "phaseCycles": 110382,
      "phaseMs": 0.073588,
      "workCycles": 77997600,
      "averageActiveTiles": 706.6152090014676,
      "usefulCycles": 56244672.0,
      "estimatedWorkCycles": 77997600,
      "usefulUtilization": 0.7211077263915813,
      "workLaneOccupancy": 0.9607142857142857,
      "deviceUsefulUtilization": 0.34615875461736434,
      "meanCycles": 108330.0,
      "p50Cycles": 108330,
      "p95Cycles": 108330,
      "maximumCycles": 108330,
      "firstOffset": 438408,
      "lastOffset": 548790
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r61_p61_c384_p384",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 5,
      "phaseCycles": 73092,
      "phaseMs": 0.048728,
      "workCycles": 76191840,
      "averageActiveTiles": 1042.4101132818914,
      "usefulCycles": 6324480.0,
      "estimatedWorkCycles": 76191840,
      "usefulUtilization": 0.08300731416907638,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.05878238027595954,
      "meanCycles": 70548.0,
      "p50Cycles": 70548,
      "p95Cycles": 70548,
      "maximumCycles": 70548,
      "firstOffset": 50754,
      "lastOffset": 123846
    },
    {
      "name": "rearrange_row_major_to_amp_f8_o0_r60_p60_c384_p384",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 5,
      "phaseCycles": 66654,
      "phaseMs": 0.044436,
      "workCycles": 23101200,
      "averageActiveTiles": 346.58385093167703,
      "usefulCycles": 2073600.0,
      "estimatedWorkCycles": 23101200,
      "usefulUtilization": 0.08976157082748948,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.02113445032817105,
      "meanCycles": 64170.0,
      "p50Cycles": 64170,
      "p95Cycles": 64170,
      "maximumCycles": 64170,
      "firstOffset": 50808,
      "lastOffset": 117462
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 2880,
      "tileCount": 1440,
      "phaseCount": 22,
      "phaseCycles": 57630,
      "phaseMs": 0.03842,
      "workCycles": 54673560,
      "averageActiveTiles": 948.699635606455,
      "usefulCycles": 22231584.0,
      "estimatedWorkCycles": 54673560,
      "usefulUtilization": 0.406624042773143,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2620679899508861,
      "meanCycles": 18983.875,
      "p50Cycles": 18852,
      "p95Cycles": 18858,
      "maximumCycles": 25890,
      "firstOffset": 267300,
      "lastOffset": 605406
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 9,
      "phaseCycles": 44826,
      "phaseMs": 0.029884,
      "workCycles": 61454700,
      "averageActiveTiles": 1370.9610493909784,
      "usefulCycles": 37651392.0,
      "estimatedWorkCycles": 61454700,
      "usefulUtilization": 0.6126690391459074,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5706150739380678,
      "meanCycles": 42150.0,
      "p50Cycles": 42150,
      "p95Cycles": 42150,
      "maximumCycles": 42150,
      "firstOffset": 298068,
      "lastOffset": 342894
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r121_p121_c48_p224",
      "dimensions": {},
      "sampleCount": 36,
      "tileCount": 36,
      "phaseCount": 5,
      "phaseCycles": 35322,
      "phaseMs": 0.023548,
      "workCycles": 1233144,
      "averageActiveTiles": 34.9114999150671,
      "usefulCycles": 104544.0,
      "estimatedWorkCycles": 1233144,
      "usefulUtilization": 0.08477842003853564,
      "workLaneOccupancy": 0.21428571428571427,
      "deviceUsefulUtilization": 0.002010694160309008,
      "meanCycles": 34254.0,
      "p50Cycles": 34254,
      "p95Cycles": 34254,
      "maximumCycles": 34254,
      "firstOffset": 357294,
      "lastOffset": 392616
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r122_p122_c48_p224",
      "dimensions": {},
      "sampleCount": 36,
      "tileCount": 36,
      "phaseCount": 5,
      "phaseCycles": 34386,
      "phaseMs": 0.022924,
      "workCycles": 1233144,
      "averageActiveTiles": 35.86180422264875,
      "usefulCycles": 105408.0,
      "estimatedWorkCycles": 1233144,
      "usefulUtilization": 0.08547906813802768,
      "workLaneOccupancy": 0.21428571428571427,
      "deviceUsefulUtilization": 0.0020824956567258163,
      "meanCycles": 34254.0,
      "p50Cycles": 34254,
      "p95Cycles": 34254,
      "maximumCycles": 34254,
      "firstOffset": 358272,
      "lastOffset": 392658
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 10,
      "phaseCycles": 25962,
      "phaseMs": 0.017308,
      "workCycles": 17772480,
      "averageActiveTiles": 684.5574300901317,
      "usefulCycles": 4706424.0,
      "estimatedWorkCycles": 17772480,
      "usefulUtilization": 0.2648152649489548,
      "workLaneOccupancy": 0.9607142857142857,
      "deviceUsefulUtilization": 0.12315302800413983,
      "meanCycles": 12342.0,
      "p50Cycles": 12342,
      "p95Cycles": 12342,
      "maximumCycles": 12342,
      "firstOffset": 379038,
      "lastOffset": 405000
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r121_p121_c224_p224",
      "dimensions": {},
      "sampleCount": 684,
      "tileCount": 684,
      "phaseCount": 10,
      "phaseCycles": 22698,
      "phaseMs": 0.015132,
      "workCycles": 14811336,
      "averageActiveTiles": 652.5392545598731,
      "usefulCycles": 9269568.0,
      "estimatedWorkCycles": 14811336,
      "usefulUtilization": 0.6258428004063914,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2774368168810123,
      "meanCycles": 21654.0,
      "p50Cycles": 21654,
      "p95Cycles": 21654,
      "maximumCycles": 21654,
      "firstOffset": 357384,
      "lastOffset": 380082
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r122_p122_c224_p224",
      "dimensions": {},
      "sampleCount": 684,
      "tileCount": 684,
      "phaseCount": 10,
      "phaseCycles": 22644,
      "phaseMs": 0.015096,
      "workCycles": 14811336,
      "averageActiveTiles": 654.0953895071542,
      "usefulCycles": 9346176.0,
      "estimatedWorkCycles": 14811336,
      "usefulUtilization": 0.6310150549552046,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2803967650514965,
      "meanCycles": 21654.0,
      "p50Cycles": 21654,
      "p95Cycles": 21654,
      "maximumCycles": 21654,
      "firstOffset": 357480,
      "lastOffset": 380124
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 2586,
      "tileCount": 1440,
      "phaseCount": 24,
      "phaseCycles": 21870,
      "phaseMs": 0.01458,
      "workCycles": 3994992,
      "averageActiveTiles": 182.66995884773664,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1544.8538283062644,
      "p50Cycles": 1590,
      "p95Cycles": 6126,
      "maximumCycles": 6222,
      "firstOffset": 264162,
      "lastOffset": 608454
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 1752,
      "tileCount": 89,
      "phaseCount": 30,
      "phaseCycles": 15504,
      "phaseMs": 0.010336,
      "workCycles": 976212,
      "averageActiveTiles": 62.96517027863777,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 1002,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 557.1986301369863,
      "p50Cycles": 558,
      "p95Cycles": 558,
      "maximumCycles": 1002,
      "firstOffset": 0,
      "lastOffset": 418392
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 1445,
      "tileCount": 1253,
      "phaseCount": 15,
      "phaseCycles": 13428,
      "phaseMs": 0.008952,
      "workCycles": 1044426,
      "averageActiveTiles": 77.77971403038427,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 722.7861591695502,
      "p50Cycles": 642,
      "p95Cycles": 642,
      "maximumCycles": 4638,
      "firstOffset": 0,
      "lastOffset": 343536
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 92,
      "tileCount": 83,
      "phaseCount": 3,
      "phaseCycles": 2592,
      "phaseMs": 0.001728,
      "workCycles": 50130,
      "averageActiveTiles": 19.34027777777778,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 544.8913043478261,
      "p50Cycles": 186,
      "p95Cycles": 1872,
      "maximumCycles": 2148,
      "firstOffset": 0,
      "lastOffset": 2592
    }
  ],
  "samples": []
}
