{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 204096,
  "matchedSampleCount": 204096,
  "profileSpanCycles": 650028,
  "profileSpanMs": 0.433352,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 10710,
      "tileCount": 1471,
      "phaseCount": 42,
      "phaseCycles": 471474,
      "phaseMs": 0.314316,
      "workCycles": 58088178,
      "averageActiveTiles": 123.20547474516093,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5423.732773109244,
      "p50Cycles": 4098,
      "p95Cycles": 17772,
      "maximumCycles": 44718,
      "firstOffset": 3804,
      "lastOffset": 647424
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 67306,
      "tileCount": 1472,
      "phaseCount": 53,
      "phaseCycles": 433710,
      "phaseMs": 0.28914,
      "workCycles": 409901664,
      "averageActiveTiles": 945.1054022272947,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 6090.120702463376,
      "p50Cycles": 4932,
      "p95Cycles": 17142,
      "maximumCycles": 30012,
      "firstOffset": 3798,
      "lastOffset": 647832
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 30461,
      "tileCount": 1472,
      "phaseCount": 202,
      "phaseCycles": 88176,
      "phaseMs": 0.058784,
      "workCycles": 25143102,
      "averageActiveTiles": 285.1467746325531,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 825.4194543842947,
      "p50Cycles": 822,
      "p95Cycles": 1770,
      "maximumCycles": 2010,
      "firstOffset": 10920,
      "lastOffset": 648948
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 13653,
      "tileCount": 1472,
      "phaseCount": 165,
      "phaseCycles": 86952,
      "phaseMs": 0.057968,
      "workCycles": 100163970,
      "averageActiveTiles": 1151.9455561689208,
      "usefulCycles": 19384386.0,
      "estimatedWorkCycles": 94642602,
      "usefulUtilization": 0.2048167061171881,
      "workLaneOccupancy": 0.9923290559874584,
      "deviceUsefulUtilization": 0.15144839094132895,
      "meanCycles": 7336.407382992749,
      "p50Cycles": 8172,
      "p95Cycles": 16704,
      "maximumCycles": 16944,
      "firstOffset": 0,
      "lastOffset": 645594
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10681,
      "tileCount": 1472,
      "phaseCount": 169,
      "phaseCycles": 48498,
      "phaseMs": 0.032332,
      "workCycles": 32108652,
      "averageActiveTiles": 662.0613633551899,
      "usefulCycles": 13548784.0,
      "estimatedWorkCycles": 32108652,
      "usefulUtilization": 0.4219667645966576,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.1897879697414685,
      "meanCycles": 3006.1466154854415,
      "p50Cycles": 2730,
      "p95Cycles": 5574,
      "maximumCycles": 6888,
      "firstOffset": 11454,
      "lastOffset": 648678
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 16750,
      "tileCount": 1472,
      "phaseCount": 135,
      "phaseCycles": 45450,
      "phaseMs": 0.0303,
      "workCycles": 5804280,
      "averageActiveTiles": 127.70693069306931,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 2475264,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 346.52417910447764,
      "p50Cycles": 300,
      "p95Cycles": 600,
      "maximumCycles": 1500,
      "firstOffset": 110742,
      "lastOffset": 619392
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 2188,
      "tileCount": 729,
      "phaseCount": 57,
      "phaseCycles": 41622,
      "phaseMs": 0.027748,
      "workCycles": 20519064,
      "averageActiveTiles": 492.98601701023495,
      "usefulCycles": 7246656.0,
      "estimatedWorkCycles": 20519064,
      "usefulUtilization": 0.3531669865642994,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.1182787948680986,
      "meanCycles": 9378.0,
      "p50Cycles": 9378,
      "p95Cycles": 9378,
      "maximumCycles": 9378,
      "firstOffset": 23538,
      "lastOffset": 630618
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3423,
      "tileCount": 1472,
      "phaseCount": 98,
      "phaseCycles": 36816,
      "phaseMs": 0.024544,
      "workCycles": 1452150,
      "averageActiveTiles": 39.44344850065189,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 424.23312883435585,
      "p50Cycles": 414,
      "p95Cycles": 714,
      "maximumCycles": 714,
      "firstOffset": 97098,
      "lastOffset": 585738
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c80_r80_r82",
      "dimensions": {},
      "sampleCount": 795,
      "tileCount": 795,
      "phaseCount": 24,
      "phaseCycles": 36654,
      "phaseMs": 0.024436,
      "workCycles": 23530410,
      "averageActiveTiles": 641.9602226223604,
      "usefulCycles": 15645600.0,
      "estimatedWorkCycles": 23530410,
      "usefulUtilization": 0.6649097912021082,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.28997665596731836,
      "meanCycles": 29598.0,
      "p50Cycles": 29598,
      "p95Cycles": 29598,
      "maximumCycles": 29598,
      "firstOffset": 314220,
      "lastOffset": 350874
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c80_r80_r82",
      "dimensions": {},
      "sampleCount": 636,
      "tileCount": 636,
      "phaseCount": 19,
      "phaseCycles": 36114,
      "phaseMs": 0.024076,
      "workCycles": 18492336,
      "averageActiveTiles": 512.0544941020103,
      "usefulCycles": 12173040.0,
      "estimatedWorkCycles": 18492336,
      "usefulUtilization": 0.6582748658687577,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.2289895403685431,
      "meanCycles": 29076.0,
      "p50Cycles": 29076,
      "p95Cycles": 29076,
      "maximumCycles": 29076,
      "firstOffset": 314214,
      "lastOffset": 350328
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 31,
      "phaseCycles": 32334,
      "phaseMs": 0.021556,
      "workCycles": 34984800,
      "averageActiveTiles": 1081.981814807942,
      "usefulCycles": 15838080.0,
      "estimatedWorkCycles": 34984800,
      "usefulUtilization": 0.45271317829457364,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.33276319717298525,
      "meanCycles": 23511.290322580644,
      "p50Cycles": 23700,
      "p95Cycles": 23700,
      "maximumCycles": 23700,
      "firstOffset": 157398,
      "lastOffset": 594360
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 33,
      "phaseCycles": 29208,
      "phaseMs": 0.019472,
      "workCycles": 31995810,
      "averageActiveTiles": 1095.4467953985209,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 31995810,
      "usefulUtilization": 0.738501697566025,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5495851345075202,
      "meanCycles": 26334.0,
      "p50Cycles": 26334,
      "p95Cycles": 26334,
      "maximumCycles": 26334,
      "firstOffset": 419286,
      "lastOffset": 448494
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 24,
      "phaseCycles": 29064,
      "phaseMs": 0.019376,
      "workCycles": 6367086,
      "averageActiveTiles": 219.0712221304707,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6367086,
      "usefulUtilization": 0.7239707458011404,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.10774535059060066,
      "meanCycles": 26202.0,
      "p50Cycles": 26202,
      "p95Cycles": 26202,
      "maximumCycles": 26202,
      "firstOffset": 419298,
      "lastOffset": 448362
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c64_r80_r82",
      "dimensions": {},
      "sampleCount": 12,
      "tileCount": 12,
      "phaseCount": 8,
      "phaseCycles": 28284,
      "phaseMs": 0.018856,
      "workCycles": 279864,
      "averageActiveTiles": 9.894781501909206,
      "usefulCycles": 183744.0,
      "estimatedWorkCycles": 279864,
      "usefulUtilization": 0.6565474659120144,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.004413310951651879,
      "meanCycles": 23322.0,
      "p50Cycles": 23322,
      "p95Cycles": 23322,
      "maximumCycles": 23322,
      "firstOffset": 312720,
      "lastOffset": 341004
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c64_r80_r82",
      "dimensions": {},
      "sampleCount": 15,
      "tileCount": 15,
      "phaseCount": 8,
      "phaseCycles": 27342,
      "phaseMs": 0.018228,
      "workCycles": 356130,
      "averageActiveTiles": 13.025016458196182,
      "usefulCycles": 236160.0,
      "estimatedWorkCycles": 356130,
      "usefulUtilization": 0.6631286328026282,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.005867704725649026,
      "meanCycles": 23742.0,
      "p50Cycles": 23742,
      "p95Cycles": 23742,
      "maximumCycles": 23742,
      "firstOffset": 312726,
      "lastOffset": 340068
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 13410,
      "tileCount": 1458,
      "phaseCount": 149,
      "phaseCycles": 26214,
      "phaseMs": 0.017476,
      "workCycles": 18998712,
      "averageActiveTiles": 724.7544060425727,
      "usefulCycles": 3095028.0,
      "estimatedWorkCycles": 18998712,
      "usefulUtilization": 0.1629072539233186,
      "workLaneOccupancy": 0.9996279286088018,
      "deviceUsefulUtilization": 0.08020906933235145,
      "meanCycles": 1416.7570469798657,
      "p50Cycles": 1068,
      "p95Cycles": 2652,
      "maximumCycles": 2652,
      "firstOffset": 18708,
      "lastOffset": 650028
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c96_r72_r74",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 3,
      "phaseCycles": 25056,
      "phaseMs": 0.016704,
      "workCycles": 17958240,
      "averageActiveTiles": 716.7241379310345,
      "usefulCycles": 11508480.0,
      "estimatedWorkCycles": 17958240,
      "usefulUtilization": 0.6408467644936252,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3120314842578711,
      "meanCycles": 24942.0,
      "p50Cycles": 24942,
      "p95Cycles": 24942,
      "maximumCycles": 24942,
      "firstOffset": 60516,
      "lastOffset": 85572
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c96_r72_r74",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 2,
      "phaseCycles": 24852,
      "phaseMs": 0.016568,
      "workCycles": 17845920,
      "averageActiveTiles": 718.0878802510864,
      "usefulCycles": 11166336.0,
      "estimatedWorkCycles": 17845920,
      "usefulUtilization": 0.6257080610021787,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.3052400646610543,
      "meanCycles": 24786.0,
      "p50Cycles": 24786,
      "p95Cycles": 24786,
      "maximumCycles": 24786,
      "firstOffset": 60582,
      "lastOffset": 85434
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 48,
      "phaseCycles": 22746,
      "phaseMs": 0.015164,
      "workCycles": 4276800,
      "averageActiveTiles": 188.0242680031654,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4276800,
      "usefulUtilization": 0.21069958847736625,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.02691347547012566,
      "meanCycles": 2430.0,
      "p50Cycles": 2430,
      "p95Cycles": 2430,
      "maximumCycles": 2430,
      "firstOffset": 118032,
      "lastOffset": 576426
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 896,
      "tileCount": 896,
      "phaseCount": 11,
      "phaseCycles": 18876,
      "phaseMs": 0.012584,
      "workCycles": 14477568,
      "averageActiveTiles": 766.9828353464717,
      "usefulCycles": 8358714.0,
      "estimatedWorkCycles": 14477568,
      "usefulUtilization": 0.5773562244708503,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.3008303763923823,
      "meanCycles": 16158.0,
      "p50Cycles": 16158,
      "p95Cycles": 16158,
      "maximumCycles": 16158,
      "firstOffset": 192834,
      "lastOffset": 211710
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 128,
      "tileCount": 128,
      "phaseCount": 9,
      "phaseCycles": 18246,
      "phaseMs": 0.012164,
      "workCycles": 2072832,
      "averageActiveTiles": 113.60473528444591,
      "usefulCycles": 1207224.0,
      "estimatedWorkCycles": 2072832,
      "usefulUtilization": 0.5824032048907003,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.044948207826372905,
      "meanCycles": 16194.0,
      "p50Cycles": 16194,
      "p95Cycles": 16194,
      "maximumCycles": 16194,
      "firstOffset": 193518,
      "lastOffset": 211764
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 224,
      "tileCount": 224,
      "phaseCount": 9,
      "phaseCycles": 18108,
      "phaseMs": 0.012072,
      "workCycles": 3033408,
      "averageActiveTiles": 167.51756129887343,
      "usefulCycles": 1834560.0,
      "estimatedWorkCycles": 3033408,
      "usefulUtilization": 0.6047851129818342,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.06882617339441612,
      "meanCycles": 13542.0,
      "p50Cycles": 13542,
      "p95Cycles": 13542,
      "maximumCycles": 13542,
      "firstOffset": 134592,
      "lastOffset": 152700
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 1008,
      "tileCount": 1008,
      "phaseCount": 8,
      "phaseCycles": 17238,
      "phaseMs": 0.011492,
      "workCycles": 10983168,
      "averageActiveTiles": 637.1486251305256,
      "usefulCycles": 6524154.0,
      "estimatedWorkCycles": 10983168,
      "usefulUtilization": 0.5940138582966226,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.25711624532756244,
      "meanCycles": 10896.0,
      "p50Cycles": 10896,
      "p95Cycles": 10896,
      "maximumCycles": 10896,
      "firstOffset": 131952,
      "lastOffset": 149190
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c64_r72_r74",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 29,
      "phaseCycles": 16800,
      "phaseMs": 0.0112,
      "workCycles": 12048480,
      "averageActiveTiles": 717.1714285714286,
      "usefulCycles": 7672320.0,
      "estimatedWorkCycles": 12048480,
      "usefulUtilization": 0.6367873789888849,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.31024844720496897,
      "meanCycles": 16734.0,
      "p50Cycles": 16734,
      "p95Cycles": 16734,
      "maximumCycles": 16734,
      "firstOffset": 515808,
      "lastOffset": 532608
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c64_r72_r74",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 36,
      "phaseCycles": 16686,
      "phaseMs": 0.011124,
      "workCycles": 11966400,
      "averageActiveTiles": 717.1521035598705,
      "usefulCycles": 7444224.0,
      "estimatedWorkCycles": 11966400,
      "usefulUtilization": 0.6220938628158845,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.30308146897425076,
      "meanCycles": 16620.0,
      "p50Cycles": 16620,
      "p95Cycles": 16620,
      "maximumCycles": 16620,
      "firstOffset": 515814,
      "lastOffset": 532500
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 35,
      "phaseCycles": 15630,
      "phaseMs": 0.01042,
      "workCycles": 313920,
      "averageActiveTiles": 20.084452975047984,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 313920,
      "usefulUtilization": 0.1019367991845056,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0013908592728587722,
      "meanCycles": 1962.0,
      "p50Cycles": 1962,
      "p95Cycles": 1962,
      "maximumCycles": 1962,
      "firstOffset": 119610,
      "lastOffset": 575496
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 176,
      "tileCount": 176,
      "phaseCount": 26,
      "phaseCycles": 15516,
      "phaseMs": 0.010344,
      "workCycles": 2072928,
      "averageActiveTiles": 133.59938128383604,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 2072928,
      "usefulUtilization": 0.21735438953982,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.019727182039297443,
      "meanCycles": 11778.0,
      "p50Cycles": 11778,
      "p95Cycles": 11778,
      "maximumCycles": 11778,
      "firstOffset": 558480,
      "lastOffset": 573996
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 40,
      "phaseCycles": 15396,
      "phaseMs": 0.010264,
      "workCycles": 18503076,
      "averageActiveTiles": 1201.8106001558847,
      "usefulCycles": 9412848.0,
      "estimatedWorkCycles": 18503076,
      "usefulUtilization": 0.508718009913595,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.41534150598122604,
      "meanCycles": 12570.02445652174,
      "p50Cycles": 12474,
      "p95Cycles": 12918,
      "maximumCycles": 13404,
      "firstOffset": 375540,
      "lastOffset": 390936
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 2432,
      "tileCount": 1243,
      "phaseCount": 124,
      "phaseCycles": 14934,
      "phaseMs": 0.009956,
      "workCycles": 541158,
      "averageActiveTiles": 36.23664122137404,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 222.515625,
      "p50Cycles": 198,
      "p95Cycles": 288,
      "maximumCycles": 1068,
      "firstOffset": 92868,
      "lastOffset": 640824
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 7,
      "phaseCycles": 14064,
      "phaseMs": 0.009376,
      "workCycles": 434304,
      "averageActiveTiles": 30.880546075085324,
      "usefulCycles": 264960.0,
      "estimatedWorkCycles": 434304,
      "usefulUtilization": 0.610079575596817,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.012798634812286689,
      "meanCycles": 13572.0,
      "p50Cycles": 13572,
      "p95Cycles": 13572,
      "maximumCycles": 13572,
      "firstOffset": 138684,
      "lastOffset": 152748
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 8,
      "phaseCycles": 12534,
      "phaseMs": 0.008356,
      "workCycles": 146592,
      "averageActiveTiles": 11.69554810914313,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 146592,
      "usefulUtilization": 0.10914647456887143,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0008672064159399477,
      "meanCycles": 9162.0,
      "p50Cycles": 9162,
      "p95Cycles": 9162,
      "maximumCycles": 9162,
      "firstOffset": 557706,
      "lastOffset": 570240
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 144,
      "tileCount": 144,
      "phaseCount": 6,
      "phaseCycles": 11700,
      "phaseMs": 0.0078,
      "workCycles": 1572480,
      "averageActiveTiles": 134.4,
      "usefulCycles": 942264.0,
      "estimatedWorkCycles": 1572480,
      "usefulUtilization": 0.5992216117216117,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.054711538461538464,
      "meanCycles": 10920.0,
      "p50Cycles": 10920,
      "p95Cycles": 10920,
      "maximumCycles": 10920,
      "firstOffset": 138402,
      "lastOffset": 150102
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 43,
      "phaseCycles": 9372,
      "phaseMs": 0.006248,
      "workCycles": 4397280,
      "averageActiveTiles": 469.1933418693982,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2955.1612903225805,
      "p50Cycles": 2970,
      "p95Cycles": 2970,
      "maximumCycles": 2970,
      "firstOffset": 219474,
      "lastOffset": 608598
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c48_r72_r74",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 23,
      "phaseCycles": 8664,
      "phaseMs": 0.005776,
      "workCycles": 6160320,
      "averageActiveTiles": 711.0249307479224,
      "usefulCycles": 3836160.0,
      "estimatedWorkCycles": 6160320,
      "usefulUtilization": 0.6227208976157083,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.30079489341201976,
      "meanCycles": 8556.0,
      "p50Cycles": 8556,
      "p95Cycles": 8556,
      "maximumCycles": 8556,
      "firstOffset": 246852,
      "lastOffset": 255516
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c48_r72_r74",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 18,
      "phaseCycles": 8538,
      "phaseMs": 0.005692,
      "workCycles": 6104160,
      "averageActiveTiles": 714.9402670414617,
      "usefulCycles": 3722112.0,
      "estimatedWorkCycles": 6104160,
      "usefulUtilization": 0.6097664543524416,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.29615936936661674,
      "meanCycles": 8478.0,
      "p50Cycles": 8478,
      "p95Cycles": 8478,
      "maximumCycles": 8478,
      "firstOffset": 246912,
      "lastOffset": 255450
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 33,
      "phaseCycles": 8400,
      "phaseMs": 0.0056,
      "workCycles": 856704,
      "averageActiveTiles": 101.98857142857143,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 856704,
      "usefulUtilization": 0.015316842223218288,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.001061238354037267,
      "meanCycles": 1164.0,
      "p50Cycles": 1164,
      "p95Cycles": 1164,
      "maximumCycles": 1164,
      "firstOffset": 578502,
      "lastOffset": 586902
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 640,
      "tileCount": 640,
      "phaseCount": 37,
      "phaseCycles": 6570,
      "phaseMs": 0.00438,
      "workCycles": 852480,
      "averageActiveTiles": 129.75342465753425,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 852480,
      "usefulUtilization": 0.015392736486486487,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.0013568344252531269,
      "meanCycles": 1332.0,
      "p50Cycles": 1332,
      "p95Cycles": 1332,
      "maximumCycles": 1332,
      "firstOffset": 596178,
      "lastOffset": 602748
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c16_r144_r144",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 1,
      "phaseCycles": 5592,
      "phaseMs": 0.003728,
      "workCycles": 5899776,
      "averageActiveTiles": 1055.038626609442,
      "usefulCycles": 3857868.0,
      "estimatedWorkCycles": 5899776,
      "usefulUtilization": 0.6539007582660765,
      "workLaneOccupancy": 0.9100118885869565,
      "deviceUsefulUtilization": 0.4686756507743982,
      "meanCycles": 4008.0,
      "p50Cycles": 4008,
      "p95Cycles": 4008,
      "maximumCycles": 4008,
      "firstOffset": 4140,
      "lastOffset": 9732
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 880,
      "tileCount": 880,
      "phaseCount": 8,
      "phaseCycles": 4650,
      "phaseMs": 0.0031,
      "workCycles": 1351680,
      "averageActiveTiles": 290.68387096774194,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 1351680,
      "usefulUtilization": 0.3333333333333333,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.06582515194015895,
      "meanCycles": 1536.0,
      "p50Cycles": 1536,
      "p95Cycles": 1536,
      "maximumCycles": 1536,
      "firstOffset": 116496,
      "lastOffset": 121146
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80",
      "dimensions": {},
      "sampleCount": 1360,
      "tileCount": 1360,
      "phaseCount": 8,
      "phaseCycles": 4380,
      "phaseMs": 0.00292,
      "workCycles": 1721760,
      "averageActiveTiles": 393.09589041095893,
      "usefulCycles": 391680.0,
      "estimatedWorkCycles": 1721760,
      "usefulUtilization": 0.22748815165876776,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.06075044669446099,
      "meanCycles": 1266.0,
      "p50Cycles": 1266,
      "p95Cycles": 1266,
      "maximumCycles": 1266,
      "firstOffset": 115230,
      "lastOffset": 119610
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r9_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 1,
      "phaseCycles": 4182,
      "phaseMs": 0.002788,
      "workCycles": 132096,
      "averageActiveTiles": 31.58680057388809,
      "usefulCycles": 5184.0,
      "estimatedWorkCycles": 132096,
      "usefulUtilization": 0.03924418604651163,
      "workLaneOccupancy": 0.5625,
      "deviceUsefulUtilization": 0.0008421183956085085,
      "meanCycles": 4128.0,
      "p50Cycles": 4128,
      "p95Cycles": 4128,
      "maximumCycles": 4128,
      "firstOffset": 13512,
      "lastOffset": 17694
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 5,
      "phaseCycles": 4056,
      "phaseMs": 0.002704,
      "workCycles": 141600,
      "averageActiveTiles": 34.9112426035503,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 141600,
      "usefulUtilization": 0.11299435028248588,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0026798730812108737,
      "meanCycles": 1770.0,
      "p50Cycles": 1770,
      "p95Cycles": 1770,
      "maximumCycles": 1770,
      "firstOffset": 116952,
      "lastOffset": 121008
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r16_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 1,
      "phaseCycles": 3768,
      "phaseMs": 0.002512,
      "workCycles": 5322240,
      "averageActiveTiles": 1412.4840764331211,
      "usefulCycles": 414720.0,
      "estimatedWorkCycles": 5322240,
      "usefulUtilization": 0.07792207792207792,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07477153143173636,
      "meanCycles": 3696.0,
      "p50Cycles": 3696,
      "p95Cycles": 3696,
      "maximumCycles": 3696,
      "firstOffset": 13494,
      "lastOffset": 17262
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 6,
      "phaseCycles": 3618,
      "phaseMs": 0.002412,
      "workCycles": 141792,
      "averageActiveTiles": 39.19071310116086,
      "usefulCycles": 28224.0,
      "estimatedWorkCycles": 141792,
      "usefulUtilization": 0.1990521327014218,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.0052995890114644175,
      "meanCycles": 1266.0,
      "p50Cycles": 1266,
      "p95Cycles": 1266,
      "maximumCycles": 1266,
      "firstOffset": 113172,
      "lastOffset": 118566
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c256_p256",
      "dimensions": {},
      "sampleCount": 1076,
      "tileCount": 1076,
      "phaseCount": 60,
      "phaseCycles": 2670,
      "phaseMs": 0.00178,
      "workCycles": 2433912,
      "averageActiveTiles": 911.5775280898877,
      "usefulCycles": 137728.0,
      "estimatedWorkCycles": 2433912,
      "usefulUtilization": 0.05658709106984969,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.035043152581012864,
      "meanCycles": 2262.0,
      "p50Cycles": 2262,
      "p95Cycles": 2262,
      "maximumCycles": 2262,
      "firstOffset": 631440,
      "lastOffset": 634110
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c128_p256",
      "dimensions": {},
      "sampleCount": 269,
      "tileCount": 269,
      "phaseCount": 25,
      "phaseCycles": 2298,
      "phaseMs": 0.001532,
      "workCycles": 608478,
      "averageActiveTiles": 264.78590078328983,
      "usefulCycles": 17216.0,
      "estimatedWorkCycles": 608478,
      "usefulUtilization": 0.028293545534924844,
      "workLaneOccupancy": 0.25,
      "deviceUsefulUtilization": 0.005089491807620993,
      "meanCycles": 2262.0,
      "p50Cycles": 2262,
      "p95Cycles": 2262,
      "maximumCycles": 2262,
      "firstOffset": 631896,
      "lastOffset": 634194
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c224_p224",
      "dimensions": {},
      "sampleCount": 1368,
      "tileCount": 1368,
      "phaseCount": 68,
      "phaseCycles": 2232,
      "phaseMs": 0.001488,
      "workCycles": 2766096,
      "averageActiveTiles": 1239.2903225806451,
      "usefulCycles": 153216.0,
      "estimatedWorkCycles": 2766096,
      "usefulUtilization": 0.05539070227497527,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.04663394109396914,
      "meanCycles": 2022.0,
      "p50Cycles": 2022,
      "p95Cycles": 2022,
      "maximumCycles": 2022,
      "firstOffset": 641376,
      "lastOffset": 643608
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k256_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1345,
      "tileCount": 1345,
      "phaseCount": 61,
      "phaseCycles": 2172,
      "phaseMs": 0.001448,
      "workCycles": 2259600,
      "averageActiveTiles": 1040.3314917127072,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2259600,
      "usefulUtilization": 0.017142857142857144,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.012115661782368484,
      "meanCycles": 1680.0,
      "p50Cycles": 1680,
      "p95Cycles": 1680,
      "maximumCycles": 1680,
      "firstOffset": 635688,
      "lastOffset": 637860
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c48_p224",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 17,
      "phaseCycles": 2052,
      "phaseMs": 0.001368,
      "workCycles": 145584,
      "averageActiveTiles": 70.94736842105263,
      "usefulCycles": 1728.0,
      "estimatedWorkCycles": 145584,
      "usefulUtilization": 0.011869436201780416,
      "workLaneOccupancy": 0.10714285714285714,
      "deviceUsefulUtilization": 0.0005720823798627002,
      "meanCycles": 2022.0,
      "p50Cycles": 2022,
      "p95Cycles": 2022,
      "maximumCycles": 2022,
      "firstOffset": 641208,
      "lastOffset": 643260
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 30,
      "phaseCycles": 2046,
      "phaseMs": 0.001364,
      "workCycles": 725760,
      "averageActiveTiles": 354.72140762463346,
      "usefulCycles": 10368.0,
      "estimatedWorkCycles": 725760,
      "usefulUtilization": 0.014285714285714285,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.003442560244804284,
      "meanCycles": 840.0,
      "p50Cycles": 840,
      "p95Cycles": 840,
      "maximumCycles": 840,
      "firstOffset": 545730,
      "lastOffset": 547776
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 77,
      "phaseCycles": 1890,
      "phaseMs": 0.00126,
      "workCycles": 2177280,
      "averageActiveTiles": 1152.0,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2177280,
      "usefulUtilization": 0.01779100529100529,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.013923395445134575,
      "meanCycles": 1512.0,
      "p50Cycles": 1512,
      "p95Cycles": 1512,
      "maximumCycles": 1512,
      "firstOffset": 645216,
      "lastOffset": 647106
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c64_p64",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 58,
      "phaseCycles": 1392,
      "phaseMs": 0.000928,
      "workCycles": 1065312,
      "averageActiveTiles": 765.3103448275862,
      "usefulCycles": 41472.0,
      "estimatedWorkCycles": 1065312,
      "usefulUtilization": 0.038929440389294405,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.020239880059970013,
      "meanCycles": 822.0,
      "p50Cycles": 822,
      "p95Cycles": 822,
      "maximumCycles": 822,
      "firstOffset": 610314,
      "lastOffset": 611706
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 58,
      "phaseCycles": 1152,
      "phaseMs": 0.000768,
      "workCycles": 870912,
      "averageActiveTiles": 756.0,
      "usefulCycles": 10368.0,
      "estimatedWorkCycles": 870912,
      "usefulUtilization": 0.011904761904761904,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.006114130434782609,
      "meanCycles": 672.0,
      "p50Cycles": 672,
      "p95Cycles": 672,
      "maximumCycles": 672,
      "firstOffset": 614448,
      "lastOffset": 615600
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1076,
      "tileCount": 1076,
      "phaseCount": 63,
      "phaseCycles": 864,
      "phaseMs": 0.000576,
      "workCycles": 839280,
      "averageActiveTiles": 971.3888888888889,
      "usefulCycles": 13988.0,
      "estimatedWorkCycles": 839280,
      "usefulUtilization": 0.016666666666666666,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.010998515499194847,
      "meanCycles": 780.0,
      "p50Cycles": 780,
      "p95Cycles": 780,
      "maximumCycles": 780,
      "firstOffset": 639810,
      "lastOffset": 640674
    }
  ],
  "samples": []
}
