{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 202925,
  "matchedSampleCount": 202925,
  "profileSpanCycles": 737892,
  "profileSpanMs": 0.491928,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 14269,
      "tileCount": 1468,
      "phaseCount": 44,
      "phaseCycles": 589026,
      "phaseMs": 0.392684,
      "workCycles": 87650982,
      "averageActiveTiles": 148.80664350979413,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 6142.755764244165,
      "p50Cycles": 2568,
      "p95Cycles": 19332,
      "maximumCycles": 63984,
      "firstOffset": 3228,
      "lastOffset": 736614
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 63747,
      "tileCount": 1472,
      "phaseCount": 53,
      "phaseCycles": 461238,
      "phaseMs": 0.307492,
      "workCycles": 448216872,
      "averageActiveTiles": 971.7691777347053,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 7031.183773354041,
      "p50Cycles": 5676,
      "p95Cycles": 19272,
      "maximumCycles": 35220,
      "firstOffset": 3228,
      "lastOffset": 736812
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 28196,
      "tileCount": 1472,
      "phaseCount": 207,
      "phaseCycles": 110694,
      "phaseMs": 0.073796,
      "workCycles": 21545052,
      "averageActiveTiles": 194.63613203967694,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 764.1173216059016,
      "p50Cycles": 558,
      "p95Cycles": 1398,
      "maximumCycles": 10578,
      "firstOffset": 10764,
      "lastOffset": 736332
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 13215,
      "tileCount": 1472,
      "phaseCount": 169,
      "phaseCycles": 67956,
      "phaseMs": 0.045304,
      "workCycles": 50991078,
      "averageActiveTiles": 750.3543174995585,
      "usefulCycles": 7716903.0,
      "estimatedWorkCycles": 46213986,
      "usefulUtilization": 0.16698198246738552,
      "workLaneOccupancy": 0.9390482260928483,
      "deviceUsefulUtilization": 0.07714493609355925,
      "meanCycles": 3858.575709421112,
      "p50Cycles": 3654,
      "p95Cycles": 8214,
      "maximumCycles": 9102,
      "firstOffset": 0,
      "lastOffset": 729924
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10659,
      "tileCount": 1472,
      "phaseCount": 181,
      "phaseCycles": 58074,
      "phaseMs": 0.038716,
      "workCycles": 39338838,
      "averageActiveTiles": 677.3915693770017,
      "usefulCycles": 21041112.0,
      "estimatedWorkCycles": 39338838,
      "usefulUtilization": 0.5348686710065,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.24613826661935073,
      "meanCycles": 3690.668730650155,
      "p50Cycles": 2658,
      "p95Cycles": 8994,
      "maximumCycles": 9576,
      "firstOffset": 11298,
      "lastOffset": 736062
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 15136,
      "tileCount": 1472,
      "phaseCount": 136,
      "phaseCycles": 48018,
      "phaseMs": 0.032012,
      "workCycles": 5601120,
      "averageActiveTiles": 116.64625765337998,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 2901888,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 370.05285412262157,
      "p50Cycles": 300,
      "p95Cycles": 600,
      "maximumCycles": 3150,
      "firstOffset": 126426,
      "lastOffset": 727650
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c192_r36_r38",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 26,
      "phaseCycles": 45198,
      "phaseMs": 0.030132,
      "workCycles": 16249680,
      "averageActiveTiles": 359.5221027479092,
      "usefulCycles": 7359840.0,
      "estimatedWorkCycles": 16249680,
      "usefulUtilization": 0.4529221498515663,
      "workLaneOccupancy": 0.9340277777777778,
      "deviceUsefulUtilization": 0.11062195903243122,
      "meanCycles": 45138.0,
      "p50Cycles": 45138,
      "p95Cycles": 45138,
      "maximumCycles": 45138,
      "firstOffset": 488466,
      "lastOffset": 533664
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c192_r36_r38",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 31,
      "phaseCycles": 44832,
      "phaseMs": 0.029888,
      "workCycles": 48366720,
      "averageActiveTiles": 1078.8436830835117,
      "usefulCycles": 20878704.0,
      "estimatedWorkCycles": 48366720,
      "usefulUtilization": 0.4316750029772538,
      "workLaneOccupancy": 0.9322980967078189,
      "deviceUsefulUtilization": 0.31637897425751793,
      "meanCycles": 44784.0,
      "p50Cycles": 44784,
      "p95Cycles": 44784,
      "maximumCycles": 44784,
      "firstOffset": 488460,
      "lastOffset": 533292
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c240_r72_r74",
      "dimensions": {},
      "sampleCount": 680,
      "tileCount": 680,
      "phaseCount": 27,
      "phaseCycles": 34770,
      "phaseMs": 0.02318,
      "workCycles": 23594640,
      "averageActiveTiles": 678.5918895599655,
      "usefulCycles": 13586400.0,
      "estimatedWorkCycles": 23594640,
      "usefulUtilization": 0.5758256960055335,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.2654556026559628,
      "meanCycles": 34698.0,
      "p50Cycles": 34698,
      "p95Cycles": 34698,
      "maximumCycles": 34698,
      "firstOffset": 351060,
      "lastOffset": 385830
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c240_r72_r74",
      "dimensions": {},
      "sampleCount": 680,
      "tileCount": 680,
      "phaseCount": 20,
      "phaseCycles": 34530,
      "phaseMs": 0.02302,
      "workCycles": 23443680,
      "averageActiveTiles": 678.9365768896612,
      "usefulCycles": 13182480.0,
      "estimatedWorkCycles": 23443680,
      "usefulUtilization": 0.5623042116254786,
      "workLaneOccupancy": 0.8975,
      "deviceUsefulUtilization": 0.25935386998073506,
      "meanCycles": 34476.0,
      "p50Cycles": 34476,
      "p95Cycles": 34476,
      "maximumCycles": 34476,
      "firstOffset": 351054,
      "lastOffset": 385584
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 33,
      "phaseCycles": 34248,
      "phaseMs": 0.022832,
      "workCycles": 37251936,
      "averageActiveTiles": 1087.7112824106516,
      "usefulCycles": 18466080.0,
      "estimatedWorkCycles": 37251936,
      "usefulUtilization": 0.4957079277705191,
      "workLaneOccupancy": 0.9974763406940063,
      "deviceUsefulUtilization": 0.3662955881904878,
      "meanCycles": 25034.90322580645,
      "p50Cycles": 25236,
      "p95Cycles": 25236,
      "maximumCycles": 25236,
      "firstOffset": 173082,
      "lastOffset": 679266
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c224_r72_r74",
      "dimensions": {},
      "sampleCount": 40,
      "tileCount": 40,
      "phaseCount": 14,
      "phaseCycles": 32808,
      "phaseMs": 0.021872,
      "workCycles": 1287840,
      "averageActiveTiles": 39.253840526700806,
      "usefulCycles": 723744.0,
      "estimatedWorkCycles": 1287840,
      "usefulUtilization": 0.5619828550130451,
      "workLaneOccupancy": 0.8975,
      "deviceUsefulUtilization": 0.014986403104226964,
      "meanCycles": 32196.0,
      "p50Cycles": 32196,
      "p95Cycles": 32196,
      "maximumCycles": 32196,
      "firstOffset": 350418,
      "lastOffset": 383226
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c224_r72_r74",
      "dimensions": {},
      "sampleCount": 40,
      "tileCount": 40,
      "phaseCount": 11,
      "phaseCycles": 32520,
      "phaseMs": 0.02168,
      "workCycles": 1296240,
      "averageActiveTiles": 39.859778597785976,
      "usefulCycles": 745920.0,
      "estimatedWorkCycles": 1296240,
      "usefulUtilization": 0.575448990927606,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.015582384084710412,
      "meanCycles": 32406.0,
      "p50Cycles": 32406,
      "p95Cycles": 32406,
      "maximumCycles": 32406,
      "firstOffset": 350424,
      "lastOffset": 382944
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 59,
      "phaseCycles": 31614,
      "phaseMs": 0.021076,
      "workCycles": 20509686,
      "averageActiveTiles": 648.7532738660087,
      "usefulCycles": 10077696.0,
      "estimatedWorkCycles": 20509686,
      "usefulUtilization": 0.491362763915547,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.21655788162096595,
      "meanCycles": 9378.0,
      "p50Cycles": 9378,
      "p95Cycles": 9378,
      "maximumCycles": 9378,
      "firstOffset": 23118,
      "lastOffset": 559278
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 1834,
      "tileCount": 788,
      "phaseCount": 116,
      "phaseCycles": 30132,
      "phaseMs": 0.020088,
      "workCycles": 524232,
      "averageActiveTiles": 17.397849462365592,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 285.84078516902946,
      "p50Cycles": 288,
      "p95Cycles": 288,
      "maximumCycles": 288,
      "firstOffset": 100854,
      "lastOffset": 728430
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3344,
      "tileCount": 1472,
      "phaseCount": 108,
      "phaseCycles": 28710,
      "phaseMs": 0.01914,
      "workCycles": 1390944,
      "averageActiveTiles": 48.44806687565308,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 415.95215311004785,
      "p50Cycles": 414,
      "p95Cycles": 714,
      "maximumCycles": 714,
      "firstOffset": 108414,
      "lastOffset": 670122
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c320_r66_r68",
      "dimensions": {},
      "sampleCount": 168,
      "tileCount": 168,
      "phaseCount": 3,
      "phaseCycles": 26694,
      "phaseMs": 0.017796,
      "workCycles": 4468464,
      "averageActiveTiles": 167.39581928523265,
      "usefulCycles": 2741760.0,
      "estimatedWorkCycles": 4468464,
      "usefulUtilization": 0.6135799684186781,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.06977630537394823,
      "meanCycles": 26598.0,
      "p50Cycles": 26598,
      "p95Cycles": 26598,
      "maximumCycles": 26598,
      "firstOffset": 58428,
      "lastOffset": 85122
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c320_r66_r68",
      "dimensions": {},
      "sampleCount": 756,
      "tileCount": 756,
      "phaseCount": 3,
      "phaseCycles": 26454,
      "phaseMs": 0.017636,
      "workCycles": 19949328,
      "averageActiveTiles": 754.1138580176911,
      "usefulCycles": 11954880.0,
      "estimatedWorkCycles": 19949328,
      "usefulUtilization": 0.5992622909403265,
      "workLaneOccupancy": 0.9983164983164983,
      "deviceUsefulUtilization": 0.3070054335499522,
      "meanCycles": 26388.0,
      "p50Cycles": 26388,
      "p95Cycles": 26388,
      "maximumCycles": 26388,
      "firstOffset": 58500,
      "lastOffset": 84954
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c304_r66_r68",
      "dimensions": {},
      "sampleCount": 96,
      "tileCount": 96,
      "phaseCount": 2,
      "phaseCycles": 25380,
      "phaseMs": 0.01692,
      "workCycles": 2427264,
      "averageActiveTiles": 95.63687943262411,
      "usefulCycles": 1488384.0,
      "estimatedWorkCycles": 2427264,
      "usefulUtilization": 0.6131941148552444,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.03983965464076472,
      "meanCycles": 25284.0,
      "p50Cycles": 25284,
      "p95Cycles": 25284,
      "maximumCycles": 25284,
      "firstOffset": 58044,
      "lastOffset": 83424
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c304_r66_r68",
      "dimensions": {},
      "sampleCount": 432,
      "tileCount": 432,
      "phaseCount": 3,
      "phaseCycles": 25152,
      "phaseMs": 0.016768,
      "workCycles": 10837152,
      "averageActiveTiles": 430.86641221374043,
      "usefulCycles": 6489792.0,
      "estimatedWorkCycles": 10837152,
      "usefulUtilization": 0.5988466342448643,
      "workLaneOccupancy": 0.9983164983164983,
      "deviceUsefulUtilization": 0.1752872967142383,
      "meanCycles": 25086.0,
      "p50Cycles": 25086,
      "p95Cycles": 25086,
      "maximumCycles": 25086,
      "firstOffset": 58116,
      "lastOffset": 83268
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 50,
      "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": 133716,
      "lastOffset": 658674
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c256_r40_r42",
      "dimensions": {},
      "sampleCount": 405,
      "tileCount": 405,
      "phaseCount": 28,
      "phaseCycles": 21948,
      "phaseMs": 0.014632,
      "workCycles": 8852490,
      "averageActiveTiles": 403.33925642427556,
      "usefulCycles": 4354560.0,
      "estimatedWorkCycles": 8852490,
      "usefulUtilization": 0.4919022783420258,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.13478498585589654,
      "meanCycles": 21858.0,
      "p50Cycles": 21858,
      "p95Cycles": 21858,
      "maximumCycles": 21858,
      "firstOffset": 583452,
      "lastOffset": 605400
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 45,
      "phaseCycles": 21720,
      "phaseMs": 0.01448,
      "workCycles": 26532684,
      "averageActiveTiles": 1221.578453038674,
      "usefulCycles": 10197252.0,
      "estimatedWorkCycles": 26532684,
      "usefulUtilization": 0.3843279481261677,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.31894479642085033,
      "meanCycles": 18198.0,
      "p50Cycles": 18198,
      "p95Cycles": 18198,
      "maximumCycles": 18198,
      "firstOffset": 412380,
      "lastOffset": 434100
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r38_p38_c80_p384",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 16,
      "phaseCycles": 21360,
      "phaseMs": 0.01424,
      "workCycles": 576540,
      "averageActiveTiles": 26.991573033707866,
      "usefulCycles": 45600.0,
      "estimatedWorkCycles": 576540,
      "usefulUtilization": 0.07909251743157457,
      "workLaneOccupancy": 0.20833333333333334,
      "deviceUsefulUtilization": 0.0014502931118710307,
      "meanCycles": 19218.0,
      "p50Cycles": 19218,
      "p95Cycles": 19218,
      "maximumCycles": 19218,
      "firstOffset": 439230,
      "lastOffset": 460590
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c256_r40_r42",
      "dimensions": {},
      "sampleCount": 1053,
      "tileCount": 1053,
      "phaseCount": 37,
      "phaseCycles": 21126,
      "phaseMs": 0.014084,
      "workCycles": 22169862,
      "averageActiveTiles": 1049.411246804885,
      "usefulCycles": 10761984.0,
      "estimatedWorkCycles": 22169862,
      "usefulUtilization": 0.4854330622355701,
      "workLaneOccupancy": 0.9980769230769231,
      "deviceUsefulUtilization": 0.3460726325278145,
      "meanCycles": 21054.0,
      "p50Cycles": 21054,
      "p95Cycles": 21054,
      "maximumCycles": 21054,
      "firstOffset": 583512,
      "lastOffset": 604638
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 12221,
      "tileCount": 1458,
      "phaseCount": 185,
      "phaseCycles": 19860,
      "phaseMs": 0.01324,
      "workCycles": 15277356,
      "averageActiveTiles": 769.2525679758309,
      "usefulCycles": 2311700.0,
      "estimatedWorkCycles": 15277356,
      "usefulUtilization": 0.15131545013417244,
      "workLaneOccupancy": 0.999037135122614,
      "deviceUsefulUtilization": 0.07907595012916502,
      "meanCycles": 1250.0904999590869,
      "p50Cycles": 1068,
      "p95Cycles": 2220,
      "maximumCycles": 2220,
      "firstOffset": 18288,
      "lastOffset": 737892
    },
    {
      "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": 210054,
      "lastOffset": 228930
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r36_p36_c80_p384",
      "dimensions": {},
      "sampleCount": 84,
      "tileCount": 84,
      "phaseCount": 16,
      "phaseCycles": 18594,
      "phaseMs": 0.012396,
      "workCycles": 1386504,
      "averageActiveTiles": 74.56727976766699,
      "usefulCycles": 120960.0,
      "estimatedWorkCycles": 1386504,
      "usefulUtilization": 0.08724100327153762,
      "workLaneOccupancy": 0.20833333333333334,
      "deviceUsefulUtilization": 0.004419377919946126,
      "meanCycles": 16506.0,
      "p50Cycles": 16506,
      "p95Cycles": 16506,
      "maximumCycles": 16506,
      "firstOffset": 438984,
      "lastOffset": 457578
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r35_p36_c80_p384",
      "dimensions": {},
      "sampleCount": 6,
      "tileCount": 6,
      "phaseCount": 3,
      "phaseCycles": 18252,
      "phaseMs": 0.012168,
      "workCycles": 100584,
      "averageActiveTiles": 5.510848126232742,
      "usefulCycles": 8400.0,
      "estimatedWorkCycles": 100584,
      "usefulUtilization": 0.08351228823669768,
      "workLaneOccupancy": 0.2025462962962963,
      "deviceUsefulUtilization": 0.00031265185947460195,
      "meanCycles": 16764.0,
      "p50Cycles": 16764,
      "p95Cycles": 16764,
      "maximumCycles": 16764,
      "firstOffset": 438852,
      "lastOffset": 457104
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 128,
      "tileCount": 128,
      "phaseCount": 10,
      "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": 210738,
      "lastOffset": 228984
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 224,
      "tileCount": 224,
      "phaseCount": 8,
      "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": 150276,
      "lastOffset": 168384
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r74_p74_c32_p160",
      "dimensions": {},
      "sampleCount": 90,
      "tileCount": 90,
      "phaseCount": 10,
      "phaseCycles": 17940,
      "phaseMs": 0.01196,
      "workCycles": 1418040,
      "averageActiveTiles": 79.04347826086956,
      "usefulCycles": 106560.0,
      "estimatedWorkCycles": 1418040,
      "usefulUtilization": 0.07514597613607514,
      "workLaneOccupancy": 0.2,
      "deviceUsefulUtilization": 0.0040351897629780425,
      "meanCycles": 15756.0,
      "p50Cycles": 15756,
      "p95Cycles": 15756,
      "maximumCycles": 15756,
      "firstOffset": 316644,
      "lastOffset": 334584
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 1008,
      "tileCount": 1008,
      "phaseCount": 9,
      "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": 147636,
      "lastOffset": 164874
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 34,
      "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": 135294,
      "lastOffset": 657744
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 176,
      "tileCount": 176,
      "phaseCount": 31,
      "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": 640728,
      "lastOffset": 656244
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r72_p72_c32_p160",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 9,
      "phaseCycles": 15396,
      "phaseMs": 0.010264,
      "workCycles": 1048464,
      "averageActiveTiles": 68.09976617303195,
      "usefulCycles": 82944.0,
      "estimatedWorkCycles": 1048464,
      "usefulUtilization": 0.07911001236093942,
      "workLaneOccupancy": 0.2,
      "deviceUsefulUtilization": 0.0036599003693788337,
      "meanCycles": 14562.0,
      "p50Cycles": 14562,
      "p95Cycles": 14562,
      "maximumCycles": 14562,
      "firstOffset": 311712,
      "lastOffset": 327108
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r38_p38_c384_p384",
      "dimensions": {},
      "sampleCount": 330,
      "tileCount": 330,
      "phaseCount": 26,
      "phaseCycles": 15384,
      "phaseMs": 0.010256,
      "workCycles": 3958020,
      "averageActiveTiles": 257.28159126365057,
      "usefulCycles": 2407680.0,
      "estimatedWorkCycles": 3958020,
      "usefulUtilization": 0.608304152076038,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.10632164417011464,
      "meanCycles": 11994.0,
      "p50Cycles": 11994,
      "p95Cycles": 11994,
      "maximumCycles": 11994,
      "firstOffset": 438852,
      "lastOffset": 454236
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r71_p72_c32_p160",
      "dimensions": {},
      "sampleCount": 18,
      "tileCount": 18,
      "phaseCount": 4,
      "phaseCycles": 14670,
      "phaseMs": 0.00978,
      "workCycles": 263844,
      "averageActiveTiles": 17.98527607361963,
      "usefulCycles": 20448.0,
      "estimatedWorkCycles": 263844,
      "usefulUtilization": 0.07750034111065629,
      "workLaneOccupancy": 0.19722222222222222,
      "deviceUsefulUtilization": 0.000946919178447586,
      "meanCycles": 14658.0,
      "p50Cycles": 14658,
      "p95Cycles": 14658,
      "maximumCycles": 14658,
      "firstOffset": 312258,
      "lastOffset": 326928
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r35_p36_c384_p384",
      "dimensions": {},
      "sampleCount": 66,
      "tileCount": 66,
      "phaseCount": 13,
      "phaseCycles": 14286,
      "phaseMs": 0.009524,
      "workCycles": 765864,
      "averageActiveTiles": 53.60940781184377,
      "usefulCycles": 443520.0,
      "estimatedWorkCycles": 765864,
      "usefulUtilization": 0.5791106514994829,
      "workLaneOccupancy": 0.9722222222222222,
      "deviceUsefulUtilization": 0.021090882530175485,
      "meanCycles": 11604.0,
      "p50Cycles": 11604,
      "p95Cycles": 11604,
      "maximumCycles": 11604,
      "firstOffset": 438828,
      "lastOffset": 453114
    },
    {
      "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": 154368,
      "lastOffset": 168432
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r36_p36_c384_p384",
      "dimensions": {},
      "sampleCount": 924,
      "tileCount": 924,
      "phaseCount": 24,
      "phaseCycles": 13446,
      "phaseMs": 0.008964,
      "workCycles": 9530136,
      "averageActiveTiles": 708.7710843373494,
      "usefulCycles": 6386688.0,
      "estimatedWorkCycles": 9530136,
      "usefulUtilization": 0.6701570680628273,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.322682032477737,
      "meanCycles": 10314.0,
      "p50Cycles": 10314,
      "p95Cycles": 10314,
      "maximumCycles": 10314,
      "firstOffset": 438696,
      "lastOffset": 452142
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r74_p74_c160_p160",
      "dimensions": {},
      "sampleCount": 630,
      "tileCount": 630,
      "phaseCount": 27,
      "phaseCycles": 13020,
      "phaseMs": 0.00868,
      "workCycles": 6339060,
      "averageActiveTiles": 486.8709677419355,
      "usefulCycles": 3729600.0,
      "estimatedWorkCycles": 6339060,
      "usefulUtilization": 0.588352216259193,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.19460028050490882,
      "meanCycles": 10062.0,
      "p50Cycles": 10062,
      "p95Cycles": 10062,
      "maximumCycles": 10062,
      "firstOffset": 316038,
      "lastOffset": 329058
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k736_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 32,
      "phaseCycles": 12546,
      "phaseMs": 0.008364,
      "workCycles": 644160,
      "averageActiveTiles": 51.34385461501674,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 644160,
      "usefulUtilization": 0.020370715350223546,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0007105373962946791,
      "meanCycles": 8052.0,
      "p50Cycles": 8052,
      "p95Cycles": 8052,
      "maximumCycles": 8052,
      "firstOffset": 686046,
      "lastOffset": 698592
    },
    {
      "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": 639954,
      "lastOffset": 652488
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 144,
      "tileCount": 144,
      "phaseCount": 8,
      "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": 154086,
      "lastOffset": 165786
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c96_r36_r38",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 20,
      "phaseCycles": 11640,
      "phaseMs": 0.00776,
      "workCycles": 4177440,
      "averageActiveTiles": 358.88659793814435,
      "usefulCycles": 1969920.0,
      "estimatedWorkCycles": 4177440,
      "usefulUtilization": 0.47156153050672184,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.11497086508292245,
      "meanCycles": 11604.0,
      "p50Cycles": 11604,
      "p95Cycles": 11604,
      "maximumCycles": 11604,
      "firstOffset": 273840,
      "lastOffset": 285480
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c96_r36_r38",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 24,
      "phaseCycles": 11550,
      "phaseMs": 0.0077,
      "workCycles": 12402720,
      "averageActiveTiles": 1073.8285714285714,
      "usefulCycles": 5588352.0,
      "estimatedWorkCycles": 12402720,
      "usefulUtilization": 0.45057471264367815,
      "workLaneOccupancy": 0.9981481481481481,
      "deviceUsefulUtilization": 0.32869565217391306,
      "meanCycles": 11484.0,
      "p50Cycles": 11484,
      "p95Cycles": 11484,
      "maximumCycles": 11484,
      "firstOffset": 273846,
      "lastOffset": 285396
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 37,
      "phaseCycles": 10860,
      "phaseMs": 0.00724,
      "workCycles": 856704,
      "averageActiveTiles": 78.88618784530387,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 856704,
      "usefulUtilization": 0.015316842223218288,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0008208473456641844,
      "meanCycles": 1164.0,
      "p50Cycles": 1164,
      "p95Cycles": 1164,
      "maximumCycles": 1164,
      "firstOffset": 660750,
      "lastOffset": 671610
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r72_p72_c160_p160",
      "dimensions": {},
      "sampleCount": 504,
      "tileCount": 504,
      "phaseCount": 16,
      "phaseCycles": 10842,
      "phaseMs": 0.007228,
      "workCycles": 4690224,
      "averageActiveTiles": 432.5976757055894,
      "usefulCycles": 2903040.0,
      "estimatedWorkCycles": 4690224,
      "usefulUtilization": 0.6189555125725339,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.18190130170111402,
      "meanCycles": 9306.0,
      "p50Cycles": 9306,
      "p95Cycles": 9306,
      "maximumCycles": 9306,
      "firstOffset": 317544,
      "lastOffset": 328386
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r71_p72_c160_p160",
      "dimensions": {},
      "sampleCount": 126,
      "tileCount": 126,
      "phaseCount": 8,
      "phaseCycles": 10176,
      "phaseMs": 0.006784,
      "workCycles": 1239840,
      "averageActiveTiles": 121.83962264150944,
      "usefulCycles": 715680.0,
      "estimatedWorkCycles": 1239840,
      "usefulUtilization": 0.5772357723577236,
      "workLaneOccupancy": 0.9861111111111112,
      "deviceUsefulUtilization": 0.04777866078753076,
      "meanCycles": 9840.0,
      "p50Cycles": 9840,
      "p95Cycles": 9840,
      "maximumCycles": 9840,
      "firstOffset": 317616,
      "lastOffset": 327792
    },
    {
      "name": "attention_merge_v72_p80_k768_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 45,
      "phaseCycles": 8688,
      "phaseMs": 0.005792,
      "workCycles": 3722400,
      "averageActiveTiles": 428.45303867403317,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2501.6129032258063,
      "p50Cycles": 2514,
      "p95Cycles": 2514,
      "maximumCycles": 2514,
      "firstOffset": 236694,
      "lastOffset": 700572
    },
    {
      "name": "cast_f32_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 45,
      "phaseCycles": 8550,
      "phaseMs": 0.0057,
      "workCycles": 4586400,
      "averageActiveTiles": 536.421052631579,
      "usefulCycles": 420480.0,
      "estimatedWorkCycles": 4586400,
      "usefulUtilization": 0.09167974882260596,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.033409610983981694,
      "meanCycles": 3082.2580645161293,
      "p50Cycles": 3138,
      "p95Cycles": 3138,
      "maximumCycles": 3138,
      "firstOffset": 239208,
      "lastOffset": 701172
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r38_p38_c192_p192",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 20,
      "phaseCycles": 7500,
      "phaseMs": 0.005,
      "workCycles": 2322000,
      "averageActiveTiles": 309.6,
      "usefulCycles": 1313280.0,
      "estimatedWorkCycles": 2322000,
      "usefulUtilization": 0.5655813953488372,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.11895652173913043,
      "meanCycles": 6450.0,
      "p50Cycles": 6450,
      "p95Cycles": 6450,
      "maximumCycles": 6450,
      "firstOffset": 256380,
      "lastOffset": 263880
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r68_p68_c96_p96",
      "dimensions": {},
      "sampleCount": 264,
      "tileCount": 264,
      "phaseCount": 3,
      "phaseCycles": 7038,
      "phaseMs": 0.004692,
      "workCycles": 1620432,
      "averageActiveTiles": 230.24040920716112,
      "usefulCycles": 861696.0,
      "estimatedWorkCycles": 1620432,
      "usefulUtilization": 0.5317693059628543,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0831758034026465,
      "meanCycles": 6138.0,
      "p50Cycles": 6138,
      "p95Cycles": 6138,
      "maximumCycles": 6138,
      "firstOffset": 39378,
      "lastOffset": 46416
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r66_p66_c96_p96",
      "dimensions": {},
      "sampleCount": 1056,
      "tileCount": 1056,
      "phaseCount": 3,
      "phaseCycles": 6642,
      "phaseMs": 0.004428,
      "workCycles": 5962176,
      "averageActiveTiles": 897.6476964769648,
      "usefulCycles": 3345408.0,
      "estimatedWorkCycles": 5962176,
      "usefulUtilization": 0.5611052072263549,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3421703782255214,
      "meanCycles": 5646.0,
      "p50Cycles": 5646,
      "p95Cycles": 5646,
      "maximumCycles": 5646,
      "firstOffset": 38916,
      "lastOffset": 45558
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r35_p36_c192_p192",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 7,
      "phaseCycles": 6612,
      "phaseMs": 0.004408,
      "workCycles": 446688,
      "averageActiveTiles": 67.55716878402905,
      "usefulCycles": 241920.0,
      "estimatedWorkCycles": 446688,
      "usefulUtilization": 0.5415860735009671,
      "workLaneOccupancy": 0.9722222222222222,
      "deviceUsefulUtilization": 0.024855993056103528,
      "meanCycles": 6204.0,
      "p50Cycles": 6204,
      "p95Cycles": 6204,
      "maximumCycles": 6204,
      "firstOffset": 255882,
      "lastOffset": 262494
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r36_p36_c192_p192",
      "dimensions": {},
      "sampleCount": 1008,
      "tileCount": 1008,
      "phaseCount": 21,
      "phaseCycles": 6600,
      "phaseMs": 0.0044,
      "workCycles": 5606496,
      "averageActiveTiles": 849.4690909090909,
      "usefulCycles": 3483648.0,
      "estimatedWorkCycles": 5606496,
      "usefulUtilization": 0.6213592233009708,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3585770750988142,
      "meanCycles": 5562.0,
      "p50Cycles": 5562,
      "p95Cycles": 5562,
      "maximumCycles": 5562,
      "firstOffset": 255678,
      "lastOffset": 262278
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r65_p66_c96_p96",
      "dimensions": {},
      "sampleCount": 132,
      "tileCount": 132,
      "phaseCount": 2,
      "phaseCycles": 6534,
      "phaseMs": 0.004356,
      "workCycles": 787248,
      "averageActiveTiles": 120.48484848484848,
      "usefulCycles": 411840.0,
      "estimatedWorkCycles": 787248,
      "usefulUtilization": 0.5231388329979879,
      "workLaneOccupancy": 0.9848484848484849,
      "deviceUsefulUtilization": 0.04281949934123847,
      "meanCycles": 5964.0,
      "p50Cycles": 5964,
      "p95Cycles": 5964,
      "maximumCycles": 5964,
      "firstOffset": 39198,
      "lastOffset": 45732
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k320_c16_r72_r72",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 1,
      "phaseCycles": 6306,
      "phaseMs": 0.004204,
      "workCycles": 7100928,
      "averageActiveTiles": 1126.058991436727,
      "usefulCycles": 3857868.0,
      "estimatedWorkCycles": 7100928,
      "usefulUtilization": 0.5432906797534068,
      "workLaneOccupancy": 0.9100118885869565,
      "deviceUsefulUtilization": 0.4156096161006081,
      "meanCycles": 4824.0,
      "p50Cycles": 4824,
      "p95Cycles": 4824,
      "maximumCycles": 4824,
      "firstOffset": 3744,
      "lastOffset": 10050
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r40_p40_c128_p128",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 26,
      "phaseCycles": 6258,
      "phaseMs": 0.004172,
      "workCycles": 4473144,
      "averageActiveTiles": 714.7881112176415,
      "usefulCycles": 2488320.0,
      "estimatedWorkCycles": 4473144,
      "usefulUtilization": 0.5562798783137766,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.270123806744758,
      "meanCycles": 4602.0,
      "p50Cycles": 4602,
      "p95Cycles": 4602,
      "maximumCycles": 4602,
      "firstOffset": 564510,
      "lastOffset": 570768
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r42_p42_c128_p128",
      "dimensions": {},
      "sampleCount": 405,
      "tileCount": 405,
      "phaseCount": 28,
      "phaseCycles": 5808,
      "phaseMs": 0.003872,
      "workCycles": 1863810,
      "averageActiveTiles": 320.9039256198347,
      "usefulCycles": 1088640.0,
      "estimatedWorkCycles": 1863810,
      "usefulUtilization": 0.5840938722294654,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.12733560905497665,
      "meanCycles": 4602.0,
      "p50Cycles": 4602,
      "p95Cycles": 4602,
      "maximumCycles": 4602,
      "firstOffset": 565524,
      "lastOffset": 571332
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r39_p40_c128_p128",
      "dimensions": {},
      "sampleCount": 81,
      "tileCount": 81,
      "phaseCount": 19,
      "phaseCycles": 5322,
      "phaseMs": 0.003548,
      "workCycles": 406782,
      "averageActiveTiles": 76.43404735062006,
      "usefulCycles": 202176.0,
      "estimatedWorkCycles": 406782,
      "usefulUtilization": 0.4970131421744325,
      "workLaneOccupancy": 0.975,
      "deviceUsefulUtilization": 0.02580755845301701,
      "meanCycles": 5022.0,
      "p50Cycles": 5022,
      "p95Cycles": 5022,
      "maximumCycles": 5022,
      "firstOffset": 565716,
      "lastOffset": 571038
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 880,
      "tileCount": 880,
      "phaseCount": 10,
      "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": 132180,
      "lastOffset": 136830
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80",
      "dimensions": {},
      "sampleCount": 1360,
      "tileCount": 1360,
      "phaseCount": 10,
      "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": 130914,
      "lastOffset": 135294
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r9_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 1,
      "phaseCycles": 4218,
      "phaseMs": 0.002812,
      "workCycles": 132096,
      "averageActiveTiles": 31.317211948790895,
      "usefulCycles": 5184.0,
      "estimatedWorkCycles": 132096,
      "usefulUtilization": 0.03924418604651163,
      "workLaneOccupancy": 0.5625,
      "deviceUsefulUtilization": 0.0008349310408806976,
      "meanCycles": 4128.0,
      "p50Cycles": 4128,
      "p95Cycles": 4128,
      "maximumCycles": 4128,
      "firstOffset": 13056,
      "lastOffset": 17274
    },
    {
      "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": 132636,
      "lastOffset": 136692
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1728,
      "tileCount": 864,
      "phaseCount": 73,
      "phaseCycles": 3912,
      "phaseMs": 0.002608,
      "workCycles": 1451520,
      "averageActiveTiles": 371.04294478527606,
      "usefulCycles": 20736.0,
      "estimatedWorkCycles": 1451520,
      "usefulUtilization": 0.014285714285714285,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.003600960256068285,
      "meanCycles": 840.0,
      "p50Cycles": 840,
      "p95Cycles": 840,
      "maximumCycles": 840,
      "firstOffset": 627990,
      "lastOffset": 707076
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r16_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 1,
      "phaseCycles": 3792,
      "phaseMs": 0.002528,
      "workCycles": 5322240,
      "averageActiveTiles": 1403.5443037974683,
      "usefulCycles": 414720.0,
      "estimatedWorkCycles": 5322240,
      "usefulUtilization": 0.07792207792207792,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07429829389102917,
      "meanCycles": 3696.0,
      "p50Cycles": 3696,
      "p95Cycles": 3696,
      "maximumCycles": 3696,
      "firstOffset": 13050,
      "lastOffset": 16842
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 7,
      "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": 128856,
      "lastOffset": 134250
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 87,
      "phaseCycles": 3294,
      "phaseMs": 0.002196,
      "workCycles": 2177280,
      "averageActiveTiles": 660.983606557377,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2177280,
      "usefulUtilization": 0.01779100529100529,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.007988833452126396,
      "meanCycles": 1512.0,
      "p50Cycles": 1512,
      "p95Cycles": 1512,
      "maximumCycles": 1512,
      "firstOffset": 731202,
      "lastOffset": 734496
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k256_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1345,
      "tileCount": 1345,
      "phaseCount": 72,
      "phaseCycles": 2892,
      "phaseMs": 0.001928,
      "workCycles": 2259600,
      "averageActiveTiles": 781.3278008298755,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2259600,
      "usefulUtilization": 0.017142857142857144,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.009099314450658489,
      "meanCycles": 1680.0,
      "p50Cycles": 1680,
      "p95Cycles": 1680,
      "maximumCycles": 1680,
      "firstOffset": 721146,
      "lastOffset": 724038
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 4,
      "tileCount": 4,
      "phaseCount": 3,
      "phaseCycles": 2556,
      "phaseMs": 0.001704,
      "workCycles": 10224,
      "averageActiveTiles": 4.0,
      "usefulCycles": 2304.0,
      "estimatedWorkCycles": 10224,
      "usefulUtilization": 0.22535211267605634,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.000612369871402327,
      "meanCycles": 2556.0,
      "p50Cycles": 2556,
      "p95Cycles": 2556,
      "maximumCycles": 2556,
      "firstOffset": 712302,
      "lastOffset": 714858
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c256_p256",
      "dimensions": {},
      "sampleCount": 1076,
      "tileCount": 1076,
      "phaseCount": 70,
      "phaseCycles": 2520,
      "phaseMs": 0.00168,
      "workCycles": 2433912,
      "averageActiveTiles": 965.8380952380952,
      "usefulCycles": 137728.0,
      "estimatedWorkCycles": 2433912,
      "usefulUtilization": 0.05658709106984969,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.03712905452035887,
      "meanCycles": 2262.0,
      "p50Cycles": 2262,
      "p95Cycles": 2262,
      "maximumCycles": 2262,
      "firstOffset": 715674,
      "lastOffset": 718194
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c128_p256",
      "dimensions": {},
      "sampleCount": 269,
      "tileCount": 269,
      "phaseCount": 16,
      "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": 715866,
      "lastOffset": 718164
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 4,
      "tileCount": 4,
      "phaseCount": 3,
      "phaseCycles": 1764,
      "phaseMs": 0.001176,
      "workCycles": 6984,
      "averageActiveTiles": 3.9591836734693877,
      "usefulCycles": 2304.0,
      "estimatedWorkCycles": 6984,
      "usefulUtilization": 0.32989690721649484,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0008873114463176575,
      "meanCycles": 1746.0,
      "p50Cycles": 1746,
      "p95Cycles": 1746,
      "maximumCycles": 1746,
      "firstOffset": 710262,
      "lastOffset": 712026
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c96_p96",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 59,
      "phaseCycles": 1614,
      "phaseMs": 0.001076,
      "workCycles": 917568,
      "averageActiveTiles": 568.5055762081785,
      "usefulCycles": 41472.0,
      "estimatedWorkCycles": 917568,
      "usefulUtilization": 0.04519774011299435,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.01745595603685146,
      "meanCycles": 1062.0,
      "p50Cycles": 1062,
      "p95Cycles": 1062,
      "maximumCycles": 1062,
      "firstOffset": 702918,
      "lastOffset": 704532
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 269,
      "tileCount": 269,
      "phaseCount": 52,
      "phaseCycles": 918,
      "phaseMs": 0.000612,
      "workCycles": 225960,
      "averageActiveTiles": 246.1437908496732,
      "usefulCycles": 12912.0,
      "estimatedWorkCycles": 225960,
      "usefulUtilization": 0.05714285714285714,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.009555271383915886,
      "meanCycles": 840.0,
      "p50Cycles": 840,
      "p95Cycles": 840,
      "maximumCycles": 840,
      "firstOffset": 726390,
      "lastOffset": 727308
    }
  ],
  "samples": []
}
