{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 246792,
  "matchedSampleCount": 246792,
  "profileSpanCycles": 13291890,
  "profileSpanMs": 8.86126,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 11,
      "phaseCycles": 12529884,
      "phaseMs": 8.353256,
      "workCycles": 18439570152,
      "averageActiveTiles": 1471.647315489912,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 12526881.89673913,
      "p50Cycles": 12526884,
      "p95Cycles": 12526896,
      "maximumCycles": 12526920,
      "firstOffset": 542184,
      "lastOffset": 13072068
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 26254,
      "tileCount": 1457,
      "phaseCount": 70,
      "phaseCycles": 603564,
      "phaseMs": 0.402376,
      "workCycles": 133977156,
      "averageActiveTiles": 221.97671829333757,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5103.114039765369,
      "p50Cycles": 2202,
      "p95Cycles": 19674,
      "maximumCycles": 38640,
      "firstOffset": 0,
      "lastOffset": 13291164
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 91506,
      "tileCount": 1472,
      "phaseCount": 80,
      "phaseCycles": 514674,
      "phaseMs": 0.343116,
      "workCycles": 535161768,
      "averageActiveTiles": 1039.8072721761737,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5848.378991541538,
      "p50Cycles": 4440,
      "p95Cycles": 14838,
      "maximumCycles": 31020,
      "firstOffset": 0,
      "lastOffset": 13291740
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 28987,
      "tileCount": 1472,
      "phaseCount": 235,
      "phaseCycles": 121854,
      "phaseMs": 0.081236,
      "workCycles": 23403648,
      "averageActiveTiles": 192.06302624452212,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 807.3842757098009,
      "p50Cycles": 576,
      "p95Cycles": 1572,
      "maximumCycles": 10578,
      "firstOffset": 38994,
      "lastOffset": 13287924
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10440,
      "tileCount": 1472,
      "phaseCount": 155,
      "phaseCycles": 63858,
      "phaseMs": 0.042572,
      "workCycles": 42663456,
      "averageActiveTiles": 668.0988443108146,
      "usefulCycles": 23746848.0,
      "estimatedWorkCycles": 42663456,
      "usefulUtilization": 0.5566086347997686,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.25262879459452836,
      "meanCycles": 4086.537931034483,
      "p50Cycles": 3690,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 39258,
      "lastOffset": 13287654
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 11652,
      "tileCount": 1458,
      "phaseCount": 165,
      "phaseCycles": 63036,
      "phaseMs": 0.042024,
      "workCycles": 56437416,
      "averageActiveTiles": 895.3203883495146,
      "usefulCycles": 10457856.0,
      "estimatedWorkCycles": 56437416,
      "usefulUtilization": 0.18530004988180182,
      "workLaneOccupancy": 0.9751656629539366,
      "deviceUsefulUtilization": 0.11270578303081469,
      "meanCycles": 4843.581874356334,
      "p50Cycles": 4512,
      "p95Cycles": 9162,
      "maximumCycles": 9906,
      "firstOffset": 18306,
      "lastOffset": 13278276
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 17964,
      "tileCount": 1472,
      "phaseCount": 113,
      "phaseCycles": 61950,
      "phaseMs": 0.0413,
      "workCycles": 6292224,
      "averageActiveTiles": 101.56939467312348,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 3543024,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 350.2685370741483,
      "p50Cycles": 306,
      "p95Cycles": 606,
      "maximumCycles": 3150,
      "firstOffset": 28926,
      "lastOffset": 13278894
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 4177,
      "tileCount": 1472,
      "phaseCount": 130,
      "phaseCycles": 52704,
      "phaseMs": 0.035136,
      "workCycles": 1816584,
      "averageActiveTiles": 34.46766848816029,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 434.90160402202537,
      "p50Cycles": 420,
      "p95Cycles": 720,
      "maximumCycles": 5430,
      "firstOffset": 0,
      "lastOffset": 13284558
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 28,
      "phaseCycles": 32700,
      "phaseMs": 0.0218,
      "workCycles": 34993632,
      "averageActiveTiles": 1070.1416513761467,
      "usefulCycles": 15838080.0,
      "estimatedWorkCycles": 34993632,
      "usefulUtilization": 0.4525989185689556,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.32903869166334265,
      "meanCycles": 23517.225806451614,
      "p50Cycles": 23706,
      "p95Cycles": 23706,
      "maximumCycles": 23706,
      "firstOffset": 202398,
      "lastOffset": 13211448
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 28,
      "phaseCycles": 30648,
      "phaseMs": 0.020432,
      "workCycles": 23797800,
      "averageActiveTiles": 776.4878621769773,
      "usefulCycles": 15350400.0,
      "estimatedWorkCycles": 23797800,
      "usefulUtilization": 0.6450344149459194,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.34025909911136837,
      "meanCycles": 30510.0,
      "p50Cycles": 30510,
      "p95Cycles": 30510,
      "maximumCycles": 30510,
      "firstOffset": 366474,
      "lastOffset": 397122
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 37,
      "phaseCycles": 30432,
      "phaseMs": 0.020288,
      "workCycles": 20518434,
      "averageActiveTiles": 674.238761829653,
      "usefulCycles": 7243344.0,
      "estimatedWorkCycles": 20518434,
      "usefulUtilization": 0.35301641441057346,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.1616965694006309,
      "meanCycles": 9382.0,
      "p50Cycles": 9384,
      "p95Cycles": 9384,
      "maximumCycles": 9384,
      "firstOffset": 58038,
      "lastOffset": 13087170
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 21,
      "phaseCycles": 30066,
      "phaseMs": 0.020044,
      "workCycles": 18720000,
      "averageActiveTiles": 622.6302135302335,
      "usefulCycles": 11943360.0,
      "estimatedWorkCycles": 18720000,
      "usefulUtilization": 0.638,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.26986282352737023,
      "meanCycles": 30000.0,
      "p50Cycles": 30000,
      "p95Cycles": 30000,
      "maximumCycles": 30000,
      "firstOffset": 366540,
      "lastOffset": 396606
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 11,
      "phaseCycles": 27624,
      "phaseMs": 0.018416,
      "workCycles": 824940,
      "averageActiveTiles": 29.863162467419635,
      "usefulCycles": 531360.0,
      "estimatedWorkCycles": 824940,
      "usefulUtilization": 0.6441195723325333,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.013067559400143543,
      "meanCycles": 27498.0,
      "p50Cycles": 27498,
      "p95Cycles": 27498,
      "maximumCycles": 27498,
      "firstOffset": 365712,
      "lastOffset": 393336
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 13,
      "phaseCycles": 27096,
      "phaseMs": 0.018064,
      "workCycles": 649008,
      "averageActiveTiles": 23.952170062001773,
      "usefulCycles": 413424.0,
      "estimatedWorkCycles": 649008,
      "usefulUtilization": 0.637009096960284,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.010365319443909577,
      "meanCycles": 27042.0,
      "p50Cycles": 27042,
      "p95Cycles": 27042,
      "maximumCycles": 27042,
      "firstOffset": 365772,
      "lastOffset": 392868
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 32,
      "phaseCycles": 27024,
      "phaseMs": 0.018016,
      "workCycles": 32666490,
      "averageActiveTiles": 1208.795515097691,
      "usefulCycles": 23628960.0,
      "estimatedWorkCycles": 32666490,
      "usefulUtilization": 0.7233394221417728,
      "workLaneOccupancy": 0.9962962962962963,
      "deviceUsefulUtilization": 0.5940009846320179,
      "meanCycles": 26886.0,
      "p50Cycles": 26886,
      "p95Cycles": 26886,
      "maximumCycles": 26886,
      "firstOffset": 487686,
      "lastOffset": 514710
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 19,
      "phaseCycles": 26820,
      "phaseMs": 0.01788,
      "workCycles": 6508512,
      "averageActiveTiles": 242.67382550335572,
      "usefulCycles": 4609584.0,
      "estimatedWorkCycles": 6508512,
      "usefulUtilization": 0.7082393026240099,
      "workLaneOccupancy": 0.9879938271604938,
      "deviceUsefulUtilization": 0.11676028596440036,
      "meanCycles": 26784.0,
      "p50Cycles": 26784,
      "p95Cycles": 26784,
      "maximumCycles": 26784,
      "firstOffset": 487764,
      "lastOffset": 514584
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 15364,
      "tileCount": 1472,
      "phaseCount": 213,
      "phaseCycles": 20850,
      "phaseMs": 0.0139,
      "workCycles": 20358648,
      "averageActiveTiles": 976.4339568345324,
      "usefulCycles": 3096104.0,
      "estimatedWorkCycles": 20358648,
      "usefulUtilization": 0.15207807512561738,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.10087920967573767,
      "meanCycles": 1325.0877375683415,
      "p50Cycles": 1074,
      "p95Cycles": 2658,
      "maximumCycles": 2658,
      "firstOffset": 45570,
      "lastOffset": 13290834
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 896,
      "tileCount": 896,
      "phaseCount": 16,
      "phaseCycles": 20232,
      "phaseMs": 0.013488,
      "workCycles": 14859264,
      "averageActiveTiles": 734.4436536180308,
      "usefulCycles": 8358714.0,
      "estimatedWorkCycles": 14859264,
      "usefulUtilization": 0.5625254386758322,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.28066796089277424,
      "meanCycles": 16584.0,
      "p50Cycles": 16584,
      "p95Cycles": 16584,
      "maximumCycles": 16584,
      "firstOffset": 240060,
      "lastOffset": 260292
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 224,
      "tileCount": 224,
      "phaseCount": 6,
      "phaseCycles": 20082,
      "phaseMs": 0.013388,
      "workCycles": 3116736,
      "averageActiveTiles": 155.20047804003585,
      "usefulCycles": 1834560.0,
      "estimatedWorkCycles": 3116736,
      "usefulUtilization": 0.5886157826649417,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.06206076824151414,
      "meanCycles": 13914.0,
      "p50Cycles": 13914,
      "p95Cycles": 13914,
      "maximumCycles": 13914,
      "firstOffset": 175830,
      "lastOffset": 195912
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r182_p182_c12_p64",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 1,
      "phaseCycles": 19980,
      "phaseMs": 0.01332,
      "workCycles": 1436400,
      "averageActiveTiles": 71.89189189189189,
      "usefulCycles": 78624.0,
      "estimatedWorkCycles": 1436400,
      "usefulUtilization": 0.05473684210526316,
      "workLaneOccupancy": 0.1875,
      "deviceUsefulUtilization": 0.002673325499412456,
      "meanCycles": 19950.0,
      "p50Cycles": 19950,
      "p95Cycles": 19950,
      "maximumCycles": 19950,
      "firstOffset": 3540,
      "lastOffset": 23520
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r184_p184_c12_p64",
      "dimensions": {},
      "sampleCount": 36,
      "tileCount": 36,
      "phaseCount": 1,
      "phaseCycles": 19968,
      "phaseMs": 0.013312,
      "workCycles": 718200,
      "averageActiveTiles": 35.96754807692308,
      "usefulCycles": 39744.0,
      "estimatedWorkCycles": 718200,
      "usefulUtilization": 0.05533834586466165,
      "workLaneOccupancy": 0.1875,
      "deviceUsefulUtilization": 0.0013521634615384615,
      "meanCycles": 19950.0,
      "p50Cycles": 19950,
      "p95Cycles": 19950,
      "maximumCycles": 19950,
      "firstOffset": 3522,
      "lastOffset": 23490
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r181_p182_c12_p64",
      "dimensions": {},
      "sampleCount": 36,
      "tileCount": 36,
      "phaseCount": 1,
      "phaseCycles": 19968,
      "phaseMs": 0.013312,
      "workCycles": 718200,
      "averageActiveTiles": 35.96754807692308,
      "usefulCycles": 39096.0,
      "estimatedWorkCycles": 718200,
      "usefulUtilization": 0.05443609022556391,
      "workLaneOccupancy": 0.18646978021978022,
      "deviceUsefulUtilization": 0.0013301173181438127,
      "meanCycles": 19950.0,
      "p50Cycles": 19950,
      "p95Cycles": 19950,
      "maximumCycles": 19950,
      "firstOffset": 3606,
      "lastOffset": 23574
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 4,
      "phaseCycles": 18966,
      "phaseMs": 0.012644,
      "workCycles": 18294984,
      "averageActiveTiles": 964.6200569440051,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18294984,
      "usefulUtilization": 0.8296674104771012,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5436914570237816,
      "meanCycles": 18822.0,
      "p50Cycles": 18822,
      "p95Cycles": 18822,
      "maximumCycles": 18822,
      "firstOffset": 96840,
      "lastOffset": 115806
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 1,
      "phaseCycles": 18762,
      "phaseMs": 0.012508,
      "workCycles": 9083340,
      "averageActiveTiles": 484.134953629677,
      "usefulCycles": 7496064.0,
      "estimatedWorkCycles": 9083340,
      "usefulUtilization": 0.8252541466024612,
      "workLaneOccupancy": 0.9958677685950413,
      "deviceUsefulUtilization": 0.27142281114000083,
      "meanCycles": 18690.0,
      "p50Cycles": 18690,
      "p95Cycles": 18690,
      "maximumCycles": 18690,
      "firstOffset": 96912,
      "lastOffset": 115674
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 144,
      "tileCount": 144,
      "phaseCount": 4,
      "phaseCycles": 18306,
      "phaseMs": 0.012204,
      "workCycles": 1617408,
      "averageActiveTiles": 88.35398230088495,
      "usefulCycles": 942264.0,
      "estimatedWorkCycles": 1617408,
      "usefulUtilization": 0.5825765669515669,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.034968043264503444,
      "meanCycles": 11232.0,
      "p50Cycles": 11232,
      "p95Cycles": 11232,
      "maximumCycles": 11232,
      "firstOffset": 175302,
      "lastOffset": 193608
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 128,
      "tileCount": 128,
      "phaseCount": 9,
      "phaseCycles": 18024,
      "phaseMs": 0.012016,
      "workCycles": 2128896,
      "averageActiveTiles": 118.11451398135819,
      "usefulCycles": 1207224.0,
      "estimatedWorkCycles": 2128896,
      "usefulUtilization": 0.5670657467532467,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.04550183089214381,
      "meanCycles": 16632.0,
      "p50Cycles": 16632,
      "p95Cycles": 16632,
      "maximumCycles": 16632,
      "firstOffset": 242310,
      "lastOffset": 260334
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 1008,
      "tileCount": 1008,
      "phaseCount": 13,
      "phaseCycles": 17970,
      "phaseMs": 0.01198,
      "workCycles": 11297664,
      "averageActiveTiles": 628.6958263772955,
      "usefulCycles": 6524154.0,
      "estimatedWorkCycles": 11297664,
      "usefulUtilization": 0.5774781406138473,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.24664272882340132,
      "meanCycles": 11208.0,
      "p50Cycles": 11208,
      "p95Cycles": 11208,
      "maximumCycles": 11208,
      "firstOffset": 175254,
      "lastOffset": 193224
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 37,
      "phaseCycles": 17490,
      "phaseMs": 0.01166,
      "workCycles": 4282080,
      "averageActiveTiles": 244.83018867924528,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4282080,
      "usefulUtilization": 0.21043978627209206,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.03500136724090785,
      "meanCycles": 2433.0,
      "p50Cycles": 2430,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 160488,
      "lastOffset": 13193898
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 8,
      "phaseCycles": 16650,
      "phaseMs": 0.0111,
      "workCycles": 13369860,
      "averageActiveTiles": 802.9945945945946,
      "usefulCycles": 8501760.0,
      "estimatedWorkCycles": 13369860,
      "usefulUtilization": 0.6358899794014298,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.34688601645123385,
      "meanCycles": 16506.0,
      "p50Cycles": 16506,
      "p95Cycles": 16506,
      "maximumCycles": 16506,
      "firstOffset": 13114986,
      "lastOffset": 13131636
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 2496,
      "tileCount": 1472,
      "phaseCount": 90,
      "phaseCycles": 16464,
      "phaseMs": 0.010976,
      "workCycles": 19121568,
      "averageActiveTiles": 1161.4169096209912,
      "usefulCycles": 9425760.0,
      "estimatedWorkCycles": 19121568,
      "usefulUtilization": 0.4929386544032372,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.38893158194955,
      "meanCycles": 7660.884615384615,
      "p50Cycles": 12396,
      "p95Cycles": 12924,
      "maximumCycles": 13410,
      "firstOffset": 429780,
      "lastOffset": 13268778
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r80_r82",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 7,
      "phaseCycles": 16314,
      "phaseMs": 0.010876,
      "workCycles": 10528704,
      "averageActiveTiles": 645.3784479588084,
      "usefulCycles": 6614784.0,
      "estimatedWorkCycles": 10528704,
      "usefulUtilization": 0.6282619399310685,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.27545293186433634,
      "meanCycles": 16248.0,
      "p50Cycles": 16248,
      "p95Cycles": 16248,
      "maximumCycles": 16248,
      "firstOffset": 13115058,
      "lastOffset": 13131372
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 3645,
      "tileCount": 1246,
      "phaseCount": 152,
      "phaseCycles": 15018,
      "phaseMs": 0.010012,
      "workCycles": 745002,
      "averageActiveTiles": 49.607271274470634,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 204.39012345679012,
      "p50Cycles": 150,
      "p95Cycles": 294,
      "maximumCycles": 336,
      "firstOffset": 42174,
      "lastOffset": 13291890
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 7,
      "phaseCycles": 14394,
      "phaseMs": 0.009596,
      "workCycles": 446208,
      "averageActiveTiles": 30.999583159649855,
      "usefulCycles": 264960.0,
      "estimatedWorkCycles": 446208,
      "usefulUtilization": 0.5938037865748709,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.012505210504376824,
      "meanCycles": 13944.0,
      "p50Cycles": 13944,
      "p95Cycles": 13944,
      "maximumCycles": 13944,
      "firstOffset": 181914,
      "lastOffset": 196308
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 26,
      "phaseCycles": 14184,
      "phaseMs": 0.009456,
      "workCycles": 314400,
      "averageActiveTiles": 22.165820642978005,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 314400,
      "usefulUtilization": 0.10178117048346055,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.001532651609897251,
      "meanCycles": 1965.0,
      "p50Cycles": 1962,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 160716,
      "lastOffset": 13193154
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k736_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 22,
      "phaseCycles": 12888,
      "phaseMs": 0.008592,
      "workCycles": 671520,
      "averageActiveTiles": 52.10428305400372,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 671520,
      "usefulUtilization": 0.01954074338813438,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0006916823536555744,
      "meanCycles": 8394.0,
      "p50Cycles": 8394,
      "p95Cycles": 8394,
      "maximumCycles": 8394,
      "firstOffset": 13218228,
      "lastOffset": 13231116
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 176,
      "tileCount": 176,
      "phaseCount": 13,
      "phaseCycles": 12276,
      "phaseMs": 0.008184,
      "workCycles": 2072928,
      "averageActiveTiles": 168.86021505376345,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 2072928,
      "usefulUtilization": 0.21735438953982,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.024933769674302633,
      "meanCycles": 11778.0,
      "p50Cycles": 11778,
      "p95Cycles": 11778,
      "maximumCycles": 11778,
      "firstOffset": 13179192,
      "lastOffset": 13191468
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r182_p182_c64_p64",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 1,
      "phaseCycles": 12060,
      "phaseMs": 0.00804,
      "workCycles": 7383312,
      "averageActiveTiles": 612.2149253731343,
      "usefulCycles": 3773952.0,
      "estimatedWorkCycles": 7383312,
      "usefulUtilization": 0.5111462173073548,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2125892277741726,
      "meanCycles": 11394.0,
      "p50Cycles": 11394,
      "p95Cycles": 11394,
      "maximumCycles": 11394,
      "firstOffset": 6912,
      "lastOffset": 18972
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r184_p184_c64_p64",
      "dimensions": {},
      "sampleCount": 324,
      "tileCount": 324,
      "phaseCount": 2,
      "phaseCycles": 11988,
      "phaseMs": 0.007992,
      "workCycles": 3691656,
      "averageActiveTiles": 307.94594594594594,
      "usefulCycles": 1907712.0,
      "estimatedWorkCycles": 3691656,
      "usefulUtilization": 0.5167632087063366,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.10810810810810811,
      "meanCycles": 11394.0,
      "p50Cycles": 11394,
      "p95Cycles": 11394,
      "maximumCycles": 11394,
      "firstOffset": 6942,
      "lastOffset": 18930
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r181_p182_c64_p64",
      "dimensions": {},
      "sampleCount": 324,
      "tileCount": 324,
      "phaseCount": 1,
      "phaseCycles": 11868,
      "phaseMs": 0.007912,
      "workCycles": 3759696,
      "averageActiveTiles": 316.7927199191102,
      "usefulCycles": 1876608.0,
      "estimatedWorkCycles": 3759696,
      "usefulUtilization": 0.4991382281971734,
      "workLaneOccupancy": 0.9945054945054945,
      "deviceUsefulUtilization": 0.10742075878137776,
      "meanCycles": 11604.0,
      "p50Cycles": 11604,
      "p95Cycles": 11604,
      "maximumCycles": 11604,
      "firstOffset": 6870,
      "lastOffset": 18738
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 6,
      "phaseCycles": 10956,
      "phaseMs": 0.007304,
      "workCycles": 146592,
      "averageActiveTiles": 13.380065717415116,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 146592,
      "usefulUtilization": 0.10914647456887143,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0009921107354318459,
      "meanCycles": 9162.0,
      "p50Cycles": 9162,
      "p95Cycles": 9162,
      "maximumCycles": 9162,
      "firstOffset": 13177668,
      "lastOffset": 13188624
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 36,
      "phaseCycles": 8778,
      "phaseMs": 0.005852,
      "workCycles": 4406112,
      "averageActiveTiles": 501.9494190020506,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2961.0967741935483,
      "p50Cycles": 2976,
      "p95Cycles": 2976,
      "maximumCycles": 2976,
      "firstOffset": 266958,
      "lastOffset": 13233324
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 25,
      "phaseCycles": 8178,
      "phaseMs": 0.005452,
      "workCycles": 927360,
      "averageActiveTiles": 113.3969185619956,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 927360,
      "usefulUtilization": 0.014149844720496895,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0010900467319531724,
      "meanCycles": 1260.0,
      "p50Cycles": 1260,
      "p95Cycles": 1260,
      "maximumCycles": 1260,
      "firstOffset": 13195812,
      "lastOffset": 13203990
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 30,
      "phaseCycles": 7386,
      "phaseMs": 0.004924,
      "workCycles": 10437120,
      "averageActiveTiles": 1413.0950446791226,
      "usefulCycles": 7558272.0,
      "estimatedWorkCycles": 10437120,
      "usefulUtilization": 0.7241721854304636,
      "workLaneOccupancy": 0.9986301369863013,
      "deviceUsefulUtilization": 0.6951930208738035,
      "meanCycles": 7248.0,
      "p50Cycles": 7248,
      "p95Cycles": 7248,
      "maximumCycles": 7248,
      "firstOffset": 290472,
      "lastOffset": 297858
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80",
      "dimensions": {},
      "sampleCount": 1360,
      "tileCount": 1360,
      "phaseCount": 14,
      "phaseCycles": 7092,
      "phaseMs": 0.004728,
      "workCycles": 1729920,
      "averageActiveTiles": 243.92554991539762,
      "usefulCycles": 391680.0,
      "estimatedWorkCycles": 1729920,
      "usefulUtilization": 0.22641509433962265,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.03751931141028471,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 155094,
      "lastOffset": 162606
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 880,
      "tileCount": 880,
      "phaseCount": 13,
      "phaseCycles": 5202,
      "phaseMs": 0.003468,
      "workCycles": 1356960,
      "averageActiveTiles": 260.8535178777393,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 1356960,
      "usefulUtilization": 0.3320363164721141,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.05884024539056885,
      "meanCycles": 1542.0,
      "p50Cycles": 1542,
      "p95Cycles": 1542,
      "maximumCycles": 1542,
      "firstOffset": 158946,
      "lastOffset": 164148
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 61,
      "phaseCycles": 4776,
      "phaseMs": 0.003184,
      "workCycles": 2324160,
      "averageActiveTiles": 486.63316582914575,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2324160,
      "usefulUtilization": 0.016666666666666666,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.00550988638846406,
      "meanCycles": 1614.0,
      "p50Cycles": 1614,
      "p95Cycles": 1614,
      "maximumCycles": 1614,
      "firstOffset": 13279956,
      "lastOffset": 13286172
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 7,
      "phaseCycles": 4296,
      "phaseMs": 0.002864,
      "workCycles": 142080,
      "averageActiveTiles": 33.072625698324025,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 142080,
      "usefulUtilization": 0.11261261261261261,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.002530159501254959,
      "meanCycles": 1776.0,
      "p50Cycles": 1776,
      "p95Cycles": 1776,
      "maximumCycles": 1776,
      "firstOffset": 159810,
      "lastOffset": 164106
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 1,
      "phaseCycles": 4128,
      "phaseMs": 0.002752,
      "workCycles": 4348080,
      "averageActiveTiles": 1053.3139534883721,
      "usefulCycles": 2884140.0,
      "estimatedWorkCycles": 4348080,
      "usefulUtilization": 0.6633134624937904,
      "workLaneOccupancy": 0.9170673076923077,
      "deviceUsefulUtilization": 0.4746449222699697,
      "meanCycles": 4026.0,
      "p50Cycles": 4026,
      "p95Cycles": 4026,
      "maximumCycles": 4026,
      "firstOffset": 31866,
      "lastOffset": 35994
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c32_r182_r184",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 4,
      "phaseCycles": 4098,
      "phaseMs": 0.002732,
      "workCycles": 1458000,
      "averageActiveTiles": 355.78330893118596,
      "usefulCycles": 973728.0,
      "estimatedWorkCycles": 1458000,
      "usefulUtilization": 0.6678518518518518,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.16142020497803808,
      "meanCycles": 4050.0,
      "p50Cycles": 4050,
      "p95Cycles": 4050,
      "maximumCycles": 4050,
      "firstOffset": 31812,
      "lastOffset": 35910
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c224_p224",
      "dimensions": {},
      "sampleCount": 1368,
      "tileCount": 1368,
      "phaseCount": 59,
      "phaseCycles": 4080,
      "phaseMs": 0.00272,
      "workCycles": 2766096,
      "averageActiveTiles": 677.964705882353,
      "usefulCycles": 153216.0,
      "estimatedWorkCycles": 2766096,
      "usefulUtilization": 0.05539070227497527,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.02551150895140665,
      "meanCycles": 2022.0,
      "p50Cycles": 2022,
      "p95Cycles": 2022,
      "maximumCycles": 2022,
      "firstOffset": 13272174,
      "lastOffset": 13276254
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 15,
      "phaseCycles": 2994,
      "phaseMs": 0.001996,
      "workCycles": 95232,
      "averageActiveTiles": 31.80761523046092,
      "usefulCycles": 2304.0,
      "estimatedWorkCycles": 95232,
      "usefulUtilization": 0.024193548387096774,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0005227846998344516,
      "meanCycles": 2976.0,
      "p50Cycles": 2976,
      "p95Cycles": 2976,
      "maximumCycles": 2976,
      "firstOffset": 13247820,
      "lastOffset": 13250814
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c48_p224",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 7,
      "phaseCycles": 2046,
      "phaseMs": 0.001364,
      "workCycles": 145584,
      "averageActiveTiles": 71.15542521994135,
      "usefulCycles": 1728.0,
      "estimatedWorkCycles": 145584,
      "usefulUtilization": 0.011869436201780416,
      "workLaneOccupancy": 0.10714285714285714,
      "deviceUsefulUtilization": 0.000573760040800714,
      "meanCycles": 2022.0,
      "p50Cycles": 2022,
      "p95Cycles": 2022,
      "maximumCycles": 2022,
      "firstOffset": 13274226,
      "lastOffset": 13276272
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 3,
      "phaseCycles": 2004,
      "phaseMs": 0.001336,
      "workCycles": 142464,
      "averageActiveTiles": 71.08982035928143,
      "usefulCycles": 28224.0,
      "estimatedWorkCycles": 142464,
      "usefulUtilization": 0.19811320754716982,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.009567820879979172,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 155244,
      "lastOffset": 157248
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 15,
      "phaseCycles": 1980,
      "phaseMs": 0.00132,
      "workCycles": 39168,
      "averageActiveTiles": 19.78181818181818,
      "usefulCycles": 2304.0,
      "estimatedWorkCycles": 39168,
      "usefulUtilization": 0.058823529411764705,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0007905138339920949,
      "meanCycles": 1224.0,
      "p50Cycles": 1224,
      "p95Cycles": 1224,
      "maximumCycles": 1224,
      "firstOffset": 13245426,
      "lastOffset": 13247406
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k32_c112_r2_r2",
      "dimensions": {},
      "sampleCount": 1044,
      "tileCount": 1044,
      "phaseCount": 51,
      "phaseCycles": 1752,
      "phaseMs": 0.001168,
      "workCycles": 1722600,
      "averageActiveTiles": 983.2191780821918,
      "usefulCycles": 29232.0,
      "estimatedWorkCycles": 1722600,
      "usefulUtilization": 0.01696969696969697,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.01133487194758785,
      "meanCycles": 1650.0,
      "p50Cycles": 1650,
      "p95Cycles": 1650,
      "maximumCycles": 1650,
      "firstOffset": 13256706,
      "lastOffset": 13258458
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 2592,
      "tileCount": 1296,
      "phaseCount": 55,
      "phaseCycles": 1692,
      "phaseMs": 0.001128,
      "workCycles": 1928448,
      "averageActiveTiles": 1139.7446808510638,
      "usefulCycles": 20736.0,
      "estimatedWorkCycles": 1928448,
      "usefulUtilization": 0.010752688172043012,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.008325624421831638,
      "meanCycles": 744.0,
      "p50Cycles": 744,
      "p95Cycles": 744,
      "maximumCycles": 744,
      "firstOffset": 13158948,
      "lastOffset": 13240938
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k32_c96_r2_r2",
      "dimensions": {},
      "sampleCount": 396,
      "tileCount": 396,
      "phaseCount": 41,
      "phaseCycles": 1620,
      "phaseMs": 0.00108,
      "workCycles": 582120,
      "averageActiveTiles": 359.3333333333333,
      "usefulCycles": 9504.0,
      "estimatedWorkCycles": 582120,
      "usefulUtilization": 0.0163265306122449,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.003985507246376811,
      "meanCycles": 1470.0,
      "p50Cycles": 1470,
      "p95Cycles": 1470,
      "maximumCycles": 1470,
      "firstOffset": 13256598,
      "lastOffset": 13258218
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p1_c72_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 5,
      "phaseCycles": 1614,
      "phaseMs": 0.001076,
      "workCycles": 12288,
      "averageActiveTiles": 7.613382899628252,
      "usefulCycles": 576.0,
      "estimatedWorkCycles": 12288,
      "usefulUtilization": 0.046875,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.00024244383384515922,
      "meanCycles": 768.0,
      "p50Cycles": 768,
      "p95Cycles": 768,
      "maximumCycles": 768,
      "firstOffset": 13169052,
      "lastOffset": 13171698
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c64_p64",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 41,
      "phaseCycles": 1410,
      "phaseMs": 0.00094,
      "workCycles": 1065312,
      "averageActiveTiles": 755.5404255319149,
      "usefulCycles": 41472.0,
      "estimatedWorkCycles": 1065312,
      "usefulUtilization": 0.038929440389294405,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.01998149861239593,
      "meanCycles": 822.0,
      "p50Cycles": 822,
      "p95Cycles": 822,
      "maximumCycles": 822,
      "firstOffset": 13235862,
      "lastOffset": 13237272
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c32_p32",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 54,
      "phaseCycles": 1170,
      "phaseMs": 0.00078,
      "workCycles": 838080,
      "averageActiveTiles": 716.3076923076923,
      "usefulCycles": 23040.0,
      "estimatedWorkCycles": 838080,
      "usefulUtilization": 0.027491408934707903,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.013377926421404682,
      "meanCycles": 582.0,
      "p50Cycles": 582,
      "p95Cycles": 582,
      "maximumCycles": 582,
      "firstOffset": 13252500,
      "lastOffset": 13253670
    }
  ],
  "samples": []
}
