{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 247416,
  "matchedSampleCount": 247416,
  "profileSpanCycles": 14374044,
  "profileSpanMs": 9.582696,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 11,
      "phaseCycles": 13552794,
      "phaseMs": 9.035196,
      "workCycles": 19945411500,
      "averageActiveTiles": 1471.6826286889625,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 13549871.942934783,
      "p50Cycles": 13549872,
      "p95Cycles": 13549890,
      "maximumCycles": 13549908,
      "firstOffset": 589722,
      "lastOffset": 14142516
    },
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 26254,
      "tileCount": 1457,
      "phaseCount": 70,
      "phaseCycles": 650406,
      "phaseMs": 0.433604,
      "workCycles": 138880782,
      "averageActiveTiles": 213.52936781026006,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5289.890378608974,
      "p50Cycles": 2250,
      "p95Cycles": 19674,
      "maximumCycles": 38640,
      "firstOffset": 0,
      "lastOffset": 14373318
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 91506,
      "tileCount": 1472,
      "phaseCount": 80,
      "phaseCycles": 584370,
      "phaseMs": 0.38958,
      "workCycles": 616903044,
      "averageActiveTiles": 1055.6719954823143,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 6741.66769392171,
      "p50Cycles": 5256,
      "p95Cycles": 18060,
      "maximumCycles": 32082,
      "firstOffset": 0,
      "lastOffset": 14373894
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 28990,
      "tileCount": 1472,
      "phaseCount": 235,
      "phaseCycles": 161310,
      "phaseMs": 0.10754,
      "workCycles": 23405646,
      "averageActiveTiles": 145.09730332899386,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 807.3696447050708,
      "p50Cycles": 576,
      "p95Cycles": 1572,
      "maximumCycles": 10578,
      "firstOffset": 44598,
      "lastOffset": 14370072
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 18588,
      "tileCount": 1472,
      "phaseCount": 141,
      "phaseCycles": 93378,
      "phaseMs": 0.062252,
      "workCycles": 6734382,
      "averageActiveTiles": 72.11957848743816,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 3981930,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 362.2972885732731,
      "p50Cycles": 306,
      "p95Cycles": 606,
      "maximumCycles": 5088,
      "firstOffset": 34464,
      "lastOffset": 14361042
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10440,
      "tileCount": 1472,
      "phaseCount": 157,
      "phaseCycles": 88326,
      "phaseMs": 0.058884,
      "workCycles": 42663456,
      "averageActiveTiles": 483.02262074587327,
      "usefulCycles": 23746848.0,
      "estimatedWorkCycles": 42663456,
      "usefulUtilization": 0.5566086347997686,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.18264576189590145,
      "meanCycles": 4086.537931034483,
      "p50Cycles": 3690,
      "p95Cycles": 7218,
      "maximumCycles": 7218,
      "firstOffset": 44892,
      "lastOffset": 14369802
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 4174,
      "tileCount": 1472,
      "phaseCount": 128,
      "phaseCycles": 71790,
      "phaseMs": 0.04786,
      "workCycles": 1814988,
      "averageActiveTiles": 25.281905557877142,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 434.83181600383324,
      "p50Cycles": 420,
      "p95Cycles": 720,
      "maximumCycles": 5430,
      "firstOffset": 0,
      "lastOffset": 14366706
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 11652,
      "tileCount": 1458,
      "phaseCount": 167,
      "phaseCycles": 69774,
      "phaseMs": 0.046516,
      "workCycles": 56439294,
      "averageActiveTiles": 808.8871786052111,
      "usefulCycles": 10457856.0,
      "estimatedWorkCycles": 56439294,
      "usefulUtilization": 0.18529388408012332,
      "workLaneOccupancy": 0.9750871045052666,
      "deviceUsefulUtilization": 0.10182190700161141,
      "meanCycles": 4843.743048403708,
      "p50Cycles": 4512,
      "p95Cycles": 9162,
      "maximumCycles": 9906,
      "firstOffset": 18276,
      "lastOffset": 14360424
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 28,
      "phaseCycles": 35196,
      "phaseMs": 0.023464,
      "workCycles": 34993632,
      "averageActiveTiles": 994.2502557108762,
      "usefulCycles": 15838080.0,
      "estimatedWorkCycles": 34993632,
      "usefulUtilization": 0.4525989185689556,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.3057042055174254,
      "meanCycles": 23517.225806451614,
      "p50Cycles": 23706,
      "p95Cycles": 23706,
      "maximumCycles": 23706,
      "firstOffset": 211068,
      "lastOffset": 14293578
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 29,
      "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": 387798,
      "lastOffset": 418446
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 38,
      "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": 65958,
      "lastOffset": 14157252
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 20,
      "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": 387864,
      "lastOffset": 417930
    },
    {
      "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": 387036,
      "lastOffset": 414660
    },
    {
      "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": 387096,
      "lastOffset": 414192
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k160_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 33,
      "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": 520716,
      "lastOffset": 547740
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 15364,
      "tileCount": 1472,
      "phaseCount": 213,
      "phaseCycles": 26916,
      "phaseMs": 0.017944,
      "workCycles": 20358648,
      "averageActiveTiles": 756.377173428444,
      "usefulCycles": 3096104.0,
      "estimatedWorkCycles": 20358648,
      "usefulUtilization": 0.15207807512561738,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07814428301898983,
      "meanCycles": 1325.0877375683415,
      "p50Cycles": 1074,
      "p95Cycles": 2658,
      "maximumCycles": 2658,
      "firstOffset": 53568,
      "lastOffset": 14372982
    },
    {
      "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": 520794,
      "lastOffset": 547614
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 37,
      "phaseCycles": 22200,
      "phaseMs": 0.0148,
      "workCycles": 4282080,
      "averageActiveTiles": 192.8864864864865,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4282080,
      "usefulUtilization": 0.21043978627209206,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.027575401488444968,
      "meanCycles": 2433.0,
      "p50Cycles": 2430,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 166734,
      "lastOffset": 14273664
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 3645,
      "tileCount": 1246,
      "phaseCount": 152,
      "phaseCycles": 20496,
      "phaseMs": 0.013664,
      "workCycles": 745002,
      "averageActiveTiles": 36.348653395784545,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 204.39012345679012,
      "p50Cycles": 150,
      "p95Cycles": 294,
      "maximumCycles": 336,
      "firstOffset": 47994,
      "lastOffset": 14374044
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 224,
      "tileCount": 224,
      "phaseCount": 7,
      "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": 184224,
      "lastOffset": 204306
    },
    {
      "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": 9060,
      "lastOffset": 29040
    },
    {
      "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": 9042,
      "lastOffset": 29010
    },
    {
      "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": 9060,
      "lastOffset": 29028
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c64_r242_r244",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 5,
      "phaseCycles": 18960,
      "phaseMs": 0.01264,
      "workCycles": 18294984,
      "averageActiveTiles": 964.9253164556962,
      "usefulCycles": 15178752.0,
      "estimatedWorkCycles": 18294984,
      "usefulUtilization": 0.8296674104771012,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.5438635112823336,
      "meanCycles": 18822.0,
      "p50Cycles": 18822,
      "p95Cycles": 18822,
      "maximumCycles": 18822,
      "firstOffset": 102150,
      "lastOffset": 121110
    },
    {
      "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": 102222,
      "lastOffset": 120984
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 896,
      "tileCount": 896,
      "phaseCount": 16,
      "phaseCycles": 18486,
      "phaseMs": 0.012324,
      "workCycles": 14859264,
      "averageActiveTiles": 803.8117494320026,
      "usefulCycles": 8358714.0,
      "estimatedWorkCycles": 14859264,
      "usefulUtilization": 0.5625254386758322,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.30717700880572374,
      "meanCycles": 16584.0,
      "p50Cycles": 16584,
      "p95Cycles": 16584,
      "maximumCycles": 16584,
      "firstOffset": 256704,
      "lastOffset": 275190
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 144,
      "tileCount": 144,
      "phaseCount": 4,
      "phaseCycles": 18300,
      "phaseMs": 0.0122,
      "workCycles": 1617408,
      "averageActiveTiles": 88.38295081967213,
      "usefulCycles": 942264.0,
      "estimatedWorkCycles": 1617408,
      "usefulUtilization": 0.5825765669515669,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.03497950819672131,
      "meanCycles": 11232.0,
      "p50Cycles": 11232,
      "p95Cycles": 11232,
      "maximumCycles": 11232,
      "firstOffset": 183696,
      "lastOffset": 201996
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 128,
      "tileCount": 128,
      "phaseCount": 9,
      "phaseCycles": 18000,
      "phaseMs": 0.012,
      "workCycles": 2128896,
      "averageActiveTiles": 118.272,
      "usefulCycles": 1207224.0,
      "estimatedWorkCycles": 2128896,
      "usefulUtilization": 0.5670657467532467,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.0455625,
      "meanCycles": 16632.0,
      "p50Cycles": 16632,
      "p95Cycles": 16632,
      "maximumCycles": 16632,
      "firstOffset": 256938,
      "lastOffset": 274938
    },
    {
      "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": 183642,
      "lastOffset": 201612
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 26,
      "phaseCycles": 17592,
      "phaseMs": 0.011728,
      "workCycles": 314400,
      "averageActiveTiles": 17.871759890859483,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 314400,
      "usefulUtilization": 0.10178117048346055,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0012357395654151097,
      "meanCycles": 1965.0,
      "p50Cycles": 1962,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 166980,
      "lastOffset": 14273028
    },
    {
      "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": 14186088,
      "lastOffset": 14202738
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 2496,
      "tileCount": 1472,
      "phaseCount": 90,
      "phaseCycles": 16554,
      "phaseMs": 0.011036,
      "workCycles": 19121568,
      "averageActiveTiles": 1155.102573396158,
      "usefulCycles": 9425760.0,
      "estimatedWorkCycles": 19121568,
      "usefulUtilization": 0.4929386544032372,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.38681705721984966,
      "meanCycles": 7660.884615384615,
      "p50Cycles": 12396,
      "p95Cycles": 12924,
      "maximumCycles": 13410,
      "firstOffset": 456606,
      "lastOffset": 14350926
    },
    {
      "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": 14186160,
      "lastOffset": 14202474
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 176,
      "tileCount": 176,
      "phaseCount": 13,
      "phaseCycles": 14814,
      "phaseMs": 0.009876,
      "workCycles": 2072928,
      "averageActiveTiles": 139.93033616848928,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 2072928,
      "usefulUtilization": 0.21735438953982,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.020662005975546047,
      "meanCycles": 11778.0,
      "p50Cycles": 11778,
      "p95Cycles": 11778,
      "maximumCycles": 11778,
      "firstOffset": 14256420,
      "lastOffset": 14271234
    },
    {
      "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": 190308,
      "lastOffset": 204702
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r181_p182_c64_p64",
      "dimensions": {},
      "sampleCount": 324,
      "tileCount": 324,
      "phaseCount": 1,
      "phaseCycles": 13650,
      "phaseMs": 0.0091,
      "workCycles": 3759696,
      "averageActiveTiles": 275.4356043956044,
      "usefulCycles": 1876608.0,
      "estimatedWorkCycles": 3759696,
      "usefulUtilization": 0.4991382281971734,
      "workLaneOccupancy": 0.9945054945054945,
      "deviceUsefulUtilization": 0.09339703774486383,
      "meanCycles": 11604.0,
      "p50Cycles": 11604,
      "p95Cycles": 11604,
      "maximumCycles": 11604,
      "firstOffset": 6978,
      "lastOffset": 20628
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r182_p182_c64_p64",
      "dimensions": {},
      "sampleCount": 648,
      "tileCount": 648,
      "phaseCount": 1,
      "phaseCycles": 13542,
      "phaseMs": 0.009028,
      "workCycles": 7383312,
      "averageActiveTiles": 545.2157731501994,
      "usefulCycles": 3773952.0,
      "estimatedWorkCycles": 7383312,
      "usefulUtilization": 0.5111462173073548,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.1893240353682264,
      "meanCycles": 11394.0,
      "p50Cycles": 11394,
      "p95Cycles": 11394,
      "maximumCycles": 11394,
      "firstOffset": 6882,
      "lastOffset": 20424
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r184_p184_c64_p64",
      "dimensions": {},
      "sampleCount": 324,
      "tileCount": 324,
      "phaseCount": 2,
      "phaseCycles": 13488,
      "phaseMs": 0.008992,
      "workCycles": 3691656,
      "averageActiveTiles": 273.69928825622776,
      "usefulCycles": 1907712.0,
      "estimatedWorkCycles": 3691656,
      "usefulUtilization": 0.5167632087063366,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.09608540925266904,
      "meanCycles": 11394.0,
      "p50Cycles": 11394,
      "p95Cycles": 11394,
      "maximumCycles": 11394,
      "firstOffset": 6942,
      "lastOffset": 20430
    },
    {
      "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": 14300358,
      "lastOffset": 14313246
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 7,
      "phaseCycles": 11880,
      "phaseMs": 0.00792,
      "workCycles": 146592,
      "averageActiveTiles": 12.33939393939394,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 146592,
      "usefulUtilization": 0.10914647456887143,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0009149465671204802,
      "meanCycles": 9162.0,
      "p50Cycles": 9162,
      "p95Cycles": 9162,
      "maximumCycles": 9162,
      "firstOffset": 14256438,
      "lastOffset": 14268318
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 36,
      "phaseCycles": 9816,
      "phaseMs": 0.006544,
      "workCycles": 4406112,
      "averageActiveTiles": 448.8704156479218,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2961.0967741935483,
      "p50Cycles": 2976,
      "p95Cycles": 2976,
      "maximumCycles": 2976,
      "firstOffset": 282786,
      "lastOffset": 14315454
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80",
      "dimensions": {},
      "sampleCount": 1360,
      "tileCount": 1360,
      "phaseCount": 14,
      "phaseCycles": 8538,
      "phaseMs": 0.005692,
      "workCycles": 1729920,
      "averageActiveTiles": 202.61419536191144,
      "usefulCycles": 391680.0,
      "estimatedWorkCycles": 1729920,
      "usefulUtilization": 0.22641509433962265,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.031165021846069237,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 162750,
      "lastOffset": 171288
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 880,
      "tileCount": 880,
      "phaseCount": 13,
      "phaseCycles": 7638,
      "phaseMs": 0.005092,
      "workCycles": 1356960,
      "averageActiveTiles": 177.65907305577377,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 1356960,
      "usefulUtilization": 0.3320363164721141,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.04007422840033244,
      "meanCycles": 1542.0,
      "p50Cycles": 1542,
      "p95Cycles": 1542,
      "maximumCycles": 1542,
      "firstOffset": 165192,
      "lastOffset": 172830
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c48_r146_r146",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 31,
      "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": 308172,
      "lastOffset": 315558
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 7,
      "phaseCycles": 7260,
      "phaseMs": 0.00484,
      "workCycles": 142080,
      "averageActiveTiles": 19.570247933884296,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 142080,
      "usefulUtilization": 0.11261261261261261,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0014971852916516948,
      "meanCycles": 1776.0,
      "p50Cycles": 1776,
      "p95Cycles": 1776,
      "maximumCycles": 1776,
      "firstOffset": 165204,
      "lastOffset": 172464
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k224_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 60,
      "phaseCycles": 4704,
      "phaseMs": 0.003136,
      "workCycles": 2324160,
      "averageActiveTiles": 494.0816326530612,
      "usefulCycles": 38736.0,
      "estimatedWorkCycles": 2324160,
      "usefulUtilization": 0.016666666666666666,
      "workLaneOccupancy": 0.48035714285714287,
      "deviceUsefulUtilization": 0.005594221384205856,
      "meanCycles": 1614.0,
      "p50Cycles": 1614,
      "p95Cycles": 1614,
      "maximumCycles": 1614,
      "firstOffset": 14362098,
      "lastOffset": 14368320
    },
    {
      "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": 37404,
      "lastOffset": 41532
    },
    {
      "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": 37350,
      "lastOffset": 41448
    },
    {
      "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": 14354322,
      "lastOffset": 14358402
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 25,
      "phaseCycles": 3768,
      "phaseMs": 0.002512,
      "workCycles": 927360,
      "averageActiveTiles": 246.11464968152868,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 927360,
      "usefulUtilization": 0.014149844720496895,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0023658179867072832,
      "meanCycles": 1260.0,
      "p50Cycles": 1260,
      "p95Cycles": 1260,
      "maximumCycles": 1260,
      "firstOffset": 14282352,
      "lastOffset": 14286120
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 3,
      "phaseCycles": 3360,
      "phaseMs": 0.00224,
      "workCycles": 142464,
      "averageActiveTiles": 42.4,
      "usefulCycles": 28224.0,
      "estimatedWorkCycles": 142464,
      "usefulUtilization": 0.19811320754716982,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.005706521739130435,
      "meanCycles": 1272.0,
      "p50Cycles": 1272,
      "p95Cycles": 1272,
      "maximumCycles": 1272,
      "firstOffset": 162756,
      "lastOffset": 166116
    },
    {
      "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": 14329956,
      "lastOffset": 14332950
    },
    {
      "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": 14356374,
      "lastOffset": 14358420
    },
    {
      "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": 14327562,
      "lastOffset": 14329542
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k32_c112_r2_r2",
      "dimensions": {},
      "sampleCount": 1044,
      "tileCount": 1044,
      "phaseCount": 50,
      "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": 14338842,
      "lastOffset": 14340594
    },
    {
      "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": 14235732,
      "lastOffset": 14323068
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p1_c72_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 5,
      "phaseCycles": 1638,
      "phaseMs": 0.001092,
      "workCycles": 12288,
      "averageActiveTiles": 7.501831501831502,
      "usefulCycles": 576.0,
      "estimatedWorkCycles": 12288,
      "usefulUtilization": 0.046875,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.00023889154323936931,
      "meanCycles": 768.0,
      "p50Cycles": 768,
      "p95Cycles": 768,
      "maximumCycles": 768,
      "firstOffset": 14247222,
      "lastOffset": 14249340
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k32_c96_r2_r2",
      "dimensions": {},
      "sampleCount": 396,
      "tileCount": 396,
      "phaseCount": 40,
      "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": 14338734,
      "lastOffset": 14340354
    },
    {
      "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": 14317992,
      "lastOffset": 14319402
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c32_p32",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 53,
      "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": 14334636,
      "lastOffset": 14335806
    }
  ],
  "samples": []
}
