{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 241056,
  "matchedSampleCount": 129720,
  "profileSpanCycles": 16985046,
  "profileSpanMs": 11.323364,
  "atOffset": null,
  "groups": [
    {
      "name": "repeat-remainder",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 9,
      "phaseCycles": 16056654,
      "phaseMs": 10.704436,
      "workCycles": 23629763826,
      "averageActiveTiles": 1471.6493128643115,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 16052828.686141305,
      "p50Cycles": 16052856,
      "p95Cycles": 16052928,
      "maximumCycles": 16052952,
      "firstOffset": 721830,
      "lastOffset": 16778484
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 22066,
      "tileCount": 1472,
      "phaseCount": 145,
      "phaseCycles": 116364,
      "phaseMs": 0.077576,
      "workCycles": 81713442,
      "averageActiveTiles": 702.2226977415695,
      "usefulCycles": 26314664.0,
      "estimatedWorkCycles": 81713442,
      "usefulUtilization": 0.3220359264758422,
      "workLaneOccupancy": 0.9895200907294152,
      "deviceUsefulUtilization": 0.15362835398068728,
      "meanCycles": 3703.1379497870025,
      "p50Cycles": 2154,
      "p95Cycles": 8748,
      "maximumCycles": 10386,
      "firstOffset": 24612,
      "lastOffset": 16980204
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 36202,
      "tileCount": 1472,
      "phaseCount": 184,
      "phaseCycles": 99942,
      "phaseMs": 0.066628,
      "workCycles": 12926658,
      "averageActiveTiles": 129.34159812691362,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 7751250,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 357.07027236064306,
      "p50Cycles": 306,
      "p95Cycles": 744,
      "maximumCycles": 3282,
      "firstOffset": 3048,
      "lastOffset": 16972662
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c192_r72_r76",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 14,
      "phaseCycles": 81498,
      "phaseMs": 0.054332,
      "workCycles": 23915520,
      "averageActiveTiles": 293.4491643966723,
      "usefulCycles": 14719680.0,
      "estimatedWorkCycles": 23915520,
      "usefulUtilization": 0.6154865125240848,
      "workLaneOccupancy": 0.9340277777777778,
      "deviceUsefulUtilization": 0.122699730161423,
      "meanCycles": 66432.0,
      "p50Cycles": 66432,
      "p95Cycles": 66432,
      "maximumCycles": 66432,
      "firstOffset": 621756,
      "lastOffset": 703254
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c192_r72_r76",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 38,
      "phaseCycles": 79902,
      "phaseMs": 0.053268,
      "workCycles": 70035840,
      "averageActiveTiles": 876.5217391304348,
      "usefulCycles": 41757408.0,
      "estimatedWorkCycles": 70035840,
      "usefulUtilization": 0.5962291306850892,
      "workLaneOccupancy": 0.9322980967078189,
      "deviceUsefulUtilization": 0.35503246912250114,
      "meanCycles": 64848.0,
      "p50Cycles": 64848,
      "p95Cycles": 64848,
      "maximumCycles": 64848,
      "firstOffset": 621774,
      "lastOffset": 701676
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 9598,
      "tileCount": 1472,
      "phaseCount": 109,
      "phaseCycles": 54480,
      "phaseMs": 0.03632,
      "workCycles": 36810930,
      "averageActiveTiles": 675.6778634361234,
      "usefulCycles": 1972820.0,
      "estimatedWorkCycles": 34682022,
      "usefulUtilization": 0.056883073311008225,
      "workLaneOccupancy": 0.9384120980109366,
      "deviceUsefulUtilization": 0.024600421774245037,
      "meanCycles": 3835.2708897687016,
      "p50Cycles": 3642,
      "p95Cycles": 10950,
      "maximumCycles": 10950,
      "firstOffset": 0,
      "lastOffset": 16971816
    },
    {
      "name": "attention_softmax_d72_p768_tail_f8_sm4",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 15,
      "phaseCycles": 53706,
      "phaseMs": 0.035804,
      "workCycles": 71468532,
      "averageActiveTiles": 1330.736454027483,
      "usefulCycles": 33499008.0,
      "estimatedWorkCycles": 71468532,
      "usefulUtilization": 0.46872388535978327,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.4237418214141728,
      "meanCycles": 48551.99184782609,
      "p50Cycles": 48552,
      "p95Cycles": 48552,
      "maximumCycles": 48552,
      "firstOffset": 271566,
      "lastOffset": 325272
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c80_r160_r164",
      "dimensions": {},
      "sampleCount": 795,
      "tileCount": 795,
      "phaseCount": 12,
      "phaseCycles": 49878,
      "phaseMs": 0.033252,
      "workCycles": 38951820,
      "averageActiveTiles": 780.9418982316853,
      "usefulCycles": 31291200.0,
      "estimatedWorkCycles": 38951820,
      "usefulUtilization": 0.8033308841538085,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.42619208259457425,
      "meanCycles": 48996.0,
      "p50Cycles": 48996,
      "p95Cycles": 48996,
      "maximumCycles": 48996,
      "firstOffset": 485808,
      "lastOffset": 535686
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 12227,
      "tileCount": 1472,
      "phaseCount": 120,
      "phaseCycles": 49146,
      "phaseMs": 0.032764,
      "workCycles": 16046712,
      "averageActiveTiles": 326.51104871200096,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1312.3997709986097,
      "p50Cycles": 834,
      "p95Cycles": 3018,
      "maximumCycles": 10578,
      "firstOffset": 94686,
      "lastOffset": 16974846
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c80_r160_r164",
      "dimensions": {},
      "sampleCount": 636,
      "tileCount": 636,
      "phaseCount": 15,
      "phaseCycles": 48336,
      "phaseMs": 0.032224,
      "workCycles": 30577608,
      "averageActiveTiles": 632.6052631578947,
      "usefulCycles": 24346080.0,
      "estimatedWorkCycles": 30577608,
      "usefulUtilization": 0.7962061649819044,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.34217677345537756,
      "meanCycles": 48078.0,
      "p50Cycles": 48078,
      "p95Cycles": 48078,
      "maximumCycles": 48078,
      "firstOffset": 485232,
      "lastOffset": 533568
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k288_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 1,
      "phaseCycles": 41034,
      "phaseMs": 0.027356,
      "workCycles": 29350080,
      "averageActiveTiles": 715.2624652727006,
      "usefulCycles": 23016960.0,
      "estimatedWorkCycles": 29350080,
      "usefulUtilization": 0.7842213717986458,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.38106257589146647,
      "meanCycles": 40764.0,
      "p50Cycles": 40764,
      "p95Cycles": 40764,
      "maximumCycles": 40764,
      "firstOffset": 131652,
      "lastOffset": 172686
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k288_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 3,
      "phaseCycles": 40434,
      "phaseMs": 0.026956,
      "workCycles": 28922400,
      "averageActiveTiles": 715.2990057872088,
      "usefulCycles": 22332672.0,
      "estimatedWorkCycles": 28922400,
      "usefulUtilization": 0.7721583271097834,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.3752201655515913,
      "meanCycles": 40170.0,
      "p50Cycles": 40170,
      "p95Cycles": 40170,
      "maximumCycles": 40170,
      "firstOffset": 131274,
      "lastOffset": 171708
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c64_r160_r164",
      "dimensions": {},
      "sampleCount": 15,
      "tileCount": 15,
      "phaseCount": 9,
      "phaseCycles": 40128,
      "phaseMs": 0.026752,
      "workCycles": 589140,
      "averageActiveTiles": 14.68151913875598,
      "usefulCycles": 472320.0,
      "estimatedWorkCycles": 589140,
      "usefulUtilization": 0.8017109685304002,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.007996151445808196,
      "meanCycles": 39276.0,
      "p50Cycles": 39276,
      "p95Cycles": 39276,
      "maximumCycles": 39276,
      "firstOffset": 485826,
      "lastOffset": 525954
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 7057,
      "tileCount": 1472,
      "phaseCount": 84,
      "phaseCycles": 39798,
      "phaseMs": 0.026532,
      "workCycles": 9451128,
      "averageActiveTiles": 237.47746117895372,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1339.2557744083888,
      "p50Cycles": 906,
      "p95Cycles": 3060,
      "maximumCycles": 3060,
      "firstOffset": 23706,
      "lastOffset": 16902102
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c64_r160_r164",
      "dimensions": {},
      "sampleCount": 12,
      "tileCount": 12,
      "phaseCount": 11,
      "phaseCycles": 38784,
      "phaseMs": 0.025856,
      "workCycles": 462528,
      "averageActiveTiles": 11.925742574257425,
      "usefulCycles": 367488.0,
      "estimatedWorkCycles": 462528,
      "usefulUtilization": 0.7945205479452054,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.00643698880757641,
      "meanCycles": 38544.0,
      "p50Cycles": 38544,
      "p95Cycles": 38544,
      "maximumCycles": 38544,
      "firstOffset": 485250,
      "lastOffset": 524034
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c576_r20_r24",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 22,
      "phaseCycles": 29790,
      "phaseMs": 0.01986,
      "workCycles": 41489280,
      "averageActiveTiles": 1392.725075528701,
      "usefulCycles": 14929920.0,
      "estimatedWorkCycles": 41489280,
      "usefulUtilization": 0.35985006247396917,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.340470248259556,
      "meanCycles": 28812.0,
      "p50Cycles": 28812,
      "p95Cycles": 28812,
      "maximumCycles": 28812,
      "firstOffset": 393600,
      "lastOffset": 423390
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 4,
      "phaseCycles": 28518,
      "phaseMs": 0.019012,
      "workCycles": 19699200,
      "averageActiveTiles": 690.763728171681,
      "usefulCycles": 15344640.0,
      "estimatedWorkCycles": 19699200,
      "usefulUtilization": 0.7789473684210526,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.36553572572014015,
      "meanCycles": 27360.0,
      "p50Cycles": 27360,
      "p95Cycles": 27360,
      "maximumCycles": 27360,
      "firstOffset": 16809798,
      "lastOffset": 16838316
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c96_r144_r148",
      "dimensions": {},
      "sampleCount": 720,
      "tileCount": 720,
      "phaseCount": 7,
      "phaseCycles": 27660,
      "phaseMs": 0.01844,
      "workCycles": 19418400,
      "averageActiveTiles": 702.0390455531453,
      "usefulCycles": 14888448.0,
      "estimatedWorkCycles": 19418400,
      "usefulUtilization": 0.7667185761957731,
      "workLaneOccupancy": 0.9972222222222222,
      "deviceUsefulUtilization": 0.365670093369801,
      "meanCycles": 26970.0,
      "p50Cycles": 26970,
      "p95Cycles": 26970,
      "maximumCycles": 26970,
      "firstOffset": 16810188,
      "lastOffset": 16837848
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 11220,
      "tileCount": 1472,
      "phaseCount": 153,
      "phaseCycles": 27420,
      "phaseMs": 0.01828,
      "workCycles": 20423970,
      "averageActiveTiles": 744.8566739606127,
      "usefulCycles": 4623400.0,
      "estimatedWorkCycles": 20423970,
      "usefulUtilization": 0.22637126866128376,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.11454765642342943,
      "meanCycles": 1820.3181818181818,
      "p50Cycles": 1374,
      "p95Cycles": 4314,
      "maximumCycles": 4314,
      "firstOffset": 91458,
      "lastOffset": 16985046
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c576_r20_r24",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 2,
      "phaseCycles": 27108,
      "phaseMs": 0.018072,
      "workCycles": 637056,
      "averageActiveTiles": 23.50066401062417,
      "usefulCycles": 186624.0,
      "estimatedWorkCycles": 637056,
      "usefulUtilization": 0.29294755877034356,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.0046769443963277325,
      "meanCycles": 26544.0,
      "p50Cycles": 26544,
      "p95Cycles": 26544,
      "maximumCycles": 26544,
      "firstOffset": 393618,
      "lastOffset": 420726
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c112_r121_r122",
      "dimensions": {},
      "sampleCount": 384,
      "tileCount": 384,
      "phaseCount": 4,
      "phaseCycles": 23304,
      "phaseMs": 0.015536,
      "workCycles": 8831232,
      "averageActiveTiles": 378.9577754891864,
      "usefulCycles": 5902848.0,
      "estimatedWorkCycles": 8831232,
      "usefulUtilization": 0.6684059483433342,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.17207719518201764,
      "meanCycles": 22998.0,
      "p50Cycles": 22998,
      "p95Cycles": 22998,
      "maximumCycles": 22998,
      "firstOffset": 241032,
      "lastOffset": 264336
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c112_r121_r122",
      "dimensions": {},
      "sampleCount": 384,
      "tileCount": 384,
      "phaseCount": 4,
      "phaseCycles": 23232,
      "phaseMs": 0.015488,
      "workCycles": 8815104,
      "averageActiveTiles": 379.43801652892563,
      "usefulCycles": 5854464.0,
      "estimatedWorkCycles": 8815104,
      "usefulUtilization": 0.6641400940930475,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.17119565217391305,
      "meanCycles": 22956.0,
      "p50Cycles": 22956,
      "p95Cycles": 22956,
      "maximumCycles": 22956,
      "firstOffset": 241050,
      "lastOffset": 264282
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 4374,
      "tileCount": 1458,
      "phaseCount": 34,
      "phaseCycles": 22866,
      "phaseMs": 0.015244,
      "workCycles": 29463264,
      "averageActiveTiles": 1288.5184990816058,
      "usefulCycles": 11967264.0,
      "estimatedWorkCycles": 29463264,
      "usefulUtilization": 0.40617577197149646,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.35554687232610405,
      "meanCycles": 6736.0,
      "p50Cycles": 6738,
      "p95Cycles": 6738,
      "maximumCycles": 6738,
      "firstOffset": 99942,
      "lastOffset": 16788216
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c96_r121_r122",
      "dimensions": {},
      "sampleCount": 288,
      "tileCount": 288,
      "phaseCount": 3,
      "phaseCycles": 20574,
      "phaseMs": 0.013716,
      "workCycles": 5693760,
      "averageActiveTiles": 276.745406824147,
      "usefulCycles": 3702456.0,
      "estimatedWorkCycles": 5693760,
      "usefulUtilization": 0.6502655538694992,
      "workLaneOccupancy": 0.878125,
      "deviceUsefulUtilization": 0.12225407965308684,
      "meanCycles": 19770.0,
      "p50Cycles": 19770,
      "p95Cycles": 19770,
      "maximumCycles": 19770,
      "firstOffset": 240462,
      "lastOffset": 261036
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c96_r121_r122",
      "dimensions": {},
      "sampleCount": 288,
      "tileCount": 288,
      "phaseCount": 5,
      "phaseCycles": 20016,
      "phaseMs": 0.013344,
      "workCycles": 5683392,
      "averageActiveTiles": 283.9424460431655,
      "usefulCycles": 3672108.0,
      "estimatedWorkCycles": 5683392,
      "usefulUtilization": 0.6461120401337793,
      "workLaneOccupancy": 0.878125,
      "deviceUsefulUtilization": 0.12463222356897091,
      "meanCycles": 19734.0,
      "p50Cycles": 19734,
      "p95Cycles": 19734,
      "maximumCycles": 19734,
      "firstOffset": 240444,
      "lastOffset": 260460
    },
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 30,
      "phaseCycles": 17220,
      "phaseMs": 0.01148,
      "workCycles": 24284448,
      "averageActiveTiles": 1410.2466898954704,
      "usefulCycles": 17256888.0,
      "estimatedWorkCycles": 24284448,
      "usefulUtilization": 0.7106147934678194,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.6808030980154522,
      "meanCycles": 16656.0,
      "p50Cycles": 16656,
      "p95Cycles": 16656,
      "maximumCycles": 16656,
      "firstOffset": 562110,
      "lastOffset": 579330
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c80_r104_r105",
      "dimensions": {},
      "sampleCount": 1152,
      "tileCount": 1152,
      "phaseCount": 14,
      "phaseCycles": 14058,
      "phaseMs": 0.009372,
      "workCycles": 13727232,
      "averageActiveTiles": 976.4711907810499,
      "usefulCycles": 8188128.0,
      "estimatedWorkCycles": 13727232,
      "usefulUtilization": 0.596487915407855,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.3956883594054445,
      "meanCycles": 11916.0,
      "p50Cycles": 11916,
      "p95Cycles": 11916,
      "maximumCycles": 11916,
      "firstOffset": 332976,
      "lastOffset": 347034
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c80_r104_r105",
      "dimensions": {},
      "sampleCount": 192,
      "tileCount": 192,
      "phaseCount": 6,
      "phaseCycles": 13878,
      "phaseMs": 0.009252,
      "workCycles": 2297088,
      "averageActiveTiles": 165.5201037613489,
      "usefulCycles": 1377810.0,
      "estimatedWorkCycles": 2297088,
      "usefulUtilization": 0.5998072342026078,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.06744575790898324,
      "meanCycles": 11964.0,
      "p50Cycles": 11964,
      "p95Cycles": 11964,
      "maximumCycles": 11964,
      "firstOffset": 333498,
      "lastOffset": 347376
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c32_p32_b64x16",
      "dimensions": {},
      "sampleCount": 704,
      "tileCount": 704,
      "phaseCount": 6,
      "phaseCycles": 13158,
      "phaseMs": 0.008772,
      "workCycles": 7569408,
      "averageActiveTiles": 575.2704058367533,
      "usefulCycles": 720896.0,
      "estimatedWorkCycles": 7569408,
      "usefulUtilization": 0.09523809523809523,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.03721987615403424,
      "meanCycles": 10752.0,
      "p50Cycles": 10752,
      "p95Cycles": 10752,
      "maximumCycles": 10752,
      "firstOffset": 211656,
      "lastOffset": 224814
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k736_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 160,
      "phaseCount": 14,
      "phaseCycles": 12384,
      "phaseMs": 0.008256,
      "workCycles": 1257600,
      "averageActiveTiles": 101.55038759689923,
      "usefulCycles": 26244.0,
      "estimatedWorkCycles": 1257600,
      "usefulUtilization": 0.020868320610687024,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0014396644337714864,
      "meanCycles": 7860.0,
      "p50Cycles": 7860,
      "p95Cycles": 7860,
      "maximumCycles": 7860,
      "firstOffset": 16917396,
      "lastOffset": 16929780
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c32_p32_b64x16",
      "dimensions": {},
      "sampleCount": 64,
      "tileCount": 64,
      "phaseCount": 1,
      "phaseCycles": 11778,
      "phaseMs": 0.007852,
      "workCycles": 651264,
      "averageActiveTiles": 55.29495669893021,
      "usefulCycles": 25600.0,
      "estimatedWorkCycles": 651264,
      "usefulUtilization": 0.03930817610062893,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0014765923202433424,
      "meanCycles": 10176.0,
      "p50Cycles": 10176,
      "p95Cycles": 10176,
      "maximumCycles": 10176,
      "firstOffset": 211410,
      "lastOffset": 223188
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 27,
      "phaseCycles": 11520,
      "phaseMs": 0.00768,
      "workCycles": 1792896,
      "averageActiveTiles": 155.63333333333333,
      "usefulCycles": 26244.0,
      "estimatedWorkCycles": 1792896,
      "usefulUtilization": 0.01463777040051403,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.0015476392663043479,
      "meanCycles": 1218.0,
      "p50Cycles": 1218,
      "p95Cycles": 1218,
      "maximumCycles": 1218,
      "firstOffset": 16891782,
      "lastOffset": 16903320
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r128_p128_c16_p16_b128x16",
      "dimensions": {},
      "sampleCount": 800,
      "tileCount": 800,
      "phaseCount": 22,
      "phaseCycles": 10992,
      "phaseMs": 0.007328,
      "workCycles": 3408000,
      "averageActiveTiles": 310.04366812227073,
      "usefulCycles": 819200.0,
      "estimatedWorkCycles": 3408000,
      "usefulUtilization": 0.2403755868544601,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.05062970698057085,
      "meanCycles": 4260.0,
      "p50Cycles": 4260,
      "p95Cycles": 4260,
      "maximumCycles": 4260,
      "firstOffset": 16879272,
      "lastOffset": 16890264
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r89_p128_c16_p16_b128x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 160,
      "phaseCount": 11,
      "phaseCycles": 9798,
      "phaseMs": 0.006532,
      "workCycles": 606720,
      "averageActiveTiles": 61.92284139620331,
      "usefulCycles": 113920.0,
      "estimatedWorkCycles": 606720,
      "usefulUtilization": 0.1877637130801688,
      "workLaneOccupancy": 0.6953125,
      "deviceUsefulUtilization": 0.007898683848522768,
      "meanCycles": 3792.0,
      "p50Cycles": 3792,
      "p95Cycles": 3792,
      "maximumCycles": 3792,
      "firstOffset": 16879200,
      "lastOffset": 16888998
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_p8_f16",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 23,
      "phaseCycles": 9084,
      "phaseMs": 0.006056,
      "workCycles": 6438528,
      "averageActiveTiles": 708.7767503302509,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 4374.0,
      "p50Cycles": 4374,
      "p95Cycles": 4374,
      "maximumCycles": 4374,
      "firstOffset": 356826,
      "lastOffset": 365910
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k64_c48_r240_r244",
      "dimensions": {},
      "sampleCount": 1200,
      "tileCount": 1200,
      "phaseCount": 4,
      "phaseCycles": 8004,
      "phaseMs": 0.005336,
      "workCycles": 8704800,
      "averageActiveTiles": 1087.5562218890555,
      "usefulCycles": 6456240.0,
      "estimatedWorkCycles": 8704800,
      "usefulUtilization": 0.7416873449131514,
      "workLaneOccupancy": 0.91875,
      "deviceUsefulUtilization": 0.5479800860439346,
      "meanCycles": 7254.0,
      "p50Cycles": 7254,
      "p95Cycles": 7254,
      "maximumCycles": 7254,
      "firstOffset": 14760,
      "lastOffset": 22764
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c48_r240_r244",
      "dimensions": {},
      "sampleCount": 240,
      "tileCount": 240,
      "phaseCount": 2,
      "phaseCycles": 7914,
      "phaseMs": 0.005276,
      "workCycles": 1726560,
      "averageActiveTiles": 218.1652767247915,
      "usefulCycles": 1259496.0,
      "estimatedWorkCycles": 1726560,
      "usefulUtilization": 0.7294829024186822,
      "workLaneOccupancy": 0.91109375,
      "deviceUsefulUtilization": 0.10811673863598906,
      "meanCycles": 7194.0,
      "p50Cycles": 7194,
      "p95Cycles": 7194,
      "maximumCycles": 7194,
      "firstOffset": 14772,
      "lastOffset": 22686
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 6,
      "phaseCycles": 6138,
      "phaseMs": 0.004092,
      "workCycles": 195072,
      "averageActiveTiles": 31.781036168132943,
      "usefulCycles": 43392.0,
      "estimatedWorkCycles": 195072,
      "usefulUtilization": 0.22244094488188976,
      "workLaneOccupancy": 0.9970588235294118,
      "deviceUsefulUtilization": 0.004802584045220791,
      "meanCycles": 6096.0,
      "p50Cycles": 6096,
      "p95Cycles": 6096,
      "maximumCycles": 6096,
      "firstOffset": 16904484,
      "lastOffset": 16910622
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 356,
      "tileCount": 208,
      "phaseCount": 43,
      "phaseCycles": 5418,
      "phaseMs": 0.003612,
      "workCycles": 95784,
      "averageActiveTiles": 17.678848283499445,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 269.0561797752809,
      "p50Cycles": 288,
      "p95Cycles": 294,
      "maximumCycles": 300,
      "firstOffset": 197496,
      "lastOffset": 16982814
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 8,
      "phaseCycles": 4248,
      "phaseMs": 0.002832,
      "workCycles": 54174,
      "averageActiveTiles": 12.752824858757062,
      "usefulCycles": 25824.0,
      "estimatedWorkCycles": 54174,
      "usefulUtilization": 0.47668623324842174,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0041298206828788995,
      "meanCycles": 3385.875,
      "p50Cycles": 3222,
      "p95Cycles": 3462,
      "maximumCycles": 4164,
      "firstOffset": 16963044,
      "lastOffset": 16967292
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 352,
      "tileCount": 352,
      "phaseCount": 3,
      "phaseCycles": 4026,
      "phaseMs": 0.002684,
      "workCycles": 857472,
      "averageActiveTiles": 212.98360655737704,
      "usefulCycles": 180224.0,
      "estimatedWorkCycles": 857472,
      "usefulUtilization": 0.21018062397372742,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.030411023996198622,
      "meanCycles": 2436.0,
      "p50Cycles": 2436,
      "p95Cycles": 2436,
      "maximumCycles": 2436,
      "firstOffset": 211866,
      "lastOffset": 215892
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 1,
      "phaseCycles": 3894,
      "phaseMs": 0.002596,
      "workCycles": 62976,
      "averageActiveTiles": 16.172573189522343,
      "usefulCycles": 6400.0,
      "estimatedWorkCycles": 62976,
      "usefulUtilization": 0.1016260162601626,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0011165449632656707,
      "meanCycles": 1968.0,
      "p50Cycles": 1968,
      "p95Cycles": 1968,
      "maximumCycles": 1968,
      "firstOffset": 210846,
      "lastOffset": 214740
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c32_r4_r4",
      "dimensions": {},
      "sampleCount": 184,
      "tileCount": 184,
      "phaseCount": 21,
      "phaseCycles": 3468,
      "phaseMs": 0.002312,
      "workCycles": 442704,
      "averageActiveTiles": 127.65397923875433,
      "usefulCycles": 17216.0,
      "estimatedWorkCycles": 442704,
      "usefulUtilization": 0.038888286530051684,
      "workLaneOccupancy": 0.48731884057971014,
      "deviceUsefulUtilization": 0.003372448723735018,
      "meanCycles": 2406.0,
      "p50Cycles": 2406,
      "p95Cycles": 2406,
      "maximumCycles": 2406,
      "firstOffset": 16973784,
      "lastOffset": 16977252
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c32_r4_r4",
      "dimensions": {},
      "sampleCount": 680,
      "tileCount": 680,
      "phaseCount": 30,
      "phaseCycles": 3336,
      "phaseMs": 0.002224,
      "workCycles": 1411680,
      "averageActiveTiles": 423.1654676258993,
      "usefulCycles": 48960.0,
      "estimatedWorkCycles": 1411680,
      "usefulUtilization": 0.03468208092485549,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.009970284641851736,
      "meanCycles": 2076.0,
      "p50Cycles": 2076,
      "p95Cycles": 2076,
      "maximumCycles": 2076,
      "firstOffset": 16956486,
      "lastOffset": 16959822
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 4,
      "phaseCycles": 2916,
      "phaseMs": 0.001944,
      "workCycles": 13440,
      "averageActiveTiles": 4.609053497942387,
      "usefulCycles": 2016.0,
      "estimatedWorkCycles": 13440,
      "usefulUtilization": 0.15,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0004696725711218465,
      "meanCycles": 1680.0,
      "p50Cycles": 1680,
      "p95Cycles": 1680,
      "maximumCycles": 1680,
      "firstOffset": 16947654,
      "lastOffset": 16950570
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k160_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 792,
      "tileCount": 792,
      "phaseCount": 26,
      "phaseCycles": 2550,
      "phaseMs": 0.0017,
      "workCycles": 978912,
      "averageActiveTiles": 383.8870588235294,
      "usefulCycles": 28512.0,
      "estimatedWorkCycles": 978912,
      "usefulUtilization": 0.02912621359223301,
      "workLaneOccupancy": 0.45,
      "deviceUsefulUtilization": 0.007595907928388747,
      "meanCycles": 1236.0,
      "p50Cycles": 1236,
      "p95Cycles": 1236,
      "maximumCycles": 1236,
      "firstOffset": 16956000,
      "lastOffset": 16958550
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 1288,
      "tileCount": 1288,
      "phaseCount": 39,
      "phaseCycles": 2454,
      "phaseMs": 0.001636,
      "workCycles": 1808352,
      "averageActiveTiles": 736.8997555012224,
      "usefulCycles": 60256.0,
      "estimatedWorkCycles": 1808352,
      "usefulUtilization": 0.033320946364424626,
      "workLaneOccupancy": 0.48731884057971014,
      "deviceUsefulUtilization": 0.016680840508840934,
      "meanCycles": 1404.0,
      "p50Cycles": 1404,
      "p95Cycles": 1404,
      "maximumCycles": 1404,
      "firstOffset": 16973460,
      "lastOffset": 16975914
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r4_r4",
      "dimensions": {},
      "sampleCount": 2592,
      "tileCount": 1296,
      "phaseCount": 40,
      "phaseCycles": 2352,
      "phaseMs": 0.001568,
      "workCycles": 1928448,
      "averageActiveTiles": 819.9183673469388,
      "usefulCycles": 41472.0,
      "estimatedWorkCycles": 1928448,
      "usefulUtilization": 0.021505376344086023,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.011978704525288377,
      "meanCycles": 744.0,
      "p50Cycles": 744,
      "p95Cycles": 744,
      "maximumCycles": 744,
      "firstOffset": 16867662,
      "lastOffset": 16942542
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 4,
      "phaseCycles": 2202,
      "phaseMs": 0.001468,
      "workCycles": 17568,
      "averageActiveTiles": 7.978201634877384,
      "usefulCycles": 3456.0,
      "estimatedWorkCycles": 17568,
      "usefulUtilization": 0.19672131147540983,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.0010662243809975121,
      "meanCycles": 2196.0,
      "p50Cycles": 2196,
      "p95Cycles": 2196,
      "maximumCycles": 2196,
      "firstOffset": 16950846,
      "lastOffset": 16953048
    },
    {
      "name": "attention_merge_v72_p80_k768_out16_f16",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 6,
      "phaseCycles": 1638,
      "phaseMs": 0.001092,
      "workCycles": 51072,
      "averageActiveTiles": 31.17948717948718,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1596.0,
      "p50Cycles": 1596,
      "p95Cycles": 1596,
      "maximumCycles": 1596,
      "firstOffset": 16930392,
      "lastOffset": 16932030
    }
  ],
  "samples": []
}
