{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 206695,
  "matchedSampleCount": 206695,
  "profileSpanCycles": 715350,
  "profileSpanMs": 0.4769,
  "atOffset": null,
  "groups": [
    {
      "name": "sync",
      "dimensions": {},
      "sampleCount": 11126,
      "tileCount": 1456,
      "phaseCount": 43,
      "phaseCycles": 570318,
      "phaseMs": 0.380212,
      "workCycles": 59267046,
      "averageActiveTiles": 103.91929765499248,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 5326.896099227036,
      "p50Cycles": 3012,
      "p95Cycles": 17778,
      "maximumCycles": 63966,
      "firstOffset": 2844,
      "lastOffset": 713958
    },
    {
      "name": "exchange",
      "dimensions": {},
      "sampleCount": 65418,
      "tileCount": 1472,
      "phaseCount": 52,
      "phaseCycles": 449436,
      "phaseMs": 0.299624,
      "workCycles": 443524992,
      "averageActiveTiles": 986.8479427549195,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 6779.861689443273,
      "p50Cycles": 5094,
      "p95Cycles": 19350,
      "maximumCycles": 31572,
      "firstOffset": 2832,
      "lastOffset": 714168
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 28217,
      "tileCount": 1472,
      "phaseCount": 201,
      "phaseCycles": 99204,
      "phaseMs": 0.066136,
      "workCycles": 21082962,
      "averageActiveTiles": 212.5212894641345,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 747.1723429138463,
      "p50Cycles": 558,
      "p95Cycles": 1326,
      "maximumCycles": 2010,
      "firstOffset": 9372,
      "lastOffset": 713580
    },
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 13815,
      "tileCount": 1472,
      "phaseCount": 160,
      "phaseCycles": 63150,
      "phaseMs": 0.0421,
      "workCycles": 59057586,
      "averageActiveTiles": 935.1953444180523,
      "usefulCycles": 10298734.0,
      "estimatedWorkCycles": 54888474,
      "usefulUtilization": 0.18763017532606208,
      "workLaneOccupancy": 0.9764424681432039,
      "deviceUsefulUtilization": 0.11079053926124824,
      "meanCycles": 4274.888599348535,
      "p50Cycles": 3654,
      "p95Cycles": 8238,
      "maximumCycles": 9102,
      "firstOffset": 0,
      "lastOffset": 708762
    },
    {
      "name": "reduce_sum_f16",
      "dimensions": {},
      "sampleCount": 10564,
      "tileCount": 1472,
      "phaseCount": 167,
      "phaseCycles": 57750,
      "phaseMs": 0.0385,
      "workCycles": 36486618,
      "averageActiveTiles": 631.8029090909091,
      "usefulCycles": 18024296.0,
      "estimatedWorkCycles": 36486618,
      "usefulUtilization": 0.49399744311736427,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.21203058535667232,
      "meanCycles": 3453.8638773191974,
      "p50Cycles": 2658,
      "p95Cycles": 7212,
      "maximumCycles": 7968,
      "firstOffset": 9906,
      "lastOffset": 713298
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 15270,
      "tileCount": 1472,
      "phaseCount": 137,
      "phaseCycles": 45606,
      "phaseMs": 0.030404,
      "workCycles": 5563332,
      "averageActiveTiles": 121.98684383633733,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 2905092,
      "usefulUtilization": 0.0,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": 0.0,
      "meanCycles": 364.33084479371314,
      "p50Cycles": 300,
      "p95Cycles": 600,
      "maximumCycles": 1500,
      "firstOffset": 120174,
      "lastOffset": 706014
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k384_c192_r36_r38",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 27,
      "phaseCycles": 45180,
      "phaseMs": 0.03012,
      "workCycles": 16249680,
      "averageActiveTiles": 359.66533864541833,
      "usefulCycles": 7359840.0,
      "estimatedWorkCycles": 16249680,
      "usefulUtilization": 0.4529221498515663,
      "workLaneOccupancy": 0.9340277777777778,
      "deviceUsefulUtilization": 0.11066603152606963,
      "meanCycles": 45138.0,
      "p50Cycles": 45138,
      "p95Cycles": 45138,
      "maximumCycles": 45138,
      "firstOffset": 479286,
      "lastOffset": 524466
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k384_c192_r36_r38",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 33,
      "phaseCycles": 44850,
      "phaseMs": 0.0299,
      "workCycles": 48366720,
      "averageActiveTiles": 1078.410702341137,
      "usefulCycles": 20878704.0,
      "estimatedWorkCycles": 48366720,
      "usefulUtilization": 0.4316750029772538,
      "workLaneOccupancy": 0.9322980967078189,
      "deviceUsefulUtilization": 0.316251999418351,
      "meanCycles": 44784.0,
      "p50Cycles": 44784,
      "p95Cycles": 44784,
      "maximumCycles": 44784,
      "firstOffset": 479274,
      "lastOffset": 524124
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 780,
      "tileCount": 780,
      "phaseCount": 28,
      "phaseCycles": 36954,
      "phaseMs": 0.024636,
      "workCycles": 23180040,
      "averageActiveTiles": 627.2674135411593,
      "usefulCycles": 15350400.0,
      "estimatedWorkCycles": 23180040,
      "usefulUtilization": 0.6622249141934181,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.2821957262966179,
      "meanCycles": 29718.0,
      "p50Cycles": 29718,
      "p95Cycles": 29718,
      "maximumCycles": 29718,
      "firstOffset": 341472,
      "lastOffset": 378426
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c160_r80_r82",
      "dimensions": {},
      "sampleCount": 624,
      "tileCount": 624,
      "phaseCount": 26,
      "phaseCycles": 36438,
      "phaseMs": 0.024292,
      "workCycles": 18225792,
      "averageActiveTiles": 500.1863988144245,
      "usefulCycles": 11943360.0,
      "estimatedWorkCycles": 18225792,
      "usefulUtilization": 0.6552999178307313,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.22267126769235174,
      "meanCycles": 29208.0,
      "p50Cycles": 29208,
      "p95Cycles": 29208,
      "maximumCycles": 29208,
      "firstOffset": 341454,
      "lastOffset": 377892
    },
    {
      "name": "attention_softmax_d72_p768_tail_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 34,
      "phaseCycles": 34254,
      "phaseMs": 0.022836,
      "workCycles": 37251936,
      "averageActiveTiles": 1087.5207566999475,
      "usefulCycles": 18466080.0,
      "estimatedWorkCycles": 37251936,
      "usefulUtilization": 0.4957079277705191,
      "workLaneOccupancy": 0.9974763406940063,
      "deviceUsefulUtilization": 0.3662314271135583,
      "meanCycles": 25034.90322580645,
      "p50Cycles": 25236,
      "p95Cycles": 25236,
      "maximumCycles": 25236,
      "firstOffset": 166830,
      "lastOffset": 667764
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 13,
      "phaseCycles": 33456,
      "phaseMs": 0.022304,
      "workCycles": 631728,
      "averageActiveTiles": 18.88235294117647,
      "usefulCycles": 413424.0,
      "estimatedWorkCycles": 631728,
      "usefulUtilization": 0.6544335536813312,
      "workLaneOccupancy": 0.996875,
      "deviceUsefulUtilization": 0.00839486775622232,
      "meanCycles": 26322.0,
      "p50Cycles": 26322,
      "p95Cycles": 26322,
      "maximumCycles": 26322,
      "firstOffset": 340698,
      "lastOffset": 374154
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 3771,
      "tileCount": 1472,
      "phaseCount": 120,
      "phaseCycles": 32388,
      "phaseMs": 0.021592,
      "workCycles": 1589406,
      "averageActiveTiles": 49.07391626528344,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 421.4813046937152,
      "p50Cycles": 414,
      "p95Cycles": 714,
      "maximumCycles": 714,
      "firstOffset": 100176,
      "lastOffset": 706836
    },
    {
      "name": "layer_norm_f16",
      "dimensions": {},
      "sampleCount": 2187,
      "tileCount": 729,
      "phaseCount": 59,
      "phaseCycles": 32142,
      "phaseMs": 0.021428,
      "workCycles": 20509686,
      "averageActiveTiles": 638.0961358969572,
      "usefulCycles": 7243344.0,
      "estimatedWorkCycles": 20509686,
      "usefulUtilization": 0.3531669865642994,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.1530940825088669,
      "meanCycles": 9378.0,
      "p50Cycles": 9378,
      "p95Cycles": 9378,
      "maximumCycles": 9378,
      "firstOffset": 21990,
      "lastOffset": 550050
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c144_r80_r82",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 15,
      "phaseCycles": 30438,
      "phaseMs": 0.020292,
      "workCycles": 803340,
      "averageActiveTiles": 26.392667060910703,
      "usefulCycles": 531360.0,
      "estimatedWorkCycles": 803340,
      "usefulUtilization": 0.6614384942863545,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.011859460571311034,
      "meanCycles": 26778.0,
      "p50Cycles": 26778,
      "p95Cycles": 26778,
      "maximumCycles": 26778,
      "firstOffset": 340704,
      "lastOffset": 371142
    },
    {
      "name": "add_f16",
      "dimensions": {},
      "sampleCount": 14221,
      "tileCount": 1458,
      "phaseCount": 182,
      "phaseCycles": 26058,
      "phaseMs": 0.017372,
      "workCycles": 19418460,
      "averageActiveTiles": 745.2014736357356,
      "usefulCycles": 3096104.0,
      "estimatedWorkCycles": 19418460,
      "usefulUtilization": 0.15944127392182492,
      "workLaneOccupancy": 0.9995596403255821,
      "deviceUsefulUtilization": 0.08071730454137425,
      "meanCycles": 1365.4778144996835,
      "p50Cycles": 1068,
      "p95Cycles": 2652,
      "maximumCycles": 2652,
      "firstOffset": 17160,
      "lastOffset": 715350
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 2301,
      "tileCount": 856,
      "phaseCount": 158,
      "phaseCycles": 23550,
      "phaseMs": 0.0157,
      "workCycles": 591444,
      "averageActiveTiles": 25.1143949044586,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 257.03780964797915,
      "p50Cycles": 288,
      "p95Cycles": 288,
      "maximumCycles": 1068,
      "firstOffset": 120468,
      "lastOffset": 714318
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r64_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 1760,
      "tileCount": 880,
      "phaseCount": 54,
      "phaseCycles": 22746,
      "phaseMs": 0.015164,
      "workCycles": 4276800,
      "averageActiveTiles": 188.0242680031654,
      "usefulCycles": 901120.0,
      "estimatedWorkCycles": 4276800,
      "usefulUtilization": 0.21069958847736625,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.02691347547012566,
      "meanCycles": 2430.0,
      "p50Cycles": 2430,
      "p95Cycles": 2430,
      "maximumCycles": 2430,
      "firstOffset": 127464,
      "lastOffset": 649446
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c256_r40_r42",
      "dimensions": {},
      "sampleCount": 405,
      "tileCount": 405,
      "phaseCount": 32,
      "phaseCycles": 21948,
      "phaseMs": 0.014632,
      "workCycles": 8852490,
      "averageActiveTiles": 403.33925642427556,
      "usefulCycles": 4354560.0,
      "estimatedWorkCycles": 8852490,
      "usefulUtilization": 0.4919022783420258,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.13478498585589654,
      "meanCycles": 21858.0,
      "p50Cycles": 21858,
      "p95Cycles": 21858,
      "maximumCycles": 21858,
      "firstOffset": 574224,
      "lastOffset": 596172
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r38_p38_c80_p384",
      "dimensions": {},
      "sampleCount": 30,
      "tileCount": 30,
      "phaseCount": 15,
      "phaseCycles": 21696,
      "phaseMs": 0.014464,
      "workCycles": 576540,
      "averageActiveTiles": 26.573561946902654,
      "usefulCycles": 45600.0,
      "estimatedWorkCycles": 576540,
      "usefulUtilization": 0.07909251743157457,
      "workLaneOccupancy": 0.20833333333333334,
      "deviceUsefulUtilization": 0.001427832820315506,
      "meanCycles": 19218.0,
      "p50Cycles": 19218,
      "p95Cycles": 19218,
      "maximumCycles": 19218,
      "firstOffset": 429732,
      "lastOffset": 451428
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k128_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 5,
      "phaseCycles": 21318,
      "phaseMs": 0.014212,
      "workCycles": 25726410,
      "averageActiveTiles": 1206.7928511117366,
      "usefulCycles": 18973440.0,
      "estimatedWorkCycles": 25726410,
      "usefulUtilization": 0.7375082648531217,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.6046329494976689,
      "meanCycles": 21174.0,
      "p50Cycles": 21174,
      "p95Cycles": 21174,
      "maximumCycles": 21174,
      "firstOffset": 65310,
      "lastOffset": 86628
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c256_r40_r42",
      "dimensions": {},
      "sampleCount": 1053,
      "tileCount": 1053,
      "phaseCount": 39,
      "phaseCycles": 21126,
      "phaseMs": 0.014084,
      "workCycles": 22169862,
      "averageActiveTiles": 1049.411246804885,
      "usefulCycles": 10761984.0,
      "estimatedWorkCycles": 22169862,
      "usefulUtilization": 0.4854330622355701,
      "workLaneOccupancy": 0.9980769230769231,
      "deviceUsefulUtilization": 0.3460726325278145,
      "meanCycles": 21054.0,
      "p50Cycles": 21054,
      "p95Cycles": 21054,
      "maximumCycles": 21054,
      "firstOffset": 574284,
      "lastOffset": 595410
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c128_r120_r122",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 2,
      "phaseCycles": 21090,
      "phaseMs": 0.01406,
      "workCycles": 5117580,
      "averageActiveTiles": 242.6543385490754,
      "usefulCycles": 3701376.0,
      "estimatedWorkCycles": 5117580,
      "usefulUtilization": 0.72326685660019,
      "workLaneOccupancy": 0.9916666666666667,
      "deviceUsefulUtilization": 0.11922815263776362,
      "meanCycles": 21060.0,
      "p50Cycles": 21060,
      "p95Cycles": 21060,
      "maximumCycles": 21060,
      "firstOffset": 65394,
      "lastOffset": 86484
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 896,
      "tileCount": 896,
      "phaseCount": 18,
      "phaseCycles": 18876,
      "phaseMs": 0.012584,
      "workCycles": 14477568,
      "averageActiveTiles": 766.9828353464717,
      "usefulCycles": 8358714.0,
      "estimatedWorkCycles": 14477568,
      "usefulUtilization": 0.5773562244708503,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.3008303763923823,
      "meanCycles": 16158.0,
      "p50Cycles": 16158,
      "p95Cycles": 16158,
      "maximumCycles": 16158,
      "firstOffset": 203802,
      "lastOffset": 222678
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r36_p36_c80_p384",
      "dimensions": {},
      "sampleCount": 84,
      "tileCount": 84,
      "phaseCount": 20,
      "phaseCycles": 18672,
      "phaseMs": 0.012448,
      "workCycles": 1386504,
      "averageActiveTiles": 74.25578406169666,
      "usefulCycles": 120960.0,
      "estimatedWorkCycles": 1386504,
      "usefulUtilization": 0.08724100327153762,
      "workLaneOccupancy": 0.20833333333333334,
      "deviceUsefulUtilization": 0.004400916508326813,
      "meanCycles": 16506.0,
      "p50Cycles": 16506,
      "p95Cycles": 16506,
      "maximumCycles": 16506,
      "firstOffset": 429618,
      "lastOffset": 448290
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r35_p36_c80_p384",
      "dimensions": {},
      "sampleCount": 6,
      "tileCount": 6,
      "phaseCount": 3,
      "phaseCycles": 18270,
      "phaseMs": 0.01218,
      "workCycles": 100584,
      "averageActiveTiles": 5.505418719211822,
      "usefulCycles": 8400.0,
      "estimatedWorkCycles": 100584,
      "usefulUtilization": 0.08351228823669768,
      "workLaneOccupancy": 0.2025462962962963,
      "deviceUsefulUtilization": 0.00031234382808595704,
      "meanCycles": 16764.0,
      "p50Cycles": 16764,
      "p95Cycles": 16764,
      "maximumCycles": 16764,
      "firstOffset": 429540,
      "lastOffset": 447810
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k96_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 128,
      "tileCount": 128,
      "phaseCount": 10,
      "phaseCycles": 18246,
      "phaseMs": 0.012164,
      "workCycles": 2072832,
      "averageActiveTiles": 113.60473528444591,
      "usefulCycles": 1207224.0,
      "estimatedWorkCycles": 2072832,
      "usefulUtilization": 0.5824032048907003,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.044948207826372905,
      "meanCycles": 16194.0,
      "p50Cycles": 16194,
      "p95Cycles": 16194,
      "maximumCycles": 16194,
      "firstOffset": 204486,
      "lastOffset": 222732
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 224,
      "tileCount": 224,
      "phaseCount": 14,
      "phaseCycles": 18108,
      "phaseMs": 0.012072,
      "workCycles": 3033408,
      "averageActiveTiles": 167.51756129887343,
      "usefulCycles": 1834560.0,
      "estimatedWorkCycles": 3033408,
      "usefulUtilization": 0.6047851129818342,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.06882617339441612,
      "meanCycles": 13542.0,
      "p50Cycles": 13542,
      "p95Cycles": 13542,
      "maximumCycles": 13542,
      "firstOffset": 144024,
      "lastOffset": 162132
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 1008,
      "tileCount": 1008,
      "phaseCount": 15,
      "phaseCycles": 17238,
      "phaseMs": 0.011492,
      "workCycles": 10983168,
      "averageActiveTiles": 637.1486251305256,
      "usefulCycles": 6524154.0,
      "estimatedWorkCycles": 10983168,
      "usefulUtilization": 0.5940138582966226,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.25711624532756244,
      "meanCycles": 10896.0,
      "p50Cycles": 10896,
      "p95Cycles": 10896,
      "maximumCycles": 10896,
      "firstOffset": 141384,
      "lastOffset": 158622
    },
    {
      "name": "gelu_tanh_approx_f16",
      "dimensions": {},
      "sampleCount": 1621,
      "tileCount": 1458,
      "phaseCount": 88,
      "phaseCycles": 17082,
      "phaseMs": 0.011388,
      "workCycles": 18324012,
      "averageActiveTiles": 1072.7088162978573,
      "usefulCycles": 9425760.0,
      "estimatedWorkCycles": 18324012,
      "usefulUtilization": 0.5143939002004583,
      "workLaneOccupancy": 0.9986320109439124,
      "deviceUsefulUtilization": 0.374860646599777,
      "meanCycles": 11304.140653917335,
      "p50Cycles": 12474,
      "p95Cycles": 12474,
      "maximumCycles": 12474,
      "firstOffset": 407544,
      "lastOffset": 705720
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r38_p38_c384_p384",
      "dimensions": {},
      "sampleCount": 330,
      "tileCount": 330,
      "phaseCount": 27,
      "phaseCycles": 15756,
      "phaseMs": 0.010504,
      "workCycles": 3958020,
      "averageActiveTiles": 251.20715917745622,
      "usefulCycles": 2407680.0,
      "estimatedWorkCycles": 3958020,
      "usefulUtilization": 0.608304152076038,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.1038113844829299,
      "meanCycles": 11994.0,
      "p50Cycles": 11994,
      "p95Cycles": 11994,
      "maximumCycles": 11994,
      "firstOffset": 429312,
      "lastOffset": 445068
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 160,
      "tileCount": 80,
      "phaseCount": 36,
      "phaseCycles": 15630,
      "phaseMs": 0.01042,
      "workCycles": 313920,
      "averageActiveTiles": 20.084452975047984,
      "usefulCycles": 32000.0,
      "estimatedWorkCycles": 313920,
      "usefulUtilization": 0.1019367991845056,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0013908592728587722,
      "meanCycles": 1962.0,
      "p50Cycles": 1962,
      "p95Cycles": 1962,
      "maximumCycles": 1962,
      "firstOffset": 129042,
      "lastOffset": 648516
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 176,
      "tileCount": 176,
      "phaseCount": 29,
      "phaseCycles": 15516,
      "phaseMs": 0.010344,
      "workCycles": 2072928,
      "averageActiveTiles": 133.59938128383604,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 2072928,
      "usefulUtilization": 0.21735438953982,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.019727182039297443,
      "meanCycles": 11778.0,
      "p50Cycles": 11778,
      "p95Cycles": 11778,
      "maximumCycles": 11778,
      "firstOffset": 631500,
      "lastOffset": 647016
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r35_p36_c384_p384",
      "dimensions": {},
      "sampleCount": 66,
      "tileCount": 66,
      "phaseCount": 16,
      "phaseCycles": 14622,
      "phaseMs": 0.009748,
      "workCycles": 765864,
      "averageActiveTiles": 52.37751333606894,
      "usefulCycles": 443520.0,
      "estimatedWorkCycles": 765864,
      "usefulUtilization": 0.5791106514994829,
      "workLaneOccupancy": 0.9722222222222222,
      "deviceUsefulUtilization": 0.02060623360867781,
      "meanCycles": 11604.0,
      "p50Cycles": 11604,
      "p95Cycles": 11604,
      "maximumCycles": 11604,
      "firstOffset": 429594,
      "lastOffset": 444216
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r36_p36_c384_p384",
      "dimensions": {},
      "sampleCount": 924,
      "tileCount": 924,
      "phaseCount": 26,
      "phaseCycles": 14430,
      "phaseMs": 0.00962,
      "workCycles": 9530136,
      "averageActiveTiles": 660.4390852390852,
      "usefulCycles": 6386688.0,
      "estimatedWorkCycles": 9530136,
      "usefulUtilization": 0.6701570680628273,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.30067793546054417,
      "meanCycles": 10314.0,
      "p50Cycles": 10314,
      "p95Cycles": 10314,
      "maximumCycles": 10314,
      "firstOffset": 429168,
      "lastOffset": 443598
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c80_r91_r92",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 8,
      "phaseCycles": 14064,
      "phaseMs": 0.009376,
      "workCycles": 434304,
      "averageActiveTiles": 30.880546075085324,
      "usefulCycles": 264960.0,
      "estimatedWorkCycles": 434304,
      "usefulUtilization": 0.610079575596817,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.012798634812286689,
      "meanCycles": 13572.0,
      "p50Cycles": 13572,
      "p95Cycles": 13572,
      "maximumCycles": 13572,
      "firstOffset": 148116,
      "lastOffset": 162180
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r122_p122_c128_p128",
      "dimensions": {},
      "sampleCount": 1215,
      "tileCount": 1215,
      "phaseCount": 5,
      "phaseCycles": 13482,
      "phaseMs": 0.008988,
      "workCycles": 16205670,
      "averageActiveTiles": 1202.022696929239,
      "usefulCycles": 9486720.0,
      "estimatedWorkCycles": 16205670,
      "usefulUtilization": 0.5853951117109012,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.4780286759157137,
      "meanCycles": 13338.0,
      "p50Cycles": 13338,
      "p95Cycles": 13338,
      "maximumCycles": 13338,
      "firstOffset": 40650,
      "lastOffset": 54132
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r79_p80_c192_p192",
      "dimensions": {},
      "sampleCount": 162,
      "tileCount": 162,
      "phaseCount": 16,
      "phaseCycles": 13344,
      "phaseMs": 0.008896,
      "workCycles": 2155896,
      "averageActiveTiles": 161.56294964028777,
      "usefulCycles": 1228608.0,
      "estimatedWorkCycles": 2155896,
      "usefulUtilization": 0.569882777276826,
      "workLaneOccupancy": 0.9875,
      "deviceUsefulUtilization": 0.0625488739443228,
      "meanCycles": 13308.0,
      "p50Cycles": 13308,
      "p95Cycles": 13308,
      "maximumCycles": 13308,
      "firstOffset": 311982,
      "lastOffset": 325326
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r119_p120_c128_p128",
      "dimensions": {},
      "sampleCount": 243,
      "tileCount": 243,
      "phaseCount": 2,
      "phaseCycles": 13176,
      "phaseMs": 0.008784,
      "workCycles": 3193020,
      "averageActiveTiles": 242.3360655737705,
      "usefulCycles": 1850688.0,
      "estimatedWorkCycles": 3193020,
      "usefulUtilization": 0.5796042617960426,
      "workLaneOccupancy": 0.9916666666666667,
      "deviceUsefulUtilization": 0.09542052744119743,
      "meanCycles": 13140.0,
      "p50Cycles": 13140,
      "p95Cycles": 13140,
      "maximumCycles": 13140,
      "firstOffset": 40518,
      "lastOffset": 53694
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r82_p82_c192_p192",
      "dimensions": {},
      "sampleCount": 810,
      "tileCount": 810,
      "phaseCount": 28,
      "phaseCycles": 12828,
      "phaseMs": 0.008552,
      "workCycles": 10259460,
      "averageActiveTiles": 799.7708138447147,
      "usefulCycles": 6376320.0,
      "estimatedWorkCycles": 10259460,
      "usefulUtilization": 0.6215063950734249,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3376784479603042,
      "meanCycles": 12666.0,
      "p50Cycles": 12666,
      "p95Cycles": 12666,
      "maximumCycles": 12666,
      "firstOffset": 312222,
      "lastOffset": 325050
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r80_p80_c192_p192",
      "dimensions": {},
      "sampleCount": 486,
      "tileCount": 486,
      "phaseCount": 19,
      "phaseCycles": 12726,
      "phaseMs": 0.008484,
      "workCycles": 6155676,
      "averageActiveTiles": 483.7086280056577,
      "usefulCycles": 3732480.0,
      "estimatedWorkCycles": 6155676,
      "usefulUtilization": 0.6063477025106585,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.19924973863846013,
      "meanCycles": 12666.0,
      "p50Cycles": 12666,
      "p95Cycles": 12666,
      "maximumCycles": 12666,
      "firstOffset": 312072,
      "lastOffset": 324798
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c80_p80",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 7,
      "phaseCycles": 12534,
      "phaseMs": 0.008356,
      "workCycles": 146592,
      "averageActiveTiles": 11.69554810914313,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 146592,
      "usefulUtilization": 0.10914647456887143,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0008672064159399477,
      "meanCycles": 9162.0,
      "p50Cycles": 9162,
      "p95Cycles": 9162,
      "maximumCycles": 9162,
      "firstOffset": 630726,
      "lastOffset": 643260
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k80_c64_r91_r92",
      "dimensions": {},
      "sampleCount": 144,
      "tileCount": 144,
      "phaseCount": 10,
      "phaseCycles": 11700,
      "phaseMs": 0.0078,
      "workCycles": 1572480,
      "averageActiveTiles": 134.4,
      "usefulCycles": 942264.0,
      "estimatedWorkCycles": 1572480,
      "usefulUtilization": 0.5992216117216117,
      "workLaneOccupancy": 0.8890625,
      "deviceUsefulUtilization": 0.054711538461538464,
      "meanCycles": 10920.0,
      "p50Cycles": 10920,
      "p95Cycles": 10920,
      "maximumCycles": 10920,
      "firstOffset": 147834,
      "lastOffset": 159534
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k192_c96_r36_r38",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 26,
      "phaseCycles": 11652,
      "phaseMs": 0.007768,
      "workCycles": 4177440,
      "averageActiveTiles": 358.5169927909372,
      "usefulCycles": 1969920.0,
      "estimatedWorkCycles": 4177440,
      "usefulUtilization": 0.47156153050672184,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.11485246048448484,
      "meanCycles": 11604.0,
      "p50Cycles": 11604,
      "p95Cycles": 11604,
      "maximumCycles": 11604,
      "firstOffset": 267564,
      "lastOffset": 279216
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k192_c96_r36_r38",
      "dimensions": {},
      "sampleCount": 1080,
      "tileCount": 1080,
      "phaseCount": 27,
      "phaseCycles": 11550,
      "phaseMs": 0.0077,
      "workCycles": 12402720,
      "averageActiveTiles": 1073.8285714285714,
      "usefulCycles": 5588352.0,
      "estimatedWorkCycles": 12402720,
      "usefulUtilization": 0.45057471264367815,
      "workLaneOccupancy": 0.9981481481481481,
      "deviceUsefulUtilization": 0.32869565217391306,
      "meanCycles": 11484.0,
      "p50Cycles": 11484,
      "p95Cycles": 11484,
      "maximumCycles": 11484,
      "firstOffset": 267570,
      "lastOffset": 279120
    },
    {
      "name": "attention_merge_v72_p80_k768_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 47,
      "phaseCycles": 8688,
      "phaseMs": 0.005792,
      "workCycles": 3722400,
      "averageActiveTiles": 428.45303867403317,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 2501.6129032258063,
      "p50Cycles": 2514,
      "p95Cycles": 2514,
      "maximumCycles": 2514,
      "firstOffset": 230442,
      "lastOffset": 681774
    },
    {
      "name": "cast_f32_f16",
      "dimensions": {},
      "sampleCount": 1488,
      "tileCount": 1472,
      "phaseCount": 47,
      "phaseCycles": 8550,
      "phaseMs": 0.0057,
      "workCycles": 4586400,
      "averageActiveTiles": 536.421052631579,
      "usefulCycles": 420480.0,
      "estimatedWorkCycles": 4586400,
      "usefulUtilization": 0.09167974882260596,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.033409610983981694,
      "meanCycles": 3082.2580645161293,
      "p50Cycles": 3138,
      "p95Cycles": 3138,
      "maximumCycles": 3138,
      "firstOffset": 232956,
      "lastOffset": 682374
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k80_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 736,
      "tileCount": 736,
      "phaseCount": 39,
      "phaseCycles": 8400,
      "phaseMs": 0.0056,
      "workCycles": 856704,
      "averageActiveTiles": 101.98857142857143,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 856704,
      "usefulUtilization": 0.015316842223218288,
      "workLaneOccupancy": 0.8914402173913043,
      "deviceUsefulUtilization": 0.001061238354037267,
      "meanCycles": 1164.0,
      "p50Cycles": 1164,
      "p95Cycles": 1164,
      "maximumCycles": 1164,
      "firstOffset": 651522,
      "lastOffset": 659922
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r38_p38_c192_p192",
      "dimensions": {},
      "sampleCount": 360,
      "tileCount": 360,
      "phaseCount": 26,
      "phaseCycles": 7500,
      "phaseMs": 0.005,
      "workCycles": 2322000,
      "averageActiveTiles": 309.6,
      "usefulCycles": 1313280.0,
      "estimatedWorkCycles": 2322000,
      "usefulUtilization": 0.5655813953488372,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.11895652173913043,
      "meanCycles": 6450.0,
      "p50Cycles": 6450,
      "p95Cycles": 6450,
      "maximumCycles": 6450,
      "firstOffset": 250128,
      "lastOffset": 257628
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r35_p36_c192_p192",
      "dimensions": {},
      "sampleCount": 72,
      "tileCount": 72,
      "phaseCount": 10,
      "phaseCycles": 6612,
      "phaseMs": 0.004408,
      "workCycles": 446688,
      "averageActiveTiles": 67.55716878402905,
      "usefulCycles": 241920.0,
      "estimatedWorkCycles": 446688,
      "usefulUtilization": 0.5415860735009671,
      "workLaneOccupancy": 0.9722222222222222,
      "deviceUsefulUtilization": 0.024855993056103528,
      "meanCycles": 6204.0,
      "p50Cycles": 6204,
      "p95Cycles": 6204,
      "maximumCycles": 6204,
      "firstOffset": 249630,
      "lastOffset": 256242
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r36_p36_c192_p192",
      "dimensions": {},
      "sampleCount": 1008,
      "tileCount": 1008,
      "phaseCount": 22,
      "phaseCycles": 6600,
      "phaseMs": 0.0044,
      "workCycles": 5606496,
      "averageActiveTiles": 849.4690909090909,
      "usefulCycles": 3483648.0,
      "estimatedWorkCycles": 5606496,
      "usefulUtilization": 0.6213592233009708,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.3585770750988142,
      "meanCycles": 5562.0,
      "p50Cycles": 5562,
      "p95Cycles": 5562,
      "maximumCycles": 5562,
      "firstOffset": 249426,
      "lastOffset": 256026
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k96_c16_r1_r1",
      "dimensions": {},
      "sampleCount": 640,
      "tileCount": 640,
      "phaseCount": 40,
      "phaseCycles": 6570,
      "phaseMs": 0.00438,
      "workCycles": 852480,
      "averageActiveTiles": 129.75342465753425,
      "usefulCycles": 13122.0,
      "estimatedWorkCycles": 852480,
      "usefulUtilization": 0.015392736486486487,
      "workLaneOccupancy": 0.854296875,
      "deviceUsefulUtilization": 0.0013568344252531269,
      "meanCycles": 1332.0,
      "p50Cycles": 1332,
      "p95Cycles": 1332,
      "maximumCycles": 1332,
      "firstOffset": 669582,
      "lastOffset": 676152
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r40_p40_c128_p128",
      "dimensions": {},
      "sampleCount": 972,
      "tileCount": 972,
      "phaseCount": 28,
      "phaseCycles": 6258,
      "phaseMs": 0.004172,
      "workCycles": 4473144,
      "averageActiveTiles": 714.7881112176415,
      "usefulCycles": 2488320.0,
      "estimatedWorkCycles": 4473144,
      "usefulUtilization": 0.5562798783137766,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.270123806744758,
      "meanCycles": 4602.0,
      "p50Cycles": 4602,
      "p95Cycles": 4602,
      "maximumCycles": 4602,
      "firstOffset": 555282,
      "lastOffset": 561540
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r42_p42_c128_p128",
      "dimensions": {},
      "sampleCount": 405,
      "tileCount": 405,
      "phaseCount": 32,
      "phaseCycles": 5808,
      "phaseMs": 0.003872,
      "workCycles": 1863810,
      "averageActiveTiles": 320.9039256198347,
      "usefulCycles": 1088640.0,
      "estimatedWorkCycles": 1863810,
      "usefulUtilization": 0.5840938722294654,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.12733560905497665,
      "meanCycles": 4602.0,
      "p50Cycles": 4602,
      "p95Cycles": 4602,
      "maximumCycles": 4602,
      "firstOffset": 556296,
      "lastOffset": 562104
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r39_p40_c128_p128",
      "dimensions": {},
      "sampleCount": 81,
      "tileCount": 81,
      "phaseCount": 21,
      "phaseCycles": 5322,
      "phaseMs": 0.003548,
      "workCycles": 406782,
      "averageActiveTiles": 76.43404735062006,
      "usefulCycles": 202176.0,
      "estimatedWorkCycles": 406782,
      "usefulUtilization": 0.4970131421744325,
      "workLaneOccupancy": 0.975,
      "deviceUsefulUtilization": 0.02580755845301701,
      "meanCycles": 5022.0,
      "p50Cycles": 5022,
      "p95Cycles": 5022,
      "maximumCycles": 5022,
      "firstOffset": 556488,
      "lastOffset": 561810
    },
    {
      "name": "gemm_f8_init_small_rows_k160_c16_r144_r144",
      "dimensions": {},
      "sampleCount": 1472,
      "tileCount": 1472,
      "phaseCount": 1,
      "phaseCycles": 4998,
      "phaseMs": 0.003332,
      "workCycles": 6138240,
      "averageActiveTiles": 1228.139255702281,
      "usefulCycles": 3857868.0,
      "estimatedWorkCycles": 6138240,
      "usefulUtilization": 0.6284974194557398,
      "workLaneOccupancy": 0.9100118885869565,
      "deviceUsefulUtilization": 0.5243765984654731,
      "meanCycles": 4170.0,
      "p50Cycles": 4170,
      "p95Cycles": 4170,
      "maximumCycles": 4170,
      "firstOffset": 3162,
      "lastOffset": 8160
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r64_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 880,
      "tileCount": 880,
      "phaseCount": 13,
      "phaseCycles": 4650,
      "phaseMs": 0.0031,
      "workCycles": 1351680,
      "averageActiveTiles": 290.68387096774194,
      "usefulCycles": 450560.0,
      "estimatedWorkCycles": 1351680,
      "usefulUtilization": 0.3333333333333333,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.06582515194015895,
      "meanCycles": 1536.0,
      "p50Cycles": 1536,
      "p95Cycles": 1536,
      "maximumCycles": 1536,
      "firstOffset": 125928,
      "lastOffset": 130578
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r8_p8_c72_p80",
      "dimensions": {},
      "sampleCount": 1360,
      "tileCount": 1360,
      "phaseCount": 13,
      "phaseCycles": 4380,
      "phaseMs": 0.00292,
      "workCycles": 1721760,
      "averageActiveTiles": 393.09589041095893,
      "usefulCycles": 391680.0,
      "estimatedWorkCycles": 1721760,
      "usefulUtilization": 0.22748815165876776,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.06075044669446099,
      "meanCycles": 1266.0,
      "p50Cycles": 1266,
      "p95Cycles": 1266,
      "maximumCycles": 1266,
      "firstOffset": 124662,
      "lastOffset": 129042
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r9_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 1,
      "phaseCycles": 4182,
      "phaseMs": 0.002788,
      "workCycles": 132096,
      "averageActiveTiles": 31.58680057388809,
      "usefulCycles": 5184.0,
      "estimatedWorkCycles": 132096,
      "usefulUtilization": 0.03924418604651163,
      "workLaneOccupancy": 0.5625,
      "deviceUsefulUtilization": 0.0008421183956085085,
      "meanCycles": 4128.0,
      "p50Cycles": 4128,
      "p95Cycles": 4128,
      "maximumCycles": 4128,
      "firstOffset": 11964,
      "lastOffset": 16146
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o1_r0_p64_c16_p16",
      "dimensions": {},
      "sampleCount": 80,
      "tileCount": 80,
      "phaseCount": 11,
      "phaseCycles": 4056,
      "phaseMs": 0.002704,
      "workCycles": 141600,
      "averageActiveTiles": 34.9112426035503,
      "usefulCycles": 16000.0,
      "estimatedWorkCycles": 141600,
      "usefulUtilization": 0.11299435028248588,
      "workLaneOccupancy": 0.390625,
      "deviceUsefulUtilization": 0.0026798730812108737,
      "meanCycles": 1770.0,
      "p50Cycles": 1770,
      "p95Cycles": 1770,
      "maximumCycles": 1770,
      "firstOffset": 126384,
      "lastOffset": 130440
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c32_r2_r2",
      "dimensions": {},
      "sampleCount": 1940,
      "tileCount": 1300,
      "phaseCount": 89,
      "phaseCycles": 3912,
      "phaseMs": 0.002608,
      "workCycles": 3270840,
      "averageActiveTiles": 836.1042944785277,
      "usefulCycles": 61732.0,
      "estimatedWorkCycles": 3270840,
      "usefulUtilization": 0.01887343923884996,
      "workLaneOccupancy": 0.4971971649484536,
      "deviceUsefulUtilization": 0.010720219836400818,
      "meanCycles": 1686.0,
      "p50Cycles": 1686,
      "p95Cycles": 1686,
      "maximumCycles": 1686,
      "firstOffset": 700740,
      "lastOffset": 710448
    },
    {
      "name": "unpack_amp_to_row_major_f16_o1_r16_p16_c36_p36",
      "dimensions": {},
      "sampleCount": 1440,
      "tileCount": 1440,
      "phaseCount": 1,
      "phaseCycles": 3768,
      "phaseMs": 0.002512,
      "workCycles": 5322240,
      "averageActiveTiles": 1412.4840764331211,
      "usefulCycles": 414720.0,
      "estimatedWorkCycles": 5322240,
      "usefulUtilization": 0.07792207792207792,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.07477153143173636,
      "meanCycles": 3696.0,
      "p50Cycles": 3696,
      "p95Cycles": 3696,
      "maximumCycles": 3696,
      "firstOffset": 11946,
      "lastOffset": 15714
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r7_p7_c72_p80",
      "dimensions": {},
      "sampleCount": 112,
      "tileCount": 112,
      "phaseCount": 7,
      "phaseCycles": 3618,
      "phaseMs": 0.002412,
      "workCycles": 141792,
      "averageActiveTiles": 39.19071310116086,
      "usefulCycles": 28224.0,
      "estimatedWorkCycles": 141792,
      "usefulUtilization": 0.1990521327014218,
      "workLaneOccupancy": 0.9,
      "deviceUsefulUtilization": 0.0052995890114644175,
      "meanCycles": 1266.0,
      "p50Cycles": 1266,
      "p95Cycles": 1266,
      "maximumCycles": 1266,
      "firstOffset": 122604,
      "lastOffset": 127998
    },
    {
      "name": "layer_norm_apply",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 16,
      "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": 693786,
      "lastOffset": 696780
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 864,
      "tileCount": 864,
      "phaseCount": 36,
      "phaseCycles": 2520,
      "phaseMs": 0.00168,
      "workCycles": 725760,
      "averageActiveTiles": 288.0,
      "usefulCycles": 10368.0,
      "estimatedWorkCycles": 725760,
      "usefulUtilization": 0.014285714285714285,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.0027950310559006213,
      "meanCycles": 840.0,
      "p50Cycles": 840,
      "p95Cycles": 840,
      "maximumCycles": 840,
      "firstOffset": 618762,
      "lastOffset": 621282
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k128_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 989,
      "tileCount": 827,
      "phaseCount": 78,
      "phaseCycles": 2202,
      "phaseMs": 0.001468,
      "workCycles": 996912,
      "averageActiveTiles": 452.7302452316076,
      "usefulCycles": 15740.0,
      "estimatedWorkCycles": 996912,
      "usefulUtilization": 0.01578875567753222,
      "workLaneOccupancy": 0.4973458038422649,
      "deviceUsefulUtilization": 0.004856010346325475,
      "meanCycles": 1008.0,
      "p50Cycles": 1008,
      "p95Cycles": 1008,
      "maximumCycles": 1008,
      "firstOffset": 700734,
      "lastOffset": 709410
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c128_p128",
      "dimensions": {},
      "sampleCount": 1467,
      "tileCount": 1467,
      "phaseCount": 78,
      "phaseCycles": 1416,
      "phaseMs": 0.000944,
      "workCycles": 1910034,
      "averageActiveTiles": 1348.8940677966102,
      "usefulCycles": 93888.0,
      "estimatedWorkCycles": 1910034,
      "usefulUtilization": 0.04915514592933948,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.04504421518054532,
      "meanCycles": 1302.0,
      "p50Cycles": 1302,
      "p95Cycles": 1302,
      "maximumCycles": 1302,
      "firstOffset": 697506,
      "lastOffset": 698922
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o0_r1_p2_c64_p64",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 60,
      "phaseCycles": 1392,
      "phaseMs": 0.000928,
      "workCycles": 1065312,
      "averageActiveTiles": 765.3103448275862,
      "usefulCycles": 41472.0,
      "estimatedWorkCycles": 1065312,
      "usefulUtilization": 0.038929440389294405,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.020239880059970013,
      "meanCycles": 822.0,
      "p50Cycles": 822,
      "p95Cycles": 822,
      "maximumCycles": 822,
      "firstOffset": 684090,
      "lastOffset": 685482
    },
    {
      "name": "layer_norm_moments",
      "dimensions": {},
      "sampleCount": 32,
      "tileCount": 32,
      "phaseCount": 16,
      "phaseCycles": 1374,
      "phaseMs": 0.000916,
      "workCycles": 39168,
      "averageActiveTiles": 28.50655021834061,
      "usefulCycles": 2304.0,
      "estimatedWorkCycles": 39168,
      "usefulUtilization": 0.058823529411764705,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.001139168407062844,
      "meanCycles": 1224.0,
      "p50Cycles": 1224,
      "p95Cycles": 1224,
      "maximumCycles": 1224,
      "firstOffset": 691998,
      "lastOffset": 693372
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k64_c16_r2_r2",
      "dimensions": {},
      "sampleCount": 1296,
      "tileCount": 1296,
      "phaseCount": 60,
      "phaseCycles": 1092,
      "phaseMs": 0.000728,
      "workCycles": 870912,
      "averageActiveTiles": 797.5384615384615,
      "usefulCycles": 10368.0,
      "estimatedWorkCycles": 870912,
      "usefulUtilization": 0.011904761904761904,
      "workLaneOccupancy": 0.5,
      "deviceUsefulUtilization": 0.006450071667462972,
      "meanCycles": 672.0,
      "p50Cycles": 672,
      "p95Cycles": 672,
      "maximumCycles": 672,
      "firstOffset": 687588,
      "lastOffset": 688680
    }
  ],
  "samples": []
}
