{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 8021,
  "matchedSampleCount": 529,
  "profileSpanCycles": 20910,
  "profileSpanMs": 0.01394,
  "atOffset": null,
  "groups": [
    {
      "name": "cast_f16_f8",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 4,
      "phaseCount": 2,
      "phaseCycles": 5922,
      "phaseMs": 0.003948,
      "workCycles": 23676,
      "averageActiveTiles": 3.9979736575481257,
      "usefulCycles": 720.0,
      "estimatedWorkCycles": 18396,
      "usefulUtilization": 0.03913894324853229,
      "workLaneOccupancy": 0.75,
      "deviceUsefulUtilization": 0.00008259548037531386,
      "meanCycles": 2959.5,
      "p50Cycles": 2646,
      "p95Cycles": 3276,
      "maximumCycles": 3276,
      "firstOffset": 0,
      "lastOffset": 5922
    },
    {
      "name": "fill_zero_u64",
      "dimensions": {},
      "sampleCount": 152,
      "tileCount": 64,
      "phaseCount": 11,
      "phaseCycles": 4110,
      "phaseMs": 0.00274,
      "workCycles": 38712,
      "averageActiveTiles": 9.418978102189781,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 254.68421052631578,
      "p50Cycles": 246,
      "p95Cycles": 330,
      "maximumCycles": 330,
      "firstOffset": 5916,
      "lastOffset": 16488
    },
    {
      "name": "gemm_f8_init_large_rows_interleaved_k96_c16_r2_r3",
      "dimensions": {},
      "sampleCount": 40,
      "tileCount": 40,
      "phaseCount": 6,
      "phaseCycles": 3582,
      "phaseMs": 0.002388,
      "workCycles": 36000,
      "averageActiveTiles": 10.050251256281408,
      "usefulCycles": 641.25,
      "estimatedWorkCycles": 36000,
      "usefulUtilization": 0.0178125,
      "workLaneOccupancy": 0.4453125,
      "deviceUsefulUtilization": 0.00012161691610225038,
      "meanCycles": 900.0,
      "p50Cycles": 900,
      "p95Cycles": 900,
      "maximumCycles": 900,
      "firstOffset": 7632,
      "lastOffset": 11586
    },
    {
      "name": "rearrange_row_major_to_amp_f16_o2_r0_p64_c16_p16_b64x16",
      "dimensions": {},
      "sampleCount": 20,
      "tileCount": 20,
      "phaseCount": 4,
      "phaseCycles": 2568,
      "phaseMs": 0.001712,
      "workCycles": 36120,
      "averageActiveTiles": 14.065420560747663,
      "usefulCycles": 3040.0,
      "estimatedWorkCycles": 36120,
      "usefulUtilization": 0.08416389811738649,
      "workLaneOccupancy": 0.296875,
      "deviceUsefulUtilization": 0.0008042123797914127,
      "meanCycles": 1806.0,
      "p50Cycles": 1806,
      "p95Cycles": 1806,
      "maximumCycles": 1806,
      "firstOffset": 8118,
      "lastOffset": 10686
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 57,
      "tileCount": 36,
      "phaseCount": 11,
      "phaseCycles": 2388,
      "phaseMs": 0.001592,
      "workCycles": 22962,
      "averageActiveTiles": 9.615577889447236,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 402.8421052631579,
      "p50Cycles": 378,
      "p95Cycles": 630,
      "maximumCycles": 630,
      "firstOffset": 6162,
      "lastOffset": 20910
    },
    {
      "name": "attention_merge_v72_p80_k64_out16_f16",
      "dimensions": {},
      "sampleCount": 64,
      "tileCount": 64,
      "phaseCount": 9,
      "phaseCycles": 2016,
      "phaseMs": 0.001344,
      "workCycles": 102144,
      "averageActiveTiles": 50.666666666666664,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 1596.0,
      "p50Cycles": 1596,
      "p95Cycles": 1596,
      "maximumCycles": 1596,
      "firstOffset": 16842,
      "lastOffset": 18858
    },
    {
      "name": "attention_softmax_d72_p64_tail_f16",
      "dimensions": {},
      "sampleCount": 64,
      "tileCount": 64,
      "phaseCount": 7,
      "phaseCycles": 1728,
      "phaseMs": 0.001152,
      "workCycles": 85608,
      "averageActiveTiles": 49.541666666666664,
      "usefulCycles": 3128.0,
      "estimatedWorkCycles": 85608,
      "usefulUtilization": 0.03653864124848145,
      "workLaneOccupancy": 0.92,
      "deviceUsefulUtilization": 0.0012297453703703704,
      "meanCycles": 1337.625,
      "p50Cycles": 1338,
      "p95Cycles": 1338,
      "maximumCycles": 1338,
      "firstOffset": 12156,
      "lastOffset": 13884
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k32_c32_r3_r4",
      "dimensions": {},
      "sampleCount": 16,
      "tileCount": 16,
      "phaseCount": 3,
      "phaseCycles": 1140,
      "phaseMs": 0.00076,
      "workCycles": 17184,
      "averageActiveTiles": 15.073684210526316,
      "usefulCycles": 608.0,
      "estimatedWorkCycles": 17184,
      "usefulUtilization": 0.035381750465549346,
      "workLaneOccupancy": 0.59375,
      "deviceUsefulUtilization": 0.00036231884057971015,
      "meanCycles": 1074.0,
      "p50Cycles": 1074,
      "p95Cycles": 1074,
      "maximumCycles": 1074,
      "firstOffset": 15108,
      "lastOffset": 16248
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k32_c32_r3_r4",
      "dimensions": {},
      "sampleCount": 24,
      "tileCount": 24,
      "phaseCount": 3,
      "phaseCycles": 1092,
      "phaseMs": 0.000728,
      "workCycles": 25632,
      "averageActiveTiles": 23.47252747252747,
      "usefulCycles": 684.0,
      "estimatedWorkCycles": 25632,
      "usefulUtilization": 0.026685393258426966,
      "workLaneOccupancy": 0.59375,
      "deviceUsefulUtilization": 0.0004255255613951266,
      "meanCycles": 1068.0,
      "p50Cycles": 1068,
      "p95Cycles": 1068,
      "maximumCycles": 1068,
      "firstOffset": 15108,
      "lastOffset": 16200
    },
    {
      "name": "gemm_f16_init_large_rows_interleaved_k32_c16_r3_r4",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 1,
      "phaseCycles": 1044,
      "phaseMs": 0.000696,
      "workCycles": 5856,
      "averageActiveTiles": 5.609195402298851,
      "usefulCycles": 76.0,
      "estimatedWorkCycles": 5856,
      "usefulUtilization": 0.012978142076502733,
      "workLaneOccupancy": 0.296875,
      "deviceUsefulUtilization": 0.000049454439446943195,
      "meanCycles": 732.0,
      "p50Cycles": 732,
      "p95Cycles": 732,
      "maximumCycles": 732,
      "firstOffset": 14862,
      "lastOffset": 15906
    },
    {
      "name": "gemm_f16_init_small_rows_interleaved_k32_c16_r3_r4",
      "dimensions": {},
      "sampleCount": 12,
      "tileCount": 12,
      "phaseCount": 3,
      "phaseCycles": 1008,
      "phaseMs": 0.000672,
      "workCycles": 8784,
      "averageActiveTiles": 8.714285714285714,
      "usefulCycles": 85.5,
      "estimatedWorkCycles": 8784,
      "usefulUtilization": 0.00973360655737705,
      "workLaneOccupancy": 0.296875,
      "deviceUsefulUtilization": 0.00005762325310559006,
      "meanCycles": 732.0,
      "p50Cycles": 732,
      "p95Cycles": 732,
      "maximumCycles": 732,
      "firstOffset": 14850,
      "lastOffset": 15858
    },
    {
      "name": "gemm_f8_init_small_rows_interleaved_k96_c16_r2_r3",
      "dimensions": {},
      "sampleCount": 8,
      "tileCount": 8,
      "phaseCount": 2,
      "phaseCycles": 912,
      "phaseMs": 0.000608,
      "workCycles": 7200,
      "averageActiveTiles": 7.894736842105263,
      "usefulCycles": 85.5,
      "estimatedWorkCycles": 7200,
      "usefulUtilization": 0.011875,
      "workLaneOccupancy": 0.4453125,
      "deviceUsefulUtilization": 0.00006368885869565217,
      "meanCycles": 900.0,
      "p50Cycles": 900,
      "p95Cycles": 900,
      "maximumCycles": 900,
      "firstOffset": 7620,
      "lastOffset": 8532
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 52,
      "tileCount": 28,
      "phaseCount": 7,
      "phaseCycles": 810,
      "phaseMs": 0.00054,
      "workCycles": 10968,
      "averageActiveTiles": 13.540740740740741,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 210.92307692307693,
      "p50Cycles": 198,
      "p95Cycles": 288,
      "maximumCycles": 288,
      "firstOffset": 8484,
      "lastOffset": 17094
    },
    {
      "name": "copy_u64",
      "dimensions": {},
      "sampleCount": 4,
      "tileCount": 4,
      "phaseCount": 1,
      "phaseCycles": 276,
      "phaseMs": 0.000184,
      "workCycles": 1080,
      "averageActiveTiles": 3.9130434782608696,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 270.0,
      "p50Cycles": 270,
      "p95Cycles": 270,
      "maximumCycles": 270,
      "firstOffset": 12228,
      "lastOffset": 12504
    }
  ],
  "samples": []
}
