{
  "clockHz": 1500000000,
  "tileCount": 1472,
  "sampleCount": 187394,
  "matchedSampleCount": 2181,
  "profileSpanCycles": 10131456,
  "profileSpanMs": 6.754304,
  "atOffset": null,
  "groups": [
    {
      "name": "bias_gelu_f16",
      "dimensions": {},
      "sampleCount": 1458,
      "tileCount": 1458,
      "phaseCount": 35,
      "phaseCycles": 9078,
      "phaseMs": 0.006052,
      "workCycles": 12474648,
      "averageActiveTiles": 1374.1625908790481,
      "usefulCycles": 8628444.0,
      "estimatedWorkCycles": 12474648,
      "usefulUtilization": 0.6916783543712015,
      "workLaneOccupancy": 1.0,
      "deviceUsefulUtilization": 0.6457055159631024,
      "meanCycles": 8556.0,
      "p50Cycles": 8556,
      "p95Cycles": 8556,
      "maximumCycles": 8556,
      "firstOffset": 323346,
      "lastOffset": 332424
    },
    {
      "name": "copy_strided_u64",
      "dimensions": {},
      "sampleCount": 453,
      "tileCount": 453,
      "phaseCount": 24,
      "phaseCycles": 618,
      "phaseMs": 0.000412,
      "workCycles": 173952,
      "averageActiveTiles": 281.4757281553398,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 384.0,
      "p50Cycles": 384,
      "p95Cycles": 384,
      "maximumCycles": 384,
      "firstOffset": 323250,
      "lastOffset": 323868
    },
    {
      "name": "static_copy_u32",
      "dimensions": {},
      "sampleCount": 270,
      "tileCount": 261,
      "phaseCount": 26,
      "phaseCycles": 594,
      "phaseMs": 0.000396,
      "workCycles": 78570,
      "averageActiveTiles": 132.27272727272728,
      "usefulCycles": 0.0,
      "estimatedWorkCycles": 0,
      "usefulUtilization": null,
      "workLaneOccupancy": null,
      "deviceUsefulUtilization": null,
      "meanCycles": 291.0,
      "p50Cycles": 294,
      "p95Cycles": 294,
      "maximumCycles": 294,
      "firstOffset": 323244,
      "lastOffset": 323838
    }
  ],
  "samples": [
    {
      "physicalTile": 0,
      "offset": 323760,
      "duration": 8556,
      "phase": 82,
      "epoch": 1,
      "kind": "compute",
      "operation": "operation.20",
      "kernel": "bias_gelu_f16",
      "metadata": [
        {
          "name": "reason",
          "value": "OperatorKernel"
        },
        {
          "name": "value",
          "value": "422"
        },
        {
          "name": "kernelSpec",
          "value": "BiasGelu"
        },
        {
          "name": "outputElements",
          "value": "2152"
        },
        {
          "name": "input0Elements",
          "value": "2152"
        },
        {
          "name": "input1Elements",
          "value": "2152"
        },
        {
          "name": "usefulCycles",
          "value": "5918"
        },
        {
          "name": "physicalWorkCycles",
          "value": "5918"
        },
        {
          "name": "workEstimateBasis",
          "value": "bias add, MIX GeLU and optional FP8 conversion"
        },
        {
          "name": "invocations",
          "value": "1"
        },
        {
          "name": "uniformInvocations",
          "value": "true"
        },
        {
          "name": "arguments",
          "value": "[1, 2152]"
        }
      ]
    },
    {
      "physicalTile": 1,
      "offset": 323436,
      "duration": 8556,
      "phase": 71,
      "epoch": 1,
      "kind": "compute",
      "operation": "operation.20",
      "kernel": "bias_gelu_f16",
      "metadata": [
        {
          "name": "reason",
          "value": "OperatorKernel"
        },
        {
          "name": "value",
          "value": "422"
        },
        {
          "name": "kernelSpec",
          "value": "BiasGelu"
        },
        {
          "name": "outputElements",
          "value": "2152"
        },
        {
          "name": "input0Elements",
          "value": "2152"
        },
        {
          "name": "input1Elements",
          "value": "2152"
        },
        {
          "name": "usefulCycles",
          "value": "5918"
        },
        {
          "name": "physicalWorkCycles",
          "value": "5918"
        },
        {
          "name": "workEstimateBasis",
          "value": "bias add, MIX GeLU and optional FP8 conversion"
        },
        {
          "name": "invocations",
          "value": "1"
        },
        {
          "name": "uniformInvocations",
          "value": "true"
        },
        {
          "name": "arguments",
          "value": "[1, 2152]"
        }
      ]
    },
    {
      "physicalTile": 2,
      "offset": 323394,
      "duration": 8556,
      "phase": 74,
      "epoch": 1,
      "kind": "compute",
      "operation": "operation.20",
      "kernel": "bias_gelu_f16",
      "metadata": [
        {
          "name": "reason",
          "value": "OperatorKernel"
        },
        {
          "name": "value",
          "value": "422"
        },
        {
          "name": "kernelSpec",
          "value": "BiasGelu"
        },
        {
          "name": "outputElements",
          "value": "2152"
        },
        {
          "name": "input0Elements",
          "value": "2152"
        },
        {
          "name": "input1Elements",
          "value": "2152"
        },
        {
          "name": "usefulCycles",
          "value": "5918"
        },
        {
          "name": "physicalWorkCycles",
          "value": "5918"
        },
        {
          "name": "workEstimateBasis",
          "value": "bias add, MIX GeLU and optional FP8 conversion"
        },
        {
          "name": "invocations",
          "value": "1"
        },
        {
          "name": "uniformInvocations",
          "value": "true"
        },
        {
          "name": "arguments",
          "value": "[1, 2152]"
        }
      ]
    },
    {
      "physicalTile": 3,
      "offset": 323700,
      "duration": 8556,
      "phase": 73,
      "epoch": 1,
      "kind": "compute",
      "operation": "operation.20",
      "kernel": "bias_gelu_f16",
      "metadata": [
        {
          "name": "reason",
          "value": "OperatorKernel"
        },
        {
          "name": "value",
          "value": "422"
        },
        {
          "name": "kernelSpec",
          "value": "BiasGelu"
        },
        {
          "name": "outputElements",
          "value": "2152"
        },
        {
          "name": "input0Elements",
          "value": "2152"
        },
        {
          "name": "input1Elements",
          "value": "2152"
        },
        {
          "name": "usefulCycles",
          "value": "5918"
        },
        {
          "name": "physicalWorkCycles",
          "value": "5918"
        },
        {
          "name": "workEstimateBasis",
          "value": "bias add, MIX GeLU and optional FP8 conversion"
        },
        {
          "name": "invocations",
          "value": "1"
        },
        {
          "name": "uniformInvocations",
          "value": "true"
        },
        {
          "name": "arguments",
          "value": "[1, 2152]"
        }
      ]
    }
  ]
}
