{
  "tables": [
    "ANALYSIS_DETAILS",
    "ANALYSIS_FILE",
    "CUPTI_ACTIVITY_KIND_CUDA_EVENT",
    "CUPTI_ACTIVITY_KIND_KERNEL",
    "CUPTI_ACTIVITY_KIND_MEMSET",
    "CUPTI_ACTIVITY_KIND_RUNTIME",
    "CUPTI_ACTIVITY_KIND_SYNCHRONIZATION",
    "DIAGNOSTIC_EVENT",
    "ENUM_CUDA_DEV_MEM_EVENT_OPER",
    "ENUM_CUDA_FUNC_CACHE_CONFIG",
    "ENUM_CUDA_KERNEL_LAUNCH_TYPE",
    "ENUM_CUDA_MEMCPY_OPER",
    "ENUM_CUDA_MEMPOOL_OPER",
    "ENUM_CUDA_MEMPOOL_TYPE",
    "ENUM_CUDA_MEM_KIND",
    "ENUM_CUDA_SHARED_MEM_LIMIT_CONFIG",
    "ENUM_CUDA_UNIF_MEM_ACCESS_TYPE",
    "ENUM_CUDA_UNIF_MEM_MIGRATION",
    "ENUM_CUPTI_JIT_ENTRY_TYPE",
    "ENUM_CUPTI_JIT_OPERATION_TYPE",
    "ENUM_CUPTI_OVERHEAD_TYPE",
    "ENUM_CUPTI_STREAM_TYPE",
    "ENUM_CUPTI_SYNC_TYPE",
    "ENUM_DIAGNOSTIC_SEVERITY_LEVEL",
    "ENUM_DIAGNOSTIC_SOURCE_TYPE",
    "ENUM_DIAGNOSTIC_TIMESTAMP_SOURCE",
    "ENUM_NSYS_EVENT_CLASS",
    "ENUM_NSYS_EVENT_TYPE",
    "ENUM_NSYS_GENERIC_EVENT_FIELD_ETW_FLAGS",
    "ENUM_NSYS_GENERIC_EVENT_FIELD_ETW_PROPERTY",
    "ENUM_NSYS_GENERIC_EVENT_FIELD_ETW_TYPE",
    "ENUM_NSYS_GENERIC_EVENT_FIELD_TYPE",
    "ENUM_NSYS_GENERIC_EVENT_GROUP",
    "ENUM_NSYS_GENERIC_EVENT_SOURCE",
    "ENUM_SAMPLING_THREAD_STATE",
    "ENUM_SCHEDULING_THREAD_BLOCK",
    "ENUM_STACK_UNWIND_METHOD",
    "META_DATA_CAPTURE",
    "META_DATA_EXPORT",
    "NVTX_EVENTS",
    "OSRT_API",
    "OSRT_CALLCHAINS",
    "PROCESSES",
    "PROFILER_OVERHEAD",
    "ProcessStreams",
    "StringIds",
    "TARGET_INFO_CUDA_CONTEXT_INFO",
    "TARGET_INFO_CUDA_DEVICE",
    "TARGET_INFO_CUDA_STREAM",
    "TARGET_INFO_GPU",
    "TARGET_INFO_SESSION_START_TIME",
    "TARGET_INFO_SYSTEM_ENV",
    "ThreadNames"
  ],
  "activities": [
    {
      "start": 7849301,
      "end": 7859733,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 7,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 7944725,
      "end": 7950293,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 12,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 56,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 7982325,
      "end": 7984629,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 17,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 57,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8217300,
      "end": 8228084,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 37,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 8234772,
      "end": 8240308,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 42,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 59,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8247636,
      "end": 8250004,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 47,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 60,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8314356,
      "end": 8325012,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 67,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 8330132,
      "end": 8335604,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 72,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 62,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8338964,
      "end": 8341236,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 77,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 63,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8509139,
      "end": 8519763,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 97,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 8524851,
      "end": 8530387,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 102,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 65,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8536019,
      "end": 8538323,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 107,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 66,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8621619,
      "end": 8632051,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 127,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 8636243,
      "end": 8641747,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 132,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 68,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8646099,
      "end": 8648403,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 137,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 69,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8827442,
      "end": 8837842,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 157,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 8840914,
      "end": 8846578,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 162,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 71,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8850226,
      "end": 8852466,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 167,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 72,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8909618,
      "end": 8919986,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 187,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 8924274,
      "end": 8929842,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 192,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 74,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8933106,
      "end": 8935538,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 197,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 75,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 8991474,
      "end": 9001874,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 217,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 9005554,
      "end": 9011186,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 222,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 77,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 9014546,
      "end": 9016850,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 227,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 78,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 9072498,
      "end": 9082865,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 247,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 9086449,
      "end": 9092209,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 252,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 80,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 9095025,
      "end": 9097393,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 257,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 81,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 9170033,
      "end": 9180497,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 277,
      "globalPid": 281526382100480,
      "value": 0,
      "bytes": 33554432,
      "graphNodeId": null,
      "memKind": 2,
      "activity_table": "CUPTI_ACTIVITY_KIND_MEMSET"
    },
    {
      "start": 9183473,
      "end": 9189137,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 282,
      "globalPid": 281526382100480,
      "demangledName": 112,
      "shortName": 117,
      "mangledName": 112,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 168,
      "gridX": 2,
      "gridY": 64,
      "gridZ": 1,
      "blockX": 384,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 213268,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 83,
      "sharedMemoryExecuted": 233472,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 1,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 2,
      "clusterY": 1,
      "clusterZ": 1,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 16,
      "maxActiveClusters": 66,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    },
    {
      "start": 9195697,
      "end": 9198001,
      "deviceId": 0,
      "contextId": 1,
      "greenContextId": 0,
      "streamId": 13,
      "correlationId": 287,
      "globalPid": 281526382100480,
      "demangledName": 113,
      "shortName": 119,
      "mangledName": 113,
      "launchType": 0,
      "cacheConfig": 0,
      "registersPerThread": 26,
      "gridX": 512,
      "gridY": 1,
      "gridZ": 1,
      "blockX": 256,
      "blockY": 1,
      "blockZ": 1,
      "staticSharedMemory": 0,
      "dynamicSharedMemory": 0,
      "localMemoryPerThread": 0,
      "localMemoryTotal": 155713536,
      "gridId": 84,
      "sharedMemoryExecuted": 32768,
      "graphNodeId": null,
      "sharedMemoryLimitConfig": 0,
      "qmdBulkReleaseDone": null,
      "qmdPreexitDone": null,
      "qmdLastCtaDone": null,
      "graphId": null,
      "clusterX": 0,
      "clusterY": 0,
      "clusterZ": 0,
      "clusterSchedulingPolicy": 1,
      "maxPotentialClusterSize": 0,
      "maxActiveClusters": 0,
      "sharedMemoryRequestedPercentage": null,
      "tensorSizeMinusOneElements": "[]",
      "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL"
    }
  ],
  "string_ids": {
    "112": "nvjet_hsh_64x128_64x8_2x1_v_bz_TNN",
    "113": "elementwise_r2_w8_b256_gs_False_2ba83ece",
    "117": "nvjet_hsh_64x128_64x8_2x1_v_bz_1...",
    "119": "elementwise_r2_w8_b256_gs_Fals_2..."
  },
  "metadata_scope": "Only CUDA activity records and their referenced names. Process environment and unrelated string-table entries omitted.",
  "kernels": [
    {
      "name": "nvjet_hsh_64x128_64x8_2x1_v_bz_TNN",
      "graph_node_id": null,
      "count": 10,
      "count_per_invocation": 1.0,
      "median_duration_ns": 5568.0,
      "grid": {
        "gridX": 2,
        "gridY": 64,
        "gridZ": 1
      },
      "block": {
        "blockX": 384,
        "blockY": 1,
        "blockZ": 1
      },
      "registers_per_thread": 168,
      "static_shared_bytes": 0,
      "dynamic_shared_bytes": 213268
    },
    {
      "name": "elementwise_r2_w8_b256_gs_False_2ba83ece",
      "graph_node_id": null,
      "count": 10,
      "count_per_invocation": 1.0,
      "median_duration_ns": 2304.0,
      "grid": {
        "gridX": 512,
        "gridY": 1,
        "gridZ": 1
      },
      "block": {
        "blockX": 256,
        "blockY": 1,
        "blockZ": 1
      },
      "registers_per_thread": 26,
      "static_shared_bytes": 0,
      "dynamic_shared_bytes": 0
    }
  ],
  "kernel_grouping": "Name, graph node ID when present, launch dimensions and resource use. Equal names alone do not identify the same kernel.",
  "kernel_activity_count": 20,
  "sum_kernel_activity_ns": 79104,
  "sum_kernel_activity_ns_per_invocation": 7910.4,
  "duration_scope": "Sum of captured kernel durations / 10 invocations; overlapping kernels double-count work. This is not critical-path or wall-clock time.",
  "interpretation": "Captured profiler range only. Activity counts are not semantic operation counts; resolve library/internal graph kernels from correlation IDs."
}
