{ "tables": [ "ANALYSIS_DETAILS", "ANALYSIS_FILE", "CUDA_GRAPH_NODE_EVENTS", "CUPTI_ACTIVITY_KIND_CUDA_EVENT", "CUPTI_ACTIVITY_KIND_KERNEL", "CUPTI_ACTIVITY_KIND_MEMCPY", "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", "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": 9290634, "end": 9291434, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 8, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9339274, "end": 9342058, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 8, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9342154, "end": 9349226, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 8, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9349354, "end": 9390634, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 8, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9390730, "end": 9394186, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 8, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9394282, "end": 9398026, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 8, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9412778, "end": 9413546, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 17, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9422953, "end": 9425705, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 17, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9425801, "end": 9432937, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 17, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9433033, "end": 9474505, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 17, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9474601, "end": 9478089, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 17, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9478185, "end": 9481961, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 17, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9492169, "end": 9492969, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 26, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9501801, "end": 9504553, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 26, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9504649, "end": 9511689, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 26, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9511817, "end": 9553033, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 26, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9553161, "end": 9556617, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 26, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9556713, "end": 9560489, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 26, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9569929, "end": 9570729, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 35, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9580840, "end": 9583560, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 35, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9583688, "end": 9590792, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 35, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9590920, "end": 9632264, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 35, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9632360, "end": 9635816, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 35, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9635912, "end": 9639688, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 35, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9649128, "end": 9649896, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 44, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9658344, "end": 9661096, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 44, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9661192, "end": 9668328, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 44, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9668456, "end": 9709672, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 44, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9709800, "end": 9713256, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 44, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9713352, "end": 9717128, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 44, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9726504, "end": 9727304, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 53, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9735111, "end": 9737863, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 53, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9737959, "end": 9745031, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 53, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9745159, "end": 9786471, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 53, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9786567, "end": 9790023, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 53, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9790151, "end": 9793895, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 53, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9803015, "end": 9803815, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 62, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9809863, "end": 9812615, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 62, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9812711, "end": 9819751, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 62, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9819847, "end": 9861063, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 62, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9861159, "end": 9864615, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 62, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9864711, "end": 9868487, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 62, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9877575, "end": 9878375, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 71, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9885830, "end": 9888582, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 71, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9888678, "end": 9895750, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 71, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9895846, "end": 9937158, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 71, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9937286, "end": 9940742, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 71, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9940838, "end": 9944614, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 71, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9963430, "end": 9964198, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 80, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 9971782, "end": 9974534, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 80, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9974662, "end": 9981702, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 80, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 9981830, "end": 10023046, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 80, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 10023142, "end": 10026598, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 80, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 10026694, "end": 10030470, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 80, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 10039845, "end": 10040613, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 89, "globalPid": 281487022751744, "bytes": 8, "copyKind": 1, "deprecatedSrcId": null, "srcKind": 1, "dstKind": 2, "srcDeviceId": null, "srcContextId": null, "dstDeviceId": null, "dstContextId": null, "migrationCause": null, "graphNodeId": 3264175144960, "virtualAddress": null, "copyCount": 1, "activity_table": "CUPTI_ACTIVITY_KIND_MEMCPY" }, { "start": 10048709, "end": 10051429, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 89, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 879, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144961, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 10051557, "end": 10058597, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 89, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 3072, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 880, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144962, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 10058725, "end": 10099845, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 89, "globalPid": 281487022751744, "demangledName": 57, "shortName": 63, "mangledName": 57, "launchType": 0, "cacheConfig": 0, "registersPerThread": 121, "gridX": 16, "gridY": 16, "gridZ": 1, "blockX": 128, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 12672, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 881, "sharedMemoryExecuted": 135168, "graphNodeId": 3264175144966, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 10099941, "end": 10103429, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 89, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 882, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144964, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" }, { "start": 10103525, "end": 10107269, "deviceId": 0, "contextId": 1, "greenContextId": 0, "streamId": 13, "correlationId": 89, "globalPid": 281487022751744, "demangledName": 56, "shortName": 56, "mangledName": 56, "launchType": 0, "cacheConfig": 0, "registersPerThread": 16, "gridX": 4096, "gridY": 1, "gridZ": 1, "blockX": 256, "blockY": 1, "blockZ": 1, "staticSharedMemory": 0, "dynamicSharedMemory": 0, "localMemoryPerThread": 0, "localMemoryTotal": 155713536, "gridId": 883, "sharedMemoryExecuted": 32768, "graphNodeId": 3264175144965, "sharedMemoryLimitConfig": 0, "qmdBulkReleaseDone": null, "qmdPreexitDone": null, "qmdLastCtaDone": null, "graphId": 760, "clusterX": 0, "clusterY": 0, "clusterZ": 0, "clusterSchedulingPolicy": 1, "maxPotentialClusterSize": 0, "maxActiveClusters": 0, "sharedMemoryRequestedPercentage": null, "tensorSizeMinusOneElements": "[]", "activity_table": "CUPTI_ACTIVITY_KIND_KERNEL" } ], "string_ids": { "56": "k", "57": "sm80_xmma_gemm_f32f32_f32f32_f32_nn_n_tilesize64x64x8_stage3_warpsize1x4x1_ffma_aligna4_alignc4_execute_kernel__5x_cublas", "63": "sm80_xmma_gemm_f32f32_f32f32_f_1..." }, "metadata_scope": "Only CUDA activity records and their referenced names. Process environment and unrelated string-table entries omitted.", "kernels": [ { "name": "k", "graph_node_id": 3264175144961, "count": 10, "count_per_invocation": 1.0, "median_duration_ns": 2752.0, "grid": { "gridX": 3072, "gridY": 1, "gridZ": 1 }, "block": { "blockX": 256, "blockY": 1, "blockZ": 1 }, "registers_per_thread": 16, "static_shared_bytes": 0, "dynamic_shared_bytes": 0 }, { "name": "k", "graph_node_id": 3264175144962, "count": 10, "count_per_invocation": 1.0, "median_duration_ns": 7072.0, "grid": { "gridX": 3072, "gridY": 1, "gridZ": 1 }, "block": { "blockX": 256, "blockY": 1, "blockZ": 1 }, "registers_per_thread": 16, "static_shared_bytes": 0, "dynamic_shared_bytes": 0 }, { "name": "sm80_xmma_gemm_f32f32_f32f32_f32_nn_n_tilesize64x64x8_stage3_warpsize1x4x1_ffma_aligna4_alignc4_execute_kernel__5x_cublas", "graph_node_id": 3264175144966, "count": 10, "count_per_invocation": 1.0, "median_duration_ns": 41248.0, "grid": { "gridX": 16, "gridY": 16, "gridZ": 1 }, "block": { "blockX": 128, "blockY": 1, "blockZ": 1 }, "registers_per_thread": 121, "static_shared_bytes": 0, "dynamic_shared_bytes": 12672 }, { "name": "k", "graph_node_id": 3264175144964, "count": 10, "count_per_invocation": 1.0, "median_duration_ns": 3456.0, "grid": { "gridX": 4096, "gridY": 1, "gridZ": 1 }, "block": { "blockX": 256, "blockY": 1, "blockZ": 1 }, "registers_per_thread": 16, "static_shared_bytes": 0, "dynamic_shared_bytes": 0 }, { "name": "k", "graph_node_id": 3264175144965, "count": 10, "count_per_invocation": 1.0, "median_duration_ns": 3776.0, "grid": { "gridX": 4096, "gridY": 1, "gridZ": 1 }, "block": { "blockX": 256, "blockY": 1, "blockZ": 1 }, "registers_per_thread": 16, "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": 50, "sum_kernel_activity_ns": 583232, "sum_kernel_activity_ns_per_invocation": 58323.2, "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." }