NCCL API

The following sections describe the NCCL methods and operations.

• Communicator Creation and Management Functions
  • ncclGetLastError
  • ncclGetErrorString
  • ncclGetVersion
  • ncclGetUniqueId
  • ncclCommInitRank
  • ncclCommInitAll
  • ncclCommInitRankConfig
  • ncclCommInitRankScalable
  • ncclCommSplit
  • ncclCommShrink
  • ncclCommGetUniqueId
  • ncclCommGrow
  • ncclCommRevoke
  • ncclCommFinalize
  • ncclCommDestroy
  • ncclCommAbort
  • ncclCommGetAsyncError
  • ncclCommCount
  • ncclCommCuDevice
  • ncclCommUserRank
  • ncclCommRegister
  • ncclCommDeregister
  • ncclCommWindowRegister
  • ncclCommWindowDeregister
  • ncclMemAlloc
  • ncclMemFree
  • ncclCommSuspend
  • ncclCommResume
  • ncclCommMemStats
• Collective Communication Functions
  • ncclAllReduce
  • ncclBroadcast
  • ncclReduce
  • ncclAllGather
  • ncclReduceScatter
  • ncclAlltoAll
  • ncclGather
  • ncclScatter
• Group Calls
  • ncclGroupStart
  • ncclGroupEnd
  • ncclGroupSimulateEnd
• Point To Point Communication Functions
  • Two-Sided Point-to-Point Operations
    • ncclSend
    • ncclRecv
  • One-Sided Point-to-Point Operations (RMA)
    • ncclPutSignal
    • ncclSignal
    • ncclWaitSignal
• Types
  • ncclComm_t
  • ncclResult_t
  • ncclDataType_t
  • ncclRedOp_t
  • ncclScalarResidence_t
  • ncclConfig_t
  • ncclSimInfo_t
  • ncclCommMemStat_t
  • ncclWindow_t
• User Defined Reduction Operators
  • ncclRedOpCreatePreMulSum
  • ncclRedOpDestroy
• NCCL API Supported Flags
  • Window Registration Flags
  • NCCL Communicator CTA Policy Flags
  • Communicator Shrink Flags
• Device API
  • Device API – Host-Side Setup
    • Host-Side Setup
      • ncclDevComm
      • ncclDevCommCreate
      • ncclDevCommDestroy
      • ncclDevCommRequirements
      • ncclCommQueryProperties
      • ncclCommProperties_t
      • ncclGinType_t
      • ncclGinConnectionType_t
    • Host-Accessible Device Pointer Functions
      • ncclGetLsaMultimemDevicePointer
      • ncclGetMultimemDevicePointer
      • ncclGetLsaDevicePointer
      • ncclGetPeerDevicePointer
  • Device API – Memory and LSA
    • LSA
      • ncclLsaBarrierSession
      • ncclGetPeerPointer
      • ncclGetLsaPointer
      • ncclGetLocalPointer
    • Multimem
      • ncclGetLsaMultimemPointer
  • Device API – GIN
    • GIN
      • ncclGin
      • Signals and Counters
      • ncclGinBarrierSession
  • Device API – Remote Reduce and Copy: Building Blocks for Custom Communication Kernels
    • Compile-Time Requirements
    • API Overview
    • ReduceSum — N Sources to One Destination
    • Copy (Broadcast) — One Source to N Destinations
    • ReduceSumCopy
      • LSA ReduceSumCopy (ncclLsaReduceSumCopy)
      • Multimem ReduceSumCopy (ncclMultimemReduceSumCopy)
      • Mixed LSA and Multimem ReduceSumCopy
    • Lambda-Based (Custom Layouts)
    • Custom Reduction Operators