Ke Wen
e51f2a83da
Add new API for creating a reduction operation which multiplies the input by a rank-specific scalar before doing an inter-rank summation (see: ncclRedOpCreatePreMulSum).
Improve CollNet (SHARP) performance of ncclAllReduce when captured in a CUDA Graph via user buffer registration.
Add environment variable NCCL_NET_PLUGIN="<suffix>" to allow user to choose among multiple NCCL net plugins by substituting into "libnccl-net-<suffix>.so".
Fix memory leak of NVB connections.
Fix topology detection of IB Virtual Functions (SR-IOV).
[ROCm/rccl commit: e11238b302]
60 lines
1.2 KiB
C
60 lines
1.2 KiB
C
/*************************************************************************
|
|
* Copyright (c) 2019-2021, NVIDIA CORPORATION. All rights reserved.
|
|
*
|
|
* See LICENSE.txt for license information
|
|
************************************************************************/
|
|
|
|
#ifndef NCCL_INFO_H_
|
|
#define NCCL_INFO_H_
|
|
|
|
#include "nccl.h"
|
|
#include "devcomm.h"
|
|
#include "collectives.h"
|
|
|
|
typedef enum {
|
|
ncclPatternRing,
|
|
ncclPatternRingTwice,
|
|
ncclPatternPipelineFrom,
|
|
ncclPatternPipelineTo,
|
|
ncclPatternTreeUp,
|
|
ncclPatternTreeDown,
|
|
ncclPatternTreeUpDown,
|
|
ncclPatternCollTreeUpDown
|
|
} ncclPattern_t;
|
|
|
|
// Used to pass NCCL call information between functions
|
|
struct ncclInfo {
|
|
ncclFunc_t coll;
|
|
const char* opName;
|
|
// NCCL Coll Args
|
|
const void* sendbuff;
|
|
void* recvbuff;
|
|
size_t count;
|
|
ncclDataType_t datatype;
|
|
ncclRedOp_t op;
|
|
int root;
|
|
ncclComm_t comm;
|
|
cudaStream_t stream;
|
|
// Algorithm details
|
|
int chunkSteps;
|
|
int sliceSteps;
|
|
// Computed later
|
|
ncclDevRedOpFull opFull;
|
|
int algorithm;
|
|
int protocol;
|
|
ncclPattern_t pattern;
|
|
int nChannels;
|
|
int nThreads;
|
|
size_t nBytes;
|
|
int nstepsPerLoop;
|
|
int nchunksPerLoop;
|
|
ssize_t sendbytes;
|
|
ssize_t recvbytes;
|
|
int recvChunkSize;
|
|
int sendChunkSize;
|
|
uint32_t delta;
|
|
int channelId;
|
|
};
|
|
|
|
#endif
|