Sylvain Jeaugey
19ab67d172
Optimize CUDA graph launch; avoid launching a CPU callback for intra-node operations. Simplify kernel common code to improve the latency of send/recv operations. Strengthen CUDA streams semantics. Change NET API to v6, to add dmabuf support. Add ncclGetLastError() function. Add ncclRemoteError code and use it for remote network errors. Support the use of a different NCCL_NET parameter per communicator. Add support for SHM and P2P transfers using cudaMemcpy.
112 řádky
4.6 KiB
C
112 řádky
4.6 KiB
C
/*************************************************************************
|
|
* Copyright (c) 2016-2022, NVIDIA CORPORATION. All rights reserved.
|
|
*
|
|
* See LICENSE.txt for license information
|
|
************************************************************************/
|
|
|
|
#ifndef NCCL_GRAPH_H_
|
|
#define NCCL_GRAPH_H_
|
|
|
|
#include "nccl.h"
|
|
#include "devcomm.h"
|
|
#include <limits.h>
|
|
#include <stdlib.h>
|
|
#include <ctype.h>
|
|
#include <stdio.h>
|
|
#include <sched.h>
|
|
|
|
ncclResult_t ncclTopoCudaPath(int cudaDev, char** path);
|
|
|
|
struct ncclTopoSystem;
|
|
// Build the topology
|
|
ncclResult_t ncclTopoGetSystem(struct ncclComm* comm, struct ncclTopoSystem** system);
|
|
ncclResult_t ncclTopoSortSystem(struct ncclTopoSystem* system);
|
|
ncclResult_t ncclTopoPrint(struct ncclTopoSystem* system);
|
|
|
|
ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm* comm);
|
|
void ncclTopoFree(struct ncclTopoSystem* system);
|
|
ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm* comm);
|
|
ncclResult_t ncclTopoComputeP2pChannels(struct ncclComm* comm);
|
|
ncclResult_t ncclTopoGetNvbGpus(struct ncclTopoSystem* system, int rank, int* nranks, int** ranks);
|
|
|
|
// Query topology
|
|
ncclResult_t ncclTopoGetNetDev(struct ncclComm* comm, int rank, struct ncclTopoGraph* graph, int channelId, int peerRank, int* net, int* proxyRank);
|
|
ncclResult_t ncclTopoCheckP2p(struct ncclTopoSystem* system, int64_t id1, int64_t id2, int* p2p, int *read, int* intermediateRank);
|
|
ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* topo, int64_t busId, int netDev, int read, int* useGdr);
|
|
int ncclPxnDisable(struct ncclComm* comm);
|
|
ncclResult_t ncclTopoGetPxnRanks(struct ncclComm* comm, int** intermediateRanks, int* nranks);
|
|
ncclResult_t ncclTopoGetLocalRank(struct ncclTopoSystem* system, int rank, int* localRank);
|
|
|
|
// Find CPU affinity
|
|
ncclResult_t ncclTopoGetCpuAffinity(struct ncclTopoSystem* system, int rank, cpu_set_t* affinity);
|
|
|
|
#define NCCL_TOPO_CPU_ARCH_X86 1
|
|
#define NCCL_TOPO_CPU_ARCH_POWER 2
|
|
#define NCCL_TOPO_CPU_ARCH_ARM 3
|
|
#define NCCL_TOPO_CPU_VENDOR_INTEL 1
|
|
#define NCCL_TOPO_CPU_VENDOR_AMD 2
|
|
#define NCCL_TOPO_CPU_VENDOR_ZHAOXIN 3
|
|
#define NCCL_TOPO_CPU_TYPE_BDW 1
|
|
#define NCCL_TOPO_CPU_TYPE_SKL 2
|
|
#define NCCL_TOPO_CPU_TYPE_YONGFENG 1
|
|
ncclResult_t ncclTopoCpuType(struct ncclTopoSystem* system, int* arch, int* vendor, int* model);
|
|
ncclResult_t ncclTopoGetNetCount(struct ncclTopoSystem* system, int* count);
|
|
ncclResult_t ncclTopoGetLocalNet(struct ncclTopoSystem* system, int rank, int* id);
|
|
|
|
#define NCCL_TOPO_MAX_NODES 256
|
|
|
|
// Init search. Needs to be done before calling ncclTopoCompute
|
|
ncclResult_t ncclTopoSearchInit(struct ncclTopoSystem* system);
|
|
|
|
#define NCCL_TOPO_PATTERN_BALANCED_TREE 1 // Spread NIC traffic between two GPUs (Tree parent + one child on first GPU, second child on second GPU)
|
|
#define NCCL_TOPO_PATTERN_SPLIT_TREE 2 // Spread NIC traffic between two GPUs (Tree parent on first GPU, tree children on the second GPU)
|
|
#define NCCL_TOPO_PATTERN_TREE 3 // All NIC traffic going to/from the same GPU
|
|
#define NCCL_TOPO_PATTERN_RING 4 // Ring
|
|
struct ncclTopoGraph {
|
|
// Input / output
|
|
int id; // ring : 0, tree : 1, collnet : 2
|
|
int pattern;
|
|
int crossNic;
|
|
int collNet;
|
|
int minChannels;
|
|
int maxChannels;
|
|
// Output
|
|
int nChannels;
|
|
float speedIntra;
|
|
float speedInter;
|
|
float latencyInter;
|
|
int typeIntra;
|
|
int typeInter;
|
|
int sameChannels;
|
|
int nHops;
|
|
int intra[MAXCHANNELS*NCCL_TOPO_MAX_NODES];
|
|
int inter[MAXCHANNELS*2];
|
|
};
|
|
ncclResult_t ncclTopoCompute(struct ncclTopoSystem* system, struct ncclTopoGraph* graph);
|
|
|
|
ncclResult_t ncclTopoPrintGraph(struct ncclTopoSystem* system, struct ncclTopoGraph* graph);
|
|
ncclResult_t ncclTopoDumpGraphs(struct ncclTopoSystem* system, int ngraphs, struct ncclTopoGraph** graphs);
|
|
|
|
struct ncclTopoRanks {
|
|
int ringRecv[MAXCHANNELS];
|
|
int ringSend[MAXCHANNELS];
|
|
int ringPrev[MAXCHANNELS];
|
|
int ringNext[MAXCHANNELS];
|
|
int treeToParent[MAXCHANNELS];
|
|
int treeToChild0[MAXCHANNELS];
|
|
int treeToChild1[MAXCHANNELS];
|
|
};
|
|
|
|
ncclResult_t ncclTopoPreset(struct ncclComm* comm,
|
|
struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph,
|
|
struct ncclTopoRanks* topoRanks);
|
|
|
|
ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePatterns,
|
|
struct ncclTopoRanks** allTopoRanks, int* rings, struct ncclTopoGraph* collNetGraph);
|
|
|
|
ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCompCap, struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph);
|
|
#include "info.h"
|
|
ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int protocol, int numPipeOps, float* time);
|
|
|
|
#endif
|