AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.18.1-1

Browse Source 
Add support for IB SHARP to NVLS (NVLink SHARP algorithm).
Add NVLS+Tree algorithm.
Add support for memory management using cuMem* functions.
Use all NICs for Send/Receive operations on systems with more than
one NIC per GPU (#804).
Add ncclCommSplit primitive, with resource sharing option in config.
Fix alltoallv hang (#788)
Increase number of channels on H100 when we're not limited by NVLink.
Improve error reporting in case of IB failure, printing local and
remote ID (#779).
Add build option to allow compilation against RDMA includes instead
of dynamically loading IB verbs symbols (#802).
Fix context creation for progress thread (#803).
NET/IB: add option to use multiple QPs in round-robin mode.
Fix tree performance issue when NVB is disabled on HCM topologies.
This commit is contained in:
Sylvain Jeaugey
2023-04-03 05:32:07 -07:00
parent 9b7d5edbfc
commit d97a32fac8
64 changed files with 4758 additions and 3131 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/graph.h
+9 -7
View File
@@ -53,9 +53,11 @@ ncclResult_t ncclTopoGetCpuAffinity(struct ncclTopoSystem* system, int rank, cpu
#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 ncclTopoGetGpuCount(struct ncclTopoSystem* system, int* count);
ncclResult_t ncclTopoGetNvsCount(struct ncclTopoSystem* system, int* count);
ncclResult_t ncclTopoGetLocalNet(struct ncclTopoSystem* system, int rank, int* id);
ncclResult_t ncclTopoGetNvsCount(struct ncclTopoSystem* system, int* count);
ncclResult_t ncclTopoGetLocalNet(struct ncclTopoSystem* system, int rank, int channelId, int* id);
ncclResult_t ncclTopoGetLocalGpu(struct ncclTopoSystem* system, int net, int* gpuIndex);
#define NCCL_TOPO_MAX_NODES 256
@@ -66,6 +68,7 @@ ncclResult_t ncclTopoSearchInit(struct ncclTopoSystem* system);
#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
#define NCCL_TOPO_PATTERN_NVLS 5 // NVLS+SHARP and NVLS+Tree
struct ncclTopoGraph {
// Input / output
int id; // ring : 0, tree : 1, collnet : 2
@@ -99,16 +102,15 @@ struct ncclTopoRanks {
int treeToParent[MAXCHANNELS];
int treeToChild0[MAXCHANNELS];
int treeToChild1[MAXCHANNELS];
int nvlsHeads[MAXCHANNELS];
};
ncclResult_t ncclTopoPreset(struct ncclComm* comm,
struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph,
struct ncclTopoRanks* topoRanks);
ncclResult_t ncclTopoPreset(struct ncclComm* comm, struct ncclTopoGraph** graphs, struct ncclTopoRanks* topoRanks);
ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePatterns,
struct ncclTopoRanks** allTopoRanks, int* rings, struct ncclTopoGraph* collNetGraph);
struct ncclTopoRanks** allTopoRanks, int* rings, struct ncclTopoGraph** graphs);
ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCompCap, struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph);
ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCompCap, struct ncclTopoGraph** graphs);
#include "info.h"
ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int protocol, int numPipeOps, float* time);