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

2.12.12-1

Browse Source 
Improve allreduce performance when we have more than one network interface per
GPU and we need to use PXN to close rings.
Add support for PCI Gen5 on 5.4 kernels.
Fix crash when setting NCCL_SET_THREAD_NAME.
Fix random crash in init due to uninitialized struct.
Fix hang on cubemesh topologies.
Add P2P_DIRECT_DISABLE parameter to disable direct access to pointers within a
process.
This commit is contained in:
Sylvain Jeaugey
2022-05-03 01:30:26 -07:00
parent 9bfc1c6e35
commit 7aa1c46fd5
12 changed files with 94 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/channel.h
+31
View File
@@ -10,5 +10,36 @@
ncclResult_t initChannel(struct ncclComm* comm, int channelid);
ncclResult_t freeChannel(struct ncclChannel* channel, int nRanks);
static ncclResult_t ncclChannelComputeBase(struct ncclComm* comm, int peer, int coll, int*channelBase) {
int p2pGroupSize = NCCL_MAX_WORK_ELEMENTS_P2P/2;
int peerNode = comm->rankToNode[peer];
int peerIndex = comm->rankToLocalRank[peer];
int nsteps = comm->maxLocalRanks;
int rankIndex = comm->rankToLocalRank[comm->rank];
int step, delta;
if (coll == ncclFuncSend) {
step = (nsteps + peerIndex - rankIndex)%nsteps;
delta = (comm->nNodes + peerNode - comm->node) % comm->nNodes;
} else if (coll == ncclFuncRecv) {
step = (nsteps + rankIndex - peerIndex)%nsteps;
delta = (comm->nNodes + comm->node - peerNode) % comm->nNodes;
} else {
return ncclInternalError;
}
*channelBase = comm->nNodes > 1 ? delta+(step/p2pGroupSize) : step;
return ncclSuccess;
}
static ncclResult_t ncclChannelComputeFromBase(struct ncclComm* comm, int base, int channelInc, int*channelId) {
*channelId = (base+comm->p2pChannels[channelInc]) % comm->p2pnChannels;
return ncclSuccess;
}
static ncclResult_t ncclChannelCompute(struct ncclComm* comm, int peer, int channelInc, int coll, int*channelId) {
int base;
NCCLCHECK(ncclChannelComputeBase(comm, peer, coll, &base));
NCCLCHECK(ncclChannelComputeFromBase(comm, base, channelInc, channelId));
return ncclSuccess;
}
#endif