topo_expl: fix build after switching to rocm-smi-lib (#405)

* topo_expl: fix build after switching to rocm-smi-lib

* Use minimal of 4 channels for gfx908

[ROCm/rccl commit: 135d47d125]
This commit is contained in:
Wenkai Du
2021-07-27 08:30:08 -07:00
کامیت شده توسط GitHub
والد 3d382b5ba3
کامیت 8fbeb14175
4فایلهای تغییر یافته به همراه20 افزوده شده و 3 حذف شده
@@ -865,6 +865,8 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, ncclUniqueId* comm
int idx;
NCCLCHECK(ncclTopoIdToIndex(comm->topo, GPU, myInfo->busId, &idx));
allGather3Data[rank].nc = 2;
if (comm->topo->nodes[GPU].nodes[idx].gpu.gcn == 908)
allGather3Data[rank].nc = std::max(4/ringGraph.nChannels, 2);
if (comm->topo->nodes[GPU].count == comm->topo->nRanks && (comm->topo->type & RCCL_TOPO_CR8G))
allGather3Data[rank].nc = 4;
if (comm->topo->nodes[GPU].count == comm->topo->nRanks && comm->topo->nodes[GPU].nodes[idx].gpu.gcn == 910)