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

Fix incorrect CR8 detection

Browse Source 
Also change level of ring graph print to help debugging


[ROCm/rccl commit: f54dc58113]
This commit is contained in:
Wenkai Du
2020-02-21 09:51:24 -08:00
parent cf4bce4ad3
commit 93d448e2fe
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/graph/connect.cc
+2 -2
View File
@@ -95,8 +95,8 @@ static ncclResult_t connectRings(struct ncclComm* comm, int* ringRecv, int* ring
channel1->ring.next = nextRecvRank;
}
}
TRACE(NCCL_GRAPH, "Ring %d : %d -> %d -> %d", c, channel0->ring.prev, comm->rank, channel0->ring.next);
TRACE(NCCL_GRAPH, "Ring %d : %d -> %d -> %d", c+nChannels, channel1->ring.prev, comm->rank, channel1->ring.next);
INFO(NCCL_GRAPH, "Ring %d : %d -> %d -> %d", c, channel0->ring.prev, comm->rank, channel0->ring.next);
INFO(NCCL_GRAPH, "Ring %d : %d -> %d -> %d", c+nChannels, channel1->ring.prev, comm->rank, channel1->ring.next);
}
return ncclSuccess;
}
projects/rccl/src/graph/search.cc
+1 -1
View File
@@ -451,7 +451,7 @@ static void parseChordalRing(struct ncclTopoSystem* system, char **str) {
int ngpus = system->nodes[GPU].count;
// single node CR8G only
if (ngpus != 8 && system->nodes[NET].count != 0)
if (ngpus != 8 || system->nodes[NET].count != 0)
return;
// validate chordal ring and calculate distance
for (i=0; i<ngpus; i++) {