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

Limit max channels for ring graph on single node Rome (#347)

Browse Source 
* Limit max channels for ring graph on single node Rome
* Partially revert "Use non-temporal access for streaming data (#341)"
This commit is contained in:
Wenkai Du authored and GitHub committed 2021-04-14 10:14:54 -07:00
1 parent 1fe031402a
commit a79f74082e
4 files changed
+13 -10
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/graph/search.cc
+4
View File
@@ -766,6 +766,10 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
}
if (graph->nChannels) return ncclSuccess;
if ((graph->pattern == NCCL_TOPO_PATTERN_RING) && (system->type & RCCL_TOPO_4P2H_ROME) && (ngpus == system->nRanks)) {
// limit single node max channels when searching ring graph on Rome
graph->maxChannels = 2;
}
if (ngpus == 1) if (graph->pattern != NCCL_TOPO_PATTERN_RING) graph->pattern = NCCL_TOPO_PATTERN_TREE;
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)