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

introduce a hw topology aware bintree

Browse Source 
for hayabusa architecture.
This commit is contained in:
Edgar Gabriel
2022-10-03 15:12:57 +00:00
parent ef71550738
commit e645b02cd8
4 changed files with 204 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/enqueue.cc
+1 -1
View File
@@ -1224,7 +1224,7 @@ comp_next:
NCCLCHECK(getPatternInfo(info));
NCCLCHECK(getLoopInfo(info));
if (info->comm->topo->pivotA2ANumBiRings == 3 ) {
if (ncclTypeSize(info->datatype)*info->count > 65536) {
if (ncclTypeSize(info->datatype)*info->count > 131072) {
work->pad_0 = 1;
} else {
work->pad_0 = 2;