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

Fix model matching with PXN enable (#1295)

Browse Source 
[ROCm/rccl commit: eff56735b0]
This commit is contained in:
Wenkai Du
2024-08-16 06:16:00 +08:00
committed by GitHub
parent 38e189bb1e
commit e770054b7f
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/graph/rome_models.cc
+6 -1
View File
@@ -1475,6 +1475,8 @@ static bool permuteNetIds(int *n, int *g, int s, int last, struct rcclRomeModel*
// match gdr level
for (i = 0; i < ref->nNics; i++) {
for (j = 0; j < ref->nGpus; j++) {
// enabling PXN override paths over PHB and SYS
if (topo->gdrLevel[n[i]*ref->nGpus+g[j]] == PATH_PXN) continue;
if (ref->gdrLevel[i*ref->nGpus+j] != topo->gdrLevel[n[i]*ref->nGpus+g[j]]) break;
}
if (j < ref->nGpus) break;
@@ -1565,8 +1567,11 @@ ncclResult_t parseRome4P2H(struct ncclTopoSystem* system, struct ncclTopoGraph*
if (!ignore_numa && romeTopoModels[i].nicNuma[j] != romeTopo.nicNuma[k]) continue;
int g;
// check GDR
for (g = 0; g < ngpus; g++)
for (g = 0; g < ngpus; g++) {
// enabling PXN override paths over PHB and SYS
if (romeTopo.gdrLevel[k*ngpus+g] == PATH_PXN) continue;
if (romeTopoModels[i].gdrLevel[j*ngpus+g] != romeTopo.gdrLevel[k*ngpus+g]) break;
}
if (g >= ngpus) break;
}
if (k < nnets) {