Enable LL128 protocol support (#605)
* Enable LL128 protocol support * Use shared memory object directly when possible
This commit is contained in:
@@ -165,6 +165,7 @@ struct ncclTopoSystem {
|
||||
|
||||
bool pivotA2AEnabled;
|
||||
int pivotA2ANumBiRings;
|
||||
bool ll128Enabled;
|
||||
};
|
||||
|
||||
ncclResult_t ncclTopoGetNode(struct ncclTopoSystem* system, struct ncclTopoNode** node, int type, uint64_t id);
|
||||
@@ -209,4 +210,8 @@ static ncclResult_t ncclTopoRankToIndex(struct ncclTopoSystem* system, int rank,
|
||||
static float ncclTopoXGMISpeed(int gcn) {
|
||||
return gcn == 910 ? MI200_XGMI_WIDTH : VEGA_XGMI_WIDTH;
|
||||
}
|
||||
|
||||
#define ncclGetKernelIndex(p_comm) \
|
||||
(((p_comm)->topo->ll128Enabled ? 1 : 0)*2 + ((p_comm)->hostDevComm.collTraceThread ? 1 : 0))
|
||||
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user