Fixed issue with attomicEXCH creating errors on multi node runs (#587)

Tento commit je obsažen v:
akolliasAMD
2022-07-22 11:32:49 -06:00
odevzdal GitHub
rodič 686dbc8bc6
revize 254208e7dd
+1 -1
Zobrazit soubor
@@ -117,7 +117,7 @@ private:
if (flags & (Recv*RoleWaitRecv | Send*RoleWaitSend)) {
if (isSendNotRecv && (flags & SizesFifoEnabled))
atomicExch_system((unsigned long long *)(connSizesFifoPtr+step%NCCL_STEPS), nelts*sizeof(T));
__atomic_store_n((connSizesFifoPtr+step%NCCL_STEPS), nelts*sizeof(T), __ATOMIC_SEQ_CST);
void **ptrs = isSendNotRecv ? (ncclShmem->groups[group].dsts + Dst)
: (ncclShmem->groups[group].srcs + Src);