Use bit reversal based mapping for multi-node (#1572)

[ROCm/rccl commit: 85eb1f16bc]
Tento commit je obsažen v:
Bertan Dogancay
2025-02-26 09:48:03 -05:00
odevzdal GitHub
rodič acf5822a6c
revize a9d09c6551
4 změnil soubory, kde provedl 24 přidání a 10 odebrání
+1 -1
Zobrazit soubor
@@ -163,7 +163,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
struct ncclDevWorkP2p* work = &works[workIx];
size_t bytes = isSend ? work->sendBytes : work->recvBytes;
int nParts = isSend ? work->nSendChannels : work->nRecvChannels;
int part = ncclP2pChannelToPart(work->nP2pChannels, work->channelBase, ncclShmem.channelId, ncclShmem.comm.p2pnChannelsPerPeer);
int part = ncclP2pChannelToPart(work->nP2pChannels, work->channelBase, ncclShmem.channelId, ncclShmem.comm.p2pnChannelsPerPeer, ncclShmem.comm.nNodes);
hasWork = (part < nParts);
if (nParts != 0) {
size_t partBeg, partEnd;