Merge remote-tracking branch 'nccl/master' into develop

[ROCm/rccl commit: a79d9e3586]
Αυτή η υποβολή περιλαμβάνεται σε:
Wenkai Du
2022-09-09 01:20:52 +00:00
γονέας fe99249cde c9a50a9ec5
υποβολή 7874a99c75
90 αρχεία άλλαξαν με 5517 προσθήκες και 3115 διαγραφές
@@ -106,6 +106,7 @@ static void *bootstrapRoot(void* args) {
do {
struct ncclSocket sock;
sock.abortFlag = NULL;
/* bootstrap root thread always uses blocking ncclSocketAccept. */
NCCLCHECKGOTO(ncclSocketAccept(&sock, listenSock), res, out);
NCCLCHECKGOTO(bootstrapNetRecv(&sock, &info, sizeof(info)), res, out);
close(sock.fd);
@@ -140,6 +141,7 @@ static void *bootstrapRoot(void* args) {
int next = (r+1) % nranks;
struct ncclSocket sock;
sock.abortFlag = NULL;
sock.asyncFlag = 0;
memcpy(&sock.addr, rankAddressesRoot+r, sizeof(union ncclSocketAddress));
NCCLCHECKGOTO(ncclSocketConnect(&sock), res, out);
NCCLCHECKGOTO(bootstrapNetSend(&sock, rankAddresses+next, sizeof(union ncclSocketAddress)), res, out);
@@ -289,7 +291,7 @@ ncclResult_t bootstrapInit(ncclUniqueId * id, struct ncclComm* comm) {
NCCLCHECK(bootstrapAllGather(state, state->peerProxyAddresses, sizeof(union ncclSocketAddress)));
NCCLCHECK(ncclProxyInit(comm, proxySocket, state->peerProxyAddresses));
TRACE(NCCL_INIT, "rank %d nranks %d - DONE", rank, nranks);
TRACE(NCCL_INIT, "rank %d nranks %d virtualId %d", rank, nranks, virtualId);
return ncclSuccess;
}
@@ -324,6 +326,7 @@ ncclResult_t bootstrapSend(void* commState, int peer, int tag, void* data, int s
struct bootstrapState* state = (struct bootstrapState*)commState;
struct ncclSocket sock;
sock.abortFlag = state->abortFlag;
sock.asyncFlag = 0;
memcpy(&sock.addr, state->peerCommAddresses+peer, sizeof(union ncclSocketAddress));
NCCLCHECK(ncclSocketConnect(&sock));
NCCLCHECK(bootstrapNetSend(&sock, &state->rank, sizeof(int)));