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

2.12.12-1

Browse Source 
Improve allreduce performance when we have more than one network interface per
GPU and we need to use PXN to close rings.
Add support for PCI Gen5 on 5.4 kernels.
Fix crash when setting NCCL_SET_THREAD_NAME.
Fix random crash in init due to uninitialized struct.
Fix hang on cubemesh topologies.
Add P2P_DIRECT_DISABLE parameter to disable direct access to pointers within a
process.


[ROCm/rccl commit: 7aa1c46fd5]
This commit is contained in:
Sylvain Jeaugey
2022-05-03 01:30:26 -07:00
parent e89ff21d35
commit 1c5734046d
12 changed files with 94 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/proxy.cc
+2 -4
View File
@@ -960,12 +960,10 @@ void* ncclProxyService(void* _args) {
struct pollfd pollfds[NCCL_MAX_LOCAL_RANKS+1];
struct ncclProxyLocalPeer peers[NCCL_MAX_LOCAL_RANKS];
memset(&peers, 0, sizeof(struct ncclProxyLocalPeer)*NCCL_MAX_LOCAL_RANKS);
for (int s=0; s<NCCL_MAX_LOCAL_RANKS; s++) {
peers[s].sock.fd = pollfds[s].fd = -1;
peers[s].sock.abortFlag = NULL;
peers[s].sock.asyncFlag = 0;
pollfds[s].events = POLLHUP|POLLIN;
peers[s].asyncOps.type = 0;
}
pollfds[NCCL_MAX_LOCAL_RANKS].fd = comm->proxyState.listenSock->fd;
pollfds[NCCL_MAX_LOCAL_RANKS].events = POLLIN;
@@ -1066,13 +1064,13 @@ void* ncclProxyService(void* _args) {
ncclResult_t ncclProxyInit(struct ncclComm* comm, struct ncclSocket* sock, union ncclSocketAddress* peerAddresses) {
comm->proxyState.listenSock = sock;
comm->proxyState.peerAddresses = peerAddresses;
ncclSetThreadName(comm->proxyState.thread, "NCCL Service %2d", comm->cudaDev);
return ncclSuccess;
}
ncclResult_t ncclProxyCreate(struct ncclComm* comm) {
// comm->proxyState.thread is pthread_join()'d by commFree() in init.cc
pthread_create(&comm->proxyState.thread, NULL, ncclProxyService, comm);
ncclSetThreadName(comm->proxyState.thread, "NCCL Service %2d", comm->cudaDev);
return ncclSuccess;
}