yezhengmao/nccl
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix random deadlock during ncclCommInitRank.

Browse Source
This commit is contained in:
Sylvain Jeaugey 2016-04-19 10:47:27 -07:00
parent 9de361a1b9
commit dba3ec9428
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/core.cu
View File

@ -188,7 +188,7 @@ static void syncRingDirect(RankGather* gather, int* ringDirectOk) {
swapped = __sync_bool_compare_and_swap(&gather->bar, bar_tmp, bar_tmp+1);
} while(!swapped);
while (gather->bar != 2*ndev) // Wait for all ranks to arrive at this second barrier
while (gather->bar < 2*ndev) // Wait for all ranks to arrive at this second barrier
sched_yield();
__sync_synchronize();
@ -203,7 +203,7 @@ static ncclResult_t closeGather(RankGather* gather, int ndev) {
swapped = __sync_bool_compare_and_swap(&gather->bar, bar_tmp, bar_tmp+1);
} while(!swapped);
while (gather->bar != 3*ndev) // Wait for all ranks to arrive at this third barrier
while (gather->bar < 3*ndev) // Wait for all ranks to arrive at this third barrier
sched_yield();
__sync_synchronize();