Fix hang in corner cases of alltoallv using point to point send/recv.
Harmonize error messages.
Fix missing NVTX section in the license.
Update README.


[ROCm/rccl commit: 911d61f214]
This commit is contained in:
Sylvain Jeaugey committed 2021-02-09 15:36:48 -08:00
1 förälder 1db601566f
incheckning fc7bdb38a5
19 files changed
+53 -57

No files matched your search

+1 -1
Visa fil
@@ -25,7 +25,7 @@ extern pthread_mutex_t ncclDebugOutputLock;
extern FILE *ncclDebugFile;
extern ncclResult_t getHostName(char* hostname, int maxlen, const char delim);
void ncclDebugLog(ncclDebugLogLevel level, unsigned long flags, const char *filefunc, int line, const char *fmt, ...);
void ncclDebugLog(ncclDebugLogLevel level, unsigned long flags, const char *filefunc, int line, const char *fmt, ...) __attribute__ ((format (printf, 5, 6)));
// Let code temporarily downgrade WARN into INFO
extern thread_local int ncclDebugNoWarn;
+1 -1
Visa fil
@@ -45,7 +45,7 @@ static ncclResult_t shmOpen(const char* shmname, const int shmsize, void** shmPt
*shmPtr = ptr;
return ncclSuccess;
sysError:
WARN("Error while %s shared memory segment %s (size %d)\n", create ? "creating" : "attaching to", shmname, shmsize);
WARN("Error while %s shared memory segment %s (size %d)", create ? "creating" : "attaching to", shmname, shmsize);
cudaError:
if (fd != -1) close(fd);
if (create) shm_unlink(shmname);
+1 -1
Visa fil
@@ -367,7 +367,7 @@ static ncclResult_t connectAddress(int* fd, union socketAddress* remoteAddr) {
/* IPv4/IPv6 support */
int family = remoteAddr->sa.sa_family;
if (family != AF_INET && family != AF_INET6) {
WARN("Error : connecting to address with family %d is neither AF_INET(%d) nor AF_INET6(%d)\n", family, AF_INET, AF_INET6);
WARN("Error : connecting to address with family %d is neither AF_INET(%d) nor AF_INET6(%d)", family, AF_INET, AF_INET6);
return ncclInternalError;
}
int salen = (family == AF_INET) ? sizeof(sockaddr_in) : sizeof(sockaddr_in6);