2.8.4-1

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]

projects/rccl/src/include/debug.h

@@ -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;

projects/rccl/src/include/shm.h

@@ -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);

projects/rccl/src/include/socket.h

@@ -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);