2.13.4-1

Optimize CUDA graph launch; avoid launching a CPU callback for
intra-node operations.
Simplify kernel common code to improve the latency of send/recv
operations.
Strengthen CUDA streams semantics.
Change NET API to v6, to add dmabuf support.
Add ncclGetLastError() function.
Add ncclRemoteError code and use it for remote network errors.
Support the use of a different NCCL_NET parameter per communicator.
Add support for SHM and P2P transfers using cudaMemcpy.

src/misc/gdrwrap.cc

@@ -57,7 +57,7 @@ ncclResult_t wrap_gdr_symbols(void) {
 
   if (__sync_bool_compare_and_swap(&gdrState, gdrUninitialized, gdrInitializing) == false) {
     // Another thread raced in front of us. Wait for it to be done.
-    while (gdrState == gdrInitializing) pthread_yield();
+    while (gdrState == gdrInitializing) sched_yield();
     return (gdrState == gdrInitialized) ? ncclSuccess : ncclSystemError;
   }