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.


[ROCm/rccl commit: 19ab67d172]

projects/rccl/src/include/checks.h

@@ -9,7 +9,7 @@
 
 #include "debug.h"
 
-// Check CUDA calls
+// Check CUDA RT calls
 #define CUDACHECK(cmd) do {                                 \
     cudaError_t err = cmd;                                  \
     if( err != cudaSuccess ) {                              \
@@ -142,9 +142,9 @@
   if (tmpAbortFlag) NEQCHECKGOTO(*tmpAbortFlag, 0, res, label); \
 } while (!(cond));
 
-#define NCCLCHECKTHREAD(a) do { \
-  if ((args->ret = (a)) != ncclSuccess) { \
-    INFO(NCCL_INIT,"%s:%d -> %d [Async thread]", __FILE__, __LINE__, args->ret); \
+#define NCCLCHECKTHREAD(a, args) do { \
+  if (((args)->ret = (a)) != ncclSuccess) { \
+    INFO(NCCL_INIT,"%s:%d -> %d [Async thread]", __FILE__, __LINE__, (args)->ret); \
     return args; \
   } \
 } while(0)