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/include/p2p.h

@@ -9,19 +9,4 @@
 #ifndef NCCL_P2P_H_
 #define NCCL_P2P_H_
 
-struct ncclP2Pinfo {
-  void* buff;
-  ssize_t nbytes;
-};
-
-typedef ncclRecyclableList<struct ncclP2Pinfo> ncclP2Plist;
-
-static ncclResult_t ncclSaveP2pInfo(ncclP2Plist* &p2p, void* buff, ssize_t nBytes) {
-  if (p2p == NULL) p2p = new ncclP2Plist();
-  struct ncclP2Pinfo* next;
-  NCCLCHECK(p2p->getNewElem(&next));
-  next->buff = buff;
-  next->nbytes = nBytes;
-  return ncclSuccess;
-}
 #endif