2.19.1-1

Add local user buffer registration for NVLink SHARP.
Add tuning plugin support.
Increase net API to v7 to allow for device-side packet reordering;
remove support for v4 plugins.
Add support for RoCE ECE.
Add support for C2C links.
Better detect SHM allocation failures to avoid crash with Bus Error.
Fix missing thread unlocks in bootstrap (Fixes #936).
Disable network flush by default on H100.
Move device code from src/collectives/device to src/device.


[ROCm/rccl commit: f9c3dc251e]

projects/rccl/src/include/proxy.h

@@ -7,10 +7,11 @@
 #ifndef NCCL_PROXY_H_
 #define NCCL_PROXY_H_
 
-#include "devcomm.h"
+#include "device.h"
 #include "info.h"
 #include "socket.h"
 #include "ipcsocket.h"
+#include "nccl_net.h"
 #include <pthread.h>
 #include "shm.h"
 #include "p2p.h"
@@ -65,6 +66,8 @@ struct ncclProxySubArgs {
   uint64_t end;
   void* requests[NCCL_STEPS];
   void* profilingEvents[NCCL_STEPS];
+  void* recvRequestsCache[NCCL_STEPS];
+  int recvRequestsSubCount;
 };
 
 struct ncclProxyArgs {
@@ -146,7 +149,7 @@ struct ncclProxyProgressState {
   char opsPoolShmSuffix[6];
 
   pthread_t thread;
-  bool stop;
+  volatile int stop;
   struct ncclProxyPeer** localPeers;
   struct ncclSharedNetComms* netComms[NCCL_MAX_NETDEVS];
   struct ncclProxyArgs* active;
@@ -157,11 +160,12 @@ struct ncclProxyProgressState {
 
 // Expected proxy response fifo
 struct ncclExpectedProxyResponse {
-  void*    opId;
-  int      respSize;
-  bool     done;
-  void*    respBuff;
-  struct   ncclExpectedProxyResponse* next;
+  void*                             opId;
+  int                               respSize;
+  bool                              done;
+  void*                             respBuff;
+  ncclResult_t                      res;
+  struct ncclExpectedProxyResponse* next;
 };
 
 struct ncclProxyAsyncOp {
@@ -181,7 +185,16 @@ struct ncclProxyLocalPeer {
   int asyncOpCounter;
 };
 
+// Common response header for all proxyOps
+// We pack this into a struct to reduce the number of blocking send and recv calls
+struct ncclProxyRpcResponseHeader {
+  void* opId;
+  ncclResult_t res;
+  int respSize;
+};
+
 struct ncclProxyState {
+  int internalRefCount;
   int refCount;
   int tpRank;
   int tpnRanks;
@@ -196,11 +209,13 @@ struct ncclProxyState {
   ncclNet_t* ncclNet;
   ncclCollNet_t* ncclCollNet;
   volatile uint32_t* abortFlag;
+  volatile uint32_t* abortFlagRefCount;
   // Service thread
   pthread_t thread;
   struct ncclSocket* listenSock;
-  int stop;
+  volatile int stop;
   CUcontext cudaCtx;
+  ncclResult_t asyncResult;
 
   // Used by main thread
   union ncclSocketAddress* peerAddresses;
@@ -233,8 +248,11 @@ struct ncclProxyConnection {
   struct ncclProxyArgs *proxyAppend;
   struct ncclProxyArgs **proxyAppendPtr;
   void* transportResources;
+  ncclNetDeviceHandle_t* netDeviceHandle;
+  void* mhandles[NCCL_NUM_PROTOCOLS];
   proxyConnectState state;
   struct ncclCollNetSharedRes* collNet;
+  int needsProxyProgress;
 };
 
 typedef ncclResult_t (*threadFunc_t)(struct ncclProxyArgs*);
@@ -260,7 +278,7 @@ enum ncclProxyMsgType {
   ncclProxyMsgClose = 6,
   ncclProxyMsgAbort = 7,
   ncclProxyMsgStop = 8,
-  ncclProxyMsgConvertFd = 9, // cuMem API support (UDS)
+  ncclProxyMsgGetFd = 9, // cuMem API support (UDS)
 };
 
 // This function is called by a client of the proxy that needs to invoke any of the non-progress proxyOp types
@@ -272,9 +290,10 @@ ncclResult_t ncclProxyCallAsync(struct ncclComm* comm, struct ncclProxyConnector
 ncclResult_t ncclProxyCallBlocking(struct ncclComm* comm, struct ncclProxyConnector* proxyConn, int type, void* reqBuff, int reqSize, void* respBuff, int respSize);
 ncclResult_t ncclPollProxyResponse(struct ncclComm* comm, struct ncclProxyConnector* proxyConn, void* respBuff, void* opId);
 
-ncclResult_t ncclProxyClientConvertFdBlocking(struct ncclComm* comm, struct ncclProxyConnector* proxyConn, int fd, int* convertedFd);
+ncclResult_t ncclProxyClientGetFdBlocking(struct ncclComm* comm, struct ncclProxyConnector* proxyConn, void *handle, int* convertedFd);
 
 ncclResult_t ncclProxyStop(struct ncclComm* comm);
 ncclResult_t ncclProxyShmUnlink(struct ncclComm* comm);
-ncclResult_t ncclProxyDestroy(struct ncclComm* comm);
+ncclResult_t ncclProxyDestroy(struct ncclProxyState *proxyState);
+ncclResult_t ncclProxyTryDetach(struct ncclProxyState *proxyState);
 #endif