2.20.3-1

Add support for alternating rings, allow for cross-nic rings without
cross-rail communication.
Add support for user buffer registration for network send/recv.
Optimize aggregated operations to better utilize all channels.
Add flattening for BCM PCI gen5 switches.
Add support for inter-node NVLink communication
Add support for port fusion in NET/IB.
Add support for ReduceScatter and AllGather using Collnet.
Update net API to v8.
Fix hang during A2A connection.


[ROCm/rccl commit: b6475625fb]

projects/rccl/src/misc/argcheck.cc

@@ -7,7 +7,7 @@
 #include "argcheck.h"
 #include "comm.h"
 
-static ncclResult_t CudaPtrCheck(const void* pointer, struct ncclComm* comm, const char* ptrname, const char* opname) {
+ncclResult_t CudaPtrCheck(const void* pointer, struct ncclComm* comm, const char* ptrname, const char* opname) {
   cudaPointerAttributes attr;
   cudaError_t err = cudaPointerGetAttributes(&attr, pointer);
   if (err != cudaSuccess || attr.devicePointer == NULL) {

projects/rccl/src/misc/cudawrap.cc

@@ -14,6 +14,9 @@
 // This env var (NCCL_CUMEM_ENABLE) toggles cuMem API usage
 NCCL_PARAM(CuMemEnable, "CUMEM_ENABLE", -2);
 
+// Handle type used for cuMemCreate()
+CUmemAllocationHandleType ncclCuMemHandleType = CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR;
+
 static int ncclCuMemSupported = 0;
 
 // Determine whether CUMEM & VMM RDMA is supported on this platform

projects/rccl/src/misc/ipcsocket.cc

@@ -132,7 +132,7 @@ ncclResult_t ncclIpcSocketRecvMsg(ncclIpcSocket *handle, void *hdr, int hdrLen,
       WARN("UDS: Receiving data over socket failed : %d", errno);
       return ncclSystemError;
     }
-    if (handle->abortFlag && *handle->abortFlag) return ncclInternalError;
+    if (handle->abortFlag && __atomic_load_n(handle->abortFlag, __ATOMIC_RELAXED)) return ncclInternalError;
   }
 
   if (recvFd != NULL) {
@@ -221,7 +221,7 @@ ncclResult_t ncclIpcSocketSendMsg(ncclIpcSocket *handle, void *hdr, int hdrLen,
       WARN("UDS: Sending data over socket %s failed : %s (%d)", temp, strerror(errno), errno);
       return ncclSystemError;
     }
-    if (handle->abortFlag && *handle->abortFlag) return ncclInternalError;
+    if (handle->abortFlag && __atomic_load_n(handle->abortFlag, __ATOMIC_RELAXED)) return ncclInternalError;
   }
 
   return ncclSuccess;

projects/rccl/src/misc/nvmlwrap.cc

@@ -39,6 +39,8 @@ namespace {
   NCCL_NVML_FN(nvmlDeviceGetCudaComputeCapability, nvmlReturn_t, (nvmlDevice_t device, int* major, int* minor))
   NCCL_NVML_FN(nvmlDeviceGetP2PStatus, nvmlReturn_t, (nvmlDevice_t device1, nvmlDevice_t device2, nvmlGpuP2PCapsIndex_t p2pIndex, nvmlGpuP2PStatus_t* p2pStatus))
   NCCL_NVML_FN(nvmlDeviceGetFieldValues, nvmlReturn_t, (nvmlDevice_t device, int valuesCount, nvmlFieldValue_t *values))
+  // MNNVL support
+  NCCL_NVML_FN(nvmlDeviceGetGpuFabricInfoV, nvmlReturn_t, (nvmlDevice_t device, nvmlGpuFabricInfoV_t *gpuFabricInfo))
 
   std::mutex lock; // NVML has had some thread safety bugs
   bool initialized = false;
@@ -82,7 +84,9 @@ ncclResult_t ncclNvmlEnsureInitialized() {
       {(void**)&pfn_nvmlDeviceGetNvLinkCapability, "nvmlDeviceGetNvLinkCapability"},
       {(void**)&pfn_nvmlDeviceGetCudaComputeCapability, "nvmlDeviceGetCudaComputeCapability"},
       {(void**)&pfn_nvmlDeviceGetP2PStatus, "nvmlDeviceGetP2PStatus"},
-      {(void**)&pfn_nvmlDeviceGetFieldValues, "nvmlDeviceGetFieldValues"}
+      {(void**)&pfn_nvmlDeviceGetFieldValues, "nvmlDeviceGetFieldValues"},
+      // MNNVL support
+      {(void**)&pfn_nvmlDeviceGetGpuFabricInfoV, "nvmlDeviceGetGpuFabricInfoV"},
     };
     for(Symbol sym: symbols) {
       *sym.ppfn = dlsym(libhandle, sym.name);
@@ -269,3 +273,12 @@ ncclResult_t ncclNvmlDeviceGetFieldValues(nvmlDevice_t device, int valuesCount,
   NVMLTRY(nvmlDeviceGetFieldValues, device, valuesCount, values);
   return ncclSuccess;
 }
+
+// MNNVL support
+ncclResult_t ncclNvmlDeviceGetGpuFabricInfoV(nvmlDevice_t device, nvmlGpuFabricInfoV_t *gpuFabricInfo) {
+  NCCLCHECK(ncclNvmlEnsureInitialized());
+  std::lock_guard<std::mutex> locked(lock);
+  gpuFabricInfo->version = nvmlGpuFabricInfo_v2;
+  NVMLTRY(nvmlDeviceGetGpuFabricInfoV, device, gpuFabricInfo);
+  return ncclSuccess;
+}

projects/rccl/src/misc/shmutils.cc

@@ -169,7 +169,7 @@ ncclResult_t ncclShmemAllgather(struct ncclComm *comm, struct ncclShmemCollBuff
   int curRound = shmem->round;
   size_t mycnt;
 
-  if (comm == NULL || shmem == NULL || sendbuff == NULL || recvbuff == NULL) {
+  if (comm == NULL || shmem == NULL || sendbuff == NULL || recvbuff == NULL || shmem->maxTypeSize < typeSize) {
     ret = ncclInvalidArgument;
     goto exit;
   }
@@ -184,7 +184,7 @@ ncclResult_t ncclShmemAllgather(struct ncclComm *comm, struct ncclShmemCollBuff
     uint64_t t0 = clockNano();
     while(__atomic_load_n(shmem->cnt[curRound], __ATOMIC_ACQUIRE) != comm->localRanks + 1) {
       if (clockNano() - t0 >= 5 * 1000) sched_yield();
-      if (*comm->abortFlag == 1) {
+      if (__atomic_load_n(comm->abortFlag, __ATOMIC_RELAXED) == 1) {
         ret = ncclInternalError;
         goto exit;
       }

projects/rccl/src/misc/socket.cc

@@ -34,7 +34,7 @@ static ncclResult_t socketProgressOpt(int op, struct ncclSocket* sock, void* ptr
       }
     }
     (*offset) += bytes;
-    if (sock->abortFlag && *sock->abortFlag != 0) {
+    if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED)) {
       INFO(NCCL_NET, "socketProgressOpt: abort called");
       return ncclInternalError;
     }
@@ -529,6 +529,8 @@ static ncclResult_t socketPollConnect(struct ncclSocket* sock) {
     sock->state = ncclSocketStateConnecting;
   } else if (ret != EINPROGRESS) {
     sock->state = ncclSocketStateError;
+    char line[SOCKET_NAME_MAXLEN+1];
+    WARN("socketPollConnect: Connect to %s returned %d(%s) errno %d(%s)", ncclSocketToString(&sock->addr, line), ret, strerror(ret), errno, strerror(errno));
     return ncclSystemError;
   }
   return ncclSuccess;
@@ -618,12 +620,12 @@ ncclResult_t ncclSocketConnect(struct ncclSocket* sock) {
   do {
     NCCLCHECK(socketProgressState(sock));
   } while (sock->asyncFlag == 0 &&
-      (sock->abortFlag == NULL || *sock->abortFlag == 0) &&
+      (sock->abortFlag == NULL || __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED) == 0) &&
       (sock->state == ncclSocketStateConnecting ||
        sock->state == ncclSocketStateConnectPolling ||
        sock->state == ncclSocketStateConnected));
 
-  if (sock->abortFlag && *sock->abortFlag != 0) return ncclInternalError;
+  if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED)) return ncclInternalError;
 
   switch (sock->state) {
     case ncclSocketStateConnecting:
@@ -665,11 +667,11 @@ ncclResult_t ncclSocketAccept(struct ncclSocket* sock, struct ncclSocket* listen
   do {
     NCCLCHECKGOTO(socketProgressState(sock), ret, exit);
   } while (sock->asyncFlag == 0 &&
-      (sock->abortFlag == NULL || *sock->abortFlag == 0) &&
+      (sock->abortFlag == NULL || __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED) == 0) &&
       (sock->state == ncclSocketStateAccepting ||
        sock->state == ncclSocketStateAccepted));
 
-  if (sock->abortFlag && *sock->abortFlag != 0) return ncclInternalError;
+  if (sock->abortFlag && __atomic_load_n(sock->abortFlag, __ATOMIC_RELAXED)) return ncclInternalError;
 
   switch (sock->state) {
     case ncclSocketStateAccepting:

projects/rccl/src/misc/tuner.cc

@@ -30,25 +30,25 @@ ncclResult_t ncclLoadTunerPlugin(ncclTuner_t** tuner) {
     if (name) {
       INFO(NCCL_TUNING, "NCCL_TUNER_PLUGIN set to %s", name);
       tunerPluginLib = dlopen(name, RTLD_LAZY | RTLD_LOCAL);
-    }
-    if (tunerPluginLib == nullptr) {
-      // dlopen does not guarantee to set errno, but dlerror only gives us a
-      // string, so checking errno doesn't hurt to try to provide a better
-      // error message
-      if (errno == ENOENT) {
-        INFO(NCCL_TUNING, "Tuner: no plugin found '%s', using default tuner instead.", name);
+      if (tunerPluginLib == nullptr) {
+        // dlopen does not guarantee to set errno, but dlerror only gives us a
+        // string, so checking errno doesn't hurt to try to provide a better
+        // error message
+        if (errno == ENOENT) {
+          INFO(NCCL_TUNING, "Tuner: no plugin found '%s', using default tuner instead.", name);
+        } else {
+          INFO(NCCL_TUNING, "Tuner: plugin load '%s' returned error (%d : %s), using default tuner instead.", name, errno, dlerror());
+        }
       } else {
-        INFO(NCCL_TUNING, "Tuner: plugin load '%s' returned error (%d : %s), using default tuner instead.", name, errno, dlerror());
-      }
-    } else {
-      tunerSymbol = (ncclTuner_t*)dlsym(tunerPluginLib, NCCL_TUNER_PLUGIN_SYMBOL);
-      if (tunerSymbol == nullptr) {
-        INFO(NCCL_TUNING, "Tuner: failed to find " NCCL_TUNER_PLUGIN_SYMBOL " in plugin (%s), using default tuner instead.", name);
-        dlclose(tunerPluginLib);
-        tunerPluginLib = nullptr;
-      } else {
-        INFO(NCCL_TUNING, "Opened tuner: '%s'", tunerSymbol->name);
-        tunerPluginRefCount = 0;
+        tunerSymbol = (ncclTuner_t*)dlsym(tunerPluginLib, NCCL_TUNER_PLUGIN_SYMBOL);
+        if (tunerSymbol == nullptr) {
+          INFO(NCCL_TUNING, "Tuner: failed to find " NCCL_TUNER_PLUGIN_SYMBOL " in plugin (%s), using default tuner instead.", name);
+          dlclose(tunerPluginLib);
+          tunerPluginLib = nullptr;
+        } else {
+          INFO(NCCL_TUNING, "Opened tuner: '%s'", tunerSymbol->name);
+          tunerPluginRefCount = 0;
+        }
       }
     }
   }

projects/rccl/src/misc/utils.cc

@@ -291,3 +291,79 @@ void ncclMemoryStackDestruct(struct ncclMemoryStack* me) {
     h = h1;
   }
 }
+
+const char* ncclOpToString(ncclRedOp_t op) {
+  switch (op) {
+    case ncclSum:
+      return "ncclSum";
+    case ncclProd:
+      return "ncclProd";
+    case ncclMax:
+      return "ncclMax";
+    case ncclMin:
+      return "ncclMin";
+    case ncclAvg:
+      return "ncclAvg";
+    default:
+      return "Unknown";
+  }
+}
+
+const char* ncclDatatypeToString(ncclDataType_t type) {
+  switch (type) {
+    case ncclInt8: // ncclChar
+      return "ncclInt8";
+    case ncclInt32: // ncclInt
+      return "ncclInt32";
+    case ncclUint32:
+      return "ncclUint32";
+    case ncclInt64:
+      return "ncclInt64";
+    case ncclUint64:
+      return "ncclUint64";
+    case ncclFloat16: // ncclHalf
+      return "ncclFloat16";
+    case ncclFloat32: // ncclFloat
+      return "ncclFloat32";
+    case ncclFloat64: // ncclDouble
+      return "ncclFloat64";
+#if defined(__CUDA_BF16_TYPES_EXIST__)
+    case ncclBfloat16:
+      return "ncclBfloat16";
+#endif
+    default:
+      return "Unknown";
+  }
+}
+
+const char* ncclAlgoToString(int algo) {
+  switch (algo) {
+    case NCCL_ALGO_TREE:
+      return "TREE";
+    case NCCL_ALGO_RING:
+      return "RING";
+    case NCCL_ALGO_COLLNET_DIRECT:
+      return "COLLNET_DIRECT";
+    case NCCL_ALGO_COLLNET_CHAIN:
+      return "COLLNET_CHAIN";
+    case NCCL_ALGO_NVLS:
+      return "NVLS";
+    case NCCL_ALGO_NVLS_TREE:
+      return "NVLS_TREE";
+    default:
+      return "Unknown";
+  }
+}
+
+const char* ncclProtoToString(int proto) {
+  switch (proto) {
+    case NCCL_PROTO_LL:
+      return "LL";
+    case NCCL_PROTO_LL128:
+      return "LL128";
+    case NCCL_PROTO_SIMPLE:
+      return "SIMPLE";
+    default:
+      return "Unknown";
+  }
+}