NCCL AllReduce

paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc

@@ -50,10 +50,6 @@ void NCCLAllReduceOpHandle::RunImpl() {
 
       auto &lod_tensor = s->FindVar(var_name)->Get<LoDTensor>();
       void *buffer = const_cast<void *>(lod_tensor.data<void>());
-      uintptr_t buf = reinterpret_cast<uintptr_t>(buffer);
-      if (buf % sizeof(float) != 0) {
-        VLOG(3) << "Buffer is not aligned " << buf;
-      }
 
       if (dtype == -1) {
         dtype = platform::ToNCCLDataType(lod_tensor.type());

paddle/fluid/platform/nccl_helper.h

@@ -36,12 +36,10 @@ inline ncclDataType_t ToNCCLDataType(std::type_index type) {
 
 class NCCLGroupGuard {
  public:
-  inline NCCLGroupGuard() {
-    mutex().lock();
-    PADDLE_ENFORCE(dynload::ncclGroupStart());
-  }
+  inline NCCLGroupGuard() { PADDLE_ENFORCE(dynload::ncclGroupStart()); }
 
   inline ~NCCLGroupGuard() {
+    mutex().lock();
     PADDLE_ENFORCE(dynload::ncclGroupEnd());
     mutex().unlock();
   }