diff --git a/Grid/communicator/Communicator_base.h b/Grid/communicator/Communicator_base.h
index cd682dd0..3f38edd3 100644
--- a/Grid/communicator/Communicator_base.h
+++ b/Grid/communicator/Communicator_base.h
@@ -128,6 +128,34 @@ public:
   void GlobalXOR(uint32_t &);
   void GlobalXOR(uint64_t &);
   
+  template<class obj> void GlobalSumP2P(obj &o)
+  {
+    std::vector<obj> column;
+    obj accum = o;
+    int source,dest;
+    for(int d=0;d<_ndimension;d++){
+      column.resize(_processors[d]);
+      column[0] = accum;
+      std::vector<CommsRequest_t> list;
+      for(int p=1;p<_processors[d];p++){
+	ShiftedRanks(d,p,source,dest);
+	SendToRecvFromBegin(list,
+			    &column[0],
+			    dest,
+			    &column[p],
+			    source,
+			    sizeof(obj),d*100+p);
+
+      }
+      CommsComplete(list);
+      for(int p=1;p<_processors[d];p++){
+	accum = accum + column[p];
+      }
+    }
+    Broadcast(0,accum);
+    o=accum;
+  }
+
   template<class obj> void GlobalSum(obj &o){
     typedef typename obj::scalar_type scalar_type;
     int words = sizeof(obj)/sizeof(scalar_type);