From 9117f61109a24253901fb9c728ae7885be7ce790 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Wed, 31 Jul 2019 01:22:54 +0100
Subject: [PATCH] GPU friendly

---
 Grid/lattice/Lattice_reduction.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/Grid/lattice/Lattice_reduction.h b/Grid/lattice/Lattice_reduction.h
index 7ec25410..96654cef 100644
--- a/Grid/lattice/Lattice_reduction.h
+++ b/Grid/lattice/Lattice_reduction.h
@@ -128,7 +128,6 @@ axpy_norm_fast(Lattice<vobj> &z,sobj a,const Lattice<vobj> &x,const Lattice<vobj
 template<class sobj,class vobj> strong_inline RealD 
 axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Lattice<vobj> &y) 
 {
-  const int pad = 8;
   z.Checkerboard() = x.Checkerboard();
   conformable(z,x);
   conformable(x,y);
@@ -148,14 +147,15 @@ axpby_norm_fast(Lattice<vobj> &z,sobj a,sobj b,const Lattice<vobj> &x,const Latt
   
   typedef decltype(innerProduct(x_v[0],y_v[0])) inner_t;
   Lattice<inner_t> inner_tmp(grid);
+  auto inner_tmp_v = inner_tmp.View();
 
   accelerator_for( ss, sites, nsimd,{
       auto tmp = a*x_v(ss)+b*y_v(ss);
-      coalescedWrite(inner_tmp[ss],innerProduct(tmp,tmp));
+      coalescedWrite(inner_tmp_v[ss],innerProduct(tmp,tmp));
       coalescedWrite(z_v[ss],tmp);
   })
   
-  nrm = TensorRemove(sum(inner_tmp));
+  nrm = real(TensorRemove(sum(inner_tmp)));
 
   z.Grid()->GlobalSum(nrm);
   return nrm;