From 5011753f4fafad0586161de7686ad1d8c3cc3774 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Thu, 30 Apr 2020 10:23:48 -0400
Subject: [PATCH 1/2] Clean up warning

---
 Grid/lattice/Lattice_peekpoke.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Grid/lattice/Lattice_peekpoke.h b/Grid/lattice/Lattice_peekpoke.h
index feca2f44..8f649bd7 100644
--- a/Grid/lattice/Lattice_peekpoke.h
+++ b/Grid/lattice/Lattice_peekpoke.h
@@ -156,7 +156,7 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
 // Peek a scalar object from the SIMD array
 //////////////////////////////////////////////////////////
 template<class vobj,class sobj>
-accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
+inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site){
         
   GridBase *grid = l.Grid();
 
@@ -185,7 +185,7 @@ accelerator_inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate
 };
 
 template<class vobj,class sobj>
-accelerator_inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
+inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site){
 
   GridBase *grid=l.Grid();
 

From 9b2d2d0fc3bcea50f9394badf2833b5de3ce452f Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Thu, 30 Apr 2020 12:31:07 -0400
Subject: [PATCH 2/2] Basis rotate stack passig to GPU reduction

---
 Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h b/Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
index 8bee43cc..a44b245c 100644
--- a/Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
+++ b/Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
@@ -108,7 +108,6 @@ void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, i
   }
 
   // Block the loop to keep storage footprint down
-  vobj zz=Zero();
   for(uint64_t s=0;s<oSites;s+=siteBlock){
 
     // remaining work in this block
@@ -116,7 +115,8 @@ void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, i
 
     // zero out the accumulators
     accelerator_for(ss,siteBlock*nrot,vobj::Nsimd(),{
-	auto z=coalescedRead(zz);
+	auto z=coalescedRead(Bp[ss]);
+	z=Zero();
 	coalescedWrite(Bp[ss],z);
     });
 
@@ -158,12 +158,12 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
   for(int k=0;k<basis.size();k++){
     basis_v[k] = basis[k].View();
   }
-  vobj zz=Zero();
   Vector<double> Qt_jv(Nm);
   double * Qt_j = & Qt_jv[0];
   for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
   accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
-    auto B=coalescedRead(zz);
+    auto B=coalescedRead(basis_v[k0][ss]);
+    B=Zero();
     for(int k=k0; k<k1; ++k){
       B +=Qt_j[k] * coalescedRead(basis_v[k][ss]);
     }