diff --git a/Grid/algorithms/GeneralCoarsenedMatrix.h b/Grid/algorithms/GeneralCoarsenedMatrix.h
index 3e4bc1df..21c75c8b 100644
--- a/Grid/algorithms/GeneralCoarsenedMatrix.h
+++ b/Grid/algorithms/GeneralCoarsenedMatrix.h
@@ -344,8 +344,8 @@ public:
 	    int ss = sss/nbasis;
 	    int b  = sss%nbasis;
 
-	    auto SE = Stencil_v.GetEntry(point,ss);
-	    auto nbr = coalescedRead(in_v[SE->_offset]);
+	    auto SE  = Stencil_v.GetEntry(point,ss);
+	    auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd);
 	    auto res = out_v(ss)(b);
 	    for(int bb=0;bb<nbasis;bb++) {
 	      res = res + coalescedRead(A_v[ss](b,bb))*nbr(bb);
diff --git a/Grid/qcd/utils/WilsonLoops.h b/Grid/qcd/utils/WilsonLoops.h
index 78e25a8d..665ab2d6 100644
--- a/Grid/qcd/utils/WilsonLoops.h
+++ b/Grid/qcd/utils/WilsonLoops.h
@@ -487,7 +487,7 @@ public:
     for(int mu=0;mu<Nd;mu++){
       { //view scope
 	autoView( gStaple_v , gStaple, AcceleratorWrite);
-	auto gStencil_v = gStencil.View();
+	auto gStencil_v = gStencil.View(AcceleratorRead);
 	
 	accelerator_for(ss, ggrid->oSites(), ggrid->Nsimd(), {
 	    decltype(coalescedRead(Ug_dirs_v[0][0])) stencil_ss;
@@ -1199,7 +1199,7 @@ public:
 
       { //view scope
 	autoView( gStaple_v , gStaple, AcceleratorWrite);
-	auto gStencil_v = gStencil.View();
+	auto gStencil_v = gStencil.View(AcceleratorRead);
 
 	accelerator_for(ss, ggrid->oSites(), ggrid->Nsimd(), {
 	    decltype(coalescedRead(Ug_dirs_v[0][0])) stencil_ss;