diff --git a/Grid/qcd/action/fermion/WilsonImpl.h b/Grid/qcd/action/fermion/WilsonImpl.h
index 52e1ee00..d7941d1f 100644
--- a/Grid/qcd/action/fermion/WilsonImpl.h
+++ b/Grid/qcd/action/fermion/WilsonImpl.h
@@ -106,11 +106,15 @@ public:
 			    const _SpinorField & phi,
 			    int mu)
   {
+    const int Nsimd = SiteHalfSpinor::Nsimd();
     autoView( out_v, out, AcceleratorWrite);
     autoView( phi_v, phi, AcceleratorRead);
     autoView( Umu_v, Umu, AcceleratorRead);
-    accelerator_for(sss,out.Grid()->oSites(),1,{
-	multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
+    typedef decltype(coalescedRead(out_v[0]))   calcSpinor;
+    accelerator_for(sss,out.Grid()->oSites(),Nsimd,{
+	calcSpinor tmp;
+	multLink(tmp,Umu_v[sss],phi_v(sss),mu);
+	coalescedWrite(out_v[sss],tmp);
     });
   }
 					   
diff --git a/benchmarks/Benchmark_ITT.cc b/benchmarks/Benchmark_ITT.cc
index 032535b3..5d602ce9 100644
--- a/benchmarks/Benchmark_ITT.cc
+++ b/benchmarks/Benchmark_ITT.cc
@@ -445,7 +445,7 @@ public:
 	// 1344= 3*(2*8+6)*2*8 + 8*3*2*2 + 3*4*2*8
 	// 1344 = Nc* (6+(Nc-1)*8)*2*Nd + Nd*Nc*2*2  + Nd*Nc*Ns*2
 	//	double flops=(1344.0*volume)/2;
-#if 1
+#if 0
 	double fps = Nc* (6+(Nc-1)*8)*Ns*Nd + Nd*Nc*Ns  + Nd*Nc*Ns*2;
 #else
 	double fps = Nc* (6+(Nc-1)*8)*Ns*Nd + 2*Nd*Nc*Ns  + 2*Nd*Nc*Ns*2;
diff --git a/tests/debug/Test_cayley_mres.cc b/tests/debug/Test_cayley_mres.cc
index 5282c756..ea88885e 100644
--- a/tests/debug/Test_cayley_mres.cc
+++ b/tests/debug/Test_cayley_mres.cc
@@ -117,8 +117,8 @@ int main (int argc, char ** argv)
   else
   {
     std::cout<<GridLogMessage <<"Using cold configuration"<<std::endl;
-    SU<Nc>::ColdConfiguration(Umu);
-    //  SU<Nc>::HotConfiguration(RNG4,Umu);
+    //SU<Nc>::ColdConfiguration(Umu);
+    SU<Nc>::HotConfiguration(RNG4,Umu);
   }
 
   RealD mass=0.3;