diff --git a/benchmarks/Benchmark_su3.cc b/benchmarks/Benchmark_su3.cc
index 3312e3c8..d1accefd 100644
--- a/benchmarks/Benchmark_su3.cc
+++ b/benchmarks/Benchmark_su3.cc
@@ -22,7 +22,7 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=24;lat+=2){
+  for(int lat=2;lat<=32;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
@@ -54,7 +54,7 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=24;lat+=2){
+  for(int lat=2;lat<=32;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
@@ -85,7 +85,7 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=24;lat+=2){
+  for(int lat=2;lat<=32;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
@@ -116,7 +116,7 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
-  for(int lat=2;lat<=24;lat+=2){
+  for(int lat=2;lat<=32;lat+=2){
 
       std::vector<int> latt_size  ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
       int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
diff --git a/lib/Make.inc b/lib/Make.inc
index 8b2bcf40..8d595ae5 100644
--- a/lib/Make.inc
+++ b/lib/Make.inc
@@ -1,4 +1,4 @@
 
-HFILES=./algorithms/approx/bigfloat.h ./algorithms/approx/bigfloat_double.h ./algorithms/approx/Chebyshev.h ./algorithms/approx/MultiShiftFunction.h ./algorithms/approx/Remez.h ./algorithms/approx/Zolotarev.h ./algorithms/CoarsenedMatrix.h ./algorithms/iterative/AdefGeneric.h ./algorithms/iterative/ConjugateGradient.h ./algorithms/iterative/ConjugateGradientMultiShift.h ./algorithms/iterative/ConjugateResidual.h ./algorithms/iterative/NormalEquations.h ./algorithms/iterative/PrecConjugateResidual.h ./algorithms/iterative/PrecGeneralisedConjugateResidual.h ./algorithms/iterative/SchurRedBlack.h ./algorithms/LinearOperator.h ./algorithms/Preconditioner.h ./algorithms/SparseMatrix.h ./Algorithms.h ./AlignedAllocator.h ./cartesian/Cartesian_base.h ./cartesian/Cartesian_full.h ./cartesian/Cartesian_red_black.h ./Cartesian.h ./communicator/Communicator_base.h ./Communicator.h ./Config.h ./cshift/Cshift_common.h ./cshift/Cshift_mpi.h ./cshift/Cshift_none.h ./Cshift.h ./Grid.h ./Init.h ./lattice/Lattice_arith.h ./lattice/Lattice_base.h ./lattice/Lattice_comparison.h ./lattice/Lattice_comparison_utils.h ./lattice/Lattice_conformable.h ./lattice/Lattice_coordinate.h ./lattice/Lattice_ET.h ./lattice/Lattice_local.h ./lattice/Lattice_overload.h ./lattice/Lattice_peekpoke.h ./lattice/Lattice_reality.h ./lattice/Lattice_reduction.h ./lattice/Lattice_rng.h ./lattice/Lattice_trace.h ./lattice/Lattice_transfer.h ./lattice/Lattice_transpose.h ./lattice/Lattice_unary.h ./lattice/Lattice_where.h ./Lattice.h ./Log.h ./MacroMagic.h ./Old/Tensor_peek.h ./Old/Tensor_poke.h ./parallelIO/NerscIO.h ./parameterIO/XMLReader.h ./qcd/action/ActionBase.h ./qcd/action/Actions.h ./qcd/action/fermion/CayleyFermion5D.h ./qcd/action/fermion/ContinuedFractionFermion5D.h ./qcd/action/fermion/DomainWallFermion.h ./qcd/action/fermion/FermionImplTypedefs.h ./qcd/action/fermion/FermionOperator.h ./qcd/action/fermion/g5HermitianLinop.h ./qcd/action/fermion/MobiusFermion.h ./qcd/action/fermion/MobiusZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h ./qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonContfracTanhFermion.h ./qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h ./qcd/action/fermion/PartialFractionFermion5D.h ./qcd/action/fermion/ScaledShamirFermion.h ./qcd/action/fermion/ShamirZolotarevFermion.h ./qcd/action/fermion/WilsonCompressor.h ./qcd/action/fermion/WilsonFermion.h ./qcd/action/fermion/WilsonFermion5D.h ./qcd/action/fermion/WilsonKernels.h ./qcd/action/gauge/WilsonGaugeAction.h ./qcd/action/pseudofermion/TwoFlavour.h ./qcd/action/pseudofermion/TwoFlavourEvenOdd.h ./qcd/hmc/HMC.h ./qcd/hmc/integrators/Integrator.h ./qcd/hmc/integrators/Integrator_algorithm.h ./qcd/hmc/integrators/Integrator_base.h ./qcd/QCD.h ./qcd/spin/Dirac.h ./qcd/spin/TwoSpinor.h ./qcd/utils/CovariantCshift.h ./qcd/utils/LinalgUtils.h ./qcd/utils/SpaceTimeGrid.h ./qcd/utils/SUn.h ./qcd/utils/WilsonLoops.h ./simd/Grid_avx.h ./simd/Grid_avx512.h ./simd/Grid_empty.h ./simd/Grid_neon.h ./simd/Grid_qpx.h ./simd/Grid_sse4.h ./simd/Grid_vector_types.h ./simd/Grid_vector_unops.h ./Simd.h ./stencil/Lebesgue.h ./Stencil.h ./tensors/Tensor_arith.h ./tensors/Tensor_arith_add.h ./tensors/Tensor_arith_mac.h ./tensors/Tensor_arith_mul.h ./tensors/Tensor_arith_scalar.h ./tensors/Tensor_arith_sub.h ./tensors/Tensor_class.h ./tensors/Tensor_determinant.h ./tensors/Tensor_exp.h ./tensors/Tensor_extract_merge.h ./tensors/Tensor_index.h ./tensors/Tensor_inner.h ./tensors/Tensor_logical.h ./tensors/Tensor_outer.h ./tensors/Tensor_reality.h ./tensors/Tensor_Ta.h ./tensors/Tensor_trace.h ./tensors/Tensor_traits.h ./tensors/Tensor_transpose.h ./tensors/Tensor_unary.h ./Tensors.h ./Threads.h ./Timer.h
+HFILES=./algorithms/approx/bigfloat.h ./algorithms/approx/bigfloat_double.h ./algorithms/approx/Chebyshev.h ./algorithms/approx/MultiShiftFunction.h ./algorithms/approx/Remez.h ./algorithms/approx/Zolotarev.h ./algorithms/CoarsenedMatrix.h ./algorithms/iterative/AdefGeneric.h ./algorithms/iterative/ConjugateGradient.h ./algorithms/iterative/ConjugateGradientMultiShift.h ./algorithms/iterative/ConjugateResidual.h ./algorithms/iterative/NormalEquations.h ./algorithms/iterative/PrecConjugateResidual.h ./algorithms/iterative/PrecGeneralisedConjugateResidual.h ./algorithms/iterative/SchurRedBlack.h ./algorithms/LinearOperator.h ./algorithms/Preconditioner.h ./algorithms/SparseMatrix.h ./Algorithms.h ./AlignedAllocator.h ./cartesian/Cartesian_base.h ./cartesian/Cartesian_full.h ./cartesian/Cartesian_red_black.h ./Cartesian.h ./communicator/Communicator_base.h ./Communicator.h ./Config.h ./cshift/Cshift_common.h ./cshift/Cshift_mpi.h ./cshift/Cshift_none.h ./Cshift.h ./Grid.h ./Init.h ./lattice/Lattice_arith.h ./lattice/Lattice_base.h ./lattice/Lattice_comparison.h ./lattice/Lattice_comparison_utils.h ./lattice/Lattice_conformable.h ./lattice/Lattice_coordinate.h ./lattice/Lattice_ET.h ./lattice/Lattice_local.h ./lattice/Lattice_overload.h ./lattice/Lattice_peekpoke.h ./lattice/Lattice_reality.h ./lattice/Lattice_reduction.h ./lattice/Lattice_rng.h ./lattice/Lattice_trace.h ./lattice/Lattice_transfer.h ./lattice/Lattice_transpose.h ./lattice/Lattice_unary.h ./lattice/Lattice_where.h ./Lattice.h ./Log.h ./MacroMagic.h ./Old/Tensor_peek.h ./Old/Tensor_poke.h ./parallelIO/NerscIO.h ./qcd/action/ActionBase.h ./qcd/action/Actions.h ./qcd/action/fermion/CayleyFermion5D.h ./qcd/action/fermion/ContinuedFractionFermion5D.h ./qcd/action/fermion/DomainWallFermion.h ./qcd/action/fermion/FermionImplTypedefs.h ./qcd/action/fermion/FermionOperator.h ./qcd/action/fermion/g5HermitianLinop.h ./qcd/action/fermion/MobiusFermion.h ./qcd/action/fermion/MobiusZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonCayleyTanhFermion.h ./qcd/action/fermion/OverlapWilsonCayleyZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonContfracTanhFermion.h ./qcd/action/fermion/OverlapWilsonContfracZolotarevFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionTanhFermion.h ./qcd/action/fermion/OverlapWilsonPartialFractionZolotarevFermion.h ./qcd/action/fermion/PartialFractionFermion5D.h ./qcd/action/fermion/ScaledShamirFermion.h ./qcd/action/fermion/ShamirZolotarevFermion.h ./qcd/action/fermion/WilsonCompressor.h ./qcd/action/fermion/WilsonFermion.h ./qcd/action/fermion/WilsonFermion5D.h ./qcd/action/fermion/WilsonKernels.h ./qcd/action/gauge/WilsonGaugeAction.h ./qcd/action/pseudofermion/TwoFlavour.h ./qcd/action/pseudofermion/TwoFlavourEvenOdd.h ./qcd/hmc/HMC.h ./qcd/hmc/integrators/Integrator.h ./qcd/hmc/integrators/Integrator_algorithm.h ./qcd/hmc/integrators/Integrator_base.h ./qcd/QCD.h ./qcd/spin/Dirac.h ./qcd/spin/TwoSpinor.h ./qcd/utils/CovariantCshift.h ./qcd/utils/LinalgUtils.h ./qcd/utils/SpaceTimeGrid.h ./qcd/utils/SUn.h ./qcd/utils/WilsonLoops.h ./simd/Grid_avx.h ./simd/Grid_avx512.h ./simd/Grid_empty.h ./simd/Grid_neon.h ./simd/Grid_qpx.h ./simd/Grid_sse4.h ./simd/Grid_vector_types.h ./simd/Grid_vector_unops.h ./Simd.h ./stencil/Lebesgue.h ./Stencil.h ./tensors/Tensor_arith.h ./tensors/Tensor_arith_add.h ./tensors/Tensor_arith_mac.h ./tensors/Tensor_arith_mul.h ./tensors/Tensor_arith_scalar.h ./tensors/Tensor_arith_sub.h ./tensors/Tensor_class.h ./tensors/Tensor_determinant.h ./tensors/Tensor_exp.h ./tensors/Tensor_extract_merge.h ./tensors/Tensor_index.h ./tensors/Tensor_inner.h ./tensors/Tensor_logical.h ./tensors/Tensor_outer.h ./tensors/Tensor_reality.h ./tensors/Tensor_Ta.h ./tensors/Tensor_trace.h ./tensors/Tensor_traits.h ./tensors/Tensor_transpose.h ./tensors/Tensor_unary.h ./Tensors.h ./Threads.h ./Timer.h
 
 CCFILES=./algorithms/approx/MultiShiftFunction.cc ./algorithms/approx/Remez.cc ./algorithms/approx/Zolotarev.cc ./Init.cc ./Log.cc ./qcd/action/fermion/CayleyFermion5D.cc ./qcd/action/fermion/ContinuedFractionFermion5D.cc ./qcd/action/fermion/PartialFractionFermion5D.cc ./qcd/action/fermion/WilsonFermion.cc ./qcd/action/fermion/WilsonFermion5D.cc ./qcd/action/fermion/WilsonKernels.cc ./qcd/action/fermion/WilsonKernelsHand.cc ./qcd/hmc/HMC.cc ./qcd/hmc/integrators/Integrator.cc ./qcd/spin/Dirac.cc ./qcd/utils/SpaceTimeGrid.cc ./stencil/Lebesgue.cc ./stencil/Stencil_common.cc
diff --git a/lib/qcd/action/fermion/CayleyFermion5D.cc b/lib/qcd/action/fermion/CayleyFermion5D.cc
index 1ec7f930..76066adf 100644
--- a/lib/qcd/action/fermion/CayleyFermion5D.cc
+++ b/lib/qcd/action/fermion/CayleyFermion5D.cc
@@ -158,12 +158,14 @@ namespace QCD {
  template<class Impl>
   void CayleyFermion5D<Impl>::Meooe       (const FermionField &psi, FermionField &chi)
   {
+    int Ls=this->Ls;
+
     FermionField tmp(psi._grid);
     // Assemble the 5d matrix
     Meooe5D(psi,tmp); 
-    std::cout << "Meooe Test replacement norm2 tmp = " <<norm2(tmp)<<std::endl;
 
-    int Ls=this->Ls;
+#if 0
+    std::cout << "Meooe Test replacement norm2 tmp = " <<norm2(tmp)<<std::endl;
     for(int s=0;s<Ls;s++){
       if ( s==0 ) {
 	//	tmp = bs psi[s] + cs[s] psi[s+1}
@@ -179,6 +181,8 @@ namespace QCD {
       }
     }
     std::cout << "Meooe Test replacement norm2 tmp old = " <<norm2(tmp)<<std::endl;
+#endif
+
     // Apply 4d dslash
     if ( psi.checkerboard == Odd ) {
       this->DhopEO(tmp,chi,DaggerNo);
@@ -199,8 +203,8 @@ namespace QCD {
     }
 
     Meooe5D(tmp,chi); 
+#if 0
     std::cout << "Meooe Test replacement norm2 chi new = " <<norm2(chi)<<std::endl;
-
     // Assemble the 5d matrix
     int Ls=this->Ls;
     for(int s=0;s<Ls;s++){
@@ -216,6 +220,7 @@ namespace QCD {
       }
     }
     std::cout << "Meooe Test replacement norm2 chi old = " <<norm2(chi)<<std::endl;
+#endif
 
   }
 
diff --git a/scripts/configure-commands b/scripts/configure-commands
index 4cb4b44f..d50da271 100755
--- a/scripts/configure-commands
+++ b/scripts/configure-commands
@@ -79,5 +79,8 @@ CXX=clang++ ../../configure --enable-simd=AVX CXXFLAGS="-mavx -O3 -I/opt/local/i
 clang-avx2-mpi)
 CXX=clang++ ../../configure --enable-simd=AVX2 CXXFLAGS="-mavx2 -mfma -O3 -I/opt/local/include/openmpi-mp/ -std=c++11" LDFLAGS=-L/opt/local/lib/openmpi-mp/ LIBS="-lmpi -lmpi_cxx -lgmp -lmpfr" --enable-comms=mpi
 ;;
+clang-avx2)
+CXX=clang++ ../../configure --enable-simd=AVX2 CXXFLAGS="-mavx2 -mfma -O3 -std=c++11" LDFLAGS="-L/usr/local/lib/" LIBS="-lgmp -lmpfr" --enable-comms=none
+;;
 esac
 echo -e $NORMAL
diff --git a/tests/Make.inc b/tests/Make.inc
index d5b926e6..111eed35 100644
--- a/tests/Make.inc
+++ b/tests/Make.inc
@@ -1,5 +1,5 @@
 
-bin_PROGRAMS = Test_GaugeAction Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_contfrac_force Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_hdcr Test_gamma Test_hmc_EOWilsonFermionGauge Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_partfrac_force Test_quenched_update Test_remez Test_rng Test_rng_fixed Test_simd Test_stencil Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi
+bin_PROGRAMS = Test_GaugeAction Test_cayley_cg Test_cayley_coarsen_support Test_cayley_even_odd Test_cayley_ldop_cr Test_cf_coarsen_support Test_cf_cr_unprec Test_cheby Test_contfrac_cg Test_contfrac_even_odd Test_contfrac_force Test_cshift Test_cshift_red_black Test_dwf_cg_prec Test_dwf_cg_schur Test_dwf_cg_unprec Test_dwf_cr_unprec Test_dwf_even_odd Test_dwf_force Test_dwf_fpgcr Test_dwf_hdcr Test_gamma Test_gparity Test_hmc_EOWilsonFermionGauge Test_hmc_WilsonFermionGauge Test_hmc_WilsonGauge Test_lie_generators Test_main Test_multishift_sqrt Test_nersc_io Test_partfrac_force Test_quenched_update Test_remez Test_rng Test_rng_fixed Test_simd Test_stencil Test_wilson_cg_prec Test_wilson_cg_schur Test_wilson_cg_unprec Test_wilson_cr_unprec Test_wilson_even_odd Test_wilson_force Test_wilson_force_phiMdagMphi Test_wilson_force_phiMphi
 
 
 Test_GaugeAction_SOURCES=Test_GaugeAction.cc
@@ -90,6 +90,10 @@ Test_gamma_SOURCES=Test_gamma.cc
 Test_gamma_LDADD=-lGrid
 
 
+Test_gparity_SOURCES=Test_gparity.cc
+Test_gparity_LDADD=-lGrid
+
+
 Test_hmc_EOWilsonFermionGauge_SOURCES=Test_hmc_EOWilsonFermionGauge.cc
 Test_hmc_EOWilsonFermionGauge_LDADD=-lGrid
 
diff --git a/tests/Test_gparity.cc b/tests/Test_gparity.cc
new file mode 100644
index 00000000..46c019c6
--- /dev/null
+++ b/tests/Test_gparity.cc
@@ -0,0 +1,73 @@
+#include <Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+template<class d>
+struct scal {
+  d internal;
+};
+
+  Gamma::GammaMatrix Gmu [] = {
+    Gamma::GammaX,
+    Gamma::GammaY,
+    Gamma::GammaZ,
+    Gamma::GammaT
+  };
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  const int Ls=4;
+  const int L=4;
+  std::vector<int> latt_2f(Nd,L);
+  std::vector<int> latt_1f(Nd,L); latt_1f[0] = 2L;
+
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi(); //node layout
+
+  GridCartesian         * UGrid_1f   = SpaceTimeGrid::makeFourDimGrid(latt_1f, simd_layout, mpi_layout);
+  GridRedBlackCartesian * UrbGrid_1f = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_1f);
+  GridCartesian         * FGrid_1f   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid_1f);
+  GridRedBlackCartesian * FrbGrid_1f = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid_1f);
+
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5_1f(FGrid_1f);  RNG5_1f.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4_1f(UGrid_1f);  RNG4_1f.SeedFixedIntegers(seeds4);
+
+  LatticeFermion    src_1f(FGrid_1f); random(RNG5_1f,src_1f);
+  LatticeFermion result_1f(FGrid_1f); result_1f=zero;
+  LatticeGaugeField Umu_1f(UGrid_1f); random(RNG4_1f,Umu_1f);
+
+  //Coordinate grid for reference
+  LatticeInteger xcoor_1f(UGrid_1f);
+  LatticeCoordinate(xcoor_1f,0);
+
+  //Copy-conjugate the gauge field
+  //First C-shift the lattice by Lx/2
+  {
+    LatticeGaugeField Umu_shift = conjugate( Cshift(Umu_1f,0,L) );
+    Umu_1f = where( xcoor_1f >= Integer(L), Umu_shift, Umu_1f );
+  }
+
+  //Make the gauge field antiperiodic in x-direction
+  Umu_1f = where(xcoor_1f == Integer(L-1), -Umu_1f, Umu_1f);
+    
+  RealD mass=0.1;
+  RealD M5=1.8;
+  DomainWallFermionR Ddwf(Umu_1f,*FGrid_1f,*FrbGrid_1f,*UGrid_1f,*UrbGrid_1f,mass,M5);
+
+  LatticeFermion    src_o_1f(FrbGrid_1f);
+  LatticeFermion result_o_1f(FrbGrid_1f);
+  pickCheckerboard(Odd,src_o_1f,src_1f);
+  result_o_1f=zero;
+
+  SchurDiagMooeeOperator<DomainWallFermionR,LatticeFermion> HermOpEO(Ddwf);
+  ConjugateGradient<LatticeFermion> CG(1.0e-8,10000);
+  CG(HermOpEO,src_o_1f,result_o_1f);
+
+  Grid_finalize();
+}