for dwf_precondition

lib/algorithms/Algorithms.h

@@ -49,6 +49,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
+#include <Grid/algorithms/iterative/InexactPrecConjugateGradient.h>
 #include <Grid/algorithms/CoarsenedMatrix.h>
 #include <Grid/algorithms/FFT.h>
 

lib/algorithms/approx/MultiShiftFunction.cc

@@ -53,4 +53,10 @@ void MultiShiftFunction::csv(std::ostream &out)
   }
   return;
 }
+
+RealD __InverseApproximation(RealD x)
+{
+  return 1.0/x;
+}
+
 }

lib/algorithms/approx/MultiShiftFunction.h

@@ -63,5 +63,7 @@ public:
   }
 
 };
+RealD __InverseApproximation(RealD x);
+
 }
 #endif

lib/algorithms/iterative/InexactPrecConjugateGradient.h

@@ -0,0 +1,278 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./lib/algorithms/iterative/InexactPrecConjugateGradient.h
+
+    Copyright (C) 2015
+
+Author: Christopher Kelly <ckelly@phys.columbia.edu>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#ifndef GRID_INEXACT_PREC_CONJUGATE_GRADIENT_H_
+#define GRID_INEXACT_PREC_CONJUGATE_GRADIENT_H_
+
+namespace Grid {
+
+//Inexact preconditioned CG based on Golub, Ye, SIAM J. Sci. Comput., 21(4), 1305–1320. 
+//(https://pdfs.semanticscholar.org/d2a9/d5bab02146a7fe3a244677432d21e33a2d98.pdf)
+template <class Field>
+class InexactPreconditionedConjugateGradient : public OperatorFunction<Field> {
+ public:
+  bool ErrorOnNoConverge;  // throw an assert when the CG fails to converge.
+                           // Defaults true.
+  RealD Tolerance;
+  Integer MaxIterations;
+  Integer IterationsToComplete; //Number of iterations the CG took to finish. Filled in upon completion
+  
+  LinearOperatorBase<Field> &Prec;
+
+  InexactPreconditionedConjugateGradient(LinearOperatorBase<Field> &_Prec, RealD tol, Integer maxit, bool err_on_no_conv = true)
+    : Prec(_Prec),
+    Tolerance(tol),
+    MaxIterations(maxit),
+    ErrorOnNoConverge(err_on_no_conv){};
+
+  void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
+
+    psi.checkerboard = src.checkerboard;
+    conformable(psi, src);
+
+    Real ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq; //inner stopping condition
+
+    Field p(src);
+    Field r(src);
+    Field rnm1(src);
+    Field mmp(src);
+    Field z(src);
+
+    //Initialize variables
+    Linop.HermOp(psi, mmp);   
+    r = src - mmp;
+    
+    Real cp = norm2(r);
+    
+    p = zero;
+    Real alpha = 0, beta = 0;
+    
+    Real z_nm1_dot_r_nm1;
+
+    int n;
+    for(n=1; n <= MaxIterations; n++) {
+      //Check stopping condition
+      if (cp <= rsq) {
+        Linop.HermOp(psi, mmp);
+        r = mmp - src;
+
+        RealD srcnorm = sqrt(norm2(src));
+        RealD resnorm = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage << "InexactPreconditionedConjugateGradient Converged on iteration " << n << std::endl;
+        std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<<std::endl;
+  	std::cout << GridLogMessage << "\tTrue residual " << true_residual<<std::endl;
+  	std::cout << GridLogMessage << "\tTarget " << Tolerance << std::endl;
+
+        if (ErrorOnNoConverge) assert(true_residual / Tolerance < 10000.0);
+
+  	IterationsToComplete = n;	
+
+        return;
+      }
+      
+      std::cout << GridLogIterative << std::setprecision(8)
+		<< "InexactPreconditionedConjugateGradient: n=" << n << " residual " << cp << " target " << rsq << std::endl;
+
+      //Apply preconditioner to current residual
+      Prec.HermOp(r, z);
+
+      //Update beta and store appropriate variables for next iteration
+      Real z_n_dot_r_n  = sqrt(norm(innerProduct(z,r)));
+
+      if(n>1){
+	//  z^T_n ( r_n - r_{n-1} )
+	//  -----------------------
+	//     z^T_{n-1} r_{n-1}
+
+	Real z_n_dot_r_nm1 = sqrt(norm(innerProduct(z,rnm1)));
+	beta = ( z_n_dot_r_n - z_n_dot_r_nm1 ) / z_nm1_dot_r_nm1;
+	std::cout << GridLogIterative << "beta " << beta << std::endl;
+      }
+
+      z_nm1_dot_r_nm1 = z_n_dot_r_n; //for next iteration
+      rnm1 = r;
+
+      axpy(p, beta, p, z); //p = beta * p + z
+
+      //Compute alpha
+      Linop.HermOp(p, mmp);      
+      alpha = z_n_dot_r_n / sqrt(norm(innerProduct(p, mmp)));
+
+      std::cout << GridLogIterative << "alpha " << alpha << std::endl;
+
+      //Update residual and solution
+      cp = axpy_norm(r, -alpha, mmp, r);
+      axpy(psi, alpha, p, psi);
+    }
+    std::cout << GridLogMessage << "InexactPreconditionedConjugateGradient did NOT converge"
+              << std::endl;
+
+    if (ErrorOnNoConverge) assert(0);
+    IterationsToComplete = n;
+  }
+};
+
+
+template<class Field>
+class PolynomialPreconditioner :  public LinearOperatorBase<Field> {
+  Chebyshev<Field> Cheby;
+  LinearOperatorBase<Field> &linop;
+public:
+  int InnerIterations;
+  int order;
+  PolynomialPreconditioner(LinearOperatorBase<Field> &_linop,RealD lo, RealD hi, int _order) 
+    : linop(_linop), Cheby(lo,hi,_order,__InverseApproximation) 
+    {
+      InnerIterations=0;
+      order = _order;
+    };
+
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ 
+    HermOp(in,out);
+    n1 = 0; n2 = norm2(out); 
+  }
+  void HermOp(const Field &in, Field &out){ 
+    Cheby(linop,in,out); 
+    InnerIterations+=order;
+  }
+};
+
+template<class Field>
+class DoNothingLinearOperator :  public LinearOperatorBase<Field> {
+public:
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ out = in; n1 = 0; n2 = norm2(out); }
+  void HermOp(const Field &in, Field &out){ out = in; }
+};
+
+template<class Field>
+class FixedIterConjugateGradientPreconditioner :  public LinearOperatorBase<Field> {
+public:
+  LinearOperatorBase<Field> &linop;
+  ConjugateGradient<Field> CG;
+
+  FixedIterConjugateGradientPreconditioner (LinearOperatorBase<Field> &_linop, Integer _iter): linop(_linop), CG(1e-20, _iter){
+    CG.ErrorOnNoConverge = false;
+  }
+
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ 
+    out = zero;
+    CG(linop,in,out);    
+    n2 = norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+    out = zero;
+    CG(linop,in,out);    
+  }
+};
+
+template<class Field>
+class SloppyConjugateGradientPreconditioner :  public LinearOperatorBase<Field> {
+public:
+  LinearOperatorBase<Field> &linop;
+  ConjugateGradient<Field> CG;
+  int InnerIterations;
+
+  SloppyConjugateGradientPreconditioner (LinearOperatorBase<Field> &_linop, Real _resid, Integer max_iter): linop(_linop), CG(_resid, max_iter), InnerIterations(0){
+  }
+
+  void ResetCounters(){ InnerIterations = 0; }
+
+  void OpDiag (const Field &in, Field &out){ assert(0); }
+  void OpDir  (const Field &in, Field &out,int dir,int disp){ assert(0); }
+  void Op     (const Field &in, Field &out){ assert(0); }
+  void AdjOp  (const Field &in, Field &out){ assert(0); }
+  void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){ 
+    out = zero;
+    CG(linop,in,out);    
+    InnerIterations += CG.IterationsToComplete;
+    n2 = norm2(out);
+  }
+  void HermOp(const Field &in, Field &out){
+    out = zero;
+    CG(linop,in,out);
+    InnerIterations += CG.IterationsToComplete;
+  }
+};
+
+
+template<class FieldH, class FieldL>
+class SloppyConjugateGradientLowerPrecPreconditioner :  public LinearOperatorBase<FieldH> {
+public:
+  LinearOperatorBase<FieldL> &linop;
+  ConjugateGradient<FieldL> CG;
+  GridBase* L_grid; //lower-prec Grid
+  int InnerIterations;
+  FieldL tmp_l1;
+  FieldL tmp_l2;
+
+  SloppyConjugateGradientLowerPrecPreconditioner (LinearOperatorBase<FieldL> &_linop, GridBase* _L_grid, Real _resid, Integer max_iter): linop(_linop), CG(_resid, max_iter), InnerIterations(0), L_grid(_L_grid), tmp_l1(_L_grid), tmp_l2(_L_grid){
+    CG.ErrorOnNoConverge = false;
+  }
+
+  void ResetCounters(){ InnerIterations = 0; }
+
+  void OpDiag (const FieldH &in, FieldH &out){ assert(0); }
+  void OpDir  (const FieldH &in, FieldH &out,int dir,int disp){ assert(0); }
+  void Op     (const FieldH &in, FieldH &out){ assert(0); }
+  void AdjOp  (const FieldH &in, FieldH &out){ assert(0); }
+  void HermOpAndNorm(const FieldH &in, FieldH &out,RealD &n1,RealD &n2){ 
+    precisionChange(tmp_l1, in);
+    tmp_l2 = zero;
+    CG(linop,tmp_l1,tmp_l2);    
+    InnerIterations += CG.IterationsToComplete;
+    precisionChange(out, tmp_l2);
+    n2 = norm2(tmp_l2);
+  }
+  void HermOp(const FieldH &in, FieldH &out){
+    precisionChange(tmp_l1, in);
+    tmp_l2 = zero;
+    CG(linop,tmp_l1,tmp_l2);    
+    InnerIterations += CG.IterationsToComplete;
+    precisionChange(out, tmp_l2);
+  }
+};
+
+}
+
+
+
+#endif

tests/solver/Test_compression.cc

@@ -0,0 +1,963 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_cg_prec.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include<bitset>
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+//Preconditioning:   M psi = chi
+//               =  M P^-1 P psi = chi
+//               =  M P^-1 psi' = chi
+
+//Solve for psi' using M P^-1 as operator, then apply  P^-1 psi' = psi 
+
+//Inexact preconditioned CG requires slight modification because we want to avoid computing P^-1 exactly
+
+
+/////////////////////////////////////////////////////////////
+// Base classes for iterative processes based on operators
+// single input vec, single output vec.
+/////////////////////////////////////////////////////////////
+
+//The compressor
+#if 0
+
+
+//Basic copy of WilsonCompressor for demonstration
+template<class _Hspinor,class _Spinor, class projector>
+class WilsonTestCompressorTemplate
+{
+ public:
+  
+  int mu,dag;  
+
+  void Point(int p) { mu=p; };
+
+  WilsonTestCompressorTemplate(int _dag=0){
+    //printf("WilsonTestCompressorTemplate constructor\n");
+    dag = _dag;
+  }
+
+  typedef _Spinor         SiteSpinor;
+  typedef _Hspinor     SiteHalfSpinor;
+  typedef _Hspinor SiteHalfCommSpinor;
+  typedef typename SiteHalfSpinor::vector_type     vComplexIn;
+  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexIn); //number of complex numbers in SiteHalfSpinor
+
+  inline int CommDatumSize(void) {
+    //printf("WilsonTestCompressorTemplate CommDatumSize\n");
+    return sizeof(SiteHalfCommSpinor);
+  }
+
+  /*****************************************************/
+  /* Compress includes precision change if mpi data is not same */
+  /*****************************************************/
+  inline void Compress(SiteHalfSpinor *buf,Integer o,const SiteSpinor &in) {
+    //printf("WilsonTestCompressorTemplate Compress\n");
+    projector::Proj(buf[o],in,mu,dag);
+  }
+
+  /*****************************************************/
+  /* Exchange includes precision change if mpi data is not same */
+  /*****************************************************/
+  inline void Exchange(SiteHalfSpinor *mp,
+                       SiteHalfSpinor *vp0,
+                       SiteHalfSpinor *vp1,
+		       Integer type,Integer o){
+    //printf("WilsonTestCompressorTemplate Exchange\n");
+    exchange(mp[2*o],mp[2*o+1],vp0[o],vp1[o],type);
+  }
+
+  /*****************************************************/
+  /* Have a decompression step if mpi data is not same */
+  /*****************************************************/
+  inline void Decompress(SiteHalfSpinor *out,
+			 SiteHalfSpinor *in, Integer o) {    
+    //printf("WilsonTestCompressorTemplate Decompress\n");
+    assert(0);
+  }
+
+  /*****************************************************/
+  /* Compress Exchange                                 */
+  /*****************************************************/
+  inline void CompressExchange(SiteHalfSpinor *out0,
+			       SiteHalfSpinor *out1,
+			       const SiteSpinor *in,
+			       Integer j,Integer k, Integer m,Integer type){
+    //printf("WilsonTestCompressorTemplate CompressExchange\n");
+    SiteHalfSpinor temp1, temp2,temp3,temp4;
+    projector::Proj(temp1,in[k],mu,dag);
+    projector::Proj(temp2,in[m],mu,dag);
+    exchange(out0[j],out1[j],temp1,temp2,type);
+  }
+
+  /*****************************************************/
+  /* Pass the info to the stencil */
+  /*****************************************************/
+  inline bool DecompressionStep(void) { return false; }
+
+};
+
+#elif 0
+
+
+//Compressor that unpacks vectorized data to scalar
+template<class _Hspinor,class _Spinor, class projector>
+class WilsonTestCompressorTemplate
+{
+ public:
+  
+  int mu,dag;  
+
+  void Point(int p) { mu=p; };
+
+  WilsonTestCompressorTemplate(int _dag=0){
+    dag = _dag;
+  }
+
+  typedef _Spinor         SiteSpinor;
+  typedef _Hspinor     SiteHalfSpinor;
+  typedef _Hspinor SiteHalfCommSpinor;
+  typedef typename SiteHalfSpinor::vector_type     vComplexIn;
+  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexIn); //number of complex numbers in SiteHalfSpinor
+
+  typedef typename SiteHalfSpinor::scalar_object ScalarSiteHalfSpinor;
+
+  constexpr static int Nsimd = vComplexIn::Nsimd();
+
+  inline int CommDatumSize(void) {
+    return Nsimd*sizeof(ScalarSiteHalfSpinor);
+  }
+
+  /*****************************************************/
+  /* Compress includes precision change if mpi data is not same */
+  /*****************************************************/
+  inline void Compress(SiteHalfSpinor *buf,Integer o,const SiteSpinor &in) {
+    SiteHalfSpinor hsp;
+    projector::Proj(hsp,in,mu,dag);
+    
+    ScalarSiteHalfSpinor* to = (ScalarSiteHalfSpinor*)buf + o*Nsimd;
+    
+    std::vector<ScalarSiteHalfSpinor*> extract_args(Nsimd);
+    for(int i=0;i<Nsimd;i++) extract_args[i] = to+i;
+    extract1(hsp,extract_args,0);
+  }
+
+  /*****************************************************/
+  /* Exchange includes precision change if mpi data is not same */
+  /*****************************************************/
+  inline void Exchange(SiteHalfSpinor *mp,
+                       SiteHalfSpinor *vp0,
+                       SiteHalfSpinor *vp1,
+		       Integer type,Integer o){
+    ScalarSiteHalfSpinor* vpp0 = (ScalarSiteHalfSpinor*)vp0 + o*Nsimd;
+    ScalarSiteHalfSpinor* vpp1 = (ScalarSiteHalfSpinor*)vp1 + o*Nsimd;
+    
+    std::vector<ScalarSiteHalfSpinor*> merge_args0(Nsimd), merge_args1(Nsimd);
+    for(int i=0;i<Nsimd;i++){
+      merge_args0[i] = vpp0+i;
+      merge_args1[i] = vpp1+i;
+    }
+
+    SiteHalfSpinor vt0,vt1;
+    merge1(vt0,merge_args0,0);
+    merge1(vt1,merge_args1,0);
+
+    exchange(mp[2*o],mp[2*o+1],vt0,vt1,type);
+  }
+
+  /*****************************************************/
+  /* Have a decompression step if mpi data is not same */
+  /*****************************************************/
+  inline void Decompress(SiteHalfSpinor *out,
+			 SiteHalfSpinor *in, Integer o) {    
+    ScalarSiteHalfSpinor* hin = (ScalarSiteHalfSpinor*)in + o*Nsimd;
+    std::vector<ScalarSiteHalfSpinor*> merge_args(Nsimd);
+    for(int i=0;i<Nsimd;i++) merge_args[i] = hin+i;
+    merge1(out[o],merge_args,0);
+  }
+
+  /*****************************************************/
+  /* Compress Exchange                                 */
+  /*****************************************************/
+  inline void CompressExchange(SiteHalfSpinor *out0,
+			       SiteHalfSpinor *out1,
+			       const SiteSpinor *in,
+			       Integer j,Integer k, Integer m,Integer type){
+    SiteHalfSpinor temp1, temp2,temp3,temp4;
+    projector::Proj(temp1,in[k],mu,dag);
+    projector::Proj(temp2,in[m],mu,dag);
+    exchange(temp3,temp4,temp1,temp2,type);
+
+    ScalarSiteHalfSpinor* hout0 = (ScalarSiteHalfSpinor*)out0 + j*Nsimd;
+    ScalarSiteHalfSpinor* hout1 = (ScalarSiteHalfSpinor*)out1 + j*Nsimd;
+
+    std::vector<ScalarSiteHalfSpinor*> extract_args0(Nsimd), extract_args1(Nsimd);
+    for(int i=0;i<Nsimd;i++){
+      extract_args0[i] = hout0+i;
+      extract_args1[i] = hout1+i;
+    }
+    extract1(temp3,extract_args0,0);
+    extract1(temp4,extract_args1,0);
+  }
+
+  /*****************************************************/
+  /* Pass the info to the stencil */
+  /*****************************************************/
+  inline bool DecompressionStep(void) { return true; }
+
+};
+
+#else
+
+//Access elements of std::complex
+template<typename T>
+inline T & cmplx_reim(std::complex<T> &c, const int reim){
+  return reinterpret_cast<T(&)[2]>(c)[reim];
+}
+
+template<typename T>
+inline const T & cmplx_reim(const std::complex<T> &c, const int reim){
+  return reinterpret_cast<const T(&)[2]>(c)[reim];
+}
+
+
+//Pack and unpack float/double to fixed point representation of SZ bits
+template<int SZ>
+struct signedIntMap{};
+
+template<>
+struct signedIntMap<8>{ typedef int8_t type; };
+template<>
+struct signedIntMap<16>{ typedef int16_t type; };
+
+
+template<typename T, int SZ>
+inline typename signedIntMap<SZ>::type packN(T val){
+  return typename signedIntMap<SZ>::type( (1<<(SZ-2) ) * val );
+}
+template<typename T, int SZ>
+inline T unpackN(typename signedIntMap<SZ>::type val){
+  return T(val)/(1<<(SZ-2));
+}
+
+template<typename T>
+struct getHalfSpinorColors{
+  //template <typename vtype> using iImplHalfSpinor        = iScalar<iVector<iVector<vtype, Dimension>, Nhs> >;
+  enum { value = sizeof(typename T::element::element)/sizeof(typename T::element::element::element) };
+};
+
+//Compressor that compresses to a single magnitude and Nhs*Dimension fixed point integers of size packSize bits
+template<class _Hspinor,class _Spinor, class projector, int packSize = 16>
+class WilsonTestCompressorTemplate
+{
+ public:
+  
+  int mu,dag;  
+
+  void Point(int p) { mu=p; };
+
+  WilsonTestCompressorTemplate(int _dag=0){
+    dag = _dag;
+  }
+
+  typedef _Spinor         SiteSpinor;
+  typedef _Hspinor     SiteHalfSpinor;
+  typedef _Hspinor SiteHalfCommSpinor;
+  typedef typename SiteHalfSpinor::vector_type     vComplexIn;
+  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexIn); //number of complex numbers in SiteHalfSpinor
+
+  typedef typename SiteHalfSpinor::scalar_object ScalarSiteHalfSpinor;
+
+  constexpr static int Nsimd = vComplexIn::Nsimd();
+  constexpr static int Dimension = getHalfSpinorColors<SiteHalfSpinor>::value;
+ 
+  typedef typename ScalarSiteHalfSpinor::scalar_type stype; //std::complex
+  typedef typename stype::value_type srtype; //float/double
+
+  //Pack and unpack *scalar* SiteHalfSpinor objects
+  void packSpinor(void* tov, const ScalarSiteHalfSpinor &from){
+    uint8_t* to = (uint8_t*)tov;
+    typedef typename signedIntMap<packSize>::type packedType;
+
+    srtype max = 0;
+    for(int s=0;s<Nhs;s++)
+      for(int c=0;c<Dimension;c++)
+	for(int reim=0;reim<2;reim++)
+	  if(fabs(cmplx_reim( from()(s)(c), reim )) > max )
+	    max =  fabs(cmplx_reim( from()(s)(c), reim )) ;
+  
+    *( (srtype*)to ) = max; //copy the normalization to the buffer
+    to += sizeof(srtype);
+  
+    packedType *top = (packedType*)to;
+    packedType p;
+    srtype q;
+    for(int s=0;s<Nhs;s++)
+      for(int c=0;c<Dimension;c++)
+	for(int reim=0;reim<2;reim++){
+	  q = cmplx_reim( from()(s)(c), reim );
+	  if(max != 0.) q /= max;
+	  *(top++) = packN<srtype,packSize>(q);
+	}
+  }
+
+  void packSpinor(void* tov, const SiteHalfSpinor &from){
+    uint8_t* to = (uint8_t*)tov;
+    std::vector<ScalarSiteHalfSpinor> extracted(Nsimd);
+    extract(from,extracted);
+
+    static const int incr = sizeof(srtype) + Nhs*Dimension*2*sizeof(typename signedIntMap<packSize>::type);
+
+    for(int i=0;i<Nsimd;i++){
+      packSpinor((void*)to, extracted[i]);
+      to += incr;
+    }
+  }
+
+
+  void unpackSpinor(ScalarSiteHalfSpinor &to, void* fromv){
+    uint8_t* from = (uint8_t*)fromv;
+    typedef typename signedIntMap<packSize>::type packedType;
+
+    srtype norm = *( (srtype*)from ); 
+    from += sizeof(srtype);
+
+    packedType *fromp = (packedType*)from;
+    srtype q;
+    for(int s=0;s<Nhs;s++)
+      for(int c=0;c<Dimension;c++)
+	for(int reim=0;reim<2;reim++){
+	  q = unpackN<srtype,packSize>(*(fromp++) );
+	  if(norm != 0.) q *= norm;
+	  cmplx_reim( to()(s)(c), reim ) = q;
+	}
+  }
+
+  void unpackSpinor(SiteHalfSpinor &to, void* fromv){
+    uint8_t* from = (uint8_t*)fromv;
+    std::vector<ScalarSiteHalfSpinor> unpacked(Nsimd);
+
+    static const int incr = sizeof(srtype) + Nhs*Dimension*2*sizeof(typename signedIntMap<packSize>::type);
+
+    for(int i=0;i<Nsimd;i++){
+      unpackSpinor(unpacked[i],(void*)from);
+      from += incr;
+    }
+
+    merge(to,unpacked);
+  }
+
+  inline int CommDatumSize(void) {
+    return Nsimd*(  sizeof(srtype) + Nhs*Dimension*2*sizeof(typename signedIntMap<packSize>::type) );
+  }
+
+  /*****************************************************/
+  /* Compress includes precision change if mpi data is not same */
+  /*****************************************************/
+  void Compress(SiteHalfSpinor *buf,Integer o,const SiteSpinor &in) {
+    SiteHalfSpinor hsp;
+    projector::Proj(hsp,in,mu,dag);
+    
+    uint8_t* to = (uint8_t*)buf + o*CommDatumSize();
+    packSpinor(to, hsp);
+  }
+
+  /*****************************************************/
+  /* Exchange includes precision change if mpi data is not same */
+  /*****************************************************/
+  void Exchange(SiteHalfSpinor *mp,
+                       SiteHalfSpinor *vp0,
+                       SiteHalfSpinor *vp1,
+		       Integer type,Integer o){
+    uint8_t* vpp0 = (uint8_t*)vp0 + o*CommDatumSize();
+    uint8_t* vpp1 = (uint8_t*)vp1 + o*CommDatumSize();
+
+    SiteHalfSpinor vt0, vt1;
+    unpackSpinor(vt0, vpp0);
+    unpackSpinor(vt1, vpp1);
+
+    exchange(mp[2*o],mp[2*o+1],vt0,vt1,type);
+  }
+
+  /*****************************************************/
+  /* Have a decompression step if mpi data is not same */
+  /*****************************************************/
+  void Decompress(SiteHalfSpinor *out,
+			 SiteHalfSpinor *in, Integer o) {    
+    uint8_t* hin = (uint8_t*)in + o*CommDatumSize();
+    unpackSpinor(out[o],hin);
+  }
+
+  /*****************************************************/
+  /* Compress Exchange                                 */
+  /*****************************************************/
+  void CompressExchange(SiteHalfSpinor *out0,
+			       SiteHalfSpinor *out1,
+			       const SiteSpinor *in,
+			       Integer j,Integer k, Integer m,Integer type){
+    SiteHalfSpinor temp1, temp2,temp3,temp4;
+    projector::Proj(temp1,in[k],mu,dag);
+    projector::Proj(temp2,in[m],mu,dag);
+    exchange(temp3,temp4,temp1,temp2,type);
+
+    uint8_t* hout0 = (uint8_t*)out0 + j*CommDatumSize();
+    uint8_t* hout1 = (uint8_t*)out1 + j*CommDatumSize();
+    packSpinor(hout0, temp3);
+    packSpinor(hout1, temp4);
+  }
+
+  /*****************************************************/
+  /* Pass the info to the stencil */
+  /*****************************************************/
+  inline bool DecompressionStep(void) { return true; }
+
+};
+
+
+#endif
+
+
+
+template<typename HS,typename S, int packSize> using WilsonTestCompressor = WilsonTestCompressorTemplate<HS,S,WilsonProjector,packSize>;
+
+
+template<class vobj,class cobj>
+class WilsonStencilBasic : public CartesianStencil<vobj,cobj> {
+public:
+  double timer0;
+  double timer1;
+  double timer2;
+  double timer3;
+  double timer4;
+  double timer5;
+  double timer6;
+  uint64_t callsi;
+  void ZeroCountersi(void)
+  {
+    timer0=0;
+    timer1=0;
+    timer2=0;
+    timer3=0;
+    timer4=0;
+    timer5=0;
+    timer6=0;
+    callsi=0;
+  }
+  void Reporti(int calls)
+  {
+    if ( timer0 ) std::cout << GridLogMessage << " timer0 (HaloGatherOpt) " <<timer0/calls <<std::endl;
+    if ( timer1 ) std::cout << GridLogMessage << " timer1 (Communicate)   " <<timer1/calls <<std::endl;
+    if ( timer2 ) std::cout << GridLogMessage << " timer2 (CommsMerge )   " <<timer2/calls <<std::endl;
+    if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
+    if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
+  }
+
+
+  std::vector<int> same_node;
+  std::vector<int> surface_list;
+
+  WilsonStencilBasic(GridBase *grid,
+		int npoints,
+		int checkerboard,
+		const std::vector<int> &directions,
+		const std::vector<int> &distances)  
+    : CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) ,
+    same_node(npoints)
+  { 
+    ZeroCountersi();
+    surface_list.resize(0);
+  };
+
+  void BuildSurfaceList(int Ls,int vol4){
+
+    // find same node for SHM
+    // Here we know the distance is 1 for WilsonStencil
+    for(int point=0;point<this->_npoints;point++){
+      same_node[point] = this->SameNode(point);
+    }
+    
+    for(int site = 0 ;site< vol4;site++){
+      int local = 1;
+      for(int point=0;point<this->_npoints;point++){
+	if( (!this->GetNodeLocal(site*Ls,point)) && (!same_node[point]) ){ 
+	  local = 0;
+	}
+      }
+      if(local == 0) { 
+	surface_list.push_back(site);
+      }
+    }
+  }
+
+
+  template < class compressor>
+  void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress) 
+  {
+    std::vector<std::vector<CommsRequest_t> > reqs;
+    this->HaloExchangeOptGather(source,compress);
+    double t1=usecond();
+    this->Communicate();
+    double t2=usecond(); timer1 += t2-t1;
+    this->CommsMerge(compress);
+    double t3=usecond(); timer2 += t3-t2;
+    this->CommsMergeSHM(compress);
+    double t4=usecond(); timer3 += t4-t3;
+  }
+  
+  template <class compressor>
+  void HaloExchangeOptGather(const Lattice<vobj> &source,compressor &compress){
+    this->Prepare();
+    double t0=usecond();
+    this->HaloGatherOpt(source,compress);
+    double t1=usecond();
+    timer0 += t1-t0;
+    callsi++;
+  }
+
+  template <class compressor>
+  void HaloGatherOpt(const Lattice<vobj> &source,compressor &compress)
+  {
+    this->halogtime-=usecond();
+    this->HaloGather(source,compress);
+    this->halogtime+=usecond();
+  }
+};
+
+
+
+
+//This is hideous
+template<class S, int packSize = 16>
+class WilsonCompressedCommsImpl: public WilsonImpl<S,FundamentalRepresentation,CoeffReal>{
+public:
+  typedef WilsonImpl<S,FundamentalRepresentation,CoeffReal> WilsonBase;
+
+#define INHERIT_BASE(TYPE) typedef typename WilsonBase::TYPE TYPE
+
+  INHERIT_BASE(Gimpl);
+  INHERIT_GIMPL_TYPES(Gimpl);
+
+  INHERIT_BASE(Coeff_t);
+  
+  INHERIT_BASE(SiteSpinor);
+  INHERIT_BASE(SitePropagator);
+  INHERIT_BASE(SiteHalfSpinor);
+  INHERIT_BASE(SiteHalfCommSpinor);    
+  INHERIT_BASE(SiteDoubledGaugeField);
+
+  INHERIT_BASE(FermionField);
+  INHERIT_BASE(PropagatorField);
+  INHERIT_BASE(DoubledGaugeField);
+
+  //typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
+  typedef WilsonTestCompressor<SiteHalfSpinor, SiteSpinor, packSize> Compressor;  
+
+  INHERIT_BASE(ImplParams);
+  //INHERIT_BASE(StencilImpl);
+  typedef WilsonStencilBasic<SiteSpinor, SiteHalfSpinor> StencilImpl;
+
+  WilsonCompressedCommsImpl(const ImplParams &p = ImplParams()) : WilsonBase(p){}
+  
+  inline void multLink(SiteHalfSpinor &phi,
+		       const SiteDoubledGaugeField &U,
+		       const SiteHalfSpinor &chi,
+		       int mu,
+		       StencilEntry *SE,
+		       StencilImpl &St) {
+    mult(&phi(), &U(mu), &chi());
+  }
+
+#undef INHERIT_BASE
+};
+
+typedef WilsonCompressedCommsImpl<vComplexF,8> WilsonCompressedComms8ImplF;
+typedef WilsonCompressedCommsImpl<vComplexD,8> WilsonCompressedComms8ImplD;
+typedef WilsonCompressedCommsImpl<vComplexF,16> WilsonCompressedComms16ImplF;
+typedef WilsonCompressedCommsImpl<vComplexD,16> WilsonCompressedComms16ImplD;
+
+
+#define TO_INSTANTIATE \
+  DOIT(WilsonCompressedComms8ImplF)\
+  DOIT(WilsonCompressedComms8ImplD)\
+  DOIT(WilsonCompressedComms16ImplF)\
+  DOIT(WilsonCompressedComms16ImplD)
+
+#include "InstantiateImpl.impl"
+
+#undef TO_INSTANTIATE
+
+typedef DomainWallFermion<WilsonCompressedComms8ImplD> DomainWallFermionCompressedComms8D;
+typedef DomainWallFermion<WilsonCompressedComms8ImplF> DomainWallFermionCompressedComms8F;
+typedef DomainWallFermion<WilsonCompressedComms16ImplD> DomainWallFermionCompressedComms16D;
+typedef DomainWallFermion<WilsonCompressedComms16ImplF> DomainWallFermionCompressedComms16F;
+
+template<typename T>
+T parse(const std::string &name, std::istream &in){
+  std::string p;
+  in >> p;
+  assert(p==name);
+  char eq;
+  in >> eq;
+  assert(eq == '=');
+  T out;
+  in >> out;
+  return out;
+}
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int Ls=8;
+  RealD mass=0.1;
+  RealD outer_tol = 1e-8;
+  RealD inner_tol_full = 1e-5;
+  RealD inner_tol_half = 1e-5;
+  RealD inner_tol_16c = 1e-5;
+  RealD inner_tol_8c = 1e-5;
+
+  RealD relup_delta_full = 0.1;
+  RealD relup_delta_half = 0.1;
+  RealD relup_delta_16c = 0.1;
+  RealD relup_delta_8c = 0.1;
+
+  std::string config_file = "";
+
+  for(int i=1;i<argc;i++){
+    if(std::string(argv[i]) == "--params"){
+      std::ifstream f(argv[i+1]);
+      f.exceptions ( std::ifstream::failbit | std::ifstream::badbit );
+      Ls = parse<int>("Ls", f);
+#define PARSEIT(NM) NM = parse<RealD>(#NM, f)
+      PARSEIT(mass);
+      PARSEIT(outer_tol);
+      PARSEIT(inner_tol_full);
+      PARSEIT(inner_tol_half);
+      PARSEIT(inner_tol_16c);
+      PARSEIT(inner_tol_8c);
+      PARSEIT(relup_delta_full);
+      PARSEIT(relup_delta_half);
+      PARSEIT(relup_delta_16c);
+      PARSEIT(relup_delta_8c);
+#undef PARSEIT
+
+      //f >> outer_tol >> inner_tol_full >> inner_tol_half >> inner_tol_16c >> inner_tol_8c;      
+    }else if(std::string(argv[i]) == "--config"){
+      config_file = argv[i+1];
+    }
+  }
+  
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  GridCartesian         * UGrid_f   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid_f = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_f);
+  GridCartesian         * FGrid_f   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid_f);
+  GridRedBlackCartesian * FrbGrid_f = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid_f);
+  
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+
+  LatticeFermionD    src(FGrid); random(RNG5,src);
+  LatticeFermionD result(FGrid); result=zero;
+  LatticeGaugeFieldD Umu(UGrid);
+  LatticeGaugeFieldF Umu_f(UGrid_f); 
+
+  if(config_file.size() > 0){
+    FieldMetaData header;
+    NerscIO::readConfiguration(Umu,header,config_file);
+  }else{
+    SU3::HotConfiguration(RNG4,Umu);
+  }
+
+  precisionChange(Umu_f,Umu);
+  
+  RealD M5=1.8;
+
+  LatticeFermionD    src_o(FrbGrid);
+  pickCheckerboard(Odd,src_o,src);
+
+  if(0){ //Test preconditioned CG
+    LatticeFermionD result_o(FrbGrid);
+    LatticeFermionD result_o_2(FrbGrid);
+    result_o.checkerboard = Odd;
+    result_o = zero;
+    result_o_2.checkerboard = Odd;
+    result_o_2 = zero;
+
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+    //DoNothingLinearOperator<LatticeFermionD> Prec;
+    //FixedIterConjugateGradientPreconditioner<LatticeFermionD> Prec(HermOpEO, 20);
+    SloppyConjugateGradientPreconditioner<LatticeFermionD> Prec(HermOpEO, 1e-2, 1000);
+
+    std::cout << "Preconditioned CG" << std::endl;
+    InexactPreconditionedConjugateGradient<LatticeFermionD> pCG(Prec,1.0e-8,10000);
+    pCG(HermOpEO,src_o,result_o);
+
+    std::cout << "Starting regular CG" << std::endl;
+    ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+    CG(HermOpEO,src_o,result_o_2);
+
+    LatticeFermionD diff_o(FrbGrid);
+    RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
+
+    std::cout << "pCG HermOp applications " << pCG.IterationsToComplete << "(outer) + " << Prec.InnerIterations << "(inner) = " << pCG.IterationsToComplete + Prec.InnerIterations << std::endl;
+    std::cout << "CG HermOp applications " << CG.IterationsToComplete << std::endl;
+    std::cout << "Diff between results: " << diff << std::endl;
+  }
+
+  if(0){ //Test compressor
+    LatticeFermionD result_o(FrbGrid);
+    LatticeFermionD result_o_2(FrbGrid);
+    result_o.checkerboard = Odd;
+    result_o = zero;
+    result_o_2.checkerboard = Odd;
+    result_o_2 = zero;
+
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+
+    DomainWallFermionCompressedComms16D DdwfC(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionCompressedComms16D,LatticeFermionD> HermOpEOC(DdwfC);
+
+    std::cout << "Starting regular CG with compressed operator" << std::endl;
+    Integer iter1;
+    {
+      ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+      CG.ErrorOnNoConverge = false;
+      CG(HermOpEOC,src_o,result_o);
+      iter1 = CG.IterationsToComplete;
+    }
+    Integer iter2;
+    {
+      std::cout << "Starting regular CG" << std::endl;
+      ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+      CG(HermOpEO,src_o,result_o_2);
+      iter2 = CG.IterationsToComplete;
+    }
+
+    LatticeFermionD diff_o(FrbGrid);
+    RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
+
+    std::cout << "CG HermOp CC applications " << iter1 << std::endl;
+    std::cout << "CG HermOp applications " << iter2 << std::endl;
+    std::cout << "Diff between results: " << diff << std::endl;
+  }
+  
+  if(1){ //Compare mixed prec restarted single/single internal with same but with single/compressed
+    LatticeFermionD result_o_full(FrbGrid);
+    LatticeFermionD result_o_half(FrbGrid);
+    LatticeFermionD result_o_16(FrbGrid);
+    LatticeFermionD result_o_8(FrbGrid);
+    result_o_full.checkerboard = Odd;
+    result_o_full = zero;
+    result_o_16 = result_o_8 = result_o_half = result_o_full;
+
+    //Std
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+
+    DomainWallFermionF Ddwf_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionF,LatticeFermionF> HermOpEO_f(Ddwf_f);
+
+    //1/2 prec
+    DomainWallFermionFH Ddwfhalf_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionFH,LatticeFermionF> HermOpEOhalf_f(Ddwfhalf_f);
+
+    //16
+    DomainWallFermionCompressedComms16F DdwfC16_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionCompressedComms16F,LatticeFermionF> HermOpEOC16_f(DdwfC16_f);
+
+    //8
+    DomainWallFermionCompressedComms8F DdwfC8_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionCompressedComms8F,LatticeFermionF> HermOpEOC8_f(DdwfC8_f);
+
+    #define ALGORITHM_MIXEDCG
+    //#define ALGORITHM_RELUP
+    //#define ALGORITHM_SLOPPY_PREC_CG
+
+#ifdef ALGORITHM_MIXEDCG
+    std::cout << "Starting mixed CG with single/compressed-16 inner\n";
+
+    Integer inner_16, outer_16, patchup_16;
+    {
+      MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(outer_tol, 10000, 50, FrbGrid_f, HermOpEOC16_f, HermOpEO);
+      mCG.InnerTolerance = inner_tol_16c;
+      mCG(src_o,result_o_16);
+      inner_16 = mCG.TotalInnerIterations; outer_16 = mCG.TotalOuterIterations; patchup_16 = mCG.TotalFinalStepIterations;
+    }
+
+    std::cout << "Starting mixed CG with single/compressed-8 inner\n";
+    Integer inner_8, outer_8, patchup_8;
+    {
+      MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(outer_tol, 10000, 50, FrbGrid_f, HermOpEOC8_f, HermOpEO);
+      mCG.InnerTolerance = inner_tol_8c;
+      mCG(src_o,result_o_8);
+      inner_8 = mCG.TotalInnerIterations; outer_8 = mCG.TotalOuterIterations; patchup_8 = mCG.TotalFinalStepIterations;
+    }
+
+    std::cout << "Starting mixed CG with single/half inner\n";
+    Integer inner_half, outer_half, patchup_half;
+    {
+      MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(outer_tol, 10000, 50, FrbGrid_f, HermOpEOhalf_f, HermOpEO);
+      mCG.InnerTolerance = inner_tol_half;
+      mCG(src_o,result_o_half);
+      inner_half = mCG.TotalInnerIterations; outer_half = mCG.TotalOuterIterations; patchup_half = mCG.TotalFinalStepIterations;
+    }
+
+    std::cout << "Starting mixed CG with single/single inner\n";
+    Integer inner_full, outer_full, patchup_full;
+    {
+      MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(outer_tol, 10000, 50, FrbGrid_f, HermOpEO_f, HermOpEO);
+      mCG.InnerTolerance = inner_tol_full;
+      mCG(src_o,result_o_full);
+      inner_full = mCG.TotalInnerIterations; outer_full = mCG.TotalOuterIterations; patchup_full = mCG.TotalFinalStepIterations;
+    }
+#elif defined(ALGORITHM_RELUP)
+    std::cout << "Starting relup CG with single/compressed-16 inner\n";    
+    Integer inner_16, outer_16, patchup_16;
+    {
+      ConjugateGradientReliableUpdate<LatticeFermionD,LatticeFermionF> relup(outer_tol, 2000, relup_delta_16c, FrbGrid_f, HermOpEOC16_f, HermOpEO);
+      relup(src_o,result_o_16);
+      inner_16 = relup.IterationsToComplete; outer_16 = relup.ReliableUpdatesPerformed; patchup_16 = relup.IterationsToCleanup;
+    }
+
+    std::cout << "Starting relup CG with single/compressed-8 inner\n";
+    Integer inner_8, outer_8, patchup_8;
+    {
+      ConjugateGradientReliableUpdate<LatticeFermionD,LatticeFermionF> relup(outer_tol, 2000, relup_delta_8c, FrbGrid_f, HermOpEOC8_f, HermOpEO);
+      relup.ErrorOnNoConverge = false;
+      relup(src_o,result_o_8);
+      inner_8 = relup.IterationsToComplete; outer_8 = relup.ReliableUpdatesPerformed; patchup_8 = relup.IterationsToCleanup;     
+    }
+
+    std::cout << "Starting relup CG with single/half inner\n";
+    Integer inner_half, outer_half, patchup_half;
+    {
+      ConjugateGradientReliableUpdate<LatticeFermionD,LatticeFermionF> relup(outer_tol, 2000, relup_delta_half, FrbGrid_f, HermOpEOhalf_f, HermOpEO);
+      relup(src_o,result_o_half);
+      inner_half = relup.IterationsToComplete; outer_half = relup.ReliableUpdatesPerformed; patchup_half = relup.IterationsToCleanup;
+    }
+
+    std::cout << "Starting relup CG with single/single inner\n";
+    Integer inner_full, outer_full, patchup_full;
+    {
+      ConjugateGradientReliableUpdate<LatticeFermionD,LatticeFermionF> relup(outer_tol, 2000, relup_delta_full, FrbGrid_f, HermOpEO_f, HermOpEO);
+      relup(src_o,result_o_full);
+      inner_full = relup.IterationsToComplete; outer_full = relup.ReliableUpdatesPerformed; patchup_full = relup.IterationsToCleanup;
+    }
+#elif defined(ALGORITHM_SLOPPY_PREC_CG)
+
+    std::cout << "Starting sloppy pCG with single/compressed-16 inner\n";    
+    Integer inner_16, outer_16;
+    {
+      SloppyConjugateGradientLowerPrecPreconditioner<LatticeFermionD,LatticeFermionF> prec(HermOpEOC16_f, FrbGrid_f, inner_tol_16c, 1000);
+      InexactPreconditionedConjugateGradient<LatticeFermionD> CG(prec, outer_tol, 100);
+      CG(HermOpEO,src_o,result_o_16);
+      inner_16 = prec.InnerIterations; outer_16 = CG.IterationsToComplete;
+    }
+
+    std::cout << "Starting sloppy pCG with single/compressed-8 inner\n";    
+    Integer inner_8, outer_8;
+    {
+      SloppyConjugateGradientLowerPrecPreconditioner<LatticeFermionD,LatticeFermionF> prec(HermOpEOC8_f, FrbGrid_f, inner_tol_8c, 1000);
+      InexactPreconditionedConjugateGradient<LatticeFermionD> CG(prec, outer_tol, 100);
+      CG(HermOpEO,src_o,result_o_8);
+      inner_8 = prec.InnerIterations; outer_8 = CG.IterationsToComplete;
+    }
+
+    std::cout << "Starting sloppy pCG with single/half inner\n";    
+    Integer inner_half, outer_half;
+    {
+      SloppyConjugateGradientLowerPrecPreconditioner<LatticeFermionD,LatticeFermionF> prec(HermOpEOhalf_f, FrbGrid_f, inner_tol_half, 1000);
+      InexactPreconditionedConjugateGradient<LatticeFermionD> CG(prec, outer_tol, 100);
+      CG(HermOpEO,src_o,result_o_half);
+      inner_half = prec.InnerIterations; outer_half = CG.IterationsToComplete;
+    }
+
+    std::cout << "Starting sloppy pCG with single/single inner\n";    
+    Integer inner_full, outer_full;
+    {
+      SloppyConjugateGradientLowerPrecPreconditioner<LatticeFermionD,LatticeFermionF> prec(HermOpEO_f, FrbGrid_f, inner_tol_full, 1000);
+      InexactPreconditionedConjugateGradient<LatticeFermionD> CG(prec, outer_tol, 100);
+      CG(HermOpEO,src_o,result_o_full);
+      inner_full = prec.InnerIterations; outer_full = CG.IterationsToComplete;
+    }
+
+#endif
+
+    std::cout << "Ls " << Ls << std::endl;
+    std::cout << "Mass " << mass << std::endl;
+    std::cout << "Outer tolerance " << outer_tol << std::endl;
+
+#if defined(ALGORITHM_MIXEDCG) || defined(ALGORITHM_SLOPPY_PREC_CG)
+    std::cout << "Inner tol full " << inner_tol_full << std::endl;
+    std::cout << "Inner tol 1/2 prec " << inner_tol_half << std::endl;
+    std::cout << "Inner tol compressed-16 " << inner_tol_16c << std::endl;
+    std::cout << "Inner tol compressed-8 " << inner_tol_8c << std::endl;
+#elif defined(ALGORITHM_RELUP)
+    std::cout << "Relup delta full " << relup_delta_full << std::endl;
+    std::cout << "Relup delta 1/2 prec " << relup_delta_half << std::endl;
+    std::cout << "Relup delta compressed-16 " << relup_delta_16c << std::endl;
+    std::cout << "Relup delta compressed-8 " << relup_delta_8c << std::endl;
+#endif
+    LatticeFermionD diff_o(FrbGrid);
+    RealD diff = axpy_norm(diff_o, -1.0, result_o_16, result_o_full);
+    std::cout << "Diff between results (s/c16): " << diff << std::endl;
+
+    diff = axpy_norm(diff_o, -1.0, result_o_8, result_o_full);
+    std::cout << "Diff between results (s/c8): " << diff << std::endl;
+
+    diff = axpy_norm(diff_o, -1.0, result_o_half, result_o_full);
+    std::cout << "Diff between results (s/h): " << diff << std::endl;
+
+#if defined(ALGORITHM_MIXEDCG) || defined(ALGORITHM_RELUP)
+    std::cout << "Iterations (s/c16) inner: " << inner_16 << " outer: " << outer_16 << " patchup: " << patchup_16 << std::endl;
+    std::cout << "Iterations (s/c8) inner: " << inner_8 << " outer: " << outer_8 << " patchup: " << patchup_8 << std::endl;
+    std::cout << "Iterations (s/h) inner: " << inner_half << " outer: " << outer_half << " patchup: " << patchup_half << std::endl;
+    std::cout << "Iterations (s/s) inner: " << inner_full << " outer: " << outer_full << " patchup: " << patchup_full << std::endl;
+#else
+    std::cout << "Iterations (s/c16) inner: " << inner_16 << " outer: " << outer_16 << std::endl;
+    std::cout << "Iterations (s/c8) inner: " << inner_8 << " outer: " << outer_8 << std::endl;
+    std::cout << "Iterations (s/h) inner: " << inner_half << " outer: " << outer_half << std::endl;
+    std::cout << "Iterations (s/s) inner: " << inner_full << " outer: " << outer_full << std::endl;
+#endif
+
+  }
+
+
+  Grid_finalize();
+}

tests/solver/Test_pcg.cc

@@ -0,0 +1,247 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_pcg.cc
+
+    Copyright (C) 2015
+
+Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License along
+    with this program; if not, write to the Free Software Foundation, Inc.,
+    51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
+
+    See the full license in the file "LICENSE" in the top level distribution directory
+    *************************************************************************************/
+    /*  END LEGAL */
+#include<bitset>
+#include <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+//Preconditioning:   M psi = chi
+//               =  M P^-1 P psi = chi
+//               =  M P^-1 psi' = chi
+
+//Solve for psi' using M P^-1 as operator, then apply  P^-1 psi' = psi 
+
+//Inexact preconditioned CG requires slight modification because we want to avoid computing P^-1 exactly
+
+
+/////////////////////////////////////////////////////////////
+// Base classes for iterative processes based on operators
+// single input vec, single output vec.
+/////////////////////////////////////////////////////////////
+
+template<typename T>
+T parse(const std::string &name, std::istream &in){
+  std::string p;
+  in >> p;
+  assert(p==name);
+  char eq;
+  in >> eq;
+  assert(eq == '=');
+  T out;
+  in >> out;
+  return out;
+}
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  int Ls=12;
+  RealD mass=0.01;
+  RealD outer_tol = 1e-8;
+  RealD inner_tol_full = 1e-5;
+  RealD inner_tol_half = 1e-5;
+  RealD inner_tol_16c = 1e-5;
+  RealD inner_tol_8c = 1e-5;
+
+  RealD relup_delta_full = 0.1;
+  RealD relup_delta_half = 0.1;
+  RealD relup_delta_16c = 0.1;
+  RealD relup_delta_8c = 0.1;
+
+  std::string config_file = "";
+  RealD lo = 1.0;
+  RealD hi = 64.0;
+  int   order=10;
+  for(int i=1;i<argc;i++){
+    if(std::string(argv[i]) == "--params"){
+      std::ifstream f(argv[i+1]);
+      f.exceptions ( std::ifstream::failbit | std::ifstream::badbit );
+      Ls = parse<int>("Ls", f);
+#define PARSEIT(NM) NM = parse<RealD>(#NM, f)
+      PARSEIT(mass);
+      PARSEIT(outer_tol);
+      PARSEIT(inner_tol_full);
+      PARSEIT(inner_tol_half);
+      PARSEIT(inner_tol_16c);
+      PARSEIT(inner_tol_8c);
+      PARSEIT(relup_delta_full);
+      PARSEIT(relup_delta_half);
+      PARSEIT(relup_delta_16c);
+      PARSEIT(relup_delta_8c);
+#undef PARSEIT
+
+      //f >> outer_tol >> inner_tol_full >> inner_tol_half >> inner_tol_16c >> inner_tol_8c;      
+    }else if(std::string(argv[i]) == "--config"){
+      config_file = argv[i+1];
+    }else if(std::string(argv[i]) == "--order"){
+      std::string ss(argv[i+1]);
+      std::stringstream f(ss);
+      f>>order; std::cout << " Order poly set to " <<order<<std::endl;
+    }else if(std::string(argv[i]) == "--lo"){
+      std::string ss(argv[i+1]);
+      std::stringstream f(ss);
+      f>>lo; std::cout << " Lo poly set to " <<lo<<std::endl;
+    }else if(std::string(argv[i]) == "--hi"){
+      std::string ss(argv[i+1]);
+      std::stringstream f(ss);
+      f>>hi; std::cout << " Hi poly set to " <<hi<<std::endl;
+    }
+  }
+  
+  GridCartesian         * UGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
+  GridCartesian         * FGrid   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
+  GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
+
+  GridCartesian         * UGrid_f   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexF::Nsimd()),GridDefaultMpi());
+  GridRedBlackCartesian * UrbGrid_f = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_f);
+  GridCartesian         * FGrid_f   = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid_f);
+  GridRedBlackCartesian * FrbGrid_f = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid_f);
+  
+  std::vector<int> seeds4({1,2,3,4});
+  std::vector<int> seeds5({5,6,7,8});
+  GridParallelRNG          RNG5(FGrid);  RNG5.SeedFixedIntegers(seeds5);
+  GridParallelRNG          RNG4(UGrid);  RNG4.SeedFixedIntegers(seeds4);
+
+  LatticeFermionD    src(FGrid); random(RNG5,src);
+  LatticeFermionD result(FGrid); result=zero;
+  LatticeGaugeFieldD Umu(UGrid);
+  LatticeGaugeFieldF Umu_f(UGrid_f); 
+
+  if(config_file.size() > 0){
+    FieldMetaData header;
+    NerscIO::readConfiguration(Umu,header,config_file);
+  }else{
+    SU3::HotConfiguration(RNG4,Umu);
+  }
+
+  precisionChange(Umu_f,Umu);
+  
+  RealD M5=1.8;
+
+  LatticeFermionD    src_o(FrbGrid);
+  pickCheckerboard(Odd,src_o,src);
+
+  //if(0){ //Test preconditioned CG
+    std::cout << "Test preconditioned CG" << std::endl;
+    LatticeFermionD result_o(FrbGrid);
+    LatticeFermionD result_o_2(FrbGrid);
+    result_o.checkerboard = Odd;
+    result_o = zero;
+    result_o_2.checkerboard = Odd;
+    result_o_2 = zero;
+
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+    //DoNothingLinearOperator<LatticeFermionD> Prec;
+    //FixedIterConjugateGradientPreconditioner<LatticeFermionD> Prec(HermOpEO, 20);
+    //    SloppyConjugateGradientPreconditioner<LatticeFermionD> Prec(HermOpEO, 1e-2, 1000);
+    PolynomialPreconditioner<LatticeFermionD> Prec(HermOpEO,lo,hi,order) ;
+
+    std::cout << "Preconditioned CG" << std::endl;
+    InexactPreconditionedConjugateGradient<LatticeFermionD> pCG(Prec,1.0e-8,10000);
+    pCG(HermOpEO,src_o,result_o);
+
+    std::cout << "Starting regular CG" << std::endl;
+    ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+    CG(HermOpEO,src_o,result_o_2);
+
+    LatticeFermionD diff_o(FrbGrid);
+    RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
+
+    std::cout << "pCG HermOp applications " << " Lo " << lo << " Hi " << hi << " Order " << order << " " <<  pCG.IterationsToComplete << "(outer) + " << Prec.InnerIterations << "(inner) = " << pCG.IterationsToComplete + Prec.InnerIterations << std::endl;
+    std::cout << "CG HermOp applications " << CG.IterationsToComplete << std::endl;
+    std::cout << "Diff between results: " << diff << std::endl;
+  //}
+
+  if(0){ //Test compressor
+    LatticeFermionD result_o(FrbGrid);
+    LatticeFermionD result_o_2(FrbGrid);
+    result_o.checkerboard = Odd;
+    result_o = zero;
+    result_o_2.checkerboard = Odd;
+    result_o_2 = zero;
+
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+
+    DomainWallFermionDF DdwfC(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionDF,LatticeFermionD> HermOpEOC(DdwfC);
+
+    std::cout << "Starting regular CG with compressed operator" << std::endl;
+    Integer iter1;
+    {
+      ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+      CG.ErrorOnNoConverge = false;
+      CG(HermOpEOC,src_o,result_o);
+      iter1 = CG.IterationsToComplete;
+    }
+    Integer iter2;
+    {
+      std::cout << "Starting regular CG" << std::endl;
+      ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
+      CG(HermOpEO,src_o,result_o_2);
+      iter2 = CG.IterationsToComplete;
+    }
+
+    LatticeFermionD diff_o(FrbGrid);
+    RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
+
+    std::cout << "CG HermOp CC applications " << iter1 << std::endl;
+    std::cout << "CG HermOp applications " << iter2 << std::endl;
+    std::cout << "Diff between results: " << diff << std::endl;
+  }
+  
+  if(1){ //Compare mixed prec restarted single/single internal with same but with single/compressed
+    LatticeFermionD result_o_full(FrbGrid);
+    LatticeFermionD result_o_half(FrbGrid);
+    LatticeFermionD result_o_16(FrbGrid);
+    LatticeFermionD result_o_8(FrbGrid);
+    result_o_full.checkerboard = Odd;
+    result_o_full = zero;
+    result_o_16 = result_o_8 = result_o_half = result_o_full;
+
+    //Std
+    DomainWallFermionD Ddwf(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
+
+    DomainWallFermionF Ddwf_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionF,LatticeFermionF> HermOpEO_f(Ddwf_f);
+
+    //1/2 prec
+    DomainWallFermionFH Ddwfhalf_f(Umu_f,*FGrid_f,*FrbGrid_f,*UGrid_f,*UrbGrid_f,mass,M5);
+    SchurDiagMooeeOperator<DomainWallFermionFH,LatticeFermionF> HermOpEOhalf_f(Ddwfhalf_f);
+
+  }
+
+
+  Grid_finalize();
+}