Hadrons: faster A2A matrix load

This reverts commit 10fc263675.

Grid/algorithms/Algorithms.h

@@ -48,6 +48,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
 #include <Grid/algorithms/iterative/BlockConjugateGradient.h>
 #include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
+#include <Grid/algorithms/iterative/MinimalResidual.h>
+#include <Grid/algorithms/iterative/GeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h>
+#include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
 #include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
 #include <Grid/algorithms/CoarsenedMatrix.h>
 #include <Grid/algorithms/FFT.h>

Grid/algorithms/CoarsenedMatrix.h

@@ -211,6 +211,7 @@ namespace Grid {
 
       for(int b=0;b<nn;b++){
 	
+	subspace[b] = zero;
 	gaussian(RNG,noise);
 	scale = std::pow(norm2(noise),-0.5); 
 	noise=noise*scale;
@@ -295,13 +296,58 @@ namespace Grid {
       return norm2(out);
     };
 
-    RealD Mdag (const CoarseVector &in, CoarseVector &out){ 
-      return M(in,out);
+    RealD Mdag (const CoarseVector &in, CoarseVector &out){
+      // // corresponds to Petrov-Galerkin coarsening
+      // return M(in,out);
+
+      // corresponds to Galerkin coarsening
+      CoarseVector tmp(Grid());
+      G5C(tmp, in);
+      M(tmp, out);
+      G5C(out, out);
+      return norm2(out);
     };
 
-    // Defer support for further coarsening for now
-    void Mdiag    (const CoarseVector &in,  CoarseVector &out){};
-    void Mdir     (const CoarseVector &in,  CoarseVector &out,int dir, int disp){};
+    void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
+
+      conformable(_grid,in._grid);
+      conformable(in._grid,out._grid);
+
+      SimpleCompressor<siteVector> compressor;
+      Stencil.HaloExchange(in,compressor);
+
+      auto point = [dir, disp](){
+        if(dir == 0 and disp == 0)
+          return 8;
+        else
+          return (4 * dir + 1 - disp) / 2;
+      }();
+
+      parallel_for(int ss=0;ss<Grid()->oSites();ss++){
+        siteVector res = zero;
+        siteVector nbr;
+        int ptype;
+        StencilEntry *SE;
+
+        SE=Stencil.GetEntry(ptype,point,ss);
+
+        if(SE->_is_local&&SE->_permute) {
+          permute(nbr,in._odata[SE->_offset],ptype);
+        } else if(SE->_is_local) {
+          nbr = in._odata[SE->_offset];
+        } else {
+          nbr = Stencil.CommBuf()[SE->_offset];
+        }
+
+        res = res + A[point]._odata[ss]*nbr;
+
+        vstream(out._odata[ss],res);
+      }
+    };
+
+    void Mdiag(const CoarseVector &in, CoarseVector &out){
+      Mdir(in, out, 0, 0); // use the self coupling (= last) point of the stencil
+    };
 
     CoarsenedMatrix(GridCartesian &CoarseGrid) 	: 
 
@@ -417,7 +463,7 @@ namespace Grid {
       std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
 #endif
       //      ForceHermitian();
-      AssertHermitian();
+      // AssertHermitian();
       // ForceDiagonal();
     }
     void ForceDiagonal(void) {

Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,244 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  bool ErrorOnNoConverge; // Throw an assert when CAGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the CAGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> s;
+
+  CommunicationAvoidingGeneralisedMinimalResidual(RealD   tol,
+                                                  Integer maxit,
+                                                  Integer restart_length,
+                                                  bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular GMRES" << std::endl;
+
+    psi.checkerboard = src.checkerboard;
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src._grid);
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "CommunicationAvoidingGeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "CAGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "CommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src._grid);
+    Field r(src._grid);
+
+    // this should probably be made a class member so that it is only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, w, i);
+
+      qrUpdate(i);
+
+      cp = std::norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(v, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, Field &w, int iter) {
+
+    MatrixTimer.Start();
+    LinOp.Op(v[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = std::conj(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &v, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + v[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h

@@ -0,0 +1,256 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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_FLEXIBLE_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_FLEXIBLE_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  bool ErrorOnNoConverge; // Throw an assert when FCAGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the FCAGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch PrecTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> s;
+
+  LinearFunction<Field> &Preconditioner;
+
+  FlexibleCommunicationAvoidingGeneralisedMinimalResidual(RealD   tol,
+                                                          Integer maxit,
+                                                          LinearFunction<Field> &Prec,
+                                                          Integer restart_length,
+                                                          bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.)
+      , Preconditioner(Prec) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular FGMRES" << std::endl;
+
+    psi.checkerboard = src.checkerboard;
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src._grid);
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    PrecTimer.Reset();
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Precon  " <<         PrecTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FCAGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src._grid);
+    Field r(src._grid);
+
+    // these should probably be made class members so that they are only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
+    std::vector<Field> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, z, w, i);
+
+      qrUpdate(i);
+
+      cp = std::norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(z, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, std::vector<Field> &z, Field &w, int iter) {
+
+    PrecTimer.Start();
+    Preconditioner(v[iter], z[iter]);
+    PrecTimer.Stop();
+
+    MatrixTimer.Start();
+    LinOp.Op(z[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = std::conj(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &z, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + z[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h

@@ -0,0 +1,254 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  bool ErrorOnNoConverge; // Throw an assert when FGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the FGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch PrecTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> s;
+
+  LinearFunction<Field> &Preconditioner;
+
+  FlexibleGeneralisedMinimalResidual(RealD   tol,
+                                     Integer maxit,
+                                     LinearFunction<Field> &Prec,
+                                     Integer restart_length,
+                                     bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.)
+      , Preconditioner(Prec) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    psi.checkerboard = src.checkerboard;
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src._grid);
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    PrecTimer.Reset();
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "FlexibleGeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Precon  " <<         PrecTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "FGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "FlexibleGeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src._grid);
+    Field r(src._grid);
+
+    // these should probably be made class members so that they are only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
+    std::vector<Field> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, z, w, i);
+
+      qrUpdate(i);
+
+      cp = std::norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(z, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, std::vector<Field> &z, Field &w, int iter) {
+
+    PrecTimer.Start();
+    Preconditioner(v[iter], z[iter]);
+    PrecTimer.Stop();
+
+    MatrixTimer.Start();
+    LinOp.Op(z[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = std::conj(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &z, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + z[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/GeneralisedMinimalResidual.h

@@ -0,0 +1,242 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/GeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field>
+class GeneralisedMinimalResidual : public OperatorFunction<Field> {
+ public:
+  bool ErrorOnNoConverge; // Throw an assert when GMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the GMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> s;
+
+  GeneralisedMinimalResidual(RealD   tol,
+                             Integer maxit,
+                             Integer restart_length,
+                             bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.) {};
+
+  void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
+
+    psi.checkerboard = src.checkerboard;
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Field r(src._grid);
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "GeneralisedMinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "GeneralisedMinimalResidual:   src " << ssq   << std::endl;
+
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "GeneralisedMinimalResidual: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "GMRES Time elapsed: Total   " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: Matrix  " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: Linalg  " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: QR      " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "GMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "GeneralisedMinimalResidual did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    Field w(src._grid);
+    Field r(src._grid);
+
+    // this should probably be made a class member so that it is only allocated once, not in every restart
+    std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, w, i);
+
+      qrUpdate(i);
+
+      cp = std::norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "GeneralisedMinimalResidual: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(v, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, Field &w, int iter) {
+
+    MatrixTimer.Start();
+    LinOp.Op(v[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = std::conj(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<Field> const &v, Field &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + v[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/MinimalResidual.h

@@ -0,0 +1,156 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/MinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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_MINIMAL_RESIDUAL_H
+#define GRID_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class Field> class MinimalResidual : public OperatorFunction<Field> {
+ public:
+  bool ErrorOnNoConverge; // throw an assert when the MR fails to converge.
+                          // Defaults true.
+  RealD   Tolerance;
+  Integer MaxIterations;
+  RealD   overRelaxParam;
+  Integer IterationsToComplete; // Number of iterations the MR took to finish.
+                                // Filled in upon completion
+
+  MinimalResidual(RealD tol, Integer maxit, Real ovrelparam = 1.0, bool err_on_no_conv = true)
+    : Tolerance(tol), MaxIterations(maxit), overRelaxParam(ovrelparam), ErrorOnNoConverge(err_on_no_conv){};
+
+  void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
+
+    psi.checkerboard = src.checkerboard;
+    conformable(psi, src);
+
+    Complex a, c;
+    Real    d;
+
+    Field Mr(src);
+    Field r(src);
+
+    // Initial residual computation & set up
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    Linop.Op(psi, Mr);
+
+    r = src - Mr;
+
+    RealD cp = norm2(r);
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "MinimalResidual: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "MinimalResidual:   src " << ssq << std::endl;
+    std::cout << GridLogIterative << "MinimalResidual:    mp " << d << std::endl;
+    std::cout << GridLogIterative << "MinimalResidual:  cp,r " << cp << std::endl;
+
+    if (cp <= rsq) {
+      return;
+    }
+
+    std::cout << GridLogIterative << "MinimalResidual: k=0 residual " << cp << " target " << rsq << std::endl;
+
+    GridStopWatch LinalgTimer;
+    GridStopWatch MatrixTimer;
+    GridStopWatch SolverTimer;
+
+    SolverTimer.Start();
+    int k;
+    for (k = 1; k <= MaxIterations; k++) {
+
+      MatrixTimer.Start();
+      Linop.Op(r, Mr);
+      MatrixTimer.Stop();
+
+      LinalgTimer.Start();
+
+      c = innerProduct(Mr, r);
+
+      d = norm2(Mr);
+
+      a = c / d;
+
+      a = a * overRelaxParam;
+
+      psi = psi + r * a;
+
+      r = r - Mr * a;
+
+      cp = norm2(r);
+
+      LinalgTimer.Stop();
+
+      std::cout << GridLogIterative << "MinimalResidual: Iteration " << k
+                << " residual " << cp << " target " << rsq << std::endl;
+      std::cout << GridLogDebug << "a = " << a << " c = " << c << " d = " << d << std::endl;
+
+      // Stopping condition
+      if (cp <= rsq) {
+        SolverTimer.Stop();
+
+        Linop.Op(psi, Mr);
+        r = src - Mr;
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "MinimalResidual Converged on iteration " << k
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "MR Time elapsed: Total   " << SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MR Time elapsed: Matrix  " << MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MR Time elapsed: Linalg  " << LinalgTimer.Elapsed() << std::endl;
+
+        if (ErrorOnNoConverge)
+          assert(true_residual / Tolerance < 10000.0);
+
+        IterationsToComplete = k;
+
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "MinimalResidual did NOT converge"
+              << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+
+    IterationsToComplete = k;
+  }
+};
+} // namespace Grid
+#endif

Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h

@@ -0,0 +1,273 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./lib/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h
+
+Copyright (C) 2015
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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_MIXED_PRECISION_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+#define GRID_MIXED_PRECISION_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
+
+namespace Grid {
+
+template<class FieldD, class FieldF, typename std::enable_if<getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
+class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction<FieldD> {
+ public:
+  bool ErrorOnNoConverge; // Throw an assert when MPFGMRES fails to converge,
+                          // defaults to true
+
+  RealD   Tolerance;
+
+  Integer MaxIterations;
+  Integer RestartLength;
+  Integer MaxNumberOfRestarts;
+  Integer IterationCount; // Number of iterations the MPFGMRES took to finish,
+                          // filled in upon completion
+
+  GridStopWatch MatrixTimer;
+  GridStopWatch PrecTimer;
+  GridStopWatch LinalgTimer;
+  GridStopWatch QrTimer;
+  GridStopWatch CompSolutionTimer;
+  GridStopWatch ChangePrecTimer;
+
+  Eigen::MatrixXcd H;
+
+  std::vector<std::complex<double>> y;
+  std::vector<std::complex<double>> gamma;
+  std::vector<std::complex<double>> c;
+  std::vector<std::complex<double>> s;
+
+  GridBase* SinglePrecGrid;
+
+  LinearFunction<FieldF> &Preconditioner;
+
+  MixedPrecisionFlexibleGeneralisedMinimalResidual(RealD   tol,
+                                                   Integer maxit,
+                                                   GridBase * sp_grid,
+                                                   LinearFunction<FieldF> &Prec,
+                                                   Integer restart_length,
+                                                   bool    err_on_no_conv = true)
+      : Tolerance(tol)
+      , MaxIterations(maxit)
+      , RestartLength(restart_length)
+      , MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
+      , ErrorOnNoConverge(err_on_no_conv)
+      , H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
+      , y(RestartLength + 1, 0.)
+      , gamma(RestartLength + 1, 0.)
+      , c(RestartLength + 1, 0.)
+      , s(RestartLength + 1, 0.)
+      , SinglePrecGrid(sp_grid)
+      , Preconditioner(Prec) {};
+
+  void operator()(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi) {
+
+    psi.checkerboard = src.checkerboard;
+    conformable(psi, src);
+
+    RealD guess = norm2(psi);
+    assert(std::isnan(guess) == 0);
+
+    RealD cp;
+    RealD ssq = norm2(src);
+    RealD rsq = Tolerance * Tolerance * ssq;
+
+    FieldD r(src._grid);
+
+    std::cout << std::setprecision(4) << std::scientific;
+    std::cout << GridLogIterative << "MPFGMRES: guess " << guess << std::endl;
+    std::cout << GridLogIterative << "MPFGMRES:   src " << ssq   << std::endl;
+
+    PrecTimer.Reset();
+    MatrixTimer.Reset();
+    LinalgTimer.Reset();
+    QrTimer.Reset();
+    CompSolutionTimer.Reset();
+    ChangePrecTimer.Reset();
+
+    GridStopWatch SolverTimer;
+    SolverTimer.Start();
+
+    IterationCount = 0;
+
+    for (int k=0; k<MaxNumberOfRestarts; k++) {
+
+      cp = outerLoopBody(LinOp, src, psi, rsq);
+
+      // Stopping condition
+      if (cp <= rsq) {
+
+        SolverTimer.Stop();
+
+        LinOp.Op(psi,r);
+        axpy(r,-1.0,src,r);
+
+        RealD srcnorm       = sqrt(ssq);
+        RealD resnorm       = sqrt(norm2(r));
+        RealD true_residual = resnorm / srcnorm;
+
+        std::cout << GridLogMessage        << "MPFGMRES: Converged on iteration " << IterationCount
+                  << " computed residual " << sqrt(cp / ssq)
+                  << " true residual "     << true_residual
+                  << " target "            << Tolerance << std::endl;
+
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Total      " <<       SolverTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Precon     " <<         PrecTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Matrix     " <<       MatrixTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: Linalg     " <<       LinalgTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: QR         " <<           QrTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: CompSol    " << CompSolutionTimer.Elapsed() << std::endl;
+        std::cout << GridLogMessage << "MPFGMRES Time elapsed: PrecChange " <<   ChangePrecTimer.Elapsed() << std::endl;
+        return;
+      }
+    }
+
+    std::cout << GridLogMessage << "MPFGMRES did NOT converge" << std::endl;
+
+    if (ErrorOnNoConverge)
+      assert(0);
+  }
+
+  RealD outerLoopBody(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi, RealD rsq) {
+
+    RealD cp = 0;
+
+    FieldD w(src._grid);
+    FieldD r(src._grid);
+
+    // these should probably be made class members so that they are only allocated once, not in every restart
+    std::vector<FieldD> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
+    std::vector<FieldD> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
+
+    MatrixTimer.Start();
+    LinOp.Op(psi, w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    r = src - w;
+
+    gamma[0] = sqrt(norm2(r));
+
+    v[0] = (1. / gamma[0]) * r;
+    LinalgTimer.Stop();
+
+    for (int i=0; i<RestartLength; i++) {
+
+      IterationCount++;
+
+      arnoldiStep(LinOp, v, z, w, i);
+
+      qrUpdate(i);
+
+      cp = std::norm(gamma[i+1]);
+
+      std::cout << GridLogIterative << "MPFGMRES: Iteration " << IterationCount
+                << " residual " << cp << " target " << rsq << std::endl;
+
+      if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
+
+        computeSolution(z, psi, i);
+
+        return cp;
+      }
+    }
+
+    assert(0); // Never reached
+    return cp;
+  }
+
+  void arnoldiStep(LinearOperatorBase<FieldD> &LinOp, std::vector<FieldD> &v, std::vector<FieldD> &z, FieldD &w, int iter) {
+
+    FieldF v_f(SinglePrecGrid);
+    FieldF z_f(SinglePrecGrid);
+
+    ChangePrecTimer.Start();
+    precisionChange(v_f, v[iter]);
+    precisionChange(z_f, z[iter]);
+    ChangePrecTimer.Stop();
+
+    PrecTimer.Start();
+    Preconditioner(v_f, z_f);
+    PrecTimer.Stop();
+
+    ChangePrecTimer.Start();
+    precisionChange(z[iter], z_f);
+    ChangePrecTimer.Stop();
+
+    MatrixTimer.Start();
+    LinOp.Op(z[iter], w);
+    MatrixTimer.Stop();
+
+    LinalgTimer.Start();
+    for (int i = 0; i <= iter; ++i) {
+      H(iter, i) = innerProduct(v[i], w);
+      w = w - H(iter, i) * v[i];
+    }
+
+    H(iter, iter + 1) = sqrt(norm2(w));
+    v[iter + 1] = (1. / H(iter, iter + 1)) * w;
+    LinalgTimer.Stop();
+  }
+
+  void qrUpdate(int iter) {
+
+    QrTimer.Start();
+    for (int i = 0; i < iter ; ++i) {
+      auto tmp       = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
+      H(iter, i)     = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
+      H(iter, i + 1) = tmp;
+    }
+
+    // Compute new Givens Rotation
+    ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
+    c[iter]     = H(iter, iter) / nu;
+    s[iter]     = H(iter, iter + 1) / nu;
+
+    // Apply new Givens rotation
+    H(iter, iter)     = nu;
+    H(iter, iter + 1) = 0.;
+
+    gamma[iter + 1] = -s[iter] * gamma[iter];
+    gamma[iter]     = std::conj(c[iter]) * gamma[iter];
+    QrTimer.Stop();
+  }
+
+  void computeSolution(std::vector<FieldD> const &z, FieldD &psi, int iter) {
+
+    CompSolutionTimer.Start();
+    for (int i = iter; i >= 0; i--) {
+      y[i] = gamma[i];
+      for (int k = i + 1; k <= iter; k++)
+        y[i] = y[i] - H(k, i) * y[k];
+      y[i] = y[i] / H(i, i);
+    }
+
+    for (int i = 0; i <= iter; i++)
+      psi = psi + z[i] * y[i];
+    CompSolutionTimer.Stop();
+  }
+};
+}
+#endif

Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h

@@ -139,8 +139,11 @@ namespace Grid {
       MatTimer.Start();
       Linop.HermOpAndNorm(psi,Az,zAz,zAAz); 
       MatTimer.Stop();
+
+      LinalgTimer.Start();
       r=src-Az;
-      
+      LinalgTimer.Stop();
+
       /////////////////////
       // p = Prec(r)
       /////////////////////
@@ -152,8 +155,10 @@ namespace Grid {
       Linop.HermOp(z,tmp); 
       MatTimer.Stop();
 
+      LinalgTimer.Start();
       ttmp=tmp;
       tmp=tmp-r;
+      LinalgTimer.Stop();
 
       /*
       std::cout<<GridLogMessage<<r<<std::endl;
@@ -166,12 +171,14 @@ namespace Grid {
       Linop.HermOpAndNorm(z,Az,zAz,zAAz); 
       MatTimer.Stop();
 
+      LinalgTimer.Start();
       //p[0],q[0],qq[0] 
       p[0]= z;
       q[0]= Az;
       qq[0]= zAAz;
 
       cp =norm2(r);
+      LinalgTimer.Stop();
 
       for(int k=0;k<nstep;k++){
 
@@ -181,12 +188,14 @@ namespace Grid {
 	int peri_k = k %mmax;
 	int peri_kp= kp%mmax;
 
+        LinalgTimer.Start();
 	rq= real(innerProduct(r,q[peri_k])); // what if rAr not real?
 	a = rq/qq[peri_k];
 
 	axpy(psi,a,p[peri_k],psi);         
 
-	cp = axpy_norm(r,-a,q[peri_k],r);  
+	cp = axpy_norm(r,-a,q[peri_k],r);
+        LinalgTimer.Stop();
 
 	if((k==nstep-1)||(cp<rsq)){
 	  return cp;
@@ -202,6 +211,8 @@ namespace Grid {
 	Linop.HermOpAndNorm(z,Az,zAz,zAAz);
 	Linop.HermOp(z,tmp);
 	MatTimer.Stop();
+
+        LinalgTimer.Start();
         tmp=tmp-r;
 	std::cout<<GridLogMessage<< " Preconditioner resid " <<sqrt(norm2(tmp)/norm2(r))<<std::endl; 
 
@@ -219,9 +230,9 @@ namespace Grid {
 
 	}
 	qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
-
-
+        LinalgTimer.Stop();
       }
+
       assert(0); // never reached
       return cp;
     }

Grid/communicator/Communicator_mpi3.cc

@@ -50,8 +50,6 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
       assert(0);
   }
 
-  Grid_quiesce_nodes();
-
   // Never clean up as done once.
   MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
 
@@ -124,10 +122,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   // split the communicator
   //////////////////////////////////////////////////////////////////////////////////////////////////////
   //  int Nparent = parent._processors ; 
-  //  std::cout << " splitting from communicator "<<parent.communicator <<std::endl;
   int Nparent;
   MPI_Comm_size(parent.communicator,&Nparent);
-  //  std::cout << " Parent size  "<<Nparent <<std::endl;
 
   int childsize=1;
   for(int d=0;d<processors.size();d++) {
@@ -136,8 +132,6 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
   int Nchild = Nparent/childsize;
   assert (childsize * Nchild == Nparent);
 
-  //  std::cout << " child size  "<<childsize <<std::endl;
-
   std::vector<int> ccoor(_ndimension); // coor within subcommunicator
   std::vector<int> scoor(_ndimension); // coor of split within parent
   std::vector<int> ssize(_ndimension); // coor of split within parent

Grid/communicator/SharedMemoryMPI.cc

@@ -413,7 +413,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -455,7 +455,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
     assert(((uint64_t)ptr&0x3F)==0);
     close(fd);
     WorldShmCommBufs[r] =ptr;
-    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
+    //    std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
   }
   _ShmAlloc=1;
   _ShmAllocBytes  = bytes;
@@ -499,7 +499,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
       
-      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
+      //      std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
       if ( ptr == (void * )MAP_FAILED ) {       
 	perror("failed mmap");     
 	assert(0);    

Grid/lattice/Lattice_transfer.h

@@ -464,8 +464,10 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    if ( d!=orthog ) {
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    }
   }
 
   // the above should guarantee that the operations are local
@@ -485,7 +487,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
 
 
 template<class vobj>
-void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
+void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
 {
   typedef typename vobj::scalar_object sobj;
 
@@ -499,8 +501,10 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slic
   assert(orthog>=0);
 
   for(int d=0;d<nh;d++){
-    assert(lg->_processors[d]  == hg->_processors[d]);
-    assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    if ( d!=orthog ) {
+      assert(lg->_processors[d]  == hg->_processors[d]);
+      assert(lg->_ldimensions[d] == hg->_ldimensions[d]);
+    }
   }
 
   // the above should guarantee that the operations are local

Grid/log/Log.cc

@@ -59,6 +59,7 @@ void GridLogTimestamp(int on){
 }
 
 Colours GridLogColours(0);
+GridLogger GridLogMG     (1, "MG"    , GridLogColours, "NORMAL");
 GridLogger GridLogIRL    (1, "IRL"   , GridLogColours, "NORMAL");
 GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
 GridLogger GridLogError  (1, "Error" , GridLogColours, "RED");

Grid/log/Log.h

@@ -169,6 +169,7 @@ public:
 
 void GridLogConfigure(std::vector<std::string> &logstreams);
 
+extern GridLogger GridLogMG;
 extern GridLogger GridLogIRL;
 extern GridLogger GridLogSolver;
 extern GridLogger GridLogError;

Grid/parallelIO/BinaryIO.cc

@@ -0,0 +1,3 @@
+#include <Grid/GridCore.h>
+
+int Grid::BinaryIO::latticeWriteMaxRetry = -1;

Grid/parallelIO/BinaryIO.h

@@ -81,6 +81,7 @@ inline void removeWhitespace(std::string &key)
 ///////////////////////////////////////////////////////////////////////////////////////////////////
 class BinaryIO {
  public:
+  static int latticeWriteMaxRetry;
 
   /////////////////////////////////////////////////////////////////////////////
   // more byte manipulation helpers
@@ -370,7 +371,7 @@ PARALLEL_CRITICAL
 #endif
       } else {
 	std::cout << GridLogMessage <<"IOobject: C++ read I/O " << file << " : "
-                  << iodata.size() * sizeof(fobj) << " bytes" << std::endl;
+                  << iodata.size() * sizeof(fobj) << " bytes and offset " << offset << std::endl;
         std::ifstream fin;
 	fin.open(file, std::ios::binary | std::ios::in);
         if (control & BINARYIO_MASTER_APPEND)
@@ -582,7 +583,9 @@ PARALLEL_CRITICAL
     typedef typename vobj::scalar_object sobj;
     typedef typename vobj::Realified::scalar_type word;    word w=0;
     GridBase *grid = Umu._grid;
-    uint64_t lsites = grid->lSites();
+    uint64_t lsites = grid->lSites(), offsetCopy = offset;
+    int attemptsLeft = std::max(0, BinaryIO::latticeWriteMaxRetry);
+    bool checkWrite = (BinaryIO::latticeWriteMaxRetry >= 0);
 
     std::vector<sobj> scalardata(lsites); 
     std::vector<fobj>     iodata(lsites); // Munge, checksum, byte order in here
@@ -597,9 +600,35 @@ PARALLEL_CRITICAL
 
     grid->Barrier();
     timer.Stop();
+    while (attemptsLeft >= 0)
+    {
+      grid->Barrier();
+      IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
+	             nersc_csum,scidac_csuma,scidac_csumb);
+      if (checkWrite)
+      {
+        std::vector<fobj> ckiodata(lsites);
+        uint32_t          cknersc_csum, ckscidac_csuma, ckscidac_csumb;
+        uint64_t          ckoffset = offsetCopy;
 
-    IOobject(w,grid,iodata,file,offset,format,BINARYIO_WRITE|BINARYIO_LEXICOGRAPHIC,
-	     nersc_csum,scidac_csuma,scidac_csumb);
+        std::cout << GridLogMessage << "writeLatticeObject: read back object" << std::endl;
+        grid->Barrier();
+        IOobject(w,grid,ckiodata,file,ckoffset,format,BINARYIO_READ|BINARYIO_LEXICOGRAPHIC,
+	               cknersc_csum,ckscidac_csuma,ckscidac_csumb);
+        if ((cknersc_csum != nersc_csum) or (ckscidac_csuma != scidac_csuma) or (ckscidac_csumb != scidac_csumb))
+        {
+          std::cout << GridLogMessage << "writeLatticeObject: read test checksum failure, re-writing (" << attemptsLeft << " attempt(s) remaining)" << std::endl;
+          offset = offsetCopy;
+        }
+        else
+        {
+          std::cout << GridLogMessage << "writeLatticeObject: read test checksum correct" << std::endl;
+          break;
+        }
+      }
+      attemptsLeft--;
+    }
+    
 
     std::cout<<GridLogMessage<<"writeLatticeObject: unvectorize overhead "<<timer.Elapsed()  <<std::endl;
   }
@@ -725,5 +754,6 @@ PARALLEL_CRITICAL
     std::cout << GridLogMessage << "RNG state overhead " << timer.Elapsed() << std::endl;
   }
 };
+
 }
 #endif

Grid/perfmon/Timer.h

@@ -64,16 +64,20 @@ inline std::ostream& operator<< (std::ostream & stream, const GridMillisecs & no
 {
   GridSecs second(1);
   auto     secs       = now/second ; 
-  auto     subseconds = now%second ; 
+  auto     subseconds = now%second ;
+  auto     fill       = stream.fill();
   stream << secs<<"."<<std::setw(3)<<std::setfill('0')<<subseconds.count()<<" s";
+  stream.fill(fill);
   return stream;
 }
 inline std::ostream& operator<< (std::ostream & stream, const GridUsecs & now)
 {
   GridSecs second(1);
   auto     seconds    = now/second ; 
-  auto     subseconds = now%second ; 
+  auto     subseconds = now%second ;
+  auto     fill       = stream.fill();
   stream << seconds<<"."<<std::setw(6)<<std::setfill('0')<<subseconds.count()<<" s";
+  stream.fill(fill);
   return stream;
 }
 

Grid/qcd/action/ActionParams.h

@@ -44,12 +44,15 @@ namespace QCD {
   
   struct WilsonImplParams {
     bool overlapCommsCompute;
+    std::vector<Real> twist_n_2pi_L;
     std::vector<Complex> boundary_phases;
     WilsonImplParams() : overlapCommsCompute(false) {
       boundary_phases.resize(Nd, 1.0);
+      twist_n_2pi_L.resize(Nd, 0.0);
     };
-    WilsonImplParams(const std::vector<Complex> phi)
-      : boundary_phases(phi), overlapCommsCompute(false) {}
+    WilsonImplParams(const std::vector<Complex> phi) : boundary_phases(phi), overlapCommsCompute(false) {
+      twist_n_2pi_L.resize(Nd, 0.0);
+    }
   };
 
   struct StaggeredImplParams {

Grid/qcd/action/fermion/FermionOperatorImpl.h

@@ -240,16 +240,30 @@ namespace QCD {
       GaugeLinkField tmp(GaugeGrid);
 
       Lattice<iScalar<vInteger> > coor(GaugeGrid);
+      ////////////////////////////////////////////////////
+      // apply any boundary phase or twists
+      ////////////////////////////////////////////////////
       for (int mu = 0; mu < Nd; mu++) {
 
-	      auto pha = Params.boundary_phases[mu];
-	      scalar_type phase( real(pha),imag(pha) );
+	////////// boundary phase /////////////
+	auto pha = Params.boundary_phases[mu];
+	scalar_type phase( real(pha),imag(pha) );
 
-        int Lmu = GaugeGrid->GlobalDimensions()[mu] - 1;
+	int L   = GaugeGrid->GlobalDimensions()[mu];
+        int Lmu = L - 1;
 
         LatticeCoordinate(coor, mu);
 
         U = PeekIndex<LorentzIndex>(Umu, mu);
+
+	// apply any twists
+	RealD theta = Params.twist_n_2pi_L[mu] * 2*M_PI / L;
+	if ( theta != 0.0) { 
+	  scalar_type twphase(::cos(theta),::sin(theta));
+	  U = twphase*U;
+	  std::cout << GridLogMessage << " Twist ["<<mu<<"] "<< Params.twist_n_2pi_L[mu]<< " phase"<<phase <<std::endl;
+	}
+
         tmp = where(coor == Lmu, phase * U, U);
         PokeIndex<LorentzIndex>(Uds, tmp, mu);
 

Grid/qcd/action/gauge/Photon.h

@@ -4,9 +4,11 @@
  
  Source file: ./lib/qcd/action/gauge/Photon.h
  
- Copyright (C) 2015
+Copyright (C) 2015-2018
  
  Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+ Author: Antonin Portelli <antonin.portelli@me.com>
+ Author: James Harrison <J.Harrison@soton.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
@@ -30,8 +32,9 @@
 
 namespace Grid{
 namespace QCD{
+
   template <class S>
-  class QedGimpl
+  class QedGImpl
   {
   public:
     typedef S Simd;
@@ -43,27 +46,27 @@ namespace QCD{
     
     typedef iImplGaugeLink<Simd>  SiteLink;
     typedef iImplGaugeField<Simd> SiteField;
-    typedef SiteField             SiteComplex;
+    typedef SiteLink              SiteComplex;
     
     typedef Lattice<SiteLink>  LinkField;
     typedef Lattice<SiteField> Field;
     typedef Field              ComplexField;
   };
   
-  typedef QedGimpl<vComplex> QedGimplR;
+  typedef QedGImpl<vComplex> QedGImplR;
   
-  template<class Gimpl>
+  template <class GImpl>
   class Photon
   {
   public:
-    INHERIT_GIMPL_TYPES(Gimpl);
+    INHERIT_GIMPL_TYPES(GImpl);
+    typedef typename SiteGaugeLink::scalar_object ScalarSite;
+    typedef typename ScalarSite::scalar_type      ScalarComplex;
     GRID_SERIALIZABLE_ENUM(Gauge, undef, feynman, 1, coulomb, 2, landau, 3);
-    GRID_SERIALIZABLE_ENUM(ZmScheme, undef, qedL, 1, qedTL, 2, qedInf, 3);
+    GRID_SERIALIZABLE_ENUM(ZmScheme, undef, qedL, 1, qedTL, 2);
   public:
-    Photon(Gauge gauge, ZmScheme zmScheme);
-    Photon(Gauge gauge, ZmScheme zmScheme, std::vector<Real> improvements);
-    Photon(Gauge gauge, ZmScheme zmScheme, Real G0);
-    Photon(Gauge gauge, ZmScheme zmScheme, std::vector<Real> improvements, Real G0);
+    Photon(GridBase *grid, Gauge gauge, ZmScheme zmScheme, std::vector<Real> improvement);
+    Photon(GridBase *grid, Gauge gauge, ZmScheme zmScheme);
     virtual ~Photon(void) = default;
     void FreePropagator(const GaugeField &in, GaugeField &out);
     void MomentumSpacePropagator(const GaugeField &in, GaugeField &out);
@@ -73,345 +76,255 @@ namespace QCD{
                          const GaugeLinkField &weight);
     void UnitField(GaugeField &out);
   private:
-    void infVolPropagator(GaugeLinkField &out);
-    void invKHatSquared(GaugeLinkField &out);
+    void makeSpatialNorm(LatticeInteger &spNrm);
+    void makeKHat(std::vector<GaugeLinkField> &khat);
+    void makeInvKHatSquared(GaugeLinkField &out);
     void zmSub(GaugeLinkField &out);
+    void transverseProjectSpatial(GaugeField &out);
+    void gaugeTransform(GaugeField &out);
   private:
-    Gauge    gauge_;
-    ZmScheme zmScheme_;
-    std::vector<Real>  improvement_;
-    Real     G0_;
+    GridBase          *grid_;
+    Gauge             gauge_;
+    ZmScheme          zmScheme_;
+    std::vector<Real> improvement_;
   };
 
-  typedef Photon<QedGimplR>  PhotonR;
+  typedef Photon<QedGImplR>  PhotonR;
   
-  template<class Gimpl>
-  Photon<Gimpl>::Photon(Gauge gauge, ZmScheme zmScheme)
-  : gauge_(gauge), zmScheme_(zmScheme), improvement_(std::vector<Real>()),
-    G0_(0.15493339023106021408483720810737508876916113364521)
-  {}
-
-  template<class Gimpl>
-  Photon<Gimpl>::Photon(Gauge gauge, ZmScheme zmScheme,
+  template<class GImpl>
+  Photon<GImpl>::Photon(GridBase *grid, Gauge gauge, ZmScheme zmScheme,
                         std::vector<Real> improvements)
-  : gauge_(gauge), zmScheme_(zmScheme), improvement_(improvements),
-    G0_(0.15493339023106021408483720810737508876916113364521)
+  : grid_(grid), gauge_(gauge), zmScheme_(zmScheme), improvement_(improvements)
   {}
 
-  template<class Gimpl>
-  Photon<Gimpl>::Photon(Gauge gauge, ZmScheme zmScheme, Real G0)
-  : gauge_(gauge), zmScheme_(zmScheme), improvement_(std::vector<Real>()), G0_(G0)
+  template<class GImpl>
+  Photon<GImpl>::Photon(GridBase *grid, Gauge gauge, ZmScheme zmScheme)
+  : Photon(grid, gauge, zmScheme, std::vector<Real>())
   {}
 
-  template<class Gimpl>
-  Photon<Gimpl>::Photon(Gauge gauge, ZmScheme zmScheme,
-                        std::vector<Real> improvements, Real G0)
-  : gauge_(gauge), zmScheme_(zmScheme), improvement_(improvements), G0_(G0)
-  {}
-
-  template<class Gimpl>
-  void Photon<Gimpl>::FreePropagator (const GaugeField &in,GaugeField &out)
+  template<class GImpl>
+  void Photon<GImpl>::FreePropagator(const GaugeField &in, GaugeField &out)
   {
-    FFT theFFT(in._grid);
+    FFT        theFFT(dynamic_cast<GridCartesian *>(grid_));
+    GaugeField in_k(grid_);
+    GaugeField prop_k(grid_);
     
-    GaugeField in_k(in._grid);
-    GaugeField prop_k(in._grid);
-    
-    theFFT.FFT_all_dim(in_k,in,FFT::forward);
-    MomentumSpacePropagator(prop_k,in_k);
-    theFFT.FFT_all_dim(out,prop_k,FFT::backward);
+    theFFT.FFT_all_dim(in_k, in, FFT::forward);
+    MomentumSpacePropagator(prop_k, in_k);
+    theFFT.FFT_all_dim(out, prop_k, FFT::backward);
   }
 
-  template<class Gimpl>
-  void Photon<Gimpl>::infVolPropagator(GaugeLinkField &out)
+  template<class GImpl>
+  void Photon<GImpl>::makeSpatialNorm(LatticeInteger &spNrm)
   {
-    auto               *grid = dynamic_cast<GridCartesian *>(out._grid);
-    LatticeReal        xmu(grid);
-    GaugeLinkField     one(grid);
-    const unsigned int nd    = grid->_ndimension;
-    std::vector<int>   &l    = grid->_fdimensions;
-    std::vector<int>   x0(nd,0);
-    TComplex           Tone  = Complex(1.0,0.0);
-    TComplex           Tzero = Complex(G0_,0.0);
-    FFT                fft(grid);
+    LatticeInteger   coor(grid_);
+    std::vector<int> l = grid_->FullDimensions();
+
+    spNrm = zero;
+    for(int mu = 0; mu < grid_->Nd() - 1; mu++)
+    {
+      LatticeCoordinate(coor, mu);
+      coor  = where(coor < Integer(l[mu]/2), coor, coor - Integer(l[mu]));
+      spNrm = spNrm + coor*coor;
+    }
+  }
+
+  template<class GImpl>
+  void Photon<GImpl>::makeKHat(std::vector<GaugeLinkField> &khat)
+  {
+    const unsigned int nd = grid_->Nd();
+    std::vector<int>   l  = grid_->FullDimensions();
+    Complex            ci(0., 1.);
+
+    khat.resize(nd, grid_);
+    for (unsigned int mu = 0; mu < nd; ++mu)
+    {
+      Real piL = M_PI/l[mu];
+
+      LatticeCoordinate(khat[mu], mu);
+      khat[mu] = exp(piL*ci*khat[mu])*2.*sin(piL*khat[mu]);
+    }
+  }
+
+  template<class GImpl>
+  void Photon<GImpl>::makeInvKHatSquared(GaugeLinkField &out)
+  {
+    std::vector<GaugeLinkField> khat;
+    GaugeLinkField              lone(grid_);
+    const unsigned int          nd = grid_->Nd();
+    std::vector<int>            zm(nd, 0);
+    ScalarSite                  one = ScalarComplex(1., 0.), z = ScalarComplex(0., 0.);
     
-    one = Complex(1.0,0.0);
     out = zero;
+    makeKHat(khat);
     for(int mu = 0; mu < nd; mu++)
     {
-      LatticeCoordinate(xmu,mu);
-      Real lo2 = l[mu]/2.0;
-      xmu = where(xmu < lo2, xmu, xmu-double(l[mu]));
-      out = out + toComplex(4*M_PI*M_PI*xmu*xmu);
+      out = out + khat[mu]*conjugate(khat[mu]);
     }
-    pokeSite(Tone, out, x0);
-    out = one/out;
-    pokeSite(Tzero, out, x0);
-    fft.FFT_all_dim(out, out, FFT::forward);
+    lone = ScalarComplex(1., 0.);
+    pokeSite(one, out, zm);
+    out = lone/out;
+    pokeSite(z, out, zm);
   }
   
-  template<class Gimpl>
-  void Photon<Gimpl>::invKHatSquared(GaugeLinkField &out)
+  template<class GImpl>
+  void Photon<GImpl>::zmSub(GaugeLinkField &out)
   {
-    GridBase           *grid = out._grid;
-    GaugeLinkField     kmu(grid), one(grid);
-    const unsigned int nd    = grid->_ndimension;
-    std::vector<int>   &l    = grid->_fdimensions;
-    std::vector<int>   zm(nd,0);
-    TComplex           Tone = Complex(1.0,0.0);
-    TComplex           Tzero= Complex(0.0,0.0);
-    
-    one = Complex(1.0,0.0);
-    out = zero;
-    for(int mu = 0; mu < nd; mu++)
-    {
-      Real twoPiL = M_PI*2./l[mu];
-      
-      LatticeCoordinate(kmu,mu);
-      kmu = 2.*sin(.5*twoPiL*kmu);
-      out = out + kmu*kmu;
-    }
-    pokeSite(Tone, out, zm);
-    out = one/out;
-    pokeSite(Tzero, out, zm);
-  }
-  
-  template<class Gimpl>
-  void Photon<Gimpl>::zmSub(GaugeLinkField &out)
-  {
-    GridBase           *grid = out._grid;
-    const unsigned int nd    = grid->_ndimension;
-    std::vector<int>   &l    = grid->_fdimensions;
-    
     switch (zmScheme_)
     {
       case ZmScheme::qedTL:
       {
-        std::vector<int> zm(nd,0);
-        TComplex         Tzero = Complex(0.0,0.0);
-        
-        pokeSite(Tzero, out, zm);
+        std::vector<int> zm(grid_->Nd(), 0);
+        ScalarSite       z = ScalarComplex(0., 0.);
         
+        pokeSite(z, out, zm);
         break;
       }
       case ZmScheme::qedL:
       {
-        LatticeInteger spNrm(grid), coor(grid);
-        GaugeLinkField z(grid);
-        
-        spNrm = zero;
-        for(int d = 0; d < grid->_ndimension - 1; d++)
-        {
-          LatticeCoordinate(coor,d);
-          coor = where(coor < Integer(l[d]/2), coor, coor-Integer(l[d]));
-          spNrm = spNrm + coor*coor;
-        }
-        out = where(spNrm == Integer(0), 0.*out, out);
+        LatticeInteger spNrm(grid_);
 
-        // IR improvement
+        makeSpatialNorm(spNrm);
+        out = where(spNrm == Integer(0), 0.*out, out);
         for(int i = 0; i < improvement_.size(); i++)
         {
-          Real f = sqrt(improvement_[i]+1);
-          out = where(spNrm == Integer(i+1), f*out, out);
+          Real f = sqrt(improvement_[i] + 1);
+          out = where(spNrm == Integer(i + 1), f*out, out);
         }
+        break;
       }
       default:
+        assert(0);
         break;
     }
   }
 
-  template<class Gimpl>
-  void Photon<Gimpl>::MomentumSpacePropagator(const GaugeField &in,
-                                               GaugeField &out)
+  template<class GImpl>
+  void Photon<GImpl>::transverseProjectSpatial(GaugeField &out)
   {
-  GridBase           *grid = out._grid;
-    LatticeComplex     momProp(grid);
-    
-    switch (zmScheme_)
+    const unsigned int          nd = grid_->Nd();
+    GaugeLinkField              invKHat(grid_), cst(grid_), spdiv(grid_);
+    LatticeInteger              spNrm(grid_);
+    std::vector<GaugeLinkField> khat, a(nd, grid_), aProj(nd, grid_);
+
+    invKHat = zero;
+    makeSpatialNorm(spNrm);
+    makeKHat(khat);
+    for (unsigned int mu = 0; mu < nd; ++mu)
     {
-      case ZmScheme::qedTL:
-      case ZmScheme::qedL:
+      a[mu] = peekLorentz(out, mu);
+      if (mu < nd - 1)
       {
-        invKHatSquared(momProp);
-        zmSub(momProp);
-        break;
+        invKHat += khat[mu]*conjugate(khat[mu]);
       }
-      case ZmScheme::qedInf:
-      {
-        infVolPropagator(momProp);
+    }
+    cst     = ScalarComplex(1., 0.);
+    invKHat = where(spNrm == Integer(0), cst, invKHat);
+    invKHat = cst/invKHat;
+    cst     = zero;
+    invKHat = where(spNrm == Integer(0), cst, invKHat);
+    spdiv   = zero;
+    for (unsigned int nu = 0; nu < nd - 1; ++nu)
+    {
+      spdiv += conjugate(khat[nu])*a[nu];
+    }
+    spdiv *= invKHat;
+    for (unsigned int mu = 0; mu < nd; ++mu)
+    {
+      aProj[mu] = a[mu] - khat[mu]*spdiv;
+      pokeLorentz(out, aProj[mu], mu);
+    }
+  }
+
+  template<class GImpl>
+  void Photon<GImpl>::gaugeTransform(GaugeField &out)
+  {
+    switch (gauge_)
+    {
+      case Gauge::feynman:
+        break;
+      case Gauge::coulomb:
+        transverseProjectSpatial(out);
+        break;
+      case Gauge::landau:
+        assert(0);
         break;
-      }
       default:
+        assert(0);
         break;
     }
+  }
+
+  template<class GImpl>
+  void Photon<GImpl>::MomentumSpacePropagator(const GaugeField &in,
+                                              GaugeField &out)
+  {
+    LatticeComplex momProp(grid_);
+    
+    makeInvKHatSquared(momProp);
+    zmSub(momProp);
     
     out = in*momProp;
   }
   
-  template<class Gimpl>
-  void Photon<Gimpl>::StochasticWeight(GaugeLinkField &weight)
+  template<class GImpl>
+  void Photon<GImpl>::StochasticWeight(GaugeLinkField &weight)
   {
-    auto               *grid     = dynamic_cast<GridCartesian *>(weight._grid);
-    const unsigned int nd        = grid->_ndimension;
-    std::vector<int>   latt_size = grid->_fdimensions;
-    
-    switch (zmScheme_)
+    const unsigned int nd  = grid_->Nd();
+    std::vector<int>   l   = grid_->FullDimensions();
+    Integer            vol = 1;
+
+    for(unsigned int mu = 0; mu < nd; mu++)
     {
-      case ZmScheme::qedTL:
-      case ZmScheme::qedL:
-      {
-        Integer vol = 1;
-        for(int d = 0; d < nd; d++)
-        {
-          vol = vol * latt_size[d];
-        }
-        invKHatSquared(weight);
-        weight = sqrt(vol)*sqrt(weight);
-        zmSub(weight);
-        break;
-      }
-      case ZmScheme::qedInf:
-      {
-        infVolPropagator(weight);
-        weight = sqrt(real(weight));
-        break;
-      }
-      default:
-        break;
+      vol = vol*l[mu];
     }
+    makeInvKHatSquared(weight);
+    weight = sqrt(vol)*sqrt(weight);
+    zmSub(weight);
   }
   
-  template<class Gimpl>
-  void Photon<Gimpl>::StochasticField(GaugeField &out, GridParallelRNG &rng)
+  template<class GImpl>
+  void Photon<GImpl>::StochasticField(GaugeField &out, GridParallelRNG &rng)
   {
-    auto           *grid = dynamic_cast<GridCartesian *>(out._grid);
-    GaugeLinkField weight(grid);
+    GaugeLinkField weight(grid_);
     
     StochasticWeight(weight);
     StochasticField(out, rng, weight);
   }
   
-  template<class Gimpl>
-  void Photon<Gimpl>::StochasticField(GaugeField &out, GridParallelRNG &rng,
+  template<class GImpl>
+  void Photon<GImpl>::StochasticField(GaugeField &out, GridParallelRNG &rng,
                                       const GaugeLinkField &weight)
   {
-    auto               *grid = dynamic_cast<GridCartesian *>(out._grid);
-    const unsigned int nd = grid->_ndimension;
-    GaugeLinkField     r(grid);
-    GaugeField         aTilde(grid);
-    FFT                fft(grid);
+    const unsigned int nd = grid_->Nd();
+    GaugeLinkField     r(grid_);
+    GaugeField         aTilde(grid_);
+    FFT                fft(dynamic_cast<GridCartesian *>(grid_));
     
-    switch (zmScheme_)
+    for(unsigned int mu = 0; mu < nd; mu++)
     {
-      case ZmScheme::qedTL:
-      case ZmScheme::qedL:
-      {
-        for(int mu = 0; mu < nd; mu++)
-        {
-          gaussian(rng, r);
-          r = weight*r;
-          pokeLorentz(aTilde, r, mu);
-        }
-        break;
-      }
-      case ZmScheme::qedInf:
-      {
-        Complex                    shift(1., 1.); // This needs to be a GaugeLink element?
-        for(int mu = 0; mu < nd; mu++)
-        {
-          bernoulli(rng, r);
-          r = weight*(2.*r - shift);
-          pokeLorentz(aTilde, r, mu);
-        }
-        break;
-      }
-      default:
-        break;
+      gaussian(rng, r);
+      r = weight*r;
+      pokeLorentz(aTilde, r, mu);
     }
-
+    gaugeTransform(aTilde);
     fft.FFT_all_dim(out, aTilde, FFT::backward);
-    
     out = real(out);
   }
 
-  template<class Gimpl>
-  void Photon<Gimpl>::UnitField(GaugeField &out)
+  template<class GImpl>
+  void Photon<GImpl>::UnitField(GaugeField &out)
   {
-    auto               *grid = dynamic_cast<GridCartesian *>(out._grid);
-    const unsigned int nd = grid->_ndimension;
-    GaugeLinkField     r(grid);
+    const unsigned int nd = grid_->Nd();
+    GaugeLinkField     r(grid_);
     
-    r = Complex(1.0,0.0);
-
-    for(int mu = 0; mu < nd; mu++)
+    r = ScalarComplex(1., 0.);
+    for(unsigned int mu = 0; mu < nd; mu++)
     {
       pokeLorentz(out, r, mu);
     }
-    
     out = real(out);
   }
-//  template<class Gimpl>
-//  void Photon<Gimpl>::FeynmanGaugeMomentumSpacePropagator_L(GaugeField &out,
-//                                                            const GaugeField &in)
-//  {
-//    
-//    FeynmanGaugeMomentumSpacePropagator_TL(out,in);
-//    
-//    GridBase *grid = out._grid;
-//    LatticeInteger     coor(grid);
-//    GaugeField zz(grid); zz=zero;
-//    
-//    // xyzt
-//    for(int d = 0; d < grid->_ndimension-1;d++){
-//      LatticeCoordinate(coor,d);
-//      out = where(coor==Integer(0),zz,out);
-//    }
-//  }
-//  
-//  template<class Gimpl>
-//  void Photon<Gimpl>::FeynmanGaugeMomentumSpacePropagator_TL(GaugeField &out,
-//                                                             const GaugeField &in)
-//  {
-//    
-//    // what type LatticeComplex
-//    GridBase *grid = out._grid;
-//    int nd = grid->_ndimension;
-//    
-//    typedef typename GaugeField::vector_type vector_type;
-//    typedef typename GaugeField::scalar_type ScalComplex;
-//    typedef Lattice<iSinglet<vector_type> > LatComplex;
-//    
-//    std::vector<int> latt_size   = grid->_fdimensions;
-//    
-//    LatComplex denom(grid); denom= zero;
-//    LatComplex   one(grid); one = ScalComplex(1.0,0.0);
-//    LatComplex   kmu(grid);
-//    
-//    ScalComplex ci(0.0,1.0);
-//    // momphase = n * 2pi / L
-//    for(int mu=0;mu<Nd;mu++) {
-//      
-//      LatticeCoordinate(kmu,mu);
-//      
-//      RealD TwoPiL =  M_PI * 2.0/ latt_size[mu];
-//      
-//      kmu = TwoPiL * kmu ;
-//      
-//      denom = denom + 4.0*sin(kmu*0.5)*sin(kmu*0.5); // Wilson term
-//    }
-//    std::vector<int> zero_mode(nd,0);
-//    TComplexD Tone = ComplexD(1.0,0.0);
-//    TComplexD Tzero= ComplexD(0.0,0.0);
-//    
-//    pokeSite(Tone,denom,zero_mode);
-//    
-//    denom= one/denom;
-//    
-//    pokeSite(Tzero,denom,zero_mode);
-//    
-//    out = zero;
-//    out = in*denom;
-//  };
   
 }}
 #endif

Grid/qcd/utils/LinalgUtils.h

@@ -173,6 +173,39 @@ void G5R5(Lattice<vobj> &z,const Lattice<vobj> &x)
     }
   }
 }
+}
 
-}}
+// I explicitly need these outside the QCD namespace
+template<typename vobj>
+void G5C(Lattice<vobj> &z, const Lattice<vobj> &x)
+{
+  GridBase *grid = x._grid;
+  z.checkerboard = x.checkerboard;
+  conformable(x, z);
+
+  QCD::Gamma G5(QCD::Gamma::Algebra::Gamma5);
+  z = G5 * x;
+}
+
+template<class CComplex, int nbasis>
+void G5C(Lattice<iVector<CComplex, nbasis>> &z, const Lattice<iVector<CComplex, nbasis>> &x)
+{
+  GridBase *grid = x._grid;
+  z.checkerboard = x.checkerboard;
+  conformable(x, z);
+
+  static_assert(nbasis % 2 == 0, "");
+  int nb = nbasis / 2;
+
+  parallel_for(int ss = 0; ss < grid->oSites(); ss++) {
+    for(int n = 0; n < nb; ++n) {
+      z._odata[ss](n) = x._odata[ss](n);
+    }
+    for(int n = nb; n < nbasis; ++n) {
+      z._odata[ss](n) = -x._odata[ss](n);
+    }
+  }
+}
+
+}
 #endif 

Grid/qcd/utils/WilsonLoops.h

@@ -6,10 +6,12 @@
 
     Copyright (C) 2015
 
-Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
-Author: Peter Boyle <paboyle@ph.ed.ac.uk>
-Author: neo <cossu@post.kek.jp>
-Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
+    Author: Peter Boyle <paboyle@ph.ed.ac.uk>
+    Author: neo <cossu@post.kek.jp>
+    Author: paboyle <paboyle@ph.ed.ac.uk>
+    Author: James Harrison <J.Harrison@soton.ac.uk>
+    Author: Antonin Portelli <antonin.portelli@me.com>
 
     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
@@ -645,6 +647,184 @@ static void StapleMult(GaugeMat &staple, const GaugeLorentz &Umu, int mu) {
       }
     }
   }
+
+  //////////////////////////////////////////////////
+  // Wilson loop of size (R1, R2), oriented in mu,nu plane
+  //////////////////////////////////////////////////
+  static void wilsonLoop(GaugeMat &wl, const std::vector<GaugeMat> &U,
+                           const int Rmu, const int Rnu,
+                           const int mu, const int nu) {
+    wl = U[nu];
+
+    for(int i = 0; i < Rnu-1; i++){
+      wl = Gimpl::CovShiftForward(U[nu], nu, wl);
+    }
+
+    for(int i = 0; i < Rmu; i++){
+      wl = Gimpl::CovShiftForward(U[mu], mu, wl);
+    }
+
+    for(int i = 0; i < Rnu; i++){
+      wl = Gimpl::CovShiftBackward(U[nu], nu, wl);
+    }
+
+    for(int i = 0; i < Rmu; i++){
+      wl = Gimpl::CovShiftBackward(U[mu], mu, wl);
+    }
+  }
+  //////////////////////////////////////////////////
+  // trace of Wilson Loop oriented in mu,nu plane
+  //////////////////////////////////////////////////
+  static void traceWilsonLoop(LatticeComplex &wl,
+                                const std::vector<GaugeMat> &U,
+                                const int Rmu, const int Rnu,
+                                const int mu, const int nu) {
+    GaugeMat sp(U[0]._grid);
+    wilsonLoop(sp, U, Rmu, Rnu, mu, nu);
+    wl = trace(sp);
+  }
+  //////////////////////////////////////////////////
+  // sum over all planes of Wilson loop
+  //////////////////////////////////////////////////
+  static void siteWilsonLoop(LatticeComplex &Wl,
+                            const std::vector<GaugeMat> &U,
+                            const int R1, const int R2) {
+    LatticeComplex siteWl(U[0]._grid);
+    Wl = zero;
+    for (int mu = 1; mu < U[0]._grid->_ndimension; mu++) {
+      for (int nu = 0; nu < mu; nu++) {
+        traceWilsonLoop(siteWl, U, R1, R2, mu, nu);
+        Wl = Wl + siteWl;
+        traceWilsonLoop(siteWl, U, R2, R1, mu, nu);
+        Wl = Wl + siteWl;
+      }
+    }
+  }
+  //////////////////////////////////////////////////
+  // sum over planes of Wilson loop with length R1
+  // in the time direction
+  //////////////////////////////////////////////////
+  static void siteTimelikeWilsonLoop(LatticeComplex &Wl,
+                            const std::vector<GaugeMat> &U,
+                            const int R1, const int R2) {
+    LatticeComplex siteWl(U[0]._grid);
+
+    int ndim = U[0]._grid->_ndimension;
+
+    Wl = zero;
+    for (int nu = 0; nu < ndim - 1; nu++) {
+      traceWilsonLoop(siteWl, U, R1, R2, ndim-1, nu);
+      Wl = Wl + siteWl;
+    }
+  }
+  //////////////////////////////////////////////////
+  // sum Wilson loop over all planes orthogonal to the time direction
+  //////////////////////////////////////////////////
+  static void siteSpatialWilsonLoop(LatticeComplex &Wl,
+                            const std::vector<GaugeMat> &U,
+                            const int R1, const int R2) {
+    LatticeComplex siteWl(U[0]._grid);
+
+    Wl = zero;
+    for (int mu = 1; mu < U[0]._grid->_ndimension - 1; mu++) {
+      for (int nu = 0; nu < mu; nu++) {
+        traceWilsonLoop(siteWl, U, R1, R2, mu, nu);
+        Wl = Wl + siteWl;
+        traceWilsonLoop(siteWl, U, R2, R1, mu, nu);
+        Wl = Wl + siteWl;
+      }
+    }
+  }
+  //////////////////////////////////////////////////
+  // sum over all x,y,z,t and over all planes of Wilson loop
+  //////////////////////////////////////////////////
+  static Real sumWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    std::vector<GaugeMat> U(4, Umu._grid);
+
+    for (int mu = 0; mu < Umu._grid->_ndimension; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    LatticeComplex Wl(Umu._grid);
+
+    siteWilsonLoop(Wl, U, R1, R2);
+
+    TComplex Tp = sum(Wl);
+    Complex p = TensorRemove(Tp);
+    return p.real();
+  }
+  //////////////////////////////////////////////////
+  // sum over all x,y,z,t and over all planes of timelike Wilson loop
+  //////////////////////////////////////////////////
+  static Real sumTimelikeWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    std::vector<GaugeMat> U(4, Umu._grid);
+
+    for (int mu = 0; mu < Umu._grid->_ndimension; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    LatticeComplex Wl(Umu._grid);
+
+    siteTimelikeWilsonLoop(Wl, U, R1, R2);
+
+    TComplex Tp = sum(Wl);
+    Complex p = TensorRemove(Tp);
+    return p.real();
+  }
+  //////////////////////////////////////////////////
+  // sum over all x,y,z,t and over all planes of spatial Wilson loop
+  //////////////////////////////////////////////////
+  static Real sumSpatialWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    std::vector<GaugeMat> U(4, Umu._grid);
+
+    for (int mu = 0; mu < Umu._grid->_ndimension; mu++) {
+      U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
+    }
+
+    LatticeComplex Wl(Umu._grid);
+
+    siteSpatialWilsonLoop(Wl, U, R1, R2);
+
+    TComplex Tp = sum(Wl);
+    Complex p = TensorRemove(Tp);
+    return p.real();
+  }
+  //////////////////////////////////////////////////
+  // average over all x,y,z,t and over all planes of Wilson loop
+  //////////////////////////////////////////////////
+  static Real avgWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    int ndim = Umu._grid->_ndimension;
+    Real sumWl = sumWilsonLoop(Umu, R1, R2);
+    Real vol = Umu._grid->gSites();
+    Real faces = 1.0 * ndim * (ndim - 1);
+    return sumWl / vol / faces / Nc; // Nc dependent... FIXME
+  }
+  //////////////////////////////////////////////////
+  // average over all x,y,z,t and over all planes of timelike Wilson loop
+  //////////////////////////////////////////////////
+  static Real avgTimelikeWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    int ndim = Umu._grid->_ndimension;
+    Real sumWl = sumTimelikeWilsonLoop(Umu, R1, R2);
+    Real vol = Umu._grid->gSites();
+    Real faces = 1.0 * (ndim - 1);
+    return sumWl / vol / faces / Nc; // Nc dependent... FIXME
+  }
+  //////////////////////////////////////////////////
+  // average over all x,y,z,t and over all planes of spatial Wilson loop
+  //////////////////////////////////////////////////
+  static Real avgSpatialWilsonLoop(const GaugeLorentz &Umu,
+                            const int R1, const int R2) {
+    int ndim = Umu._grid->_ndimension;
+    Real sumWl = sumSpatialWilsonLoop(Umu, R1, R2);
+    Real vol = Umu._grid->gSites();
+    Real faces = 1.0 * (ndim - 1) * (ndim - 2);
+    return sumWl / vol / faces / Nc; // Nc dependent... FIXME
+  }
 };
 
 typedef WilsonLoops<PeriodicGimplR> ColourWilsonLoops;

Grid/serialisation/Hdf5IO.cc

@@ -61,9 +61,9 @@ Group & Hdf5Writer::getGroup(void)
 }
 
 // Reader implementation ///////////////////////////////////////////////////////
-Hdf5Reader::Hdf5Reader(const std::string &fileName)
+Hdf5Reader::Hdf5Reader(const std::string &fileName, const bool readOnly)
 : fileName_(fileName)
-, file_(fileName.c_str(), H5F_ACC_RDWR)
+, file_(fileName.c_str(), readOnly ? H5F_ACC_RDONLY : H5F_ACC_RDWR)
 {
   group_ = file_.openGroup("/");
   readSingleAttribute(dataSetThres_, HDF5_GRID_GUARD "dataset_threshold",

Grid/serialisation/Hdf5IO.h

@@ -54,7 +54,7 @@ namespace Grid
   class Hdf5Reader: public Reader<Hdf5Reader>
   {
   public:
-    Hdf5Reader(const std::string &fileName);
+    Hdf5Reader(const std::string &fileName, const bool readOnly = true);
     virtual ~Hdf5Reader(void) = default;
     bool push(const std::string &s);
     void pop(void);
@@ -123,9 +123,12 @@ namespace Grid
     
     if (flatx.size() > dataSetThres_)
     {
-      H5NS::DataSet dataSet;
+      H5NS::DataSet           dataSet;
+      H5NS::DSetCreatPropList plist;
       
-      dataSet = group_.createDataSet(s, Hdf5Type<Element>::type(), dataSpace);
+      plist.setChunk(dim.size(), dim.data());
+      plist.setFletcher32();
+      dataSet = group_.createDataSet(s, Hdf5Type<Element>::type(), dataSpace, plist);
       dataSet.write(flatx.data(), Hdf5Type<Element>::type());
     }
     else

Grid/threads/Threads.h

@@ -47,6 +47,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #else
 #define PARALLEL_FOR_LOOP
 #define PARALLEL_FOR_LOOP_INTERN
+#define PARALLEL_FOR_LOOP_REDUCE(op, var)
 #define PARALLEL_NESTED_LOOP2
 #define PARALLEL_NESTED_LOOP5
 #define PARALLEL_REGION
@@ -58,6 +59,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define parallel_for_internal PARALLEL_FOR_LOOP_INTERN for
 #define parallel_for_nest2 PARALLEL_NESTED_LOOP2 for
 #define parallel_for_nest5 PARALLEL_NESTED_LOOP5 for
+#define parallel_critical PARALLEL_CRITICAL
 
 namespace Grid {
 

Grid/util/Sha.h

@@ -28,16 +28,31 @@
 extern "C" {
 #include <openssl/sha.h>
 }
+#ifdef USE_IPP
+#include "ipp.h"
+#endif
 
 #pragma once
 
 class GridChecksum
 {
 public:
-  static inline uint32_t crc32(void *data,size_t bytes)
+  static inline uint32_t crc32(const void *data, size_t bytes)
   {
     return ::crc32(0L,(unsigned char *)data,bytes);
   }
+
+#ifdef USE_IPP
+  static inline uint32_t crc32c(const void* data, size_t bytes)
+  {
+      uint32_t crc32c = ~(uint32_t)0;
+      ippsCRC32C_8u(reinterpret_cast<const unsigned char *>(data), bytes, &crc32c);
+      ippsSwapBytes_32u_I(&crc32c, 1);
+  
+      return ~crc32c;
+  }
+#endif
+
   template <typename T>
   static inline std::string sha256_string(const std::vector<T> &hash)
   {

Hadrons/A2AMatrix.hpp

@@ -32,11 +32,19 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #include <Hadrons/Global.hpp>
 #include <Hadrons/TimerArray.hpp>
 #include <Grid/Eigen/unsupported/CXX11/Tensor>
+#ifdef USE_MKL
+#include "mkl.h"
+#include "mkl_cblas.h"
+#endif
 
 #ifndef HADRONS_A2AM_NAME 
 #define HADRONS_A2AM_NAME "a2aMatrix"
 #endif
 
+#ifndef HADRONS_A2AM_IO_TYPE
+#define HADRONS_A2AM_IO_TYPE ComplexF
+#endif
+
 #define HADRONS_A2AM_PARALLEL_IO
 
 BEGIN_HADRONS_NAMESPACE
@@ -51,6 +59,15 @@ BEGIN_HADRONS_NAMESPACE
 template <typename T>
 using A2AMatrixSet = Eigen::TensorMap<Eigen::Tensor<T, 5, Eigen::RowMajor>>;
 
+template <typename T>
+using A2AMatrix = Eigen::Matrix<T, -1, -1, Eigen::RowMajor>;
+
+template <typename T>
+using A2AMatrixMap = Eigen::Map<A2AMatrix<T>>;
+
+template <typename T>
+using A2AMatrixTr = Eigen::Matrix<T, -1, -1, Eigen::ColMajor>;
+
 /******************************************************************************
  *                      Abstract class for A2A kernels                        *
  ******************************************************************************/
@@ -76,10 +93,15 @@ public:
     // constructors
     A2AMatrixIo(void) = default;
     A2AMatrixIo(std::string filename, std::string dataname, 
-                const unsigned int nt, const unsigned int ni,
-                const unsigned int nj);
+                const unsigned int nt, const unsigned int ni = 0,
+                const unsigned int nj = 0);
     // destructor
     ~A2AMatrixIo(void) = default;
+    // access
+    unsigned int getNi(void) const;
+    unsigned int getNj(void) const;
+    unsigned int getNt(void) const;
+    size_t       getSize(void) const;
     // file allocation
     template <typename MetadataType>
     void initFile(const MetadataType &d, const unsigned int chunkSize);
@@ -88,9 +110,11 @@ public:
                    const unsigned int blockSizei, const unsigned int blockSizej);
     void saveBlock(const A2AMatrixSet<T> &m, const unsigned int ext, const unsigned int str,
                    const unsigned int i, const unsigned int j);
+    template <template <class> class Vec, typename VecT>
+    void load(Vec<VecT> &v, double *tRead = nullptr, const bool useCache = true);
 private:
-    std::string  filename_, dataname_;
-    unsigned int nt_, ni_, nj_;
+    std::string  filename_{""}, dataname_{""};
+    unsigned int nt_{0}, ni_{0}, nj_{0};
 };
 
 /******************************************************************************
@@ -136,6 +160,226 @@ private:
     std::vector<IoHelper> nodeIo_;
 };
 
+/******************************************************************************
+ *                       A2A matrix contraction kernels                       *
+ ******************************************************************************/
+class A2AContraction
+{
+public:
+    // accTrMul(acc, a, b): acc += tr(a*b)
+    template <typename C, typename MatLeft, typename MatRight>
+    static inline void accTrMul(C &acc, const MatLeft &a, const MatRight &b)
+    {
+        if ((MatLeft::Options == Eigen::RowMajor) and
+            (MatRight::Options == Eigen::ColMajor))
+        {
+            parallel_for (unsigned int r = 0; r < a.rows(); ++r)
+            {
+                C tmp;
+#ifdef USE_MKL
+                dotuRow(tmp, r, a, b);
+#else
+                tmp = a.row(r).conjugate().dot(b.col(r));
+#endif
+                parallel_critical
+                {
+                    acc += tmp;
+                }
+            }
+        }
+        else
+        {
+            parallel_for (unsigned int c = 0; c < a.cols(); ++c)
+            {
+                C tmp;
+#ifdef USE_MKL 
+                dotuCol(tmp, c, a, b);
+#else
+                tmp = a.col(c).conjugate().dot(b.row(c));
+#endif
+                parallel_critical
+                {
+                    acc += tmp;
+                }
+            }
+        }
+    }
+
+    template <typename MatLeft, typename MatRight>
+    static inline double accTrMulFlops(const MatLeft &a, const MatRight &b)
+    {
+        double n = a.rows()*a.cols();
+
+        return 8.*n;
+    }
+
+    // mul(res, a, b): res = a*b
+#ifdef USE_MKL
+    template <template <class, int...> class Mat, int... Opts>
+    static inline void mul(Mat<ComplexD, Opts...> &res, 
+                           const Mat<ComplexD, Opts...> &a, 
+                           const Mat<ComplexD, Opts...> &b)
+    {
+        static const ComplexD one(1., 0.), zero(0., 0.);
+
+        if ((res.rows() != a.rows()) or (res.cols() != b.cols()))
+        {
+            res.resize(a.rows(), b.cols());
+        }
+        if (Mat<ComplexD, Opts...>::Options == Eigen::RowMajor)
+        {
+            cblas_zgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, a.rows(), b.cols(),
+                        a.cols(), &one, a.data(), a.cols(), b.data(), b.cols(), &zero,
+                        res.data(), res.cols());
+        }
+        else if (Mat<ComplexD, Opts...>::Options == Eigen::ColMajor)
+        {
+            cblas_zgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, a.rows(), b.cols(),
+                        a.cols(), &one, a.data(), a.rows(), b.data(), b.rows(), &zero,
+                        res.data(), res.rows());
+        }
+    }
+
+    template <template <class, int...> class Mat, int... Opts>
+    static inline void mul(Mat<ComplexF, Opts...> &res, 
+                           const Mat<ComplexF, Opts...> &a, 
+                           const Mat<ComplexF, Opts...> &b)
+    {
+        static const ComplexF one(1., 0.), zero(0., 0.);
+
+        if ((res.rows() != a.rows()) or (res.cols() != b.cols()))
+        {
+            res.resize(a.rows(), b.cols());
+        }
+        if (Mat<ComplexF, Opts...>::Options == Eigen::RowMajor)
+        {
+            cblas_cgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, a.rows(), b.cols(),
+                        a.cols(), &one, a.data(), a.cols(), b.data(), b.cols(), &zero,
+                        res.data(), res.cols());
+        }
+        else if (Mat<ComplexF, Opts...>::Options == Eigen::ColMajor)
+        {
+            cblas_cgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, a.rows(), b.cols(),
+                        a.cols(), &one, a.data(), a.rows(), b.data(), b.rows(), &zero,
+                        res.data(), res.rows());
+        }
+    }
+#else
+    template <typename Mat>
+    static inline void mul(Mat &res, const Mat &a, const Mat &b)
+    {
+        res = a*b;
+    }
+#endif
+    template <typename Mat>
+    static inline double mulFlops(const Mat &a, const Mat &b)
+    {
+        double nr = a.rows(), nc = a.cols();
+
+        return nr*nr*(6.*nc + 2.*(nc - 1.));
+    }
+private:
+    template <typename C, typename MatLeft, typename MatRight>
+    static inline void makeDotRowPt(C * &aPt, unsigned int &aInc, C * &bPt, 
+                                    unsigned int &bInc, const unsigned int aRow, 
+                                    const MatLeft &a, const MatRight &b)
+    {
+        if (MatLeft::Options == Eigen::RowMajor)
+        {
+            aPt  = a.data() + aRow*a.cols();
+            aInc = 1;
+        }
+        else if (MatLeft::Options == Eigen::ColMajor)
+        {
+            aPt  = a.data() + aRow;
+            aInc = a.rows();
+        }
+        if (MatRight::Options == Eigen::RowMajor)
+        {
+            bPt  = b.data() + aRow;
+            bInc = b.cols();
+        }
+        else if (MatRight::Options == Eigen::ColMajor)
+        {
+            bPt  = b.data() + aRow*b.rows();
+            bInc = 1;
+        }
+    }
+
+#ifdef USE_MKL
+    template <typename C, typename MatLeft, typename MatRight>
+    static inline void makeDotColPt(C * &aPt, unsigned int &aInc, C * &bPt, 
+                                    unsigned int &bInc, const unsigned int aCol, 
+                                    const MatLeft &a, const MatRight &b)
+    {
+        if (MatLeft::Options == Eigen::RowMajor)
+        {
+            aPt  = a.data() + aCol;
+            aInc = a.cols();
+        }
+        else if (MatLeft::Options == Eigen::ColMajor)
+        {
+            aPt  = a.data() + aCol*a.rows();
+            aInc = 1;
+        }
+        if (MatRight::Options == Eigen::RowMajor)
+        {
+            bPt  = b.data() + aCol*b.cols();
+            bInc = 1;
+        }
+        else if (MatRight::Options == Eigen::ColMajor)
+        {
+            bPt  = b.data() + aCol;
+            bInc = b.rows();
+        }
+    }
+
+    template <typename MatLeft, typename MatRight>
+    static inline void dotuRow(ComplexF &res, const unsigned int aRow,
+                               const MatLeft &a, const MatRight &b)
+    {
+        const ComplexF *aPt, *bPt;
+        unsigned int   aInc, bInc;
+
+        makeDotRowPt(aPt, aInc, bPt, bInc, aRow, a, b);
+        cblas_cdotu_sub(a.cols(), aPt, aInc, bPt, bInc, &res);
+    }
+
+    template <typename MatLeft, typename MatRight>
+    static inline void dotuCol(ComplexF &res, const unsigned int aCol,
+                               const MatLeft &a, const MatRight &b)
+    {
+        const ComplexF *aPt, *bPt;
+        unsigned int   aInc, bInc;
+
+        makeDotColPt(aPt, aInc, bPt, bInc, aCol, a, b);
+        cblas_cdotu_sub(a.rows(), aPt, aInc, bPt, bInc, &res);
+    }
+
+    template <typename MatLeft, typename MatRight>
+    static inline void dotuRow(ComplexD &res, const unsigned int aRow,
+                               const MatLeft &a, const MatRight &b)
+    {
+        const ComplexD *aPt, *bPt;
+        unsigned int   aInc, bInc;
+
+        makeDotRowPt(aPt, aInc, bPt, bInc, aRow, a, b);
+        cblas_zdotu_sub(a.cols(), aPt, aInc, bPt, bInc, &res);
+    }
+
+    template <typename MatLeft, typename MatRight>
+    static inline void dotuCol(ComplexD &res, const unsigned int aCol,
+                               const MatLeft &a, const MatRight &b)
+    {
+        const ComplexD *aPt, *bPt;
+        unsigned int   aInc, bInc;
+
+        makeDotColPt(aPt, aInc, bPt, bInc, aCol, a, b);
+        cblas_zdotu_sub(a.rows(), aPt, aInc, bPt, bInc, &res);
+    }
+#endif
+};
+
 /******************************************************************************
  *                     A2AMatrixIo template implementation                    *
  ******************************************************************************/
@@ -148,6 +392,31 @@ A2AMatrixIo<T>::A2AMatrixIo(std::string filename, std::string dataname,
 , nt_(nt), ni_(ni), nj_(nj)
 {}
 
+// access //////////////////////////////////////////////////////////////////////
+template <typename T>
+unsigned int A2AMatrixIo<T>::getNt(void) const
+{
+    return nt_;
+}
+
+template <typename T>
+unsigned int A2AMatrixIo<T>::getNi(void) const
+{
+    return ni_;
+}
+
+template <typename T>
+unsigned int A2AMatrixIo<T>::getNj(void) const
+{
+    return nj_;
+}
+
+template <typename T>
+size_t A2AMatrixIo<T>::getSize(void) const
+{
+    return nt_*ni_*nj_*sizeof(T);
+}
+
 // file allocation /////////////////////////////////////////////////////////////
 template <typename T>
 template <typename MetadataType>
@@ -171,11 +440,12 @@ void A2AMatrixIo<T>::initFile(const MetadataType &d, const unsigned int chunkSiz
     }
 
     // create the dataset
-    Hdf5Reader reader(filename_);
+    Hdf5Reader reader(filename_, false);
 
     push(reader, dataname_);
     auto &group = reader.getGroup();
     plist.setChunk(chunk.size(), chunk.data());
+    plist.setFletcher32();
     dataset = group.createDataSet(HADRONS_A2AM_NAME, Hdf5Type<T>::type(), dataspace, plist);
 #else
     HADRONS_ERROR(Implementation, "all-to-all matrix I/O needs HDF5 library");
@@ -191,7 +461,7 @@ void A2AMatrixIo<T>::saveBlock(const T *data,
                                const unsigned int blockSizej)
 {
 #ifdef HAVE_HDF5
-    Hdf5Reader           reader(filename_);
+    Hdf5Reader           reader(filename_, false);
     std::vector<hsize_t> count = {nt_, blockSizei, blockSizej},
                          offset = {0, static_cast<hsize_t>(i),
                                    static_cast<hsize_t>(j)},
@@ -226,6 +496,99 @@ void A2AMatrixIo<T>::saveBlock(const A2AMatrixSet<T> &m,
     saveBlock(m.data() + offset, i, j, blockSizei, blockSizej);
 }
 
+template <typename T>
+template <template <class> class Vec, typename VecT>
+void A2AMatrixIo<T>::load(Vec<VecT> &v, double *tRead, const bool useCache)
+{
+#ifdef HAVE_HDF5
+    Hdf5Reader           reader(filename_);
+    std::vector<hsize_t> hdim;
+    H5NS::DataSet        dataset;
+    H5NS::DataSpace      dataspace;
+    H5NS::CompType       datatype;
+    
+    push(reader, dataname_);
+    auto &group = reader.getGroup();
+    dataset     = group.openDataSet(HADRONS_A2AM_NAME);
+    datatype    = dataset.getCompType();
+    dataspace   = dataset.getSpace();
+    hdim.resize(dataspace.getSimpleExtentNdims());
+    dataspace.getSimpleExtentDims(hdim.data());
+    if ((nt_*ni_*nj_ != 0) and
+        ((hdim[0] != nt_) or (hdim[1] != ni_) or (hdim[2] != nj_)))
+    {
+        HADRONS_ERROR(Size, "all-to-all matrix size mismatch (got "
+            + std::to_string(hdim[0]) + "x" + std::to_string(hdim[1]) + "x"
+            + std::to_string(hdim[2]) + ", expected "
+            + std::to_string(nt_) + "x" + std::to_string(ni_) + "x"
+            + std::to_string(nj_));
+    }
+    else if (ni_*nj_ == 0)
+    {
+        if (hdim[0] != nt_)
+        {
+            HADRONS_ERROR(Size, "all-to-all time size mismatch (got "
+                + std::to_string(hdim[0]) + ", expected "
+                + std::to_string(nt_) + ")");
+        }
+        ni_ = hdim[1];
+        nj_ = hdim[2];
+    }
+
+    if (useCache)
+    {
+        std::vector<T> buf(nt_*ni_*nj_);
+        T              *pt;
+
+        dataset.read(buf.data(), datatype);
+        pt = buf.data();
+        for (unsigned int t = 0; t < nt_; ++t)
+        {
+            A2AMatrixMap<T> bufMap(pt, ni_, nj_);
+
+            v[t]  = bufMap.template cast<VecT>();
+            pt   += ni_*nj_;
+        }
+    }
+    // if useCache = false, do I/O timeslice per timeslice (much slower)
+    else
+    {
+        A2AMatrix<T>         buf(ni_, nj_);
+        std::vector<hsize_t> count    = {1, static_cast<hsize_t>(ni_),
+                                        static_cast<hsize_t>(nj_)},
+                             stride   = {1, 1, 1},
+                             block    = {1, 1, 1},
+                             memCount = {static_cast<hsize_t>(ni_),
+                                         static_cast<hsize_t>(nj_)};
+        H5NS::DataSpace      memspace(memCount.size(), memCount.data());
+
+        std::cout << "Loading timeslice";
+        std::cout.flush();
+        *tRead = 0.;
+        for (unsigned int tp1 = nt_; tp1 > 0; --tp1)
+        {
+            unsigned int         t      = tp1 - 1;
+            std::vector<hsize_t> offset = {static_cast<hsize_t>(t), 0, 0};
+            
+            if (t % 10 == 0)
+            {
+                std::cout << " " << t;
+                std::cout.flush();
+            }
+            dataspace.selectHyperslab(H5S_SELECT_SET, count.data(), offset.data(),
+                                      stride.data(), block.data());
+            if (tRead) *tRead -= usecond();    
+            dataset.read(buf.data(), datatype, memspace, dataspace);
+            if (tRead) *tRead += usecond();
+            v[t] = buf.template cast<VecT>();
+        }
+        std::cout << std::endl;
+    }
+#else
+    HADRONS_ERROR(Implementation, "all-to-all matrix I/O needs HDF5 library");
+#endif
+}
+
 /******************************************************************************
  *               A2AMatrixBlockComputation template implementation            *
  ******************************************************************************/

Hadrons/Application.cc

@@ -108,6 +108,9 @@ void Application::run(void)
         HADRONS_ERROR(Definition, "run id is empty");
     }
     LOG(Message) << "RUN ID '" << getPar().runId << "'" << std::endl;
+    BinaryIO::latticeWriteMaxRetry = getPar().parallelWriteMaxRetry;
+    LOG(Message) << "Attempt(s) for resilient parallel I/O: " 
+                 << BinaryIO::latticeWriteMaxRetry << std::endl;
     vm().setRunId(getPar().runId);
     vm().printContent();
     env().printContent();

Hadrons/Application.hpp

@@ -41,14 +41,6 @@ BEGIN_HADRONS_NAMESPACE
 class Application
 {
 public:
-    class TrajRange: Serializable
-    {
-    public:
-        GRID_SERIALIZABLE_CLASS_MEMBERS(TrajRange,
-                                        unsigned int, start,
-                                        unsigned int, end,
-                                        unsigned int, step);
-    };
     class GlobalPar: Serializable
     {
     public:
@@ -56,7 +48,9 @@ public:
                                         TrajRange,                  trajCounter,
                                         VirtualMachine::GeneticPar, genetic,
                                         std::string,                runId,
-                                        std::string,                graphFile);
+                                        std::string,                graphFile,
+                                        int,                        parallelWriteMaxRetry);
+        GlobalPar(void): parallelWriteMaxRetry{-1} {}
     };
 public:
     // constructors

Hadrons/DiskVector.hpp

@@ -29,6 +29,7 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #define Hadrons_DiskVector_hpp_
 
 #include <Hadrons/Global.hpp>
+#include <Hadrons/A2AMatrix.hpp>
 #include <deque>
 #include <sys/stat.h>
 #include <ftw.h>
@@ -59,14 +60,18 @@ public:
         : master_(master), cmaster_(master), i_(i) {}
 
         // operator=: somebody is trying to store a vector element
-        // write to disk and cache
+        // write to cache and tag as modified
         T &operator=(const T &obj) const
         {
+            auto &cache    = *master_.cachePtr_;
+            auto &modified = *master_.modifiedPtr_;
+            auto &index    = *master_.indexPtr_;
+
             DV_DEBUG_MSG(&master_, "writing to " << i_);
             master_.cacheInsert(i_, obj);
-            master_.save(master_.filename(i_), obj);
+            modified[index.at(i_)] = true;
             
-            return master_.cachePtr_->at(i_);
+            return cache[index.at(i_)];
         }
 
         // implicit cast to const object reference and redirection
@@ -83,6 +88,7 @@ public:
 public:
     DiskVectorBase(const std::string dirname, const unsigned int size = 0,
                    const unsigned int cacheSize = 1, const bool clean = true);
+    DiskVectorBase(DiskVectorBase<T> &&v) = default;
     virtual ~DiskVectorBase(void);
     const T & operator[](const unsigned int i) const;
     RwAccessHelper operator[](const unsigned int i);
@@ -97,14 +103,17 @@ private:
     void cacheInsert(const unsigned int i, const T &obj) const;
     void clean(void);
 private:
-    std::string                                dirname_;
-    unsigned int                               size_, cacheSize_;
-    double                                     access_{0.}, hit_{0.};
-    bool                                       clean_;
+    std::string                                           dirname_;
+    unsigned int                                          size_, cacheSize_;
+    double                                                access_{0.}, hit_{0.};
+    bool                                                  clean_;
     // using pointers to allow modifications when class is const
     // semantic: const means data unmodified, but cache modification allowed
-    std::unique_ptr<std::map<unsigned int, T>> cachePtr_;
-    std::unique_ptr<std::deque<unsigned int>>  loadsPtr_;                
+    std::unique_ptr<std::vector<T>>                       cachePtr_;
+    std::unique_ptr<std::vector<bool>>                    modifiedPtr_;
+    std::unique_ptr<std::map<unsigned int, unsigned int>> indexPtr_;
+    std::unique_ptr<std::stack<unsigned int>>             freePtr_;
+    std::unique_ptr<std::deque<unsigned int>>             loadsPtr_;                
 };
 
 /******************************************************************************
@@ -135,7 +144,7 @@ private:
  *                      Specialisation for Eigen matrices                     *
  ******************************************************************************/
 template <typename T>
-using EigenDiskVectorMat = Eigen::Matrix<T, Eigen::Dynamic, Eigen::Dynamic>;
+using EigenDiskVectorMat = A2AMatrix<T>;
 
 template <typename T>
 class EigenDiskVector: public DiskVectorBase<EigenDiskVectorMat<T>>
@@ -152,24 +161,31 @@ public:
 private:
     virtual void load(EigenDiskVectorMat<T> &obj, const std::string filename) const
     {
-        std::ifstream              f(filename, std::ios::binary);
-        std::vector<unsigned char> hash(SHA256_DIGEST_LENGTH);
-        Eigen::Index               nRow, nCol;
-        size_t                     matSize;
-        double                     t;
+        std::ifstream f(filename, std::ios::binary);
+        uint32_t      crc, check;
+        Eigen::Index  nRow, nCol;
+        size_t        matSize;
+        double        tRead, tHash;
 
-        f.read(reinterpret_cast<char *>(hash.data()), hash.size()*sizeof(unsigned char));
-        f.read(reinterpret_cast<char *>(&nRow), sizeof(Eigen::Index));
-        f.read(reinterpret_cast<char *>(&nCol), sizeof(Eigen::Index));
+        f.read(reinterpret_cast<char *>(&crc), sizeof(crc));
+        f.read(reinterpret_cast<char *>(&nRow), sizeof(nRow));
+        f.read(reinterpret_cast<char *>(&nCol), sizeof(nCol));
         obj.resize(nRow, nCol);
         matSize = nRow*nCol*sizeof(T);
-        t  = -usecond();
+        tRead  = -usecond();
         f.read(reinterpret_cast<char *>(obj.data()), matSize);
-        t += usecond();
-        DV_DEBUG_MSG(this, "Eigen read " << matSize/t*1.0e6/1024/1024 << " MB/s");
-        auto check = GridChecksum::sha256(obj.data(), matSize);
-        DV_DEBUG_MSG(this, "Eigen sha256 " << GridChecksum::sha256_string(check));
-        if (hash != check)
+        tRead += usecond();
+        tHash  = -usecond();
+#ifdef USE_IPP
+        check  = GridChecksum::crc32c(obj.data(), matSize);
+#else
+        check  = GridChecksum::crc32(obj.data(), matSize);
+#endif
+        tHash += usecond();
+        DV_DEBUG_MSG(this, "Eigen read " << tRead/1.0e6 << " sec " << matSize/tRead*1.0e6/1024/1024 << " MB/s");
+        DV_DEBUG_MSG(this, "Eigen crc32 " << std::hex << check << std::dec 
+                     << " " << tHash/1.0e6 << " sec " << matSize/tHash*1.0e6/1024/1024 << " MB/s");
+        if (crc != check)
         {
             HADRONS_ERROR(Io, "checksum failed")
         }
@@ -177,24 +193,31 @@ private:
 
     virtual void save(const std::string filename, const EigenDiskVectorMat<T> &obj) const
     {
-        std::ofstream              f(filename, std::ios::binary);
-        std::vector<unsigned char> hash(SHA256_DIGEST_LENGTH);
-        Eigen::Index               nRow, nCol;
-        size_t                     matSize;
-        double                     t;
+        std::ofstream f(filename, std::ios::binary);
+        uint32_t      crc;
+        Eigen::Index  nRow, nCol;
+        size_t        matSize;
+        double        tWrite, tHash;
         
         nRow    = obj.rows();
         nCol    = obj.cols();
         matSize = nRow*nCol*sizeof(T);
-        hash    = GridChecksum::sha256(obj.data(), matSize);
-        DV_DEBUG_MSG(this, "Eigen sha256 " << GridChecksum::sha256_string(hash));
-        f.write(reinterpret_cast<char *>(hash.data()), hash.size()*sizeof(unsigned char));
-        f.write(reinterpret_cast<char *>(&nRow), sizeof(Eigen::Index));
-        f.write(reinterpret_cast<char *>(&nCol), sizeof(Eigen::Index));
-        t  = -usecond();
+        tHash   = -usecond();
+#ifdef USE_IPP
+        crc     = GridChecksum::crc32c(obj.data(), matSize);
+#else
+        crc     = GridChecksum::crc32(obj.data(), matSize);
+#endif
+        tHash  += usecond();
+        f.write(reinterpret_cast<char *>(&crc), sizeof(crc));
+        f.write(reinterpret_cast<char *>(&nRow), sizeof(nRow));
+        f.write(reinterpret_cast<char *>(&nCol), sizeof(nCol));
+        tWrite = -usecond();
         f.write(reinterpret_cast<const char *>(obj.data()), matSize);
-        t += usecond();
-        DV_DEBUG_MSG(this, "Eigen write " << matSize/t*1.0e6/1024/1024 << " MB/s");
+        tWrite += usecond();
+        DV_DEBUG_MSG(this, "Eigen write " << tWrite/1.0e6 << " sec " << matSize/tWrite*1.0e6/1024/1024 << " MB/s");
+        DV_DEBUG_MSG(this, "Eigen crc32 " << std::hex << crc << std::dec
+                     << " " << tHash/1.0e6 << " sec " << matSize/tHash*1.0e6/1024/1024 << " MB/s");
     }
 };
 
@@ -207,7 +230,10 @@ DiskVectorBase<T>::DiskVectorBase(const std::string dirname,
                                   const unsigned int cacheSize,
                                   const bool clean)
 : dirname_(dirname), size_(size), cacheSize_(cacheSize), clean_(clean)
-, cachePtr_(new std::map<unsigned int, T>())
+, cachePtr_(new std::vector<T>(size))
+, modifiedPtr_(new std::vector<bool>(size, false))
+, indexPtr_(new std::map<unsigned int, unsigned int>())
+, freePtr_(new std::stack<unsigned int>)
 , loadsPtr_(new std::deque<unsigned int>())
 {
     struct stat s;
@@ -217,6 +243,10 @@ DiskVectorBase<T>::DiskVectorBase(const std::string dirname,
         HADRONS_ERROR(Io, "directory '" + dirname + "' already exists")
     }
     mkdir(dirname);
+    for (unsigned int i = 0; i < cacheSize_; ++i)
+    {
+        freePtr_->push(i);
+    }
 }
 
 template <typename T>
@@ -231,8 +261,10 @@ DiskVectorBase<T>::~DiskVectorBase(void)
 template <typename T>
 const T & DiskVectorBase<T>::operator[](const unsigned int i) const
 {
-    auto &cache  = *cachePtr_;
-    auto &loads  = *loadsPtr_;
+    auto &cache   = *cachePtr_;
+    auto &index   = *indexPtr_;
+    auto &freeInd = *freePtr_;
+    auto &loads   = *loadsPtr_;
 
     DV_DEBUG_MSG(this, "accessing " << i << " (RO)");
 
@@ -241,7 +273,7 @@ const T & DiskVectorBase<T>::operator[](const unsigned int i) const
         HADRONS_ERROR(Size, "index out of range");
     }
     const_cast<double &>(access_)++;
-    if (cache.find(i) == cache.end())
+    if (index.find(i) == index.end())
     {
         // cache miss
         DV_DEBUG_MSG(this, "cache miss");
@@ -268,7 +300,7 @@ const T & DiskVectorBase<T>::operator[](const unsigned int i) const
     DV_DEBUG_MSG(this, "in cache: " << msg);
 #endif
 
-    return cache.at(i);
+    return cache[index.at(i)];
 }
 
 template <typename T>
@@ -306,13 +338,24 @@ std::string DiskVectorBase<T>::filename(const unsigned int i) const
 template <typename T>
 void DiskVectorBase<T>::evict(void) const
 {
-    auto &cache = *cachePtr_;
-    auto &loads = *loadsPtr_;
+    auto &cache    = *cachePtr_;
+    auto &modified = *modifiedPtr_;
+    auto &index    = *indexPtr_;
+    auto &freeInd  = *freePtr_;
+    auto &loads    = *loadsPtr_;
 
-    if (cache.size() >= cacheSize_)
+    if (index.size() >= cacheSize_)
     {
-        DV_DEBUG_MSG(this, "evicting " << loads.front());
-        cache.erase(loads.front());
+        unsigned int i = loads.front();
+        
+        DV_DEBUG_MSG(this, "evicting " << i);
+        if (modified[index.at(i)])
+        {
+            DV_DEBUG_MSG(this, "element " << i << " modified, saving to disk");
+            save(filename(i), cache[index.at(i)]);
+        }
+        freeInd.push(index.at(i));
+        index.erase(i);
         loads.pop_front();
     }
 }
@@ -320,30 +363,44 @@ void DiskVectorBase<T>::evict(void) const
 template <typename T>
 void DiskVectorBase<T>::fetch(const unsigned int i) const
 {
-    auto &cache = *cachePtr_;
-    auto &loads = *loadsPtr_;
+    auto &cache    = *cachePtr_;
+    auto &modified = *modifiedPtr_;
+    auto &index    = *indexPtr_;
+    auto &freeInd  = *freePtr_;
+    auto &loads    = *loadsPtr_;
+
     struct stat s;
 
     DV_DEBUG_MSG(this, "loading " << i << " from disk");
 
     evict();
+    
     if(stat(filename(i).c_str(), &s) != 0)
     {
         HADRONS_ERROR(Io, "disk vector element " + std::to_string(i) + " uninitialised");
     }
-    load(cache[i], filename(i));
+    index[i] = freeInd.top();
+    freeInd.pop();
+    load(cache[index.at(i)], filename(i));
     loads.push_back(i);
+    modified[index.at(i)] = false;
 }
 
 template <typename T>
 void DiskVectorBase<T>::cacheInsert(const unsigned int i, const T &obj) const
 {
-    auto &cache = *cachePtr_;
-    auto &loads = *loadsPtr_;
+    auto &cache    = *cachePtr_;
+    auto &modified = *modifiedPtr_;
+    auto &index    = *indexPtr_;
+    auto &freeInd  = *freePtr_;
+    auto &loads    = *loadsPtr_;
 
     evict();
-    cache[i] = obj;
+    index[i] = freeInd.top();
+    freeInd.pop();
+    cache[index.at(i)] = obj;
     loads.push_back(i);
+    modified[index.at(i)] = false;
 
 #ifdef DV_DEBUG
     std::string msg;

Hadrons/Global.cc

@@ -166,7 +166,13 @@ std::string Hadrons::dirname(const std::string &s)
 
 void Hadrons::makeFileDir(const std::string filename, GridBase *g)
 {
-    if (g->IsBoss())
+    bool doIt = true;
+
+    if (g)
+    {
+        doIt = g->IsBoss();
+    }
+    if (doIt)
     {
         std::string dir    = dirname(filename);
         int         status = mkdir(dir);

Hadrons/Global.hpp

@@ -32,6 +32,7 @@ See the full license in the file "LICENSE" in the top level distribution directo
 
 #include <set>
 #include <stack>
+#include <regex>
 #include <Grid/Grid.h>
 #include <cxxabi.h>
 
@@ -217,15 +218,15 @@ typedef XmlReader ResultReader;
 typedef XmlWriter ResultWriter;
 #endif
 
-#define RESULT_FILE_NAME(name) \
-name + "." + std::to_string(vm().getTrajectory()) + "." + resultFileExt
+#define RESULT_FILE_NAME(name, traj) \
+name + "." + std::to_string(traj) + "." + resultFileExt
 
 // recursive mkdir
 #define MAX_PATH_LENGTH 512u
 int         mkdir(const std::string dirName);
 std::string basename(const std::string &s);
 std::string dirname(const std::string &s);
-void        makeFileDir(const std::string filename, GridBase *g);
+void        makeFileDir(const std::string filename, GridBase *g = nullptr);
 
 // default Schur convention
 #ifndef HADRONS_DEFAULT_SCHUR 
@@ -248,6 +249,47 @@ void        makeFileDir(const std::string filename, GridBase *g);
 // pretty print time profile
 void printTimeProfile(const std::map<std::string, GridTime> &timing, GridTime total);
 
+// token replacement utility
+template <typename T>
+void tokenReplace(std::string &str, const std::string token,
+                  const T &x, const std::string mark = "@")
+{
+    std::string fullToken = mark + token + mark;
+    
+    auto pos = str.find(fullToken);
+    if (pos != std::string::npos)
+    {
+        str.replace(pos, fullToken.size(), std::to_string(x));
+    }
+}
+
+// trajectory range
+class TrajRange: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(TrajRange,
+                                    unsigned int, start,
+                                    unsigned int, end,
+                                    unsigned int, step,
+                                    std::string,  exclude);
+
+    inline std::vector<unsigned int> getTrajectoryList(void)
+    {
+        std::vector<unsigned int> excVec = strToVec<unsigned int>(exclude);
+        std::vector<unsigned int> list;
+
+        for (unsigned int t = start; t < end; t += step)
+        {
+            if (std::find(excVec.begin(), excVec.end(), t) == excVec.end())
+            {
+                list.push_back(t);
+            }
+        }
+
+        return list;
+    }
+};
+
 END_HADRONS_NAMESPACE
 
 #include <Hadrons/Exceptions.hpp>

Hadrons/Makefile.am

@@ -5,17 +5,17 @@ lib_LIBRARIES = libHadrons.a
 include modules.inc
 
 libHadrons_a_SOURCES = \
-    $(modules_cc)      \
     Application.cc     \
     Environment.cc     \
 	Exceptions.cc      \
     Global.cc          \
     Module.cc		   \
 	TimerArray.cc      \
-	VirtualMachine.cc
+	VirtualMachine.cc  \
+	$(modules_cc)
+	
 libHadrons_adir = $(includedir)/Hadrons
 nobase_libHadrons_a_HEADERS = \
-	$(modules_hpp)            \
 	A2AVectors.hpp            \
 	A2AMatrix.hpp             \
 	Application.hpp           \
@@ -33,4 +33,6 @@ nobase_libHadrons_a_HEADERS = \
 	ModuleFactory.hpp         \
 	Solver.hpp                \
 	TimerArray.hpp            \
-	VirtualMachine.hpp
+	VirtualMachine.hpp        \
+	Utilities/Contractor.hpp  \
+	$(modules_hpp)

Hadrons/Module.hpp

@@ -144,7 +144,7 @@ if (env().getGrid()->IsBoss() and !ioStem.empty())\
 {\
     makeFileDir(ioStem, env().getGrid());\
     {\
-        ResultWriter _writer(RESULT_FILE_NAME(ioStem));\
+        ResultWriter _writer(RESULT_FILE_NAME(ioStem, vm().getTrajectory()));\
         write(_writer, name, result);\
     }\
 }

Hadrons/Modules.hpp

@@ -24,9 +24,11 @@
 #include <Hadrons/Modules/MSolver/Guesser.hpp>
 #include <Hadrons/Modules/MSolver/RBPrecCG.hpp>
 #include <Hadrons/Modules/MSolver/A2AVectors.hpp>
+#include <Hadrons/Modules/MSolver/A2AAslashVectors.hpp>
 #include <Hadrons/Modules/MGauge/UnitEm.hpp>
 #include <Hadrons/Modules/MGauge/StoutSmearing.hpp>
 #include <Hadrons/Modules/MGauge/Unit.hpp>
+#include <Hadrons/Modules/MGauge/Electrify.hpp>
 #include <Hadrons/Modules/MGauge/Random.hpp>
 #include <Hadrons/Modules/MGauge/GaugeFix.hpp>
 #include <Hadrons/Modules/MGauge/FundtoHirep.hpp>

Hadrons/Modules/MAction/DWF.hpp

@@ -49,7 +49,8 @@ public:
                                     unsigned int, Ls,
                                     double      , mass,
                                     double      , M5,
-                                    std::string , boundary);
+                                    std::string , boundary,
+                                    std::string , twist);
 };
 
 template <typename FImpl>
@@ -119,8 +120,9 @@ void TDWF<FImpl>::setup(void)
     auto &grb4 = *envGetRbGrid(FermionField);
     auto &g5   = *envGetGrid(FermionField, par().Ls);
     auto &grb5 = *envGetRbGrid(FermionField, par().Ls);
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename DomainWallFermion<FImpl>::ImplParams implParams(boundary);
+    typename DomainWallFermion<FImpl>::ImplParams implParams;
+    implParams.boundary_phases = strToVec<Complex>(par().boundary);
+    implParams.twist_n_2pi_L   = strToVec<Real>(par().twist);
     envCreateDerived(FMat, DomainWallFermion<FImpl>, getName(), par().Ls, U, g5,
                      grb5, g4, grb4, par().mass, par().M5, implParams);
 }

Hadrons/Modules/MAction/MobiusDWF.hpp

@@ -49,7 +49,8 @@ public:
                                     double      , M5,
                                     double      , b,
                                     double      , c,
-                                    std::string , boundary);
+                                    std::string , boundary,
+                                    std::string , twist);
 };
 
 template <typename FImpl>
@@ -119,8 +120,9 @@ void TMobiusDWF<FImpl>::setup(void)
     auto &grb4 = *envGetRbGrid(FermionField);
     auto &g5   = *envGetGrid(FermionField, par().Ls);
     auto &grb5 = *envGetRbGrid(FermionField, par().Ls);
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename MobiusFermion<FImpl>::ImplParams implParams(boundary);
+    typename MobiusFermion<FImpl>::ImplParams implParams;
+    implParams.boundary_phases = strToVec<Complex>(par().boundary);
+    implParams.twist_n_2pi_L   = strToVec<Real>(par().twist);
     envCreateDerived(FMat, MobiusFermion<FImpl>, getName(), par().Ls, U, g5,
                      grb5, g4, grb4, par().mass, par().M5, par().b, par().c,
                      implParams);

Hadrons/Modules/MAction/ScaledDWF.hpp

@@ -48,7 +48,8 @@ public:
                                     double      , mass,
                                     double      , M5,
                                     double      , scale,
-                                    std::string , boundary);
+                                    std::string , boundary,
+                                    std::string , twist);
 };
 
 template <typename FImpl>
@@ -118,8 +119,9 @@ void TScaledDWF<FImpl>::setup(void)
     auto &grb4 = *envGetRbGrid(FermionField);
     auto &g5   = *envGetGrid(FermionField, par().Ls);
     auto &grb5 = *envGetRbGrid(FermionField, par().Ls);
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename MobiusFermion<FImpl>::ImplParams implParams(boundary);
+    typename ScaledShamirFermion<FImpl>::ImplParams implParams;
+    implParams.boundary_phases = strToVec<Complex>(par().boundary);
+    implParams.twist_n_2pi_L   = strToVec<Real>(par().twist);
     envCreateDerived(FMat, ScaledShamirFermion<FImpl>, getName(), par().Ls, U, g5,
                      grb5, g4, grb4, par().mass, par().M5, par().scale,
                      implParams);

Hadrons/Modules/MAction/Wilson.hpp

@@ -47,7 +47,9 @@ public:
     GRID_SERIALIZABLE_CLASS_MEMBERS(WilsonPar,
                                     std::string, gauge,
                                     double     , mass,
-                                    std::string, boundary);
+                                    std::string, boundary,
+                                    std::string, string,
+                                    std::string, twist);
 };
 
 template <typename FImpl>
@@ -113,8 +115,9 @@ void TWilson<FImpl>::setup(void)
     auto &U      = envGet(GaugeField, par().gauge);
     auto &grid   = *envGetGrid(FermionField);
     auto &gridRb = *envGetRbGrid(FermionField);
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename WilsonFermion<FImpl>::ImplParams implParams(boundary);
+    typename WilsonFermion<FImpl>::ImplParams implParams;
+    implParams.boundary_phases = strToVec<Complex>(par().boundary);
+    implParams.twist_n_2pi_L   = strToVec<Real>(par().twist);
     envCreateDerived(FMat, WilsonFermion<FImpl>, getName(), 1, U, grid, gridRb,
                      par().mass, implParams);
 }

Hadrons/Modules/MAction/WilsonClover.hpp

@@ -51,7 +51,8 @@ public:
 				                    double     , csw_r,
 				                    double     , csw_t,
 				                    WilsonAnisotropyCoefficients ,clover_anisotropy,
-                                    std::string, boundary
+                                    std::string, boundary,
+                                    std::string, twist
 				    );
 };
 
@@ -119,8 +120,9 @@ void TWilsonClover<FImpl>::setup(void)
     auto &U      = envGet(GaugeField, par().gauge);
     auto &grid   = *envGetGrid(FermionField);
     auto &gridRb = *envGetRbGrid(FermionField);
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename WilsonCloverFermion<FImpl>::ImplParams implParams(boundary);
+    typename WilsonCloverFermion<FImpl>::ImplParams implParams;
+    implParams.boundary_phases = strToVec<Complex>(par().boundary);
+    implParams.twist_n_2pi_L   = strToVec<Real>(par().twist);
     envCreateDerived(FMat, WilsonCloverFermion<FImpl>, getName(), 1, U, grid,
                      gridRb, par().mass, par().csw_r, par().csw_t, 
                      par().clover_anisotropy, implParams); 

Hadrons/Modules/MAction/ZMobiusDWF.hpp

@@ -50,7 +50,8 @@ public:
                                     double                           , b,
                                     double                           , c,
                                     std::vector<std::complex<double>>, omega,
-                                    std::string                      , boundary);
+                                    std::string                      , boundary,
+                                    std::string                      , twist);
 };
 
 template <typename FImpl>
@@ -127,8 +128,9 @@ void TZMobiusDWF<FImpl>::setup(void)
     auto &g5   = *envGetGrid(FermionField, par().Ls);
     auto &grb5 = *envGetRbGrid(FermionField, par().Ls);
     auto omega = par().omega;
-    std::vector<Complex> boundary = strToVec<Complex>(par().boundary);
-    typename ZMobiusFermion<FImpl>::ImplParams implParams(boundary);
+    typename ZMobiusFermion<FImpl>::ImplParams implParams;
+    implParams.boundary_phases = strToVec<Complex>(par().boundary);
+    implParams.twist_n_2pi_L   = strToVec<Real>(par().twist);
     envCreateDerived(FMat, ZMobiusFermion<FImpl>, getName(), par().Ls, U, g5,
                      grb5, g4, grb4, par().mass, par().M5, omega,
                      par().b, par().c, implParams);

Hadrons/Modules/MContraction/A2AAslashField.hpp

@@ -33,10 +33,6 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #include <Hadrons/ModuleFactory.hpp>
 #include <Hadrons/A2AMatrix.hpp>
 
-#ifndef ASF_IO_TYPE
-#define ASF_IO_TYPE ComplexF
-#endif
-
 BEGIN_HADRONS_NAMESPACE
 
 /******************************************************************************
@@ -113,7 +109,7 @@ public:
     typedef A2AMatrixBlockComputation<Complex, 
                                       FermionField, 
                                       A2AAslashFieldMetadata, 
-                                      ASF_IO_TYPE> Computation;
+                                      HADRONS_A2AM_IO_TYPE> Computation;
     typedef AslashFieldKernel<Complex, FImpl> Kernel;
 public:
     // constructor
@@ -196,7 +192,7 @@ void TA2AAslashField<FImpl, PhotonImpl>::execute(void)
         LOG(Message) << "  " << name << std::endl;
     }
     LOG(Message) << "A-slash field size: " << nt << "*" << N_i << "*" << N_j 
-                 << " (filesize " << sizeString(nt*N_i*N_j*sizeof(ASF_IO_TYPE)) 
+                 << " (filesize " << sizeString(nt*N_i*N_j*sizeof(HADRONS_A2AM_IO_TYPE)) 
                  << "/EM field)" << std::endl;
     
     // preparing "B" complexified fields

Hadrons/Modules/MContraction/A2AMesonField.hpp

@@ -35,10 +35,6 @@ See the full license in the file "LICENSE" in the top level distribution directo
 #include <Hadrons/ModuleFactory.hpp>
 #include <Hadrons/A2AMatrix.hpp>
 
-#ifndef MF_IO_TYPE
-#define MF_IO_TYPE ComplexF
-#endif
-
 BEGIN_HADRONS_NAMESPACE
 
 /******************************************************************************
@@ -118,7 +114,7 @@ public:
     typedef A2AMatrixBlockComputation<Complex, 
                                       FermionField, 
                                       A2AMesonFieldMetadata, 
-                                      MF_IO_TYPE> Computation;
+                                      HADRONS_A2AM_IO_TYPE> Computation;
     typedef MesonFieldKernel<Complex, FImpl> Kernel;
 public:
     // constructor
@@ -248,7 +244,7 @@ void TA2AMesonField<FImpl>::execute(void)
         LOG(Message) << "  " << g << std::endl;
     }
     LOG(Message) << "Meson field size: " << nt << "*" << N_i << "*" << N_j 
-                 << " (filesize " << sizeString(nt*N_i*N_j*sizeof(MF_IO_TYPE)) 
+                 << " (filesize " << sizeString(nt*N_i*N_j*sizeof(HADRONS_A2AM_IO_TYPE)) 
                  << "/momentum/bilinear)" << std::endl;
 
     auto &ph = envGet(std::vector<ComplexField>, momphName_);

Hadrons/Modules/MGauge/Electrify.cc

@@ -0,0 +1,34 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: Hadrons/Modules/MGauge/Electrify.cc
+
+Copyright (C) 2015-2018
+
+Author: Vera Guelpers <Vera.Guelpers@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 <Hadrons/Modules/MGauge/Electrify.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MGauge;
+
+template class Grid::Hadrons::MGauge::TElectrify<GIMPL>;

Hadrons/Modules/MGauge/Electrify.hpp

@@ -0,0 +1,151 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: Hadrons/Modules/MGauge/Electrify.hpp
+
+Copyright (C) 2015-2018
+
+Author: Vera Guelpers <Vera.Guelpers@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 */
+
+#ifndef Hadrons_MGauge_Electrify_hpp_
+#define Hadrons_MGauge_Electrify_hpp_
+
+#include <Hadrons/Global.hpp>
+#include <Hadrons/Module.hpp>
+#include <Hadrons/ModuleFactory.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                              Electrify gauge                               *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MGauge)
+
+/****************************************************************************
+*  Electrify a gauge field:
+*
+*  Ue_mu(x) = U_mu(x)*exp(ieqA_mu(x))
+*
+*  with
+*
+*  - gauge: U_mu(x): gauge field
+*  - emField: A_mu(x): electromagnetic photon field
+*  - e: value for the elementary charge
+*  - q: charge in units of e
+*
+*****************************************************************************/
+
+
+class ElectrifyPar: Serializable
+{
+public:
+    GRID_SERIALIZABLE_CLASS_MEMBERS(ElectrifyPar,
+                                    std::string, gauge,
+				    std::string, emField,
+				    double, e,
+				    double, charge);
+};
+
+template <typename GImpl>
+class TElectrify: public Module<ElectrifyPar>
+{
+public:
+    GAUGE_TYPE_ALIASES(GImpl,);
+public:
+    typedef PhotonR::GaugeField     EmField;
+public:
+    // constructor
+    TElectrify(const std::string name);
+    // destructor
+    virtual ~TElectrify(void) {};
+    // dependencies/products
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+protected:
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+};
+
+MODULE_REGISTER_TMP(Electrify, TElectrify<GIMPL>, MGauge);
+
+/******************************************************************************
+*                            TElectrify implementation                             *
+******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename GImpl>
+TElectrify<GImpl>::TElectrify(const std::string name)
+: Module<ElectrifyPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename GImpl>
+std::vector<std::string> TElectrify<GImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().gauge, par().emField};
+
+    return in;
+}
+
+template <typename GImpl>
+std::vector<std::string> TElectrify<GImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+    
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename GImpl>
+void TElectrify<GImpl>::setup(void)
+{
+    envCreateLat(GaugeField, getName());
+    envTmpLat(LatticeComplex, "eiAmu");
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename GImpl>
+void TElectrify<GImpl>::execute(void)
+{
+    LOG(Message) << "Electrify the gauge field " << par().gauge << " using the photon field " 
+                  << par().emField << " with charge e*q= " << par().e << "*" << par().charge << std::endl;
+    
+    auto &Ue = envGet(GaugeField, getName());
+    auto &U = envGet(GaugeField, par().gauge);
+    auto &A = envGet(EmField,  par().emField);
+    envGetTmp(LatticeComplex, eiAmu);
+
+    Complex i(0.0,1.0);
+
+    for(unsigned int mu = 0; mu < env().getNd(); mu++)
+    {
+	eiAmu = exp(i * (Real)(par().e * par().charge) * PeekIndex<LorentzIndex>(A, mu));
+	PokeIndex<LorentzIndex>(Ue, PeekIndex<LorentzIndex>(U, mu) * eiAmu, mu);
+    }
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MGauge_Electrify_hpp_

Hadrons/Modules/MGauge/StochEm.cc

@@ -70,7 +70,7 @@ void TStochEm::execute(void)
     LOG(Message) << "Generating stochastic EM potential..." << std::endl;
 
     std::vector<Real> improvements = strToVec<Real>(par().improvement);
-    PhotonR photon(par().gauge, par().zmScheme, improvements, par().G0_qedInf);
+    PhotonR photon(envGetGrid(EmField), par().gauge, par().zmScheme, improvements);
     auto    &a = envGet(EmField, getName());
     auto    &w = envGet(EmComp, "_" + getName() + "_weight");
     

Hadrons/Modules/MGauge/StochEm.hpp

@@ -47,8 +47,7 @@ public:
     GRID_SERIALIZABLE_CLASS_MEMBERS(StochEmPar,
                                     PhotonR::Gauge,    gauge,
                                     PhotonR::ZmScheme, zmScheme,
-                                    std::string,       improvement,
-                                    Real,              G0_qedInf);
+                                    std::string,       improvement);
 };
 
 class TStochEm: public Module<StochEmPar>

Hadrons/Modules/MGauge/UnitEm.cc

@@ -62,7 +62,7 @@ void TUnitEm::setup(void)
 // execution ///////////////////////////////////////////////////////////////////
 void TUnitEm::execute(void)
 {
-    PhotonR photon(0, 0); // Just chose arbitrary input values here
+    PhotonR photon(envGetGrid(EmField), 0, 0); // Just chose arbitrary input values here
     auto    &a = envGet(EmField, getName());
     LOG(Message) << "Generating unit EM potential..." << std::endl;
     photon.UnitField(a);

Hadrons/Modules/MScalar/ChargedProp.cc

@@ -146,7 +146,7 @@ void TChargedProp::execute(void)
         std::vector<int>    siteCoor;
 
         LOG(Message) << "Saving momentum-projected propagator to '"
-                     << RESULT_FILE_NAME(par().output) << "'..."
+                     << RESULT_FILE_NAME(par().output, vm().getTrajectory()) << "'..."
                      << std::endl;
         result.projection.resize(par().outputMom.size());
         result.lattice_size = env().getGrid()->_fdimensions;

Hadrons/Modules/MScalar/ScalarVP.cc

@@ -462,7 +462,7 @@ void TScalarVP::execute(void)
     if (!par().output.empty())
     {
         LOG(Message) << "Saving momentum-projected HVP to '"
-                     << RESULT_FILE_NAME(par().output) << "'..."
+                     << RESULT_FILE_NAME(par().output, vm().getTrajectory()) << "'..."
                      << std::endl;
         saveResult(par().output, "HVP", outputData);
     }

Hadrons/Modules/MScalar/VPCounterTerms.cc

@@ -239,7 +239,7 @@ void TVPCounterTerms::execute(void)
     if (!par().output.empty())
     {
         LOG(Message) << "Saving momentum-projected correlators to '"
-                     << RESULT_FILE_NAME(par().output) << "'..."
+                     << RESULT_FILE_NAME(par().output, vm().getTrajectory()) << "'..."
                      << std::endl;
         saveResult(par().output, "scalar_loops", outputData);
     }

Hadrons/Modules/MSolver/A2AAslashVectors.cc

@@ -0,0 +1,35 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: Hadrons/Modules/MSolver/A2AAslashVectors.cc
+
+Copyright (C) 2015-2018
+
+Author: Vera Guelpers <Vera.Guelpers@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 <Hadrons/Modules/MSolver/A2AAslashVectors.hpp>
+
+using namespace Grid;
+using namespace Hadrons;
+using namespace MSolver;
+
+template class Grid::Hadrons::MSolver::TA2AAslashVectors<FIMPL>;
+template class Grid::Hadrons::MSolver::TA2AAslashVectors<ZFIMPL>;

Hadrons/Modules/MSolver/A2AAslashVectors.hpp

@@ -0,0 +1,194 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: Hadrons/Modules/MSolver/A2AAslashVectors.hpp
+
+Copyright (C) 2015-2018
+
+Author: Vera Guelpers <Vera.Guelpers@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 */
+#ifndef Hadrons_MSolver_A2AAslashVectors_hpp_
+#define Hadrons_MSolver_A2AAslashVectors_hpp_
+
+#include <Hadrons/Global.hpp>
+#include <Hadrons/Module.hpp>
+#include <Hadrons/ModuleFactory.hpp>
+#include <Hadrons/Solver.hpp>
+#include <Hadrons/A2AVectors.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+/******************************************************************************
+ *                       Create all-to-all V & W vectors                      *
+ ******************************************************************************/
+BEGIN_MODULE_NAMESPACE(MSolver)
+
+/****************************************************************************
+*  Calculate a sequential propagator on an insertion of i*g_mu*A_mu 
+*  on an A2A vector
+*
+*  vv_i(y) = S(y,x) * i * g_mu*A_mu(x) * v_i(x)
+*
+*  with
+*
+*  - vector: A2A vector v_i(x)
+*  - emField: A_mu(x): electromagnetic photon field
+*  - solver: the solver for calculating the sequential propagator
+*
+*****************************************************************************/
+
+class A2AAslashVectorsPar: Serializable
+{
+public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(A2AAslashVectorsPar,
+                                  std::string, vector,
+                                  std::string, emField,
+                                  std::string, solver,
+                                  std::string, output,
+                                  bool,        multiFile);
+};
+
+template <typename FImpl>
+class TA2AAslashVectors : public Module<A2AAslashVectorsPar>
+{
+public:
+    FERM_TYPE_ALIASES(FImpl,);
+    SOLVER_TYPE_ALIASES(FImpl,);
+public:
+    typedef PhotonR::GaugeField EmField;
+public:
+    // constructor
+    TA2AAslashVectors(const std::string name);
+    // destructor
+    virtual ~TA2AAslashVectors(void) {};
+    // dependency relation
+    virtual std::vector<std::string> getInput(void);
+    virtual std::vector<std::string> getOutput(void);
+    // setup
+    virtual void setup(void);
+    // execution
+    virtual void execute(void);
+private:
+    unsigned int Ls_;
+};
+
+MODULE_REGISTER_TMP(A2AAslashVectors, TA2AAslashVectors<FIMPL>, MSolver);
+MODULE_REGISTER_TMP(ZA2AAslashVectors, TA2AAslashVectors<ZFIMPL>, MSolver);
+
+/******************************************************************************
+ *                       TA2AAslashVectors implementation                       *
+ ******************************************************************************/
+// constructor /////////////////////////////////////////////////////////////////
+template <typename FImpl>
+TA2AAslashVectors<FImpl>::TA2AAslashVectors(const std::string name)
+: Module<A2AAslashVectorsPar>(name)
+{}
+
+// dependencies/products ///////////////////////////////////////////////////////
+template <typename FImpl>
+std::vector<std::string> TA2AAslashVectors<FImpl>::getInput(void)
+{
+    std::vector<std::string> in = {par().vector, par().emField, par().solver};
+
+    return in;
+}
+
+template <typename FImpl>
+std::vector<std::string> TA2AAslashVectors<FImpl>::getOutput(void)
+{
+    std::vector<std::string> out = {getName()};
+
+    return out;
+}
+
+// setup ///////////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TA2AAslashVectors<FImpl>::setup(void)
+{
+    Ls_  = env().getObjectLs(par().solver);
+    auto &vvector = envGet(std::vector<FermionField>, par().vector);
+    unsigned int Nmodes = vvector.size();
+    envCreate(std::vector<FermionField>, getName(), 1, 
+              Nmodes, envGetGrid(FermionField));
+   
+    envTmpLat(FermionField, "v4dtmp");
+    envTmpLat(FermionField, "v5dtmp", Ls_);
+    envTmpLat(FermionField, "v5dtmp_sol", Ls_);
+}
+
+// execution ///////////////////////////////////////////////////////////////////
+template <typename FImpl>
+void TA2AAslashVectors<FImpl>::execute(void)
+{
+    auto &solver = envGet(Solver, par().solver);
+    auto &stoch_photon = envGet(EmField,  par().emField);
+    auto &vvector = envGet(std::vector<FermionField>, par().vector);
+    auto &Aslashv = envGet(std::vector<FermionField>, getName());
+    unsigned int Nmodes = vvector.size();
+    auto &mat = solver.getFMat();
+    envGetTmp(FermionField, v4dtmp);
+    envGetTmp(FermionField, v5dtmp);
+    envGetTmp(FermionField, v5dtmp_sol);
+
+    Complex ci(0.0,1.0);
+
+    startTimer("Seq Aslash");
+    LOG(Message) << "Calculate Sequential propagator on Aslash * v with the A2A vector " 
+                 << par().vector << " and the photon field " << par().emField << std::endl;
+    for(unsigned int i=0; i<Nmodes; i++)
+    {
+        v4dtmp = zero;
+        startTimer("Multiply Aslash");
+        for(unsigned int mu=0;mu<=3;mu++)
+        {
+            Gamma gmu(Gamma::gmu[mu]);
+            v4dtmp +=  ci * PeekIndex<LorentzIndex>(stoch_photon, mu) * (gmu * vvector[i]);
+        }
+        stopTimer("Multiply Aslash");
+
+        startTimer("Inversion");
+        if (Ls_ == 1)
+        {
+            solver(Aslashv[i], v4dtmp);
+        }
+        else
+        {
+            mat.ImportPhysicalFermionSource(v4dtmp, v5dtmp);
+            solver(v5dtmp_sol, v5dtmp);
+            mat.ExportPhysicalFermionSolution(v5dtmp_sol, v4dtmp);
+            Aslashv[i] = v4dtmp;
+        }
+        stopTimer("Inversion");
+    }
+    stopTimer("Seq Aslash");
+    if (!par().output.empty())
+    {
+        startTimer("I/O");
+        A2AVectorsIo::write(par().output, Aslashv, par().multiFile, vm().getTrajectory());
+        stopTimer("I/O");
+    }
+}
+
+END_MODULE_NAMESPACE
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_MSolver_A2AAslashVectors_hpp_

Hadrons/Utilities/Contractor.cc

@@ -0,0 +1,442 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: Hadrons/Utilities/Contractor.cc
+
+Copyright (C) 2015-2018
+
+
+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 <Hadrons/Global.hpp>
+#include <Hadrons/A2AMatrix.hpp>
+#include <Hadrons/DiskVector.hpp>
+#include <Hadrons/TimerArray.hpp>
+#include <Hadrons/Utilities/Contractor.hpp>
+
+#ifdef GRID_COMMS_MPI3
+#define GET_RANK(rank, nMpi) \
+MPI_Comm_size(MPI_COMM_WORLD, &(nMpi));\
+MPI_Comm_rank(MPI_COMM_WORLD, &(rank))
+#define BARRIER() MPI_Barrier(MPI_COMM_WORLD)
+#define GLOBAL_DSUM(x) MPI_Allreduce(MPI_IN_PLACE, &x, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD)
+#define GLOBAL_DMAX(x) MPI_Allreduce(MPI_IN_PLACE, &x, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD)
+#define INIT() MPI_Init(NULL, NULL)
+#define FINALIZE() MPI_Finalize()
+#else
+#define GET_RANK(rank, nMpi) (nMpi) = 1; (rank) = 0
+#define BARRIER()
+#define GLOBAL_DSUM(x)
+#define GLOBAL_DMAX(x)
+#define INIT()
+#define FINALIZE()
+#endif
+
+using namespace Grid;
+using namespace QCD;
+using namespace Hadrons;
+
+#define TIME_MOD(t) (((t) + par.global.nt) % par.global.nt)
+
+struct ContractorPar
+{
+    Contractor::GlobalPar                  global;
+    std::vector<Contractor::A2AMatrixPar>  a2aMatrix;
+    std::vector<Contractor::ProductPar>    product;
+};
+
+void makeTimeSeq(std::vector<std::vector<unsigned int>> &timeSeq, 
+                 const std::vector<std::set<unsigned int>> &times,
+                 std::vector<unsigned int> &current,
+                 const unsigned int depth)
+{
+    if (depth > 0)
+    {
+        for (auto t: times[times.size() - depth])
+        {
+            current[times.size() - depth] = t;
+            makeTimeSeq(timeSeq, times, current, depth - 1);
+        }
+    }
+    else
+    {
+        timeSeq.push_back(current);
+    }
+}
+
+void makeTimeSeq(std::vector<std::vector<unsigned int>> &timeSeq, 
+                 const std::vector<std::set<unsigned int>> &times)
+{
+    std::vector<unsigned int> current(times.size());
+
+    makeTimeSeq(timeSeq, times, current, times.size());
+}
+
+void saveCorrelator(const Contractor::CorrelatorResult &result, const std::string dir, 
+                    const unsigned int dt, const unsigned int traj)
+{
+    std::string              fileStem = "", filename;
+    std::vector<std::string> terms = strToVec<std::string>(result.contraction.terms);
+
+    for (unsigned int i = 0; i < terms.size() - 1; i++)
+    {
+        fileStem += terms[i] + "_" + std::to_string(result.times[i]) + "_";
+    }
+    fileStem += terms.back();
+    if (!result.contraction.translationAverage)
+    {
+        fileStem += "_dt_" + std::to_string(dt);
+    }
+    filename = dir + "/" + RESULT_FILE_NAME(fileStem, traj);
+    std::cout << "Saving correlator to '" << filename << "'" << std::endl;
+    makeFileDir(dir);
+    ResultWriter writer(filename);
+    write(writer, fileStem, result);
+}
+
+void printPerf(const double bytes, const double usec)
+{
+    double maxt;
+
+    maxt = usec;
+    GLOBAL_DMAX(maxt);
+    std::cout << maxt/1.0e6 << " sec " << bytes/maxt*1.0e6/1024/1024/1024 << " GB/s";
+}
+
+void printPerf(const double bytes, const double busec, 
+               const double flops, const double fusec)
+{
+    double maxt;
+
+    printPerf(bytes, busec);
+    std::cout << " ";
+    maxt = fusec;
+    GLOBAL_DMAX(maxt);
+    std::cout << flops/fusec/1.0e3 << " GFlop/s";
+}
+
+std::set<unsigned int> parseTimeRange(const std::string str, const unsigned int nt)
+{
+    std::regex               rex("([0-9]+)|(([0-9]+)\\.\\.([0-9]+))");
+    std::smatch              sm;
+    std::vector<std::string> rstr = strToVec<std::string>(str);
+    std::set<unsigned int>   tSet;
+
+    for (auto &s: rstr)
+    {
+        std::regex_match(s, sm, rex);
+        if (sm[1].matched)
+        {
+            unsigned int t;
+            
+            t = std::stoi(sm[1].str());
+            if (t >= nt)
+            {
+                HADRONS_ERROR(Range, "time out of range (from expression '" + str + "')");
+            }
+            tSet.insert(t);
+        }
+        else if (sm[2].matched)
+        {
+            unsigned int ta, tb;
+
+            ta = std::stoi(sm[3].str());
+            tb = std::stoi(sm[4].str());
+            if ((ta >= nt) or (tb >= nt))
+            {
+                HADRONS_ERROR(Range, "time out of range (from expression '" + str + "')");
+            }
+            for (unsigned int ti = ta; ti <= tb; ++ti)
+            {
+                tSet.insert(ti);
+            }
+        }
+    }
+
+    return tSet;
+}
+
+int main(int argc, char* argv[])
+{
+    // MPI init
+    int nMpi, rank;
+
+    INIT();
+    GET_RANK(rank, nMpi);
+    if (rank != 0)
+    {
+        std::cout.setstate(std::ios::badbit);
+    }
+
+    // parse command line
+    std::string   parFilename;
+
+    if (argc != 2)
+    {
+        std::cerr << "usage: " << argv[0] << " <parameter file>";
+        std::cerr << std::endl;
+        
+        return EXIT_FAILURE;
+    }
+    parFilename = argv[1];
+
+    // parse parameter file
+    ContractorPar par;
+    unsigned int  nMat, nCont;
+    XmlReader     reader(parFilename);
+
+    read(reader, "global",    par.global);
+    read(reader, "a2aMatrix", par.a2aMatrix);
+    read(reader, "product",   par.product);
+    nMat  = par.a2aMatrix.size();
+    nCont = par.product.size();
+
+    // create diskvectors
+    std::map<std::string, EigenDiskVector<ComplexD>> a2aMat;
+    unsigned int                                     cacheSize;
+
+    for (auto &p: par.a2aMatrix)
+    {
+        std::string dirName = par.global.diskVectorDir + "/" + p.name + "." + std::to_string(rank);
+
+        a2aMat.emplace(p.name, EigenDiskVector<ComplexD>(dirName, par.global.nt, p.cacheSize));
+    }
+
+    // trajectory loop
+    std::vector<unsigned int> tList = par.global.trajCounter.getTrajectoryList();
+    unsigned int              indi, inde, indPerRank;
+
+    indPerRank = tList.size()/nMpi;
+    indi       = rank*indPerRank;
+    BARRIER();
+    for (unsigned int tInd = indi; tInd < indi + indPerRank; tInd++)
+    {
+        unsigned int traj;
+
+        if (tInd < tList.size())
+        {
+            traj = tList[tInd];
+        }
+        else
+        {
+            traj = tList.back();
+        }
+        if (nMpi > 1)
+        {
+            if (rank == 0)
+            {
+                std::cout << ":::::::: Trajectories ";
+                for (unsigned int r = 0; r < nMpi - 1; ++r)
+                {
+                    std::cout << tList[tInd + r*indPerRank] << " ";
+                }
+                if (tInd + (nMpi - 1)*indPerRank < tList.size())
+                {
+                    std::cout << tList[tInd + (nMpi - 1)*indPerRank];
+                }
+                std::cout << std::endl;
+            }
+        }
+        else
+        {
+            std::cout << ":::::::: Trajectory " << traj << std::endl;
+        }
+    
+        // load data
+        for (auto &p: par.a2aMatrix)
+        {
+            std::string filename = p.file;
+            double      t;
+
+            tokenReplace(filename, "traj", traj);
+            std::cout << "======== Loading '" << p.file << "'" << std::endl;
+
+            BARRIER();
+            A2AMatrixIo<HADRONS_A2AM_IO_TYPE> a2aIo(filename, p.dataset, par.global.nt);
+
+            a2aIo.load(a2aMat.at(p.name), &t);
+            GLOBAL_DMAX(t);
+            std::cout << "Read "  << nMpi*a2aIo.getSize() << " bytes in " << t/1.0e6 
+                      << " sec, " << nMpi*a2aIo.getSize()/t*1.0e6/1024/1024 
+                      << " MB/s"  << std::endl;
+        }
+
+        // contract
+        EigenDiskVector<ComplexD>::Matrix buf;
+
+        for (auto &p: par.product)
+        {
+            std::vector<std::string>               term = strToVec<std::string>(p.terms);
+            std::vector<std::set<unsigned int>>    times;
+            std::vector<std::vector<unsigned int>> timeSeq;
+            std::set<unsigned int>                 translations;
+            std::vector<A2AMatrixTr<ComplexD>>     lastTerm(par.global.nt);
+            A2AMatrix<ComplexD>                    prod, buf, tmp;
+            TimerArray                             tAr;
+            double                                 fusec, busec, flops, bytes, tusec;
+            Contractor::CorrelatorResult           result;             
+
+            BARRIER();
+            tAr.startTimer("Total");
+            std::cout << "======== Contraction tr(";
+            for (unsigned int g = 0; g < term.size(); ++g)
+            {
+                std::cout << term[g] << ((g == term.size() - 1) ? ')' : '*');
+            }
+            std::cout << std::endl;
+            if (term.size() != p.times.size() + 1)
+            {
+                HADRONS_ERROR(Size, "number of terms (" + std::to_string(term.size()) 
+                            + ") different from number of times (" 
+                            + std::to_string(p.times.size() + 1) + ")");
+            }
+            for (auto &s: p.times)
+            {
+                times.push_back(parseTimeRange(s, par.global.nt));
+            }
+            for (auto &m: par.a2aMatrix)
+            {
+                if (std::find(result.a2aMatrix.begin(), result.a2aMatrix.end(), m) == result.a2aMatrix.end())
+                {
+                    result.a2aMatrix.push_back(m);
+                    tokenReplace(result.a2aMatrix.back().file, "traj", traj);
+                }
+            }
+            result.contraction = p;
+            result.correlator.resize(par.global.nt, 0.);
+
+            translations = parseTimeRange(p.translations, par.global.nt);
+            makeTimeSeq(timeSeq, times);
+            std::cout << timeSeq.size()*translations.size()*(term.size() - 2) << " A*B, "
+                    << timeSeq.size()*translations.size()*par.global.nt << " tr(A*B)"
+                    << std::endl;
+
+            std::cout << "* Caching transposed last term" << std::endl;
+            for (unsigned int t = 0; t < par.global.nt; ++t)
+            {
+                tAr.startTimer("Disk vector overhead");
+                const A2AMatrix<ComplexD> &ref = a2aMat.at(term.back())[t];
+                tAr.stopTimer("Disk vector overhead");
+
+                tAr.startTimer("Transpose caching");
+                lastTerm[t].resize(ref.rows(), ref.cols());
+                parallel_for (unsigned int j = 0; j < ref.cols(); ++j)
+                for (unsigned int i = 0; i < ref.rows(); ++i)
+                {
+                    lastTerm[t](i, j) = ref(i, j);
+                }
+                tAr.stopTimer("Transpose caching");
+            }
+            bytes  = par.global.nt*lastTerm[0].rows()*lastTerm[0].cols();
+            bytes *= sizeof(ComplexD)*nMpi;
+            printPerf(bytes, tAr.getDTimer("Transpose caching"));
+            std::cout << std::endl;
+            for (unsigned int i = 0; i < timeSeq.size(); ++i)
+            {
+                unsigned int dti = 0;
+                auto         &t = timeSeq[i];
+
+                result.times = t;
+                for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast)
+                {
+                    result.correlator[tLast] = 0.;
+                }
+                for (auto &dt: translations)
+                {
+                    std::cout << "* Step " << i*translations.size() + dti + 1
+                            << "/" << timeSeq.size()*translations.size()
+                            << " -- positions= " << t << ", dt= " << dt << std::endl;
+                    if (term.size() > 2)
+                    {
+                        std::cout << std::setw(10) << "products ";
+                    }
+                    flops  = 0.;
+                    bytes  = 0.;
+                    fusec  = tAr.getDTimer("A*B algebra");
+                    busec  = tAr.getDTimer("A*B total");
+                    tAr.startTimer("Linear algebra");
+                    tAr.startTimer("Disk vector overhead");
+                    prod = a2aMat.at(term[0])[TIME_MOD(t[0] + dt)];
+                    tAr.stopTimer("Disk vector overhead");
+                    for (unsigned int j = 1; j < term.size() - 1; ++j)
+                    {
+                        tAr.startTimer("Disk vector overhead");
+                        const A2AMatrix<ComplexD> &ref = a2aMat.at(term[j])[TIME_MOD(t[j] + dt)];
+                        tAr.stopTimer("Disk vector overhead");
+                        
+                        tAr.startTimer("A*B total");
+                        tAr.startTimer("A*B algebra");
+                        A2AContraction::mul(tmp, prod, ref);
+                        tAr.stopTimer("A*B algebra");
+                        flops += A2AContraction::mulFlops(prod, ref);
+                        prod   = tmp;
+                        tAr.stopTimer("A*B total");
+                        bytes += 3.*tmp.rows()*tmp.cols()*sizeof(ComplexD);
+                    }
+                    if (term.size() > 2)
+                    {
+                        printPerf(bytes*nMpi, tAr.getDTimer("A*B total") - busec,
+                                  flops*nMpi, tAr.getDTimer("A*B algebra") - fusec);
+                        std::cout << std::endl;
+                    }
+                    std::cout << std::setw(10) << "traces ";
+                    flops  = 0.;
+                    bytes  = 0.;
+                    fusec  = tAr.getDTimer("tr(A*B)");
+                    busec  = tAr.getDTimer("tr(A*B)");
+                    for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast)
+                    {
+                        tAr.startTimer("tr(A*B)");
+                        A2AContraction::accTrMul(result.correlator[TIME_MOD(tLast - dt)], prod, lastTerm[tLast]);
+                        tAr.stopTimer("tr(A*B)");
+                        flops += A2AContraction::accTrMulFlops(prod, lastTerm[tLast]);
+                        bytes += 2.*prod.rows()*prod.cols()*sizeof(ComplexD);
+                    }
+                    tAr.stopTimer("Linear algebra");
+                    printPerf(bytes*nMpi, tAr.getDTimer("tr(A*B)") - busec,
+                              flops*nMpi, tAr.getDTimer("tr(A*B)") - fusec);
+                    std::cout << std::endl;
+                    if (!p.translationAverage)
+                    {
+                        saveCorrelator(result, par.global.output, dt, traj);
+                        for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast)
+                        {
+                            result.correlator[tLast] = 0.;
+                        }
+                    }
+                    dti++;
+                }
+                if (p.translationAverage)
+                {
+                    for (unsigned int tLast = 0; tLast < par.global.nt; ++tLast)
+                    {
+                        result.correlator[tLast] /= translations.size();
+                    }
+                    saveCorrelator(result, par.global.output, 0, traj);
+                }
+            }
+            tAr.stopTimer("Total");
+            printTimeProfile(tAr.getTimings(), tAr.getTimer("Total"));
+        }
+    }
+    
+    FINALIZE();
+
+    return EXIT_SUCCESS;
+}

Hadrons/Utilities/Contractor.hpp

@@ -0,0 +1,53 @@
+#ifndef  Hadrons_Contractor_hpp_
+#define Hadrons_Contractor_hpp_
+
+#include <Hadrons/Global.hpp>
+
+BEGIN_HADRONS_NAMESPACE
+
+namespace Contractor
+{   
+    class GlobalPar: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(GlobalPar,
+                                        TrajRange, trajCounter,
+                                        unsigned int, nt,
+                                        std::string, diskVectorDir,
+                                        std::string, output);
+    };
+
+    class A2AMatrixPar: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(A2AMatrixPar,
+                                        std::string, file,
+                                        std::string, dataset,
+                                        unsigned int, cacheSize,
+                                        std::string, name);
+    };
+
+    class ProductPar: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(ProductPar,
+                                        std::string, terms,
+                                        std::vector<std::string>, times,
+                                        std::string, translations,
+                                        bool, translationAverage);
+    };
+
+    class CorrelatorResult: Serializable
+    {
+    public:
+        GRID_SERIALIZABLE_CLASS_MEMBERS(CorrelatorResult,
+                                        std::vector<Contractor::A2AMatrixPar>,  a2aMatrix,
+                                        ProductPar, contraction,
+                                        std::vector<unsigned int>, times,
+                                        std::vector<ComplexD>, correlator);
+    };
+}
+
+END_HADRONS_NAMESPACE
+
+#endif // Hadrons_Contractor_hpp_

Hadrons/Utilities/ContractorBenchmark.cc

@@ -0,0 +1,434 @@
+#include <Hadrons/Global.hpp>
+#include <Hadrons/A2AMatrix.hpp>
+#ifdef USE_MKL
+#include "mkl.h"
+#include "mkl_cblas.h"
+#endif
+
+using namespace Grid;
+using namespace Hadrons;
+
+#ifdef GRID_COMMS_MPI3
+#define GET_RANK(rank, nMpi) \
+MPI_Comm_size(MPI_COMM_WORLD, &(nMpi));\
+MPI_Comm_rank(MPI_COMM_WORLD, &(rank))
+#define BARRIER() MPI_Barrier(MPI_COMM_WORLD)
+#define INIT() MPI_Init(NULL, NULL)
+#define FINALIZE() MPI_Finalize()
+#else
+#define GET_RANK(rank, nMpi) (nMpi) = 1; (rank) = 0
+#define BARRIER()
+#define INIT()
+#define FINALIZE()
+#endif
+
+template <typename Function, typename MatLeft, typename MatRight>
+inline void trBenchmark(const std::string name, const MatLeft &left,
+                        const MatRight &right, const ComplexD ref, Function fn)
+{
+    double       t, flops, bytes, n = left[0].rows()*left[0].cols();
+    unsigned int nMat = left.size();
+    int          nMpi, rank;
+    ComplexD     buf;
+
+    t = 0.;
+    GET_RANK(rank, nMpi);
+    t = -usecond();
+    BARRIER();
+    for (unsigned int i = rank*nMat/nMpi; i < (rank+1)*nMat/nMpi; ++i)
+    {
+        fn(buf, left[i], right[i]);      
+    }
+    BARRIER();
+    t += usecond();
+    flops = nMat*(6.*n + 2.*(n - 1.));
+    bytes = nMat*(2.*n*sizeof(ComplexD));
+
+    if (rank == 0)
+    {
+        std::cout << std::setw(34) << name << ": diff= "
+                  << std::setw(12) << std::norm(buf-ref)
+                  << std::setw(10) << t/1.0e6 << " sec "
+                  << std::setw(10) << flops/t/1.0e3 << " GFlop/s " 
+                  << std::setw(10) << bytes/t*1.0e6/1024/1024/1024 << " GB/s "
+                  << std::endl;
+    }
+    ::sleep(1);
+}
+
+template <typename Function, typename MatV, typename Mat>
+inline void mulBenchmark(const std::string name, const MatV &left,
+                         const MatV &right, const Mat &ref, Function fn)
+{
+    double       t, flops, bytes;
+    double       nr = left[0].rows(), nc = left[0].cols(), n = nr*nc;
+    unsigned int nMat = left.size();
+    int          nMpi, rank;
+    Mat          buf(left[0].rows(), left[0].rows());
+
+    t = 0.;
+    GET_RANK(rank, nMpi);
+    t = -usecond();
+    BARRIER();
+    for (unsigned int i = rank*nMat/nMpi; i < (rank+1)*nMat/nMpi; ++i)
+    {
+        fn(buf, left[i], right[i]);
+    }
+    BARRIER();
+    t += usecond();
+    flops = nMat*(nr*nr*(6.*nc + 2.*(nc - 1.)));
+    bytes = nMat*(2*nc*nr*sizeof(ComplexD));
+
+    if (rank == 0)
+    {
+        std::cout << std::setw(34) << name << ": diff= "
+                  << std::setw(12) << (buf-ref).squaredNorm()
+                  << std::setw(10) << t/1.0e6 << " sec "
+                  << std::setw(10) << flops/t/1.0e3 << " GFlop/s " 
+                  << std::setw(10) << bytes/t*1.0e6/1024/1024/1024 << " GB/s "
+                  << std::endl;
+    }
+    ::sleep(1);
+}
+
+#ifdef USE_MKL
+template <typename MatLeft, typename MatRight>
+static inline void zdotuRow(ComplexD &res, const unsigned int aRow,
+                            const MatLeft &a, const MatRight &b)
+{
+    const ComplexD *aPt, *bPt;
+    unsigned int   aInc, bInc;
+
+    if (MatLeft::Options == Eigen::RowMajor)
+    {
+        aPt  = a.data() + aRow*a.cols();
+        aInc = 1;
+    }
+    else if (MatLeft::Options == Eigen::ColMajor)
+    {
+        aPt  = a.data() + aRow;
+        aInc = a.rows();
+    }
+    if (MatRight::Options == Eigen::RowMajor)
+    {
+        bPt  = b.data() + aRow;
+        bInc = b.cols();
+    }
+    else if (MatRight::Options == Eigen::ColMajor)
+    {
+        bPt  = b.data() + aRow*b.rows();
+        bInc = 1;
+    }
+    cblas_zdotu_sub(a.cols(), aPt, aInc, bPt, bInc, &res);
+}
+
+template <typename MatLeft, typename MatRight>
+static inline void zdotuCol(ComplexD &res, const unsigned int aCol,
+                            const MatLeft &a, const MatRight &b)
+{
+    const ComplexD *aPt, *bPt;
+    unsigned int   aInc, bInc;
+
+    if (MatLeft::Options == Eigen::RowMajor)
+    {
+        aPt  = a.data() + aCol;
+        aInc = a.cols();
+    }
+    else if (MatLeft::Options == Eigen::ColMajor)
+    {
+        aPt  = a.data() + aCol*a.rows();
+        aInc = 1;
+    }
+    if (MatRight::Options == Eigen::RowMajor)
+    {
+        bPt  = b.data() + aCol*b.cols();
+        bInc = 1;
+    }
+    else if (MatRight::Options == Eigen::ColMajor)
+    {
+        bPt  = b.data() + aCol;
+        bInc = b.rows();
+    }
+    cblas_zdotu_sub(a.rows(), aPt, aInc, bPt, bInc, &res);
+}
+#endif
+
+template <typename MatLeft, typename MatRight>
+void fullTrBenchmark(const unsigned int ni, const unsigned int nj, const unsigned int nMat)
+{
+    std::vector<MatLeft>  left;
+    std::vector<MatRight> right;
+    MatRight              buf;
+    ComplexD              ref;
+    int                   rank, nMpi;
+
+    left.resize(nMat, MatLeft::Random(ni, nj));
+    right.resize(nMat, MatRight::Random(nj, ni));
+    GET_RANK(rank, nMpi);
+    if (rank == 0)
+    {
+        std::cout << "==== tr(A*B) benchmarks" << std::endl;
+        std::cout << "A matrices use ";
+        if (MatLeft::Options == Eigen::RowMajor)
+        {
+            std::cout << "row-major ordering" << std::endl;
+        }
+        else if (MatLeft::Options == Eigen::ColMajor)
+        {
+            std::cout << "col-major ordering" << std::endl;
+        }
+        std::cout << "B matrices use ";
+        if (MatRight::Options == Eigen::RowMajor)
+        {
+            std::cout << "row-major ordering" << std::endl;
+        }
+        else if (MatRight::Options == Eigen::ColMajor)
+        {
+            std::cout << "col-major ordering" << std::endl;
+        }
+        std::cout << std::endl;
+    }
+    BARRIER();
+    ref = (left.back()*right.back()).trace();
+    trBenchmark("Hadrons A2AContraction::accTrMul", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    { 
+        res = 0.;
+        A2AContraction::accTrMul(res, a, b);
+    });
+    trBenchmark("Naive loop rows first", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    { 
+        auto nr = a.rows(), nc = a.cols();
+        
+        res = 0.;
+        parallel_for (unsigned int i = 0; i < nr; ++i)
+        {
+            ComplexD tmp = 0.;
+
+            for (unsigned int j = 0; j < nc; ++j)
+            {
+                tmp += a(i, j)*b(j, i);
+            }
+            parallel_critical
+            {
+                res += tmp;
+            }
+        }
+    });
+    trBenchmark("Naive loop cols first", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    {
+        auto nr = a.rows(), nc = a.cols();
+        
+        res = 0.;
+        parallel_for (unsigned int j = 0; j < nc; ++j)
+        {
+            ComplexD tmp = 0.;
+
+            for (unsigned int i = 0; i < nr; ++i)
+            {
+                tmp += a(i, j)*b(j, i);
+            }        
+            parallel_critical
+            {
+                res += tmp;
+            }
+        }
+    });
+    trBenchmark("Eigen tr(A*B)", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    { 
+        res = (a*b).trace();
+    });
+    trBenchmark("Eigen row-wise dot", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    {
+        res = 0.;
+        parallel_for (unsigned int r = 0; r < a.rows(); ++r)
+        {
+            ComplexD tmp;
+
+            tmp = a.row(r).conjugate().dot(b.col(r));
+            parallel_critical
+            {
+                res += tmp;
+            }
+        }
+    });
+    trBenchmark("Eigen col-wise dot", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    {
+        res = 0.;
+        parallel_for (unsigned int c = 0; c < a.cols(); ++c)
+        {
+            ComplexD tmp;
+
+            tmp = a.col(c).conjugate().dot(b.row(c));
+            parallel_critical
+            {
+                res += tmp;
+            }
+        }
+    });
+    trBenchmark("Eigen Hadamard", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    { 
+        res = a.cwiseProduct(b.transpose()).sum();
+    });
+#ifdef USE_MKL
+    trBenchmark("MKL row-wise zdotu", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    {
+        res = 0.;
+        parallel_for (unsigned int r = 0; r < a.rows(); ++r)
+        {
+            ComplexD tmp;
+
+            zdotuRow(tmp, r, a, b);
+            parallel_critical
+            {
+                res += tmp;
+            }
+        }
+    });
+    trBenchmark("MKL col-wise zdotu", left, right, ref,
+    [](ComplexD &res, const MatLeft &a, const MatRight &b)
+    {
+        res = 0.;
+        parallel_for (unsigned int c = 0; c < a.cols(); ++c)
+        {
+            ComplexD tmp;
+
+            zdotuCol(tmp, c, a, b);
+            parallel_critical
+            {
+                res += tmp;
+            }
+        }
+    });
+#endif
+    BARRIER();
+    if (rank == 0)
+    {
+        std::cout << std::endl;
+    }
+}
+
+template <typename Mat>
+void fullMulBenchmark(const unsigned int ni, const unsigned int nj, const unsigned int nMat)
+{
+    std::vector<Mat> left, right;
+    Mat              ref;
+    int              rank, nMpi;
+
+    left.resize(nMat, Mat::Random(ni, nj));
+    right.resize(nMat, Mat::Random(nj, ni));
+    GET_RANK(rank, nMpi);
+    if (rank == 0)
+    {
+        std::cout << "==== A*B benchmarks" << std::endl;
+        std::cout << "all matrices use ";
+        if (Mat::Options == Eigen::RowMajor)
+        {
+            std::cout << "row-major ordering" << std::endl;
+        }
+        else if (Mat::Options == Eigen::ColMajor)
+        {
+            std::cout << "col-major ordering" << std::endl;
+        }
+        std::cout << std::endl;
+    }
+    BARRIER();
+    ref = left.back()*right.back();
+    mulBenchmark("Hadrons A2AContraction::mul", left, right, ref,
+    [](Mat &res, const Mat &a, const Mat &b)
+    { 
+        A2AContraction::mul(res, a, b);
+    });
+    mulBenchmark("Eigen A*B", left, right, ref,
+    [](Mat &res, const Mat &a, const Mat &b)
+    { 
+        res = a*b;
+    });
+#ifdef USE_MKL
+    mulBenchmark("MKL A*B", left, right, ref,
+    [](Mat &res, const Mat &a, const Mat &b)
+    {
+        const ComplexD one(1., 0.), zero(0., 0.);
+        if (Mat::Options == Eigen::RowMajor)
+        {
+            cblas_zgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, a.rows(), b.cols(),
+                        a.cols(), &one, a.data(), a.cols(), b.data(), b.cols(), &zero,
+                        res.data(), res.cols());
+        }
+        else if (Mat::Options == Eigen::ColMajor)
+        {
+            cblas_zgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, a.rows(), b.cols(),
+                        a.cols(), &one, a.data(), a.rows(), b.data(), b.rows(), &zero,
+                        res.data(), res.rows());
+        }
+    });
+#endif
+    BARRIER();
+    if (rank == 0)
+    {
+        std::cout << std::endl;
+    }
+}
+
+int main(int argc, char *argv[])
+{
+    // parse command line
+    Eigen::Index ni, nj, nMat;
+    int          nMpi, rank;
+
+    if (argc != 4)
+    {
+        std::cerr << "usage: " << argv[0] << " <Ni> <Nj> <#matrices>";
+        std::cerr << std::endl;
+        
+        return EXIT_FAILURE;
+    }
+    ni   = std::stoi(argv[1]);
+    nj   = std::stoi(argv[2]);
+    nMat = std::stoi(argv[3]);
+
+    INIT();
+    GET_RANK(rank, nMpi);
+    if (rank == 0)
+    {
+        std::cout << "\n*** ALL-TO-ALL MATRIX CONTRACTION BENCHMARK ***\n" << std::endl;
+        std::cout << nMat << " couples of " << ni << "x" << nj << " matrices\n" << std::endl;
+
+        std::cout << nMpi << " MPI processes" << std::endl;
+#ifdef GRID_OMP
+        #pragma omp parallel
+        {
+            #pragma omp single
+            std::cout << omp_get_num_threads() << " threads\n" << std::endl; 
+        }
+#else
+        std::cout << "Single-threaded\n" << std::endl; 
+#endif
+
+#ifdef EIGEN_USE_MKL_ALL
+        std::cout << "Eigen uses the MKL" << std::endl;
+#endif
+        std::cout << "Eigen uses " << Eigen::nbThreads() << " threads" << std::endl;
+#ifdef USE_MKL
+        std::cout << "MKL   uses " << mkl_get_max_threads() << " threads" << std::endl;
+#endif
+        std::cout << std::endl;
+    }
+
+    fullTrBenchmark<A2AMatrix<ComplexD>, A2AMatrix<ComplexD>>(ni, nj, nMat);
+    fullTrBenchmark<A2AMatrix<ComplexD>, A2AMatrixTr<ComplexD>>(ni, nj, nMat);
+    fullTrBenchmark<A2AMatrixTr<ComplexD>, A2AMatrix<ComplexD>>(ni, nj, nMat);
+    fullTrBenchmark<A2AMatrixTr<ComplexD>, A2AMatrixTr<ComplexD>>(ni, nj, nMat);
+    fullMulBenchmark<A2AMatrix<ComplexD>>(ni, nj, nMat);
+    fullMulBenchmark<A2AMatrixTr<ComplexD>>(ni, nj, nMat);
+    FINALIZE();
+
+    return EXIT_SUCCESS;
+}

Hadrons/Utilities/Makefile.am

@@ -1,4 +1,4 @@
-bin_PROGRAMS = HadronsXmlRun HadronsFermionEP64To32
+bin_PROGRAMS = HadronsXmlRun HadronsFermionEP64To32 HadronsContractor HadronsContractorBenchmark
 
 HadronsXmlRun_SOURCES = HadronsXmlRun.cc
 HadronsXmlRun_LDADD   = ../libHadrons.a ../../Grid/libGrid.a
@@ -6,3 +6,9 @@ HadronsXmlRun_LDADD   = ../libHadrons.a ../../Grid/libGrid.a
 HadronsFermionEP64To32_SOURCES  = EigenPackCast.cc
 HadronsFermionEP64To32_CXXFLAGS = $(AM_CXXFLAGS) -DFIN=WilsonImplD::FermionField -DFOUT=WilsonImplF::FermionField
 HadronsFermionEP64To32_LDADD    = ../libHadrons.a ../../Grid/libGrid.a
+
+HadronsContractor_SOURCES = Contractor.cc Contractor.hpp
+HadronsContractor_LDADD   = ../libHadrons.a ../../Grid/libGrid.a
+
+HadronsContractorBenchmark_SOURCES = ContractorBenchmark.cc
+HadronsContractorBenchmark_LDADD   = ../libHadrons.a ../../Grid/libGrid.a

Hadrons/modules.inc

@@ -20,11 +20,13 @@ modules_cc =\
   Modules/MSink/Point.cc \
   Modules/MSink/Smear.cc \
   Modules/MSolver/A2AVectors.cc \
+  Modules/MSolver/A2AAslashVectors.cc \
   Modules/MSolver/RBPrecCG.cc \
   Modules/MSolver/MixedPrecisionRBPrecCG.cc \
   Modules/MSolver/LocalCoherenceLanczos.cc \
   Modules/MGauge/StoutSmearing.cc \
   Modules/MGauge/Unit.cc \
+  Modules/MGauge/Electrify.cc \
   Modules/MGauge/UnitEm.cc \
   Modules/MGauge/StochEm.cc \
   Modules/MGauge/Random.cc \
@@ -95,9 +97,11 @@ modules_hpp =\
   Modules/MSolver/Guesser.hpp \
   Modules/MSolver/RBPrecCG.hpp \
   Modules/MSolver/A2AVectors.hpp \
+  Modules/MSolver/A2AAslashVectors.hpp \
   Modules/MGauge/UnitEm.hpp \
   Modules/MGauge/StoutSmearing.hpp \
   Modules/MGauge/Unit.hpp \
+  Modules/MGauge/Electrify.hpp \
   Modules/MGauge/Random.hpp \
   Modules/MGauge/GaugeFix.hpp \
   Modules/MGauge/FundtoHirep.hpp \

configure.ac

@@ -123,10 +123,13 @@ case ${ac_SFW_FP16} in
       AC_MSG_ERROR(["SFW FP16 option not supported ${ac_SFW_FP16}"]);;
 esac
 
-############### MKL
+############### Intel libraries
 AC_ARG_ENABLE([mkl],
     [AC_HELP_STRING([--enable-mkl=yes|no|prefix], [enable Intel MKL for LAPACK & FFTW])],
     [ac_MKL=${enable_mkl}], [ac_MKL=no])
+AC_ARG_ENABLE([ipp],
+    [AC_HELP_STRING([--enable-ipp=yes|no|prefix], [enable Intel IPP for fast CRC32C])],
+    [ac_IPP=${enable_mkl}], [ac_IPP=no])
 
 case ${ac_MKL} in
     no)
@@ -139,6 +142,17 @@ case ${ac_MKL} in
         AC_DEFINE([USE_MKL], [1], [Define to 1 if you use the Intel MKL]);;
 esac
 
+case ${ac_IPP} in
+    no)
+        ;;
+    yes)
+        AC_DEFINE([USE_IPP], [1], [Define to 1 if you use the Intel IPP]);;
+    *)
+        AM_CXXFLAGS="-I$ac_IPP/include $AM_CXXFLAGS"
+        AM_LDFLAGS="-L$ac_IPP/lib $AM_LDFLAGS"
+        AC_DEFINE([USE_IPP], [1], [Define to 1 if you use the Intel IPP]);;
+esac
+
 ############### HDF5
 AC_ARG_WITH([hdf5],
     [AS_HELP_STRING([--with-hdf5=prefix],
@@ -170,7 +184,13 @@ AC_CHECK_FUNCS([gettimeofday])
 
 if test "${ac_MKL}x" != "nox"; then
     AC_SEARCH_LIBS([mkl_set_interface_layer], [mkl_rt], [],
-                   [AC_MSG_ERROR("MKL enabled but library not found")])
+                   [AC_MSG_ERROR("Intel MKL enabled but library not found")])
+fi
+
+if test "${ac_IPP}x" != "nox"; then
+     AC_SEARCH_LIBS([ippsCRC32C_8u], [ippdc], 
+                   [LIBS="${LIBS} -lippdc -lippvm -lipps -lippcore"],
+                   [AC_MSG_ERROR("Intel IPP enabled but library not found")])
 fi
 
 AC_SEARCH_LIBS([__gmpf_init], [gmp],
@@ -485,6 +505,7 @@ DX_INIT_DOXYGEN([$PACKAGE_NAME], [doxygen.cfg])
 
 ############### Ouput
 cwd=`pwd -P`; cd ${srcdir}; abs_srcdir=`pwd -P`; cd ${cwd}
+GRID_CXX="$CXX"
 GRID_CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
 GRID_LDFLAGS="$AM_LDFLAGS $LDFLAGS"
 GRID_LIBS=$LIBS
@@ -497,6 +518,7 @@ AM_LDFLAGS="-L${cwd}/Grid $AM_LDFLAGS"
 AC_SUBST([AM_CFLAGS])
 AC_SUBST([AM_CXXFLAGS])
 AC_SUBST([AM_LDFLAGS])
+AC_SUBST([GRID_CXX])
 AC_SUBST([GRID_CXXFLAGS])
 AC_SUBST([GRID_LDFLAGS])
 AC_SUBST([GRID_LIBS])

grid-config.in

@@ -61,6 +61,10 @@ while test $# -gt 0; do
       echo @GRID_CXXFLAGS@
     ;;
     
+    --cxx)
+      echo @GRID_CXX@
+    ;;
+    
     --ldflags)
       echo @GRID_LDFLAGS@
     ;;

tests/core/Test_qed.cc

@@ -0,0 +1,138 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid 
+
+Source file: tests/core/Test_qed.cc
+
+Copyright (C) 2015-2018
+
+Author: Antonin Portelli <antonin.portelli@me.com>
+Author: James Harrison <J.Harrison@soton.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
+*************************************************************************************/
+
+#include <Grid/Grid.h>
+
+using namespace Grid;
+using namespace QCD;
+
+typedef PeriodicGaugeImpl<QedGImplR>  QedPeriodicGImplR;
+typedef PhotonR::GaugeField           EmField;
+typedef PhotonR::GaugeLinkField       EmComp;
+
+const int NCONFIGS = 20;
+const int NWILSON  = 10;
+
+int main(int argc, char *argv[])
+{
+  // initialization
+  Grid_init(&argc, &argv);
+  std::cout << GridLogMessage << "Grid initialized" << std::endl;
+  
+  // QED stuff
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(4, vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian    grid(latt_size,simd_layout,mpi_layout);
+  GridParallelRNG  pRNG(&grid);
+  PhotonR          photon(&grid, PhotonR::Gauge::coulomb, PhotonR::ZmScheme::qedL);
+  EmField          a(&grid);
+  EmField          expA(&grid);
+
+  Complex imag_unit(0, 1);
+
+  Real wlA;
+  std::vector<Real> logWlAvg(NWILSON, 0.0), logWlTime(NWILSON, 0.0), logWlSpace(NWILSON, 0.0);
+
+  pRNG.SeedFixedIntegers({1, 2, 3, 4});
+
+  std::cout << GridLogMessage << "Wilson loop calculation beginning" << std::endl;
+  for(int ic = 0; ic < NCONFIGS; ic++){
+      std::cout << GridLogMessage << "Configuration " << ic <<std::endl;
+      photon.StochasticField(a, pRNG);
+
+      // Exponentiate photon field
+      expA = exp(imag_unit*a);
+
+      // Calculate zero-modes
+      std::vector<EmField::vector_object::scalar_object> zm;
+
+      std::cout << GridLogMessage << "Total zero-mode norm 2 " 
+                << std::sqrt(norm2(sum(a))) << std::endl;
+
+      std::cout << GridLogMessage << "Spatial zero-mode norm 2" << std::endl;
+      sliceSum(a, zm, grid.Nd() - 1);
+      for (unsigned int t = 0; t < latt_size.back(); ++t)
+      {
+        std::cout << GridLogMessage << "t = " << t << " " << std::sqrt(norm2(zm[t])) << std::endl;
+      }
+
+      // Calculate divergence
+      EmComp diva(&grid), amu(&grid);
+
+      diva = zero;
+      for (unsigned int mu = 0; mu < grid.Nd(); ++mu)
+      {
+        amu   = peekLorentz(a, mu);
+        diva += amu - Cshift(amu, mu, -1);
+        if (mu == grid.Nd() - 2)
+        {
+          std::cout << GridLogMessage << "Spatial divergence norm 2 " << std::sqrt(norm2(diva)) << std::endl;
+        }
+      }
+      std::cout << GridLogMessage << "Total divergence norm 2 " << std::sqrt(norm2(diva)) << std::endl;
+
+      // Calculate Wilson loops
+      for(int iw=1; iw<=NWILSON; iw++){
+          wlA = WilsonLoops<QedPeriodicGImplR>::avgWilsonLoop(expA, iw, iw) * 3;
+          logWlAvg[iw-1] -= 2*log(wlA);
+          wlA = WilsonLoops<QedPeriodicGImplR>::avgTimelikeWilsonLoop(expA, iw, iw) * 3;
+          logWlTime[iw-1] -= 2*log(wlA);
+          wlA = WilsonLoops<QedPeriodicGImplR>::avgSpatialWilsonLoop(expA, iw, iw) * 3;
+          logWlSpace[iw-1] -= 2*log(wlA);
+      }
+  }
+  std::cout << GridLogMessage << "Wilson loop calculation completed" << std::endl;
+  
+  // Calculate Wilson loops
+  // From A. Portelli's PhD thesis:
+  // size  -2*log(W)
+  // 1     0.500000000(1)
+  // 2     1.369311535(1) 
+  // 3     2.305193057(1) 
+  // 4     3.261483854(1) 
+  // 5     4.228829967(1) 
+  // 6     5.203604529(1) 
+  // 7     6.183728249(1) 
+  // 8     7.167859805(1) 
+  // 9     8.155091868(1) 
+  // 10    9.144788116(1)
+
+  for(int iw=1; iw<=10; iw++){
+      std::cout << GridLogMessage << iw << 'x' << iw << " Wilson loop" << std::endl;
+      std::cout << GridLogMessage << "-2*log(W) average: " << logWlAvg[iw-1]/NCONFIGS << std::endl;
+      std::cout << GridLogMessage << "-2*log(W) timelike: " << logWlTime[iw-1]/NCONFIGS << std::endl;
+      std::cout << GridLogMessage << "-2*log(W) spatial: " << logWlSpace[iw-1]/NCONFIGS << std::endl;
+  }
+
+  // epilogue
+  std::cout << GridLogMessage << "Grid is finalizing now" << std::endl;
+  Grid_finalize();
+  
+  return EXIT_SUCCESS;
+}

tests/debug/Test_split_laplacian.cc

@@ -0,0 +1,104 @@
+    /*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/Test_dwf_mrhs_cg.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 <Grid/Grid.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+
+int main (int argc, char ** argv)
+{
+  typedef LatticeComplex ComplexField; 
+
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  int nd   = latt_size.size();
+  int ndm1 = nd-1;
+
+  std::vector<int> simd_layout = GridDefaultSimd(nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  std::vector<int> mpi_split (mpi_layout.size(),1);
+
+  std::cout << " Full " << GridCmdVectorIntToString(latt_size)  << " subgrid"         <<std::endl;
+  std::cout << " Full " << GridCmdVectorIntToString(mpi_layout) << " sub communicator"<<std::endl;
+  std::cout << " Full " << GridCmdVectorIntToString(simd_layout)<< " simd layout "    <<std::endl;
+
+  GridCartesian         * GridN = new GridCartesian(latt_size,
+						    simd_layout,
+						    mpi_layout);
+
+  std::vector<int> latt_m  = latt_size;   latt_m[nd-1] = 1;
+  std::vector<int> mpi_m   = mpi_layout;  mpi_m [nd-1] = 1;
+  std::vector<int> simd_m  = GridDefaultSimd(ndm1,vComplex::Nsimd()); simd_m.push_back(1);
+
+
+  std::cout << " Requesting " << GridCmdVectorIntToString(latt_m)<< " subgrid"         <<std::endl;
+  std::cout << " Requesting " << GridCmdVectorIntToString(mpi_m) << " sub communicator"<<std::endl;
+  std::cout << " Requesting " << GridCmdVectorIntToString(simd_m)<< " simd layout "    <<std::endl;
+  GridCartesian         * Grid_m = new GridCartesian(latt_m,
+						     simd_m,
+						     mpi_m,
+						     *GridN); 
+
+  Complex C(1.0);
+  Complex tmp;
+
+  ComplexField Full(GridN); Full = C;
+  ComplexField Full_cpy(GridN);
+  ComplexField Split(Grid_m);Split= C;
+
+  std::cout << GridLogMessage<< " Full  volume "<< norm2(Full) <<std::endl;
+  std::cout << GridLogMessage<< " Split volume "<< norm2(Split) <<std::endl;
+
+  tmp=C;
+  GridN->GlobalSum(tmp);
+  std::cout << GridLogMessage<< " Full  nodes "<< tmp <<std::endl;
+
+  tmp=C;
+  Grid_m->GlobalSum(tmp);
+  std::cout << GridLogMessage<< " Split nodes "<< tmp <<std::endl;
+  GridN->Barrier();
+
+  auto local_latt = GridN->LocalDimensions();
+
+  Full_cpy = zero;
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          RNG(GridN);  RNG.SeedFixedIntegers(seeds);
+
+  random(RNG,Full);
+  for(int t=0;t<local_latt[nd-1];t++){
+    ExtractSliceLocal(Split,Full,0,t,Tp);
+    InsertSliceLocal (Split,Full_cpy,0,t,Tp);
+  }
+  Full_cpy = Full_cpy - Full;
+  std::cout << " NormFull " << norm2(Full)<<std::endl;
+  std::cout << " NormDiff " << norm2(Full_cpy)<<std::endl;
+  Grid_finalize();
+}

tests/hadrons/Test_QED.cc

@@ -72,6 +72,7 @@ int main(int argc, char *argv[])
     
     // set fermion boundary conditions to be periodic space, antiperiodic time.
     std::string boundary = "1 1 1 -1";
+    std::string twist    = "0. 0. 0. 0.";
 
     //stochastic photon field
     MGauge::StochEm::Par photonPar;
@@ -90,6 +91,7 @@ int main(int argc, char *argv[])
         actionPar.M5    = 1.8;
         actionPar.mass  = mass[i];
         actionPar.boundary = boundary;
+        actionPar.twist = "0. 0. 0. 0.";
         application.createModule<MAction::DWF>("DWF_" + flavour[i], actionPar);
 
         

tests/hadrons/Test_hadrons.hpp

@@ -126,6 +126,7 @@ inline void makeWilsonAction(Application &application, std::string actionName,
         actionPar.gauge = gaugeField;
         actionPar.mass  = mass;
         actionPar.boundary = boundary;
+        actionPar.twist = "0. 0. 0. 0.";
         application.createModule<MAction::Wilson>(actionName, actionPar);
     }
 }
@@ -154,6 +155,7 @@ inline void makeDWFAction(Application &application, std::string actionName,
         actionPar.M5    = M5;
         actionPar.mass  = mass;
         actionPar.boundary = boundary;
+        actionPar.twist = "0. 0. 0. 0.";
         application.createModule<MAction::DWF>(actionName, actionPar);
     }
 }

tests/hadrons/Test_hadrons_meson_3pt.cc

@@ -66,6 +66,7 @@ int main(int argc, char *argv[])
     
     // set fermion boundary conditions to be periodic space, antiperiodic time.
     std::string boundary = "1 1 1 -1";
+    std::string twist = "0. 0. 0. 0.";
 
     // sink
     MSink::Point::Par sinkPar;
@@ -80,6 +81,7 @@ int main(int argc, char *argv[])
         actionPar.M5    = 1.8;
         actionPar.mass  = mass[i];
         actionPar.boundary = boundary;
+        actionPar.twist = twist;
         application.createModule<MAction::DWF>("DWF_" + flavour[i], actionPar);
         
         // solvers

tests/hadrons/Test_hadrons_spectrum.cc

@@ -72,6 +72,7 @@ int main(int argc, char *argv[])
     
     // set fermion boundary conditions to be periodic space, antiperiodic time.
     std::string boundary = "1 1 1 -1";
+    std::string twist = "0. 0. 0. 0.";
 
     for (unsigned int i = 0; i < flavour.size(); ++i)
     {
@@ -82,6 +83,7 @@ int main(int argc, char *argv[])
         actionPar.M5    = 1.8;
         actionPar.mass  = mass[i];
         actionPar.boundary = boundary;
+        actionPar.twist = twist;
         application.createModule<MAction::DWF>("DWF_" + flavour[i], actionPar);
         
         // solvers

tests/solver/Test_multigrid_common.h

@@ -0,0 +1,670 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/solver/Test_multigrid_common.h
+
+    Copyright (C) 2015-2018
+
+    Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+    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_TEST_MULTIGRID_COMMON_H
+#define GRID_TEST_MULTIGRID_COMMON_H
+
+namespace Grid {
+
+// TODO: Can think about having one parameter struct per level and then a
+// vector of these structs. How well would that work together with the
+// serialization strategy of Grid?
+
+// clang-format off
+struct MultiGridParams : Serializable {
+public:
+  GRID_SERIALIZABLE_CLASS_MEMBERS(MultiGridParams,
+                                  int,                           nLevels,
+                                  std::vector<std::vector<int>>, blockSizes,           // size == nLevels - 1
+                                  std::vector<double>,           smootherTol,          // size == nLevels - 1
+                                  std::vector<int>,              smootherMaxOuterIter, // size == nLevels - 1
+                                  std::vector<int>,              smootherMaxInnerIter, // size == nLevels - 1
+                                  bool,                          kCycle,
+                                  std::vector<double>,           kCycleTol,            // size == nLevels - 1
+                                  std::vector<int>,              kCycleMaxOuterIter,   // size == nLevels - 1
+                                  std::vector<int>,              kCycleMaxInnerIter,   // size == nLevels - 1
+                                  double,                        coarseSolverTol,
+                                  int,                           coarseSolverMaxOuterIter,
+                                  int,                           coarseSolverMaxInnerIter);
+
+  // constructor with default values
+  MultiGridParams(int                           _nLevels                  = 2,
+                  std::vector<std::vector<int>> _blockSizes               = {{4, 4, 4, 4}},
+                  std::vector<double>           _smootherTol              = {1e-14},
+                  std::vector<int>              _smootherMaxOuterIter     = {4},
+                  std::vector<int>              _smootherMaxInnerIter     = {4},
+                  bool                          _kCycle                   = true,
+                  std::vector<double>           _kCycleTol                = {1e-1},
+                  std::vector<int>              _kCycleMaxOuterIter       = {2},
+                  std::vector<int>              _kCycleMaxInnerIter       = {5},
+                  double                        _coarseSolverTol          = 5e-2,
+                  int                           _coarseSolverMaxOuterIter = 10,
+                  int                           _coarseSolverMaxInnerIter = 500)
+  : nLevels(_nLevels)
+  , blockSizes(_blockSizes)
+  , smootherTol(_smootherTol)
+  , smootherMaxOuterIter(_smootherMaxOuterIter)
+  , smootherMaxInnerIter(_smootherMaxInnerIter)
+  , kCycle(_kCycle)
+  , kCycleTol(_kCycleTol)
+  , kCycleMaxOuterIter(_kCycleMaxOuterIter)
+  , kCycleMaxInnerIter(_kCycleMaxInnerIter)
+  , coarseSolverTol(_coarseSolverTol)
+  , coarseSolverMaxOuterIter(_coarseSolverMaxOuterIter)
+  , coarseSolverMaxInnerIter(_coarseSolverMaxInnerIter)
+  {}
+};
+// clang-format on
+
+void checkParameterValidity(MultiGridParams const &params) {
+
+  auto correctSize = params.nLevels - 1;
+
+  assert(correctSize == params.blockSizes.size());
+  assert(correctSize == params.smootherTol.size());
+  assert(correctSize == params.smootherMaxOuterIter.size());
+  assert(correctSize == params.smootherMaxInnerIter.size());
+  assert(correctSize == params.kCycleTol.size());
+  assert(correctSize == params.kCycleMaxOuterIter.size());
+  assert(correctSize == params.kCycleMaxInnerIter.size());
+}
+
+struct LevelInfo {
+public:
+  std::vector<std::vector<int>> Seeds;
+  std::vector<GridCartesian *>  Grids;
+  std::vector<GridParallelRNG>  PRNGs;
+
+  LevelInfo(GridCartesian *FineGrid, MultiGridParams const &mgParams) {
+
+    auto nCoarseLevels = mgParams.blockSizes.size();
+
+    assert(nCoarseLevels == mgParams.nLevels - 1);
+
+    // set up values for finest grid
+    Grids.push_back(FineGrid);
+    Seeds.push_back({1, 2, 3, 4});
+    PRNGs.push_back(GridParallelRNG(Grids.back()));
+    PRNGs.back().SeedFixedIntegers(Seeds.back());
+
+    // set up values for coarser grids
+    for(int level = 1; level < mgParams.nLevels; ++level) {
+      auto Nd  = Grids[level - 1]->_ndimension;
+      auto tmp = Grids[level - 1]->_fdimensions;
+      assert(tmp.size() == Nd);
+
+      Seeds.push_back(std::vector<int>(Nd));
+
+      for(int d = 0; d < Nd; ++d) {
+        tmp[d] /= mgParams.blockSizes[level - 1][d];
+        Seeds[level][d] = (level)*Nd + d + 1;
+      }
+
+      Grids.push_back(QCD::SpaceTimeGrid::makeFourDimGrid(tmp, Grids[level - 1]->_simd_layout, GridDefaultMpi()));
+      PRNGs.push_back(GridParallelRNG(Grids[level]));
+
+      PRNGs[level].SeedFixedIntegers(Seeds[level]);
+    }
+
+    std::cout << GridLogMessage << "Constructed " << mgParams.nLevels << " levels" << std::endl;
+
+    for(int level = 0; level < mgParams.nLevels; ++level) {
+      std::cout << GridLogMessage << "level = " << level << ":" << std::endl;
+      Grids[level]->show_decomposition();
+    }
+  }
+};
+
+template<class Field> class MultiGridPreconditionerBase : public LinearFunction<Field> {
+public:
+  virtual ~MultiGridPreconditionerBase()               = default;
+  virtual void setup()                                 = 0;
+  virtual void operator()(Field const &in, Field &out) = 0;
+  virtual void runChecks(RealD tolerance)              = 0;
+  virtual void reportTimings()                         = 0;
+};
+
+template<class Fobj, class CComplex, int nBasis, int nCoarserLevels, class Matrix>
+class MultiGridPreconditioner : public MultiGridPreconditionerBase<Lattice<Fobj>> {
+public:
+  /////////////////////////////////////////////
+  // Type Definitions
+  /////////////////////////////////////////////
+
+  // clang-format off
+  typedef Aggregation<Fobj, CComplex, nBasis>                                                                         Aggregates;
+  typedef CoarsenedMatrix<Fobj, CComplex, nBasis>                                                                     CoarseDiracMatrix;
+  typedef typename Aggregates::CoarseVector                                                                           CoarseVector;
+  typedef typename Aggregates::siteVector                                                                             CoarseSiteVector;
+  typedef Matrix                                                                                                      FineDiracMatrix;
+  typedef typename Aggregates::FineField                                                                              FineVector;
+  typedef MultiGridPreconditioner<CoarseSiteVector, iScalar<CComplex>, nBasis, nCoarserLevels - 1, CoarseDiracMatrix> NextPreconditionerLevel;
+  // clang-format on
+
+  /////////////////////////////////////////////
+  // Member Data
+  /////////////////////////////////////////////
+
+  int _CurrentLevel;
+  int _NextCoarserLevel;
+
+  MultiGridParams &_MultiGridParams;
+  LevelInfo &      _LevelInfo;
+
+  FineDiracMatrix & _FineMatrix;
+  FineDiracMatrix & _SmootherMatrix;
+  Aggregates        _Aggregates;
+  CoarseDiracMatrix _CoarseMatrix;
+
+  std::unique_ptr<NextPreconditionerLevel> _NextPreconditionerLevel;
+
+  GridStopWatch _SetupTotalTimer;
+  GridStopWatch _SetupCreateSubspaceTimer;
+  GridStopWatch _SetupProjectToChiralitiesTimer;
+  GridStopWatch _SetupCoarsenOperatorTimer;
+  GridStopWatch _SetupNextLevelTimer;
+  GridStopWatch _SolveTotalTimer;
+  GridStopWatch _SolveRestrictionTimer;
+  GridStopWatch _SolveProlongationTimer;
+  GridStopWatch _SolveSmootherTimer;
+  GridStopWatch _SolveNextLevelTimer;
+
+  /////////////////////////////////////////////
+  // Member Functions
+  /////////////////////////////////////////////
+
+  MultiGridPreconditioner(MultiGridParams &mgParams, LevelInfo &LvlInfo, FineDiracMatrix &FineMat, FineDiracMatrix &SmootherMat)
+    : _CurrentLevel(mgParams.nLevels - (nCoarserLevels + 1)) // _Level = 0 corresponds to finest
+    , _NextCoarserLevel(_CurrentLevel + 1)                   // incremented for instances on coarser levels
+    , _MultiGridParams(mgParams)
+    , _LevelInfo(LvlInfo)
+    , _FineMatrix(FineMat)
+    , _SmootherMatrix(SmootherMat)
+    , _Aggregates(_LevelInfo.Grids[_NextCoarserLevel], _LevelInfo.Grids[_CurrentLevel], 0)
+    , _CoarseMatrix(*_LevelInfo.Grids[_NextCoarserLevel]) {
+
+    _NextPreconditionerLevel
+      = std::unique_ptr<NextPreconditionerLevel>(new NextPreconditionerLevel(_MultiGridParams, _LevelInfo, _CoarseMatrix, _CoarseMatrix));
+
+    resetTimers();
+  }
+
+  void setup() {
+
+    _SetupTotalTimer.Start();
+
+    static_assert((nBasis & 0x1) == 0, "MG Preconditioner only supports an even number of basis vectors");
+    int nb = nBasis / 2;
+
+    MdagMLinearOperator<FineDiracMatrix, FineVector> fineMdagMOp(_FineMatrix);
+
+    _SetupCreateSubspaceTimer.Start();
+    _Aggregates.CreateSubspace(_LevelInfo.PRNGs[_CurrentLevel], fineMdagMOp, nb);
+    _SetupCreateSubspaceTimer.Stop();
+
+    _SetupProjectToChiralitiesTimer.Start();
+    FineVector tmp1(_Aggregates.subspace[0]._grid);
+    FineVector tmp2(_Aggregates.subspace[0]._grid);
+    for(int n = 0; n < nb; n++) {
+      auto tmp1 = _Aggregates.subspace[n];
+      G5C(tmp2, _Aggregates.subspace[n]);
+      axpby(_Aggregates.subspace[n], 0.5, 0.5, tmp1, tmp2);
+      axpby(_Aggregates.subspace[n + nb], 0.5, -0.5, tmp1, tmp2);
+      std::cout << GridLogMG << " Level " << _CurrentLevel << ": Chirally doubled vector " << n << ". "
+                << "norm2(vec[" << n << "]) = " << norm2(_Aggregates.subspace[n]) << ". "
+                << "norm2(vec[" << n + nb << "]) = " << norm2(_Aggregates.subspace[n + nb]) << std::endl;
+    }
+    _SetupProjectToChiralitiesTimer.Stop();
+
+    _SetupCoarsenOperatorTimer.Start();
+    _CoarseMatrix.CoarsenOperator(_LevelInfo.Grids[_CurrentLevel], fineMdagMOp, _Aggregates);
+    _SetupCoarsenOperatorTimer.Stop();
+
+    _SetupNextLevelTimer.Start();
+    _NextPreconditionerLevel->setup();
+    _SetupNextLevelTimer.Stop();
+
+    _SetupTotalTimer.Stop();
+  }
+
+  virtual void operator()(FineVector const &in, FineVector &out) {
+
+    conformable(_LevelInfo.Grids[_CurrentLevel], in._grid);
+    conformable(in, out);
+
+    // TODO: implement a W-cycle
+    if(_MultiGridParams.kCycle)
+      kCycle(in, out);
+    else
+      vCycle(in, out);
+  }
+
+  void vCycle(FineVector const &in, FineVector &out) {
+
+    _SolveTotalTimer.Start();
+
+    RealD inputNorm = norm2(in);
+
+    CoarseVector coarseSrc(_LevelInfo.Grids[_NextCoarserLevel]);
+    CoarseVector coarseSol(_LevelInfo.Grids[_NextCoarserLevel]);
+    coarseSol = zero;
+
+    FineVector fineTmp(in._grid);
+
+    auto maxSmootherIter = _MultiGridParams.smootherMaxOuterIter[_CurrentLevel] * _MultiGridParams.smootherMaxInnerIter[_CurrentLevel];
+
+    TrivialPrecon<FineVector>                      fineTrivialPreconditioner;
+    FlexibleGeneralisedMinimalResidual<FineVector> fineFGMRES(_MultiGridParams.smootherTol[_CurrentLevel],
+                                                              maxSmootherIter,
+                                                              fineTrivialPreconditioner,
+                                                              _MultiGridParams.smootherMaxInnerIter[_CurrentLevel],
+                                                              false);
+
+    MdagMLinearOperator<FineDiracMatrix, FineVector> fineMdagMOp(_FineMatrix);
+    MdagMLinearOperator<FineDiracMatrix, FineVector> fineSmootherMdagMOp(_SmootherMatrix);
+
+    _SolveRestrictionTimer.Start();
+    _Aggregates.ProjectToSubspace(coarseSrc, in);
+    _SolveRestrictionTimer.Stop();
+
+    _SolveNextLevelTimer.Start();
+    (*_NextPreconditionerLevel)(coarseSrc, coarseSol);
+    _SolveNextLevelTimer.Stop();
+
+    _SolveProlongationTimer.Start();
+    _Aggregates.PromoteFromSubspace(coarseSol, out);
+    _SolveProlongationTimer.Stop();
+
+    fineMdagMOp.Op(out, fineTmp);
+    fineTmp                                = in - fineTmp;
+    auto r                                 = norm2(fineTmp);
+    auto residualAfterCoarseGridCorrection = std::sqrt(r / inputNorm);
+
+    _SolveSmootherTimer.Start();
+    fineFGMRES(fineSmootherMdagMOp, in, out);
+    _SolveSmootherTimer.Stop();
+
+    fineMdagMOp.Op(out, fineTmp);
+    fineTmp                        = in - fineTmp;
+    r                              = norm2(fineTmp);
+    auto residualAfterPostSmoother = std::sqrt(r / inputNorm);
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": V-cycle: Input norm = " << std::sqrt(inputNorm)
+              << " Coarse residual = " << residualAfterCoarseGridCorrection << " Post-Smoother residual = " << residualAfterPostSmoother
+              << std::endl;
+
+    _SolveTotalTimer.Stop();
+  }
+
+  void kCycle(FineVector const &in, FineVector &out) {
+
+    _SolveTotalTimer.Start();
+
+    RealD inputNorm = norm2(in);
+
+    CoarseVector coarseSrc(_LevelInfo.Grids[_NextCoarserLevel]);
+    CoarseVector coarseSol(_LevelInfo.Grids[_NextCoarserLevel]);
+    coarseSol = zero;
+
+    FineVector fineTmp(in._grid);
+
+    auto smootherMaxIter = _MultiGridParams.smootherMaxOuterIter[_CurrentLevel] * _MultiGridParams.smootherMaxInnerIter[_CurrentLevel];
+    auto kCycleMaxIter   = _MultiGridParams.kCycleMaxOuterIter[_CurrentLevel] * _MultiGridParams.kCycleMaxInnerIter[_CurrentLevel];
+
+    TrivialPrecon<FineVector>                        fineTrivialPreconditioner;
+    FlexibleGeneralisedMinimalResidual<FineVector>   fineFGMRES(_MultiGridParams.smootherTol[_CurrentLevel],
+                                                              smootherMaxIter,
+                                                              fineTrivialPreconditioner,
+                                                              _MultiGridParams.smootherMaxInnerIter[_CurrentLevel],
+                                                              false);
+    FlexibleGeneralisedMinimalResidual<CoarseVector> coarseFGMRES(_MultiGridParams.kCycleTol[_CurrentLevel],
+                                                                  kCycleMaxIter,
+                                                                  *_NextPreconditionerLevel,
+                                                                  _MultiGridParams.kCycleMaxInnerIter[_CurrentLevel],
+                                                                  false);
+
+    MdagMLinearOperator<FineDiracMatrix, FineVector>     fineMdagMOp(_FineMatrix);
+    MdagMLinearOperator<FineDiracMatrix, FineVector>     fineSmootherMdagMOp(_SmootherMatrix);
+    MdagMLinearOperator<CoarseDiracMatrix, CoarseVector> coarseMdagMOp(_CoarseMatrix);
+
+    _SolveRestrictionTimer.Start();
+    _Aggregates.ProjectToSubspace(coarseSrc, in);
+    _SolveRestrictionTimer.Stop();
+
+    _SolveNextLevelTimer.Start();
+    coarseFGMRES(coarseMdagMOp, coarseSrc, coarseSol);
+    _SolveNextLevelTimer.Stop();
+
+    _SolveProlongationTimer.Start();
+    _Aggregates.PromoteFromSubspace(coarseSol, out);
+    _SolveProlongationTimer.Stop();
+
+    fineMdagMOp.Op(out, fineTmp);
+    fineTmp                                = in - fineTmp;
+    auto r                                 = norm2(fineTmp);
+    auto residualAfterCoarseGridCorrection = std::sqrt(r / inputNorm);
+
+    _SolveSmootherTimer.Start();
+    fineFGMRES(fineSmootherMdagMOp, in, out);
+    _SolveSmootherTimer.Stop();
+
+    fineMdagMOp.Op(out, fineTmp);
+    fineTmp                        = in - fineTmp;
+    r                              = norm2(fineTmp);
+    auto residualAfterPostSmoother = std::sqrt(r / inputNorm);
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": K-cycle: Input norm = " << std::sqrt(inputNorm)
+              << " Coarse residual = " << residualAfterCoarseGridCorrection << " Post-Smoother residual = " << residualAfterPostSmoother
+              << std::endl;
+
+    _SolveTotalTimer.Stop();
+  }
+
+  void runChecks(RealD tolerance) {
+
+    std::vector<FineVector>   fineTmps(7, _LevelInfo.Grids[_CurrentLevel]);
+    std::vector<CoarseVector> coarseTmps(4, _LevelInfo.Grids[_NextCoarserLevel]);
+
+    MdagMLinearOperator<FineDiracMatrix, FineVector>     fineMdagMOp(_FineMatrix);
+    MdagMLinearOperator<CoarseDiracMatrix, CoarseVector> coarseMdagMOp(_CoarseMatrix);
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": MG correctness check: 0 == (M - (Mdiag + Σ_μ Mdir_μ)) * v" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+
+    random(_LevelInfo.PRNGs[_CurrentLevel], fineTmps[0]);
+
+    fineMdagMOp.Op(fineTmps[0], fineTmps[1]);     //     M * v
+    fineMdagMOp.OpDiag(fineTmps[0], fineTmps[2]); // Mdiag * v
+
+    fineTmps[4] = zero;
+    for(int dir = 0; dir < 4; dir++) { //       Σ_μ Mdir_μ * v
+      for(auto disp : {+1, -1}) {
+        fineMdagMOp.OpDir(fineTmps[0], fineTmps[3], dir, disp);
+        fineTmps[4] = fineTmps[4] + fineTmps[3];
+      }
+    }
+
+    fineTmps[5] = fineTmps[2] + fineTmps[4]; // (Mdiag + Σ_μ Mdir_μ) * v
+
+    fineTmps[6]    = fineTmps[1] - fineTmps[5];
+    auto deviation = std::sqrt(norm2(fineTmps[6]) / norm2(fineTmps[1]));
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(M * v)                    = " << norm2(fineTmps[1]) << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(Mdiag * v)                = " << norm2(fineTmps[2]) << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(Σ_μ Mdir_μ * v)           = " << norm2(fineTmps[4]) << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2((Mdiag + Σ_μ Mdir_μ) * v) = " << norm2(fineTmps[5]) << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": relative deviation              = " << deviation;
+
+    if(deviation > tolerance) {
+      std::cout << " > " << tolerance << " -> check failed" << std::endl;
+      abort();
+    } else {
+      std::cout << " < " << tolerance << " -> check passed" << std::endl;
+    }
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": MG correctness check: 0 == (1 - P R) v" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+
+    for(auto i = 0; i < _Aggregates.subspace.size(); ++i) {
+      _Aggregates.ProjectToSubspace(coarseTmps[0], _Aggregates.subspace[i]); //   R v_i
+      _Aggregates.PromoteFromSubspace(coarseTmps[0], fineTmps[0]);           // P R v_i
+
+      fineTmps[1] = _Aggregates.subspace[i] - fineTmps[0]; // v_i - P R v_i
+      deviation   = std::sqrt(norm2(fineTmps[1]) / norm2(_Aggregates.subspace[i]));
+
+      std::cout << GridLogMG << " Level " << _CurrentLevel << ": Vector " << i << ": norm2(v_i) = " << norm2(_Aggregates.subspace[i])
+                << " | norm2(R v_i) = " << norm2(coarseTmps[0]) << " | norm2(P R v_i) = " << norm2(fineTmps[0])
+                << " | relative deviation = " << deviation;
+
+      if(deviation > tolerance) {
+        std::cout << " > " << tolerance << " -> check failed" << std::endl;
+        abort();
+      } else {
+        std::cout << " < " << tolerance << " -> check passed" << std::endl;
+      }
+    }
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": MG correctness check: 0 == (1 - R P) v_c" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+
+    random(_LevelInfo.PRNGs[_NextCoarserLevel], coarseTmps[0]);
+
+    _Aggregates.PromoteFromSubspace(coarseTmps[0], fineTmps[0]); //   P v_c
+    _Aggregates.ProjectToSubspace(coarseTmps[1], fineTmps[0]);   // R P v_c
+
+    coarseTmps[2] = coarseTmps[0] - coarseTmps[1]; // v_c - R P v_c
+    deviation     = std::sqrt(norm2(coarseTmps[2]) / norm2(coarseTmps[0]));
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(v_c) = " << norm2(coarseTmps[0])
+              << " | norm2(R P v_c) = " << norm2(coarseTmps[1]) << " | norm2(P v_c) = " << norm2(fineTmps[0])
+              << " | relative deviation = " << deviation;
+
+    if(deviation > tolerance) {
+      std::cout << " > " << tolerance << " -> check failed" << std::endl;
+      abort();
+    } else {
+      std::cout << " < " << tolerance << " -> check passed" << std::endl;
+    }
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": MG correctness check: 0 == (R D P - D_c) v_c" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+
+    random(_LevelInfo.PRNGs[_NextCoarserLevel], coarseTmps[0]);
+
+    _Aggregates.PromoteFromSubspace(coarseTmps[0], fineTmps[0]); //     P v_c
+    fineMdagMOp.Op(fineTmps[0], fineTmps[1]);                    //   D P v_c
+    _Aggregates.ProjectToSubspace(coarseTmps[1], fineTmps[1]);   // R D P v_c
+
+    coarseMdagMOp.Op(coarseTmps[0], coarseTmps[2]); // D_c v_c
+
+    coarseTmps[3] = coarseTmps[1] - coarseTmps[2]; // R D P v_c - D_c v_c
+    deviation     = std::sqrt(norm2(coarseTmps[3]) / norm2(coarseTmps[1]));
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": norm2(R D P v_c) = " << norm2(coarseTmps[1])
+              << " | norm2(D_c v_c) = " << norm2(coarseTmps[2]) << " | relative deviation = " << deviation;
+
+    if(deviation > tolerance) {
+      std::cout << " > " << tolerance << " -> check failed" << std::endl;
+      abort();
+    } else {
+      std::cout << " < " << tolerance << " -> check passed" << std::endl;
+    }
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": MG correctness check: 0 == |(Im(v_c^dag D_c^dag D_c v_c)|" << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": **************************************************" << std::endl;
+
+    random(_LevelInfo.PRNGs[_NextCoarserLevel], coarseTmps[0]);
+
+    coarseMdagMOp.Op(coarseTmps[0], coarseTmps[1]);    //         D_c v_c
+    coarseMdagMOp.AdjOp(coarseTmps[1], coarseTmps[2]); // D_c^dag D_c v_c
+
+    auto dot  = innerProduct(coarseTmps[0], coarseTmps[2]); //v_c^dag D_c^dag D_c v_c
+    deviation = std::abs(imag(dot)) / std::abs(real(dot));
+
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Re(v_c^dag D_c^dag D_c v_c) = " << real(dot)
+              << " | Im(v_c^dag D_c^dag D_c v_c) = " << imag(dot) << " | relative deviation = " << deviation;
+
+    if(deviation > tolerance) {
+      std::cout << " > " << tolerance << " -> check failed" << std::endl;
+      abort();
+    } else {
+      std::cout << " < " << tolerance << " -> check passed" << std::endl;
+    }
+
+    _NextPreconditionerLevel->runChecks(tolerance);
+  }
+
+  void reportTimings() {
+
+    // clang-format off
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Sum   total            " <<                _SetupTotalTimer.Elapsed() + _SolveTotalTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Setup total            " <<                _SetupTotalTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Setup create subspace  " <<       _SetupCreateSubspaceTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Setup project chiral   " << _SetupProjectToChiralitiesTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Setup coarsen operator " <<      _SetupCoarsenOperatorTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Setup next level       " <<            _SetupNextLevelTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Solve total            " <<                _SolveTotalTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Solve restriction      " <<          _SolveRestrictionTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Solve prolongation     " <<         _SolveProlongationTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Solve smoother         " <<             _SolveSmootherTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Solve next level       " <<            _SolveNextLevelTimer.Elapsed() << std::endl;
+    // clang-format on
+
+    _NextPreconditionerLevel->reportTimings();
+  }
+
+  void resetTimers() {
+
+    _SetupTotalTimer.Reset();
+    _SetupCreateSubspaceTimer.Reset();
+    _SetupProjectToChiralitiesTimer.Reset();
+    _SetupCoarsenOperatorTimer.Reset();
+    _SetupNextLevelTimer.Reset();
+    _SolveTotalTimer.Reset();
+    _SolveRestrictionTimer.Reset();
+    _SolveProlongationTimer.Reset();
+    _SolveSmootherTimer.Reset();
+    _SolveNextLevelTimer.Reset();
+
+    _NextPreconditionerLevel->resetTimers();
+  }
+};
+
+// Specialization for the coarsest level
+template<class Fobj, class CComplex, int nBasis, class Matrix>
+class MultiGridPreconditioner<Fobj, CComplex, nBasis, 0, Matrix> : public MultiGridPreconditionerBase<Lattice<Fobj>> {
+public:
+  /////////////////////////////////////////////
+  // Type Definitions
+  /////////////////////////////////////////////
+
+  typedef Matrix        FineDiracMatrix;
+  typedef Lattice<Fobj> FineVector;
+
+  /////////////////////////////////////////////
+  // Member Data
+  /////////////////////////////////////////////
+
+  int _CurrentLevel;
+
+  MultiGridParams &_MultiGridParams;
+  LevelInfo &      _LevelInfo;
+
+  FineDiracMatrix &_FineMatrix;
+  FineDiracMatrix &_SmootherMatrix;
+
+  GridStopWatch _SolveTotalTimer;
+  GridStopWatch _SolveSmootherTimer;
+
+  /////////////////////////////////////////////
+  // Member Functions
+  /////////////////////////////////////////////
+
+  MultiGridPreconditioner(MultiGridParams &mgParams, LevelInfo &LvlInfo, FineDiracMatrix &FineMat, FineDiracMatrix &SmootherMat)
+    : _CurrentLevel(mgParams.nLevels - (0 + 1))
+    , _MultiGridParams(mgParams)
+    , _LevelInfo(LvlInfo)
+    , _FineMatrix(FineMat)
+    , _SmootherMatrix(SmootherMat) {
+
+    resetTimers();
+  }
+
+  void setup() {}
+
+  virtual void operator()(FineVector const &in, FineVector &out) {
+
+    _SolveTotalTimer.Start();
+
+    conformable(_LevelInfo.Grids[_CurrentLevel], in._grid);
+    conformable(in, out);
+
+    auto coarseSolverMaxIter = _MultiGridParams.coarseSolverMaxOuterIter * _MultiGridParams.coarseSolverMaxInnerIter;
+
+    // On the coarsest level we only have what I above call the fine level, no coarse one
+    TrivialPrecon<FineVector>                      fineTrivialPreconditioner;
+    FlexibleGeneralisedMinimalResidual<FineVector> fineFGMRES(
+      _MultiGridParams.coarseSolverTol, coarseSolverMaxIter, fineTrivialPreconditioner, _MultiGridParams.coarseSolverMaxInnerIter, false);
+
+    MdagMLinearOperator<FineDiracMatrix, FineVector> fineMdagMOp(_FineMatrix);
+
+    _SolveSmootherTimer.Start();
+    fineFGMRES(fineMdagMOp, in, out);
+    _SolveSmootherTimer.Stop();
+
+    _SolveTotalTimer.Stop();
+  }
+
+  void runChecks(RealD tolerance) {}
+
+  void reportTimings() {
+
+    // clang-format off
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Solve total            " <<    _SolveTotalTimer.Elapsed() << std::endl;
+    std::cout << GridLogMG << " Level " << _CurrentLevel << ": Time elapsed: Solve smoother         " << _SolveSmootherTimer.Elapsed() << std::endl;
+    // clang-format on
+  }
+
+  void resetTimers() {
+
+    _SolveTotalTimer.Reset();
+    _SolveSmootherTimer.Reset();
+  }
+};
+
+template<class Fobj, class CComplex, int nBasis, int nLevels, class Matrix>
+using NLevelMGPreconditioner = MultiGridPreconditioner<Fobj, CComplex, nBasis, nLevels - 1, Matrix>;
+
+template<class Fobj, class CComplex, int nBasis, class Matrix>
+std::unique_ptr<MultiGridPreconditionerBase<Lattice<Fobj>>>
+createMGInstance(MultiGridParams &mgParams, LevelInfo &levelInfo, Matrix &FineMat, Matrix &SmootherMat) {
+
+#define CASE_FOR_N_LEVELS(nLevels)                                                                                     \
+  case nLevels:                                                                                                        \
+    return std::unique_ptr<NLevelMGPreconditioner<Fobj, CComplex, nBasis, nLevels, Matrix>>(                           \
+      new NLevelMGPreconditioner<Fobj, CComplex, nBasis, nLevels, Matrix>(mgParams, levelInfo, FineMat, SmootherMat)); \
+    break;
+
+  switch(mgParams.nLevels) {
+    CASE_FOR_N_LEVELS(2);
+    CASE_FOR_N_LEVELS(3);
+    CASE_FOR_N_LEVELS(4);
+    default:
+      std::cout << GridLogError << "We currently only support nLevels ∈ {2, 3, 4}" << std::endl;
+      exit(EXIT_FAILURE);
+      break;
+  }
+#undef CASE_FOR_N_LEVELS
+}
+
+}
+#endif

tests/solver/Test_staggered_cagmres_unprec.cc

@@ -0,0 +1,72 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_staggered_cagmres_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField;
+  typedef typename ImprovedStaggeredFermionR::ComplexField ComplexField;
+  typename ImprovedStaggeredFermionR::ImplParams params;
+
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
+
+  MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp(Ds);
+  CommunicationAvoidingGeneralisedMinimalResidual<FermionField> CAGMRES(1.0e-8, 10000, 25);
+  CAGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_staggered_fcagmres_prec.cc

@@ -0,0 +1,75 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_staggered_fcagmres_prec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField;
+  typedef typename ImprovedStaggeredFermionR::ComplexField ComplexField;
+  typename ImprovedStaggeredFermionR::ImplParams params;
+
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
+
+  MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp(Ds);
+
+  TrivialPrecon<FermionField> simple;
+
+  FlexibleCommunicationAvoidingGeneralisedMinimalResidual<FermionField> FCAGMRES(1.0e-8, 10000, simple, 25);
+  FCAGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_staggered_fgmres_prec.cc

@@ -0,0 +1,75 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_staggered_fgmres_prec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField;
+  typedef typename ImprovedStaggeredFermionR::ComplexField ComplexField;
+  typename ImprovedStaggeredFermionR::ImplParams params;
+
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
+
+  MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp(Ds);
+
+  TrivialPrecon<FermionField> simple;
+
+  FlexibleGeneralisedMinimalResidual<FermionField> FGMRES(1.0e-8, 10000, simple, 25);
+  FGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_staggered_gmres_unprec.cc

@@ -0,0 +1,72 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_staggered_gmres_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField;
+  typedef typename ImprovedStaggeredFermionR::ComplexField ComplexField;
+  typename ImprovedStaggeredFermionR::ImplParams params;
+
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
+
+  MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp(Ds);
+  GeneralisedMinimalResidual<FermionField> GMRES(1.0e-8, 10000, 25);
+  GMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_staggered_mr_unprec.cc

@@ -0,0 +1,72 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_staggered_mr_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  typedef typename ImprovedStaggeredFermionR::FermionField FermionField;
+  typedef typename ImprovedStaggeredFermionR::ComplexField ComplexField;
+  typename ImprovedStaggeredFermionR::ImplParams params;
+
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  RealD c1=9.0/8.0;
+  RealD c2=-1.0/24.0;
+  RealD u0=1.0;
+  ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0);
+
+  MdagMLinearOperator<ImprovedStaggeredFermionR,FermionField> HermOp(Ds);
+  MinimalResidual<FermionField> MR(1.0e-8,10000,0.8);
+  MR(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilson_cagmres_unprec.cc

@@ -0,0 +1,65 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilson_cagmres_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  LatticeFermion src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  LatticeFermion result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  WilsonFermionR Dw(Umu,Grid,RBGrid,mass);
+
+  MdagMLinearOperator<WilsonFermionR,LatticeFermion> HermOp(Dw);
+  CommunicationAvoidingGeneralisedMinimalResidual<LatticeFermion> CAGMRES(1.0e-8, 10000, 25);
+  CAGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilson_fcagmres_prec.cc

@@ -0,0 +1,68 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilson_fcagmres_prec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  LatticeFermion src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  LatticeFermion result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  WilsonFermionR Dw(Umu,Grid,RBGrid,mass);
+
+  MdagMLinearOperator<WilsonFermionR,LatticeFermion> HermOp(Dw);
+
+  TrivialPrecon<LatticeFermion> simple;
+
+  FlexibleCommunicationAvoidingGeneralisedMinimalResidual<LatticeFermion> FCAGMRES(1.0e-8, 10000, simple, 25);
+  FCAGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilson_fgmres_prec.cc

@@ -0,0 +1,68 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilson_fgmres_prec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  LatticeFermion src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  LatticeFermion result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  WilsonFermionR Dw(Umu,Grid,RBGrid,mass);
+
+  MdagMLinearOperator<WilsonFermionR,LatticeFermion> HermOp(Dw);
+
+  TrivialPrecon<LatticeFermion> simple;
+
+  FlexibleGeneralisedMinimalResidual<LatticeFermion> FGMRES(1.0e-8, 10000, simple, 25);
+  FGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilson_gmres_unprec.cc

@@ -0,0 +1,65 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilson_gmres_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  LatticeFermion src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  LatticeFermion result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  WilsonFermionR Dw(Umu,Grid,RBGrid,mass);
+
+  MdagMLinearOperator<WilsonFermionR,LatticeFermion> HermOp(Dw);
+  GeneralisedMinimalResidual<LatticeFermion> GMRES(1.0e-8, 10000, 25);
+  GMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilson_mg.cc

@@ -0,0 +1,114 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/solver/Test_wilson_mg.cc
+
+    Copyright (C) 2015-2018
+
+    Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+    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 <Grid/Grid.h>
+#include <Test_multigrid_common.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main(int argc, char **argv) {
+
+  Grid_init(&argc, &argv);
+
+  GridCartesian *        FGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplex::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid);
+
+  std::vector<int> fSeeds({1, 2, 3, 4});
+  GridParallelRNG  fPRNG(FGrid);
+  fPRNG.SeedFixedIntegers(fSeeds);
+
+  // clang-format off
+  LatticeFermion    src(FGrid); gaussian(fPRNG, src);
+  LatticeFermion result(FGrid); result = zero;
+  LatticeGaugeField Umu(FGrid); SU3::HotConfiguration(fPRNG, Umu);
+  // clang-format on
+
+  RealD mass = -0.25;
+
+  MultiGridParams mgParams;
+  std::string     inputXml{"./mg_params.xml"};
+
+  if(GridCmdOptionExists(argv, argv + argc, "--inputxml")) {
+    inputXml = GridCmdOptionPayload(argv, argv + argc, "--inputxml");
+    assert(inputXml.length() != 0);
+  }
+
+  {
+    XmlWriter writer("mg_params_template.xml");
+    write(writer, "Params", mgParams);
+    std::cout << GridLogMessage << "Written mg_params_template.xml" << std::endl;
+
+    XmlReader reader(inputXml);
+    read(reader, "Params", mgParams);
+    std::cout << GridLogMessage << "Read in " << inputXml << std::endl;
+  }
+
+  checkParameterValidity(mgParams);
+  std::cout << mgParams << std::endl;
+
+  LevelInfo levelInfo(FGrid, mgParams);
+
+  // Note: We do chiral doubling, so actually only nbasis/2 full basis vectors are used
+  const int nbasis = 40;
+
+  WilsonFermionR Dw(Umu, *FGrid, *FrbGrid, mass);
+
+  MdagMLinearOperator<WilsonFermionR, LatticeFermion> MdagMOpDw(Dw);
+
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+  std::cout << GridLogMessage << "Testing Multigrid for Wilson" << std::endl;
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+  TrivialPrecon<LatticeFermion> TrivialPrecon;
+  auto MGPreconDw = createMGInstance<vSpinColourVector, vTComplex, nbasis, WilsonFermionR>(mgParams, levelInfo, Dw, Dw);
+
+  MGPreconDw->setup();
+
+  if(GridCmdOptionExists(argv, argv + argc, "--runchecks")) {
+    RealD toleranceForMGChecks = (getPrecision<LatticeFermion>::value == 1) ? 1e-6 : 1e-13;
+    MGPreconDw->runChecks(toleranceForMGChecks);
+  }
+
+  std::vector<std::unique_ptr<OperatorFunction<LatticeFermion>>> solversDw;
+
+  solversDw.emplace_back(new ConjugateGradient<LatticeFermion>(1.0e-12, 50000, false));
+  solversDw.emplace_back(new FlexibleGeneralisedMinimalResidual<LatticeFermion>(1.0e-12, 50000, TrivialPrecon, 100, false));
+  solversDw.emplace_back(new FlexibleGeneralisedMinimalResidual<LatticeFermion>(1.0e-12, 50000, *MGPreconDw, 100, false));
+
+  for(auto const &solver : solversDw) {
+    std::cout << std::endl << "Starting with a new solver" << std::endl;
+    result = zero;
+    (*solver)(MdagMOpDw, src, result);
+  }
+
+  MGPreconDw->reportTimings();
+
+  Grid_finalize();
+}

tests/solver/Test_wilson_mg_mp.cc

@@ -0,0 +1,166 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/solver/Test_wilson_mg_mp.cc
+
+    Copyright (C) 2015-2018
+
+    Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+    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 <Grid/Grid.h>
+#include <Test_multigrid_common.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main(int argc, char **argv) {
+
+  Grid_init(&argc, &argv);
+
+  // clang-format off
+  GridCartesian         *FGrid_d   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexD::Nsimd()), GridDefaultMpi());
+  GridCartesian         *FGrid_f   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian *FrbGrid_d = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid_d);
+  GridRedBlackCartesian *FrbGrid_f = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid_f);
+  // clang-format on
+
+  std::vector<int> fSeeds({1, 2, 3, 4});
+  GridParallelRNG  fPRNG(FGrid_d);
+  fPRNG.SeedFixedIntegers(fSeeds);
+
+  // clang-format off
+  LatticeFermionD       src_d(FGrid_d); gaussian(fPRNG, src_d);
+  LatticeFermionD resultMGD_d(FGrid_d); resultMGD_d = zero;
+  LatticeFermionD resultMGF_d(FGrid_d); resultMGF_d = zero;
+  LatticeGaugeFieldD    Umu_d(FGrid_d); SU3::HotConfiguration(fPRNG, Umu_d);
+  LatticeGaugeFieldF    Umu_f(FGrid_f); precisionChange(Umu_f, Umu_d);
+  // clang-format on
+
+  RealD mass = -0.25;
+
+  MultiGridParams mgParams;
+  std::string     inputXml{"./mg_params.xml"};
+
+  if(GridCmdOptionExists(argv, argv + argc, "--inputxml")) {
+    inputXml = GridCmdOptionPayload(argv, argv + argc, "--inputxml");
+    assert(inputXml.length() != 0);
+  }
+
+  {
+    XmlWriter writer("mg_params_template.xml");
+    write(writer, "Params", mgParams);
+    std::cout << GridLogMessage << "Written mg_params_template.xml" << std::endl;
+
+    XmlReader reader(inputXml);
+    read(reader, "Params", mgParams);
+    std::cout << GridLogMessage << "Read in " << inputXml << std::endl;
+  }
+
+  checkParameterValidity(mgParams);
+  std::cout << mgParams << std::endl;
+
+  LevelInfo levelInfo_d(FGrid_d, mgParams);
+  LevelInfo levelInfo_f(FGrid_f, mgParams);
+
+  // Note: We do chiral doubling, so actually only nbasis/2 full basis vectors are used
+  const int nbasis = 40;
+
+  WilsonFermionD Dw_d(Umu_d, *FGrid_d, *FrbGrid_d, mass);
+  WilsonFermionF Dw_f(Umu_f, *FGrid_f, *FrbGrid_f, mass);
+
+  MdagMLinearOperator<WilsonFermionD, LatticeFermionD> MdagMOpDw_d(Dw_d);
+  MdagMLinearOperator<WilsonFermionF, LatticeFermionF> MdagMOpDw_f(Dw_f);
+
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+  std::cout << GridLogMessage << "Testing single-precision Multigrid for Wilson" << std::endl;
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+  auto MGPreconDw_f = createMGInstance<vSpinColourVectorF, vTComplexF, nbasis, WilsonFermionF>(mgParams, levelInfo_f, Dw_f, Dw_f);
+
+  MGPreconDw_f->setup();
+
+  if(GridCmdOptionExists(argv, argv + argc, "--runchecks")) {
+    MGPreconDw_f->runChecks(1e-6);
+  }
+
+  MixedPrecisionFlexibleGeneralisedMinimalResidual<LatticeFermionD, LatticeFermionF> MPFGMRESPREC(1.0e-12, 50000, FGrid_f, *MGPreconDw_f, 100, false);
+
+  std::cout << std::endl << "Starting with a new solver" << std::endl;
+  MPFGMRESPREC(MdagMOpDw_d, src_d, resultMGF_d);
+
+  MGPreconDw_f->reportTimings();
+
+  if(GridCmdOptionExists(argv, argv + argc, "--docomparison")) {
+
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+    std::cout << GridLogMessage << "Testing double-precision Multigrid for Wilson" << std::endl;
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+    auto MGPreconDw_d = createMGInstance<vSpinColourVectorD, vTComplexD, nbasis, WilsonFermionD>(mgParams, levelInfo_d, Dw_d, Dw_d);
+
+    MGPreconDw_d->setup();
+
+    if(GridCmdOptionExists(argv, argv + argc, "--runchecks")) {
+      MGPreconDw_d->runChecks(1e-13);
+    }
+
+    FlexibleGeneralisedMinimalResidual<LatticeFermionD> FGMRESPREC(1.0e-12, 50000, *MGPreconDw_d, 100, false);
+
+    std::cout << std::endl << "Starting with a new solver" << std::endl;
+    FGMRESPREC(MdagMOpDw_d, src_d, resultMGD_d);
+
+    MGPreconDw_d->reportTimings();
+
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+    std::cout << GridLogMessage << "Comparing single-precision Multigrid with double-precision one for Wilson" << std::endl;
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+    LatticeFermionD diffFullSolver(FGrid_d);
+
+    RealD deviationFullSolver = axpy_norm(diffFullSolver, -1.0, resultMGF_d, resultMGD_d);
+
+    // clang-format off
+    LatticeFermionF src_f(FGrid_f);    precisionChange(src_f, src_d);
+    LatticeFermionF resMGF_f(FGrid_f); resMGF_f = zero;
+    LatticeFermionD resMGD_d(FGrid_d); resMGD_d = zero;
+    // clang-format on
+
+    (*MGPreconDw_f)(src_f, resMGF_f);
+    (*MGPreconDw_d)(src_d, resMGD_d);
+
+    LatticeFermionD diffOnlyMG(FGrid_d);
+    LatticeFermionD resMGF_d(FGrid_d);
+    precisionChange(resMGF_d, resMGF_f);
+
+    RealD deviationOnlyPrec = axpy_norm(diffOnlyMG, -1.0, resMGF_d, resMGD_d);
+
+    // clang-format off
+    std::cout << GridLogMessage << "Absolute difference between FGMRES preconditioned by double and single precicision MG: " << deviationFullSolver                      << std::endl;
+    std::cout << GridLogMessage << "Relative deviation  between FGMRES preconditioned by double and single precicision MG: " << deviationFullSolver / norm2(resultMGD_d) << std::endl;
+    std::cout << GridLogMessage << "Absolute difference between one iteration of MG Prec in double and single precision:   " << deviationOnlyPrec                        << std::endl;
+    std::cout << GridLogMessage << "Relative deviation  between one iteration of MG Prec in double and single precision:   " << deviationOnlyPrec / norm2(resMGD_d)      << std::endl;
+    // clang-format on
+  }
+
+  Grid_finalize();
+}

tests/solver/Test_wilson_mr_unprec.cc

@@ -0,0 +1,65 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilson_mr_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  LatticeFermion src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  LatticeFermion result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass=0.5;
+  WilsonFermionR Dw(Umu,Grid,RBGrid,mass);
+
+  MdagMLinearOperator<WilsonFermionR,LatticeFermion> HermOp(Dw);
+  MinimalResidual<LatticeFermion> MR(1.0e-8,10000,0.8);
+  MR(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilsonclover_cagmres_unprec.cc

@@ -0,0 +1,71 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilsonclover_cagmres_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  typedef typename WilsonCloverFermionR::FermionField FermionField;
+  typename WilsonCloverFermionR::ImplParams params;
+  WilsonAnisotropyCoefficients anis;
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass  = 0.5;
+  RealD csw_r = 1.0;
+  RealD csw_t = 1.0;
+  WilsonCloverFermionR Dwc(Umu,Grid,RBGrid,mass,csw_r,csw_t,anis,params);
+
+  MdagMLinearOperator<WilsonCloverFermionR,FermionField> HermOp(Dwc);
+  CommunicationAvoidingGeneralisedMinimalResidual<FermionField> CAGMRES(1.0e-8, 10000, 25);
+  CAGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilsonclover_fcagmres_prec.cc

@@ -0,0 +1,74 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilsonclover_fcagmres_prec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  typedef typename WilsonCloverFermionR::FermionField FermionField;
+  typename WilsonCloverFermionR::ImplParams params;
+  WilsonAnisotropyCoefficients anis;
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass  = 0.5;
+  RealD csw_r = 1.0;
+  RealD csw_t = 1.0;
+  WilsonCloverFermionR Dwc(Umu,Grid,RBGrid,mass,csw_r,csw_t,anis,params);
+
+  MdagMLinearOperator<WilsonCloverFermionR,FermionField> HermOp(Dwc);
+
+  TrivialPrecon<FermionField> simple;
+
+  FlexibleCommunicationAvoidingGeneralisedMinimalResidual<FermionField> FCAGMRES(1.0e-8, 10000, simple, 25);
+  FCAGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilsonclover_fgmres_prec.cc

@@ -0,0 +1,74 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilsonclover_fgmres_prec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  typedef typename WilsonCloverFermionR::FermionField FermionField;
+  typename WilsonCloverFermionR::ImplParams params;
+  WilsonAnisotropyCoefficients anis;
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass  = 0.5;
+  RealD csw_r = 1.0;
+  RealD csw_t = 1.0;
+  WilsonCloverFermionR Dwc(Umu,Grid,RBGrid,mass,csw_r,csw_t,anis,params);
+
+  MdagMLinearOperator<WilsonCloverFermionR,FermionField> HermOp(Dwc);
+
+  TrivialPrecon<FermionField> simple;
+
+  FlexibleGeneralisedMinimalResidual<FermionField> FGMRES(1.0e-8, 10000, simple, 25);
+  FGMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilsonclover_gmres_unprec.cc

@@ -0,0 +1,71 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilsonclover_gmres_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  typedef typename WilsonCloverFermionR::FermionField FermionField;
+  typename WilsonCloverFermionR::ImplParams params;
+  WilsonAnisotropyCoefficients anis;
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass  = 0.5;
+  RealD csw_r = 1.0;
+  RealD csw_t = 1.0;
+  WilsonCloverFermionR Dwc(Umu,Grid,RBGrid,mass,csw_r,csw_t,anis,params);
+
+  MdagMLinearOperator<WilsonCloverFermionR,FermionField> HermOp(Dwc);
+  GeneralisedMinimalResidual<FermionField> GMRES(1.0e-8, 10000, 25);
+  GMRES(HermOp,src,result);
+
+  Grid_finalize();
+}

tests/solver/Test_wilsonclover_mg.cc

@@ -0,0 +1,117 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/solver/Test_wilsonclover_mg.cc
+
+    Copyright (C) 2015-2018
+
+    Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+    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 <Grid/Grid.h>
+#include <Test_multigrid_common.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main(int argc, char **argv) {
+
+  Grid_init(&argc, &argv);
+
+  GridCartesian *        FGrid   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplex::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid);
+
+  std::vector<int> fSeeds({1, 2, 3, 4});
+  GridParallelRNG  fPRNG(FGrid);
+  fPRNG.SeedFixedIntegers(fSeeds);
+
+  // clang-format off
+  LatticeFermion    src(FGrid); gaussian(fPRNG, src);
+  LatticeFermion result(FGrid); result = zero;
+  LatticeGaugeField Umu(FGrid); SU3::HotConfiguration(fPRNG, Umu);
+  // clang-format on
+
+  RealD mass  = -0.25;
+  RealD csw_r = 1.0;
+  RealD csw_t = 1.0;
+
+  MultiGridParams mgParams;
+  std::string     inputXml{"./mg_params.xml"};
+
+  if(GridCmdOptionExists(argv, argv + argc, "--inputxml")) {
+    inputXml = GridCmdOptionPayload(argv, argv + argc, "--inputxml");
+    assert(inputXml.length() != 0);
+  }
+
+  {
+    XmlWriter writer("mg_params_template.xml");
+    write(writer, "Params", mgParams);
+    std::cout << GridLogMessage << "Written mg_params_template.xml" << std::endl;
+
+    XmlReader reader(inputXml);
+    read(reader, "Params", mgParams);
+    std::cout << GridLogMessage << "Read in " << inputXml << std::endl;
+  }
+
+  checkParameterValidity(mgParams);
+  std::cout << mgParams << std::endl;
+
+  LevelInfo levelInfo(FGrid, mgParams);
+
+  // Note: We do chiral doubling, so actually only nbasis/2 full basis vectors are used
+  const int nbasis = 40;
+
+  WilsonCloverFermionR Dwc(Umu, *FGrid, *FrbGrid, mass, csw_r, csw_t);
+
+  MdagMLinearOperator<WilsonCloverFermionR, LatticeFermion> MdagMOpDwc(Dwc);
+
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+  std::cout << GridLogMessage << "Testing Multigrid for Wilson Clover" << std::endl;
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+  TrivialPrecon<LatticeFermion> TrivialPrecon;
+  auto MGPreconDwc = createMGInstance<vSpinColourVector, vTComplex, nbasis, WilsonCloverFermionR>(mgParams, levelInfo, Dwc, Dwc);
+
+  MGPreconDwc->setup();
+
+  if(GridCmdOptionExists(argv, argv + argc, "--runchecks")) {
+    RealD toleranceForMGChecks = (getPrecision<LatticeFermion>::value == 1) ? 1e-6 : 1e-13;
+    MGPreconDwc->runChecks(toleranceForMGChecks);
+  }
+
+  std::vector<std::unique_ptr<OperatorFunction<LatticeFermion>>> solversDwc;
+
+  solversDwc.emplace_back(new ConjugateGradient<LatticeFermion>(1.0e-12, 50000, false));
+  solversDwc.emplace_back(new FlexibleGeneralisedMinimalResidual<LatticeFermion>(1.0e-12, 50000, TrivialPrecon, 100, false));
+  solversDwc.emplace_back(new FlexibleGeneralisedMinimalResidual<LatticeFermion>(1.0e-12, 50000, *MGPreconDwc, 100, false));
+
+  for(auto const &solver : solversDwc) {
+    std::cout << std::endl << "Starting with a new solver" << std::endl;
+    result = zero;
+    (*solver)(MdagMOpDwc, src, result);
+    std::cout << std::endl;
+  }
+
+  MGPreconDwc->reportTimings();
+
+  Grid_finalize();
+}

tests/solver/Test_wilsonclover_mg_mp.cc

@@ -0,0 +1,169 @@
+/*************************************************************************************
+
+    Grid physics library, www.github.com/paboyle/Grid 
+
+    Source file: ./tests/solver/Test_wilsonclover_mg_mp.cc
+
+    Copyright (C) 2015-2018
+
+    Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+    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 <Grid/Grid.h>
+#include <Test_multigrid_common.h>
+
+using namespace std;
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main(int argc, char **argv) {
+
+  Grid_init(&argc, &argv);
+
+  // clang-format off
+  GridCartesian         *FGrid_d   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexD::Nsimd()), GridDefaultMpi());
+  GridCartesian         *FGrid_f   = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
+  GridRedBlackCartesian *FrbGrid_d = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid_d);
+  GridRedBlackCartesian *FrbGrid_f = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid_f);
+  // clang-format on
+
+  std::vector<int> fSeeds({1, 2, 3, 4});
+  GridParallelRNG  fPRNG(FGrid_d);
+  fPRNG.SeedFixedIntegers(fSeeds);
+
+  // clang-format off
+  LatticeFermionD       src_d(FGrid_d); gaussian(fPRNG, src_d);
+  LatticeFermionD resultMGD_d(FGrid_d); resultMGD_d = zero;
+  LatticeFermionD resultMGF_d(FGrid_d); resultMGF_d = zero;
+  LatticeGaugeFieldD    Umu_d(FGrid_d); SU3::HotConfiguration(fPRNG, Umu_d);
+  LatticeGaugeFieldF    Umu_f(FGrid_f); precisionChange(Umu_f, Umu_d);
+  // clang-format on
+
+  RealD mass  = -0.25;
+  RealD csw_r = 1.0;
+  RealD csw_t = 1.0;
+
+  MultiGridParams mgParams;
+  std::string     inputXml{"./mg_params.xml"};
+
+  if(GridCmdOptionExists(argv, argv + argc, "--inputxml")) {
+    inputXml = GridCmdOptionPayload(argv, argv + argc, "--inputxml");
+    assert(inputXml.length() != 0);
+  }
+
+  {
+    XmlWriter writer("mg_params_template.xml");
+    write(writer, "Params", mgParams);
+    std::cout << GridLogMessage << "Written mg_params_template.xml" << std::endl;
+
+    XmlReader reader(inputXml);
+    read(reader, "Params", mgParams);
+    std::cout << GridLogMessage << "Read in " << inputXml << std::endl;
+  }
+
+  checkParameterValidity(mgParams);
+  std::cout << mgParams << std::endl;
+
+  LevelInfo levelInfo_d(FGrid_d, mgParams);
+  LevelInfo levelInfo_f(FGrid_f, mgParams);
+
+  // Note: We do chiral doubling, so actually only nbasis/2 full basis vectors are used
+  const int nbasis = 40;
+
+  WilsonCloverFermionD Dwc_d(Umu_d, *FGrid_d, *FrbGrid_d, mass, csw_r, csw_t);
+  WilsonCloverFermionF Dwc_f(Umu_f, *FGrid_f, *FrbGrid_f, mass, csw_r, csw_t);
+
+  MdagMLinearOperator<WilsonCloverFermionD, LatticeFermionD> MdagMOpDwc_d(Dwc_d);
+  MdagMLinearOperator<WilsonCloverFermionF, LatticeFermionF> MdagMOpDwc_f(Dwc_f);
+
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+  std::cout << GridLogMessage << "Testing single-precision Multigrid for Wilson Clover" << std::endl;
+  std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+  auto MGPreconDwc_f = createMGInstance<vSpinColourVectorF, vTComplexF, nbasis, WilsonCloverFermionF>(mgParams, levelInfo_f, Dwc_f, Dwc_f);
+
+  MGPreconDwc_f->setup();
+
+  if(GridCmdOptionExists(argv, argv + argc, "--runchecks")) {
+    MGPreconDwc_f->runChecks(1e-6);
+  }
+
+  MixedPrecisionFlexibleGeneralisedMinimalResidual<LatticeFermionD, LatticeFermionF> MPFGMRESPREC(
+    1.0e-12, 50000, FGrid_f, *MGPreconDwc_f, 100, false);
+
+  std::cout << std::endl << "Starting with a new solver" << std::endl;
+  MPFGMRESPREC(MdagMOpDwc_d, src_d, resultMGF_d);
+
+  MGPreconDwc_f->reportTimings();
+
+  if(GridCmdOptionExists(argv, argv + argc, "--docomparison")) {
+
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+    std::cout << GridLogMessage << "Testing double-precision Multigrid for Wilson Clover" << std::endl;
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+    auto MGPreconDwc_d = createMGInstance<vSpinColourVectorD, vTComplexD, nbasis, WilsonCloverFermionD>(mgParams, levelInfo_d, Dwc_d, Dwc_d);
+
+    MGPreconDwc_d->setup();
+
+    if(GridCmdOptionExists(argv, argv + argc, "--runchecks")) {
+      MGPreconDwc_d->runChecks(1e-13);
+    }
+
+    FlexibleGeneralisedMinimalResidual<LatticeFermionD> FGMRESPREC(1.0e-12, 50000, *MGPreconDwc_d, 100, false);
+
+    std::cout << std::endl << "Starting with a new solver" << std::endl;
+    FGMRESPREC(MdagMOpDwc_d, src_d, resultMGD_d);
+
+    MGPreconDwc_d->reportTimings();
+
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+    std::cout << GridLogMessage << "Comparing single-precision Multigrid with double-precision one for Wilson Clover" << std::endl;
+    std::cout << GridLogMessage << "**************************************************" << std::endl;
+
+    LatticeFermionD diffFullSolver(FGrid_d);
+
+    RealD deviationFullSolver = axpy_norm(diffFullSolver, -1.0, resultMGF_d, resultMGD_d);
+
+    // clang-format off
+    LatticeFermionF src_f(FGrid_f);    precisionChange(src_f, src_d);
+    LatticeFermionF resMGF_f(FGrid_f); resMGF_f = zero;
+    LatticeFermionD resMGD_d(FGrid_d); resMGD_d = zero;
+    // clang-format on
+
+    (*MGPreconDwc_f)(src_f, resMGF_f);
+    (*MGPreconDwc_d)(src_d, resMGD_d);
+
+    LatticeFermionD diffOnlyMG(FGrid_d);
+    LatticeFermionD resMGF_d(FGrid_d);
+    precisionChange(resMGF_d, resMGF_f);
+
+    RealD deviationOnlyPrec = axpy_norm(diffOnlyMG, -1.0, resMGF_d, resMGD_d);
+
+    // clang-format off
+    std::cout << GridLogMessage << "Absolute difference between FGMRES preconditioned by double and single precicision MG: " << deviationFullSolver                      << std::endl;
+    std::cout << GridLogMessage << "Relative deviation  between FGMRES preconditioned by double and single precicision MG: " << deviationFullSolver / norm2(resultMGD_d) << std::endl;
+    std::cout << GridLogMessage << "Absolute difference between one iteration of MG Prec in double and single precision:   " << deviationOnlyPrec                        << std::endl;
+    std::cout << GridLogMessage << "Relative deviation  between one iteration of MG Prec in double and single precision:   " << deviationOnlyPrec / norm2(resMGD_d)      << std::endl;
+    // clang-format on
+  }
+
+  Grid_finalize();
+}

tests/solver/Test_wilsonclover_mr_unprec.cc

@@ -0,0 +1,71 @@
+/*************************************************************************************
+
+Grid physics library, www.github.com/paboyle/Grid
+
+Source file: ./tests/solver/Test_wilsonclover_mr_unprec.cc
+
+Copyright (C) 2015-2018
+
+Author: Daniel Richtmann <daniel.richtmann@ur.de>
+
+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 <Grid/Grid.h>
+
+using namespace Grid;
+using namespace Grid::QCD;
+
+int main (int argc, char ** argv)
+{
+  Grid_init(&argc,&argv);
+
+  std::vector<int> latt_size   = GridDefaultLatt();
+  std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
+  std::vector<int> mpi_layout  = GridDefaultMpi();
+  GridCartesian               Grid(latt_size,simd_layout,mpi_layout);
+  GridRedBlackCartesian     RBGrid(&Grid);
+
+  std::vector<int> seeds({1,2,3,4});
+  GridParallelRNG          pRNG(&Grid);  pRNG.SeedFixedIntegers(seeds);
+
+  typedef typename WilsonCloverFermionR::FermionField FermionField;
+  typename WilsonCloverFermionR::ImplParams params;
+  WilsonAnisotropyCoefficients anis;
+
+  FermionField src(&Grid); random(pRNG,src);
+  RealD nrm = norm2(src);
+  FermionField result(&Grid); result=zero;
+  LatticeGaugeField Umu(&Grid); SU3::HotConfiguration(pRNG,Umu);
+
+  double volume=1;
+  for(int mu=0;mu<Nd;mu++){
+    volume=volume*latt_size[mu];
+  }
+
+  RealD mass  = 0.5;
+  RealD csw_r = 1.0;
+  RealD csw_t = 1.0;
+  WilsonCloverFermionR Dwc(Umu,Grid,RBGrid,mass,csw_r,csw_t,anis,params);
+
+  MdagMLinearOperator<WilsonCloverFermionR,FermionField> HermOp(Dwc);
+  MinimalResidual<FermionField> MR(1.0e-8,10000,0.8);
+  MR(HermOp,src,result);
+
+  Grid_finalize();
+}