Added support for non-periodic gauge field implementations in the random gauge shift performed at the start of the HMC trajectory

(The above required exposing the gauge implementation to the HMC class through the Integrator class)
Made the random shift optional (default on) through a parameter in HMCparameters
Modified ConjugateBC::CshiftLink such that it supports any shift in  -L < shift < L rather than just +-1
Added a tester for the BC-respecting Cshift
Fixed a missing system header include in SSE4 intrinsics wrapper
Fixed sumD_cpu for single-prec types performing an incorrect conversion to a single-prec data type at the end, that fails to compile on some systems

Grid/lattice/Lattice_reduction.h

@@ -91,10 +91,7 @@ inline typename vobj::scalar_objectD sumD_cpu(const vobj *arg, Integer osites)
   for(int i=0;i<nthread;i++){
     ssum = ssum+sumarray[i];
   } 
-  
-  typedef typename vobj::scalar_object ssobj;
-  ssobj ret = ssum;
-  return ret;
+  return ssum;
 }
 /*
 Threaded max, don't use for now

Grid/qcd/hmc/HMC.h

@@ -53,6 +53,7 @@ struct HMCparameters: Serializable {
                                   Integer, Trajectories, /* @brief Number of sweeps in this run */
                                   bool, MetropolisTest,
                                   Integer, NoMetropolisUntil,
+				  bool, PerformRandomShift, /* @brief Randomly shift the gauge configuration at the start of a trajectory */
                                   std::string, StartingType,
                                   IntegratorParameters, MD)
 
@@ -63,6 +64,7 @@ struct HMCparameters: Serializable {
     StartTrajectory   = 0;
     Trajectories      = 10;
     StartingType      = "HotStart";
+    PerformRandomShift = true;
     /////////////////////////////////
   }
 
@@ -83,6 +85,7 @@ struct HMCparameters: Serializable {
     std::cout << GridLogMessage << "[HMC parameters] Start trajectory        : " << StartTrajectory << "\n";
     std::cout << GridLogMessage << "[HMC parameters] Metropolis test (on/off): " << std::boolalpha << MetropolisTest << "\n";
     std::cout << GridLogMessage << "[HMC parameters] Thermalization trajs    : " << NoMetropolisUntil << "\n";
+    std::cout << GridLogMessage << "[HMC parameters] Doing random shift      : " << std::boolalpha << PerformRandomShift << "\n";
     std::cout << GridLogMessage << "[HMC parameters] Starting type           : " << StartingType << "\n";
     MD.print_parameters();
   }
@@ -95,6 +98,7 @@ private:
   const HMCparameters Params;
 
   typedef typename IntegratorType::Field Field;
+  typedef typename IntegratorType::FieldImplementation FieldImplementation;
   typedef std::vector< HmcObservable<Field> * > ObsListType;
 
   //pass these from the resource manager
@@ -138,26 +142,37 @@ private:
 
     GridBase *Grid = U.Grid();
 
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    // Mainly for DDHMC perform a random translation of U modulo volume
-    //////////////////////////////////////////////////////////////////////////////////////////////////////
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
-    std::cout << GridLogMessage << "Random shifting gauge field by [";
-    for(int d=0;d<Grid->Nd();d++) {
+    if(Params.PerformRandomShift){
+      //////////////////////////////////////////////////////////////////////////////////////////////////////
+      // Mainly for DDHMC perform a random translation of U modulo volume
+      //////////////////////////////////////////////////////////////////////////////////////////////////////
+      std::cout << GridLogMessage << "--------------------------------------------------\n";
+      std::cout << GridLogMessage << "Random shifting gauge field by [";
 
-      int L = Grid->GlobalDimensions()[d];
+      std::vector<typename FieldImplementation::GaugeLinkField> Umu(Grid->Nd(), U.Grid());
+      for(int mu=0;mu<Grid->Nd();mu++) Umu[mu] = PeekIndex<LorentzIndex>(U, mu);
 
-      RealD rn_uniform;  random(sRNG, rn_uniform);
+      for(int d=0;d<Grid->Nd();d++) {
 
-      int shift = (int) (rn_uniform*L);
+	int L = Grid->GlobalDimensions()[d];
 
-      std::cout << shift;
-      if(d<Grid->Nd()-1) std::cout <<",";
-      else               std::cout <<"]\n";
+	RealD rn_uniform;  random(sRNG, rn_uniform);
+
+	int shift = (int) (rn_uniform*L);
+
+	std::cout << shift;
+	if(d<Grid->Nd()-1) std::cout <<",";
+	else               std::cout <<"]\n";
       
-      U = Cshift(U,d,shift);
+	//shift all fields together in a way that respects the gauge BCs
+	for(int mu=0; mu < Grid->Nd(); mu++)
+	  Umu[mu] = FieldImplementation::CshiftLink(Umu[mu],d,shift);
+      }
+
+      for(int mu=0;mu<Grid->Nd();mu++) PokeIndex<LorentzIndex>(U,Umu[mu],mu);
+      
+      std::cout << GridLogMessage << "--------------------------------------------------\n";
     }
-    std::cout << GridLogMessage << "--------------------------------------------------\n";
 
     TheIntegrator.reset_timer();
     

Grid/qcd/hmc/integrators/Integrator.h

@@ -63,10 +63,10 @@ public:
 };
 
 /*! @brief Class for Molecular Dynamics management */
-template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy>
+template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy>
 class Integrator {
 protected:
-
+  typedef FieldImplementation_ FieldImplementation;
   typedef typename FieldImplementation::Field MomentaField;  //for readability
   typedef typename FieldImplementation::Field Field;
 

Grid/qcd/hmc/integrators/Integrator_algorithm.h

@@ -92,10 +92,11 @@ NAMESPACE_BEGIN(Grid);
  *  P 1/2                            P 1/2
  */
 
-template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
-class LeapFrog : public Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy> 
+template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
+class LeapFrog : public Integrator<FieldImplementation_, SmearingPolicy, RepresentationPolicy> 
 {
 public:
+  typedef FieldImplementation_ FieldImplementation;
   typedef LeapFrog<FieldImplementation, SmearingPolicy, RepresentationPolicy> Algorithm;
   INHERIT_FIELD_TYPES(FieldImplementation);
 
@@ -135,13 +136,14 @@ public:
   }
 };
 
-template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
-class MinimumNorm2 : public Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy> 
+template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
+class MinimumNorm2 : public Integrator<FieldImplementation_, SmearingPolicy, RepresentationPolicy> 
 {
 private:
   const RealD lambda = 0.1931833275037836;
 
 public:
+  typedef FieldImplementation_ FieldImplementation;
   INHERIT_FIELD_TYPES(FieldImplementation);
 
   MinimumNorm2(GridBase* grid, IntegratorParameters Par, ActionSet<Field, RepresentationPolicy>& Aset, SmearingPolicy& Sm)
@@ -192,8 +194,8 @@ public:
   }
 };
 
-template <class FieldImplementation, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
-class ForceGradient : public Integrator<FieldImplementation, SmearingPolicy, RepresentationPolicy> 
+template <class FieldImplementation_, class SmearingPolicy, class RepresentationPolicy = Representations<FundamentalRepresentation> >
+class ForceGradient : public Integrator<FieldImplementation_, SmearingPolicy, RepresentationPolicy> 
 {
 private:
   const RealD lambda = 1.0 / 6.0;
@@ -202,6 +204,7 @@ private:
   const RealD theta = 0.0;
 
 public:
+  typedef FieldImplementation_ FieldImplementation;
   INHERIT_FIELD_TYPES(FieldImplementation);
 
   // Looks like dH scales as dt^4. tested wilson/wilson 2 level.

Grid/qcd/utils/CovariantCshift.h

@@ -227,26 +227,38 @@ namespace ConjugateBC {
   //shift = -1
   //Out(x) = U_\mu(x-mu)  | x_\mu != 0
   //       = U*_\mu(L-1)  | x_\mu == 0
+  //shift = 2
+  //Out(x) = U_\mu(x+2\hat\mu)  | x_\mu < L-2
+  //       = U*_\mu(1)  | x_\mu == L-1
+  //       = U*_\mu(0)  | x_\mu == L-2
+  //shift = -2
+  //Out(x) = U_\mu(x-2mu)  | x_\mu > 1
+  //       = U*_\mu(L-2)  | x_\mu == 0
+  //       = U*_\mu(L-1)  | x_\mu == 1
+  //etc
   template<class gauge> Lattice<gauge>
   CshiftLink(const Lattice<gauge> &Link, int mu, int shift)
   {
     GridBase *grid = Link.Grid();
-    int Lmu = grid->GlobalDimensions()[mu] - 1;
+    int Lmu = grid->GlobalDimensions()[mu];
+    assert(abs(shift) < Lmu && "Invalid shift value");
 
     Lattice<iScalar<vInteger>> coor(grid);
     LatticeCoordinate(coor, mu);
 
     Lattice<gauge> tmp(grid);
-    if(shift == 1){
-      tmp = Cshift(Link, mu, 1);
-      tmp = where(coor == Lmu, conjugate(tmp), tmp);
+    if(shift > 0){
+      tmp = Cshift(Link, mu, shift);
+      tmp = where(coor >= Lmu-shift, conjugate(tmp), tmp);
       return tmp;
-    }else if(shift == -1){
+    }else if(shift < 0){
       tmp = Link;
-      tmp = where(coor == Lmu, conjugate(tmp), tmp);
-      return Cshift(tmp, mu, -1);
-    }else assert(0 && "Invalid shift value");
-    return tmp; //shuts up the compiler fussing about the return type
+      tmp = where(coor >= Lmu+shift, conjugate(tmp), tmp);
+      return Cshift(tmp, mu, shift);
+    }
+    
+    //shift == 0
+    return Link;
   }
 
 }

Grid/qcd/utils/GaugeFix.h

@@ -72,12 +72,12 @@ public:
 
   //Fix the gauge field Umu
   //0 < alpha < 1 is related to the step size, cf https://arxiv.org/pdf/1405.5812.pdf
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,Real alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
     GridBase *grid = Umu.Grid();
     GaugeMat xform(grid);
     SteepestDescentGaugeFix(Umu,xform,alpha,maxiter,Omega_tol,Phi_tol,Fourier,orthog,err_on_no_converge);
   }
-  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real & alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
+  static void SteepestDescentGaugeFix(GaugeLorentz &Umu,GaugeMat &xform,Real alpha,int maxiter,Real Omega_tol, Real Phi_tol,bool Fourier=false,int orthog=-1,bool err_on_no_converge=true) {
   //Fix the gauge field Umu and also return the gauge transformation from the original gauge field, xform
 
     GridBase *grid = Umu.Grid();

Grid/simd/Grid_sse4.h

@@ -35,7 +35,7 @@ Author: neo <cossu@post.kek.jp>
 */
 // Time-stamp: <2015-06-16 23:27:54 neo>
 //----------------------------------------------------------------------
-
+#include <immintrin.h>
 #include <pmmintrin.h>
 
 NAMESPACE_BEGIN(Grid);