Improve the matching of stencil coarsening

Grid/algorithms/LinearOperator.h

@@ -52,6 +52,9 @@ public:
   virtual void AdjOp  (const Field &in, Field &out) = 0; // Abstract base
   virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2)=0;
   virtual void HermOp(const Field &in, Field &out)=0;
+
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 };
 
 
@@ -76,6 +79,9 @@ class MdagMLinearOperator : public LinearOperatorBase<Field> {
 public:
   MdagMLinearOperator(Matrix &Mat): _Mat(Mat){};
 
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
+
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
     _Mat.Mdiag(in,out);
@@ -111,6 +117,8 @@ class ShiftedMdagMLinearOperator : public LinearOperatorBase<Field> {
   Matrix &_Mat;
   RealD _shift;
 public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   ShiftedMdagMLinearOperator(Matrix &Mat,RealD shift): _Mat(Mat), _shift(shift){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
@@ -151,6 +159,8 @@ template<class Matrix,class Field>
 class HermitianLinearOperator : public LinearOperatorBase<Field> {
   Matrix &_Mat;
 public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   HermitianLinearOperator(Matrix &Mat): _Mat(Mat){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
@@ -182,6 +192,8 @@ template<class Matrix,class Field>
 class NonHermitianLinearOperator : public LinearOperatorBase<Field> {
   Matrix &_Mat;
 public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   NonHermitianLinearOperator(Matrix &Mat): _Mat(Mat){};
   // Support for coarsening to a multigrid
   void OpDiag (const Field &in, Field &out) {
@@ -255,6 +267,8 @@ template<class Matrix,class Field>
   class SchurDiagMooeeOperator :  public SchurOperatorBase<Field> {
  public:
     Matrix &_Mat;
+    virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+    virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
     SchurDiagMooeeOperator (Matrix &Mat): _Mat(Mat){};
     virtual  void Mpc      (const Field &in, Field &out) {
       Field tmp(in.Grid());
@@ -281,6 +295,8 @@ template<class Matrix,class Field>
  protected:
     Matrix &_Mat;
  public:
+    virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+    virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
     SchurDiagOneOperator (Matrix &Mat): _Mat(Mat){};
     
     virtual void Mpc      (const Field &in, Field &out) {
@@ -307,6 +323,8 @@ template<class Matrix,class Field>
  protected:
     Matrix &_Mat;
  public:
+    virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+    virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
     SchurDiagTwoOperator (Matrix &Mat): _Mat(Mat){};
     
     virtual void Mpc      (const Field &in, Field &out) {
@@ -372,6 +390,8 @@ class NonHermitianSchurDiagMooeeOperator :  public NonHermitianSchurOperatorBase
 {
  public:
   Matrix& _Mat;
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
  NonHermitianSchurDiagMooeeOperator(Matrix& Mat): _Mat(Mat){};
   virtual void Mpc(const Field& in, Field& out) {
     Field tmp(in.Grid());
@@ -405,6 +425,8 @@ class NonHermitianSchurDiagOneOperator : public NonHermitianSchurOperatorBase<Fi
   Matrix &_Mat;
   
  public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   NonHermitianSchurDiagOneOperator (Matrix& Mat): _Mat(Mat){};
   virtual void Mpc(const Field& in, Field& out) {
     Field tmp(in.Grid());
@@ -435,6 +457,8 @@ class NonHermitianSchurDiagTwoOperator : public NonHermitianSchurOperatorBase<Fi
   Matrix& _Mat;
   
  public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
  NonHermitianSchurDiagTwoOperator(Matrix& Mat): _Mat(Mat){};
 
   virtual void Mpc(const Field& in, Field& out) {
@@ -475,6 +499,8 @@ class SchurStaggeredOperator :  public SchurOperatorBase<Field> {
   Field tmp;
   RealD mass;
  public:
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
   SchurStaggeredOperator (Matrix &Mat): _Mat(Mat), tmp(_Mat.RedBlackGrid()) 
   { 
     assert( _Mat.isTrivialEE() );

Grid/algorithms/SparseMatrix.h

@@ -48,6 +48,8 @@ public:
   virtual  void Mdiag    (const Field &in, Field &out)=0;
   virtual  void Mdir     (const Field &in, Field &out,int dir, int disp)=0;
   virtual  void MdirAll  (const Field &in, std::vector<Field> &out)=0;
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 };
 
 /////////////////////////////////////////////////////////////////////////////////////////////
@@ -73,6 +75,8 @@ public:
   virtual  void MooeeDag    (const Field &in, Field &out)=0;
   virtual  void MooeeInvDag (const Field &in, Field &out)=0;
 
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 };
 
 NAMESPACE_END(Grid);

Grid/lattice/Lattice.h

@@ -36,7 +36,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 #include <Grid/lattice/Lattice_local.h>
 #include <Grid/lattice/Lattice_reduction.h>
 #include <Grid/lattice/Lattice_peekpoke.h>
-//#include <Grid/lattice/Lattice_reality.h>
+#include <Grid/lattice/Lattice_reality.h>
 #include <Grid/lattice/Lattice_real_imag.h>
 #include <Grid/lattice/Lattice_comparison_utils.h>
 #include <Grid/lattice/Lattice_comparison.h>

Grid/qcd/action/fermion/FermionOperator.h

@@ -89,7 +89,8 @@ public:
   virtual void  Mdiag  (const FermionField &in, FermionField &out) { Mooee(in,out);};   // Same as Mooee applied to both CB's
   virtual void  Mdir   (const FermionField &in, FermionField &out,int dir,int disp)=0;   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
   virtual void  MdirAll(const FermionField &in, std::vector<FermionField> &out)=0;   // case by case Wilson, Clover, Cayley, ContFrac, PartFrac
-
+  virtual std::vector<int> Directions(void)   =0;
+  virtual std::vector<int> Displacements(void)=0;
 
   virtual void  MomentumSpacePropagator(FermionField &out,const FermionField &in,RealD _m,std::vector<double> twist) { assert(0);};
 

Grid/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -44,6 +44,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
   typedef StaggeredKernels<Impl> Kernels;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 

Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -49,6 +49,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
   typedef StaggeredKernels<Impl> Kernels;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 

Grid/qcd/action/fermion/NaiveStaggeredFermion.h

@@ -47,6 +47,9 @@ public:
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   ////////////////////////////////////////
   // Performance monitoring
   ////////////////////////////////////////

Grid/qcd/action/fermion/WilsonFermion.h

@@ -63,6 +63,9 @@ public:
   INHERIT_IMPL_TYPES(Impl);
   typedef WilsonKernels<Impl> Kernels;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   ///////////////////////////////////////////////////////////////
   // Implement the abstract base
   ///////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/WilsonFermion5D.h

@@ -72,6 +72,9 @@ public:
   typedef WilsonKernels<Impl> Kernels;
   PmuStat stat;
 
+  virtual std::vector<int> Directions(void)   { return this->directions; };
+  virtual std::vector<int> Displacements(void){ return this->displacements;};
+
   FermionField _tmp;
   FermionField &tmp(void) { return _tmp; }
 

Grid/qcd/action/fermion/g5HermitianLinop.h

@@ -79,6 +79,8 @@ public:
     _Mat.M(in,tmp);
     G5R5(out,tmp);
   }
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
 };
 
 
@@ -127,6 +129,8 @@ public:
     _Mat.M(in,tmp);
     out=g5*tmp;
   }
+  virtual std::vector<int> Directions(void)   { return _Mat.Directions();};
+  virtual std::vector<int> Displacements(void){ return _Mat.Displacements();};
 };
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h

@@ -133,14 +133,14 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
   pickCheckerboard(Even, CloverTermEven, CloverTerm);
   pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
 
-  pickCheckerboard(Even, CloverTermDagEven, closure(adj(CloverTerm)));
+  pickCheckerboard(Even, CloverTermDagEven, adj(CloverTerm));
-  pickCheckerboard(Odd, CloverTermDagOdd, closure(adj(CloverTerm)));
+  pickCheckerboard(Odd, CloverTermDagOdd, adj(CloverTerm));
 
   pickCheckerboard(Even, CloverTermInvEven, CloverTermInv);
   pickCheckerboard(Odd, CloverTermInvOdd, CloverTermInv);
 
-  pickCheckerboard(Even, CloverTermInvDagEven, closure(adj(CloverTermInv)));
+  pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
-  pickCheckerboard(Odd, CloverTermInvDagOdd, closure(adj(CloverTermInv)));
+  pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
 }
 
 template <class Impl>

benchmarks/Benchmark_su3.cc

@@ -191,7 +191,7 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking SU3xSU3  CovShiftForward(z,x,y)"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
+  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GB/s (incl Cshift)\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
   for(int lat=LMIN;lat<=LMAX;lat+=LADD){
@@ -216,15 +216,16 @@ int main (int argc, char ** argv)
 	
 	
 	    double bytes=3*vol*Nc*Nc*sizeof(Complex);
+	    double ncbytes=5*vol*Nc*Nc*sizeof(Complex);
 	    double flops=Nc*Nc*(6+8+8)*vol;
-	    std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
+	    std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<ncbytes/time<<"\t\t" << flops/time<<std::endl;
       }
   }
 #if 1
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
   std::cout<<GridLogMessage << "= Benchmarking SU3xSU3  z= x * Cshift(y)"<<std::endl;
   std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GFlop/s"<<std::endl;
+  std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s\t\t GB/s (incl Cshift)\t\t GFlop/s"<<std::endl;
   std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
 
   for(int lat=LMIN;lat<=LMAX;lat+=LADD){
@@ -258,10 +259,11 @@ int main (int argc, char ** argv)
 	tmult  = tmult /Nloop;
 	
 	double bytes=3*vol*Nc*Nc*sizeof(Complex);
+	double ncbytes=5*vol*Nc*Nc*sizeof(Complex);
 	double flops=Nc*Nc*(6+8+8)*vol;
 	std::cout<<GridLogMessage<<std::setprecision(3) << "total us "<<time<<" shift "<<tshift <<" mult "<<tmult<<std::endl;
 	time = time * 1000; // convert to NS for GB/s
-	std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t" << flops/time<<std::endl;
+	std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t" <<ncbytes/time<<"\t\t" << flops/time<<std::endl;
       }
     }
 #endif