Merge branch 'feature/staggered-comms-compute' of https://github.com/paboyle/Grid into feature/staggered-comms-compute

lib/qcd/action/fermion/ImprovedStaggeredFermion.cc

@@ -84,15 +84,6 @@ ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GaugeField &_Uthin, Gau
 {
   ImportGauge(_Uthin,_Ufat);
 }
-template <class Impl>
-ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GaugeField &_Utriple, GaugeField &_Ufat, GridCartesian &Fgrid,
-							 GridRedBlackCartesian &Hgrid, RealD _mass,
-							 const ImplParams &p)
-  : ImprovedStaggeredFermion(Fgrid,Hgrid,_mass,1.0,1.0,1.0,p)
-{
-  ImportGaugeSimple(_Utriple,_Ufat);
-}
-
 
   ////////////////////////////////////////////////////////////
   // Momentum space propagator should be 
@@ -106,11 +97,6 @@ ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GaugeField &_Utriple, G
   // of above link to implmement fourier based solver.
   ////////////////////////////////////////////////////////////
 template <class Impl>
-void ImprovedStaggeredFermion<Impl>::ImportGauge(const GaugeField &_Uthin) 
-{
-  ImportGauge(_Uthin,_Uthin);
-};
-template <class Impl>
 void ImprovedStaggeredFermion<Impl>::ImportGaugeSimple(const GaugeField &_Utriple,const GaugeField &_Ufat) 
 {
   /////////////////////////////////////////////////////////////////
@@ -133,6 +119,20 @@ void ImprovedStaggeredFermion<Impl>::ImportGaugeSimple(const GaugeField &_Utripl
     PokeIndex<LorentzIndex>(Umu, -U, mu+4);
 
   }
+  CopyGaugeCheckerboards();
+}
+template <class Impl>
+void ImprovedStaggeredFermion<Impl>::ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U) 
+{
+
+  Umu   = _U;
+  UUUmu = _UUU;
+  CopyGaugeCheckerboards();
+}
+
+template <class Impl>
+void ImprovedStaggeredFermion<Impl>::CopyGaugeCheckerboards(void)
+{
   pickCheckerboard(Even, UmuEven,  Umu);
   pickCheckerboard(Odd,  UmuOdd ,  Umu);
   pickCheckerboard(Even, UUUmuEven,UUUmu);
@@ -168,10 +168,7 @@ void ImprovedStaggeredFermion<Impl>::ImportGauge(const GaugeField &_Uthin,const
     PokeIndex<LorentzIndex>(UUUmu, U*(-0.5*c2/u0/u0/u0), mu+4);
   }
 
-  pickCheckerboard(Even, UmuEven, Umu);
+  CopyGaugeCheckerboards();
-  pickCheckerboard(Odd,  UmuOdd , Umu);
-  pickCheckerboard(Even, UUUmuEven, UUUmu);
-  pickCheckerboard(Odd,   UUUmuOdd, UUUmu);
 }
 
 /////////////////////////////

lib/qcd/action/fermion/ImprovedStaggeredFermion.h

@@ -122,30 +122,29 @@ class ImprovedStaggeredFermion : public StaggeredKernels<Impl>, public ImprovedS
   void DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
                     const FermionField &in, FermionField &out, int dag);
 
-  // Constructor
+  //////////////////////////////////////////////////////////////////////////
+  // Grid own interface Constructor
+  //////////////////////////////////////////////////////////////////////////
   ImprovedStaggeredFermion(GaugeField &_Uthin, GaugeField &_Ufat, GridCartesian &Fgrid,
 			   GridRedBlackCartesian &Hgrid, RealD _mass,
 			   RealD _c1, RealD _c2,RealD _u0,
 			   const ImplParams &p = ImplParams());
 
   //////////////////////////////////////////////////////////////////////////
-  // MILC constructor no coefficients; premultiply links by desired scaling
+  // MILC constructor no gauge fields
-  //////////////////////////////////////////////////////////////////////////
-  ImprovedStaggeredFermion(GaugeField &_Utriple, GaugeField &_Ufat, GridCartesian &Fgrid,
-			   GridRedBlackCartesian &Hgrid, RealD _mass,
-			   const ImplParams &p = ImplParams());
-
-  //////////////////////////////////////////////////////////////////////////
-  // A don't initialise the gauge field constructor; largely internal
   //////////////////////////////////////////////////////////////////////////
   ImprovedStaggeredFermion(GridCartesian &Fgrid, GridRedBlackCartesian &Hgrid, RealD _mass,
-			   RealD _c1, RealD _c2,RealD _u0,
+			   RealD _c1=1.0, RealD _c2=1.0,RealD _u0=1.0,
 			   const ImplParams &p = ImplParams());
 
   // DoubleStore impl dependent
-  void ImportGaugeSimple(const GaugeField &_Utriple, const GaugeField &_Ufat);
+  void ImportGauge      (const GaugeField &_Uthin ) { assert(0); }
-  void ImportGauge(const GaugeField &_Uthin, const GaugeField &_Ufat);
+  void ImportGauge      (const GaugeField &_Uthin  ,const GaugeField &_Ufat);
-  void ImportGauge(const GaugeField &_Uthin);
+  void ImportGaugeSimple(const GaugeField &_UUU    ,const GaugeField &_U);
+  void ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U);
+  DoubledGaugeField &GetU(void)   { return Umu ; } ;
+  DoubledGaugeField &GetUUU(void) { return UUUmu; };
+  void CopyGaugeCheckerboards(void);
 
   ///////////////////////////////////////////////////////////////
   // Data members require to support the functionality

lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc

@@ -128,7 +128,14 @@ ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GridCartesian
   StencilEven.BuildSurfaceList(LLs,vol4);
   StencilOdd.BuildSurfaceList(LLs,vol4);
 }
-
+template <class Impl>
+void ImprovedStaggeredFermion5D<Impl>::CopyGaugeCheckerboards(void)
+{
+  pickCheckerboard(Even, UmuEven,  Umu);
+  pickCheckerboard(Odd,  UmuOdd ,  Umu);
+  pickCheckerboard(Even, UUUmuEven,UUUmu);
+  pickCheckerboard(Odd,  UUUmuOdd, UUUmu);
+}
 template<class Impl>
 ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GaugeField &_Uthin,GaugeField &_Ufat,
 							     GridCartesian         &FiveDimGrid,
@@ -144,26 +151,7 @@ ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GaugeField &_Uthin,
 {
   ImportGauge(_Uthin,_Ufat);
 }
-template<class Impl>
-ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GaugeField &_Utriple,GaugeField &_Ufat,
-							     GridCartesian         &FiveDimGrid,
-							     GridRedBlackCartesian &FiveDimRedBlackGrid,
-							     GridCartesian         &FourDimGrid,
-							     GridRedBlackCartesian &FourDimRedBlackGrid,
-							     RealD _mass,
-							     const ImplParams &p) :
-  ImprovedStaggeredFermion5D(FiveDimGrid,FiveDimRedBlackGrid,
-			     FourDimGrid,FourDimRedBlackGrid,
-			     _mass,1.0,1.0,1.0,p)
-{
-  ImportGaugeSimple(_Utriple,_Ufat);
-}
 
-template <class Impl>
-void ImprovedStaggeredFermion5D<Impl>::ImportGauge(const GaugeField &_Uthin) 
-{
-  ImportGauge(_Uthin,_Uthin);
-};
 ///////////////////////////////////////////////////
 // For MILC use; pass three link U's and 1 link U
 ///////////////////////////////////////////////////
@@ -188,10 +176,17 @@ void ImprovedStaggeredFermion5D<Impl>::ImportGaugeSimple(const GaugeField &_Utri
     Impl::InsertGaugeField(Umu,-U,mu+4);
 
   }
-  pickCheckerboard(Even, UmuEven,  Umu);
+  CopyGaugeCheckerboards();
-  pickCheckerboard(Odd,  UmuOdd ,  Umu);
+}
-  pickCheckerboard(Even, UUUmuEven,UUUmu);
+template <class Impl>
-  pickCheckerboard(Odd,  UUUmuOdd, UUUmu);
+void ImprovedStaggeredFermion5D<Impl>::ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U) 
+{
+  /////////////////////////////////////////////////////////////////
+  // Trivial import; phases and fattening and such like preapplied
+  /////////////////////////////////////////////////////////////////
+  Umu   = _U;
+  UUUmu = _UUU;
+  CopyGaugeCheckerboards();
 }
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::ImportGauge(const GaugeField &_Uthin,const GaugeField &_Ufat)
@@ -221,10 +216,7 @@ void ImprovedStaggeredFermion5D<Impl>::ImportGauge(const GaugeField &_Uthin,cons
     PokeIndex<LorentzIndex>(UUUmu, U*(-0.5*c2/u0/u0/u0), mu+4);
   }
 
-  pickCheckerboard(Even, UmuEven, Umu);
+  CopyGaugeCheckerboards();
-  pickCheckerboard(Odd,  UmuOdd , Umu);
-  pickCheckerboard(Even, UUUmuEven, UUUmu);
-  pickCheckerboard(Odd,  UUUmuOdd, UUUmu);
 }
 template<class Impl>
 void ImprovedStaggeredFermion5D<Impl>::DhopDir(const FermionField &in, FermionField &out,int dir5,int disp)

lib/qcd/action/fermion/ImprovedStaggeredFermion5D.h

@@ -139,15 +139,9 @@ namespace QCD {
     
     
     // Constructors
-    // -- No Gauge field 
+    ////////////////////////////////////////////////////////////////////////////////////////////////
-    ImprovedStaggeredFermion5D(GridCartesian         &FiveDimGrid,
+    // Grid internal interface -- Thin link and fat link, with coefficients
-			       GridRedBlackCartesian &FiveDimRedBlackGrid,
+    ////////////////////////////////////////////////////////////////////////////////////////////////
-			       GridCartesian         &FourDimGrid,
-			       GridRedBlackCartesian &FourDimRedBlackGrid,
-			       double _mass,
-			       RealD _c1, RealD _c2,RealD _u0,
-			       const ImplParams &p= ImplParams());
-    // -- Thin link and fat link, with coefficients
     ImprovedStaggeredFermion5D(GaugeField &_Uthin,
 			       GaugeField &_Ufat,
 			       GridCartesian         &FiveDimGrid,
@@ -160,19 +154,24 @@ namespace QCD {
     ////////////////////////////////////////////////////////////////////////////////////////////////
     // MILC constructor ; triple links, no rescale factors; must be externally pre multiplied
     ////////////////////////////////////////////////////////////////////////////////////////////////
-    ImprovedStaggeredFermion5D(GaugeField &_Utriple,
+    ImprovedStaggeredFermion5D(GridCartesian         &FiveDimGrid,
-			       GaugeField &_Ufat,
-			       GridCartesian         &FiveDimGrid,
 			       GridRedBlackCartesian &FiveDimRedBlackGrid,
 			       GridCartesian         &FourDimGrid,
 			       GridRedBlackCartesian &FourDimRedBlackGrid,
 			       double _mass,
+			       RealD _c1=1.0, RealD _c2=1.0,RealD _u0=1.0,
 			       const ImplParams &p= ImplParams());
 
-    // DoubleStore
+    // DoubleStore gauge field in operator
-    void ImportGauge(const GaugeField &_U);
+    void ImportGauge      (const GaugeField &_Uthin ) { assert(0); }
-    void ImportGauge(const GaugeField &_Uthin,const GaugeField &_Ufat);
+    void ImportGauge      (const GaugeField &_Uthin  ,const GaugeField &_Ufat);
-    void ImportGaugeSimple(const GaugeField &_Uthin,const GaugeField &_Ufat);
+    void ImportGaugeSimple(const GaugeField &_UUU,const GaugeField &_U);
+    void ImportGaugeSimple(const DoubledGaugeField &_UUU,const DoubledGaugeField &_U);
+    // Give a reference; can be used to do an assignment or copy back out after import
+    // if Carleton wants to cache them and not use the ImportSimple
+    DoubledGaugeField &GetU(void)   { return Umu ; } ;
+    DoubledGaugeField &GetUUU(void) { return UUUmu; };
+    void CopyGaugeCheckerboards(void);
     
     ///////////////////////////////////////////////////////////////
     // Data members require to support the functionality

tests/core/Test_staggered5Dvec.cc

@@ -141,6 +141,7 @@ int main (int argc, char ** argv)
   t1=usecond();
  
   std::cout<<GridLogMessage << "Called Ds ASM"<<std::endl;
+  std::cout<<GridLogMessage << "norm src "<< norm2(src)<<std::endl;
   std::cout<<GridLogMessage << "norm result "<< norm2(tmp)<<std::endl;
   std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
 
@@ -160,7 +161,8 @@ int main (int argc, char ** argv)
   localConvert(sresult,tmp);
  
   std::cout<<GridLogMessage << "Called sDs unroll"<<std::endl;
-  std::cout<<GridLogMessage << "norm result "<< norm2(sresult)<<std::endl;
+  std::cout<<GridLogMessage << "norm ssrc "<< norm2(ssrc)<<std::endl;
+  std::cout<<GridLogMessage << "norm sresult "<< norm2(sresult)<<std::endl;
   std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
 
 
@@ -181,6 +183,7 @@ int main (int argc, char ** argv)
   localConvert(sresult,tmp);
  
   std::cout<<GridLogMessage << "Called sDs asm"<<std::endl;
+  std::cout<<GridLogMessage << "norm ssrc   "<< norm2(ssrc)<<std::endl;
   std::cout<<GridLogMessage << "norm result "<< norm2(sresult)<<std::endl;
   std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)*extra<<std::endl;