Remove D2 lattice fields; remove WilsonCompressor based half precision project; add sComplex lexLattice layouts

Fix staggered instantiations for new types

Grid/qcd/action/fermion/CloverHelpers.h

@@ -143,6 +143,8 @@ public:
   static int getNMAX(Lattice<iImplClover<vComplexD2>> &t, RealD R) {return getNMAX(1e-12,R);}
   static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
   static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+  static int getNMAX(Lattice<iImplClover<sComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<sComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
 
   static void Instantiate(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
     GridBase* grid = Clover.Grid();
@@ -268,6 +270,8 @@ public:
 
   static int getNMAX(Lattice<iImplClover<vComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
   static int getNMAX(Lattice<iImplClover<vComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
+  static int getNMAX(Lattice<iImplClover<sComplexD>> &t, RealD R) {return getNMAX(1e-12,R);}
+  static int getNMAX(Lattice<iImplClover<sComplexF>> &t, RealD R) {return getNMAX(1e-6,R);}
 
   static void InstantiateClover(CloverField& Clover, CloverField& CloverInv, RealD csw_t, RealD diag_mass) {
 

Grid/qcd/action/fermion/DomainWallVec5dImpl.h

@@ -40,15 +40,12 @@ public:
   static const int Dimension = Representation::Dimension;
   static const bool isFundamental = Representation::isFundamental;
   static const bool LsVectorised=true;
-  static const int Nhcs = Options::Nhcs;
       
   typedef typename Options::_Coeff_t Coeff_t;      
-  typedef typename Options::template PrecisionMapper<Simd>::LowerPrecVector SimdL;
   
   template <typename vtype> using iImplSpinor            = iScalar<iVector<iVector<vtype, Dimension>, Ns> >;
   template <typename vtype> using iImplPropagator        = iScalar<iMatrix<iMatrix<vtype, Dimension>, Ns> >;
   template <typename vtype> using iImplHalfSpinor        = iScalar<iVector<iVector<vtype, Dimension>, Nhs> >;
-  template <typename vtype> using iImplHalfCommSpinor    = iScalar<iVector<iVector<vtype, Dimension>, Nhcs> >;
   template <typename vtype> using iImplDoubledGaugeField = iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>;
   template <typename vtype> using iImplGaugeField        = iVector<iScalar<iMatrix<vtype, Dimension> >, Nd>;
   template <typename vtype> using iImplGaugeLink         = iScalar<iScalar<iMatrix<vtype, Dimension> > >;
@@ -56,7 +53,6 @@ public:
   typedef iImplSpinor<Simd>            SiteSpinor;
   typedef iImplPropagator<Simd>        SitePropagator;
   typedef iImplHalfSpinor<Simd>        SiteHalfSpinor;
-  typedef iImplHalfCommSpinor<SimdL>   SiteHalfCommSpinor;
   typedef Lattice<SiteSpinor>          FermionField;
   typedef Lattice<SitePropagator>      PropagatorField;
 
@@ -68,7 +64,7 @@ public:
   typedef iImplGaugeLink<typename Simd::scalar_type>          SiteScalarGaugeLink;  // scalar
   typedef Lattice<SiteDoubledGaugeField>                      DoubledGaugeField;
       
-  typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
+  typedef WilsonCompressor<SiteHalfSpinor, SiteSpinor> Compressor;
   typedef WilsonImplParams ImplParams;
   typedef WilsonStencil<SiteSpinor, SiteHalfSpinor,ImplParams> StencilImpl;
   typedef typename StencilImpl::View_type StencilView;
@@ -201,16 +197,8 @@ typedef DomainWallVec5dImpl<vComplex ,FundamentalRepresentation, CoeffReal> Doma
 typedef DomainWallVec5dImpl<vComplexF,FundamentalRepresentation, CoeffReal> DomainWallVec5dImplF; // Float
 typedef DomainWallVec5dImpl<vComplexD,FundamentalRepresentation, CoeffReal> DomainWallVec5dImplD; // Double
  
-typedef DomainWallVec5dImpl<vComplex ,FundamentalRepresentation, CoeffRealHalfComms> DomainWallVec5dImplRL; // Real.. whichever prec
-typedef DomainWallVec5dImpl<vComplexF,FundamentalRepresentation, CoeffRealHalfComms> DomainWallVec5dImplFH; // Float
-typedef DomainWallVec5dImpl<vComplexD,FundamentalRepresentation, CoeffRealHalfComms> DomainWallVec5dImplDF; // Double
- 
 typedef DomainWallVec5dImpl<vComplex ,FundamentalRepresentation,CoeffComplex> ZDomainWallVec5dImplR; // Real.. whichever prec
 typedef DomainWallVec5dImpl<vComplexF,FundamentalRepresentation,CoeffComplex> ZDomainWallVec5dImplF; // Float
 typedef DomainWallVec5dImpl<vComplexD,FundamentalRepresentation,CoeffComplex> ZDomainWallVec5dImplD; // Double
- 
-typedef DomainWallVec5dImpl<vComplex ,FundamentalRepresentation,CoeffComplexHalfComms> ZDomainWallVec5dImplRL; // Real.. whichever prec
-typedef DomainWallVec5dImpl<vComplexF,FundamentalRepresentation,CoeffComplexHalfComms> ZDomainWallVec5dImplFH; // Float
-typedef DomainWallVec5dImpl<vComplexD,FundamentalRepresentation,CoeffComplexHalfComms> ZDomainWallVec5dImplDF; // Double
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/Fermion.h

@@ -114,7 +114,7 @@ NAMESPACE_CHECK(DWFutils);
 // Cayley 5d
 NAMESPACE_BEGIN(Grid);
 
-typedef WilsonFermion<WilsonImplD2> WilsonFermionD2;
+//typedef WilsonFermion<WilsonImplD2> WilsonFermionD2;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
@@ -138,7 +138,7 @@ typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplF> SpWilsonTwoIndexSymmetricF
 typedef WilsonFermion<SpWilsonTwoIndexSymmetricImplD> SpWilsonTwoIndexSymmetricFermionD;
 
 // Twisted mass fermion
-typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2;
+//typedef WilsonTMFermion<WilsonImplD2> WilsonTMFermionD2;
 typedef WilsonTMFermion<WilsonImplF> WilsonTMFermionF;
 typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 
@@ -146,11 +146,11 @@ typedef WilsonTMFermion<WilsonImplD> WilsonTMFermionD;
 template <typename WImpl> using WilsonClover = WilsonCloverFermion<WImpl, CloverHelpers<WImpl>>;
 template <typename WImpl> using WilsonExpClover = WilsonCloverFermion<WImpl, ExpCloverHelpers<WImpl>>;
 
-typedef WilsonClover<WilsonImplD2> WilsonCloverFermionD2;
+//typedef WilsonClover<WilsonImplD2> WilsonCloverFermionD2;
 typedef WilsonClover<WilsonImplF> WilsonCloverFermionF;
 typedef WilsonClover<WilsonImplD> WilsonCloverFermionD;
 
-typedef WilsonExpClover<WilsonImplD2> WilsonExpCloverFermionD2;
+//typedef WilsonExpClover<WilsonImplD2> WilsonExpCloverFermionD2;
 typedef WilsonExpClover<WilsonImplF> WilsonExpCloverFermionF;
 typedef WilsonExpClover<WilsonImplD> WilsonExpCloverFermionD;
 
@@ -168,15 +168,15 @@ template <typename WImpl> using CompactWilsonClover = CompactWilsonCloverFermion
 template <typename WImpl> using CompactWilsonClover5D = CompactWilsonCloverFermion5D<WImpl, CompactCloverHelpers<WImpl>>;
 template <typename WImpl> using CompactWilsonExpClover = CompactWilsonCloverFermion<WImpl, CompactExpCloverHelpers<WImpl>>;
 
-typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2;
+//typedef CompactWilsonClover<WilsonImplD2> CompactWilsonCloverFermionD2;
 typedef CompactWilsonClover<WilsonImplF> CompactWilsonCloverFermionF;
 typedef CompactWilsonClover<WilsonImplD> CompactWilsonCloverFermionD;
 
-typedef CompactWilsonClover5D<WilsonImplD2> CompactWilsonCloverFermion5DD2;
+//typedef CompactWilsonClover5D<WilsonImplD2> CompactWilsonCloverFermion5DD2;
 typedef CompactWilsonClover5D<WilsonImplF> CompactWilsonCloverFermion5DF;
 typedef CompactWilsonClover5D<WilsonImplD> CompactWilsonCloverFermion5DD;
 
-typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2;
+//typedef CompactWilsonExpClover<WilsonImplD2> CompactWilsonExpCloverFermionD2;
 typedef CompactWilsonExpClover<WilsonImplF> CompactWilsonExpCloverFermionF;
 typedef CompactWilsonExpClover<WilsonImplD> CompactWilsonExpCloverFermionD;
 
@@ -192,56 +192,56 @@ typedef CompactWilsonClover<WilsonTwoIndexAntiSymmetricImplD> CompactWilsonClove
 // Domain Wall fermions
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
-typedef DomainWallFermion<WilsonImplD2> DomainWallFermionD2;
+//typedef DomainWallFermion<WilsonImplD2> DomainWallFermionD2;
 
-typedef DomainWallEOFAFermion<WilsonImplD2> DomainWallEOFAFermionD2;
+//typedef DomainWallEOFAFermion<WilsonImplD2> DomainWallEOFAFermionD2;
 typedef DomainWallEOFAFermion<WilsonImplF> DomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<WilsonImplD> DomainWallEOFAFermionD;
 
-typedef MobiusFermion<WilsonImplD2> MobiusFermionD2;
+//typedef MobiusFermion<WilsonImplD2> MobiusFermionD2;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
-typedef MobiusEOFAFermion<WilsonImplD2> MobiusEOFAFermionD2;
+//typedef MobiusEOFAFermion<WilsonImplD2> MobiusEOFAFermionD2;
 typedef MobiusEOFAFermion<WilsonImplF> MobiusEOFAFermionF;
 typedef MobiusEOFAFermion<WilsonImplD> MobiusEOFAFermionD;
 
-typedef ZMobiusFermion<ZWilsonImplD2> ZMobiusFermionD2;
+//typedef ZMobiusFermion<ZWilsonImplD2> ZMobiusFermionD2;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
-typedef ScaledShamirFermion<WilsonImplD2> ScaledShamirFermionD2;
+//typedef ScaledShamirFermion<WilsonImplD2> ScaledShamirFermionD2;
 typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF;
 typedef ScaledShamirFermion<WilsonImplD> ScaledShamirFermionD;
 
-typedef MobiusZolotarevFermion<WilsonImplD2> MobiusZolotarevFermionD2;
+//typedef MobiusZolotarevFermion<WilsonImplD2> MobiusZolotarevFermionD2;
 typedef MobiusZolotarevFermion<WilsonImplF> MobiusZolotarevFermionF;
 typedef MobiusZolotarevFermion<WilsonImplD> MobiusZolotarevFermionD;
-typedef ShamirZolotarevFermion<WilsonImplD2> ShamirZolotarevFermionD2;
+//typedef ShamirZolotarevFermion<WilsonImplD2> ShamirZolotarevFermionD2;
 typedef ShamirZolotarevFermion<WilsonImplF> ShamirZolotarevFermionF;
 typedef ShamirZolotarevFermion<WilsonImplD> ShamirZolotarevFermionD;
 
-typedef OverlapWilsonCayleyTanhFermion<WilsonImplD2> OverlapWilsonCayleyTanhFermionD2;
+//typedef OverlapWilsonCayleyTanhFermion<WilsonImplD2> OverlapWilsonCayleyTanhFermionD2;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplF> OverlapWilsonCayleyTanhFermionF;
 typedef OverlapWilsonCayleyTanhFermion<WilsonImplD> OverlapWilsonCayleyTanhFermionD;
-typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD2> OverlapWilsonCayleyZolotarevFermionD2;
+//typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD2> OverlapWilsonCayleyZolotarevFermionD2;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplF> OverlapWilsonCayleyZolotarevFermionF;
 typedef OverlapWilsonCayleyZolotarevFermion<WilsonImplD> OverlapWilsonCayleyZolotarevFermionD;
 
 // Continued fraction
-typedef OverlapWilsonContFracTanhFermion<WilsonImplD2> OverlapWilsonContFracTanhFermionD2;
+//typedef OverlapWilsonContFracTanhFermion<WilsonImplD2> OverlapWilsonContFracTanhFermionD2;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplF> OverlapWilsonContFracTanhFermionF;
 typedef OverlapWilsonContFracTanhFermion<WilsonImplD> OverlapWilsonContFracTanhFermionD;
-typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD2> OverlapWilsonContFracZolotarevFermionD2;
+//typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD2> OverlapWilsonContFracZolotarevFermionD2;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplF> OverlapWilsonContFracZolotarevFermionF;
 typedef OverlapWilsonContFracZolotarevFermion<WilsonImplD> OverlapWilsonContFracZolotarevFermionD;
 
 // Partial fraction
-typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD2> OverlapWilsonPartialFractionTanhFermionD2;
+//typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD2> OverlapWilsonPartialFractionTanhFermionD2;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplF> OverlapWilsonPartialFractionTanhFermionF;
 typedef OverlapWilsonPartialFractionTanhFermion<WilsonImplD> OverlapWilsonPartialFractionTanhFermionD;
 
-typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD2> OverlapWilsonPartialFractionZolotarevFermionD2;
+//typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD2> OverlapWilsonPartialFractionZolotarevFermionD2;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplF> OverlapWilsonPartialFractionZolotarevFermionF;
 typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonPartialFractionZolotarevFermionD;
 
@@ -252,19 +252,19 @@ typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
-typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionD2;
+//typedef DomainWallEOFAFermion<GparityWilsonImplR> GparityDomainWallEOFAFermionD2;
 typedef DomainWallEOFAFermion<GparityWilsonImplF> GparityDomainWallEOFAFermionF;
 typedef DomainWallEOFAFermion<GparityWilsonImplD> GparityDomainWallEOFAFermionD;
 
-typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionD2;
+//typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionD2;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
 
-typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionD2;
+//typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionD2;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
-typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionD2;
+//typedef MobiusEOFAFermion<GparityWilsonImplR> GparityMobiusEOFAFermionD2;
 typedef MobiusEOFAFermion<GparityWilsonImplF> GparityMobiusEOFAFermionF;
 typedef MobiusEOFAFermion<GparityWilsonImplD> GparityMobiusEOFAFermionD;
 

Grid/qcd/action/fermion/FermionOperatorImpl.h

@@ -88,51 +88,14 @@ NAMESPACE_BEGIN(Grid);
 //  }
 //////////////////////////////////////////////
 
-template <class T> struct SamePrecisionMapper {
-  typedef T HigherPrecVector ;
-  typedef T LowerPrecVector ;
-};
-template <class T> struct LowerPrecisionMapper {  };
-template <> struct LowerPrecisionMapper<vRealF> {
-  typedef vRealF HigherPrecVector ;
-  typedef vRealH LowerPrecVector ;
-};
-template <> struct LowerPrecisionMapper<vRealD> {
-  typedef vRealD HigherPrecVector ;
-  typedef vRealF LowerPrecVector ;
-};
-template <> struct LowerPrecisionMapper<vComplexF> {
-  typedef vComplexF HigherPrecVector ;
-  typedef vComplexH LowerPrecVector ;
-};
-template <> struct LowerPrecisionMapper<vComplexD> {
-  typedef vComplexD HigherPrecVector ;
-  typedef vComplexF LowerPrecVector ;
-};
 
 struct CoeffReal {
 public:
   typedef RealD _Coeff_t;
-  static const int Nhcs = 2;
-  template<class Simd> using PrecisionMapper = SamePrecisionMapper<Simd>;
-};
-struct CoeffRealHalfComms {
-public:
-  typedef RealD _Coeff_t;
-  static const int Nhcs = 1;
-  template<class Simd> using PrecisionMapper = LowerPrecisionMapper<Simd>;
 };
 struct CoeffComplex {
 public:
   typedef ComplexD _Coeff_t;
-  static const int Nhcs = 2;
-  template<class Simd> using PrecisionMapper = SamePrecisionMapper<Simd>;
-};
-struct CoeffComplexHalfComms {
-public:
-  typedef ComplexD _Coeff_t;
-  static const int Nhcs = 1;
-  template<class Simd> using PrecisionMapper = LowerPrecisionMapper<Simd>;
 };
 
 ////////////////////////////////////////////////////////////////////////

Grid/qcd/action/fermion/GparityWilsonImpl.h

@@ -48,7 +48,6 @@ public:
 
  static const int Dimension = Representation::Dimension;
  static const bool isFundamental = Representation::isFundamental;
- static const int Nhcs = Options::Nhcs;
  static const bool LsVectorised=false;
  static const bool isGparity=true;
 
@@ -56,18 +55,15 @@ public:
  INHERIT_GIMPL_TYPES(Gimpl);
  
  typedef typename Options::_Coeff_t Coeff_t;
- typedef typename Options::template PrecisionMapper<Simd>::LowerPrecVector SimdL;
       
  template <typename vtype> using iImplSpinor            = iVector<iVector<iVector<vtype, Dimension>, Ns>,   Ngp>;
  template <typename vtype> using iImplPropagator        = iMatrix<iMatrix<iMatrix<vtype, Dimension>, Ns>,   Ngp>;
  template <typename vtype> using iImplHalfSpinor        = iVector<iVector<iVector<vtype, Dimension>, Nhs>,  Ngp>;
- template <typename vtype> using iImplHalfCommSpinor    = iVector<iVector<iVector<vtype, Dimension>, Nhcs>, Ngp>;
  template <typename vtype> using iImplDoubledGaugeField = iVector<iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>, Ngp>;
 
   typedef iImplSpinor<Simd>            SiteSpinor;
   typedef iImplPropagator<Simd>        SitePropagator;
   typedef iImplHalfSpinor<Simd>        SiteHalfSpinor;
-  typedef iImplHalfCommSpinor<SimdL>   SiteHalfCommSpinor;
   typedef iImplDoubledGaugeField<Simd> SiteDoubledGaugeField;
 
   typedef Lattice<SiteSpinor> FermionField;
@@ -75,7 +71,7 @@ public:
   typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
  
   typedef GparityWilsonImplParams ImplParams;
-  typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
+  typedef WilsonCompressor<SiteHalfSpinor, SiteSpinor> Compressor;
   typedef WilsonStencil<SiteSpinor, SiteHalfSpinor, ImplParams> StencilImpl;
   typedef typename StencilImpl::View_type StencilView;
       
@@ -408,9 +404,5 @@ public:
 typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffReal> GparityWilsonImplR;  // Real.. whichever prec
 typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffReal> GparityWilsonImplF;  // Float
 typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffReal> GparityWilsonImplD;  // Double
- 
-//typedef GparityWilsonImpl<vComplex , FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec
-//typedef GparityWilsonImpl<vComplexF, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
-//typedef GparityWilsonImpl<vComplexD, FundamentalRepresentation,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
 
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/StaggeredImpl.h

@@ -176,4 +176,13 @@ typedef StaggeredImpl<vComplex,  FundamentalRepresentation > StaggeredImplR;   /
 typedef StaggeredImpl<vComplexF, FundamentalRepresentation > StaggeredImplF;  // Float
 typedef StaggeredImpl<vComplexD, FundamentalRepresentation > StaggeredImplD;  // Double
 
+/////////////////////////////////////////////////////////////////////////////
+// Lexicographic staggered implementations
+/////////////////////////////////////////////////////////////////////////////
+typedef StaggeredImpl<sComplexF, FundamentalRepresentation > lexStaggeredImplF;
+typedef StaggeredImpl<sComplexD, FundamentalRepresentation > lexStaggeredImplD;
+/////////////////////////////////////////////////////////////////////////////
+// End lexicographic staggered implementations
+/////////////////////////////////////////////////////////////////////////////
+
 NAMESPACE_END(Grid);

Grid/qcd/action/fermion/WilsonCompressor.h

@@ -38,7 +38,7 @@ NAMESPACE_BEGIN(Grid);
 
 
 //Could make FaceGather a template param, but then behaviour is runtime not compile time
-template<class _HCspinor,class _Hspinor,class _Spinor, class projector>
+template<class _Hspinor,class _Spinor, class projector>
 class WilsonCompressorTemplate : public FaceGatherSimple
 {
 public:
@@ -53,13 +53,11 @@ public:
 
   typedef _Spinor         SiteSpinor;
   typedef _Hspinor     SiteHalfSpinor;
-  typedef _HCspinor SiteHalfCommSpinor;
+  typedef typename SiteHalfSpinor::vector_type vector_type;    
-  typedef typename SiteHalfCommSpinor::vector_type vComplexLow;
+  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vector_type);
-  typedef typename SiteHalfSpinor::vector_type     vComplexHigh;
-  constexpr static int Nw=sizeof(SiteHalfSpinor)/sizeof(vComplexHigh);
 
   accelerator_inline int CommDatumSize(void) const {
-    return sizeof(SiteHalfCommSpinor);
+    return sizeof(SiteHalfSpinor);
   }
 
   /*****************************************************/
@@ -159,7 +157,7 @@ public:
     spProj(result,in);							\
   }									\
   };									\
-  template<typename HCS,typename HS,typename S> using Compressor = WilsonCompressorTemplate<HCS,HS,S,Projector>;
+  template<typename HS,typename S> using Compressor = WilsonCompressorTemplate<HS,S,Projector>;
 
 DECLARE_PROJ(WilsonXpProjector,WilsonXpCompressor,spProjXp);
 DECLARE_PROJ(WilsonYpProjector,WilsonYpCompressor,spProjYp);
@@ -188,7 +186,7 @@ public:
     }
   }
 };
-template<typename HCS,typename HS,typename S> using WilsonCompressor = WilsonCompressorTemplate<HCS,HS,S,WilsonProjector>;
+template<typename HS,typename S> using WilsonCompressor = WilsonCompressorTemplate<HS,S,WilsonProjector>;
 
 // Fast comms buffer manipulation which should inline right through (avoid direction
 // dependent logic that prevents inlining
@@ -237,7 +235,6 @@ public:
     // Use types to select the write direction by directon compressor
     typedef typename compressor::SiteSpinor         SiteSpinor;
     typedef typename compressor::SiteHalfSpinor     SiteHalfSpinor;
-    typedef typename compressor::SiteHalfCommSpinor SiteHalfCommSpinor;
 
     this->_grid->StencilBarrier();
 
@@ -245,14 +242,14 @@ public:
     
     this->u_comm_offset=0;
       
-    WilsonXpCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> XpCompress; 
+    WilsonXpCompressor<SiteHalfSpinor,SiteSpinor> XpCompress; 
-    WilsonYpCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> YpCompress; 
+    WilsonYpCompressor<SiteHalfSpinor,SiteSpinor> YpCompress; 
-    WilsonZpCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> ZpCompress; 
+    WilsonZpCompressor<SiteHalfSpinor,SiteSpinor> ZpCompress; 
-    WilsonTpCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> TpCompress;
+    WilsonTpCompressor<SiteHalfSpinor,SiteSpinor> TpCompress;
-    WilsonXmCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> XmCompress; 
+    WilsonXmCompressor<SiteHalfSpinor,SiteSpinor> XmCompress; 
-    WilsonYmCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> YmCompress; 
+    WilsonYmCompressor<SiteHalfSpinor,SiteSpinor> YmCompress; 
-    WilsonZmCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> ZmCompress; 
+    WilsonZmCompressor<SiteHalfSpinor,SiteSpinor> ZmCompress; 
-    WilsonTmCompressor<SiteHalfCommSpinor,SiteHalfSpinor,SiteSpinor> TmCompress;
+    WilsonTmCompressor<SiteHalfSpinor,SiteSpinor> TmCompress;
 
     int dag = compress.dag;
     int face_idx=0;

Grid/qcd/action/fermion/WilsonImpl.h

@@ -42,7 +42,6 @@ public:
   static const bool isFundamental = Representation::isFundamental;
   static const bool LsVectorised=false;
   static const bool isGparity=false;
-  static const int Nhcs = Options::Nhcs;
 
   typedef PeriodicGaugeImpl<GaugeImplTypes<S, Dimension > > Gimpl;
   INHERIT_GIMPL_TYPES(Gimpl);
@@ -51,25 +50,22 @@ public:
   constexpr bool is_fundamental() const{return Dimension == Nc ? 1 : 0;}
     
   typedef typename Options::_Coeff_t Coeff_t;
-  typedef typename Options::template PrecisionMapper<Simd>::LowerPrecVector SimdL;
       
   template <typename vtype> using iImplSpinor            = iScalar<iVector<iVector<vtype, Dimension>, Ns> >;
   template <typename vtype> using iImplPropagator        = iScalar<iMatrix<iMatrix<vtype, Dimension>, Ns> >;
   template <typename vtype> using iImplHalfSpinor        = iScalar<iVector<iVector<vtype, Dimension>, Nhs> >;
-  template <typename vtype> using iImplHalfCommSpinor    = iScalar<iVector<iVector<vtype, Dimension>, Nhcs> >;
   template <typename vtype> using iImplDoubledGaugeField = iVector<iScalar<iMatrix<vtype, Dimension> >, Nds>;
     
   typedef iImplSpinor<Simd>            SiteSpinor;
   typedef iImplPropagator<Simd>        SitePropagator;
   typedef iImplHalfSpinor<Simd>        SiteHalfSpinor;
-  typedef iImplHalfCommSpinor<SimdL>   SiteHalfCommSpinor;
   typedef iImplDoubledGaugeField<Simd> SiteDoubledGaugeField;
     
   typedef Lattice<SiteSpinor>            FermionField;
   typedef Lattice<SitePropagator>        PropagatorField;
   typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
     
-  typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
+  typedef WilsonCompressor<SiteHalfSpinor, SiteSpinor> Compressor;
   typedef WilsonImplParams ImplParams;
   typedef WilsonStencil<SiteSpinor, SiteHalfSpinor,ImplParams> StencilImpl;
   typedef const typename StencilImpl::View_type StencilView;
@@ -242,12 +238,19 @@ public:
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffReal > WilsonImplR;  // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF;  // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD;  // Double
-typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffReal > WilsonImplD2;  // Double
+
+/////////////////////////////////////////////////////////////////////////////
+// Lexicographic Wilson implementations
+/////////////////////////////////////////////////////////////////////////////
+typedef WilsonImpl<sComplexF, FundamentalRepresentation, CoeffReal > lexWilsonImplF;
+typedef WilsonImpl<sComplexD, FundamentalRepresentation, CoeffReal > lexWilsonImplD;
+/////////////////////////////////////////////////////////////////////////////
+// End lexicographic Wilson implementations
+/////////////////////////////////////////////////////////////////////////////
 
 typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplex > ZWilsonImplR; // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplex > ZWilsonImplF; // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplex > ZWilsonImplD; // Double
-typedef WilsonImpl<vComplexD2, FundamentalRepresentation, CoeffComplex > ZWilsonImplD2; // Double
 
 typedef WilsonImpl<vComplex,  AdjointRepresentation, CoeffReal > WilsonAdjImplR;   // Real.. whichever prec
 typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF;  // Float

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

@@ -684,7 +684,7 @@ void StaggeredKernels<Impl>::DhopSiteAsm(StencilView &st,
 #ifdef STAG_VEC5D
   // This is the single precision 5th direction vectorised kernel
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView &st,
+template <> inline void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView &st,
 								    DoubledGaugeFieldView &U,
 								    DoubledGaugeFieldView &UUU,
 								    SiteSpinor *buf, int sF,
@@ -739,7 +739,7 @@ template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilView
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView &st, 
+template <> inline void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView &st, 
 								    DoubledGaugeFieldView &U,
 								    DoubledGaugeFieldView &UUU,
 								    SiteSpinor *buf, int sF,
@@ -825,7 +825,7 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilView
   // This is the single precision 5th direction vectorised kernel
 
 #include <Grid/simd/Intel512single.h>
-template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilView &st, 
+template <> inline void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilView &st, 
 							       DoubledGaugeFieldView &U,
 							       DoubledGaugeFieldView &UUU,
 							       SiteSpinor *buf, int sF,
@@ -894,7 +894,7 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilView &st,
 }
 
 #include <Grid/simd/Intel512double.h>
-template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilView &st, 
+template <> inline void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilView &st, 
 							       DoubledGaugeFieldView &U,
 							       DoubledGaugeFieldView &UUU,
 							       SiteSpinor *buf, int sF,

Grid/qcd/action/fermion/instantiation/StaggeredKernelsInstantiation.cc.master

@@ -29,6 +29,7 @@ directory
 #include <Grid/qcd/action/fermion/FermionCore.h>
 #include <Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h>
 #include <Grid/qcd/action/fermion/implementation/StaggeredKernelsHand.h>
+#include <Grid/qcd/action/fermion/implementation/StaggeredKernelsAsm.h>
 
 NAMESPACE_BEGIN(Grid);
 

Grid/qcd/action/fermion/instantiation/generate_instantiations.sh

@@ -2,14 +2,15 @@
 
 STAG_IMPL_LIST=" \
 	   StaggeredImplF \
-	   StaggeredImplD  "
+	   StaggeredImplD \
-
+	   lexStaggeredImplF \
-STAG5_IMPL_LIST=""
+	   lexStaggeredImplD  "
 
 WILSON_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
-	   WilsonImplD2 \
+	   lexWilsonImplF \
+	   lexWilsonImplD \
 	   SpWilsonImplF \
 	   SpWilsonImplD \
 	   WilsonAdjImplF \
@@ -32,9 +33,10 @@ COMPACT_WILSON_IMPL_LIST=" \
 DWF_IMPL_LIST=" \
 	   WilsonImplF \
 	   WilsonImplD \
-	   WilsonImplD2 \
+	   lexWilsonImplF \
+	   lexWilsonImplD \
 	   ZWilsonImplF \
-	   ZWilsonImplD2 "
+	   ZWilsonImplD "
 
 GDWF_IMPL_LIST=" \
 	   GparityWilsonImplF \
@@ -109,8 +111,3 @@ do
   ln -f -s ../$f.cc.master $impl/$f$impl.cc
 done
 done
-
-CC_LIST=" \
-  ImprovedStaggeredFermion5DInstantiation \
-  StaggeredKernelsInstantiation "
-