Working half precision comms

benchmarks/Benchmark_dwf.cc

@@ -152,9 +152,6 @@ int main (int argc, char ** argv)
 
   RealD NP = UGrid->_Nprocessors;
 
-  std::cout << GridLogMessage << "Creating action operator " << std::endl;
-  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
-
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
   std::cout << GridLogMessage<< "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" <<std::endl;
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
@@ -168,6 +165,7 @@ int main (int argc, char ** argv)
   if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3   WilsonKernels" <<std::endl;
   std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
 
+  DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
   int ncall =1000;
   if (1) {
     FGrid->Barrier();
@@ -206,6 +204,33 @@ int main (int argc, char ** argv)
     Dw.Report();
   }
 
+  DomainWallFermionRL DwH(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
+  if (1) {
+    FGrid->Barrier();
+    DwH.ZeroCounters();
+    DwH.Dhop(src,result,0);
+    double t0=usecond();
+    for(int i=0;i<ncall;i++){
+      __SSC_START;
+      DwH.Dhop(src,result,0);
+      __SSC_STOP;
+    }
+    double t1=usecond();
+    FGrid->Barrier();
+    
+    double volume=Ls;  for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
+    double flops=1344*volume*ncall;
+
+    std::cout<<GridLogMessage << "Called half prec comms Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s =   "<< flops/(t1-t0)<<std::endl;
+    std::cout<<GridLogMessage << "mflop/s per rank =  "<< flops/(t1-t0)/NP<<std::endl;
+    err = ref-result; 
+    std::cout<<GridLogMessage << "norm diff   "<< norm2(err)<<std::endl;
+
+    assert (norm2(err)< 1.0e-3 );
+    DwH.Report();
+  }
+
   if (1)
   {
 

lib/qcd/action/fermion/CayleyFermion5Dcache.cc

@@ -237,6 +237,13 @@ void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &
   INSTANTIATE_DPERP(GparityWilsonImplD);
   INSTANTIATE_DPERP(ZWilsonImplF);
   INSTANTIATE_DPERP(ZWilsonImplD);
+
+  INSTANTIATE_DPERP(WilsonImplFH);
+  INSTANTIATE_DPERP(WilsonImplDF);
+  INSTANTIATE_DPERP(GparityWilsonImplFH);
+  INSTANTIATE_DPERP(GparityWilsonImplDF);
+  INSTANTIATE_DPERP(ZWilsonImplFH);
+  INSTANTIATE_DPERP(ZWilsonImplDF);
 #endif
 
 }}

lib/qcd/action/fermion/CayleyFermion5Ddense.cc

@@ -137,6 +137,20 @@ template void CayleyFermion5D<WilsonImplF>::MooeeInternal(const FermionField &ps
 template void CayleyFermion5D<WilsonImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
 template void CayleyFermion5D<ZWilsonImplF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
 template void CayleyFermion5D<ZWilsonImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+
+INSTANTIATE_DPERP(GparityWilsonImplFH);
+INSTANTIATE_DPERP(GparityWilsonImplDF);
+INSTANTIATE_DPERP(WilsonImplFH);
+INSTANTIATE_DPERP(WilsonImplDF);
+INSTANTIATE_DPERP(ZWilsonImplFH);
+INSTANTIATE_DPERP(ZWilsonImplDF);
+
+template void CayleyFermion5D<GparityWilsonImplFH>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<GparityWilsonImplDF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<WilsonImplFH>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<WilsonImplDF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<ZWilsonImplFH>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<ZWilsonImplDF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
 #endif
 
 }}

lib/qcd/action/fermion/CayleyFermion5Dssp.cc

@@ -37,7 +37,6 @@ namespace Grid {
 namespace QCD {
 
   // FIXME -- make a version of these routines with site loop outermost for cache reuse.
-
   // Pminus fowards
   // Pplus  backwards
 template<class Impl>  
@@ -152,6 +151,13 @@ void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &
   INSTANTIATE_DPERP(GparityWilsonImplD);
   INSTANTIATE_DPERP(ZWilsonImplF);
   INSTANTIATE_DPERP(ZWilsonImplD);
+
+  INSTANTIATE_DPERP(WilsonImplFH);
+  INSTANTIATE_DPERP(WilsonImplDF);
+  INSTANTIATE_DPERP(GparityWilsonImplFH);
+  INSTANTIATE_DPERP(GparityWilsonImplDF);
+  INSTANTIATE_DPERP(ZWilsonImplFH);
+  INSTANTIATE_DPERP(ZWilsonImplDF);
 #endif
 
 }

lib/qcd/action/fermion/CayleyFermion5Dvec.cc

@@ -808,10 +808,21 @@ INSTANTIATE_DPERP(DomainWallVec5dImplF);
 INSTANTIATE_DPERP(ZDomainWallVec5dImplD);
 INSTANTIATE_DPERP(ZDomainWallVec5dImplF);
 
+INSTANTIATE_DPERP(DomainWallVec5dImplDF);
+INSTANTIATE_DPERP(DomainWallVec5dImplFH);
+INSTANTIATE_DPERP(ZDomainWallVec5dImplDF);
+INSTANTIATE_DPERP(ZDomainWallVec5dImplFH);
+
 template void CayleyFermion5D<DomainWallVec5dImplF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
 template void CayleyFermion5D<DomainWallVec5dImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
 template void CayleyFermion5D<ZDomainWallVec5dImplF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
 template void CayleyFermion5D<ZDomainWallVec5dImplD>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
 
+template void CayleyFermion5D<DomainWallVec5dImplFH>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<DomainWallVec5dImplDF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<ZDomainWallVec5dImplFH>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+template void CayleyFermion5D<ZDomainWallVec5dImplDF>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv);
+
+
 
 }}

lib/qcd/action/fermion/Fermion.h

@@ -89,6 +89,10 @@ typedef WilsonFermion<WilsonImplR> WilsonFermionR;
 typedef WilsonFermion<WilsonImplF> WilsonFermionF;
 typedef WilsonFermion<WilsonImplD> WilsonFermionD;
 
+typedef WilsonFermion<WilsonImplRL> WilsonFermionRL;
+typedef WilsonFermion<WilsonImplFH> WilsonFermionFH;
+typedef WilsonFermion<WilsonImplDF> WilsonFermionDF;
+
 typedef WilsonFermion<WilsonAdjImplR> WilsonAdjFermionR;
 typedef WilsonFermion<WilsonAdjImplF> WilsonAdjFermionF;
 typedef WilsonFermion<WilsonAdjImplD> WilsonAdjFermionD;
@@ -105,27 +109,50 @@ typedef DomainWallFermion<WilsonImplR> DomainWallFermionR;
 typedef DomainWallFermion<WilsonImplF> DomainWallFermionF;
 typedef DomainWallFermion<WilsonImplD> DomainWallFermionD;
 
+typedef DomainWallFermion<WilsonImplRL> DomainWallFermionRL;
+typedef DomainWallFermion<WilsonImplFH> DomainWallFermionFH;
+typedef DomainWallFermion<WilsonImplDF> DomainWallFermionDF;
+
 typedef MobiusFermion<WilsonImplR> MobiusFermionR;
 typedef MobiusFermion<WilsonImplF> MobiusFermionF;
 typedef MobiusFermion<WilsonImplD> MobiusFermionD;
 
+typedef MobiusFermion<WilsonImplRL> MobiusFermionRL;
+typedef MobiusFermion<WilsonImplFH> MobiusFermionFH;
+typedef MobiusFermion<WilsonImplDF> MobiusFermionDF;
+
 typedef ZMobiusFermion<ZWilsonImplR> ZMobiusFermionR;
 typedef ZMobiusFermion<ZWilsonImplF> ZMobiusFermionF;
 typedef ZMobiusFermion<ZWilsonImplD> ZMobiusFermionD;
 
+typedef ZMobiusFermion<ZWilsonImplRL> ZMobiusFermionRL;
+typedef ZMobiusFermion<ZWilsonImplFH> ZMobiusFermionFH;
+typedef ZMobiusFermion<ZWilsonImplDF> ZMobiusFermionDF;
+
 // Ls vectorised 
 typedef DomainWallFermion<DomainWallVec5dImplR> DomainWallFermionVec5dR;
 typedef DomainWallFermion<DomainWallVec5dImplF> DomainWallFermionVec5dF;
 typedef DomainWallFermion<DomainWallVec5dImplD> DomainWallFermionVec5dD;
 
+typedef DomainWallFermion<DomainWallVec5dImplRL> DomainWallFermionVec5dRL;
+typedef DomainWallFermion<DomainWallVec5dImplFH> DomainWallFermionVec5dFH;
+typedef DomainWallFermion<DomainWallVec5dImplDF> DomainWallFermionVec5dDF;
+
 typedef MobiusFermion<DomainWallVec5dImplR> MobiusFermionVec5dR;
 typedef MobiusFermion<DomainWallVec5dImplF> MobiusFermionVec5dF;
 typedef MobiusFermion<DomainWallVec5dImplD> MobiusFermionVec5dD;
 
+typedef MobiusFermion<DomainWallVec5dImplRL> MobiusFermionVec5dRL;
+typedef MobiusFermion<DomainWallVec5dImplFH> MobiusFermionVec5dFH;
+typedef MobiusFermion<DomainWallVec5dImplDF> MobiusFermionVec5dDF;
+
 typedef ZMobiusFermion<ZDomainWallVec5dImplR> ZMobiusFermionVec5dR;
 typedef ZMobiusFermion<ZDomainWallVec5dImplF> ZMobiusFermionVec5dF;
 typedef ZMobiusFermion<ZDomainWallVec5dImplD> ZMobiusFermionVec5dD;
 
+typedef ZMobiusFermion<ZDomainWallVec5dImplRL> ZMobiusFermionVec5dRL;
+typedef ZMobiusFermion<ZDomainWallVec5dImplFH> ZMobiusFermionVec5dFH;
+typedef ZMobiusFermion<ZDomainWallVec5dImplDF> ZMobiusFermionVec5dDF;
 
 typedef ScaledShamirFermion<WilsonImplR> ScaledShamirFermionR;
 typedef ScaledShamirFermion<WilsonImplF> ScaledShamirFermionF;
@@ -166,17 +193,35 @@ typedef OverlapWilsonPartialFractionZolotarevFermion<WilsonImplD> OverlapWilsonP
 typedef WilsonFermion<GparityWilsonImplR>     GparityWilsonFermionR;
 typedef WilsonFermion<GparityWilsonImplF>     GparityWilsonFermionF;
 typedef WilsonFermion<GparityWilsonImplD>     GparityWilsonFermionD;
+
+typedef WilsonFermion<GparityWilsonImplRL>     GparityWilsonFermionRL;
+typedef WilsonFermion<GparityWilsonImplFH>     GparityWilsonFermionFH;
+typedef WilsonFermion<GparityWilsonImplDF>     GparityWilsonFermionDF;
+
 typedef DomainWallFermion<GparityWilsonImplR> GparityDomainWallFermionR;
 typedef DomainWallFermion<GparityWilsonImplF> GparityDomainWallFermionF;
 typedef DomainWallFermion<GparityWilsonImplD> GparityDomainWallFermionD;
 
+typedef DomainWallFermion<GparityWilsonImplRL> GparityDomainWallFermionRL;
+typedef DomainWallFermion<GparityWilsonImplFH> GparityDomainWallFermionFH;
+typedef DomainWallFermion<GparityWilsonImplDF> GparityDomainWallFermionDF;
+
 typedef WilsonTMFermion<GparityWilsonImplR> GparityWilsonTMFermionR;
 typedef WilsonTMFermion<GparityWilsonImplF> GparityWilsonTMFermionF;
 typedef WilsonTMFermion<GparityWilsonImplD> GparityWilsonTMFermionD;
+
+typedef WilsonTMFermion<GparityWilsonImplRL> GparityWilsonTMFermionRL;
+typedef WilsonTMFermion<GparityWilsonImplFH> GparityWilsonTMFermionFH;
+typedef WilsonTMFermion<GparityWilsonImplDF> GparityWilsonTMFermionDF;
+
 typedef MobiusFermion<GparityWilsonImplR> GparityMobiusFermionR;
 typedef MobiusFermion<GparityWilsonImplF> GparityMobiusFermionF;
 typedef MobiusFermion<GparityWilsonImplD> GparityMobiusFermionD;
 
+typedef MobiusFermion<GparityWilsonImplRL> GparityMobiusFermionRL;
+typedef MobiusFermion<GparityWilsonImplFH> GparityMobiusFermionFH;
+typedef MobiusFermion<GparityWilsonImplDF> GparityMobiusFermionDF;
+
 typedef ImprovedStaggeredFermion<StaggeredImplR> ImprovedStaggeredFermionR;
 typedef ImprovedStaggeredFermion<StaggeredImplF> ImprovedStaggeredFermionF;
 typedef ImprovedStaggeredFermion<StaggeredImplD> ImprovedStaggeredFermionD;

lib/qcd/action/fermion/FermionCore.h

@@ -55,7 +55,14 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
   template class A<ZWilsonImplF>;		\
   template class A<ZWilsonImplD>;		\
   template class A<GparityWilsonImplF>;		\
-  template class A<GparityWilsonImplD>;		
+  template class A<GparityWilsonImplD>;		\
+  template class A<WilsonImplFH>;		\
+  template class A<WilsonImplDF>;		\
+  template class A<ZWilsonImplFH>;		\
+  template class A<ZWilsonImplDF>;		\
+  template class A<GparityWilsonImplFH>;		\
+  template class A<GparityWilsonImplDF>;		
+
 
 #define AdjointFermOpTemplateInstantiate(A) \
   template class A<WilsonAdjImplF>; \
@@ -69,7 +76,11 @@ Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
   template class A<DomainWallVec5dImplF>;	\
   template class A<DomainWallVec5dImplD>;	\
   template class A<ZDomainWallVec5dImplF>;	\
-  template class A<ZDomainWallVec5dImplD>;	
+  template class A<ZDomainWallVec5dImplD>;	\
+  template class A<DomainWallVec5dImplFH>;	\
+  template class A<DomainWallVec5dImplDF>;	\
+  template class A<ZDomainWallVec5dImplFH>;	\
+  template class A<ZDomainWallVec5dImplDF>;	
 
 #define FermOpTemplateInstantiate(A) \
  FermOp4dVecTemplateInstantiate(A) \

lib/qcd/action/fermion/FermionOperatorImpl.h

@@ -874,17 +874,17 @@ typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffReal > WilsonImplR
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffReal > WilsonImplF;  // Float
 typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffReal > WilsonImplD;  // Double
 
-typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRH;  // Real.. whichever prec
+typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffRealHalfComms > WilsonImplRL;  // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplFH;  // Float
-typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDH;  // Double
+typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffRealHalfComms > WilsonImplDF;  // Double
 
 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<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRH; // Real.. whichever prec
+typedef WilsonImpl<vComplex,  FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplRL; // Real.. whichever prec
 typedef WilsonImpl<vComplexF, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplFH; // Float
-typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDH; // Double
+typedef WilsonImpl<vComplexD, FundamentalRepresentation, CoeffComplexHalfComms > ZWilsonImplDF; // Double
  
 typedef WilsonImpl<vComplex,  AdjointRepresentation, CoeffReal > WilsonAdjImplR;   // Real.. whichever prec
 typedef WilsonImpl<vComplexF, AdjointRepresentation, CoeffReal > WilsonAdjImplF;  // Float
@@ -898,25 +898,25 @@ typedef DomainWallVec5dImpl<vComplex ,Nc, CoeffReal> DomainWallVec5dImplR; // Re
 typedef DomainWallVec5dImpl<vComplexF,Nc, CoeffReal> DomainWallVec5dImplF; // Float
 typedef DomainWallVec5dImpl<vComplexD,Nc, CoeffReal> DomainWallVec5dImplD; // Double
  
-typedef DomainWallVec5dImpl<vComplex ,Nc, CoeffRealHalfComms> DomainWallVec5dImplRH; // Real.. whichever prec
+typedef DomainWallVec5dImpl<vComplex ,Nc, CoeffRealHalfComms> DomainWallVec5dImplRL; // Real.. whichever prec
 typedef DomainWallVec5dImpl<vComplexF,Nc, CoeffRealHalfComms> DomainWallVec5dImplFH; // Float
-typedef DomainWallVec5dImpl<vComplexD,Nc, CoeffRealHalfComms> DomainWallVec5dImplDH; // Double
+typedef DomainWallVec5dImpl<vComplexD,Nc, CoeffRealHalfComms> DomainWallVec5dImplDF; // Double
  
 typedef DomainWallVec5dImpl<vComplex ,Nc,CoeffComplex> ZDomainWallVec5dImplR; // Real.. whichever prec
 typedef DomainWallVec5dImpl<vComplexF,Nc,CoeffComplex> ZDomainWallVec5dImplF; // Float
 typedef DomainWallVec5dImpl<vComplexD,Nc,CoeffComplex> ZDomainWallVec5dImplD; // Double
  
-typedef DomainWallVec5dImpl<vComplex ,Nc,CoeffComplexHalfComms> ZDomainWallVec5dImplRH; // Real.. whichever prec
+typedef DomainWallVec5dImpl<vComplex ,Nc,CoeffComplexHalfComms> ZDomainWallVec5dImplRL; // Real.. whichever prec
 typedef DomainWallVec5dImpl<vComplexF,Nc,CoeffComplexHalfComms> ZDomainWallVec5dImplFH; // Float
-typedef DomainWallVec5dImpl<vComplexD,Nc,CoeffComplexHalfComms> ZDomainWallVec5dImplDH; // Double
+typedef DomainWallVec5dImpl<vComplexD,Nc,CoeffComplexHalfComms> ZDomainWallVec5dImplDF; // Double
  
 typedef GparityWilsonImpl<vComplex , Nc,CoeffReal> GparityWilsonImplR;  // Real.. whichever prec
 typedef GparityWilsonImpl<vComplexF, Nc,CoeffReal> GparityWilsonImplF;  // Float
 typedef GparityWilsonImpl<vComplexD, Nc,CoeffReal> GparityWilsonImplD;  // Double
  
-typedef GparityWilsonImpl<vComplex , Nc,CoeffRealHalfComms> GparityWilsonImplRH;  // Real.. whichever prec
+typedef GparityWilsonImpl<vComplex , Nc,CoeffRealHalfComms> GparityWilsonImplRL;  // Real.. whichever prec
 typedef GparityWilsonImpl<vComplexF, Nc,CoeffRealHalfComms> GparityWilsonImplFH;  // Float
-typedef GparityWilsonImpl<vComplexD, Nc,CoeffRealHalfComms> GparityWilsonImplDH;  // Double
+typedef GparityWilsonImpl<vComplexD, Nc,CoeffRealHalfComms> GparityWilsonImplDF;  // Double
 
 typedef StaggeredImpl<vComplex,  FundamentalRepresentation > StaggeredImplR;   // Real.. whichever prec
 typedef StaggeredImpl<vComplexF, FundamentalRepresentation > StaggeredImplF;  // Float

lib/qcd/action/fermion/WilsonFermion5D.cc

@@ -679,7 +679,6 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHw(FermionField &out,const Fe
 
 }
 
-
 FermOpTemplateInstantiate(WilsonFermion5D);
 GparityFermOpTemplateInstantiate(WilsonFermion5D);
   

lib/qcd/action/fermion/WilsonKernelsAsm.cc

@@ -112,5 +112,16 @@ INSTANTIATE_ASM(DomainWallVec5dImplD);
 INSTANTIATE_ASM(ZDomainWallVec5dImplF);
 INSTANTIATE_ASM(ZDomainWallVec5dImplD);
 
+INSTANTIATE_ASM(WilsonImplFH);
+INSTANTIATE_ASM(WilsonImplDF);
+INSTANTIATE_ASM(ZWilsonImplFH);
+INSTANTIATE_ASM(ZWilsonImplDF);
+INSTANTIATE_ASM(GparityWilsonImplFH);
+INSTANTIATE_ASM(GparityWilsonImplDF);
+INSTANTIATE_ASM(DomainWallVec5dImplFH);
+INSTANTIATE_ASM(DomainWallVec5dImplDF);
+INSTANTIATE_ASM(ZDomainWallVec5dImplFH);
+INSTANTIATE_ASM(ZDomainWallVec5dImplDF);
+
 }}
 

lib/qcd/action/fermion/WilsonKernelsHand.cc

@@ -829,6 +829,36 @@ WilsonKernels<GparityWilsonImplD>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder
   assert(0);
 }
 
+template<> void 
+WilsonKernels<GparityWilsonImplFH>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
+							SiteHalfSpinor *buf,
+							int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
+{
+  assert(0);
+}
+
+template<> void 
+WilsonKernels<GparityWilsonImplFH>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,
+							   SiteHalfSpinor *buf,
+							   int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
+{
+  assert(0);
+}
+
+template<> void 
+WilsonKernels<GparityWilsonImplDF>::HandDhopSite(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
+							int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
+{
+  assert(0);
+}
+
+template<> void 
+WilsonKernels<GparityWilsonImplDF>::HandDhopSiteDag(StencilImpl &st,LebesgueOrder &lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
+							   int sF,int sU,const FermionField &in, FermionField &out,int internal,int external)
+{
+  assert(0);
+}
+
 
 
 ////////////// Wilson ; uses this implementation /////////////////////
@@ -850,5 +880,15 @@ INSTANTIATE_THEM(DomainWallVec5dImplF);
 INSTANTIATE_THEM(DomainWallVec5dImplD);
 INSTANTIATE_THEM(ZDomainWallVec5dImplF);
 INSTANTIATE_THEM(ZDomainWallVec5dImplD);
+INSTANTIATE_THEM(WilsonImplFH);
+INSTANTIATE_THEM(WilsonImplDF);
+INSTANTIATE_THEM(ZWilsonImplFH);
+INSTANTIATE_THEM(ZWilsonImplDF);
+INSTANTIATE_THEM(GparityWilsonImplFH);
+INSTANTIATE_THEM(GparityWilsonImplDF);
+INSTANTIATE_THEM(DomainWallVec5dImplFH);
+INSTANTIATE_THEM(DomainWallVec5dImplDF);
+INSTANTIATE_THEM(ZDomainWallVec5dImplFH);
+INSTANTIATE_THEM(ZDomainWallVec5dImplDF);
 
 }}

lib/stencil/Stencil.h

@@ -901,8 +901,9 @@ class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal
     int reduced_buffer_size = buffer_size;
     if (cbmask != 0x3) reduced_buffer_size=buffer_size>>1;
 
-    int bytes = (reduced_buffer_size*sizeof(cobj))/simd_layout;
+    int datum_bytes = compress.CommDatumSize();
-    assert(bytes*simd_layout == reduced_buffer_size*sizeof(cobj));
+    int bytes = (reduced_buffer_size*datum_bytes)/simd_layout;
+    assert(bytes*simd_layout == reduced_buffer_size*datum_bytes);
 
     std::vector<cobj *> rpointers(maxl);
     std::vector<cobj *> spointers(maxl);

lib/tensors/Tensor_traits.h

@@ -142,6 +142,17 @@ namespace Grid {
     typedef vRealD Realified;
     enum { TensorLevel = 0 };
   };
+  template<> class GridTypeMapper<vComplexH> {
+  public:
+    typedef ComplexF  scalar_type;
+    typedef vComplexH vector_type;
+    typedef vComplexD vector_typeD;
+    typedef vComplexH tensor_reduced;
+    typedef ComplexF  scalar_object;
+    typedef vComplexH Complexified;
+    typedef vRealH Realified;
+    enum { TensorLevel = 0 };
+  };
   template<> class GridTypeMapper<vComplexF> {
   public:
     typedef ComplexF  scalar_type;