diff --git a/lib/qcd/action/fermion/FermionOperatorImpl.h b/lib/qcd/action/fermion/FermionOperatorImpl.h
index 1d90cace..580f5e30 100644
--- a/lib/qcd/action/fermion/FermionOperatorImpl.h
+++ b/lib/qcd/action/fermion/FermionOperatorImpl.h
@@ -26,7 +26,7 @@ namespace Grid {
// and Methods:
// void ImportGauge(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
// void DoubleStore(GridBase *GaugeGrid,DoubledGaugeField &Uds,const GaugeField &Umu)
- // void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,CartesianStencil &St)
+ // void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,StencilImpl &St)
// void InsertForce4D(GaugeField &mat,const FermionField &Btilde,const FermionField &A,int mu)
// void InsertForce5D(GaugeField &mat,const FermionField &Btilde,const FermionField &A,int mu)
//
@@ -101,6 +101,7 @@ namespace Grid {
typedef typename Impl::SiteSpinor SiteSpinor; \
typedef typename Impl::SiteHalfSpinor SiteHalfSpinor; \
typedef typename Impl::Compressor Compressor; \
+ typedef typename Impl::StencilImpl StencilImpl; \
typedef typename Impl::ImplParams ImplParams;
///////
@@ -112,7 +113,6 @@ namespace Grid {
typedef ImplGauge<S,Nrepresentation> Gimpl;
-
INHERIT_GIMPL_TYPES(Gimpl);
template<typename vtype> using iImplSpinor = iScalar<iVector<iVector<vtype, Nrepresentation>, Ns> >;
@@ -128,10 +128,11 @@ namespace Grid {
typedef WilsonCompressor<SiteHalfSpinor,SiteSpinor> Compressor;
typedef WilsonImplParams ImplParams;
+ typedef CartesianStencil<SiteSpinor,SiteHalfSpinor,Compressor> StencilImpl;
ImplParams Params;
WilsonImpl(const ImplParams &p= ImplParams()) : Params(p) {};
- inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,CartesianStencil &St){
+ inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,StencilImpl &St){
mult(&phi(),&U(mu),&chi());
}
@@ -198,13 +199,15 @@ PARALLEL_FOR_LOOP
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef WilsonCompressor<SiteHalfSpinor,SiteSpinor> Compressor;
+ typedef CartesianStencil<SiteSpinor,SiteHalfSpinor,Compressor> StencilImpl;
typedef GparityWilsonImplParams ImplParams;
ImplParams Params;
GparityWilsonImpl(const ImplParams &p= ImplParams()) : Params(p) {};
+
// provide the multiply by link that is differentiated between Gparity (with flavour index) and non-Gparity
- inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,CartesianStencil &St){
+ inline void multLink(SiteHalfSpinor &phi,const SiteDoubledGaugeField &U,const SiteHalfSpinor &chi,int mu,StencilEntry *SE,StencilImpl &St){
typedef SiteHalfSpinor vobj;
typedef typename SiteHalfSpinor::scalar_object sobj;
diff --git a/lib/qcd/action/fermion/WilsonFermion.cc b/lib/qcd/action/fermion/WilsonFermion.cc
index 053526cd..5e37238c 100644
--- a/lib/qcd/action/fermion/WilsonFermion.cc
+++ b/lib/qcd/action/fermion/WilsonFermion.cc
@@ -109,7 +109,7 @@ namespace QCD {
///////////////////////////////////
template<class Impl>
- void WilsonFermion<Impl>::DerivInternal(CartesianStencil & st,
+ void WilsonFermion<Impl>::DerivInternal(StencilImpl & st,
DoubledGaugeField & U,
GaugeField &mat,
const FermionField &A,
@@ -123,7 +123,7 @@ namespace QCD {
FermionField Atilde(B._grid);
Atilde = A;
- st.HaloExchange<SiteSpinor,SiteHalfSpinor,Compressor>(B,comm_buf,compressor);
+ st.HaloExchange(B,comm_buf,compressor);
for(int mu=0;mu<Nd;mu++){
@@ -242,7 +242,7 @@ PARALLEL_FOR_LOOP
Compressor compressor(dag);
- Stencil.HaloExchange<SiteSpinor,SiteHalfSpinor,Compressor>(in,comm_buf,compressor);
+ Stencil.HaloExchange(in,comm_buf,compressor);
PARALLEL_FOR_LOOP
for(int sss=0;sss<in._grid->oSites();sss++){
@@ -253,13 +253,13 @@ PARALLEL_FOR_LOOP
template<class Impl>
- void WilsonFermion<Impl>::DhopInternal(CartesianStencil & st,DoubledGaugeField & U,
+ void WilsonFermion<Impl>::DhopInternal(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) {
assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor(dag);
- st.HaloExchange<SiteSpinor,SiteHalfSpinor,Compressor>(in,comm_buf,compressor);
+ st.HaloExchange(in,comm_buf,compressor);
if ( dag == DaggerYes ) {
if( HandOptDslash ) {
diff --git a/lib/qcd/action/fermion/WilsonFermion.h b/lib/qcd/action/fermion/WilsonFermion.h
index bab8ff15..9a8c848a 100644
--- a/lib/qcd/action/fermion/WilsonFermion.h
+++ b/lib/qcd/action/fermion/WilsonFermion.h
@@ -73,14 +73,14 @@ namespace Grid {
///////////////////////////////////////////////////////////////
// Extra methods added by derived
///////////////////////////////////////////////////////////////
- void DerivInternal(CartesianStencil & st,
+ void DerivInternal(StencilImpl & st,
DoubledGaugeField & U,
GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag);
- void DhopInternal(CartesianStencil & st,DoubledGaugeField & U,
+ void DhopInternal(StencilImpl & st,DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag) ;
@@ -108,9 +108,9 @@ namespace Grid {
GridBase * _cbgrid;
//Defines the stencils for even and odd
- CartesianStencil Stencil;
- CartesianStencil StencilEven;
- CartesianStencil StencilOdd;
+ StencilImpl Stencil;
+ StencilImpl StencilEven;
+ StencilImpl StencilOdd;
// Copy of the gauge field , with even and odd subsets
DoubledGaugeField Umu;
diff --git a/lib/qcd/action/fermion/WilsonFermion5D.cc b/lib/qcd/action/fermion/WilsonFermion5D.cc
index 66ca67d5..8f22bc32 100644
--- a/lib/qcd/action/fermion/WilsonFermion5D.cc
+++ b/lib/qcd/action/fermion/WilsonFermion5D.cc
@@ -68,6 +68,8 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
comm_buf.resize(Stencil._unified_buffer_size); // this is always big enough to contain EO
ImportGauge(_Umu);
+ commtime=0;
+ dslashtime=0;
}
template<class Impl>
void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
@@ -85,7 +87,7 @@ void WilsonFermion5D<Impl>::DhopDir(const FermionField &in, FermionField &out,in
// assert( (dir>=0)&&(dir<4) ); //must do x,y,z or t;
Compressor compressor(DaggerNo);
- Stencil.HaloExchange<SiteSpinor,SiteHalfSpinor,Compressor>(in,comm_buf,compressor);
+ Stencil.HaloExchange(in,comm_buf,compressor);
int skip = (disp==1) ? 0 : 1;
@@ -105,7 +107,7 @@ PARALLEL_FOR_LOOP
};
template<class Impl>
-void WilsonFermion5D<Impl>::DerivInternal(CartesianStencil & st,
+void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
DoubledGaugeField & U,
GaugeField &mat,
const FermionField &A,
@@ -122,7 +124,7 @@ void WilsonFermion5D<Impl>::DerivInternal(CartesianStencil & st,
FermionField Btilde(B._grid);
FermionField Atilde(B._grid);
- st.HaloExchange<SiteSpinor,SiteHalfSpinor,Compressor>(B,comm_buf,compressor);
+ st.HaloExchange(B,comm_buf,compressor);
Atilde=A;
@@ -194,6 +196,27 @@ void WilsonFermion5D<Impl>::DhopDerivEO(GaugeField &mat,
DerivInternal(StencilOdd,UmuEven,mat,A,B,dag);
}
+
+template<class Impl>
+void WilsonFermion5D<Impl>::Report(void)
+{
+ std::cout<<GridLogMessage << "********************"<<std::endl;
+ std::cout<<GridLogMessage << "Halo time "<<commtime <<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Dslash time "<<dslashtime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil All time "<<Stencil.halotime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "********************"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil nosplice time "<<Stencil.nosplicetime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil gather time "<<Stencil.gathertime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil comm time "<<Stencil.commtime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil scattertime "<<Stencil.scattertime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "********************"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil splice time "<<Stencil.splicetime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil comm time "<<Stencil.commstime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil gathremtime "<<Stencil.gathermtime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil merge time "<<Stencil.mergetime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "Stencil buf time "<<Stencil.buftime<<" us"<<std::endl;
+ std::cout<<GridLogMessage << "********************"<<std::endl;
+}
template<class Impl>
void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
const FermionField &A,
@@ -212,7 +235,7 @@ void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
}
template<class Impl>
-void WilsonFermion5D<Impl>::DhopInternal(CartesianStencil & st, LebesgueOrder &lo,
+void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
@@ -220,13 +243,16 @@ void WilsonFermion5D<Impl>::DhopInternal(CartesianStencil & st, LebesgueOrder &l
Compressor compressor(dag);
- st.HaloExchange<SiteSpinor,SiteHalfSpinor,Compressor>(in,comm_buf,compressor);
+ commtime -=usecond();
+ st.HaloExchange(in,comm_buf,compressor);
+ commtime +=usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
// Not loop ordering and data layout.
// Designed to create
// - per thread reuse in L1 cache for U
// - 8 linear access unit stride streams per thread for Fermion for hw prefetchable.
+ dslashtime -=usecond();
if ( dag == DaggerYes ) {
if( this->HandOptDslash ) {
PARALLEL_FOR_LOOP
@@ -274,6 +300,7 @@ PARALLEL_FOR_LOOP
}
}
}
+ dslashtime +=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
diff --git a/lib/qcd/action/fermion/WilsonFermion5D.h b/lib/qcd/action/fermion/WilsonFermion5D.h
index b0d25309..e95fc2b3 100644
--- a/lib/qcd/action/fermion/WilsonFermion5D.h
+++ b/lib/qcd/action/fermion/WilsonFermion5D.h
@@ -31,7 +31,8 @@ namespace Grid {
public:
INHERIT_IMPL_TYPES(Impl);
typedef WilsonKernels<Impl> Kernels;
-
+ double commtime;
+ double dslashtime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////
@@ -72,14 +73,14 @@ namespace Grid {
///////////////////////////////////////////////////////////////
// New methods added
///////////////////////////////////////////////////////////////
- void DerivInternal(CartesianStencil & st,
+ void DerivInternal(StencilImpl & st,
DoubledGaugeField & U,
GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag);
- void DhopInternal(CartesianStencil & st,
+ void DhopInternal(StencilImpl & st,
LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
@@ -97,6 +98,7 @@ namespace Grid {
// DoubleStore
void ImportGauge(const GaugeField &_Umu);
+ void Report(void);
///////////////////////////////////////////////////////////////
// Data members require to support the functionality
///////////////////////////////////////////////////////////////
@@ -112,9 +114,9 @@ namespace Grid {
int Ls;
//Defines the stencils for even and odd
- CartesianStencil Stencil;
- CartesianStencil StencilEven;
- CartesianStencil StencilOdd;
+ StencilImpl Stencil;
+ StencilImpl StencilEven;
+ StencilImpl StencilOdd;
// Copy of the gauge field , with even and odd subsets
DoubledGaugeField Umu;
diff --git a/lib/qcd/action/fermion/WilsonKernels.cc b/lib/qcd/action/fermion/WilsonKernels.cc
index a897921a..a37e7f9e 100644
--- a/lib/qcd/action/fermion/WilsonKernels.cc
+++ b/lib/qcd/action/fermion/WilsonKernels.cc
@@ -3,7 +3,7 @@ namespace Grid {
namespace QCD {
template<class Impl>
-void WilsonKernels<Impl>::DiracOptDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+void WilsonKernels<Impl>::DiracOptDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
@@ -122,7 +122,7 @@ void WilsonKernels<Impl>::DiracOptDhopSite(CartesianStencil &st,DoubledGaugeFiel
};
template<class Impl>
-void WilsonKernels<Impl>::DiracOptDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
+void WilsonKernels<Impl>::DiracOptDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
@@ -241,7 +241,7 @@ void WilsonKernels<Impl>::DiracOptDhopSiteDag(CartesianStencil &st,DoubledGaugeF
}
template<class Impl>
-void WilsonKernels<Impl>::DiracOptDhopDir(CartesianStencil &st,DoubledGaugeField &U,
+void WilsonKernels<Impl>::DiracOptDhopDir(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,int dir,int gamma)
{
diff --git a/lib/qcd/action/fermion/WilsonKernels.h b/lib/qcd/action/fermion/WilsonKernels.h
index 78d7c6a0..9696fad5 100644
--- a/lib/qcd/action/fermion/WilsonKernels.h
+++ b/lib/qcd/action/fermion/WilsonKernels.h
@@ -17,36 +17,36 @@ namespace Grid {
typedef FermionOperator<Impl> Base;
public:
- void DiracOptDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+ void DiracOptDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out);
- void DiracOptDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
+ void DiracOptDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in,FermionField &out);
- void DiracOptDhopDir(CartesianStencil &st,DoubledGaugeField &U,
+ void DiracOptDhopDir(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out,int dirdisp,int gamma);
#define HANDOPT
#ifdef HANDOPT
- void DiracOptHandDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+ void DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out);
- void DiracOptHandDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
+ void DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out);
#else
- void DiracOptHandDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+ void DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
DiracOptDhopSite(st,U,buf,sF,sU,in,out); // will template override for Wilson Nc=3
}
- void DiracOptHandDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
+ void DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int sF,int sU,const FermionField &in, FermionField &out)
{
diff --git a/lib/qcd/action/fermion/WilsonKernelsAsm.cc b/lib/qcd/action/fermion/WilsonKernelsAsm.cc
index 9ccd18c6..fedfdd00 100644
--- a/lib/qcd/action/fermion/WilsonKernelsAsm.cc
+++ b/lib/qcd/action/fermion/WilsonKernelsAsm.cc
@@ -73,7 +73,7 @@ namespace Grid {
namespace QCD {
template<class Impl>
-void WilsonKernels<Impl >::DiracOptAsmDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+void WilsonKernels<Impl >::DiracOptAsmDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out,uint64_t *timers)
{
diff --git a/lib/qcd/action/fermion/WilsonKernelsHand.cc b/lib/qcd/action/fermion/WilsonKernelsHand.cc
index fee1b0fb..b263ea83 100644
--- a/lib/qcd/action/fermion/WilsonKernelsHand.cc
+++ b/lib/qcd/action/fermion/WilsonKernelsHand.cc
@@ -282,7 +282,7 @@ namespace QCD {
#ifdef HANDOPT
template<class Impl>
-void WilsonKernels<Impl >::DiracOptHandDhopSite(CartesianStencil &st,DoubledGaugeField &U,
+void WilsonKernels<Impl >::DiracOptHandDhopSite(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out)
{
@@ -526,7 +526,7 @@ void WilsonKernels<Impl >::DiracOptHandDhopSite(CartesianStencil &st,DoubledGaug
}
template<class Impl>
-void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(CartesianStencil &st,DoubledGaugeField &U,
+void WilsonKernels<Impl >::DiracOptHandDhopSiteDag(StencilImpl &st,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int sU,const FermionField &in, FermionField &out)
{