portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Compiles GPU and CPU, still gives good performance on CPU

Browse Source
This commit is contained in:
Peter Boyle 2019-06-05 13:28:16 +01:00
parent 18d3cde29a
commit 0ee6e77cbc
71 changed files with 1512 additions and 33769 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

668
Grid/qcd/action/fermion/CayleyFermion5D.cc
View File

@ -1,668 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid_Eigen_Dense.h>
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/CayleyFermion5D.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
CayleyFermion5D<Impl>::CayleyFermion5D(GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mass,RealD _M5,const ImplParams &p) :
WilsonFermion5D<Impl>(_Umu,
FiveDimGrid,
FiveDimRedBlackGrid,
FourDimGrid,
FourDimRedBlackGrid,_M5,p),
mass(_mass)
{
}
///////////////////////////////////////////////////////////////
// Physical surface field utilities
///////////////////////////////////////////////////////////////
template<class Impl>
void CayleyFermion5D<Impl>::ExportPhysicalFermionSolution(const FermionField &solution5d,FermionField &exported4d)
{
int Ls = this->Ls;
FermionField tmp(this->FermionGrid());
tmp = solution5d;
conformable(solution5d.Grid(),this->FermionGrid());
conformable(exported4d.Grid(),this->GaugeGrid());
axpby_ssp_pminus(tmp, 0., solution5d, 1., solution5d, 0, 0);
axpby_ssp_pplus (tmp, 1., tmp , 1., solution5d, 0, Ls-1);
ExtractSlice(exported4d, tmp, 0, 0);
}
template<class Impl>
void CayleyFermion5D<Impl>::P(const FermionField &psi, FermionField &chi)
{
int Ls= this->Ls;
chi=Zero();
for(int s=0;s<Ls;s++){
axpby_ssp_pminus(chi,1.0,chi,1.0,psi,s,s);
axpby_ssp_pplus (chi,1.0,chi,1.0,psi,s,(s+1)%Ls);
}
}
template<class Impl>
void CayleyFermion5D<Impl>::Pdag(const FermionField &psi, FermionField &chi)
{
int Ls= this->Ls;
chi=Zero();
for(int s=0;s<Ls;s++){
axpby_ssp_pminus(chi,1.0,chi,1.0,psi,s,s);
axpby_ssp_pplus (chi,1.0,chi,1.0,psi,s,(s-1+Ls)%Ls);
}
}
template<class Impl>
void CayleyFermion5D<Impl>::ExportPhysicalFermionSource(const FermionField &solution5d,FermionField &exported4d)
{
int Ls = this->Ls;
FermionField tmp(this->FermionGrid());
tmp = solution5d;
conformable(solution5d.Grid(),this->FermionGrid());
conformable(exported4d.Grid(),this->GaugeGrid());
axpby_ssp_pplus (tmp, 0., solution5d, 1., solution5d, 0, 0);
axpby_ssp_pminus(tmp, 1., tmp , 1., solution5d, 0, Ls-1);
ExtractSlice(exported4d, tmp, 0, 0);
}
template<class Impl>
void CayleyFermion5D<Impl>::ImportUnphysicalFermion(const FermionField &input4d,FermionField &imported5d)
{
int Ls = this->Ls;
FermionField tmp(this->FermionGrid());
conformable(imported5d.Grid(),this->FermionGrid());
conformable(input4d.Grid() ,this->GaugeGrid());
tmp = Zero();
InsertSlice(input4d, tmp, 0 , 0);
InsertSlice(input4d, tmp, Ls-1, 0);
axpby_ssp_pplus (tmp, 0., tmp, 1., tmp, 0, 0);
axpby_ssp_pminus(tmp, 0., tmp, 1., tmp, Ls-1, Ls-1);
imported5d=tmp;
}
template<class Impl>
void CayleyFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d)
{
int Ls = this->Ls;
FermionField tmp(this->FermionGrid());
conformable(imported5d.Grid(),this->FermionGrid());
conformable(input4d.Grid() ,this->GaugeGrid());
tmp = Zero();
InsertSlice(input4d, tmp, 0 , 0);
InsertSlice(input4d, tmp, Ls-1, 0);
axpby_ssp_pplus (tmp, 0., tmp, 1., tmp, 0, 0);
axpby_ssp_pminus(tmp, 0., tmp, 1., tmp, Ls-1, Ls-1);
Dminus(tmp,imported5d);
}
template<class Impl>
void CayleyFermion5D<Impl>::Dminus(const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
FermionField tmp_f(this->FermionGrid());
this->DW(psi,tmp_f,DaggerNo);
for(int s=0;s<Ls;s++){
axpby_ssp(chi,Coeff_t(1.0),psi,-cs[s],tmp_f,s,s);// chi = (1-c[s] D_W) psi
}
}
template<class Impl>
void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
FermionField tmp_f(this->FermionGrid());
this->DW(psi,tmp_f,DaggerYes);
for(int s=0;s<Ls;s++){
axpby_ssp(chi,Coeff_t(1.0),psi,conjugate(-cs[s]),tmp_f,s,s);// chi = (1-c[s] D_W) psi
}
}
template<class Impl> void CayleyFermion5D<Impl>::CayleyReport(void)
{
this->Report();
Coordinate latt = GridDefaultLatt();
RealD volume = this->Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = this->_FourDimGrid->_Nprocessors;
if ( M5Dcalls > 0 ) {
std::cout << GridLogMessage << "#### M5D calls report " << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D Number of M5D Calls : " << M5Dcalls << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls : " << M5Dtime / M5Dcalls << " us" << std::endl;
// Flops = 10.0*(Nc*Ns) *Ls*vol
RealD mflops = 10.0*(Nc*Ns)*volume*M5Dcalls/M5Dtime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
// Bytes = sizeof(Real) * (Nc*Ns*Nreim) * Ls * vol * (read+write) (/2 for red black counting)
// read = 2 ( psi[ss+s+1] and psi[ss+s-1] count as 1 )
// write = 1
RealD Gbytes = sizeof(Real) * (Nc*Ns*2) * volume * 3 /2. * 1.e-9;
std::cout << GridLogMessage << "Average bandwidth (GB/s) : " << Gbytes/M5Dtime*M5Dcalls*1.e6 << std::endl;
}
if ( MooeeInvCalls > 0 ) {
std::cout << GridLogMessage << "#### MooeeInv calls report " << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D Number of MooeeInv Calls : " << MooeeInvCalls << std::endl;
std::cout << GridLogMessage << "CayleyFermion5D ComputeTime/Calls : " << MooeeInvTime / MooeeInvCalls << " us" << std::endl;
#ifdef GRID_NVCC
RealD mflops = ( -16.*Nc*Ns+this->Ls*(1.+18.*Nc*Ns) )*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
#else
// Flops = MADD * Ls *Ls *4dvol * spin/colour/complex
RealD mflops = 2.0*24*this->Ls*volume*MooeeInvCalls/MooeeInvTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
#endif
}
}
template<class Impl> void CayleyFermion5D<Impl>::CayleyZeroCounters(void)
{
this->ZeroCounters();
M5Dflops=0;
M5Dcalls=0;
M5Dtime=0;
MooeeInvFlops=0;
MooeeInvCalls=0;
MooeeInvTime=0;
}
template<class Impl>
void CayleyFermion5D<Impl>::M5D (const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
Vector<Coeff_t> diag (Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1]=mass;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] =mass;
M5D(psi,chi,chi,lower,diag,upper);
}
template<class Impl>
void CayleyFermion5D<Impl>::Meooe5D (const FermionField &psi, FermionField &Din)
{
int Ls=this->Ls;
Vector<Coeff_t> diag = bs;
Vector<Coeff_t> upper= cs;
Vector<Coeff_t> lower= cs;
upper[Ls-1]=-mass*upper[Ls-1];
lower[0] =-mass*lower[0];
M5D(psi,psi,Din,lower,diag,upper);
}
// FIXME Redunant with the above routine; check this and eliminate
template<class Impl> void CayleyFermion5D<Impl>::Meo5D (const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
Vector<Coeff_t> diag = beo;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int i=0;i<Ls;i++) {
upper[i]=-ceo[i];
lower[i]=-ceo[i];
}
upper[Ls-1]=-mass*upper[Ls-1];
lower[0] =-mass*lower[0];
M5D(psi,psi,chi,lower,diag,upper);
}
template<class Impl>
void CayleyFermion5D<Impl>::Mooee (const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
Vector<Coeff_t> diag = bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int i=0;i<Ls;i++) {
upper[i]=-cee[i];
lower[i]=-cee[i];
}
upper[Ls-1]=-mass*upper[Ls-1];
lower[0] =-mass*lower[0];
M5D(psi,psi,chi,lower,diag,upper);
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeDag (const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
Vector<Coeff_t> diag = bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for (int s=0;s<Ls;s++){
// Assemble the 5d matrix
if ( s==0 ) {
upper[s] = -cee[s+1] ;
lower[s] = mass*cee[Ls-1];
} else if ( s==(Ls-1)) {
upper[s] = mass*cee[0];
lower[s] = -cee[s-1];
} else {
upper[s]=-cee[s+1];
lower[s]=-cee[s-1];
}
}
// Conjugate the terms
for (int s=0;s<Ls;s++){
diag[s] =conjugate(diag[s]);
upper[s]=conjugate(upper[s]);
lower[s]=conjugate(lower[s]);
}
M5Ddag(psi,psi,chi,lower,diag,upper);
}
template<class Impl>
void CayleyFermion5D<Impl>::M5Ddag (const FermionField &psi, FermionField &chi)
{
int Ls=this->Ls;
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0);
Vector<Coeff_t> lower(Ls,-1.0);
upper[Ls-1]=-mass*upper[Ls-1];
lower[0] =-mass*lower[0];
M5Ddag(psi,chi,chi,lower,diag,upper);
}
template<class Impl>
void CayleyFermion5D<Impl>::MeooeDag5D (const FermionField &psi, FermionField &Din)
{
int Ls=this->Ls;
Vector<Coeff_t> diag =bs;
Vector<Coeff_t> upper=cs;
Vector<Coeff_t> lower=cs;
for (int s=0;s<Ls;s++){
if ( s== 0 ) {
upper[s] = cs[s+1];
lower[s] =-mass*cs[Ls-1];
} else if ( s==(Ls-1) ) {
upper[s] =-mass*cs[0];
lower[s] = cs[s-1];
} else {
upper[s] = cs[s+1];
lower[s] = cs[s-1];
}
upper[s] = conjugate(upper[s]);
lower[s] = conjugate(lower[s]);
diag[s] = conjugate(diag[s]);
}
M5Ddag(psi,psi,Din,lower,diag,upper);
}
template<class Impl>
RealD CayleyFermion5D<Impl>::M (const FermionField &psi, FermionField &chi)
{
FermionField Din(psi.Grid());
// Assemble Din
Meooe5D(psi,Din);
this->DW(Din,chi,DaggerNo);
// ((b D_W + D_w hop terms +1) on s-diag
axpby(chi,1.0,1.0,chi,psi);
M5D(psi,chi);
return(norm2(chi));
}
template<class Impl>
RealD CayleyFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
{
// Under adjoint
//D1+ D1- P- -> D1+^dag P+ D2-^dag
//D2- P+ D2+ P-D1-^dag D2+dag
FermionField Din(psi.Grid());
// Apply Dw
this->DW(psi,Din,DaggerYes);
MeooeDag5D(Din,chi);
M5Ddag(psi,chi);
// ((b D_W + D_w hop terms +1) on s-diag
axpby (chi,1.0,1.0,chi,psi);
return norm2(chi);
}
// half checkerboard operations
template<class Impl>
void CayleyFermion5D<Impl>::Meooe (const FermionField &psi, FermionField &chi)
{
Meooe5D(psi,this->tmp());
if ( psi.Checkerboard() == Odd ) {
this->DhopEO(this->tmp(),chi,DaggerNo);
} else {
this->DhopOE(this->tmp(),chi,DaggerNo);
}
}
template<class Impl>
void CayleyFermion5D<Impl>::MeooeDag (const FermionField &psi, FermionField &chi)
{
// Apply 4d dslash
if ( psi.Checkerboard() == Odd ) {
this->DhopEO(psi,this->tmp(),DaggerYes);
} else {
this->DhopOE(psi,this->tmp(),DaggerYes);
}
MeooeDag5D(this->tmp(),chi);
}
template<class Impl>
void CayleyFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,int dir,int disp){
Meo5D(psi,this->tmp());
// Apply 4d dslash fragment
this->DhopDir(this->tmp(),chi,dir,disp);
}
// force terms; five routines; default to Dhop on diagonal
template<class Impl>
void CayleyFermion5D<Impl>::MDeriv (GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
FermionField Din(V.Grid());
if ( dag == DaggerNo ) {
// U d/du [D_w D5] V = U d/du DW D5 V
Meooe5D(V,Din);
this->DhopDeriv(mat,U,Din,dag);
} else {
// U d/du [D_w D5]^dag V = U D5^dag d/du DW^dag Y // implicit adj on U in call
Meooe5D(U,Din);
this->DhopDeriv(mat,Din,V,dag);
}
};
template<class Impl>
void CayleyFermion5D<Impl>::MoeDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
FermionField Din(V.Grid());
if ( dag == DaggerNo ) {
// U d/du [D_w D5] V = U d/du DW D5 V
Meooe5D(V,Din);
this->DhopDerivOE(mat,U,Din,dag);
} else {
// U d/du [D_w D5]^dag V = U D5^dag d/du DW^dag Y // implicit adj on U in call
Meooe5D(U,Din);
this->DhopDerivOE(mat,Din,V,dag);
}
};
template<class Impl>
void CayleyFermion5D<Impl>::MeoDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
FermionField Din(V.Grid());
if ( dag == DaggerNo ) {
// U d/du [D_w D5] V = U d/du DW D5 V
Meooe5D(V,Din);
this->DhopDerivEO(mat,U,Din,dag);
} else {
// U d/du [D_w D5]^dag V = U D5^dag d/du DW^dag Y // implicit adj on U in call
Meooe5D(U,Din);
this->DhopDerivEO(mat,Din,V,dag);
}
};
// Tanh
template<class Impl>
void CayleyFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD b,RealD c)
{
Vector<Coeff_t> gamma(this->Ls);
for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
SetCoefficientsInternal(1.0,gamma,b,c);
}
//Zolo
template<class Impl>
void CayleyFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata,RealD b,RealD c)
{
Vector<Coeff_t> gamma(this->Ls);
for(int s=0;s<this->Ls;s++) gamma[s] = zdata->gamma[s];
SetCoefficientsInternal(zolo_hi,gamma,b,c);
}
//Zolo
template<class Impl>
void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,Vector<Coeff_t> & gamma,RealD b,RealD c)
{
int Ls=this->Ls;
///////////////////////////////////////////////////////////
// The Cayley coeffs (unprec)
///////////////////////////////////////////////////////////
assert(gamma.size()==Ls);
omega.resize(Ls);
bs.resize(Ls);
cs.resize(Ls);
as.resize(Ls);
//
// Ts = ( [bs+cs]Dw )^-1 ( (bs+cs) Dw )
// -(g5 ------- -1 ) ( g5 --------- + 1 )
// ( {2+(bs-cs)Dw} ) ( 2+(bs-cs) Dw )
//
// bs = 1/2( (1/omega_s + 1)*b + (1/omega - 1)*c ) = 1/2( 1/omega(b+c) + (b-c) )
// cs = 1/2( (1/omega_s - 1)*b + (1/omega + 1)*c ) = 1/2( 1/omega(b+c) - (b-c) )
//
// bs+cs = 0.5*( 1/omega(b+c) + (b-c) + 1/omega(b+c) - (b-c) ) = 1/omega(b+c)
// bs-cs = 0.5*( 1/omega(b+c) + (b-c) - 1/omega(b+c) + (b-c) ) = b-c
//
// So
//
// Ts = ( [b+c]Dw/omega_s )^-1 ( (b+c) Dw /omega_s )
// -(g5 ------- -1 ) ( g5 --------- + 1 )
// ( {2+(b-c)Dw} ) ( 2+(b-c) Dw )
//
// Ts = ( [b+c]Dw )^-1 ( (b+c) Dw )
// -(g5 ------- -omega_s) ( g5 --------- + omega_s )
// ( {2+(b-c)Dw} ) ( 2+(b-c) Dw )
//
double bpc = b+c;
double bmc = b-c;
_b = b;
_c = c;
_gamma = gamma; // Save the parameters so we can change mass later.
_zolo_hi= zolo_hi;
for(int i=0; i < Ls; i++){
as[i] = 1.0;
omega[i] = _gamma[i]*_zolo_hi; //NB reciprocal relative to Chroma NEF code
assert(omega[i]!=Coeff_t(0.0));
bs[i] = 0.5*(bpc/omega[i] + bmc);
cs[i] = 0.5*(bpc/omega[i] - bmc);
}
////////////////////////////////////////////////////////
// Constants for the preconditioned matrix Cayley form
////////////////////////////////////////////////////////
bee.resize(Ls);
cee.resize(Ls);
beo.resize(Ls);
ceo.resize(Ls);
for(int i=0;i<Ls;i++){
bee[i]=as[i]*(bs[i]*(4.0-this->M5) +1.0);
assert(bee[i]!=Coeff_t(0.0));
cee[i]=as[i]*(1.0-cs[i]*(4.0-this->M5));
beo[i]=as[i]*bs[i];
ceo[i]=-as[i]*cs[i];
}
aee.resize(Ls);
aeo.resize(Ls);
for(int i=0;i<Ls;i++){
aee[i]=cee[i];
aeo[i]=ceo[i];
}
//////////////////////////////////////////
// LDU decomposition of eeoo
//////////////////////////////////////////
dee.resize(Ls);
lee.resize(Ls);
leem.resize(Ls);
uee.resize(Ls);
ueem.resize(Ls);
for(int i=0;i<Ls;i++){
dee[i] = bee[i];
if ( i < Ls-1 ) {
assert(bee[i]!=Coeff_t(0.0));
assert(bee[0]!=Coeff_t(0.0));
lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
leem[i]=mass*cee[Ls-1]/bee[0];
for(int j=0;j<i;j++) {
assert(bee[j+1]!=Coeff_t(0.0));
leem[i]*= aee[j]/bee[j+1];
}
uee[i] =-aee[i]/bee[i]; // up-diag entry on the ith row
ueem[i]=mass;
for(int j=1;j<=i;j++) ueem[i]*= cee[j]/bee[j];
ueem[i]*= aee[0]/bee[0];
} else {
lee[i] =0.0;
leem[i]=0.0;
uee[i] =0.0;
ueem[i]=0.0;
}
}
{
Coeff_t delta_d=mass*cee[Ls-1];
for(int j=0;j<Ls-1;j++) {
assert(bee[j] != Coeff_t(0.0));
delta_d *= cee[j]/bee[j];
}
dee[Ls-1] += delta_d;
}
int inv=1;
this->MooeeInternalCompute(0,inv,MatpInv,MatmInv);
this->MooeeInternalCompute(1,inv,MatpInvDag,MatmInvDag);
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternalCompute(int dag, int inv,
Vector<iSinglet<Simd> > & Matp,
Vector<iSinglet<Simd> > & Matm)
{
int Ls=this->Ls;
GridBase *grid = this->FermionRedBlackGrid();
int LLs = grid->_rdimensions[0];
if ( LLs == Ls ) {
return; // Not vectorised in 5th direction
}
Eigen::MatrixXcd Pplus = Eigen::MatrixXcd::Zero(Ls,Ls);
Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
for(int s=0;s<Ls;s++){
Pplus(s,s) = bee[s];
Pminus(s,s)= bee[s];
}
for(int s=0;s<Ls-1;s++){
Pminus(s,s+1) = -cee[s];
}
for(int s=0;s<Ls-1;s++){
Pplus(s+1,s) = -cee[s+1];
}
Pplus (0,Ls-1) = mass*cee[0];
Pminus(Ls-1,0) = mass*cee[Ls-1];
Eigen::MatrixXcd PplusMat ;
Eigen::MatrixXcd PminusMat;
if ( inv ) {
PplusMat =Pplus.inverse();
PminusMat=Pminus.inverse();
} else {
PplusMat =Pplus;
PminusMat=Pminus;
}
if(dag){
PplusMat.adjointInPlace();
PminusMat.adjointInPlace();
}
typedef typename SiteHalfSpinor::scalar_type scalar_type;
const int Nsimd=Simd::Nsimd();
Matp.resize(Ls*LLs);
Matm.resize(Ls*LLs);
for(int s2=0;s2<Ls;s2++){
for(int s1=0;s1<LLs;s1++){
int istride = LLs;
int ostride = 1;
Simd Vp;
Simd Vm;
scalar_type *sp = (scalar_type *)&Vp;
scalar_type *sm = (scalar_type *)&Vm;
for(int l=0;l<Nsimd;l++){
if ( switcheroo<Coeff_t>::iscomplex() ) {
sp[l] = PplusMat (l*istride+s1*ostride,s2);
sm[l] = PminusMat(l*istride+s1*ostride,s2);
} else {
// if real
scalar_type tmp;
tmp = PplusMat (l*istride+s1*ostride,s2);
sp[l] = scalar_type(tmp.real(),tmp.real());
tmp = PminusMat(l*istride+s1*ostride,s2);
sm[l] = scalar_type(tmp.real(),tmp.real());
}
}
Matp[LLs*s2+s1] = Vp;
Matm[LLs*s2+s1] = Vm;
}}
}
FermOpTemplateInstantiate(CayleyFermion5D);
GparityFermOpTemplateInstantiate(CayleyFermion5D);
NAMESPACE_END(Grid);

247
Grid/qcd/action/fermion/CayleyFermion5Dcache.cc
View File

@ -1,247 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/CayleyFermion5D.h>
NAMESPACE_BEGIN(Grid);
// Pminus fowards
// Pplus backwards..
template<class Impl>
void CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
Vector<Coeff_t> &lower,
Vector<Coeff_t> &diag,
Vector<Coeff_t> &upper)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls =this->Ls;
assert(phi.Checkerboard() == psi.Checkerboard());
// 10 = 3 complex mult + 2 complex add
// Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
M5Dcalls++;
M5Dtime-=usecond();
thread_loop( (int ss=0;ss<grid->oSites();ss+=Ls),{ // adds Ls
for(int s=0;s<Ls;s++){
auto tmp = psi[0];
if ( s==0 ) {
spProj5m(tmp,psi[ss+s+1]);
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
spProj5p(tmp,psi[ss+Ls-1]);
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else if ( s==(Ls-1)) {
spProj5m(tmp,psi[ss+0]);
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
spProj5p(tmp,psi[ss+s-1]);
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else {
spProj5m(tmp,psi[ss+s+1]);
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
spProj5p(tmp,psi[ss+s-1]);
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
}
}
});
M5Dtime+=usecond();
}
template<class Impl>
void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
Vector<Coeff_t> &lower,
Vector<Coeff_t> &diag,
Vector<Coeff_t> &upper)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls =this->Ls;
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
M5Dcalls++;
M5Dtime-=usecond();
thread_loop( (int ss=0;ss<grid->oSites();ss+=Ls),{ // adds Ls
auto tmp = psi[0];
for(int s=0;s<Ls;s++){
if ( s==0 ) {
spProj5p(tmp,psi[ss+s+1]);
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
spProj5m(tmp,psi[ss+Ls-1]);
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else if ( s==(Ls-1)) {
spProj5p(tmp,psi[ss+0]);
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
spProj5m(tmp,psi[ss+s-1]);
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else {
spProj5p(tmp,psi[ss+s+1]);
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
spProj5m(tmp,psi[ss+s-1]);
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
}
}
});
M5Dtime+=usecond();
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi_i, FermionField &chi_i)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
int Ls=this->Ls;
MooeeInvCalls++;
MooeeInvTime-=usecond();
thread_loop((int ss=0;ss<grid->oSites();ss+=Ls),{ // adds Ls
auto tmp = psi[0];
// flops = 12*2*Ls + 12*2*Ls + 3*12*Ls + 12*2*Ls = 12*Ls * (9) = 108*Ls flops
// Apply (L^{\prime})^{-1}
chi[ss]=psi[ss]; // chi[0]=psi[0]
for(int s=1;s<Ls;s++){
spProj5p(tmp,chi[ss+s-1]);
chi[ss+s] = psi[ss+s]-lee[s-1]*tmp;
}
// L_m^{-1}
for (int s=0;s<Ls-1;s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
spProj5m(tmp,chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - leem[s]*tmp;
}
// U_m^{-1} D^{-1}
for (int s=0;s<Ls-1;s++){
// Chi[s] + 1/d chi[s]
spProj5p(tmp,chi[ss+Ls-1]);
chi[ss+s] = (1.0/dee[s])*chi[ss+s]-(ueem[s]/dee[Ls-1])*tmp;
}
chi[ss+Ls-1]= (1.0/dee[Ls-1])*chi[ss+Ls-1];
// Apply U^{-1}
for (int s=Ls-2;s>=0;s--){
spProj5m(tmp,chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - uee[s]*tmp;
}
});
MooeeInvTime+=usecond();
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi_i)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
int Ls=this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
assert(psi.Checkerboard() == psi.Checkerboard());
MooeeInvCalls++;
MooeeInvTime-=usecond();
thread_loop((int ss=0;ss<grid->oSites();ss+=Ls),{ // adds Ls
auto tmp = psi[0];
// Apply (U^{\prime})^{-dagger}
chi[ss]=psi[ss];
for (int s=1;s<Ls;s++){
spProj5m(tmp,chi[ss+s-1]);
chi[ss+s] = psi[ss+s]-conjugate(uee[s-1])*tmp;
}
// U_m^{-\dagger}
for (int s=0;s<Ls-1;s++){
spProj5p(tmp,chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - conjugate(ueem[s])*tmp;
}
// L_m^{-\dagger} D^{-dagger}
for (int s=0;s<Ls-1;s++){
spProj5m(tmp,chi[ss+Ls-1]);
chi[ss+s] = conjugate(1.0/dee[s])*chi[ss+s]-conjugate(leem[s]/dee[Ls-1])*tmp;
}
chi[ss+Ls-1]= conjugate(1.0/dee[Ls-1])*chi[ss+Ls-1];
// Apply L^{-dagger}
for (int s=Ls-2;s>=0;s--){
spProj5p(tmp,chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - conjugate(lee[s])*tmp;
}
});
MooeeInvTime+=usecond();
}
#ifdef CAYLEY_DPERP_CACHE
INSTANTIATE_DPERP(WilsonImplF);
INSTANTIATE_DPERP(WilsonImplD);
INSTANTIATE_DPERP(GparityWilsonImplF);
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
NAMESPACE_END(Grid);

284
Grid/qcd/action/fermion/CayleyFermion5Dgpu.cc
View File

@ -1,284 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/CayleyFermion5D.h>
NAMESPACE_BEGIN(Grid);
// Pminus fowards
// Pplus backwards..
template<class Impl>
void CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
Vector<Coeff_t> &lower,
Vector<Coeff_t> &diag,
Vector<Coeff_t> &upper)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
Coeff_t *lower_v = &lower[0];
Coeff_t *diag_v = &diag[0];
Coeff_t *upper_v = &upper[0];
int Ls =this->Ls;
assert(phi.Checkerboard() == psi.Checkerboard());
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites4d = nsimd * grid->oSites() / Ls;
// 10 = 3 complex mult + 2 complex add
// Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
M5Dcalls++;
M5Dtime-=usecond();
accelerator_loopN( sss, sites4d ,{
uint64_t lane = sss % nsimd;
uint64_t ss = Ls * (sss / nsimd);
for(int s=0;s<Ls;s++){
auto res = extractLane(lane,phi[ss+s]);
res = diag_v[s]*res;
auto tmp = extractLane(lane,psi[ss+(s+1)%Ls]);
spProj5m(tmp,tmp);
res += upper_v[s]*tmp;
tmp = extractLane(lane,psi[ss+(s+Ls-1)%Ls]);
spProj5p(tmp,tmp);
res += lower_v[s]*tmp;
insertLane(lane,chi[ss+s],res);
}
});
M5Dtime+=usecond();
}
template<class Impl>
void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
Vector<Coeff_t> &lower,
Vector<Coeff_t> &diag,
Vector<Coeff_t> &upper)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
Coeff_t *lower_v = &lower[0];
Coeff_t *diag_v = &diag[0];
Coeff_t *upper_v = &upper[0];
int Ls =this->Ls;
assert(phi.Checkerboard() == psi.Checkerboard());
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites4d = nsimd * grid->oSites() / Ls;
// 10 = 3 complex mult + 2 complex add
// Flops = 10.0*(Nc*Ns) *Ls*vol (/2 for red black counting)
M5Dcalls++;
M5Dtime-=usecond();
accelerator_loopN( sss, sites4d ,{
uint64_t lane = sss % nsimd;
uint64_t ss = Ls * (sss / nsimd);
for(int s=0;s<Ls;s++){
auto res = extractLane(lane,phi[ss+s]);
res = diag_v[s]*res;
auto tmp = extractLane(lane,psi[ss+(s+1)%Ls]);
spProj5p(tmp,tmp);
res += upper_v[s]*tmp;
tmp = extractLane(lane,psi[ss+(s+Ls-1)%Ls]);
spProj5m(tmp,tmp);
res += lower_v[s]*tmp;
insertLane(lane,chi[ss+s],res);
}
});
M5Dtime+=usecond();
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi_i, FermionField &chi_i)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
Coeff_t *lee_v = &lee[0];
Coeff_t *leem_v = &leem[0];
Coeff_t *uee_v = &uee[0];
Coeff_t *ueem_v = &ueem[0];
Coeff_t *dee_v = &dee[0];
int Ls=this->Ls;
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites4d = nsimd * grid->oSites() / Ls;
typedef typename SiteSpinor::scalar_object ScalarSiteSpinor;
MooeeInvCalls++;
MooeeInvTime-=usecond();
accelerator_loopN( sss, sites4d ,{
uint64_t lane = sss % nsimd;
uint64_t ss = Ls * (sss / nsimd);
ScalarSiteSpinor res, tmp, acc;
// X = Nc*Ns
// flops = 2X + (Ls-2)(4X + 4X) + 6X + 1 + 2X + (Ls-1)(10X + 1) = -16X + Ls(1+18X) = -192 + 217*Ls flops
// Apply (L^{\prime})^{-1} L_m^{-1}
res = extractLane(lane,psi[ss]);
spProj5m(tmp,res);
acc = leem_v[0]*tmp;
spProj5p(tmp,res);
insertLane(lane,chi[ss],res);
for(int s=1;s<Ls-1;s++){
res = extractLane(lane,psi[ss+s]);
res -= lee_v[s-1]*tmp;
spProj5m(tmp,res);
acc += leem_v[s]*tmp;
spProj5p(tmp,res);
insertLane(lane,chi[ss+s],res);
}
res = extractLane(lane,psi[ss+Ls-1]);
res = res - lee_v[Ls-2]*tmp - acc;
// Apply U_m^{-1} D^{-1} U^{-1}
res = (1.0/dee_v[Ls-1])*res;
insertLane(lane,chi[ss+Ls-1],res);
spProj5p(acc,res);
spProj5m(tmp,res);
for (int s=Ls-2;s>=0;s--){
res = extractLane(lane,chi[ss+s]);
res = (1.0/dee_v[s])*res - uee_v[s]*tmp - ueem_v[s]*acc;
spProj5m(tmp,res);
insertLane(lane,chi[ss+s],res);
}
});
MooeeInvTime+=usecond();
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi_i)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
Coeff_t *lee_v = &lee[0];
Coeff_t *leem_v = &leem[0];
Coeff_t *uee_v = &uee[0];
Coeff_t *ueem_v = &ueem[0];
Coeff_t *dee_v = &dee[0];
int Ls=this->Ls;
const uint64_t nsimd = grid->Nsimd();
const uint64_t sites4d = nsimd * grid->oSites() / Ls;
typedef typename SiteSpinor::scalar_object ScalarSiteSpinor;
MooeeInvCalls++;
MooeeInvTime-=usecond();
accelerator_loopN( sss, sites4d ,{
uint64_t lane = sss % nsimd;
uint64_t ss = Ls * (sss / nsimd);
ScalarSiteSpinor res, tmp, acc;
// X = Nc*Ns
// flops = 2X + (Ls-2)(4X + 4X) + 6X + 1 + 2X + (Ls-1)(10X + 1) = -16X + Ls(1+18X) = -192 + 217*Ls flops
// Apply (U^{\prime})^{-dagger} U_m^{-\dagger}
res = extractLane(lane,psi[ss]);
spProj5p(tmp,res);
acc = conjugate(ueem_v[0])*tmp;
spProj5m(tmp,res);
insertLane(lane,chi[ss],res);
for(int s=1;s<Ls-1;s++){
res = extractLane(lane,psi[ss+s]);
res -= conjugate(uee_v[s-1])*tmp;
spProj5p(tmp,res);
acc += conjugate(ueem_v[s])*tmp;
spProj5m(tmp,res);
insertLane(lane,chi[ss+s],res);
}
res = extractLane(lane,psi[ss+Ls-1]);
res = res - conjugate(uee_v[Ls-2])*tmp - acc;
// Apply L_m^{-\dagger} D^{-dagger} L^{-dagger}
res = conjugate(1.0/dee_v[Ls-1])*res;
insertLane(lane,chi[ss+Ls-1],res);
spProj5m(acc,res);
spProj5p(tmp,res);
for (int s=Ls-2;s>=0;s--){
res = extractLane(lane,chi[ss+s]);
res = conjugate(1.0/dee_v[s])*res - conjugate(lee_v[s])*tmp - conjugate(leem_v[s])*acc;
spProj5p(tmp,res);
insertLane(lane,chi[ss+s],res);
}
});
MooeeInvTime+=usecond();
}
#ifdef CAYLEY_DPERP_GPU
INSTANTIATE_DPERP(WilsonImplF);
INSTANTIATE_DPERP(WilsonImplD);
INSTANTIATE_DPERP(GparityWilsonImplF);
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
NAMESPACE_END(Grid);

838
Grid/qcd/action/fermion/CayleyFermion5Dvec.cc
View File

@ -1,838 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/CayleyFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/CayleyFermion5D.h>
NAMESPACE_BEGIN(Grid);
/*
* Dense matrix versions of routines
*/
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
{
this->MooeeInternal(psi,chi,DaggerYes,InverseYes);
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInv(const FermionField &psi, FermionField &chi)
{
this->MooeeInternal(psi,chi,DaggerNo,InverseYes);
}
template<class Impl>
void CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
Vector<Coeff_t> &lower,
Vector<Coeff_t> &diag,
Vector<Coeff_t> &upper)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd= Simd::Nsimd();
Vector<iSinglet<Simd> > u(LLs);
Vector<iSinglet<Simd> > l(LLs);
Vector<iSinglet<Simd> > d(LLs);
assert(Ls/LLs==nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type * u_p = (scalar_type *)&u[0];
scalar_type * l_p = (scalar_type *)&l[0];
scalar_type * d_p = (scalar_type *)&d[0];
for(int o=0;o<LLs;o++){ // outer
for(int i=0;i<nsimd;i++){ //inner
int s = o+i*LLs;
int ss = o*nsimd+i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
M5Dcalls++;
M5Dtime-=usecond();
assert(Nc==3);
thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0;v<LLs;v++){
int vp=(v+1)%LLs;
int vm=(v+LLs-1)%LLs;
spProj5m(hp,psi[ss+vp]);
spProj5p(hm,psi[ss+vm]);
if ( vp<=v ) rotate(hp,hp,1);
if ( vm>=v ) rotate(hm,hm,nsimd-1);
hp=0.5*hp;
hm=0.5*hm;
spRecon5m(fp,hp);
spRecon5p(fm,hm);
chi[ss+v] = d[v]*phi[ss+v];
chi[ss+v] = chi[ss+v] +u[v]*fp;
chi[ss+v] = chi[ss+v] +l[v]*fm;
}
#else
for(int v=0;v<LLs;v++){
vprefetch(psi[ss+v+LLs]);
int vp= (v==LLs-1) ? 0 : v+1;
int vm= (v==0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(2)(0);
Simd hp_01 = psi[ss+vp]()(2)(1);
Simd hp_02 = psi[ss+vp]()(2)(2);
Simd hp_10 = psi[ss+vp]()(3)(0);
Simd hp_11 = psi[ss+vp]()(3)(1);
Simd hp_12 = psi[ss+vp]()(3)(2);
Simd hm_00 = psi[ss+vm]()(0)(0);
Simd hm_01 = psi[ss+vm]()(0)(1);
Simd hm_02 = psi[ss+vm]()(0)(2);
Simd hm_10 = psi[ss+vm]()(1)(0);
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
if ( vp<=v ) {
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if ( vm>=v ) {
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
// Can force these to real arithmetic and save 2x.
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_12);
vstream(chi[ss+v]()(0)(0),p_00);
vstream(chi[ss+v]()(0)(1),p_01);
vstream(chi[ss+v]()(0)(2),p_02);
vstream(chi[ss+v]()(1)(0),p_10);
vstream(chi[ss+v]()(1)(1),p_11);
vstream(chi[ss+v]()(1)(2),p_12);
vstream(chi[ss+v]()(2)(0),p_20);
vstream(chi[ss+v]()(2)(1),p_21);
vstream(chi[ss+v]()(2)(2),p_22);
vstream(chi[ss+v]()(3)(0),p_30);
vstream(chi[ss+v]()(3)(1),p_31);
vstream(chi[ss+v]()(3)(2),p_32);
}
#endif
});
M5Dtime+=usecond();
}
template<class Impl>
void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
Vector<Coeff_t> &lower,
Vector<Coeff_t> &diag,
Vector<Coeff_t> &upper)
{
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi=psi_i.View();
auto phi=phi_i.View();
auto chi=chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd= Simd::Nsimd();
Vector<iSinglet<Simd> > u(LLs);
Vector<iSinglet<Simd> > l(LLs);
Vector<iSinglet<Simd> > d(LLs);
assert(Ls/LLs==nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type * u_p = (scalar_type *)&u[0];
scalar_type * l_p = (scalar_type *)&l[0];
scalar_type * d_p = (scalar_type *)&d[0];
for(int o=0;o<LLs;o++){ // outer
for(int i=0;i<nsimd;i++){ //inner
int s = o+i*LLs;
int ss = o*nsimd+i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
M5Dcalls++;
M5Dtime-=usecond();
thread_loop( (int ss=0;ss<grid->oSites();ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0;v<LLs;v++){
int vp=(v+1)%LLs;
int vm=(v+LLs-1)%LLs;
spProj5p(hp,psi[ss+vp]);
spProj5m(hm,psi[ss+vm]);
if ( vp<=v ) rotate(hp,hp,1);
if ( vm>=v ) rotate(hm,hm,nsimd-1);
hp=hp*0.5;
hm=hm*0.5;
spRecon5p(fp,hp);
spRecon5m(fm,hm);
chi[ss+v] = d[v]*phi[ss+v]+u[v]*fp;
chi[ss+v] = chi[ss+v] +l[v]*fm;
}
#else
for(int v=0;v<LLs;v++){
vprefetch(psi[ss+v+LLs]);
int vp= (v==LLs-1) ? 0 : v+1;
int vm= (v==0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if ( vp<=v ) {
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if ( vm>=v ) {
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(),hp_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(),hm_12);
vstream(chi[ss+v]()(0)(0),p_00);
vstream(chi[ss+v]()(0)(1),p_01);
vstream(chi[ss+v]()(0)(2),p_02);
vstream(chi[ss+v]()(1)(0),p_10);
vstream(chi[ss+v]()(1)(1),p_11);
vstream(chi[ss+v]()(1)(2),p_12);
vstream(chi[ss+v]()(2)(0),p_20);
vstream(chi[ss+v]()(2)(1),p_21);
vstream(chi[ss+v]()(2)(2),p_22);
vstream(chi[ss+v]()(3)(0),p_30);
vstream(chi[ss+v]()(3)(1),p_31);
vstream(chi[ss+v]()(3)(2),p_32);
}
#endif
});
M5Dtime+=usecond();
}
#ifdef AVX512
#include <simd/Intel512common.h>
#include <simd/Intel512avx.h>
#include <simd/Intel512single.h>
#endif
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternalAsm(const FermionField &psi_i, FermionField &chi_i,
int LLs, int site,
Vector<iSinglet<Simd> > &Matp,
Vector<iSinglet<Simd> > &Matm)
{
auto psi = psi_i.View();
auto chi = chi_i.View();
#ifndef AVX512
{
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
SiteHalfSpinor SiteChiP;
SiteHalfSpinor SiteChiM;
// Ls*Ls * 2 * 12 * vol flops
for(int s1=0;s1<LLs;s1++){
for(int s2=0;s2<LLs;s2++){
for(int l=0; l<Simd::Nsimd();l++){ // simd lane
int s=s2+l*LLs;
int lex=s2+LLs*site;
if ( s2==0 && l==0) {
SiteChiP=Zero();
SiteChiM=Zero();
}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vbroadcast(BcastP()(sp )(co),psi[lex]()(sp)(co),l);
}}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vbroadcast(BcastM()(sp )(co),psi[lex]()(sp+2)(co),l);
}}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
SiteChiP()(sp)(co)=real_madd(Matp[LLs*s+s1]()()(),BcastP()(sp)(co),SiteChiP()(sp)(co)); // 1100 us.
SiteChiM()(sp)(co)=real_madd(Matm[LLs*s+s1]()()(),BcastM()(sp)(co),SiteChiM()(sp)(co)); // each found by commenting out
}}
}}
{
int lex = s1+LLs*site;
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vstream(chi[lex]()(sp)(co), SiteChiP()(sp)(co));
vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
}}
}
}
}
#else
{
// pointers
// MASK_REGS;
#define Chi_00 %%zmm1
#define Chi_01 %%zmm2
#define Chi_02 %%zmm3
#define Chi_10 %%zmm4
#define Chi_11 %%zmm5
#define Chi_12 %%zmm6
#define Chi_20 %%zmm7
#define Chi_21 %%zmm8
#define Chi_22 %%zmm9
#define Chi_30 %%zmm10
#define Chi_31 %%zmm11
#define Chi_32 %%zmm12
#define BCAST0 %%zmm13
#define BCAST1 %%zmm14
#define BCAST2 %%zmm15
#define BCAST3 %%zmm16
#define BCAST4 %%zmm17
#define BCAST5 %%zmm18
#define BCAST6 %%zmm19
#define BCAST7 %%zmm20
#define BCAST8 %%zmm21
#define BCAST9 %%zmm22
#define BCAST10 %%zmm23
#define BCAST11 %%zmm24
int incr=LLs*LLs*sizeof(iSinglet<Simd>);
for(int s1=0;s1<LLs;s1++){
for(int s2=0;s2<LLs;s2++){
int lex=s2+LLs*site;
uint64_t a0 = (uint64_t)&Matp[LLs*s2+s1]; // should be cacheable
uint64_t a1 = (uint64_t)&Matm[LLs*s2+s1];
uint64_t a2 = (uint64_t)&psi[lex];
for(int l=0; l<Simd::Nsimd();l++){ // simd lane
if ( (s2+l)==0 ) {
asm (
VPREFETCH1(0,%2) VPREFETCH1(0,%1)
VPREFETCH1(12,%2) VPREFETCH1(13,%2)
VPREFETCH1(14,%2) VPREFETCH1(15,%2)
VBCASTCDUP(0,%2,BCAST0)
VBCASTCDUP(1,%2,BCAST1)
VBCASTCDUP(2,%2,BCAST2)
VBCASTCDUP(3,%2,BCAST3)
VBCASTCDUP(4,%2,BCAST4) VMULMEM (0,%0,BCAST0,Chi_00)
VBCASTCDUP(5,%2,BCAST5) VMULMEM (0,%0,BCAST1,Chi_01)
VBCASTCDUP(6,%2,BCAST6) VMULMEM (0,%0,BCAST2,Chi_02)
VBCASTCDUP(7,%2,BCAST7) VMULMEM (0,%0,BCAST3,Chi_10)
VBCASTCDUP(8,%2,BCAST8) VMULMEM (0,%0,BCAST4,Chi_11)
VBCASTCDUP(9,%2,BCAST9) VMULMEM (0,%0,BCAST5,Chi_12)
VBCASTCDUP(10,%2,BCAST10) VMULMEM (0,%1,BCAST6,Chi_20)
VBCASTCDUP(11,%2,BCAST11) VMULMEM (0,%1,BCAST7,Chi_21)
VMULMEM (0,%1,BCAST8,Chi_22)
VMULMEM (0,%1,BCAST9,Chi_30)
VMULMEM (0,%1,BCAST10,Chi_31)
VMULMEM (0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
} else {
asm (
VBCASTCDUP(0,%2,BCAST0) VMADDMEM (0,%0,BCAST0,Chi_00)
VBCASTCDUP(1,%2,BCAST1) VMADDMEM (0,%0,BCAST1,Chi_01)
VBCASTCDUP(2,%2,BCAST2) VMADDMEM (0,%0,BCAST2,Chi_02)
VBCASTCDUP(3,%2,BCAST3) VMADDMEM (0,%0,BCAST3,Chi_10)
VBCASTCDUP(4,%2,BCAST4) VMADDMEM (0,%0,BCAST4,Chi_11)
VBCASTCDUP(5,%2,BCAST5) VMADDMEM (0,%0,BCAST5,Chi_12)
VBCASTCDUP(6,%2,BCAST6) VMADDMEM (0,%1,BCAST6,Chi_20)
VBCASTCDUP(7,%2,BCAST7) VMADDMEM (0,%1,BCAST7,Chi_21)
VBCASTCDUP(8,%2,BCAST8) VMADDMEM (0,%1,BCAST8,Chi_22)
VBCASTCDUP(9,%2,BCAST9) VMADDMEM (0,%1,BCAST9,Chi_30)
VBCASTCDUP(10,%2,BCAST10) VMADDMEM (0,%1,BCAST10,Chi_31)
VBCASTCDUP(11,%2,BCAST11) VMADDMEM (0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
}
a0 = a0+incr;
a1 = a1+incr;
a2 = a2+sizeof(typename Simd::scalar_type);
}}
{
int lexa = s1+LLs*site;
asm (
VSTORE(0,%0,Chi_00) VSTORE(1 ,%0,Chi_01) VSTORE(2 ,%0,Chi_02)
VSTORE(3,%0,Chi_10) VSTORE(4 ,%0,Chi_11) VSTORE(5 ,%0,Chi_12)
VSTORE(6,%0,Chi_20) VSTORE(7 ,%0,Chi_21) VSTORE(8 ,%0,Chi_22)
VSTORE(9,%0,Chi_30) VSTORE(10,%0,Chi_31) VSTORE(11,%0,Chi_32)
: : "r" ((uint64_t)&chi[lexa]) : "memory" );
}
}
}
#undef Chi_00
#undef Chi_01
#undef Chi_02
#undef Chi_10
#undef Chi_11
#undef Chi_12
#undef Chi_20
#undef Chi_21
#undef Chi_22
#undef Chi_30
#undef Chi_31
#undef Chi_32
#undef BCAST0
#undef BCAST1
#undef BCAST2
#undef BCAST3
#undef BCAST4
#undef BCAST5
#undef BCAST6
#undef BCAST7
#undef BCAST8
#undef BCAST9
#undef BCAST10
#undef BCAST11
#endif
};
// Z-mobius version
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi_i, FermionField &chi_i,
int LLs, int site, Vector<iSinglet<Simd> > &Matp, Vector<iSinglet<Simd> > &Matm)
{
#ifndef AVX512
{
auto psi = psi_i.View();
auto chi = chi_i.View();
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
SiteHalfSpinor SiteChiP;
SiteHalfSpinor SiteChiM;
// Ls*Ls * 2 * 12 * vol flops
for(int s1=0;s1<LLs;s1++){
for(int s2=0;s2<LLs;s2++){
for(int l=0; l<Simd::Nsimd();l++){ // simd lane
int s=s2+l*LLs;
int lex=s2+LLs*site;
if ( s2==0 && l==0) {
SiteChiP=Zero();
SiteChiM=Zero();
}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vbroadcast(BcastP()(sp )(co),psi[lex]()(sp)(co),l);
}}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vbroadcast(BcastM()(sp )(co),psi[lex]()(sp+2)(co),l);
}}
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
SiteChiP()(sp)(co)=SiteChiP()(sp)(co)+ Matp[LLs*s+s1]()()()*BcastP()(sp)(co);
SiteChiM()(sp)(co)=SiteChiM()(sp)(co)+ Matm[LLs*s+s1]()()()*BcastM()(sp)(co);
}}
}}
{
int lex = s1+LLs*site;
for(int sp=0;sp<2;sp++){
for(int co=0;co<Nc;co++){
vstream(chi[lex]()(sp)(co), SiteChiP()(sp)(co));
vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
}}
}
}
}
#else
{
auto psi = psi_i.View();
auto chi = chi_i.View();
// pointers
// MASK_REGS;
#define Chi_00 %zmm0
#define Chi_01 %zmm1
#define Chi_02 %zmm2
#define Chi_10 %zmm3
#define Chi_11 %zmm4
#define Chi_12 %zmm5
#define Chi_20 %zmm6
#define Chi_21 %zmm7
#define Chi_22 %zmm8
#define Chi_30 %zmm9
#define Chi_31 %zmm10
#define Chi_32 %zmm11
#define pChi_00 %%zmm0
#define pChi_01 %%zmm1
#define pChi_02 %%zmm2
#define pChi_10 %%zmm3
#define pChi_11 %%zmm4
#define pChi_12 %%zmm5
#define pChi_20 %%zmm6
#define pChi_21 %%zmm7
#define pChi_22 %%zmm8
#define pChi_30 %%zmm9
#define pChi_31 %%zmm10
#define pChi_32 %%zmm11
#define BCAST_00 %zmm12
#define SHUF_00 %zmm13
#define BCAST_01 %zmm14
#define SHUF_01 %zmm15
#define BCAST_02 %zmm16
#define SHUF_02 %zmm17
#define BCAST_10 %zmm18
#define SHUF_10 %zmm19
#define BCAST_11 %zmm20
#define SHUF_11 %zmm21
#define BCAST_12 %zmm22
#define SHUF_12 %zmm23
#define Mp %zmm24
#define Mps %zmm25
#define Mm %zmm26
#define Mms %zmm27
#define N 8
int incr=LLs*LLs*sizeof(iSinglet<Simd>);
for(int s1=0;s1<LLs;s1++){
for(int s2=0;s2<LLs;s2++){
int lex=s2+LLs*site;
uint64_t a0 = (uint64_t)&Matp[LLs*s2+s1]; // should be cacheable
uint64_t a1 = (uint64_t)&Matm[LLs*s2+s1];
uint64_t a2 = (uint64_t)&psi[lex];
for(int l=0; l<Simd::Nsimd();l++){ // simd lane
if ( (s2+l)==0 ) {
LOAD64(%r8,a0);
LOAD64(%r9,a1);
LOAD64(%r10,a2);
asm (
VLOAD(0,%r8,Mp)// i r
VLOAD(0,%r9,Mm)
VSHUF(Mp,Mps) // r i
VSHUF(Mm,Mms)
VPREFETCH1(12,%r10) VPREFETCH1(13,%r10)
VPREFETCH1(14,%r10) VPREFETCH1(15,%r10)
VMULIDUP(0*N,%r10,Mps,Chi_00)
VMULIDUP(1*N,%r10,Mps,Chi_01)
VMULIDUP(2*N,%r10,Mps,Chi_02)
VMULIDUP(3*N,%r10,Mps,Chi_10)
VMULIDUP(4*N,%r10,Mps,Chi_11)
VMULIDUP(5*N,%r10,Mps,Chi_12)
VMULIDUP(6*N ,%r10,Mms,Chi_20)
VMULIDUP(7*N ,%r10,Mms,Chi_21)
VMULIDUP(8*N ,%r10,Mms,Chi_22)
VMULIDUP(9*N ,%r10,Mms,Chi_30)
VMULIDUP(10*N,%r10,Mms,Chi_31)
VMULIDUP(11*N,%r10,Mms,Chi_32)
VMADDSUBRDUP(0*N,%r10,Mp,Chi_00)
VMADDSUBRDUP(1*N,%r10,Mp,Chi_01)
VMADDSUBRDUP(2*N,%r10,Mp,Chi_02)
VMADDSUBRDUP(3*N,%r10,Mp,Chi_10)
VMADDSUBRDUP(4*N,%r10,Mp,Chi_11)
VMADDSUBRDUP(5*N,%r10,Mp,Chi_12)
VMADDSUBRDUP(6*N ,%r10,Mm,Chi_20)
VMADDSUBRDUP(7*N ,%r10,Mm,Chi_21)
VMADDSUBRDUP(8*N ,%r10,Mm,Chi_22)
VMADDSUBRDUP(9*N ,%r10,Mm,Chi_30)
VMADDSUBRDUP(10*N,%r10,Mm,Chi_31)
VMADDSUBRDUP(11*N,%r10,Mm,Chi_32)
);
} else {
LOAD64(%r8,a0);
LOAD64(%r9,a1);
LOAD64(%r10,a2);
asm (
VLOAD(0,%r8,Mp)
VSHUF(Mp,Mps)
VLOAD(0,%r9,Mm)
VSHUF(Mm,Mms)
VMADDSUBIDUP(0*N,%r10,Mps,Chi_00) // Mri * Pii +- Cir
VMADDSUBIDUP(1*N,%r10,Mps,Chi_01)
VMADDSUBIDUP(2*N,%r10,Mps,Chi_02)
VMADDSUBIDUP(3*N,%r10,Mps,Chi_10)
VMADDSUBIDUP(4*N,%r10,Mps,Chi_11)
VMADDSUBIDUP(5*N,%r10,Mps,Chi_12)
VMADDSUBIDUP(6 *N,%r10,Mms,Chi_20)
VMADDSUBIDUP(7 *N,%r10,Mms,Chi_21)
VMADDSUBIDUP(8 *N,%r10,Mms,Chi_22)
VMADDSUBIDUP(9 *N,%r10,Mms,Chi_30)
VMADDSUBIDUP(10*N,%r10,Mms,Chi_31)
VMADDSUBIDUP(11*N,%r10,Mms,Chi_32)
VMADDSUBRDUP(0*N,%r10,Mp,Chi_00) // Cir = Mir * Prr +- ( Mri * Pii +- Cir)
VMADDSUBRDUP(1*N,%r10,Mp,Chi_01) // Ci = MiPr + Ci + MrPi ; Cr = MrPr - ( MiPi - Cr)
VMADDSUBRDUP(2*N,%r10,Mp,Chi_02)
VMADDSUBRDUP(3*N,%r10,Mp,Chi_10)
VMADDSUBRDUP(4*N,%r10,Mp,Chi_11)
VMADDSUBRDUP(5*N,%r10,Mp,Chi_12)
VMADDSUBRDUP(6 *N,%r10,Mm,Chi_20)
VMADDSUBRDUP(7 *N,%r10,Mm,Chi_21)
VMADDSUBRDUP(8 *N,%r10,Mm,Chi_22)
VMADDSUBRDUP(9 *N,%r10,Mm,Chi_30)
VMADDSUBRDUP(10*N,%r10,Mm,Chi_31)
VMADDSUBRDUP(11*N,%r10,Mm,Chi_32)
);
}
a0 = a0+incr;
a1 = a1+incr;
a2 = a2+sizeof(typename Simd::scalar_type);
}}
{
int lexa = s1+LLs*site;
/*
SiteSpinor tmp;
asm (
VSTORE(0,%0,pChi_00) VSTORE(1 ,%0,pChi_01) VSTORE(2 ,%0,pChi_02)
VSTORE(3,%0,pChi_10) VSTORE(4 ,%0,pChi_11) VSTORE(5 ,%0,pChi_12)
VSTORE(6,%0,pChi_20) VSTORE(7 ,%0,pChi_21) VSTORE(8 ,%0,pChi_22)
VSTORE(9,%0,pChi_30) VSTORE(10,%0,pChi_31) VSTORE(11,%0,pChi_32)
: : "r" ((uint64_t)&tmp) : "memory" );
*/
asm (
VSTORE(0,%0,pChi_00) VSTORE(1 ,%0,pChi_01) VSTORE(2 ,%0,pChi_02)
VSTORE(3,%0,pChi_10) VSTORE(4 ,%0,pChi_11) VSTORE(5 ,%0,pChi_12)
VSTORE(6,%0,pChi_20) VSTORE(7 ,%0,pChi_21) VSTORE(8 ,%0,pChi_22)
VSTORE(9,%0,pChi_30) VSTORE(10,%0,pChi_31) VSTORE(11,%0,pChi_32)
: : "r" ((uint64_t)&chi[lexa]) : "memory" );
// if ( 1 || (site==0) ) {
// std::cout<<site << " s1 "<<s1<<"\n\t"<<tmp << "\n't" << chi[lexa] <<"\n\t"<<tmp-chi[lexa]<<std::endl;
// }
}
}
}
#undef Chi_00
#undef Chi_01
#undef Chi_02
#undef Chi_10
#undef Chi_11
#undef Chi_12
#undef Chi_20
#undef Chi_21
#undef Chi_22
#undef Chi_30
#undef Chi_31
#undef Chi_32
#undef BCAST0
#undef BCAST1
#undef BCAST2
#undef BCAST3
#undef BCAST4
#undef BCAST5
#undef BCAST6
#undef BCAST7
#undef BCAST8
#undef BCAST9
#undef BCAST10
#undef BCAST11
#endif
};
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInternal(const FermionField &psi, FermionField &chi,int dag, int inv)
{
chi.Checkerboard()=psi.Checkerboard();
int Ls=this->Ls;
int LLs = psi.Grid()->_rdimensions[0];
int vol = psi.Grid()->oSites()/LLs;
Vector<iSinglet<Simd> > Matp;
Vector<iSinglet<Simd> > Matm;
Vector<iSinglet<Simd> > *_Matp;
Vector<iSinglet<Simd> > *_Matm;
// MooeeInternalCompute(dag,inv,Matp,Matm);
if ( inv && dag ) {
_Matp = &MatpInvDag;
_Matm = &MatmInvDag;
}
if ( inv && (!dag) ) {
_Matp = &MatpInv;
_Matm = &MatmInv;
}
if ( !inv ) {
MooeeInternalCompute(dag,inv,Matp,Matm);
_Matp = &Matp;
_Matm = &Matm;
}
assert(_Matp->size()==Ls*LLs);
MooeeInvCalls++;
MooeeInvTime-=usecond();
if ( switcheroo<Coeff_t>::iscomplex() ) {
thread_loop( (auto site=0;site<vol;site++),{
MooeeInternalZAsm(psi,chi,LLs,site,*_Matp,*_Matm);
});
} else {
thread_loop( (auto site=0;site<vol;site++),{
MooeeInternalAsm(psi,chi,LLs,site,*_Matp,*_Matm);
});
}
MooeeInvTime+=usecond();
}
INSTANTIATE_DPERP(DomainWallVec5dImplD);
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);
NAMESPACE_END(Grid);

320
Grid/qcd/action/fermion/ContinuedFractionFermion5D.cc
View File

@ -1,320 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ContinuedFractionFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/ContinuedFractionFermion5D.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
void ContinuedFractionFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD scale)
{
SetCoefficientsZolotarev(1.0/scale,zdata);
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata)
{
// How to check Ls matches??
// std::cout<<GridLogMessage << Ls << " Ls"<<std::endl;
// std::cout<<GridLogMessage << zdata->n << " - n"<<std::endl;
// std::cout<<GridLogMessage << zdata->da << " -da "<<std::endl;
// std::cout<<GridLogMessage << zdata->db << " -db"<<std::endl;
// std::cout<<GridLogMessage << zdata->dn << " -dn"<<std::endl;
// std::cout<<GridLogMessage << zdata->dd << " -dd"<<std::endl;
int Ls = this->Ls;
assert(zdata->db==Ls);// Beta has Ls coeffs
R=(1+this->mass)/(1-this->mass);
Beta.resize(Ls);
cc.resize(Ls);
cc_d.resize(Ls);
sqrt_cc.resize(Ls);
for(int i=0; i < Ls ; i++){
Beta[i] = zdata -> beta[i];
cc[i] = 1.0/Beta[i];
cc_d[i]=std::sqrt(cc[i]);
}
cc_d[Ls-1]=1.0;
for(int i=0; i < Ls-1 ; i++){
sqrt_cc[i]= std::sqrt(cc[i]*cc[i+1]);
}
sqrt_cc[Ls-2]=std::sqrt(cc[Ls-2]);
ZoloHiInv =1.0/zolo_hi;
dw_diag = (4.0-this->M5)*ZoloHiInv;
See.resize(Ls);
Aee.resize(Ls);
int sign=1;
for(int s=0;s<Ls;s++){
Aee[s] = sign * Beta[s] * dw_diag;
sign = - sign;
}
Aee[Ls-1] += R;
See[0] = Aee[0];
for(int s=1;s<Ls;s++){
See[s] = Aee[s] - 1.0/See[s-1];
}
for(int s=0;s<Ls;s++){
std::cout<<GridLogMessage <<"s = "<<s<<" Beta "<<Beta[s]<<" Aee "<<Aee[s] <<" See "<<See[s] <<std::endl;
}
}
template<class Impl>
RealD ContinuedFractionFermion5D<Impl>::M (const FermionField &psi, FermionField &chi)
{
int Ls = this->Ls;
FermionField D(psi.Grid());
this->DW(psi,D,DaggerNo);
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==0 ) {
ag5xpby_ssp(chi,cc[0]*Beta[0]*sign*ZoloHiInv,D,sqrt_cc[0],psi,s,s+1); // Multiplies Dw by G5 so Hw
} else if ( s==(Ls-1) ){
RealD R=(1.0+mass)/(1.0-mass);
ag5xpby_ssp(chi,Beta[s]*ZoloHiInv,D,sqrt_cc[s-1],psi,s,s-1);
ag5xpby_ssp(chi,R,psi,1.0,chi,s,s);
} else {
ag5xpby_ssp(chi,cc[s]*Beta[s]*sign*ZoloHiInv,D,sqrt_cc[s],psi,s,s+1);
axpby_ssp(chi,1.0,chi,sqrt_cc[s-1],psi,s,s-1);
}
sign=-sign;
}
return norm2(chi);
}
template<class Impl>
RealD ContinuedFractionFermion5D<Impl>::Mdag (const FermionField &psi, FermionField &chi)
{
// This matrix is already hermitian. (g5 Dw) = Dw dag g5 = (g5 Dw)dag
// The rest of matrix is symmetric.
// Can ignore "dag"
return M(psi,chi);
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,int dir,int disp){
int Ls = this->Ls;
this->DhopDir(psi,chi,dir,disp); // Dslash on diagonal. g5 Dslash is hermitian
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==(Ls-1) ){
ag5xpby_ssp(chi,Beta[s]*ZoloHiInv,chi,0.0,chi,s,s);
} else {
ag5xpby_ssp(chi,cc[s]*Beta[s]*sign*ZoloHiInv,chi,0.0,chi,s,s);
}
sign=-sign;
}
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::Meooe (const FermionField &psi, FermionField &chi)
{
int Ls = this->Ls;
// Apply 4d dslash
if ( psi.Checkerboard() == Odd ) {
this->DhopEO(psi,chi,DaggerNo); // Dslash on diagonal. g5 Dslash is hermitian
} else {
this->DhopOE(psi,chi,DaggerNo); // Dslash on diagonal. g5 Dslash is hermitian
}
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==(Ls-1) ){
ag5xpby_ssp(chi,Beta[s]*ZoloHiInv,chi,0.0,chi,s,s);
} else {
ag5xpby_ssp(chi,cc[s]*Beta[s]*sign*ZoloHiInv,chi,0.0,chi,s,s);
}
sign=-sign;
}
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MeooeDag (const FermionField &psi, FermionField &chi)
{
this->Meooe(psi,chi);
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::Mooee (const FermionField &psi, FermionField &chi)
{
int Ls = this->Ls;
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==0 ) {
ag5xpby_ssp(chi,cc[0]*Beta[0]*sign*dw_diag,psi,sqrt_cc[0],psi,s,s+1); // Multiplies Dw by G5 so Hw
} else if ( s==(Ls-1) ){
// Drop the CC here.
double R=(1+mass)/(1-mass);
ag5xpby_ssp(chi,Beta[s]*dw_diag,psi,sqrt_cc[s-1],psi,s,s-1);
ag5xpby_ssp(chi,R,psi,1.0,chi,s,s);
} else {
ag5xpby_ssp(chi,cc[s]*Beta[s]*sign*dw_diag,psi,sqrt_cc[s],psi,s,s+1);
axpby_ssp(chi,1.0,chi,sqrt_cc[s-1],psi,s,s-1);
}
sign=-sign;
}
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MooeeDag (const FermionField &psi, FermionField &chi)
{
this->Mooee(psi,chi);
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MooeeInv (const FermionField &psi, FermionField &chi)
{
int Ls = this->Ls;
// Apply Linv
axpby_ssp(chi,1.0/cc_d[0],psi,0.0,psi,0,0);
for(int s=1;s<Ls;s++){
axpbg5y_ssp(chi,1.0/cc_d[s],psi,-1.0/See[s-1],chi,s,s-1);
}
// Apply Dinv
for(int s=0;s<Ls;s++){
ag5xpby_ssp(chi,1.0/See[s],chi,0.0,chi,s,s); //only appearance of See[0]
}
// Apply Uinv = (Linv)^T
axpby_ssp(chi,1.0/cc_d[Ls-1],chi,0.0,chi,Ls-1,Ls-1);
for(int s=Ls-2;s>=0;s--){
axpbg5y_ssp(chi,1.0/cc_d[s],chi,-1.0*cc_d[s+1]/See[s]/cc_d[s],chi,s,s+1);
}
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
{
this->MooeeInv(psi,chi);
}
// force terms; five routines; default to Dhop on diagonal
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MDeriv (GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
int Ls = this->Ls;
FermionField D(V.Grid());
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==(Ls-1) ){
ag5xpby_ssp(D,Beta[s]*ZoloHiInv,U,0.0,U,s,s);
} else {
ag5xpby_ssp(D,cc[s]*Beta[s]*sign*ZoloHiInv,U,0.0,U,s,s);
}
sign=-sign;
}
this->DhopDeriv(mat,D,V,DaggerNo);
};
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MoeDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
int Ls = this->Ls;
FermionField D(V.Grid());
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==(Ls-1) ){
ag5xpby_ssp(D,Beta[s]*ZoloHiInv,U,0.0,U,s,s);
} else {
ag5xpby_ssp(D,cc[s]*Beta[s]*sign*ZoloHiInv,U,0.0,U,s,s);
}
sign=-sign;
}
this->DhopDerivOE(mat,D,V,DaggerNo);
};
template<class Impl>
void ContinuedFractionFermion5D<Impl>::MeoDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
int Ls = this->Ls;
FermionField D(V.Grid());
int sign=1;
for(int s=0;s<Ls;s++){
if ( s==(Ls-1) ){
ag5xpby_ssp(D,Beta[s]*ZoloHiInv,U,0.0,U,s,s);
} else {
ag5xpby_ssp(D,cc[s]*Beta[s]*sign*ZoloHiInv,U,0.0,U,s,s);
}
sign=-sign;
}
this->DhopDerivEO(mat,D,V,DaggerNo);
};
// Constructors
template<class Impl>
ContinuedFractionFermion5D<Impl>::ContinuedFractionFermion5D(
GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mass,RealD M5,const ImplParams &p) :
WilsonFermion5D<Impl>(_Umu,
FiveDimGrid, FiveDimRedBlackGrid,
FourDimGrid, FourDimRedBlackGrid,M5,p),
mass(_mass)
{
int Ls = this->Ls;
assert((Ls&0x1)==1); // Odd Ls required
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::ExportPhysicalFermionSolution(const FermionField &solution5d,FermionField &exported4d)
{
int Ls = this->Ls;
conformable(solution5d.Grid(),this->FermionGrid());
conformable(exported4d.Grid(),this->GaugeGrid());
ExtractSlice(exported4d, solution5d, Ls-1, Ls-1);
}
template<class Impl>
void ContinuedFractionFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d)
{
int Ls = this->Ls;
conformable(imported5d.Grid(),this->FermionGrid());
conformable(input4d.Grid() ,this->GaugeGrid());
FermionField tmp(this->FermionGrid());
tmp=Zero();
InsertSlice(input4d, tmp, Ls-1, Ls-1);
tmp=Gamma(Gamma::Algebra::Gamma5)*tmp;
this->Dminus(tmp,imported5d);
}
FermOpTemplateInstantiate(ContinuedFractionFermion5D);
NAMESPACE_END(Grid);

433
Grid/qcd/action/fermion/DomainWallEOFAFermion.cc
View File

@ -1,433 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermion.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid_Eigen_Dense.h>
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
DomainWallEOFAFermion<Impl>::DomainWallEOFAFermion(
GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mq1, RealD _mq2, RealD _mq3,
RealD _shift, int _pm, RealD _M5, const ImplParams &p) :
AbstractEOFAFermion<Impl>(_Umu, FiveDimGrid, FiveDimRedBlackGrid,
FourDimGrid, FourDimRedBlackGrid, _mq1, _mq2, _mq3,
_shift, _pm, _M5, 1.0, 0.0, p)
{
RealD eps = 1.0;
Approx::zolotarev_data *zdata = Approx::higham(eps,this->Ls);
assert(zdata->n == this->Ls);
std::cout << GridLogMessage << "DomainWallEOFAFermion with Ls=" << this->Ls << std::endl;
this->SetCoefficientsTanh(zdata, 1.0, 0.0);
Approx::zolotarev_free(zdata);
}
/***************************************************************
* Additional EOFA operators only called outside the inverter.
* Since speed is not essential, simple axpby-style
* implementations should be fine.
***************************************************************/
template<class Impl>
void DomainWallEOFAFermion<Impl>::Omega(const FermionField& psi, FermionField& Din, int sign, int dag)
{
int Ls = this->Ls;
Din = Zero();
if((sign == 1) && (dag == 0)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, Ls-1, 0); }
else if((sign == -1) && (dag == 0)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, 0); }
else if((sign == 1 ) && (dag == 1)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, Ls-1); }
else if((sign == -1) && (dag == 1)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, 0); }
}
// This is just the identity for DWF
template<class Impl>
void DomainWallEOFAFermion<Impl>::Dtilde(const FermionField& psi, FermionField& chi){ chi = psi; }
// This is just the identity for DWF
template<class Impl>
void DomainWallEOFAFermion<Impl>::DtildeInv(const FermionField& psi, FermionField& chi){ chi = psi; }
/*****************************************************************************************************/
template<class Impl>
RealD DomainWallEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->Meooe5D(psi, Din);
this->DW(Din, chi, DaggerNo);
axpby(chi, 1.0, 1.0, chi, psi);
this->M5D(psi, chi);
return(norm2(chi));
}
template<class Impl>
RealD DomainWallEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->DW(psi, Din, DaggerYes);
this->MeooeDag5D(Din, chi);
this->M5Ddag(psi, chi);
axpby(chi, 1.0, 1.0, chi, psi);
return(norm2(chi));
}
/********************************************************************
* Performance critical fermion operators called inside the inverter
********************************************************************/
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
int pm = this->pm;
RealD shift = this->shift;
RealD mq1 = this->mq1;
RealD mq2 = this->mq2;
RealD mq3 = this->mq3;
// coefficients for shift operator ( = shift*\gamma_{5}*R_{5}*\Delta_{\pm}(mq2,mq3)*P_{\pm} )
Coeff_t shiftp(0.0), shiftm(0.0);
if(shift != 0.0){
if(pm == 1){ shiftp = shift*(mq3-mq2); }
else{ shiftm = -shift*(mq3-mq2); }
}
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftm;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = mq1 + shiftp;
#if(0)
std::cout << GridLogMessage << "DomainWallEOFAFermion::M5D(FF&,FF&):" << std::endl;
for(int i=0; i<diag.size(); ++i){
std::cout << GridLogMessage << "diag[" << i << "] =" << diag[i] << std::endl;
}
for(int i=0; i<upper.size(); ++i){
std::cout << GridLogMessage << "upper[" << i << "] =" << upper[i] << std::endl;
}
for(int i=0; i<lower.size(); ++i){
std::cout << GridLogMessage << "lower[" << i << "] =" << lower[i] << std::endl;
}
#endif
this->M5D(psi, chi, chi, lower, diag, upper);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
int pm = this->pm;
RealD shift = this->shift;
RealD mq1 = this->mq1;
RealD mq2 = this->mq2;
RealD mq3 = this->mq3;
// coefficients for shift operator ( = shift*\gamma_{5}*R_{5}*\Delta_{\pm}(mq2,mq3)*P_{\pm} )
Coeff_t shiftp(0.0), shiftm(0.0);
if(shift != 0.0){
if(pm == 1){ shiftp = shift*(mq3-mq2); }
else{ shiftm = -shift*(mq3-mq2); }
}
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftp;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = mq1 + shiftm;
#if(0)
std::cout << GridLogMessage << "DomainWallEOFAFermion::M5Ddag(FF&,FF&):" << std::endl;
for(int i=0; i<diag.size(); ++i){
std::cout << GridLogMessage << "diag[" << i << "] =" << diag[i] << std::endl;
}
for(int i=0; i<upper.size(); ++i){
std::cout << GridLogMessage << "upper[" << i << "] =" << upper[i] << std::endl;
}
for(int i=0; i<lower.size(); ++i){
std::cout << GridLogMessage << "lower[" << i << "] =" << lower[i] << std::endl;
}
#endif
this->M5Ddag(psi, chi, chi, lower, diag, upper);
}
// half checkerboard operations
template<class Impl>
void DomainWallEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
upper[s] = -this->cee[s];
lower[s] = -this->cee[s];
}
upper[Ls-1] = this->dm;
lower[0] = this->dp;
this->M5D(psi, psi, chi, lower, diag, upper);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
upper[s] = -this->cee[s];
lower[s] = -this->cee[s];
}
upper[Ls-1] = this->dp;
lower[0] = this->dm;
this->M5Ddag(psi, psi, chi, lower, diag, upper);
}
/****************************************************************************************/
//Zolo
template<class Impl>
void DomainWallEOFAFermion<Impl>::SetCoefficientsInternal(RealD zolo_hi, Vector<Coeff_t>& gamma, RealD b, RealD c)
{
int Ls = this->Ls;
int pm = this->pm;
RealD mq1 = this->mq1;
RealD mq2 = this->mq2;
RealD mq3 = this->mq3;
RealD shift = this->shift;
////////////////////////////////////////////////////////
// Constants for the preconditioned matrix Cayley form
////////////////////////////////////////////////////////
this->bs.resize(Ls);
this->cs.resize(Ls);
this->aee.resize(Ls);
this->aeo.resize(Ls);
this->bee.resize(Ls);
this->beo.resize(Ls);
this->cee.resize(Ls);
this->ceo.resize(Ls);
for(int i=0; i<Ls; ++i){
this->bee[i] = 4.0 - this->M5 + 1.0;
this->cee[i] = 1.0;
}
for(int i=0; i<Ls; ++i){
this->aee[i] = this->cee[i];
this->bs[i] = this->beo[i] = 1.0;
this->cs[i] = this->ceo[i] = 0.0;
}
//////////////////////////////////////////
// EOFA shift terms
//////////////////////////////////////////
if(pm == 1){
this->dp = mq1*this->cee[0] + shift*(mq3-mq2);
this->dm = mq1*this->cee[Ls-1];
} else if(this->pm == -1) {
this->dp = mq1*this->cee[0];
this->dm = mq1*this->cee[Ls-1] - shift*(mq3-mq2);
} else {
this->dp = mq1*this->cee[0];
this->dm = mq1*this->cee[Ls-1];
}
//////////////////////////////////////////
// LDU decomposition of eeoo
//////////////////////////////////////////
this->dee.resize(Ls+1);
this->lee.resize(Ls);
this->leem.resize(Ls);
this->uee.resize(Ls);
this->ueem.resize(Ls);
for(int i=0; i<Ls; ++i){
if(i < Ls-1){
this->lee[i] = -this->cee[i+1]/this->bee[i]; // sub-diag entry on the ith column
this->leem[i] = this->dm/this->bee[i];
for(int j=0; j<i; j++){ this->leem[i] *= this->aee[j]/this->bee[j]; }
this->dee[i] = this->bee[i];
this->uee[i] = -this->aee[i]/this->bee[i]; // up-diag entry on the ith row
this->ueem[i] = this->dp / this->bee[0];
for(int j=1; j<=i; j++){ this->ueem[i] *= this->cee[j]/this->bee[j]; }
} else {
this->lee[i] = 0.0;
this->leem[i] = 0.0;
this->uee[i] = 0.0;
this->ueem[i] = 0.0;
}
}
{
Coeff_t delta_d = 1.0 / this->bee[0];
for(int j=1; j<Ls-1; j++){ delta_d *= this->cee[j] / this->bee[j]; }
this->dee[Ls-1] = this->bee[Ls-1] + this->cee[0] * this->dm * delta_d;
this->dee[Ls] = this->bee[Ls-1] + this->cee[Ls-1] * this->dp * delta_d;
}
int inv = 1;
this->MooeeInternalCompute(0, inv, this->MatpInv, this->MatmInv);
this->MooeeInternalCompute(1, inv, this->MatpInvDag, this->MatmInvDag);
}
// Recompute Cayley-form coefficients for different shift
template<class Impl>
void DomainWallEOFAFermion<Impl>::RefreshShiftCoefficients(RealD new_shift)
{
this->shift = new_shift;
Approx::zolotarev_data *zdata = Approx::higham(1.0, this->Ls);
this->SetCoefficientsTanh(zdata, 1.0, 0.0);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternalCompute(int dag, int inv,
Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
int Ls = this->Ls;
GridBase* grid = this->FermionRedBlackGrid();
int LLs = grid->_rdimensions[0];
if(LLs == Ls){ return; } // Not vectorised in 5th direction
Eigen::MatrixXcd Pplus = Eigen::MatrixXcd::Zero(Ls,Ls);
Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
for(int s=0; s<Ls; s++){
Pplus(s,s) = this->bee[s];
Pminus(s,s) = this->bee[s];
}
for(int s=0; s<Ls-1; s++){
Pminus(s,s+1) = -this->cee[s];
}
for(int s=0; s<Ls-1; s++){
Pplus(s+1,s) = -this->cee[s+1];
}
Pplus (0,Ls-1) = this->dp;
Pminus(Ls-1,0) = this->dm;
Eigen::MatrixXcd PplusMat ;
Eigen::MatrixXcd PminusMat;
#if(0)
std::cout << GridLogMessage << "Pplus:" << std::endl;
for(int s=0; s<Ls; ++s){
for(int ss=0; ss<Ls; ++ss){
std::cout << Pplus(s,ss) << "\t";
}
std::cout << std::endl;
}
std::cout << GridLogMessage << "Pminus:" << std::endl;
for(int s=0; s<Ls; ++s){
for(int ss=0; ss<Ls; ++ss){
std::cout << Pminus(s,ss) << "\t";
}
std::cout << std::endl;
}
#endif
if(inv) {
PplusMat = Pplus.inverse();
PminusMat = Pminus.inverse();
} else {
PplusMat = Pplus;
PminusMat = Pminus;
}
if(dag){
PplusMat.adjointInPlace();
PminusMat.adjointInPlace();
}
typedef typename SiteHalfSpinor::scalar_type scalar_type;
const int Nsimd = Simd::Nsimd();
Matp.resize(Ls*LLs);
Matm.resize(Ls*LLs);
for(int s2=0; s2<Ls; s2++){
for(int s1=0; s1<LLs; s1++){
int istride = LLs;
int ostride = 1;
Simd Vp;
Simd Vm;
scalar_type *sp = (scalar_type*) &Vp;
scalar_type *sm = (scalar_type*) &Vm;
for(int l=0; l<Nsimd; l++){
if(switcheroo<Coeff_t>::iscomplex()) {
sp[l] = PplusMat (l*istride+s1*ostride,s2);
sm[l] = PminusMat(l*istride+s1*ostride,s2);
} else {
// if real
scalar_type tmp;
tmp = PplusMat (l*istride+s1*ostride,s2);
sp[l] = scalar_type(tmp.real(),tmp.real());
tmp = PminusMat(l*istride+s1*ostride,s2);
sm[l] = scalar_type(tmp.real(),tmp.real());
}
}
Matp[LLs*s2+s1] = Vp;
Matm[LLs*s2+s1] = Vm;
}}
}
FermOpTemplateInstantiate(DomainWallEOFAFermion);
GparityFermOpTemplateInstantiate(DomainWallEOFAFermion);
NAMESPACE_END(Grid);

255
Grid/qcd/action/fermion/DomainWallEOFAFermioncache.cc
View File

@ -1,255 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermioncache.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
// FIXME -- make a version of these routines with site loop outermost for cache reuse.
// Pminus fowards
// Pplus backwards..
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
int Ls = this->Ls;
GridBase* grid = psi_i.Grid();
auto phi = phi_i.View();
auto psi = psi_i.View();
auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
for(int s=0; s<Ls; s++){
auto tmp = psi[0];
if(s==0) {
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5m(tmp, psi[ss+0]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
auto tmp = psi[0];
for(int s=0; s<Ls; s++){
if(s==0) {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5p(tmp, psi[ss+0]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionField& chi_i)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi=psi_i.View();
auto chi=chi_i.View();
int Ls = this->Ls;
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
auto tmp1 = psi[0];
auto tmp2 = psi[0];
// flops = 12*2*Ls + 12*2*Ls + 3*12*Ls + 12*2*Ls = 12*Ls * (9) = 108*Ls flops
// Apply (L^{\prime})^{-1}
chi[ss] = psi[ss]; // chi[0]=psi[0]
for(int s=1; s<Ls; s++){
spProj5p(tmp1, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - this->lee[s-1]*tmp1;
}
// L_m^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
spProj5m(tmp1, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - this->leem[s]*tmp1;
}
// U_m^{-1} D^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[s] + 1/d chi[s]
spProj5p(tmp1, chi[ss+Ls-1]);
chi[ss+s] = (1.0/this->dee[s])*chi[ss+s] - (this->ueem[s]/this->dee[Ls])*tmp1;
}
spProj5m(tmp2, chi[ss+Ls-1]);
chi[ss+Ls-1] = (1.0/this->dee[Ls])*tmp1 + (1.0/this->dee[Ls-1])*tmp2;
// Apply U^{-1}
for(int s=Ls-2; s>=0; s--){
spProj5m(tmp1, chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - this->uee[s]*tmp1;
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, FermionField& chi_i)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
assert(psi.Checkerboard() == psi.Checkerboard());
Vector<Coeff_t> ueec(Ls);
Vector<Coeff_t> deec(Ls+1);
Vector<Coeff_t> leec(Ls);
Vector<Coeff_t> ueemc(Ls);
Vector<Coeff_t> leemc(Ls);
for(int s=0; s<ueec.size(); s++){
ueec[s] = conjugate(this->uee[s]);
deec[s] = conjugate(this->dee[s]);
leec[s] = conjugate(this->lee[s]);
ueemc[s] = conjugate(this->ueem[s]);
leemc[s] = conjugate(this->leem[s]);
}
deec[Ls] = conjugate(this->dee[Ls]);
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
auto tmp1 = psi[0];
auto tmp2 = psi[0];
// Apply (U^{\prime})^{-dagger}
chi[ss] = psi[ss];
for(int s=1; s<Ls; s++){
spProj5m(tmp1, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - ueec[s-1]*tmp1;
}
// U_m^{-\dagger}
for(int s=0; s<Ls-1; s++){
spProj5p(tmp1, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - ueemc[s]*tmp1;
}
// L_m^{-\dagger} D^{-dagger}
for(int s=0; s<Ls-1; s++){
spProj5m(tmp1, chi[ss+Ls-1]);
chi[ss+s] = (1.0/deec[s])*chi[ss+s] - (leemc[s]/deec[Ls-1])*tmp1;
}
spProj5p(tmp2, chi[ss+Ls-1]);
chi[ss+Ls-1] = (1.0/deec[Ls-1])*tmp1 + (1.0/deec[Ls])*tmp2;
// Apply L^{-dagger}
for(int s=Ls-2; s>=0; s--){
spProj5p(tmp1, chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - leec[s]*tmp1;
}
});
this->MooeeInvTime += usecond();
}
#ifdef DOMAIN_WALL_EOFA_DPERP_CACHE
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplF);
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplD);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplF);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplD);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplF);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplD);
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplFH);
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplDF);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplFH);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplDF);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplFH);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplDF);
#endif
NAMESPACE_END(Grid);

613
Grid/qcd/action/fermion/DomainWallEOFAFermionvec.cc
View File

@ -1,613 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermionvec.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
/*
* Dense matrix versions of routines
*/
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd> > u(LLs);
Vector<iSinglet<Simd> > l(LLs);
Vector<iSinglet<Simd> > d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0;o<LLs;o++){ // outer
for(int i=0;i<nsimd;i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
assert(Nc == 3);
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5m(hp, psi[ss+vp]);
spProj5p(hm, psi[ss+vm]);
if (vp <= v){ rotate(hp, hp, 1); }
if (vm >= v){ rotate(hm, hm, nsimd-1); }
hp = 0.5*hp;
hm = 0.5*hm;
spRecon5m(fp, hp);
spRecon5p(fm, hm);
chi[ss+v] = d[v]*phi[ss+v];
chi[ss+v] = chi[ss+v] + u[v]*fp;
chi[ss+v] = chi[ss+v] + l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v==LLs-1) ? 0 : v+1;
int vm = (v==0) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(2)(0);
Simd hp_01 = psi[ss+vp]()(2)(1);
Simd hp_02 = psi[ss+vp]()(2)(2);
Simd hp_10 = psi[ss+vp]()(3)(0);
Simd hp_11 = psi[ss+vp]()(3)(1);
Simd hp_12 = psi[ss+vp]()(3)(2);
Simd hm_00 = psi[ss+vm]()(0)(0);
Simd hm_01 = psi[ss+vm]()(0)(1);
Simd hm_02 = psi[ss+vm]()(0)(2);
Simd hm_10 = psi[ss+vm]()(1)(0);
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
if(vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
// Can force these to real arithmetic and save 2x.
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd> > u(LLs);
Vector<iSinglet<Simd> > l(LLs);
Vector<iSinglet<Simd> > d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5p(hp, psi[ss+vp]);
spProj5m(hm, psi[ss+vm]);
if(vp <= v){ rotate(hp, hp, 1); }
if(vm >= v){ rotate(hm, hm, nsimd-1); }
hp = hp*0.5;
hm = hm*0.5;
spRecon5p(fp, hp);
spRecon5m(fm, hm);
chi[ss+v] = d[v]*phi[ss+v]+u[v]*fp;
chi[ss+v] = chi[ss+v] +l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if (vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
#ifdef AVX512
#include<simd/Intel512common.h>
#include<simd/Intel512avx.h>
#include<simd/Intel512single.h>
#endif
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternalAsm(const FermionField& psi_i, FermionField& chi_i,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
#ifndef AVX512
{
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
SiteHalfSpinor SiteChiP;
SiteHalfSpinor SiteChiM;
// Ls*Ls * 2 * 12 * vol flops
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
for(int l=0; l < Simd::Nsimd(); l++){ // simd lane
int s = s2 + l*LLs;
int lex = s2 + LLs*site;
if( s2==0 && l==0 ){
SiteChiP=Zero();
SiteChiM=Zero();
}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastP()(sp)(co), psi[lex]()(sp)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastM()(sp)(co), psi[lex]()(sp+2)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
SiteChiP()(sp)(co) = real_madd(Matp[LLs*s+s1]()()(), BcastP()(sp)(co), SiteChiP()(sp)(co)); // 1100 us.
SiteChiM()(sp)(co) = real_madd(Matm[LLs*s+s1]()()(), BcastM()(sp)(co), SiteChiM()(sp)(co)); // each found by commenting out
}}
}}
{
int lex = s1 + LLs*site;
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vstream(chi[lex]()(sp)(co), SiteChiP()(sp)(co));
vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
}}
}
}
}
#else
{
// pointers
// MASK_REGS;
#define Chi_00 %%zmm1
#define Chi_01 %%zmm2
#define Chi_02 %%zmm3
#define Chi_10 %%zmm4
#define Chi_11 %%zmm5
#define Chi_12 %%zmm6
#define Chi_20 %%zmm7
#define Chi_21 %%zmm8
#define Chi_22 %%zmm9
#define Chi_30 %%zmm10
#define Chi_31 %%zmm11
#define Chi_32 %%zmm12
#define BCAST0 %%zmm13
#define BCAST1 %%zmm14
#define BCAST2 %%zmm15
#define BCAST3 %%zmm16
#define BCAST4 %%zmm17
#define BCAST5 %%zmm18
#define BCAST6 %%zmm19
#define BCAST7 %%zmm20
#define BCAST8 %%zmm21
#define BCAST9 %%zmm22
#define BCAST10 %%zmm23
#define BCAST11 %%zmm24
int incr = LLs*LLs*sizeof(iSinglet<Simd>);
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
int lex = s2 + LLs*site;
uint64_t a0 = (uint64_t) &Matp[LLs*s2+s1]; // should be cacheable
uint64_t a1 = (uint64_t) &Matm[LLs*s2+s1];
uint64_t a2 = (uint64_t) &psi[lex];
for(int l=0; l<Simd::Nsimd(); l++){ // simd lane
if((s2+l)==0) {
asm(
VPREFETCH1(0,%2) VPREFETCH1(0,%1)
VPREFETCH1(12,%2) VPREFETCH1(13,%2)
VPREFETCH1(14,%2) VPREFETCH1(15,%2)
VBCASTCDUP(0,%2,BCAST0)
VBCASTCDUP(1,%2,BCAST1)
VBCASTCDUP(2,%2,BCAST2)
VBCASTCDUP(3,%2,BCAST3)
VBCASTCDUP(4,%2,BCAST4) VMULMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(5,%2,BCAST5) VMULMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(6,%2,BCAST6) VMULMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(7,%2,BCAST7) VMULMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(8,%2,BCAST8) VMULMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(9,%2,BCAST9) VMULMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(10,%2,BCAST10) VMULMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(11,%2,BCAST11) VMULMEM(0,%1,BCAST7,Chi_21)
VMULMEM(0,%1,BCAST8,Chi_22)
VMULMEM(0,%1,BCAST9,Chi_30)
VMULMEM(0,%1,BCAST10,Chi_31)
VMULMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
} else {
asm(
VBCASTCDUP(0,%2,BCAST0) VMADDMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(1,%2,BCAST1) VMADDMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(2,%2,BCAST2) VMADDMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(3,%2,BCAST3) VMADDMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(4,%2,BCAST4) VMADDMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(5,%2,BCAST5) VMADDMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(6,%2,BCAST6) VMADDMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(7,%2,BCAST7) VMADDMEM(0,%1,BCAST7,Chi_21)
VBCASTCDUP(8,%2,BCAST8) VMADDMEM(0,%1,BCAST8,Chi_22)
VBCASTCDUP(9,%2,BCAST9) VMADDMEM(0,%1,BCAST9,Chi_30)
VBCASTCDUP(10,%2,BCAST10) VMADDMEM(0,%1,BCAST10,Chi_31)
VBCASTCDUP(11,%2,BCAST11) VMADDMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
}
a0 = a0 + incr;
a1 = a1 + incr;
a2 = a2 + sizeof(typename Simd::scalar_type);
}
}
{
int lexa = s1+LLs*site;
asm (
VSTORE(0,%0,Chi_00) VSTORE(1 ,%0,Chi_01) VSTORE(2 ,%0,Chi_02)
VSTORE(3,%0,Chi_10) VSTORE(4 ,%0,Chi_11) VSTORE(5 ,%0,Chi_12)
VSTORE(6,%0,Chi_20) VSTORE(7 ,%0,Chi_21) VSTORE(8 ,%0,Chi_22)
VSTORE(9,%0,Chi_30) VSTORE(10,%0,Chi_31) VSTORE(11,%0,Chi_32)
: : "r" ((uint64_t)&chi[lexa]) : "memory" );
}
}
}
#undef Chi_00
#undef Chi_01
#undef Chi_02
#undef Chi_10
#undef Chi_11
#undef Chi_12
#undef Chi_20
#undef Chi_21
#undef Chi_22
#undef Chi_30
#undef Chi_31
#undef Chi_32
#undef BCAST0
#undef BCAST1
#undef BCAST2
#undef BCAST3
#undef BCAST4
#undef BCAST5
#undef BCAST6
#undef BCAST7
#undef BCAST8
#undef BCAST9
#undef BCAST10
#undef BCAST11
#endif
};
// Z-mobius version
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternalZAsm(const FermionField& psi, FermionField& chi,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
std::cout << "Error: zMobius not implemented for EOFA" << std::endl;
exit(-1);
};
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv)
{
chi.Checkerboard() = psi.Checkerboard();
int Ls = this->Ls;
int LLs = psi.Grid()->_rdimensions[0];
int vol = psi.Grid()->oSites()/LLs;
Vector<iSinglet<Simd> > Matp;
Vector<iSinglet<Simd> > Matm;
Vector<iSinglet<Simd> > *_Matp;
Vector<iSinglet<Simd> > *_Matm;
// MooeeInternalCompute(dag,inv,Matp,Matm);
if(inv && dag){
_Matp = &this->MatpInvDag;
_Matm = &this->MatmInvDag;
}
if(inv && (!dag)){
_Matp = &this->MatpInv;
_Matm = &this->MatmInv;
}
if(!inv){
MooeeInternalCompute(dag, inv, Matp, Matm);
_Matp = &Matp;
_Matm = &Matm;
}
assert(_Matp->size() == Ls*LLs);
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
if(switcheroo<Coeff_t>::iscomplex()){
thread_loop((auto site=0; site<vol; site++),{
MooeeInternalZAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
} else {
thread_loop((auto site=0; site<vol; site++){
MooeeInternalAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
}
this->MooeeInvTime += usecond();
}
#ifdef DOMAIN_WALL_EOFA_DPERP_VEC
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplD);
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplF);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplD);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplF);
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplDF);
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplFH);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplDF);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplFH);
template void DomainWallEOFAFermion<DomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<DomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<DomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<DomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
#endif
NAMESPACE_END(Grid);

267
Grid/qcd/action/fermion/DomainWallVec5dImpl.h Normal file
View File

@ -0,0 +1,267 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/FermionOperatorImpl.h
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#pragma once
NAMESPACE_BEGIN(Grid);
template<class S,class Representation = FundamentalRepresentation, class Options=CoeffReal>
class DomainWallVec5dImpl : public PeriodicGaugeImpl< GaugeImplTypes< S,Representation::Dimension> > {
public:
typedef PeriodicGaugeImpl<GaugeImplTypes<S, Representation::Dimension> > Gimpl;
INHERIT_GIMPL_TYPES(Gimpl);
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> > >;
typedef iImplSpinor<Simd> SiteSpinor;
typedef iImplPropagator<Simd> SitePropagator;
typedef iImplHalfSpinor<Simd> SiteHalfSpinor;
typedef iImplHalfCommSpinor<SimdL> SiteHalfCommSpinor;
typedef Lattice<SiteSpinor> FermionField;
typedef Lattice<SitePropagator> PropagatorField;
/////////////////////////////////////////////////
// Make the doubled gauge field a *scalar*
/////////////////////////////////////////////////
typedef iImplDoubledGaugeField<typename Simd::scalar_type> SiteDoubledGaugeField; // This is a scalar
typedef iImplGaugeField<typename Simd::scalar_type> SiteScalarGaugeField; // scalar
typedef iImplGaugeLink<typename Simd::scalar_type> SiteScalarGaugeLink; // scalar
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef WilsonCompressor<SiteHalfCommSpinor,SiteHalfSpinor, SiteSpinor> Compressor;
typedef WilsonImplParams ImplParams;
typedef WilsonStencil<SiteSpinor, SiteHalfSpinor,ImplParams> StencilImpl;
typedef typename StencilImpl::View_type StencilView;
ImplParams Params;
DomainWallVec5dImpl(const ImplParams &p = ImplParams()) : Params(p){};
template <class ref>
static accelerator_inline void loadLinkElement(Simd &reg, ref &memory)
{
vsplat(reg, memory);
}
template<class _Spinor>
static accelerator_inline void multLink(_Spinor &phi, const SiteDoubledGaugeField &U,
const _Spinor &chi, int mu, StencilEntry *SE,
StencilView &St)
{
#ifdef GPU_VEC
// Gauge link is scalarised
mult(&phi(), &U(mu), &chi());
#else
SiteGaugeLink UU;
for (int i = 0; i < Dimension; i++) {
for (int j = 0; j < Dimension; j++) {
vsplat(UU()()(i, j), U(mu)()(i, j));
}
}
mult(&phi(), &UU(), &chi());
#endif
}
#ifdef GPU_VEC
static accelerator_inline void copyLinkGpu(int lane,
SiteDoubledGaugeField & UU,
const SiteDoubledGaugeField &U)
{
UU = U;
}
static accelerator_inline void multLinkGpu(int lane,
typename SiteHalfSpinor::scalar_object &phi,
const SiteDoubledGaugeField &U,
const typename SiteHalfSpinor::scalar_object &chi,
int mu)
{
#if 1
typedef typename ExtractTypeMap<typename Simd::scalar_type>::extract_type extract_type;
SiteScalarGaugeLink U_l;
extract_type * U_mem = (extract_type *) &U(mu);
extract_type * U_stack= (extract_type *) &U_l;
for(int w=0;w<(sizeof(U_l)/sizeof(extract_type)) ;w++) U_stack[w] = U_mem[w];
phi() = U_l() * chi();
#else
auto U_l = U(mu);
phi() = U_l * chi();
#endif
}
#else
static accelerator_inline void multLinkGpu(int lane,
SiteHalfSpinor &phi,
const SiteDoubledGaugeField &U,
const SiteHalfSpinor &chi,
int mu)
{
auto U_l = U(mu);
phi() = U_l * chi();
}
#endif
static accelerator_inline void multLinkProp(SitePropagator &phi,
const SiteDoubledGaugeField &U,
const SitePropagator &chi,int mu)
{
SiteGaugeLink UU;
for (int i = 0; i < Dimension; i++) {
for (int j = 0; j < Dimension; j++) {
vsplat(UU()()(i, j), U(mu)()(i, j));
}
}
mult(&phi(), &UU(), &chi());
}
inline void DoubleStore(GridBase *GaugeGrid, DoubledGaugeField &Uds,const GaugeField &Umu)
{
SiteScalarGaugeField ScalarUmu;
SiteDoubledGaugeField ScalarUds;
GaugeLinkField U(Umu.Grid());
GaugeField Uadj(Umu.Grid());
for (int mu = 0; mu < Nd; mu++) {
U = PeekIndex<LorentzIndex>(Umu, mu);
U = adj(Cshift(U, mu, -1));
PokeIndex<LorentzIndex>(Uadj, U, mu);
}
for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
Coordinate lcoor;
GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
peekLocalSite(ScalarUmu, Umu, lcoor);
for (int mu = 0; mu < 4; mu++) ScalarUds(mu) = ScalarUmu(mu);
peekLocalSite(ScalarUmu, Uadj, lcoor);
for (int mu = 0; mu < 4; mu++) ScalarUds(mu + 4) = ScalarUmu(mu);
pokeLocalSite(ScalarUds, Uds, lcoor);
}
}
inline void InsertForce4D(GaugeField &mat, FermionField &Btilde,FermionField &A, int mu)
{
assert(0);
}
inline void outerProductImpl(PropagatorField &mat, const FermionField &Btilde, const FermionField &A){
assert(0);
}
inline void TraceSpinImpl(GaugeLinkField &mat, PropagatorField&P) {
assert(0);
}
inline void extractLinkField(std::vector<GaugeLinkField> &mat, DoubledGaugeField &Uds){
assert(0);
}
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde, int mu) {
assert(0);
// Following lines to be revised after Peter's addition of half prec
// missing put lane...
/*
typedef decltype(traceIndex<SpinIndex>(outerProduct(Btilde[0], Atilde[0]))) result_type;
unsigned int LLs = Btilde.Grid()->_rdimensions[0];
conformable(Atilde.Grid(),Btilde.Grid());
GridBase* grid = mat.Grid();
GridBase* Bgrid = Btilde.Grid();
unsigned int dimU = grid->Nd();
unsigned int dimF = Bgrid->Nd();
GaugeLinkField tmp(grid);
tmp = Zero();
// FIXME
// Current implementation works, thread safe, probably suboptimal
// Passing through the local coordinate for grid transformation
// the force grid is in general very different from the Ls vectorized grid
for (int so = 0; so < grid->oSites(); so++) {
std::vector<typename result_type::scalar_object> vres(Bgrid->Nsimd());
std::vector<int> ocoor; grid->oCoorFromOindex(ocoor,so);
for (int si = 0; si < tmp.Grid()->iSites(); si++){
typename result_type::scalar_object scalar_object; scalar_object = Zero();
std::vector<int> local_coor;
std::vector<int> icoor; grid->iCoorFromIindex(icoor,si);
grid->InOutCoorToLocalCoor(ocoor, icoor, local_coor);
for (int s = 0; s < LLs; s++) {
std::vector<int> slocal_coor(dimF);
slocal_coor[0] = s;
for (int s4d = 1; s4d< dimF; s4d++) slocal_coor[s4d] = local_coor[s4d-1];
int sF = Bgrid->oIndexReduced(slocal_coor);
assert(sF < Bgrid->oSites());
extract(traceIndex<SpinIndex>(outerProduct(Btilde[sF], Atilde[sF])), vres);
// sum across the 5d dimension
for (auto v : vres) scalar_object += v;
}
tmp[so].putlane(scalar_object, si);
}
}
PokeIndex<LorentzIndex>(mat, tmp, mu);
*/
}
};
typedef DomainWallVec5dImpl<vComplex ,FundamentalRepresentation, CoeffReal> DomainWallVec5dImplR; // Real.. whichever prec
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);

625
Grid/qcd/action/fermion/ImprovedStaggeredFermion.cc
View File

@ -1,625 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
Copyright (C) 2015
Author: Azusa Yamaguchi, Peter Boyle
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid.h>
NAMESPACE_BEGIN(Grid);
const std::vector<int>
ImprovedStaggeredFermionStatic::directions({0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3});
const std::vector<int>
ImprovedStaggeredFermionStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1, 3, 3, 3, 3, -3, -3, -3, -3});
/////////////////////////////////
// Constructor and gauge import
/////////////////////////////////
template <class Impl>
ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GridCartesian &Fgrid, GridRedBlackCartesian &Hgrid,
RealD _mass,
RealD _c1, RealD _c2,RealD _u0,
const ImplParams &p)
: Kernels(p),
_grid(&Fgrid),
_cbgrid(&Hgrid),
Stencil(&Fgrid, npoint, Even, directions, displacements,p),
StencilEven(&Hgrid, npoint, Even, directions, displacements,p), // source is Even
StencilOdd(&Hgrid, npoint, Odd, directions, displacements,p), // source is Odd
mass(_mass),
Lebesgue(_grid),
LebesgueEvenOdd(_cbgrid),
Umu(&Fgrid),
UmuEven(&Hgrid),
UmuOdd(&Hgrid),
UUUmu(&Fgrid),
UUUmuEven(&Hgrid),
UUUmuOdd(&Hgrid) ,
_tmp(&Hgrid)
{
int vol4;
int LLs=1;
c1=_c1;
c2=_c2;
u0=_u0;
vol4= _grid->oSites();
Stencil.BuildSurfaceList(LLs,vol4);
vol4= _cbgrid->oSites();
StencilEven.BuildSurfaceList(LLs,vol4);
StencilOdd.BuildSurfaceList(LLs,vol4);
}
template <class Impl>
ImprovedStaggeredFermion<Impl>::ImprovedStaggeredFermion(GaugeField &_Uthin, GaugeField &_Ufat, GridCartesian &Fgrid,
GridRedBlackCartesian &Hgrid, RealD _mass,
RealD _c1, RealD _c2,RealD _u0,
const ImplParams &p)
: ImprovedStaggeredFermion(Fgrid,Hgrid,_mass,_c1,_c2,_u0,p)
{
ImportGauge(_Uthin,_Ufat);
}
////////////////////////////////////////////////////////////
// Momentum space propagator should be
// https://arxiv.org/pdf/hep-lat/9712010.pdf
//
// mom space action.
// gamma_mu i ( c1 sin pmu + c2 sin 3 pmu ) + m
//
// must track through staggered flavour/spin reduction in literature to
// turn to free propagator for the one component chi field, a la page 4/5
// of above link to implmement fourier based solver.
////////////////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion<Impl>::ImportGaugeSimple(const GaugeField &_Utriple,const GaugeField &_Ufat)
{
/////////////////////////////////////////////////////////////////
// Trivial import; phases and fattening and such like preapplied
/////////////////////////////////////////////////////////////////
GaugeLinkField U(GaugeGrid());
for (int mu = 0; mu < Nd; mu++) {
U = PeekIndex<LorentzIndex>(_Utriple, mu);
PokeIndex<LorentzIndex>(UUUmu, U, mu );
U = adj( Cshift(U, mu, -3));
PokeIndex<LorentzIndex>(UUUmu, -U, mu+4 );
U = PeekIndex<LorentzIndex>(_Ufat, mu);
PokeIndex<LorentzIndex>(Umu, U, mu);
U = adj( Cshift(U, mu, -1));
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);
pickCheckerboard(Odd, UUUmuOdd, UUUmu);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::ImportGauge(const GaugeField &_Uthin,const GaugeField &_Ufat)
{
GaugeLinkField U(GaugeGrid());
////////////////////////////////////////////////////////
// Double Store should take two fields for Naik and one hop separately.
////////////////////////////////////////////////////////
Impl::DoubleStore(GaugeGrid(), UUUmu, Umu, _Uthin, _Ufat );
////////////////////////////////////////////////////////
// Apply scale factors to get the right fermion Kinetic term
// Could pass coeffs into the double store to save work.
// 0.5 ( U p(x+mu) - Udag(x-mu) p(x-mu) )
////////////////////////////////////////////////////////
for (int mu = 0; mu < Nd; mu++) {
U = PeekIndex<LorentzIndex>(Umu, mu);
PokeIndex<LorentzIndex>(Umu, U*( 0.5*c1/u0), mu );
U = PeekIndex<LorentzIndex>(Umu, mu+4);
PokeIndex<LorentzIndex>(Umu, U*(-0.5*c1/u0), mu+4);
U = PeekIndex<LorentzIndex>(UUUmu, mu);
PokeIndex<LorentzIndex>(UUUmu, U*( 0.5*c2/u0/u0/u0), mu );
U = PeekIndex<LorentzIndex>(UUUmu, mu+4);
PokeIndex<LorentzIndex>(UUUmu, U*(-0.5*c2/u0/u0/u0), mu+4);
}
CopyGaugeCheckerboards();
}
/////////////////////////////
// Implement the interface
/////////////////////////////
template <class Impl>
RealD ImprovedStaggeredFermion<Impl>::M(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);
return axpy_norm(out, mass, in, out);
}
template <class Impl>
RealD ImprovedStaggeredFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerYes);
return axpy_norm(out, mass, in, out);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerNo);
} else {
DhopOE(in, out, DaggerNo);
}
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerYes);
} else {
DhopOE(in, out, DaggerYes);
}
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
typename FermionField::scalar_type scal(mass);
out = scal * in;
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Mooee(in, out);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
out = (1.0 / (mass)) * in;
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::MooeeInvDag(const FermionField &in,
FermionField &out) {
out.Checkerboard() = in.Checkerboard();
MooeeInv(in, out);
}
///////////////////////////////////
// Internal
///////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U, DoubledGaugeField &UUU,
GaugeField & mat,
const FermionField &A, const FermionField &B, int dag) {
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor;
FermionField Btilde(B.Grid());
FermionField Atilde(B.Grid());
Atilde = A;
st.HaloExchange(B, compressor);
for (int mu = 0; mu < Nd; mu++) {
////////////////////////
// Call the single hop
////////////////////////
auto U_v = U.View();
auto UUU_v = UUU.View();
auto B_v = B.View();
auto Btilde_v = Btilde.View();
thread_loop( (int sss = 0; sss < B.Grid()->oSites(); sss++), {
Kernels::DhopDirKernel(st, U_v, UUU_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu,1);
});
// Force in three link terms
//
// Impl::InsertForce4D(mat, Btilde, Atilde, mu);
//
// dU_ac(x)/dt = i p_ab U_bc(x)
//
// => dS_f/dt = dS_f/dU_ac(x) . dU_ac(x)/dt = i p_ab U_bc(x) dS_f/dU_ac(x)
//
// One link: form fragments S_f = A U B
//
// write Btilde = U(x) B(x+mu)
//
// mat+= TraceIndex<SpinIndex>(outerProduct(Btilde,A));
//
// Three link: form fragments S_f = A UUU B
//
// mat+= outer ( A, UUUB) <-- Best take DhopDeriv with one linke or identity matrix
// mat+= outer ( AU, UUB) <-- and then use covariant cshift?
// mat+= outer ( AUU, UB) <-- Returned from call to DhopDir
assert(0);// need to figure out the force interface with a blasted three link term.
}
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _grid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
mat.Checkerboard() = U.Checkerboard();
DerivInternal(Stencil, Umu, UUUmu, mat, U, V, dag);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
assert(V.Checkerboard() == Even);
assert(U.Checkerboard() == Odd);
mat.Checkerboard() = Odd;
DerivInternal(StencilEven, UmuOdd, UUUmuOdd, mat, U, V, dag);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
assert(V.Checkerboard() == Odd);
assert(U.Checkerboard() == Even);
mat.Checkerboard() = Even;
DerivInternal(StencilOdd, UmuEven, UUUmuEven, mat, U, V, dag);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=2;
conformable(in.Grid(), _grid); // verifies full grid
conformable(in.Grid(), out.Grid());
out.Checkerboard() = in.Checkerboard();
DhopInternal(Stencil, Lebesgue, Umu, UUUmu, in, out, dag);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
assert(in.Checkerboard() == Even);
out.Checkerboard() = Odd;
DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, UUUmuOdd, in, out, dag);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag)
{
DhopCalls+=1;
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
assert(in.Checkerboard() == Odd);
out.Checkerboard() = Even;
DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, UUUmuEven, in, out, dag);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) {
DhopDir(in, out, dir, disp);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp) {
Compressor compressor;
Stencil.HaloExchange(in, compressor);
auto Umu_v = Umu.View();
auto UUUmu_v = UUUmu.View();
auto in_v = in.View();
auto out_v = out.View();
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++) , {
Kernels::DhopDirKernel(Stencil, Umu_v, UUUmu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dir, disp);
});
};
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
DoubledGaugeField &UUU,
const FermionField &in,
FermionField &out, int dag)
{
#ifdef GRID_OMP
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
else
#endif
DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
DoubledGaugeField &UUU,
const FermionField &in,
FermionField &out, int dag)
{
#ifdef GRID_OMP
Compressor compressor;
int len = U.Grid()->oSites();
const int LLs = 1;
DhopTotalTime -= usecond();
DhopFaceTime -= usecond();
st.Prepare();
st.HaloGather(in,compressor);
st.CommsMergeSHM(compressor);
DhopFaceTime += usecond();
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
DhopComputeTime -= usecond();
#pragma omp parallel
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = len;
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
// Interior = 1; Exterior = 0; must implement for staggered
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0);
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
// Interior = 1; Exterior = 0;
int sU = ss;
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,1,0);
}
}
} else {
st.CommunicateThreaded();
}
}
DhopComputeTime += usecond();
// First to enter, last to leave timing
DhopFaceTime -= usecond();
st.CommsMerge(compressor);
DhopFaceTime -= usecond();
DhopComputeTime2 -= usecond();
{
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
thread_loop( (int ss = 0; ss < sz; ss++) ,{
int sU = st.surface_list[ss];
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
});
} else {
int sz=st.surface_list.size();
thread_loop( (int ss = 0; ss < sz; ss++) ,{
int sU = st.surface_list[ss];
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),1,sU,in_v,out_v,0,1);
});
}
}
DhopComputeTime2 += usecond();
#else
assert(0);
#endif
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
DoubledGaugeField &UUU,
const FermionField &in,
FermionField &out, int dag)
{
assert((dag == DaggerNo) || (dag == DaggerYes));
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
Compressor compressor;
st.HaloExchange(in, compressor);
DhopCommTime += usecond();
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
DhopComputeTime -= usecond();
if (dag == DaggerYes) {
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++), {
Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
});
} else {
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++), {
Kernels::DhopSite(st, lo, U_v, UUU_v, st.CommBuf(), 1, sss, in_v, out_v);
});
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
};
////////////////////////////////////////////////////////////////
// Reporting
////////////////////////////////////////////////////////////////
template<class Impl>
void ImprovedStaggeredFermion<Impl>::Report(void)
{
Coordinate latt = _grid->GlobalDimensions();
RealD volume = 1; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_grid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void ImprovedStaggeredFermion<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopFaceTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
////////////////////////////////////////////////////////
// Conserved current - not yet implemented.
////////////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)
{
assert(0);
}
template <class Impl>
void ImprovedStaggeredFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx)
{
assert(0);
}
FermOpStaggeredTemplateInstantiate(ImprovedStaggeredFermion);
//AdjointFermOpTemplateInstantiate(ImprovedStaggeredFermion);
//TwoIndexFermOpTemplateInstantiate(ImprovedStaggeredFermion);
NAMESPACE_END(Grid);

672
Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
View File

@ -1,672 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h>
#include <Grid/perfmon/PerfCount.h>
NAMESPACE_BEGIN(Grid);
// S-direction is INNERMOST and takes no part in the parity.
const std::vector<int>
ImprovedStaggeredFermion5DStatic::directions({1,2,3,4,1,2,3,4,1,2,3,4,1,2,3,4});
const std::vector<int>
ImprovedStaggeredFermion5DStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1, 3, 3, 3, 3, -3, -3, -3, -3});
// 5d lattice for DWF.
template<class Impl>
ImprovedStaggeredFermion5D<Impl>::ImprovedStaggeredFermion5D(GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mass,
RealD _c1,RealD _c2, RealD _u0,
const ImplParams &p) :
Kernels(p),
_FiveDimGrid (&FiveDimGrid),
_FiveDimRedBlackGrid(&FiveDimRedBlackGrid),
_FourDimGrid (&FourDimGrid),
_FourDimRedBlackGrid(&FourDimRedBlackGrid),
Stencil (&FiveDimGrid,npoint,Even,directions,displacements,p),
StencilEven(&FiveDimRedBlackGrid,npoint,Even,directions,displacements,p), // source is Even
StencilOdd (&FiveDimRedBlackGrid,npoint,Odd ,directions,displacements,p), // source is Odd
mass(_mass),
c1(_c1),
c2(_c2),
u0(_u0),
Umu(&FourDimGrid),
UmuEven(&FourDimRedBlackGrid),
UmuOdd (&FourDimRedBlackGrid),
UUUmu(&FourDimGrid),
UUUmuEven(&FourDimRedBlackGrid),
UUUmuOdd(&FourDimRedBlackGrid),
Lebesgue(&FourDimGrid),
LebesgueEvenOdd(&FourDimRedBlackGrid),
_tmp(&FiveDimRedBlackGrid)
{
// some assertions
assert(FiveDimGrid._ndimension==5);
assert(FourDimGrid._ndimension==4);
assert(FourDimRedBlackGrid._ndimension==4);
assert(FiveDimRedBlackGrid._ndimension==5);
assert(FiveDimRedBlackGrid._checker_dim==1); // Don't checker the s direction
// extent of fifth dim and not spread out
Ls=FiveDimGrid._fdimensions[0];
assert(FiveDimRedBlackGrid._fdimensions[0]==Ls);
assert(FiveDimGrid._processors[0] ==1);
assert(FiveDimRedBlackGrid._processors[0] ==1);
// Other dimensions must match the decomposition of the four-D fields
for(int d=0;d<4;d++){
assert(FiveDimGrid._processors[d+1] ==FourDimGrid._processors[d]);
assert(FiveDimRedBlackGrid._processors[d+1] ==FourDimGrid._processors[d]);
assert(FourDimRedBlackGrid._processors[d] ==FourDimGrid._processors[d]);
assert(FiveDimGrid._fdimensions[d+1] ==FourDimGrid._fdimensions[d]);
assert(FiveDimRedBlackGrid._fdimensions[d+1]==FourDimGrid._fdimensions[d]);
assert(FourDimRedBlackGrid._fdimensions[d] ==FourDimGrid._fdimensions[d]);
assert(FiveDimGrid._simd_layout[d+1] ==FourDimGrid._simd_layout[d]);
assert(FiveDimRedBlackGrid._simd_layout[d+1]==FourDimGrid._simd_layout[d]);
assert(FourDimRedBlackGrid._simd_layout[d] ==FourDimGrid._simd_layout[d]);
}
if (Impl::LsVectorised) {
int nsimd = Simd::Nsimd();
// Dimension zero of the five-d is the Ls direction
assert(FiveDimGrid._simd_layout[0] ==nsimd);
assert(FiveDimRedBlackGrid._simd_layout[0]==nsimd);
for(int d=0;d<4;d++){
assert(FourDimGrid._simd_layout[d]==1);
assert(FourDimRedBlackGrid._simd_layout[d]==1);
assert(FiveDimRedBlackGrid._simd_layout[d+1]==1);
}
} else {
// Dimension zero of the five-d is the Ls direction
assert(FiveDimRedBlackGrid._simd_layout[0]==1);
assert(FiveDimGrid._simd_layout[0] ==1);
}
int LLs = FiveDimGrid._rdimensions[0];
int vol4= FourDimGrid.oSites();
Stencil.BuildSurfaceList(LLs,vol4);
vol4=FourDimRedBlackGrid.oSites();
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,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mass,
RealD _c1,RealD _c2, RealD _u0,
const ImplParams &p) :
ImprovedStaggeredFermion5D(FiveDimGrid,FiveDimRedBlackGrid,
FourDimGrid,FourDimRedBlackGrid,
_mass,_c1,_c2,_u0,p)
{
ImportGauge(_Uthin,_Ufat);
}
///////////////////////////////////////////////////
// For MILC use; pass three link U's and 1 link U
///////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::ImportGaugeSimple(const GaugeField &_Utriple,const GaugeField &_Ufat)
{
/////////////////////////////////////////////////////////////////
// Trivial import; phases and fattening and such like preapplied
/////////////////////////////////////////////////////////////////
for (int mu = 0; mu < Nd; mu++) {
auto U = PeekIndex<LorentzIndex>(_Utriple, mu);
Impl::InsertGaugeField(UUUmu,U,mu);
U = adj( Cshift(U, mu, -3));
Impl::InsertGaugeField(UUUmu,-U,mu+4);
U = PeekIndex<LorentzIndex>(_Ufat, mu);
Impl::InsertGaugeField(Umu,U,mu);
U = adj( Cshift(U, mu, -1));
Impl::InsertGaugeField(Umu,-U,mu+4);
}
CopyGaugeCheckerboards();
}
template <class Impl>
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)
{
////////////////////////////////////////////////////////
// Double Store should take two fields for Naik and one hop separately.
////////////////////////////////////////////////////////
Impl::DoubleStore(GaugeGrid(), UUUmu, Umu, _Uthin, _Ufat );
////////////////////////////////////////////////////////
// Apply scale factors to get the right fermion Kinetic term
// Could pass coeffs into the double store to save work.
// 0.5 ( U p(x+mu) - Udag(x-mu) p(x-mu) )
////////////////////////////////////////////////////////
for (int mu = 0; mu < Nd; mu++) {
auto U = PeekIndex<LorentzIndex>(Umu, mu);
PokeIndex<LorentzIndex>(Umu, U*( 0.5*c1/u0), mu );
U = PeekIndex<LorentzIndex>(Umu, mu+4);
PokeIndex<LorentzIndex>(Umu, U*(-0.5*c1/u0), mu+4);
U = PeekIndex<LorentzIndex>(UUUmu, mu);
PokeIndex<LorentzIndex>(UUUmu, U*( 0.5*c2/u0/u0/u0), mu );
U = PeekIndex<LorentzIndex>(UUUmu, mu+4);
PokeIndex<LorentzIndex>(UUUmu, U*(-0.5*c2/u0/u0/u0), mu+4);
}
CopyGaugeCheckerboards();
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopDir(const FermionField &in, FermionField &out,int dir5,int disp)
{
int dir = dir5-1; // Maps to the ordering above in "directions" that is passed to stencil
// we drop off the innermost fifth dimension
Compressor compressor;
Stencil.HaloExchange(in,compressor);
auto Umu_v = Umu.View();
auto UUUmu_v = UUUmu.View();
auto in_v = in.View();
auto out_v = out.View();
thread_loop( (int ss=0;ss<Umu.Grid()->oSites();ss++),{
for(int s=0;s<Ls;s++){
int sU=ss;
int sF = s+Ls*sU;
Kernels::DhopDirKernel(Stencil, Umu_v, UUUmu_v, Stencil.CommBuf(), sF, sU, in_v, out_v, dir, disp);
}
});
};
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DerivInternal(StencilImpl & st,
DoubledGaugeField & U,
DoubledGaugeField & UUU,
GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
{
// No force terms in multi-rhs solver staggered
assert(0);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopDeriv(GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
{
assert(0);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopDerivEO(GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
{
assert(0);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
{
assert(0);
}
/*CHANGE */
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,DoubledGaugeField & UUU,
const FermionField &in, FermionField &out,int dag)
{
#ifdef GRID_OMP
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
else
#endif
DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,DoubledGaugeField & UUU,
const FermionField &in, FermionField &out,int dag)
{
#ifdef GRID_OMP
// assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor;
int LLs = in.Grid()->_rdimensions[0];
int len = U.Grid()->oSites();
DhopFaceTime-=usecond();
st.Prepare();
st.HaloGather(in,compressor);
// st.HaloExchangeOptGather(in,compressor); // Wilson compressor
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
double ctime=0;
double ptime=0;
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
double start = usecond();
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = U.Grid()->oSites(); // 4d vol
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
// Interior = 1; Exterior = 0; must implement for staggered
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<---------
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
// Interior = 1; Exterior = 0;
int sU = ss;
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,1,0); //<------------
}
}
ptime = usecond() - start;
} else {
double start = usecond();
st.CommunicateThreaded();
ctime = usecond() - start;
}
}
DhopCommTime += ctime;
DhopComputeTime+=ptime;
// First to enter, last to leave timing
st.CollateThreads();
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
DhopComputeTime2-=usecond();
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
thread_loop( (int ss = 0; ss < sz; ss++) ,{
int sU = st.surface_list[ss];
Kernels::DhopSiteDag(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1); //<----------
});
} else {
int sz=st.surface_list.size();
thread_loop( (int ss = 0; ss < sz; ss++) ,{
int sU = st.surface_list[ss];
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v,0,1);//<----------
});
}
DhopComputeTime2+=usecond();
#else
assert(0);
#endif
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,DoubledGaugeField & UUU,
const FermionField &in, FermionField &out,int dag)
{
Compressor compressor;
int LLs = in.Grid()->_rdimensions[0];
//double t1=usecond();
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
st.HaloExchange(in,compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
auto U_v = U.View();
auto UUU_v = UUU.View();
auto in_v = in.View();
auto out_v = out.View();
if (dag == DaggerYes) {
thread_loop( (int ss = 0; ss < U.Grid()->oSites(); ss++), {
int sU=ss;
Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), LLs, sU,in_v, out_v);
});
} else {
thread_loop( (int ss = 0; ss < U.Grid()->oSites(); ss++) ,{
int sU=ss;
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v);
});
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
//double t2=usecond();
//std::cout << __FILE__ << " " << __func__ << " Total Time " << DhopTotalTime << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Total Time Org " << t2-t1 << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Comml Time " << DhopCommTime << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Compute Time " << DhopComputeTime << std::endl;
}
/*CHANGE END*/
/* ORG
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,DoubledGaugeField & UUU,
const FermionField &in, FermionField &out,int dag)
{
Compressor compressor;
int LLs = in.Grid()->_rdimensions[0];
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
st.HaloExchange(in,compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
auto U_v = U.View();
auto UUU_v = UUU.View();
auto out_v = out.View();
auto in_v = in.View();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
if (dag == DaggerYes) {
thread_loop( (int ss = 0; ss < U.Grid()->oSites(); ss++), {
int sU=ss;
Kernels::DhopSiteDag(st, lo, U_v, UUU_v, st.CommBuf(), LLs, sU,in_v, out_v);
});
} else {
thread_loop( (int ss = 0; ss < U.Grid()->oSites(); ss++) ,{
int sU=ss;
Kernels::DhopSite(st,lo,U_v,UUU_v,st.CommBuf(),LLs,sU,in_v,out_v);
});
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
}
*/
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in.Grid(),FermionRedBlackGrid()); // verifies half grid
conformable(in.Grid(),out.Grid()); // drops the cb check
assert(in.Checkerboard()==Even);
out.Checkerboard() = Odd;
DhopInternal(StencilEven,LebesgueEvenOdd,UmuOdd,UUUmuOdd,in,out,dag);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=1;
conformable(in.Grid(),FermionRedBlackGrid()); // verifies half grid
conformable(in.Grid(),out.Grid()); // drops the cb check
assert(in.Checkerboard()==Odd);
out.Checkerboard() = Even;
DhopInternal(StencilOdd,LebesgueEvenOdd,UmuEven,UUUmuEven,in,out,dag);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
{
DhopCalls+=2;
conformable(in.Grid(),FermionGrid()); // verifies full grid
conformable(in.Grid(),out.Grid());
out.Checkerboard() = in.Checkerboard();
DhopInternal(Stencil,Lebesgue,Umu,UUUmu,in,out,dag);
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::Report(void)
{
Coordinate latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_FourDimGrid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void ImprovedStaggeredFermion5D<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopFaceTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
/////////////////////////////////////////////////////////////////////////
// Implement the general interface. Here we use SAME mass on all slices
/////////////////////////////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) {
DhopDir(in, out, dir, disp);
}
template <class Impl>
RealD ImprovedStaggeredFermion5D<Impl>::M(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);
return axpy_norm(out, mass, in, out);
}
template <class Impl>
RealD ImprovedStaggeredFermion5D<Impl>::Mdag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerYes);
return axpy_norm(out, mass, in, out);
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::Meooe(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerNo);
} else {
DhopOE(in, out, DaggerNo);
}
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerYes);
} else {
DhopOE(in, out, DaggerYes);
}
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::Mooee(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
typename FermionField::scalar_type scal(mass);
out = scal * in;
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Mooee(in, out);
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
out = (1.0 / (mass)) * in;
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::MooeeInvDag(const FermionField &in,
FermionField &out) {
out.Checkerboard() = in.Checkerboard();
MooeeInv(in, out);
}
////////////////////////////////////////////////////////
// Conserved current - not yet implemented.
////////////////////////////////////////////////////////
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)
{
assert(0);
}
template <class Impl>
void ImprovedStaggeredFermion5D<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx)
{
assert(0);
}
FermOpStaggeredTemplateInstantiate(ImprovedStaggeredFermion5D);
FermOpStaggeredVec5dTemplateInstantiate(ImprovedStaggeredFermion5D);
NAMESPACE_END(Grid);

497
Grid/qcd/action/fermion/MobiusEOFAFermion.cc
View File

@ -1,497 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/MobiusEOFAFermion.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid_Eigen_Dense.h>
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/MobiusEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
MobiusEOFAFermion<Impl>::MobiusEOFAFermion(
GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mq1, RealD _mq2, RealD _mq3,
RealD _shift, int _pm, RealD _M5,
RealD _b, RealD _c, const ImplParams &p) :
AbstractEOFAFermion<Impl>(_Umu, FiveDimGrid, FiveDimRedBlackGrid,
FourDimGrid, FourDimRedBlackGrid, _mq1, _mq2, _mq3,
_shift, _pm, _M5, _b, _c, p)
{
int Ls = this->Ls;
RealD eps = 1.0;
Approx::zolotarev_data *zdata = Approx::higham(eps, this->Ls);
assert(zdata->n == this->Ls);
std::cout << GridLogMessage << "MobiusEOFAFermion (b=" << _b <<
",c=" << _c << ") with Ls=" << Ls << std::endl;
this->SetCoefficientsTanh(zdata, _b, _c);
std::cout << GridLogMessage << "EOFA parameters: (mq1=" << _mq1 <<
",mq2=" << _mq2 << ",mq3=" << _mq3 << ",shift=" << _shift <<
",pm=" << _pm << ")" << std::endl;
Approx::zolotarev_free(zdata);
if(_shift != 0.0){
SetCoefficientsPrecondShiftOps();
} else {
Mooee_shift.resize(Ls, 0.0);
MooeeInv_shift_lc.resize(Ls, 0.0);
MooeeInv_shift_norm.resize(Ls, 0.0);
MooeeInvDag_shift_lc.resize(Ls, 0.0);
MooeeInvDag_shift_norm.resize(Ls, 0.0);
}
}
/****************************************************************
* Additional EOFA operators only called outside the inverter.
* Since speed is not essential, simple axpby-style
* implementations should be fine.
***************************************************************/
template<class Impl>
void MobiusEOFAFermion<Impl>::Omega(const FermionField& psi, FermionField& Din, int sign, int dag)
{
int Ls = this->Ls;
RealD alpha = this->alpha;
Din = Zero();
if((sign == 1) && (dag == 0)) { // \Omega_{+}
for(int s=0; s<Ls; ++s){
axpby_ssp(Din, 0.0, psi, 2.0*std::pow(1.0-alpha,Ls-s-1)/std::pow(1.0+alpha,Ls-s), psi, s, 0);
}
} else if((sign == -1) && (dag == 0)) { // \Omega_{-}
for(int s=0; s<Ls; ++s){
axpby_ssp(Din, 0.0, psi, 2.0*std::pow(1.0-alpha,s)/std::pow(1.0+alpha,s+1), psi, s, 0);
}
} else if((sign == 1 ) && (dag == 1)) { // \Omega_{+}^{\dagger}
for(int sp=0; sp<Ls; ++sp){
axpby_ssp(Din, 1.0, Din, 2.0*std::pow(1.0-alpha,Ls-sp-1)/std::pow(1.0+alpha,Ls-sp), psi, 0, sp);
}
} else if((sign == -1) && (dag == 1)) { // \Omega_{-}^{\dagger}
for(int sp=0; sp<Ls; ++sp){
axpby_ssp(Din, 1.0, Din, 2.0*std::pow(1.0-alpha,sp)/std::pow(1.0+alpha,sp+1), psi, 0, sp);
}
}
}
// This is the operator relating the usual Ddwf to TWQCD's EOFA Dirac operator (arXiv:1706.05843, Eqn. 6).
// It also relates the preconditioned and unpreconditioned systems described in Appendix B.2.
template<class Impl>
void MobiusEOFAFermion<Impl>::Dtilde(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
RealD b = 0.5 * ( 1.0 + this->alpha );
RealD c = 0.5 * ( 1.0 - this->alpha );
RealD mq1 = this->mq1;
for(int s=0; s<Ls; ++s){
if(s == 0) {
axpby_ssp_pminus(chi, b, psi, -c, psi, s, s+1);
axpby_ssp_pplus (chi, 1.0, chi, mq1*c, psi, s, Ls-1);
} else if(s == (Ls-1)) {
axpby_ssp_pminus(chi, b, psi, mq1*c, psi, s, 0);
axpby_ssp_pplus (chi, 1.0, chi, -c, psi, s, s-1);
} else {
axpby_ssp_pminus(chi, b, psi, -c, psi, s, s+1);
axpby_ssp_pplus (chi, 1.0, chi, -c, psi, s, s-1);
}
}
}
template<class Impl>
void MobiusEOFAFermion<Impl>::DtildeInv(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
RealD m = this->mq1;
RealD c = 0.5 * this->alpha;
RealD d = 0.5;
RealD DtInv_p(0.0), DtInv_m(0.0);
RealD N = std::pow(c+d,Ls) + m*std::pow(c-d,Ls);
FermionField tmp(this->FermionGrid());
for(int s=0; s<Ls; ++s){
for(int sp=0; sp<Ls; ++sp){
DtInv_p = m * std::pow(-1.0,s-sp+1) * std::pow(c-d,Ls+s-sp) / std::pow(c+d,s-sp+1) / N;
DtInv_p += (s < sp) ? 0.0 : std::pow(-1.0,s-sp) * std::pow(c-d,s-sp) / std::pow(c+d,s-sp+1);
DtInv_m = m * std::pow(-1.0,sp-s+1) * std::pow(c-d,Ls+sp-s) / std::pow(c+d,sp-s+1) / N;
DtInv_m += (s > sp) ? 0.0 : std::pow(-1.0,sp-s) * std::pow(c-d,sp-s) / std::pow(c+d,sp-s+1);
if(sp == 0){
axpby_ssp_pplus (tmp, 0.0, tmp, DtInv_p, psi, s, sp);
axpby_ssp_pminus(tmp, 0.0, tmp, DtInv_m, psi, s, sp);
} else {
axpby_ssp_pplus (tmp, 1.0, tmp, DtInv_p, psi, s, sp);
axpby_ssp_pminus(tmp, 1.0, tmp, DtInv_m, psi, s, sp);
}
}}
}
/*****************************************************************************************************/
template<class Impl>
RealD MobiusEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->Meooe5D(psi, Din);
this->DW(Din, chi, DaggerNo);
axpby(chi, 1.0, 1.0, chi, psi);
this->M5D(psi, chi);
return(norm2(chi));
}
template<class Impl>
RealD MobiusEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->DW(psi, Din, DaggerYes);
this->MeooeDag5D(Din, chi);
this->M5Ddag(psi, chi);
axpby(chi, 1.0, 1.0, chi, psi);
return(norm2(chi));
}
/********************************************************************
* Performance critical fermion operators called inside the inverter
********************************************************************/
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = this->mq1;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = this->mq1;
// no shift term
if(this->shift == 0.0){ this->M5D(psi, chi, chi, lower, diag, upper); }
// fused M + shift operation
else{ this->M5D_shift(psi, chi, chi, lower, diag, upper, Mooee_shift); }
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = this->mq1;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = this->mq1;
// no shift term
if(this->shift == 0.0){ this->M5Ddag(psi, chi, chi, lower, diag, upper); }
// fused M + shift operation
else{ this->M5Ddag_shift(psi, chi, chi, lower, diag, upper, Mooee_shift); }
}
// half checkerboard operations
template<class Impl>
void MobiusEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
// coefficients of Mooee
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
upper[s] = -this->cee[s];
lower[s] = -this->cee[s];
}
upper[Ls-1] *= -this->mq1;
lower[0] *= -this->mq1;
// no shift term
if(this->shift == 0.0){ this->M5D(psi, psi, chi, lower, diag, upper); }
// fused M + shift operation
else { this->M5D_shift(psi, psi, chi, lower, diag, upper, Mooee_shift); }
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
// coefficients of MooeeDag
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
if(s==0) {
upper[s] = -this->cee[s+1];
lower[s] = this->mq1*this->cee[Ls-1];
} else if(s==(Ls-1)) {
upper[s] = this->mq1*this->cee[0];
lower[s] = -this->cee[s-1];
} else {
upper[s] = -this->cee[s+1];
lower[s] = -this->cee[s-1];
}
}
// no shift term
if(this->shift == 0.0){ this->M5Ddag(psi, psi, chi, lower, diag, upper); }
// fused M + shift operation
else{ this->M5Ddag_shift(psi, psi, chi, lower, diag, upper, Mooee_shift); }
}
/****************************************************************************************/
// Computes coefficients for applying Cayley preconditioned shift operators
// (Mooee + \Delta) --> Mooee_shift
// (Mooee + \Delta)^{-1} --> MooeeInv_shift_lc, MooeeInv_shift_norm
// (Mooee + \Delta)^{-dag} --> MooeeInvDag_shift_lc, MooeeInvDag_shift_norm
// For the latter two cases, the operation takes the form
// [ (Mooee + \Delta)^{-1} \psi ]_{i} = Mooee_{ij} \psi_{j} +
// ( MooeeInv_shift_norm )_{i} ( \sum_{j} [ MooeeInv_shift_lc ]_{j} P_{pm} \psi_{j} )
template<class Impl>
void MobiusEOFAFermion<Impl>::SetCoefficientsPrecondShiftOps()
{
int Ls = this->Ls;
int pm = this->pm;
RealD alpha = this->alpha;
RealD k = this->k;
RealD mq1 = this->mq1;
RealD shift = this->shift;
// Initialize
Mooee_shift.resize(Ls);
MooeeInv_shift_lc.resize(Ls);
MooeeInv_shift_norm.resize(Ls);
MooeeInvDag_shift_lc.resize(Ls);
MooeeInvDag_shift_norm.resize(Ls);
// Construct Mooee_shift
int idx(0);
Coeff_t N = ( (pm == 1) ? 1.0 : -1.0 ) * (2.0*shift*k) *
( std::pow(alpha+1.0,Ls) + mq1*std::pow(alpha-1.0,Ls) );
for(int s=0; s<Ls; ++s){
idx = (pm == 1) ? (s) : (Ls-1-s);
Mooee_shift[idx] = N * std::pow(-1.0,s) * std::pow(alpha-1.0,s) / std::pow(alpha+1.0,Ls+s+1);
}
// Tridiagonal solve for MooeeInvDag_shift_lc
{
Coeff_t m(0.0);
Vector<Coeff_t> d = Mooee_shift;
Vector<Coeff_t> u(Ls,0.0);
Vector<Coeff_t> y(Ls,0.0);
Vector<Coeff_t> q(Ls,0.0);
if(pm == 1){ u[0] = 1.0; }
else{ u[Ls-1] = 1.0; }
// Tridiagonal matrix algorithm + Sherman-Morrison formula
//
// We solve
// ( Mooee' + u \otimes v ) MooeeInvDag_shift_lc = Mooee_shift
// where Mooee' is the tridiagonal part of Mooee_{+}, and
// u = (1,0,...,0) and v = (0,...,0,mq1*cee[0]) are chosen
// so that the outer-product u \otimes v gives the (0,Ls-1)
// entry of Mooee_{+}.
//
// We do this as two solves: Mooee'*y = d and Mooee'*q = u,
// and then construct the solution to the original system
// MooeeInvDag_shift_lc = y - <v,y> / ( 1 + <v,q> ) q
if(pm == 1){
for(int s=1; s<Ls; ++s){
m = -this->cee[s] / this->bee[s-1];
d[s] -= m*d[s-1];
u[s] -= m*u[s-1];
}
}
y[Ls-1] = d[Ls-1] / this->bee[Ls-1];
q[Ls-1] = u[Ls-1] / this->bee[Ls-1];
for(int s=Ls-2; s>=0; --s){
if(pm == 1){
y[s] = d[s] / this->bee[s];
q[s] = u[s] / this->bee[s];
} else {
y[s] = ( d[s] + this->cee[s]*y[s+1] ) / this->bee[s];
q[s] = ( u[s] + this->cee[s]*q[s+1] ) / this->bee[s];
}
}
// Construct MooeeInvDag_shift_lc
for(int s=0; s<Ls; ++s){
if(pm == 1){
MooeeInvDag_shift_lc[s] = y[s] - mq1*this->cee[0]*y[Ls-1] /
(1.0+mq1*this->cee[0]*q[Ls-1]) * q[s];
} else {
MooeeInvDag_shift_lc[s] = y[s] - mq1*this->cee[Ls-1]*y[0] /
(1.0+mq1*this->cee[Ls-1]*q[0]) * q[s];
}
}
// Compute remaining coefficients
N = (pm == 1) ? (1.0 + MooeeInvDag_shift_lc[Ls-1]) : (1.0 + MooeeInvDag_shift_lc[0]);
for(int s=0; s<Ls; ++s){
// MooeeInv_shift_lc
if(pm == 1){ MooeeInv_shift_lc[s] = pow(this->bee[s],s) * pow(this->cee[s],Ls-1-s); }
else { MooeeInv_shift_lc[s] = pow(this->bee[s],Ls-1-s) * pow(this->cee[s],s); }
// MooeeInv_shift_norm
MooeeInv_shift_norm[s] = -MooeeInvDag_shift_lc[s] /
( pow(this->bee[s],Ls) + mq1*pow(this->cee[s],Ls) ) / N;
// MooeeInvDag_shift_norm
if(pm == 1){ MooeeInvDag_shift_norm[s] = -pow(this->bee[s],s) * pow(this->cee[s],(Ls-1-s)) /
( pow(this->bee[s],Ls) + mq1*pow(this->cee[s],Ls) ) / N; }
else{ MooeeInvDag_shift_norm[s] = -pow(this->bee[s],(Ls-1-s)) * pow(this->cee[s],s) /
( pow(this->bee[s],Ls) + mq1*pow(this->cee[s],Ls) ) / N; }
}
}
}
// Recompute coefficients for a different value of shift constant
template<class Impl>
void MobiusEOFAFermion<Impl>::RefreshShiftCoefficients(RealD new_shift)
{
this->shift = new_shift;
if(new_shift != 0.0){
SetCoefficientsPrecondShiftOps();
} else {
int Ls = this->Ls;
Mooee_shift.resize(Ls,0.0);
MooeeInv_shift_lc.resize(Ls,0.0);
MooeeInv_shift_norm.resize(Ls,0.0);
MooeeInvDag_shift_lc.resize(Ls,0.0);
MooeeInvDag_shift_norm.resize(Ls,0.0);
}
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternalCompute(int dag, int inv,
Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
int Ls = this->Ls;
GridBase* grid = this->FermionRedBlackGrid();
int LLs = grid->_rdimensions[0];
if(LLs == Ls){ return; } // Not vectorised in 5th direction
Eigen::MatrixXcd Pplus = Eigen::MatrixXcd::Zero(Ls,Ls);
Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
for(int s=0; s<Ls; s++){
Pplus(s,s) = this->bee[s];
Pminus(s,s) = this->bee[s];
}
for(int s=0; s<Ls-1; s++){
Pminus(s,s+1) = -this->cee[s];
Pplus(s+1,s) = -this->cee[s+1];
}
Pplus (0,Ls-1) = this->mq1*this->cee[0];
Pminus(Ls-1,0) = this->mq1*this->cee[Ls-1];
if(this->shift != 0.0){
RealD c = 0.5 * this->alpha;
RealD d = 0.5;
RealD N = this->shift * this->k * ( std::pow(c+d,Ls) + this->mq1*std::pow(c-d,Ls) );
if(this->pm == 1) {
for(int s=0; s<Ls; ++s){
Pplus(s,Ls-1) += N * std::pow(-1.0,s) * std::pow(c-d,s) / std::pow(c+d,Ls+s+1);
}
} else {
for(int s=0; s<Ls; ++s){
Pminus(s,0) += N * std::pow(-1.0,s+1) * std::pow(c-d,Ls-1-s) / std::pow(c+d,2*Ls-s);
}
}
}
Eigen::MatrixXcd PplusMat ;
Eigen::MatrixXcd PminusMat;
if(inv) {
PplusMat = Pplus.inverse();
PminusMat = Pminus.inverse();
} else {
PplusMat = Pplus;
PminusMat = Pminus;
}
if(dag){
PplusMat.adjointInPlace();
PminusMat.adjointInPlace();
}
typedef typename SiteHalfSpinor::scalar_type scalar_type;
const int Nsimd = Simd::Nsimd();
Matp.resize(Ls*LLs);
Matm.resize(Ls*LLs);
for(int s2=0; s2<Ls; s2++){
for(int s1=0; s1<LLs; s1++){
int istride = LLs;
int ostride = 1;
Simd Vp;
Simd Vm;
scalar_type *sp = (scalar_type*) &Vp;
scalar_type *sm = (scalar_type*) &Vm;
for(int l=0; l<Nsimd; l++){
if(switcheroo<Coeff_t>::iscomplex()) {
sp[l] = PplusMat (l*istride+s1*ostride,s2);
sm[l] = PminusMat(l*istride+s1*ostride,s2);
} else {
// if real
scalar_type tmp;
tmp = PplusMat (l*istride+s1*ostride,s2);
sp[l] = scalar_type(tmp.real(),tmp.real());
tmp = PminusMat(l*istride+s1*ostride,s2);
sm[l] = scalar_type(tmp.real(),tmp.real());
}
}
Matp[LLs*s2+s1] = Vp;
Matm[LLs*s2+s1] = Vm;
}}
}
FermOpTemplateInstantiate(MobiusEOFAFermion);
GparityFermOpTemplateInstantiate(MobiusEOFAFermion);
NAMESPACE_END(Grid);

998
Grid/qcd/action/fermion/MobiusEOFAFermionvec.cc
View File

@ -1,998 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/MobiusEOFAFermionvec.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/MobiusEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
/*
* Dense matrix versions of routines
*/
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
assert(Nc == 3);
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5m(hp, psi[ss+vp]);
spProj5p(hm, psi[ss+vm]);
if (vp <= v){ rotate(hp, hp, 1); }
if (vm >= v){ rotate(hm, hm, nsimd-1); }
hp = 0.5*hp;
hm = 0.5*hm;
spRecon5m(fp, hp);
spRecon5p(fm, hm);
chi[ss+v] = d[v]*phi[ss+v];
chi[ss+v] = chi[ss+v] + u[v]*fp;
chi[ss+v] = chi[ss+v] + l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(2)(0);
Simd hp_01 = psi[ss+vp]()(2)(1);
Simd hp_02 = psi[ss+vp]()(2)(2);
Simd hp_10 = psi[ss+vp]()(3)(0);
Simd hp_11 = psi[ss+vp]()(3)(1);
Simd hp_12 = psi[ss+vp]()(3)(2);
Simd hm_00 = psi[ss+vm]()(0)(0);
Simd hm_01 = psi[ss+vm]()(0)(1);
Simd hm_02 = psi[ss+vm]()(0)(2);
Simd hm_10 = psi[ss+vm]()(1)(0);
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
if(vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
// Can force these to real arithmetic and save 2x.
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField& psi_i, const FermionField& phi_i,
FermionField& chi_i, Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
Vector<Coeff_t>& shift_coeffs)
{
#if 0
auto & psi = psi_i;
auto & phi = phi_i;
auto & chi = chi_i;
this->M5D(psi, phi, chi, lower, diag, upper);
// FIXME: possible gain from vectorizing shift operation as well?
Coeff_t one(1.0);
int Ls = this->Ls;
for(int s=0; s<Ls; s++){
if(this->pm == 1){ axpby_ssp_pplus(chi, one, chi, shift_coeffs[s], psi, s, Ls-1); }
else{ axpby_ssp_pminus(chi, one, chi, shift_coeffs[s], psi, s, 0); }
}
#else
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
Vector<iSinglet<Simd>> s(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
scalar_type* s_p = (scalar_type*) &s[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
s_p[ss] = shift_coeffs[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
assert(Nc == 3);
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
int vs = (this->pm == 1) ? LLs-1 : 0;
Simd hs_00 = (this->pm == 1) ? psi[ss+vs]()(2)(0) : psi[ss+vs]()(0)(0);
Simd hs_01 = (this->pm == 1) ? psi[ss+vs]()(2)(1) : psi[ss+vs]()(0)(1);
Simd hs_02 = (this->pm == 1) ? psi[ss+vs]()(2)(2) : psi[ss+vs]()(0)(2);
Simd hs_10 = (this->pm == 1) ? psi[ss+vs]()(3)(0) : psi[ss+vs]()(1)(0);
Simd hs_11 = (this->pm == 1) ? psi[ss+vs]()(3)(1) : psi[ss+vs]()(1)(1);
Simd hs_12 = (this->pm == 1) ? psi[ss+vs]()(3)(2) : psi[ss+vs]()(1)(2);
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(2)(0);
Simd hp_01 = psi[ss+vp]()(2)(1);
Simd hp_02 = psi[ss+vp]()(2)(2);
Simd hp_10 = psi[ss+vp]()(3)(0);
Simd hp_11 = psi[ss+vp]()(3)(1);
Simd hp_12 = psi[ss+vp]()(3)(2);
Simd hm_00 = psi[ss+vm]()(0)(0);
Simd hm_01 = psi[ss+vm]()(0)(1);
Simd hm_02 = psi[ss+vm]()(0)(2);
Simd hm_10 = psi[ss+vm]()(1)(0);
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
if(vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(this->pm == 1 && vs <= v){
hs_00.v = Optimization::Rotate::tRotate<2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2>(hs_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
if(this->pm == -1 && vs >= v){
hs_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_12.v);
}
// Can force these to real arithmetic and save 2x.
Simd p_00 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00);
Simd p_01 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01);
Simd p_02 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02);
Simd p_10 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10);
Simd p_11 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11);
Simd p_12 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12);
Simd p_20 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_21 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_22 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_30 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_31 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_32 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
});
this->M5Dtime += usecond();
#endif
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5p(hp, psi[ss+vp]);
spProj5m(hm, psi[ss+vm]);
if(vp <= v){ rotate(hp, hp, 1); }
if(vm >= v){ rotate(hm, hm, nsimd-1); }
hp = hp*0.5;
hm = hm*0.5;
spRecon5p(fp, hp);
spRecon5m(fm, hm);
chi[ss+v] = d[v]*phi[ss+v]+u[v]*fp;
chi[ss+v] = chi[ss+v] +l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if (vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
Vector<Coeff_t>& shift_coeffs)
{
#if 0
auto & psi = psi_i;
auto & phi = phi_i;
auto & chi = chi_i;
this->M5Ddag(psi, phi, chi, lower, diag, upper);
// FIXME: possible gain from vectorizing shift operation as well?
Coeff_t one(1.0);
int Ls = this->Ls;
for(int s=0; s<Ls; s++){
if(this->pm == 1){ axpby_ssp_pplus(chi, one, chi, shift_coeffs[s], psi, Ls-1, s); }
else{ axpby_ssp_pminus(chi, one, chi, shift_coeffs[s], psi, 0, s); }
}
#else
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
Vector<iSinglet<Simd>> s(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
scalar_type* s_p = (scalar_type*) &s[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
s_p[ss] = shift_coeffs[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
int vs = (this->pm == 1) ? LLs-1 : 0;
Simd hs_00 = (this->pm == 1) ? psi[ss+vs]()(0)(0) : psi[ss+vs]()(2)(0);
Simd hs_01 = (this->pm == 1) ? psi[ss+vs]()(0)(1) : psi[ss+vs]()(2)(1);
Simd hs_02 = (this->pm == 1) ? psi[ss+vs]()(0)(2) : psi[ss+vs]()(2)(2);
Simd hs_10 = (this->pm == 1) ? psi[ss+vs]()(1)(0) : psi[ss+vs]()(3)(0);
Simd hs_11 = (this->pm == 1) ? psi[ss+vs]()(1)(1) : psi[ss+vs]()(3)(1);
Simd hs_12 = (this->pm == 1) ? psi[ss+vs]()(1)(2) : psi[ss+vs]()(3)(2);
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if (vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(this->pm == 1 && vs <= v){
hs_00.v = Optimization::Rotate::tRotate<2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2>(hs_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
if(this->pm == -1 && vs >= v){
hs_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_12.v);
}
Simd p_00 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_01 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_02 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_10 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_11 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_12 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
Simd p_20 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00);
Simd p_21 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01);
Simd p_22 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02);
Simd p_30 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10);
Simd p_31 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11);
Simd p_32 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
});
this->M5Dtime += usecond();
#endif
}
#ifdef AVX512
#include<simd/Intel512common.h>
#include<simd/Intel512avx.h>
#include<simd/Intel512single.h>
#endif
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternalAsm(const FermionField& psi_i, FermionField& chi_i,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
auto psi = psi_i.View();
auto chi = chi_i.View();
#ifndef AVX512
{
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
SiteHalfSpinor SiteChiP;
SiteHalfSpinor SiteChiM;
// Ls*Ls * 2 * 12 * vol flops
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
for(int l=0; l < Simd::Nsimd(); l++){ // simd lane
int s = s2 + l*LLs;
int lex = s2 + LLs*site;
if( s2==0 && l==0 ){
SiteChiP=Zero();
SiteChiM=Zero();
}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastP()(sp)(co), psi[lex]()(sp)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastM()(sp)(co), psi[lex]()(sp+2)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
SiteChiP()(sp)(co) = real_madd(Matp[LLs*s+s1]()()(), BcastP()(sp)(co), SiteChiP()(sp)(co)); // 1100 us.
SiteChiM()(sp)(co) = real_madd(Matm[LLs*s+s1]()()(), BcastM()(sp)(co), SiteChiM()(sp)(co)); // each found by commenting out
}}
}}
{
int lex = s1 + LLs*site;
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vstream(chi[lex]()(sp)(co), SiteChiP()(sp)(co));
vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
}}
}
}
}
#else
{
// pointers
// MASK_REGS;
#define Chi_00 %%zmm1
#define Chi_01 %%zmm2
#define Chi_02 %%zmm3
#define Chi_10 %%zmm4
#define Chi_11 %%zmm5
#define Chi_12 %%zmm6
#define Chi_20 %%zmm7
#define Chi_21 %%zmm8
#define Chi_22 %%zmm9
#define Chi_30 %%zmm10
#define Chi_31 %%zmm11
#define Chi_32 %%zmm12
#define BCAST0 %%zmm13
#define BCAST1 %%zmm14
#define BCAST2 %%zmm15
#define BCAST3 %%zmm16
#define BCAST4 %%zmm17
#define BCAST5 %%zmm18
#define BCAST6 %%zmm19
#define BCAST7 %%zmm20
#define BCAST8 %%zmm21
#define BCAST9 %%zmm22
#define BCAST10 %%zmm23
#define BCAST11 %%zmm24
int incr = LLs*LLs*sizeof(iSinglet<Simd>);
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
int lex = s2 + LLs*site;
uint64_t a0 = (uint64_t) &Matp[LLs*s2+s1]; // should be cacheable
uint64_t a1 = (uint64_t) &Matm[LLs*s2+s1];
uint64_t a2 = (uint64_t) &psi[lex];
for(int l=0; l<Simd::Nsimd(); l++){ // simd lane
if((s2+l)==0) {
asm(
VPREFETCH1(0,%2) VPREFETCH1(0,%1)
VPREFETCH1(12,%2) VPREFETCH1(13,%2)
VPREFETCH1(14,%2) VPREFETCH1(15,%2)
VBCASTCDUP(0,%2,BCAST0)
VBCASTCDUP(1,%2,BCAST1)
VBCASTCDUP(2,%2,BCAST2)
VBCASTCDUP(3,%2,BCAST3)
VBCASTCDUP(4,%2,BCAST4) VMULMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(5,%2,BCAST5) VMULMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(6,%2,BCAST6) VMULMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(7,%2,BCAST7) VMULMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(8,%2,BCAST8) VMULMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(9,%2,BCAST9) VMULMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(10,%2,BCAST10) VMULMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(11,%2,BCAST11) VMULMEM(0,%1,BCAST7,Chi_21)
VMULMEM(0,%1,BCAST8,Chi_22)
VMULMEM(0,%1,BCAST9,Chi_30)
VMULMEM(0,%1,BCAST10,Chi_31)
VMULMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
} else {
asm(
VBCASTCDUP(0,%2,BCAST0) VMADDMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(1,%2,BCAST1) VMADDMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(2,%2,BCAST2) VMADDMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(3,%2,BCAST3) VMADDMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(4,%2,BCAST4) VMADDMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(5,%2,BCAST5) VMADDMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(6,%2,BCAST6) VMADDMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(7,%2,BCAST7) VMADDMEM(0,%1,BCAST7,Chi_21)
VBCASTCDUP(8,%2,BCAST8) VMADDMEM(0,%1,BCAST8,Chi_22)
VBCASTCDUP(9,%2,BCAST9) VMADDMEM(0,%1,BCAST9,Chi_30)
VBCASTCDUP(10,%2,BCAST10) VMADDMEM(0,%1,BCAST10,Chi_31)
VBCASTCDUP(11,%2,BCAST11) VMADDMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
}
a0 = a0 + incr;
a1 = a1 + incr;
a2 = a2 + sizeof(typename Simd::scalar_type);
}
}
{
int lexa = s1+LLs*site;
asm (
VSTORE(0,%0,Chi_00) VSTORE(1 ,%0,Chi_01) VSTORE(2 ,%0,Chi_02)
VSTORE(3,%0,Chi_10) VSTORE(4 ,%0,Chi_11) VSTORE(5 ,%0,Chi_12)
VSTORE(6,%0,Chi_20) VSTORE(7 ,%0,Chi_21) VSTORE(8 ,%0,Chi_22)
VSTORE(9,%0,Chi_30) VSTORE(10,%0,Chi_31) VSTORE(11,%0,Chi_32)
: : "r" ((uint64_t)&chi[lexa]) : "memory" );
}
}
}
#undef Chi_00
#undef Chi_01
#undef Chi_02
#undef Chi_10
#undef Chi_11
#undef Chi_12
#undef Chi_20
#undef Chi_21
#undef Chi_22
#undef Chi_30
#undef Chi_31
#undef Chi_32
#undef BCAST0
#undef BCAST1
#undef BCAST2
#undef BCAST3
#undef BCAST4
#undef BCAST5
#undef BCAST6
#undef BCAST7
#undef BCAST8
#undef BCAST9
#undef BCAST10
#undef BCAST11
#endif
};
// Z-mobius version
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternalZAsm(const FermionField& psi, FermionField& chi,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
std::cout << "Error: zMobius not implemented for EOFA" << std::endl;
exit(-1);
};
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv)
{
chi.Checkerboard() = psi.Checkerboard();
int Ls = this->Ls;
int LLs = psi.Grid()->_rdimensions[0];
int vol = psi.Grid()->oSites()/LLs;
Vector<iSinglet<Simd>> Matp;
Vector<iSinglet<Simd>> Matm;
Vector<iSinglet<Simd>>* _Matp;
Vector<iSinglet<Simd>>* _Matm;
// MooeeInternalCompute(dag,inv,Matp,Matm);
if(inv && dag){
_Matp = &this->MatpInvDag;
_Matm = &this->MatmInvDag;
}
if(inv && (!dag)){
_Matp = &this->MatpInv;
_Matm = &this->MatmInv;
}
if(!inv){
MooeeInternalCompute(dag, inv, Matp, Matm);
_Matp = &Matp;
_Matm = &Matm;
}
assert(_Matp->size() == Ls*LLs);
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
if(switcheroo<Coeff_t>::iscomplex()){
thread_loop( (auto site=0; site<vol; site++),{
MooeeInternalZAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
} else {
thread_loop( (auto site=0; site<vol; site++),{
MooeeInternalAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
}
this->MooeeInvTime += usecond();
}
#ifdef MOBIUS_EOFA_DPERP_VEC
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplD);
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplF);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplD);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplF);
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplDF);
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplFH);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplDF);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplFH);
template void MobiusEOFAFermion<DomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<DomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<DomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<DomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
#endif
NAMESPACE_END(Grid);

452
Grid/qcd/action/fermion/PartialFractionFermion5D.cc
View File

@ -1,452 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/PartialFractionFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/PartialFractionFermion5D.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
void PartialFractionFermion5D<Impl>::Mdir (const FermionField &psi, FermionField &chi,int dir,int disp){
// this does both dag and undag but is trivial; make a common helper routing
int Ls = this->Ls;
this->DhopDir(psi,chi,dir,disp);
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
ag5xpby_ssp(chi,-scale,chi,0.0,chi,s,s);
ag5xpby_ssp(chi, scale,chi,0.0,chi,s+1,s+1);
}
ag5xpby_ssp(chi,p[nblock]*scale/amax,chi,0.0,chi,Ls-1,Ls-1);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::Meooe_internal(const FermionField &psi, FermionField &chi,int dag)
{
int Ls = this->Ls;
if ( psi.Checkerboard() == Odd ) {
this->DhopEO(psi,chi,DaggerNo);
} else {
this->DhopOE(psi,chi,DaggerNo);
}
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
ag5xpby_ssp(chi,-scale,chi,0.0,chi,s,s);
ag5xpby_ssp(chi, scale,chi,0.0,chi,s+1,s+1);
}
ag5xpby_ssp(chi,p[nblock]*scale/amax,chi,0.0,chi,Ls-1,Ls-1);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::Mooee_internal(const FermionField &psi, FermionField &chi,int dag)
{
// again dag and undag are trivially related
int sign = dag ? (-1) : 1;
int Ls = this->Ls;
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
RealD pp = p[nblock-1-b];
RealD qq = q[nblock-1-b];
// Do each 2x2 block aligned at s and multiplies Dw site diagonal by G5 so Hw
ag5xpby_ssp(chi,-dw_diag*scale,psi,amax*sqrt(qq)*scale,psi, s ,s+1);
ag5xpby_ssp(chi, dw_diag*scale,psi,amax*sqrt(qq)*scale,psi, s+1,s);
axpby_ssp (chi, 1.0, chi,sqrt(amax*pp)*scale*sign,psi,s+1,Ls-1);
}
{
RealD R=(1+mass)/(1-mass);
//R g5 psi[Ls-1] + p[0] H
ag5xpbg5y_ssp(chi,R*scale,psi,p[nblock]*scale*dw_diag/amax,psi,Ls-1,Ls-1);
for(int b=0;b<nblock;b++){
int s = 2*b+1;
RealD pp = p[nblock-1-b];
axpby_ssp(chi,1.0,chi,-sqrt(amax*pp)*scale*sign,psi,Ls-1,s);
}
}
}
template<class Impl>
void PartialFractionFermion5D<Impl>::MooeeInv_internal(const FermionField &psi, FermionField &chi,int dag)
{
int sign = dag ? (-1) : 1;
int Ls = this->Ls;
FermionField tmp(psi.Grid());
///////////////////////////////////////////////////////////////////////////////////////
//Linv
///////////////////////////////////////////////////////////////////////////////////////
int nblock=(Ls-1)/2;
axpy(chi,0.0,psi,psi); // Identity piece
for(int b=0;b<nblock;b++){
int s = 2*b;
RealD pp = p[nblock-1-b];
RealD qq = q[nblock-1-b];
RealD coeff1=sign*sqrt(amax*amax*amax*pp*qq) / ( dw_diag*dw_diag + amax*amax* qq);
RealD coeff2=sign*sqrt(amax*pp)*dw_diag / ( dw_diag*dw_diag + amax*amax* qq); // Implicit g5 here
axpby_ssp (chi,1.0,chi,coeff1,psi,Ls-1,s);
axpbg5y_ssp(chi,1.0,chi,coeff2,psi,Ls-1,s+1);
}
///////////////////////////////////////////////////////////////////////////////////////
//Dinv (note D isn't really diagonal -- just diagonal enough that we can still invert)
// Compute Seeinv (coeff of gamma5)
///////////////////////////////////////////////////////////////////////////////////////
RealD R=(1+mass)/(1-mass);
RealD Seeinv = R + p[nblock]*dw_diag/amax;
for(int b=0;b<nblock;b++){
Seeinv += p[nblock-1-b]*dw_diag/amax / ( dw_diag*dw_diag/amax/amax + q[nblock-1-b]);
}
Seeinv = 1.0/Seeinv;
for(int b=0;b<nblock;b++){
int s = 2*b;
RealD pp = p[nblock-1-b];
RealD qq = q[nblock-1-b];
RealD coeff1=dw_diag / ( dw_diag*dw_diag + amax*amax* qq); // Implicit g5 here
RealD coeff2=amax*sqrt(qq) / ( dw_diag*dw_diag + amax*amax* qq);
ag5xpby_ssp (tmp,-coeff1,chi,coeff2,chi,s,s+1);
ag5xpby_ssp (tmp, coeff1,chi,coeff2,chi,s+1,s);
}
ag5xpby_ssp (tmp, Seeinv,chi,0.0,chi,Ls-1,Ls-1);
///////////////////////////////////////////////////////////////////////////////////////
// Uinv
///////////////////////////////////////////////////////////////////////////////////////
for(int b=0;b<nblock;b++){
int s = 2*b;
RealD pp = p[nblock-1-b];
RealD qq = q[nblock-1-b];
RealD coeff1=-sign*sqrt(amax*amax*amax*pp*qq) / ( dw_diag*dw_diag + amax*amax* qq);
RealD coeff2=-sign*sqrt(amax*pp)*dw_diag / ( dw_diag*dw_diag + amax*amax* qq); // Implicit g5 here
axpby_ssp (chi,1.0/scale,tmp,coeff1/scale,tmp,s,Ls-1);
axpbg5y_ssp(chi,1.0/scale,tmp,coeff2/scale,tmp,s+1,Ls-1);
}
axpby_ssp (chi, 1.0/scale,tmp,0.0,tmp,Ls-1,Ls-1);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::M_internal(const FermionField &psi, FermionField &chi,int dag)
{
FermionField D(psi.Grid());
int Ls = this->Ls;
int sign = dag ? (-1) : 1;
// For partial frac Hw case (b5=c5=1) chroma quirkily computes
//
// Conventions for partfrac appear to be a mess.
// Tony's Nara lectures have
//
// BlockDiag( H/p_i 1 | 1 )
// ( 1 p_i H / q_i^2 | 0 )
// ---------------------------------
// ( -1 0 | R +p0 H )
//
//Chroma ( -2H 2sqrt(q_i) | 0 )
// (2 sqrt(q_i) 2H | 2 sqrt(p_i) )
// ---------------------------------
// ( 0 -2 sqrt(p_i) | 2 R gamma_5 + p0 2H
//
// Edwards/Joo/Kennedy/Wenger
//
// Here, the "beta's" selected by chroma to scale the unphysical bulk constraint fields
// incorporate the approx scale factor. This is obtained by propagating the
// scale on "H" out to the off diagonal elements as follows:
//
// BlockDiag( H/p_i 1 | 1 )
// ( 1 p_i H / q_i^2 | 0 )
// ---------------------------------
// ( -1 0 | R + p_0 H )
//
// becomes:
// BlockDiag( H/ sp_i 1 | 1 )
// ( 1 sp_i H / s^2q_i^2 | 0 )
// ---------------------------------
// ( -1 0 | R + p_0/s H )
//
//
// This is implemented in Chroma by
// p0' = p0/approxMax
// p_i' = p_i*approxMax
// q_i' = q_i*approxMax*approxMax
//
// After the equivalence transform is applied the matrix becomes
//
//Chroma ( -2H sqrt(q'_i) | 0 )
// (sqrt(q'_i) 2H | sqrt(p'_i) )
// ---------------------------------
// ( 0 -sqrt(p'_i) | 2 R gamma_5 + p'0 2H
//
// = ( -2H sqrt(q_i)amax | 0 )
// (sqrt(q_i)amax 2H | sqrt(p_i*amax) )
// ---------------------------------
// ( 0 -sqrt(p_i)*amax | 2 R gamma_5 + p0/amax 2H
//
this->DW(psi,D,DaggerNo);
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
double pp = p[nblock-1-b];
double qq = q[nblock-1-b];
// Do each 2x2 block aligned at s and
ag5xpby_ssp(chi,-1.0*scale,D,amax*sqrt(qq)*scale,psi, s ,s+1); // Multiplies Dw by G5 so Hw
ag5xpby_ssp(chi, 1.0*scale,D,amax*sqrt(qq)*scale,psi, s+1,s);
// Pick up last column
axpby_ssp (chi, 1.0, chi,sqrt(amax*pp)*scale*sign,psi,s+1,Ls-1);
}
{
double R=(1+this->mass)/(1-this->mass);
//R g5 psi[Ls] + p[0] H
ag5xpbg5y_ssp(chi,R*scale,psi,p[nblock]*scale/amax,D,Ls-1,Ls-1);
for(int b=0;b<nblock;b++){
int s = 2*b+1;
double pp = p[nblock-1-b];
axpby_ssp(chi,1.0,chi,-sqrt(amax*pp)*scale*sign,psi,Ls-1,s);
}
}
}
template<class Impl>
RealD PartialFractionFermion5D<Impl>::M (const FermionField &in, FermionField &out)
{
M_internal(in,out,DaggerNo);
return norm2(out);
}
template<class Impl>
RealD PartialFractionFermion5D<Impl>::Mdag (const FermionField &in, FermionField &out)
{
M_internal(in,out,DaggerYes);
return norm2(out);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::Meooe (const FermionField &in, FermionField &out)
{
Meooe_internal(in,out,DaggerNo);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::MeooeDag (const FermionField &in, FermionField &out)
{
Meooe_internal(in,out,DaggerYes);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::Mooee (const FermionField &in, FermionField &out)
{
Mooee_internal(in,out,DaggerNo);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::MooeeDag (const FermionField &in, FermionField &out)
{
Mooee_internal(in,out,DaggerYes);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::MooeeInv (const FermionField &in, FermionField &out)
{
MooeeInv_internal(in,out,DaggerNo);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::MooeeInvDag (const FermionField &in, FermionField &out)
{
MooeeInv_internal(in,out,DaggerYes);
}
// force terms; five routines; default to Dhop on diagonal
template<class Impl>
void PartialFractionFermion5D<Impl>::MDeriv (GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
int Ls = this->Ls;
FermionField D(V.Grid());
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
ag5xpby_ssp(D,-scale,U,0.0,U,s,s);
ag5xpby_ssp(D, scale,U,0.0,U,s+1,s+1);
}
ag5xpby_ssp(D,p[nblock]*scale/amax,U,0.0,U,Ls-1,Ls-1);
this->DhopDeriv(mat,D,V,DaggerNo);
};
template<class Impl>
void PartialFractionFermion5D<Impl>::MoeDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
int Ls = this->Ls;
FermionField D(V.Grid());
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
ag5xpby_ssp(D,-scale,U,0.0,U,s,s);
ag5xpby_ssp(D, scale,U,0.0,U,s+1,s+1);
}
ag5xpby_ssp(D,p[nblock]*scale/amax,U,0.0,U,Ls-1,Ls-1);
this->DhopDerivOE(mat,D,V,DaggerNo);
};
template<class Impl>
void PartialFractionFermion5D<Impl>::MeoDeriv(GaugeField &mat,const FermionField &U,const FermionField &V,int dag)
{
int Ls = this->Ls;
FermionField D(V.Grid());
int nblock=(Ls-1)/2;
for(int b=0;b<nblock;b++){
int s = 2*b;
ag5xpby_ssp(D,-scale,U,0.0,U,s,s);
ag5xpby_ssp(D, scale,U,0.0,U,s+1,s+1);
}
ag5xpby_ssp(D,p[nblock]*scale/amax,U,0.0,U,Ls-1,Ls-1);
this->DhopDerivEO(mat,D,V,DaggerNo);
};
template<class Impl>
void PartialFractionFermion5D<Impl>::SetCoefficientsTanh(Approx::zolotarev_data *zdata,RealD scale){
SetCoefficientsZolotarev(1.0/scale,zdata);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::SetCoefficientsZolotarev(RealD zolo_hi,Approx::zolotarev_data *zdata){
// check on degree matching
// std::cout<<GridLogMessage << Ls << " Ls"<<std::endl;
// std::cout<<GridLogMessage << zdata->n << " - n"<<std::endl;
// std::cout<<GridLogMessage << zdata->da << " -da "<<std::endl;
// std::cout<<GridLogMessage << zdata->db << " -db"<<std::endl;
// std::cout<<GridLogMessage << zdata->dn << " -dn"<<std::endl;
// std::cout<<GridLogMessage << zdata->dd << " -dd"<<std::endl;
int Ls = this->Ls;
assert(Ls == (2*zdata->da -1) );
// Part frac
// RealD R;
R=(1+mass)/(1-mass);
dw_diag = (4.0-this->M5);
// std::vector<RealD> p;
// std::vector<RealD> q;
p.resize(zdata->da);
q.resize(zdata->dd);
for(int n=0;n<zdata->da;n++){
p[n] = zdata -> alpha[n];
}
for(int n=0;n<zdata->dd;n++){
q[n] = -zdata -> ap[n];
}
scale= part_frac_chroma_convention ? 2.0 : 1.0; // Chroma conventions annoy me
amax=zolo_hi;
}
template<class Impl>
void PartialFractionFermion5D<Impl>::ExportPhysicalFermionSolution(const FermionField &solution5d,FermionField &exported4d)
{
int Ls = this->Ls;
conformable(solution5d.Grid(),this->FermionGrid());
conformable(exported4d.Grid(),this->GaugeGrid());
ExtractSlice(exported4d, solution5d, Ls-1, Ls-1);
}
template<class Impl>
void PartialFractionFermion5D<Impl>::ImportPhysicalFermionSource(const FermionField &input4d,FermionField &imported5d)
{
int Ls = this->Ls;
conformable(imported5d.Grid(),this->FermionGrid());
conformable(input4d.Grid() ,this->GaugeGrid());
FermionField tmp(this->FermionGrid());
tmp=Zero();
InsertSlice(input4d, tmp, Ls-1, Ls-1);
tmp=Gamma(Gamma::Algebra::Gamma5)*tmp;
this->Dminus(tmp,imported5d);
}
// Constructors
template<class Impl>
PartialFractionFermion5D<Impl>::PartialFractionFermion5D(GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mass,RealD M5,
const ImplParams &p) :
WilsonFermion5D<Impl>(_Umu,
FiveDimGrid, FiveDimRedBlackGrid,
FourDimGrid, FourDimRedBlackGrid,M5,p),
mass(_mass)
{
int Ls = this->Ls;
assert((Ls&0x1)==1); // Odd Ls required
int nrational=Ls-1;
Approx::zolotarev_data *zdata = Approx::higham(1.0,nrational);
// NB: chroma uses a cast to "float" for the zolotarev range(!?).
// this creates a real difference in the operator which I do not like but we can replicate here
// to demonstrate compatibility
// RealD eps = (zolo_lo / zolo_hi);
// zdata = bfm_zolotarev(eps,nrational,0);
SetCoefficientsTanh(zdata,1.0);
Approx::zolotarev_free(zdata);
}
FermOpTemplateInstantiate(PartialFractionFermion5D);
NAMESPACE_END(Grid);

294
Grid/qcd/action/fermion/StaggeredKernels.cc
View File

@ -1,294 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015
Author: Azusa Yamaguchi, Peter Boyle
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
NAMESPACE_BEGIN(Grid);
int StaggeredKernelsStatic::Opt= StaggeredKernelsStatic::OptGeneric;
int StaggeredKernelsStatic::Comms = StaggeredKernelsStatic::CommsAndCompute;
#define GENERIC_STENCIL_LEG(U,Dir,skew,multLink) \
SE = st.GetEntry(ptype, Dir+skew, sF); \
if (SE->_is_local ) { \
if (SE->_permute) { \
chi_p = &chi; \
permute(chi, in[SE->_offset], ptype); \
} else { \
chi_p = &in[SE->_offset]; \
} \
} else { \
chi_p = &buf[SE->_offset]; \
} \
multLink(Uchi, U[sU], *chi_p, Dir);
#define GENERIC_STENCIL_LEG_INT(U,Dir,skew,multLink) \
SE = st.GetEntry(ptype, Dir+skew, sF); \
if (SE->_is_local ) { \
if (SE->_permute) { \
chi_p = &chi; \
permute(chi, in[SE->_offset], ptype); \
} else { \
chi_p = &in[SE->_offset]; \
} \
} else if ( st.same_node[Dir] ) { \
chi_p = &buf[SE->_offset]; \
} \
if (SE->_is_local || st.same_node[Dir] ) { \
multLink(Uchi, U[sU], *chi_p, Dir); \
}
#define GENERIC_STENCIL_LEG_EXT(U,Dir,skew,multLink) \
SE = st.GetEntry(ptype, Dir+skew, sF); \
if ((!SE->_is_local) && (!st.same_node[Dir]) ) { \
nmu++; \
chi_p = &buf[SE->_offset]; \
multLink(Uchi, U[sU], *chi_p, Dir); \
}
template <class Impl>
StaggeredKernels<Impl>::StaggeredKernels(const ImplParams &p) : Base(p){};
////////////////////////////////////////////////////////////////////////////////////
// Generic implementation; move to different file?
// Int, Ext, Int+Ext cases for comms overlap
////////////////////////////////////////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteGeneric(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out, int dag) {
const SiteSpinor *chi_p;
SiteSpinor chi;
SiteSpinor Uchi;
StencilEntry *SE;
int ptype;
int skew;
for(int s=0;s<LLs;s++){
int sF=LLs*sU+s;
skew = 0;
GENERIC_STENCIL_LEG(U,Xp,skew,Impl::multLink);
GENERIC_STENCIL_LEG(U,Yp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Zp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Tp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Xm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(U,Tm,skew,Impl::multLinkAdd);
skew=8;
GENERIC_STENCIL_LEG(UUU,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Yp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Zp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Tp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Xm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG(UUU,Tm,skew,Impl::multLinkAdd);
if ( dag ) {
Uchi = - Uchi;
}
vstream(out[sF], Uchi);
}
};
///////////////////////////////////////////////////
// Only contributions from interior of our node
///////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteGenericInt(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag) {
const SiteSpinor *chi_p;
SiteSpinor chi;
SiteSpinor Uchi;
StencilEntry *SE;
int ptype;
int skew ;
for(int s=0;s<LLs;s++){
int sF=LLs*sU+s;
skew = 0;
Uchi=Zero();
GENERIC_STENCIL_LEG_INT(U,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Yp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Zp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Tp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Xm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(U,Tm,skew,Impl::multLinkAdd);
skew=8;
GENERIC_STENCIL_LEG_INT(UUU,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Yp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Zp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Tp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Xm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_INT(UUU,Tm,skew,Impl::multLinkAdd);
if ( dag ) {
Uchi = - Uchi;
}
vstream(out[sF], Uchi);
}
};
///////////////////////////////////////////////////
// Only contributions from exterior of our node
///////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteGenericExt(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag) {
const SiteSpinor *chi_p;
// SiteSpinor chi;
SiteSpinor Uchi;
StencilEntry *SE;
int ptype;
int nmu=0;
int skew ;
for(int s=0;s<LLs;s++){
int sF=LLs*sU+s;
skew = 0;
Uchi=Zero();
GENERIC_STENCIL_LEG_EXT(U,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Yp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Zp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Tp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Xm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(U,Tm,skew,Impl::multLinkAdd);
skew=8;
GENERIC_STENCIL_LEG_EXT(UUU,Xp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Yp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Zp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Tp,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Xm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Ym,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Zm,skew,Impl::multLinkAdd);
GENERIC_STENCIL_LEG_EXT(UUU,Tm,skew,Impl::multLinkAdd);
if ( nmu ) {
if ( dag ) {
out[sF] = out[sF] - Uchi;
} else {
out[sF] = out[sF] + Uchi;
}
}
}
};
////////////////////////////////////////////////////////////////////////////////////
// Driving / wrapping routine to select right kernel
////////////////////////////////////////////////////////////////////////////////////
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,
int interior,int exterior)
{
int dag=1;
DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
};
template <class Impl>
void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,
int interior,int exterior)
{
int dag=0;
DhopSite(st,lo,U,UUU,buf,LLs,sU,in,out,dag,interior,exterior);
};
template <class Impl>
void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionFieldView &in, FermionFieldView &out,
int dag,int interior,int exterior)
{
switch(Opt) {
#ifdef AVX512
case OptInlineAsm:
if ( interior && exterior ) {
DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else {
std::cout << GridLogError << "Cannot overlap comms and compute with Staggered assembly"<<std::endl;
assert(0);
}
break;
#endif
case OptHandUnroll:
if ( interior && exterior ) {
DhopSiteHand (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( interior ) {
DhopSiteHandInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( exterior ) {
DhopSiteHandExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
}
break;
case OptGeneric:
if ( interior && exterior ) {
DhopSiteGeneric (st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( interior ) {
DhopSiteGenericInt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
} else if ( exterior ) {
DhopSiteGenericExt(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
}
break;
default:
std::cout<<"Oops Opt = "<<Opt<<std::endl;
assert(0);
break;
}
};
template <class Impl>
void StaggeredKernels<Impl>::DhopDirKernel( StencilImpl &st, DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU, SiteSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int disp)
{
// Disp should be either +1,-1,+3,-3
// What about "dag" ?
// Because we work out pU . dS/dU
// U
assert(0);
}
FermOpStaggeredTemplateInstantiate(StaggeredKernels);
FermOpStaggeredVec5dTemplateInstantiate(StaggeredKernels);
NAMESPACE_END(Grid);

972
Grid/qcd/action/fermion/StaggeredKernelsAsm.cc
View File

@ -1,972 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/StaggerdKernelsHand.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid.h>
#ifdef AVX512
#include <simd/Intel512common.h>
#include <simd/Intel512avx.h>
#endif
// Interleave operations from two directions
// This looks just like a 2 spin multiply and reuse same sequence from the Wilson
// Kernel. But the spin index becomes a mu index instead.
#define Chi_00 %zmm0
#define Chi_01 %zmm1
#define Chi_02 %zmm2
#define Chi_10 %zmm3
#define Chi_11 %zmm4
#define Chi_12 %zmm5
#define Chi_20 %zmm6
#define Chi_21 %zmm7
#define Chi_22 %zmm8
#define Chi_30 %zmm9
#define Chi_31 %zmm10
#define Chi_32 %zmm11
#define UChi_00 %zmm12
#define UChi_01 %zmm13
#define UChi_02 %zmm14
#define UChi_10 %zmm15
#define UChi_11 %zmm16
#define UChi_12 %zmm17
#define UChi_20 %zmm18
#define UChi_21 %zmm19
#define UChi_22 %zmm20
#define UChi_30 %zmm21
#define UChi_31 %zmm22
#define UChi_32 %zmm23
#define pChi_00 %%zmm0
#define pChi_01 %%zmm1
#define pChi_02 %%zmm2
#define pChi_10 %%zmm3
#define pChi_11 %%zmm4
#define pChi_12 %%zmm5
#define pChi_20 %%zmm6
#define pChi_21 %%zmm7
#define pChi_22 %%zmm8
#define pChi_30 %%zmm9
#define pChi_31 %%zmm10
#define pChi_32 %%zmm11
#define pUChi_00 %%zmm12
#define pUChi_01 %%zmm13
#define pUChi_02 %%zmm14
#define pUChi_10 %%zmm15
#define pUChi_11 %%zmm16
#define pUChi_12 %%zmm17
#define pUChi_20 %%zmm18
#define pUChi_21 %%zmm19
#define pUChi_22 %%zmm20
#define pUChi_30 %%zmm21
#define pUChi_31 %%zmm22
#define pUChi_32 %%zmm23
#define T0 %zmm24
#define T1 %zmm25
#define T2 %zmm26
#define T3 %zmm27
#define Z00 %zmm26
#define Z10 %zmm27
#define Z0 Z00
#define Z1 %zmm28
#define Z2 %zmm29
#define Z3 %zmm30
#define Z4 %zmm31
#define Z5 Chi_31
#define Z6 Chi_32
#define MULT_ADD_LS(g0,g1,g2,g3) \
asm ( "movq %0, %%r8 \n\t" \
"movq %1, %%r9 \n\t" \
"movq %2, %%r10 \n\t" \
"movq %3, %%r11 \n\t" : : "r"(g0), "r"(g1), "r"(g2), "r"(g3) : "%r8","%r9","%r10","%r11" );\
asm ( \
VSHUF(Chi_00,T0) VSHUF(Chi_10,T1) \
VSHUF(Chi_20,T2) VSHUF(Chi_30,T3) \
VMADDSUBIDUP(0,%r8,T0,UChi_00) VMADDSUBIDUP(0,%r9,T1,UChi_10) \
VMADDSUBIDUP(3,%r8,T0,UChi_01) VMADDSUBIDUP(3,%r9,T1,UChi_11) \
VMADDSUBIDUP(6,%r8,T0,UChi_02) VMADDSUBIDUP(6,%r9,T1,UChi_12) \
VMADDSUBIDUP(0,%r10,T2,UChi_20) VMADDSUBIDUP(0,%r11,T3,UChi_30) \
VMADDSUBIDUP(3,%r10,T2,UChi_21) VMADDSUBIDUP(3,%r11,T3,UChi_31) \
VMADDSUBIDUP(6,%r10,T2,UChi_22) VMADDSUBIDUP(6,%r11,T3,UChi_32) \
VMADDSUBRDUP(0,%r8,Chi_00,UChi_00) VMADDSUBRDUP(0,%r9,Chi_10,UChi_10) \
VMADDSUBRDUP(3,%r8,Chi_00,UChi_01) VMADDSUBRDUP(3,%r9,Chi_10,UChi_11) \
VMADDSUBRDUP(6,%r8,Chi_00,UChi_02) VMADDSUBRDUP(6,%r9,Chi_10,UChi_12) \
VMADDSUBRDUP(0,%r10,Chi_20,UChi_20) VMADDSUBRDUP(0,%r11,Chi_30,UChi_30) \
VMADDSUBRDUP(3,%r10,Chi_20,UChi_21) VMADDSUBRDUP(3,%r11,Chi_30,UChi_31) \
VMADDSUBRDUP(6,%r10,Chi_20,UChi_22) VMADDSUBRDUP(6,%r11,Chi_30,UChi_32) \
VSHUF(Chi_01,T0) VSHUF(Chi_11,T1) \
VSHUF(Chi_21,T2) VSHUF(Chi_31,T3) \
VMADDSUBIDUP(1,%r8,T0,UChi_00) VMADDSUBIDUP(1,%r9,T1,UChi_10) \
VMADDSUBIDUP(4,%r8,T0,UChi_01) VMADDSUBIDUP(4,%r9,T1,UChi_11) \
VMADDSUBIDUP(7,%r8,T0,UChi_02) VMADDSUBIDUP(7,%r9,T1,UChi_12) \
VMADDSUBIDUP(1,%r10,T2,UChi_20) VMADDSUBIDUP(1,%r11,T3,UChi_30) \
VMADDSUBIDUP(4,%r10,T2,UChi_21) VMADDSUBIDUP(4,%r11,T3,UChi_31) \
VMADDSUBIDUP(7,%r10,T2,UChi_22) VMADDSUBIDUP(7,%r11,T3,UChi_32) \
VMADDSUBRDUP(1,%r8,Chi_01,UChi_00) VMADDSUBRDUP(1,%r9,Chi_11,UChi_10) \
VMADDSUBRDUP(4,%r8,Chi_01,UChi_01) VMADDSUBRDUP(4,%r9,Chi_11,UChi_11) \
VMADDSUBRDUP(7,%r8,Chi_01,UChi_02) VMADDSUBRDUP(7,%r9,Chi_11,UChi_12) \
VMADDSUBRDUP(1,%r10,Chi_21,UChi_20) VMADDSUBRDUP(1,%r11,Chi_31,UChi_30) \
VMADDSUBRDUP(4,%r10,Chi_21,UChi_21) VMADDSUBRDUP(4,%r11,Chi_31,UChi_31) \
VMADDSUBRDUP(7,%r10,Chi_21,UChi_22) VMADDSUBRDUP(7,%r11,Chi_31,UChi_32) \
VSHUF(Chi_02,T0) VSHUF(Chi_12,T1) \
VSHUF(Chi_22,T2) VSHUF(Chi_32,T3) \
VMADDSUBIDUP(2,%r8,T0,UChi_00) VMADDSUBIDUP(2,%r9,T1,UChi_10) \
VMADDSUBIDUP(5,%r8,T0,UChi_01) VMADDSUBIDUP(5,%r9,T1,UChi_11) \
VMADDSUBIDUP(8,%r8,T0,UChi_02) VMADDSUBIDUP(8,%r9,T1,UChi_12) \
VMADDSUBIDUP(2,%r10,T2,UChi_20) VMADDSUBIDUP(2,%r11,T3,UChi_30) \
VMADDSUBIDUP(5,%r10,T2,UChi_21) VMADDSUBIDUP(5,%r11,T3,UChi_31) \
VMADDSUBIDUP(8,%r10,T2,UChi_22) VMADDSUBIDUP(8,%r11,T3,UChi_32) \
VMADDSUBRDUP(2,%r8,Chi_02,UChi_00) VMADDSUBRDUP(2,%r9,Chi_12,UChi_10) \
VMADDSUBRDUP(5,%r8,Chi_02,UChi_01) VMADDSUBRDUP(5,%r9,Chi_12,UChi_11) \
VMADDSUBRDUP(8,%r8,Chi_02,UChi_02) VMADDSUBRDUP(8,%r9,Chi_12,UChi_12) \
VMADDSUBRDUP(2,%r10,Chi_22,UChi_20) VMADDSUBRDUP(2,%r11,Chi_32,UChi_30) \
VMADDSUBRDUP(5,%r10,Chi_22,UChi_21) VMADDSUBRDUP(5,%r11,Chi_32,UChi_31) \
VMADDSUBRDUP(8,%r10,Chi_22,UChi_22) VMADDSUBRDUP(8,%r11,Chi_32,UChi_32) );
#define MULT_LS(g0,g1,g2,g3) \
asm ( "movq %0, %%r8 \n\t" \
"movq %1, %%r9 \n\t" \
"movq %2, %%r10 \n\t" \
"movq %3, %%r11 \n\t" : : "r"(g0), "r"(g1), "r"(g2), "r"(g3) : "%r8","%r9","%r10","%r11" );\
asm ( \
VSHUF(Chi_00,T0) VSHUF(Chi_10,T1) \
VSHUF(Chi_20,T2) VSHUF(Chi_30,T3) \
VMULIDUP(0,%r8,T0,UChi_00) VMULIDUP(0,%r9,T1,UChi_10) \
VMULIDUP(3,%r8,T0,UChi_01) VMULIDUP(3,%r9,T1,UChi_11) \
VMULIDUP(6,%r8,T0,UChi_02) VMULIDUP(6,%r9,T1,UChi_12) \
VMULIDUP(0,%r10,T2,UChi_20) VMULIDUP(0,%r11,T3,UChi_30) \
VMULIDUP(3,%r10,T2,UChi_21) VMULIDUP(3,%r11,T3,UChi_31) \
VMULIDUP(6,%r10,T2,UChi_22) VMULIDUP(6,%r11,T3,UChi_32) \
VMADDSUBRDUP(0,%r8,Chi_00,UChi_00) VMADDSUBRDUP(0,%r9,Chi_10,UChi_10) \
VMADDSUBRDUP(3,%r8,Chi_00,UChi_01) VMADDSUBRDUP(3,%r9,Chi_10,UChi_11) \
VMADDSUBRDUP(6,%r8,Chi_00,UChi_02) VMADDSUBRDUP(6,%r9,Chi_10,UChi_12) \
VMADDSUBRDUP(0,%r10,Chi_20,UChi_20) VMADDSUBRDUP(0,%r11,Chi_30,UChi_30) \
VMADDSUBRDUP(3,%r10,Chi_20,UChi_21) VMADDSUBRDUP(3,%r11,Chi_30,UChi_31) \
VMADDSUBRDUP(6,%r10,Chi_20,UChi_22) VMADDSUBRDUP(6,%r11,Chi_30,UChi_32) \
VSHUF(Chi_01,T0) VSHUF(Chi_11,T1) \
VSHUF(Chi_21,T2) VSHUF(Chi_31,T3) \
VMADDSUBIDUP(1,%r8,T0,UChi_00) VMADDSUBIDUP(1,%r9,T1,UChi_10) \
VMADDSUBIDUP(4,%r8,T0,UChi_01) VMADDSUBIDUP(4,%r9,T1,UChi_11) \
VMADDSUBIDUP(7,%r8,T0,UChi_02) VMADDSUBIDUP(7,%r9,T1,UChi_12) \
VMADDSUBIDUP(1,%r10,T2,UChi_20) VMADDSUBIDUP(1,%r11,T3,UChi_30) \
VMADDSUBIDUP(4,%r10,T2,UChi_21) VMADDSUBIDUP(4,%r11,T3,UChi_31) \
VMADDSUBIDUP(7,%r10,T2,UChi_22) VMADDSUBIDUP(7,%r11,T3,UChi_32) \
VMADDSUBRDUP(1,%r8,Chi_01,UChi_00) VMADDSUBRDUP(1,%r9,Chi_11,UChi_10) \
VMADDSUBRDUP(4,%r8,Chi_01,UChi_01) VMADDSUBRDUP(4,%r9,Chi_11,UChi_11) \
VMADDSUBRDUP(7,%r8,Chi_01,UChi_02) VMADDSUBRDUP(7,%r9,Chi_11,UChi_12) \
VMADDSUBRDUP(1,%r10,Chi_21,UChi_20) VMADDSUBRDUP(1,%r11,Chi_31,UChi_30) \
VMADDSUBRDUP(4,%r10,Chi_21,UChi_21) VMADDSUBRDUP(4,%r11,Chi_31,UChi_31) \
VMADDSUBRDUP(7,%r10,Chi_21,UChi_22) VMADDSUBRDUP(7,%r11,Chi_31,UChi_32) \
VSHUF(Chi_02,T0) VSHUF(Chi_12,T1) \
VSHUF(Chi_22,T2) VSHUF(Chi_32,T3) \
VMADDSUBIDUP(2,%r8,T0,UChi_00) VMADDSUBIDUP(2,%r9,T1,UChi_10) \
VMADDSUBIDUP(5,%r8,T0,UChi_01) VMADDSUBIDUP(5,%r9,T1,UChi_11) \
VMADDSUBIDUP(8,%r8,T0,UChi_02) VMADDSUBIDUP(8,%r9,T1,UChi_12) \
VMADDSUBIDUP(2,%r10,T2,UChi_20) VMADDSUBIDUP(2,%r11,T3,UChi_30) \
VMADDSUBIDUP(5,%r10,T2,UChi_21) VMADDSUBIDUP(5,%r11,T3,UChi_31) \
VMADDSUBIDUP(8,%r10,T2,UChi_22) VMADDSUBIDUP(8,%r11,T3,UChi_32) \
VMADDSUBRDUP(2,%r8,Chi_02,UChi_00) VMADDSUBRDUP(2,%r9,Chi_12,UChi_10) \
VMADDSUBRDUP(5,%r8,Chi_02,UChi_01) VMADDSUBRDUP(5,%r9,Chi_12,UChi_11) \
VMADDSUBRDUP(8,%r8,Chi_02,UChi_02) VMADDSUBRDUP(8,%r9,Chi_12,UChi_12) \
VMADDSUBRDUP(2,%r10,Chi_22,UChi_20) VMADDSUBRDUP(2,%r11,Chi_32,UChi_30) \
VMADDSUBRDUP(5,%r10,Chi_22,UChi_21) VMADDSUBRDUP(5,%r11,Chi_32,UChi_31) \
VMADDSUBRDUP(8,%r10,Chi_22,UChi_22) VMADDSUBRDUP(8,%r11,Chi_32,UChi_32) );
#define MULT_ADD_XYZTa(g0,g1) \
asm ( "movq %0, %%r8 \n\t" \
"movq %1, %%r9 \n\t" : : "r"(g0), "r"(g1) : "%r8","%r9");\
__asm__ ( \
VSHUF(Chi_00,T0) \
VSHUF(Chi_10,T1) \
VMOVIDUP(0,%r8,Z0 ) \
VMOVIDUP(3,%r8,Z1 ) \
VMOVIDUP(6,%r8,Z2 ) \
VMADDSUB(Z0,T0,UChi_00) \
VMADDSUB(Z1,T0,UChi_01) \
VMADDSUB(Z2,T0,UChi_02) \
\
VMOVIDUP(0,%r9,Z0 ) \
VMOVIDUP(3,%r9,Z1 ) \
VMOVIDUP(6,%r9,Z2 ) \
VMADDSUB(Z0,T1,UChi_10) \
VMADDSUB(Z1,T1,UChi_11) \
VMADDSUB(Z2,T1,UChi_12) \
\
\
VMOVRDUP(0,%r8,Z3 ) \
VMOVRDUP(3,%r8,Z4 ) \
VMOVRDUP(6,%r8,Z5 ) \
VMADDSUB(Z3,Chi_00,UChi_00)/*rr * ir = ri rr*/ \
VMADDSUB(Z4,Chi_00,UChi_01) \
VMADDSUB(Z5,Chi_00,UChi_02) \
\
VMOVRDUP(0,%r9,Z3 ) \
VMOVRDUP(3,%r9,Z4 ) \
VMOVRDUP(6,%r9,Z5 ) \
VMADDSUB(Z3,Chi_10,UChi_10) \
VMADDSUB(Z4,Chi_10,UChi_11)\
VMADDSUB(Z5,Chi_10,UChi_12) \
\
\
VMOVIDUP(1,%r8,Z0 ) \
VMOVIDUP(4,%r8,Z1 ) \
VMOVIDUP(7,%r8,Z2 ) \
VSHUF(Chi_01,T0) \
VMADDSUB(Z0,T0,UChi_00) \
VMADDSUB(Z1,T0,UChi_01) \
VMADDSUB(Z2,T0,UChi_02) \
\
VMOVIDUP(1,%r9,Z0 ) \
VMOVIDUP(4,%r9,Z1 ) \
VMOVIDUP(7,%r9,Z2 ) \
VSHUF(Chi_11,T1) \
VMADDSUB(Z0,T1,UChi_10) \
VMADDSUB(Z1,T1,UChi_11) \
VMADDSUB(Z2,T1,UChi_12) \
\
VMOVRDUP(1,%r8,Z3 ) \
VMOVRDUP(4,%r8,Z4 ) \
VMOVRDUP(7,%r8,Z5 ) \
VMADDSUB(Z3,Chi_01,UChi_00) \
VMADDSUB(Z4,Chi_01,UChi_01) \
VMADDSUB(Z5,Chi_01,UChi_02) \
\
VMOVRDUP(1,%r9,Z3 ) \
VMOVRDUP(4,%r9,Z4 ) \
VMOVRDUP(7,%r9,Z5 ) \
VMADDSUB(Z3,Chi_11,UChi_10) \
VMADDSUB(Z4,Chi_11,UChi_11) \
VMADDSUB(Z5,Chi_11,UChi_12) \
\
VSHUF(Chi_02,T0) \
VSHUF(Chi_12,T1) \
VMOVIDUP(2,%r8,Z0 ) \
VMOVIDUP(5,%r8,Z1 ) \
VMOVIDUP(8,%r8,Z2 ) \
VMADDSUB(Z0,T0,UChi_00) \
VMADDSUB(Z1,T0,UChi_01) \
VMADDSUB(Z2,T0,UChi_02) \
VMOVIDUP(2,%r9,Z0 ) \
VMOVIDUP(5,%r9,Z1 ) \
VMOVIDUP(8,%r9,Z2 ) \
VMADDSUB(Z0,T1,UChi_10) \
VMADDSUB(Z1,T1,UChi_11) \
VMADDSUB(Z2,T1,UChi_12) \
/*55*/ \
VMOVRDUP(2,%r8,Z3 ) \
VMOVRDUP(5,%r8,Z4 ) \
VMOVRDUP(8,%r8,Z5 ) \
VMADDSUB(Z3,Chi_02,UChi_00) \
VMADDSUB(Z4,Chi_02,UChi_01) \
VMADDSUB(Z5,Chi_02,UChi_02) \
VMOVRDUP(2,%r9,Z3 ) \
VMOVRDUP(5,%r9,Z4 ) \
VMOVRDUP(8,%r9,Z5 ) \
VMADDSUB(Z3,Chi_12,UChi_10) \
VMADDSUB(Z4,Chi_12,UChi_11) \
VMADDSUB(Z5,Chi_12,UChi_12) \
/*61 insns*/ );
#define MULT_ADD_XYZT(g0,g1) \
asm ( "movq %0, %%r8 \n\t" \
"movq %1, %%r9 \n\t" : : "r"(g0), "r"(g1) : "%r8","%r9");\
__asm__ ( \
VSHUFMEM(0,%r8,Z00) VSHUFMEM(0,%r9,Z10) \
VRDUP(Chi_00,T0) VIDUP(Chi_00,Chi_00) \
VRDUP(Chi_10,T1) VIDUP(Chi_10,Chi_10) \
VMUL(Z00,Chi_00,Z1) VMUL(Z10,Chi_10,Z2) \
VSHUFMEM(3,%r8,Z00) VSHUFMEM(3,%r9,Z10) \
VMUL(Z00,Chi_00,Z3) VMUL(Z10,Chi_10,Z4) \
VSHUFMEM(6,%r8,Z00) VSHUFMEM(6,%r9,Z10) \
VMUL(Z00,Chi_00,Z5) VMUL(Z10,Chi_10,Z6) \
VMADDMEM(0,%r8,T0,UChi_00) VMADDMEM(0,%r9,T1,UChi_10) \
VMADDMEM(3,%r8,T0,UChi_01) VMADDMEM(3,%r9,T1,UChi_11) \
VMADDMEM(6,%r8,T0,UChi_02) VMADDMEM(6,%r9,T1,UChi_12) \
VSHUFMEM(1,%r8,Z00) VSHUFMEM(1,%r9,Z10) \
VRDUP(Chi_01,T0) VIDUP(Chi_01,Chi_01) \
VRDUP(Chi_11,T1) VIDUP(Chi_11,Chi_11) \
VMADD(Z00,Chi_01,Z1) VMADD(Z10,Chi_11,Z2) \
VSHUFMEM(4,%r8,Z00) VSHUFMEM(4,%r9,Z10) \
VMADD(Z00,Chi_01,Z3) VMADD(Z10,Chi_11,Z4) \
VSHUFMEM(7,%r8,Z00) VSHUFMEM(7,%r9,Z10) \
VMADD(Z00,Chi_01,Z5) VMADD(Z10,Chi_11,Z6) \
VMADDMEM(1,%r8,T0,UChi_00) VMADDMEM(1,%r9,T1,UChi_10) \
VMADDMEM(4,%r8,T0,UChi_01) VMADDMEM(4,%r9,T1,UChi_11) \
VMADDMEM(7,%r8,T0,UChi_02) VMADDMEM(7,%r9,T1,UChi_12) \
VSHUFMEM(2,%r8,Z00) VSHUFMEM(2,%r9,Z10) \
VRDUP(Chi_02,T0) VIDUP(Chi_02,Chi_02) \
VRDUP(Chi_12,T1) VIDUP(Chi_12,Chi_12) \
VMADD(Z00,Chi_02,Z1) VMADD(Z10,Chi_12,Z2) \
VSHUFMEM(5,%r8,Z00) VSHUFMEM(5,%r9,Z10) \
VMADD(Z00,Chi_02,Z3) VMADD(Z10,Chi_12,Z4) \
VSHUFMEM(8,%r8,Z00) VSHUFMEM(8,%r9,Z10) \
VMADD(Z00,Chi_02,Z5) VMADD(Z10,Chi_12,Z6) \
VMADDSUBMEM(2,%r8,T0,Z1) VMADDSUBMEM(2,%r9,T1,Z2) \
VMADDSUBMEM(5,%r8,T0,Z3) VMADDSUBMEM(5,%r9,T1,Z4) \
VMADDSUBMEM(8,%r8,T0,Z5) VMADDSUBMEM(8,%r9,T1,Z6) \
VADD(Z1,UChi_00,UChi_00) VADD(Z2,UChi_10,UChi_10) \
VADD(Z3,UChi_01,UChi_01) VADD(Z4,UChi_11,UChi_11) \
VADD(Z5,UChi_02,UChi_02) VADD(Z6,UChi_12,UChi_12) );
#define MULT_XYZT(g0,g1) \
asm ( "movq %0, %%r8 \n\t" \
"movq %1, %%r9 \n\t" : : "r"(g0), "r"(g1) : "%r8","%r9" ); \
__asm__ ( \
VSHUF(Chi_00,T0) \
VSHUF(Chi_10,T1) \
VMOVIDUP(0,%r8,Z0 ) \
VMOVIDUP(3,%r8,Z1 ) \
VMOVIDUP(6,%r8,Z2 ) \
/*6*/ \
VMUL(Z0,T0,UChi_00) \
VMUL(Z1,T0,UChi_01) \
VMUL(Z2,T0,UChi_02) \
VMOVIDUP(0,%r9,Z0 ) \
VMOVIDUP(3,%r9,Z1 ) \
VMOVIDUP(6,%r9,Z2 ) \
VMUL(Z0,T1,UChi_10) \
VMUL(Z1,T1,UChi_11) \
VMUL(Z2,T1,UChi_12) \
VMOVRDUP(0,%r8,Z3 ) \
VMOVRDUP(3,%r8,Z4 ) \
VMOVRDUP(6,%r8,Z5 ) \
/*18*/ \
VMADDSUB(Z3,Chi_00,UChi_00) \
VMADDSUB(Z4,Chi_00,UChi_01)\
VMADDSUB(Z5,Chi_00,UChi_02) \
VMOVRDUP(0,%r9,Z3 ) \
VMOVRDUP(3,%r9,Z4 ) \
VMOVRDUP(6,%r9,Z5 ) \
VMADDSUB(Z3,Chi_10,UChi_10) \
VMADDSUB(Z4,Chi_10,UChi_11)\
VMADDSUB(Z5,Chi_10,UChi_12) \
VMOVIDUP(1,%r8,Z0 ) \
VMOVIDUP(4,%r8,Z1 ) \
VMOVIDUP(7,%r8,Z2 ) \
/*28*/ \
VSHUF(Chi_01,T0) \
VMADDSUB(Z0,T0,UChi_00) \
VMADDSUB(Z1,T0,UChi_01) \
VMADDSUB(Z2,T0,UChi_02) \
VMOVIDUP(1,%r9,Z0 ) \
VMOVIDUP(4,%r9,Z1 ) \
VMOVIDUP(7,%r9,Z2 ) \
VSHUF(Chi_11,T1) \
VMADDSUB(Z0,T1,UChi_10) \
VMADDSUB(Z1,T1,UChi_11) \
VMADDSUB(Z2,T1,UChi_12) \
VMOVRDUP(1,%r8,Z3 ) \
VMOVRDUP(4,%r8,Z4 ) \
VMOVRDUP(7,%r8,Z5 ) \
/*38*/ \
VMADDSUB(Z3,Chi_01,UChi_00) \
VMADDSUB(Z4,Chi_01,UChi_01) \
VMADDSUB(Z5,Chi_01,UChi_02) \
VMOVRDUP(1,%r9,Z3 ) \
VMOVRDUP(4,%r9,Z4 ) \
VMOVRDUP(7,%r9,Z5 ) \
VMADDSUB(Z3,Chi_11,UChi_10) \
VMADDSUB(Z4,Chi_11,UChi_11) \
VMADDSUB(Z5,Chi_11,UChi_12) \
/*48*/ \
VSHUF(Chi_02,T0) \
VSHUF(Chi_12,T1) \
VMOVIDUP(2,%r8,Z0 ) \
VMOVIDUP(5,%r8,Z1 ) \
VMOVIDUP(8,%r8,Z2 ) \
VMADDSUB(Z0,T0,UChi_00) \
VMADDSUB(Z1,T0,UChi_01) \
VMADDSUB(Z2,T0,UChi_02) \
VMOVIDUP(2,%r9,Z0 ) \
VMOVIDUP(5,%r9,Z1 ) \
VMOVIDUP(8,%r9,Z2 ) \
VMADDSUB(Z0,T1,UChi_10) \
VMADDSUB(Z1,T1,UChi_11) \
VMADDSUB(Z2,T1,UChi_12) \
/*55*/ \
VMOVRDUP(2,%r8,Z3 ) \
VMOVRDUP(5,%r8,Z4 ) \
VMOVRDUP(8,%r8,Z5 ) \
VMADDSUB(Z3,Chi_02,UChi_00) \
VMADDSUB(Z4,Chi_02,UChi_01) \
VMADDSUB(Z5,Chi_02,UChi_02) \
VMOVRDUP(2,%r9,Z3 ) \
VMOVRDUP(5,%r9,Z4 ) \
VMOVRDUP(8,%r9,Z5 ) \
VMADDSUB(Z3,Chi_12,UChi_10) \
VMADDSUB(Z4,Chi_12,UChi_11) \
VMADDSUB(Z5,Chi_12,UChi_12) \
/*61 insns*/ );
#define MULT_XYZTa(g0,g1) \
asm ( "movq %0, %%r8 \n\t" \
"movq %1, %%r9 \n\t" : : "r"(g0), "r"(g1) : "%r8","%r9" ); \
__asm__ ( \
VSHUFMEM(0,%r8,Z00) VSHUFMEM(0,%r9,Z10) \
VRDUP(Chi_00,T0) VIDUP(Chi_00,Chi_00) \
VRDUP(Chi_10,T1) VIDUP(Chi_10,Chi_10) \
VMUL(Z00,Chi_00,Z1) VMUL(Z10,Chi_10,Z2) \
VSHUFMEM(3,%r8,Z00) VSHUFMEM(3,%r9,Z10) \
VMUL(Z00,Chi_00,Z3) VMUL(Z10,Chi_10,Z4) \
VSHUFMEM(6,%r8,Z00) VSHUFMEM(6,%r9,Z10) \
VMUL(Z00,Chi_00,Z5) VMUL(Z10,Chi_10,Z6) \
VMULMEM(0,%r8,T0,UChi_00) VMULMEM(0,%r9,T1,UChi_10) \
VMULMEM(3,%r8,T0,UChi_01) VMULMEM(3,%r9,T1,UChi_11) \
VMULMEM(6,%r8,T0,UChi_02) VMULMEM(6,%r9,T1,UChi_12) \
VSHUFMEM(1,%r8,Z00) VSHUFMEM(1,%r9,Z10) \
VRDUP(Chi_01,T0) VIDUP(Chi_01,Chi_01) \
VRDUP(Chi_11,T1) VIDUP(Chi_11,Chi_11) \
VMADD(Z00,Chi_01,Z1) VMADD(Z10,Chi_11,Z2) \
VSHUFMEM(4,%r8,Z00) VSHUFMEM(4,%r9,Z10) \
VMADD(Z00,Chi_01,Z3) VMADD(Z10,Chi_11,Z4) \
VSHUFMEM(7,%r8,Z00) VSHUFMEM(7,%r9,Z10) \
VMADD(Z00,Chi_01,Z5) VMADD(Z10,Chi_11,Z6) \
VMADDMEM(1,%r8,T0,UChi_00) VMADDMEM(1,%r9,T1,UChi_10) \
VMADDMEM(4,%r8,T0,UChi_01) VMADDMEM(4,%r9,T1,UChi_11) \
VMADDMEM(7,%r8,T0,UChi_02) VMADDMEM(7,%r9,T1,UChi_12) \
VSHUFMEM(2,%r8,Z00) VSHUFMEM(2,%r9,Z10) \
VRDUP(Chi_02,T0) VIDUP(Chi_02,Chi_02) \
VRDUP(Chi_12,T1) VIDUP(Chi_12,Chi_12) \
VMADD(Z00,Chi_02,Z1) VMADD(Z10,Chi_12,Z2) \
VSHUFMEM(5,%r8,Z00) VSHUFMEM(5,%r9,Z10) \
VMADD(Z00,Chi_02,Z3) VMADD(Z10,Chi_12,Z4) \
VSHUFMEM(8,%r8,Z00) VSHUFMEM(8,%r9,Z10) \
VMADD(Z00,Chi_02,Z5) VMADD(Z10,Chi_12,Z6) \
VMADDSUBMEM(2,%r8,T0,Z1) VMADDSUBMEM(2,%r9,T1,Z2) \
VMADDSUBMEM(5,%r8,T0,Z3) VMADDSUBMEM(5,%r9,T1,Z4) \
VMADDSUBMEM(8,%r8,T0,Z5) VMADDSUBMEM(8,%r9,T1,Z6) \
VADD(Z1,UChi_00,UChi_00) VADD(Z2,UChi_10,UChi_10) \
VADD(Z3,UChi_01,UChi_01) VADD(Z4,UChi_11,UChi_11) \
VADD(Z5,UChi_02,UChi_02) VADD(Z6,UChi_12,UChi_12) );
#define LOAD_CHI(a0,a1,a2,a3) \
asm ( \
"movq %0, %%r8 \n\t" \
VLOAD(0,%%r8,pChi_00) \
VLOAD(1,%%r8,pChi_01) \
VLOAD(2,%%r8,pChi_02) \
: : "r" (a0) : "%r8" ); \
asm ( \
"movq %0, %%r8 \n\t" \
VLOAD(0,%%r8,pChi_10) \
VLOAD(1,%%r8,pChi_11) \
VLOAD(2,%%r8,pChi_12) \
: : "r" (a1) : "%r8" ); \
asm ( \
"movq %0, %%r8 \n\t" \
VLOAD(0,%%r8,pChi_20) \
VLOAD(1,%%r8,pChi_21) \
VLOAD(2,%%r8,pChi_22) \
: : "r" (a2) : "%r8" ); \
asm ( \
"movq %0, %%r8 \n\t" \
VLOAD(0,%%r8,pChi_30) \
VLOAD(1,%%r8,pChi_31) \
VLOAD(2,%%r8,pChi_32) \
: : "r" (a3) : "%r8" );
#define LOAD_CHIa(a0,a1) \
asm ( \
"movq %0, %%r8 \n\t" \
VLOAD(0,%%r8,pChi_00) \
VLOAD(1,%%r8,pChi_01) \
VLOAD(2,%%r8,pChi_02) \
: : "r" (a0) : "%r8" ); \
asm ( \
"movq %0, %%r8 \n\t" \
VLOAD(0,%%r8,pChi_10) \
VLOAD(1,%%r8,pChi_11) \
VLOAD(2,%%r8,pChi_12) \
: : "r" (a1) : "%r8" );
#define PF_CHI(a0)
#define PF_CHIa(a0) \
asm ( \
"movq %0, %%r8 \n\t" \
VPREFETCH1(0,%%r8) \
VPREFETCH1(1,%%r8) \
VPREFETCH1(2,%%r8) \
: : "r" (a0) : "%r8" ); \
#define PF_GAUGE_XYZT(a0)
#define PF_GAUGE_XYZTa(a0) \
asm ( \
"movq %0, %%r8 \n\t" \
VPREFETCH1(0,%%r8) \
VPREFETCH1(1,%%r8) \
VPREFETCH1(2,%%r8) \
VPREFETCH1(3,%%r8) \
VPREFETCH1(4,%%r8) \
VPREFETCH1(5,%%r8) \
VPREFETCH1(6,%%r8) \
VPREFETCH1(7,%%r8) \
VPREFETCH1(8,%%r8) \
: : "r" (a0) : "%r8" ); \
#define PF_GAUGE_LS(a0)
#define PF_GAUGE_LSa(a0) \
asm ( \
"movq %0, %%r8 \n\t" \
VPREFETCH1(0,%%r8) \
VPREFETCH1(1,%%r8) \
: : "r" (a0) : "%r8" ); \
#define REDUCE(out) \
asm ( \
VADD(UChi_00,UChi_10,UChi_00) \
VADD(UChi_01,UChi_11,UChi_01) \
VADD(UChi_02,UChi_12,UChi_02) \
VADD(UChi_30,UChi_20,UChi_30) \
VADD(UChi_31,UChi_21,UChi_31) \
VADD(UChi_32,UChi_22,UChi_32) \
VADD(UChi_00,UChi_30,UChi_00) \
VADD(UChi_01,UChi_31,UChi_01) \
VADD(UChi_02,UChi_32,UChi_02) ); \
asm ( \
VSTORE(0,%0,pUChi_00) \
VSTORE(1,%0,pUChi_01) \
VSTORE(2,%0,pUChi_02) \
: : "r" (out) : "memory" );
#define nREDUCE(out) \
asm ( \
VADD(UChi_00,UChi_10,UChi_00) \
VADD(UChi_01,UChi_11,UChi_01) \
VADD(UChi_02,UChi_12,UChi_02) \
VADD(UChi_30,UChi_20,UChi_30) \
VADD(UChi_31,UChi_21,UChi_31) \
VADD(UChi_32,UChi_22,UChi_32) \
VADD(UChi_00,UChi_30,UChi_00) \
VADD(UChi_01,UChi_31,UChi_01) \
VADD(UChi_02,UChi_32,UChi_02) ); \
asm (VZERO(Chi_00) \
VSUB(UChi_00,Chi_00,UChi_00) \
VSUB(UChi_01,Chi_00,UChi_01) \
VSUB(UChi_02,Chi_00,UChi_02) ); \
asm ( \
VSTORE(0,%0,pUChi_00) \
VSTORE(1,%0,pUChi_01) \
VSTORE(2,%0,pUChi_02) \
: : "r" (out) : "memory" );
#define REDUCEa(out) \
asm ( \
VADD(UChi_00,UChi_10,UChi_00) \
VADD(UChi_01,UChi_11,UChi_01) \
VADD(UChi_02,UChi_12,UChi_02) ); \
asm ( \
VSTORE(0,%0,pUChi_00) \
VSTORE(1,%0,pUChi_01) \
VSTORE(2,%0,pUChi_02) \
: : "r" (out) : "memory" );
// FIXME is sign right in the VSUB ?
#define nREDUCEa(out) \
asm ( \
VADD(UChi_00,UChi_10,UChi_00) \
VADD(UChi_01,UChi_11,UChi_01) \
VADD(UChi_02,UChi_12,UChi_02) ); \
asm (VZERO(Chi_00) \
VSUB(UChi_00,Chi_00,UChi_00) \
VSUB(UChi_01,Chi_00,UChi_01) \
VSUB(UChi_02,Chi_00,UChi_02) ); \
asm ( \
VSTORE(0,%0,pUChi_00) \
VSTORE(1,%0,pUChi_01) \
VSTORE(2,%0,pUChi_02) \
: : "r" (out) : "memory" );
#define PERMUTE_DIR(dir) \
permute##dir(Chi_0,Chi_0);\
permute##dir(Chi_1,Chi_1);\
permute##dir(Chi_2,Chi_2);
NAMESPACE_BEGIN(Grid);
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
assert(0);
};
//#define CONDITIONAL_MOVE(l,o,out) if ( l ) { out = (uint64_t) &in[o] ; } else { out =(uint64_t) &buf[o]; }
#define CONDITIONAL_MOVE(l,o,out) { const SiteSpinor *ptr = l? in_p : buf; out = (uint64_t) &ptr[o]; }
#define PREPARE_XYZT(X,Y,Z,T,skew,UU) \
PREPARE(X,Y,Z,T,skew,UU); \
PF_GAUGE_XYZT(gauge0); \
PF_GAUGE_XYZT(gauge1); \
PF_GAUGE_XYZT(gauge2); \
PF_GAUGE_XYZT(gauge3);
#define PREPARE_LS(X,Y,Z,T,skew,UU) \
PREPARE(X,Y,Z,T,skew,UU); \
PF_GAUGE_LS(gauge0); \
PF_GAUGE_LS(gauge1); \
PF_GAUGE_LS(gauge2); \
PF_GAUGE_LS(gauge3);
#define PREPARE(X,Y,Z,T,skew,UU) \
SE0=st.GetEntry(ptype,X+skew,sF); \
o0 = SE0->_offset; \
l0 = SE0->_is_local; \
p0 = SE0->_permute; \
CONDITIONAL_MOVE(l0,o0,addr0); \
PF_CHI(addr0); \
\
SE1=st.GetEntry(ptype,Y+skew,sF); \
o1 = SE1->_offset; \
l1 = SE1->_is_local; \
p1 = SE1->_permute; \
CONDITIONAL_MOVE(l1,o1,addr1); \
PF_CHI(addr1); \
\
SE2=st.GetEntry(ptype,Z+skew,sF); \
o2 = SE2->_offset; \
l2 = SE2->_is_local; \
p2 = SE2->_permute; \
CONDITIONAL_MOVE(l2,o2,addr2); \
PF_CHI(addr2); \
\
SE3=st.GetEntry(ptype,T+skew,sF); \
o3 = SE3->_offset; \
l3 = SE3->_is_local; \
p3 = SE3->_permute; \
CONDITIONAL_MOVE(l3,o3,addr3); \
PF_CHI(addr3); \
\
gauge0 =(uint64_t)&UU[sU]( X ); \
gauge1 =(uint64_t)&UU[sU]( Y ); \
gauge2 =(uint64_t)&UU[sU]( Z ); \
gauge3 =(uint64_t)&UU[sU]( T );
// This is the single precision 5th direction vectorised kernel
#include <simd/Intel512single.h>
template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
#ifdef AVX512
uint64_t gauge0,gauge1,gauge2,gauge3;
uint64_t addr0,addr1,addr2,addr3;
const SiteSpinor *in_p; in_p = &in[0];
int o0,o1,o2,o3; // offsets
int l0,l1,l2,l3; // local
int p0,p1,p2,p3; // perm
int ptype;
StencilEntry *SE0;
StencilEntry *SE1;
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_LS(gauge0,gauge1,gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
PREPARE(Xp,Yp,Zp,Tp,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
addr0 = (uint64_t) &out[sF];
if ( dag ) {
nREDUCE(addr0);
} else {
REDUCE(addr0);
}
}
#else
assert(0);
#endif
}
#include <simd/Intel512double.h>
template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionFieldView &in, FermionFieldView &out, int dag)
{
#ifdef AVX512
uint64_t gauge0,gauge1,gauge2,gauge3;
uint64_t addr0,addr1,addr2,addr3;
const SiteSpinor *in_p; in_p = &in[0];
int o0,o1,o2,o3; // offsets
int l0,l1,l2,l3; // local
int p0,p1,p2,p3; // perm
int ptype;
StencilEntry *SE0;
StencilEntry *SE1;
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_LS(gauge0,gauge1,gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
PREPARE(Xp,Yp,Zp,Tp,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
addr0 = (uint64_t) &out[sF];
if ( dag ) {
nREDUCE(addr0);
} else {
REDUCE(addr0);
}
}
#else
assert(0);
#endif
}
#define PERMUTE_DIR3 __asm__ ( \
VPERM3(Chi_00,Chi_00) \
VPERM3(Chi_01,Chi_01) \
VPERM3(Chi_02,Chi_02) );
#define PERMUTE_DIR2 __asm__ ( \
VPERM2(Chi_10,Chi_10) \
VPERM2(Chi_11,Chi_11) \
VPERM2(Chi_12,Chi_12) );
#define PERMUTE_DIR1 __asm__ ( \
VPERM1(Chi_00,Chi_00) \
VPERM1(Chi_01,Chi_01) \
VPERM1(Chi_02,Chi_02) );
#define PERMUTE_DIR0 __asm__ ( \
VPERM0(Chi_10,Chi_10) \
VPERM0(Chi_11,Chi_11) \
VPERM0(Chi_12,Chi_12) );
#define PERMUTE01 \
if ( p0 ) { PERMUTE_DIR3; }\
if ( p1 ) { PERMUTE_DIR2; }
#define PERMUTE23 \
if ( p2 ) { PERMUTE_DIR1; }\
if ( p3 ) { PERMUTE_DIR0; }
// This is the single precision 5th direction vectorised kernel
#include <simd/Intel512single.h>
template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
#ifdef AVX512
uint64_t gauge0,gauge1,gauge2,gauge3;
uint64_t addr0,addr1,addr2,addr3;
const SiteSpinor *in_p; in_p = &in[0];
int o0,o1,o2,o3; // offsets
int l0,l1,l2,l3; // local
int p0,p1,p2,p3; // perm
int ptype;
StencilEntry *SE0;
StencilEntry *SE1;
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,0,U);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xp,Yp,Zp,Tp,8,UUU);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,8,UUU);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
addr0 = (uint64_t) &out[sF];
if ( dag ) {
nREDUCEa(addr0);
} else {
REDUCEa(addr0);
}
}
#else
assert(0);
#endif
}
#include <simd/Intel512double.h>
template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U,
DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionFieldView &in, FermionFieldView &out,int dag)
{
#ifdef AVX512
uint64_t gauge0,gauge1,gauge2,gauge3;
uint64_t addr0,addr1,addr2,addr3;
const SiteSpinor *in_p; in_p = &in[0];
int o0,o1,o2,o3; // offsets
int l0,l1,l2,l3; // local
int p0,p1,p2,p3; // perm
int ptype;
StencilEntry *SE0;
StencilEntry *SE1;
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,0,U);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xp,Yp,Zp,Tp,8,UUU);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,8,UUU);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
addr0 = (uint64_t) &out[sF];
if ( dag ) {
nREDUCEa(addr0);
} else {
REDUCEa(addr0);
}
}
#else
assert(0);
#endif
}
#define KERNEL_INSTANTIATE(CLASS,FUNC,IMPL) \
template void CLASS<IMPL>::FUNC(StencilImpl &st, LebesgueOrder &lo, \
DoubledGaugeFieldView &U, \
DoubledGaugeFieldView &UUU, \
SiteSpinor *buf, int LLs, \
int sU, const FermionFieldView &in, FermionFieldView &out,int dag);
//KERNEL_INSTANTIATE(StaggeredKernels,DhopSiteAsm,StaggeredImplD);
//KERNEL_INSTANTIATE(StaggeredKernels,DhopSiteAsm,StaggeredImplF);
//KERNEL_INSTANTIATE(StaggeredKernels,DhopSiteAsm,StaggeredVec5dImplD);
//KERNEL_INSTANTIATE(StaggeredKernels,DhopSiteAsm,StaggeredVec5dImplF);
NAMESPACE_END(Grid);

396
Grid/qcd/action/fermion/StaggeredKernelsHand.cc
View File

@ -1,396 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/StaggerdKernelsHand.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid.h>
NAMESPACE_BEGIN(Grid);
#define LOAD_CHI(b) \
const SiteSpinor & ref (b[offset]); \
Chi_0=ref()()(0);\
Chi_1=ref()()(1);\
Chi_2=ref()()(2);
// To splat or not to splat depends on the implementation
#define MULT(A,UChi) \
auto & ref(U[sU](A)); \
Impl::loadLinkElement(U_00,ref()(0,0)); \
Impl::loadLinkElement(U_10,ref()(1,0)); \
Impl::loadLinkElement(U_20,ref()(2,0)); \
Impl::loadLinkElement(U_01,ref()(0,1)); \
Impl::loadLinkElement(U_11,ref()(1,1)); \
Impl::loadLinkElement(U_21,ref()(2,1)); \
Impl::loadLinkElement(U_02,ref()(0,2)); \
Impl::loadLinkElement(U_12,ref()(1,2)); \
Impl::loadLinkElement(U_22,ref()(2,2)); \
UChi ## _0 = U_00*Chi_0; \
UChi ## _1 = U_10*Chi_0;\
UChi ## _2 = U_20*Chi_0;\
UChi ## _0 += U_01*Chi_1;\
UChi ## _1 += U_11*Chi_1;\
UChi ## _2 += U_21*Chi_1;\
UChi ## _0 += U_02*Chi_2;\
UChi ## _1 += U_12*Chi_2;\
UChi ## _2 += U_22*Chi_2;
#define MULT_ADD(U,A,UChi) \
auto & ref(U[sU](A)); \
Impl::loadLinkElement(U_00,ref()(0,0)); \
Impl::loadLinkElement(U_10,ref()(1,0)); \
Impl::loadLinkElement(U_20,ref()(2,0)); \
Impl::loadLinkElement(U_01,ref()(0,1)); \
Impl::loadLinkElement(U_11,ref()(1,1)); \
Impl::loadLinkElement(U_21,ref()(2,1)); \
Impl::loadLinkElement(U_02,ref()(0,2)); \
Impl::loadLinkElement(U_12,ref()(1,2)); \
Impl::loadLinkElement(U_22,ref()(2,2)); \
UChi ## _0 += U_00*Chi_0; \
UChi ## _1 += U_10*Chi_0;\
UChi ## _2 += U_20*Chi_0;\
UChi ## _0 += U_01*Chi_1;\
UChi ## _1 += U_11*Chi_1;\
UChi ## _2 += U_21*Chi_1;\
UChi ## _0 += U_02*Chi_2;\
UChi ## _1 += U_12*Chi_2;\
UChi ## _2 += U_22*Chi_2;
#define PERMUTE_DIR(dir) \
permute##dir(Chi_0,Chi_0); \
permute##dir(Chi_1,Chi_1); \
permute##dir(Chi_2,Chi_2);
#define HAND_STENCIL_LEG_BASE(Dir,Perm,skew) \
SE=st.GetEntry(ptype,Dir+skew,sF); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHI(in); \
if ( perm) { \
PERMUTE_DIR(Perm); \
} \
} else { \
LOAD_CHI(buf); \
}
#define HAND_STENCIL_LEG_BEGIN(Dir,Perm,skew,even) \
HAND_STENCIL_LEG_BASE(Dir,Perm,skew) \
{ \
MULT(Dir,even); \
}
#define HAND_STENCIL_LEG(U,Dir,Perm,skew,even) \
HAND_STENCIL_LEG_BASE(Dir,Perm,skew) \
{ \
MULT_ADD(U,Dir,even); \
}
#define HAND_STENCIL_LEG_INT(U,Dir,Perm,skew,even) \
SE=st.GetEntry(ptype,Dir+skew,sF); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHI(in); \
if ( perm) { \
PERMUTE_DIR(Perm); \
} \
} else if ( st.same_node[Dir] ) { \
LOAD_CHI(buf); \
} \
if (SE->_is_local || st.same_node[Dir] ) { \
MULT_ADD(U,Dir,even); \
}
#define HAND_STENCIL_LEG_EXT(U,Dir,Perm,skew,even) \
SE=st.GetEntry(ptype,Dir+skew,sF); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ((!SE->_is_local) && (!st.same_node[Dir]) ) { \
nmu++; \
{ LOAD_CHI(buf); } \
{ MULT_ADD(U,Dir,even); } \
}
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
Simd even_0; // 12 regs on knc
Simd even_1;
Simd even_2;
Simd odd_0; // 12 regs on knc
Simd odd_1;
Simd odd_2;
Simd Chi_0; // two spinor; 6 regs
Simd Chi_1;
Simd Chi_2;
Simd U_00; // two rows of U matrix
Simd U_10;
Simd U_20;
Simd U_01;
Simd U_11;
Simd U_21; // 2 reg left.
Simd U_02;
Simd U_12;
Simd U_22;
SiteSpinor result;
int offset,local,perm, ptype;
StencilEntry *SE;
int skew;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
skew = 0;
HAND_STENCIL_LEG_BEGIN(Xp,3,skew,even);
HAND_STENCIL_LEG_BEGIN(Yp,2,skew,odd);
HAND_STENCIL_LEG (U,Zp,1,skew,even);
HAND_STENCIL_LEG (U,Tp,0,skew,odd);
HAND_STENCIL_LEG (U,Xm,3,skew,even);
HAND_STENCIL_LEG (U,Ym,2,skew,odd);
HAND_STENCIL_LEG (U,Zm,1,skew,even);
HAND_STENCIL_LEG (U,Tm,0,skew,odd);
skew = 8;
HAND_STENCIL_LEG(UUU,Xp,3,skew,even);
HAND_STENCIL_LEG(UUU,Yp,2,skew,odd);
HAND_STENCIL_LEG(UUU,Zp,1,skew,even);
HAND_STENCIL_LEG(UUU,Tp,0,skew,odd);
HAND_STENCIL_LEG(UUU,Xm,3,skew,even);
HAND_STENCIL_LEG(UUU,Ym,2,skew,odd);
HAND_STENCIL_LEG(UUU,Zm,1,skew,even);
HAND_STENCIL_LEG(UUU,Tm,0,skew,odd);
if ( dag ) {
result()()(0) = - even_0 - odd_0;
result()()(1) = - even_1 - odd_1;
result()()(2) = - even_2 - odd_2;
} else {
result()()(0) = even_0 + odd_0;
result()()(1) = even_1 + odd_1;
result()()(2) = even_2 + odd_2;
}
vstream(out[sF],result);
}
}
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
Simd even_0; // 12 regs on knc
Simd even_1;
Simd even_2;
Simd odd_0; // 12 regs on knc
Simd odd_1;
Simd odd_2;
Simd Chi_0; // two spinor; 6 regs
Simd Chi_1;
Simd Chi_2;
Simd U_00; // two rows of U matrix
Simd U_10;
Simd U_20;
Simd U_01;
Simd U_11;
Simd U_21; // 2 reg left.
Simd U_02;
Simd U_12;
Simd U_22;
SiteSpinor result;
int offset, ptype, local, perm;
StencilEntry *SE;
int skew;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
even_0 = Zero(); even_1 = Zero(); even_2 = Zero();
odd_0 = Zero(); odd_1 = Zero(); odd_2 = Zero();
skew = 0;
HAND_STENCIL_LEG_INT(U,Xp,3,skew,even);
HAND_STENCIL_LEG_INT(U,Yp,2,skew,odd);
HAND_STENCIL_LEG_INT(U,Zp,1,skew,even);
HAND_STENCIL_LEG_INT(U,Tp,0,skew,odd);
HAND_STENCIL_LEG_INT(U,Xm,3,skew,even);
HAND_STENCIL_LEG_INT(U,Ym,2,skew,odd);
HAND_STENCIL_LEG_INT(U,Zm,1,skew,even);
HAND_STENCIL_LEG_INT(U,Tm,0,skew,odd);
skew = 8;
HAND_STENCIL_LEG_INT(UUU,Xp,3,skew,even);
HAND_STENCIL_LEG_INT(UUU,Yp,2,skew,odd);
HAND_STENCIL_LEG_INT(UUU,Zp,1,skew,even);
HAND_STENCIL_LEG_INT(UUU,Tp,0,skew,odd);
HAND_STENCIL_LEG_INT(UUU,Xm,3,skew,even);
HAND_STENCIL_LEG_INT(UUU,Ym,2,skew,odd);
HAND_STENCIL_LEG_INT(UUU,Zm,1,skew,even);
HAND_STENCIL_LEG_INT(UUU,Tm,0,skew,odd);
// Assume every site must be connected to at least one interior point. No 1^4 subvols.
if ( dag ) {
result()()(0) = - even_0 - odd_0;
result()()(1) = - even_1 - odd_1;
result()()(2) = - even_2 - odd_2;
} else {
result()()(0) = even_0 + odd_0;
result()()(1) = even_1 + odd_1;
result()()(2) = even_2 + odd_2;
}
vstream(out[sF],result);
}
}
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeFieldView &U, DoubledGaugeFieldView &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionFieldView &in, FermionFieldView &out,int dag)
{
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
Simd even_0; // 12 regs on knc
Simd even_1;
Simd even_2;
Simd odd_0; // 12 regs on knc
Simd odd_1;
Simd odd_2;
Simd Chi_0; // two spinor; 6 regs
Simd Chi_1;
Simd Chi_2;
Simd U_00; // two rows of U matrix
Simd U_10;
Simd U_20;
Simd U_01;
Simd U_11;
Simd U_21; // 2 reg left.
Simd U_02;
Simd U_12;
Simd U_22;
SiteSpinor result;
int offset, ptype, local, perm;
StencilEntry *SE;
int skew;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
even_0 = Zero(); even_1 = Zero(); even_2 = Zero();
odd_0 = Zero(); odd_1 = Zero(); odd_2 = Zero();
int nmu=0;
skew = 0;
HAND_STENCIL_LEG_EXT(U,Xp,3,skew,even);
HAND_STENCIL_LEG_EXT(U,Yp,2,skew,odd);
HAND_STENCIL_LEG_EXT(U,Zp,1,skew,even);
HAND_STENCIL_LEG_EXT(U,Tp,0,skew,odd);
HAND_STENCIL_LEG_EXT(U,Xm,3,skew,even);
HAND_STENCIL_LEG_EXT(U,Ym,2,skew,odd);
HAND_STENCIL_LEG_EXT(U,Zm,1,skew,even);
HAND_STENCIL_LEG_EXT(U,Tm,0,skew,odd);
skew = 8;
HAND_STENCIL_LEG_EXT(UUU,Xp,3,skew,even);
HAND_STENCIL_LEG_EXT(UUU,Yp,2,skew,odd);
HAND_STENCIL_LEG_EXT(UUU,Zp,1,skew,even);
HAND_STENCIL_LEG_EXT(UUU,Tp,0,skew,odd);
HAND_STENCIL_LEG_EXT(UUU,Xm,3,skew,even);
HAND_STENCIL_LEG_EXT(UUU,Ym,2,skew,odd);
HAND_STENCIL_LEG_EXT(UUU,Zm,1,skew,even);
HAND_STENCIL_LEG_EXT(UUU,Tm,0,skew,odd);
// Add sum of all exterior connected stencil legs
if ( nmu ) {
if ( dag ) {
result()()(0) = - even_0 - odd_0;
result()()(1) = - even_1 - odd_1;
result()()(2) = - even_2 - odd_2;
} else {
result()()(0) = even_0 + odd_0;
result()()(1) = even_1 + odd_1;
result()()(2) = even_2 + odd_2;
}
out[sF] = out[sF] + result;
}
}
}
#define DHOP_SITE_HAND_INSTANTIATE(IMPL) \
template void StaggeredKernels<IMPL>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, \
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
SiteSpinor *buf, int LLs, int sU, \
const FermionFieldView &in, FermionFieldView &out, int dag); \
\
template void StaggeredKernels<IMPL>::DhopSiteHandInt(StencilImpl &st, LebesgueOrder &lo, \
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
SiteSpinor *buf, int LLs, int sU, \
const FermionFieldView &in, FermionFieldView &out, int dag); \
\
template void StaggeredKernels<IMPL>::DhopSiteHandExt(StencilImpl &st, LebesgueOrder &lo, \
DoubledGaugeFieldView &U,DoubledGaugeFieldView &UUU, \
SiteSpinor *buf, int LLs, int sU, \
const FermionFieldView &in, FermionFieldView &out, int dag); \
DHOP_SITE_HAND_INSTANTIATE(StaggeredImplD);
DHOP_SITE_HAND_INSTANTIATE(StaggeredImplF);
DHOP_SITE_HAND_INSTANTIATE(StaggeredVec5dImplD);
DHOP_SITE_HAND_INSTANTIATE(StaggeredVec5dImplF);
NAMESPACE_END(Grid);

203
Grid/qcd/action/fermion/StaggeredVec5dImpl.h Normal file
View File

@ -0,0 +1,203 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/FermionOperatorImpl.h
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#pragma once
NAMESPACE_BEGIN(Grid);
template <class S, class Representation = FundamentalRepresentation >
class StaggeredVec5dImpl : public PeriodicGaugeImpl<GaugeImplTypes<S, Representation::Dimension > > {
public:
static const int Dimension = Representation::Dimension;
static const bool isFundamental = Representation::isFundamental;
static const bool LsVectorised=true;
typedef RealD Coeff_t ;
typedef PeriodicGaugeImpl<GaugeImplTypes<S, Dimension > > Gimpl;
//Necessary?
constexpr bool is_fundamental() const{return Dimension == Nc ? 1 : 0;}
INHERIT_GIMPL_TYPES(Gimpl);
template <typename vtype> using iImplSpinor = iScalar<iScalar<iVector<vtype, Dimension> > >;
template <typename vtype> using iImplHalfSpinor = iScalar<iScalar<iVector<vtype, Dimension> > >;
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> > >;
template <typename vtype> using iImplPropagator = iScalar<iScalar<iMatrix<vtype, Dimension> > >;
// Make the doubled gauge field a *scalar*
typedef iImplDoubledGaugeField<typename Simd::scalar_type> SiteDoubledGaugeField; // This is a scalar
typedef iImplGaugeField<typename Simd::scalar_type> SiteScalarGaugeField; // scalar
typedef iImplGaugeLink<typename Simd::scalar_type> SiteScalarGaugeLink; // scalar
typedef iImplPropagator<Simd> SitePropagator;
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef Lattice<SitePropagator> PropagatorField;
typedef iImplSpinor<Simd> SiteSpinor;
typedef iImplHalfSpinor<Simd> SiteHalfSpinor;
typedef Lattice<SiteSpinor> FermionField;
typedef StaggeredImplParams ImplParams;
typedef SimpleCompressor<SiteSpinor> Compressor;
typedef CartesianStencil<SiteSpinor, SiteSpinor, ImplParams> StencilImpl;
typedef typename StencilImpl::View_type StencilView;
ImplParams Params;
StaggeredVec5dImpl(const ImplParams &p = ImplParams()) : Params(p){};
template <class ref>
static accelerator_inline void loadLinkElement(Simd &reg, ref &memory)
{
vsplat(reg, memory);
}
static accelerator_inline void multLink(SiteHalfSpinor &phi,
const SiteDoubledGaugeField &U,
const SiteHalfSpinor &chi,
int mu)
{
SiteGaugeLink UU;
for (int i = 0; i < Dimension; i++) {
for (int j = 0; j < Dimension; j++) {
vsplat(UU()()(i, j), U(mu)()(i, j));
}
}
mult(&phi(), &UU(), &chi());
}
static accelerator_inline void multLinkAdd(SiteHalfSpinor &phi,
const SiteDoubledGaugeField &U,
const SiteHalfSpinor &chi,
int mu)
{
SiteGaugeLink UU;
for (int i = 0; i < Dimension; i++) {
for (int j = 0; j < Dimension; j++) {
vsplat(UU()()(i, j), U(mu)()(i, j));
}
}
mac(&phi(), &UU(), &chi());
}
inline void InsertGaugeField(DoubledGaugeField &U_ds,const GaugeLinkField &U,int mu)
{
GridBase *GaugeGrid = U_ds.Grid();
thread_loop( (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++), {
SiteScalarGaugeLink ScalarU;
SiteDoubledGaugeField ScalarUds;
Coordinate lcoor;
GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
peekLocalSite(ScalarUds, U_ds, lcoor);
peekLocalSite(ScalarU, U, lcoor);
ScalarUds(mu) = ScalarU();
});
}
inline void DoubleStore(GridBase *GaugeGrid,
DoubledGaugeField &UUUds, // for Naik term
DoubledGaugeField &Uds,
const GaugeField &Uthin,
const GaugeField &Ufat)
{
GridBase * InputGrid = Uthin.Grid();
conformable(InputGrid,Ufat.Grid());
GaugeLinkField U(InputGrid);
GaugeLinkField UU(InputGrid);
GaugeLinkField UUU(InputGrid);
GaugeLinkField Udag(InputGrid);
GaugeLinkField UUUdag(InputGrid);
for (int mu = 0; mu < Nd; mu++) {
// Staggered Phase.
Lattice<iScalar<vInteger> > coor(InputGrid);
Lattice<iScalar<vInteger> > x(InputGrid); LatticeCoordinate(x,0);
Lattice<iScalar<vInteger> > y(InputGrid); LatticeCoordinate(y,1);
Lattice<iScalar<vInteger> > z(InputGrid); LatticeCoordinate(z,2);
Lattice<iScalar<vInteger> > t(InputGrid); LatticeCoordinate(t,3);
Lattice<iScalar<vInteger> > lin_z(InputGrid); lin_z=x+y;
Lattice<iScalar<vInteger> > lin_t(InputGrid); lin_t=x+y+z;
ComplexField phases(InputGrid); phases=1.0;
if ( mu == 1 ) phases = where( mod(x ,2)==(Integer)0, phases,-phases);
if ( mu == 2 ) phases = where( mod(lin_z,2)==(Integer)0, phases,-phases);
if ( mu == 3 ) phases = where( mod(lin_t,2)==(Integer)0, phases,-phases);
// 1 hop based on fat links
U = PeekIndex<LorentzIndex>(Ufat, mu);
Udag = adj( Cshift(U, mu, -1));
U = U *phases;
Udag = Udag *phases;
InsertGaugeField(Uds,U,mu);
InsertGaugeField(Uds,Udag,mu+4);
// 3 hop based on thin links. Crazy huh ?
U = PeekIndex<LorentzIndex>(Uthin, mu);
UU = Gimpl::CovShiftForward(U,mu,U);
UUU= Gimpl::CovShiftForward(U,mu,UU);
UUUdag = adj( Cshift(UUU, mu, -3));
UUU = UUU *phases;
UUUdag = UUUdag *phases;
InsertGaugeField(UUUds,UUU,mu);
InsertGaugeField(UUUds,UUUdag,mu+4);
}
}
inline void InsertForce4D(GaugeField &mat, FermionField &Btilde, FermionField &A,int mu){
assert(0);
}
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
assert (0);
}
};
typedef StaggeredVec5dImpl<vComplex, FundamentalRepresentation > StaggeredVec5dImplR; // Real.. whichever prec
typedef StaggeredVec5dImpl<vComplexF, FundamentalRepresentation > StaggeredVec5dImplF; // Float
typedef StaggeredVec5dImpl<vComplexD, FundamentalRepresentation > StaggeredVec5dImplD; // Double
NAMESPACE_END(Grid);

242
Grid/qcd/action/fermion/WilsonCloverFermion.cc
View File

@ -1,242 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
Copyright (C) 2017
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
#include <Grid/qcd/spin/Dirac.h>
NAMESPACE_BEGIN(Grid);
// *NOT* EO
template <class Impl>
RealD WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
// Wilson term
out.Checkerboard() = in.Checkerboard();
this->Dhop(in, out, DaggerNo);
// Clover term
Mooee(in, temp);
out += temp;
return norm2(out);
}
template <class Impl>
RealD WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
// Wilson term
out.Checkerboard() = in.Checkerboard();
this->Dhop(in, out, DaggerYes);
// Clover term
MooeeDag(in, temp);
out += temp;
return norm2(out);
}
template <class Impl>
void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
{
WilsonFermion<Impl>::ImportGauge(_Umu);
GridBase *grid = _Umu.Grid();
typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
// Compute the field strength terms mu>nu
WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
// Compute the Clover Operator acting on Colour and Spin
// multiply here by the clover coefficients for the anisotropy
CloverTerm = fillCloverYZ(Bx) * csw_r;
CloverTerm += fillCloverXZ(By) * csw_r;
CloverTerm += fillCloverXY(Bz) * csw_r;
CloverTerm += fillCloverXT(Ex) * csw_t;
CloverTerm += fillCloverYT(Ey) * csw_t;
CloverTerm += fillCloverZT(Ez) * csw_t;
CloverTerm += diag_mass;
int lvol = _Umu.Grid()->lSites();
int DimRep = Impl::Dimension;
Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
Coordinate lcoor;
typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
for (int site = 0; site < lvol; site++)
{
grid->LocalIndexToLocalCoor(site, lcoor);
EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
peekLocalSite(Qx, CloverTerm, lcoor);
Qxinv = Zero();
//if (csw!=0){
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++){
auto zz = Qx()(j, k)(a, b);
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
}
// if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
EigenInvCloverOp = EigenCloverOp.inverse();
//std::cout << EigenInvCloverOp << std::endl;
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++)
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
// if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
// }
pokeLocalSite(Qxinv, CloverTermInv, lcoor);
}
// Separate the even and odd parts
pickCheckerboard(Even, CloverTermEven, CloverTerm);
pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
pickCheckerboard(Even, CloverTermDagEven, adj(CloverTerm));
pickCheckerboard(Odd, CloverTermDagOdd, adj(CloverTerm));
pickCheckerboard(Even, CloverTermInvEven, CloverTermInv);
pickCheckerboard(Odd, CloverTermInvOdd, CloverTermInv);
pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
}
template <class Impl>
void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerNo, InverseNo);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerYes, InverseNo);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerNo, InverseYes);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerYes, InverseYes);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
{
out.Checkerboard() = in.Checkerboard();
CloverFieldType *Clover;
assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
if (dag)
{
if (in.Grid()->_isCheckerBoarded)
{
if (in.Checkerboard() == Odd)
{
Clover = (inv) ? &CloverTermInvDagOdd : &CloverTermDagOdd;
}
else
{
Clover = (inv) ? &CloverTermInvDagEven : &CloverTermDagEven;
}
out = *Clover * in;
}
else
{
Clover = (inv) ? &CloverTermInv : &CloverTerm;
out = adj(*Clover) * in;
}
}
else
{
if (in.Grid()->_isCheckerBoarded)
{
if (in.Checkerboard() == Odd)
{
// std::cout << "Calling clover term Odd" << std::endl;
Clover = (inv) ? &CloverTermInvOdd : &CloverTermOdd;
}
else
{
// std::cout << "Calling clover term Even" << std::endl;
Clover = (inv) ? &CloverTermInvEven : &CloverTermEven;
}
out = *Clover * in;
// std::cout << GridLogMessage << "*Clover.Checkerboard() " << (*Clover).Checkerboard() << std::endl;
}
else
{
Clover = (inv) ? &CloverTermInv : &CloverTerm;
out = *Clover * in;
}
}
} // MooeeInternal
// Derivative parts
template <class Impl>
void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
{
assert(0);
}
// Derivative parts
template <class Impl>
void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
{
assert(0); // not implemented yet
}
FermOpTemplateInstantiate(WilsonCloverFermion);
AdjointFermOpTemplateInstantiate(WilsonCloverFermion);
TwoIndexFermOpTemplateInstantiate(WilsonCloverFermion);
//GparityFermOpTemplateInstantiate(WilsonCloverFermion);
NAMESPACE_END(Grid);

596
Grid/qcd/action/fermion/WilsonFermion.cc
View File

@ -1,596 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonFermion.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/WilsonFermion.h>
NAMESPACE_BEGIN(Grid);
const std::vector<int> WilsonFermionStatic::directions({0, 1, 2, 3, 0, 1, 2, 3});
const std::vector<int> WilsonFermionStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1});
int WilsonFermionStatic::HandOptDslash;
/////////////////////////////////
// Constructor and gauge import
/////////////////////////////////
template <class Impl>
WilsonFermion<Impl>::WilsonFermion(GaugeField &_Umu, GridCartesian &Fgrid,
GridRedBlackCartesian &Hgrid, RealD _mass,
const ImplParams &p,
const WilsonAnisotropyCoefficients &anis)
:
Kernels(p),
_grid(&Fgrid),
_cbgrid(&Hgrid),
Stencil(&Fgrid, npoint, Even, directions, displacements,p),
StencilEven(&Hgrid, npoint, Even, directions,displacements,p), // source is Even
StencilOdd(&Hgrid, npoint, Odd, directions,displacements,p), // source is Odd
mass(_mass),
Lebesgue(_grid),
LebesgueEvenOdd(_cbgrid),
Umu(&Fgrid),
UmuEven(&Hgrid),
UmuOdd(&Hgrid),
_tmp(&Hgrid),
anisotropyCoeff(anis)
{
// Allocate the required comms buffer
ImportGauge(_Umu);
if (anisotropyCoeff.isAnisotropic){
diag_mass = mass + 1.0 + (Nd-1)*(anisotropyCoeff.nu / anisotropyCoeff.xi_0);
} else {
diag_mass = 4.0 + mass;
}
}
template <class Impl>
void WilsonFermion<Impl>::ImportGauge(const GaugeField &_Umu)
{
GaugeField HUmu(_Umu.Grid());
//Here multiply the anisotropy coefficients
if (anisotropyCoeff.isAnisotropic)
{
for (int mu = 0; mu < Nd; mu++)
{
GaugeLinkField U_dir = (-0.5)*PeekIndex<LorentzIndex>(_Umu, mu);
if (mu != anisotropyCoeff.t_direction)
U_dir *= (anisotropyCoeff.nu / anisotropyCoeff.xi_0);
PokeIndex<LorentzIndex>(HUmu, U_dir, mu);
}
}
else
{
HUmu = _Umu * (-0.5);
}
Impl::DoubleStore(GaugeGrid(), Umu, HUmu);
pickCheckerboard(Even, UmuEven, Umu);
pickCheckerboard(Odd, UmuOdd, Umu);
}
/////////////////////////////
// Implement the interface
/////////////////////////////
template <class Impl>
RealD WilsonFermion<Impl>::M(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerNo);
return axpy_norm(out, diag_mass, in, out);
}
template <class Impl>
RealD WilsonFermion<Impl>::Mdag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Dhop(in, out, DaggerYes);
return axpy_norm(out, diag_mass, in, out);
}
template <class Impl>
void WilsonFermion<Impl>::Meooe(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerNo);
} else {
DhopOE(in, out, DaggerNo);
}
}
template <class Impl>
void WilsonFermion<Impl>::MeooeDag(const FermionField &in, FermionField &out) {
if (in.Checkerboard() == Odd) {
DhopEO(in, out, DaggerYes);
} else {
DhopOE(in, out, DaggerYes);
}
}
template <class Impl>
void WilsonFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
typename FermionField::scalar_type scal(diag_mass);
out = scal * in;
}
template <class Impl>
void WilsonFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
Mooee(in, out);
}
template<class Impl>
void WilsonFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
out = (1.0/(diag_mass))*in;
}
template<class Impl>
void WilsonFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) {
out.Checkerboard() = in.Checkerboard();
MooeeInv(in,out);
}
template<class Impl>
void WilsonFermion<Impl>::MomentumSpacePropagator(FermionField &out, const FermionField &in,RealD _m,std::vector<double> twist)
{
typedef typename FermionField::vector_type vector_type;
typedef typename FermionField::scalar_type ScalComplex;
typedef Lattice<iSinglet<vector_type> > LatComplex;
// what type LatticeComplex
conformable(_grid,out.Grid());
Gamma::Algebra Gmu [] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT
};
Coordinate latt_size = _grid->_fdimensions;
FermionField num (_grid); num = Zero();
LatComplex wilson(_grid); wilson= Zero();
LatComplex one (_grid); one = ScalComplex(1.0,0.0);
LatComplex denom(_grid); denom= Zero();
LatComplex kmu(_grid);
ScalComplex ci(0.0,1.0);
// momphase = n * 2pi / L
for(int mu=0;mu<Nd;mu++) {
LatticeCoordinate(kmu,mu);
RealD TwoPiL = M_PI * 2.0/ latt_size[mu];
kmu = TwoPiL * kmu;
kmu = kmu + TwoPiL * one * twist[mu];//momentum for twisted boundary conditions
wilson = wilson + 2.0*sin(kmu*0.5)*sin(kmu*0.5); // Wilson term
num = num - sin(kmu)*ci*(Gamma(Gmu[mu])*in); // derivative term
denom=denom + sin(kmu)*sin(kmu);
}
wilson = wilson + _m; // 2 sin^2 k/2 + m
num = num + wilson*in; // -i gmu sin k + 2 sin^2 k/2 + m
denom= denom+wilson*wilson; // sin^2 k + (2 sin^2 k/2 + m)^2
denom= one/denom;
out = num*denom; // [ -i gmu sin k + 2 sin^2 k/2 + m] / [ sin^2 k + (2 sin^2 k/2 + m)^2 ]
}
///////////////////////////////////
// Internal
///////////////////////////////////
template <class Impl>
void WilsonFermion<Impl>::DerivInternal(StencilImpl &st, DoubledGaugeField &U,
GaugeField &mat, const FermionField &A,
const FermionField &B, int dag) {
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor(dag);
FermionField Btilde(B.Grid());
FermionField Atilde(B.Grid());
Atilde = A;
st.HaloExchange(B, compressor);
for (int mu = 0; mu < Nd; mu++) {
////////////////////////////////////////////////////////////////////////
// Flip gamma (1+g)<->(1-g) if dag
////////////////////////////////////////////////////////////////////////
int gamma = mu;
if (!dag) gamma += Nd;
////////////////////////
// Call the single hop
////////////////////////
auto U_v = U.View();
auto B_v = B.View();
auto Btilde_v = Btilde.View();
auto st_v = st.View();
thread_loop( (int sss = 0; sss < B.Grid()->oSites(); sss++) ,{
Kernels::DhopDirK(st_v, U_v, st.CommBuf(), sss, sss, B_v, Btilde_v, mu, gamma);
});
//////////////////////////////////////////////////
// spin trace outer product
//////////////////////////////////////////////////
Impl::InsertForce4D(mat, Btilde, Atilde, mu);
}
}
template <class Impl>
void WilsonFermion<Impl>::DhopDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _grid);
conformable(U.Grid(), V.Grid());
conformable(U.Grid(), mat.Grid());
mat.Checkerboard() = U.Checkerboard();
DerivInternal(Stencil, Umu, mat, U, V, dag);
}
template <class Impl>
void WilsonFermion<Impl>::DhopDerivOE(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
//conformable(U.Grid(), mat.Grid()); not general, leaving as a comment (Guido)
// Motivation: look at the SchurDiff operator
assert(V.Checkerboard() == Even);
assert(U.Checkerboard() == Odd);
mat.Checkerboard() = Odd;
DerivInternal(StencilEven, UmuOdd, mat, U, V, dag);
}
template <class Impl>
void WilsonFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionField &U, const FermionField &V, int dag) {
conformable(U.Grid(), _cbgrid);
conformable(U.Grid(), V.Grid());
//conformable(U.Grid(), mat.Grid());
assert(V.Checkerboard() == Odd);
assert(U.Checkerboard() == Even);
mat.Checkerboard() = Even;
DerivInternal(StencilOdd, UmuEven, mat, U, V, dag);
}
template <class Impl>
void WilsonFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) {
conformable(in.Grid(), _grid); // verifies full grid
conformable(in.Grid(), out.Grid());
out.Checkerboard() = in.Checkerboard();
DhopInternal(Stencil, Lebesgue, Umu, in, out, dag);
}
template <class Impl>
void WilsonFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) {
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
assert(in.Checkerboard() == Even);
out.Checkerboard() = Odd;
DhopInternal(StencilEven, LebesgueEvenOdd, UmuOdd, in, out, dag);
}
template <class Impl>
void WilsonFermion<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag) {
conformable(in.Grid(), _cbgrid); // verifies half grid
conformable(in.Grid(), out.Grid()); // drops the cb check
assert(in.Checkerboard() == Odd);
out.Checkerboard() = Even;
DhopInternal(StencilOdd, LebesgueEvenOdd, UmuEven, in, out, dag);
}
template <class Impl>
void WilsonFermion<Impl>::Mdir(const FermionField &in, FermionField &out, int dir, int disp) {
DhopDir(in, out, dir, disp);
}
template <class Impl>
void WilsonFermion<Impl>::DhopDir(const FermionField &in, FermionField &out, int dir, int disp) {
int skip = (disp == 1) ? 0 : 1;
int dirdisp = dir + skip * 4;
int gamma = dir + (1 - skip) * 4;
DhopDirDisp(in, out, dirdisp, gamma, DaggerNo);
};
template <class Impl>
void WilsonFermion<Impl>::DhopDirDisp(const FermionField &in, FermionField &out,int dirdisp, int gamma, int dag)
{
Compressor compressor(dag);
Stencil.HaloExchange(in, compressor);
auto in_v = in.View();
auto out_v = in.View();
auto Umu_v = Umu.View();
auto Stencil_v = Stencil.View();
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++) ,{
Kernels::DhopDirK(Stencil_v, Umu_v, Stencil.CommBuf(), sss, sss, in_v, out_v, dirdisp, gamma);
});
};
template <class Impl>
void WilsonFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag) {
#ifdef GRID_OMP
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,in,out,dag);
else
#endif
DhopInternalSerial(st,lo,U,in,out,dag);
}
template <class Impl>
void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag) {
assert((dag == DaggerNo) || (dag == DaggerYes));
#ifdef GRID_OMP
Compressor compressor(dag);
int len = U.Grid()->oSites();
const int LLs = 1;
st.Prepare();
st.HaloGather(in,compressor);
st.CommsMergeSHM(compressor);
#pragma omp parallel
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = len;
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
for (int sss = myblock; sss < myblock+myn; ++sss) {
Kernels::DhopSiteDag(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,1,0);
// Kernels::DhopSiteDag(st_v, lo, U_v, st.CommBuf(), sss, sss, 1, 1, in_v, out_v);
}
} else {
for (int sss = myblock; sss < myblock+myn; ++sss) {
Kernels::DhopSite(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,1,0);
// Kernels::DhopSite(st_v, lo, U_v, st.CommBuf(), sss, sss, 1, 1, in_v, out_v);
}
}
} else {
st.CommunicateThreaded();
}
} //pragma
{
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++) ,{
Kernels::DhopSiteDag(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,0,1);
});
} else {
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++) ,{
Kernels::DhopSite(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,0,1);
});
}
}
#else
assert(0);
#endif
};
template <class Impl>
void WilsonFermion<Impl>::DhopInternalSerial(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
const FermionField &in,
FermionField &out, int dag) {
assert((dag == DaggerNo) || (dag == DaggerYes));
Compressor compressor(dag);
st.HaloExchange(in, compressor);
int Opt = WilsonKernelsStatic::Opt;
auto U_v = U.View();
auto in_v = in.View();
auto out_v= out.View();
auto st_v = st.View();
if (dag == DaggerYes) {
accelerator_loop( sss,in_v, {
Kernels::DhopSiteDag(Opt,st_v, U_v, st.CommBuf(), sss, sss, 1, 1, in_v, out_v);
});
} else {
accelerator_loop( sss,in_v, {
Kernels::DhopSite(Opt,st_v, U_v, st.CommBuf(), sss, sss, 1, 1, in_v, out_v);
});
}
};
/*Change ends */
/*******************************************************************************
* Conserved current utilities for Wilson fermions, for contracting propagators
* to make a conserved current sink or inserting the conserved current
* sequentially.
******************************************************************************/
template <class Impl>
void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
PropagatorField &q_in_2,
PropagatorField &q_out,
Current curr_type,
unsigned int mu)
{
Gamma g5(Gamma::Algebra::Gamma5);
conformable(_grid, q_in_1.Grid());
conformable(_grid, q_in_2.Grid());
conformable(_grid, q_out.Grid());
PropagatorField tmp1(_grid), tmp2(_grid);
q_out = Zero();
// Forward, need q1(x + mu), q2(x). Backward, need q1(x), q2(x + mu).
// Inefficient comms method but not performance critical.
tmp1 = Cshift(q_in_1, mu, 1);
tmp2 = Cshift(q_in_2, mu, 1);
auto tmp1_v = tmp1.View();
auto tmp2_v = tmp2.View();
auto q_in_1_v=q_in_1.View();
auto q_in_2_v=q_in_2.View();
auto q_out_v = q_out.View();
auto Umu_v = Umu.View();
thread_loop( (unsigned int sU = 0; sU < Umu.Grid()->oSites(); ++sU), {
Kernels::ContractConservedCurrentSiteFwd(tmp1_v[sU],
q_in_2_v[sU],
q_out_v[sU],
Umu_v, sU, mu);
Kernels::ContractConservedCurrentSiteBwd(q_in_1_v[sU],
tmp2_v[sU],
q_out_v[sU],
Umu_v, sU, mu);
});
}
template <class Impl>
void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
PropagatorField &q_out,
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx)
{
conformable(_grid, q_in.Grid());
conformable(_grid, q_out.Grid());
// Lattice<iSinglet<Simd>> ph(_grid), coor(_grid);
Complex i(0.0,1.0);
PropagatorField tmpFwd(_grid), tmpBwd(_grid), tmp(_grid);
unsigned int tshift = (mu == Tp) ? 1 : 0;
unsigned int LLt = GridDefaultLatt()[Tp];
q_out = Zero();
LatticeInteger coords(_grid);
LatticeCoordinate(coords, Tp);
// Need q(x + mu) and q(x - mu).
tmp = Cshift(q_in, mu, 1);
tmpFwd = tmp*lattice_cmplx;
tmp = lattice_cmplx*q_in;
tmpBwd = Cshift(tmp, mu, -1);
auto coords_v = coords.View();
auto tmpFwd_v = tmpFwd.View();
auto tmpBwd_v = tmpBwd.View();
auto Umu_v = Umu.View();
auto q_out_v = q_out.View();
thread_loop( (unsigned int sU = 0; sU < Umu.Grid()->oSites(); ++sU), {
// Compute the sequential conserved current insertion only if our simd
// object contains a timeslice we need.
vInteger t_mask = ((coords_v[sU] >= tmin) &&
(coords_v[sU] <= tmax));
Integer timeSlices = Reduce(t_mask);
if (timeSlices > 0) {
Kernels::SeqConservedCurrentSiteFwd(tmpFwd_v[sU],
q_out_v[sU],
Umu_v, sU, mu, t_mask);
}
// Repeat for backward direction.
t_mask = ((coords_v[sU] >= (tmin + tshift)) &&
(coords_v[sU] <= (tmax + tshift)));
//if tmax = LLt-1 (last timeslice) include timeslice 0 if the time is shifted (mu=3)
unsigned int t0 = 0;
if((tmax==LLt-1) && (tshift==1)) t_mask = (t_mask || (coords_v[sU] == t0 ));
timeSlices = Reduce(t_mask);
if (timeSlices > 0) {
Kernels::SeqConservedCurrentSiteBwd(tmpBwd_v[sU],
q_out_v[sU],
Umu_v, sU, mu, t_mask);
}
});
}
FermOpTemplateInstantiate(WilsonFermion);
AdjointFermOpTemplateInstantiate(WilsonFermion);
TwoIndexFermOpTemplateInstantiate(WilsonFermion);
GparityFermOpTemplateInstantiate(WilsonFermion);
NAMESPACE_END(Grid);

1119
Grid/qcd/action/fermion/WilsonFermion5D.cc
View File

File diff suppressed because it is too large Load Diff

4
Grid/qcd/action/fermion/WilsonFermion5D.h
View File

@ -227,8 +227,8 @@ public:
Current curr_type,
unsigned int mu,
unsigned int tmin,
unsigned int tmax,
ComplexField &lattice_cmplx);
unsigned int tmax,
ComplexField &lattice_cmplx);
};
NAMESPACE_END(Grid);

21
Grid/qcd/action/fermion/WilsonImpl.h
View File

@ -90,15 +90,7 @@ public:
auto UU = coalescedRead(U(mu));
mult(&phi(), &UU, &chi());
}
#ifdef GPU_VEC
static accelerator_inline void copyLinkGpu(int lane,
SiteDoubledGaugeField & UU,
const SiteDoubledGaugeField &U)
{
auto U_l = extractLane(lane,U);
insertLane(lane,UU,U_l);
}
static accelerator_inline void multLinkGpu(int lane,
typename SiteHalfSpinor::scalar_object &phi,
const SiteDoubledGaugeField &U,
@ -108,17 +100,6 @@ public:
auto U_l = extractLane(lane,U(mu));
phi() = U_l * chi();
}
#else
static accelerator_inline void multLinkGpu(int lane,
SiteHalfSpinor &phi,
const SiteDoubledGaugeField &U,
const SiteHalfSpinor &chi,
int mu)
{
auto U_l = U(mu);
phi() = U_l * chi();
}
#endif
static accelerator_inline void multLinkProp(SitePropagator &phi,
const SiteDoubledGaugeField &U,

445
Grid/qcd/action/fermion/WilsonKernels.cc
View File

@ -1,445 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
NAMESPACE_BEGIN(Grid);
int WilsonKernelsStatic::Opt = WilsonKernelsStatic::OptGeneric;
int WilsonKernelsStatic::Comms = WilsonKernelsStatic::CommsAndCompute;
////////////////////////////////////////////
// Generic implementation; move to different file?
////////////////////////////////////////////
#define GENERIC_STENCIL_LEG(Dir,spProj,Recon) \
SE = st.GetEntry(ptype, Dir, sF); \
if (SE->_is_local) { \
chi_p = &chi; \
if (SE->_permute) { \
spProj(tmp, in[SE->_offset]); \
permute(chi, tmp, ptype); \
} else { \
spProj(chi, in[SE->_offset]); \
} \
} else { \
chi_p = &buf[SE->_offset]; \
} \
Impl::multLink(Uchi, U[sU], *chi_p, Dir, SE, st); \
Recon(result, Uchi);
#define GENERIC_STENCIL_LEG_INT(Dir,spProj,Recon) \
SE = st.GetEntry(ptype, Dir, sF); \
if (SE->_is_local) { \
chi_p = &chi; \
if (SE->_permute) { \
spProj(tmp, in[SE->_offset]); \
permute(chi, tmp, ptype); \
} else { \
spProj(chi, in[SE->_offset]); \
} \
} else if ( st.same_node[Dir] ) { \
chi_p = &buf[SE->_offset]; \
} \
if (SE->_is_local || st.same_node[Dir] ) { \
Impl::multLink(Uchi, U[sU], *chi_p, Dir, SE, st); \
Recon(result, Uchi); \
}
#define GENERIC_STENCIL_LEG_EXT(Dir,spProj,Recon) \
SE = st.GetEntry(ptype, Dir, sF); \
if ((!SE->_is_local) && (!st.same_node[Dir]) ) { \
chi_p = &buf[SE->_offset]; \
Impl::multLink(Uchi, U[sU], *chi_p, Dir, SE, st); \
Recon(result, Uchi); \
nmu++; \
}
#define GENERIC_DHOPDIR_LEG(Dir,spProj,Recon) \
if (gamma == Dir) { \
if (SE->_is_local && SE->_permute) { \
spProj(tmp, in[SE->_offset]); \
permute(chi, tmp, ptype); \
} else if (SE->_is_local) { \
spProj(chi, in[SE->_offset]); \
} else { \
chi = buf[SE->_offset]; \
} \
Impl::multLink(Uchi, U[sU], chi, dir, SE, st); \
Recon(result, Uchi); \
}
////////////////////////////////////////////////////////////////////
// All legs kernels ; comms then compute
////////////////////////////////////////////////////////////////////
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
GENERIC_STENCIL_LEG(Xp,spProjXp,spReconXp);
GENERIC_STENCIL_LEG(Yp,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG(Zp,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG(Tp,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG(Xm,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG(Ym,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG(Zm,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG(Tm,spProjTm,accumReconTm);
vstream(out[sF], result);
};
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
GENERIC_STENCIL_LEG(Xm,spProjXp,spReconXp);
GENERIC_STENCIL_LEG(Ym,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG(Zm,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG(Tm,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG(Xp,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG(Yp,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG(Zp,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG(Tp,spProjTm,accumReconTm);
vstream(out[sF], result);
};
////////////////////////////////////////////////////////////////////
// Interior kernels
////////////////////////////////////////////////////////////////////
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
result=Zero();
GENERIC_STENCIL_LEG_INT(Xp,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_INT(Yp,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_INT(Zp,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_INT(Tp,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_INT(Xm,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_INT(Ym,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_INT(Zm,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_INT(Tm,spProjTm,accumReconTm);
vstream(out[sF], result);
};
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
result=Zero();
GENERIC_STENCIL_LEG_INT(Xm,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_INT(Ym,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_INT(Zm,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_INT(Tm,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_INT(Xp,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_INT(Yp,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_INT(Zp,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_INT(Tp,spProjTm,accumReconTm);
vstream(out[sF], result);
};
////////////////////////////////////////////////////////////////////
// Exterior kernels
////////////////////////////////////////////////////////////////////
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
// SiteHalfSpinor tmp;
// SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
int nmu=0;
result=Zero();
GENERIC_STENCIL_LEG_EXT(Xp,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_EXT(Yp,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_EXT(Zp,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_EXT(Tp,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_EXT(Xm,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_EXT(Ym,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_EXT(Zm,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_EXT(Tm,spProjTm,accumReconTm);
if ( nmu ) {
out[sF] = out[sF] + result;
}
};
template <class Impl>
void WilsonKernels<Impl>::GenericDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
// SiteHalfSpinor tmp;
// SiteHalfSpinor chi;
SiteHalfSpinor *chi_p;
SiteHalfSpinor Uchi;
SiteSpinor result;
StencilEntry *SE;
int ptype;
int nmu=0;
result=Zero();
GENERIC_STENCIL_LEG_EXT(Xm,spProjXp,accumReconXp);
GENERIC_STENCIL_LEG_EXT(Ym,spProjYp,accumReconYp);
GENERIC_STENCIL_LEG_EXT(Zm,spProjZp,accumReconZp);
GENERIC_STENCIL_LEG_EXT(Tm,spProjTp,accumReconTp);
GENERIC_STENCIL_LEG_EXT(Xp,spProjXm,accumReconXm);
GENERIC_STENCIL_LEG_EXT(Yp,spProjYm,accumReconYm);
GENERIC_STENCIL_LEG_EXT(Zp,spProjZm,accumReconZm);
GENERIC_STENCIL_LEG_EXT(Tp,spProjTm,accumReconTm);
if ( nmu ) {
out[sF] = out[sF] + result;
}
};
template <class Impl>
void WilsonKernels<Impl>::DhopDirK( StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, int sF,
int sU, const FermionFieldView &in, FermionFieldView &out, int dir, int gamma)
{
SiteHalfSpinor tmp;
SiteHalfSpinor chi;
SiteSpinor result;
SiteHalfSpinor Uchi;
StencilEntry *SE;
int ptype;
SE = st.GetEntry(ptype, dir, sF);
GENERIC_DHOPDIR_LEG(Xp,spProjXp,spReconXp);
GENERIC_DHOPDIR_LEG(Yp,spProjYp,spReconYp);
GENERIC_DHOPDIR_LEG(Zp,spProjZp,spReconZp);
GENERIC_DHOPDIR_LEG(Tp,spProjTp,spReconTp);
GENERIC_DHOPDIR_LEG(Xm,spProjXm,spReconXm);
GENERIC_DHOPDIR_LEG(Ym,spProjYm,spReconYm);
GENERIC_DHOPDIR_LEG(Zm,spProjZm,spReconZm);
GENERIC_DHOPDIR_LEG(Tm,spProjTm,spReconTm);
vstream(out[sF], result);
}
/*******************************************************************************
* Conserved current utilities for Wilson fermions, for contracting propagators
* to make a conserved current sink or inserting the conserved current
* sequentially. Common to both 4D and 5D.
******************************************************************************/
// N.B. Functions below assume a -1/2 factor within U.
#define WilsonCurrentFwd(expr, mu) ((expr - Gamma::gmu[mu]*expr))
#define WilsonCurrentBwd(expr, mu) ((expr + Gamma::gmu[mu]*expr))
/*******************************************************************************
* Name: ContractConservedCurrentSiteFwd
* Operation: (1/2) * q2[x] * U(x) * (g[mu] - 1) * q1[x + mu]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in_1 shifted in +ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::ContractConservedCurrentSiteFwd(
const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
bool switch_sign)
{
SitePropagator result, tmp;
Gamma g5(Gamma::Algebra::Gamma5);
Impl::multLinkProp(tmp, U[sU], q_in_1, mu);
result = g5 * adj(q_in_2) * g5 * WilsonCurrentFwd(tmp, mu);
if (switch_sign) {
q_out -= result;
} else {
q_out += result;
}
}
/*******************************************************************************
* Name: ContractConservedCurrentSiteBwd
* Operation: (1/2) * q2[x + mu] * U^dag(x) * (g[mu] + 1) * q1[x]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in_2 shifted in +ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::ContractConservedCurrentSiteBwd(
const SitePropagator &q_in_1,
const SitePropagator &q_in_2,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
bool switch_sign)
{
SitePropagator result, tmp;
Gamma g5(Gamma::Algebra::Gamma5);
Impl::multLinkProp(tmp, U[sU], q_in_1, mu + Nd);
result = g5 * adj(q_in_2) * g5 * WilsonCurrentBwd(tmp, mu);
if (switch_sign) {
q_out += result;
} else {
q_out -= result;
}
}
// G-parity requires more specialised implementation.
#define NO_CURR_SITE(Impl) \
template <> \
void WilsonKernels<Impl>::ContractConservedCurrentSiteFwd( \
const SitePropagator &q_in_1, \
const SitePropagator &q_in_2, \
SitePropagator &q_out, \
DoubledGaugeFieldView &U, \
unsigned int sU, \
unsigned int mu, \
bool switch_sign) \
{ \
assert(0); \
} \
template <> \
void WilsonKernels<Impl>::ContractConservedCurrentSiteBwd( \
const SitePropagator &q_in_1, \
const SitePropagator &q_in_2, \
SitePropagator &q_out, \
DoubledGaugeFieldView &U, \
unsigned int mu, \
unsigned int sU, \
bool switch_sign) \
{ \
assert(0); \
}
NO_CURR_SITE(GparityWilsonImplF);
NO_CURR_SITE(GparityWilsonImplD);
NO_CURR_SITE(GparityWilsonImplFH);
NO_CURR_SITE(GparityWilsonImplDF);
/*******************************************************************************
* Name: SeqConservedCurrentSiteFwd
* Operation: (1/2) * U(x) * (g[mu] - 1) * q[x + mu]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in shifted in +ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::SeqConservedCurrentSiteFwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
vInteger t_mask,
bool switch_sign)
{
SitePropagator result;
Impl::multLinkProp(result, U[sU], q_in, mu);
result = WilsonCurrentFwd(result, mu);
// Zero any unwanted timeslice entries.
result = predicatedWhere(t_mask, result, 0.*result);
if (switch_sign) {
q_out -= result;
} else {
q_out += result;
}
}
/*******************************************************************************
* Name: SeqConservedCurrentSiteFwd
* Operation: (1/2) * U^dag(x) * (g[mu] + 1) * q[x - mu]
* Notes: - DoubledGaugeField U assumed to contain -1/2 factor.
* - Pass in q_in shifted in -ve mu direction.
******************************************************************************/
template<class Impl>
void WilsonKernels<Impl>::SeqConservedCurrentSiteBwd(const SitePropagator &q_in,
SitePropagator &q_out,
DoubledGaugeFieldView &U,
unsigned int sU,
unsigned int mu,
vInteger t_mask,
bool switch_sign)
{
SitePropagator result;
Impl::multLinkProp(result, U[sU], q_in, mu + Nd);
result = WilsonCurrentBwd(result, mu);
// Zero any unwanted timeslice entries.
result = predicatedWhere(t_mask, result, 0.*result);
if (switch_sign) {
q_out += result;
} else {
q_out -= result;
}
}
FermOpTemplateInstantiate(WilsonKernels);
AdjointFermOpTemplateInstantiate(WilsonKernels);
TwoIndexFermOpTemplateInstantiate(WilsonKernels);
NAMESPACE_END(Grid);

2
Grid/qcd/action/fermion/WilsonKernels.h
View File

@ -107,7 +107,7 @@ private:
int Ls,int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
static accelerator void GenericDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
int sF, int sU, const FermionFieldView &in, FermionFieldView &out);
static accelerator void GenericDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor * buf,
int sF, int sU, const FermionFieldView &in, FermionFieldView &out);

125
Grid/qcd/action/fermion/WilsonKernelsAsm.cc
View File

@ -1,125 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernelsAsm.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////
// Default to no assembler implementation
///////////////////////////////////////////////////////////
template<class Impl> void
WilsonKernels<Impl >::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
{
assert(0);
}
template<class Impl> void
WilsonKernels<Impl >::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
{
assert(0);
}
template<class Impl> void
WilsonKernels<Impl >::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
{
assert(0);
}
template<class Impl> void
WilsonKernels<Impl >::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
{
assert(0);
}
template<class Impl> void
WilsonKernels<Impl >::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
{
assert(0);
}
template<class Impl> void
WilsonKernels<Impl >::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
{
assert(0);
}
#include <qcd/action/fermion/WilsonKernelsAsmAvx512.h>
#include <qcd/action/fermion/WilsonKernelsAsmQPX.h>
#define INSTANTIATE_ASM(A) \
template void WilsonKernels<A>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,\
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out);\
\
template void WilsonKernels<A>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,\
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out);\
template void WilsonKernels<A>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,\
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out);\
\
template void WilsonKernels<A>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,\
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out);\
template void WilsonKernels<A>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,\
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out);\
\
template void WilsonKernels<A>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,\
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out);\
//INSTANTIATE_ASM(WilsonImplF);
//INSTANTIATE_ASM(WilsonImplD);
INSTANTIATE_ASM(GparityWilsonImplF);
INSTANTIATE_ASM(GparityWilsonImplD);
//INSTANTIATE_ASM(ZWilsonImplF);
//INSTANTIATE_ASM(ZWilsonImplD);
//INSTANTIATE_ASM(DomainWallVec5dImplF);
//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);
NAMESPACE_END(Grid);

650
Grid/qcd/action/fermion/WilsonKernelsAsmAvx512.h
View File

@ -1,650 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernelsAsmAvx512.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#if defined(AVX512)
///////////////////////////////////////////////////////////
// If we are AVX512 specialise the single precision routine
///////////////////////////////////////////////////////////
#include <simd/Intel512wilson.h>
#include <simd/Intel512single.h>
static Vector<vComplexF> signsF;
template<typename vtype>
int setupSigns(Vector<vtype>& signs ){
Vector<vtype> bother(2);
signs = bother;
vrsign(signs[0]);
visign(signs[1]);
return 1;
}
static int signInitF = setupSigns(signsF);
#define MAYBEPERM(A,perm) if (perm) { A ; }
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
#define COMPLEX_SIGNS(isigns) vComplexF *isigns = &signsF[0];
/////////////////////////////////////////////////////////////////
// XYZT vectorised, undag Kernel, single
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
// XYZT vectorised, dag Kernel, single
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef MAYBEPERM
#undef MULT_2SPIN
#define MAYBEPERM(A,B)
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
/////////////////////////////////////////////////////////////////
// Ls vectorised, undag Kernel, single
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
#undef MULT_2SPIN
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LSNOPF(ptr,pf)
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
// Ls vectorised, dag Kernel, single
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplFH>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef COMPLEX_SIGNS
#undef MAYBEPERM
#undef MULT_2SPIN
///////////////////////////////////////////////////////////
// If we are AVX512 specialise the double precision routine
///////////////////////////////////////////////////////////
#include <simd/Intel512double.h>
static Vector<vComplexD> signsD;
static int signInitD = setupSigns(signsD);
#define MAYBEPERM(A,perm) if (perm) { A ; }
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
#define COMPLEX_SIGNS(isigns) vComplexD *isigns = &signsD[0];
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
/////////////////////////////////////////////////////////////////
// XYZT vectorised, undag Kernel, single
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
// XYZT vectorised, dag Kernel, single
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<WilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZWilsonImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef MAYBEPERM
#undef MULT_2SPIN
#define MAYBEPERM(A,B)
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
/////////////////////////////////////////////////////////////////
// Ls vectorised, undag Kernel, single
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSite(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
#undef MULT_2SPIN
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LSNOPF(ptr,pf)
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
// Ls vectorised, dag Kernel, single
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#define INTERIOR
#undef EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef INTERIOR_AND_EXTERIOR
#undef INTERIOR
#define EXTERIOR
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplD>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<DomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
template<> void
WilsonKernels<ZDomainWallVec5dImplDF>::AsmDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionFieldView &in, FermionFieldView &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef COMPLEX_SIGNS
#undef MAYBEPERM
#undef MULT_2SPIN
#endif //AVX512

198
Grid/qcd/action/fermion/WilsonKernelsAsmBody.h
View File

@ -1,198 +0,0 @@
#ifdef KERNEL_DAG
#define DIR0_PROJMEM(base) XP_PROJMEM(base);
#define DIR1_PROJMEM(base) YP_PROJMEM(base);
#define DIR2_PROJMEM(base) ZP_PROJMEM(base);
#define DIR3_PROJMEM(base) TP_PROJMEM(base);
#define DIR4_PROJMEM(base) XM_PROJMEM(base);
#define DIR5_PROJMEM(base) YM_PROJMEM(base);
#define DIR6_PROJMEM(base) ZM_PROJMEM(base);
#define DIR7_PROJMEM(base) TM_PROJMEM(base);
#define DIR0_RECON XP_RECON
#define DIR1_RECON YP_RECON_ACCUM
#define DIR2_RECON ZP_RECON_ACCUM
#define DIR3_RECON TP_RECON_ACCUM
#define DIR4_RECON XM_RECON_ACCUM
#define DIR5_RECON YM_RECON_ACCUM
#define DIR6_RECON ZM_RECON_ACCUM
#define DIR7_RECON TM_RECON_ACCUM
#else
#define DIR0_PROJMEM(base) XM_PROJMEM(base);
#define DIR1_PROJMEM(base) YM_PROJMEM(base);
#define DIR2_PROJMEM(base) ZM_PROJMEM(base);
#define DIR3_PROJMEM(base) TM_PROJMEM(base);
#define DIR4_PROJMEM(base) XP_PROJMEM(base);
#define DIR5_PROJMEM(base) YP_PROJMEM(base);
#define DIR6_PROJMEM(base) ZP_PROJMEM(base);
#define DIR7_PROJMEM(base) TP_PROJMEM(base);
#define DIR0_RECON XM_RECON
#define DIR1_RECON YM_RECON_ACCUM
#define DIR2_RECON ZM_RECON_ACCUM
#define DIR3_RECON TM_RECON_ACCUM
#define DIR4_RECON XP_RECON_ACCUM
#define DIR5_RECON YP_RECON_ACCUM
#define DIR6_RECON ZP_RECON_ACCUM
#define DIR7_RECON TP_RECON_ACCUM
#endif
////////////////////////////////////////////////////////////////////////////////
// Comms then compute kernel
////////////////////////////////////////////////////////////////////////////////
#ifdef INTERIOR_AND_EXTERIOR
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
basep = st.GetPFInfo(nent,plocal); nent++; \
if ( local ) { \
LOAD64(%r10,isigns); \
PROJ(base); \
MAYBEPERM(PERMUTE_DIR,perm); \
} else { \
LOAD_CHI(base); \
} \
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
PREFETCH_CHIMU(base); \
MULT_2SPIN_DIR_PF(Dir,basep); \
LOAD64(%r10,isigns); \
RECON; \
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
PF_GAUGE(Xp); \
PREFETCH1_CHIMU(base); \
ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
#define RESULT(base,basep) SAVE_RESULT(base,basep);
#endif
////////////////////////////////////////////////////////////////////////////////
// Pre comms kernel -- prefetch like normal because it is mostly right
////////////////////////////////////////////////////////////////////////////////
#ifdef INTERIOR
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
basep = st.GetPFInfo(nent,plocal); nent++; \
if ( local ) { \
LOAD64(%r10,isigns); \
PROJ(base); \
MAYBEPERM(PERMUTE_DIR,perm); \
}else if ( st.same_node[Dir] ) {LOAD_CHI(base);} \
if ( local || st.same_node[Dir] ) { \
MULT_2SPIN_DIR_PF(Dir,basep); \
LOAD64(%r10,isigns); \
RECON; \
} \
base = st.GetInfo(ptype,local,perm,NxtDir,ent,plocal); ent++; \
PREFETCH_CHIMU(base); \
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
PF_GAUGE(Xp); \
PREFETCH1_CHIMU(base); \
{ ZERO_PSI; } \
ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON)
#define RESULT(base,basep) SAVE_RESULT(base,basep);
#endif
////////////////////////////////////////////////////////////////////////////////
// Post comms kernel
////////////////////////////////////////////////////////////////////////////////
#ifdef EXTERIOR
#define ASM_LEG(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \
MULT_2SPIN_DIR_PF(Dir,base); \
LOAD64(%r10,isigns); \
RECON; \
nmu++; \
}
#define ASM_LEG_XP(Dir,NxtDir,PERMUTE_DIR,PROJ,RECON) \
nmu=0; \
{ ZERO_PSI;} \
base = st.GetInfo(ptype,local,perm,Dir,ent,plocal); ent++; \
if((!local)&&(!st.same_node[Dir]) ) { \
LOAD_CHI(base); \
MULT_2SPIN_DIR_PF(Dir,base); \
LOAD64(%r10,isigns); \
RECON; \
nmu++; \
}
#define RESULT(base,basep) if (nmu){ ADD_RESULT(base,base);}
#endif
{
int nmu;
int local,perm, ptype;
uint64_t base;
uint64_t basep;
const uint64_t plocal =(uint64_t) & in[0];
COMPLEX_SIGNS(isigns);
MASK_REGS;
int nmax=U.oSites();
for(int site=0;site<Ns;site++) {
#ifndef EXTERIOR
// int sU =lo.Reorder(ssU);
int sU =ssU;
int ssn=ssU+1; if(ssn>=nmax) ssn=0;
// int sUn=lo.Reorder(ssn);
int sUn=ssn;
LOCK_GAUGE(0);
#else
int sU =ssU;
int ssn=ssU+1; if(ssn>=nmax) ssn=0;
int sUn=ssn;
#endif
for(int s=0;s<Ls;s++) {
ss =sU*Ls+s;
ssn=sUn*Ls+s;
int ent=ss*8;// 2*Ndim
int nent=ssn*8;
ASM_LEG_XP(Xp,Yp,PERMUTE_DIR3,DIR0_PROJMEM,DIR0_RECON);
ASM_LEG(Yp,Zp,PERMUTE_DIR2,DIR1_PROJMEM,DIR1_RECON);
ASM_LEG(Zp,Tp,PERMUTE_DIR1,DIR2_PROJMEM,DIR2_RECON);
ASM_LEG(Tp,Xm,PERMUTE_DIR0,DIR3_PROJMEM,DIR3_RECON);
ASM_LEG(Xm,Ym,PERMUTE_DIR3,DIR4_PROJMEM,DIR4_RECON);
ASM_LEG(Ym,Zm,PERMUTE_DIR2,DIR5_PROJMEM,DIR5_RECON);
ASM_LEG(Zm,Tm,PERMUTE_DIR1,DIR6_PROJMEM,DIR6_RECON);
ASM_LEG(Tm,Xp,PERMUTE_DIR0,DIR7_PROJMEM,DIR7_RECON);
#ifdef EXTERIOR
if (nmu==0) break;
// if (nmu!=0) std::cout << "EXT "<<sU<<std::endl;
#endif
base = (uint64_t) &out[ss];
basep= st.GetPFInfo(nent,plocal); nent++;
RESULT(base,basep);
}
ssU++;
UNLOCK_GAUGE(0);
}
}
#undef DIR0_PROJMEM
#undef DIR1_PROJMEM
#undef DIR2_PROJMEM
#undef DIR3_PROJMEM
#undef DIR4_PROJMEM
#undef DIR5_PROJMEM
#undef DIR6_PROJMEM
#undef DIR7_PROJMEM
#undef DIR0_RECON
#undef DIR1_RECON
#undef DIR2_RECON
#undef DIR3_RECON
#undef DIR4_RECON
#undef DIR5_RECON
#undef DIR6_RECON
#undef DIR7_RECON
#undef ASM_LEG
#undef ASM_LEG_XP
#undef RESULT

161
Grid/qcd/action/fermion/WilsonKernelsAsmBody.h.ab
View File

@ -1,161 +0,0 @@
{
int locala,perma, ptypea;
int localb,permb, ptypeb;
uint64_t basea, baseb;
const uint64_t plocal =(uint64_t) & in._odata[0];
// vComplexF isigns[2] = { signs[0], signs[1] };
vComplexF *isigns = &signs[0];
MASK_REGS;
for(int site=0;site<Ns;site++) {
int sU=lo.Reorder(ssU);
for(int s=0;s<Ls;s++) {
ss=sU*Ls+s;
////////////////////////////////
// Xp
////////////////////////////////
int ent=ss*8;// 2*Ndim
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns);
XM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR3,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFXP(Xp,baseb);
}
LOAD64(%r10,isigns);
XM_RECON;
////////////////////////////////
// Yp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Zp,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR2,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFYP(Yp,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_RECON_ACCUM;
////////////////////////////////
// Zp
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Tp,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR1,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFZP(Zp,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_RECON_ACCUM;
////////////////////////////////
// Tp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xm,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR0,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFTP(Tp,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_RECON_ACCUM;
////////////////////////////////
// Xm
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Ym,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR3,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFXM(Xm,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_RECON_ACCUM;
////////////////////////////////
// Ym
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Zm,ent,plocal); ent++;
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR2,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFYM(Ym,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_RECON_ACCUM;
////////////////////////////////
// Zm
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Tm,ent,plocal); ent++;
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR1,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFZM(Zm,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_RECON_ACCUM;
////////////////////////////////
// Tm
////////////////////////////////
basea = (uint64_t)&out._odata[ss];
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR0,permb);
} else {
LOAD_CHI(baseb);
}
baseb = st.GetInfo(ptypeb,localb,permb,Xp,ent,plocal);
{
MULT_2SPIN_DIR_PFTM(Tm,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_RECON_ACCUM;
SAVE_RESULT(&out._odata[ss],baseb);
}
ssU++;
}
}

187
Grid/qcd/action/fermion/WilsonKernelsAsmBody.h.abc
View File

@ -1,187 +0,0 @@
{
int locala,perma, ptypea;
int localb,permb, ptypeb;
int localc,permc, ptypec;
uint64_t basea, baseb, basec;
uint64_t basex;
const uint64_t plocal =(uint64_t) & in._odata[0];
// vComplexF isigns[2] = { signs[0], signs[1] };
vComplexF *isigns = &signs[0];
MASK_REGS;
for(int site=0;site<Ns;site++) {
int sU=lo.Reorder(ssU);
for(int s=0;s<Ls;s++) {
ss =sU*Ls+s;
////////////////////////////////
// Xp
////////////////////////////////
int ent=ss*8;// 2*Ndim
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(basea);
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(baseb);
basec = st.GetInfo(ptypec,localc,permc,Zp,ent,plocal); ent++;
PREFETCH_CHIMU(basec);
basex = basea;
label(FX(XP) );
if ( locala ) {
LOAD64(%r10,isigns);
XM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR3,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFXP(Xp,baseb);
}
LOAD64(%r10,isigns);
XM_RECON;
////////////////////////////////
// Yp
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(basea);
label(FX(YP) );
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR2,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFYP(Yp,basec);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YM_RECON_ACCUM;
////////////////////////////////
// Zp
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(baseb);
label(FX(ZP) );
if ( localc ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_PROJMEM(basec);
MAYBEPERM(PERMUTE_DIR1,permc);
} else {
LOAD_CHI(basec);
}
{
MULT_2SPIN_DIR_PFZP(Zp,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZM_RECON_ACCUM;
////////////////////////////////
// Tp
////////////////////////////////
basec = st.GetInfo(ptypec,localc,permc,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(basec);
label(FX(TP) );
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR0,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFTP(Tp,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TM_RECON_ACCUM;
////////////////////////////////
// Xm
////////////////////////////////
basea = st.GetInfo(ptypea,locala,perma,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(basea);
label(FX(XM) );
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR3,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFXM(Xm,basec);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
XP_RECON_ACCUM;
////////////////////////////////
// Ym
////////////////////////////////
baseb = st.GetInfo(ptypeb,localb,permb,Xp,ent,plocal); ent++;
PREFETCH_CHIMU(baseb);
label(FX(YM) );
if ( localc ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_PROJMEM(basec);
MAYBEPERM(PERMUTE_DIR2,permc);
} else {
LOAD_CHI(basec);
}
{
MULT_2SPIN_DIR_PFYM(Ym,basea);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
YP_RECON_ACCUM;
////////////////////////////////
// Zm
////////////////////////////////
basec = st.GetInfo(ptypec,localc,permc,Yp,ent,plocal); ent++;
PREFETCH_CHIMU(basec);
label(FX(ZM) );
if ( locala ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_PROJMEM(basea);
MAYBEPERM(PERMUTE_DIR1,perma);
} else {
LOAD_CHI(basea);
}
{
MULT_2SPIN_DIR_PFZM(Zm,baseb);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
ZP_RECON_ACCUM;
////////////////////////////////
// Tm
////////////////////////////////
basea = (uint64_t)&out._odata[ss];
PREFETCH_CHIMU(basea);
label(FX(TM) );
if ( localb ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_PROJMEM(baseb);
MAYBEPERM(PERMUTE_DIR0,permb);
} else {
LOAD_CHI(baseb);
}
{
MULT_2SPIN_DIR_PFTM(Tm,basec);
}
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
TP_RECON_ACCUM;
// PREFETCH_CHIMU(basex);
label(FX(SAV) );
SAVE_RESULT(&out._odata[ss]);
}
ssU++;
}
}

150
Grid/qcd/action/fermion/WilsonKernelsAsmQPX.h
View File

@ -1,150 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernelsAsmQPX.h
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#if defined(QPX)
///////////////////////////////////////////////////////////
// If we are QPX specialise the single precision routine
///////////////////////////////////////////////////////////
#include <simd/IBM_qpx.h>
#include <simd/IBM_qpx_single.h>
#define MAYBEPERM(A,perm) if (perm) { A ; }
#define MULT_2SPIN(ptr,pf) MULT_2SPIN_QPX(ptr,pf)
#define COMPLEX_SIGNS(isigns)
#define INTERIOR_AND_EXTERIOR
#undef INTERIOR
#undef EXTERIOR
/////////////////////////////////////////////////////////////////
// XYZT vectorised, undag Kernel, single
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSite(StencilView &st, DoubledGaugeField &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
// XYZT vectorised, dag Kernel, single
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
template<> void
WilsonKernels<WilsonImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef MAYBEPERM
#undef MULT_2SPIN
#define MAYBEPERM(A,B)
#define MULT_2SPIN(ptr,pf) MULT_2SPIN_QPX_LS(ptr,pf)
/////////////////////////////////////////////////////////////////
// Ls vectorised, undag Kernel, single
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSite(StencilView &st, DoubledGaugeField &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
// Ls vectorised, dag Kernel, single
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
template<> void
WilsonKernels<DomainWallVec5dImplF>::AsmDhopSiteDag(StencilView &st, DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef MAYBEPERM
#undef MULT_2SPIN
///////////////////////////////////////////////////////////
// DP routines
///////////////////////////////////////////////////////////
#include <simd/IBM_qpx_double.h>
#define MAYBEPERM(A,perm) if (perm) { A ; }
#define MULT_2SPIN(ptr,pf) MULT_2SPIN_QPX(ptr,pf)
/////////////////////////////////////////////////////////////////
// XYZT Vectorised, undag Kernel, double
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSite(StencilView &st, DoubledGaugeField &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////////
// XYZT Vectorised, dag Kernel, double
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
template<> void
WilsonKernels<WilsonImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
#undef MAYBEPERM
#undef MULT_2SPIN
#define MAYBEPERM(A,B)
#define MULT_2SPIN(ptr,pf) MULT_2SPIN_QPX_LS(ptr,pf)
/////////////////////////////////////////////////////////////////
// Ls vectorised, undag Kernel, double
/////////////////////////////////////////////////////////////////
#undef KERNEL_DAG
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSite(StencilView &st, DoubledGaugeField &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
/////////////////////////////////////////////////////////////////
// Ls vectorised, dag Kernel, double
/////////////////////////////////////////////////////////////////
#define KERNEL_DAG
template<> void
WilsonKernels<DomainWallVec5dImplD>::AsmDhopSiteDag(StencilView &st, DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
/////////////////////////////////////////////////////////////////
#undef MAYBEPERM
#undef MULT_2SPIN
#endif

378
Grid/qcd/action/fermion/WilsonKernelsGpu.cc
View File

@ -1,378 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernelsGpu.cc
Copyright (C) 2018
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
NAMESPACE_BEGIN(Grid);
//////////////////////////////////////////////////////////////
// Gpu implementation; thread loop is implicit ; move to header
//////////////////////////////////////////////////////////////
accelerator_inline void synchronise(void)
{
#ifdef __CUDA_ARCH__
__syncthreads();
#endif
return;
}
accelerator_inline int get_my_lanes(int Nsimd)
{
#ifdef __CUDA_ARCH__
return 1;
#else
return Nsimd;
#endif
}
accelerator_inline int get_my_lane_offset(int Nsimd)
{
#ifdef __CUDA_ARCH__
return ( (threadIdx.x) % Nsimd);
#else
return 0;
#endif
}
accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
{
#ifdef __CUDA_ARCH__
static_assert(sizeof(StencilEntry)==sizeof(uint4),"Unexpected Stencil Entry Size");
uint4 * mem_pun = (uint4 *)mem; // force 128 bit loads
uint4 * chip_pun = (uint4 *)&chip;
* chip_pun = * mem_pun;
#else
chip = *mem;
#endif
return;
}
#ifdef GPU_VEC
#if 1
#define GPU_COALESCED_STENCIL_LEG_PROJ(Dir,spProj) \
if (SE._is_local) { \
int mask = Nsimd >> (ptype + 1); \
int plane= SE._permute ? (lane ^ mask) : lane; \
auto in_l = extractLane(plane,in[SE._offset+s]); \
spProj(chi,in_l); \
} else { \
chi = extractLane(lane,buf[SE._offset+s]); \
} \
synchronise();
#else
#define GPU_COALESCED_STENCIL_LEG_PROJ(Dir,spProj) \
{ int mask = Nsimd >> (ptype + 1); \
int plane= SE._permute ? (lane ^ mask) : lane; \
auto in_l = extractLane(plane,in[SE._offset+s]); \
spProj(chi,in_l); }
#endif
#else
#define GPU_COALESCED_STENCIL_LEG_PROJ(Dir,spProj) \
if (SE._is_local) { \
auto in_t = in[SE._offset+s]; \
if (SE._permute) { \
spProj(tmp, in_t); \
permute(chi, tmp, ptype); \
} else { \
spProj(chi, in_t); \
} \
} else { \
chi = buf[SE._offset+s]; \
} \
synchronise();
#endif
template <class Impl>
accelerator_inline void WilsonKernels<Impl>::GpuDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,
SiteHalfSpinor *buf, int Ls, int s,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
#ifdef GPU_VEC
typename SiteHalfSpinor::scalar_object chi;
typename SiteHalfSpinor::scalar_object Uchi;
typename SiteSpinor::scalar_object result;
#else
SiteHalfSpinor chi;
SiteHalfSpinor Uchi;
SiteHalfSpinor tmp;
SiteSpinor result;
#endif
typedef typename SiteSpinor::scalar_type scalar_type;
typedef typename SiteSpinor::vector_type vector_type;
constexpr int Nsimd = sizeof(vector_type)/sizeof(scalar_type);
uint64_t lane_offset= get_my_lane_offset(Nsimd);
uint64_t lanes = get_my_lanes(Nsimd);
StencilEntry *SE_mem;
StencilEntry SE;
int ptype;
uint64_t ssF = Ls * sU;
uint64_t sF = ssF + s;
#ifndef __CUDA_ARCH__
for(int lane = lane_offset;lane<lane_offset+lanes;lane++){
#else
int lane = lane_offset; {
#endif
SE_mem = st.GetEntry(ptype, Xp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Xp,spProjXp);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Xp);
spReconXp(result, Uchi);
SE_mem = st.GetEntry(ptype, Yp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Yp,spProjYp);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Yp);
accumReconYp(result, Uchi);
SE_mem = st.GetEntry(ptype, Zp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Zp,spProjZp);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Zp);
accumReconZp(result, Uchi);
SE_mem = st.GetEntry(ptype, Tp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Tp,spProjTp);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Tp);
accumReconTp(result, Uchi);
SE_mem = st.GetEntry(ptype, Xm, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Xm,spProjXm);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Xm);
accumReconXm(result, Uchi);
SE_mem = st.GetEntry(ptype, Ym, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Ym,spProjYm);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Ym);
accumReconYm(result, Uchi);
SE_mem = st.GetEntry(ptype, Zm, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Zm,spProjZm);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Zm);
accumReconZm(result, Uchi);
SE_mem = st.GetEntry(ptype, Tm, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Tm,spProjTm);
Impl::multLinkGpu(lane,Uchi,U[sU],chi,Tm);
accumReconTm(result, Uchi);
#ifdef GPU_VEC
insertLane (lane,out[sF],result);
#else
vstream(out[sF], result);
#endif
}
}
template <class Impl>
accelerator_inline void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, SiteDoubledGaugeField &U,
SiteHalfSpinor *buf, int Ls, int s,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
#ifdef GPU_VEC
typename SiteHalfSpinor::scalar_object chi;
typename SiteHalfSpinor::scalar_object Uchi;
typename SiteSpinor::scalar_object result;
#else
SiteHalfSpinor chi;
SiteHalfSpinor Uchi;
SiteHalfSpinor tmp;
SiteSpinor result;
#endif
typedef typename SiteSpinor::scalar_type scalar_type;
typedef typename SiteSpinor::vector_type vector_type;
constexpr int Nsimd = sizeof(vector_type)/sizeof(scalar_type);
uint64_t lane_offset= get_my_lane_offset(Nsimd);
uint64_t lanes = get_my_lanes(Nsimd);
// printf (" sU %d s %d Nsimd %d lanes %ld lane_off %ld\n",sU, s, Nsimd, lanes, lane_offset);
StencilEntry *SE_mem;
StencilEntry SE;
int ptype;
// Forces some degree of coalesce on the table look ups
// Could also use wide load instructions on the data structure
uint64_t ssF = Ls * sU;
uint64_t sF = ssF + s;
#ifndef __CUDA_ARCH__
for(int lane = lane_offset;lane<lane_offset+lanes;lane++){
#else
int lane = lane_offset; {
#endif
SE_mem = st.GetEntry(ptype, Xp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Xp,spProjXm);
Impl::multLinkGpu(lane,Uchi,U,chi,Xp);
spReconXm(result, Uchi);
SE_mem = st.GetEntry(ptype, Yp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Yp,spProjYm);
Impl::multLinkGpu(lane,Uchi,U,chi,Yp);
accumReconYm(result, Uchi);
SE_mem = st.GetEntry(ptype, Zp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Zp,spProjZm);
Impl::multLinkGpu(lane,Uchi,U,chi,Zp);
accumReconZm(result, Uchi);
SE_mem = st.GetEntry(ptype, Tp, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Tp,spProjTm);
Impl::multLinkGpu(lane,Uchi,U,chi,Tp);
accumReconTm(result, Uchi);
SE_mem = st.GetEntry(ptype, Xm, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Xm,spProjXp);
Impl::multLinkGpu(lane,Uchi,U,chi,Xm);
accumReconXp(result, Uchi);
SE_mem = st.GetEntry(ptype, Ym, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Ym,spProjYp);
Impl::multLinkGpu(lane,Uchi,U,chi,Ym);
accumReconYp(result, Uchi);
SE_mem = st.GetEntry(ptype, Zm, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Zm,spProjZp);
Impl::multLinkGpu(lane,Uchi,U,chi,Zm);
accumReconZp(result, Uchi);
SE_mem = st.GetEntry(ptype, Tm, ssF); get_stencil(SE_mem,SE);
GPU_COALESCED_STENCIL_LEG_PROJ(Tm,spProjTp);
Impl::multLinkGpu(lane,Uchi,U,chi,Tm);
accumReconTp(result, Uchi);
#ifdef GPU_VEC
insertLane (lane,out[sF],result);
#else
vstream(out[sF], result);
#endif
}
};
// Template specialise Gparity to empty for now
#define GPU_EMPTY(A) \
template <> \
accelerator_inline void \
WilsonKernels<A>::GpuDhopSite(StencilView &st, \
SiteDoubledGaugeField &U, \
SiteHalfSpinor *buf, int Ls, int sF, \
int sU, \
const FermionFieldView &in, \
FermionFieldView &out) { assert(0);}; \
template <> \
accelerator_inline void \
WilsonKernels<A>::GpuDhopSiteDag(StencilView &st, \
DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, int Ls,int sF, \
int sU, \
const FermionFieldView &in, \
FermionFieldView &out) { assert(0);};
GPU_EMPTY(GparityWilsonImplF);
GPU_EMPTY(GparityWilsonImplFH);
GPU_EMPTY(GparityWilsonImplD);
GPU_EMPTY(GparityWilsonImplDF);
template <class Impl>
void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
int interior,int exterior)
{
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
if ( (Opt == WilsonKernelsStatic::OptGpu) && interior && exterior ) {
const uint64_t nsimd = Simd::Nsimd();
const uint64_t NN = Nsite*Ls*nsimd;
accelerator_loopN( sss, NN, {
uint64_t cur = sss;
// uint64_t lane = cur % nsimd;
cur = cur / nsimd;
uint64_t s = cur%Ls;
// uint64_t sF = cur;
cur = cur / Ls;
uint64_t sU = cur;
WilsonKernels<Impl>::GpuDhopSite(st_v,U_v[sU],buf,Ls,s,sU,in_v,out_v);
});
} else {
accelerator_loop( ss, U_v, {
int sU = ss;
int sF = Ls * sU;
WilsonKernels<Impl>::DhopSite(Opt,st_v,U_v,st.CommBuf(),sF,sU,Ls,1,in_v,out_v);
});
}
}
template <class Impl>
void WilsonKernels<Impl>::DhopDagKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
int interior,int exterior)
{
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
if ( (Opt == WilsonKernelsStatic::OptGpu) && interior && exterior ) {
const uint64_t nsimd = Simd::Nsimd();
const uint64_t NN = Nsite*Ls*nsimd;
accelerator_loopN( sss, NN, {
uint64_t cur = sss;
// uint64_t lane = cur % nsimd;
cur = cur / nsimd;
uint64_t s = cur%Ls;
//uint64_t sF = cur;
cur = cur / Ls;
uint64_t sU = cur;
WilsonKernels<Impl>::GpuDhopSiteDag(st_v,U_v,buf,Ls,s,sU,in_v,out_v);
});
} else {
accelerator_loop( ss, U_v, {
int sU = ss;
int sF = Ls * sU;
WilsonKernels<Impl>::DhopSiteDag(Opt,st,U_v,st.CommBuf(),sF,sU,Ls,1,in_v,out_v);
});
}
}
/*
GPU_EMPTY(DomainWallVec5dImplF);
GPU_EMPTY(DomainWallVec5dImplFH);
GPU_EMPTY(DomainWallVec5dImplD);
GPU_EMPTY(DomainWallVec5dImplDF);
GPU_EMPTY(ZDomainWallVec5dImplF);
GPU_EMPTY(ZDomainWallVec5dImplFH);
GPU_EMPTY(ZDomainWallVec5dImplD);
GPU_EMPTY(ZDomainWallVec5dImplDF);
*/
FermOpTemplateInstantiate(WilsonKernels);
AdjointFermOpTemplateInstantiate(WilsonKernels);
TwoIndexFermOpTemplateInstantiate(WilsonKernels);
NAMESPACE_END(Grid);

654
Grid/qcd/action/fermion/WilsonKernelsHand.cc
View File

@ -1,654 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernelsHand.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#define REGISTER
#define LOAD_CHIMU \
{const SiteSpinor & ref (in[offset]); \
Chimu_00=ref()(0)(0);\
Chimu_01=ref()(0)(1);\
Chimu_02=ref()(0)(2);\
Chimu_10=ref()(1)(0);\
Chimu_11=ref()(1)(1);\
Chimu_12=ref()(1)(2);\
Chimu_20=ref()(2)(0);\
Chimu_21=ref()(2)(1);\
Chimu_22=ref()(2)(2);\
Chimu_30=ref()(3)(0);\
Chimu_31=ref()(3)(1);\
Chimu_32=ref()(3)(2);}
#define LOAD_CHI\
{const SiteHalfSpinor &ref(buf[offset]); \
Chi_00 = ref()(0)(0);\
Chi_01 = ref()(0)(1);\
Chi_02 = ref()(0)(2);\
Chi_10 = ref()(1)(0);\
Chi_11 = ref()(1)(1);\
Chi_12 = ref()(1)(2);}
// To splat or not to splat depends on the implementation
#define MULT_2SPIN(A)\
{auto & ref(U[sU](A)); \
Impl::loadLinkElement(U_00,ref()(0,0)); \
Impl::loadLinkElement(U_10,ref()(1,0)); \
Impl::loadLinkElement(U_20,ref()(2,0)); \
Impl::loadLinkElement(U_01,ref()(0,1)); \
Impl::loadLinkElement(U_11,ref()(1,1)); \
Impl::loadLinkElement(U_21,ref()(2,1)); \
UChi_00 = U_00*Chi_00;\
UChi_10 = U_00*Chi_10;\
UChi_01 = U_10*Chi_00;\
UChi_11 = U_10*Chi_10;\
UChi_02 = U_20*Chi_00;\
UChi_12 = U_20*Chi_10;\
UChi_00+= U_01*Chi_01;\
UChi_10+= U_01*Chi_11;\
UChi_01+= U_11*Chi_01;\
UChi_11+= U_11*Chi_11;\
UChi_02+= U_21*Chi_01;\
UChi_12+= U_21*Chi_11;\
Impl::loadLinkElement(U_00,ref()(0,2)); \
Impl::loadLinkElement(U_10,ref()(1,2)); \
Impl::loadLinkElement(U_20,ref()(2,2)); \
UChi_00+= U_00*Chi_02;\
UChi_10+= U_00*Chi_12;\
UChi_01+= U_10*Chi_02;\
UChi_11+= U_10*Chi_12;\
UChi_02+= U_20*Chi_02;\
UChi_12+= U_20*Chi_12;}
#define PERMUTE_DIR(dir) \
permute##dir(Chi_00,Chi_00);\
permute##dir(Chi_01,Chi_01);\
permute##dir(Chi_02,Chi_02);\
permute##dir(Chi_10,Chi_10);\
permute##dir(Chi_11,Chi_11);\
permute##dir(Chi_12,Chi_12);
// hspin(0)=fspin(0)+timesI(fspin(3));
// hspin(1)=fspin(1)+timesI(fspin(2));
#define XP_PROJ \
Chi_00 = Chimu_00+timesI(Chimu_30);\
Chi_01 = Chimu_01+timesI(Chimu_31);\
Chi_02 = Chimu_02+timesI(Chimu_32);\
Chi_10 = Chimu_10+timesI(Chimu_20);\
Chi_11 = Chimu_11+timesI(Chimu_21);\
Chi_12 = Chimu_12+timesI(Chimu_22);
#define YP_PROJ \
Chi_00 = Chimu_00-Chimu_30;\
Chi_01 = Chimu_01-Chimu_31;\
Chi_02 = Chimu_02-Chimu_32;\
Chi_10 = Chimu_10+Chimu_20;\
Chi_11 = Chimu_11+Chimu_21;\
Chi_12 = Chimu_12+Chimu_22;
#define ZP_PROJ \
Chi_00 = Chimu_00+timesI(Chimu_20); \
Chi_01 = Chimu_01+timesI(Chimu_21); \
Chi_02 = Chimu_02+timesI(Chimu_22); \
Chi_10 = Chimu_10-timesI(Chimu_30); \
Chi_11 = Chimu_11-timesI(Chimu_31); \
Chi_12 = Chimu_12-timesI(Chimu_32);
#define TP_PROJ \
Chi_00 = Chimu_00+Chimu_20; \
Chi_01 = Chimu_01+Chimu_21; \
Chi_02 = Chimu_02+Chimu_22; \
Chi_10 = Chimu_10+Chimu_30; \
Chi_11 = Chimu_11+Chimu_31; \
Chi_12 = Chimu_12+Chimu_32;
// hspin(0)=fspin(0)-timesI(fspin(3));
// hspin(1)=fspin(1)-timesI(fspin(2));
#define XM_PROJ \
Chi_00 = Chimu_00-timesI(Chimu_30);\
Chi_01 = Chimu_01-timesI(Chimu_31);\
Chi_02 = Chimu_02-timesI(Chimu_32);\
Chi_10 = Chimu_10-timesI(Chimu_20);\
Chi_11 = Chimu_11-timesI(Chimu_21);\
Chi_12 = Chimu_12-timesI(Chimu_22);
#define YM_PROJ \
Chi_00 = Chimu_00+Chimu_30;\
Chi_01 = Chimu_01+Chimu_31;\
Chi_02 = Chimu_02+Chimu_32;\
Chi_10 = Chimu_10-Chimu_20;\
Chi_11 = Chimu_11-Chimu_21;\
Chi_12 = Chimu_12-Chimu_22;
#define ZM_PROJ \
Chi_00 = Chimu_00-timesI(Chimu_20); \
Chi_01 = Chimu_01-timesI(Chimu_21); \
Chi_02 = Chimu_02-timesI(Chimu_22); \
Chi_10 = Chimu_10+timesI(Chimu_30); \
Chi_11 = Chimu_11+timesI(Chimu_31); \
Chi_12 = Chimu_12+timesI(Chimu_32);
#define TM_PROJ \
Chi_00 = Chimu_00-Chimu_20; \
Chi_01 = Chimu_01-Chimu_21; \
Chi_02 = Chimu_02-Chimu_22; \
Chi_10 = Chimu_10-Chimu_30; \
Chi_11 = Chimu_11-Chimu_31; \
Chi_12 = Chimu_12-Chimu_32;
// fspin(0)=hspin(0);
// fspin(1)=hspin(1);
// fspin(2)=timesMinusI(hspin(1));
// fspin(3)=timesMinusI(hspin(0));
#define XP_RECON\
result_00 = UChi_00;\
result_01 = UChi_01;\
result_02 = UChi_02;\
result_10 = UChi_10;\
result_11 = UChi_11;\
result_12 = UChi_12;\
result_20 = timesMinusI(UChi_10);\
result_21 = timesMinusI(UChi_11);\
result_22 = timesMinusI(UChi_12);\
result_30 = timesMinusI(UChi_00);\
result_31 = timesMinusI(UChi_01);\
result_32 = timesMinusI(UChi_02);
#define XP_RECON_ACCUM\
result_00+=UChi_00;\
result_01+=UChi_01;\
result_02+=UChi_02;\
result_10+=UChi_10;\
result_11+=UChi_11;\
result_12+=UChi_12;\
result_20-=timesI(UChi_10);\
result_21-=timesI(UChi_11);\
result_22-=timesI(UChi_12);\
result_30-=timesI(UChi_00);\
result_31-=timesI(UChi_01);\
result_32-=timesI(UChi_02);
#define XM_RECON\
result_00 = UChi_00;\
result_01 = UChi_01;\
result_02 = UChi_02;\
result_10 = UChi_10;\
result_11 = UChi_11;\
result_12 = UChi_12;\
result_20 = timesI(UChi_10);\
result_21 = timesI(UChi_11);\
result_22 = timesI(UChi_12);\
result_30 = timesI(UChi_00);\
result_31 = timesI(UChi_01);\
result_32 = timesI(UChi_02);
#define XM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= timesI(UChi_10);\
result_21+= timesI(UChi_11);\
result_22+= timesI(UChi_12);\
result_30+= timesI(UChi_00);\
result_31+= timesI(UChi_01);\
result_32+= timesI(UChi_02);
#define YP_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= UChi_10;\
result_21+= UChi_11;\
result_22+= UChi_12;\
result_30-= UChi_00;\
result_31-= UChi_01;\
result_32-= UChi_02;
#define YM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20-= UChi_10;\
result_21-= UChi_11;\
result_22-= UChi_12;\
result_30+= UChi_00;\
result_31+= UChi_01;\
result_32+= UChi_02;
#define ZP_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20-= timesI(UChi_00); \
result_21-= timesI(UChi_01); \
result_22-= timesI(UChi_02); \
result_30+= timesI(UChi_10); \
result_31+= timesI(UChi_11); \
result_32+= timesI(UChi_12);
#define ZM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= timesI(UChi_00); \
result_21+= timesI(UChi_01); \
result_22+= timesI(UChi_02); \
result_30-= timesI(UChi_10); \
result_31-= timesI(UChi_11); \
result_32-= timesI(UChi_12);
#define TP_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= UChi_00; \
result_21+= UChi_01; \
result_22+= UChi_02; \
result_30+= UChi_10; \
result_31+= UChi_11; \
result_32+= UChi_12;
#define TM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20-= UChi_00; \
result_21-= UChi_01; \
result_22-= UChi_02; \
result_30-= UChi_10; \
result_31-= UChi_11; \
result_32-= UChi_12;
#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU; \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
LOAD_CHI; \
} \
MULT_2SPIN(DIR); \
RECON;
#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU; \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else if ( st.same_node[DIR] ) { \
LOAD_CHI; \
} \
if (local || st.same_node[DIR] ) { \
MULT_2SPIN(DIR); \
RECON; \
}
#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
if((!SE->_is_local)&&(!st.same_node[DIR]) ) { \
LOAD_CHI; \
MULT_2SPIN(DIR); \
RECON; \
nmu++; \
}
#define HAND_RESULT(ss) \
{ \
SiteSpinor & ref (out[ss]); \
vstream(ref()(0)(0),result_00); \
vstream(ref()(0)(1),result_01); \
vstream(ref()(0)(2),result_02); \
vstream(ref()(1)(0),result_10); \
vstream(ref()(1)(1),result_11); \
vstream(ref()(1)(2),result_12); \
vstream(ref()(2)(0),result_20); \
vstream(ref()(2)(1),result_21); \
vstream(ref()(2)(2),result_22); \
vstream(ref()(3)(0),result_30); \
vstream(ref()(3)(1),result_31); \
vstream(ref()(3)(2),result_32); \
}
#define HAND_RESULT_EXT(ss) \
if (nmu){ \
SiteSpinor & ref (out[ss]); \
ref()(0)(0)+=result_00; \
ref()(0)(1)+=result_01; \
ref()(0)(2)+=result_02; \
ref()(1)(0)+=result_10; \
ref()(1)(1)+=result_11; \
ref()(1)(2)+=result_12; \
ref()(2)(0)+=result_20; \
ref()(2)(1)+=result_21; \
ref()(2)(2)+=result_22; \
ref()(3)(0)+=result_30; \
ref()(3)(1)+=result_31; \
ref()(3)(2)+=result_32; \
}
#define HAND_DECLARATIONS(a) \
Simd result_00; \
Simd result_01; \
Simd result_02; \
Simd result_10; \
Simd result_11; \
Simd result_12; \
Simd result_20; \
Simd result_21; \
Simd result_22; \
Simd result_30; \
Simd result_31; \
Simd result_32; \
Simd Chi_00; \
Simd Chi_01; \
Simd Chi_02; \
Simd Chi_10; \
Simd Chi_11; \
Simd Chi_12; \
Simd UChi_00; \
Simd UChi_01; \
Simd UChi_02; \
Simd UChi_10; \
Simd UChi_11; \
Simd UChi_12; \
Simd U_00; \
Simd U_10; \
Simd U_20; \
Simd U_01; \
Simd U_11; \
Simd U_21;
#define ZERO_RESULT \
result_00=Zero(); \
result_01=Zero(); \
result_02=Zero(); \
result_10=Zero(); \
result_11=Zero(); \
result_12=Zero(); \
result_20=Zero(); \
result_21=Zero(); \
result_22=Zero(); \
result_30=Zero(); \
result_31=Zero(); \
result_32=Zero();
#define Chimu_00 Chi_00
#define Chimu_01 Chi_01
#define Chimu_02 Chi_02
#define Chimu_10 Chi_10
#define Chimu_11 Chi_11
#define Chimu_12 Chi_12
#define Chimu_20 UChi_00
#define Chimu_21 UChi_01
#define Chimu_22 UChi_02
#define Chimu_30 UChi_10
#define Chimu_31 UChi_11
#define Chimu_32 UChi_12
NAMESPACE_BEGIN(Grid);
template<class Impl> void
WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
int offset,local,perm, ptype;
StencilEntry *SE;
HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON);
HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM);
HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM);
HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM);
HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM);
HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM);
HAND_RESULT(ss);
#else
assert(0);
#endif
}
template<class Impl>
void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
StencilEntry *SE;
int offset,local,perm, ptype;
HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON);
HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM);
HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM);
HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM);
HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM);
HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM);
HAND_RESULT(ss);
#else
assert(0);
#endif
}
template<class Impl> void
WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
int offset,local,perm, ptype;
StencilEntry *SE;
ZERO_RESULT;
HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
HAND_STENCIL_LEG_INT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
HAND_STENCIL_LEG_INT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
HAND_STENCIL_LEG_INT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
HAND_RESULT(ss);
#else
assert(0);
#endif
}
template<class Impl>
void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
StencilEntry *SE;
int offset,local,perm, ptype;
ZERO_RESULT;
HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
HAND_STENCIL_LEG_INT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
HAND_STENCIL_LEG_INT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
HAND_STENCIL_LEG_INT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
HAND_RESULT(ss);
#else
assert(0);
#endif
}
template<class Impl> void
WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
int offset,local,perm, ptype;
StencilEntry *SE;
int nmu=0;
ZERO_RESULT;
HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xp,XM_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(YM_PROJ,2,Yp,YM_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tp,TM_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xm,XP_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(YP_PROJ,2,Ym,YP_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tm,TP_RECON_ACCUM);
HAND_RESULT_EXT(ss);
#else
assert(0);
#endif
}
template<class Impl>
void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
StencilEntry *SE;
int offset,local,perm, ptype;
int nmu=0;
ZERO_RESULT;
HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(YP_PROJ,2,Yp,YP_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tp,TP_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xm,XM_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(YM_PROJ,2,Ym,YM_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM);
HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tm,TM_RECON_ACCUM);
HAND_RESULT_EXT(ss);
#else
assert(0);
#endif
}
////////////// Wilson ; uses this implementation /////////////////////
#define INSTANTIATE_THEM(A) \
template void WilsonKernels<A>::HandDhopSite(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteDag(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out);\
template void WilsonKernels<A>::HandDhopSiteInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteDagInt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteDagExt(StencilView &st,DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out);
INSTANTIATE_THEM(WilsonImplF);
INSTANTIATE_THEM(WilsonImplD);
INSTANTIATE_THEM(ZWilsonImplF);
INSTANTIATE_THEM(ZWilsonImplD);
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(DomainWallVec5dImplFH);
INSTANTIATE_THEM(DomainWallVec5dImplDF);
INSTANTIATE_THEM(ZDomainWallVec5dImplFH);
INSTANTIATE_THEM(ZDomainWallVec5dImplDF);
INSTANTIATE_THEM(WilsonTwoIndexAntiSymmetricImplF);
INSTANTIATE_THEM(WilsonTwoIndexAntiSymmetricImplD);
NAMESPACE_END(Grid);

943
Grid/qcd/action/fermion/WilsonKernelsHandGparity.cc
View File

@ -1,943 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernelsHand.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#define REGISTER
#define LOAD_CHIMU_BODY(F) \
Chimu_00=ref(F)(0)(0); \
Chimu_01=ref(F)(0)(1); \
Chimu_02=ref(F)(0)(2); \
Chimu_10=ref(F)(1)(0); \
Chimu_11=ref(F)(1)(1); \
Chimu_12=ref(F)(1)(2); \
Chimu_20=ref(F)(2)(0); \
Chimu_21=ref(F)(2)(1); \
Chimu_22=ref(F)(2)(2); \
Chimu_30=ref(F)(3)(0); \
Chimu_31=ref(F)(3)(1); \
Chimu_32=ref(F)(3)(2)
#define LOAD_CHIMU(DIR,F,PERM) \
{ const SiteSpinor & ref (in[offset]); LOAD_CHIMU_BODY(F); }
#define LOAD_CHI_BODY(F) \
Chi_00 = ref(F)(0)(0);\
Chi_01 = ref(F)(0)(1);\
Chi_02 = ref(F)(0)(2);\
Chi_10 = ref(F)(1)(0);\
Chi_11 = ref(F)(1)(1);\
Chi_12 = ref(F)(1)(2)
#define LOAD_CHI(DIR,F,PERM) \
{const SiteHalfSpinor &ref(buf[offset]); LOAD_CHI_BODY(F); }
//G-parity implementations using in-place intrinsic ops
//1l 1h -> 1h 1l
//0l 0h , 1h 1l -> 0l 1h 0h,1l
//0h,1l -> 1l,0h
//if( (distance == 1 && !perm_will_occur) || (distance == -1 && perm_will_occur) )
//Pulled fermion through forwards face, GPBC on upper component
//Need 0= 0l 1h 1= 1l 0h
//else if( (distance == -1 && !perm) || (distance == 1 && perm) )
//Pulled fermion through backwards face, GPBC on lower component
//Need 0= 1l 0h 1= 0l 1h
//1l 1h -> 1h 1l
//0l 0h , 1h 1l -> 0l 1h 0h,1l
#define DO_TWIST_0L_1H(INTO,S,C,F, PERM, tmp1, tmp2, tmp3) \
permute##PERM(tmp1, ref(1)(S)(C)); \
exchange##PERM(tmp2,tmp3, ref(0)(S)(C), tmp1); \
INTO = tmp2;
//0l 0h -> 0h 0l
//1l 1h, 0h 0l -> 1l 0h, 1h 0l
#define DO_TWIST_1L_0H(INTO,S,C,F, PERM, tmp1, tmp2, tmp3) \
permute##PERM(tmp1, ref(0)(S)(C)); \
exchange##PERM(tmp2,tmp3, ref(1)(S)(C), tmp1); \
INTO = tmp2;
#define LOAD_CHI_SETUP(DIR,F) \
g = F; \
direction = st._directions[DIR]; \
distance = st._distances[DIR]; \
sl = st._simd_layout[direction]; \
inplace_twist = 0; \
if(SE->_around_the_world && st.parameters.twists[DIR % 4]){ \
if(sl == 1){ \
g = (F+1) % 2; \
}else{ \
inplace_twist = 1; \
} \
}
#define LOAD_CHIMU_GPARITY_INPLACE_TWIST(DIR,F,PERM) \
{ const SiteSpinor &ref(in[offset]); \
LOAD_CHI_SETUP(DIR,F); \
if(!inplace_twist){ \
LOAD_CHIMU_BODY(g); \
}else{ \
if( ( F==0 && ((distance == 1 && !perm) || (distance == -1 && perm)) ) || \
( F==1 && ((distance == -1 && !perm) || (distance == 1 && perm)) ) ){ \
DO_TWIST_0L_1H(Chimu_00,0,0,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chimu_01,0,1,F,PERM, U_11,U_20,U_21); \
DO_TWIST_0L_1H(Chimu_02,0,2,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chimu_10,1,0,F,PERM, U_11,U_20,U_21); \
DO_TWIST_0L_1H(Chimu_11,1,1,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chimu_12,1,2,F,PERM, U_11,U_20,U_21); \
DO_TWIST_0L_1H(Chimu_20,2,0,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chimu_21,2,1,F,PERM, U_11,U_20,U_21); \
DO_TWIST_0L_1H(Chimu_22,2,2,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chimu_30,3,0,F,PERM, U_11,U_20,U_21); \
DO_TWIST_0L_1H(Chimu_31,3,1,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chimu_32,3,2,F,PERM, U_11,U_20,U_21); \
}else{ \
DO_TWIST_1L_0H(Chimu_00,0,0,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chimu_01,0,1,F,PERM, U_11,U_20,U_21); \
DO_TWIST_1L_0H(Chimu_02,0,2,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chimu_10,1,0,F,PERM, U_11,U_20,U_21); \
DO_TWIST_1L_0H(Chimu_11,1,1,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chimu_12,1,2,F,PERM, U_11,U_20,U_21); \
DO_TWIST_1L_0H(Chimu_20,2,0,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chimu_21,2,1,F,PERM, U_11,U_20,U_21); \
DO_TWIST_1L_0H(Chimu_22,2,2,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chimu_30,3,0,F,PERM, U_11,U_20,U_21); \
DO_TWIST_1L_0H(Chimu_31,3,1,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chimu_32,3,2,F,PERM, U_11,U_20,U_21); \
} \
} \
}
#define LOAD_CHI_GPARITY_INPLACE_TWIST(DIR,F,PERM) \
{ const SiteHalfSpinor &ref(buf[offset]); \
LOAD_CHI_SETUP(DIR,F); \
if(!inplace_twist){ \
LOAD_CHI_BODY(g); \
}else{ \
if( ( F==0 && ((distance == 1 && !perm) || (distance == -1 && perm)) ) || \
( F==1 && ((distance == -1 && !perm) || (distance == 1 && perm)) ) ){ \
DO_TWIST_0L_1H(Chi_00,0,0,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chi_01,0,1,F,PERM, U_11,U_20,U_21); \
DO_TWIST_0L_1H(Chi_02,0,2,F,PERM, UChi_00,UChi_01,UChi_02); \
DO_TWIST_0L_1H(Chi_10,1,0,F,PERM, UChi_10,UChi_11,UChi_12); \
DO_TWIST_0L_1H(Chi_11,1,1,F,PERM, U_00,U_01,U_10); \
DO_TWIST_0L_1H(Chi_12,1,2,F,PERM, U_11,U_20,U_21); \
}else{ \
DO_TWIST_1L_0H(Chi_00,0,0,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chi_01,0,1,F,PERM, U_11,U_20,U_21); \
DO_TWIST_1L_0H(Chi_02,0,2,F,PERM, UChi_00,UChi_01,UChi_02); \
DO_TWIST_1L_0H(Chi_10,1,0,F,PERM, UChi_10,UChi_11,UChi_12); \
DO_TWIST_1L_0H(Chi_11,1,1,F,PERM, U_00,U_01,U_10); \
DO_TWIST_1L_0H(Chi_12,1,2,F,PERM, U_11,U_20,U_21); \
} \
} \
}
#define LOAD_CHI_GPARITY(DIR,F,PERM) LOAD_CHI_GPARITY_INPLACE_TWIST(DIR,F,PERM)
#define LOAD_CHIMU_GPARITY(DIR,F,PERM) LOAD_CHIMU_GPARITY_INPLACE_TWIST(DIR,F,PERM)
// To splat or not to splat depends on the implementation
#define MULT_2SPIN_BODY \
Impl::loadLinkElement(U_00,ref()(0,0)); \
Impl::loadLinkElement(U_10,ref()(1,0)); \
Impl::loadLinkElement(U_20,ref()(2,0)); \
Impl::loadLinkElement(U_01,ref()(0,1)); \
Impl::loadLinkElement(U_11,ref()(1,1)); \
Impl::loadLinkElement(U_21,ref()(2,1)); \
UChi_00 = U_00*Chi_00; \
UChi_10 = U_00*Chi_10; \
UChi_01 = U_10*Chi_00; \
UChi_11 = U_10*Chi_10; \
UChi_02 = U_20*Chi_00; \
UChi_12 = U_20*Chi_10; \
UChi_00+= U_01*Chi_01; \
UChi_10+= U_01*Chi_11; \
UChi_01+= U_11*Chi_01; \
UChi_11+= U_11*Chi_11; \
UChi_02+= U_21*Chi_01; \
UChi_12+= U_21*Chi_11; \
Impl::loadLinkElement(U_00,ref()(0,2)); \
Impl::loadLinkElement(U_10,ref()(1,2)); \
Impl::loadLinkElement(U_20,ref()(2,2)); \
UChi_00+= U_00*Chi_02; \
UChi_10+= U_00*Chi_12; \
UChi_01+= U_10*Chi_02; \
UChi_11+= U_10*Chi_12; \
UChi_02+= U_20*Chi_02; \
UChi_12+= U_20*Chi_12
#define MULT_2SPIN(A,F) \
{auto & ref(U[sU](A)); MULT_2SPIN_BODY; }
#define MULT_2SPIN_GPARITY(A,F) \
{auto & ref(U[sU](F)(A)); MULT_2SPIN_BODY; }
#define PERMUTE_DIR(dir) \
permute##dir(Chi_00,Chi_00);\
permute##dir(Chi_01,Chi_01);\
permute##dir(Chi_02,Chi_02);\
permute##dir(Chi_10,Chi_10);\
permute##dir(Chi_11,Chi_11);\
permute##dir(Chi_12,Chi_12);
// hspin(0)=fspin(0)+timesI(fspin(3));
// hspin(1)=fspin(1)+timesI(fspin(2));
#define XP_PROJ \
Chi_00 = Chimu_00+timesI(Chimu_30);\
Chi_01 = Chimu_01+timesI(Chimu_31);\
Chi_02 = Chimu_02+timesI(Chimu_32);\
Chi_10 = Chimu_10+timesI(Chimu_20);\
Chi_11 = Chimu_11+timesI(Chimu_21);\
Chi_12 = Chimu_12+timesI(Chimu_22);
#define YP_PROJ \
Chi_00 = Chimu_00-Chimu_30;\
Chi_01 = Chimu_01-Chimu_31;\
Chi_02 = Chimu_02-Chimu_32;\
Chi_10 = Chimu_10+Chimu_20;\
Chi_11 = Chimu_11+Chimu_21;\
Chi_12 = Chimu_12+Chimu_22;
#define ZP_PROJ \
Chi_00 = Chimu_00+timesI(Chimu_20); \
Chi_01 = Chimu_01+timesI(Chimu_21); \
Chi_02 = Chimu_02+timesI(Chimu_22); \
Chi_10 = Chimu_10-timesI(Chimu_30); \
Chi_11 = Chimu_11-timesI(Chimu_31); \
Chi_12 = Chimu_12-timesI(Chimu_32);
#define TP_PROJ \
Chi_00 = Chimu_00+Chimu_20; \
Chi_01 = Chimu_01+Chimu_21; \
Chi_02 = Chimu_02+Chimu_22; \
Chi_10 = Chimu_10+Chimu_30; \
Chi_11 = Chimu_11+Chimu_31; \
Chi_12 = Chimu_12+Chimu_32;
// hspin(0)=fspin(0)-timesI(fspin(3));
// hspin(1)=fspin(1)-timesI(fspin(2));
#define XM_PROJ \
Chi_00 = Chimu_00-timesI(Chimu_30);\
Chi_01 = Chimu_01-timesI(Chimu_31);\
Chi_02 = Chimu_02-timesI(Chimu_32);\
Chi_10 = Chimu_10-timesI(Chimu_20);\
Chi_11 = Chimu_11-timesI(Chimu_21);\
Chi_12 = Chimu_12-timesI(Chimu_22);
#define YM_PROJ \
Chi_00 = Chimu_00+Chimu_30;\
Chi_01 = Chimu_01+Chimu_31;\
Chi_02 = Chimu_02+Chimu_32;\
Chi_10 = Chimu_10-Chimu_20;\
Chi_11 = Chimu_11-Chimu_21;\
Chi_12 = Chimu_12-Chimu_22;
#define ZM_PROJ \
Chi_00 = Chimu_00-timesI(Chimu_20); \
Chi_01 = Chimu_01-timesI(Chimu_21); \
Chi_02 = Chimu_02-timesI(Chimu_22); \
Chi_10 = Chimu_10+timesI(Chimu_30); \
Chi_11 = Chimu_11+timesI(Chimu_31); \
Chi_12 = Chimu_12+timesI(Chimu_32);
#define TM_PROJ \
Chi_00 = Chimu_00-Chimu_20; \
Chi_01 = Chimu_01-Chimu_21; \
Chi_02 = Chimu_02-Chimu_22; \
Chi_10 = Chimu_10-Chimu_30; \
Chi_11 = Chimu_11-Chimu_31; \
Chi_12 = Chimu_12-Chimu_32;
// fspin(0)=hspin(0);
// fspin(1)=hspin(1);
// fspin(2)=timesMinusI(hspin(1));
// fspin(3)=timesMinusI(hspin(0));
#define XP_RECON\
result_00 = UChi_00;\
result_01 = UChi_01;\
result_02 = UChi_02;\
result_10 = UChi_10;\
result_11 = UChi_11;\
result_12 = UChi_12;\
result_20 = timesMinusI(UChi_10);\
result_21 = timesMinusI(UChi_11);\
result_22 = timesMinusI(UChi_12);\
result_30 = timesMinusI(UChi_00);\
result_31 = timesMinusI(UChi_01);\
result_32 = timesMinusI(UChi_02);
#define XP_RECON_ACCUM\
result_00+=UChi_00;\
result_01+=UChi_01;\
result_02+=UChi_02;\
result_10+=UChi_10;\
result_11+=UChi_11;\
result_12+=UChi_12;\
result_20-=timesI(UChi_10);\
result_21-=timesI(UChi_11);\
result_22-=timesI(UChi_12);\
result_30-=timesI(UChi_00);\
result_31-=timesI(UChi_01);\
result_32-=timesI(UChi_02);
#define XM_RECON\
result_00 = UChi_00;\
result_01 = UChi_01;\
result_02 = UChi_02;\
result_10 = UChi_10;\
result_11 = UChi_11;\
result_12 = UChi_12;\
result_20 = timesI(UChi_10);\
result_21 = timesI(UChi_11);\
result_22 = timesI(UChi_12);\
result_30 = timesI(UChi_00);\
result_31 = timesI(UChi_01);\
result_32 = timesI(UChi_02);
#define XM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= timesI(UChi_10);\
result_21+= timesI(UChi_11);\
result_22+= timesI(UChi_12);\
result_30+= timesI(UChi_00);\
result_31+= timesI(UChi_01);\
result_32+= timesI(UChi_02);
#define YP_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= UChi_10;\
result_21+= UChi_11;\
result_22+= UChi_12;\
result_30-= UChi_00;\
result_31-= UChi_01;\
result_32-= UChi_02;
#define YM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20-= UChi_10;\
result_21-= UChi_11;\
result_22-= UChi_12;\
result_30+= UChi_00;\
result_31+= UChi_01;\
result_32+= UChi_02;
#define ZP_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20-= timesI(UChi_00); \
result_21-= timesI(UChi_01); \
result_22-= timesI(UChi_02); \
result_30+= timesI(UChi_10); \
result_31+= timesI(UChi_11); \
result_32+= timesI(UChi_12);
#define ZM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= timesI(UChi_00); \
result_21+= timesI(UChi_01); \
result_22+= timesI(UChi_02); \
result_30-= timesI(UChi_10); \
result_31-= timesI(UChi_11); \
result_32-= timesI(UChi_12);
#define TP_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20+= UChi_00; \
result_21+= UChi_01; \
result_22+= UChi_02; \
result_30+= UChi_10; \
result_31+= UChi_11; \
result_32+= UChi_12;
#define TM_RECON_ACCUM\
result_00+= UChi_00;\
result_01+= UChi_01;\
result_02+= UChi_02;\
result_10+= UChi_10;\
result_11+= UChi_11;\
result_12+= UChi_12;\
result_20-= UChi_00; \
result_21-= UChi_01; \
result_22-= UChi_02; \
result_30-= UChi_10; \
result_31-= UChi_11; \
result_32-= UChi_12;
#define HAND_STENCIL_LEG(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU_IMPL(DIR,F,PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else { \
LOAD_CHI_IMPL(DIR,F,PERM); \
} \
MULT_2SPIN_IMPL(DIR,F); \
RECON;
#define HAND_STENCIL_LEG_INT(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
local = SE->_is_local; \
perm = SE->_permute; \
if ( local ) { \
LOAD_CHIMU_IMPL(DIR,F,PERM); \
PROJ; \
if ( perm) { \
PERMUTE_DIR(PERM); \
} \
} else if ( st.same_node[DIR] ) { \
LOAD_CHI_IMPL(DIR,F,PERM); \
} \
if (local || st.same_node[DIR] ) { \
MULT_2SPIN_IMPL(DIR,F); \
RECON; \
}
#define HAND_STENCIL_LEG_EXT(PROJ,PERM,DIR,RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
SE=st.GetEntry(ptype,DIR,ss); \
offset = SE->_offset; \
perm = SE->_permute; \
if((!SE->_is_local)&&(!st.same_node[DIR]) ) { \
LOAD_CHI_IMPL(DIR,F,PERM); \
MULT_2SPIN_IMPL(DIR,F); \
RECON; \
nmu++; \
}
#define HAND_RESULT(ss,F) \
{ \
SiteSpinor & ref (out[ss]); \
vstream(ref(F)(0)(0),result_00); \
vstream(ref(F)(0)(1),result_01); \
vstream(ref(F)(0)(2),result_02); \
vstream(ref(F)(1)(0),result_10); \
vstream(ref(F)(1)(1),result_11); \
vstream(ref(F)(1)(2),result_12); \
vstream(ref(F)(2)(0),result_20); \
vstream(ref(F)(2)(1),result_21); \
vstream(ref(F)(2)(2),result_22); \
vstream(ref(F)(3)(0),result_30); \
vstream(ref(F)(3)(1),result_31); \
vstream(ref(F)(3)(2),result_32); \
}
#define HAND_RESULT_EXT(ss,F) \
if (nmu){ \
SiteSpinor & ref (out[ss]); \
ref(F)(0)(0)+=result_00; \
ref(F)(0)(1)+=result_01; \
ref(F)(0)(2)+=result_02; \
ref(F)(1)(0)+=result_10; \
ref(F)(1)(1)+=result_11; \
ref(F)(1)(2)+=result_12; \
ref(F)(2)(0)+=result_20; \
ref(F)(2)(1)+=result_21; \
ref(F)(2)(2)+=result_22; \
ref(F)(3)(0)+=result_30; \
ref(F)(3)(1)+=result_31; \
ref(F)(3)(2)+=result_32; \
}
#define HAND_DECLARATIONS(a) \
Simd result_00; \
Simd result_01; \
Simd result_02; \
Simd result_10; \
Simd result_11; \
Simd result_12; \
Simd result_20; \
Simd result_21; \
Simd result_22; \
Simd result_30; \
Simd result_31; \
Simd result_32; \
Simd Chi_00; \
Simd Chi_01; \
Simd Chi_02; \
Simd Chi_10; \
Simd Chi_11; \
Simd Chi_12; \
Simd UChi_00; \
Simd UChi_01; \
Simd UChi_02; \
Simd UChi_10; \
Simd UChi_11; \
Simd UChi_12; \
Simd U_00; \
Simd U_10; \
Simd U_20; \
Simd U_01; \
Simd U_11; \
Simd U_21;
#define ZERO_RESULT \
result_00=Zero(); \
result_01=Zero(); \
result_02=Zero(); \
result_10=Zero(); \
result_11=Zero(); \
result_12=Zero(); \
result_20=Zero(); \
result_21=Zero(); \
result_22=Zero(); \
result_30=Zero(); \
result_31=Zero(); \
result_32=Zero();
#define Chimu_00 Chi_00
#define Chimu_01 Chi_01
#define Chimu_02 Chi_02
#define Chimu_10 Chi_10
#define Chimu_11 Chi_11
#define Chimu_12 Chi_12
#define Chimu_20 UChi_00
#define Chimu_21 UChi_01
#define Chimu_22 UChi_02
#define Chimu_30 UChi_10
#define Chimu_31 UChi_11
#define Chimu_32 UChi_12
NAMESPACE_BEGIN(Grid);
template<class Impl> void accelerator
WilsonKernels<Impl>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
int offset,local,perm, ptype;
StencilEntry *SE;
#define HAND_DOP_SITE(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
HAND_STENCIL_LEG(XM_PROJ,3,Xp,XM_RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_RESULT(ss,F)
HAND_DOP_SITE(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
#endif
}
template<class Impl> accelerator
void WilsonKernels<Impl>::HandDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
StencilEntry *SE;
int offset,local,perm, ptype;
#define HAND_DOP_SITE_DAG(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
HAND_STENCIL_LEG(XP_PROJ,3,Xp,XP_RECON,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_RESULT(ss,F)
HAND_DOP_SITE_DAG(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
#endif
}
template<class Impl> void accelerator
WilsonKernels<Impl>::HandDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
int offset,local,perm, ptype;
StencilEntry *SE;
#define HAND_DOP_SITE_INT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
ZERO_RESULT; \
HAND_STENCIL_LEG_INT(XM_PROJ,3,Xp,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_RESULT(ss,F)
HAND_DOP_SITE_INT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
#endif
}
template<class Impl> accelerator
void WilsonKernels<Impl>::HandDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
StencilEntry *SE;
int offset,local,perm, ptype;
#define HAND_DOP_SITE_DAG_INT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
ZERO_RESULT; \
HAND_STENCIL_LEG_INT(XP_PROJ,3,Xp,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_INT(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_RESULT(ss,F)
HAND_DOP_SITE_DAG_INT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
#endif
}
template<class Impl> void accelerator
WilsonKernels<Impl>::HandDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
// T==0, Z==1, Y==2, Z==3 expect 1,2,2,2 simd layout etc...
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
int offset, perm, ptype;
StencilEntry *SE;
int nmu=0;
#define HAND_DOP_SITE_EXT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
ZERO_RESULT; \
HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xp,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(YM_PROJ,2,Yp,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zp,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tp,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xm,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(YP_PROJ,2,Ym,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zm,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tm,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_RESULT_EXT(ss,F)
HAND_DOP_SITE_EXT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
perm++;
#endif
}
template<class Impl>
accelerator void WilsonKernels<Impl>::HandDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,
int ss,int sU,const FermionFieldView &in, FermionFieldView &out)
{
#ifndef GRID_NVCC
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
HAND_DECLARATIONS(ignore);
StencilEntry *SE;
int offset, perm, ptype;
int nmu=0;
#define HAND_DOP_SITE_DAG_EXT(F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL) \
ZERO_RESULT; \
HAND_STENCIL_LEG_EXT(XP_PROJ,3,Xp,XP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(YP_PROJ,2,Yp,YP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(ZP_PROJ,1,Zp,ZP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(TP_PROJ,0,Tp,TP_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(XM_PROJ,3,Xm,XM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(YM_PROJ,2,Ym,YM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(ZM_PROJ,1,Zm,ZM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_STENCIL_LEG_EXT(TM_PROJ,0,Tm,TM_RECON_ACCUM,F,LOAD_CHI_IMPL,LOAD_CHIMU_IMPL,MULT_2SPIN_IMPL); \
HAND_RESULT_EXT(ss,F)
HAND_DOP_SITE_DAG_EXT(, LOAD_CHI,LOAD_CHIMU,MULT_2SPIN);
perm++;
#endif
}
////////////////////////////////////////////////
// Specialise Gparity to simple implementation
////////////////////////////////////////////////
#define HAND_SPECIALISE_EMPTY(IMPL) \
template<> void \
WilsonKernels<IMPL>::HandDhopSite(StencilView &st, \
\
DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, \
int sF,int sU, \
const FermionFieldView &in, \
FermionFieldView &out){ assert(0); } \
template<> void \
WilsonKernels<IMPL>::HandDhopSiteDag(StencilView &st, \
\
DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, \
int sF,int sU, \
const FermionFieldView &in, \
FermionFieldView &out){ assert(0); } \
template<> void \
WilsonKernels<IMPL>::HandDhopSiteInt(StencilView &st, \
\
DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, \
int sF,int sU, \
const FermionFieldView &in, \
FermionFieldView &out){ assert(0); } \
template<> void \
WilsonKernels<IMPL>::HandDhopSiteExt(StencilView &st, \
\
DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, \
int sF,int sU, \
const FermionFieldView &in, \
FermionFieldView &out){ assert(0); } \
template<> void \
WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilView &st, \
\
DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, \
int sF,int sU, \
const FermionFieldView &in, \
FermionFieldView &out){ assert(0); } \
template<> void \
WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilView &st, \
\
DoubledGaugeFieldView &U, \
SiteHalfSpinor *buf, \
int sF,int sU, \
const FermionFieldView &in, \
FermionFieldView &out){ assert(0); } \
#ifdef GRID_NVCC
#define HAND_SPECIALISE_GPARITY(IMPL) HAND_SPECIALISE_EMPTY(IMPL)
#else
#define HAND_SPECIALISE_GPARITY(IMPL) \
template<> void \
WilsonKernels<IMPL>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
{ \
typedef IMPL Impl; \
typedef typename Simd::scalar_type S; \
typedef typename Simd::vector_type V; \
\
HAND_DECLARATIONS(ignore); \
\
int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
StencilEntry *SE; \
HAND_DOP_SITE(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
HAND_DOP_SITE(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
} \
\
template<> \
void WilsonKernels<IMPL>::HandDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
{ \
typedef IMPL Impl; \
typedef typename Simd::scalar_type S; \
typedef typename Simd::vector_type V; \
\
HAND_DECLARATIONS(ignore); \
\
StencilEntry *SE; \
int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
HAND_DOP_SITE_DAG(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
HAND_DOP_SITE_DAG(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
} \
\
template<> void \
WilsonKernels<IMPL>::HandDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
{ \
typedef IMPL Impl; \
typedef typename Simd::scalar_type S; \
typedef typename Simd::vector_type V; \
\
HAND_DECLARATIONS(ignore); \
\
int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
StencilEntry *SE; \
HAND_DOP_SITE_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
HAND_DOP_SITE_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
} \
\
template<> \
void WilsonKernels<IMPL>::HandDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
{ \
typedef IMPL Impl; \
typedef typename Simd::scalar_type S; \
typedef typename Simd::vector_type V; \
\
HAND_DECLARATIONS(ignore); \
\
StencilEntry *SE; \
int offset,local,perm, ptype, g, direction, distance, sl, inplace_twist; \
HAND_DOP_SITE_DAG_INT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
HAND_DOP_SITE_DAG_INT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
} \
\
template<> void \
WilsonKernels<IMPL>::HandDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
{ \
typedef IMPL Impl; \
typedef typename Simd::scalar_type S; \
typedef typename Simd::vector_type V; \
\
HAND_DECLARATIONS(ignore); \
\
int offset,perm, ptype, g, direction, distance, sl, inplace_twist; \
StencilEntry *SE; \
int nmu=0; \
HAND_DOP_SITE_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
nmu = 0; \
HAND_DOP_SITE_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
} \
template<> \
void WilsonKernels<IMPL>::HandDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out) \
{ \
typedef IMPL Impl; \
typedef typename Simd::scalar_type S; \
typedef typename Simd::vector_type V; \
\
HAND_DECLARATIONS(ignore); \
\
StencilEntry *SE; \
int offset,perm, ptype, g, direction, distance, sl, inplace_twist; \
int nmu=0; \
HAND_DOP_SITE_DAG_EXT(0, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
nmu = 0; \
HAND_DOP_SITE_DAG_EXT(1, LOAD_CHI_GPARITY,LOAD_CHIMU_GPARITY,MULT_2SPIN_GPARITY); \
}
#endif
HAND_SPECIALISE_GPARITY(GparityWilsonImplF);
HAND_SPECIALISE_GPARITY(GparityWilsonImplD);
HAND_SPECIALISE_GPARITY(GparityWilsonImplFH);
HAND_SPECIALISE_GPARITY(GparityWilsonImplDF);
////////////// Wilson ; uses this implementation /////////////////////
#define INSTANTIATE_THEM(A) \
template void WilsonKernels<A>::HandDhopSite(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteDag(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out);\
template void WilsonKernels<A>::HandDhopSiteInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteDagInt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf,\
int ss,int sU,const FermionFieldView &in, FermionFieldView &out); \
template void WilsonKernels<A>::HandDhopSiteDagExt(StencilView &st, DoubledGaugeFieldView &U,SiteHalfSpinor *buf, \
int ss,int sU,const FermionFieldView &in, FermionFieldView &out);
//INSTANTIATE_THEM(GparityWilsonImplF);
//INSTANTIATE_THEM(GparityWilsonImplD);
//INSTANTIATE_THEM(GparityWilsonImplFH);
//INSTANTIATE_THEM(GparityWilsonImplDF);
//INSTANTIATE_THEM(DomainWallVec5dImplFH);
//INSTANTIATE_THEM(DomainWallVec5dImplDF);
//INSTANTIATE_THEM(ZDomainWallVec5dImplFH);
//INSTANTIATE_THEM(ZDomainWallVec5dImplDF);
NAMESPACE_END(Grid);

97
Grid/qcd/action/fermion/WilsonTMFermion.cc
View File

@ -1,97 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonTMFermion.cc
Copyright (C) 2015
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/WilsonTMFermion.h>
NAMESPACE_BEGIN(Grid);
/*
* BF sequence
*
void bfmbase<Float>::MooeeInv(Fermion_t psi,
Fermion_t chi,
int dag, int cb)
double m = this->mass;
double tm = this->twistedmass;
double mtil = 4.0+this->mass;
double sq = mtil*mtil + tm*tm;
double a = mtil/sq;
double b = -tm /sq;
if(dag) b=-b;
axpibg5x(chi,psi,a,b);
void bfmbase<Float>::Mooee(Fermion_t psi,
Fermion_t chi,
int dag,int cb)
double a = 4.0+this->mass;
double b = this->twistedmass;
if(dag) b=-b;
axpibg5x(chi,psi,a,b);
*/
template<class Impl>
void WilsonTMFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
RealD a = 4.0+this->mass;
RealD b = this->mu;
out.Checkerboard() = in.Checkerboard();
axpibg5x(out,in,a,b);
}
template<class Impl>
void WilsonTMFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
RealD a = 4.0+this->mass;
RealD b = -this->mu;
out.Checkerboard() = in.Checkerboard();
axpibg5x(out,in,a,b);
}
template<class Impl>
void WilsonTMFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
RealD m = this->mass;
RealD tm = this->mu;
RealD mtil = 4.0+m;
RealD sq = mtil*mtil+tm*tm;
RealD a = mtil/sq;
RealD b = -tm /sq;
axpibg5x(out,in,a,b);
}
template<class Impl>
void WilsonTMFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) {
RealD m = this->mass;
RealD tm = this->mu;
RealD mtil = 4.0+m;
RealD sq = mtil*mtil+tm*tm;
RealD a = mtil/sq;
RealD b = tm /sq;
axpibg5x(out,in,a,b);
}
FermOpTemplateInstantiate(WilsonTMFermion);
NAMESPACE_END(Grid);

0
Grid/qcd/action/fermion/CayleyFermion5Ddense.cc → Grid/qcd/action/fermion/deprecated/CayleyFermion5Ddense.h
View File

0
Grid/qcd/action/fermion/CayleyFermion5Dssp.cc → Grid/qcd/action/fermion/deprecated/CayleyFermion5Dssp.h
View File

0
Grid/qcd/action/fermion/DomainWallEOFAFermiondense.cc → Grid/qcd/action/fermion/deprecated/DomainWallEOFAFermiondense.h
View File

0
Grid/qcd/action/fermion/DomainWallEOFAFermionssp.cc → Grid/qcd/action/fermion/deprecated/DomainWallEOFAFermionssp.h
View File

0
Grid/qcd/action/fermion/MobiusEOFAFermiondense.cc → Grid/qcd/action/fermion/deprecated/MobiusEOFAFermiondense.h
View File

0
Grid/qcd/action/fermion/MobiusEOFAFermionssp.cc → Grid/qcd/action/fermion/deprecated/MobiusEOFAFermionssp.h
View File

433
Grid/qcd/action/fermion/implementation/DomainWallEOFAFermion.h
View File

@ -1,433 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermion.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid_Eigen_Dense.h>
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
DomainWallEOFAFermion<Impl>::DomainWallEOFAFermion(
GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mq1, RealD _mq2, RealD _mq3,
RealD _shift, int _pm, RealD _M5, const ImplParams &p) :
AbstractEOFAFermion<Impl>(_Umu, FiveDimGrid, FiveDimRedBlackGrid,
FourDimGrid, FourDimRedBlackGrid, _mq1, _mq2, _mq3,
_shift, _pm, _M5, 1.0, 0.0, p)
{
RealD eps = 1.0;
Approx::zolotarev_data *zdata = Approx::higham(eps,this->Ls);
assert(zdata->n == this->Ls);
std::cout << GridLogMessage << "DomainWallEOFAFermion with Ls=" << this->Ls << std::endl;
this->SetCoefficientsTanh(zdata, 1.0, 0.0);
Approx::zolotarev_free(zdata);
}
/***************************************************************
* Additional EOFA operators only called outside the inverter.
* Since speed is not essential, simple axpby-style
* implementations should be fine.
***************************************************************/
template<class Impl>
void DomainWallEOFAFermion<Impl>::Omega(const FermionField& psi, FermionField& Din, int sign, int dag)
{
int Ls = this->Ls;
Din = Zero();
if((sign == 1) && (dag == 0)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, Ls-1, 0); }
else if((sign == -1) && (dag == 0)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, 0); }
else if((sign == 1 ) && (dag == 1)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, Ls-1); }
else if((sign == -1) && (dag == 1)){ axpby_ssp(Din, 0.0, psi, 1.0, psi, 0, 0); }
}
// This is just the identity for DWF
template<class Impl>
void DomainWallEOFAFermion<Impl>::Dtilde(const FermionField& psi, FermionField& chi){ chi = psi; }
// This is just the identity for DWF
template<class Impl>
void DomainWallEOFAFermion<Impl>::DtildeInv(const FermionField& psi, FermionField& chi){ chi = psi; }
/*****************************************************************************************************/
template<class Impl>
RealD DomainWallEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->Meooe5D(psi, Din);
this->DW(Din, chi, DaggerNo);
axpby(chi, 1.0, 1.0, chi, psi);
this->M5D(psi, chi);
return(norm2(chi));
}
template<class Impl>
RealD DomainWallEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->DW(psi, Din, DaggerYes);
this->MeooeDag5D(Din, chi);
this->M5Ddag(psi, chi);
axpby(chi, 1.0, 1.0, chi, psi);
return(norm2(chi));
}
/********************************************************************
* Performance critical fermion operators called inside the inverter
********************************************************************/
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
int pm = this->pm;
RealD shift = this->shift;
RealD mq1 = this->mq1;
RealD mq2 = this->mq2;
RealD mq3 = this->mq3;
// coefficients for shift operator ( = shift*\gamma_{5}*R_{5}*\Delta_{\pm}(mq2,mq3)*P_{\pm} )
Coeff_t shiftp(0.0), shiftm(0.0);
if(shift != 0.0){
if(pm == 1){ shiftp = shift*(mq3-mq2); }
else{ shiftm = -shift*(mq3-mq2); }
}
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftm;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = mq1 + shiftp;
#if(0)
std::cout << GridLogMessage << "DomainWallEOFAFermion::M5D(FF&,FF&):" << std::endl;
for(int i=0; i<diag.size(); ++i){
std::cout << GridLogMessage << "diag[" << i << "] =" << diag[i] << std::endl;
}
for(int i=0; i<upper.size(); ++i){
std::cout << GridLogMessage << "upper[" << i << "] =" << upper[i] << std::endl;
}
for(int i=0; i<lower.size(); ++i){
std::cout << GridLogMessage << "lower[" << i << "] =" << lower[i] << std::endl;
}
#endif
this->M5D(psi, chi, chi, lower, diag, upper);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
int pm = this->pm;
RealD shift = this->shift;
RealD mq1 = this->mq1;
RealD mq2 = this->mq2;
RealD mq3 = this->mq3;
// coefficients for shift operator ( = shift*\gamma_{5}*R_{5}*\Delta_{\pm}(mq2,mq3)*P_{\pm} )
Coeff_t shiftp(0.0), shiftm(0.0);
if(shift != 0.0){
if(pm == 1){ shiftp = shift*(mq3-mq2); }
else{ shiftm = -shift*(mq3-mq2); }
}
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = mq1 + shiftp;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = mq1 + shiftm;
#if(0)
std::cout << GridLogMessage << "DomainWallEOFAFermion::M5Ddag(FF&,FF&):" << std::endl;
for(int i=0; i<diag.size(); ++i){
std::cout << GridLogMessage << "diag[" << i << "] =" << diag[i] << std::endl;
}
for(int i=0; i<upper.size(); ++i){
std::cout << GridLogMessage << "upper[" << i << "] =" << upper[i] << std::endl;
}
for(int i=0; i<lower.size(); ++i){
std::cout << GridLogMessage << "lower[" << i << "] =" << lower[i] << std::endl;
}
#endif
this->M5Ddag(psi, chi, chi, lower, diag, upper);
}
// half checkerboard operations
template<class Impl>
void DomainWallEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
upper[s] = -this->cee[s];
lower[s] = -this->cee[s];
}
upper[Ls-1] = this->dm;
lower[0] = this->dp;
this->M5D(psi, psi, chi, lower, diag, upper);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
upper[s] = -this->cee[s];
lower[s] = -this->cee[s];
}
upper[Ls-1] = this->dp;
lower[0] = this->dm;
this->M5Ddag(psi, psi, chi, lower, diag, upper);
}
/****************************************************************************************/
//Zolo
template<class Impl>
void DomainWallEOFAFermion<Impl>::SetCoefficientsInternal(RealD zolo_hi, Vector<Coeff_t>& gamma, RealD b, RealD c)
{
int Ls = this->Ls;
int pm = this->pm;
RealD mq1 = this->mq1;
RealD mq2 = this->mq2;
RealD mq3 = this->mq3;
RealD shift = this->shift;
////////////////////////////////////////////////////////
// Constants for the preconditioned matrix Cayley form
////////////////////////////////////////////////////////
this->bs.resize(Ls);
this->cs.resize(Ls);
this->aee.resize(Ls);
this->aeo.resize(Ls);
this->bee.resize(Ls);
this->beo.resize(Ls);
this->cee.resize(Ls);
this->ceo.resize(Ls);
for(int i=0; i<Ls; ++i){
this->bee[i] = 4.0 - this->M5 + 1.0;
this->cee[i] = 1.0;
}
for(int i=0; i<Ls; ++i){
this->aee[i] = this->cee[i];
this->bs[i] = this->beo[i] = 1.0;
this->cs[i] = this->ceo[i] = 0.0;
}
//////////////////////////////////////////
// EOFA shift terms
//////////////////////////////////////////
if(pm == 1){
this->dp = mq1*this->cee[0] + shift*(mq3-mq2);
this->dm = mq1*this->cee[Ls-1];
} else if(this->pm == -1) {
this->dp = mq1*this->cee[0];
this->dm = mq1*this->cee[Ls-1] - shift*(mq3-mq2);
} else {
this->dp = mq1*this->cee[0];
this->dm = mq1*this->cee[Ls-1];
}
//////////////////////////////////////////
// LDU decomposition of eeoo
//////////////////////////////////////////
this->dee.resize(Ls+1);
this->lee.resize(Ls);
this->leem.resize(Ls);
this->uee.resize(Ls);
this->ueem.resize(Ls);
for(int i=0; i<Ls; ++i){
if(i < Ls-1){
this->lee[i] = -this->cee[i+1]/this->bee[i]; // sub-diag entry on the ith column
this->leem[i] = this->dm/this->bee[i];
for(int j=0; j<i; j++){ this->leem[i] *= this->aee[j]/this->bee[j]; }
this->dee[i] = this->bee[i];
this->uee[i] = -this->aee[i]/this->bee[i]; // up-diag entry on the ith row
this->ueem[i] = this->dp / this->bee[0];
for(int j=1; j<=i; j++){ this->ueem[i] *= this->cee[j]/this->bee[j]; }
} else {
this->lee[i] = 0.0;
this->leem[i] = 0.0;
this->uee[i] = 0.0;
this->ueem[i] = 0.0;
}
}
{
Coeff_t delta_d = 1.0 / this->bee[0];
for(int j=1; j<Ls-1; j++){ delta_d *= this->cee[j] / this->bee[j]; }
this->dee[Ls-1] = this->bee[Ls-1] + this->cee[0] * this->dm * delta_d;
this->dee[Ls] = this->bee[Ls-1] + this->cee[Ls-1] * this->dp * delta_d;
}
int inv = 1;
this->MooeeInternalCompute(0, inv, this->MatpInv, this->MatmInv);
this->MooeeInternalCompute(1, inv, this->MatpInvDag, this->MatmInvDag);
}
// Recompute Cayley-form coefficients for different shift
template<class Impl>
void DomainWallEOFAFermion<Impl>::RefreshShiftCoefficients(RealD new_shift)
{
this->shift = new_shift;
Approx::zolotarev_data *zdata = Approx::higham(1.0, this->Ls);
this->SetCoefficientsTanh(zdata, 1.0, 0.0);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternalCompute(int dag, int inv,
Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
int Ls = this->Ls;
GridBase* grid = this->FermionRedBlackGrid();
int LLs = grid->_rdimensions[0];
if(LLs == Ls){ return; } // Not vectorised in 5th direction
Eigen::MatrixXcd Pplus = Eigen::MatrixXcd::Zero(Ls,Ls);
Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
for(int s=0; s<Ls; s++){
Pplus(s,s) = this->bee[s];
Pminus(s,s) = this->bee[s];
}
for(int s=0; s<Ls-1; s++){
Pminus(s,s+1) = -this->cee[s];
}
for(int s=0; s<Ls-1; s++){
Pplus(s+1,s) = -this->cee[s+1];
}
Pplus (0,Ls-1) = this->dp;
Pminus(Ls-1,0) = this->dm;
Eigen::MatrixXcd PplusMat ;
Eigen::MatrixXcd PminusMat;
#if(0)
std::cout << GridLogMessage << "Pplus:" << std::endl;
for(int s=0; s<Ls; ++s){
for(int ss=0; ss<Ls; ++ss){
std::cout << Pplus(s,ss) << "\t";
}
std::cout << std::endl;
}
std::cout << GridLogMessage << "Pminus:" << std::endl;
for(int s=0; s<Ls; ++s){
for(int ss=0; ss<Ls; ++ss){
std::cout << Pminus(s,ss) << "\t";
}
std::cout << std::endl;
}
#endif
if(inv) {
PplusMat = Pplus.inverse();
PminusMat = Pminus.inverse();
} else {
PplusMat = Pplus;
PminusMat = Pminus;
}
if(dag){
PplusMat.adjointInPlace();
PminusMat.adjointInPlace();
}
typedef typename SiteHalfSpinor::scalar_type scalar_type;
const int Nsimd = Simd::Nsimd();
Matp.resize(Ls*LLs);
Matm.resize(Ls*LLs);
for(int s2=0; s2<Ls; s2++){
for(int s1=0; s1<LLs; s1++){
int istride = LLs;
int ostride = 1;
Simd Vp;
Simd Vm;
scalar_type *sp = (scalar_type*) &Vp;
scalar_type *sm = (scalar_type*) &Vm;
for(int l=0; l<Nsimd; l++){
if(switcheroo<Coeff_t>::iscomplex()) {
sp[l] = PplusMat (l*istride+s1*ostride,s2);
sm[l] = PminusMat(l*istride+s1*ostride,s2);
} else {
// if real
scalar_type tmp;
tmp = PplusMat (l*istride+s1*ostride,s2);
sp[l] = scalar_type(tmp.real(),tmp.real());
tmp = PminusMat(l*istride+s1*ostride,s2);
sm[l] = scalar_type(tmp.real(),tmp.real());
}
}
Matp[LLs*s2+s1] = Vp;
Matm[LLs*s2+s1] = Vm;
}}
}
FermOpTemplateInstantiate(DomainWallEOFAFermion);
GparityFermOpTemplateInstantiate(DomainWallEOFAFermion);
NAMESPACE_END(Grid);

255
Grid/qcd/action/fermion/implementation/DomainWallEOFAFermioncache.h
View File

@ -1,255 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermioncache.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
// FIXME -- make a version of these routines with site loop outermost for cache reuse.
// Pminus fowards
// Pplus backwards..
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
int Ls = this->Ls;
GridBase* grid = psi_i.Grid();
auto phi = phi_i.View();
auto psi = psi_i.View();
auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
for(int s=0; s<Ls; s++){
auto tmp = psi[0];
if(s==0) {
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5m(tmp, psi[ss+0]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
auto tmp = psi[0];
for(int s=0; s<Ls; s++){
if(s==0) {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5p(tmp, psi[ss+0]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi_i, FermionField& chi_i)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi=psi_i.View();
auto chi=chi_i.View();
int Ls = this->Ls;
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
auto tmp1 = psi[0];
auto tmp2 = psi[0];
// flops = 12*2*Ls + 12*2*Ls + 3*12*Ls + 12*2*Ls = 12*Ls * (9) = 108*Ls flops
// Apply (L^{\prime})^{-1}
chi[ss] = psi[ss]; // chi[0]=psi[0]
for(int s=1; s<Ls; s++){
spProj5p(tmp1, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - this->lee[s-1]*tmp1;
}
// L_m^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
spProj5m(tmp1, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - this->leem[s]*tmp1;
}
// U_m^{-1} D^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[s] + 1/d chi[s]
spProj5p(tmp1, chi[ss+Ls-1]);
chi[ss+s] = (1.0/this->dee[s])*chi[ss+s] - (this->ueem[s]/this->dee[Ls])*tmp1;
}
spProj5m(tmp2, chi[ss+Ls-1]);
chi[ss+Ls-1] = (1.0/this->dee[Ls])*tmp1 + (1.0/this->dee[Ls-1])*tmp2;
// Apply U^{-1}
for(int s=Ls-2; s>=0; s--){
spProj5m(tmp1, chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - this->uee[s]*tmp1;
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi_i, FermionField& chi_i)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
assert(psi.Checkerboard() == psi.Checkerboard());
Vector<Coeff_t> ueec(Ls);
Vector<Coeff_t> deec(Ls+1);
Vector<Coeff_t> leec(Ls);
Vector<Coeff_t> ueemc(Ls);
Vector<Coeff_t> leemc(Ls);
for(int s=0; s<ueec.size(); s++){
ueec[s] = conjugate(this->uee[s]);
deec[s] = conjugate(this->dee[s]);
leec[s] = conjugate(this->lee[s]);
ueemc[s] = conjugate(this->ueem[s]);
leemc[s] = conjugate(this->leem[s]);
}
deec[Ls] = conjugate(this->dee[Ls]);
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=Ls),{ // adds Ls
auto tmp1 = psi[0];
auto tmp2 = psi[0];
// Apply (U^{\prime})^{-dagger}
chi[ss] = psi[ss];
for(int s=1; s<Ls; s++){
spProj5m(tmp1, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - ueec[s-1]*tmp1;
}
// U_m^{-\dagger}
for(int s=0; s<Ls-1; s++){
spProj5p(tmp1, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - ueemc[s]*tmp1;
}
// L_m^{-\dagger} D^{-dagger}
for(int s=0; s<Ls-1; s++){
spProj5m(tmp1, chi[ss+Ls-1]);
chi[ss+s] = (1.0/deec[s])*chi[ss+s] - (leemc[s]/deec[Ls-1])*tmp1;
}
spProj5p(tmp2, chi[ss+Ls-1]);
chi[ss+Ls-1] = (1.0/deec[Ls-1])*tmp1 + (1.0/deec[Ls])*tmp2;
// Apply L^{-dagger}
for(int s=Ls-2; s>=0; s--){
spProj5p(tmp1, chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - leec[s]*tmp1;
}
});
this->MooeeInvTime += usecond();
}
#ifdef DOMAIN_WALL_EOFA_DPERP_CACHE
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplF);
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplD);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplF);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplD);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplF);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplD);
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplFH);
INSTANTIATE_DPERP_DWF_EOFA(WilsonImplDF);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplFH);
INSTANTIATE_DPERP_DWF_EOFA(GparityWilsonImplDF);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplFH);
INSTANTIATE_DPERP_DWF_EOFA(ZWilsonImplDF);
#endif
NAMESPACE_END(Grid);

613
Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionvec.h
View File

@ -1,613 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermionvec.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
/*
* Dense matrix versions of routines
*/
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd> > u(LLs);
Vector<iSinglet<Simd> > l(LLs);
Vector<iSinglet<Simd> > d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0;o<LLs;o++){ // outer
for(int i=0;i<nsimd;i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
assert(Nc == 3);
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5m(hp, psi[ss+vp]);
spProj5p(hm, psi[ss+vm]);
if (vp <= v){ rotate(hp, hp, 1); }
if (vm >= v){ rotate(hm, hm, nsimd-1); }
hp = 0.5*hp;
hm = 0.5*hm;
spRecon5m(fp, hp);
spRecon5p(fm, hm);
chi[ss+v] = d[v]*phi[ss+v];
chi[ss+v] = chi[ss+v] + u[v]*fp;
chi[ss+v] = chi[ss+v] + l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v==LLs-1) ? 0 : v+1;
int vm = (v==0) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(2)(0);
Simd hp_01 = psi[ss+vp]()(2)(1);
Simd hp_02 = psi[ss+vp]()(2)(2);
Simd hp_10 = psi[ss+vp]()(3)(0);
Simd hp_11 = psi[ss+vp]()(3)(1);
Simd hp_12 = psi[ss+vp]()(3)(2);
Simd hm_00 = psi[ss+vm]()(0)(0);
Simd hm_01 = psi[ss+vm]()(0)(1);
Simd hm_02 = psi[ss+vm]()(0)(2);
Simd hm_10 = psi[ss+vm]()(1)(0);
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
if(vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
// Can force these to real arithmetic and save 2x.
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
template<class Impl>
void DomainWallEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd> > u(LLs);
Vector<iSinglet<Simd> > l(LLs);
Vector<iSinglet<Simd> > d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop((int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5p(hp, psi[ss+vp]);
spProj5m(hm, psi[ss+vm]);
if(vp <= v){ rotate(hp, hp, 1); }
if(vm >= v){ rotate(hm, hm, nsimd-1); }
hp = hp*0.5;
hm = hm*0.5;
spRecon5p(fp, hp);
spRecon5m(fm, hm);
chi[ss+v] = d[v]*phi[ss+v]+u[v]*fp;
chi[ss+v] = chi[ss+v] +l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if (vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
#ifdef AVX512
#include<simd/Intel512common.h>
#include<simd/Intel512avx.h>
#include<simd/Intel512single.h>
#endif
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternalAsm(const FermionField& psi_i, FermionField& chi_i,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
#ifndef AVX512
{
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
SiteHalfSpinor SiteChiP;
SiteHalfSpinor SiteChiM;
// Ls*Ls * 2 * 12 * vol flops
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
for(int l=0; l < Simd::Nsimd(); l++){ // simd lane
int s = s2 + l*LLs;
int lex = s2 + LLs*site;
if( s2==0 && l==0 ){
SiteChiP=Zero();
SiteChiM=Zero();
}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastP()(sp)(co), psi[lex]()(sp)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastM()(sp)(co), psi[lex]()(sp+2)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
SiteChiP()(sp)(co) = real_madd(Matp[LLs*s+s1]()()(), BcastP()(sp)(co), SiteChiP()(sp)(co)); // 1100 us.
SiteChiM()(sp)(co) = real_madd(Matm[LLs*s+s1]()()(), BcastM()(sp)(co), SiteChiM()(sp)(co)); // each found by commenting out
}}
}}
{
int lex = s1 + LLs*site;
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vstream(chi[lex]()(sp)(co), SiteChiP()(sp)(co));
vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
}}
}
}
}
#else
{
// pointers
// MASK_REGS;
#define Chi_00 %%zmm1
#define Chi_01 %%zmm2
#define Chi_02 %%zmm3
#define Chi_10 %%zmm4
#define Chi_11 %%zmm5
#define Chi_12 %%zmm6
#define Chi_20 %%zmm7
#define Chi_21 %%zmm8
#define Chi_22 %%zmm9
#define Chi_30 %%zmm10
#define Chi_31 %%zmm11
#define Chi_32 %%zmm12
#define BCAST0 %%zmm13
#define BCAST1 %%zmm14
#define BCAST2 %%zmm15
#define BCAST3 %%zmm16
#define BCAST4 %%zmm17
#define BCAST5 %%zmm18
#define BCAST6 %%zmm19
#define BCAST7 %%zmm20
#define BCAST8 %%zmm21
#define BCAST9 %%zmm22
#define BCAST10 %%zmm23
#define BCAST11 %%zmm24
int incr = LLs*LLs*sizeof(iSinglet<Simd>);
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
int lex = s2 + LLs*site;
uint64_t a0 = (uint64_t) &Matp[LLs*s2+s1]; // should be cacheable
uint64_t a1 = (uint64_t) &Matm[LLs*s2+s1];
uint64_t a2 = (uint64_t) &psi[lex];
for(int l=0; l<Simd::Nsimd(); l++){ // simd lane
if((s2+l)==0) {
asm(
VPREFETCH1(0,%2) VPREFETCH1(0,%1)
VPREFETCH1(12,%2) VPREFETCH1(13,%2)
VPREFETCH1(14,%2) VPREFETCH1(15,%2)
VBCASTCDUP(0,%2,BCAST0)
VBCASTCDUP(1,%2,BCAST1)
VBCASTCDUP(2,%2,BCAST2)
VBCASTCDUP(3,%2,BCAST3)
VBCASTCDUP(4,%2,BCAST4) VMULMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(5,%2,BCAST5) VMULMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(6,%2,BCAST6) VMULMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(7,%2,BCAST7) VMULMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(8,%2,BCAST8) VMULMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(9,%2,BCAST9) VMULMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(10,%2,BCAST10) VMULMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(11,%2,BCAST11) VMULMEM(0,%1,BCAST7,Chi_21)
VMULMEM(0,%1,BCAST8,Chi_22)
VMULMEM(0,%1,BCAST9,Chi_30)
VMULMEM(0,%1,BCAST10,Chi_31)
VMULMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
} else {
asm(
VBCASTCDUP(0,%2,BCAST0) VMADDMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(1,%2,BCAST1) VMADDMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(2,%2,BCAST2) VMADDMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(3,%2,BCAST3) VMADDMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(4,%2,BCAST4) VMADDMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(5,%2,BCAST5) VMADDMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(6,%2,BCAST6) VMADDMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(7,%2,BCAST7) VMADDMEM(0,%1,BCAST7,Chi_21)
VBCASTCDUP(8,%2,BCAST8) VMADDMEM(0,%1,BCAST8,Chi_22)
VBCASTCDUP(9,%2,BCAST9) VMADDMEM(0,%1,BCAST9,Chi_30)
VBCASTCDUP(10,%2,BCAST10) VMADDMEM(0,%1,BCAST10,Chi_31)
VBCASTCDUP(11,%2,BCAST11) VMADDMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
}
a0 = a0 + incr;
a1 = a1 + incr;
a2 = a2 + sizeof(typename Simd::scalar_type);
}
}
{
int lexa = s1+LLs*site;
asm (
VSTORE(0,%0,Chi_00) VSTORE(1 ,%0,Chi_01) VSTORE(2 ,%0,Chi_02)
VSTORE(3,%0,Chi_10) VSTORE(4 ,%0,Chi_11) VSTORE(5 ,%0,Chi_12)
VSTORE(6,%0,Chi_20) VSTORE(7 ,%0,Chi_21) VSTORE(8 ,%0,Chi_22)
VSTORE(9,%0,Chi_30) VSTORE(10,%0,Chi_31) VSTORE(11,%0,Chi_32)
: : "r" ((uint64_t)&chi[lexa]) : "memory" );
}
}
}
#undef Chi_00
#undef Chi_01
#undef Chi_02
#undef Chi_10
#undef Chi_11
#undef Chi_12
#undef Chi_20
#undef Chi_21
#undef Chi_22
#undef Chi_30
#undef Chi_31
#undef Chi_32
#undef BCAST0
#undef BCAST1
#undef BCAST2
#undef BCAST3
#undef BCAST4
#undef BCAST5
#undef BCAST6
#undef BCAST7
#undef BCAST8
#undef BCAST9
#undef BCAST10
#undef BCAST11
#endif
};
// Z-mobius version
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternalZAsm(const FermionField& psi, FermionField& chi,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
std::cout << "Error: zMobius not implemented for EOFA" << std::endl;
exit(-1);
};
template<class Impl>
void DomainWallEOFAFermion<Impl>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv)
{
chi.Checkerboard() = psi.Checkerboard();
int Ls = this->Ls;
int LLs = psi.Grid()->_rdimensions[0];
int vol = psi.Grid()->oSites()/LLs;
Vector<iSinglet<Simd> > Matp;
Vector<iSinglet<Simd> > Matm;
Vector<iSinglet<Simd> > *_Matp;
Vector<iSinglet<Simd> > *_Matm;
// MooeeInternalCompute(dag,inv,Matp,Matm);
if(inv && dag){
_Matp = &this->MatpInvDag;
_Matm = &this->MatmInvDag;
}
if(inv && (!dag)){
_Matp = &this->MatpInv;
_Matm = &this->MatmInv;
}
if(!inv){
MooeeInternalCompute(dag, inv, Matp, Matm);
_Matp = &Matp;
_Matm = &Matm;
}
assert(_Matp->size() == Ls*LLs);
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
if(switcheroo<Coeff_t>::iscomplex()){
thread_loop((auto site=0; site<vol; site++),{
MooeeInternalZAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
} else {
thread_loop((auto site=0; site<vol; site++){
MooeeInternalAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
}
this->MooeeInvTime += usecond();
}
#ifdef DOMAIN_WALL_EOFA_DPERP_VEC
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplD);
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplF);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplD);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplF);
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplDF);
INSTANTIATE_DPERP_DWF_EOFA(DomainWallVec5dImplFH);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplDF);
INSTANTIATE_DPERP_DWF_EOFA(ZDomainWallVec5dImplFH);
template void DomainWallEOFAFermion<DomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<DomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<DomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<DomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void DomainWallEOFAFermion<ZDomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
#endif
NAMESPACE_END(Grid);

497
Grid/qcd/action/fermion/implementation/MobiusEOFAFermion.h
View File

@ -1,497 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/MobiusEOFAFermion.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid_Eigen_Dense.h>
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/MobiusEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
MobiusEOFAFermion<Impl>::MobiusEOFAFermion(
GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _mq1, RealD _mq2, RealD _mq3,
RealD _shift, int _pm, RealD _M5,
RealD _b, RealD _c, const ImplParams &p) :
AbstractEOFAFermion<Impl>(_Umu, FiveDimGrid, FiveDimRedBlackGrid,
FourDimGrid, FourDimRedBlackGrid, _mq1, _mq2, _mq3,
_shift, _pm, _M5, _b, _c, p)
{
int Ls = this->Ls;
RealD eps = 1.0;
Approx::zolotarev_data *zdata = Approx::higham(eps, this->Ls);
assert(zdata->n == this->Ls);
std::cout << GridLogMessage << "MobiusEOFAFermion (b=" << _b <<
",c=" << _c << ") with Ls=" << Ls << std::endl;
this->SetCoefficientsTanh(zdata, _b, _c);
std::cout << GridLogMessage << "EOFA parameters: (mq1=" << _mq1 <<
",mq2=" << _mq2 << ",mq3=" << _mq3 << ",shift=" << _shift <<
",pm=" << _pm << ")" << std::endl;
Approx::zolotarev_free(zdata);
if(_shift != 0.0){
SetCoefficientsPrecondShiftOps();
} else {
Mooee_shift.resize(Ls, 0.0);
MooeeInv_shift_lc.resize(Ls, 0.0);
MooeeInv_shift_norm.resize(Ls, 0.0);
MooeeInvDag_shift_lc.resize(Ls, 0.0);
MooeeInvDag_shift_norm.resize(Ls, 0.0);
}
}
/****************************************************************
* Additional EOFA operators only called outside the inverter.
* Since speed is not essential, simple axpby-style
* implementations should be fine.
***************************************************************/
template<class Impl>
void MobiusEOFAFermion<Impl>::Omega(const FermionField& psi, FermionField& Din, int sign, int dag)
{
int Ls = this->Ls;
RealD alpha = this->alpha;
Din = Zero();
if((sign == 1) && (dag == 0)) { // \Omega_{+}
for(int s=0; s<Ls; ++s){
axpby_ssp(Din, 0.0, psi, 2.0*std::pow(1.0-alpha,Ls-s-1)/std::pow(1.0+alpha,Ls-s), psi, s, 0);
}
} else if((sign == -1) && (dag == 0)) { // \Omega_{-}
for(int s=0; s<Ls; ++s){
axpby_ssp(Din, 0.0, psi, 2.0*std::pow(1.0-alpha,s)/std::pow(1.0+alpha,s+1), psi, s, 0);
}
} else if((sign == 1 ) && (dag == 1)) { // \Omega_{+}^{\dagger}
for(int sp=0; sp<Ls; ++sp){
axpby_ssp(Din, 1.0, Din, 2.0*std::pow(1.0-alpha,Ls-sp-1)/std::pow(1.0+alpha,Ls-sp), psi, 0, sp);
}
} else if((sign == -1) && (dag == 1)) { // \Omega_{-}^{\dagger}
for(int sp=0; sp<Ls; ++sp){
axpby_ssp(Din, 1.0, Din, 2.0*std::pow(1.0-alpha,sp)/std::pow(1.0+alpha,sp+1), psi, 0, sp);
}
}
}
// This is the operator relating the usual Ddwf to TWQCD's EOFA Dirac operator (arXiv:1706.05843, Eqn. 6).
// It also relates the preconditioned and unpreconditioned systems described in Appendix B.2.
template<class Impl>
void MobiusEOFAFermion<Impl>::Dtilde(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
RealD b = 0.5 * ( 1.0 + this->alpha );
RealD c = 0.5 * ( 1.0 - this->alpha );
RealD mq1 = this->mq1;
for(int s=0; s<Ls; ++s){
if(s == 0) {
axpby_ssp_pminus(chi, b, psi, -c, psi, s, s+1);
axpby_ssp_pplus (chi, 1.0, chi, mq1*c, psi, s, Ls-1);
} else if(s == (Ls-1)) {
axpby_ssp_pminus(chi, b, psi, mq1*c, psi, s, 0);
axpby_ssp_pplus (chi, 1.0, chi, -c, psi, s, s-1);
} else {
axpby_ssp_pminus(chi, b, psi, -c, psi, s, s+1);
axpby_ssp_pplus (chi, 1.0, chi, -c, psi, s, s-1);
}
}
}
template<class Impl>
void MobiusEOFAFermion<Impl>::DtildeInv(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
RealD m = this->mq1;
RealD c = 0.5 * this->alpha;
RealD d = 0.5;
RealD DtInv_p(0.0), DtInv_m(0.0);
RealD N = std::pow(c+d,Ls) + m*std::pow(c-d,Ls);
FermionField tmp(this->FermionGrid());
for(int s=0; s<Ls; ++s){
for(int sp=0; sp<Ls; ++sp){
DtInv_p = m * std::pow(-1.0,s-sp+1) * std::pow(c-d,Ls+s-sp) / std::pow(c+d,s-sp+1) / N;
DtInv_p += (s < sp) ? 0.0 : std::pow(-1.0,s-sp) * std::pow(c-d,s-sp) / std::pow(c+d,s-sp+1);
DtInv_m = m * std::pow(-1.0,sp-s+1) * std::pow(c-d,Ls+sp-s) / std::pow(c+d,sp-s+1) / N;
DtInv_m += (s > sp) ? 0.0 : std::pow(-1.0,sp-s) * std::pow(c-d,sp-s) / std::pow(c+d,sp-s+1);
if(sp == 0){
axpby_ssp_pplus (tmp, 0.0, tmp, DtInv_p, psi, s, sp);
axpby_ssp_pminus(tmp, 0.0, tmp, DtInv_m, psi, s, sp);
} else {
axpby_ssp_pplus (tmp, 1.0, tmp, DtInv_p, psi, s, sp);
axpby_ssp_pminus(tmp, 1.0, tmp, DtInv_m, psi, s, sp);
}
}}
}
/*****************************************************************************************************/
template<class Impl>
RealD MobiusEOFAFermion<Impl>::M(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->Meooe5D(psi, Din);
this->DW(Din, chi, DaggerNo);
axpby(chi, 1.0, 1.0, chi, psi);
this->M5D(psi, chi);
return(norm2(chi));
}
template<class Impl>
RealD MobiusEOFAFermion<Impl>::Mdag(const FermionField& psi, FermionField& chi)
{
FermionField Din(psi.Grid());
this->DW(psi, Din, DaggerYes);
this->MeooeDag5D(Din, chi);
this->M5Ddag(psi, chi);
axpby(chi, 1.0, 1.0, chi, psi);
return(norm2(chi));
}
/********************************************************************
* Performance critical fermion operators called inside the inverter
********************************************************************/
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = this->mq1;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = this->mq1;
// no shift term
if(this->shift == 0.0){ this->M5D(psi, chi, chi, lower, diag, upper); }
// fused M + shift operation
else{ this->M5D_shift(psi, chi, chi, lower, diag, upper, Mooee_shift); }
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
Vector<Coeff_t> diag(Ls,1.0);
Vector<Coeff_t> upper(Ls,-1.0); upper[Ls-1] = this->mq1;
Vector<Coeff_t> lower(Ls,-1.0); lower[0] = this->mq1;
// no shift term
if(this->shift == 0.0){ this->M5Ddag(psi, chi, chi, lower, diag, upper); }
// fused M + shift operation
else{ this->M5Ddag_shift(psi, chi, chi, lower, diag, upper, Mooee_shift); }
}
// half checkerboard operations
template<class Impl>
void MobiusEOFAFermion<Impl>::Mooee(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
// coefficients of Mooee
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
upper[s] = -this->cee[s];
lower[s] = -this->cee[s];
}
upper[Ls-1] *= -this->mq1;
lower[0] *= -this->mq1;
// no shift term
if(this->shift == 0.0){ this->M5D(psi, psi, chi, lower, diag, upper); }
// fused M + shift operation
else { this->M5D_shift(psi, psi, chi, lower, diag, upper, Mooee_shift); }
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeDag(const FermionField& psi, FermionField& chi)
{
int Ls = this->Ls;
// coefficients of MooeeDag
Vector<Coeff_t> diag = this->bee;
Vector<Coeff_t> upper(Ls);
Vector<Coeff_t> lower(Ls);
for(int s=0; s<Ls; s++){
if(s==0) {
upper[s] = -this->cee[s+1];
lower[s] = this->mq1*this->cee[Ls-1];
} else if(s==(Ls-1)) {
upper[s] = this->mq1*this->cee[0];
lower[s] = -this->cee[s-1];
} else {
upper[s] = -this->cee[s+1];
lower[s] = -this->cee[s-1];
}
}
// no shift term
if(this->shift == 0.0){ this->M5Ddag(psi, psi, chi, lower, diag, upper); }
// fused M + shift operation
else{ this->M5Ddag_shift(psi, psi, chi, lower, diag, upper, Mooee_shift); }
}
/****************************************************************************************/
// Computes coefficients for applying Cayley preconditioned shift operators
// (Mooee + \Delta) --> Mooee_shift
// (Mooee + \Delta)^{-1} --> MooeeInv_shift_lc, MooeeInv_shift_norm
// (Mooee + \Delta)^{-dag} --> MooeeInvDag_shift_lc, MooeeInvDag_shift_norm
// For the latter two cases, the operation takes the form
// [ (Mooee + \Delta)^{-1} \psi ]_{i} = Mooee_{ij} \psi_{j} +
// ( MooeeInv_shift_norm )_{i} ( \sum_{j} [ MooeeInv_shift_lc ]_{j} P_{pm} \psi_{j} )
template<class Impl>
void MobiusEOFAFermion<Impl>::SetCoefficientsPrecondShiftOps()
{
int Ls = this->Ls;
int pm = this->pm;
RealD alpha = this->alpha;
RealD k = this->k;
RealD mq1 = this->mq1;
RealD shift = this->shift;
// Initialize
Mooee_shift.resize(Ls);
MooeeInv_shift_lc.resize(Ls);
MooeeInv_shift_norm.resize(Ls);
MooeeInvDag_shift_lc.resize(Ls);
MooeeInvDag_shift_norm.resize(Ls);
// Construct Mooee_shift
int idx(0);
Coeff_t N = ( (pm == 1) ? 1.0 : -1.0 ) * (2.0*shift*k) *
( std::pow(alpha+1.0,Ls) + mq1*std::pow(alpha-1.0,Ls) );
for(int s=0; s<Ls; ++s){
idx = (pm == 1) ? (s) : (Ls-1-s);
Mooee_shift[idx] = N * std::pow(-1.0,s) * std::pow(alpha-1.0,s) / std::pow(alpha+1.0,Ls+s+1);
}
// Tridiagonal solve for MooeeInvDag_shift_lc
{
Coeff_t m(0.0);
Vector<Coeff_t> d = Mooee_shift;
Vector<Coeff_t> u(Ls,0.0);
Vector<Coeff_t> y(Ls,0.0);
Vector<Coeff_t> q(Ls,0.0);
if(pm == 1){ u[0] = 1.0; }
else{ u[Ls-1] = 1.0; }
// Tridiagonal matrix algorithm + Sherman-Morrison formula
//
// We solve
// ( Mooee' + u \otimes v ) MooeeInvDag_shift_lc = Mooee_shift
// where Mooee' is the tridiagonal part of Mooee_{+}, and
// u = (1,0,...,0) and v = (0,...,0,mq1*cee[0]) are chosen
// so that the outer-product u \otimes v gives the (0,Ls-1)
// entry of Mooee_{+}.
//
// We do this as two solves: Mooee'*y = d and Mooee'*q = u,
// and then construct the solution to the original system
// MooeeInvDag_shift_lc = y - <v,y> / ( 1 + <v,q> ) q
if(pm == 1){
for(int s=1; s<Ls; ++s){
m = -this->cee[s] / this->bee[s-1];
d[s] -= m*d[s-1];
u[s] -= m*u[s-1];
}
}
y[Ls-1] = d[Ls-1] / this->bee[Ls-1];
q[Ls-1] = u[Ls-1] / this->bee[Ls-1];
for(int s=Ls-2; s>=0; --s){
if(pm == 1){
y[s] = d[s] / this->bee[s];
q[s] = u[s] / this->bee[s];
} else {
y[s] = ( d[s] + this->cee[s]*y[s+1] ) / this->bee[s];
q[s] = ( u[s] + this->cee[s]*q[s+1] ) / this->bee[s];
}
}
// Construct MooeeInvDag_shift_lc
for(int s=0; s<Ls; ++s){
if(pm == 1){
MooeeInvDag_shift_lc[s] = y[s] - mq1*this->cee[0]*y[Ls-1] /
(1.0+mq1*this->cee[0]*q[Ls-1]) * q[s];
} else {
MooeeInvDag_shift_lc[s] = y[s] - mq1*this->cee[Ls-1]*y[0] /
(1.0+mq1*this->cee[Ls-1]*q[0]) * q[s];
}
}
// Compute remaining coefficients
N = (pm == 1) ? (1.0 + MooeeInvDag_shift_lc[Ls-1]) : (1.0 + MooeeInvDag_shift_lc[0]);
for(int s=0; s<Ls; ++s){
// MooeeInv_shift_lc
if(pm == 1){ MooeeInv_shift_lc[s] = pow(this->bee[s],s) * pow(this->cee[s],Ls-1-s); }
else { MooeeInv_shift_lc[s] = pow(this->bee[s],Ls-1-s) * pow(this->cee[s],s); }
// MooeeInv_shift_norm
MooeeInv_shift_norm[s] = -MooeeInvDag_shift_lc[s] /
( pow(this->bee[s],Ls) + mq1*pow(this->cee[s],Ls) ) / N;
// MooeeInvDag_shift_norm
if(pm == 1){ MooeeInvDag_shift_norm[s] = -pow(this->bee[s],s) * pow(this->cee[s],(Ls-1-s)) /
( pow(this->bee[s],Ls) + mq1*pow(this->cee[s],Ls) ) / N; }
else{ MooeeInvDag_shift_norm[s] = -pow(this->bee[s],(Ls-1-s)) * pow(this->cee[s],s) /
( pow(this->bee[s],Ls) + mq1*pow(this->cee[s],Ls) ) / N; }
}
}
}
// Recompute coefficients for a different value of shift constant
template<class Impl>
void MobiusEOFAFermion<Impl>::RefreshShiftCoefficients(RealD new_shift)
{
this->shift = new_shift;
if(new_shift != 0.0){
SetCoefficientsPrecondShiftOps();
} else {
int Ls = this->Ls;
Mooee_shift.resize(Ls,0.0);
MooeeInv_shift_lc.resize(Ls,0.0);
MooeeInv_shift_norm.resize(Ls,0.0);
MooeeInvDag_shift_lc.resize(Ls,0.0);
MooeeInvDag_shift_norm.resize(Ls,0.0);
}
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternalCompute(int dag, int inv,
Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
int Ls = this->Ls;
GridBase* grid = this->FermionRedBlackGrid();
int LLs = grid->_rdimensions[0];
if(LLs == Ls){ return; } // Not vectorised in 5th direction
Eigen::MatrixXcd Pplus = Eigen::MatrixXcd::Zero(Ls,Ls);
Eigen::MatrixXcd Pminus = Eigen::MatrixXcd::Zero(Ls,Ls);
for(int s=0; s<Ls; s++){
Pplus(s,s) = this->bee[s];
Pminus(s,s) = this->bee[s];
}
for(int s=0; s<Ls-1; s++){
Pminus(s,s+1) = -this->cee[s];
Pplus(s+1,s) = -this->cee[s+1];
}
Pplus (0,Ls-1) = this->mq1*this->cee[0];
Pminus(Ls-1,0) = this->mq1*this->cee[Ls-1];
if(this->shift != 0.0){
RealD c = 0.5 * this->alpha;
RealD d = 0.5;
RealD N = this->shift * this->k * ( std::pow(c+d,Ls) + this->mq1*std::pow(c-d,Ls) );
if(this->pm == 1) {
for(int s=0; s<Ls; ++s){
Pplus(s,Ls-1) += N * std::pow(-1.0,s) * std::pow(c-d,s) / std::pow(c+d,Ls+s+1);
}
} else {
for(int s=0; s<Ls; ++s){
Pminus(s,0) += N * std::pow(-1.0,s+1) * std::pow(c-d,Ls-1-s) / std::pow(c+d,2*Ls-s);
}
}
}
Eigen::MatrixXcd PplusMat ;
Eigen::MatrixXcd PminusMat;
if(inv) {
PplusMat = Pplus.inverse();
PminusMat = Pminus.inverse();
} else {
PplusMat = Pplus;
PminusMat = Pminus;
}
if(dag){
PplusMat.adjointInPlace();
PminusMat.adjointInPlace();
}
typedef typename SiteHalfSpinor::scalar_type scalar_type;
const int Nsimd = Simd::Nsimd();
Matp.resize(Ls*LLs);
Matm.resize(Ls*LLs);
for(int s2=0; s2<Ls; s2++){
for(int s1=0; s1<LLs; s1++){
int istride = LLs;
int ostride = 1;
Simd Vp;
Simd Vm;
scalar_type *sp = (scalar_type*) &Vp;
scalar_type *sm = (scalar_type*) &Vm;
for(int l=0; l<Nsimd; l++){
if(switcheroo<Coeff_t>::iscomplex()) {
sp[l] = PplusMat (l*istride+s1*ostride,s2);
sm[l] = PminusMat(l*istride+s1*ostride,s2);
} else {
// if real
scalar_type tmp;
tmp = PplusMat (l*istride+s1*ostride,s2);
sp[l] = scalar_type(tmp.real(),tmp.real());
tmp = PminusMat(l*istride+s1*ostride,s2);
sm[l] = scalar_type(tmp.real(),tmp.real());
}
}
Matp[LLs*s2+s1] = Vp;
Matm[LLs*s2+s1] = Vm;
}}
}
FermOpTemplateInstantiate(MobiusEOFAFermion);
GparityFermOpTemplateInstantiate(MobiusEOFAFermion);
NAMESPACE_END(Grid);

0
Grid/qcd/action/fermion/MobiusEOFAFermioncache.cc → Grid/qcd/action/fermion/implementation/MobiusEOFAFermionCache.h
View File

445
Grid/qcd/action/fermion/implementation/MobiusEOFAFermioncache.h
View File

@ -1,445 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/MobiusEOFAFermioncache.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/MobiusEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
for(int s=0; s<Ls; s++){
auto tmp = psi[0];
if(s==0){
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5m(tmp, psi[ss+0]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper,
Vector<Coeff_t> &shift_coeffs)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int shift_s = (this->pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
for(int s=0; s<Ls; s++){
auto tmp = psi[0];
if(s==0){
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5m(tmp, psi[ss+0]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5m(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5p(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
if(this->pm == 1){ spProj5p(tmp, psi[ss+shift_s]); }
else{ spProj5m(tmp, psi[ss+shift_s]); }
chi[ss+s] = chi[ss+s] + shift_coeffs[s]*tmp;
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
auto tmp = psi[0];
for(int s=0; s<Ls; s++){
if(s==0) {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5p(tmp, psi[ss+0]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField &psi_i, const FermionField &phi_i, FermionField &chi_i,
Vector<Coeff_t> &lower, Vector<Coeff_t> &diag, Vector<Coeff_t> &upper,
Vector<Coeff_t> &shift_coeffs)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
int shift_s = (this->pm == 1) ? (Ls-1) : 0; // s-component modified by shift operator
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
assert(phi.Checkerboard() == psi.Checkerboard());
// Flops = 6.0*(Nc*Ns) *Ls*vol
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
chi[ss+Ls-1] = Zero();
auto tmp = psi[0];
for(int s=0; s<Ls; s++){
if(s==0) {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+Ls-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else if(s==(Ls-1)) {
spProj5p(tmp, psi[ss+0]);
chi[ss+s] = chi[ss+s] + diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
} else {
spProj5p(tmp, psi[ss+s+1]);
chi[ss+s] = diag[s]*phi[ss+s] + upper[s]*tmp;
spProj5m(tmp, psi[ss+s-1]);
chi[ss+s] = chi[ss+s] + lower[s]*tmp;
}
if(this->pm == 1){ spProj5p(tmp, psi[ss+s]); }
else{ spProj5m(tmp, psi[ss+s]); }
chi[ss+shift_s] = chi[ss+shift_s] + shift_coeffs[s]*tmp;
}
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField &psi_i, FermionField &chi_i)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
if(this->shift != 0.0){ MooeeInv_shift(psi_i,chi_i); return; }
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
auto tmp = psi[0];
// Apply (L^{\prime})^{-1}
chi[ss] = psi[ss]; // chi[0]=psi[0]
for(int s=1; s<Ls; s++){
spProj5p(tmp, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - this->lee[s-1]*tmp;
}
// L_m^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
spProj5m(tmp, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - this->leem[s]*tmp;
}
// U_m^{-1} D^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[s] + 1/d chi[s]
spProj5p(tmp, chi[ss+Ls-1]);
chi[ss+s] = (1.0/this->dee[s])*chi[ss+s] - (this->ueem[s]/this->dee[Ls-1])*tmp;
}
chi[ss+Ls-1] = (1.0/this->dee[Ls-1])*chi[ss+Ls-1];
// Apply U^{-1}
for(int s=Ls-2; s>=0; s--){
spProj5m(tmp, chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - this->uee[s]*tmp;
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField &psi_i, FermionField &chi_i)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
auto tmp1 = psi[0];
auto tmp2 = psi[0];
auto tmp2_spProj = psi[0];
// Apply (L^{\prime})^{-1} and accumulate MooeeInv_shift_lc[j]*psi[j] in tmp2
chi[ss] = psi[ss]; // chi[0]=psi[0]
tmp2 = MooeeInv_shift_lc[0]*psi[ss];
for(int s=1; s<Ls; s++){
spProj5p(tmp1, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - this->lee[s-1]*tmp1;
tmp2 = tmp2 + MooeeInv_shift_lc[s]*psi[ss+s];
}
if(this->pm == 1){ spProj5p(tmp2_spProj, tmp2);}
else{ spProj5m(tmp2_spProj, tmp2); }
// L_m^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
spProj5m(tmp1, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - this->leem[s]*tmp1;
}
// U_m^{-1} D^{-1}
for(int s=0; s<Ls-1; s++){ // Chi[s] + 1/d chi[s]
spProj5p(tmp1, chi[ss+Ls-1]);
chi[ss+s] = (1.0/this->dee[s])*chi[ss+s] - (this->ueem[s]/this->dee[Ls-1])*tmp1;
}
// chi[ss+Ls-1] = (1.0/this->dee[Ls-1])*chi[ss+Ls-1] + MooeeInv_shift_norm[Ls-1]*tmp2_spProj;
chi[ss+Ls-1] = (1.0/this->dee[Ls-1])*chi[ss+Ls-1];
spProj5m(tmp1, chi[ss+Ls-1]);
chi[ss+Ls-1] = chi[ss+Ls-1] + MooeeInv_shift_norm[Ls-1]*tmp2_spProj;
// Apply U^{-1} and add shift term
for(int s=Ls-2; s>=0; s--){
chi[ss+s] = chi[ss+s] - this->uee[s]*tmp1;
spProj5m(tmp1, chi[ss+s]);
chi[ss+s] = chi[ss+s] + MooeeInv_shift_norm[s]*tmp2_spProj;
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField &psi_i, FermionField &chi_i)
{
if(this->shift != 0.0){ MooeeInvDag_shift(psi_i,chi_i); return; }
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
int Ls = this->Ls;
auto psi = psi_i.View();
auto chi = chi_i.View();
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
auto tmp = psi[0];
// Apply (U^{\prime})^{-dag}
chi[ss] = psi[ss];
for(int s=1; s<Ls; s++){
spProj5m(tmp, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - this->uee[s-1]*tmp;
}
// U_m^{-\dag}
for(int s=0; s<Ls-1; s++){
spProj5p(tmp, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - this->ueem[s]*tmp;
}
// L_m^{-\dag} D^{-dag}
for(int s=0; s<Ls-1; s++){
spProj5m(tmp, chi[ss+Ls-1]);
chi[ss+s] = (1.0/this->dee[s])*chi[ss+s] - (this->leem[s]/this->dee[Ls-1])*tmp;
}
chi[ss+Ls-1] = (1.0/this->dee[Ls-1])*chi[ss+Ls-1];
// Apply L^{-dag}
for(int s=Ls-2; s>=0; s--){
spProj5p(tmp, chi[ss+s+1]);
chi[ss+s] = chi[ss+s] - this->lee[s]*tmp;
}
});
this->MooeeInvTime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField &psi_i, FermionField &chi_i)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase *grid = psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=Ls),{
auto tmp1 = psi[0];
auto tmp2 = psi[0];
auto tmp2_spProj = psi[0];
// Apply (U^{\prime})^{-dag} and accumulate MooeeInvDag_shift_lc[j]*psi[j] in tmp2
chi[ss] = psi[ss];
tmp2 = MooeeInvDag_shift_lc[0]*psi[ss];
for(int s=1; s<Ls; s++){
spProj5m(tmp1, chi[ss+s-1]);
chi[ss+s] = psi[ss+s] - this->uee[s-1]*tmp1;
tmp2 = tmp2 + MooeeInvDag_shift_lc[s]*psi[ss+s];
}
if(this->pm == 1){ spProj5p(tmp2_spProj, tmp2);}
else{ spProj5m(tmp2_spProj, tmp2); }
// U_m^{-\dag}
for(int s=0; s<Ls-1; s++){
spProj5p(tmp1, chi[ss+s]);
chi[ss+Ls-1] = chi[ss+Ls-1] - this->ueem[s]*tmp1;
}
// L_m^{-\dag} D^{-dag}
for(int s=0; s<Ls-1; s++){
spProj5m(tmp1, chi[ss+Ls-1]);
chi[ss+s] = (1.0/this->dee[s])*chi[ss+s] - (this->leem[s]/this->dee[Ls-1])*tmp1;
}
chi[ss+Ls-1] = (1.0/this->dee[Ls-1])*chi[ss+Ls-1];
spProj5p(tmp1, chi[ss+Ls-1]);
chi[ss+Ls-1] = chi[ss+Ls-1] + MooeeInvDag_shift_norm[Ls-1]*tmp2_spProj;
// Apply L^{-dag}
for(int s=Ls-2; s>=0; s--){
chi[ss+s] = chi[ss+s] - this->lee[s]*tmp1;
spProj5p(tmp1, chi[ss+s]);
chi[ss+s] = chi[ss+s] + MooeeInvDag_shift_norm[s]*tmp2_spProj;
}
});
this->MooeeInvTime += usecond();
}
#ifdef MOBIUS_EOFA_DPERP_CACHE
INSTANTIATE_DPERP_MOBIUS_EOFA(WilsonImplF);
INSTANTIATE_DPERP_MOBIUS_EOFA(WilsonImplD);
INSTANTIATE_DPERP_MOBIUS_EOFA(GparityWilsonImplF);
INSTANTIATE_DPERP_MOBIUS_EOFA(GparityWilsonImplD);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZWilsonImplF);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZWilsonImplD);
INSTANTIATE_DPERP_MOBIUS_EOFA(WilsonImplFH);
INSTANTIATE_DPERP_MOBIUS_EOFA(WilsonImplDF);
INSTANTIATE_DPERP_MOBIUS_EOFA(GparityWilsonImplFH);
INSTANTIATE_DPERP_MOBIUS_EOFA(GparityWilsonImplDF);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZWilsonImplFH);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZWilsonImplDF);
#endif
NAMESPACE_END(Grid);

998
Grid/qcd/action/fermion/implementation/MobiusEOFAFermionvec.h
View File

@ -1,998 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/MobiusEOFAFermionvec.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/MobiusEOFAFermion.h>
NAMESPACE_BEGIN(Grid);
/*
* Dense matrix versions of routines
*/
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInv(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInv_shift(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerNo, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInvDag(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInvDag_shift(const FermionField& psi, FermionField& chi)
{
this->MooeeInternal(psi, chi, DaggerYes, InverseYes);
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
assert(Nc == 3);
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5m(hp, psi[ss+vp]);
spProj5p(hm, psi[ss+vm]);
if (vp <= v){ rotate(hp, hp, 1); }
if (vm >= v){ rotate(hm, hm, nsimd-1); }
hp = 0.5*hp;
hm = 0.5*hm;
spRecon5m(fp, hp);
spRecon5p(fm, hm);
chi[ss+v] = d[v]*phi[ss+v];
chi[ss+v] = chi[ss+v] + u[v]*fp;
chi[ss+v] = chi[ss+v] + l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(2)(0);
Simd hp_01 = psi[ss+vp]()(2)(1);
Simd hp_02 = psi[ss+vp]()(2)(2);
Simd hp_10 = psi[ss+vp]()(3)(0);
Simd hp_11 = psi[ss+vp]()(3)(1);
Simd hp_12 = psi[ss+vp]()(3)(2);
Simd hm_00 = psi[ss+vm]()(0)(0);
Simd hm_01 = psi[ss+vm]()(0)(1);
Simd hm_02 = psi[ss+vm]()(0)(2);
Simd hm_10 = psi[ss+vm]()(1)(0);
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
if(vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
// Can force these to real arithmetic and save 2x.
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5D_shift(const FermionField& psi_i, const FermionField& phi_i,
FermionField& chi_i, Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
Vector<Coeff_t>& shift_coeffs)
{
#if 0
auto & psi = psi_i;
auto & phi = phi_i;
auto & chi = chi_i;
this->M5D(psi, phi, chi, lower, diag, upper);
// FIXME: possible gain from vectorizing shift operation as well?
Coeff_t one(1.0);
int Ls = this->Ls;
for(int s=0; s<Ls; s++){
if(this->pm == 1){ axpby_ssp_pplus(chi, one, chi, shift_coeffs[s], psi, s, Ls-1); }
else{ axpby_ssp_pminus(chi, one, chi, shift_coeffs[s], psi, s, 0); }
}
#else
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
const int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
Vector<iSinglet<Simd>> s(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
scalar_type* s_p = (scalar_type*) &s[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
s_p[ss] = shift_coeffs[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
assert(Nc == 3);
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
int vs = (this->pm == 1) ? LLs-1 : 0;
Simd hs_00 = (this->pm == 1) ? psi[ss+vs]()(2)(0) : psi[ss+vs]()(0)(0);
Simd hs_01 = (this->pm == 1) ? psi[ss+vs]()(2)(1) : psi[ss+vs]()(0)(1);
Simd hs_02 = (this->pm == 1) ? psi[ss+vs]()(2)(2) : psi[ss+vs]()(0)(2);
Simd hs_10 = (this->pm == 1) ? psi[ss+vs]()(3)(0) : psi[ss+vs]()(1)(0);
Simd hs_11 = (this->pm == 1) ? psi[ss+vs]()(3)(1) : psi[ss+vs]()(1)(1);
Simd hs_12 = (this->pm == 1) ? psi[ss+vs]()(3)(2) : psi[ss+vs]()(1)(2);
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(2)(0);
Simd hp_01 = psi[ss+vp]()(2)(1);
Simd hp_02 = psi[ss+vp]()(2)(2);
Simd hp_10 = psi[ss+vp]()(3)(0);
Simd hp_11 = psi[ss+vp]()(3)(1);
Simd hp_12 = psi[ss+vp]()(3)(2);
Simd hm_00 = psi[ss+vm]()(0)(0);
Simd hm_01 = psi[ss+vm]()(0)(1);
Simd hm_02 = psi[ss+vm]()(0)(2);
Simd hm_10 = psi[ss+vm]()(1)(0);
Simd hm_11 = psi[ss+vm]()(1)(1);
Simd hm_12 = psi[ss+vm]()(1)(2);
if(vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(this->pm == 1 && vs <= v){
hs_00.v = Optimization::Rotate::tRotate<2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2>(hs_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
if(this->pm == -1 && vs >= v){
hs_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_12.v);
}
// Can force these to real arithmetic and save 2x.
Simd p_00 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00);
Simd p_01 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01);
Simd p_02 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02);
Simd p_10 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10);
Simd p_11 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11);
Simd p_12 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12);
Simd p_20 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_21 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_22 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_30 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_31 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_32 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
});
this->M5Dtime += usecond();
#endif
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag(const FermionField& psi_i, const FermionField& phi_i,FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper)
{
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
#if 0
alignas(64) SiteHalfSpinor hp;
alignas(64) SiteHalfSpinor hm;
alignas(64) SiteSpinor fp;
alignas(64) SiteSpinor fm;
for(int v=0; v<LLs; v++){
int vp = (v+1)%LLs;
int vm = (v+LLs-1)%LLs;
spProj5p(hp, psi[ss+vp]);
spProj5m(hm, psi[ss+vm]);
if(vp <= v){ rotate(hp, hp, 1); }
if(vm >= v){ rotate(hm, hm, nsimd-1); }
hp = hp*0.5;
hm = hm*0.5;
spRecon5p(fp, hp);
spRecon5m(fm, hm);
chi[ss+v] = d[v]*phi[ss+v]+u[v]*fp;
chi[ss+v] = chi[ss+v] +l[v]*fm;
}
#else
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if (vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
Simd p_00 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_01 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_02 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_10 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_11 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_12 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
Simd p_20 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00);
Simd p_21 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01);
Simd p_22 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02);
Simd p_30 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10);
Simd p_31 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11);
Simd p_32 = switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
#endif
});
this->M5Dtime += usecond();
}
template<class Impl>
void MobiusEOFAFermion<Impl>::M5Ddag_shift(const FermionField& psi_i, const FermionField& phi_i, FermionField& chi_i,
Vector<Coeff_t>& lower, Vector<Coeff_t>& diag, Vector<Coeff_t>& upper,
Vector<Coeff_t>& shift_coeffs)
{
#if 0
auto & psi = psi_i;
auto & phi = phi_i;
auto & chi = chi_i;
this->M5Ddag(psi, phi, chi, lower, diag, upper);
// FIXME: possible gain from vectorizing shift operation as well?
Coeff_t one(1.0);
int Ls = this->Ls;
for(int s=0; s<Ls; s++){
if(this->pm == 1){ axpby_ssp_pplus(chi, one, chi, shift_coeffs[s], psi, Ls-1, s); }
else{ axpby_ssp_pminus(chi, one, chi, shift_coeffs[s], psi, 0, s); }
}
#else
chi_i.Checkerboard() = psi_i.Checkerboard();
GridBase* grid = psi_i.Grid();
auto psi = psi_i.View();
auto phi = phi_i.View();
auto chi = chi_i.View();
int Ls = this->Ls;
int LLs = grid->_rdimensions[0];
int nsimd = Simd::Nsimd();
Vector<iSinglet<Simd>> u(LLs);
Vector<iSinglet<Simd>> l(LLs);
Vector<iSinglet<Simd>> d(LLs);
Vector<iSinglet<Simd>> s(LLs);
assert(Ls/LLs == nsimd);
assert(phi.Checkerboard() == psi.Checkerboard());
// just directly address via type pun
typedef typename Simd::scalar_type scalar_type;
scalar_type* u_p = (scalar_type*) &u[0];
scalar_type* l_p = (scalar_type*) &l[0];
scalar_type* d_p = (scalar_type*) &d[0];
scalar_type* s_p = (scalar_type*) &s[0];
for(int o=0; o<LLs; o++){ // outer
for(int i=0; i<nsimd; i++){ //inner
int s = o + i*LLs;
int ss = o*nsimd + i;
u_p[ss] = upper[s];
l_p[ss] = lower[s];
d_p[ss] = diag[s];
s_p[ss] = shift_coeffs[s];
}}
this->M5Dcalls++;
this->M5Dtime -= usecond();
thread_loop( (int ss=0; ss<grid->oSites(); ss+=LLs),{ // adds LLs
int vs = (this->pm == 1) ? LLs-1 : 0;
Simd hs_00 = (this->pm == 1) ? psi[ss+vs]()(0)(0) : psi[ss+vs]()(2)(0);
Simd hs_01 = (this->pm == 1) ? psi[ss+vs]()(0)(1) : psi[ss+vs]()(2)(1);
Simd hs_02 = (this->pm == 1) ? psi[ss+vs]()(0)(2) : psi[ss+vs]()(2)(2);
Simd hs_10 = (this->pm == 1) ? psi[ss+vs]()(1)(0) : psi[ss+vs]()(3)(0);
Simd hs_11 = (this->pm == 1) ? psi[ss+vs]()(1)(1) : psi[ss+vs]()(3)(1);
Simd hs_12 = (this->pm == 1) ? psi[ss+vs]()(1)(2) : psi[ss+vs]()(3)(2);
for(int v=0; v<LLs; v++){
vprefetch(psi[ss+v+LLs]);
int vp = (v == LLs-1) ? 0 : v+1;
int vm = (v == 0 ) ? LLs-1 : v-1;
Simd hp_00 = psi[ss+vp]()(0)(0);
Simd hp_01 = psi[ss+vp]()(0)(1);
Simd hp_02 = psi[ss+vp]()(0)(2);
Simd hp_10 = psi[ss+vp]()(1)(0);
Simd hp_11 = psi[ss+vp]()(1)(1);
Simd hp_12 = psi[ss+vp]()(1)(2);
Simd hm_00 = psi[ss+vm]()(2)(0);
Simd hm_01 = psi[ss+vm]()(2)(1);
Simd hm_02 = psi[ss+vm]()(2)(2);
Simd hm_10 = psi[ss+vm]()(3)(0);
Simd hm_11 = psi[ss+vm]()(3)(1);
Simd hm_12 = psi[ss+vm]()(3)(2);
if (vp <= v){
hp_00.v = Optimization::Rotate::tRotate<2>(hp_00.v);
hp_01.v = Optimization::Rotate::tRotate<2>(hp_01.v);
hp_02.v = Optimization::Rotate::tRotate<2>(hp_02.v);
hp_10.v = Optimization::Rotate::tRotate<2>(hp_10.v);
hp_11.v = Optimization::Rotate::tRotate<2>(hp_11.v);
hp_12.v = Optimization::Rotate::tRotate<2>(hp_12.v);
}
if(this->pm == 1 && vs <= v){
hs_00.v = Optimization::Rotate::tRotate<2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2>(hs_12.v);
}
if(vm >= v){
hm_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_00.v);
hm_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_01.v);
hm_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_02.v);
hm_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_10.v);
hm_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_11.v);
hm_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hm_12.v);
}
if(this->pm == -1 && vs >= v){
hs_00.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_00.v);
hs_01.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_01.v);
hs_02.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_02.v);
hs_10.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_10.v);
hs_11.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_11.v);
hs_12.v = Optimization::Rotate::tRotate<2*Simd::Nsimd()-2>(hs_12.v);
}
Simd p_00 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_00);
Simd p_01 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_01);
Simd p_02 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(0)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_02);
Simd p_10 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(0)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_10);
Simd p_11 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(1)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_11);
Simd p_12 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(1)(2)) + switcheroo<Coeff_t>::mult(u[v]()()(), hp_12);
Simd p_20 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_00)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_00);
Simd p_21 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_01)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_01);
Simd p_22 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(2)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_02)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_02);
Simd p_30 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(0)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_10)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_10);
Simd p_31 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(1)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_11)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_11);
Simd p_32 = (this->pm == 1) ? switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
: switcheroo<Coeff_t>::mult(d[v]()()(), phi[ss+v]()(3)(2)) + switcheroo<Coeff_t>::mult(l[v]()()(), hm_12)
+ switcheroo<Coeff_t>::mult(s[v]()()(), hs_12);
vstream(chi[ss+v]()(0)(0), p_00);
vstream(chi[ss+v]()(0)(1), p_01);
vstream(chi[ss+v]()(0)(2), p_02);
vstream(chi[ss+v]()(1)(0), p_10);
vstream(chi[ss+v]()(1)(1), p_11);
vstream(chi[ss+v]()(1)(2), p_12);
vstream(chi[ss+v]()(2)(0), p_20);
vstream(chi[ss+v]()(2)(1), p_21);
vstream(chi[ss+v]()(2)(2), p_22);
vstream(chi[ss+v]()(3)(0), p_30);
vstream(chi[ss+v]()(3)(1), p_31);
vstream(chi[ss+v]()(3)(2), p_32);
}
});
this->M5Dtime += usecond();
#endif
}
#ifdef AVX512
#include<simd/Intel512common.h>
#include<simd/Intel512avx.h>
#include<simd/Intel512single.h>
#endif
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternalAsm(const FermionField& psi_i, FermionField& chi_i,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
auto psi = psi_i.View();
auto chi = chi_i.View();
#ifndef AVX512
{
SiteHalfSpinor BcastP;
SiteHalfSpinor BcastM;
SiteHalfSpinor SiteChiP;
SiteHalfSpinor SiteChiM;
// Ls*Ls * 2 * 12 * vol flops
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
for(int l=0; l < Simd::Nsimd(); l++){ // simd lane
int s = s2 + l*LLs;
int lex = s2 + LLs*site;
if( s2==0 && l==0 ){
SiteChiP=Zero();
SiteChiM=Zero();
}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastP()(sp)(co), psi[lex]()(sp)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vbroadcast(BcastM()(sp)(co), psi[lex]()(sp+2)(co), l);
}}
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
SiteChiP()(sp)(co) = real_madd(Matp[LLs*s+s1]()()(), BcastP()(sp)(co), SiteChiP()(sp)(co)); // 1100 us.
SiteChiM()(sp)(co) = real_madd(Matm[LLs*s+s1]()()(), BcastM()(sp)(co), SiteChiM()(sp)(co)); // each found by commenting out
}}
}}
{
int lex = s1 + LLs*site;
for(int sp=0; sp<2; sp++){
for(int co=0; co<Nc; co++){
vstream(chi[lex]()(sp)(co), SiteChiP()(sp)(co));
vstream(chi[lex]()(sp+2)(co), SiteChiM()(sp)(co));
}}
}
}
}
#else
{
// pointers
// MASK_REGS;
#define Chi_00 %%zmm1
#define Chi_01 %%zmm2
#define Chi_02 %%zmm3
#define Chi_10 %%zmm4
#define Chi_11 %%zmm5
#define Chi_12 %%zmm6
#define Chi_20 %%zmm7
#define Chi_21 %%zmm8
#define Chi_22 %%zmm9
#define Chi_30 %%zmm10
#define Chi_31 %%zmm11
#define Chi_32 %%zmm12
#define BCAST0 %%zmm13
#define BCAST1 %%zmm14
#define BCAST2 %%zmm15
#define BCAST3 %%zmm16
#define BCAST4 %%zmm17
#define BCAST5 %%zmm18
#define BCAST6 %%zmm19
#define BCAST7 %%zmm20
#define BCAST8 %%zmm21
#define BCAST9 %%zmm22
#define BCAST10 %%zmm23
#define BCAST11 %%zmm24
int incr = LLs*LLs*sizeof(iSinglet<Simd>);
for(int s1=0; s1<LLs; s1++){
for(int s2=0; s2<LLs; s2++){
int lex = s2 + LLs*site;
uint64_t a0 = (uint64_t) &Matp[LLs*s2+s1]; // should be cacheable
uint64_t a1 = (uint64_t) &Matm[LLs*s2+s1];
uint64_t a2 = (uint64_t) &psi[lex];
for(int l=0; l<Simd::Nsimd(); l++){ // simd lane
if((s2+l)==0) {
asm(
VPREFETCH1(0,%2) VPREFETCH1(0,%1)
VPREFETCH1(12,%2) VPREFETCH1(13,%2)
VPREFETCH1(14,%2) VPREFETCH1(15,%2)
VBCASTCDUP(0,%2,BCAST0)
VBCASTCDUP(1,%2,BCAST1)
VBCASTCDUP(2,%2,BCAST2)
VBCASTCDUP(3,%2,BCAST3)
VBCASTCDUP(4,%2,BCAST4) VMULMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(5,%2,BCAST5) VMULMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(6,%2,BCAST6) VMULMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(7,%2,BCAST7) VMULMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(8,%2,BCAST8) VMULMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(9,%2,BCAST9) VMULMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(10,%2,BCAST10) VMULMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(11,%2,BCAST11) VMULMEM(0,%1,BCAST7,Chi_21)
VMULMEM(0,%1,BCAST8,Chi_22)
VMULMEM(0,%1,BCAST9,Chi_30)
VMULMEM(0,%1,BCAST10,Chi_31)
VMULMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
} else {
asm(
VBCASTCDUP(0,%2,BCAST0) VMADDMEM(0,%0,BCAST0,Chi_00)
VBCASTCDUP(1,%2,BCAST1) VMADDMEM(0,%0,BCAST1,Chi_01)
VBCASTCDUP(2,%2,BCAST2) VMADDMEM(0,%0,BCAST2,Chi_02)
VBCASTCDUP(3,%2,BCAST3) VMADDMEM(0,%0,BCAST3,Chi_10)
VBCASTCDUP(4,%2,BCAST4) VMADDMEM(0,%0,BCAST4,Chi_11)
VBCASTCDUP(5,%2,BCAST5) VMADDMEM(0,%0,BCAST5,Chi_12)
VBCASTCDUP(6,%2,BCAST6) VMADDMEM(0,%1,BCAST6,Chi_20)
VBCASTCDUP(7,%2,BCAST7) VMADDMEM(0,%1,BCAST7,Chi_21)
VBCASTCDUP(8,%2,BCAST8) VMADDMEM(0,%1,BCAST8,Chi_22)
VBCASTCDUP(9,%2,BCAST9) VMADDMEM(0,%1,BCAST9,Chi_30)
VBCASTCDUP(10,%2,BCAST10) VMADDMEM(0,%1,BCAST10,Chi_31)
VBCASTCDUP(11,%2,BCAST11) VMADDMEM(0,%1,BCAST11,Chi_32)
: : "r" (a0), "r" (a1), "r" (a2) );
}
a0 = a0 + incr;
a1 = a1 + incr;
a2 = a2 + sizeof(typename Simd::scalar_type);
}
}
{
int lexa = s1+LLs*site;
asm (
VSTORE(0,%0,Chi_00) VSTORE(1 ,%0,Chi_01) VSTORE(2 ,%0,Chi_02)
VSTORE(3,%0,Chi_10) VSTORE(4 ,%0,Chi_11) VSTORE(5 ,%0,Chi_12)
VSTORE(6,%0,Chi_20) VSTORE(7 ,%0,Chi_21) VSTORE(8 ,%0,Chi_22)
VSTORE(9,%0,Chi_30) VSTORE(10,%0,Chi_31) VSTORE(11,%0,Chi_32)
: : "r" ((uint64_t)&chi[lexa]) : "memory" );
}
}
}
#undef Chi_00
#undef Chi_01
#undef Chi_02
#undef Chi_10
#undef Chi_11
#undef Chi_12
#undef Chi_20
#undef Chi_21
#undef Chi_22
#undef Chi_30
#undef Chi_31
#undef Chi_32
#undef BCAST0
#undef BCAST1
#undef BCAST2
#undef BCAST3
#undef BCAST4
#undef BCAST5
#undef BCAST6
#undef BCAST7
#undef BCAST8
#undef BCAST9
#undef BCAST10
#undef BCAST11
#endif
};
// Z-mobius version
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternalZAsm(const FermionField& psi, FermionField& chi,
int LLs, int site, Vector<iSinglet<Simd> >& Matp, Vector<iSinglet<Simd> >& Matm)
{
std::cout << "Error: zMobius not implemented for EOFA" << std::endl;
exit(-1);
};
template<class Impl>
void MobiusEOFAFermion<Impl>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv)
{
chi.Checkerboard() = psi.Checkerboard();
int Ls = this->Ls;
int LLs = psi.Grid()->_rdimensions[0];
int vol = psi.Grid()->oSites()/LLs;
Vector<iSinglet<Simd>> Matp;
Vector<iSinglet<Simd>> Matm;
Vector<iSinglet<Simd>>* _Matp;
Vector<iSinglet<Simd>>* _Matm;
// MooeeInternalCompute(dag,inv,Matp,Matm);
if(inv && dag){
_Matp = &this->MatpInvDag;
_Matm = &this->MatmInvDag;
}
if(inv && (!dag)){
_Matp = &this->MatpInv;
_Matm = &this->MatmInv;
}
if(!inv){
MooeeInternalCompute(dag, inv, Matp, Matm);
_Matp = &Matp;
_Matm = &Matm;
}
assert(_Matp->size() == Ls*LLs);
this->MooeeInvCalls++;
this->MooeeInvTime -= usecond();
if(switcheroo<Coeff_t>::iscomplex()){
thread_loop( (auto site=0; site<vol; site++),{
MooeeInternalZAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
} else {
thread_loop( (auto site=0; site<vol; site++),{
MooeeInternalAsm(psi, chi, LLs, site, *_Matp, *_Matm);
});
}
this->MooeeInvTime += usecond();
}
#ifdef MOBIUS_EOFA_DPERP_VEC
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplD);
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplF);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplD);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplF);
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplDF);
INSTANTIATE_DPERP_MOBIUS_EOFA(DomainWallVec5dImplFH);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplDF);
INSTANTIATE_DPERP_MOBIUS_EOFA(ZDomainWallVec5dImplFH);
template void MobiusEOFAFermion<DomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<DomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplD>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<DomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<DomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplFH>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
template void MobiusEOFAFermion<ZDomainWallVec5dImplDF>::MooeeInternal(const FermionField& psi, FermionField& chi, int dag, int inv);
#endif
NAMESPACE_END(Grid);

242
Grid/qcd/action/fermion/implementation/WilsonCloverFermion.h
View File

@ -1,242 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
Copyright (C) 2017
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
#include <Grid/qcd/spin/Dirac.h>
NAMESPACE_BEGIN(Grid);
// *NOT* EO
template <class Impl>
RealD WilsonCloverFermion<Impl>::M(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
// Wilson term
out.Checkerboard() = in.Checkerboard();
this->Dhop(in, out, DaggerNo);
// Clover term
Mooee(in, temp);
out += temp;
return norm2(out);
}
template <class Impl>
RealD WilsonCloverFermion<Impl>::Mdag(const FermionField &in, FermionField &out)
{
FermionField temp(out.Grid());
// Wilson term
out.Checkerboard() = in.Checkerboard();
this->Dhop(in, out, DaggerYes);
// Clover term
MooeeDag(in, temp);
out += temp;
return norm2(out);
}
template <class Impl>
void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
{
WilsonFermion<Impl>::ImportGauge(_Umu);
GridBase *grid = _Umu.Grid();
typename Impl::GaugeLinkField Bx(grid), By(grid), Bz(grid), Ex(grid), Ey(grid), Ez(grid);
// Compute the field strength terms mu>nu
WilsonLoops<Impl>::FieldStrength(Bx, _Umu, Zdir, Ydir);
WilsonLoops<Impl>::FieldStrength(By, _Umu, Zdir, Xdir);
WilsonLoops<Impl>::FieldStrength(Bz, _Umu, Ydir, Xdir);
WilsonLoops<Impl>::FieldStrength(Ex, _Umu, Tdir, Xdir);
WilsonLoops<Impl>::FieldStrength(Ey, _Umu, Tdir, Ydir);
WilsonLoops<Impl>::FieldStrength(Ez, _Umu, Tdir, Zdir);
// Compute the Clover Operator acting on Colour and Spin
// multiply here by the clover coefficients for the anisotropy
CloverTerm = fillCloverYZ(Bx) * csw_r;
CloverTerm += fillCloverXZ(By) * csw_r;
CloverTerm += fillCloverXY(Bz) * csw_r;
CloverTerm += fillCloverXT(Ex) * csw_t;
CloverTerm += fillCloverYT(Ey) * csw_t;
CloverTerm += fillCloverZT(Ez) * csw_t;
CloverTerm += diag_mass;
int lvol = _Umu.Grid()->lSites();
int DimRep = Impl::Dimension;
Eigen::MatrixXcd EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
Eigen::MatrixXcd EigenInvCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
Coordinate lcoor;
typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
for (int site = 0; site < lvol; site++)
{
grid->LocalIndexToLocalCoor(site, lcoor);
EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
peekLocalSite(Qx, CloverTerm, lcoor);
Qxinv = Zero();
//if (csw!=0){
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++){
auto zz = Qx()(j, k)(a, b);
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
}
// if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
EigenInvCloverOp = EigenCloverOp.inverse();
//std::cout << EigenInvCloverOp << std::endl;
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++)
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
// if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
// }
pokeLocalSite(Qxinv, CloverTermInv, lcoor);
}
// Separate the even and odd parts
pickCheckerboard(Even, CloverTermEven, CloverTerm);
pickCheckerboard(Odd, CloverTermOdd, CloverTerm);
pickCheckerboard(Even, CloverTermDagEven, adj(CloverTerm));
pickCheckerboard(Odd, CloverTermDagOdd, adj(CloverTerm));
pickCheckerboard(Even, CloverTermInvEven, CloverTermInv);
pickCheckerboard(Odd, CloverTermInvOdd, CloverTermInv);
pickCheckerboard(Even, CloverTermInvDagEven, adj(CloverTermInv));
pickCheckerboard(Odd, CloverTermInvDagOdd, adj(CloverTermInv));
}
template <class Impl>
void WilsonCloverFermion<Impl>::Mooee(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerNo, InverseNo);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerYes, InverseNo);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerNo, InverseYes);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out)
{
this->MooeeInternal(in, out, DaggerYes, InverseYes);
}
template <class Impl>
void WilsonCloverFermion<Impl>::MooeeInternal(const FermionField &in, FermionField &out, int dag, int inv)
{
out.Checkerboard() = in.Checkerboard();
CloverFieldType *Clover;
assert(in.Checkerboard() == Odd || in.Checkerboard() == Even);
if (dag)
{
if (in.Grid()->_isCheckerBoarded)
{
if (in.Checkerboard() == Odd)
{
Clover = (inv) ? &CloverTermInvDagOdd : &CloverTermDagOdd;
}
else
{
Clover = (inv) ? &CloverTermInvDagEven : &CloverTermDagEven;
}
out = *Clover * in;
}
else
{
Clover = (inv) ? &CloverTermInv : &CloverTerm;
out = adj(*Clover) * in;
}
}
else
{
if (in.Grid()->_isCheckerBoarded)
{
if (in.Checkerboard() == Odd)
{
// std::cout << "Calling clover term Odd" << std::endl;
Clover = (inv) ? &CloverTermInvOdd : &CloverTermOdd;
}
else
{
// std::cout << "Calling clover term Even" << std::endl;
Clover = (inv) ? &CloverTermInvEven : &CloverTermEven;
}
out = *Clover * in;
// std::cout << GridLogMessage << "*Clover.Checkerboard() " << (*Clover).Checkerboard() << std::endl;
}
else
{
Clover = (inv) ? &CloverTermInv : &CloverTerm;
out = *Clover * in;
}
}
} // MooeeInternal
// Derivative parts
template <class Impl>
void WilsonCloverFermion<Impl>::MooDeriv(GaugeField &mat, const FermionField &X, const FermionField &Y, int dag)
{
assert(0);
}
// Derivative parts
template <class Impl>
void WilsonCloverFermion<Impl>::MeeDeriv(GaugeField &mat, const FermionField &U, const FermionField &V, int dag)
{
assert(0); // not implemented yet
}
FermOpTemplateInstantiate(WilsonCloverFermion);
AdjointFermOpTemplateInstantiate(WilsonCloverFermion);
TwoIndexFermOpTemplateInstantiate(WilsonCloverFermion);
//GparityFermOpTemplateInstantiate(WilsonCloverFermion);
NAMESPACE_END(Grid);

130
Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
View File

@ -386,11 +386,9 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
const FermionField &in, FermionField &out,int dag)
{
DhopTotalTime-=usecond();
#ifdef GRID_OMP
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,in,out,dag);
else
#endif
DhopInternalSerialComms(st,lo,U,in,out,dag);
DhopTotalTime+=usecond();
}
@ -401,111 +399,70 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
#ifdef GRID_OMP
Compressor compressor(dag);
int LLs = in.Grid()->_rdimensions[0];
int len = U.Grid()->oSites();
/////////////////////////////
// Start comms // Gather intranode and extra node differentiated??
/////////////////////////////
DhopFaceTime-=usecond();
st.HaloExchangeOptGather(in,compressor);
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
double ctime=0;
double ptime=0;
DhopCommTime -=usecond();
std::vector<std::vector<CommsRequest_t> > requests;
st.CommunicateBegin(requests);
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
int Opt = WilsonKernelsStatic::Opt;
#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
double start = usecond();
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = U.Grid()->oSites();
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
/////////////////////////////
// Overlap with comms
/////////////////////////////
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
// do the compute
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSiteDag(Opt,st,U_v,st.CommBuf(),sF,sU,LLs,1,in_v,out_v,1,0);
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSite(Opt,st,U_v,st.CommBuf(),sF,sU,LLs,1,in_v,out_v,1,0);
}
}
ptime = usecond() - start;
}
{
double start = usecond();
st.CommunicateThreaded();
ctime = usecond() - start;
}
/////////////////////////////
// do the compute interior
/////////////////////////////
int Opt = WilsonKernelsStatic::Opt; // Why pass this. Kernels should know
DhopComputeTime-=usecond();
if (dag == DaggerYes) {
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
} else {
Kernels::DhopKernel (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,1,0);
}
DhopCommTime += ctime;
DhopComputeTime+=ptime;
DhopComputeTime+=usecond();
// First to enter, last to leave timing
st.CollateThreads();
/////////////////////////////
// Complete comms
/////////////////////////////
st.CommunicateComplete(requests);
DhopCommTime +=usecond();
/////////////////////////////
// do the compute exterior
/////////////////////////////
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
thread_loop( (int ss = 0; ss < sz; ss++) ,{
int sU = st.surface_list[ss];
int sF = LLs * sU;
Kernels::DhopSiteDag(Opt,st,U_v,st.CommBuf(),sF,sU,LLs,1,in_v,out_v,0,1);
});
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
} else {
int sz=st.surface_list.size();
thread_loop( (int ss = 0; ss < sz; ss++) ,{
int sU = st.surface_list[ss];
int sF = LLs * sU;
Kernels::DhopSite(Opt,st,U_v,st.CommBuf(),sF,sU,LLs,1,in_v,out_v,0,1);
});
Kernels::DhopKernel (Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out,0,1);
}
DhopComputeTime2+=usecond();
#else
assert(0);
#endif
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
DoubledGaugeField & U,
const FermionField &in,
FermionField &out,int dag)
{
// assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor(dag);
int LLs = in.Grid()->_rdimensions[0];
@ -515,24 +472,11 @@ void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOr
DhopCommTime+=usecond();
DhopComputeTime-=usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
auto U_v = U.View();
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U_v.size(),in,out);
// parallel_for (int ss = 0; ss < U.Grid()->oSites(); ss++) {
// int sU = ss;
// int sF = LLs * sU;
// Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out);
// }
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
} else {
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U_v.size(),in,out);
// parallel_for (int ss = 0; ss < U.Grid()->oSites(); ss++) {
// int sU = ss;
// int sF = LLs * sU;
// Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out);
// }
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),LLs,U.oSites(),in,out);
}
DhopComputeTime+=usecond();
}

93
Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
View File

@ -375,78 +375,47 @@ void WilsonFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st, LebesgueO
const FermionField &in,
FermionField &out, int dag) {
assert((dag == DaggerNo) || (dag == DaggerYes));
#ifdef GRID_OMP
Compressor compressor(dag);
int len = U.Grid()->oSites();
const int LLs = 1;
/////////////////////////////
// Start comms // Gather intranode and extra node differentiated??
/////////////////////////////
std::vector<std::vector<CommsRequest_t> > requests;
st.Prepare();
st.HaloGather(in,compressor);
st.CommunicateBegin(requests);
/////////////////////////////
// Overlap with comms
/////////////////////////////
st.CommsMergeSHM(compressor);
#pragma omp parallel
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = len;
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
for (int sss = myblock; sss < myblock+myn; ++sss) {
Kernels::DhopSiteDag(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,1,0);
// Kernels::DhopSiteDag(st_v, lo, U_v, st.CommBuf(), sss, sss, 1, 1, in_v, out_v);
}
} else {
for (int sss = myblock; sss < myblock+myn; ++sss) {
Kernels::DhopSite(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,1,0);
// Kernels::DhopSite(st_v, lo, U_v, st.CommBuf(), sss, sss, 1, 1, in_v, out_v);
}
}
/////////////////////////////
// do the compute interior
/////////////////////////////
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
} else {
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,1,0);
}
} else {
st.CommunicateThreaded();
}
} //pragma
/////////////////////////////
// Complete comms
/////////////////////////////
st.CommunicateComplete(requests);
st.CommsMerge(compressor);
{
auto U_v = U.View();
auto in_v = in.View();
auto out_v = out.View();
auto st_v = st.View();
int Opt = WilsonKernelsStatic::Opt;
if (dag == DaggerYes) {
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++) ,{
Kernels::DhopSiteDag(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,0,1);
});
} else {
thread_loop( (int sss = 0; sss < in.Grid()->oSites(); sss++) ,{
Kernels::DhopSite(Opt,st_v,U_v,st.CommBuf(),sss,sss,1,1,in_v,out_v,0,1);
});
}
/////////////////////////////
// do the compute exterior
/////////////////////////////
if (dag == DaggerYes) {
Kernels::DhopDagKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
} else {
Kernels::DhopKernel(Opt,st,U,st.CommBuf(),1,U.oSites(),in,out,0,1);
}
#else
assert(0);
#endif
};

106
Grid/qcd/action/fermion/implementation/WilsonKernelsGpuImplementation.h
View File

@ -73,7 +73,7 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
return;
}
#ifdef GPU_VEC
#if 1
#define GPU_COALESCED_STENCIL_LEG_PROJ(Dir,spProj) \
if (SE._is_local) { \
int mask = Nsimd >> (ptype + 1); \
@ -96,7 +96,7 @@ accelerator_inline void get_stencil(StencilEntry * mem, StencilEntry &chip)
spProj(chi, in_t); \
} \
} else { \
chi = buf[SE._offset+s]; \
chi = (buf[SE._offset+s]; \
} \
synchronise();
#endif
@ -106,15 +106,9 @@ accelerator_inline void WilsonKernels<Impl>::GpuDhopSiteDag(StencilView &st, Sit
SiteHalfSpinor *buf, int Ls, int s,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
#ifdef __CUDA_ARCH__
typename SiteHalfSpinor::scalar_object chi;
typename SiteHalfSpinor::scalar_object Uchi;
typename SiteSpinor::scalar_object result;
#else
SiteHalfSpinor chi;
SiteHalfSpinor Uchi;
SiteSpinor result;
#endif
typedef typename SiteSpinor::scalar_type scalar_type;
typedef typename SiteSpinor::vector_type vector_type;
@ -173,11 +167,7 @@ accelerator_inline void WilsonKernels<Impl>::GpuDhopSiteDag(StencilView &st, Sit
GPU_COALESCED_STENCIL_LEG_PROJ(Tm,spProjTm);
Impl::multLinkGpu(lane,Uchi,U,chi,Tm);
accumReconTm(result, Uchi);
#ifdef GPU_VEC
insertLane (lane,out[sF],result);
#else
vstream(out[sF], result);
#endif
insertLane (lane,out[sF],result);
}
}
@ -186,15 +176,10 @@ accelerator_inline void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, SiteDo
SiteHalfSpinor *buf, int Ls, int s,
int sU, const FermionFieldView &in, FermionFieldView &out)
{
#ifdef __CUDA_ARCH__
typename SiteHalfSpinor::scalar_object chi;
typename SiteHalfSpinor::scalar_object Uchi;
typename SiteSpinor::scalar_object result;
#else
SiteHalfSpinor chi;
SiteHalfSpinor Uchi;
SiteSpinor result;
#endif
typedef typename SiteSpinor::scalar_type scalar_type;
typedef typename SiteSpinor::vector_type vector_type;
constexpr int Nsimd = sizeof(vector_type)/sizeof(scalar_type);
@ -255,11 +240,7 @@ accelerator_inline void WilsonKernels<Impl>::GpuDhopSite(StencilView &st, SiteDo
Impl::multLinkGpu(lane,Uchi,U,chi,Tm);
accumReconTp(result, Uchi);
#ifdef GPU_VEC
insertLane (lane,out[sF],result);
#else
vstream(out[sF], result);
#endif
}
};
@ -287,6 +268,25 @@ GPU_EMPTY(GparityWilsonImplFH);
GPU_EMPTY(GparityWilsonImplD);
GPU_EMPTY(GparityWilsonImplDF);
#define KERNEL_CALL(A) \
const uint64_t nsimd = Simd::Nsimd(); \
const uint64_t NN = Nsite*Ls*nsimd;\
accelerator_loopN( sss, NN, { \
uint64_t cur = sss; \
cur = cur / nsimd; \
uint64_t s = cur%Ls; \
cur = cur / Ls; \
uint64_t sU = cur; \
WilsonKernels<Impl>::A(st_v,U_v[sU],buf,Ls,s,sU,in_v,out_v);\
});
#define HOST_CALL(A) \
accelerator_loopN( ss, Ls*Nsite, { \
int sF = ss; \
int sU = ss/Ls; \
WilsonKernels<Impl>::A(st_v,U_v,st.CommBuf(),sF,sU,in_v,out_v); \
});
template <class Impl>
void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
int Ls, int Nsite, const FermionField &in, FermionField &out,
@ -297,25 +297,18 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
auto out_v = out.View();
auto st_v = st.View();
if ( (Opt == WilsonKernelsStatic::OptGpu) && interior && exterior ) {
#define KERNEL_CALL(A) \
const uint64_t nsimd = Simd::Nsimd(); \
const uint64_t NN = Nsite*Ls*nsimd;\
accelerator_loopN( sss, NN, { \
uint64_t cur = sss; \
cur = cur / nsimd; \
uint64_t s = cur%Ls; \
cur = cur / Ls; \
uint64_t sU = cur;
WilsonKernels<Impl>::GpuDhopSite(st_v,U_v[sU],buf,Ls,s,sU,in_v,out_v);
});
} else {
accelerator_loop( ss, U_v, {
int sU = ss;
int sF = Ls * sU;
WilsonKernels<Impl>::GenericDhopSite(Opt,st_v,U_v,st.CommBuf(),sF,sU,Ls,1,in_v,out_v);
});
}
if( interior && exterior ) {
if (Opt == WilsonKernelsStatic::OptGpu) {
KERNEL_CALL(GpuDhopSite);
} else {
HOST_CALL(GenericDhopSite);
}
} else if( interior ) {
HOST_CALL(GenericDhopSiteInt);
} else if( exterior ) {
HOST_CALL(GenericDhopSiteExt);
}
}
template <class Impl>
void WilsonKernels<Impl>::DhopDagKernel(int Opt,StencilImpl &st, DoubledGaugeField &U, SiteHalfSpinor * buf,
@ -327,25 +320,16 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
auto out_v = out.View();
auto st_v = st.View();
if ( (Opt == WilsonKernelsStatic::OptGpu) && interior && exterior ) {
const uint64_t nsimd = Simd::Nsimd();
const uint64_t NN = Nsite*Ls*nsimd;
accelerator_loopN( sss, NN, {
uint64_t cur = sss;
// uint64_t lane = cur % nsimd;
cur = cur / nsimd;
uint64_t s = cur%Ls;
// uint64_t sF = cur;
cur = cur / Ls;
uint64_t sU = cur;
WilsonKernels<Impl>::GpuDhopSiteDag(st_v,U_v[sU],buf,Ls,s,sU,in_v,out_v);
});
} else {
accelerator_loop( ss, U_v, {
int sU = ss;
int sF = Ls * sU;
WilsonKernels<Impl>::GenericDhopSiteDag(Opt,st,U_v,st.CommBuf(),sF,sU,Ls,1,in_v,out_v);
});
if( interior && exterior ) {
if (Opt == WilsonKernelsStatic::OptGpu) {
KERNEL_CALL(GpuDhopSiteDag);
} else {
HOST_CALL(GenericDhopSiteDag);
}
} else if( interior ) {
HOST_CALL(GenericDhopSiteDagInt);
} else if( exterior ) {
HOST_CALL(GenericDhopSiteDagExt);
}
}

1
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@ -267,7 +267,6 @@ void WilsonKernels<Impl>::DhopDirK( StencilView &st, DoubledGaugeFieldView &U,Si
int ptype;
SE = st.GetEntry(ptype, dir, sF);
// GENERIC_DHOPDIR_LEG(Xp,spProjXp,spReconXp);
if (gamma == Xp) {
if (SE->_is_local ) {
int perm= SE->_permute;

97
Grid/qcd/action/fermion/implementation/WilsonTMFermion.h
View File

@ -1,97 +0,0 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonTMFermion.cc
Copyright (C) 2015
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/WilsonTMFermion.h>
NAMESPACE_BEGIN(Grid);
/*
* BF sequence
*
void bfmbase<Float>::MooeeInv(Fermion_t psi,
Fermion_t chi,
int dag, int cb)
double m = this->mass;
double tm = this->twistedmass;
double mtil = 4.0+this->mass;
double sq = mtil*mtil + tm*tm;
double a = mtil/sq;
double b = -tm /sq;
if(dag) b=-b;
axpibg5x(chi,psi,a,b);
void bfmbase<Float>::Mooee(Fermion_t psi,
Fermion_t chi,
int dag,int cb)
double a = 4.0+this->mass;
double b = this->twistedmass;
if(dag) b=-b;
axpibg5x(chi,psi,a,b);
*/
template<class Impl>
void WilsonTMFermion<Impl>::Mooee(const FermionField &in, FermionField &out) {
RealD a = 4.0+this->mass;
RealD b = this->mu;
out.Checkerboard() = in.Checkerboard();
axpibg5x(out,in,a,b);
}
template<class Impl>
void WilsonTMFermion<Impl>::MooeeDag(const FermionField &in, FermionField &out) {
RealD a = 4.0+this->mass;
RealD b = -this->mu;
out.Checkerboard() = in.Checkerboard();
axpibg5x(out,in,a,b);
}
template<class Impl>
void WilsonTMFermion<Impl>::MooeeInv(const FermionField &in, FermionField &out) {
RealD m = this->mass;
RealD tm = this->mu;
RealD mtil = 4.0+m;
RealD sq = mtil*mtil+tm*tm;
RealD a = mtil/sq;
RealD b = -tm /sq;
axpibg5x(out,in,a,b);
}
template<class Impl>
void WilsonTMFermion<Impl>::MooeeInvDag(const FermionField &in, FermionField &out) {
RealD m = this->mass;
RealD tm = this->mu;
RealD mtil = 4.0+m;
RealD sq = mtil*mtil+tm*tm;
RealD a = mtil/sq;
RealD b = tm /sq;
axpibg5x(out,in,a,b);
}
FermOpTemplateInstantiate(WilsonTMFermion);
NAMESPACE_END(Grid);

45
Grid/qcd/action/fermion/instantiation/CayleyFermion5DInstantiation.cc Normal file
View File

@ -0,0 +1,45 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/CayleyFermion5DImplementation.h>
#include <Grid/qcd/action/fermion/implementation/CayleyFermion5Dcache.h>
//#include <Grid/qcd/action/fermion/implementation/CayleyFermion5Dvec.h>
//#include <Grid/qcd/action/fermion/implementation/CayleyFermion5Dgpu.h>
NAMESPACE_BEGIN(Grid);
// FIXME: Break these out to parallel make accelerate
FermOpTemplateInstantiate(CayleyFermion5D);
GparityFermOpTemplateInstantiate(CayleyFermion5D);
NAMESPACE_END(Grid);

37
Grid/qcd/action/fermion/instantiation/ContinuedFractionFermion5DInstantiation.cc Normal file
View File

@ -0,0 +1,37 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ContinuedFractionFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/ContinuedFractionFermion5D.h>
#include <Grid/qcd/action/fermion/implementation/ContinuedFractionFermion5DImplementation.h>
NAMESPACE_BEGIN(Grid);
FermOpTemplateInstantiate(ContinuedFractionFermion5D);
NAMESPACE_END(Grid);

45
Grid/qcd/action/fermion/instantiation/DomainWallEOFAFermionInstantiation.cc Normal file
View File

@ -0,0 +1,45 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/DomainWallEOFAFermion.cc
Copyright (C) 2017
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: David Murphy <dmurphy@phys.columbia.edu>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid_Eigen_Dense.h>
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/DomainWallEOFAFermion.h>
#include <Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionImplementation.h>
#include <Grid/qcd/action/fermion/implementation/DomainWallEOFAFermionCache.h>
NAMESPACE_BEGIN(Grid);
FermOpTemplateInstantiate(DomainWallEOFAFermion);
GparityFermOpTemplateInstantiate(DomainWallEOFAFermion);
NAMESPACE_END(Grid);

46
Grid/qcd/action/fermion/instantiation/ImprovedStaggeredFermion5DInstantiation.cc Normal file
View File

@ -0,0 +1,46 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
Copyright (C) 2015
Author: Azusa Yamaguchi <ayamaguc@staffmail.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/ImprovedStaggeredFermion5D.h>
#include <Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermion5DImplementation.h>
#include <Grid/perfmon/PerfCount.h>
NAMESPACE_BEGIN(Grid);
// S-direction is INNERMOST and takes no part in the parity.
const std::vector<int> ImprovedStaggeredFermion5DStatic::directions({1,2,3,4,1,2,3,4,1,2,3,4,1,2,3,4});
const std::vector<int> ImprovedStaggeredFermion5DStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1, 3, 3, 3, 3, -3, -3, -3, -3});
FermOpStaggeredTemplateInstantiate(ImprovedStaggeredFermion5D);
FermOpStaggeredVec5dTemplateInstantiate(ImprovedStaggeredFermion5D);
NAMESPACE_END(Grid);

39
Grid/qcd/action/fermion/instantiation/ImprovedStaggeredFermionInstantiation.cc Normal file
View File

@ -0,0 +1,39 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
Copyright (C) 2015
Author: Azusa Yamaguchi, Peter Boyle
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid.h>
#include <Grid/qcd/action/fermion/implementation/ImprovedStaggeredFermionImplementation.h>
NAMESPACE_BEGIN(Grid);
const std::vector<int> ImprovedStaggeredFermionStatic::directions({0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3});
const std::vector<int> ImprovedStaggeredFermionStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1, 3, 3, 3, 3, -3, -3, -3, -3});
FermOpStaggeredTemplateInstantiate(ImprovedStaggeredFermion);
NAMESPACE_END(Grid);

38
Grid/qcd/action/fermion/instantiation/PartialFractionFermion5DInstantiation.cc Normal file
View File

@ -0,0 +1,38 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/PartialFractionFermion5D.cc
Copyright (C) 2015
Author: Peter Boyle <pabobyle@ph.ed.ac.uk>
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/PartialFractionFermion5D.h>
#include <Grid/qcd/action/fermion/implementation/PartialFractionFermion5DImplementation.h>
NAMESPACE_BEGIN(Grid);
FermOpTemplateInstantiate(PartialFractionFermion5D);
NAMESPACE_END(Grid);

43
Grid/qcd/action/fermion/instantiation/StaggeredKernelsInstantiation.cc Normal file
View File

@ -0,0 +1,43 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015
Author: Azusa Yamaguchi, Peter Boyle
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/StaggeredKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/StaggeredKernelsHand.h>
NAMESPACE_BEGIN(Grid);
int StaggeredKernelsStatic::Opt= StaggeredKernelsStatic::OptGeneric;
int StaggeredKernelsStatic::Comms = StaggeredKernelsStatic::CommsAndCompute;
FermOpStaggeredTemplateInstantiate(StaggeredKernels);
FermOpStaggeredVec5dTemplateInstantiate(StaggeredKernels);
NAMESPACE_END(Grid);

42
Grid/qcd/action/fermion/instantiation/WilsonCloverFermionInstantiation.cc Normal file
View File

@ -0,0 +1,42 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonCloverFermion.cc
Copyright (C) 2017
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Guido Cossu <guido.cossu@ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/Grid.h>
#include <Grid/qcd/spin/Dirac.h>
#include <Grid/qcd/action/fermion/WilsonCloverFermion.h>
#include <Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h>
NAMESPACE_BEGIN(Grid);
FermOpTemplateInstantiate(WilsonCloverFermion);
AdjointFermOpTemplateInstantiate(WilsonCloverFermion);
TwoIndexFermOpTemplateInstantiate(WilsonCloverFermion);
//GparityFermOpTemplateInstantiate(WilsonCloverFermion);
NAMESPACE_END(Grid);

40
Grid/qcd/action/fermion/instantiation/WilsonFermion5DInstantiation.cc Normal file
View File

@ -0,0 +1,40 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h>
NAMESPACE_BEGIN(Grid);
FermOpTemplateInstantiate(WilsonFermion5D);
GparityFermOpTemplateInstantiate(WilsonFermion5D);
NAMESPACE_END(Grid);

46
Grid/qcd/action/fermion/instantiation/WilsonFermionInstantiation.cc Normal file
View File

@ -0,0 +1,46 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h>
NAMESPACE_BEGIN(Grid);
const std::vector<int> WilsonFermionStatic::directions({0, 1, 2, 3, 0, 1, 2, 3});
const std::vector<int> WilsonFermionStatic::displacements({1, 1, 1, 1, -1, -1, -1, -1});
int WilsonFermionStatic::HandOptDslash;
FermOpTemplateInstantiate(WilsonFermion);
AdjointFermOpTemplateInstantiate(WilsonFermion);
TwoIndexFermOpTemplateInstantiate(WilsonFermion);
GparityFermOpTemplateInstantiate(WilsonFermion);
NAMESPACE_END(Grid);

51
Grid/qcd/action/fermion/instantiation/WilsonKernelsInstantiation.cc Normal file
View File

@ -0,0 +1,51 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonKernels.cc
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Peter Boyle <peterboyle@Peters-MacBook-Pro-2.local>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution
directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsAsmImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsGpuImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandImplementation.h>
#include <Grid/qcd/action/fermion/implementation/WilsonKernelsHandGparityImplementation.h>
NAMESPACE_BEGIN(Grid);
// Move these
int WilsonKernelsStatic::Opt = WilsonKernelsStatic::OptGeneric;
int WilsonKernelsStatic::Comms = WilsonKernelsStatic::CommsAndCompute;
// FIXME: Break these out to parallel make
FermOpTemplateInstantiate(WilsonKernels);
GparityFermOpTemplateInstantiate(WilsonKernels); // Specialisation in Gparity forces instantiation
AdjointFermOpTemplateInstantiate(WilsonKernels);
TwoIndexFermOpTemplateInstantiate(WilsonKernels);
NAMESPACE_END(Grid);

36
Grid/qcd/action/fermion/instantiation/WilsonTMFermionInstantiation.cc Normal file
View File

@ -0,0 +1,36 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/qcd/action/fermion/WilsonTMFermion.cc
Copyright (C) 2015
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/WilsonTMFermion.h>
#include <Grid/qcd/action/fermion/implementation/WilsonTMFermionImplementation.h>
NAMESPACE_BEGIN(Grid);
FermOpTemplateInstantiate(WilsonTMFermion);
NAMESPACE_END(Grid);

9
Grid/tensors/Tensor_traits.h
View File

@ -267,11 +267,12 @@ struct getVectorType{
template<typename T>
class isSIMDvectorized{
template<typename U>
static typename std::enable_if< !std::is_same< typename GridTypeMapper<typename getVectorType<U>::type>::scalar_type,
typename GridTypeMapper<typename getVectorType<U>::type>::vector_type>::value, char>::type test(void *);
static typename std::enable_if<
!std::is_same< typename GridTypeMapper<typename getVectorType<U>::type>::scalar_type,
typename GridTypeMapper<typename getVectorType<U>::type>::vector_type>::value,
char>::type test(void *);
template<typename U>
static double test(...);
template<typename U> static double test(...);
public:
enum {value = sizeof(test<T>(0)) == sizeof(char) };

1
README
View File

@ -1 +0,0 @@
README.md

407
README Normal file
View File

@ -0,0 +1,407 @@
# Grid [![Teamcity status](http://ci.cliath.ph.ed.ac.uk/app/rest/builds/aggregated/strob:(buildType:(affectedProject(id:Grid)),branch:name:develop)/statusIcon.svg)](http://ci.cliath.ph.ed.ac.uk/project.html?projectId=Grid&tab=projectOverview) [![Travis status](https://travis-ci.org/paboyle/Grid.svg?branch=develop)](https://travis-ci.org/paboyle/Grid)
**Data parallel C++ mathematical object library.**
License: GPL v2.
Last update June 2017.
_Please do not send pull requests to the `master` branch which is reserved for releases._
### Description
This library provides data parallel C++ container classes with internal memory layout
that is transformed to map efficiently to SIMD architectures. CSHIFT facilities
are provided, similar to HPF and cmfortran, and user control is given over the mapping of
array indices to both MPI tasks and SIMD processing elements.
* Identically shaped arrays then be processed with perfect data parallelisation.
* Such identically shaped arrays are called conformable arrays.
The transformation is based on the observation that Cartesian array processing involves
identical processing to be performed on different regions of the Cartesian array.
The library will both geometrically decompose into MPI tasks and across SIMD lanes.
Local vector loops are parallelised with OpenMP pragmas.
Data parallel array operations can then be specified with a SINGLE data parallel paradigm, but
optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a significant simplification
for most programmers.
The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture.
Presently SSE4, ARM NEON (128 bits) AVX, AVX2, QPX (256 bits), IMCI and AVX512 (512 bits) targets are supported.
These are presented as `vRealF`, `vRealD`, `vComplexF`, and `vComplexD` internal vector data types.
The corresponding scalar types are named `RealF`, `RealD`, `ComplexF` and `ComplexD`.
MPI, OpenMP, and SIMD parallelism are present in the library.
Please see [this paper](https://arxiv.org/abs/1512.03487) for more detail.
### Compilers
Intel ICPC v16.0.3 and later
Clang v3.5 and later (need 3.8 and later for OpenMP)
GCC v4.9.x (recommended)
GCC v6.3 and later
### Important:
Some versions of GCC appear to have a bug under high optimisation (-O2, -O3).
The safety of these compiler versions cannot be guaranteed at this time. Follow Issue 100 for details and updates.
GCC v5.x
GCC v6.1, v6.2
### Bug report
_To help us tracking and solving more efficiently issues with Grid, please report problems using the issue system of GitHub rather than sending emails to Grid developers._
When you file an issue, please go though the following checklist:
1. Check that the code is pointing to the `HEAD` of `develop` or any commit in `master` which is tagged with a version number.
2. Give a description of the target platform (CPU, network, compiler). Please give the full CPU part description, using for example `cat /proc/cpuinfo | grep 'model name' | uniq` (Linux) or `sysctl machdep.cpu.brand_string` (macOS) and the full output the `--version` option of your compiler.
3. Give the exact `configure` command used.
4. Attach `config.log`.
5. Attach `grid.config.summary`.
6. Attach the output of `make V=1`.
7. Describe the issue and any previous attempt to solve it. If relevant, show how to reproduce the issue using a minimal working example.
### Required libraries
Grid requires:
[GMP](https://gmplib.org/),
[MPFR](http://www.mpfr.org/)
Bootstrapping grid downloads and uses for internal dense matrix (non-QCD operations) the Eigen library.
Grid optionally uses:
[HDF5](https://support.hdfgroup.org/HDF5/)
[LIME](http://usqcd-software.github.io/c-lime/) for ILDG and SciDAC file format support.
[FFTW](http://www.fftw.org) either generic version or via the Intel MKL library.
LAPACK either generic version or Intel MKL library.
### Quick start
First, start by cloning the repository:
``` bash
git clone https://github.com/paboyle/Grid.git
```
Then enter the cloned directory and set up the build system:
``` bash
cd Grid
./bootstrap.sh
```
Now you can execute the `configure` script to generate makefiles (here from a build directory):
``` bash
mkdir build; cd build
../configure --enable-precision=double --enable-simd=AVX --enable-comms=mpi-auto --prefix=<path>
```
where `--enable-precision=` set the default precision,
`--enable-simd=` set the SIMD type, `--enable-
comms=`, and `<path>` should be replaced by the prefix path where you want to
install Grid. Other options are detailed in the next section, you can also use `configure
--help` to display them. Like with any other program using GNU autotool, the
`CXX`, `CXXFLAGS`, `LDFLAGS`, ... environment variables can be modified to
customise the build.
Finally, you can build, check, and install Grid:
``` bash
make; make check; make install
```
To minimise the build time, only the tests at the root of the `tests` directory are built by default. If you want to build tests in the sub-directory `<subdir>` you can execute:
``` bash
make -C tests/<subdir> tests
```
If you want to build all the tests at once just use `make tests`.
### Build configuration options
- `--prefix=<path>`: installation prefix for Grid.
- `--with-gmp=<path>`: look for GMP in the UNIX prefix `<path>`
- `--with-mpfr=<path>`: look for MPFR in the UNIX prefix `<path>`
- `--with-fftw=<path>`: look for FFTW in the UNIX prefix `<path>`
- `--enable-lapack[=<path>]`: enable LAPACK support in Lanczos eigensolver. A UNIX prefix containing the library can be specified (optional).
- `--enable-mkl[=<path>]`: use Intel MKL for FFT (and LAPACK if enabled) routines. A UNIX prefix containing the library can be specified (optional).
- `--enable-numa`: enable NUMA first touch optimisation
- `--enable-simd=<code>`: setup Grid for the SIMD target `<code>` (default: `GEN`). A list of possible SIMD targets is detailed in a section below.
- `--enable-gen-simd-width=<size>`: select the size (in bytes) of the generic SIMD vector type (default: 32 bytes).
- `--enable-precision={single|double}`: set the default precision (default: `double`).
- `--enable-precision=<comm>`: Use `<comm>` for message passing (default: `none`). A list of possible SIMD targets is detailed in a section below.
- `--enable-rng={sitmo|ranlux48|mt19937}`: choose the RNG (default: `sitmo `).
- `--disable-timers`: disable system dependent high-resolution timers.
- `--enable-chroma`: enable Chroma regression tests.
- `--enable-doxygen-doc`: enable the Doxygen documentation generation (build with `make doxygen-doc`)
### Possible communication interfaces
The following options can be use with the `--enable-comms=` option to target different communication interfaces:
| `<comm>` | Description |
| -------------- | ------------------------------------------------------------- |
| `none` | no communications |
| `mpi[-auto]` | MPI communications |
| `mpi3[-auto]` | MPI communications using MPI 3 shared memory |
| `shmem ` | Cray SHMEM communications |
For the MPI interfaces the optional `-auto` suffix instructs the `configure` scripts to determine all the necessary compilation and linking flags. This is done by extracting the informations from the MPI wrapper specified in the environment variable `MPICXX` (if not specified `configure` will scan though a list of default names). The `-auto` suffix is not supported by the Cray environment wrapper scripts. Use the standard versions instead.
### Possible SIMD types
The following options can be use with the `--enable-simd=` option to target different SIMD instruction sets:
| `<code>` | Description |
| ----------- | -------------------------------------- |
| `GEN` | generic portable vector code |
| `SSE4` | SSE 4.2 (128 bit) |
| `AVX` | AVX (256 bit) |
| `AVXFMA` | AVX (256 bit) + FMA |
| `AVXFMA4` | AVX (256 bit) + FMA4 |
| `AVX2` | AVX 2 (256 bit) |
| `AVX512` | AVX 512 bit |
| `NEONv8` | [ARM NEON](http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.den0024a/ch07s03.html) (128 bit) |
| `QPX` | IBM QPX (256 bit) |
Alternatively, some CPU codenames can be directly used:
| `<code>` | Description |
| ----------- | -------------------------------------- |
| `KNL` | [Intel Xeon Phi codename Knights Landing](http://ark.intel.com/products/codename/48999/Knights-Landing) |
| `SKL` | [Intel Skylake with AVX512 extensions](https://ark.intel.com/products/codename/37572/Skylake#@server) |
| `BGQ` | Blue Gene/Q |
#### Notes:
- We currently support AVX512 for the Intel compiler and GCC (KNL and SKL target). Support for clang will appear in future versions of Grid when the AVX512 support in the compiler will be more advanced.
- For BG/Q only [bgclang](http://trac.alcf.anl.gov/projects/llvm-bgq) is supported. We do not presently plan to support more compilers for this platform.
- BG/Q performances are currently rather poor. This is being investigated for future versions.
- The vector size for the `GEN` target can be specified with the `configure` script option `--enable-gen-simd-width`.
### Build setup for Intel Knights Landing platform
The following configuration is recommended for the Intel Knights Landing platform:
``` bash
../configure --enable-precision=double\
--enable-simd=KNL \
--enable-comms=mpi-auto \
--enable-mkl \
CXX=icpc MPICXX=mpiicpc
```
The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
``` bash
../configure --enable-precision=double\
--enable-simd=KNL \
--enable-comms=mpi \
--enable-mkl \
CXX=CC CC=cc
```
If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
``` bash
--with-gmp=<path> \
--with-mpfr=<path> \
```
where `<path>` is the UNIX prefix where GMP and MPFR are installed.
Knight's Landing with Intel Omnipath adapters with two adapters per node
presently performs better with use of more than one rank per node, using shared memory
for interior communication. This is the mpi3 communications implementation.
We recommend four ranks per node for best performance, but optimum is local volume dependent.
``` bash
../configure --enable-precision=double\
--enable-simd=KNL \
--enable-comms=mpi3-auto \
--enable-mkl \
CC=icpc MPICXX=mpiicpc
```
### Build setup for Intel Haswell Xeon platform
The following configuration is recommended for the Intel Haswell platform:
``` bash
../configure --enable-precision=double\
--enable-simd=AVX2 \
--enable-comms=mpi3-auto \
--enable-mkl \
CXX=icpc MPICXX=mpiicpc
```
The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
``` bash
--with-gmp=<path> \
--with-mpfr=<path> \
```
where `<path>` is the UNIX prefix where GMP and MPFR are installed.
If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
``` bash
../configure --enable-precision=double\
--enable-simd=AVX2 \
--enable-comms=mpi3 \
--enable-mkl \
CXX=CC CC=cc
```
Since Dual socket nodes are commonplace, we recommend MPI-3 as the default with the use of
one rank per socket. If using the Intel MPI library, threads should be pinned to NUMA domains using
```
export I_MPI_PIN=1
```
This is the default.
### Build setup for Intel Skylake Xeon platform
The following configuration is recommended for the Intel Skylake platform:
``` bash
../configure --enable-precision=double\
--enable-simd=AVX512 \
--enable-comms=mpi3 \
--enable-mkl \
CXX=mpiicpc
```
The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.
If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
``` bash
--with-gmp=<path> \
--with-mpfr=<path> \
```
where `<path>` is the UNIX prefix where GMP and MPFR are installed.
If you are working on a Cray machine that does not use the `mpiicpc` wrapper, please use:
``` bash
../configure --enable-precision=double\
--enable-simd=AVX512 \
--enable-comms=mpi3 \
--enable-mkl \
CXX=CC CC=cc
```
Since Dual socket nodes are commonplace, we recommend MPI-3 as the default with the use of
one rank per socket. If using the Intel MPI library, threads should be pinned to NUMA domains using
```
export I_MPI_PIN=1
```
This is the default.
#### Expected Skylake Gold 6148 dual socket (single prec, single node 20+20 cores) performance using NUMA MPI mapping):
mpirun -n 2 benchmarks/Benchmark_dwf --grid 16.16.16.16 --mpi 2.1.1.1 --cacheblocking 2.2.2.2 --dslash-asm --shm 1024 --threads 18
TBA
### Build setup for AMD EPYC / RYZEN
The AMD EPYC is a multichip module comprising 32 cores spread over four distinct chips each with 8 cores.
So, even with a single socket node there is a quad-chip module. Dual socket nodes with 64 cores total
are common. Each chip within the module exposes a separate NUMA domain.
There are four NUMA domains per socket and we recommend one MPI rank per NUMA domain.
MPI-3 is recommended with the use of four ranks per socket,
and 8 threads per rank.
The following configuration is recommended for the AMD EPYC platform.
``` bash
../configure --enable-precision=double\
--enable-simd=AVX2 \
--enable-comms=mpi3 \
CXX=mpicxx
```
If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:
``` bash
--with-gmp=<path> \
--with-mpfr=<path> \
```
where `<path>` is the UNIX prefix where GMP and MPFR are installed.
Using MPICH and g++ v4.9.2, best performance can be obtained using explicit GOMP_CPU_AFFINITY flags for each MPI rank.
This can be done by invoking MPI on a wrapper script omp_bind.sh to handle this.
It is recommended to run 8 MPI ranks on a single dual socket AMD EPYC, with 8 threads per rank using MPI3 and
shared memory to communicate within this node:
mpirun -np 8 ./omp_bind.sh ./Benchmark_dwf --mpi 2.2.2.1 --dslash-unroll --threads 8 --grid 16.16.16.16 --cacheblocking 4.4.4.4
Where omp_bind.sh does the following:
```
#!/bin/bash
numanode=` expr $PMI_RANK % 8 `
basecore=`expr $numanode \* 16`
core0=`expr $basecore + 0 `
core1=`expr $basecore + 2 `
core2=`expr $basecore + 4 `
core3=`expr $basecore + 6 `
core4=`expr $basecore + 8 `
core5=`expr $basecore + 10 `
core6=`expr $basecore + 12 `
core7=`expr $basecore + 14 `
export GOMP_CPU_AFFINITY="$core0 $core1 $core2 $core3 $core4 $core5 $core6 $core7"
echo GOMP_CUP_AFFINITY $GOMP_CPU_AFFINITY
$@
```
Performance:
#### Expected AMD EPYC 7601 dual socket (single prec, single node 32+32 cores) performance using NUMA MPI mapping):
mpirun -np 8 ./omp_bind.sh ./Benchmark_dwf --threads 8 --mpi 2.2.2.1 --dslash-unroll --grid 16.16.16.16 --cacheblocking 4.4.4.4
TBA
### Build setup for BlueGene/Q
To be written...
### Build setup for ARM Neon
To be written...
### Build setup for laptops, other compilers, non-cluster builds
Many versions of g++ and clang++ work with Grid, and involve merely replacing CXX (and MPICXX),
and omit the enable-mkl flag.
Single node builds are enabled with
```
--enable-comms=none
```
FFTW support that is not in the default search path may then enabled with
```
--with-fftw=<installpath>
```
BLAS will not be compiled in by default, and Lanczos will default to Eigen diagonalisation.

7
TODO
View File

@ -3,19 +3,23 @@
GPU branch code item work list
-----------------------------
1) Common source GPU and CPU generic kernels???
- coalescedRead, coalescedWrite in expressions.
- Uniform coding between GPU kernels and CPU kernels attempt
- Clean up PRAGMAS
-- Figure what to do about "multLinkGpu" etc.. in FermionOperatorImpl.
-- Gparity is the awkward one
-- Solve non-Gparity first.
-- Simplify the operator IMPL support
2) - SIMD dirs in stencil
3) Merge develop and test HMC
4) GPU accelerate EOFA
4) GPU accelerate EOFA
5) Accelerate the cshift
@ -43,7 +47,6 @@ Single GPU simd target (VGPU)
15) Staggered kernels inline for GPU
-----
Gianluca's changes
- Performance impact of construct in aligned allocator???

1
include/Grid
View File

@ -1 +0,0 @@
../lib

15432
scripts/loop.log
View File

File diff suppressed because it is too large Load Diff