portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-14 01:35:36 +00:00
Code Releases Activity
e9ef7e3852
Grid/lib/qcd/action/fermion/CayleyFermion5Dssp.cc

162 lines
5.1 KiB
C++
Raw Normal View History

Namespace, indent
2018-01-14 23:42:47 +00:00
/*************************************************************************************
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
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
Namespace, indent
2018-01-14 23:42:47 +00:00
*************************************************************************************/
/* END LEGAL */
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
Refactoring header layout
2017-02-22 18:09:33 +00:00
#include <Grid/qcd/action/fermion/FermionCore.h>
#include <Grid/qcd/action/fermion/CayleyFermion5D.h>
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
Namespace, indent
2018-01-14 23:42:47 +00:00
NAMESPACE_BEGIN(Grid);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
Namespace, indent
2018-01-14 23:42:47 +00:00
// Pminus fowards
// Pplus backwards
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
template<class Impl>
void CayleyFermion5D<Impl>::M5D(const FermionField &psi,
const FermionField &phi,
FermionField &chi,
Feature for z-Mobius prep
2016-08-15 22:31:29 +01:00
std::vector<Coeff_t> &lower,
std::vector<Coeff_t> &diag,
std::vector<Coeff_t> &upper)
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
{
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
Coeff_t one(1.0);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
int Ls=this->Ls;
for(int s=0;s<Ls;s++){
if ( s==0 ) {
axpby_ssp_pminus(chi,diag[s],phi,upper[s],psi,s,s+1);
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pplus (chi,one,chi,lower[s],psi,s,Ls-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
} else if ( s==(Ls-1)) {
axpby_ssp_pminus(chi,diag[s],phi,upper[s],psi,s,0);
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pplus (chi,one,chi,lower[s],psi,s,s-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
} else {
axpby_ssp_pminus(chi,diag[s],phi,upper[s],psi,s,s+1);
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pplus(chi,one,chi,lower[s],psi,s,s-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
}
}
template<class Impl>
void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi,
const FermionField &phi,
FermionField &chi,
Feature for z-Mobius prep
2016-08-15 22:31:29 +01:00
std::vector<Coeff_t> &lower,
std::vector<Coeff_t> &diag,
std::vector<Coeff_t> &upper)
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
{
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
Coeff_t one(1.0);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
int Ls=this->Ls;
for(int s=0;s<Ls;s++){
if ( s==0 ) {
axpby_ssp_pplus (chi,diag[s],phi,upper[s],psi,s,s+1);
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pminus(chi,one,chi,lower[s],psi,s,Ls-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
} else if ( s==(Ls-1)) {
axpby_ssp_pplus (chi,diag[s],phi,upper[s],psi,s,0);
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pminus(chi,one,chi,lower[s],psi,s,s-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
} else {
axpby_ssp_pplus (chi,diag[s],phi,upper[s],psi,s,s+1);
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pminus(chi,one,chi,lower[s],psi,s,s-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
}
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi, FermionField &chi)
{
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
Coeff_t one(1.0);
Coeff_t czero(0.0);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
chi.checkerboard=psi.checkerboard;
int Ls=this->Ls;
// Apply (L^{\prime})^{-1}
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp (chi,one,psi, czero,psi,0,0); // chi[0]=psi[0]
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
for (int s=1;s<Ls;s++){
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pplus(chi,one,psi,-lee[s-1],chi,s,s-1);// recursion Psi[s] -lee P_+ chi[s-1]
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
// L_m^{-1}
for (int s=0;s<Ls-1;s++){ // Chi[ee] = 1 - sum[s<Ls-1] -leem[s]P_- chi
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pminus(chi,one,chi,-leem[s],chi,Ls-1,s);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
// U_m^{-1} D^{-1}
for (int s=0;s<Ls-1;s++){
// Chi[s] + 1/d chi[s]
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pplus(chi,one/dee[s],chi,-ueem[s]/dee[Ls-1],chi,s,Ls-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp(chi,one/dee[Ls-1],chi,czero,chi,Ls-1,Ls-1); // Modest avoidable
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
// Apply U^{-1}
for (int s=Ls-2;s>=0;s--){
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pminus (chi,one,chi,-uee[s],chi,s,s+1); // chi[Ls]
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
}
template<class Impl>
void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi, FermionField &chi)
{
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
Coeff_t one(1.0);
Coeff_t czero(0.0);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
chi.checkerboard=psi.checkerboard;
int Ls=this->Ls;
// Apply (U^{\prime})^{-dagger}
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp (chi,one,psi, czero,psi,0,0); // chi[0]=psi[0]
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
for (int s=1;s<Ls;s++){
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pminus(chi,one,psi,-conjugate(uee[s-1]),chi,s,s-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
// U_m^{-\dagger}
for (int s=0;s<Ls-1;s++){
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pplus(chi,one,chi,-conjugate(ueem[s]),chi,Ls-1,s);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
// L_m^{-\dagger} D^{-dagger}
for (int s=0;s<Ls-1;s++){
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pminus(chi,one/conjugate(dee[s]),chi,-conjugate(leem[s]/dee[Ls-1]),chi,s,Ls-1);
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp(chi,one/conjugate(dee[Ls-1]),chi,czero,chi,Ls-1,Ls-1); // Modest avoidable
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
// Apply L^{-dagger}
for (int s=Ls-2;s>=0;s--){
Conjugate coefficients on adjoint
2017-03-30 05:44:05 +01:00
axpby_ssp_pplus (chi,one,chi,-conjugate(lee[s]),chi,s,s+1); // chi[Ls]
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
}
}
Solving the instantiation no compile issue
2016-07-15 17:19:44 +01:00
#ifdef CAYLEY_DPERP_LINALG
Namespace, indent
2018-01-14 23:42:47 +00:00
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);
Solving the instantiation no compile issue
2016-07-15 17:19:44 +01:00
#endif
Adding files for multiple implementations (cache opt) and Ls vectorisation of the 5D cayley form chiral fermions for the 5d matrix. With Ls entirely in the vector direction, s-hopping terms involve rotations. The serial dependence of the LDU inversion for Mobius and 4d even odd checkerboarding is removed by simply applying Ls^2 operations (vectorised many ways) as a dense matrix operation. This should give similar throughput but high flops (non-compulsory flops) but enable use of the KNL cache friendly kernels throughout the code. Ls is still constrained to be a multiple of Nsimd, which is as much as 8 for AVX512 with single precision.
2016-07-14 22:59:21 +01:00
Namespace, indent
2018-01-14 23:42:47 +00:00
NAMESPACE_END(Grid);
Copy Permalink