portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-15 10:15:36 +00:00
Code Releases Activity
2018077770
Grid/lib/qcd/action/fermion/CayleyFermion5Dcache.cc

257 lines
7.0 KiB
C++
Raw Normal View History

Namespace, format
2018-01-14 23:44:16 +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, format
2018-01-14 23:44:16 +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, format
2018-01-14 23:44:16 +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, format
2018-01-14 23:44:16 +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>
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
void CayleyFermion5D<Impl>::M5D(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
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
{
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
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();
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;
Hide internal data
2018-01-26 23:04:46 +00:00
assert(phi.Checkerboard() == psi.Checkerboard());
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
// Flops = 6.0*(Nc*Ns) *Ls*vol
M5Dcalls++;
M5Dtime-=usecond();
Global changes to parallel_for structure. Move the comms flags to more sensible names
2017-02-21 10:24:27 +00:00
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
thread_loop( (int ss=0;ss<grid->oSites();ss+=Ls),{ // adds 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
for(int s=0;s<Ls;s++){
Hide internal data
2018-01-26 23:04:46 +00:00
auto tmp = 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
if ( s==0 ) {
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5m(tmp,psi[ss+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
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5p(tmp,psi[ss+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
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else if ( s==(Ls-1)) {
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5m(tmp,psi[ss+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[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5p(tmp,psi[ss+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
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else {
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5m(tmp,psi[ss+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
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5p(tmp,psi[ss+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
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
}
}
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
});
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
M5Dtime+=usecond();
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>
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
void CayleyFermion5D<Impl>::M5Ddag(const FermionField &psi_i,
const FermionField &phi_i,
FermionField &chi_i,
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
{
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
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();
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;
Hide internal data
2018-01-26 23:04:46 +00:00
assert(phi.Checkerboard() == psi.Checkerboard());
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
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
// Flops = 6.0*(Nc*Ns) *Ls*vol
M5Dcalls++;
M5Dtime-=usecond();
Global changes to parallel_for structure. Move the comms flags to more sensible names
2017-02-21 10:24:27 +00:00
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
thread_loop( (int ss=0;ss<grid->oSites();ss+=Ls),{ // adds Ls
Hide internal data
2018-01-26 23:04:46 +00:00
auto tmp = 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=0;s<Ls;s++){
if ( s==0 ) {
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5p(tmp,psi[ss+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
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5m(tmp,psi[ss+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
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else if ( s==(Ls-1)) {
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5p(tmp,psi[ss+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[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5m(tmp,psi[ss+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
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
} else {
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5p(tmp,psi[ss+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
chi[ss+s]=diag[s]*phi[ss+s]+upper[s]*tmp;
Hide internal data
2018-01-26 23:04:46 +00:00
spProj5m(tmp,psi[ss+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
chi[ss+s]=chi[ss+s]+lower[s]*tmp;
}
}
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
});
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
M5Dtime+=usecond();
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>
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
void CayleyFermion5D<Impl>::MooeeInv (const FermionField &psi_i, FermionField &chi_i)
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
{
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.Grid();
auto psi = psi_i.View();
auto chi = chi_i.View();
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
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
int Ls=this->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
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
MooeeInvCalls++;
MooeeInvTime-=usecond();
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
thread_loop((int ss=0;ss<grid->oSites();ss+=Ls),{ // adds Ls
Hide internal data
2018-01-26 23:04:46 +00:00
auto tmp = 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
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
// flops = 12*2*Ls + 12*2*Ls + 3*12*Ls + 12*2*Ls = 12*Ls * (9) = 108*Ls flops
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^{\prime})^{-1}
chi[ss]=psi[ss]; // chi[0]=psi[0]
for(int s=1;s<Ls;s++){
Namespace, format
2018-01-14 23:44:16 +00:00
spProj5p(tmp,chi[ss+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
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
Namespace, format
2018-01-14 23:44:16 +00:00
spProj5m(tmp,chi[ss+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
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]
Namespace, format
2018-01-14 23:44:16 +00:00
spProj5p(tmp,chi[ss+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
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--){
Namespace, format
2018-01-14 23:44:16 +00:00
spProj5m(tmp,chi[ss+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
chi[ss+s] = chi[ss+s] - uee[s]*tmp;
}
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
});
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
MooeeInvTime+=usecond();
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>
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
void CayleyFermion5D<Impl>::MooeeInvDag (const FermionField &psi_i, FermionField &chi_i)
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
{
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
chi_i.Checkerboard()=psi_i.Checkerboard();
GridBase *grid=psi_i.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
int Ls=this->Ls;
View introduction to prepare for accelerator offload. Probably same problem exists for stencil object
2018-03-04 16:38:08 +00:00
auto psi = psi_i.View();
auto chi = chi_i.View();
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
Hide internal data
2018-01-26 23:04:46 +00:00
assert(psi.Checkerboard() == psi.Checkerboard());
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 coefficient on the dagger
2017-03-30 05:43:13 +01:00
std::vector<Coeff_t> ueec(Ls);
std::vector<Coeff_t> deec(Ls);
std::vector<Coeff_t> leec(Ls);
std::vector<Coeff_t> ueemc(Ls);
std::vector<Coeff_t> leemc(Ls);
for(int s=0;s<ueec.size();s++){
ueec[s] = conjugate(uee[s]);
deec[s] = conjugate(dee[s]);
leec[s] = conjugate(lee[s]);
ueemc[s]= conjugate(ueem[s]);
leemc[s]= conjugate(leem[s]);
}
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
MooeeInvCalls++;
MooeeInvTime-=usecond();
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
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
thread_loop((int ss=0;ss<grid->oSites();ss+=Ls),{ // adds 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
Hide internal data
2018-01-26 23:04:46 +00:00
auto tmp = 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
// Apply (U^{\prime})^{-dagger}
chi[ss]=psi[ss];
for (int s=1;s<Ls;s++){
Namespace, format
2018-01-14 23:44:16 +00:00
spProj5m(tmp,chi[ss+s-1]);
conjugate coefficient on the dagger
2017-03-30 05:43:13 +01:00
chi[ss+s] = psi[ss+s]-ueec[s-1]*tmp;
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++){
Namespace, format
2018-01-14 23:44:16 +00:00
spProj5p(tmp,chi[ss+s]);
conjugate coefficient on the dagger
2017-03-30 05:43:13 +01:00
chi[ss+Ls-1] = chi[ss+Ls-1] - ueemc[s]*tmp;
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++){
spProj5m(tmp,chi[ss+Ls-1]);
conjugate coefficient on the dagger
2017-03-30 05:43:13 +01:00
chi[ss+s] = (1.0/deec[s])*chi[ss+s]-(leemc[s]/deec[Ls-1])*tmp;
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 coefficient on the dagger
2017-03-30 05:43:13 +01:00
chi[ss+Ls-1]= (1.0/deec[Ls-1])*chi[ss+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
// Apply L^{-dagger}
for (int s=Ls-2;s>=0;s--){
spProj5p(tmp,chi[ss+s+1]);
conjugate coefficient on the dagger
2017-03-30 05:43:13 +01:00
chi[ss+s] = chi[ss+s] - leec[s]*tmp;
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
}
parallel_for elimination -> thread_loop
2018-01-28 01:01:14 +00:00
});
Serialisation in malloc fixed
2016-11-29 22:27:55 +00:00
MooeeInvTime+=usecond();
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_CACHE
Namespace, format
2018-01-14 23:44:16 +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, format
2018-01-14 23:44:16 +00:00
NAMESPACE_END(Grid);
Copy Permalink