portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 04:37:05 +01:00
Code Releases Activity

update my fork and fixed conflicts

Browse Source
This commit is contained in:
Vera Guelpers
2018-03-02 17:08:08 +00:00
parent 315a42843f 550142bd6a
commit aaf39222c3
48 changed files with 1277 additions and 164 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
lib/algorithms/Algorithms.h
View File

@ -39,6 +39,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/algorithms/approx/MultiShiftFunction.h>
#include <Grid/algorithms/approx/Forecast.h>
#include <Grid/algorithms/iterative/Deflation.h>
#include <Grid/algorithms/iterative/ConjugateGradient.h>
#include <Grid/algorithms/iterative/ConjugateResidual.h>
#include <Grid/algorithms/iterative/NormalEquations.h>

101
lib/algorithms/iterative/Deflation.h Normal file
View File

@ -0,0 +1,101 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
Copyright (C) 2015
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 */
#ifndef GRID_DEFLATION_H
#define GRID_DEFLATION_H
namespace Grid {
struct ZeroGuesser {
public:
template<class Field>
void operator()(const Field &src,Field &guess) { guess = Zero(); };
};
struct SourceGuesser {
public:
template<class Field>
void operator()(const Field &src,Field &guess) { guess = src; };
};
////////////////////////////////
// Fine grid deflation
////////////////////////////////
template<class Field>
struct DeflatedGuesser {
private:
const std::vector<Field> &evec;
const std::vector<RealD> &eval;
public:
DeflatedGuesser(const std::vector<Field> & _evec,const std::vector<RealD> & _eval) : evec(_evec), eval(_eval) {};
void operator()(const Field &src,Field &guess) {
guess = zero;
assert(evec.size()==eval.size());
auto N = evec.size();
for (int i=0;i<N;i++) {
Field& tmp = evec[i];
axpy(guess,TensorRemove(innerProduct(tmp,src)) / eval[i],tmp,guess);
}
}
};
template<class FineField, class CoarseField>
class LocalCoherenceDeflatedGuesser {
private:
const std::vector<FineField> &subspace;
const std::vector<CoarseField> &evec_coarse;
const std::vector<RealD> &eval_coarse;
public:
LocalCoherenceDeflatedGuesser(const std::vector<FineField> &_subspace,
const std::vector<CoarseField> &_evec_coarse,
const std::vector<RealD> &_eval_coarse)
: subspace(_subspace),
evec_coarse(_evec_coarse),
eval_coarse(_eval_coarse)
{
}
void operator()(const FineField &src,FineField &guess) {
int N = (int)evec_coarse.size();
CoarseField src_coarse(evec_coarse[0]._grid);
CoarseField guess_coarse(evec_coarse[0]._grid); guess_coarse = zero;
blockProject(src,src_coarse,subspace);
for (int i=0;i<N;i++) {
CoarseField & tmp = evec_coarse[i];
axpy(guess_coarse,TensorRemove(innerProduct(tmp,src_coarse)) / eval_coarse[i],tmp,guess_coarse);
}
blockPromote(guess_coarse,guess,subspace);
};
};
}
#endif

15
lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -149,19 +149,6 @@ void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, boo
basisReorderInPlace(_v,sort_vals,idx);
}
// PAB: faster to compute the inner products first then fuse loops.
// If performance critical can improve.
template<class Field>
void basisDeflate(const std::vector<Field> &_v,const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
result = zero;
assert(_v.size()==eval.size());
int N = (int)_v.size();
for (int i=0;i<N;i++) {
Field& tmp = _v[i];
axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
}
}
/////////////////////////////////////////////////////////////
// Implicitly restarted lanczos
/////////////////////////////////////////////////////////////
@ -181,6 +168,7 @@ enum IRLdiagonalisation {
template<class Field> class ImplicitlyRestartedLanczosHermOpTester : public ImplicitlyRestartedLanczosTester<Field>
{
public:
LinearFunction<Field> &_HermOp;
ImplicitlyRestartedLanczosHermOpTester(LinearFunction<Field> &HermOp) : _HermOp(HermOp) { };
int ReconstructEval(int j,RealD resid,Field &B, RealD &eval,RealD evalMaxApprox)
@ -243,6 +231,7 @@ class ImplicitlyRestartedLanczos {
/////////////////////////
public:
//////////////////////////////////////////////////////////////////
// PAB:
//////////////////////////////////////////////////////////////////

175
lib/algorithms/iterative/LocalCoherenceLanczos.h
View File

@ -28,7 +28,10 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
/* END LEGAL */
#ifndef GRID_LOCAL_COHERENCE_IRL_H
#define GRID_LOCAL_COHERENCE_IRL_H
namespace Grid {
struct LanczosParams : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
@ -70,21 +73,24 @@ public:
typedef Lattice<Fobj> FineField;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
std::vector<FineField> &subspace;
ProjectedHermOp(LinearOperatorBase<FineField>& linop, Aggregation<Fobj,CComplex,nbasis> &aggregate) :
_Linop(linop),
_Aggregate(aggregate) { };
ProjectedHermOp(LinearOperatorBase<FineField>& linop, std::vector<FineField> & _subspace) :
_Linop(linop), subspace(_subspace)
{
assert(subspace.size() >0);
};
void operator()(const CoarseField& in, CoarseField& out) {
GridBase *FineGrid = subspace[0]._grid;
int checkerboard = subspace[0].checkerboard;
FineField fin (FineGrid); fin.checkerboard= checkerboard;
FineField fout(FineGrid); fout.checkerboard = checkerboard;
GridBase *FineGrid = _Aggregate.FineGrid;
FineField fin(FineGrid);
FineField fout(FineGrid);
_Aggregate.PromoteFromSubspace(in,fin); std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
_Linop.HermOp(fin,fout); std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
_Aggregate.ProjectToSubspace(out,fout); std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
blockPromote(in,fin,subspace); std::cout<<GridLogIRL<<"ProjectedHermop : Promote to fine"<<std::endl;
_Linop.HermOp(fin,fout); std::cout<<GridLogIRL<<"ProjectedHermop : HermOp (fine) "<<std::endl;
blockProject(out,fout,subspace); std::cout<<GridLogIRL<<"ProjectedHermop : Project to coarse "<<std::endl;
}
};
@ -99,24 +105,27 @@ public:
OperatorFunction<FineField> & _poly;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
std::vector<FineField> &subspace;
ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,LinearOperatorBase<FineField>& linop,
Aggregation<Fobj,CComplex,nbasis> &aggregate) :
ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,
LinearOperatorBase<FineField>& linop,
std::vector<FineField> & _subspace) :
_poly(poly),
_Linop(linop),
_Aggregate(aggregate) { };
subspace(_subspace)
{ };
void operator()(const CoarseField& in, CoarseField& out) {
GridBase *FineGrid = _Aggregate.FineGrid;
FineField fin(FineGrid) ;fin.checkerboard =_Aggregate.checkerboard;
FineField fout(FineGrid);fout.checkerboard =_Aggregate.checkerboard;
_Aggregate.PromoteFromSubspace(in,fin); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
GridBase *FineGrid = subspace[0]._grid;
int checkerboard = subspace[0].checkerboard;
FineField fin (FineGrid); fin.checkerboard =checkerboard;
FineField fout(FineGrid);fout.checkerboard =checkerboard;
blockPromote(in,fin,subspace); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Promote to fine"<<std::endl;
_poly(_Linop,fin,fout); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
_Aggregate.ProjectToSubspace(out,fout); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
blockProject(out,fout,subspace); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
}
};
@ -132,19 +141,23 @@ class ImplicitlyRestartedLanczosSmoothedTester : public ImplicitlyRestartedLanc
LinearFunction<CoarseField> & _Poly;
OperatorFunction<FineField> & _smoother;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
RealD _coarse_relax_tol;
RealD _coarse_relax_tol;
std::vector<FineField> &_subspace;
ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField> &Poly,
OperatorFunction<FineField> &smoother,
LinearOperatorBase<FineField> &Linop,
Aggregation<Fobj,CComplex,nbasis> &Aggregate,
std::vector<FineField> &subspace,
RealD coarse_relax_tol=5.0e3)
: _smoother(smoother), _Linop(Linop),_Aggregate(Aggregate), _Poly(Poly), _coarse_relax_tol(coarse_relax_tol) { };
: _smoother(smoother), _Linop(Linop), _Poly(Poly), _subspace(subspace),
_coarse_relax_tol(coarse_relax_tol)
{ };
int TestConvergence(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
{
CoarseField v(B);
RealD eval_poly = eval;
// Apply operator
_Poly(B,v);
@ -168,14 +181,13 @@ class ImplicitlyRestartedLanczosSmoothedTester : public ImplicitlyRestartedLanc
}
int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
{
GridBase *FineGrid = _Aggregate.FineGrid;
int checkerboard = _Aggregate.checkerboard;
GridBase *FineGrid = _subspace[0]._grid;
int checkerboard = _subspace[0].checkerboard;
FineField fB(FineGrid);fB.checkerboard =checkerboard;
FineField fv(FineGrid);fv.checkerboard =checkerboard;
_Aggregate.PromoteFromSubspace(B,fv);
blockPromote(B,fv,_subspace);
_smoother(_Linop,fv,fB);
RealD eval_poly = eval;
@ -217,27 +229,65 @@ protected:
int _checkerboard;
LinearOperatorBase<FineField> & _FineOp;
// FIXME replace Aggregation with vector of fine; the code reuse is too small for
// the hassle and complexity of cross coupling.
Aggregation<Fobj,CComplex,nbasis> _Aggregate;
std::vector<RealD> evals_fine;
std::vector<RealD> evals_coarse;
std::vector<CoarseField> evec_coarse;
std::vector<RealD> &evals_fine;
std::vector<RealD> &evals_coarse;
std::vector<FineField> &subspace;
std::vector<CoarseField> &evec_coarse;
private:
std::vector<RealD> _evals_fine;
std::vector<RealD> _evals_coarse;
std::vector<FineField> _subspace;
std::vector<CoarseField> _evec_coarse;
public:
LocalCoherenceLanczos(GridBase *FineGrid,
GridBase *CoarseGrid,
LinearOperatorBase<FineField> &FineOp,
int checkerboard) :
GridBase *CoarseGrid,
LinearOperatorBase<FineField> &FineOp,
int checkerboard) :
_CoarseGrid(CoarseGrid),
_FineGrid(FineGrid),
_Aggregate(CoarseGrid,FineGrid,checkerboard),
_FineOp(FineOp),
_checkerboard(checkerboard)
_checkerboard(checkerboard),
evals_fine (_evals_fine),
evals_coarse(_evals_coarse),
subspace (_subspace),
evec_coarse(_evec_coarse)
{
evals_fine.resize(0);
evals_coarse.resize(0);
};
void Orthogonalise(void ) { _Aggregate.Orthogonalise(); }
//////////////////////////////////////////////////////////////////////////
// Alternate constructore, external storage for use by Hadrons module
//////////////////////////////////////////////////////////////////////////
LocalCoherenceLanczos(GridBase *FineGrid,
GridBase *CoarseGrid,
LinearOperatorBase<FineField> &FineOp,
int checkerboard,
std::vector<FineField> &ext_subspace,
std::vector<CoarseField> &ext_coarse,
std::vector<RealD> &ext_eval_fine,
std::vector<RealD> &ext_eval_coarse
) :
_CoarseGrid(CoarseGrid),
_FineGrid(FineGrid),
_FineOp(FineOp),
_checkerboard(checkerboard),
evals_fine (ext_eval_fine),
evals_coarse(ext_eval_coarse),
subspace (ext_subspace),
evec_coarse (ext_coarse)
{
evals_fine.resize(0);
evals_coarse.resize(0);
};
void Orthogonalise(void ) {
CoarseScalar InnerProd(_CoarseGrid);
blockOrthogonalise(InnerProd,subspace);std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
blockOrthogonalise(InnerProd,subspace);std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
};
template<typename T> static RealD normalise(T& v)
{
@ -246,43 +296,44 @@ public:
v = v * (1.0/nn);
return nn;
}
/*
void fakeFine(void)
{
int Nk = nbasis;
_Aggregate.subspace.resize(Nk,_FineGrid);
_Aggregate.subspace[0]=1.0;
_Aggregate.subspace[0].checkerboard=_checkerboard;
normalise(_Aggregate.subspace[0]);
subspace.resize(Nk,_FineGrid);
subspace[0]=1.0;
subspace[0].checkerboard=_checkerboard;
normalise(subspace[0]);
PlainHermOp<FineField> Op(_FineOp);
for(int k=1;k<Nk;k++){
_Aggregate.subspace[k].checkerboard=_checkerboard;
Op(_Aggregate.subspace[k-1],_Aggregate.subspace[k]);
normalise(_Aggregate.subspace[k]);
subspace[k].checkerboard=_checkerboard;
Op(subspace[k-1],subspace[k]);
normalise(subspace[k]);
}
}
*/
void testFine(RealD resid)
{
assert(evals_fine.size() == nbasis);
assert(_Aggregate.subspace.size() == nbasis);
assert(subspace.size() == nbasis);
PlainHermOp<FineField> Op(_FineOp);
ImplicitlyRestartedLanczosHermOpTester<FineField> SimpleTester(Op);
for(int k=0;k<nbasis;k++){
assert(SimpleTester.ReconstructEval(k,resid,_Aggregate.subspace[k],evals_fine[k],1.0)==1);
assert(SimpleTester.ReconstructEval(k,resid,subspace[k],evals_fine[k],1.0)==1);
}
}
void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax)
{
assert(evals_fine.size() == nbasis);
assert(_Aggregate.subspace.size() == nbasis);
assert(subspace.size() == nbasis);
//////////////////////////////////////////////////////////////////////////////////////////////////
// create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
//////////////////////////////////////////////////////////////////////////////////////////////////
Chebyshev<FineField> ChebySmooth(cheby_smooth);
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (ChebySmooth,_FineOp,_Aggregate);
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,_Aggregate,relax);
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (ChebySmooth,_FineOp,subspace);
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
for(int k=0;k<evec_coarse.size();k++){
if ( k < nbasis ) {
@ -302,34 +353,34 @@ public:
PlainHermOp<FineField> Op(_FineOp);
evals_fine.resize(Nm);
_Aggregate.subspace.resize(Nm,_FineGrid);
subspace.resize(Nm,_FineGrid);
ImplicitlyRestartedLanczos<FineField> IRL(ChebyOp,Op,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
FineField src(_FineGrid); src=1.0; src.checkerboard = _checkerboard;
int Nconv;
IRL.calc(evals_fine,_Aggregate.subspace,src,Nconv,false);
IRL.calc(evals_fine,subspace,src,Nconv,false);
// Shrink down to number saved
assert(Nstop>=nbasis);
assert(Nconv>=nbasis);
evals_fine.resize(nbasis);
_Aggregate.subspace.resize(nbasis,_FineGrid);
subspace.resize(nbasis,_FineGrid);
}
void calcCoarse(ChebyParams cheby_op,ChebyParams cheby_smooth,RealD relax,
int Nstop, int Nk, int Nm,RealD resid,
RealD MaxIt, RealD betastp, int MinRes)
{
Chebyshev<FineField> Cheby(cheby_op);
ProjectedHermOp<Fobj,CComplex,nbasis> Op(_FineOp,_Aggregate);
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,_Aggregate);
ProjectedHermOp<Fobj,CComplex,nbasis> Op(_FineOp,subspace);
ProjectedFunctionHermOp<Fobj,CComplex,nbasis> ChebyOp (Cheby,_FineOp,subspace);
//////////////////////////////////////////////////////////////////////////////////////////////////
// create a smoother and see if we can get a cheap convergence test and smooth inside the IRL
//////////////////////////////////////////////////////////////////////////////////////////////////
Chebyshev<FineField> ChebySmooth(cheby_smooth);
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,_Aggregate,relax);
ImplicitlyRestartedLanczosSmoothedTester<Fobj,CComplex,nbasis> ChebySmoothTester(ChebyOp,ChebySmooth,_FineOp,subspace,relax);
evals_coarse.resize(Nm);
evec_coarse.resize(Nm,_CoarseGrid);

33
lib/algorithms/iterative/SchurRedBlack.h
View File

@ -107,7 +107,12 @@ namespace Grid {
};
template<class Matrix>
void operator() (Matrix & _Matrix,const Field &in, Field &out){
void operator() (Matrix & _Matrix,const Field &in, Field &out){
ZeroGuesser guess;
(*this)(_Matrix,in,out,guess);
}
template<class Matrix, class Guesser>
void operator() (Matrix & _Matrix,const Field &in, Field &out, Guesser &guess){
// FIXME CGdiagonalMee not implemented virtual function
// FIXME use CBfactorise to control schur decomp
@ -129,7 +134,6 @@ namespace Grid {
pickCheckerboard(Odd ,src_o,in);
pickCheckerboard(Even,sol_e,out);
pickCheckerboard(Odd ,sol_o,out);
std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve checkerboards picked" <<std::endl;
/////////////////////////////////////////////////////
@ -146,6 +150,7 @@ namespace Grid {
// Call the red-black solver
//////////////////////////////////////////////////////////////
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver calling the Mpc solver" <<std::endl;
guess(src_o,sol_o);
_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver called the Mpc solver" <<std::endl;
@ -189,7 +194,12 @@ namespace Grid {
CBfactorise=cb;
};
template<class Matrix>
void operator() (Matrix & _Matrix,const Field &in, Field &out){
void operator() (Matrix & _Matrix,const Field &in, Field &out){
ZeroGuesser guess;
(*this)(_Matrix,in,out,guess);
}
template<class Matrix, class Guesser>
void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
// FIXME CGdiagonalMee not implemented virtual function
// FIXME use CBfactorise to control schur decomp
@ -225,6 +235,7 @@ namespace Grid {
// Call the red-black solver
//////////////////////////////////////////////////////////////
std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
guess(src_o,sol_o);
_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
///////////////////////////////////////////////////
@ -268,7 +279,12 @@ namespace Grid {
};
template<class Matrix>
void operator() (Matrix & _Matrix,const Field &in, Field &out){
void operator() (Matrix & _Matrix,const Field &in, Field &out){
ZeroGuesser guess;
(*this)(_Matrix,in,out,guess);
}
template<class Matrix,class Guesser>
void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
// FIXME CGdiagonalMee not implemented virtual function
// FIXME use CBfactorise to control schur decomp
@ -305,6 +321,7 @@ namespace Grid {
//////////////////////////////////////////////////////////////
std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
// _HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
guess(src_o,tmp);
_HermitianRBSolver(_HermOpEO,src_o,tmp); assert(tmp.checkerboard==Odd);
_Matrix.MooeeInv(tmp,sol_o); assert( sol_o.checkerboard ==Odd);
@ -347,7 +364,12 @@ namespace Grid {
};
template<class Matrix>
void operator() (Matrix & _Matrix,const Field &in, Field &out){
void operator() (Matrix & _Matrix,const Field &in, Field &out){
ZeroGuesser guess;
(*this)(_Matrix,in,out,guess);
}
template<class Matrix, class Guesser>
void operator() (Matrix & _Matrix,const Field &in, Field &out,Guesser &guess){
// FIXME CGdiagonalMee not implemented virtual function
// FIXME use CBfactorise to control schur decomp
@ -385,6 +407,7 @@ namespace Grid {
std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
// _HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
// _HermitianRBSolver(_HermOpEO,src_o,tmp); assert(tmp.checkerboard==Odd);
guess(src_o,tmp);
_HermitianRBSolver(src_o,tmp); assert(tmp.checkerboard==Odd);
_Matrix.MooeeInv(tmp,sol_o); assert( sol_o.checkerboard ==Odd);

28
lib/communicator/Communicator_mpi3.cc
View File

@ -44,11 +44,15 @@ void CartesianCommunicator::Init(int *argc, char ***argv)
MPI_Initialized(&flag); // needed to coexist with other libs apparently
if ( !flag ) {
MPI_Init_thread(argc,argv,MPI_THREAD_MULTIPLE,&provided);
assert (provided == MPI_THREAD_MULTIPLE);
//If only 1 comms thread we require any threading mode other than SINGLE, but for multiple comms threads we need MULTIPLE
if( (nCommThreads == 1 && provided == MPI_THREAD_SINGLE) ||
(nCommThreads > 1 && provided != MPI_THREAD_MULTIPLE) )
assert(0);
}
Grid_quiesce_nodes();
// Never clean up as done once.
MPI_Comm_dup (MPI_COMM_WORLD,&communicator_world);
GlobalSharedMemory::Init(communicator_world);
@ -85,9 +89,17 @@ void CartesianCommunicator::ProcessorCoorFromRank(int rank, std::vector<int> &c
CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
MPI_Comm optimal_comm;
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm); // Remap using the shared memory optimising routine
////////////////////////////////////////////////////
// Remap using the shared memory optimising routine
// The remap creates a comm which must be freed
////////////////////////////////////////////////////
GlobalSharedMemory::OptimalCommunicator (processors,optimal_comm);
InitFromMPICommunicator(processors,optimal_comm);
SetCommunicator(optimal_comm);
///////////////////////////////////////////////////
// Free the temp communicator
///////////////////////////////////////////////////
MPI_Comm_free(&optimal_comm);
}
//////////////////////////////////
@ -183,8 +195,8 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
} else {
srank = 0;
comm_split = parent.communicator;
// std::cout << " Inherited communicator " <<comm_split <<std::endl;
int ierr = MPI_Comm_dup (parent.communicator,&comm_split);
assert(ierr==0);
}
//////////////////////////////////////////////////////////////////////////////////////////////////////
@ -196,6 +208,11 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
// Take the right SHM buffers
//////////////////////////////////////////////////////////////////////////////////////////////////////
SetCommunicator(comm_split);
///////////////////////////////////////////////
// Free the temp communicator
///////////////////////////////////////////////
MPI_Comm_free(&comm_split);
if(0){
std::cout << " ndim " <<_ndimension<<" " << parent._ndimension << std::endl;
@ -210,6 +227,9 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors,
void CartesianCommunicator::InitFromMPICommunicator(const std::vector<int> &processors, MPI_Comm communicator_base)
{
////////////////////////////////////////////////////
// Creates communicator, and the communicator_halo
////////////////////////////////////////////////////
_ndimension = processors.size();
_processor_coor.resize(_ndimension);

1
lib/communicator/SharedMemory.h
View File

@ -133,6 +133,7 @@ class SharedMemory
public:
SharedMemory() {};
~SharedMemory();
///////////////////////////////////////////////////////////////////////////////////////
// set the buffers & sizes
///////////////////////////////////////////////////////////////////////////////////////

20
lib/communicator/SharedMemoryMPI.cc
View File

@ -182,6 +182,7 @@ void GlobalSharedMemory::OptimalCommunicator(const std::vector<int> &processors,
#ifdef GRID_MPI3_SHMMMAP
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << "SharedMemoryAllocate "<< bytes<< " MMAP implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -218,6 +219,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
assert(((uint64_t)ptr&0x3F)==0);
close(fd);
WorldShmCommBufs[r] =ptr;
std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< bytes<< "bytes)"<<std::endl;
}
_ShmAlloc=1;
_ShmAllocBytes = bytes;
@ -232,6 +234,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
////////////////////////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << "SharedMemoryAllocate "<< bytes<< " SHMOPEN implementation "<<std::endl;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
MPI_Barrier(WorldShmComm);
@ -259,7 +262,11 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
#endif
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
if ( ptr == (void * )MAP_FAILED ) { perror("failed mmap"); assert(0); }
std::cout << "Set WorldShmCommBufs["<<r<<"]="<<ptr<< "("<< size<< "bytes)"<<std::endl;
if ( ptr == (void * )MAP_FAILED ) {
perror("failed mmap");
assert(0);
}
assert(((uint64_t)ptr&0x3F)==0);
WorldShmCommBufs[r] =ptr;
@ -318,11 +325,12 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
heap_size = GlobalSharedMemory::ShmAllocBytes();
for(int r=0;r<ShmSize;r++){
uint32_t sr = (r==ShmRank) ? GlobalSharedMemory::WorldRank : 0 ;
uint32_t wsr = (r==ShmRank) ? GlobalSharedMemory::WorldShmRank : 0 ;
MPI_Allreduce(MPI_IN_PLACE,&sr,1,MPI_UINT32_T,MPI_SUM,comm);
MPI_Allreduce(MPI_IN_PLACE,&wsr,1,MPI_UINT32_T,MPI_SUM,ShmComm);
ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[sr];
ShmCommBufs[r] = GlobalSharedMemory::WorldShmCommBufs[wsr];
// std::cout << "SetCommunicator ShmCommBufs ["<< r<< "] = "<< ShmCommBufs[r]<< " wsr = "<<wsr<<std::endl;
}
ShmBufferFreeAll();
@ -391,5 +399,9 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
return (void *) remote;
}
}
SharedMemory::~SharedMemory()
{
MPI_Comm_free(&ShmComm);
};
}

2
lib/communicator/SharedMemoryNone.cc
View File

@ -122,5 +122,7 @@ void *SharedMemory::ShmBufferTranslate(int rank,void * local_p)
{
return NULL;
}
SharedMemory::~SharedMemory()
{};
}

10
lib/log/Log.h
View File

@ -86,7 +86,7 @@ protected:
Colours &Painter;
int active;
int timing_mode;
int topWidth{-1};
int topWidth{-1}, chanWidth{-1};
static int timestamp;
std::string name, topName;
std::string COLOUR;
@ -126,6 +126,7 @@ public:
}
}
void setTopWidth(const int w) {topWidth = w;}
void setChanWidth(const int w) {chanWidth = w;}
friend std::ostream& operator<< (std::ostream& stream, Logger& log){
@ -136,7 +137,12 @@ public:
stream << std::setw(log.topWidth);
}
stream << log.topName << log.background()<< " : ";
stream << log.colour() << std::left << log.name << log.background() << " : ";
stream << log.colour() << std::left;
if (log.chanWidth > 0)
{
stream << std::setw(log.chanWidth);
}
stream << log.name << log.background() << " : ";
if ( log.timestamp ) {
log.StopWatch->Stop();
GridTime now = log.StopWatch->Elapsed();

2
lib/qcd/action/fermion/CayleyFermion5D.cc
View File

@ -73,7 +73,7 @@ void CayleyFermion5D<Impl>::DminusDag(const FermionField &psi, FermionField &chi
this->DW(psi,tmp_f,DaggerYes);
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
axpby_ssp(chi,Coeff_t(1.0),psi,conjugate(-cs[s]),tmp_f,s,s);// chi = (1-c[s] D_W) psi
}
}

4
lib/qcd/action/fermion/CayleyFermion5Dvec.cc
View File

@ -469,7 +469,7 @@ void CayleyFermion5D<Impl>::MooeeInternalAsm(const FermionField &psi, FermionFie
}
a0 = a0+incr;
a1 = a1+incr;
a2 = a2+sizeof(Simd::scalar_type);
a2 = a2+sizeof(typename Simd::scalar_type);
}}
{
int lexa = s1+LLs*site;
@ -701,7 +701,7 @@ void CayleyFermion5D<Impl>::MooeeInternalZAsm(const FermionField &psi, FermionFi
}
a0 = a0+incr;
a1 = a1+incr;
a2 = a2+sizeof(Simd::scalar_type);
a2 = a2+sizeof(typename Simd::scalar_type);
}}
{
int lexa = s1+LLs*site;

2
lib/qcd/action/fermion/DomainWallEOFAFermionvec.cc
View File

@ -475,7 +475,7 @@ namespace QCD {
}
a0 = a0 + incr;
a1 = a1 + incr;
a2 = a2 + sizeof(Simd::scalar_type);
a2 = a2 + sizeof(typename Simd::scalar_type);
}
}

2
lib/qcd/action/fermion/MobiusEOFAFermionvec.cc
View File

@ -853,7 +853,7 @@ namespace QCD {
a0 = a0 + incr;
a1 = a1 + incr;
a2 = a2 + sizeof(Simd::scalar_type);
a2 = a2 + sizeof(typename Simd::scalar_type);
}
}

2
lib/simd/Grid_avx512.h
View File

@ -556,7 +556,7 @@ namespace Optimization {
v3 = _mm256_add_epi32(v1, v2);
v1 = _mm256_hadd_epi32(v3, v3);
v2 = _mm256_hadd_epi32(v1, v1);
u1 = _mm256_castsi256_si128(v2) // upper half
u1 = _mm256_castsi256_si128(v2); // upper half
u2 = _mm256_extracti128_si256(v2, 1); // lower half
ret = _mm_add_epi32(u1, u2);
return _mm_cvtsi128_si32(ret);

2
lib/simd/Intel512avx.h
View File

@ -79,7 +79,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define ZEND2f(Criir,Ciirr, tmp) "vshufps $0xb1," #Ciirr "," #Ciirr "," #tmp ";\n"\
"vsubps " #tmp "," #Ciirr "," #Criir"{%k7}" ";\n"
#define ZEND1d(Criir,Ciirr, tmp) "vshufpd $0x55," #Criir "," #Criir "," #tmp ";\n"\
#define ZEND1d(Criir,Ciirr, tmp) "vshufpd $0x55," #Criir "," #Criir "," #tmp ";\n"\
"vaddps " #tmp "," #Criir "," #Criir"{%k6}" ";\n"
#define ZEND2d(Criir,Ciirr, tmp) "vshufpd $0x55," #Ciirr "," #Ciirr "," #tmp ";\n"\

72
lib/util/Profiling.h Normal file
View File

@ -0,0 +1,72 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/util/Profiling.h
Copyright (C) 2018
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 */
#ifndef GRID_PERF_PROFILING_H
#define GRID_PERF_PROFILING_H
#include <sstream>
#include <iostream>
#include <functional>
#include <fcntl.h>
#include <sys/stat.h>
#include <sys/wait.h>
#include <sys/types.h>
#include <unistd.h>
#include <signal.h>
struct System
{
static void profile(const std::string& name,std::function<void()> body) {
std::string filename = name.find(".data") == std::string::npos ? (name + ".data") : name;
// Launch profiler
pid_t pid;
std::stringstream s;
s << getpid();
pid = fork();
if (pid == 0) {
auto fd=open("/dev/null",O_RDWR);
dup2(fd,1);
dup2(fd,2);
exit(execl("/usr/bin/perf","perf","record","-o",filename.c_str(),"-p",s.str().c_str(),nullptr));
}
// Run body
body();
// Kill profiler
kill(pid,SIGINT);
waitpid(pid,nullptr,0);
}
static void profile(std::function<void()> body) {
profile("perf.data",body);
}
};
#endif // GRID_PERF_PROFILING_H