portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Updating the feature/clover branch with the newest Hadron package

Browse Source
This commit is contained in:
Guido Cossu
2018-01-12 13:35:51 +00:00
parent e199fda9dc 7bb405e790
commit 3923683e9b
158 changed files with 9663 additions and 5976 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
lib/algorithms/CoarsenedMatrix.h
View File

@ -103,29 +103,32 @@ namespace Grid {
GridBase *CoarseGrid;
GridBase *FineGrid;
std::vector<Lattice<Fobj> > subspace;
int checkerboard;
Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid) :
CoarseGrid(_CoarseGrid),
Aggregation(GridBase *_CoarseGrid,GridBase *_FineGrid,int _checkerboard) :
CoarseGrid(_CoarseGrid),
FineGrid(_FineGrid),
subspace(nbasis,_FineGrid)
subspace(nbasis,_FineGrid),
checkerboard(_checkerboard)
{
};
void Orthogonalise(void){
CoarseScalar InnerProd(CoarseGrid);
std::cout << GridLogMessage <<" Gramm-Schmidt pass 1"<<std::endl;
blockOrthogonalise(InnerProd,subspace);
std::cout << GridLogMessage <<" Gramm-Schmidt pass 2"<<std::endl;
blockOrthogonalise(InnerProd,subspace);
// std::cout << GridLogMessage <<" Gramm-Schmidt checking orthogonality"<<std::endl;
// CheckOrthogonal();
}
void CheckOrthogonal(void){
CoarseVector iProj(CoarseGrid);
CoarseVector eProj(CoarseGrid);
Lattice<CComplex> pokey(CoarseGrid);
for(int i=0;i<nbasis;i++){
blockProject(iProj,subspace[i],subspace);
eProj=zero;
for(int ss=0;ss<CoarseGrid->oSites();ss++){
parallel_for(int ss=0;ss<CoarseGrid->oSites();ss++){
eProj._odata[ss](i)=CComplex(1.0);
}
eProj=eProj - iProj;
@ -137,6 +140,7 @@ namespace Grid {
blockProject(CoarseVec,FineVec,subspace);
}
void PromoteFromSubspace(const CoarseVector &CoarseVec,FineField &FineVec){
FineVec.checkerboard = subspace[0].checkerboard;
blockPromote(CoarseVec,FineVec,subspace);
}
void CreateSubspaceRandom(GridParallelRNG &RNG){
@ -147,6 +151,7 @@ namespace Grid {
Orthogonalise();
}
/*
virtual void CreateSubspaceLanczos(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis)
{
// Run a Lanczos with sloppy convergence
@ -195,7 +200,7 @@ namespace Grid {
std::cout << GridLogMessage <<"subspace["<<b<<"] = "<<norm2(subspace[b])<<std::endl;
}
}
*/
virtual void CreateSubspace(GridParallelRNG &RNG,LinearOperatorBase<FineField> &hermop,int nn=nbasis) {
RealD scale;

88
lib/algorithms/LinearOperator.h
View File

@ -312,20 +312,34 @@ namespace Grid {
public:
SchurStaggeredOperator (Matrix &Mat): _Mat(Mat){};
virtual void HermOpAndNorm(const Field &in, Field &out,RealD &n1,RealD &n2){
GridLogIterative.TimingMode(1);
std::cout << GridLogIterative << " HermOpAndNorm "<<std::endl;
n2 = Mpc(in,out);
std::cout << GridLogIterative << " HermOpAndNorm.Mpc "<<std::endl;
ComplexD dot= innerProduct(in,out);
std::cout << GridLogIterative << " HermOpAndNorm.innerProduct "<<std::endl;
n1 = real(dot);
}
virtual void HermOp(const Field &in, Field &out){
std::cout << GridLogIterative << " HermOp "<<std::endl;
Mpc(in,out);
}
virtual RealD Mpc (const Field &in, Field &out) {
Field tmp(in._grid);
_Mat.Meooe(in,tmp);
_Mat.MooeeInv(tmp,out);
_Mat.MeooeDag(out,tmp);
Field tmp2(in._grid);
std::cout << GridLogIterative << " HermOp.Mpc "<<std::endl;
_Mat.Mooee(in,out);
return axpy_norm(out,-1.0,tmp,out);
_Mat.Mooee(out,tmp);
std::cout << GridLogIterative << " HermOp.MooeeMooee "<<std::endl;
_Mat.Meooe(in,out);
_Mat.Meooe(out,tmp2);
std::cout << GridLogIterative << " HermOp.MeooeMeooe "<<std::endl;
RealD nn=axpy_norm(out,-1.0,tmp2,tmp);
std::cout << GridLogIterative << " HermOp.axpy_norm "<<std::endl;
return nn;
}
virtual RealD MpcDag (const Field &in, Field &out){
return Mpc(in,out);
@ -350,6 +364,14 @@ namespace Grid {
virtual void operator() (const Field &in, Field &out) = 0;
};
template<class Field> class IdentityLinearFunction : public LinearFunction<Field> {
public:
void operator() (const Field &in, Field &out){
out = in;
};
};
/////////////////////////////////////////////////////////////
// Base classes for Multishift solvers for operators
/////////////////////////////////////////////////////////////
@ -372,6 +394,64 @@ namespace Grid {
};
*/
////////////////////////////////////////////////////////////////////////////////////////////
// Hermitian operator Linear function and operator function
////////////////////////////////////////////////////////////////////////////////////////////
template<class Field>
class HermOpOperatorFunction : public OperatorFunction<Field> {
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Linop.HermOp(in,out);
};
};
template<typename Field>
class PlainHermOp : public LinearFunction<Field> {
public:
LinearOperatorBase<Field> &_Linop;
PlainHermOp(LinearOperatorBase<Field>& linop) : _Linop(linop)
{}
void operator()(const Field& in, Field& out) {
_Linop.HermOp(in,out);
}
};
template<typename Field>
class FunctionHermOp : public LinearFunction<Field> {
public:
OperatorFunction<Field> & _poly;
LinearOperatorBase<Field> &_Linop;
FunctionHermOp(OperatorFunction<Field> & poly,LinearOperatorBase<Field>& linop)
: _poly(poly), _Linop(linop) {};
void operator()(const Field& in, Field& out) {
_poly(_Linop,in,out);
}
};
template<class Field>
class Polynomial : public OperatorFunction<Field> {
private:
std::vector<RealD> Coeffs;
public:
Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Field AtoN(in._grid);
Field Mtmp(in._grid);
AtoN = in;
out = AtoN*Coeffs[0];
for(int n=1;n<Coeffs.size();n++){
Mtmp = AtoN;
Linop.HermOp(Mtmp,AtoN);
out=out+AtoN*Coeffs[n];
}
};
};
}

48
lib/algorithms/approx/Chebyshev.h
View File

@ -34,41 +34,12 @@ Author: Christoph Lehner <clehner@bnl.gov>
namespace Grid {
////////////////////////////////////////////////////////////////////////////////////////////
// Simple general polynomial with user supplied coefficients
////////////////////////////////////////////////////////////////////////////////////////////
template<class Field>
class HermOpOperatorFunction : public OperatorFunction<Field> {
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Linop.HermOp(in,out);
};
};
template<class Field>
class Polynomial : public OperatorFunction<Field> {
private:
std::vector<RealD> Coeffs;
public:
Polynomial(std::vector<RealD> &_Coeffs) : Coeffs(_Coeffs) { };
// Implement the required interface
void operator() (LinearOperatorBase<Field> &Linop, const Field &in, Field &out) {
Field AtoN(in._grid);
Field Mtmp(in._grid);
AtoN = in;
out = AtoN*Coeffs[0];
// std::cout <<"Poly in " <<norm2(in)<<" size "<< Coeffs.size()<<std::endl;
// std::cout <<"Coeffs[0]= "<<Coeffs[0]<< " 0 " <<norm2(out)<<std::endl;
for(int n=1;n<Coeffs.size();n++){
Mtmp = AtoN;
Linop.HermOp(Mtmp,AtoN);
out=out+AtoN*Coeffs[n];
// std::cout <<"Coeffs "<<n<<"= "<< Coeffs[n]<< " 0 " <<std::endl;
// std::cout << n<<" " <<norm2(out)<<std::endl;
}
};
};
struct ChebyParams : Serializable {
GRID_SERIALIZABLE_CLASS_MEMBERS(ChebyParams,
RealD, alpha,
RealD, beta,
int, Npoly);
};
////////////////////////////////////////////////////////////////////////////////////////////
// Generic Chebyshev approximations
@ -83,8 +54,10 @@ namespace Grid {
public:
void csv(std::ostream &out){
RealD diff = hi-lo;
for (RealD x=lo-0.2*diff; x<hi+0.2*diff; x+=(hi-lo)/1000) {
RealD diff = hi-lo;
RealD delta = (hi-lo)*1.0e-9;
for (RealD x=lo; x<hi; x+=delta) {
delta*=1.1;
RealD f = approx(x);
out<< x<<" "<<f<<std::endl;
}
@ -100,6 +73,7 @@ namespace Grid {
};
Chebyshev(){};
Chebyshev(ChebyParams p){ Init(p.alpha,p.beta,p.Npoly);};
Chebyshev(RealD _lo,RealD _hi,int _order, RealD (* func)(RealD) ) {Init(_lo,_hi,_order,func);};
Chebyshev(RealD _lo,RealD _hi,int _order) {Init(_lo,_hi,_order);};

753
lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockImplicitlyRestartedLanczos.h
View File

@ -1,753 +0,0 @@
/*************************************************************************************
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>
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Chulwoo Jung <chulwoo@bnl.gov>
Author: Christoph Lehner <clehner@bnl.gov>
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_BIRL_H
#define GRID_BIRL_H
#include <string.h> //memset
#include <zlib.h>
#include <sys/stat.h>
#include <Grid/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockedGrid.h>
#include <Grid/algorithms/iterative/BlockImplicitlyRestartedLanczos/FieldBasisVector.h>
#include <Grid/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockProjector.h>
namespace Grid {
/////////////////////////////////////////////////////////////
// Implicitly restarted lanczos
/////////////////////////////////////////////////////////////
template<class Field>
class BlockImplicitlyRestartedLanczos {
const RealD small = 1.0e-16;
public:
int lock;
int get;
int Niter;
int converged;
int Nminres; // Minimum number of restarts; only check for convergence after
int Nstop; // Number of evecs checked for convergence
int Nk; // Number of converged sought
int Np; // Np -- Number of spare vecs in kryloc space
int Nm; // Nm -- total number of vectors
int orth_period;
RealD OrthoTime;
RealD eresid, betastp;
SortEigen<Field> _sort;
LinearFunction<Field> &_HermOp;
LinearFunction<Field> &_HermOpTest;
/////////////////////////
// Constructor
/////////////////////////
BlockImplicitlyRestartedLanczos(
LinearFunction<Field> & HermOp,
LinearFunction<Field> & HermOpTest,
int _Nstop, // sought vecs
int _Nk, // sought vecs
int _Nm, // spare vecs
RealD _eresid, // resid in lmdue deficit
RealD _betastp, // if beta(k) < betastp: converged
int _Niter, // Max iterations
int _Nminres, int _orth_period = 1) :
_HermOp(HermOp),
_HermOpTest(HermOpTest),
Nstop(_Nstop),
Nk(_Nk),
Nm(_Nm),
eresid(_eresid),
betastp(_betastp),
Niter(_Niter),
Nminres(_Nminres),
orth_period(_orth_period)
{
Np = Nm-Nk; assert(Np>0);
};
BlockImplicitlyRestartedLanczos(
LinearFunction<Field> & HermOp,
LinearFunction<Field> & HermOpTest,
int _Nk, // sought vecs
int _Nm, // spare vecs
RealD _eresid, // resid in lmdue deficit
RealD _betastp, // if beta(k) < betastp: converged
int _Niter, // Max iterations
int _Nminres,
int _orth_period = 1) :
_HermOp(HermOp),
_HermOpTest(HermOpTest),
Nstop(_Nk),
Nk(_Nk),
Nm(_Nm),
eresid(_eresid),
betastp(_betastp),
Niter(_Niter),
Nminres(_Nminres),
orth_period(_orth_period)
{
Np = Nm-Nk; assert(Np>0);
};
/* Saad PP. 195
1. Choose an initial vector v1 of 2-norm unity. Set β1 ≡ 0, v0 ≡ 0
2. For k = 1,2,...,m Do:
3. wk:=Avk−βkv_{k−1}
4. αk:=(wk,vk) //
5. wk:=wk−αkvk // wk orthog vk
6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
7. vk+1 := wk/βk+1
8. EndDo
*/
void step(std::vector<RealD>& lmd,
std::vector<RealD>& lme,
BasisFieldVector<Field>& evec,
Field& w,int Nm,int k)
{
assert( k< Nm );
GridStopWatch gsw_op,gsw_o;
Field& evec_k = evec[k];
gsw_op.Start();
_HermOp(evec_k,w);
gsw_op.Stop();
if(k>0){
w -= lme[k-1] * evec[k-1];
}
ComplexD zalph = innerProduct(evec_k,w); // 4. αk:=(wk,vk)
RealD alph = real(zalph);
w = w - alph * evec_k;// 5. wk:=wk−αkvk
RealD beta = normalise(w); // 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
// 7. vk+1 := wk/βk+1
std::cout<<GridLogMessage << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
const RealD tiny = 1.0e-20;
if ( beta < tiny ) {
std::cout<<GridLogMessage << " beta is tiny "<<beta<<std::endl;
}
lmd[k] = alph;
lme[k] = beta;
gsw_o.Start();
if (k>0 && k % orth_period == 0) {
orthogonalize(w,evec,k); // orthonormalise
}
gsw_o.Stop();
if(k < Nm-1) {
evec[k+1] = w;
}
std::cout << GridLogMessage << "Timing: operator=" << gsw_op.Elapsed() <<
" orth=" << gsw_o.Elapsed() << std::endl;
}
void qr_decomp(std::vector<RealD>& lmd,
std::vector<RealD>& lme,
int Nk,
int Nm,
std::vector<RealD>& Qt,
RealD Dsh,
int kmin,
int kmax)
{
int k = kmin-1;
RealD x;
RealD Fden = 1.0/hypot(lmd[k]-Dsh,lme[k]);
RealD c = ( lmd[k] -Dsh) *Fden;
RealD s = -lme[k] *Fden;
RealD tmpa1 = lmd[k];
RealD tmpa2 = lmd[k+1];
RealD tmpb = lme[k];
lmd[k] = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
lme[k] = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
x =-s*lme[k+1];
lme[k+1] = c*lme[k+1];
for(int i=0; i<Nk; ++i){
RealD Qtmp1 = Qt[i+Nm*k ];
RealD Qtmp2 = Qt[i+Nm*(k+1)];
Qt[i+Nm*k ] = c*Qtmp1 - s*Qtmp2;
Qt[i+Nm*(k+1)] = s*Qtmp1 + c*Qtmp2;
}
// Givens transformations
for(int k = kmin; k < kmax-1; ++k){
RealD Fden = 1.0/hypot(x,lme[k-1]);
RealD c = lme[k-1]*Fden;
RealD s = - x*Fden;
RealD tmpa1 = lmd[k];
RealD tmpa2 = lmd[k+1];
RealD tmpb = lme[k];
lmd[k] = c*c*tmpa1 +s*s*tmpa2 -2.0*c*s*tmpb;
lmd[k+1] = s*s*tmpa1 +c*c*tmpa2 +2.0*c*s*tmpb;
lme[k] = c*s*(tmpa1-tmpa2) +(c*c-s*s)*tmpb;
lme[k-1] = c*lme[k-1] -s*x;
if(k != kmax-2){
x = -s*lme[k+1];
lme[k+1] = c*lme[k+1];
}
for(int i=0; i<Nk; ++i){
RealD Qtmp1 = Qt[i+Nm*k ];
RealD Qtmp2 = Qt[i+Nm*(k+1)];
Qt[i+Nm*k ] = c*Qtmp1 -s*Qtmp2;
Qt[i+Nm*(k+1)] = s*Qtmp1 +c*Qtmp2;
}
}
}
#ifdef USE_LAPACK_IRL
#define LAPACK_INT int
//long long
void diagonalize_lapack(std::vector<RealD>& lmd,
std::vector<RealD>& lme,
int N1,
int N2,
std::vector<RealD>& Qt,
GridBase *grid){
std::cout << GridLogMessage << "diagonalize_lapack start\n";
GridStopWatch gsw;
const int size = Nm;
// tevals.resize(size);
// tevecs.resize(size);
LAPACK_INT NN = N1;
std::vector<double> evals_tmp(NN);
std::vector<double> evec_tmp(NN*NN);
memset(&evec_tmp[0],0,sizeof(double)*NN*NN);
// double AA[NN][NN];
std::vector<double> DD(NN);
std::vector<double> EE(NN);
for (int i = 0; i< NN; i++)
for (int j = i - 1; j <= i + 1; j++)
if ( j < NN && j >= 0 ) {
if (i==j) DD[i] = lmd[i];
if (i==j) evals_tmp[i] = lmd[i];
if (j==(i-1)) EE[j] = lme[j];
}
LAPACK_INT evals_found;
LAPACK_INT lwork = ( (18*NN) > (1+4*NN+NN*NN)? (18*NN):(1+4*NN+NN*NN)) ;
LAPACK_INT liwork = 3+NN*10 ;
std::vector<LAPACK_INT> iwork(liwork);
std::vector<double> work(lwork);
std::vector<LAPACK_INT> isuppz(2*NN);
char jobz = 'V'; // calculate evals & evecs
char range = 'I'; // calculate all evals
// char range = 'A'; // calculate all evals
char uplo = 'U'; // refer to upper half of original matrix
char compz = 'I'; // Compute eigenvectors of tridiagonal matrix
std::vector<int> ifail(NN);
LAPACK_INT info;
// int total = QMP_get_number_of_nodes();
// int node = QMP_get_node_number();
// GridBase *grid = evec[0]._grid;
int total = grid->_Nprocessors;
int node = grid->_processor;
int interval = (NN/total)+1;
double vl = 0.0, vu = 0.0;
LAPACK_INT il = interval*node+1 , iu = interval*(node+1);
if (iu > NN) iu=NN;
double tol = 0.0;
if (1) {
memset(&evals_tmp[0],0,sizeof(double)*NN);
if ( il <= NN){
std::cout << GridLogMessage << "dstegr started" << std::endl;
gsw.Start();
dstegr(&jobz, &range, &NN,
(double*)&DD[0], (double*)&EE[0],
&vl, &vu, &il, &iu, // these four are ignored if second parameteris 'A'
&tol, // tolerance
&evals_found, &evals_tmp[0], (double*)&evec_tmp[0], &NN,
&isuppz[0],
&work[0], &lwork, &iwork[0], &liwork,
&info);
gsw.Stop();
std::cout << GridLogMessage << "dstegr completed in " << gsw.Elapsed() << std::endl;
for (int i = iu-1; i>= il-1; i--){
evals_tmp[i] = evals_tmp[i - (il-1)];
if (il>1) evals_tmp[i-(il-1)]=0.;
for (int j = 0; j< NN; j++){
evec_tmp[i*NN + j] = evec_tmp[(i - (il-1)) * NN + j];
if (il>1) evec_tmp[(i-(il-1)) * NN + j]=0.;
}
}
}
{
// QMP_sum_double_array(evals_tmp,NN);
// QMP_sum_double_array((double *)evec_tmp,NN*NN);
grid->GlobalSumVector(&evals_tmp[0],NN);
grid->GlobalSumVector(&evec_tmp[0],NN*NN);
}
}
// cheating a bit. It is better to sort instead of just reversing it, but the document of the routine says evals are sorted in increasing order. qr gives evals in decreasing order.
for(int i=0;i<NN;i++){
for(int j=0;j<NN;j++)
Qt[(NN-1-i)*N2+j]=evec_tmp[i*NN + j];
lmd [NN-1-i]=evals_tmp[i];
}
std::cout << GridLogMessage << "diagonalize_lapack complete\n";
}
#undef LAPACK_INT
#endif
void diagonalize(std::vector<RealD>& lmd,
std::vector<RealD>& lme,
int N2,
int N1,
std::vector<RealD>& Qt,
GridBase *grid)
{
#ifdef USE_LAPACK_IRL
const int check_lapack=0; // just use lapack if 0, check against lapack if 1
if(!check_lapack)
return diagonalize_lapack(lmd,lme,N2,N1,Qt,grid);
std::vector <RealD> lmd2(N1);
std::vector <RealD> lme2(N1);
std::vector<RealD> Qt2(N1*N1);
for(int k=0; k<N1; ++k){
lmd2[k] = lmd[k];
lme2[k] = lme[k];
}
for(int k=0; k<N1*N1; ++k)
Qt2[k] = Qt[k];
// diagonalize_lapack(lmd2,lme2,Nm2,Nm,Qt,grid);
#endif
int Niter = 10000*N1;
int kmin = 1;
int kmax = N2;
// (this should be more sophisticated)
for(int iter=0; ; ++iter){
if ( (iter+1)%(100*N1)==0)
std::cout<<GridLogMessage << "[QL method] Not converged - iteration "<<iter+1<<"\n";
// determination of 2x2 leading submatrix
RealD dsub = lmd[kmax-1]-lmd[kmax-2];
RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
// (Dsh: shift)
// transformation
qr_decomp(lmd,lme,N2,N1,Qt,Dsh,kmin,kmax);
// Convergence criterion (redef of kmin and kamx)
for(int j=kmax-1; j>= kmin; --j){
RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
if(fabs(lme[j-1])+dds > dds){
kmax = j+1;
goto continued;
}
}
Niter = iter;
#ifdef USE_LAPACK_IRL
if(check_lapack){
const double SMALL=1e-8;
diagonalize_lapack(lmd2,lme2,N2,N1,Qt2,grid);
std::vector <RealD> lmd3(N2);
for(int k=0; k<N2; ++k) lmd3[k]=lmd[k];
_sort.push(lmd3,N2);
_sort.push(lmd2,N2);
for(int k=0; k<N2; ++k){
if (fabs(lmd2[k] - lmd3[k]) >SMALL) std::cout<<GridLogMessage <<"lmd(qr) lmd(lapack) "<< k << ": " << lmd2[k] <<" "<< lmd3[k] <<std::endl;
// if (fabs(lme2[k] - lme[k]) >SMALL) std::cout<<GridLogMessage <<"lme(qr)-lme(lapack) "<< k << ": " << lme2[k] - lme[k] <<std::endl;
}
for(int k=0; k<N1*N1; ++k){
// if (fabs(Qt2[k] - Qt[k]) >SMALL) std::cout<<GridLogMessage <<"Qt(qr)-Qt(lapack) "<< k << ": " << Qt2[k] - Qt[k] <<std::endl;
}
}
#endif
return;
continued:
for(int j=0; j<kmax-1; ++j){
RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
if(fabs(lme[j])+dds > dds){
kmin = j+1;
break;
}
}
}
std::cout<<GridLogMessage << "[QL method] Error - Too many iteration: "<<Niter<<"\n";
abort();
}
#if 1
template<typename T>
static RealD normalise(T& v)
{
RealD nn = norm2(v);
nn = sqrt(nn);
v = v * (1.0/nn);
return nn;
}
void orthogonalize(Field& w,
BasisFieldVector<Field>& evec,
int k)
{
double t0=-usecond()/1e6;
evec.orthogonalize(w,k);
normalise(w);
t0+=usecond()/1e6;
OrthoTime +=t0;
}
void setUnit_Qt(int Nm, std::vector<RealD> &Qt) {
for(int i=0; i<Qt.size(); ++i) Qt[i] = 0.0;
for(int k=0; k<Nm; ++k) Qt[k + k*Nm] = 1.0;
}
/* Rudy Arthur's thesis pp.137
------------------------
Require: M > K P = M − K †
Compute the factorization AVM = VM HM + fM eM
repeat
Q=I
for i = 1,...,P do
QiRi =HM −θiI Q = QQi
H M = Q †i H M Q i
end for
βK =HM(K+1,K) σK =Q(M,K)
r=vK+1βK +rσK
VK =VM(1:M)Q(1:M,1:K)
HK =HM(1:K,1:K)
→AVK =VKHK +fKe†K † Extend to an M = K + P step factorization AVM = VMHM + fMeM
until convergence
*/
void calc(std::vector<RealD>& eval,
BasisFieldVector<Field>& evec,
const Field& src,
int& Nconv,
bool reverse,
int SkipTest)
{
GridBase *grid = evec._v[0]._grid;//evec.get(0 + evec_offset)._grid;
assert(grid == src._grid);
std::cout<<GridLogMessage << " -- Nk = " << Nk << " Np = "<< Np << std::endl;
std::cout<<GridLogMessage << " -- Nm = " << Nm << std::endl;
std::cout<<GridLogMessage << " -- size of eval = " << eval.size() << std::endl;
std::cout<<GridLogMessage << " -- size of evec = " << evec.size() << std::endl;
assert(Nm <= evec.size() && Nm <= eval.size());
// quickly get an idea of the largest eigenvalue to more properly normalize the residuum
RealD evalMaxApprox = 0.0;
{
auto src_n = src;
auto tmp = src;
const int _MAX_ITER_IRL_MEVAPP_ = 50;
for (int i=0;i<_MAX_ITER_IRL_MEVAPP_;i++) {
_HermOpTest(src_n,tmp);
RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
RealD vden = norm2(src_n);
RealD na = vnum/vden;
if (fabs(evalMaxApprox/na - 1.0) < 0.05)
i=_MAX_ITER_IRL_MEVAPP_;
evalMaxApprox = na;
std::cout << GridLogMessage << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
src_n = tmp;
}
}
std::vector<RealD> lme(Nm);
std::vector<RealD> lme2(Nm);
std::vector<RealD> eval2(Nm);
std::vector<RealD> eval2_copy(Nm);
std::vector<RealD> Qt(Nm*Nm);
Field f(grid);
Field v(grid);
int k1 = 1;
int k2 = Nk;
Nconv = 0;
RealD beta_k;
// Set initial vector
evec[0] = src;
normalise(evec[0]);
std:: cout<<GridLogMessage <<"norm2(evec[0])= " << norm2(evec[0])<<std::endl;
// Initial Nk steps
OrthoTime=0.;
double t0=usecond()/1e6;
for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
double t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL::Initial steps: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
std::cout<<GridLogMessage <<"IRL::Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
t1=usecond()/1e6;
// Restarting loop begins
for(int iter = 0; iter<Niter; ++iter){
std::cout<<GridLogMessage<<"\n Restart iteration = "<< iter << std::endl;
//
// Rudy does a sort first which looks very different. Getting fed up with sorting out the algo defs.
// We loop over
//
OrthoTime=0.;
for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL:: "<<Np <<" steps: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
std::cout<<GridLogMessage <<"IRL::Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
f *= lme[Nm-1];
t1=usecond()/1e6;
// getting eigenvalues
for(int k=0; k<Nm; ++k){
eval2[k] = eval[k+k1-1];
lme2[k] = lme[k+k1-1];
}
setUnit_Qt(Nm,Qt);
diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL:: diagonalize: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
// sorting
eval2_copy = eval2;
_sort.push(eval2,Nm);
t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL:: eval sorting: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
// Implicitly shifted QR transformations
setUnit_Qt(Nm,Qt);
for(int ip=0; ip<k2; ++ip){
std::cout<<GridLogMessage << "eval "<< ip << " "<< eval2[ip] << std::endl;
}
for(int ip=k2; ip<Nm; ++ip){
std::cout<<GridLogMessage << "qr_decomp "<< ip << " "<< eval2[ip] << std::endl;
qr_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
}
t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL::qr_decomp: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
assert(k2<Nm);
assert(k2<Nm);
assert(k1>0);
evec.rotate(Qt,k1-1,k2+1,0,Nm,Nm);
t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL::QR rotation: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
fflush(stdout);
// Compressed vector f and beta(k2)
f *= Qt[Nm-1+Nm*(k2-1)];
f += lme[k2-1] * evec[k2];
beta_k = norm2(f);
beta_k = sqrt(beta_k);
std::cout<<GridLogMessage<<" beta(k) = "<<beta_k<<std::endl;
RealD betar = 1.0/beta_k;
evec[k2] = betar * f;
lme[k2-1] = beta_k;
// Convergence test
for(int k=0; k<Nm; ++k){
eval2[k] = eval[k];
lme2[k] = lme[k];
std::cout<<GridLogMessage << "eval2[" << k << "] = " << eval2[k] << std::endl;
}
setUnit_Qt(Nm,Qt);
diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL::diagonalize: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
Nconv = 0;
if (iter >= Nminres) {
std::cout << GridLogMessage << "Rotation to test convergence " << std::endl;
Field ev0_orig(grid);
ev0_orig = evec[0];
evec.rotate(Qt,0,Nk,0,Nk,Nm);
{
std::cout << GridLogMessage << "Test convergence" << std::endl;
Field B(grid);
for(int j = 0; j<Nk; j+=SkipTest){
B=evec[j];
//std::cout << "Checkerboard: " << evec[j].checkerboard << std::endl;
B.checkerboard = evec[0].checkerboard;
_HermOpTest(B,v);
RealD vnum = real(innerProduct(B,v)); // HermOp.
RealD vden = norm2(B);
RealD vv0 = norm2(v);
eval2[j] = vnum/vden;
v -= eval2[j]*B;
RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
std::cout.precision(13);
std::cout<<GridLogMessage << "[" << std::setw(3)<< std::setiosflags(std::ios_base::right) <<j<<"] "
<<"eval = "<<std::setw(25)<< std::setiosflags(std::ios_base::left)<< eval2[j] << " (" << eval2_copy[j] << ")"
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25)<< std::setiosflags(std::ios_base::right)<< vv
<<" "<< vnum/(sqrt(vden)*sqrt(vv0))
<< " norm(B["<<j<<"])="<< vden <<std::endl;
// change the criteria as evals are supposed to be sorted, all evals smaller(larger) than Nstop should have converged
if((vv<eresid*eresid) && (j == Nconv) ){
Nconv+=SkipTest;
}
}
// test if we converged, if so, terminate
t1=usecond()/1e6;
std::cout<<GridLogMessage <<"IRL::convergence testing: "<<t1-t0<< "seconds"<<std::endl; t0=t1;
std::cout<<GridLogMessage<<" #modes converged: "<<Nconv<<std::endl;
if( Nconv>=Nstop || beta_k < betastp){
goto converged;
}
std::cout << GridLogMessage << "Rotate back" << std::endl;
//B[j] +=Qt[k+_Nm*j] * _v[k]._odata[ss];
{
Eigen::MatrixXd qm = Eigen::MatrixXd::Zero(Nk,Nk);
for (int k=0;k<Nk;k++)
for (int j=0;j<Nk;j++)
qm(j,k) = Qt[k+Nm*j];
GridStopWatch timeInv;
timeInv.Start();
Eigen::MatrixXd qmI = qm.inverse();
timeInv.Stop();
std::vector<RealD> QtI(Nm*Nm);
for (int k=0;k<Nk;k++)
for (int j=0;j<Nk;j++)
QtI[k+Nm*j] = qmI(j,k);
RealD res_check_rotate_inverse = (qm*qmI - Eigen::MatrixXd::Identity(Nk,Nk)).norm(); // sqrt( |X|^2 )
assert(res_check_rotate_inverse < 1e-7);
evec.rotate(QtI,0,Nk,0,Nk,Nm);
axpy(ev0_orig,-1.0,evec[0],ev0_orig);
std::cout << GridLogMessage << "Rotation done (in " << timeInv.Elapsed() << " = " << timeInv.useconds() << " us" <<
", error = " << res_check_rotate_inverse <<
"); | evec[0] - evec[0]_orig | = " << ::sqrt(norm2(ev0_orig)) << std::endl;
}
}
} else {
std::cout << GridLogMessage << "iter < Nminres: do not yet test for convergence\n";
} // end of iter loop
}
std::cout<<GridLogMessage<<"\n NOT converged.\n";
abort();
converged:
if (SkipTest == 1) {
eval = eval2;
} else {
// test quickly
for (int j=0;j<Nstop;j+=SkipTest) {
std::cout<<GridLogMessage << "Eigenvalue[" << j << "] = " << eval2[j] << " (" << eval2_copy[j] << ")" << std::endl;
}
eval2_copy.resize(eval2.size());
eval = eval2_copy;
}
evec.sortInPlace(eval,reverse);
{
// test
for (int j=0;j<Nstop;j++) {
std::cout<<GridLogMessage << " |e[" << j << "]|^2 = " << norm2(evec[j]) << std::endl;
}
}
//_sort.push(eval,evec,Nconv);
//evec.sort(eval,Nconv);
std::cout<<GridLogMessage << "\n Converged\n Summary :\n";
std::cout<<GridLogMessage << " -- Iterations = "<< Nconv << "\n";
std::cout<<GridLogMessage << " -- beta(k) = "<< beta_k << "\n";
std::cout<<GridLogMessage << " -- Nconv = "<< Nconv << "\n";
}
#endif
};
}
#endif

143
lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockProjector.h
View File

@ -1,143 +0,0 @@
namespace Grid {
/*
BlockProjector
If _HP_BLOCK_PROJECTORS_ is defined, we assume that _evec is a basis that is not
fully orthonormalized (to the precision of the coarse field) and we allow for higher-precision
coarse field than basis field.
*/
//#define _HP_BLOCK_PROJECTORS_
template<typename Field>
class BlockProjector {
public:
BasisFieldVector<Field>& _evec;
BlockedGrid<Field>& _bgrid;
BlockProjector(BasisFieldVector<Field>& evec, BlockedGrid<Field>& bgrid) : _evec(evec), _bgrid(bgrid) {
}
void createOrthonormalBasis(RealD thres = 0.0) {
GridStopWatch sw;
sw.Start();
int cnt = 0;
#pragma omp parallel shared(cnt)
{
int lcnt = 0;
#pragma omp for
for (int b=0;b<_bgrid._o_blocks;b++) {
for (int i=0;i<_evec._Nm;i++) {
auto nrm0 = _bgrid.block_sp(b,_evec._v[i],_evec._v[i]);
// |i> -= <j|i> |j>
for (int j=0;j<i;j++) {
_bgrid.block_caxpy(b,_evec._v[i],-_bgrid.block_sp(b,_evec._v[j],_evec._v[i]),_evec._v[j],_evec._v[i]);
}
auto nrm = _bgrid.block_sp(b,_evec._v[i],_evec._v[i]);
auto eps = nrm/nrm0;
if (Reduce(eps).real() < thres) {
lcnt++;
}
// TODO: if norm is too small, remove this eigenvector/mark as not needed; in practice: set it to zero norm here and return a mask
// that is then used later to decide not to write certain eigenvectors to disk (add a norm calculation before subtraction step and look at nrm/nrm0 < eps to decide)
_bgrid.block_cscale(b,1.0 / sqrt(nrm),_evec._v[i]);
}
}
#pragma omp critical
{
cnt += lcnt;
}
}
sw.Stop();
std::cout << GridLogMessage << "Gram-Schmidt to create blocked basis took " << sw.Elapsed() << " (" << ((RealD)cnt / (RealD)_bgrid._o_blocks / (RealD)_evec._Nm)
<< " below threshold)" << std::endl;
}
template<typename CoarseField>
void coarseToFine(const CoarseField& in, Field& out) {
out = zero;
out.checkerboard = _evec._v[0].checkerboard;
int Nbasis = sizeof(in._odata[0]._internal._internal) / sizeof(in._odata[0]._internal._internal[0]);
assert(Nbasis == _evec._Nm);
#pragma omp parallel for
for (int b=0;b<_bgrid._o_blocks;b++) {
for (int j=0;j<_evec._Nm;j++) {
_bgrid.block_caxpy(b,out,in._odata[b]._internal._internal[j],_evec._v[j],out);
}
}
}
template<typename CoarseField>
void fineToCoarse(const Field& in, CoarseField& out) {
out = zero;
int Nbasis = sizeof(out._odata[0]._internal._internal) / sizeof(out._odata[0]._internal._internal[0]);
assert(Nbasis == _evec._Nm);
Field tmp(_bgrid._grid);
tmp = in;
#pragma omp parallel for
for (int b=0;b<_bgrid._o_blocks;b++) {
for (int j=0;j<_evec._Nm;j++) {
// |rhs> -= <j|rhs> |j>
auto c = _bgrid.block_sp(b,_evec._v[j],tmp);
_bgrid.block_caxpy(b,tmp,-c,_evec._v[j],tmp); // may make this more numerically stable
out._odata[b]._internal._internal[j] = c;
}
}
}
template<typename CoarseField>
void deflateFine(BasisFieldVector<CoarseField>& _coef,const std::vector<RealD>& eval,int N,const Field& src_orig,Field& result) {
result = zero;
for (int i=0;i<N;i++) {
Field tmp(result._grid);
coarseToFine(_coef._v[i],tmp);
axpy(result,TensorRemove(innerProduct(tmp,src_orig)) / eval[i],tmp,result);
}
}
template<typename CoarseField>
void deflateCoarse(BasisFieldVector<CoarseField>& _coef,const std::vector<RealD>& eval,int N,const Field& src_orig,Field& result) {
CoarseField src_coarse(_coef._v[0]._grid);
CoarseField result_coarse = src_coarse;
result_coarse = zero;
fineToCoarse(src_orig,src_coarse);
for (int i=0;i<N;i++) {
axpy(result_coarse,TensorRemove(innerProduct(_coef._v[i],src_coarse)) / eval[i],_coef._v[i],result_coarse);
}
coarseToFine(result_coarse,result);
}
template<typename CoarseField>
void deflate(BasisFieldVector<CoarseField>& _coef,const std::vector<RealD>& eval,int N,const Field& src_orig,Field& result) {
// Deflation on coarse Grid is much faster, so use it by default. Deflation on fine Grid is kept for legacy reasons for now.
deflateCoarse(_coef,eval,N,src_orig,result);
}
};
}

401
lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/BlockedGrid.h
View File

@ -1,401 +0,0 @@
namespace Grid {
template<typename Field>
class BlockedGrid {
public:
GridBase* _grid;
typedef typename Field::scalar_type Coeff_t;
typedef typename Field::vector_type vCoeff_t;
std::vector<int> _bs; // block size
std::vector<int> _nb; // number of blocks
std::vector<int> _l; // local dimensions irrespective of cb
std::vector<int> _l_cb; // local dimensions of checkerboarded vector
std::vector<int> _l_cb_o; // local dimensions of inner checkerboarded vector
std::vector<int> _bs_cb; // block size in checkerboarded vector
std::vector<int> _nb_o; // number of blocks of simd o-sites
int _nd, _blocks, _cf_size, _cf_block_size, _cf_o_block_size, _o_blocks, _block_sites;
BlockedGrid(GridBase* grid, const std::vector<int>& block_size) :
_grid(grid), _bs(block_size), _nd((int)_bs.size()),
_nb(block_size), _l(block_size), _l_cb(block_size), _nb_o(block_size),
_l_cb_o(block_size), _bs_cb(block_size) {
_blocks = 1;
_o_blocks = 1;
_l = grid->FullDimensions();
_l_cb = grid->LocalDimensions();
_l_cb_o = grid->_rdimensions;
_cf_size = 1;
_block_sites = 1;
for (int i=0;i<_nd;i++) {
_l[i] /= grid->_processors[i];
assert(!(_l[i] % _bs[i])); // lattice must accommodate choice of blocksize
int r = _l[i] / _l_cb[i];
assert(!(_bs[i] % r)); // checkerboarding must accommodate choice of blocksize
_bs_cb[i] = _bs[i] / r;
_block_sites *= _bs_cb[i];
_nb[i] = _l[i] / _bs[i];
_nb_o[i] = _nb[i] / _grid->_simd_layout[i];
if (_nb[i] % _grid->_simd_layout[i]) { // simd must accommodate choice of blocksize
std::cout << GridLogMessage << "Problem: _nb[" << i << "] = " << _nb[i] << " _grid->_simd_layout[" << i << "] = " << _grid->_simd_layout[i] << std::endl;
assert(0);
}
_blocks *= _nb[i];
_o_blocks *= _nb_o[i];
_cf_size *= _l[i];
}
_cf_size *= 12 / 2;
_cf_block_size = _cf_size / _blocks;
_cf_o_block_size = _cf_size / _o_blocks;
std::cout << GridLogMessage << "BlockedGrid:" << std::endl;
std::cout << GridLogMessage << " _l = " << _l << std::endl;
std::cout << GridLogMessage << " _l_cb = " << _l_cb << std::endl;
std::cout << GridLogMessage << " _l_cb_o = " << _l_cb_o << std::endl;
std::cout << GridLogMessage << " _bs = " << _bs << std::endl;
std::cout << GridLogMessage << " _bs_cb = " << _bs_cb << std::endl;
std::cout << GridLogMessage << " _nb = " << _nb << std::endl;
std::cout << GridLogMessage << " _nb_o = " << _nb_o << std::endl;
std::cout << GridLogMessage << " _blocks = " << _blocks << std::endl;
std::cout << GridLogMessage << " _o_blocks = " << _o_blocks << std::endl;
std::cout << GridLogMessage << " sizeof(vCoeff_t) = " << sizeof(vCoeff_t) << std::endl;
std::cout << GridLogMessage << " _cf_size = " << _cf_size << std::endl;
std::cout << GridLogMessage << " _cf_block_size = " << _cf_block_size << std::endl;
std::cout << GridLogMessage << " _block_sites = " << _block_sites << std::endl;
std::cout << GridLogMessage << " _grid->oSites() = " << _grid->oSites() << std::endl;
// _grid->Barrier();
//abort();
}
void block_to_coor(int b, std::vector<int>& x0) {
std::vector<int> bcoor;
bcoor.resize(_nd);
x0.resize(_nd);
assert(b < _o_blocks);
Lexicographic::CoorFromIndex(bcoor,b,_nb_o);
int i;
for (i=0;i<_nd;i++) {
x0[i] = bcoor[i]*_bs_cb[i];
}
//std::cout << GridLogMessage << "Map block b -> " << x0 << std::endl;
}
void block_site_to_o_coor(const std::vector<int>& x0, std::vector<int>& coor, int i) {
Lexicographic::CoorFromIndex(coor,i,_bs_cb);
for (int j=0;j<_nd;j++)
coor[j] += x0[j];
}
int block_site_to_o_site(const std::vector<int>& x0, int i) {
std::vector<int> coor; coor.resize(_nd);
block_site_to_o_coor(x0,coor,i);
Lexicographic::IndexFromCoor(coor,i,_l_cb_o);
return i;
}
vCoeff_t block_sp(int b, const Field& x, const Field& y) {
std::vector<int> x0;
block_to_coor(b,x0);
vCoeff_t ret = 0.0;
for (int i=0;i<_block_sites;i++) { // only odd sites
int ss = block_site_to_o_site(x0,i);
ret += TensorRemove(innerProduct(x._odata[ss],y._odata[ss]));
}
return ret;
}
vCoeff_t block_sp(int b, const Field& x, const std::vector< ComplexD >& y) {
std::vector<int> x0;
block_to_coor(b,x0);
constexpr int nsimd = sizeof(vCoeff_t) / sizeof(Coeff_t);
int lsize = _cf_o_block_size / _block_sites;
std::vector< ComplexD > ret(nsimd);
for (int i=0;i<nsimd;i++)
ret[i] = 0.0;
for (int i=0;i<_block_sites;i++) { // only odd sites
int ss = block_site_to_o_site(x0,i);
int n = lsize / nsimd;
for (int l=0;l<n;l++) {
for (int j=0;j<nsimd;j++) {
int t = lsize * i + l*nsimd + j;
ret[j] += conjugate(((Coeff_t*)&x._odata[ss]._internal)[l*nsimd + j]) * y[t];
}
}
}
vCoeff_t vret;
for (int i=0;i<nsimd;i++)
((Coeff_t*)&vret)[i] = (Coeff_t)ret[i];
return vret;
}
template<class T>
void vcaxpy(iScalar<T>& r,const vCoeff_t& a,const iScalar<T>& x,const iScalar<T>& y) {
vcaxpy(r._internal,a,x._internal,y._internal);
}
template<class T,int N>
void vcaxpy(iVector<T,N>& r,const vCoeff_t& a,const iVector<T,N>& x,const iVector<T,N>& y) {
for (int i=0;i<N;i++)
vcaxpy(r._internal[i],a,x._internal[i],y._internal[i]);
}
void vcaxpy(vCoeff_t& r,const vCoeff_t& a,const vCoeff_t& x,const vCoeff_t& y) {
r = a*x + y;
}
void block_caxpy(int b, Field& ret, const vCoeff_t& a, const Field& x, const Field& y) {
std::vector<int> x0;
block_to_coor(b,x0);
for (int i=0;i<_block_sites;i++) { // only odd sites
int ss = block_site_to_o_site(x0,i);
vcaxpy(ret._odata[ss],a,x._odata[ss],y._odata[ss]);
}
}
void block_caxpy(int b, std::vector< ComplexD >& ret, const vCoeff_t& a, const Field& x, const std::vector< ComplexD >& y) {
std::vector<int> x0;
block_to_coor(b,x0);
constexpr int nsimd = sizeof(vCoeff_t) / sizeof(Coeff_t);
int lsize = _cf_o_block_size / _block_sites;
for (int i=0;i<_block_sites;i++) { // only odd sites
int ss = block_site_to_o_site(x0,i);
int n = lsize / nsimd;
for (int l=0;l<n;l++) {
vCoeff_t r = a* ((vCoeff_t*)&x._odata[ss]._internal)[l];
for (int j=0;j<nsimd;j++) {
int t = lsize * i + l*nsimd + j;
ret[t] = y[t] + ((Coeff_t*)&r)[j];
}
}
}
}
void block_set(int b, Field& ret, const std::vector< ComplexD >& x) {
std::vector<int> x0;
block_to_coor(b,x0);
int lsize = _cf_o_block_size / _block_sites;
for (int i=0;i<_block_sites;i++) { // only odd sites
int ss = block_site_to_o_site(x0,i);
for (int l=0;l<lsize;l++)
((Coeff_t*)&ret._odata[ss]._internal)[l] = (Coeff_t)x[lsize * i + l]; // convert precision
}
}
void block_get(int b, const Field& ret, std::vector< ComplexD >& x) {
std::vector<int> x0;
block_to_coor(b,x0);
int lsize = _cf_o_block_size / _block_sites;
for (int i=0;i<_block_sites;i++) { // only odd sites
int ss = block_site_to_o_site(x0,i);
for (int l=0;l<lsize;l++)
x[lsize * i + l] = (ComplexD)((Coeff_t*)&ret._odata[ss]._internal)[l];
}
}
template<class T>
void vcscale(iScalar<T>& r,const vCoeff_t& a,const iScalar<T>& x) {
vcscale(r._internal,a,x._internal);
}
template<class T,int N>
void vcscale(iVector<T,N>& r,const vCoeff_t& a,const iVector<T,N>& x) {
for (int i=0;i<N;i++)
vcscale(r._internal[i],a,x._internal[i]);
}
void vcscale(vCoeff_t& r,const vCoeff_t& a,const vCoeff_t& x) {
r = a*x;
}
void block_cscale(int b, const vCoeff_t& a, Field& ret) {
std::vector<int> x0;
block_to_coor(b,x0);
for (int i=0;i<_block_sites;i++) { // only odd sites
int ss = block_site_to_o_site(x0,i);
vcscale(ret._odata[ss],a,ret._odata[ss]);
}
}
void getCanonicalBlockOffset(int cb, std::vector<int>& x0) {
const int ndim = 5;
assert(_nb.size() == ndim);
std::vector<int> _nbc = { _nb[1], _nb[2], _nb[3], _nb[4], _nb[0] };
std::vector<int> _bsc = { _bs[1], _bs[2], _bs[3], _bs[4], _bs[0] };
x0.resize(ndim);
assert(cb >= 0);
assert(cb < _nbc[0]*_nbc[1]*_nbc[2]*_nbc[3]*_nbc[4]);
Lexicographic::CoorFromIndex(x0,cb,_nbc);
int i;
for (i=0;i<ndim;i++) {
x0[i] *= _bsc[i];
}
//if (cb < 2)
// std::cout << GridLogMessage << "Map: " << cb << " To: " << x0 << std::endl;
}
void pokeBlockOfVectorCanonical(int cb,Field& v,const std::vector<float>& buf) {
std::vector<int> _bsc = { _bs[1], _bs[2], _bs[3], _bs[4], _bs[0] };
std::vector<int> ldim = v._grid->LocalDimensions();
std::vector<int> cldim = { ldim[1], ldim[2], ldim[3], ldim[4], ldim[0] };
const int _nbsc = _bs_cb[0]*_bs_cb[1]*_bs_cb[2]*_bs_cb[3]*_bs_cb[4];
// take canonical block cb of v and put it in canonical ordering in buf
std::vector<int> cx0;
getCanonicalBlockOffset(cb,cx0);
#pragma omp parallel
{
std::vector<int> co0,cl0;
co0=cx0; cl0=cx0;
#pragma omp for
for (int i=0;i<_nbsc;i++) {
Lexicographic::CoorFromIndex(co0,2*i,_bsc); // 2* for eo
for (int j=0;j<(int)_bsc.size();j++)
cl0[j] = cx0[j] + co0[j];
std::vector<int> l0 = { cl0[4], cl0[0], cl0[1], cl0[2], cl0[3] };
int oi = v._grid->oIndex(l0);
int ii = v._grid->iIndex(l0);
int lti = i;
//if (cb < 2 && i<2)
// std::cout << GridLogMessage << "Map: " << cb << ", " << i << " To: " << cl0 << ", " << cx0 << ", " << oi << ", " << ii << std::endl;
for (int s=0;s<4;s++)
for (int c=0;c<3;c++) {
Coeff_t& ld = ((Coeff_t*)&v._odata[oi]._internal._internal[s]._internal[c])[ii];
int ti = 12*lti + 3*s + c;
ld = Coeff_t(buf[2*ti+0], buf[2*ti+1]);
}
}
}
}
void peekBlockOfVectorCanonical(int cb,const Field& v,std::vector<float>& buf) {
std::vector<int> _bsc = { _bs[1], _bs[2], _bs[3], _bs[4], _bs[0] };
std::vector<int> ldim = v._grid->LocalDimensions();
std::vector<int> cldim = { ldim[1], ldim[2], ldim[3], ldim[4], ldim[0] };
const int _nbsc = _bs_cb[0]*_bs_cb[1]*_bs_cb[2]*_bs_cb[3]*_bs_cb[4];
// take canonical block cb of v and put it in canonical ordering in buf
std::vector<int> cx0;
getCanonicalBlockOffset(cb,cx0);
buf.resize(_cf_block_size * 2);
#pragma omp parallel
{
std::vector<int> co0,cl0;
co0=cx0; cl0=cx0;
#pragma omp for
for (int i=0;i<_nbsc;i++) {
Lexicographic::CoorFromIndex(co0,2*i,_bsc); // 2* for eo
for (int j=0;j<(int)_bsc.size();j++)
cl0[j] = cx0[j] + co0[j];
std::vector<int> l0 = { cl0[4], cl0[0], cl0[1], cl0[2], cl0[3] };
int oi = v._grid->oIndex(l0);
int ii = v._grid->iIndex(l0);
int lti = i;
//if (cb < 2 && i<2)
// std::cout << GridLogMessage << "Map: " << cb << ", " << i << " To: " << cl0 << ", " << cx0 << ", " << oi << ", " << ii << std::endl;
for (int s=0;s<4;s++)
for (int c=0;c<3;c++) {
Coeff_t& ld = ((Coeff_t*)&v._odata[oi]._internal._internal[s]._internal[c])[ii];
int ti = 12*lti + 3*s + c;
buf[2*ti+0] = ld.real();
buf[2*ti+1] = ld.imag();
}
}
}
}
int globalToLocalCanonicalBlock(int slot,const std::vector<int>& src_nodes,int nb) {
// processor coordinate
int _nd = (int)src_nodes.size();
std::vector<int> _src_nodes = src_nodes;
std::vector<int> pco(_nd);
Lexicographic::CoorFromIndex(pco,slot,_src_nodes);
std::vector<int> cpco = { pco[1], pco[2], pco[3], pco[4], pco[0] };
// get local block
std::vector<int> _nbc = { _nb[1], _nb[2], _nb[3], _nb[4], _nb[0] };
assert(_nd == 5);
std::vector<int> c_src_local_blocks(_nd);
for (int i=0;i<_nd;i++) {
assert(_grid->_fdimensions[i] % (src_nodes[i] * _bs[i]) == 0);
c_src_local_blocks[(i+4) % 5] = _grid->_fdimensions[i] / src_nodes[i] / _bs[i];
}
std::vector<int> cbcoor(_nd); // coordinate of block in slot in canonical form
Lexicographic::CoorFromIndex(cbcoor,nb,c_src_local_blocks);
// cpco, cbcoor
std::vector<int> clbcoor(_nd);
for (int i=0;i<_nd;i++) {
int cgcoor = cpco[i] * c_src_local_blocks[i] + cbcoor[i]; // global block coordinate
int pcoor = cgcoor / _nbc[i]; // processor coordinate in my Grid
int tpcoor = _grid->_processor_coor[(i+1)%5];
if (pcoor != tpcoor)
return -1;
clbcoor[i] = cgcoor - tpcoor * _nbc[i]; // canonical local block coordinate for canonical dimension i
}
int lnb;
Lexicographic::IndexFromCoor(clbcoor,lnb,_nbc);
//std::cout << "Mapped slot = " << slot << " nb = " << nb << " to " << lnb << std::endl;
return lnb;
}
};
}

163
lib/algorithms/iterative/BlockImplicitlyRestartedLanczos/FieldBasisVector.h
View File

@ -1,163 +0,0 @@
namespace Grid {
template<class Field>
class BasisFieldVector {
public:
int _Nm;
typedef typename Field::scalar_type Coeff_t;
typedef typename Field::vector_type vCoeff_t;
typedef typename Field::vector_object vobj;
typedef typename vobj::scalar_object sobj;
std::vector<Field> _v; // _Nfull vectors
void report(int n,GridBase* value) {
std::cout << GridLogMessage << "BasisFieldVector allocated:\n";
std::cout << GridLogMessage << " Delta N = " << n << "\n";
std::cout << GridLogMessage << " Size of full vectors (size) = " <<
((double)n*sizeof(vobj)*value->oSites() / 1024./1024./1024.) << " GB\n";
std::cout << GridLogMessage << " Size = " << _v.size() << " Capacity = " << _v.capacity() << std::endl;
value->Barrier();
if (value->IsBoss()) {
system("cat /proc/meminfo");
}
value->Barrier();
}
BasisFieldVector(int Nm,GridBase* value) : _Nm(Nm), _v(Nm,value) {
report(Nm,value);
}
~BasisFieldVector() {
}
Field& operator[](int i) {
return _v[i];
}
void orthogonalize(Field& w, int k) {
for(int j=0; j<k; ++j){
Coeff_t ip = (Coeff_t)innerProduct(_v[j],w);
w = w - ip*_v[j];
}
}
void rotate(std::vector<RealD>& Qt,int j0, int j1, int k0,int k1,int Nm) {
GridBase* grid = _v[0]._grid;
#pragma omp parallel
{
std::vector < vobj > B(Nm);
#pragma omp for
for(int ss=0;ss < grid->oSites();ss++){
for(int j=j0; j<j1; ++j) B[j]=0.;
for(int j=j0; j<j1; ++j){
for(int k=k0; k<k1; ++k){
B[j] +=Qt[k+Nm*j] * _v[k]._odata[ss];
}
}
for(int j=j0; j<j1; ++j){
_v[j]._odata[ss] = B[j];
}
}
}
}
size_t size() const {
return _Nm;
}
void resize(int n) {
if (n > _Nm)
_v.reserve(n);
_v.resize(n,_v[0]._grid);
if (n < _Nm)
_v.shrink_to_fit();
report(n - _Nm,_v[0]._grid);
_Nm = n;
}
std::vector<int> getIndex(std::vector<RealD>& sort_vals) {
std::vector<int> idx(sort_vals.size());
iota(idx.begin(), idx.end(), 0);
// sort indexes based on comparing values in v
sort(idx.begin(), idx.end(),
[&sort_vals](int i1, int i2) {return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);});
return idx;
}
void reorderInPlace(std::vector<RealD>& sort_vals, std::vector<int>& idx) {
GridStopWatch gsw;
gsw.Start();
int nswaps = 0;
for (size_t i=0;i<idx.size();i++) {
if (idx[i] != i) {
// find proper place (this could be done in logarithmic time, don't bother for now)
size_t j;
for (j=i;j<idx.size();j++)
if (idx[j]==i)
break;
assert(j!=idx.size());
Field _t(_v[0]._grid);
_t = _v[idx[j]];
_v[idx[j]] = _v[idx[i]];
_v[idx[i]] = _t;
RealD _td = sort_vals[idx[j]];
sort_vals[idx[j]] = sort_vals[idx[i]];
sort_vals[idx[i]] = _td;
int _tt = idx[i];
idx[i] = idx[j];
idx[j] = _tt;
nswaps++;
}
}
// sort values
gsw.Stop();
std::cout << GridLogMessage << "Sorted eigenspace in place in " << gsw.Elapsed() << " using " << nswaps << " swaps" << std::endl;
}
void sortInPlace(std::vector<RealD>& sort_vals, bool reverse) {
std::vector<int> idx = getIndex(sort_vals);
if (reverse)
std::reverse(idx.begin(), idx.end());
reorderInPlace(sort_vals,idx);
}
void deflate(const std::vector<RealD>& eval,const Field& src_orig,Field& result) {
result = zero;
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);
}
}
};
}

14
lib/algorithms/iterative/ConjugateGradient.h
View File

@ -78,12 +78,12 @@ class ConjugateGradient : public OperatorFunction<Field> {
cp = a;
ssq = norm2(src);
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: mp " << d << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: mmp " << b << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: cp,r " << cp << std::endl;
std::cout << GridLogIterative << std::setprecision(4) << "ConjugateGradient: p " << a << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: guess " << guess << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: src " << ssq << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: mp " << d << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: mmp " << b << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: cp,r " << cp << std::endl;
std::cout << GridLogIterative << std::setprecision(8) << "ConjugateGradient: p " << a << std::endl;
RealD rsq = Tolerance * Tolerance * ssq;
@ -92,7 +92,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
return;
}
std::cout << GridLogIterative << std::setprecision(4)
std::cout << GridLogIterative << std::setprecision(8)
<< "ConjugateGradient: k=0 residual " << cp << " target " << rsq << std::endl;
GridStopWatch LinalgTimer;

743
lib/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -7,8 +7,9 @@
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: Chulwoo Jung
Author: Guido Cossu
Author: paboyle <paboyle@ph.ed.ac.uk>
Author: Chulwoo Jung <chulwoo@bnl.gov>
Author: Christoph Lehner <clehner@bnl.gov>
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
@ -27,125 +28,288 @@ Author: Guido Cossu
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_IRL_H
#define GRID_IRL_H
#ifndef GRID_BIRL_H
#define GRID_BIRL_H
#include <string.h> //memset
//#include <zlib.h>
#include <sys/stat.h>
namespace Grid {
namespace Grid {
enum IRLdiagonalisation {
IRLdiagonaliseWithDSTEGR,
IRLdiagonaliseWithQR,
IRLdiagonaliseWithEigen
};
////////////////////////////////////////////////////////////////////////////////
// Helper class for sorting the evalues AND evectors by Field
// Use pointer swizzle on vectors
////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////
// Move following 100 LOC to lattice/Lattice_basis.h
////////////////////////////////////////////////////////
template<class Field>
class SortEigen {
private:
static bool less_lmd(RealD left,RealD right){
return left > right;
}
static bool less_pair(std::pair<RealD,Field const*>& left,
std::pair<RealD,Field const*>& right){
return left.first > (right.first);
}
public:
void push(std::vector<RealD>& lmd,std::vector<Field>& evec,int N) {
////////////////////////////////////////////////////////////////////////
// PAB: FIXME: VERY VERY VERY wasteful: takes a copy of the entire vector set.
// : The vector reorder should be done by pointer swizzle somehow
////////////////////////////////////////////////////////////////////////
std::vector<Field> cpy(lmd.size(),evec[0]._grid);
for(int i=0;i<lmd.size();i++) cpy[i] = evec[i];
std::vector<std::pair<RealD, Field const*> > emod(lmd.size());
void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k)
{
for(int j=0; j<k; ++j){
auto ip = innerProduct(basis[j],w);
w = w - ip*basis[j];
}
}
for(int i=0;i<lmd.size();++i) emod[i] = std::pair<RealD,Field const*>(lmd[i],&cpy[i]);
partial_sort(emod.begin(),emod.begin()+N,emod.end(),less_pair);
typename std::vector<std::pair<RealD, Field const*> >::iterator it = emod.begin();
for(int i=0;i<N;++i){
lmd[i]=it->first;
evec[i]=*(it->second);
++it;
template<class Field>
void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm)
{
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0]._grid;
parallel_region
{
std::vector < vobj > B(Nm); // Thread private
parallel_for_internal(int ss=0;ss < grid->oSites();ss++){
for(int j=j0; j<j1; ++j) B[j]=0.;
for(int j=j0; j<j1; ++j){
for(int k=k0; k<k1; ++k){
B[j] +=Qt(j,k) * basis[k]._odata[ss];
}
}
for(int j=j0; j<j1; ++j){
basis[j]._odata[ss] = B[j];
}
}
}
void push(std::vector<RealD>& lmd,int N) {
std::partial_sort(lmd.begin(),lmd.begin()+N,lmd.end(),less_lmd);
}
// Extract a single rotated vector
template<class Field>
void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm)
{
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0]._grid;
result.checkerboard = basis[0].checkerboard;
parallel_for(int ss=0;ss < grid->oSites();ss++){
vobj B = zero;
for(int k=k0; k<k1; ++k){
B +=Qt(j,k) * basis[k]._odata[ss];
}
result._odata[ss] = B;
}
bool saturated(RealD lmd, RealD thrs) {
return fabs(lmd) > fabs(thrs);
}
template<class Field>
void basisReorderInPlace(std::vector<Field> &_v,std::vector<RealD>& sort_vals, std::vector<int>& idx)
{
int vlen = idx.size();
assert(vlen>=1);
assert(vlen<=sort_vals.size());
assert(vlen<=_v.size());
for (size_t i=0;i<vlen;i++) {
if (idx[i] != i) {
//////////////////////////////////////
// idx[i] is a table of desired sources giving a permutation.
// Swap v[i] with v[idx[i]].
// Find j>i for which _vnew[j] = _vold[i],
// track the move idx[j] => idx[i]
// track the move idx[i] => i
//////////////////////////////////////
size_t j;
for (j=i;j<idx.size();j++)
if (idx[j]==i)
break;
assert(idx[i] > i); assert(j!=idx.size()); assert(idx[j]==i);
std::swap(_v[i]._odata,_v[idx[i]]._odata); // should use vector move constructor, no data copy
std::swap(sort_vals[i],sort_vals[idx[i]]);
idx[j] = idx[i];
idx[i] = i;
}
}
};
}
inline std::vector<int> basisSortGetIndex(std::vector<RealD>& sort_vals)
{
std::vector<int> idx(sort_vals.size());
std::iota(idx.begin(), idx.end(), 0);
// sort indexes based on comparing values in v
std::sort(idx.begin(), idx.end(), [&sort_vals](int i1, int i2) {
return ::fabs(sort_vals[i1]) < ::fabs(sort_vals[i2]);
});
return idx;
}
template<class Field>
void basisSortInPlace(std::vector<Field> & _v,std::vector<RealD>& sort_vals, bool reverse)
{
std::vector<int> idx = basisSortGetIndex(sort_vals);
if (reverse)
std::reverse(idx.begin(), idx.end());
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
/////////////////////////////////////////////////////////////
template<class Field> class ImplicitlyRestartedLanczosTester
{
public:
virtual int TestConvergence(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
virtual int ReconstructEval(int j,RealD resid,Field &evec, RealD &eval,RealD evalMaxApprox)=0;
};
enum IRLdiagonalisation {
IRLdiagonaliseWithDSTEGR,
IRLdiagonaliseWithQR,
IRLdiagonaliseWithEigen
};
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)
{
return TestConvergence(j,resid,B,eval,evalMaxApprox);
}
int TestConvergence(int j,RealD eresid,Field &B, RealD &eval,RealD evalMaxApprox)
{
Field v(B);
RealD eval_poly = eval;
// Apply operator
_HermOp(B,v);
RealD vnum = real(innerProduct(B,v)); // HermOp.
RealD vden = norm2(B);
RealD vv0 = norm2(v);
eval = vnum/vden;
v -= eval*B;
RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
std::cout.precision(13);
std::cout<<GridLogIRL << "[" << std::setw(3)<<j<<"] "
<<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
<<std::endl;
int conv=0;
if( (vv<eresid*eresid) ) conv = 1;
return conv;
}
};
template<class Field>
class ImplicitlyRestartedLanczos {
private:
int MaxIter; // Max iterations
int Nstop; // Number of evecs checked for convergence
int Nk; // Number of converged sought
int Nm; // Nm -- total number of vectors
RealD eresid;
private:
const RealD small = 1.0e-8;
int MaxIter;
int MinRestart; // Minimum number of restarts; only check for convergence after
int Nstop; // Number of evecs checked for convergence
int Nk; // Number of converged sought
// int Np; // Np -- Number of spare vecs in krylov space // == Nm - Nk
int Nm; // Nm -- total number of vectors
IRLdiagonalisation diagonalisation;
////////////////////////////////////
int orth_period;
RealD OrthoTime;
RealD eresid, betastp;
////////////////////////////////
// Embedded objects
////////////////////////////////////
SortEigen<Field> _sort;
LinearOperatorBase<Field> &_Linop;
OperatorFunction<Field> &_poly;
////////////////////////////////
LinearFunction<Field> &_PolyOp;
LinearFunction<Field> &_HermOp;
ImplicitlyRestartedLanczosTester<Field> &_Tester;
// Default tester provided (we need a ref to something in default case)
ImplicitlyRestartedLanczosHermOpTester<Field> SimpleTester;
/////////////////////////
// Constructor
/////////////////////////
public:
ImplicitlyRestartedLanczos(LinearOperatorBase<Field> &Linop, // op
OperatorFunction<Field> & poly, // polynomial
int _Nstop, // really sought vecs
int _Nk, // sought vecs
int _Nm, // total vecs
RealD _eresid, // resid in lmd deficit
int _MaxIter, // Max iterations
IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen ) :
_Linop(Linop), _poly(poly),
Nstop(_Nstop), Nk(_Nk), Nm(_Nm),
eresid(_eresid), MaxIter(_MaxIter),
diagonalisation(_diagonalisation)
{ };
//////////////////////////////////////////////////////////////////
// PAB:
//////////////////////////////////////////////////////////////////
// Too many options & knobs.
// Eliminate:
// orth_period
// betastp
// MinRestart
//
// Do we really need orth_period
// What is the theoretical basis & guarantees of betastp ?
// Nstop=Nk viable?
// MinRestart avoidable with new convergence test?
// Could cut to PolyOp, HermOp, Tester, Nk, Nm, resid, maxiter (+diagonalisation)
// HermOp could be eliminated if we dropped the Power method for max eval.
// -- also: The eval, eval2, eval2_copy stuff is still unnecessarily unclear
//////////////////////////////////////////////////////////////////
ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
LinearFunction<Field> & HermOp,
ImplicitlyRestartedLanczosTester<Field> & Tester,
int _Nstop, // sought vecs
int _Nk, // sought vecs
int _Nm, // spare vecs
RealD _eresid, // resid in lmdue deficit
int _MaxIter, // Max iterations
RealD _betastp=0.0, // if beta(k) < betastp: converged
int _MinRestart=1, int _orth_period = 1,
IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
SimpleTester(HermOp), _PolyOp(PolyOp), _HermOp(HermOp), _Tester(Tester),
Nstop(_Nstop) , Nk(_Nk), Nm(_Nm),
eresid(_eresid), betastp(_betastp),
MaxIter(_MaxIter) , MinRestart(_MinRestart),
orth_period(_orth_period), diagonalisation(_diagonalisation) { };
ImplicitlyRestartedLanczos(LinearFunction<Field> & PolyOp,
LinearFunction<Field> & HermOp,
int _Nstop, // sought vecs
int _Nk, // sought vecs
int _Nm, // spare vecs
RealD _eresid, // resid in lmdue deficit
int _MaxIter, // Max iterations
RealD _betastp=0.0, // if beta(k) < betastp: converged
int _MinRestart=1, int _orth_period = 1,
IRLdiagonalisation _diagonalisation= IRLdiagonaliseWithEigen) :
SimpleTester(HermOp), _PolyOp(PolyOp), _HermOp(HermOp), _Tester(SimpleTester),
Nstop(_Nstop) , Nk(_Nk), Nm(_Nm),
eresid(_eresid), betastp(_betastp),
MaxIter(_MaxIter) , MinRestart(_MinRestart),
orth_period(_orth_period), diagonalisation(_diagonalisation) { };
////////////////////////////////
// Helpers
////////////////////////////////
static RealD normalise(Field& v)
template<typename T> static RealD normalise(T& v)
{
RealD nn = norm2(v);
nn = sqrt(nn);
v = v * (1.0/nn);
return nn;
}
void orthogonalize(Field& w, std::vector<Field>& evec, int k)
void orthogonalize(Field& w, std::vector<Field>& evec,int k)
{
typedef typename Field::scalar_type MyComplex;
MyComplex ip;
for(int j=0; j<k; ++j){
ip = innerProduct(evec[j],w);
w = w - ip * evec[j];
}
OrthoTime-=usecond()/1e6;
basisOrthogonalize(evec,w,k);
normalise(w);
OrthoTime+=usecond()/1e6;
}
/* Rudy Arthur's thesis pp.137
@ -165,184 +329,238 @@ repeat
→AVK =VKHK +fKe†K † Extend to an M = K + P step factorization AVM = VMHM + fMeM
until convergence
*/
void calc(std::vector<RealD>& eval, std::vector<Field>& evec, const Field& src, int& Nconv)
void calc(std::vector<RealD>& eval, std::vector<Field>& evec, const Field& src, int& Nconv, bool reverse=false)
{
GridBase *grid = src._grid;
assert(grid == evec[0]._grid);
GridBase *grid = evec[0]._grid;
assert(grid == src._grid);
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout << GridLogMessage <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 / "<< MaxIter<< std::endl;
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout << GridLogMessage <<" -- seek Nk = " << Nk <<" vectors"<< std::endl;
std::cout << GridLogMessage <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
std::cout << GridLogMessage <<" -- total Nm = " << Nm <<" vectors"<< std::endl;
std::cout << GridLogMessage <<" -- size of eval = " << eval.size() << std::endl;
std::cout << GridLogMessage <<" -- size of evec = " << evec.size() << std::endl;
GridLogIRL.TimingMode(1);
std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
std::cout << GridLogIRL <<" ImplicitlyRestartedLanczos::calc() starting iteration 0 / "<< MaxIter<< std::endl;
std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
std::cout << GridLogIRL <<" -- seek Nk = " << Nk <<" vectors"<< std::endl;
std::cout << GridLogIRL <<" -- accept Nstop = " << Nstop <<" vectors"<< std::endl;
std::cout << GridLogIRL <<" -- total Nm = " << Nm <<" vectors"<< std::endl;
std::cout << GridLogIRL <<" -- size of eval = " << eval.size() << std::endl;
std::cout << GridLogIRL <<" -- size of evec = " << evec.size() << std::endl;
if ( diagonalisation == IRLdiagonaliseWithDSTEGR ) {
std::cout << GridLogMessage << "Diagonalisation is DSTEGR "<<std::endl;
std::cout << GridLogIRL << "Diagonalisation is DSTEGR "<<std::endl;
} else if ( diagonalisation == IRLdiagonaliseWithQR ) {
std::cout << GridLogMessage << "Diagonalisation is QR "<<std::endl;
std::cout << GridLogIRL << "Diagonalisation is QR "<<std::endl;
} else if ( diagonalisation == IRLdiagonaliseWithEigen ) {
std::cout << GridLogMessage << "Diagonalisation is Eigen "<<std::endl;
std::cout << GridLogIRL << "Diagonalisation is Eigen "<<std::endl;
}
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
assert(Nm <= evec.size() && Nm <= eval.size());
assert(Nm == evec.size() && Nm == eval.size());
// quickly get an idea of the largest eigenvalue to more properly normalize the residuum
RealD evalMaxApprox = 0.0;
{
auto src_n = src;
auto tmp = src;
const int _MAX_ITER_IRL_MEVAPP_ = 50;
for (int i=0;i<_MAX_ITER_IRL_MEVAPP_;i++) {
normalise(src_n);
_HermOp(src_n,tmp);
RealD vnum = real(innerProduct(src_n,tmp)); // HermOp.
RealD vden = norm2(src_n);
RealD na = vnum/vden;
if (fabs(evalMaxApprox/na - 1.0) < 0.05)
i=_MAX_ITER_IRL_MEVAPP_;
evalMaxApprox = na;
std::cout << GridLogIRL << " Approximation of largest eigenvalue: " << evalMaxApprox << std::endl;
src_n = tmp;
}
}
std::vector<RealD> lme(Nm);
std::vector<RealD> lme2(Nm);
std::vector<RealD> eval2(Nm);
std::vector<RealD> eval2_copy(Nm);
Eigen::MatrixXd Qt = Eigen::MatrixXd::Zero(Nm,Nm);
Eigen::MatrixXd Qt = Eigen::MatrixXd::Zero(Nm,Nm);
std::vector<int> Iconv(Nm);
std::vector<Field> B(Nm,grid); // waste of space replicating
Field f(grid);
Field v(grid);
int k1 = 1;
int k2 = Nk;
Nconv = 0;
RealD beta_k;
Nconv = 0;
// Set initial vector
evec[0] = src;
std::cout << GridLogMessage <<"norm2(src)= " << norm2(src)<<std::endl;
normalise(evec[0]);
std::cout << GridLogMessage <<"norm2(evec[0])= " << norm2(evec[0]) <<std::endl;
// Initial Nk steps
OrthoTime=0.;
for(int k=0; k<Nk; ++k) step(eval,lme,evec,f,Nm,k);
std::cout<<GridLogIRL <<"Initial "<< Nk <<"steps done "<<std::endl;
std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
//////////////////////////////////
// Restarting loop begins
//////////////////////////////////
int iter;
for(iter = 0; iter<MaxIter; ++iter){
OrthoTime=0.;
std::cout<< GridLogMessage <<" **********************"<< std::endl;
std::cout<< GridLogMessage <<" Restart iteration = "<< iter << std::endl;
std::cout<< GridLogMessage <<" **********************"<< std::endl;
std::cout<<GridLogIRL <<" running "<<Nm-Nk <<" steps: "<<std::endl;
for(int k=Nk; k<Nm; ++k) step(eval,lme,evec,f,Nm,k);
f *= lme[Nm-1];
std::cout<<GridLogIRL <<" "<<Nm-Nk <<" steps done "<<std::endl;
std::cout<<GridLogIRL <<"Initial steps:OrthoTime "<<OrthoTime<< "seconds"<<std::endl;
//////////////////////////////////
// getting eigenvalues
//////////////////////////////////
for(int k=0; k<Nm; ++k){
eval2[k] = eval[k+k1-1];
lme2[k] = lme[k+k1-1];
}
Qt = Eigen::MatrixXd::Identity(Nm,Nm);
diagonalize(eval2,lme2,Nm,Nm,Qt,grid);
std::cout<<GridLogIRL <<" diagonalized "<<std::endl;
//////////////////////////////////
// sorting
_sort.push(eval2,Nm);
//////////////////////////////////
eval2_copy = eval2;
std::partial_sort(eval2.begin(),eval2.begin()+Nm,eval2.end(),std::greater<RealD>());
std::cout<<GridLogIRL <<" evals sorted "<<std::endl;
const int chunk=8;
for(int io=0; io<k2;io+=chunk){
std::cout<<GridLogIRL << "eval "<< std::setw(3) << io ;
for(int ii=0;ii<chunk;ii++){
if ( (io+ii)<k2 )
std::cout<< " "<< std::setw(12)<< eval2[io+ii];
}
std::cout << std::endl;
}
//////////////////////////////////
// Implicitly shifted QR transformations
//////////////////////////////////
Qt = Eigen::MatrixXd::Identity(Nm,Nm);
for(int ip=k2; ip<Nm; ++ip){
// Eigen replacement for qr_decomp ???
qr_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
QR_decomp(eval,lme,Nm,Nm,Qt,eval2[ip],k1,Nm);
}
for(int i=0; i<(Nk+1); ++i) B[i] = 0.0;
for(int j=k1-1; j<k2+1; ++j){
for(int k=0; k<Nm; ++k){
B[j].checkerboard = evec[k].checkerboard;
B[j] += Qt(j,k) * evec[k];
}
}
for(int j=k1-1; j<k2+1; ++j) evec[j] = B[j];
std::cout<<GridLogIRL <<"QR decomposed "<<std::endl;
assert(k2<Nm); assert(k2<Nm); assert(k1>0);
basisRotate(evec,Qt,k1-1,k2+1,0,Nm,Nm); /// big constraint on the basis
std::cout<<GridLogIRL <<"basisRotated by Qt"<<std::endl;
////////////////////////////////////////////////////
// Compressed vector f and beta(k2)
////////////////////////////////////////////////////
f *= Qt(k2-1,Nm-1);
f += lme[k2-1] * evec[k2];
beta_k = norm2(f);
beta_k = sqrt(beta_k);
std::cout<< GridLogMessage<<" beta(k) = "<<beta_k<<std::endl;
std::cout<<GridLogIRL<<" beta(k) = "<<beta_k<<std::endl;
RealD betar = 1.0/beta_k;
evec[k2] = betar * f;
lme[k2-1] = beta_k;
////////////////////////////////////////////////////
// Convergence test
////////////////////////////////////////////////////
for(int k=0; k<Nm; ++k){
eval2[k] = eval[k];
lme2[k] = lme[k];
}
Qt = Eigen::MatrixXd::Identity(Nm,Nm);
diagonalize(eval2,lme2,Nk,Nm,Qt,grid);
for(int k = 0; k<Nk; ++k) B[k]=0.0;
for(int j = 0; j<Nk; ++j){
for(int k = 0; k<Nk; ++k){
B[j].checkerboard = evec[k].checkerboard;
B[j] += Qt(j,k) * evec[k];
}
}
std::cout<<GridLogIRL <<" Diagonalized "<<std::endl;
Nconv = 0;
for(int i=0; i<Nk; ++i){
_Linop.HermOp(B[i],v);
RealD vnum = real(innerProduct(B[i],v)); // HermOp.
RealD vden = norm2(B[i]);
eval2[i] = vnum/vden;
v -= eval2[i]*B[i];
RealD vv = norm2(v);
std::cout.precision(13);
std::cout << GridLogMessage << "[" << std::setw(3)<< std::setiosflags(std::ios_base::right) <<i<<"] ";
std::cout << "eval = "<<std::setw(25)<< std::setiosflags(std::ios_base::left)<< eval2[i];
std::cout << " |H B[i] - eval[i]B[i]|^2 "<< std::setw(25)<< std::setiosflags(std::ios_base::right)<< vv<< std::endl;
// change the criteria as evals are supposed to be sorted, all evals smaller(larger) than Nstop should have converged
if((vv<eresid*eresid) && (i == Nconv) ){
Iconv[Nconv] = i;
++Nconv;
}
} // i-loop end
std::cout<< GridLogMessage <<" #modes converged: "<<Nconv<<std::endl;
if (iter >= MinRestart) {
if( Nconv>=Nstop ){
goto converged;
}
} // end of iter loop
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout<< GridLogError <<" ImplicitlyRestartedLanczos::calc() NOT converged.";
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout << GridLogIRL << "Test convergence: rotate subset of vectors to test convergence " << std::endl;
Field B(grid); B.checkerboard = evec[0].checkerboard;
// power of two search pattern; not every evalue in eval2 is assessed.
for(int jj = 1; jj<=Nstop; jj*=2){
int j = Nstop-jj;
RealD e = eval2_copy[j]; // Discard the evalue
basisRotateJ(B,evec,Qt,j,0,Nk,Nm);
if( _Tester.TestConvergence(j,eresid,B,e,evalMaxApprox) ) {
if ( j > Nconv ) {
Nconv=j+1;
jj=Nstop; // Terminate the scan
}
}
}
// Do evec[0] for good measure
{
int j=0;
RealD e = eval2_copy[0];
basisRotateJ(B,evec,Qt,j,0,Nk,Nm);
_Tester.TestConvergence(j,eresid,B,e,evalMaxApprox);
}
// test if we converged, if so, terminate
std::cout<<GridLogIRL<<" #modes converged: >= "<<Nconv<<"/"<<Nstop<<std::endl;
// if( Nconv>=Nstop || beta_k < betastp){
if( Nconv>=Nstop){
goto converged;
}
} else {
std::cout << GridLogIRL << "iter < MinRestart: do not yet test for convergence\n";
} // end of iter loop
}
std::cout<<GridLogError<<"\n NOT converged.\n";
abort();
converged:
// Sorting
eval.resize(Nconv);
evec.resize(Nconv,grid);
for(int i=0; i<Nconv; ++i){
eval[i] = eval2[Iconv[i]];
evec[i] = B[Iconv[i]];
{
Field B(grid); B.checkerboard = evec[0].checkerboard;
basisRotate(evec,Qt,0,Nk,0,Nk,Nm);
std::cout << GridLogIRL << " Rotated basis"<<std::endl;
Nconv=0;
//////////////////////////////////////////////////////////////////////
// Full final convergence test; unconditionally applied
//////////////////////////////////////////////////////////////////////
for(int j = 0; j<=Nk; j++){
B=evec[j];
if( _Tester.ReconstructEval(j,eresid,B,eval2[j],evalMaxApprox) ) {
Nconv++;
}
}
if ( Nconv < Nstop )
std::cout << GridLogIRL << "Nconv ("<<Nconv<<") < Nstop ("<<Nstop<<")"<<std::endl;
eval=eval2;
//Keep only converged
eval.resize(Nconv);// Nstop?
evec.resize(Nconv,grid);// Nstop?
basisSortInPlace(evec,eval,reverse);
}
_sort.push(eval,evec,Nconv);
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout << GridLogMessage << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout << GridLogMessage << " -- Iterations = "<< iter << "\n";
std::cout << GridLogMessage << " -- beta(k) = "<< beta_k << "\n";
std::cout << GridLogMessage << " -- Nconv = "<< Nconv << "\n";
std::cout << GridLogMessage <<"**************************************************************************"<< std::endl;
std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
std::cout << GridLogIRL << "ImplicitlyRestartedLanczos CONVERGED ; Summary :\n";
std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
std::cout << GridLogIRL << " -- Iterations = "<< iter << "\n";
std::cout << GridLogIRL << " -- beta(k) = "<< beta_k << "\n";
std::cout << GridLogIRL << " -- Nconv = "<< Nconv << "\n";
std::cout << GridLogIRL <<"**************************************************************************"<< std::endl;
}
private:
private:
/* Saad PP. 195
1. Choose an initial vector v1 of 2-norm unity. Set β1 ≡ 0, v0 ≡ 0
2. For k = 1,2,...,m Do:
@ -360,28 +578,38 @@ private:
{
const RealD tiny = 1.0e-20;
assert( k< Nm );
_poly(_Linop,evec[k],w); // 3. wk:=Avk−βkv_{k−1}
GridStopWatch gsw_op,gsw_o;
Field& evec_k = evec[k];
_PolyOp(evec_k,w); std::cout<<GridLogIRL << "PolyOp" <<std::endl;
if(k>0) w -= lme[k-1] * evec[k-1];
ComplexD zalph = innerProduct(evec[k],w); // 4. αk:=(wk,vk)
ComplexD zalph = innerProduct(evec_k,w); // 4. αk:=(wk,vk)
RealD alph = real(zalph);
w = w - alph * evec[k];// 5. wk:=wk−αkvk
w = w - alph * evec_k;// 5. wk:=wk−αkvk
RealD beta = normalise(w); // 6. βk+1 := ∥wk∥2. If βk+1 = 0 then Stop
// 7. vk+1 := wk/βk+1
lmd[k] = alph;
lme[k] = beta;
if ( k > 0 ) orthogonalize(w,evec,k); // orthonormalise
if ( k < Nm-1) evec[k+1] = w;
if ( beta < tiny ) std::cout << GridLogMessage << " beta is tiny "<<beta<<std::endl;
if (k>0 && k % orth_period == 0) {
orthogonalize(w,evec,k); // orthonormalise
std::cout<<GridLogIRL << "Orthogonalised " <<std::endl;
}
if(k < Nm-1) evec[k+1] = w;
std::cout<<GridLogIRL << "alpha[" << k << "] = " << zalph << " beta[" << k << "] = "<<beta<<std::endl;
if ( beta < tiny )
std::cout<<GridLogIRL << " beta is tiny "<<beta<<std::endl;
}
void diagonalize_Eigen(std::vector<RealD>& lmd, std::vector<RealD>& lme,
int Nk, int Nm,
Eigen::MatrixXd & Qt, // Nm x Nm
@ -404,11 +632,11 @@ private:
}
}
}
///////////////////////////////////////////////////////////////////////////
// File could end here if settle on Eigen ???
///////////////////////////////////////////////////////////////////////////
void qr_decomp(std::vector<RealD>& lmd, // Nm
///////////////////////////////////////////////////////////////////////////
// File could end here if settle on Eigen ??? !!!
///////////////////////////////////////////////////////////////////////////
void QR_decomp(std::vector<RealD>& lmd, // Nm
std::vector<RealD>& lme, // Nm
int Nk, int Nm, // Nk, Nm
Eigen::MatrixXd& Qt, // Nm x Nm matrix
@ -575,51 +803,50 @@ void diagonalize_lapack(std::vector<RealD>& lmd,
#endif
}
void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
int Nk, int Nm,
Eigen::MatrixXd & Qt,
GridBase *grid)
{
int Niter = 100*Nm;
int kmin = 1;
int kmax = Nk;
// (this should be more sophisticated)
for(int iter=0; iter<Niter; ++iter){
// determination of 2x2 leading submatrix
RealD dsub = lmd[kmax-1]-lmd[kmax-2];
RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
// (Dsh: shift)
// transformation
qr_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
// Convergence criterion (redef of kmin and kamx)
for(int j=kmax-1; j>= kmin; --j){
RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
if(fabs(lme[j-1])+dds > dds){
kmax = j+1;
goto continued;
}
}
Niter = iter;
return;
continued:
for(int j=0; j<kmax-1; ++j){
RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
if(fabs(lme[j])+dds > dds){
kmin = j+1;
break;
}
void diagonalize_QR(std::vector<RealD>& lmd, std::vector<RealD>& lme,
int Nk, int Nm,
Eigen::MatrixXd & Qt,
GridBase *grid)
{
int QRiter = 100*Nm;
int kmin = 1;
int kmax = Nk;
// (this should be more sophisticated)
for(int iter=0; iter<QRiter; ++iter){
// determination of 2x2 leading submatrix
RealD dsub = lmd[kmax-1]-lmd[kmax-2];
RealD dd = sqrt(dsub*dsub + 4.0*lme[kmax-2]*lme[kmax-2]);
RealD Dsh = 0.5*(lmd[kmax-2]+lmd[kmax-1] +dd*(dsub/fabs(dsub)));
// (Dsh: shift)
// transformation
QR_decomp(lmd,lme,Nk,Nm,Qt,Dsh,kmin,kmax); // Nk, Nm
// Convergence criterion (redef of kmin and kamx)
for(int j=kmax-1; j>= kmin; --j){
RealD dds = fabs(lmd[j-1])+fabs(lmd[j]);
if(fabs(lme[j-1])+dds > dds){
kmax = j+1;
goto continued;
}
}
QRiter = iter;
return;
continued:
for(int j=0; j<kmax-1; ++j){
RealD dds = fabs(lmd[j])+fabs(lmd[j+1]);
if(fabs(lme[j])+dds > dds){
kmin = j+1;
break;
}
}
std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<Niter<<"\n";
abort();
}
};
std::cout << GridLogError << "[QL method] Error - Too many iteration: "<<QRiter<<"\n";
abort();
}
};
}
#endif

352
lib/algorithms/iterative/LocalCoherenceLanczos.h Normal file
View File

@ -0,0 +1,352 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/LocalCoherenceLanczos.h
Copyright (C) 2015
Author: Christoph Lehner <clehner@bnl.gov>
Author: paboyle <paboyle@ph.ed.ac.uk>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_LOCAL_COHERENCE_IRL_H
#define GRID_LOCAL_COHERENCE_IRL_H
namespace Grid {
struct LanczosParams : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(LanczosParams,
ChebyParams, Cheby,/*Chebyshev*/
int, Nstop, /*Vecs in Lanczos must converge Nstop < Nk < Nm*/
int, Nk, /*Vecs in Lanczos seek converge*/
int, Nm, /*Total vecs in Lanczos include restart*/
RealD, resid, /*residual*/
int, MaxIt,
RealD, betastp, /* ? */
int, MinRes); // Must restart
};
struct LocalCoherenceLanczosParams : Serializable {
public:
GRID_SERIALIZABLE_CLASS_MEMBERS(LocalCoherenceLanczosParams,
bool, doFine,
bool, doFineRead,
bool, doCoarse,
bool, doCoarseRead,
LanczosParams, FineParams,
LanczosParams, CoarseParams,
ChebyParams, Smoother,
RealD , coarse_relax_tol,
std::vector<int>, blockSize,
std::string, config,
std::vector < std::complex<double> >, omega,
RealD, mass,
RealD, M5);
};
// Duplicate functionality; ProjectedFunctionHermOp could be used with the trivial function
template<class Fobj,class CComplex,int nbasis>
class ProjectedHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj> FineField;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
ProjectedHermOp(LinearOperatorBase<FineField>& linop, Aggregation<Fobj,CComplex,nbasis> &aggregate) :
_Linop(linop),
_Aggregate(aggregate) { };
void operator()(const CoarseField& in, CoarseField& out) {
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;
}
};
template<class Fobj,class CComplex,int nbasis>
class ProjectedFunctionHermOp : public LinearFunction<Lattice<iVector<CComplex,nbasis > > > {
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj> FineField;
OperatorFunction<FineField> & _poly;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
ProjectedFunctionHermOp(OperatorFunction<FineField> & poly,LinearOperatorBase<FineField>& linop,
Aggregation<Fobj,CComplex,nbasis> &aggregate) :
_poly(poly),
_Linop(linop),
_Aggregate(aggregate) { };
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;
_poly(_Linop,fin,fout); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Poly "<<std::endl;
_Aggregate.ProjectToSubspace(out,fout); std::cout<<GridLogIRL<<"ProjectedFunctionHermop : Project to coarse "<<std::endl;
}
};
template<class Fobj,class CComplex,int nbasis>
class ImplicitlyRestartedLanczosSmoothedTester : public ImplicitlyRestartedLanczosTester<Lattice<iVector<CComplex,nbasis > > >
{
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj> FineField;
LinearFunction<CoarseField> & _Poly;
OperatorFunction<FineField> & _smoother;
LinearOperatorBase<FineField> &_Linop;
Aggregation<Fobj,CComplex,nbasis> &_Aggregate;
RealD _coarse_relax_tol;
ImplicitlyRestartedLanczosSmoothedTester(LinearFunction<CoarseField> &Poly,
OperatorFunction<FineField> &smoother,
LinearOperatorBase<FineField> &Linop,
Aggregation<Fobj,CComplex,nbasis> &Aggregate,
RealD coarse_relax_tol=5.0e3)
: _smoother(smoother), _Linop(Linop),_Aggregate(Aggregate), _Poly(Poly), _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);
RealD vnum = real(innerProduct(B,v)); // HermOp.
RealD vden = norm2(B);
RealD vv0 = norm2(v);
eval = vnum/vden;
v -= eval*B;
RealD vv = norm2(v) / ::pow(evalMaxApprox,2.0);
std::cout.precision(13);
std::cout<<GridLogIRL << "[" << std::setw(3)<<j<<"] "
<<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
<<std::endl;
int conv=0;
if( (vv<eresid*eresid) ) conv = 1;
return conv;
}
int ReconstructEval(int j,RealD eresid,CoarseField &B, RealD &eval,RealD evalMaxApprox)
{
GridBase *FineGrid = _Aggregate.FineGrid;
int checkerboard = _Aggregate.checkerboard;
FineField fB(FineGrid);fB.checkerboard =checkerboard;
FineField fv(FineGrid);fv.checkerboard =checkerboard;
_Aggregate.PromoteFromSubspace(B,fv);
_smoother(_Linop,fv,fB);
RealD eval_poly = eval;
_Linop.HermOp(fB,fv);
RealD vnum = real(innerProduct(fB,fv)); // HermOp.
RealD vden = norm2(fB);
RealD vv0 = norm2(fv);
eval = vnum/vden;
fv -= eval*fB;
RealD vv = norm2(fv) / ::pow(evalMaxApprox,2.0);
std::cout.precision(13);
std::cout<<GridLogIRL << "[" << std::setw(3)<<j<<"] "
<<"eval = "<<std::setw(25)<< eval << " (" << eval_poly << ")"
<<" |H B[i] - eval[i]B[i]|^2 / evalMaxApprox^2 " << std::setw(25) << vv
<<std::endl;
if ( j > nbasis ) eresid = eresid*_coarse_relax_tol;
if( (vv<eresid*eresid) ) return 1;
return 0;
}
};
////////////////////////////////////////////
// Make serializable Lanczos params
////////////////////////////////////////////
template<class Fobj,class CComplex,int nbasis>
class LocalCoherenceLanczos
{
public:
typedef iVector<CComplex,nbasis > CoarseSiteVector;
typedef Lattice<CComplex> CoarseScalar; // used for inner products on fine field
typedef Lattice<CoarseSiteVector> CoarseField;
typedef Lattice<Fobj> FineField;
protected:
GridBase *_CoarseGrid;
GridBase *_FineGrid;
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;
public:
LocalCoherenceLanczos(GridBase *FineGrid,
GridBase *CoarseGrid,
LinearOperatorBase<FineField> &FineOp,
int checkerboard) :
_CoarseGrid(CoarseGrid),
_FineGrid(FineGrid),
_Aggregate(CoarseGrid,FineGrid,checkerboard),
_FineOp(FineOp),
_checkerboard(checkerboard)
{
evals_fine.resize(0);
evals_coarse.resize(0);
};
void Orthogonalise(void ) { _Aggregate.Orthogonalise(); }
template<typename T> static RealD normalise(T& v)
{
RealD nn = norm2(v);
nn = ::sqrt(nn);
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]);
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]);
}
}
void testFine(RealD resid)
{
assert(evals_fine.size() == nbasis);
assert(_Aggregate.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);
}
}
void testCoarse(RealD resid,ChebyParams cheby_smooth,RealD relax)
{
assert(evals_fine.size() == nbasis);
assert(_Aggregate.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);
for(int k=0;k<evec_coarse.size();k++){
if ( k < nbasis ) {
assert(ChebySmoothTester.ReconstructEval(k,resid,evec_coarse[k],evals_coarse[k],1.0)==1);
} else {
assert(ChebySmoothTester.ReconstructEval(k,resid*relax,evec_coarse[k],evals_coarse[k],1.0)==1);
}
}
}
void calcFine(ChebyParams cheby_parms,int Nstop,int Nk,int Nm,RealD resid,
RealD MaxIt, RealD betastp, int MinRes)
{
assert(nbasis<=Nm);
Chebyshev<FineField> Cheby(cheby_parms);
FunctionHermOp<FineField> ChebyOp(Cheby,_FineOp);
PlainHermOp<FineField> Op(_FineOp);
evals_fine.resize(Nm);
_Aggregate.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);
// Shrink down to number saved
assert(Nstop>=nbasis);
assert(Nconv>=nbasis);
evals_fine.resize(nbasis);
_Aggregate.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);
//////////////////////////////////////////////////////////////////////////////////////////////////
// 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);
evals_coarse.resize(Nm);
evec_coarse.resize(Nm,_CoarseGrid);
CoarseField src(_CoarseGrid); src=1.0;
ImplicitlyRestartedLanczos<CoarseField> IRL(ChebyOp,ChebyOp,ChebySmoothTester,Nstop,Nk,Nm,resid,MaxIt,betastp,MinRes);
int Nconv=0;
IRL.calc(evals_coarse,evec_coarse,src,Nconv,false);
assert(Nconv>=Nstop);
evals_coarse.resize(Nstop);
evec_coarse.resize (Nstop,_CoarseGrid);
for (int i=0;i<Nstop;i++){
std::cout << i << " Coarse eval = " << evals_coarse[i] << std::endl;
}
}
};
}
#endif

26
lib/algorithms/iterative/SchurRedBlack.h
View File

@ -55,7 +55,15 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
*Odd
* i) D_oo psi_o = L^{-1} eta_o
* eta_o' = (D_oo)^dag (eta_o - Moe Mee^{-1} eta_e)
*
* Wilson:
* (D_oo)^{\dag} D_oo psi_o = (D_oo)^dag L^{-1} eta_o
* Stag:
* D_oo psi_o = L^{-1} eta = (eta_o - Moe Mee^{-1} eta_e)
*
* L^-1 eta_o= (1 0 ) (e
* (-MoeMee^{-1} 1 )
*
*Even
* ii) Mee psi_e + Meo psi_o = src_e
*
@ -82,7 +90,7 @@ namespace Grid {
// Take a matrix and form a Red Black solver calling a Herm solver
// Use of RB info prevents making SchurRedBlackSolve conform to standard interface
///////////////////////////////////////////////////////////////////////////////////////////////////////
// Now make the norm reflect extra factor of Mee
template<class Field> class SchurRedBlackStaggeredSolve {
private:
OperatorFunction<Field> & _HermitianRBSolver;
@ -115,26 +123,31 @@ namespace Grid {
Field tmp(grid);
Field Mtmp(grid);
Field resid(fgrid);
std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve " <<std::endl;
pickCheckerboard(Even,src_e,in);
pickCheckerboard(Odd ,src_o,in);
pickCheckerboard(Even,sol_e,out);
pickCheckerboard(Odd ,sol_o,out);
std::cout << GridLogMessage << " SchurRedBlackStaggeredSolve checkerboards picked" <<std::endl;
/////////////////////////////////////////////////////
// src_o = Mdag * (source_o - Moe MeeInv source_e)
// src_o = (source_o - Moe MeeInv source_e)
/////////////////////////////////////////////////////
_Matrix.MooeeInv(src_e,tmp); assert( tmp.checkerboard ==Even);
_Matrix.Meooe (tmp,Mtmp); assert( Mtmp.checkerboard ==Odd);
tmp=src_o-Mtmp; assert( tmp.checkerboard ==Odd);
_Matrix.Mooee(tmp,src_o); assert(src_o.checkerboard ==Odd);
//src_o = tmp; assert(src_o.checkerboard ==Odd);
_Matrix.Mooee(tmp,src_o); // Extra factor of "m" in source from dumb choice of matrix norm.
//////////////////////////////////////////////////////////////
// Call the red-black solver
//////////////////////////////////////////////////////////////
std::cout<<GridLogMessage << "SchurRedBlack solver calling the MpcDagMp solver" <<std::endl;
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver calling the Mpc solver" <<std::endl;
_HermitianRBSolver(_HermOpEO,src_o,sol_o); assert(sol_o.checkerboard==Odd);
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver called the Mpc solver" <<std::endl;
///////////////////////////////////////////////////
// sol_e = M_ee^-1 * ( src_e - Meo sol_o )...
@ -143,15 +156,16 @@ namespace Grid {
src_e = src_e-tmp; assert( src_e.checkerboard ==Even);
_Matrix.MooeeInv(src_e,sol_e); assert( sol_e.checkerboard ==Even);
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver reconstructed other CB" <<std::endl;
setCheckerboard(out,sol_e); assert( sol_e.checkerboard ==Even);
setCheckerboard(out,sol_o); assert( sol_o.checkerboard ==Odd );
std::cout<<GridLogMessage << "SchurRedBlackStaggeredSolver inserted solution" <<std::endl;
// Verify the unprec residual
_Matrix.M(out,resid);
resid = resid-in;
RealD ns = norm2(in);
RealD nr = norm2(resid);
std::cout<<GridLogMessage << "SchurRedBlackStaggered solver true unprec resid "<< std::sqrt(nr/ns) <<" nr "<< nr <<" ns "<<ns << std::endl;
}
};