portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-14 13:57:07 +01:00
Code Releases Activity

Merge branch 'feature/wilsonmg' of https://github.com/DanielRichtmann/Grid into DanielRichtmann-feature/wilsonmg

Browse Source
This commit is contained in:
Peter Boyle
2019-01-02 14:39:59 +00:00
parent fd66325321 2881b3e8e5
commit 8a1be021d3
32 changed files with 3828 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
Grid/algorithms/Algorithms.h
View File

@ -48,6 +48,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/algorithms/iterative/ConjugateGradientMixedPrec.h>
#include <Grid/algorithms/iterative/BlockConjugateGradient.h>
#include <Grid/algorithms/iterative/ConjugateGradientReliableUpdate.h>
#include <Grid/algorithms/iterative/MinimalResidual.h>
#include <Grid/algorithms/iterative/GeneralisedMinimalResidual.h>
#include <Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h>
#include <Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h>
#include <Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h>
#include <Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h>
#include <Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h>
#include <Grid/algorithms/CoarsenedMatrix.h>
#include <Grid/algorithms/FFT.h>

58
Grid/algorithms/CoarsenedMatrix.h
View File

@ -211,6 +211,7 @@ namespace Grid {
for(int b=0;b<nn;b++){
subspace[b] = zero;
gaussian(RNG,noise);
scale = std::pow(norm2(noise),-0.5);
noise=noise*scale;
@ -295,13 +296,58 @@ namespace Grid {
return norm2(out);
};
RealD Mdag (const CoarseVector &in, CoarseVector &out){
return M(in,out);
RealD Mdag (const CoarseVector &in, CoarseVector &out){
// // corresponds to Petrov-Galerkin coarsening
// return M(in,out);
// corresponds to Galerkin coarsening
CoarseVector tmp(Grid());
G5C(tmp, in);
M(tmp, out);
G5C(out, out);
return norm2(out);
};
// Defer support for further coarsening for now
void Mdiag (const CoarseVector &in, CoarseVector &out){};
void Mdir (const CoarseVector &in, CoarseVector &out,int dir, int disp){};
void Mdir(const CoarseVector &in, CoarseVector &out, int dir, int disp){
conformable(_grid,in._grid);
conformable(in._grid,out._grid);
SimpleCompressor<siteVector> compressor;
Stencil.HaloExchange(in,compressor);
auto point = [dir, disp](){
if(dir == 0 and disp == 0)
return 8;
else
return (4 * dir + 1 - disp) / 2;
}();
parallel_for(int ss=0;ss<Grid()->oSites();ss++){
siteVector res = zero;
siteVector nbr;
int ptype;
StencilEntry *SE;
SE=Stencil.GetEntry(ptype,point,ss);
if(SE->_is_local&&SE->_permute) {
permute(nbr,in._odata[SE->_offset],ptype);
} else if(SE->_is_local) {
nbr = in._odata[SE->_offset];
} else {
nbr = Stencil.CommBuf()[SE->_offset];
}
res = res + A[point]._odata[ss]*nbr;
vstream(out._odata[ss],res);
}
};
void Mdiag(const CoarseVector &in, CoarseVector &out){
Mdir(in, out, 0, 0); // use the self coupling (= last) point of the stencil
};
CoarsenedMatrix(GridCartesian &CoarseGrid) :
@ -417,7 +463,7 @@ namespace Grid {
std::cout<<GridLogMessage<<"Computed Coarse Operator"<<std::endl;
#endif
// ForceHermitian();
AssertHermitian();
// AssertHermitian();
// ForceDiagonal();
}
void ForceDiagonal(void) {

244
Grid/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h Normal file
View File

@ -0,0 +1,244 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/CommunicationAvoidingGeneralisedMinimalResidual.h
Copyright (C) 2015
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
#define GRID_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
namespace Grid {
template<class Field>
class CommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
bool ErrorOnNoConverge; // Throw an assert when CAGMRES fails to converge,
// defaults to true
RealD Tolerance;
Integer MaxIterations;
Integer RestartLength;
Integer MaxNumberOfRestarts;
Integer IterationCount; // Number of iterations the CAGMRES took to finish,
// filled in upon completion
GridStopWatch MatrixTimer;
GridStopWatch LinalgTimer;
GridStopWatch QrTimer;
GridStopWatch CompSolutionTimer;
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
CommunicationAvoidingGeneralisedMinimalResidual(RealD tol,
Integer maxit,
Integer restart_length,
bool err_on_no_conv = true)
: Tolerance(tol)
, MaxIterations(maxit)
, RestartLength(restart_length)
, MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
, ErrorOnNoConverge(err_on_no_conv)
, H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
, y(RestartLength + 1, 0.)
, gamma(RestartLength + 1, 0.)
, c(RestartLength + 1, 0.)
, s(RestartLength + 1, 0.) {};
void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular GMRES" << std::endl;
psi.checkerboard = src.checkerboard;
conformable(psi, src);
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
RealD cp;
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: src " << ssq << std::endl;
MatrixTimer.Reset();
LinalgTimer.Reset();
QrTimer.Reset();
CompSolutionTimer.Reset();
GridStopWatch SolverTimer;
SolverTimer.Start();
IterationCount = 0;
for (int k=0; k<MaxNumberOfRestarts; k++) {
cp = outerLoopBody(LinOp, src, psi, rsq);
// Stopping condition
if (cp <= rsq) {
SolverTimer.Stop();
LinOp.Op(psi,r);
axpy(r,-1.0,src,r);
RealD srcnorm = sqrt(ssq);
RealD resnorm = sqrt(norm2(r));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "CommunicationAvoidingGeneralisedMinimalResidual: Converged on iteration " << IterationCount
<< " computed residual " << sqrt(cp / ssq)
<< " true residual " << true_residual
<< " target " << Tolerance << std::endl;
std::cout << GridLogMessage << "CAGMRES Time elapsed: Total " << SolverTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "CAGMRES Time elapsed: Matrix " << MatrixTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "CAGMRES Time elapsed: Linalg " << LinalgTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "CAGMRES Time elapsed: QR " << QrTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "CAGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
return;
}
}
std::cout << GridLogMessage << "CommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
if (ErrorOnNoConverge)
assert(0);
}
RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
// this should probably be made a class member so that it is only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
MatrixTimer.Start();
LinOp.Op(psi, w);
MatrixTimer.Stop();
LinalgTimer.Start();
r = src - w;
gamma[0] = sqrt(norm2(r));
v[0] = (1. / gamma[0]) * r;
LinalgTimer.Stop();
for (int i=0; i<RestartLength; i++) {
IterationCount++;
arnoldiStep(LinOp, v, w, i);
qrUpdate(i);
cp = std::norm(gamma[i+1]);
std::cout << GridLogIterative << "CommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
computeSolution(v, psi, i);
return cp;
}
}
assert(0); // Never reached
return cp;
}
void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, Field &w, int iter) {
MatrixTimer.Start();
LinOp.Op(v[iter], w);
MatrixTimer.Stop();
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
void qrUpdate(int iter) {
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
// Apply new Givens rotation
H(iter, iter) = nu;
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
QrTimer.Stop();
}
void computeSolution(std::vector<Field> const &v, Field &psi, int iter) {
CompSolutionTimer.Start();
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
}
for (int i = 0; i <= iter; i++)
psi = psi + v[i] * y[i];
CompSolutionTimer.Stop();
}
};
}
#endif

256
Grid/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h Normal file
View File

@ -0,0 +1,256 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/FlexibleCommunicationAvoidingGeneralisedMinimalResidual.h
Copyright (C) 2015
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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_FLEXIBLE_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
#define GRID_FLEXIBLE_COMMUNICATION_AVOIDING_GENERALISED_MINIMAL_RESIDUAL_H
namespace Grid {
template<class Field>
class FlexibleCommunicationAvoidingGeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
bool ErrorOnNoConverge; // Throw an assert when FCAGMRES fails to converge,
// defaults to true
RealD Tolerance;
Integer MaxIterations;
Integer RestartLength;
Integer MaxNumberOfRestarts;
Integer IterationCount; // Number of iterations the FCAGMRES took to finish,
// filled in upon completion
GridStopWatch MatrixTimer;
GridStopWatch PrecTimer;
GridStopWatch LinalgTimer;
GridStopWatch QrTimer;
GridStopWatch CompSolutionTimer;
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
LinearFunction<Field> &Preconditioner;
FlexibleCommunicationAvoidingGeneralisedMinimalResidual(RealD tol,
Integer maxit,
LinearFunction<Field> &Prec,
Integer restart_length,
bool err_on_no_conv = true)
: Tolerance(tol)
, MaxIterations(maxit)
, RestartLength(restart_length)
, MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
, ErrorOnNoConverge(err_on_no_conv)
, H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
, y(RestartLength + 1, 0.)
, gamma(RestartLength + 1, 0.)
, c(RestartLength + 1, 0.)
, s(RestartLength + 1, 0.)
, Preconditioner(Prec) {};
void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
std::cout << GridLogWarning << "This algorithm currently doesn't differ from regular FGMRES" << std::endl;
psi.checkerboard = src.checkerboard;
conformable(psi, src);
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
RealD cp;
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: guess " << guess << std::endl;
std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: src " << ssq << std::endl;
PrecTimer.Reset();
MatrixTimer.Reset();
LinalgTimer.Reset();
QrTimer.Reset();
CompSolutionTimer.Reset();
GridStopWatch SolverTimer;
SolverTimer.Start();
IterationCount = 0;
for (int k=0; k<MaxNumberOfRestarts; k++) {
cp = outerLoopBody(LinOp, src, psi, rsq);
// Stopping condition
if (cp <= rsq) {
SolverTimer.Stop();
LinOp.Op(psi,r);
axpy(r,-1.0,src,r);
RealD srcnorm = sqrt(ssq);
RealD resnorm = sqrt(norm2(r));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: Converged on iteration " << IterationCount
<< " computed residual " << sqrt(cp / ssq)
<< " true residual " << true_residual
<< " target " << Tolerance << std::endl;
std::cout << GridLogMessage << "FCAGMRES Time elapsed: Total " << SolverTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FCAGMRES Time elapsed: Precon " << PrecTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FCAGMRES Time elapsed: Matrix " << MatrixTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FCAGMRES Time elapsed: Linalg " << LinalgTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FCAGMRES Time elapsed: QR " << QrTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FCAGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
return;
}
}
std::cout << GridLogMessage << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual did NOT converge" << std::endl;
if (ErrorOnNoConverge)
assert(0);
}
RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
// these should probably be made class members so that they are only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<Field> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
MatrixTimer.Start();
LinOp.Op(psi, w);
MatrixTimer.Stop();
LinalgTimer.Start();
r = src - w;
gamma[0] = sqrt(norm2(r));
v[0] = (1. / gamma[0]) * r;
LinalgTimer.Stop();
for (int i=0; i<RestartLength; i++) {
IterationCount++;
arnoldiStep(LinOp, v, z, w, i);
qrUpdate(i);
cp = std::norm(gamma[i+1]);
std::cout << GridLogIterative << "FlexibleCommunicationAvoidingGeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
computeSolution(z, psi, i);
return cp;
}
}
assert(0); // Never reached
return cp;
}
void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, std::vector<Field> &z, Field &w, int iter) {
PrecTimer.Start();
Preconditioner(v[iter], z[iter]);
PrecTimer.Stop();
MatrixTimer.Start();
LinOp.Op(z[iter], w);
MatrixTimer.Stop();
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
void qrUpdate(int iter) {
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
// Apply new Givens rotation
H(iter, iter) = nu;
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
QrTimer.Stop();
}
void computeSolution(std::vector<Field> const &z, Field &psi, int iter) {
CompSolutionTimer.Start();
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
}
for (int i = 0; i <= iter; i++)
psi = psi + z[i] * y[i];
CompSolutionTimer.Stop();
}
};
}
#endif

254
Grid/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h Normal file
View File

@ -0,0 +1,254 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/FlexibleGeneralisedMinimalResidual.h
Copyright (C) 2015
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
#define GRID_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
namespace Grid {
template<class Field>
class FlexibleGeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
bool ErrorOnNoConverge; // Throw an assert when FGMRES fails to converge,
// defaults to true
RealD Tolerance;
Integer MaxIterations;
Integer RestartLength;
Integer MaxNumberOfRestarts;
Integer IterationCount; // Number of iterations the FGMRES took to finish,
// filled in upon completion
GridStopWatch MatrixTimer;
GridStopWatch PrecTimer;
GridStopWatch LinalgTimer;
GridStopWatch QrTimer;
GridStopWatch CompSolutionTimer;
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
LinearFunction<Field> &Preconditioner;
FlexibleGeneralisedMinimalResidual(RealD tol,
Integer maxit,
LinearFunction<Field> &Prec,
Integer restart_length,
bool err_on_no_conv = true)
: Tolerance(tol)
, MaxIterations(maxit)
, RestartLength(restart_length)
, MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
, ErrorOnNoConverge(err_on_no_conv)
, H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
, y(RestartLength + 1, 0.)
, gamma(RestartLength + 1, 0.)
, c(RestartLength + 1, 0.)
, s(RestartLength + 1, 0.)
, Preconditioner(Prec) {};
void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
psi.checkerboard = src.checkerboard;
conformable(psi, src);
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
RealD cp;
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: guess " << guess << std::endl;
std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: src " << ssq << std::endl;
PrecTimer.Reset();
MatrixTimer.Reset();
LinalgTimer.Reset();
QrTimer.Reset();
CompSolutionTimer.Reset();
GridStopWatch SolverTimer;
SolverTimer.Start();
IterationCount = 0;
for (int k=0; k<MaxNumberOfRestarts; k++) {
cp = outerLoopBody(LinOp, src, psi, rsq);
// Stopping condition
if (cp <= rsq) {
SolverTimer.Stop();
LinOp.Op(psi,r);
axpy(r,-1.0,src,r);
RealD srcnorm = sqrt(ssq);
RealD resnorm = sqrt(norm2(r));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "FlexibleGeneralisedMinimalResidual: Converged on iteration " << IterationCount
<< " computed residual " << sqrt(cp / ssq)
<< " true residual " << true_residual
<< " target " << Tolerance << std::endl;
std::cout << GridLogMessage << "FGMRES Time elapsed: Total " << SolverTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FGMRES Time elapsed: Precon " << PrecTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FGMRES Time elapsed: Matrix " << MatrixTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FGMRES Time elapsed: Linalg " << LinalgTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FGMRES Time elapsed: QR " << QrTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "FGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
return;
}
}
std::cout << GridLogMessage << "FlexibleGeneralisedMinimalResidual did NOT converge" << std::endl;
if (ErrorOnNoConverge)
assert(0);
}
RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
// these should probably be made class members so that they are only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<Field> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
MatrixTimer.Start();
LinOp.Op(psi, w);
MatrixTimer.Stop();
LinalgTimer.Start();
r = src - w;
gamma[0] = sqrt(norm2(r));
v[0] = (1. / gamma[0]) * r;
LinalgTimer.Stop();
for (int i=0; i<RestartLength; i++) {
IterationCount++;
arnoldiStep(LinOp, v, z, w, i);
qrUpdate(i);
cp = std::norm(gamma[i+1]);
std::cout << GridLogIterative << "FlexibleGeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
computeSolution(z, psi, i);
return cp;
}
}
assert(0); // Never reached
return cp;
}
void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, std::vector<Field> &z, Field &w, int iter) {
PrecTimer.Start();
Preconditioner(v[iter], z[iter]);
PrecTimer.Stop();
MatrixTimer.Start();
LinOp.Op(z[iter], w);
MatrixTimer.Stop();
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
void qrUpdate(int iter) {
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
// Apply new Givens rotation
H(iter, iter) = nu;
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
QrTimer.Stop();
}
void computeSolution(std::vector<Field> const &z, Field &psi, int iter) {
CompSolutionTimer.Start();
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
}
for (int i = 0; i <= iter; i++)
psi = psi + z[i] * y[i];
CompSolutionTimer.Stop();
}
};
}
#endif

242
Grid/algorithms/iterative/GeneralisedMinimalResidual.h Normal file
View File

@ -0,0 +1,242 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/GeneralisedMinimalResidual.h
Copyright (C) 2015
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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_GENERALISED_MINIMAL_RESIDUAL_H
#define GRID_GENERALISED_MINIMAL_RESIDUAL_H
namespace Grid {
template<class Field>
class GeneralisedMinimalResidual : public OperatorFunction<Field> {
public:
bool ErrorOnNoConverge; // Throw an assert when GMRES fails to converge,
// defaults to true
RealD Tolerance;
Integer MaxIterations;
Integer RestartLength;
Integer MaxNumberOfRestarts;
Integer IterationCount; // Number of iterations the GMRES took to finish,
// filled in upon completion
GridStopWatch MatrixTimer;
GridStopWatch LinalgTimer;
GridStopWatch QrTimer;
GridStopWatch CompSolutionTimer;
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
GeneralisedMinimalResidual(RealD tol,
Integer maxit,
Integer restart_length,
bool err_on_no_conv = true)
: Tolerance(tol)
, MaxIterations(maxit)
, RestartLength(restart_length)
, MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
, ErrorOnNoConverge(err_on_no_conv)
, H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
, y(RestartLength + 1, 0.)
, gamma(RestartLength + 1, 0.)
, c(RestartLength + 1, 0.)
, s(RestartLength + 1, 0.) {};
void operator()(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi) {
psi.checkerboard = src.checkerboard;
conformable(psi, src);
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
RealD cp;
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Field r(src._grid);
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "GeneralisedMinimalResidual: guess " << guess << std::endl;
std::cout << GridLogIterative << "GeneralisedMinimalResidual: src " << ssq << std::endl;
MatrixTimer.Reset();
LinalgTimer.Reset();
QrTimer.Reset();
CompSolutionTimer.Reset();
GridStopWatch SolverTimer;
SolverTimer.Start();
IterationCount = 0;
for (int k=0; k<MaxNumberOfRestarts; k++) {
cp = outerLoopBody(LinOp, src, psi, rsq);
// Stopping condition
if (cp <= rsq) {
SolverTimer.Stop();
LinOp.Op(psi,r);
axpy(r,-1.0,src,r);
RealD srcnorm = sqrt(ssq);
RealD resnorm = sqrt(norm2(r));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "GeneralisedMinimalResidual: Converged on iteration " << IterationCount
<< " computed residual " << sqrt(cp / ssq)
<< " true residual " << true_residual
<< " target " << Tolerance << std::endl;
std::cout << GridLogMessage << "GMRES Time elapsed: Total " << SolverTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "GMRES Time elapsed: Matrix " << MatrixTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "GMRES Time elapsed: Linalg " << LinalgTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "GMRES Time elapsed: QR " << QrTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "GMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
return;
}
}
std::cout << GridLogMessage << "GeneralisedMinimalResidual did NOT converge" << std::endl;
if (ErrorOnNoConverge)
assert(0);
}
RealD outerLoopBody(LinearOperatorBase<Field> &LinOp, const Field &src, Field &psi, RealD rsq) {
RealD cp = 0;
Field w(src._grid);
Field r(src._grid);
// this should probably be made a class member so that it is only allocated once, not in every restart
std::vector<Field> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
MatrixTimer.Start();
LinOp.Op(psi, w);
MatrixTimer.Stop();
LinalgTimer.Start();
r = src - w;
gamma[0] = sqrt(norm2(r));
v[0] = (1. / gamma[0]) * r;
LinalgTimer.Stop();
for (int i=0; i<RestartLength; i++) {
IterationCount++;
arnoldiStep(LinOp, v, w, i);
qrUpdate(i);
cp = std::norm(gamma[i+1]);
std::cout << GridLogIterative << "GeneralisedMinimalResidual: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
computeSolution(v, psi, i);
return cp;
}
}
assert(0); // Never reached
return cp;
}
void arnoldiStep(LinearOperatorBase<Field> &LinOp, std::vector<Field> &v, Field &w, int iter) {
MatrixTimer.Start();
LinOp.Op(v[iter], w);
MatrixTimer.Stop();
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
void qrUpdate(int iter) {
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
// Apply new Givens rotation
H(iter, iter) = nu;
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
QrTimer.Stop();
}
void computeSolution(std::vector<Field> const &v, Field &psi, int iter) {
CompSolutionTimer.Start();
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
}
for (int i = 0; i <= iter; i++)
psi = psi + v[i] * y[i];
CompSolutionTimer.Stop();
}
};
}
#endif

156
Grid/algorithms/iterative/MinimalResidual.h Normal file
View File

@ -0,0 +1,156 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/MinimalResidual.h
Copyright (C) 2015
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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_MINIMAL_RESIDUAL_H
#define GRID_MINIMAL_RESIDUAL_H
namespace Grid {
template<class Field> class MinimalResidual : public OperatorFunction<Field> {
public:
bool ErrorOnNoConverge; // throw an assert when the MR fails to converge.
// Defaults true.
RealD Tolerance;
Integer MaxIterations;
RealD overRelaxParam;
Integer IterationsToComplete; // Number of iterations the MR took to finish.
// Filled in upon completion
MinimalResidual(RealD tol, Integer maxit, Real ovrelparam = 1.0, bool err_on_no_conv = true)
: Tolerance(tol), MaxIterations(maxit), overRelaxParam(ovrelparam), ErrorOnNoConverge(err_on_no_conv){};
void operator()(LinearOperatorBase<Field> &Linop, const Field &src, Field &psi) {
psi.checkerboard = src.checkerboard;
conformable(psi, src);
Complex a, c;
Real d;
Field Mr(src);
Field r(src);
// Initial residual computation & set up
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
Linop.Op(psi, Mr);
r = src - Mr;
RealD cp = norm2(r);
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "MinimalResidual: guess " << guess << std::endl;
std::cout << GridLogIterative << "MinimalResidual: src " << ssq << std::endl;
std::cout << GridLogIterative << "MinimalResidual: mp " << d << std::endl;
std::cout << GridLogIterative << "MinimalResidual: cp,r " << cp << std::endl;
if (cp <= rsq) {
return;
}
std::cout << GridLogIterative << "MinimalResidual: k=0 residual " << cp << " target " << rsq << std::endl;
GridStopWatch LinalgTimer;
GridStopWatch MatrixTimer;
GridStopWatch SolverTimer;
SolverTimer.Start();
int k;
for (k = 1; k <= MaxIterations; k++) {
MatrixTimer.Start();
Linop.Op(r, Mr);
MatrixTimer.Stop();
LinalgTimer.Start();
c = innerProduct(Mr, r);
d = norm2(Mr);
a = c / d;
a = a * overRelaxParam;
psi = psi + r * a;
r = r - Mr * a;
cp = norm2(r);
LinalgTimer.Stop();
std::cout << GridLogIterative << "MinimalResidual: Iteration " << k
<< " residual " << cp << " target " << rsq << std::endl;
std::cout << GridLogDebug << "a = " << a << " c = " << c << " d = " << d << std::endl;
// Stopping condition
if (cp <= rsq) {
SolverTimer.Stop();
Linop.Op(psi, Mr);
r = src - Mr;
RealD srcnorm = sqrt(ssq);
RealD resnorm = sqrt(norm2(r));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "MinimalResidual Converged on iteration " << k
<< " computed residual " << sqrt(cp / ssq)
<< " true residual " << true_residual
<< " target " << Tolerance << std::endl;
std::cout << GridLogMessage << "MR Time elapsed: Total " << SolverTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MR Time elapsed: Matrix " << MatrixTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MR Time elapsed: Linalg " << LinalgTimer.Elapsed() << std::endl;
if (ErrorOnNoConverge)
assert(true_residual / Tolerance < 10000.0);
IterationsToComplete = k;
return;
}
}
std::cout << GridLogMessage << "MinimalResidual did NOT converge"
<< std::endl;
if (ErrorOnNoConverge)
assert(0);
IterationsToComplete = k;
}
};
} // namespace Grid
#endif

273
Grid/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h Normal file
View File

@ -0,0 +1,273 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
Source file: ./lib/algorithms/iterative/MixedPrecisionFlexibleGeneralisedMinimalResidual.h
Copyright (C) 2015
Author: Daniel Richtmann <daniel.richtmann@ur.de>
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_MIXED_PRECISION_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
#define GRID_MIXED_PRECISION_FLEXIBLE_GENERALISED_MINIMAL_RESIDUAL_H
namespace Grid {
template<class FieldD, class FieldF, typename std::enable_if<getPrecision<FieldD>::value == 2, int>::type = 0, typename std::enable_if< getPrecision<FieldF>::value == 1, int>::type = 0>
class MixedPrecisionFlexibleGeneralisedMinimalResidual : public OperatorFunction<FieldD> {
public:
bool ErrorOnNoConverge; // Throw an assert when MPFGMRES fails to converge,
// defaults to true
RealD Tolerance;
Integer MaxIterations;
Integer RestartLength;
Integer MaxNumberOfRestarts;
Integer IterationCount; // Number of iterations the MPFGMRES took to finish,
// filled in upon completion
GridStopWatch MatrixTimer;
GridStopWatch PrecTimer;
GridStopWatch LinalgTimer;
GridStopWatch QrTimer;
GridStopWatch CompSolutionTimer;
GridStopWatch ChangePrecTimer;
Eigen::MatrixXcd H;
std::vector<std::complex<double>> y;
std::vector<std::complex<double>> gamma;
std::vector<std::complex<double>> c;
std::vector<std::complex<double>> s;
GridBase* SinglePrecGrid;
LinearFunction<FieldF> &Preconditioner;
MixedPrecisionFlexibleGeneralisedMinimalResidual(RealD tol,
Integer maxit,
GridBase * sp_grid,
LinearFunction<FieldF> &Prec,
Integer restart_length,
bool err_on_no_conv = true)
: Tolerance(tol)
, MaxIterations(maxit)
, RestartLength(restart_length)
, MaxNumberOfRestarts(MaxIterations/RestartLength + ((MaxIterations%RestartLength == 0) ? 0 : 1))
, ErrorOnNoConverge(err_on_no_conv)
, H(Eigen::MatrixXcd::Zero(RestartLength, RestartLength + 1)) // sizes taken from DD-αAMG code base
, y(RestartLength + 1, 0.)
, gamma(RestartLength + 1, 0.)
, c(RestartLength + 1, 0.)
, s(RestartLength + 1, 0.)
, SinglePrecGrid(sp_grid)
, Preconditioner(Prec) {};
void operator()(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi) {
psi.checkerboard = src.checkerboard;
conformable(psi, src);
RealD guess = norm2(psi);
assert(std::isnan(guess) == 0);
RealD cp;
RealD ssq = norm2(src);
RealD rsq = Tolerance * Tolerance * ssq;
FieldD r(src._grid);
std::cout << std::setprecision(4) << std::scientific;
std::cout << GridLogIterative << "MPFGMRES: guess " << guess << std::endl;
std::cout << GridLogIterative << "MPFGMRES: src " << ssq << std::endl;
PrecTimer.Reset();
MatrixTimer.Reset();
LinalgTimer.Reset();
QrTimer.Reset();
CompSolutionTimer.Reset();
ChangePrecTimer.Reset();
GridStopWatch SolverTimer;
SolverTimer.Start();
IterationCount = 0;
for (int k=0; k<MaxNumberOfRestarts; k++) {
cp = outerLoopBody(LinOp, src, psi, rsq);
// Stopping condition
if (cp <= rsq) {
SolverTimer.Stop();
LinOp.Op(psi,r);
axpy(r,-1.0,src,r);
RealD srcnorm = sqrt(ssq);
RealD resnorm = sqrt(norm2(r));
RealD true_residual = resnorm / srcnorm;
std::cout << GridLogMessage << "MPFGMRES: Converged on iteration " << IterationCount
<< " computed residual " << sqrt(cp / ssq)
<< " true residual " << true_residual
<< " target " << Tolerance << std::endl;
std::cout << GridLogMessage << "MPFGMRES Time elapsed: Total " << SolverTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MPFGMRES Time elapsed: Precon " << PrecTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MPFGMRES Time elapsed: Matrix " << MatrixTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MPFGMRES Time elapsed: Linalg " << LinalgTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MPFGMRES Time elapsed: QR " << QrTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MPFGMRES Time elapsed: CompSol " << CompSolutionTimer.Elapsed() << std::endl;
std::cout << GridLogMessage << "MPFGMRES Time elapsed: PrecChange " << ChangePrecTimer.Elapsed() << std::endl;
return;
}
}
std::cout << GridLogMessage << "MPFGMRES did NOT converge" << std::endl;
if (ErrorOnNoConverge)
assert(0);
}
RealD outerLoopBody(LinearOperatorBase<FieldD> &LinOp, const FieldD &src, FieldD &psi, RealD rsq) {
RealD cp = 0;
FieldD w(src._grid);
FieldD r(src._grid);
// these should probably be made class members so that they are only allocated once, not in every restart
std::vector<FieldD> v(RestartLength + 1, src._grid); for (auto &elem : v) elem = zero;
std::vector<FieldD> z(RestartLength + 1, src._grid); for (auto &elem : z) elem = zero;
MatrixTimer.Start();
LinOp.Op(psi, w);
MatrixTimer.Stop();
LinalgTimer.Start();
r = src - w;
gamma[0] = sqrt(norm2(r));
v[0] = (1. / gamma[0]) * r;
LinalgTimer.Stop();
for (int i=0; i<RestartLength; i++) {
IterationCount++;
arnoldiStep(LinOp, v, z, w, i);
qrUpdate(i);
cp = std::norm(gamma[i+1]);
std::cout << GridLogIterative << "MPFGMRES: Iteration " << IterationCount
<< " residual " << cp << " target " << rsq << std::endl;
if ((i == RestartLength - 1) || (IterationCount == MaxIterations) || (cp <= rsq)) {
computeSolution(z, psi, i);
return cp;
}
}
assert(0); // Never reached
return cp;
}
void arnoldiStep(LinearOperatorBase<FieldD> &LinOp, std::vector<FieldD> &v, std::vector<FieldD> &z, FieldD &w, int iter) {
FieldF v_f(SinglePrecGrid);
FieldF z_f(SinglePrecGrid);
ChangePrecTimer.Start();
precisionChange(v_f, v[iter]);
precisionChange(z_f, z[iter]);
ChangePrecTimer.Stop();
PrecTimer.Start();
Preconditioner(v_f, z_f);
PrecTimer.Stop();
ChangePrecTimer.Start();
precisionChange(z[iter], z_f);
ChangePrecTimer.Stop();
MatrixTimer.Start();
LinOp.Op(z[iter], w);
MatrixTimer.Stop();
LinalgTimer.Start();
for (int i = 0; i <= iter; ++i) {
H(iter, i) = innerProduct(v[i], w);
w = w - H(iter, i) * v[i];
}
H(iter, iter + 1) = sqrt(norm2(w));
v[iter + 1] = (1. / H(iter, iter + 1)) * w;
LinalgTimer.Stop();
}
void qrUpdate(int iter) {
QrTimer.Start();
for (int i = 0; i < iter ; ++i) {
auto tmp = -s[i] * H(iter, i) + c[i] * H(iter, i + 1);
H(iter, i) = std::conj(c[i]) * H(iter, i) + std::conj(s[i]) * H(iter, i + 1);
H(iter, i + 1) = tmp;
}
// Compute new Givens Rotation
ComplexD nu = sqrt(std::norm(H(iter, iter)) + std::norm(H(iter, iter + 1)));
c[iter] = H(iter, iter) / nu;
s[iter] = H(iter, iter + 1) / nu;
// Apply new Givens rotation
H(iter, iter) = nu;
H(iter, iter + 1) = 0.;
gamma[iter + 1] = -s[iter] * gamma[iter];
gamma[iter] = std::conj(c[iter]) * gamma[iter];
QrTimer.Stop();
}
void computeSolution(std::vector<FieldD> const &z, FieldD &psi, int iter) {
CompSolutionTimer.Start();
for (int i = iter; i >= 0; i--) {
y[i] = gamma[i];
for (int k = i + 1; k <= iter; k++)
y[i] = y[i] - H(k, i) * y[k];
y[i] = y[i] / H(i, i);
}
for (int i = 0; i <= iter; i++)
psi = psi + z[i] * y[i];
CompSolutionTimer.Stop();
}
};
}
#endif

19
Grid/algorithms/iterative/PrecGeneralisedConjugateResidual.h
View File

@ -139,8 +139,11 @@ namespace Grid {
MatTimer.Start();
Linop.HermOpAndNorm(psi,Az,zAz,zAAz);
MatTimer.Stop();
LinalgTimer.Start();
r=src-Az;
LinalgTimer.Stop();
/////////////////////
// p = Prec(r)
/////////////////////
@ -152,8 +155,10 @@ namespace Grid {
Linop.HermOp(z,tmp);
MatTimer.Stop();
LinalgTimer.Start();
ttmp=tmp;
tmp=tmp-r;
LinalgTimer.Stop();
/*
std::cout<<GridLogMessage<<r<<std::endl;
@ -166,12 +171,14 @@ namespace Grid {
Linop.HermOpAndNorm(z,Az,zAz,zAAz);
MatTimer.Stop();
LinalgTimer.Start();
//p[0],q[0],qq[0]
p[0]= z;
q[0]= Az;
qq[0]= zAAz;
cp =norm2(r);
LinalgTimer.Stop();
for(int k=0;k<nstep;k++){
@ -181,12 +188,14 @@ namespace Grid {
int peri_k = k %mmax;
int peri_kp= kp%mmax;
LinalgTimer.Start();
rq= real(innerProduct(r,q[peri_k])); // what if rAr not real?
a = rq/qq[peri_k];
axpy(psi,a,p[peri_k],psi);
cp = axpy_norm(r,-a,q[peri_k],r);
cp = axpy_norm(r,-a,q[peri_k],r);
LinalgTimer.Stop();
if((k==nstep-1)||(cp<rsq)){
return cp;
@ -202,6 +211,8 @@ namespace Grid {
Linop.HermOpAndNorm(z,Az,zAz,zAAz);
Linop.HermOp(z,tmp);
MatTimer.Stop();
LinalgTimer.Start();
tmp=tmp-r;
std::cout<<GridLogMessage<< " Preconditioner resid " <<sqrt(norm2(tmp)/norm2(r))<<std::endl;
@ -219,9 +230,9 @@ namespace Grid {
}
qq[peri_kp]=norm2(q[peri_kp]); // could use axpy_norm
LinalgTimer.Stop();
}
assert(0); // never reached
return cp;
}

1
Grid/log/Log.cc
View File

@ -59,6 +59,7 @@ void GridLogTimestamp(int on){
}
Colours GridLogColours(0);
GridLogger GridLogMG (1, "MG" , GridLogColours, "NORMAL");
GridLogger GridLogIRL (1, "IRL" , GridLogColours, "NORMAL");
GridLogger GridLogSolver (1, "Solver", GridLogColours, "NORMAL");
GridLogger GridLogError (1, "Error" , GridLogColours, "RED");

1
Grid/log/Log.h
View File

@ -169,6 +169,7 @@ public:
void GridLogConfigure(std::vector<std::string> &logstreams);
extern GridLogger GridLogMG;
extern GridLogger GridLogIRL;
extern GridLogger GridLogSolver;
extern GridLogger GridLogError;

35
Grid/qcd/utils/LinalgUtils.h
View File

@ -173,6 +173,39 @@ void G5R5(Lattice<vobj> &z,const Lattice<vobj> &x)
}
}
}
}
}}
// I explicitly need these outside the QCD namespace
template<typename vobj>
void G5C(Lattice<vobj> &z, const Lattice<vobj> &x)
{
GridBase *grid = x._grid;
z.checkerboard = x.checkerboard;
conformable(x, z);
QCD::Gamma G5(QCD::Gamma::Algebra::Gamma5);
z = G5 * x;
}
template<class CComplex, int nbasis>
void G5C(Lattice<iVector<CComplex, nbasis>> &z, const Lattice<iVector<CComplex, nbasis>> &x)
{
GridBase *grid = x._grid;
z.checkerboard = x.checkerboard;
conformable(x, z);
static_assert(nbasis % 2 == 0, "");
int nb = nbasis / 2;
parallel_for(int ss = 0; ss < grid->oSites(); ss++) {
for(int n = 0; n < nb; ++n) {
z._odata[ss](n) = x._odata[ss](n);
}
for(int n = nb; n < nbasis; ++n) {
z._odata[ss](n) = -x._odata[ss](n);
}
}
}
}
#endif