/************************************************************************************* Grid physics library, www.github.com/paboyle/Grid Source file: ./lib/algorithms/iterative/MinimalResidual.h Copyright (C) 2015 Author: Azusa Yamaguchi Author: Peter Boyle Author: paboyle 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 { ///////////////////////////////////////////////////////////// // Base classes for iterative processes based on operators // single input vec, single output vec. ///////////////////////////////////////////////////////////// template class MinimalResidual : public OperatorFunction { public: bool ErrorOnNoConverge; // throw an assert when the MR fails to converge. // Defaults true. RealD Tolerance; Integer MaxIterations; Integer IterationsToComplete; // Number of iterations the MR took to finish. Filled in upon completion MinimalResidual(RealD tol, Integer maxit, bool err_on_no_conv = true) : Tolerance(tol), MaxIterations(maxit), ErrorOnNoConverge(err_on_no_conv){}; //! Minimal-residual (MR) algorithm for a generic Linear Operator /*! \ingroup invert * This subroutine uses the Minimal Residual (MR) algorithm to determine * the solution of the set of linear equations. Here we allow M to be nonhermitian. * * M . Psi = src * * Algorithm: * * Psi[0] Argument * r[0] := src - M . Psi[0] ; Initial residual * IF |r[0]| <= RsdCG |src| THEN RETURN; Converged? * FOR k FROM 1 TO MaxCG DO MR iterations * a[k-1] := / ; * ap[k-1] := MRovpar * a[k] ; Overrelaxtion step * Psi[k] += ap[k-1] r[k-1] ; New solution vector * r[k] -= ap[k-1] A . r[k-1] ; New residual * IF |r[k]| <= RsdCG |src| THEN RETURN; Converged? * Arguments: * \param M Linear Operator (Read) * \param src Source (Read) * \param psi Solution (Modify) * \param RsdCG MR residual accuracy (Read) * \param MRovpar Overrelaxation parameter (Read) * \param MaxIterations Maximum MR iterations (Read) * Local Variables: * r Residual vector * cp | r[k] |**2 * c | r[k-1] |**2 * k MR iteration counter * a a[k] * d < M.r[k], M.r[k] > * R_Aux Temporary for M.Psi * Mr Temporary for M.r * Global Variables: * MaxIterations Maximum number of MR iterations allowed * RsdCG Maximum acceptable MR residual (relative to source) * * Subroutines: * * M Apply matrix to vector * * @{ */ void operator()(LinearOperatorBase &Linop, const Field &src, Field &psi) { psi.checkerboard = src.checkerboard; conformable(psi, src); Complex a, c; RealD 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); // flopcount.addSiteFlops(4*Nc*Ns,s); // stands for "source squared" RealD rsd_sq = Tolerance * Tolerance * ssq; // flopcount.addSiteFlops(4*Nc*Ns,s); // // stands for "residual squared" /* r[0] := src - M . Psi[0] */ /* r := M . Psi */ // M(Mr, psi, isign); // flopcount.addFlops(M.nFlops()); Linop.Op(psi, Mr); // flopcount.addFlops(M.nFlops()); r = src - Mr; // flopcount.addSiteFlops(2*Nc*Ns,s); RealD cp = norm2(r); /* Cp = |r[0]|^2 */ /* 2 Nc Ns flops */ // flopcount.addSiteFlops(4*Nc*Ns, s); // auto cp = norm2(r); /* Cp = |r[0]|^2 */ /* 2 Nc Ns flops */ // // flopcount.addSiteFlops(4*Nc*Ns, s); if(cp <= rsd_sq) { /* IF |r[0]| <= Tolerance|src| THEN RETURN; */ return; } std::cout << GridLogIterative << std::setprecision(4) << "MinimalResidual: k=0 residual " << cp << " target " << rsd_sq << std::endl; GridStopWatch LinalgTimer; GridStopWatch MatrixTimer; GridStopWatch SolverTimer; SolverTimer.Start(); auto k = 0; while((k < MaxIterations) && (cp > rsd_sq)) { ++k; /* a[k-1] := < M.r[k-1], r[k-1] >/ < M.r[k-1], M.r[k-1] > ; */ MatrixTimer.Start(); // M(Mr, r, isign); /* Mr = M * r */ // flopcount.addFlops(M.nFlops()); Linop.Op(r, Mr); /* Mr = M * r */ // flopcount.addFlops(M.nFlops()); MatrixTimer.Stop(); LinalgTimer.Start(); c = innerProduct(Mr, r); /* c = < M.r, r > */ // flopcount.addSiteFlops(4*Nc*Ns,s); d = norm2(Mr); /* d = | M.r | ** 2 */ // flopcount.addSiteFlops(4*Nc*Ns,s); a = c / d; // a = a * MRovpar; /* a[k-1] *= MRovpar ; */ psi = psi + r * a; /* Psi[k] += a[k-1] r[k-1] ; */ // flopcount.addSiteFlops(4*Nc*Ns,s); r = r - Mr * a; /* r[k] -= a[k-1] M . r[k-1] ; */ // flopcount.addSiteFlops(4*Nc*Ns,s); cp = norm2(r); /* cp = | r[k] |**2 */ // flopcount.addSiteFlops(4*Nc*Ns,s); LinalgTimer.Stop(); std::cout << GridLogIterative << "MinimalResidual: Iteration " << k << " residual " << cp << " target " << rsd_sq << std::endl; } SolverTimer.Stop(); IterationsToComplete = k; // res.resid = sqrt(cp); std::cout << "InvMR: k = " << k << " cp = " << cp << std::endl; // flopcount.report("invmr", swatch.getTimeInSeconds()); std::cout << GridLogMessage << "MinimalResidual Converged on iteration " << k << std::endl; std::cout << GridLogMessage << "\tComputed residual " << sqrt(cp / ssq)<