GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-14 04:35:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

Small cleanup of IDRS.h

Browse Source
This commit is contained in:
Chris 2022-11-16 13:51:23 +00:00 committed by Rasmus Munk Larsen
parent 02805bd56c
commit 6728683938
1 changed files with 8 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
unsupported/Eigen/src/IterativeSolvers/IDRS.h
View File

@ -66,7 +66,6 @@ bool idrs(const MatrixType& A, const Rhs& b, Dest& x, const Preconditioner& prec
const RealScalar tol = relres; const RealScalar tol = relres;
const Index maxit = iter; const Index maxit = iter;
Index replacements = 0;
bool trueres = false; bool trueres = false;
FullPivLU<DenseMatrixType> lu_solver; FullPivLU<DenseMatrixType> lu_solver;
@ -88,15 +87,14 @@ bool idrs(const MatrixType& A, const Rhs& b, Dest& x, const Preconditioner& prec
} }
// from http://homepage.tudelft.nl/1w5b5/IDRS/manual.pdf // from http://homepage.tudelft.nl/1w5b5/IDRS/manual.pdf
// A peak in the residual is considered dangerously high if‖ri‖/‖b‖> C(tol/epsilon). // A peak in the residual is considered dangerously high if‖ri‖/‖b‖> C(tol/epsilon).
// With epsilon the // With epsilon the relative machine precision. The factor tol/epsilon corresponds
// relative machine precision. The factor tol/epsilon corresponds to the size of a // to the size of a finite precision number that is so large that the absolute
// finite precision number that is so large that the absolute round-off error in // round-off error in this number, when propagated through the process, makes it
// this number, when propagated through the process, makes it impossible to // impossible to achieve the required accuracy. The factor C accounts for the
// achieve the required accuracy.The factor C accounts for the accumulation of // accumulation of round-off errors. This parameter has been set to 10^{-3}.
// round-off errors. This parameter has beenset to 10−3. // mp is epsilon/C 10^3 * eps is very conservative, so normally no residual
// mp is epsilon/C // replacements will take place. It only happens if things go very wrong. Too many
// 10^3 * eps is very conservative, so normally no residual replacements will take place. // restarts may ruin the convergence.
// It only happens if things go very wrong. Too many restarts may ruin the convergence.
const RealScalar mp = RealScalar(1e3) * NumTraits<Scalar>::epsilon(); const RealScalar mp = RealScalar(1e3) * NumTraits<Scalar>::epsilon();
// Compute initial residual // Compute initial residual
@ -224,7 +222,6 @@ bool idrs(const MatrixType& A, const Rhs& b, Dest& x, const Preconditioner& prec
if (trueres && normr < normb) { if (trueres && normr < normb) {
r = b - A * x; r = b - A * x;
trueres = false; trueres = false;
replacements++;
} }
// Smoothing: // Smoothing: