Absolute stopping criteria normalized by problem size

[blegouix]

Hello,

If I am not wrong, when using:

std::shared_ptr const residual_criterion = gko::stop::ResidualNorm<double>::build()
                                                           .with_baseline(gko::stop::mode::absolute)
                                                           .with_reduction_factor(eps)

the norm of residual will be compared with eps. However, it means it does not "adapt" to the size of the problem. For example, if I call gamma the average residual (typical ||R_ij|| value) for a multiple-rhs problem, the stopping criteria will lead to norm(R)~m*n_rhs*gamma<eps and thus gamma < eps/m/n_rhs. gamma is inversely proportionnal to the size of the problem.

Am I correct ? Is their any way to divide the norm per the size of the problem m*n_rhs while using absolute stopping criteria ?

(Note : I may not know m and n_rhs when building the factory, that's why I cannot juste write .with_reduction_factor(eps*m*n_rhs))

[MarcelKoch]

The residual check is applied to each vector of the right-hand-side individually. So we check ||R_:j|| < eps and not ||R|| < eps.