<div class="gmail_quote">On Wed, May 23, 2012 at 6:08 AM, Klaij, Christiaan <span dir="ltr"><<a href="mailto:C.Klaij@marin.nl" target="_blank">C.Klaij@marin.nl</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div class="im">> > Yes, only the norms. As I understand now, GMRES will compute the<br>
> > norm of the total system residual [U,V,W,P], but I would like the<br>
> > norms of [U], [V], [W] and [P] residuals separated. Based on<br>
> > ex42, I got the impression that I have to use KSPBuildResidual.<br>
> ><br>
><br>
> Yes, the cost of building the residual (check -log_summary to see how big<br>
> it is) is a reason why people often avoid GMRES if they need the full<br>
> residual. It's directly available with GCR, for example (which is otherwise<br>
> similar in work to FGMRES) and with most of the non-residual-minimizing<br>
> methods.<br>
<br>
</div>I find GCR to be less robust than FGMRES,</blockquote><div><br></div><div>Right, it can break down in some cases that FGMRES does not.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
but ok, suppose I switch<br>
to GCR, how would I get the residual vector?</blockquote></div><br><div>KSPBuildResidual(), but it's trivially fast with GCR.</div>