flyspray/FS#1147: Compute norm of vectors with NaN entries properly

This is an attempt to solve the issue outlined in flyspray/FS#1147 (closed). I've yet to fix DenseMatrix::infinity_norm the same way.

Title changed from [WIP] flyspray/FS#1147 (closed): Computer norm of vectors with NaN entries properly to [WIP] flyspray/FS#1147 (closed): Compute norm of vectors with NaN entries properly

Added 34 commits:

• fff2581e...627fc0b0 - 31 commits from branch core:master
• f28ab282 - Clean up NaN norm test
• 2535cbc1 - Extend NaN vector test
• b3a810e3 - Fix vector norm computation with NaN values

Now that !11 (merged) has been merged, I've rebased and force-pushed the branch(*), so that only the three relevant commits remain.

(*) I felt at liberty to do so because nobody had replied here yet, but I'm not sure I'll do it again because the message above (@pipping added 34 commits) is so misleading

imho it would be easier to read if we use a SFINAE and just use different functions, instead of the the two structs.

Also I'd suggest to introduce our own trait Dune::has_nan<T>::value, because it allows to specialize the trait for other types, if they have nan. The default can just forward to std::is_floating_point.

Added 1 commit:

• f965455b - Introduce has_nan and use SFINAE

@christi That's a great idea. I lack experience with traits and it took me a bit to piece it together this way but now I've got what I believe is what you must've had in mind. The has_nan struct is certainly not in the right place yet. Should I move it to typetraits.hh?

@pipping: IMHO SFINAE is much more readable this way:

template<class T, typename std::enable_if<condition, int>::type =0>
...

but that's just a matter of taste. The traits should definitively go to typetraits.hh. Furthermore I very much like the fancy new traits pattern to derive from integral_constant<bool, result>.

Added 1 commit:

• b2edc8fe - Move has_nan and infinity_norm from fvmeta

Added 2 commits:

• b8506f60 - Use std::enable_if<condition, int>::type =0
• f541568a - Use std::integral_constant for Dune::has_nan

@carsten.graeser integral_constant is nice indeed. I've used it now.

Added 5 commits:

• 1e69ee29 - Typo
• a0b6c645 - Clean up test
• a8e32a87 - Extend test
• b8b9bb40 - Fix matrix norm computation with NaN values
• c03c5fa9 - Use integer constants

Title changed from [WIP] flyspray/FS#1147 (closed): Compute norm of vectors with NaN entries properly to flyspray/FS#1147 (closed): Compute norm of vectors with NaN entries properly

Now complete, ready to merge after a rebase if accepted.

Added 1 commit:

• 9983691c - Undo inadvertent whitespace changes

We will further have to adopt the implementations in dune-istl.

@pipping PS: thanks for all your effort on this topic.

Added 1 commit:

• 8520c610 - Do not use auto&& in loops

@christi Likewise! Thanks a lot for doing the benchmark and all the helpful comments :)