Closed lkdvos closed 7 months ago
Attention: Patch coverage is 81.79272%
with 65 lines
in your changes are missing coverage. Please review.
Project coverage is 82.05%. Comparing base (
e4932b4
) to head (7a27157
).
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
Hi Lukas; thanks for this great work. I was indeed considering switching to bangbang methods when adopting VectorInterface, so if you have time to try this out and see if there would be any unexpected consequences or complications, feel free to do so ...
@Jutho I seem to be stuck with the geneigsolve for out-of-place vectors, although I can't seem to locate what's going wrong. It produces matrices that are not positive definite, such that geneigh! fails, but I have absolutely no clue why this would be different for the out-of-place methods. I might look into it later this week, but any help is definitely welcome. (tests are ran by just commenting out modules "svdsolve.jl" and "expintegrator.jl")
Some small remaining remarks:
So, if the lights turn green, this should be ready for review.
I excluded AD rules from the new functionality, because we should probably rewrite that section anyways, using Package extensions.
Otherwise, I think everything should now work with VectorInterface
, and the tests have been reorganised to extensively test everything with regular vectors, and then do some small additional tests with MinimalVec
-type vectors.
In principle I think we can also remove/deprecate RecursiveVec
and InnerProductVec
, but this would probably require some deprecation update first?
Nightly failures seem due to Zygote failing?
I think so, it also seems to vary quite a bit depending on what version of nightly, so I think we can ignore this. We'll refactor that part anyways when we move the AD support to a package extension I guess.
rewrite algorithms in terms of VectorInterface methods.