On Thu, 22 Sep 2016 18:10:44 +0200
Dominik 'Rathann' Mierzejewski <dominik(a)greysector.net> wrote:
On Thursday, 22 September 2016 at 16:55, Dan Horák wrote:
> On Thu, 22 Sep 2016 15:41:01 +0200
> Dominik 'Rathann' Mierzejewski <dominik(a)greysector.net> wrote:
>
> > Hello,
> > I've just pushed (but not built) python-matplotlib-2.0.0b4 to
> > rawhide. I'll be attempting to rebuild all the affected packages
> > locally to test if they're compatible. In the meantime, feel free
> > to git pull and build locally for your own testing.
>
> python-matplotlib has been a pain for alternative arches recently,
> due failures in the test-suite (in 1.5.x). Is the 2.0 version going
> to be better in this regard?
tldr: no, but upstream is aware and working on it.
There are quite a few issues opened on upstream github tracker either
by me or by Debian maintainer. The main problem is that upstream
checks that rendering output matches their generated images. This
changes slightly with different versions of freetype. As a
workaround, I reintroduced a certain base tolerance in order not to
add specific tolerance value to each failing test (there are ~900
failing due to slight rendering differences). Then, there are some
arch-specific issues, mainly on ARM.
Could we disable the test-suite for alternative arches until there is
proper upstream fix? Quite a number of packages/builds are blocked on
python-matplotlib currently.
For math precision issues -ffp-contract=off often helps. And yes,
comparing rendered output can be considered fragile, we have already
met it ...
Dan