[Matplotlib-devel] A (new) cairo backend for matplotlib

[Antony Lee]

Hi all,

Following my frustration with the somewhat incomprehensible (... to me) Agg
backend codebase, and as recently mentioned on the gitter channel, I have
hacked together a mostly complete native cairo backend for matplotlib.  See
https://github.com/anntzer/mpl_cairo (which also lists the missing features
that I can see).  It is based on pybind11 and requires a C++ compiler with
C++17 support (e.g., GCC 7.1).

One of the original motivations was to get rid of the "inaccurate marker
stamping" issue (https://github.com/matplotlib/matplotlib/issues/7262,
https://github.com/matplotlib/matplotlib/issues/7233).  Yes, I could have
tried to revert the original patch too...  The repository contains a few
examples where the new cairo backend places markers much more accurately.

Removing that optimization means that the backend is a bit slower than Agg
for drawing markers (though not more than ~1.5x); for other tasks it
appears to be nearly as fast.  It is dramatically faster than the current
(somewhat buggy) pycairo/cairocffi-based cairo backend (gtk3cairo, and
qt5cairo with https://github.com/matplotlib/matplotlib/pull/8771).

Being a native backend allows us to directly call libraries such as
freetype (bypassing -- in part -- the longstanding suggestion that ft2font
needs to be rewritten); it should likewise be possible to support complex
text layout (http://matplotlib.org/devel/MEP/MEP14.html) with the
appropriate library (from a quick search,
https://github.com/HOST-Oman/libraqm seems to be an interesting
possibility).  Additionally, cairo itself has an OpenGL (cairo-gl) backend,
which we may look into (currently, the cairo calls go into an image buffer
which is then picked up by the GUI toolkit for display; the idea would be
to instead use the cairo-gl backend to directly render to the display).

I am not making a PR out of this backend *yet* mostly due to the dependency
on a very recent compiler, which I believe would not be acceptable to many,
and would probably require delving into the build scripts.  Note that even
if I remove the C++17-isms from the codebase (which I'd rather not do --
"auto [x, y] = ..." is quite nice), pybind11 requires C++11 and suggests
using a C++14 compiler.  Additionally, the backend depends on a number of
other large (current) PRs which do not seem close to get in (as listed in
the README).  But I'd appreciate if you give it a try and give me some
feedback.

Antony
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/matplotlib-devel/attachments/20170629/a2d3dc60/attachment.html>