Repository navigation
Upgrade agg to SVN version - #3777
Conversation
|
The failing image comparisons are very subtle. Haven't got to the bottom of why yet, but the difference in colors is quite minor. |
Conflicts: src/_tkagg.cpp
Conflicts: setupext.py
|
On close reading of the code, it seems that the new results are actually more correct. The image interpolation calculation is being done in 32-bit rather than 16-bit integers in the new Agg version resulting in slightly smaller rounding errors. I've updated the base images here. |
|
I am seeing a bunch of extraneous whitespaces. |
|
Are they coming from the SVN version of agg? |
|
Yes. The Agg files are included verbatim. |
There was a problem hiding this comment.
dropping a namespace? perhaps we should push up some patches to them?
There was a problem hiding this comment.
In this case, it's just the dropping of a redundant namespace specifier, since this code is in a namespace agg block. It's not actually changing the namespace in which rendering_buffer lives. Generally, agg doesn't using the namespace prefix everywhere in its internal code, which is typical C++ practice.
|
Should I trim the whitespace on Agg code now? It makes things easier now, but will make tracking Agg SVN harder in the long run, unless I can convince Agg to change their trailing whitespace policy, which seems unlikely. |
There was a problem hiding this comment.
these are getting passed into functions uninitialized. That doesn't seem right to me.
There was a problem hiding this comment.
Actually, their addresses are passed to vertex, where they should be set, but they aren't currently, due to the fault of matplotlib-side code. See #3781.
As a first step toward doing image interpolation in floating-point, and having (optional) floating-point rendering buffers for more accurate alpha blending, this upgrades our included copy of Agg to the current SVN version here: http://sourceforge.net/p/agg/svn/HEAD/tree/
This version of Agg is a fork of Agg 2.4, so it is still under a BSD license.
A few caveats: There is a regression in the alpha blending on RGBA 32-bit (8-bit-per-plane) buffers caused by doing the alpha blending calculation in 8-bits rather than 32-bits as it was before. I've communicated on the Agg mailing list about a solution, but I can't find the "right way" to do it. So, instead, I've included the one-off solution that uses the old calculation from Agg-2.4 verbatim.
The core Agg developers seem to see no need for releasing tagged versions of Agg. (Again, discussed on the mailing list). This may be a problem for some packagers, since matplotlib can no longer be built against the "system" Agg. Currently, Fedora uses a system Agg for its matplotlib package, but Debian, Arch, MacPorts, and Homebrew do not.