Improve varLib model algorithm.
This, basically means any varfont built that had an unusual master
configuration will change when rebuilt.
Here's a good test: a two-axis with 8 masters at unusual locations:
2-----------------5----------3
| |
| 7 |
| |
| 6 |
| |
| |
| |
0-----------4----------------1
Previously, the reach of master 3 (Black Extended) would
have started from A=.4, ie, the A position of master 4.
It now correctly starts from 0. Same thing with masters
after it. Ie, master 5 gets a span on the A axis from
0 to 1, whereas before it was getting from .4 to 1.
Previously, the on-axis masters always cut the space. They
don't anymore. That's more consistent, and fixes the main
issue Erik showed at TYPO Labs 2017. Same issue was also
causing the reach of master 3 to be limited, which I think
is the issue being discussed in the linked issue. Both should
be fixed.
It's hard to describe exactly what happened before / after.
Best to read the actual support values:
Before:
Sorted locations:
$ ./fonttools varLib.models 0,0 0,1 1,0 1,1 .4,0 .6,1 .5,.5 .7,.7
[{},
{'A': 0.4},
{'A': 1.0},
{'B': 1.0},
{'A': 1.0, 'B': 1.0},
{'A': 0.6, 'B': 1.0},
{'A': 0.5, 'B': 0.5},
{'A': 0.7, 'B': 0.7}]
Supports:
[{},
{'A': (0.0, 0.4, 1.0)},
{'A': (0.4, 1.0, 1.0)},
{'B': (0.0, 1.0, 1.0)},
{'A': (0.4, 1.0, 1.0), 'B': (0.0, 1.0, 1.0)},
{'A': (0.4, 0.6, 1.0), 'B': (0.0, 1.0, 1.0)},
{'A': (0.4, 0.5, 1.0), 'B': (0.0, 0.5, 1.0)},
{'A': (0.5, 0.7, 1.0), 'B': (0.5, 0.7, 1.0)}]
After:
$ ./fonttools varLib.models 0,0 0,1 1,0 1,1 .4,0 .6,1 .5,.5 .7,.7
Sorted locations:
[{},
{'A': 0.4},
{'A': 1.0},
{'B': 1.0},
{'A': 1.0, 'B': 1.0},
{'A': 0.6, 'B': 1.0},
{'A': 0.5, 'B': 0.5},
{'A': 0.7, 'B': 0.7}]
Supports:
[{},
{'A': (0, 0.4, 1.0)},
{'A': (0.4, 1.0, 1.0)},
{'B': (0, 1.0, 1.0)},
{'A': (0, 1.0, 1.0), 'B': (0, 1.0, 1.0)},
{'A': (0, 0.6, 1.0), 'B': (0, 1.0, 1.0)},
{'A': (0, 0.5, 1.0), 'B': (0, 0.5, 1.0)},
{'A': (0.5, 0.7, 1.0), 'B': (0.5, 0.7, 1.0)}]
TODO: We should add this as a test case.
There's another improvement I want to make, but that's separate.
The variation merger reuses matrix rows / entries when aligning kerning
matrices. Don't remap same DeviceTable twice. Alternative fix would
have been to not reuse objects but that slows things down.
Fixes https://github.com/fonttools/fonttools/issues/1206
While we optimize VarStore later, just deduping here reduces amount
of data we have to work with later by 8x in NotoSans-VF.ttf test
case. It's a cheap optimization, so do it.
Tests were failing on python 2.7.6 because array module doesn't accept
unicode strings (fixed in later 2.7.x releases).
An alternative fix would have been to cast array(str('h'), [...]) but
that seemed a bit too much, especially since varStore module doesn't
really do heavy text manipulation (there aren't other string literals).
We cannot use unicode_literals unconditionally, because of this and
similar issues in the stdlib. We should use it only when it makes sense
and reduces the frequency of u"".
Ouch! Shrinks NotoSans-VF.ttf GDEF VarStore further down.
Original was 1016278 bytes. Before this change it was optimizing
to 256145 bytes. With this change, it goes down to a mere 57127
bytes!!!
Right now just reflects the axes, and even that with certain limitations:
- AxisOrdering is set to the order axes are defined,
- Name-table entries are not shared with fvar.
Towards https://github.com/LettError/designSpaceDocument/issues/8
