Re: [ft-devel] Autohinter: stem darkening, first rough prototype

[Nikolaus Waxweiler]

I pushed some changes to 
https://github.com/madig/freetype2/tree/stem-darkening-autohinter. I 
have that branch installed system-wide for proper dog-fooding ;)

I changed the strategy to get the standard_width of a face. Instead of 
relying on the hard-to-follow workings of the style analyzers, first see 
if the font defines a stdVW by itself (Postscript fonts, once I figure 
out how to get at SFNT->CFF->PrivateDict without instantiating a 
CFF_Decoder also OpenType/CFF fonts). If it doesn't, in the future, set 
the style analyzers on it in some controlled way to determine the 
standard_width the hard way. Currently the code sets the standard_width 
arbitrarily like Dave suggested by looking at the weight class in the 
OS/2 table (light, regular, bold: 50, 75, 110). How well this works 
depends on how far off that value is from the "real" stdVW. I compared 
some article with the Linux Libertine font, because it comes with 
differently named .otf and .ttf fonts for quick comparison and it's 
dominant stem width is approximately 79, close to the 75 of it's weight 
class: http://postimg.org/gallery/32jipb0bi/0fb77538/ (Gamma correction 
through modified pixman). Alternating quickly between both shows that 
stem contrast is almost the same, so the default CFF darkening 
parameters seem a good fit there. If you look at the developer console 
though, you'll see that the font there (Liberation Mono) appears too 
thick vertically (stdVW =~ 88).

Another interesting thing when comparing Adobe's CFF engine with the 
emboldening autohinter: text is fairly readable at very small sizes with 
Adobe's engine while I have to skew the emboldening computation below 
9(pt?) because emboldening makes stuff less readable there: 
http://postimg.org/image/my3w9aerp/91a87a2d/. Top is Linux Libertine 
.otf -> Adobe CFF, bottom is .ttf -> autohinter.

What I now want to do:
- Find out how to access SFNT->CFF->PrivateDict.
- Find out how to to run style analyzers over the face in a controlled 
way to compute real standard_widths.

Once that is done, I think the prototype should be closer to merging to 
master. The last hurdle should be performance testing and optimization.