Re: Dual outline source files

[Antoine Leca]

Another interesting article on the OpenType mailing list...

Please note Greg's points: these are informal tests.

Greg Hitchcock <address@hidden> wrote:
> 
> Hi Antoine,
> 
> I guess I don't mind, but from a "publish" point of view, it was a very
> informal test, and I didn't go to extra lengths to confirm my results.
> 
> GregH
> 
>  -----Original Message-----
> From:   Antoine Leca [mailto:address@hidden
>
> Can I publish this article on the freetype list, for information.
> I believe this will be of interest to the TrueType users there.
> 
> En/Na Greg Hitchcock ha escrit:
> >
> > We did a quick and informal profiling test to see the impact of a non power
> > of two em. On a Pentium II, 300 MHz, we ran two versions of Palatino
> > Linotype, one with an em of 2048 and one with an em of 2049, and executed
> > every glyph at 19 PPEm.
> >
> > We examined the routine in TrueType that reads the outline data, scales the
> > outline, builds optional composites, and optionally applies the hints. I
> > looked at measurements on the routine that scales the outlines, and the
> > total time with and without hinting. Note, I did not include scan conversion
> > time in this measurement.
> >
> > For the scaling routine, there was a 5:1 performance difference on 2048 vs.
> > 2049, with the 2048 being 5x faster. This does not surprise me, as there is
> > a significant amount of code in the TrueType rasterizer to leverage this
> > performance scenario.
> >
> > The percentage impact of reading/scaling without hinting, was a 14%
> > performance degradation. With hinting, it was only a 4% performance
> > degradation (hinting takes a lot of time :) ).
> >
> > Big picture: there is an impact with a power of two em, and this is
> > important in some applications. Hinting, scan conversion, and other text
> > layout operations take a much larger percentage of the whole picture. The
> > caching of font outlines or bitmaps would reduce the impact to almost
> > immeasurable over a long period.
> >
> > GregH