I've added performance numbers to the shaping comparison suite results
at http://ec2-54-226-13-158.compute-1.amazonaws.com/index.html. The
"timings" column lists the total number of seconds spent within
hb_shape() for the "native" HarfBuzz OT and Uniscribe backends, and the
percentage delta for the HB backend time compared to Uniscribe (i.e.
negative values are good, positive are bad).
The figures should be taken with a large grain of salt, and may not
translate directly to performance differences between HarfBuzz and
Uniscribe when integrated into other codebases; much
Still, there are some interesting things that emerge, such as that HB
seems to perform particularly poorly with certain fonts such as Noto
Sans Devanagari; or that when we compare WinXP fonts with their Win8
counterparts, HarfBuzz often runs faster with the Win8 versions, while
Uniscribe runs substantially slower.
It might be worth investigating some of the specific cases that are
highlighted by these results. E.g. are the relatively poor HB results
for many WinXP fonts with "simple" scripts a result of our fallback
positioning code, and if so can we perhaps optimize that? What is the
particular characteristic of a number of the Noto Indic fonts that
results in slower performance with these?
JK
_______________________________________________
HarfBuzz mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/harfbuzz
