Behdad Esfahbod
Behdad Esfahbod
I'm not seeing a speedup, but ~3% slowdown. I ran it multiple times. Humm. ``` Comparing before to after Benchmark Time CPU Time Old Time New CPU Old CPU New...
Is there an increased setup cost per shape call? I wonder if https://github.com/harfbuzz/harfrust/issues/115 is relevant.
So, after this landed: https://github.com/harfbuzz/harfrust/pull/133 I actually see this PR speed up our Arabic fonts a good deal (5 to 9%), but slow down Roboto 2 / 3%. I'm in...
> An OpenType engine can only loop infinitely by growing the tape infinitely. [Citation Needed]
My proof that OpenType shaping is Turing Complete, albeit with a caveat: the machine never halts: https://www.youtube.com/watch?v=lK_myoawb0U
I mean, looping infinitely is the easiest thing in OpenType: just a Contextual lookup that loops back to itself. No tape needed more than one entry.
Also, for AAT `morx` fonts I think I also once proved that they are Turing Complete, despite looking like DFAs. It's because of the epsilon-transitions IIRC.
Haha. Nice one. Will measure soon.
On Mac I don't see any change whatsoever. On Linux, I actually see 5% slowdown on the ligature-heavy `BM_Shape/SourceSerifVariable-Roman.ttf/react-dom.txt/harfrust` benchmark and 1% slowdown in other benches. I'm okay merging, but...