Classdef cache

[behdad]

[layout] Use a cache for main input ClassDef of ChainContextLookups

I expected this to speed up NotoNastaliqUrdu, but didn't make a dent.
Need to profile on Linux to understand why. But it's mostly because
the cache is only effective when we enter the lookup from the main
loop (apply_forward), but NotoNastaliq mostly recurses...

[behdad]

The cache has a hit rate of ~10% on NotoNastaliqUrdu shaping. So I'm baffled that I'm not seeing any speedup whatsoever; well, I see maybe ~1% range only. But that's probably because the ClassDef2 lookups are only ~20% of shaping. So this makes sense. :'-(

[behdad]

I see ~2 percent speedup and ~1 percent overhead, adding up to net ~1 percent speedup...

[behdad]

Phew! Finally some progress!

Looks like NotoNastaliqUrdu spends most of its time in GPOS ContextFormat2 (dot positioning, etc...), and I was only applying the cache to ChainContextFormat2. Applying the cache to ContextFormat2 as well shows 16% speedup in NotoNastaliqUrdu. I observe a slowdown of less than 3% in Amiri as the overhead of new infrastructure, and much less than that for other tests.

_

Interestingly, the Gulzar font does a lot more lookups than NotoNastaliqUrdu, but is 7 times faster. Need to investigate when I have a profiler (Linux machine) around.
_
Update: I was testing wrong. Gulzar is indeed 20 times slower than NotoNastaliqUrdu. This branch speeds up Gulzar by 5% only, because Gulzar does a LOT of work still using ChainContextFormat3. Simon's Format3 to Format2 converter does not do a full job yet it seems. cc @simoncozens

[behdad]

cc @jfkthame @drott

[simoncozens]

I didn’t use format 3 to format 2 converter in Gulzar. I should try it.

[behdad]

I observe a slowdown of less than 3% in Amiri as the overhead of new infrastructure

Upon further measurement the slowdown is nothing like that. That one should have been noise. I don't see any measurable slowdown whatsoever.

[behdad]

This caching opportunity was first introduced to me by @mhosken many years ago.

[behdad]

From last commit I pushed:

[layout-cache] Cache lookahead, not input, classdef in ChainContextFo…
…rmat2

From the commit:

+    /* For ChainContextFormat2 we cache the LookaheadClassDef instead of InputClassDef.
+     * The reason is that most heavy fonts want to identify a glyph in context and apply
+     * a lookup to it. In this scenario, the length of the input sequence is one, whereas
+     * the lookahead / backtrack are typically longer.  The one glyph in input sequence is
+     * looked-up below and no input glyph is looked up in individual rules, whereas the
+     * lookahead and backtrack glyphs are tried.  Since we match lookahead before backtrack,
+     * we should cache lookahead.  This decisions showed a 20% improvement in shaping of
+     * the Gulzar font.