[Webkit-unassigned] [Bug 54456] Optimizing lightning filter to ARM-neon SIMD instruction set

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=54456





--- Comment #33 from Holger Freyther <zecke at selfish.org>  2011-03-08 01:21:03 PST ---
(In reply to comment #32)
> > Other issues:
> > - Make the code easier to understand. Naturally, it is still required to know the SVG standard to understand it.
> >   - I have added several comments.
> 
> Although the asm code itself and the neonized algorithm were also readable before, but the comments are really needed for non-ARM observers.

Readable and maintainable is surely subjective. Some people have no issue to read r0-r15 and remember what they are used for, same applies for %1 but then again a variable name is
something we were very used to from high level languages and GAS supports that. Anyway.


> I totally agree with you. The GCC's intrinsics does not offer accurate controlling of the NEON infrastructure. The most painful parts are the inaccessible register mapping and the instruction mixing.

Well, again a subjective matter. Projects like pixman have a copy of the code in intrinsics and then the handtuned result (based on the intrinsics version) as GCC still sucks.

> 
> > I soon upload the hopefully final patch.
> 
> I still think this patch is fine for landing.
> Gavin, what do you think about it?

Not so subjective. Could you post any benchmark data? How much faster than the C version? How much faster than a version compile for neon with tree vectorize and such?

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.