A fast, easy to use, free, and community supported 2D game engine
Hey all, math-neon is a math library with ARMv7 (cortex) optimized cmath functions. There is also the vfpmath library for ARMv6 (1176) devices. I've been using fat binaries for a while, which allows you to compile your app optimized for multiple CPU architectures.
I'm sure chipmunk / box3d could benefit from some well targeted optimizations, so I might play with that some time and see how it goes.
I've emailed the project owner of math-neon to verify it's okay to incorporate into cocos2d. It is LGPL'd.
Cheers,
Stu
Got an email back from Lachlan, and he is okay with it, along with adopting the "static LGPL" license that cocos2d uses! Very cool!
I think some other places that could potentially benefit are anything doing matrix math:
CCSprite's updateTransform
CCNode's transform
CCCamera's locate (which actually means gluLookAt() which also does some matrix math)
All of these are potentially called EVERY frame.
I found implementations of Matrix, Quaternion, Vector and various transformation methods (like MatrixLookAtRH) in the open source Oolong engine and it already makes use of the vfp/neon optimized codepaths. They are all licensed under a zlib or MIT (very permissive) licenses, so it should be no problem to integrate into Cocos2D.
The only technical hurdle is to import these files and swap out the various existing cocos2d implementations.
Update: I have submitted an issue
http://code.google.com/p/cocos2d-iphone/issues/detail?id=825
I'll try to do some performance tests. If the results are good ones, I'll merge it in v0.99.5
Also, iOS4 comes with the Accelerate framework. One of the components is the BLAS framework which is a "fast matrix" library optimized for ARMv7, but it also works OK on ARMv6.
As it is now, I would go with vfp library and not with the official Apple framework since it only works with iOS >= 4.
great riq!!!
Great news Riq. I'm wondering how much blas performs better than math-neon and vfp math. I'm assuming it does, since the assembly is written by apple itself. If the performance benefit is considerable, maybe in the future there could be blas for >= 4.0 and math-neon, vfp math for < 4.0.
Edit Blas is also optimized for armv6, although the performance benefit of using vfp isn't as much as neon's.
@araker: I guess that the BLAS performance should be more or less the same as math-neon... at least regarding matrix multiplication.
So, probably cocos2d will do compile-time decisions, and not runtime decisions:
if compiled using ARMv6, then use vfp
if compiled using ARMv7, then use Neon.
If you are shipping fat binaries (ARMv6 + ARMv7), then it will use vfp in iPhone 1st and 2nd gen, and Neon on iPad, iPhone 3rd and 4th gen.
The thing is that math-neon is licensed under LGPL, and cocos2d is no longer licensed under "cocos2d license" so I'm not sure I'll be able to use that library. I'll ask the authors.
Backup plan: In case we can't use math-neon, Oolong-engine includes a 4x4 matrix multiplication function that is optimized for Neon, perhaps I'll adapt that one for 3x3 matrices
Update: Backup plan 2: http://blogs.arm.com/software-enablement/coding-for-neon-part-3-matrix-multiplication/
You must log in to post.
