Hacker News new | ask | show | jobs
by feistyio 4624 days ago
I wish you all the best with your ambitious project ^w^

You should seriously consider employing techniques such as LLJS[1] to avoid thrashing the GC.

At the very least use typed arrays à la glMatrix[2] rather than a Vector prototype.

Typed arrays have the added benefit of interoperability with WebGL and Canvas2D via Function::apply.

---

[1] http://mbebenita.github.io/LLJS

[2] http://glmatrix.net

2 comments

Author of ChipmunkJS here. I did some benchmarks moving things to typed arrays about a year ago, and the change was a wash in my benchmark results. I think that might not be the case now considering all the optimizations the v8 team have been making to asmjs-like code.

I did some quick benchmarks using emscripten recently and performance absolutely blew away my hand-tweaked chipmunk code[3]. Hand optimized, ChipmunkJS takes 2.8x as long to run the tests. Emscripten compiled chipmunk is 1.4x, so twice as fast.

That said, the generated code is huge and awful, and the API is terrible. I've been playing around with porting my library to LLJS recently, but when I looked at LLJS it hadn't been updated in over a year. Its nice to see that there were some commits a few months ago, but LLJS needs a lot of work before its generally usable. I've been doing some stuff manually as a toy - still not sure where to go with all of this. Bridging from ASMJS to regular javascript makes your API really hard to use.

---

[1] https://github.com/josephg/Chipmunk-js

[2] https://docs.google.com/spreadsheet/ccc?key=0Ap2hDnCZnXiVdDV... , though this is somewhat hard to read.

[3] http://josephg.com/blog/chipmunkjs-and-emscripten

[4] http://josephg.com/blog/chipmunk-in-asm-js

A primary goal of newton is to be accessible, simple, easily read JavaScript with a minimal build step, so I don't plan on moving to something like LLJS. However, I agree that GC friendliness should be a priority. I may implement object pooling (again), but, after a branch testing just that, I found no significant performance improvement over the existing mutable-by-default Vectors.

On the other hand, the vec2 class of glMatrix looks fantastic and this is something to consider. Thanks!