Lua is a dynamically typed language, with only a small selection of primitive types. Any aggregate or otherwise complex types must use tables (i.e., dictionaries/associative arrays) in some form, which will significantly impact the performance of field/method access as well as the efficiency of data storage (and thus cache usage). LuaJIT’s foreign function interface (FFI) let’s us work around those limitations by offering us the ability to use C declarations to define new types with much greater control and efficiency. In this blog post, I’ll share how, for my Worldbuilder project, I used the FFI to implement a fairly performant fixed point number type. Public domain source code can be downloaded here. (more…)
I mentioned in my previous post that I needed to do some serious performance optimizations on my Lua code, and indicated in an edit that after a day’s worth of effort, I managed to speed things up by roughly a factor of 13. I have some more work to do, but in the hopes that the steps I’ve taken so far are helpful to anyone else going through a similar process, allow me to describe some of them before they slip my mind. (more…)
It took me three weeks of design, redesign, more redesign, lots of implementing scattered within, and three intense days of debugging a giant mass of previously untested code, but I finally have a basic modular system in place for running all procedural generation from Lua. This will enable me (and eventually anyone else) to quickly experiment with various algorithms for every stage of planet creation and presentation.
Unfortunately, I have a lot of optimizing investigations to do, because it seems to be running about 100 times slower than the prior C++ code. But at least it generates the exact same planet (give or take a few subtle discrepancies due to slightly different math here and there). Based on some of my earlier experiments at the beginning of the month, I’m pretty sure I can bring that up to within at least 10% of the speed of my C++ implementation, and quite possibly within 50% of its speed. Just need to profile and figure out the bottlenecks. (Edit: A day’s worth of investigation has gotten me up to around 13%, or 7.5 times slower than the C++ implementation. That should be acceptable for the moment.)
A cool thing about the architecture I ended up with is that not only will it naturally support a rich modularity of hooking up algorithms at different stages to each other, but that the way this modularity is exposed will also automatically enable a significant degree of concurrent execution on multiple cores with little to no effort on the part of the script writer. Right now I have only implemented a single threaded execution model, but I should be able to change these details under the hood when I get to that stage in the project, and the Lua scripts won’t know the difference. If you’re curious, allow me to provide an overview of how I’ve designed this modularity and concurrency. (more…)