I felt like writing about the design process and some implementation details that I have been going through since I started working on GeoKone.NET. I will talk about performance issues, early designs that I worked for GeoKone, show some screenshots of different versions along the development process and finally look at what is up and coming with the next version of GeoKone.NET.
I was originally thinking about waiting for 1.0 version until to write about some of this stuff, but I feel it’s maybe better to split it into couple of parts and just to show you what kind of things I am facing when developing GeoKone.NET
Performance & Design Issues
One of the issues that I have been struggling constantly with GeoKone is performance.
The first version of GeoKone used a simple model: Processing.js running in noLoop() mode and the canvas was only redrawn when user changed the input. This worked pretty well, as GeoKone was still really simple.
But this noLoop() model was too simple for taking into account ways to present visual feedback for the user when interacting with the PolyForms (the formations on the screen, based on number of points around a circle). I needed a way to run logic & drawing even when the user was not doing anything, so I could present cool animations and transition effects in the program that would run for a while after the user stopped doing anything, or before stuff happened on the screen.
So I designed to take the game engine approach, where a collection of state machines are running at 30 FPS, rendering all polyforms on each frame. This model was used before versions 0.96, and it proved to be too slow to be really used without hardware acceleration.
This design was very responsive and allowed to make some nice transition effects and other realtime animations when joggling polyforms for example, but would almost immediately raise the CPU usage to 100% or even over on multiple cores, depending on the browser.
I also designed and implemented this cool Hyper Chakana Controller for modifying and interacting with objects on the screen. Here you can see a early design image that I had in mind for GeoKone running in fullscreen:
The Hyper Chakana Controller is the Compass Looking controller, with 4 actions in each direction, allowing Context Specific Actions to be mapped to each one of these directions, so that if you select a PolyForm, the Chakana would be able to Rotate, Scale, Move etc the polyform based on the Natural Directions the user is touching.
The name and design for this was based on the South American Sacred Geometry Cube, The Chakana, which you can see a 2D -version here:
I even went so far as to implement this HyperChakana controller, as you can see in this early preview video I made:
But after testing this model for a while, I realized that I cannot run this 30 FPS all the time, as making the CPU fan scream was not an option, so I had to figure out something else.
I looked into WebGL, but since back then it was still experimental (and still is, Safari does not officialy even support it yet, you have to toggle it via the developer options) I decided to stick with Processing.js + basic 2D canvas.
I also decided get rid of the Chakana Controller for now, although I put a lot of work into designing and implementing it. Hopefully I will be able to use this design in upcoming, native versions of GeoKone.NET, as I believe this could be a very natural way to interact with objects on the screen, especially with touch screens.
So I had to find a middle road, not running the logic & drawing at 30 FPS, but still having to be able to animate transitions between polyforms. So I decided to run logic for 50 milliseconds after the user has stopped interacting, and after this call noLoop() to stop Processing.js from calling the main draw() method. This way I could still animate stuff and run logic, and the it wouldn’t take as much CPU as before.
This model worked pretty well, and is the one that is still in use with the current live version (0.97). But it proved to create unnecessary logic for handling the stopping and starting of the loop() and noLoop() methods, creating some pretty ugly state handling code that is just unnecessary.
For the next version of GeoKone.NET 0.98, I have cleaned up the code and got rid of this difficult method of determining when to run the loop and when no to, and just tell Processing.JS not to run the loop at all in the beginning, and to call redraw() manually whenever the user interacts with the polyforms. This seems to be the only acceptable model in which GeoKone is responsive, and does not hog the CPU.
Also I had foolishly pre-optimized the code, using precalculated sine and cosine tables for the polyforms, inside the PolyForm class. These were not really even used as any time any parameter of the polyform was changed, the class was re-created completely. So even when the user moved the polyform around, it was re-created, thus re-creating the sine and cosine tables also, and preventing from re-using them. Doh. For the next version I have removed all this kind of “optimizations” and just draw and calculate everything on the fly.
Premature optimization truly creates a lot of problems, as the logic of the program changes during development process so much that the optimizations are not even affecting the program in anyway, but they are making it more difficult to adapt to changes in the architecture.
I actually profiled my code and found out that this creating of these sin/cos tables was causing major slowdown, as I used the new keyword everytime the PolyForm was re-created to create the tables. For debugging I use Firefox and the excellent Firebug extension, and I could see the more I removed any use of new in loops, the more faster the initialization & drawing got. This is kind of obvious, as creating classes in performance critical loops of course takes time to allocate new objects, instead of just changing parameters of existing objects on the fly.
It’s really easy to start optimizing early, and run into problems afterwards. This also bit me in the ass when trying to optimize the drawing so that all the in-active polyforms, that is, those polyforms which are not currently being edited, are being drawn into a separate backbuffer and the active polyforms are drawn to a front buffer, and these are then combined to make up what the user sees on the screen.
This enabled me to draw more complex scenes than before, as I could copy very complex formations into the background buffer, and just move the stuff in front buffer around.
But this created problems with the z-ordering of polyforms, as whenever I would choose polyforms to be modified in the front buffer, these would rise on top of the polyforms in the backbuffer, even though logically they were behind the ones in backbuffer.
This was caused because the backbuffer was drawn in the back, and the frontbuffer always on top of the backbuffer, ignoring completely the z-ordering of the polyforms and changing the way the scene looked when editing and when you disabling Mod All.
I have enabled this Back Buffer/Front Buffer optimization for now at least three times, and yet again I have to disable it as it causes problem with the drawing logic. Better just to stick with implementing the functionality first, and then worry about optimization :) It’s kinda difficult also to let go of these optimizations, as I know that I could be drawing much more faster scenes even with current versions, but there would be some minor drawing bugs which I find unacceptable. Maybe I will find a good way to do it after the program logic is finished.
Here are a couple screenshots of the next version in action, I’m not going to write anything more now as I’m really close to releasing this and I have to write those in the release notes anyway :) Major new improvement is the Layer Style Polyform Selector, which you can see on the left side of the screenshots. Also, you can now move the PolyForms up and down in their z-ordering, which makes it more easier to edit your scenes.
That’s it for now! Continuing finishing the last tweaks on the next version, and if you want to try it out yourself early, you can check it out at the master-optimization branch from GitHub: https://github.com/inDigiNeous/GeoKone/tree/master-optimization.