Mandelbox drawing performance questions

Hi everyone.

I’m fairly new to OpenGL/GLES and I’m taking on an unusual project. I’m trying to get a realtime Mandelbox explorer working on the iPad2. I’m using shaders pulled from “boxplorer” by “Rrrola”: Here is the project he made.

His technique uses ray casting and puts all the computation in the fragment shader. This is reasonable on today’s modern desktop & laptops but can’t achieve a reasonable frame rate on iOS devices. I would like to implement screen door style partial renderings, rendering every Nth fragment. On subsequent frames, I can then render another Nth offset by 1. Think even pixels, then odd pixels, then even, etc. For sample purposes, I’ll just assume two passes. The fragment shader would look something like this:


uniform int pass; // driving app flips this on each frame: 0,1,0,1
....
main() {
  if (int(frag.x + frag.y + pass) % 2 == 1) {
    // ... expensive rendering
  }
  else {
    discard;
  }
}

I’ve run tests and discard drops my workload in half, despite general claims that it’s a terrible thing. I suspect that’s specific to my hardware, which I’m fine with for now.

To do what I want to do though, I want to draw on top of the last frame, not clear the buffers. The GL analyzer tells me that if I don’t clear the frame, the GPU has to load the frame back in from system memory; an expensive operation. Is it better to do all my drawing into a persistent offscreen buffer (texture?) and then draw that on screen in a second drawing pass?

The bigger metaquestion is how can I make this shader faster to increase interactivity on slower devices. If anyone has thoughts on that, I’d appreciate the help as well.

Charlie.

Yea rendering to a texture is probably the smartest way.

Regarding making it faster, well it’s raycasting so you basically have two angles of attack

  1. optimizing the raycaster to reduce the number of tests needed.
    Using a polygonal proxy object is a start, so can varying the step length based on density

  2. optimizing the test algorithm itself, even a few instructions less can have a huge impact due to the number of times its executed.
    I haven’t yet worked with mandelbox, but if it’s algorithm anything like mandelbrot then each time you iterate, the volume of the fractal gets smaller. This could enable you to run at a lower or even variable iterative resolution until you get right up close to the “surface”.