High-speed, Off-screen Particles in Unity

Our approach to fill-rate problems with massive numbers of particles filling the screen.

http://technology.blurst.com/off-screen-particle-optimization/

Hope someone finds it useful

 

Nice. I implemented a downsized particle buffer in a project at work last year for some hardware that was fillrate limited. We used if for all 'transparent' objects in the end including distortion objects ect.

You mention that you didn't have issue with artifacts, I'm guessing that is due to the 'close' range that the particles are rendering into. We were not so lucky and it took some crazy up scaling / full res depth buffer sampling tricks to make the scene recombine nicely.

We also didn't rerender the scene for the depth buffer, we did a downsample of the full res buffer. We could not afford the time to rerender the geometry.

One nice way we found to fix the 'bad' edges on upscale was to make the edges a bit 'softer' by sampling the full scale depth buffer and the downsampled depth buffer at the texel location and if they were different do a few different combines (multisampling of the depth buffer / heuristic) to make the edges seem softer and match up.

Oh and nice work on the premultiplied alpha. It's the only way to fly

 

Guess I didn't make that clear enough. That's what we did too. We rendered a full res depth buffer (since we don't have access to the depth buffer on the graphics card) and then downsampled it for use by the offscreen particles. We're using the depth buffer for other effects as well, so we needed it full res. Agreed that it would be too expensive to render another depth buffer at the lower res just for the particles.

 

Ahh, I read this and it seemed to indicate that you were rerendering part of the scene (with a different clip plane.

What method did you use to downsample the z-buffer? Did you just sample 4 pixels and pick one? If so what kind of heuristic did you use for pixel selection.

We took the pessimistic option and always took the pixel closest to the camera. We found that this was more appropriate given our use cases. But as you are just doing additive particles (and looking a the pictures) it seems like you took the optimistic route of picking the z value furthest from the camera. Some of the pixel bleeding gives it away.

 

Ah, all of our effects work with a depth buffer that has a reduced far clip plane.

Nothing special Just point sampling at the center of pixels. I tried doing tests for pixels that were closer and pixels that were further, neither really looked any better or worse for us.