Blurst Technology

After running into fill-rate problems with dust/dirt particles in Raptor Safari we decided to implement High-Speed, Off-Screen Particles, as outlined by NVIDIA’s GPU Gems 3. Without getting too technical, the article shows how to render particles into a smaller render target (RenderTexture in Unity) and then blend the particles back into screen. This works well for dust, dirt, and smoke like particles because they have low frequency textures. This low frequency masks the natural “blurring” that occurs from upscaling the smaller render target to the screen’s resolution.

Rendering Depth

First thing we’ll need is a depth buffer. By rendering to a separate render texture we can no longer take advantage of the GPU’s built-in z-testing. Having a depth buffer will allow us to do our own z-testing. Unity made it stupid easy to get a depth buffer of the scene with version 2.6.0. On the main camera run this line somewhere in Awake:

However, for our purposes the depth buffer given to us by Unity is a bit overkill. Unity will render everything that the main camera can see. We only have dust-like particles near the camera, so there’s no reason for Unity to render depth information at far distances. So instead of taking advantage of Unity’s one-liner solution, we render our own depth buffer with a different far-clip plane. This isn’t as straight forward as setting the camera’s far-clip plane and is slightly outside of this article’s scope. We’ll address this in a future article.

Rendering the Particles

Here comes the hard part. First we’ll outline the steps we need to take to render the particles. All of this happens in the camera’s OnRenderImage function.

1. Create/Setup the render texture that we will render our particles into
2. Create/Setup the camera that will render our particles
3. Render the particles with a replacement shader
4. Blend the particles back into the screen with a composite shader

First we’ll create and setup the render texture that will hold our particles.

The downsampleFactor determines the quality of the off-screen particles. Higher numbers will give worse quality, but better performance.

Next, we’ll create and setup the camera.

And the PostProcessingHelper.GetPPCamera() function:

Notice how we are setting the layer that the particles are on. This is how the camera determines which renderers in the scene are the particles that we wish to render off-screen.

Next we render the actual particles. Telling the camera to render is easy enough.

The replacement shader is a bit unwieldy, so here it is as a file. Don’t worry too much about what’s going on in the replacement shader. Just make sure to place this shader in a Resources folder.

Lastly, we blend the particles back into the scene.

And the Composite shader (again, place this in a Resources folder):

Don’t forget to release the particles render texture!

And after you finish doing any other post processing effects you may be doing, output to the destination RenderTexture:

PRETTY PICTURES!

Click for bigger images. These all should be pixel perfect if you want to flip through or diff them.

Off-Screen Particles Disabled

Off-Screen Particles Enabled

Off-Screen Particles RGB

Off-Screen Particles Alpha

Notes

Separate Alpha Blend Function: As of Unity 2.6.1, there is no way to blend alpha channels with a different blending function. So this replacement shader takes two passes on the particles. Once to render to the RGB channel and once to render to the Alpha channel. Apparently, this will be fixed in a future release of Unity and the separate pass will not be necessary. I’ll update this replacement shader if someone reminds me when that happens. Update: Looks like this can be accomplished with a single blending function if you premultiply the alpha into the rgb channel in the pixel shader. The new blending function for rendering the particles is One OneMinusSrcAlpha. The new blending function for blending the particle RT back into the screen buffer is One OneMinusSrcAlpha. And your particle RT will need to be cleared to (0,0,0,0) instead of (0,0,0,1).

Mixed Resolution: We decided that mixed resolution particles wasn’t necessary for us. The scene is too fast moving to notice the depth sampling artifacts. Plus the performance overhead from needing a second pass for the alpha channel made mixed resolution rendering just too expensive.

Soft Particles: Soft Particles are extremely easy to implement with the off-screen particles, but we didn’t see much of a difference in the final render. We decided to just use discard instead of soft particles in the end.

Anti-aliasing: This is untested with Anti-aliasing on directx. I really doubt it’ll work correctly as is. Shouldn’t be too difficult to get it to work though.

“one day there was no internet in the office, so I did not know what bugs to fix… so I played around with this instead :)” - Aras Pranckevičius

That day of internet outage is probably one of the happiest days for me using the Unity3d engine. On that fateful day Camera.RenderWithShader was born which lived on to become the backbone of almost all of our post processing effects. The best way to explain how this makes a world of difference in our projects is to just give examples of how we handled post processing effects before and after replacement shaders.

Jetpack Brontosaurus

What We Did

We haven’t posted a graphics postmortem for Jetpack Brontosaurus yet. This post will simplify the techniques we used in order to demonstrate how we accomplished the “multiple dimension” effect. Each object in Jetpack Brontosaurus has at least three renderers associated with it. One for the nightmare/death dimension, one for the dream/living dimension, and one for the mask that determines which dimension gets rendered to the screen. The renderers are divided into each dimension by using different layers. We have a different camera for each dimension. One camera is set to render all the renderers in the nightmare dimension, one camera is set for the dream dimension, and one for the mask. These cameras output their renders to separate render textures. Then we combine the two dimensions based on the color in the mask.

The implementation required each of the three renderers to have its own game object, material, and layer. As a result, the complexity of the scenes increased, producing redundant information and human errors.

What We Could Have Done

If we had the ability to use replacement shaders during Jetpack Brontosaurus production, everyone’s life would have been easier. Most of the renderers in Jetpack Brontosaurus used a vertex color shader. Because the renderers shared the same shader, using a replacement shader is an easy change.

An important thing to note here is that all of the tags have the same Key/Value pair.

The dream replacement shader:

The death replacement shader:

The mask replacement shader:

Now we have one shader that has texture and color information for all of dimensions in a single material. No need for more than one renderer per object anymore!

The shader used by the material for the renderers (this shader is also used for rendering the object to the scene view):

For ingame rendering the material shader is never used. The only shaders used are the replacement shaders.

Now we have a script on the scene’s camera that renders the scene with each replacement shader and then composites them. The scene’s camera should be set to render nothing in its culling mask.

Blush

Glow

The Pro Standard Assets package that comes with Unity has a simple glow effect. Their implementation uses the alpha channel of the destination render texture to decide where to render glow. This limits us a couple ways. We can not have multicolored glow effects and we can not use the alpha channel for anything else. We decided to just bite the bullet and have a separate render texture for rendering glow.

Now we’ve freed up the destination render texture’s alpha channel for something else to use and we can have glow any color we like. Other than the additional memory needed for a 32bit render texture there’s a large disadvantage to doing it our way. Glow is no longer occluded by other geometry. A workaround for this is to have objects that you don’t want to glow render black to the glow render texture. Since glow is an additive pass, anything that is black will do nothing to the original image.

The Glow Replace Shader:

For every object we wanted to glow we added two things to the object’s original shader.
- Tag: “RenderEffect” = “Glow”
- Property: _Glow_Color (”Glow Color”, COLOR) = (1,1,1,1)

Distortion

Distortion is handled in a few steps.
1. Render the scene to a render texture.
2. Render a 2 dimensional “normal” map to a render texture.
3. Draw the scene’s render texture to the screen offsetting each texels’ texture coordinate by the amount specified by the 2 dimensional “normal” map texture.

Blush used a constantly oscillating full screen normal map to distort the scene. This distortion helped established an underwater feel. We used replacement shaders to render directly to this normal map to further modify the distortion. The artists used these shaders on particles to provide distortion effects on fast-moving tentacles, enemies, and other places.

The Particle Distortion Replace Shader: