Many games, even on current "next-gen" hardware, render particles using camera facing quads. In many cases these particles are used to represent volumes of many smaller microscopic particles. These volumes typically are simulated simply by determining how much contribution they present to the view using a simple blend function. This blend function defines how much the simulated volume of particles obscures the scene behind them.

Although this method has been employed in games for many years, this article defines a method using shader technology to more physically represent these volumetric particles. This method will give a more accurate visual representation of the simulated volumes as well as potentially decreasing the necessary number of particles, which in turn will help to improve render performance.

It should first be stated that the method defined in this article is limited to particles that represent volumes of sub-particles. It is also noted that the analysis that is to follow assumes a uniform density of the particles. There are methods that would allow the user to define more complex density functions, but that will not be covered here.

Many games have tried to represent environmental effects such as dust, mist, gases, energy volumes etc. using the same method for years. Essentially they define how much a particle obscures the rendered environment behind it using the alpha channel in the texture.

The alpha channel can be used to define transparency or opacity but in either case the algorithm is the same, the resulting pixel colour (Cr) is interpolated by some function (A) between the already rendered environment colour (Ci) and the particle colour (Cp). In most cases I have seen, the alpha channel defines the A value and the function is:

This is the basic interpolation function.

So what is wrong with this function? It uses two adds and two multiplies and in most cases is handled in hardware. It is simple and easy to use.

For the most part, semi- transparent camera-facing polygons rendered using this function represent the macroscopic volumes that simulate the microscopic particles as the user intended. Assuming the art is correct, they are sorted correctly from back to front and the quads they are rendered to don't collide with any other quads in the scene, the visual effect is sufficient to simulate microscopic particle volumes.

Now the art correctness is handled by the art department, and many engines handle sorting of the particles as they are rendered, so there is no problem there. How about the problem with them colliding with other geometry? A consequence of rendering all the particles as camera facing polygons is that the particles don't collide with each other almost all of the time. We can't make the same claim about the particles' collision with other geometry in the scene.

In some cases, the particles can be rendered without considering the depth buffer, but that assumes that nothing will obscure the particles themselves from the view. In other cases, there might be the CPU time to handle each particle's collision with the world, and simply make sure they don't ever encounter this case. For most games, the artifacts caused by particles and the environment have been considered an acceptable artifact.

Fig 1. Example from the Chrome Dragon Engine using the basic modulate blend mode render method for smoke particles

Unfortunately, due to the visual aliasing problems that arise from this method of particle rendering, art teams have developed some problematic procedures:

1. Typically, on a first pass of a volumetric particle effect, the artist will create the particles the size and density that best represents the visual quality that is appropriate. When the effect is added to the game, it typically becomes immediately apparent that the actual particles are geometrically too big.
   
   When they render, the artist is required to tweak the visual look away from the intended visual quality, and retrofit it to avoid any possible collision with environmental geometry.


2. In cases when it is necessary to have particles close to geometry, the artist is required to not only artificially shrink the particles to avoid noticeable aliasing artifacts, but typically increase the number of particles to account for the now-empty space. As soon as they increase the number of particles they must change the alpha values in the function to account for the increased likelihood of overlapping particles.
   
   This decreases the intended resolution of the alpha channel (potentially causing visual effects itself), makes the effect itself more grainy and exponentially dense, and worst of all, requires the artist (and possibly a programmer) to waste time tweaking to account for the inadequacies of the basic interpolation function.