[CESCG logo] PARSEC: Enhancing Realism of Real-Time Graphics Through Multiple Layer Rendering and Particle Systems

Markus Hadwiger

Institute of Computer Graphics
Vienna University of Technology
Vienna, Austria
[CESCG logo]


One of the major goals in real-time computer graphics is to achieve high realism of rendered images without compromising rendering speed too much. Frame update rate requirements are especially demanding in the area of entertainment software. Player immersion is directly related to both the number of frames displayed per second and the visual quality of the images rendered in each frame. Traditionally, these two goals have largely excluded one another. In recent years, however, standard desktop hardware has become powerful enough to make real-time rendering of convincing virtual environments possible.
    This paper describes the rendering engine of Parsec, a three-dimensional computer game we are currently developing. In this, we are focusing on those parts of the rendering process that are not part of the standard transform-clip-draw pipeline for polygonal objects. We are rendering multiple layers to achieve different effects with techniques specifically suited to the nature of these effects. We use an image-based rendering approach to render an infinitely far-away background. Particle systems are employed to render effects that cannot easily be achieved with polygonal objects. Our software-renderer also harnesses multi-layer rendering to decouple specific rasterization tasks in order to gain significant speed-ups.

Keywords: multi-layer rendering, image-based rendering, environment mapping, computer games, real-time graphics, particle systems, special effects, flare, glare.

0. Contents

1. Introduction

The developer of contemporary three-dimensional entertainment software constantly faces the problem of two largely contradictory goals:
    First, rendering speed has to be very high to achieve the frame update rates necessary. If too few images are rendered per second the virtual environment will not be very convincing and enjoyable for the player. The feeling of immersion is easily lost when the frame rate is below, say, 20 frames per second (fps), and computer games often demand frame rates of 30fps and more. Even more important than high peak frame rates is the requirement that the frame rate be almost constant in order to facilitate smooth animation. A somewhat lower but largely constant frame rate is clearly preferable to widely varying rates with a higher peak frame rate.
    Second, the quality of the generated images is also crucial to achieving player immersion. The more aspects of reality are captured in those images, the better the feeling of "actually being there".
    The key to rendering a convincing virtual environment in real-time is to make well-considered trade-offs between visual quality and rendering speed. If different parts of a scene are rendered using different techniques well suited to these parts, significant speed-ups can be gained in contrast to rendering an entire scene with a single rendering technique.
    Hence, the idea of rendering the projected image of a scene in multiple passes and using various object representations is very important to combining high visual quality and rendering speed. In this paper we are going to look at exactly those aspects of the rendering engine of Parsec, a three-dimensional computer game we are currently developing [Parsec]. Parsec is a space-fight game specifically targeted at network gameplay. The current version can be played by four simultaneous players over an IPX network. Since the background of a Parsec screen depicts mostly unreachable parts of outer-space, image-based rendering is naturally suited to rendering this background. We use multiple layers of cubic full view panoramic images to render nebulae, stars, and planets, similar to environment mapping with cubic environment maps. Objects like spacecraft are polygonal models rendered using a graphics pipeline employing BSP trees. Special effects are rendered after polygonal objects, mostly using particle systems. Additional lens flare effects can be rendered last.

2. Related work

In this section we are going to briefly review research work concerned with those fields of computer graphics we have drawn upon in our own work, and which have proved to be very useful for our particular application.

2.1 Image-based rendering

Image-based rendering and environment mapping techniques are particularly suited to rendering an infinitely far-away background of an outer-space setting, that is, for part of our application. [Gree86] mostly considers different types of environment maps as they are used for reflection mapping [Blin76]. Nevertheless, Greene’s paper already mentions that world projections onto environment maps can be used to render a far-away background without explicitly modeling it, e.g., a sky backdrop. In recent years, the concept of image-based rendering has received a lot of attention, often also removing the requirement that the rendered background be far away. [McMi95] provides a consistent conceptual framework to all these techniques via plenoptic modeling, using the so-called plenoptic function. Since a single panoramic image is only correct for a single viewpoint, view interpolation and morphing may be used to seamlessly migrate from one viewpoint with an available panorama to the next [Chen93]. Recently, image-based techniques have also entered the marketplace for consumer-products. The work in [Chen95] still requires a specialized viewer, though, that is capable of cylindrically distorting the employed cylindrical panoramic images on-the-fly. Cylindrical panoramas also restrict the possible camera rotation, since they are not full view. Another problem may be the need to track camera motion in order to be able to reconstruct panoramic images. Video mosaics [Szel96] alleviate this problem. This work is greatly expanded upon in [Szel97], which provides a technique for creating full view panoramas from image mosaics. The algorithm presented in this work is able to automatically register images and calculate texture-maps for arbitrary all-encompassing objects. The simplest such object still yielding impressive results is a cube with the viewpoint at its center. The image-based approach can be combined effectively with the notion of rendering the projection of a scene in multiple layers [Leng97]. This helps in allocating rendering resources according to the visual importance of different parts of a scene and their frequency of change. Rendering multiple layers is an important approach to exploiting frame-to-frame coherence. We have also used the related idea of multiresolution compositing [Berm94] to render different parts of a scene in resolutions particularly suited to them, performing image-composition afterwards.

2.2 Particle systems

Particle systems are a powerful approach to rendering effects that cannot easily be modeled otherwise [Reev83]. Particles are assigned basic properties like appearance and velocity, a huge number of them behaving in a conceptually identical way [Sims90]. Their visual effect on the viewer depends largely on the creativity of the modeler, however, who has to creatively harness the basic functionality provided by such systems. More powerful primitives, like vector fields [Hilt94], can help in this task.

2.3 Special effects

We have also used work on various other special effects. [Spen97] describes in detail how glare is caused by the human visual system. This information can be used to enhance the realism of computer-generated images by adding bloom and flare lines around very bright objects in a scene in order to increase their perceived brightness. If those effects are pre-rendered and used as texture maps during run-time, they can easily be used in a real-time rendering system. The work in [Reed94] provides physically-based information about how to render realistic lightning. Nevertheless, the emphasis is on visual appearance instead of physically exact evaluation. Particle systems are also used in this work, and we have actually integrated a modified lightning effect as a "special weapon" in Parsec, utilizing our particle system.

3. Rendering layers in Parsec

In this section we are going to explain step by step how each frame in Parsec is rendered. There are four major layers: Panoramic background, polygonal objects, particles, and additional special effects. We discuss these layers in the same order as they are rendered.

3.1 Panoramic background

The visual appearance of the background is especially important in an outer-space setting like Parsec, since large areas of the screen will be dominated by this background only. In Parsec, the background depicts large-area nebulae, hundreds of stars only a few pixels in size each, and other objects that can never be reached by the player, like planets and stars bright or big enough to justify using textures instead of single points. Image-based rendering is ideally suited to be used in this case, but not only because the player can never reach anything being part of the background. Characteristic astronomical phenomena like nebulae can simply not be rendered realistically enough using another real-time rendering technique. Panoramic background rendering is done in four layers. The first two layers are responsible for nebulae and stars, respectively. Layers three and four are used for two different types of detail objects. But first we will have to consider how to represent a full view panorama.

3.1.1 Choosing a suitable world projection technique

To surround the player with a seamless full view panoramic image, a suitable representation for a world projection [Gree86] must be chosen. Cylindrical projections, like in [Chen95], are out of the question, since the player is able to look about without any restrictions. Therefore, some kind of spherical projection has to be used to achieve a full 360° view in every direction. Ideally, if we could somehow project the entire world onto the surface of a sphere centered at the current viewpoint, and sample this sphere for each pixel on the screen at run-time, we would be able to render images identical to rendering the entire world. This would work regardless of camera orientation, as long as the viewpoint (camera position) stays the same. In our case, even the viewpoint position need not stay the same, since everything being part of the background is considered being infinitely far away. Actually, with respect to panoramic background rendering, the viewpoint is always located at the world’s center. Camera translation is ignored in this stage of rendering, although it is used in all other stages, of course.
    The first problem with this ideal world projection is that normally the sphere’s surface has to be sampled at discrete points, provided that we cannot always recalculate a color for an arbitrary surface-point at the time we need it. This would only be the case if we were able to represent the entire surface procedurally. At first glance this may seem totally out of the question, but if the world consists only of procedurally represented nebulae (i.e., solid texturing functions) this would actually be possible. Still, on-the-fly evaluation of solid texturing functions is not what we desire for a real-time application, and we also do not want to restrict ourselves in terms of flexibility regarding the method employed for creating the actual nebulae. As soon as the sphere is not represented continuously, we represent the world projection by a finite set of sampling points. At run-time we could resample such a discretely sampled sphere to render an arbitrarily oriented view of the world.
    The second problem, however, is how to represent the surface of a sphere by a planar texture. A latitude-longitude projection has always problems at the poles, for instance. Furthermore, since we want to use the environment map to render the entire background, not only for reflections on objects’ surfaces like in reflection mapping, we do not want to afford spherical projection of the map at run-time, in order to be able to display it full-screen. In addition, we want to be able to exploit standard texture-mapping hardware, which is only capable of performing planar projections.
    Another approach to representing full view panoramic images, described in [Szel97], is to surround the viewpoint with an arbitrary, all-encompassing, texture-mapped object centered at the viewpoint. The textures for this object can be calculated by sampling the entire world for each texel. At run-time, it is rendered using planar texture-mapping only. Therefore, graphics accelerator hardware can naturally be exploited. For all objects other than a cube, the needed textures cannot easily be created using standard rendering packages, however, which is why we chose to use a cube surrounding the viewer. We modeled background nebulae and stars using the freely available POVRay raytracer, employing procedural techniques. In POVRay, it is particularly easy to specify six cameras with 90° viewing angles to project the entire world onto the six sides of a cube, without incurring noticeable seams at the cube’s edges.

3.1.2 Panoramic layers one and two: Nebulae and stars

As we have stated before, nebulae and stars are rendered in two independent layers. The reason for this is mainly texture-map resolution, that is, memory-consumption of the employed environment maps. The first cubic environment map (layer one) is comprised of six 256x256 RGB textures, one for each side. Since nebulae are rather diffuse phenomena and bilinear texture-filtering is used at run-time, this resolution is high enough. The second cubic environment map (layer two) is comprised of a single 256x256 alpha-only texture, used 24 times, that is, four times per side. This yields an effective resolution of 512x512 per side for the stars texture. Since this texture is used to render hundreds of small stars of just a few pixels each, its tiling and reuse are not noticeable. See Figure 1. There is still the problem, however, of how to blend these two layers efficiently and still achieve the same visual quality as if full-resolution textures had been used. Simple chroma-keying will not suffice in this case, because of bilinear filtering. Stars are simply points of varying levels of gray (non-fullbright white) and the rest of the texture is black. After bilinear filtering, these points will be surrounded by not-quite black haloes that would be very disturbing when overlaid upon a non-black area, i.e., a nebula. This is the reason why we use an alpha texture for layer two. After layer one has been drawn, stars represented by alpha-values only are drawn as effectively antialiased points. This is possible because we are exploiting the graphics accelerator’s ability to bilinearly filter the alpha channel and blend the result with the already rendered frame. No chroma-keying need be employed. If a star is overlaid upon a, say, blue area of a nebula, it will be white at its center and have varying levels of blue around it, i.e., be antialiased to the surrounding blue background. If it is overlaid upon a black area, however, it will be correctly antialiased to black.


 figure 1
Figure 1: Cubic panorama with nebulae, stars, and detail objects

3.1.3 Panoramic layer three: Detail objects

After nebulae and point-sized stars have been rendered in the first two layers, detail objects are overlaid as layer three. Detail objects are used for everything being "infinitely" far away, but needing a texture of its own, e.g., a far-away planet. That is, the main reason for layering is once again texture resolution and memory efficiency. It is simply not reasonable – and often not possible – to use environment maps of six times 1024x1024 or higher resolution, just because it is needed for certain patches, in contrast to most areas of the environment map. We therefore store just exactly those patches at high resolution, say, a single planet, and overlay them as layer three. That is, each patch in layer three has its own texture and there is no explicit cube present in this layer at all. Nebulae and stars are pre-rendered using POVRay, as already stated, but for planets we wanted to be able to use arbitrary bitmaps. Since these bitmaps have to be distorted in order to appear correctly when rendered as patches placed in the cubic panorama, we adopted the following approach: Bitmaps are placed in space at a location where they should be seen as defined when looking at them up-front. Then, they are back-projected onto the panoramic cube, effectively resampling them onto the sampling grid of at most three cube sides each. This process yields a high-resolution environment map with non-black pixels only at certain patches. We then proceed by automatically extracting four-connected areas, calculating their bounding rectangles and storing coordinates for these newly found patches. Thus, we automatically generate a texture-mapped object consisting of rectangular patches that lie on a cube’s surface, together with the corresponding – and correctly distorted – textures. See Figure 2.


 figure 2
Figure 2: Detail object. Left: original. Right: texture warped by perspective back-projection.

Note that it is impossible to distinguish these three layers after they have been composited, since there occurs no motion parallax between them. For the viewer, a panoramic background rendered in this way just looks like a single spherical panorama, albeit with different resolutions for different areas (solid angles).

3.1.4 Panoramic layer four: Rotationally invariant detail objects

Optionally, there may be a fourth panoramic layer used for detail objects that should not rotate in themselves when the camera is rotated. This is useful for textures depicting bright objects with glare [Spen97] around them, since flare lines are due to the human visual system and are not related to some effectively non-existent rotation of a very far away, spherical light-source. If such detail objects are rendered in layer three, pre-calculated flare lines will rotate when the camera is rotated, which is quite unnatural. Layer four is not related to the cubic panorama, it consists of points distributed randomly on the surface of a sphere. These points are transformed correctly when the camera is rotated. Then, textures are placed at each point – the point’s projected coordinates used as center. Thus, the location of such detail objects is correct, but their textures don’t rotate.
    This is the only background-layer used in our software-renderer, to enable the use of simple bitmap blitting instead of texture-mapping for detail objects.


 figure 3
Figure 3: Panoramic background as seen from the cockpit

3.2 Polygonal objects

After the panoramic background has filled the entire screen, as described in the preceding section, polygonal objects are rendered. These are, for example, spacecraft, extras that can be collected (e.g., energy-boosts), and laser- and missile-objects.
    Our software-renderer uses object-local BSP trees to determine correct visibility for each object, and depth-sorting to establish an inter-object drawing order. If the particle system is enabled, polygons are rendered in two independent passes. In one pass, they are rendered into the frame-buffer. In the second pass, they are rendered into the z-buffer. Note that the z-buffer is only written to during this pass, it need never be checked. The reason for this approach is that particle visibility cannot easily be accommodated by the rendering pipeline without a z-buffer or extensive sorting overhead. Nevertheless, software z-buffering all polygonal objects is much too slow for achieving the desired frame rates. Thus, we use this hybrid approach of rendering objects using BSP trees and particles using the z-buffer. Naturally, though, these two visibility determination concepts cannot be entirely independent of one another. Prior to the rendering of particles, the z-buffer has to contain valid depth-values according to visible polygons. This is achieved by the second pass, rendering only a z-footprint of the entire polygonal part of the scene. Ideally, the texture-mapper would fill the z-buffer simultaneously to rendering polygons, but since we are targeting the Intel 80x86 architecture where registers are extremely scarce, it proved to be more efficient to decouple these two tasks and optimize both routines separately for maximum performance. This is also sensible with respect to instruction-cache considerations. Polygons are converted into span-lists only once and then used in both passes, first rendering into the frame-buffer and then into the z-buffer.
    When hardware acceleration is available there is no performance-related problem with using the z-buffer for visibility determination. Therefore we will probably use hardware z-buffering without employing BSP trees in the future, if the hardware-accelerated rendering-subsystem is activated. For the time being, though, we are using BSP trees even if a hardware z-buffer is available. First, all objects in Parsec are BSP-compiled anyway, and since the user can switch between software- and hardware-rendering on-the-fly we don’t have two different sets of object data. Second, this helps us avoid problems with z-aliasing, because our hardware-renderer only supports 16-bit z-buffering.

3.3 Particles

After outer-space background and polygonal objects have been rendered and the z-buffer contains a correct footprint of the entire scene, particles can be rendered in any order. Basically, particles are just points in three-space where semi-transparent texture-maps will be placed by the renderer, i.e., particles are rendered as billboards. This yields two major ways how particles in Parsec can be animated over time: First, a particle’s position can be animated, say, along a certain trajectory. This animation is determined by the type of particle system this particle belongs to. Second, the particle itself, that is, its texture, can be animated. These two different kinds of particle animation can also be seen as particle behavior and animation of particle appearance, respectively. The visual appearance of particles is controlled via particle definitions. These definitions can be used to animate the texture of a particle by using multiple animation frames and attached two-dimensional transformations. To achieve maximum flexibility, animation is controlled entirely by tables. There may be two different tables for each particle definition. One table contains a sequence of particle textures and the other table contains a sequence of 2x2 transformation matrices. Playback speed of texture and transformation animation may be chosen separately. Additionally, animation playback may be one-shot or looped. An example for a particle animation in Parsec utilizing these capabilities is the rendering of propulsion fumes. The textures of particles that are created along the trajectory of missiles rotate automatically and are also animated to depict decrease in "fume-density". Much more important to rendering most special effects are a particle’s behavioral attributes, though. These attributes determine a particle’s life-span, positional animation, and general behavior, for instance. Particles can belong to certain particle objects (mostly identical to particle systems), for example, a particle sphere depicting a spacecraft’s protective shield. Particles can be animated by moving randomly about the sphere’s surface, rotating about a common origin, moving linearly along a certain direction vector, and so on.
    Parsec uses this functionality to render special effects that cannot easily be modeled otherwise. Particles are currently used to depict explosions, weapons projectiles, protective shields, invulnerability shields, energy fields, the aforementioned propulsion fumes, and beams of lightning. If the player enables a ship’s lightning device, two jagged beams of lightning, modeled entirely with particles, shoot forth. In contrast to [Reed94], we do not connect the particles of a lightning channel with glowing lines, but just place a single lightning particle at each segment’s boundary. Explosions are actually rendered using two different techniques: First, a translucent bitmap animation is used at different spatial positions and scales. Second, a particle sphere with automatically expanding radius is created at the exploding object’s center. These particles have a relatively short life-time, destroying themselves shortly after the bitmap animation has finished. Energy fields are autonomous particle spheres (i.e., not attached to any spacecraft) that are created at random locations from time to time, expand themselves to a specified radius, and contract once again after their lifetime is spent. If the player flies through such an energy-field, the spacecraft’s energy will be boosted. Particle spheres attached to specific spacecraft are, for example, used to depict the protective shield that is activated when a ship is hit. These particle spheres are transformed along with the objects they are attached to and rotate about the object’s center. Particles are also used to visualize the points of impact of lightning beams onto a protective shield. At the points of impact small particle spheres are created, their particles moving quickly about the sphere’s surface in a random manner.
    Since particles have a constant depth-value over their entire surface, they can be rendered very quickly using the (already filled) z-buffer to resolve visibility. Previous z-values have to be checked only against this constant depth-value, i.e., no depth-values need be interpolated. Because of the simplicity of this approach, it is perfectly feasible for software-rendering. Note that this is only true since polygons do not need to check the z-buffer, as detailed in the preceding section.


figure 4
Figure 4: Spacecraft with lightning device activated, energy fields in the background

3.4 Additional special effects

Last, there is the possibility to render additional special effects using whatever technique is appropriate. Some things we do in Parsec after everything else has already been rendered are, for example, a simple lens-flare effect and fading the entire screen to a certain color, say, blue if the player is hit. We currently model lens-flare using differently sized and colored hexagonal transparency-textures. A single ray is traced to the light-source (a sun) and this ray’s points of intersection with the lenses of an arbitrary system of lenses are used to infer positions for these textures.

4. Implementation

Our current implementation of Parsec natively supports MS-DOS and Win32 (Windows 95/NT with DirectX 3/5) as host platforms. The rendering subsystem can be changed on-the-fly between pure software rendering and hardware-accelerated rendering on boards equipped with the Voodoo Graphics accelerator by 3Dfx Interactive, using their proprietary Glide API [Glide]. Our software renderer mostly consists of hand-optimized 80x86 assembly language code. The perspective-correct texture-mapper supports subpixel and subtexel precision and linearly interpolates texture coordinates for 16-pixel subspans. That is, the necessary homogeneous division is performed only every 16 pixels. This approximation to the true perspective hyperbola has proved to be sufficient in order to be indistinguishable from full evaluation at every pixel. The software renderer supports resolutions ranging from 320x200 up to 1280x1024, 800x600 being smooth enough for satisfying gameplay on mid-range desktop PCs. We also support different color-depths: 8-bit color in palettized mode, hicolor using 16 bits per pixel, and true-color with 24 bits per pixel, although true-color has been incorporated for experimental reasons only. Limited transparency (alpha-blending) is supported even in palettized modes. With hardware-rendering we currently support 640x480 resolution and a color-depth of 16 bits per pixel. Alpha-blending is essential to displaying the command-console which can be overlaid upon the screen at any time, in order to view and alter various engine parameters.
    In order to achieve high platform portability, the entire code-base is structured in subsystems, only the system-dependent of which need to be implemented anew for each new platform, as well as making the aforementioned on-the-fly switching between software and hardware rendering-subsystems possible. The current code-base is comprised of about 15,000 lines 80x86 assembly language code (most of which is the software-renderer) and about 50,000 lines of code written in a C-like subset of C++. Since Parsec is a multiplayer-only game, we support the IPX protocol for network gameplay on local area networks.

5. Conclusions and Future work

We have reviewed how Parsec, the three-dimensional multiplayer space-game we are currently developing, renders each frame using multiple layers. This is especially important for achieving convincing realism and still being able to maintain high frame update rates. We have described rendering of layers in exactly the same order as each frame is composited from back to front. We have emphasized the ideas employed to render a panoramic outer-space background. Furthermore, we have looked at how various special effects are rendered, mostly using the integrated particle system.
    In the future we would like to supersede the IPX network code by TCP/IP and internet-gameplay. For this we are going to develop a network-server where players will be able to log in and join games hosted by others already playing. We would like to make the network-architecture scaleable via dead-reckoning and other techniques used in virtual environments. We would also like to incorporate more colorful particle effects and additional special effects, since 16bpp rendering is not performance-constrained anymore, due to contemporary hardware accelerators. Host-platform support is also likely to increase. We are also working on the integration of smooth (progressive) levels of detail technology to make rendering huge carriers possible in the future. Last but not least, we are currently working on many new objects exploiting 16bpp colors and higher polygon counts, since our old spacecraft have been designed for a single 8-bit palette.
    Future improvements and other news will be announced on the Parsec Homepage [Parsec]. As soon as we are ready to release them, we will also make self-running demos and fully playable versions for all supported host platforms freely available for download on this web-page.

6. Acknowledgments

I would like to thank Clemens Beer and Andreas Varga for their help with creating the panoramic background images and their ongoing work on the Parsec project. Stefan Poiss composed the music and created the sound-effects. Alex Mastny designed the space-ship that can be seen in the screen-shots. Alexander Fuchs rendered particle textures and background planets. Zsolt Szalavari helped me with preparing the paper and talk.

7. References

8. Screenshots

Here you can view some screenshots of the Voodoo Graphics version of Parsec that appear only in the HTML version of this paper. Click on the images to view them in their original 640x480 resolution.
screenshot 1 screenshot 2
Parsec Screenshot #1 Parsec Screenshot #2
Close-up view of a firebird. 
An object-camera is available for every spacecraft.
Chasing a firebird trying to escape. 
A sun as light-source causes lens-flare.
screenshot 3 screenshot 4
Parsec Screenshot #3 Parsec Screenshot #4
In-flight shot of a firebird with its 
lightning-device activated.
A firebird is destroyed by beams of lightning. 
The protective shield is activated as soon as the beams hit.
screenshot 5 screenshot 6
Parsec Screenshot #5 Parsec Screenshot #6
A humming-bird starting to explode. 
Semi-transparent textures are overlaid multiple times.
Being in the midst of a particle explosion.
screenshot 7 screenshot 8
Parsec Screenshot #7 Parsec Screenshot #8
Scavenger surrounded by invulnerability shield 
rendered using a particle-sphere. Planet in the background.
Flying right through an energy field.