GPGPU: General Purpose Computing on Graphics Processing Units

Global illumination on the GPU
Outside of physics, the coolest GPGPU application would have to be global illumination. Although there are countless game engines in the market place, they all use the same core rendering algorithm – rasterisation. Rasterisation is incredibly fast and deserves full credit for bringing about interactive graphics. However, for photorealism, it is an awful hack.

Raytracing, radiosity and photon mapping are fundamentally different algorithms for rendering graphics. Falling under the umbrella term global illumination, these algorithms actually model the way light behaves as it bounces around the world. For example, in photon mapping, the rendering begins with each light source shooting out a number of photons. These are then allowed to bounce around the environment, leaving a portion of their energy behind at each bounce. When every photon has exhausted its energy, the scene is lit and the rendering is complete.

Global illumination (GI) produces luscious images at agonisingly slow speeds. So far, only clusters have managed to run GI scenes at interactive speeds. A research group at Saarland University in Germany produced a raytraced Quake 3 engine that ran on a cluster of 20 Athlon XP 1800s at 20 frames per second. Similar groups have adopted other GI algorithms on SGI and Linux clusters.

Because global illumination is so sought after, efforts to port it to the GPU have been under way since the GeForce 3. Raytracing, radiosity and photon mapping have all been made to run on the GPU with varying degrees of quality. So far, few have managed to run at interactive speeds. A scene with about 10,000 elements can be rendered in realtime using radiosity on the GPU. Photon mapping is a bit slower, taking a few seconds to render each frame.

The good news is that most of the work was done on early hardware like the GeForce FX 5900. No paper has mentioned SLI acceleration. With next generation graphics processors, especially in dual or quad configurations, we can expect to see the first interactive global illumination demos on the GPU.

Vista, the ultimate GPGPU platform
Just how far will GPGPU go? A breakthrough is due early next year with the release of Microsoft’s next operating system, Windows Vista. This, perhaps more than any other technology, will bring GPGPU out of academic circles and into the mainstream.

Windows Vista will expose the GPU for all applications to use. Currently the GPU is viewed as a single resource. Vista will represent the GPU as a virtual device that can be time-shared by many applications. Like current CPUs, the graphics processor will then handle multiple threads and switch between them to give the illusion of running multiple programs.

Another major change is that the graphics processor’s memory will be fully virtualised. This means the graphics card will no longer be limited to its onboard memory but gain full access to the system memory. For GPGPU applications, this will allow much larger data structures. For games this will mean much richer content.

Windows Vista will introduce Direct3D 10. Current GPGPU applications spend a lot of their time working around various quirks in the hardware. Basic issues such as integer number support, GPU sorting and data structures take up a huge amount of processing time. However, with Direct3D 10, some of these problems simply disappear.

Direct3D 10 presents a unified shading language for all shading processors. Instruction limits for shader programs are abolished. A new stream output function allows results to be written into memory without resorting to pixel processors. And full integer support is added. These changes remove much of the current burdens in GPGPU programs.

With the release of Windows Vista and Direct3D 10 hardware you can expect to see a host of desktop applications tap into the power of graphics processors. Every program that is currently taking advantage of the SIMD unit of the CPU should enjoy a significant speed-up when accelerated by the GPU.

A potential killer app is a GPU accelerated version of Photoshop. Currently, the filters in Photoshop are essentially pixel shaders. But as they run on the CPU, each filter takes a few seconds to execute. Using the graphics processor, they can run at interactive speeds. This would also have the benefit of sending Photoshop benchmarks to the grave.

Media encoding is also a perfect match for the GPU. Most audio and video encoders today use some form of SSE optimisation. As graphics processors provide more SIMD power than CPUs, GPU media encoders should be lightening fast.

A very general future
Today, we are still in the infancy of GPGPU development. With Windows Vista, the graphics processor’s prestige has been elevated to a new level. It is undeniable that CPUs have caught up immensely in recent months. The new multi-core approach will allow CPUs to scale the number of cores as GPUs have scaled the number of pipelines. This will close the transistor gap between the two processors. But the floating point gap will remain, and with it, the motivation for GPGPU.

The next generation of graphics processors will bring unprecedented freedom to GPGPU programs. The CPU may have declined, but the GPU lives on!