GPGPU (General-Purpose computation on GPU) approaches such as NVIDIA’s CUDA have become increasingly popular the last few years, recently coming full-circle with various non-traditional rendering algorithms (perhaps this should be called GPGPUG?). However, the existing solutions are vendor-specific, often requiring reprogramming even for different GPUs from the same vendor. They also tend not to “play well” with the traditional graphics pipeline. For example, on GeForce 8000-series GPUs using CUDA there is a large delay when switching between CUDA and traditional graphics rendering.
Direct3D 11 introduces a new kind of shader called a Compute Shader. A compute shader is invoked as a regular array of threads. The threads are divided into groups. Each group has 32KB of memory shared among the threads in the group. Thus the threads can use partial results computed by other threads in the same group, improving performance. Threads can also perform random-access reads and writes to graphics resources such as textures, vertex arrays or render targets. These memory accesses are unordered, although various synchronization instructions exist to impose ordering when needed.
Pixel shaders can also perform random-access (unordered) writes. This allows them to write data structures such as linked lists that can then be processed by a compute shader, or vice-versa (pixel shaders have always had the ability to perform random access reads via texture lookups).
Several examples of compute shaders were shown at Gamefest, performing post-process operations such as finding the average luminance of a render target, or computing a luminance histogram (both used in tone mapping). For these operations, a 2X speedup was quoted over the best performance possible using pixel shaders.
Compute shaders can also perform operations such as computing summed-area tables and fast-Fourier transforms significantly faster than traditional GPU methods. Microsoft is looking into providing library functions to perform such operations.
Microsoft speculated that algorithms such as A-buffer rendering and ray tracing could also be performed efficiently, but they don’t have any hard performance numbers for those.