LOSTCIRCUITS

Floating Point Pixel Pipelines

Whatever is left after the Z-occlusion culling as visible surface has to be processed in one of the eight pixel pipelines of the R300. If eight pipelines with one texture unit each is not enough, the real significance of the R300 pipelines is the transition from an integer engine to floating point processing as part of the DX9 specifications that, coincidentally, were co-developed by ATI.

Eight 128 bit precision pixel pipelines in parallel add colors to the individual pixels

The most significant achievement of the R300 core is the move from a 32 bit signed integer pixel pipeline to a 128 bit precision floating point pixel pipeline. Conventional pixel pipelines are either 16 or 32 bit integer engines with the latter mostly configured as 8 bit precision for RGBA (red, green, blue, alpha transparency), translating into 256 levels for each channel. The result is 16.7 million colors with 256 levels of transparency and 256 levels of brightness. The main difference between any integer engine and a floating point engine is that integer operations are binary functions, meaning that 8 bit correspond to 256 levels, 10 bit to 1024 and so on.

Floating point execution units, on the contrary, use logarithmic values, that is, a 32 bit FPUnit will output values between 10^-32 and +/-10³² and, furthermore, be faster since operations like divisions can be carried out on the fly. One example would be "Red = 2/3 of Blue" which cannot be done in an integer unit.

Pixel Shader-Insider

Next Page:    => The Dynamics of Light and Pipelines =>