| 摘要 |
In accordance with the present invention, the rate of change of texture addresses when mapped to individual pixels of a polygon is used to obtain the correct level of detail (LOD) map from a set of prefiltered maps. The method comprises a first determination of perspectively correct texture address values found at four corners of a predefined span or grid of pixels. Then, a linear interpolation technique is implemented to calculate a rate of change of texture addresses for pixels between the perspectively bound span corners. This linear interpolation technique is performed in both screen directions to thereby create a level of detail value for each pixel. The YUV formats described above have Y components for every pixel sample, and UN (they are also named Cr and Cb) components for every fourth sample. Every UN sample coincides with four (2x2) Y samples. This is identical to the organization of texels in U.S. Pat. No. 4,965,745 "YIQ-Based Color Cell Texturing", incorporated herein by reference. The improvement of this algorithm is that a single 32-bit word contains four packed Y values, one value each for U and V, and optionally four one-bit Alpha components: YUV_0566: 5-bits each of four Y values, 6-bits each for U and V YUV_1544: 5-bits each of four Y values, 4-bits each for U and V, four 1-bit Alphas These components are converted from 4-, 5-, or 6-bit values to 8-bit values by the concept of color promotion. The reconstructed texels consist of Y components for every texel, and UN components repeated for every block of 2x2 texels. The combination of the YIQ-Based Color Cell Texturing concept, the packing of components into convenient 32-bit words, and color promoting the components to 8-bit values yields a compression from 96 bits down to 32 bits, or 3:1. There is a similarity between the trilinear filtering equation (performing bilinear filtering of four samples at each of two LODs, then linearly filtering those two results) and the motion compensation filtering equation (performing bilinear filtering of four samples from each of a "previous picture" and a "future picture", then averaging those two results). Thus some of the texture filtering hardware can do double duty and perform the motion compensation filtering when those primitives are sent through the pipeline. The palette RAM area is conveniently used to store correction data (used to "correct" the predicted images that fall between the "I" images in an MPEG data stream) since, during motion compensation the texture palette memory would otherwise be unused.
|
