Improved Encoding for Compressed Textures

For the past few decades, graphics hardware has supported mapping a two dimensional image, or texture, onto a three dimensional surface to add detail during rendering. The complexity of modern applications using interactive graphics hardware have created an explosion of the amount of data needed to represent these images. In order to alleviate the amount of memory required to store and transmit textures, graphics hardware manufacturers have introduced hardware decompression units into the texturing pipeline. Textures may now be stored as compressed in memory and decoded at run-time in order to access the pixel data. In order to encode images to be used with these hardware features, many compression algorithms are run offline as a preprocessing step, often times the most time-consuming step in the asset preparation pipeline. This research presents several techniques to quickly serve compressed texture data. With the goal of interactive compression rates while maintaining compression quality, three algorithms are presented in the class of endpoint compression formats. The first uses intensity dilation to estimate compression parameters for low-frequency signal-modulated compressed textures and offers up to a 3X improvement in compression speed. The second, FasTC, shows that by estimating the final compression parameters, partition-based formats can choose an approximate partitioning and offer orders of magnitude faster encoding speed. The third, SegTC, shows additional improvement over selecting a partitioning by using a global segmentation to find the boundaries between image features. This segmentation offers an additional 2X improvement over FasTC while maintaining similar compressed quality. Also presented is a case study in using texture compression to benefit two dimensional concave path rendering. Compressing pixel coverage textures used for compositing yields both an increase in rendering speed and a decrease in storage overhead. Additionally an algorithm is presented that uses a single layer of indirection to adaptively select the block size compressed for each texture, giving a 2X increase in compression ratio for textures of mixed detail. Finally, a texture storage representation that is decoded at runtime on the GPU is presented. The decoded texture is still compressed for graphics hardware but uses 2X fewer bytes for storage and network bandwidth.Doctor of Philosoph