GST: GPU-decodable Supercompressed Textures

摘要

Modern GPUs supporting compressed textures allow interactive applicationrndevelopers to save scarce GPU resources such as VRAMrnand bandwidth. Compressed textures use fixed compression ratiosrnwhose lossy representations are significantly poorer quality thanrntraditional image compression formats such as JPEG. We present arnnew method in the class of supercompressed textures that providesrnan additional layer of compression to already compressed textures.rnOur texture representation is designed for endpoint compressed formatsrnsuch as DXT and PVRTC and decoding on commodity GPUs.rnWe apply our algorithm to commonly used formats by separatingrntheir representation into two parts that are processed independentlyrnand then entropy encoded. Our method preserves the CPU-GPUrnbandwidth during the decoding phase and exploits the parallelismrnof GPUs to provide up to 3X faster decode compared to prior texturernsupercompression algorithms. Along with the gains in decodingrnspeed, our method maintains both the compression size andrnquality of current state of the art texture representations.