Efficient Floating-Point Texture Decompression

Tomi Aarnio1,  Claudio Brunelli1,  Timo Viitanen2

1Nokia Research Center, 2Tampere University of Technology


We propose a novel hardware design for decoding compressed floating-point textures in a graphics processing unit (GPU). Our decoder is based on the NXR texture format, which provides lossy, fixed-rate 6:1 compression for floating-point textures. Our design exploits the constraints of the compressed pixel blocks to produce the correct output using only fixed-point arithmetic. This results in significantly lower silicon area occupation compared to pre-existing floating-point texture decoders.