Decoding JPEG XS on a GPU

Abstract

JPEG XS is an upcoming lightweight image compression standard that is especially developed to meet the requirements of compressed video-over-IP use cases. It is designed with not only CPU, FPGA or ASIC platforms in mind, but explicitly also targets GPUs. Though not yet finished, the codec is now sufficiently mature to present a first NVIDIA CUDA-based GPU decoder architecture and preliminary performance results. On a 2014 mid-range GPU with 640 cores a 12 bit UHD 4:2:2 (4:4:4) can be decoded with 54 (42) fps. The algorithm scales very well: on a 2017 high-end GPU with 2560 cores the throughput increases to 190 (150) fps. In contrast, an optimized GPU-accelerated JPEG 2000 decoder takes 2x as long for high compression ratios that yield a PSNR of 40 dB and 3x as long for lower compression ratios with a PSNR of over 50 dB.