Authors: Jiajun Huang (University of South Florida)
Abstract: Existing fast GPU error-bounded lossy compressors achieve high throughput with pure-GPU single-kernel designs, but their compression ratios remain limited because they apply a fixed first-order predictor on independent blocks. We propose FSZ, a GPU error-bounded lossy compressor that redesigns the prediction stage with three innovations within a single CUDA kernel: (1) cross-block prediction state carries Lorenzo state across block boundaries within 256-element tiles, eliminating seven of eight boundary residuals; (2) per-tile adaptive multi-order prediction and centering selects the best strategy for each tile; and (3) a single-pass four-way evaluation exploits a property of finite differences to evaluate all variants from one data read, within a fixed predictor's bandwidth budget. On NVIDIA GH200 with eight real-world datasets, FSZ outperforms cuSZp-P by up to 10.95x and cuSZp-O by up to 2.92x in compression ratio while achieving the highest average throughput (676 GB/s compression, 785 GB/s decompression) among all evaluated compressors.
Presentation: file