SBLC and minSSIM Artifact

Artifact for Structural-Similarity-Preserving Lossy Data Compression for CPUs and GPUs. It builds the SBLC compressor (serial, OpenMP, and CUDA) and its baseline, runs them on six SDRBench single-precision datasets at three NOA error bounds, and produces filled_artifact.pdf containing the results table and the compression-ratio vs. throughput graphs.

Requirements

• Compilers: g++ with OpenMP, and nvcc. GPU binaries are built with -arch=sm_$(NV_SM) (e.g. 70 for Compute 7.0)
• Python 3 with requests, pandas, and numpy
• LaTeX: pdflatex with IEEEtran and pgfplots
• Resources: NVIDIA GPU (compute capability 7.0+), ~30 GB of network download, and ~65 GB of free disk, approx. 1 day

Running everything

python3 run_all.py

You can also run the individual steps:

1. python3 download_inputs.py - downloads and unpacks the SDRBench inputs into inputs/. Download time depends on connection speed
2. python3 run_experiment.py - builds all executables via make all and runs the baseline, SBLC GPU, and SBLC CPU (serial + OpenMP) experiments
3. python3 generate_latex.py - aggregates the CSVs, fills the LaTeX template src/empty_artifact.tex, and compiles filled_artifact.pdf with the results table and the two graphs similar to how they are in the paper

Manual usage of the compressor

make all [NV_SM=<cc>]
./sblc_compress_<ser|omp|gpu> input_file compressed_file eb ssim_b d1 d2 [d3]   # dims in row-major order (slowest first)
./sblc_decompress_<ser|omp|gpu> compressed_file decompressed_file
./float_analysis original_file decompressed_file d1 d2 [d3]   # dims in row-major order (slowest first)

eb is the absolute (not NOA) error bound and ssim_b the minSSIM target (a value between 0 and 1).

Publication

If you use minSSIM or SBLC in your work, please cite the following publication:

Alex Fallin and Martin Burtscher. "Structural-Similarity-Preserving Lossy Data Compression for CPUs and GPUs." Proceedings of the 30th Annual IEEE High-Performance Extreme Computing Conference. September 2026. [paper]

This work has been supported by the U.S. National Science Foundation (NSF) under Award CCF-2403380, by the Department of Energy (DOE), Office of Science, Advanced Scientific Computing Research (ASCR) under Award DE-SC0022223, and by an equipment donation from NVIDIA Corporation.