pSZ/cuSZ

1. The user picks an error mode: absolute (abs), or relative to the data's value range (r2r), and an error bound.
2. Prediction errors are quantized in units of that bound. In-range quant-codes are stored; out-of-range ones are gathered as outliers.
3. The in-range quant-codes go through a Huffman encoder.
4. CPU-SZ adds a DEFLATE pass to exploit repeated patterns; on the GPU, cuSZ+'s RLE plays the same role.

There are real differences between CPU-SZ and pSZ/cuSZ; the papers below carry the detail.

• cuSZ · Tian et al., '20, '21

  The basic framework: an N-D Lorenzo predictor, a histogram, and Huffman encoding. A balanced compression ratio and quality.

• cuSZ+ (hi-ratio) · Tian et al., '21

  A de-redundancy RLE pass ahead of Huffman, from compressibility research showing RLE wins on highly compressible data.

• FZ-GPU · Zhang, Tian et al., '23

  Prediction plus de-redundancy only, no histogram or Huffman: it prioritizes data-processing speed.

• SZp-CUDA / GSZ · Huang et al., '23, '24

  A single-kernel 1-D predictor plus de-redundancy pass, also prioritizing data-processing speed.

• cuSZ-i · Liu, Tian, Wu et al., '24

  A spline-interpolative 3-D predictor with QoZ-like methods, for substantially better reconstruction quality and compression ratio.

• cuSZ-Hi · Wu and Pan et al., '25

  A 2-D/3-D spline predictor feeding two encoder paths, one tuned for ratio and one for throughput.

git clone --recursive https://github.com/szcompressor/cuSZ.git cusz-latest
cd cusz-latest && mkdir build && cd build
cmake .. -DPSZ_BACKEND=cuda -DPSZ_BUILD_EXAMPLES=on -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_CUDA_ARCHITECTURES="<sm>"
make -j
make install

cusz -t f32 -m r2r -e 1e-4 -i data.f32 -l 3600-1800 -z --report cr
cusz -i data.f32.cusza -x --compare data.f32 --skip write2disk

CUDAARCHS="<sm>" pip install -e py_ext --no-build-isolation

from psz import compress_init, decompress_init

with compress_init(data.shape) as c:  # data: a float32 CuPy array
    c.compress_process("lrz..", data, 1e-3 * c.compress_extrema(data).rng)
    header, archive = c.compress_archive()

with decompress_init(header) as dc:
    xdata = dc.decompress_process(archive)

Fuller instructions, including the recommended GCC and Clang combinations, live in the wiki: build and install, command-line use, API reference, and Python.

1. PACT '20cuSZ: An efficient GPU-based error-bounded lossy compression framework for scientific data The basic framework: fine-grained N-D error-controlled construction plus coarse-grained lossless encoding. ACM arXiv

2. CLUSTER '21Optimizing error-bounded lossy compression for scientific data on GPUs Throughput via fine-grained N-D reconstruction, and ratio when the data is smooth. IEEE arXiv

3. SC '24cuSZ-i: High-ratio scientific lossy compression on GPUs with optimized multi-level interpolation Spline interpolation for high-ratio compression and high-quality reconstruction, plus synergetic lossless encoding. IEEE arXiv

4. SC '25Boosting scientific error-bounded lossy compression through optimized synergistic lossy-lossless orchestration Ratio gains from the LC framework's lossless codec composition, plus HiTP and HiCR modes atop cuSZ-i. ACM arXiv

Citing pSZ/cuSZ: BibTeX entries live in doc/psz-cusz.bib.

The fields are available from SDRBench.

Supported by the Exascale Computing Project (17-SC-20-SC), a collaboration of the DOE Office of Science and the NNSA; by the U.S. Department of Energy, Office of Science, under contract DE-AC02-06CH11357; and by the National Science Foundation under grants CSSI/OAC #2104023, #2247080, OAC #2247060, CSSI/OAC #2311875, and #2311876.