cuSZp Documentation

cuSZp is an ultra-fast, error-bounded lossy compressor for single- and double-precision floating-point arrays resident in NVIDIA GPU memory. It combines quantization, dimension-aware delta operation, lightweight fixed-length encoding, and output assembly in a small number of GPU kernels.

The current release provides:

  • 1D, 2D, and 3D processing;

  • float (f32) and double (f64) data;

  • fixed, plain, and outlier encoding modes;

  • absolute error control in the library API and absolute or value-range relative error control in the command-line program; and

  • a command-line program, C/C++ APIs, and a lightweight Python binding.

Start with Installation, then use Examples for a first run. Readers choosing an encoding mode should see Algorithms; developers integrating GPU buffers should see High-level C/C++ API.

Note

cuSZp is a lossy compressor. Keep the error bound, processing dimension, data type, array shape, and encoding mode with the compressed bytes: the current low-level format is not a self-describing archival container.

Documentation contents

User guide