{
  "status": "shipping",
  "advantages": [],
  "limitations": [],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-17",
  "slug": "zfp",
  "term": "zfp",
  "aliases": [
    "ZFP compression"
  ],
  "definition": "zfp is a block-based codec for numerical arrays with selectable compressed-size, precision, absolute-error and reversible modes.",
  "note": "Blocks: The lossy path partitions a d-dimensional array into blocks of 4^d values, aligns exponents, applies an integer transform and encodes coefficient bit planes. Edge blocks are padded when dimensions are not multiples of four. The current algorithm differs from the original 2014 method. [algorithm] [paper]\n\nControls: Fixed rate budgets compressed bits per block; fixed precision limits coefficient bit planes; fixed accuracy targets an absolute tolerance with data-dependent size. Precision does not specify the mantissa bits retained in every input value. [modes]\n\nLimits: A shared exponent can erase small values beside much larger ones. Fixed precision does not give a uniform pointwise relative-error guarantee; absolute tolerances below available block precision may also be unattainable. Zero tolerance is near-lossless, not reversible mode. [faq]\n\nAccounting: A native 5 by 3 binary32 example at 8 bits per padded value occupies 32 payload bytes, or 256/15 bits per original value. Two 4 by 4 blocks are encoded. This excludes headers, array metadata and caches. [probe]",
  "related": [
    "fpzip",
    "error-bounded-lossy-compression",
    "lossless-compression"
  ],
  "citations": [
    {
      "id": "paper",
      "title": "Lindstrom (2014): Fixed-Rate Compressed Floating-Point Arrays, sections 3-6; DOI 10.1109/TVCG.2014.2346458",
      "url": "https://vis.cs.ucdavis.edu/vis2014papers/TVCG/papers/2674_20tvcg12-lindstrom-2346458.pdf",
      "accessed": "2026-11-17"
    },
    {
      "id": "analysis",
      "title": "Diffenderfer et al. (2019): Error Analysis of ZFP Compression for Floating-Point Data, sections 4-6, arXiv v3",
      "url": "https://arxiv.org/abs/1805.00546v3",
      "accessed": "2026-11-17"
    },
    {
      "id": "modes",
      "title": "zfp 1.0.1: Compression modes",
      "url": "https://zfp.readthedocs.io/en/release1.0.1/modes.html",
      "accessed": "2026-11-17"
    },
    {
      "id": "faq",
      "title": "zfp 1.0.1: FAQ 17, 20 and 21, finite precision and relative error",
      "url": "https://zfp.readthedocs.io/en/release1.0.1/faq.html",
      "accessed": "2026-11-17"
    },
    {
      "id": "algorithm",
      "title": "zfp 1.0.1: block layout and lossy compression algorithm",
      "url": "https://zfp.readthedocs.io/en/release1.0.1/algorithm.html",
      "accessed": "2026-11-17"
    },
    {
      "id": "probe",
      "title": "Hesela: native zfp 1.0.1 error budgets and padded-block byte accounting",
      "url": "https://hesela.dev/research/zfp-error-budgets/README.md",
      "accessed": "2026-11-17"
    }
  ],
  "provenance": [
    "Source synthesis plus 56 constructed native API observations on zfp 1.0.1; no production compression ratio or speed estimate.",
    "The 2014 paper evaluates double-precision simulation fields; the published Hesela probe uses small binary32 arrays and does not reproduce that evaluation.",
    "The exact source revision, build settings, inputs, decoded values and compressed bytes are published with the probe."
  ]
}
