NeuroJSON/zmat: ZMat: A portable C-library and MATLAB/Octave/Python toolbox for zlib/gzip/lzma/lz4/zstd/blosc2 data compression · GitHub

• Copyright (C) 2019-2026 Qianqian Fang <q.fang at neu.edu>
• License: GNU General Public License version 3 (GPL v3), see License*.txt
• Version: 1.2.preview (Kylie the Opossum - preview)
• URL: https://neurojson.org/zmat
• Github: https://github.com/NeuroJSON/zmat
• PyPi: https://pypi.org/project/zmat
• Acknowledgement: This project is part of the NeuroJSON project supported by US National Institute of Health (NIH) grant U24-NS124027

• Python module — ZMat is now available as a Python C extension (pip install zmat), with pre-built wheels for Linux, macOS, and Windows. The Python API supports all compression/encoding algorithms available in the MATLAB/Octave toolbox.
• Octave 10+ compatibility — special matrices (e.g., logical, complex) are now correctly handled in GNU Octave 10 and later.
• Apple Silicon support — the blosc2 backend now automatically detects SSE2 availability on Apple Silicon (M-series) Macs.
• Updated bundled libraries:
  • C-Blosc2 updated to v2.23.1
  • Zstandard updated to v1.5.7
  • miniz updated to v3.1.1
  • LZ4 updated to v1.10.0
• New Linux MEX binaries — pre-compiled MEX files for Linux are now included in the private folder alongside the existing macOS and Windows binaries.
• Memory and safety fixes — multiple safety, memory-leak, and efficiency issues in zmatlib.c have been resolved.
• AI coding assistant Claude has been used in the development of this release.

ZMat provides both an easy-to-use C-based data compression library - libzmat, a portable mex function to enable zlib/gzip/lzma/lz4/zstd/blosc2 based data compression/decompression and base64 encoding/decoding support in MATLAB and GNU Octave, as well as a Python C extension module with the same feature set. It is fast and compact, can process a large array within a fraction of a second.

Among all the supported compression methods, or codecs, lz4 is among the fastest for both compression/decompression; lzma is the slowest but offers the highest compression ratio; zlib/gzip have excellent balance between speed and compression time; zstd can be fast, although not as fast as lz4, at the low-compression levels, and can offer higher compression ratios than zlib/gzip, although not as high as lzma, at high compression levels. We understand there are many other existing general purpose data compression algorithms. We prioritize the support of these compression algorithms because they have widespread use.

Starting in v0.9.9, we added support to a high-performance meta-compressor, called blosc2 (https://blosc.org). The blosc2 compressor is not single compression method, but a container format that supports a diverse set of strategies to losslessly compress data, especially optimized for storing and retrieving N-D numerical data of uniform binary types. Users can choose between blosc2blosclz, blosc2lz4, blosc2lz4hc, blosc2zlib and blosc2zstd, to access these blosc2 codecs.

The libzmat library, including the static library (libzmat.a) and the dynamic library libzmat.so or libzmat.dll, provides a simple interface to conveniently compress or decompress a memory buffer:

int zmat_run(
      const size_t inputsize,     /* input buffer data length */
      unsigned char *inputstr,    /* input buffer */
      size_t *outputsize,         /* output buffer data length */
      unsigned char **outputbuf,  /* output buffer */
      const int zipid,            /* 0: zlib, 1: gzip, 2: base64, 3: lzma, 4: lzip, 5: lz4, 6: lz4hc
                                     7: zstd, 8: blosc2blosclz, 9: blosc2lz4, 10: blosc2lz4hc,
                                     11: blosc2zlib, 12: blosc2zstd */
      int *status,                /* return status for error handling */
      const int clevel            /* 1 to compress (default level); 0 to decompress, -1 to -9 (-22 for zstd): setting compression level */
    );

When blosc2 codecs are used, users can set additional compression flags, including number of threads, byte-shuffling length, can be set via the flags.param interface in the following data structure, and pass on flags.iscompress as the last input to zmat_run.

union cflag {
    int iscompress;      /* combined flag used to pass on to zmat_run */
    struct settings {    /* unpacked flags */
        char clevel;     /* compression level */
        char nthread;    /* number of compression/decompression threads */
        char shuffle;    /* byte shuffle length */
        char typesize;   /* for ND-array, the byte-size for each array element */
    } param;
} flags = {0};

The zmat library is highly portable and can be directly embedded in the source code to provide maximal portability. In the test folder, we provided sample codes to call zmat_run/zmat_encode/zmat_decode for stream-level compression and decompression in C and Fortran90. The Fortran90 C-binding module can be found in the fortran90 folder.

The ZMat MATLAB function accepts 3 types of inputs: char-based strings, numerical arrays or vectors, or logical arrays/vectors. Any other input format will result in an error unless you typecast the input into int8/uint8 format. A multi-dimensional numerical array is accepted, and the original input's type/dimension info is stored in the 2nd output "info". If one calls zmat with both the encoded data (in byte vector) and the "info" structure, zmat will first decode the binary data and then restore the original input's type and size.

The pre-compiled mex binaries for MATLAB are stored inside the subfolder named private. Those precompiled for GNU Octave are stored in the subfolder named octave, with one operating system per subfolder. The PKG_ADD script should automatically select the correct mex file when one types addpath in Octave. These precompiled mex files are expected to run out-of-box across a wide-range of MATLAB (tested as old as R2008) and Octave (tested as old as v3.8).

If you do not want to compile zmat yourself, you can download the precompiled package by either clicking on the "Download ZIP" button on the above URL, or use the below git command:

git clone https://github.com/NeuroJSON/zmat.git

The installation of ZMat is no different from any other simple MATLAB toolboxes. You only need to download/unzip the package to a folder, and add the folder's path (that contains zmat.m and the "private" folder) to MATLAB's path list by using the following command:

addpath('/path/to/zmat');

For Octave, one needs to copy the zipmat.mat file inside the "octave", from the subfolder matching the OS into the "private" subfolder.

If you want to add this path permanently, you need to type "pathtool", browse to the zmat root folder and add to the list, then click "Save". Then, run "rehash" in MATLAB, and type "which zmat", if you see an output, that means ZMat is installed for MATLAB/Octave.

If you use MATLAB in a shared environment such as a Linux server, the best way to add path is to type

mkdir ~/matlab/
nano ~/matlab/startup.m

and type addpath('/path/to/zmat') in this file, save and quit the editor. MATLAB will execute this file every time it starts. For Octave, the file you need to edit is ~/.octaverc , where "~" is your home directory.

One can directly install zmat on Fedora Linux 29 or later via the below shell command

sudo dnf install octave-zmat

Similarly, the below command installs the libzmat library for developing software using this library:

sudo dnf install zmat zmat-devel zmat-static

The above command installs the dynamic library, C/Fortran90 header files and static library, respectively

Similarly, if one uses Debian (11) or Ubuntu 21.04 or newer, the command to install zmat toolbox for Octave (and optionally for MATLAB) is

sudo apt-get install octave-zmat matlab-zmat

and that for installing the development environment is

sudo apt-get install libzmat1 libzmat1-dev

A Ubuntu (16.04/18.04) user can use the same commands as Debian to install these packages but one must first run

sudo add-apt-repository ppa:fangq/ppa
sudo apt-get update

to enable the relevant PPA (personal package achieve) first.

The easiest way to install ZMat for Python is via pip:

pip install zmat

To install from source, clone the repository and build the extension:

git clone https://github.com/NeuroJSON/zmat.git
cd zmat/python
pip install .

All compression libraries (miniz, easylzma, lz4, zstd, blosc2) are embedded in the source and compiled directly into the module — no system libraries are required. The build can be customized with the following environment variables:

For example, to build without blosc2 for a smaller binary:

ZMAT_NO_BLOSC2=1 pip install .

import zmat

# Compress and decompress using the default algorithm (zlib)
data = b"Hello, ZMat! " * 1000
compressed = zmat.compress(data)
print(f"Original: {len(data)} bytes -> Compressed: {len(compressed)} bytes")

restored = zmat.decompress(compressed)
assert restored == data

# Use a different algorithm
fast      = zmat.compress(data, method='lz4')    # fastest
small     = zmat.compress(data, method='lzma')   # smallest
balanced  = zmat.compress(data, method='zstd')   # good balance

# Base64 encoding / decoding
encoded = zmat.encode(data, method='base64')
decoded = zmat.decode(encoded, method='base64')
assert decoded == data

The four main API functions are:

• zmat.compress(data, method='zlib', level=1) — compress a bytes-like object
• zmat.decompress(data, method='zlib') — decompress a bytes-like object
• zmat.encode(data, method='base64') — encode data (e.g., base64)
• zmat.decode(data, method='base64') — decode data (e.g., base64)

All functions accept bytes, bytearray, or any object supporting Python's buffer protocol (including NumPy arrays).

ZMat's zlib and gzip output is fully compatible with Python's standard library:

import zmat, zlib, gzip, io

data = b"interoperability test " * 100

# zmat -> Python zlib
assert zlib.decompress(zmat.compress(data, method='zlib')) == data

# Python zlib -> zmat
assert zmat.decompress(zlib.compress(data), method='zlib') == data

# zmat -> Python gzip
compressed = zmat.compress(data, method='gzip')
with gzip.open(io.BytesIO(compressed), 'rb') as f:
    assert f.read() == data

ZMat can capture a NumPy array's dtype, shape, and memory order alongside the compressed bytes, so the original array is restored automatically on decompression. This mirrors the [ss, info] = zmat(arr) / zmat(ss, info) pattern from the MATLAB/Octave toolbox:

import numpy as np
import zmat

arr = np.random.rand(1000, 1000)   # float64, C-contiguous

# compress — capture metadata
compressed, info = zmat.compress(arr, method='lz4', info=True)
# info: {'type': 'float64', 'shape': (1000, 1000), 'byte': 8,
#        'method': 'lz4', 'order': 'C'}

# decompress — restored to original dtype and shape
restored = zmat.decompress(compressed, info=info)
assert np.array_equal(restored, arr)

# Fortran-contiguous arrays are handled correctly too
arr_f = np.asfortranarray(np.eye(100))
compressed, info = zmat.compress(arr_f, info=True)
restored = zmat.decompress(compressed, info=info)
assert restored.flags['F_CONTIGUOUS']

For numerical arrays, blosc2 methods with byte-shuffle can achieve better compression ratios:

# blosc2 with byte-shuffle (typesize=8 for float64)
compressed, info = zmat.compress(arr, method='blosc2zstd', info=True)
restored = zmat.decompress(compressed, info=info)

The low-level zmat.zmat() interface provides full control over blosc2 multi-threading and shuffle options:

output = zmat.zmat(data, iscompress=1, method='blosc2zstd',
                   nthread=4, shuffle=1, typesize=8)

iscompress=1 compresses at the default level; iscompress=0 decompresses; a negative value (e.g., -9) sets the compression level.

ZMat provides a single mex function, zipmat.mex* -- for both compressing/encoding or decompresing/decoding data streams. The help info of the function is shown below

output=zmat(input)
   or
[output, info]=zmat(input, iscompress, method)
[output, info]=zmat(input, iscompress, method, options ...)
output=zmat(input, info)

A portable data compression/decompression toolbox for MATLAB/GNU Octave

author: Qianqian Fang <q.fang at neu.edu>
initial version created on 04/30/2019

input:
     input: a char, non-complex numeric or logical vector or array
     iscompress: (optional) if iscompress is 1, zmat compresses/encodes the input,
            if 0, it decompresses/decodes the input. Default value is 1.

            if iscompress is set to a negative integer, (-iscompress) specifies
            the compression level. For zlib/gzip, default level is 6 (1-9); for
            lzma/lzip, default level is 5 (1-9); for lz4hc, default level is 8 (1-16).
            the default compression level is used if iscompress is set to 1.

            zmat removes the trailing newline when iscompress=2 and method='base64'
            all newlines are kept when iscompress=3 and method='base64'

            if one defines iscompress as the info struct (2nd output of zmat), zmat
            will perform a decoding/decompression operation and recover the original
            input using the info stored in the info structure.
     method: (optional) compression method, currently, zmat supports the below methods
            'zlib': zlib/zip based data compression (default)
            'gzip': gzip formatted data compression
            'lzip': lzip formatted data compression
            'lzma': lzma formatted data compression
            'lz4':  lz4 formatted data compression
            'lz4hc':lz4hc (LZ4 with high-compression ratio) formatted data compression
            'zstd':  zstd formatted data compression
            'blosc2blosclz':  blosc2 meta-compressor with blosclz compression
            'blosc2lz4':  blosc2 meta-compressor with lz4 compression
            'blosc2lz4hc':  blosc2 meta-compressor with lz4hc compression
            'blosc2zlib':  blosc2 meta-compressor with zlib/zip compression
            'blosc2zstd':  blosc2 meta-compressor with zstd compression
            'base64': encode or decode use base64 format
    options: a series of ('name', value) pairs, supported options include
            'nthread': followed by an integer specifying number of threads for blosc2 meta-compressors
            'typesize': followed by an integer specifying the number of bytes per data element (used for shuffle)
            'shuffle': shuffle methods in blosc2 meta-compressor, 0 disable, 1, byte-shuffle

output:
     output: a uint8 row vector, storing the compressed or decompressed data;
            empty when an error is encountered
     info: (optional) a struct storing additional info regarding the input data, may have
           'type': the class of the input array
           'size': the dimensions of the input array
           'byte': the number of bytes per element in the input array
           'method': a copy of the 3rd input indicating the encoding method
           'status': the zlib/lzma/lz4 compression/decompression function return value,
                   including potential error codes; see documentation of the respective
                   libraries for details
           'level': a copy of the iscompress flag; if non-zero, specifying compression
                   level, see above

example:

  [ss, info]=zmat(eye(5))
  orig=zmat(ss,0)
  orig=zmat(ss,info)
  ss=char(zmat('zmat test',1,'base64'))
  orig=char(zmat(ss,0,'base64'))

-- this function is part of the zmat toolbox (https://github.com/NeuroJSON/zmat)

Under the "example" folder, you can find a demo script showing the basic utilities of ZMat. Running the "demo_zmat_basic.m" script, you can see how to compress/decompress a simple array, as well as apply base64 encoding/decoding to strings.

Please run these examples and understand how ZMat works before you use it to process your data.

Under the "c" and "f90" folders, sample C/Fortran90 units calling the compression/decompression APIs provided by zmat are also provided. You may run "make" in each of the folders to build the binary and execute the output program.

Under the "test" folder, you can run "run_zmat_test.m" script to run unit tests on the key features provided by zmat.

To recompile ZMat, you first need to check out ZMat source code, along with the needed submodules from the Github repository using the below command

git clone https://github.com/NeuroJSON/zmat.git zmat

Next, you need to make sure your system has gcc, g++, mex and mkoctfile (if compiling for Octave is needed). If not, please install gcc, MATLAB and GNU Octave and add the paths to these utilities to the system PATH environment variable.

To compile zmat, you may choose one of the three methods:

1. Method 1: please open MATLAB or Octave, and run the below commands

cd zmat/src
compilezmat

The above script utilizes the MinGW-w64 MATLAB Compiler plugin.

To install the MinGW-w64 compiler plugin for MATLAB, please follow the below steps

• If you have MATLAB R2017b or later, you may skip this step. To compile mcxlabcl in MATLAB R2017a or earlier on Windows, you must pre-install the MATLAB support for MinGW-w64 compiler https://www.mathworks.com/matlabcentral/fileexchange/52848-matlab-support-for-mingw-w64-c-c-compiler

  Note: it appears that installing the above Add On is no longer working and may give an error at the download stage. In this case, you should install MSYS2 from https://www.msys2.org/. Once you install MSYS2, run MSYS2.0 MinGW 64bit from Start menu, in the popup terminal window, type

pacman -Syu
pacman -S base-devel gcc git zlib-devel

Then, start MATLAB, and in the command window, run

setenv('MW_MINGW64_LOC','C:\msys64\usr');

• After installation of MATLAB MinGW support, you must type mex -setup C in MATLAB and select "MinGW64 Compiler (C)".
• Once you select the MingW C compiler, you should run mex -setup C++ again in MATLAB and select "MinGW64 Compiler (C++)" to compile C++.

2. Method 2: Compile with cmake (3.3 or later)

Please open a terminal, and run the below shall commands

cd zmat/src
rm -rf build
mkdir build && cd build
cmake ../
make clean
make

if MATLAB was not installed in a standard path, you may change cmake ../ to

cmake -DMatlab_ROOT_DIR=/path/to/matlab/root ../

by default, this will first compile libzmat.a and then create the .mex file that is statically linked with libzmat.a. If one prefers to create a dynamic library libzmat.so and then a dynamically linked .mex file, this can be done by

cmake -DMatlab_ROOT_DIR=/path/to/matlab/root -DSTATIC_LIB=off ../

3. Method 3: please open a terminal, and run the below shall commands

cd zmat/src
make clean
make mex

to create the mex file for MATLAB, and run make clean oct to compile the mex file for Octave.

The compiled mex files are named as zipmat.mex* under the zmat root folder. One may move those into the private folder to overwrite the existing files, or leave them in the root folder. MATLAB/Octave will use these files when zmat is called.

ZMat is an open-source project. This means you can not only use it and modify it as you wish, but also you can contribute your changes back to ZMat so that everyone else can enjoy the improvement. For anyone who want to contribute, please download ZMat source code from its source code repositories by using the following command:

git clone https://github.com/NeuroJSON/zmat.git zmat

or browsing the github site at

https://github.com/NeuroJSON/zmat

You can make changes to the files as needed. Once you are satisfied with your changes, and ready to share it with others, please submit your changes as a "pull request" on github. The project maintainer, Dr. Qianqian Fang will review the changes and choose to accept the patch.

We appreciate any suggestions and feedbacks from you. Please use the NeuroJSON forum to report any questions you may have regarding ZMat:

NeuroJSON forum

(Subscription to the mailing list is needed in order to post messages).

ZMat bundles the below open-source data compression/encoding libraries

1. miniz: https://github.com/richgel999/miniz

• Copyright (c) 2013-2014 RAD Game Tools and Valve Software
• Copyright (c) 2010-2014 Rich Geldreich and Tenacious Software LLC
• License: MIT License (https://github.com/richgel999/miniz/blob/master/LICENSE)

2. Eazylzma: https://github.com/lloyd/easylzma

• Author: Lloyd Hilaiel (lloyd)
• License: public domain

3. Original LZMA library:

• Author: Igor Pavlov
• License: public domain

4. LZ4 library: https://lz4.github.io/lz4/

• Copyright (C) 2011-2019, Yann Collet.
• License: BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)

5. C-blosc2: https://blosc.org/

• Copyright (c) 2009-2018 Francesc Alted <francesc@blosc.org>
• Copyright (c) 2019-present The Blosc Development Team <blosc@blosc.org>
• License: BSD 3-Clause License (http://www.opensource.org/licenses/bsd-license.php)

6. ZStd: https://facebook.github.io/zstd/

• Copyright (c) Meta Platforms, Inc. and affiliates. All rights reserved.
• License: BSD 3-Clause License (http://www.opensource.org/licenses/bsd-license.php)

7. base64_{encode,decode} functions in the zmat.cpp

• Copyright (c) 2005, Jouni Malinen <j@w1.fi>
• License: BSD 3-Clause License (http://www.opensource.org/licenses/bsd-license.php)
• Source http://web.mit.edu/freebsd/head/contrib/wpa/src/utils/base64.c and base64.h