npio is a C library for reading and writing .npy files.

It supports arrays with a single data type, i.e., not structured arrays.

To communicate data with Python without adding a large dependency, while eliminating the risk of silly mistakes that can happen when sharing raw data files with no encoded metadata.

For what has been done, see the CHANGELOG.

• Always saves arrays with 'fortran_order': False, i.e. up to the user to navigate around this if it is a problem.

• Ignores the byte order. The code is only run on little endian machines (AARCH64, x86_64) (to do).

• Decide for a better way to communicate errors without relying on stdout.

• Facilitate memory mapping of the data.

• Load data to a user-provided buffer.

• NPZ support using libzip.

• Supported nested arrays, and for the same reason there is no point in supporting Numpy files of version 2.0 ( > 65535 bytes header length).

Start with

#include <npio.h>

Open an npy file and print the some info to stdout:

npio_t * np = npio_load(filename);
if(np)
{
   npio_print(stdout, np);
   npio_free(np);
}

Write an array to a file:

npio_write("array.npy",
           ndim, &shape,
           array,
           NPIO_F32, NPIO_F32);

To build and install as a shared library, use the CMakeLists.txt:

mkdir build
cd build
cmake ../
make
sudo make install

Alternatively copy and add npio.h, npio.c and npio_config.h to your project.

Some self-tests can be run with ./npio --unittest. A few more cases are covered by the Python script npio_test_suite.py.

• NEP 1 — A Simple File Format for NumPy Arrays
• numpy/format.py