ANN: ndtypes/xnd 0.2.0b1 released

Previous Topic Next Topic
classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view

ANN: ndtypes/xnd 0.2.0b1 released

Stefan Krah-3


this is the first release (beta) of two Python modules for describing and
manipulating unboxed in-memory data.


ndtypes is based on the libndtypes C-library.  libndtypes has four main

   1) Typing a memory block (abstract type).

   2) Storing the exact layout information of a memory block (concrete type).

   3) Constructing the type from a close variant of the datashape type language.

   3) Dynamic type checking of memory blocks by pattern-matching an abstract
      function signature against the types of concrete input arguments.

ndtypes implements all the heavy-lifting for typing and handling memory blocks
and is meant to be used as a foundation for other modules.


xnd is based on the libxnd C-library and the ndtypes module. It defines and
manages typed memory blocks.  Because libndtypes has full layout information,
the xnd memory block is a small C-struct that contains bitmaps for optional
data, a linear index for addressing bitmaps, the type and the data pointer.

This makes it easy to implement lightweight memory views that support the
whole range of C types relevant for scientific computing.

Since libndtypes uses a standard algebraic data type to store recursive
type and layout information, mapping and unpacking nested Python values
directly to memory is straightforward.

Unless the type contains explicit pointer types like references, bytes or
strings, a single, pointer-free memory block is allocated.

List of xnd's features

   1) Missing values (NA) are supported for all data. The bitmaps are
      Arrow compatible.

   2) ndarrays with arbitrarily nested element types and optional values.

   3) Ragged arrays using offset addressing with nested element types and
      optional values. The ragged array type is compatible with the Arrow
      list type.

   4) Slicing and multi-dimensional indexing. Slicing and indexing returns
      fully typed views on the data.

   5) First class tuples and records with detailed explicit alignment
      and packing.

   6) Constructor types (Joule(float32) != Ampere(float32)).

   7) Reference types for compatibility with other memory block libraries.

   8) Buffer protocol support for importing other memory blocks while
      translating PEP-3118 format strings to datashape.



815455812cdeb0f8bd6c6d2111aca1c3ed23760d3c5a62dd45de20b37e01c0b1  ndtypes-0.2.0b1.tar.gz
fbfc4060fd95057d9882ae281b7d2b1b102283b8459ef5bfd8d4ffb21200d8df  xnd-0.2.0b1.tar.gz

Stefan Krah


        Support the Python Software Foundation: