v1.6.6

Release 1.6.6

• Added the ability to automatically publish to PyPI

Minor updates

This release brings minor updates:

• Support for Python 3.6 and NumPy<1.17 has been dropped
• Fixed some deprecation warnings from NumPy 1.20+.

Infrastructure improvements

This release is the result of changes in infrastructure, namely testing and documentation. The package is now tested with Github Actions.

• Fixed an issue regarding a deprecation of collections.Sized (in favour of collections.abc.Sized) in Python 3.10+
• Code snippets in documentation are now tested for correctness.
• Tests are now included in source distributions.

Explicit support for Python 3.9

This release only validates that thing are working well with Python 3.9.

Various improvements

• Added a changelog.
• Added the possibility to use weights in ihistogram.
• Added the function average_and_var to compute the average and variance in a single pass.
• Documentation regarding the ddof keyword in many statistical functions wrongly stated that the default value was 1. This has been corrected.

Explicit support for NumPy 1.16+

This release fixes some issues with NumPy versions above 1.16.

Benchmarking suite

This release includes some small tweaks with array streams, as well as full benchmarking capabilities. See the documentation for more information.

From this release onwards, only Python 3.6+ is supported.

Performance improvements and initial CUDA support

0.4

FIX: cmean

Performance increase for axis = -1

By reusing memory locations, streaming functions with axis = -1 (default behavior) can show drastic performance improvements.

For example, summing arrays of shape (2048, 2048) shows speedup of 3x on my machine.

Easier factory generator function for NumPy binary ufuncs

It is now much easier to roll up your own streaming reduction function using ireduce_ufunc.

Some routines were also added:

• imin, imax
• iinner, idot, itensordot, ieinsum

I'm probably forgetting some more.