pygram11

• Use scikit-build-core for packaging system
• Bump mp11 version

• Add support for Windows

• Project now uses scikit-build

• Bug fix: variance calculation for weighted histograms with density=True.
• Bump pygram11 and boost/mp11

• Bump pybind11 to 2.8.0
• Add wheels for Python 3.10
• Drop support for Python 3.6

• OpenMP configuration redesigned.
  • constants have been removed in favor of new pygram11.config module.
  • Decorators and context managers added to control toggling OpenMP thresholds
    • pygram11.without_omp decorator
    • pygram11.with_omp decorator
    • pygram11.disable_omp context manager
    • pygram11.force_omp context manager
  • See the documentation for more
• New keyword argument added to histogramming functions (cons_var) for returning variance instead of standard error.

Renamed internal Python files hist.py and misc.py to _hist.py and _misc.py, respectively.

The contents of these modules are brought in to the main pygram11 module namespace by imports in __init__.py (the submodules themselves are not meant to be part of the public API). This avoids tab completions of pygram11.hi<tab> yielding pygram11.hist when we actually want pygram11.histogram.

Two convenience functions added to the pygram11 namespace:

• bin_centers: returns an array representing the the center of histogram bins given a total number of bins and an axis range or given existing bin edges.
• bin_edges: returns an array representing bin edges given a total number of bins and an axis range.

• API change: functions calls with weights=None now return None as the second return. Previously the uncertainty was returned (which is just the square-root of the bin heights); now users can take the square-root themselves, and the back-end does not waste cycles tracking the uncertainty since it's trivial for unweighted histograms.
• More types are supported without conversions (previously np.float64 and np.float32 were the only supported array types, and we converted non-floating point input). Now signed and unsigned integer inputs (both 64 and 32 bit) are supported.
  • If unsupported array types are used TypeError is now raised. This library prioritizes performance; hidden performance hits are explicitly avoided.
• Configurable thresholds have been introduced to configure when OpenMP acceleration is used (described in the documentation).
• The back-end was refactored with some help from boost::mp11 to aid in adding more type support without introducing more boilerplate. We now vendor boost::mp11 as a submodule.
• Bumped the vendored pybind11 submodule to v2.6.2.
• C++14 now required to build from source.
• Added Apple Silicon support for Python 3.9 with libomp from Homebrew installed at /opt/homebrew.
• Documentation improvements
• Renamed master branch to main.

First release candidate for 0.11.0

• Fix GitHub action to make sure pybind11 submodule is checked out for source distribution creation.

• Vendor pybind11 2.6 series.
• Some changes to backend types (std::size_t -> pybind11::ssize_t for sizes) (https://github.com/pybind/pybind11/pull/2293)
• Improvements to CI/CD: using cibuildwheel to build wheels.
• Python 3.9 is now tested and compatible wheels are built.

Test new release triggering wheel builds.

• Fix setup.py for building binaries for conda-forge.
• Bump NumPy requirement to >=1.16 (see NEP 29)

Renamed internal Python module from histogram to hist. This avoids a clash with the module function of the same name. Some IDE features were confused.

Fixes issue with MANIFEST.in (a header was missing) to build from source.

• Repository reorganized to use PEP 517
• Python API moved outside of __init__.py to histogram.py, functions are imported in __init__.py

• Remove requirement on specific NumPy version (it's not required to build, only to use).
• Large improvements to documentation.

• Bump the bundled pybind11 version to 2.5.0.
• Renamed some backend files.

Public API changes:

• Remove previously deprecated omp function argument

Backend changed:

• moved 1D backend back to pybind11 (template flexibility)
• parallelization cutoffs consistently at array size of 5000 for all calculations
• all non-floating point inputs are converted

Other:

• Improved tests