Pyperf

------------------------

Major change: perf now supports benchmark suites. A benchmark suite is made
of multiple benchmarks. perf commands now accepts benchmark suites as well.

New features:

* New ``convert`` command
* Add new command line options to TextRunner:

* ``--fast``, ``--rigorous``
* ``--hist``, ``--stats``
* ``--json-append``
* ``--quiet``

Changes:

* Remove ``--max-time`` option of TextRunner
* Replace ``--raw`` option with ``--worker``
* Replace ``--json`` with ``--stdout``
* Replace ``--json-file`` with ``--json``
* New ``perf convert`` command to convert or modify a benchmark suite
* Remove ``perf hist_scipy`` command, replaced with an example in the doc
* Add back "Mean +- Std dev" to the stats command
* Add get_loops() method to Benchmark
* Replace ``python3 -m perf.timeit`` (with dot) CLI with ``-m perf timeit``
(without dot)
* Add :class:`perf.BenchmarkSuite` class
* name is now mandatory: it must be a non-empty string in Benchmark
and TextRunner.
* A single JSON file can now contain multiple benchmarks
* Add a dependency to the ``six`` module
:meth:`Benchmark.add_run` now raises an exception if a sample is zero.
* Benchmark.name becomes a property and is now stored in metadata
* TextRunner now uses powers of 2, rather than powers of 10, to calibrate the
number of loops

------------------------

Changes:

* The ``hist`` command now accepts multiple files
* ``hist`` and ``hist_scipy`` commands got a new ``--bins`` option
* Replace mean with median
* Add :meth:`perf.Benchmark.median` method, remove ``Benchmark.mean()`` method
* ``Benchmark.get_metadata()`` method removed: use directly the
:attr:`perf.Benchmark.metadata` attribute
* Add ``timer`` metadata. ``python_version`` now also contains the architecture
(32 or 64 bits).

------------------------

New features:

* New ``hist`` and ``hist_scipy`` commands: display an histogram (text or
graphical mode)
* New ``stats`` command: display statistics on a benchmark result
* New ``--affinity=CPU_LIST`` command line option
* Emit a warning or an error in English if the standard deviation is larger
than 10% and/or the shortest sample is shorter than 1 ms
* Emit a warning or an error if the shortest sample took less than 1 ms
* Add ``perf_version``, ``duration`` metadata. Moreover, the ``date`` metadata
is now displayed.

API:

* The API deeply changed to minimize duplications of data and make the JSON
files more compact

Changes:

* The command line interface also changed. For example, ``perf.metadata``
command becomes ``perf metadata``.
* On Python 2, ``psutil`` optional dependency is now used for CPU affinity.
It ensures that CPU affinity is set for loop calibration too.
* On Python 2, add dependency to the backported ``statistics`` module
* ``perf.mean()`` and ``perf.stdev()`` functions have been removed: use
the ``statistics`` module (which is available on Python 2.7 and Python 3)
* New optional dependency on ``boltons`` (``boltons.statsutils``) to compute
even more statistics in the ``stats`` and ``hist_scipy`` commands

------------------------

* Add ``compare`` and ``compare_to`` commands to the ``-m perf`` CLI
* TextRunner is now able to spawn child processes, parse command arguments
and more features
* If TextRunner detects isolated CPUs, it sets automatically the CPU affinity
to these isolated CPUs
* Add ``--json-file`` command line option
* Add :meth:`TextRunner.bench_sample_func` method
* Add examples of the API to the documentation. Split also the documentation
into subpages.
* Add metadata ``cpu_affinity``
* Add :func:`perf.is_significant` function
* Move metadata from :class:`~perf.Benchmark` to ``RunResult``
* Rename the ``Results`` class to :class:`~perf.Benchmark`
* Add :attr:`~TextRunner.inner_loops` attribute to
:class:`TextRunner`, used for microbenchmarks when an
instruction is manually duplicated multiple times

------------------------

* use JSON to exchange results between processes
* new ``python3 -m perf`` CLI
* new :class:`TextRunner` class
* huge enhancement of the timeit module
* timeit has a better output format in verbose mode and now also supports a
``-vv`` (very verbose) mode. Minimum and maximum are not more shown in
verbose module, only in very verbose mode.
* metadata: add ``python_implementation`` and ``aslr``

------------------------

* First public release