testbed/astropy__astropy/docs/units/performance.inc.rst

.. note that if this is changed from the default approach of using an *include*
   (in index.rst) to a separate performance page, the header needs to be changed
   from === to ***, the filename extension needs to be changed from .inc.rst to
   .rst, and a link needs to be added in the subpackage toctree

.. _astropy-units-performance:

Performance Tips
================

If you are attaching units to arrays to make |Quantity| objects, multiplying
arrays by units will result in the array being copied in memory, which will slow
things down. Furthermore, if you are multiplying an array by a composite unit,
the array will be copied for each individual multiplication - thus, in the following
case, the array is copied four successive times::

    In [1]: array = np.random.random(10000000)

    In [2]: %timeit array * u.m / u.s / u.kg / u.sr
    92.5 ms ± 2.52 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

There are several ways to speed this up. First, when you are using composite
units, you should make sure that the entire unit gets evaluated first, then
attached to the array. You can do this by using parentheses as for any other
operation::

    In [3]: %timeit array * (u.m / u.s / u.kg / u.sr)
    21.5 ms ± 886 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In this case, this has sped things up by a factor of 4x. If you
use a composite unit several times in your code, another approach is to create a
constant at the top of your code for this unit and use it subsequently::

    In [4]: UNIT_MSKGSR = u.m / u.s / u.kg / u.sr

    In [5]: %timeit array * UNIT_MSKGSR
    22.2 ms ± 551 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In this case and the case with brackets, the array is still copied once when
creating the |Quantity|. If you want to avoid any copies altogether, you can
make use of the ``<<`` operator to attach the unit to the array::

    In [6]: %timeit array << u.m / u.s / u.kg / u.sr
    47.1 µs ± 5.77 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

Note that these are now **microseconds**, so this is 2000x faster than the
original case with no brackets. Note that brackets are not needed when using
``<<`` since ``*`` and ``/`` have a higher precedence, so the unit will be
evaluated first. When using ``<<``, be aware that because the data is not being
copied, changing the original array will also change the |Quantity| object.

Note that for composite units, you will definitely see an
impact if you can pre-compute the composite unit::

    In [7]: %timeit array << UNIT_MSKGSR
    6.51 µs ± 112 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)

Which is over 10000x faster than the original example. See
:ref:`astropy-units-quantity-no-copy` for more details about the ``<<``
operator.