[SciPy-Dev] Recent Numba updates of interest
Stanley Seibert
sseibert at anaconda.com
Mon Jul 9 21:56:28 EDT 2018
Hi all,
Following up from conversation in March about Numba's progress toward being
a dependency for SciPy-related projects, I wanted to give an update on some
recent progress on items of interest. Some of these items will be not be
available until Numba 0.39, which is tagged now and will be officially
released in the next day or so:
- ppc64le (POWER8 and 9) is now supported with the release of LLVM
6.0.1. (This may be of interest to those of you applying for time on the
Summit and Sierra supercomputers in the US.) Since there is no wheel
standard for ppc64le, I would recommend using the conda packages in the
numba channel on anaconda.org.
- Testing Numba on ARMv7 (Raspberry Pi) support flushed out some bugs in
NumPy on ARM (related to alignment) and OpenBLAS. NumPy 1.16 should have
the fixes that Numba needs to pass its unit tests on ARM. Once that
happens, we should have some Berryconda-compatible conda packages in the
numba channel.
- We've made a variety of changes to compiler error messages to hide
giant tracebacks, provide suggestions and links to relevant documentation,
and optionally colorize errors similar to gcc. There's much more to do
here, and we'll be adding more heuristics to make error messages more
helpful.
- Thanks to Matthias Bussonnier, you can now see an HTML annotated
version of your function in a Jupyter notebook with the
`inspect_types(pretty=True)` method. We plan to use this as the core of an
improved Numba annotation feature that will give you more info about how
your function was optimized (or not) by ParallelAccelerator and LLVM in the
future.
- We have an open PR for support for accessing the attributes of a
scipy.LowLevelCallable (https://github.com/numba/numba/pull/2999) in
nopython mode, but we're not sure we understand how it is being used in
practice. Anyone who can comment on that PR with example code would be
much appreciated.
- There is now a documentation page on how to use and combine the
various performance features in Numba:
https://numba.pydata.org/numba-doc/dev/user/performance-tips.html
Those of you at SciPy 2018 this week can see Siu's 3 minute talk on Numba
in the Tools plenary session on Thursday, where he will speedrun all the
other Numba improvements that have happened lately.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scipy-dev/attachments/20180709/b58e29a1/attachment.html>
More information about the SciPy-Dev
mailing list