Numba 0.25 Released!

[Stanley Seibert]

Hi Numba Users,

I'm happy to announce the release of Numba 0.25 (live from the NVIDIA GPU Technology Conference where Siu and I are currently sitting!), which you can get with "conda update numba", or from PyPI:

https://pypi.python.org/pypi/numba/0.25.0

As promised, we dropped support for Python 2.6, 3.3 and NumPy 1.6 from this release.  Numba 0.25.0 also requires llvmlite 0.10.

This release was primarily focused on our CUDA compilation target, but there were also a lot of NumPy related improvements:

• Much improved support for using CUDA Python kernels and ufuncs with both Dask and Spark.  Now you can do GPU computing with your cluster without leaving Python!
  
• If you happen to be at GTC, come check out our talk on using Numba with Spark clusters: http://mygtc.gputechconf.com/quicklink/fDSKKz
  
• Added math.erf, erfc, gamma, and lgamma to the CUDA target
• Support for set objects in nopython mode!
  
• Support for numpy.nditer
• More useful NumPy functions: histogram, diff, searchsorted, NaN-aware functions, and others
  
• Improved performance for NumPy reduction functions (like sum, prod, median, etc)
• Support applying iter and next functions to iterators
• Can use the size parameter to return arrays from numpy.random functions
• Numba-compiled functions are now reported more clearly in the Python profiler

There were also a number of smaller features, and important bug fixes.  For the full list, see the release notes:

http://numba.pydata.org/numba-doc/latest/release-notes.html#version-0-25-0

Thanks again everyone for your enthusiasm, questions, and bug reports!

[Brian Merchant]

I am really liking how numba-compiled functions are now show up better in Python profiler results. Have already made some good use of that! Thank you :)