Question

• Does a reliable fast implementiaont already exist? If so, consider using that
• Start with a numpy/python prototype - if this is fast enough, stop
• See if better use of vectoriazaiton via numpy will help
• Moving to native code:
  • Most Python devleopers will use Cython as the tool of choice. Cython can also be used to access/wrap C/C++ code
  • JIT compilation with numba is improving fast and may become competitive with Cython in the near future
  • If the function is “minimal”, it is usually worth considering numexpr because there is almost no work to be done
  • Use C/C++/Fortran if you are fluent in those languages - you have seen how to call these functions from Python
• If appropriate, consider parallelization (covered in later session)
• As you optimize your code, remmeber:
  • Check that is is giving correct results!
  • Profile often - it is very hard to preidct the effect of an optimizaiton in general
  • Remember that your time is precious - stop when fast enough
  • If getting a bigger, faster machine will sovle the problem, that is sometimes the best solution

%load_ext version_information

%version_information numpy, scipy, numexpr, numba, numbapro, parakeet, cython, f2py,

Software	Version
Python	2.7.9 64bit [GCC 4.2.1 (Apple Inc. build 5577)]
IPython	3.1.0
OS	Darwin 13.4.0 x86_64 i386 64bit
numpy	1.9.2
scipy	0.15.1
numexpr	2.3.1
numba	0.17.0
numbapro	0.17.1
parakeet	0.23.2
cython	0.22
f2py	f2py
Thu Apr 09 09:52:28 2015 EDT