[SciPy-user] Automatic Differentiation with PYADOLC and Removing Boost::Python dependency

[Ravi]

On Thursday 26 March 2009 13:34:16 Rob Clewley wrote:
> Ravi's points are well taken, and I do not pretend to know much about
> boost.python myself, but I am tempted to agree with David that it
> would still be more helpful to have a source-based version that does
> not depend on boost at all. Maybe Ravi could educate me as to what it
> would mean for boost.python to be more portable than python in this
> example?

On "exotic" systems (such as one with 24 bit ints and non-IEEE 754 floating 
point systems), boost.python (among other well-written packages) compiles 
where even python does not[1]. Python would be a wonderful scripting language 
for some embedded processor work, but I have been unable to compile python for 
such platforms (perhaps because I give up if it does not compile after ~20 
hours of work). Perhaps python 2.4+ may be more portable than 2.3-, but I 
haven't tried compiling python on exotic systems in a while.

[1] Interestingly enough, python headers are much cleaner than most people 
realize. Of course, when python does not compile, a boost.python static 
library, when one can be created, is useless.

> As an example, I would love to use this code in PyDSTool to find
> derivatives more accurately to compute periodic orbit solutions of
> differential equations, but it does not appeal to me to require boost
> as a dependency. My impression from occasional reading of the boost
> website is that users would need to install an awful lot of additional
> things (often from source on non-windows machines) to get this
> interface to work.

Not really. In fact, no external dependencies are required other than your 
reasonably standards-conformant C++ compiler. (You could optionally link 
against IBM ICU for the regexp library, and optionally build an MPI interface 
for some libraries if you have OpenMPI installed.) If you have a reasonably 
standard python installation, boost.python will be built. The steps are:
  - download boost-xxx.tar.gz
  - unpack the tarfile
  - cd to the bjam directory, build it
  - run bjam on the boost sources
I currently use it on various linux flavors & Windows; in the past, I have 
used it on Solaris and a subset on HPUX. If you have a POSIX shell, there is 
even a script that does it all for you.

> Alternatively, SWIG is a lightweight additional
> dependency that easily wraps simple numpy-dependent C/C++ code, and
> this is what I use instead.  I'll be delighted to learn otherwise if I
> am mistaken in my impression.

I don't consider SWIG to be a lightweight dependency, but that's merely a 
matter of opinion. More importantly, SWIG does not provide easy methods to map 
the C++ object model to the python object model since it uses (essentially) 
opadue pointers. Overriding C++ virtual functions in python is (or at least 
used to be) a pain with SWIG; SWIG also has problems with interfaces that 
depend on Koenig lookup. SWIG works very well for C code, though.

Regards,
Ravi