[SciPy-dev] Major CVS changes to weave

eric jones eric at enthought.com
Thu Sep 12 02:10:22 CDT 2002

Hey group,

I've just checked a new version of weave into the CVS.  It has major
changes in it which are highlighted below.  The good news is there are
some big improvements and the code should be more portable.  The bad
news is that it introduced some backward incompatibility.  Sorry about
that.  I'm getting pretty happy with the interface and code generation
now, so we're close to a stage where backward compatibility will be
maintained in future releases.    

The docs and maybe a few examples still need updating.  I'll try to get
these and the web pages cleaned up tomorrow and the next day and then
make packaged release of weave when it is done.

This big change was needed even if new things weren't coming, but the
real reason I've checked this in now is in preparation of adding the
*really* cool stuff that Pat Miller is doing to optimize standard python
code and Numeric expressions.  His stuff is still way experimental, but
has huge potential.  We'll be working over the next month or so to get
it up and running so that it is ready for public consumption.



The underlying library code is significantly re-factored and simpler.
There used to be a xxx_spec.py and xxx_info.py file for every group of
type conversion classes.  The spec file held the python code that
handled the conversion and the info file had most of the C code
templates that were generated.  This proved pretty confusing in
practice, so the two files have mostly been merged into the spec file.
Also, there was quite a bit of code duplication running around.  The
re-factoring was able to trim the standard conversion code base
(excluding blitz and accelerate stuff) by about 40%.  This should be a
huge maintainability and extensibility win.

With multiple months of using Numeric arrays, I've found some of weave's
"magic variable" names unwieldy and want to change them.  The following
are the old declarations for an array x of Float32 type:
        PyArrayObject* x = convert_to_numpy(...);
        float* x_data = (float*) x->data;
        int*   _Nx = x->dimensions;
        int*   _Sx = x->strides;
        int    _Dx = x->nd;
The new declaration looks like this:
        PyArrayObject* x_array = convert_to_numpy(...);
        float* x = (float*) x->data;
        int*   Nx = x->dimensions;
        int*   Sx = x->strides;
        int    Dx = x->nd;
This is obviously not backward compatible, and will break some code
(including a lot of mine).  It also makes inline() code more readable
and natural to write.

I've switched from CXX to Gordon McMillan's SCXX for list, tuples, and
dictionaries.  I like CXX pretty well, but its use of advanced C++
(templates, etc.) caused some portability problems.  The SCXX library is
similar to CXX but doesn't use templates at all.  This, like (1) is not
API compatible change and requires repairing existing code.

I have also thought about boost python, but it also makes heavy use of
templates.  Moving to SCXX gets rid of almost all template usage for the
standard type converters which should help portability.  std::complex
and std::string from the STL are the only templates left.  Of course
blitz still uses templates in a major way so weave.blitz will continue
to be hard on compilers.

I've actually considered scrapping the C++ classes for list, tuples, and
dictionaries, and just fall back to the standard Python C API because
the classes are waaay slower than the raw API in many cases.  They are
also more convenient and less error prone in many cases, so I've decided
to stick with them.  The PyObject variable will always be made available
for variable "x" under the name "py_x" for more speedy operations.
You'll definitely want to use these for anything that needs to be

strings are converted to std::string now.  I found this to be the most
useful type in for strings in my code.  Py::String was used previously.

There are a number of reference count "errors" in some of the less
tested conversion codes such as instance, module, etc.  I've cleaned
most of these up.  I put errors in quotes here because I'm actually not
positive that objects passed into "inline" really need reference
counting applied to them.  The dictionaries passed in by inline() hold
references to these objects so it doesn't seem that they could ever be
garbage collected inadvertently.  Variables used by ext_tools, though,
definitely need the reference counting done.  I don't think this is a
major cost in speed, so it probably isn't worth getting rid of the ref
count code.  

Unicode objects are now supported.  This was necessary to support
rendering Unicode strings in the freetype wrappers for Chaco.

blitz++ was upgraded to the latest CVS.  It compiles about twice as fast
as the old blitz and looks like it supports a large number of compilers
(including SGI, Steve) though only gcc 2.95.3 (2.96 on Linux) is tested.
Compile times now take about 9 seconds on my 850 MHz PIII laptop.

More information about the Scipy-dev mailing list