Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Fetching contributors…

Cannot retrieve contributors at this time

2753 lines (2738 sloc) 124.328 kb
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
</head>
<body>
<h1>Weave Documentation</h1>
<p>
By Eric Jones eric@enthought.com
</p>
<p></p>
<h2>Outline</h2>
<dl>
<dd> <a href="#Introduction">Introduction</a> </dd>
<dd> <a href="#Requirements">Requirements</a> </dd>
<dd> <a href="#Installation">Installation</a> </dd>
<dd> <a href="#Testing">Testing</a> </dd>
<dd> <a href="#Benchmarks">Benchmarks</a> </dd>
<dd> <a href="#Inline">Inline</a>
<dl>
<dd><a href="#More%20with%20printf">More with printf</a> </dd>
<dd> <a href="#More%20examples">More examples</a>
<dl>
<dd><a href="#Binary%20search">Binary search</a> </dd>
<dd><a href="#Dictionary%20sort">Dictionary sort</a> </dd>
<dd><a href="#Numeric%20--%20cast/copy/transpose">NumPy --
cast/copy/transpose</a> </dd>
<dd><a href="#wxPython">wxPython</a></dd>
</dl>
</dd>
<dd><a href="#Keyword%20options">Keyword options</a> </dd>
<dd><a href="#Returning%20values">Returning values</a>
<dl>
<dd><a href="#The%20issue%20with%20locals%28%29"> The issue
with <code>locals()</code></a></dd>
</dl>
</dd>
<dd><a href="#inline_quick_look_at_code">A quick look at the code</a>
</dd>
<dd> <a href="#inline_technical_details">Technical Details</a>
<dl>
<dd><a href="#Converting%20Types">Converting Types</a>
<dl>
<dd><a href="#inline_numeric_argument_conversion"> NumPy
Argument Conversion</a> </dd>
<dd><a href="#inline_python_argument_conversion"> String,
List, Tuple, and Dictionary Conversion</a> </dd>
<dd><a href="#inline_callable_argument_conversion">File
Conversion</a> </dd>
<dd><a href="#inline_callable_argument_conversion">Callable,
Instance, and Module Conversion</a> </dd>
<dd><a href="#Customizing%20Conversions">Customizing
Conversions</a> </dd>
</dl>
</dd>
<dd><a href="#Compiling%20Code">Compiling Code</a> </dd>
<dd><a href="#The%20Catalog">"Cataloging" functions</a>
<dl>
<dd><a href="#function%20storage">Function Storage</a> </dd>
<dd><a href="#PYTHONCOMPILED">The PYTHONCOMPILED
evnironment variable</a></dd>
</dl>
</dd>
</dl>
</dd>
</dl>
</dd>
<dd><a href="#Blitz">Blitz</a>
<dl>
<dd><a href="#blitz_requirements">Requirements</a> </dd>
<dd><a href="#blitz_limitations">Limitations</a> </dd>
<dd><a href="#Numeric%20Efficiency">NumPy Efficiency Issues</a> </dd>
<dd><a href="#blitz_tools">The Tools</a>
<dl>
<dd><a href="#blitz_parser">Parser</a> </dd>
<dd><a href="#blitz_blitz">Blitz and NumPy</a> </dd>
</dl>
</dd>
<dd><a href="#blitz_type_conversions">Type defintions and coersion</a>
</dd>
<dd><a href="#blitz_catalog">Cataloging Compiled Functions</a> </dd>
<dd><a href="#blitz_array_sizes">Checking Array Sizes</a> </dd>
<dd><a href="#blitz_extension_module">Creating the Extension
Module</a> </dd>
</dl>
</dd>
<dd> <a href="#Extension%20Modules"> Extension Modules</a>
<dl>
<dd><a href="#A%20Simple%20Example">A Simple Example</a> </dd>
<dd><a href="#Fibonacci%20Example">Fibonacci Example</a> </dd>
</dl>
</dd>
<dd> <a href="#Type%20Factories"> Customizing Type Conversions --
Type Factories (not written)</a>
<dl>
<dd>Type Specifications </dd>
<dd>Type Information </dd>
<dd>The Conversion Process </dd>
</dl>
</dd>
</dl>
<a name="Introduction"></a>
<h1>Introduction</h1>
<p>
The <code>weave</code> package provides tools for including C/C++ code
within
in Python code. This offers both another level of optimization to those
who need it, and an easy way to modify and extend any supported
extension libraries such as wxPython and hopefully VTK soon. Inlining
C/C++ code within Python generally
results in speed ups of 1.5x to 30x speed-up over algorithms written in
pure
Python (However, it is also possible to slow things down...). Generally
algorithms that require a large number of calls to the Python API don't
benefit
as much from the conversion to C/C++ as algorithms that have inner
loops completely convertable to C.
</p>
<p> There are three basic ways to use <code>weave</code>. The <code>weave.inline()</code>
function executes C code directly within Python, and <code>weave.blitz()</code>
translates Python NumPy expressions to C++ for fast execution. <code>blitz()</code>
was the original reason <code>weave</code> was built. For those
interested in building extension
libraries, the <code>ext_tools</code> module provides classes for
building extension modules within Python. </p>
<p>Most of <code>weave's</code> functionality should work on Windows
and Unix, although some of its functionality requires <code>gcc</code>
or a similarly modern C++ compiler that handles templates well. Up to
now, most testing has been done on Windows 2000 with Microsoft's C++
compiler (MSVC) and with gcc (mingw32 2.95.2 and 2.95.3-6). All tests
also pass on Linux (RH 7.1 with gcc 2.96), and I've had reports that it
works on Debian also (thanks Pearu).
</p>
<p>The <code>inline</code> and <code>blitz</code> provide new
functionality to Python (although I've recently learned about the <a
href="http://pyinline.sourceforge.net/">PyInline</a> project which may
offer similar functionality to <code>inline</code>). On the other
hand, tools for building Python extension modules already exists (SWIG,
SIP, pycpp, CXX, and others). As of yet, I'm not sure where <code>weave</code>
fits in this spectrum. It is closest in flavor to CXX in that it makes
creating new C/C++ extension modules pretty easy. However, if you're
wrapping a gaggle of legacy functions or classes, SWIG and friends are
definitely the better choice. <code>weave</code> is set up so that you
can customize how Python types are converted to C types in <code>weave</code>.
This is great for <code>inline()</code>, but, for wrapping legacy
code, it is more flexible to specify things the other way around --
that is how C types map to Python types. This <code>weave</code> does
not do. I guess it would be possible to build such a tool on top of <code>weave</code>,
but with good tools like SWIG around, I'm not sure the effort produces
any new capabilities. Things like function overloading are probably
easily implemented in <code>weave</code> and it might be easier to mix
Python/C code in function calls, but nothing beyond this comes to mind.
So, if you're developing new extension modules or optimizing Python
functions in C, <code>weave.ext_tools()</code> might be the tool for
you. If you're wrapping legacy code, stick with SWIG.
</p>
<p>The next several sections give the basics of how to use <code>weave</code>.
We'll discuss what's happening under the covers in more detail later
on. Serious users will need to at least look at the type conversion
section to understand how Python variables map to C/C++ types and how
to customize this behavior. One other note. If you don't know C or C++
then these docs are probably of very little help to you. Further, it'd
be helpful if you know something about writing Python extensions. <code>weave</code>
does quite a bit for you, but for anything complex, you'll need to do
some conversions, reference counting, etc.
</p>
<p><em>
Note: </em><code>weave</code><em> is actually part of the <a
href="http://www.scipy.org">SciPy</a> package. However, it also works
fine as a standalone package (you can install as ``python setup.py install``)
in the scipy/weave directory. The examples here are given as if it is used
as a stand alone package. If you are using from within scipy, you can
use <code> from scipy import weave</code> and the examples will work
identically.</em>
<a name="Requirements"></a></p>
<h1>Requirements</h1>
<ul>
<li> Python
<p> I use 2.1.1. Probably 2.0 or higher should work. </p>
<p> </p>
</li>
<li> C++ compiler
<p> <code>weave</code> uses <code>distutils</code> to actually
build extension modules, so it uses whatever compiler was originally
used to build Python. <code>weave</code> itself requires a C++
compiler. If you used a C++ compiler to build Python, your probably
fine. </p>
<p> On Unix gcc is the preferred choice because I've done a little
testing with it. All testing has been done with gcc, but I expect the
majority of compilers should work for <code>inline</code> and <code>ext_tools</code>.
The one issue I'm not sure about is that I've hard coded things so that
compilations are linked with the <code>stdc++</code> library. <em>Is
this standard across Unix compilers, or is this a gcc-ism?</em> </p>
<p> For <code>blitz()</code>, you'll need a reasonably recent
version of gcc. 2.95.2 works on windows and 2.96 looks fine on Linux.
Other versions are likely to work. Its likely that KAI's C++ compiler
and maybe some others will work, but I haven't tried. My advice is to
use gcc for now unless your willing to tinker with the code some. </p>
<p> On Windows, either MSVC or gcc (<a
href="http://www.mingw.org%3Ewww.mingw.org"> mingw32</a>) should work.
Again, you'll need gcc for <code>blitz()</code> as the MSVC compiler
doesn't handle templates well. </p>
<p> I have not tried Cygwin, so please report success if it works
for you. </p>
<p> </p>
</li>
<li> NumPy
<p>The python NumPy module from <a href="http://numeric.scipy.org/">here</a>.
is required for <code>blitz()</code> to work and for numpy.distutils
which is used by weave. </p>
<p> </p>
</li>
</ul>
<p>
<a name="Installation"></a></p>
<h1>Installation</h1>
<p>
There are currently two ways to get <code>weave</code>. Fist, <code>weave</code>
is part of SciPy and installed automatically (as a sub-
package) whenever SciPy is installed. Second, since <code>weave</code>
is useful outside of the scientific community, it has been setup so
that it can be
used as a stand-alone module. </p>
<p>The stand-alone version can be downloaded from <a
href="http://www.scipy.org/Weave">here</a>.&nbsp; Instructions for
installing should be found there as well.&nbsp; setup.py file to
simplify
installation. </p>
<p><a name="Testing"></a></p>
<h1>Testing</h1>
Once <code>weave</code> is installed, fire up python and run its unit
tests.
<blockquote>
<pre><code>
&gt;&gt;&gt; import weave
&gt;&gt;&gt; weave.test()
runs long time... spews tons of output and a few warnings
.
.
.
..............................................................
................................................................
..................................................
----------------------------------------------------------------------
Ran 184 tests in 158.418s
OK
<unittest.texttestrunner
instance="" at="" 01562934="">
&gt;&gt;&gt;
</unittest.texttestrunner></code></pre>
</blockquote>
This takes a while, usually several minutes.
On Unix with remote file systems, I've had it take 15 or so minutes. In
the end, it should run about 180 tests and spew some speed results
along the way. If you get errors, they'll be reported at the end of the
output. Please report erros that you find.&nbsp;&nbsp; Some tests are
known to fail at this point. <br>
<p>If you only want to test a single module of the package, you can do
this by
running test() for that specific module. </p>
<blockquote>
<pre><code>
&gt;&gt;&gt; import weave.scalar_spec
&gt;&gt;&gt; weave.scalar_spec.test()
.......
----------------------------------------------------------------------
Ran 7 tests in 23.284s
</code></pre>
</blockquote>
<em>Testing Notes:
</em>
<ul>
<em> </em>
<li><em> Windows 1
<p> I've had some test fail on windows machines where I have msvc,
gcc-2.95.2 (in c:\gcc-2.95.2), and gcc-2.95.3-6 (in c:\gcc) all
installed. My environment has c:\gcc in the path and does not have
c:\gcc-2.95.2 in the path. The test process runs very smoothly until
the end where several test using gcc fail with cpp0 not found by g++.
If I check os.system('gcc -v') before running tests, I get
gcc-2.95.3-6. If I check after running tests (and after failure), I get
gcc-2.95.2. ??huh??. The os.environ['PATH'] still has c:\gcc first in
it and is not corrupted (msvc/distutils messes with the environment
variables, so we have to undo its work in some places). If anyone else
sees this, let me know - - it may just be an quirk on my machine
(unlikely). Testing with the gcc- 2.95.2 installation always works. </p>
<p> </p>
</em></li>
<em> </em>
<li><em> Windows 2
<p> If you run the tests from PythonWin or some other GUI tool,
you'll get a ton of DOS windows popping up periodically as <code>weave</code>
spawns the compiler multiple times. Very annoying. Anyone know how to
fix this? </p>
<p> </p>
</em></li>
<em> </em>
<li><em> wxPython
<p> wxPython tests are not enabled by default because importing
wxPython on a Unix machine without access to a X-term will cause the
program to exit. Anyone know of a safe way to detect whether wxPython
can be imported and whether a display exists on a machine? </p>
<p> </p>
<p></p>
</em></li>
<em> </em>
</ul>
<a name="Benchmarks"></a>
<h1>Benchmarks</h1>
This section has not been updated from old scipy weave and Numeric....<br>
<br>
This section has a few benchmarks&nbsp; -- thats all people want to see
anyway right? These are mostly taken from running files in the <code>weave/example</code>
directory and also from the test scripts. Without more information
about what the test actually do, their value is limited. Still, their
here for the curious. Look at the example scripts for more specifics
about what problem was actually solved by each run. These examples are
run under windows 2000 using Microsoft Visual C++ and python2.1 on a
850 MHz PIII laptop with 320 MB of RAM.
Speed up is the improvement (degredation) factor of <code>weave</code>
compared to conventional Python functions. <code>The blitz()</code>
comparisons are shown
compared to NumPy.
<p></p>
<center>
<table border="1" width="100%">
<tbody>
<tr>
<td colspan="2" width="100%">
<p align="center">inline and ext_tools</p>
</td>
</tr>
<tr>
<td>
<p align="center">Algorithm</p>
</td>
<td>
<p align="center">Speed up </p>
</td>
</tr>
<tr>
<td>binary search</td>
<td> &nbsp;&nbsp;1.50 </td>
</tr>
<tr>
<td>fibonacci (recursive)</td>
<td> &nbsp;82.10 </td>
</tr>
<tr>
<td>fibonacci (loop)</td>
<td> &nbsp;&nbsp;9.17 </td>
</tr>
<tr>
<td>return None</td>
<td> &nbsp;&nbsp;0.14 </td>
</tr>
<tr>
<td>map</td>
<td> &nbsp;&nbsp;1.20 </td>
</tr>
<tr>
<td>dictionary sort</td>
<td> &nbsp;&nbsp;2.54 </td>
</tr>
<tr>
<td>vector quantization</td>
<td> &nbsp;37.40 </td>
</tr>
<tr>
<td colspan="2" width="100%">
<p align="center">blitz -- double precision</p>
</td>
</tr>
<tr>
<td>
<p align="center">Algorithm</p>
</td>
<td>
<p align="center">Speed up </p>
</td>
</tr>
<tr>
<td>a = b + c 512x512</td>
<td> &nbsp;&nbsp;3.05 </td>
</tr>
<tr>
<td>a = b + c + d 512x512</td>
<td> &nbsp;&nbsp;4.59 </td>
</tr>
<tr>
<td>5 pt avg. filter, 2D Image 512x512</td>
<td> &nbsp;&nbsp;9.01 </td>
</tr>
<tr>
<td>Electromagnetics (FDTD) 100x100x100</td>
<td> &nbsp;&nbsp;8.61 </td>
</tr>
</tbody>
</table>
</center>
<p>
The benchmarks shown <code>blitz</code> in the best possible light.
NumPy (at least on my machine) is significantly worse for double
precision than it is for single precision calculations. If your
interested in single precision results, you can pretty much divide the
double precision speed up by 3 and you'll
be close.
<a name="Inline"></a></p>
<h1>Inline</h1>
<p>
<code>inline()</code> compiles and executes C/C++ code on the fly.
Variables in the local and global Python scope are also available in
the C/C++ code. Values are passed to the C/C++ code by assignment much
like variables are passed into a standard Python function. Values are
returned from the C/C++ code through a special argument called
return_val. Also, the contents of mutable objects can be changed within
the C/C++ code and the changes remain after the C code exits and
returns to Python. (more on this later)
</p>
<p> Here's a trivial <code>printf</code> example using <code>inline()</code>:
</p>
<blockquote>
<pre><code>
&gt;&gt;&gt; import weave
&gt;&gt;&gt; a = 1
&gt;&gt;&gt; weave.inline('printf("%d\\n",a);',['a'])
1
</code></pre>
</blockquote>
<p>In this, its most basic form, <code>inline(c_code, var_list)</code>
requires two arguments. <code>c_code</code> is a string of valid C/C++
code. <code>var_list</code> is a list of variable names that are
passed from Python into C/C++. Here we have a simple <code>printf</code>
statement that writes the Python variable <code>a</code> to the
screen. The first time you run this, there will be a pause while the
code is written to a .cpp file, compiled into an extension module,
loaded into Python, cataloged for future use, and executed. On windows
(850 MHz PIII), this takes about 1.5 seconds when using Microsoft's C++
compiler (MSVC) and 6-12 seconds using gcc (mingw32 2.95.2). All
subsequent executions of the code will happen very quickly because the
code only needs to be compiled once. If you kill and restart the
interpreter and then execute the same code fragment again, there will
be a much shorter delay in the fractions of seconds range. This is
because <code>weave</code> stores a catalog of all previously compiled
functions in an on disk cache. When it sees a string that has been
compiled, it loads the already compiled module and executes the
appropriate function. </p>
<p><em>
Note: If you try the <code>printf</code> example in a GUI shell such
as IDLE, PythonWin, PyShell, etc., you're unlikely to see the output.
This is because the C code is writing to stdout, instead of to the GUI
window. This doesn't mean that inline doesn't work in these
environments -- it only means that standard out in C is not the same as
the standard out for Python in these cases. Non input/output functions
will work as expected.
</em></p>
<p>Although effort has been made to reduce the overhead associated with
calling inline, it is still less efficient for simple code snippets
than using equivalent Python code. The simple <code>printf</code>
example is actually slower by 30% or so than using Python <code>print</code>
statement. And, it is not difficult to create code fragments that are
8-10 times slower using inline than equivalent Python. However, for
more complicated algorithms, the speed up can be worth while --
anywhwere from 1.5- 30 times faster. Algorithms that have to manipulate
Python objects (sorting a list) usually only see a factor of 2 or so
improvement. Algorithms that are highly computational or manipulate
NumPy arrays can see much larger improvements. The examples/vq.py file
shows a factor of 30 or more improvement on the vector quantization
algorithm that is used heavily in information theory and classification
problems.
</p>
<p><a name="More with printf"></a>
</p>
<h2>More with printf</h2>
<p>
MSVC users will actually see a bit of compiler output that distutils
does not
supress the first time the code executes: </p>
<blockquote>
<pre><code> <br> &gt;&gt;&gt; weave.inline(r'printf("%d\n",a);',['a'])<br> sc_e013937dbc8c647ac62438874e5795131.cpp<br> Creating library C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp<br> \Release\sc_e013937dbc8c647ac62438874e5795131.lib and object C:\DOCUME<br> ~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\sc_e013937dbc8c64<br> 7ac62438874e5795131.exp<br> 1<br> </code></pre>
</blockquote>
<p>
Nothing bad is happening, its just a bit annoying. <em> Anyone know
how to turn this off?</em> </p>
<p>This example also demonstrates using 'raw strings'. The <code>r</code>
preceeding the code string in the last example denotes that this is a
'raw string'. In raw strings, the backslash character is not
interpreted as an escape character, and so it isn't necessary to use a
double backslash to indicate that the '\n' is meant to be interpreted
in the C <code>printf</code> statement instead of by Python. If your C
code contains a lot
of strings and control characters, raw strings might make things
easier.
Most of the time, however, standard strings work just as well.
</p>
<p>The <code>printf</code> statement in these examples is formatted to
print out integers. What happens if <code>a</code> is a string? <code>inline</code>
will happily, compile a new version of the code to accept strings as
input,
and execute the code. The result? </p>
<blockquote>
<pre><code> <br> &gt;&gt;&gt; a = 'string'<br> &gt;&gt;&gt; weave.inline(r'printf("%d\n",a);',['a'])<br> 32956972<br> </code></pre>
</blockquote>
<p>In this case, the result is non-sensical, but also non-fatal. In
other situations, it might produce a compile time error because <code>a</code>
is required to be an integer at some point in the code, or it could
produce a segmentation fault. Its possible to protect against passing <code>inline</code>
arguments of the wrong data type by using asserts in Python. </p>
<blockquote>
<pre><code> <br> &gt;&gt;&gt; a = 'string'<br> &gt;&gt;&gt; def protected_printf(a): <br> ... assert(type(a) == type(1))<br> ... weave.inline(r'printf("%d\n",a);',['a'])<br> &gt;&gt;&gt; protected_printf(1)<br> 1<br> &gt;&gt;&gt; protected_printf('string')<br> AssertError...<br> </code></pre>
</blockquote>
<p>For printing strings, the format statement needs to be changed.
Also, weave
doesn't convert strings to char*. Instead it uses CXX Py::String type,
so you have to do a little more work. Here we convert it to a C++
std::string
and then ask cor the char* version. </p>
<blockquote>
<pre><code> <br> &gt;&gt;&gt; a = 'string' <br> &gt;&gt;&gt; weave.inline(r'printf("%s\n",std::string(a).c_str());',['a'])<br> string<br> </code></pre>
</blockquote>
<p><em> This is a little convoluted. Perhaps strings should convert to
std::string
objects instead of CXX objects. Or maybe to char*.
</em></p>
<p>As in this case, C/C++ code fragments often have to change to accept
different types. For the given printing task, however, C++ streams
provide a way of a single statement that works for integers and
strings. By default, the stream objects live in the std (standard)
namespace and thus require the use of <code>std::</code>. </p>
<blockquote>
<pre><code> <br> &gt;&gt;&gt; weave.inline('std::cout &lt;&lt; a &lt;&lt; std::endl;',['a'])<br> 1 <br> &gt;&gt;&gt; a = 'string'<br> &gt;&gt;&gt; weave.inline('std::cout &lt;&lt; a &lt;&lt; std::endl;',['a'])<br> string<br> </code></pre>
</blockquote>
<p>Examples using <code>printf</code> and <code>cout</code> are
included in examples/print_example.py.
<a name="More examples"></a></p>
<h2> More examples </h2>
This section shows several more advanced uses of <code>inline</code>.
It includes a few algorithms from the <a
href="http://aspn.activestate.com/ASPN/Cookbook/Python">Python Cookbook</a>
that have been re-written in inline C to improve speed as well as a
couple examples using NumPy and wxPython.
<a name="Binary search"></a>
<h3> Binary search</h3>
Lets look at the example of searching a sorted list of integers for a
value. For inspiration, we'll use Kalle Svensson's <a
href="http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/81188">
binary_search()</a> algorithm from the Python Cookbook. His recipe
follows:
<blockquote>
<pre><code>
def binary_search(seq, t):
min = 0; max = len(seq) - 1
while 1:
if max &lt; min:
return -1
m = (min + max) / 2
if seq[m] &lt; t:
min = m + 1
elif seq[m] &gt; t:
max = m - 1
else:
return m
</code></pre>
</blockquote>
This Python version works for arbitrary Python data types. The C
version below is specialized to handle integer values. There is a
little type checking done in Python to assure that we're working with
the correct data types before heading into C. The variables <code>seq</code>
and <code>t</code> don't need to be declared beacuse <code>weave</code>
handles converting and declaring them in the C code. All other
temporary variables such as <code>min, max</code>, etc. must be
declared -- it is C after all. Here's the new mixed Python/C function:
<blockquote>
<pre><code> <br> def c_int_binary_search(seq,t):<br> # do a little type checking in Python<br> assert(type(t) == type(1))<br> assert(type(seq) == type([]))<br> <br> # now the C code<br> code = """<br> #line 29 "binary_search.py"<br> int val, m, min = 0; <br> int max = seq.length() - 1;<br> PyObject *py_val; <br> for(;;)<br> {<br> if (max &lt; min ) <br> { <br> return_val = Py::new_reference_to(Py::Int(-1)); <br> break;<br> } <br> m = (min + max) /2;<br> val = py_to_int(PyList_GetItem(seq.ptr(),m),"val"); <br> if (val &lt; t) <br> min = m + 1;<br> else if (val &gt; t)<br> max = m - 1;<br> else<br> {<br> return_val = Py::new_reference_to(Py::Int(m));<br> break;<br> }<br> }<br> """<br> return inline(code,['seq','t'])<br> </code></pre>
</blockquote>
<p>We have two variables <code>seq</code> and <code>t</code> passed
in. <code>t</code> is guaranteed (by the <code>assert</code>) to be
an integer. Python integers are converted to C int types in the
transition from Python to C. <code>seq</code> is a Python list. By
default, it is translated to a CXX list object. Full documentation for
the CXX library can be found at its <a
href="http://cxx.sourceforge.net/">website</a>. The basics are that
the CXX provides C++ class equivalents for Python objects that
simplify, or at least object orientify, working with Python objects in
C/C++. For example, <code>seq.length()</code> returns the length of
the list. A little more about
CXX and its class methods, etc. is in the ** type conversions **
section.
</p>
<p><em>
Note: CXX uses templates and therefore may be a little less portable
than another alternative by Gordan McMillan called SCXX which was
inspired by
CXX. It doesn't use templates so it should compile faster and be more
portable.
SCXX has a few less features, but it appears to me that it would mesh
with
the needs of weave quite well. Hopefully xxx_spec files will be written
for SCXX in the future, and we'll be able to compare on a more
empirical
basis. Both sets of spec files will probably stick around, it just a
question
of which becomes the default.
</em></p>
<p>Most of the algorithm above looks similar in C to the original
Python code. There are two main differences. The first is the setting
of <code>return_val</code> instead of directly returning from the C
code with a <code>return</code> statement. <code>return_val</code> is
an automatically defined variable of type <code>PyObject*</code> that
is returned from the C code back to Python. You'll have to handle
reference counting issues when setting this variable. In this example,
CXX classes and functions handle the dirty work. All CXX functions and
classes live in the namespace <code>Py::</code>. The following code
converts the integer <code>m</code> to a CXX <code>Int()</code>
object and then to a <code>PyObject*</code> with an incremented
reference count using <code>Py::new_reference_to()</code>. </p>
<blockquote>
<pre><code> <br> return_val = Py::new_reference_to(Py::Int(m));<br> </code></pre>
</blockquote>
<p>The second big differences shows up in the retrieval of integer
values from the Python list. The simple Python <code>seq[i]</code>
call balloons into a C Python API call to grab the value out of the
list and then a separate call to <code>py_to_int()</code> that
converts the PyObject* to an integer. <code>py_to_int()</code>
includes both a NULL cheack and a <code>PyInt_Check()</code> call as
well as the conversion call. If either of the checks fail, an exception
is raised. The entire C++ code block is executed with in a <code>try/catch</code>
block that handles exceptions much like Python does. This removes the
need for most error checking code.
</p>
<p>It is worth note that CXX lists do have indexing operators that
result in code that looks much like Python. However, the overhead in
using them appears to be relatively high, so the standard Python API
was used on the <code>seq.ptr()</code> which is the underlying <code>PyObject*</code>
of the List object.
</p>
<p>The <code>#line</code> directive that is the first line of the C
code block isn't necessary, but it's nice for debugging. If the
compilation fails because of the syntax error in the code, the error
will be reported as an error in the Python file "binary_search.py" with
an offset from the given line number (29 here).
</p>
<p>So what was all our effort worth in terms of efficiency? Well not a
lot in this case. The examples/binary_search.py file runs both Python
and C versions of the functions As well as using the standard <code>bisect</code>
module. If we run it on a 1 million element list and run the search
3000 times (for 0-
2999), here are the results we get: </p>
<blockquote>
<pre><code> <br> C:\home\ej\wrk\scipy\weave\examples&gt; python binary_search.py<br> Binary search for 3000 items in 1000000 length list of integers:<br> speed in python: 0.159999966621<br> speed of bisect: 0.121000051498<br> speed up: 1.32<br> speed in c: 0.110000014305<br> speed up: 1.45<br> speed in c(no asserts): 0.0900000333786<br> speed up: 1.78<br> </code></pre>
</blockquote>
<p>So, we get roughly a 50-75% improvement depending on whether we use
the Python asserts in our C version. If we move down to searching a
10000 element list, the advantage evaporates. Even smaller lists might
result in the Python version being faster. I'd like to say that moving
to NumPy lists (and getting rid of the GetItem() call) offers a
substantial speed up, but my preliminary efforts didn't produce one. I
think the log(N) algorithm is to blame. Because the algorithm is nice,
there just isn't much time spent computing things, so moving to C isn't
that big of a win. If there are ways to reduce conversion overhead of
values, this may improve the C/Python speed up. Anyone have other
explanations or faster code, please let me know.
<a name="#Dictionary sort"></a></p>
<h3> Dictionary Sort</h3>
<p>
The demo in examples/dict_sort.py is another example from the Python
CookBook. <a
href="http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/52306">This
submission</a>, by Alex Martelli, demonstrates how to return the values
from a dictionary sorted by their keys: </p>
<blockquote>
<pre><code> <br> def sortedDictValues3(adict):<br> keys = adict.keys()<br> keys.sort()<br> return map(adict.get, keys)<br> </code></pre>
</blockquote>
<p>Alex provides 3 algorithms and this is the 3rd and fastest of the
set. The C version of this same algorithm follows: </p>
<blockquote>
<pre><code> <br> def c_sort(adict):<br> assert(type(adict) == type({}))<br> code = """ <br> #line 21 "dict_sort.py" <br> Py::List keys = adict.keys();<br> Py::List items(keys.length()); keys.sort(); <br> PyObject* item = NULL; <br> for(int i = 0; i &lt; keys.length();i++)<br> {<br> item = PyList_GET_ITEM(keys.ptr(),i);<br> item = PyDict_GetItem(adict.ptr(),item);<br> Py_XINCREF(item);<br> PyList_SetItem(items.ptr(),i,item); <br> } <br> return_val = Py::new_reference_to(items);<br> """ <br> return inline_tools.inline(code,['adict'],verbose=1)<br> </code></pre>
</blockquote>
<p>Like the original Python function, the C++ version can handle any
Python dictionary regardless of the key/value pair types. It uses CXX
objects for the most part to declare python types in C++, but uses
Python API calls to manipulate their contents. Again, this choice is
made for speed. The C++ version, while
more complicated, is about a factor of 2 faster than Python. </p>
<blockquote>
<pre><code> <br> C:\home\ej\wrk\scipy\weave\examples&gt; python dict_sort.py<br> Dict sort of 1000 items for 300 iterations:<br> speed in python: 0.319999933243<br> [0, 1, 2, 3, 4]<br> speed in c: 0.151000022888<br> speed up: 2.12<br> [0, 1, 2, 3, 4]<br> </code></pre>
</blockquote>
<p>
<a name="#Numeric -- cast/copy/transpose"></a></p>
<h3>NumPy -- cast/copy/transpose</h3>
CastCopyTranspose is a function called quite heavily by Linear Algebra
routines
in the NumPy library. Its needed in part because of the row-major
memory layout
of multi-demensional Python (and C) arrays vs. the col-major order of
the underlying
Fortran algorithms. For small matrices (say 100x100 or less), a
significant
portion of the common routines such as LU decompisition or singular
value decompostion
are spent in this setup routine. This shouldn't happen. Here is the
Python
version of the function using standard NumPy operations.
<blockquote>
<pre><code> <br> def _castCopyAndTranspose(type, array):<br> if a.typecode() == type:<br> cast_array = copy.copy(NumPy.transpose(a))<br> else:<br> cast_array = copy.copy(NumPy.transpose(a).astype(type))<br> return cast_array<br> </code></pre>
</blockquote>
And the following is a inline C version of the same function:
<blockquote>
<pre><code>
from weave.blitz_tools import blitz_type_factories
from weave import scalar_spec
from weave import inline
def _cast_copy_transpose(type,a_2d):
assert(len(shape(a_2d)) == 2)
new_array = zeros(shape(a_2d),type)
NumPy_type = scalar_spec.NumPy_to_blitz_type_mapping[type]
code = \
"""
for(int i = 0;i &lt; _Na_2d[0]; i++)
for(int j = 0; j &lt; _Na_2d[1]; j++)
new_array(i,j) = (%s) a_2d(j,i);
""" % NumPy_type
inline(code,['new_array','a_2d'],
type_factories = blitz_type_factories,compiler='gcc')
return new_array
</code></pre>
</blockquote>
This example uses blitz++ arrays instead of the standard representation
of NumPy arrays so that indexing is simplier to write. This is
accomplished by passing in the blitz++ "type factories" to override the
standard Python to C++ type conversions. Blitz++ arrays allow you to
write clean, fast code, but they also are sloooow to compile (20
seconds or more for this snippet). This is why they aren't the default
type used for Numeric arrays (and also because most compilers can't
compile blitz arrays...). <code>inline()</code> is also forced to use
'gcc' as the compiler because the default compiler on Windows (MSVC)
will not compile blitz code. <em> 'gcc' I think will use the standard
compiler on Unix machine instead of explicitly forcing gcc (check this)
</em>Comparisons of the Python vs inline C++ code show a factor of 3
speed
up. Also shown are the results of an "inplace" transpose routine that
can be used if the output of the linear algebra routine can overwrite
the original matrix (this is often appropriate). This provides another
factor of 2 improvement.
<blockquote>
<pre><code>
#C:\home\ej\wrk\scipy\weave\examples&gt; python cast_copy_transpose.py
# Cast/Copy/Transposing (150,150)array 1 times
# speed in python: 0.870999932289
# speed in c: 0.25
# speed up: 3.48
# inplace transpose c: 0.129999995232
# speed up: 6.70
</code></pre>
</blockquote>
<a name="#wxPython" a="">&lt;&gt;
</a>
<h3><a name="#wxPython" a="">wxPython</a></h3>
<code><a name="#wxPython" a="">inline</a></code><a name="#wxPython" a="">
knows how to handle wxPython objects. Thats nice in and of
itself, but it also demonstrates that the type conversion mechanism is
reasonably flexible. Chances are, it won't take a ton of effort to
support special types
you might have. The examples/wx_example.py borrows the scrolled window
example from the wxPython demo, accept that it mixes inline C code in
the middle
of the drawing function. </a>
<blockquote>
<pre><code>
<a name="#wxPython" a=""> def DoDrawing(self, dc):<br> <br> red = wxNamedColour("RED");<br> blue = wxNamedColour("BLUE");<br> grey_brush = wxLIGHT_GREY_BRUSH;<br> code = \<br> """<br> #line 108 "wx_example.py" <br> dc-&gt;BeginDrawing();<br> dc-&gt;SetPen(wxPen(*red,4,wxSOLID));<br> dc-&gt;DrawRectangle(5,5,50,50);<br> dc-&gt;SetBrush(*grey_brush);<br> dc-&gt;SetPen(wxPen(*blue,4,wxSOLID));<br> dc-&gt;DrawRectangle(15, 15, 50, 50);<br> """<br> inline(code,['dc','red','blue','grey_brush'])<br> <br> dc.SetFont(wxFont(14, wxSWISS, wxNORMAL, wxNORMAL))<br> dc.SetTextForeground(wxColour(0xFF, 0x20, 0xFF))<br> te = dc.GetTextExtent("Hello World")<br> dc.DrawText("Hello World", 60, 65)<br><br> dc.SetPen(wxPen(wxNamedColour('VIOLET'), 4))<br> dc.DrawLine(5, 65+te[1], 60+te[0], 65+te[1])<br> ...<br> </a></code></pre>
</blockquote>
<a name="#wxPython" a="">Here, some of the Python calls to wx objects
were just converted to C++ calls. There
isn't any benefit, it just demonstrates the capabilities. You might
want to use this
if you have a computationally intensive loop in your drawing code that
you want to speed up.
On windows, you'll have to use the MSVC compiler if you use the
standard wxPython
DLLs distributed by Robin Dunn. Thats because MSVC and gcc, while
binary
compatible in C, are not binary compatible for C++. In fact, its
probably best, no matter what platform you're on, to specify that <code>inline</code>
use the same
compiler that was used to build wxPython to be on the safe side. There
isn't currently
a way to learn this info from the library -- you just have to know.
Also, at least
on the windows platform, you'll need to install the wxWindows libraries
and link to them. I think there is a way around this, but I haven't
found it yet -- I get some
linking errors dealing with wxString. One final note. You'll probably
have to
tweak weave/wx_spec.py or weave/wx_info.py for your machine's
configuration to
point at the correct directories etc. There. That should sufficiently
scare people
into not even looking at this... :)
</a><a name="Keyword Options"></a>
<h2> Keyword Options </h2>
<p>
The basic definition of the <code>inline()</code> function has a slew
of optional variables. It also takes keyword arguments that are passed
to <code>distutils</code> as compiler options. The following is a
formatted cut/paste of the argument section of <code>inline's</code>
doc-string. It explains all of the variables. Some examples using
various options will follow. </p>
<blockquote>
<pre><code> <br> def inline(code,arg_names,local_dict = None, global_dict = None, <br> force = 0, <br> compiler='',<br> verbose = 0, <br> support_code = None,<br> customize=None, <br> type_factories = None, <br> auto_downcast=1,<br> **kw):<br> </code></pre>
</blockquote>
<code>inline</code> has quite a few options as listed below. Also, the
keyword arguments for distutils extension modules are accepted to
specify extra information needed for compiling.
<blockquote></blockquote>
<h4>inline Arguments:</h4>
<blockquote>
<dl>
<dt>code </dt>
<dd>string. A string of valid C++ code. It should not specify a
return statement. Instead it should assign results that need to be
returned to Python in the return_val. </dd>
<dt>arg_names </dt>
<dd>list of strings. A list of Python variable names that should be
transferred from Python into the C/C++ code. </dd>
<dt>local_dict </dt>
<dd>optional. dictionary. If specified, it is a dictionary of
values that should be used as the local scope for the C/C++ code. If
local_dict is not specified the local dictionary of the calling
function is used. </dd>
<dt>global_dict </dt>
<dd>optional. dictionary. If specified, it is a dictionary of
values that should be used as the global scope for the C/C++ code. If
global_dict is not specified the global dictionary of the calling
function is used. </dd>
<dt>force </dt>
<dd>optional. 0 or 1. default 0. If 1, the C++ code is compiled
every time inline is called. This is really only useful for debugging,
and probably only useful if you're editing support_code a lot. </dd>
<dt>compiler </dt>
<dd>optional. string. The name of compiler to use when compiling.
On windows, it understands 'msvc' and 'gcc' as well as all the compiler
names understood by distutils. On Unix, it'll only understand the
values understoof by distutils. (I should add 'gcc' though to this).
<p>On windows, the compiler defaults to the Microsoft C++
compiler. If this isn't available, it looks for mingw32 (the gcc
compiler). </p>
<p>On Unix, it'll probably use the same compiler that was used
when compiling Python. Cygwin's behavior should be similar.</p>
</dd>
<dt>verbose </dt>
<dd>optional. 0,1, or 2. defualt 0. Speficies how much much
information is printed during the compile phase of inlining code. 0 is
silent (except on windows with msvc where it still prints some
garbage). 1 informs you when compiling starts, finishes, and how long
it took. 2 prints out the command lines for the compilation process and
can be useful if you're having problems getting code to work. Its handy
for finding the name of the .cpp file if you need to examine it.
verbose has no affect if the compilation isn't necessary. </dd>
<dt>support_code </dt>
<dd>optional. string. A string of valid C++ code declaring extra
code that might be needed by your compiled function. This could be
declarations of functions, classes, or structures. </dd>
<dt>customize </dt>
<dd>optional. base_info.custom_info object. An alternative way to
specifiy support_code, headers, etc. needed by the function see the
weave.base_info module for more details. (not sure this'll be used
much). </dd>
<dt>type_factories </dt>
<dd>optional. list of type specification factories. These guys are
what convert Python data types to C/C++ data types. If you'd like to
use a different set of type conversions than the default, specify them
here. Look in the type conversions section of the main documentation
for examples. </dd>
<dt>auto_downcast </dt>
<dd>optional. 0 or 1. default 1. This only affects functions that
have Numeric arrays as input variables. Setting this to 1 will cause
all floating point values to be cast as float instead of double if all
the NumPy arrays are of type float. If even one of the arrays has type
double or double complex, all variables maintain there standard types. </dd>
</dl>
</blockquote>
<h4> Distutils keywords:</h4>
<blockquote> <code>inline()</code> also accepts a number of <code>distutils</code>
keywords for controlling how the code is compiled. The following
descriptions have been copied from Greg Ward's <code>distutils.extension.Extension</code>
class doc-
strings for convenience:
<dl>
<dt>sources </dt>
<dd>[string] list of source filenames, relative to the distribution
root (where the setup script lives), in Unix form (slash-separated) for
portability. Source files may be C, C++, SWIG (.i), platform-specific
resource files, or whatever else is recognized by the "build_ext"
command as source for a Python extension. Note: The module_path file is
always appended to the front of this list </dd>
<dt>include_dirs </dt>
<dd>[string] list of directories to search for C/C++ header files
(in Unix form for portability) </dd>
<dt>define_macros </dt>
<dd>[(name : string, value : string|None)] list of macros to
define; each macro is defined using a 2-tuple, where 'value' is either
the string to define it to or None to define it without a particular
value (equivalent of "#define FOO" in source or -DFOO on Unix C
compiler command line) </dd>
<dt>undef_macros </dt>
<dd>[string] list of macros to undefine explicitly </dd>
<dt>library_dirs </dt>
<dd> [string] list of directories to search for C/C++ libraries at
link time </dd>
<dt>libraries </dt>
<dd> [string] list of library names (not filenames or paths) to
link against </dd>
<dt>runtime_library_dirs </dt>
<dd>[string] list of directories to search for C/C++ libraries at
run time
(for shared extensions, this is when the extension is loaded) </dd>
<dt>extra_objects </dt>
<dd>[string] list of extra files to link with (eg. object files not
implied by 'sources', static library that must be explicitly specified,
binary resource files, etc.) </dd>
<dt>extra_compile_args </dt>
<dd>[string] any extra platform- and compiler-specific information
to use when compiling the source files in 'sources'. For platforms and
compilers where "command line" makes sense, this is typically a list of
command-line arguments, but for other platforms it could be anything. </dd>
<dt>extra_link_args </dt>
<dd>[string] any extra platform- and compiler-specific information
to use when linking object files together to create the extension (or
to create a new static Python interpreter). Similar interpretation as
for 'extra_compile_args'. </dd>
<dt>export_symbols </dt>
<dd>[string] list of symbols to be exported from a shared
extension. Not used on all platforms, and not generally necessary for
Python extensions, which typically export exactly one symbol: "init" +
extension_name. </dd>
</dl>
</blockquote>
<a name="Keyword Option Examples"></a>
<h3> Keyword Option Examples</h3>
We'll walk through several examples here to demonstrate the behavior of
<code>inline</code> and also how the various arguments are used.
In the simplest (most) cases, <code>code</code> and <code>arg_names</code>
are the only arguments that need to be specified. Here's a simple
example
run on Windows machine that has Microsoft VC++ installed.
<blockquote>
<pre><code>
&gt;&gt;&gt; from weave import inline
&gt;&gt;&gt; a = 'string'
&gt;&gt;&gt; code = """
... int l = a.length();
... return_val = Py::new_reference_to(Py::Int(l));
... """
&gt;&gt;&gt; inline(code,['a'])
sc_86e98826b65b047ffd2cd5f479c627f12.cpp
Creating
library C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\sc_86e98826b65b047ffd2cd5f479c627f12.lib
and object C:\DOCUME~ 1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\sc_86e98826b65b047ff
d2cd5f479c627f12.exp
6
&gt;&gt;&gt; inline(code,['a'])
6
</code></pre>
</blockquote>
When <code>inline</code> is first run, you'll notice that pause and
some trash printed to the screen. The "trash" is acutually part of the
compilers
output that distutils does not supress. The name of the extension file,
<code>sc_bighonkingnumber.cpp</code>, is generated from the md5 check
sum
of the C/C++ code fragment. On Unix or windows machines with only
gcc installed, the trash will not appear. On the second call, the code
fragment is not compiled since it already exists, and only the answer
is returned. Now kill the interpreter and restart, and run the same
code with
a different string.
<blockquote>
<pre><code>
&gt;&gt;&gt; from weave import inline
&gt;&gt;&gt; a = 'a longer string'
&gt;&gt;&gt; code = """
... int l = a.length();
... return_val = Py::new_reference_to(Py::Int(l));
... """
&gt;&gt;&gt; inline(code,['a'])
15
</code></pre>
</blockquote>
<p>
Notice this time, <code>inline()</code> did not recompile the code
because it
found the compiled function in the persistent catalog of functions.
There is
a short pause as it looks up and loads the function, but it is much
shorter than compiling would require.
</p>
<p>You can specify the local and global dictionaries if you'd like
(much like <code>exec</code> or <code>eval()</code> in Python), but
if they aren't specified, the "expected" ones are used -- i.e. the ones
from the function that called <code>inline() </code>. This is
accomplished through a little call frame trickery. Here is an example
where the local_dict is specified using
the same code example from above: </p>
<blockquote>
<pre><code>
&gt;&gt;&gt; a = 'a longer string'
&gt;&gt;&gt; b = 'an even longer string'
&gt;&gt;&gt; my_dict = {'a':b}
&gt;&gt;&gt; inline(code,['a'])
15
&gt;&gt;&gt; inline(code,['a'],my_dict)
21
</code></pre>
</blockquote>
<p>Everytime, the <code>code</code> is changed, <code>inline</code>
does a recompile. However, changing any of the other options in inline
does not
force a recompile. The <code>force</code> option was added so that one
could force a recompile when tinkering with other variables. In
practice,
it is just as easy to change the <code>code</code> by a single
character
(like adding a space some place) to force the recompile. <em>Note: It
also might be nice to add some methods for purging the cache and on
disk catalogs.</em>
</p>
<p>I use <code>verbose</code> sometimes for debugging. When set to 2,
it'll output all the information (including the name of the .cpp file)
that you'd
expect from running a make file. This is nice if you need to examine
the
generated code to see where things are going haywire. Note that error
messages from failed compiles are printed to the screen even if <code>verbose
</code> is set to 0.
</p>
<p>The following example demonstrates using gcc instead of the standard
msvc compiler on windows using same code fragment as above. Because the
example has already been compiled, the <code>force=1</code> flag is
needed to make <code>inline()</code> ignore the previously compiled
version and recompile using gcc. The verbose flag is added to show what
is printed out: </p>
<blockquote>
<pre><code>
&gt;&gt;&gt;inline(code,['a'],compiler='gcc',verbose=2,force=1)
running build_ext
building 'sc_86e98826b65b047ffd2cd5f479c627f13' extension
c:\gcc-2.95.2\bin\g++.exe -mno-cygwin -mdll -O2 -w -Wstrict-prototypes -IC:
\home\ej\wrk\scipy\weave -IC:\Python21\Include -c C:\DOCUME~1\eric\LOCAL
S~1\Temp\python21_compiled\sc_86e98826b65b047ffd2cd5f479c627f13.cpp -o C:\D
OCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\sc_86e98826b65b04
7ffd2cd5f479c627f13.o
skipping C:\home\ej\wrk\scipy\weave\CXX\cxxextensions.c (C:\DOCUME~1\eri
c\LOCALS~1\Temp\python21_compiled\temp\Release\cxxextensions.o up-to-date)
skipping C:\home\ej\wrk\scipy\weave\CXX\cxxsupport.cxx (C:\DOCUME~1\eric
\LOCALS~1\Temp\python21_compiled\temp\Release\cxxsupport.o up-to-date)
skipping C:\home\ej\wrk\scipy\weave\CXX\IndirectPythonInterface.cxx (C:\
DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\indirectpythonin
terface.o up-to-date)
skipping C:\home\ej\wrk\scipy\weave\CXX\cxx_extensions.cxx (C:\DOCUME~1\
eric\LOCALS~1\Temp\python21_compiled\temp\Release\cxx_extensions.o up-to-da
te)
writing C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\sc_86
e98826b65b047ffd2cd5f479c627f13.def
c:\gcc-2.95.2\bin\dllwrap.exe --driver-name g++ -mno-cygwin -mdll -static -
-output-lib C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\l
ibsc_86e98826b65b047ffd2cd5f479c627f13.a --def C:\DOCUME~1\eric\LOCALS~1\Te
mp\python21_compiled\temp\Release\sc_86e98826b65b047ffd2cd5f479c627f13.def
-s C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\sc_86e9882
6b65b047ffd2cd5f479c627f13.o C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compil
ed\temp\Release\cxxextensions.o C:\DOCUME~1\eric\LOCALS~1\Temp\python21_com
piled\temp\Release\cxxsupport.o C:\DOCUME~1\eric\LOCALS~1\Temp\python21_com
piled\temp\Release\indirectpythoninterface.o C:\DOCUME~1\eric\LOCALS~1\Temp
\python21_compiled\temp\Release\cxx_extensions.o -LC:\Python21\libs -lpytho
n21 -o C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\sc_86e98826b65b047f
fd2cd5f479c627f13.pyd
15
</code></pre>
</blockquote>
That's quite a bit of output. <code>verbose=1</code> just prints the
compile
time.
<blockquote>
<pre><code>
&gt;&gt;&gt;inline(code,['a'],compiler='gcc',verbose=1,force=1)
Compiling code...
finished compiling (sec): 6.00800001621
15
</code></pre>
</blockquote>
<p>
<em> Note: I've only used the <code>compiler</code> option for
switching between 'msvc'
and 'gcc' on windows. It may have use on Unix also, but I don't know
yet.
</em></p>
<p>The <code>support_code</code> argument is likely to be used a lot.
It allows you to specify extra code fragments such as function,
structure or class definitions that you want to use in the <code>code</code>
string. Note that changes to <code>support_code</code> do <em>not</em>
force a recompile. The catalog only relies on <code>code</code> (for
performance reasons) to determine whether recompiling is necessary. So,
if you make a change to support_code, you'll need to alter <code>code</code>
in some way or use the <code>force</code> argument to get the code to
recompile. I usually just add some inocuous whitespace to the end of
one of the lines in <code>code</code> somewhere. Here's an example of
defining a separate method for calculating
the string length: </p>
<blockquote>
<pre><code>
&gt;&gt;&gt; from weave import inline
&gt;&gt;&gt; a = 'a longer string'
&gt;&gt;&gt; support_code = """
... PyObject* length(Py::String a)
... {
... int l = a.length();
... return Py::new_reference_to(Py::Int(l));
... }
... """
&gt;&gt;&gt; inline("return_val = length(a);",['a'],
... support_code = support_code)
15
</code></pre>
</blockquote>
<p>
<code>customize</code> is a left over from a previous way of specifying
compiler options. It is a <code>custom_info</code> object that can
specify quite a bit of information about how a file is compiled. These <code>info</code>
objects are the standard way of defining compile information for type
conversion classes. However, I don't think they are as handy here,
especially since we've exposed all the keyword arguments that distutils
can handle. Between these keywords, and the <code>support_code</code>
option, I think <code>customize</code> may be obsolete. We'll see if
anyone cares to use it. If not, it'll get axed in the next version.
</p>
<p>The <code>type_factories</code> variable is important to people who
want to
customize the way arguments are converted from Python to C. We'll talk
about
this in the next chapter **xx** of this document when we discuss type
conversions.
</p>
<p><code>auto_downcast</code> handles one of the big type conversion
issues that
is common when using NumPy arrays in conjunction with Python scalar
values.
If you have an array of single precision values and multiply that array
by a Python scalar, the result is upcast to a double precision array
because the
scalar value is double precision. This is not usually the desired
behavior
because it can double your memory usage. <code>auto_downcast</code>
goes
some distance towards changing the casting precedence of arrays and
scalars.
If your only using single precision arrays, it will automatically
downcast all
scalar values from double to single precision when they are passed into
the
C++ code. This is the default behavior. If you want all values to keep
there
default type, set <code>auto_downcast</code> to 0.
</p>
<p><a name="Returning Values"></a>
</p>
<h3> Returning Values</h3>
Python variables in the local and global scope transfer seemlessly from
Python into the C++ snippets. And, if <code>inline</code> were to
completely live up
to its name, any modifications to variables in the C++ code would be
reflected
in the Python variables when control was passed back to Python. For
example,
the desired behavior would be something like:
<blockquote>
<pre><code>
# THIS DOES NOT WORK
&gt;&gt;&gt; a = 1
&gt;&gt;&gt; weave.inline("a++;",['a'])
&gt;&gt;&gt; a
2
</code></pre>
</blockquote>
Instead you get:
<blockquote>
<pre><code>
&gt;&gt;&gt; a = 1
&gt;&gt;&gt; weave.inline("a++;",['a'])
&gt;&gt;&gt; a
1
</code></pre>
</blockquote>
Variables are passed into C++ as if you are calling a Python function.
Python's calling convention is sometimes called "pass by assignment".
This means its as if a <code>c_a = a</code> assignment is made right
before <code>inline</code> call is made and the <code>c_a</code>
variable is used within the C++ code. Thus, any changes made to <code>c_a</code>
are not reflected in Python's <code>a</code> variable. Things do get a
little more confusing, however, when looking at variables with mutable
types. Changes made in C++ to the contents of mutable types <em>are</em>
reflected in the Python variables.
<blockquote>
<pre><code>
&gt;&gt;&gt; a= [1,2]
&gt;&gt;&gt; weave.inline("PyList_SetItem(a.ptr(),0,PyInt_FromLong(3));",['a'])
&gt;&gt;&gt; print a
[3, 2]
</code></pre>
</blockquote>
So modifications to the contents of mutable types in C++ are seen when
control
is returned to Python. Modifications to immutable types such as tuples,
strings, and numbers do not alter the Python variables.
If you need to make changes to an immutable variable, you'll need to
assign
the new value to the "magic" variable <code>return_val</code> in C++.
This
value is returned by the <code>inline()</code> function:
<blockquote>
<pre><code>
&gt;&gt;&gt; a = 1
&gt;&gt;&gt; a = weave.inline("return_val = Py::new_reference_to(Py::Int(a+1));",['a'])
&gt;&gt;&gt; a
2
</code></pre>
</blockquote>
The <code>return_val</code> variable can also be used to return newly
created values. This is possible by returning a tuple. The following
trivial example illustrates how this can be done:
<blockquote>
<pre><code> <br> # python version<br> def multi_return():<br> return 1, '2nd'<br> <br> # C version.<br> def c_multi_return(): <br> code = """<br> py::tuple results(2);<br> results[0] = 1;<br> results[1] = "2nd";<br> return_val = results; <br> """<br> return inline_tools.inline(code)<br> </code></pre>
</blockquote>
<p> The example is available in <code>examples/tuple_return.py</code>.
It also
has the dubious honor of demonstrating how much <code>inline()</code>
can slow things down. The C version here is about 7-10 times slower
than
the Python
version. Of course, something so trivial has no reason to be written in
C anyway.
<a name="The issue with locals()"></a></p>
<h4> The issue with <code>locals()</code></h4>
<p>
<code>inline</code> passes the <code>locals()</code> and <code>globals()</code>
dictionaries from Python into the C++ function from the calling
function. It extracts the variables that are used in the C++ code from
these dictionaries, converts then to C++ variables, and then calculates
using them. It seems like it would be trivial, then, after the
calculations were finished to then insert the new values back into the <code>locals()</code>
and <code>globals()</code> dictionaries so that the modified values
were reflected in Python. Unfortunately, as pointed out by the Python
manual, the locals() dictionary is not writable. </p>
<p><em>
I suspect <code>locals()</code> is not writable because there are some
optimizations done to speed lookups of the local namespace. I'm
guessing local lookups don't always look at a dictionary to find
values. Can someone "in the know" confirm or correct this? Another
thing I'd like to know is whether there is a way to write to the local
namespace of another stack frame from C/C++. If so, it would be
possible to have some clean up code in compiled functions that wrote
final values of variables in C++ back to the correct Python stack
frame. I think this goes a long way toward making <code>inline</code>
truely live up to its name. I don't think we'll get to the point of
creating variables in Python for variables created in C -- although I
suppose with a C/C++ parser you could do that also.
</em></p>
<p><a name="inline_quick_look_at_code"></a>
</p>
<h3>A quick look at the code</h3>
<code>weave</code> generates a C++ file holding an extension function
for each <code>inline</code> code snippet. These file names are
generated using from the md5 signature of the code snippet and saved to
a location specified by the PYTHONCOMPILED environment variable
(discussed later). The cpp files are generally about 200-400 lines long
and include quite a few functions to support type conversions, etc.
However, the actual compiled function is pretty simple. Below is the
familiar <code>printf</code> example:
<blockquote>
<pre><code>
&gt;&gt;&gt; import weave
&gt;&gt;&gt; a = 1
&gt;&gt;&gt; weave.inline('printf("%d\\n",a);',['a'])
1
</code></pre>
</blockquote>
And here is the extension function generated by <code>inline</code>:
<blockquote>
<pre><code>
static PyObject* compiled_func(PyObject*self, PyObject* args)<br>{<br> py::object return_val;<br> int exception_occurred = 0;<br> PyObject *py__locals = NULL;<br> PyObject *py__globals = NULL;<br> PyObject *py_a;<br> py_a = NULL;<br><br> if(!PyArg_ParseTuple(args,"OO:compiled_func",&amp;py__locals,&amp;py__globals))<br> return NULL;<br> try<br> {<br> PyObject* raw_locals = py_to_raw_dict(py__locals,"_locals");<br> PyObject* raw_globals = py_to_raw_dict(py__globals,"_globals");<br> /* argument conversion code */<br> py_a = get_variable("a",raw_locals,raw_globals);<br> int a = convert_to_int(py_a,"a");<br> /* inline code */<br> /* NDARRAY API VERSION 90907 */<br> printf("%d\n",a); /*I would like to fill in changed locals and globals here...*/<br><br> }<br> catch(...)<br> {<br> return_val = py::object();<br> exception_occurred = 1;<br> }<br> /* cleanup code */<br> if(!(PyObject*)return_val &amp;&amp; !exception_occurred)<br> {<br><br> return_val = Py_None;<br> }<br><br> return return_val.disown();<br>}<br> </code><br></pre>
</blockquote>
Every inline function takes exactly two arguments -- the local and
global
dictionaries for the current scope. All variable values are looked up
out
of these dictionaries. The lookups, along with all <code>inline</code>
code execution, are done within a C++ <code>try</code> block. If the
variables
aren't found, or there is an error converting a Python variable to the
appropriate type in C++, an exception is raised. The C++ exception
is automatically converted to a Python exception by SCXX and returned
to
Python.
The <code>py_to_int()</code> function illustrates how the conversions
and
exception handling works. py_to_int first checks that the given
PyObject*
pointer is not NULL and is a Python integer. If all is well, it calls
the
Python API to convert the value to an <code>int</code>. Otherwise, it
calls
<code>handle_bad_type()</code> which gathers information about what
went wrong
and then raises a SCXX TypeError which returns to Python as a
TypeError.
<blockquote>
<pre><code>
int py_to_int(PyObject* py_obj,char* name)
{
if (!py_obj || !PyInt_Check(py_obj))
handle_bad_type(py_obj,"int", name);
return (int) PyInt_AsLong(py_obj);
}
</code></pre>
</blockquote>
<blockquote>
<pre><code>
void handle_bad_type(PyObject* py_obj, char* good_type, char* var_name)
{
char msg[500];
sprintf(msg,"received '%s' type instead of '%s' for variable '%s'",
find_type(py_obj),good_type,var_name);
throw Py::TypeError(msg);
}
char* find_type(PyObject* py_obj)
{
if(py_obj == NULL) return "C NULL value";
if(PyCallable_Check(py_obj)) return "callable";
if(PyString_Check(py_obj)) return "string";
if(PyInt_Check(py_obj)) return "int";
if(PyFloat_Check(py_obj)) return "float";
if(PyDict_Check(py_obj)) return "dict";
if(PyList_Check(py_obj)) return "list";
if(PyTuple_Check(py_obj)) return "tuple";
if(PyFile_Check(py_obj)) return "file";
if(PyModule_Check(py_obj)) return "module";
//should probably do more interagation (and thinking) on these.
if(PyCallable_Check(py_obj) &amp;&amp; PyInstance_Check(py_obj)) return "callable";
if(PyInstance_Check(py_obj)) return "instance";
if(PyCallable_Check(py_obj)) return "callable";
return "unknown type";
}
</code></pre>
</blockquote>
Since the <code>inline</code> is also executed within the <code>try/catch</code>
block, you can use CXX exceptions within your code. It is usually a bad
idea
to directly <code>return</code> from your code, even if an error
occurs. This
skips the clean up section of the extension function. In this simple
example,
there isn't any clean up code, but in more complicated examples, there
may
be some reference counting that needs to be taken care of here on
converted
variables. To avoid this, either uses exceptions or set <code>return_val</code>
to NULL and use <code>if/then's</code> to skip code
after errors.
<a name="inline_technical_details"></a>
<h2> Technical Details </h2>
<p>
There are several main steps to using C/C++ code withing Python:
</p>
<ol>
<li>Type conversion </li>
<li>Generating C/C++ code </li>
<li>Compile the code to an extension module </li>
<li>Catalog (and cache) the function for future use</li>
</ol>
<p>
Items 1 and 2 above are related, but most easily discussed separately.
Type conversions are customizable by the user if needed. Understanding
them is pretty important for anything beyond trivial uses of <code>inline</code>.
Generating the C/C++ code is handled by <code>ext_function</code> and <code>ext_module</code>
classes and . For the most part, compiling the code is handled by
distutils. Some customizations were needed, but they were relatively
minor and do not require changes to distutils itself. Cataloging is
pretty simple in concept, but surprisingly required the most code to
implement (and still likely needs some work). So, this section covers
items 1 and 4 from the list. Item 2 is covered later in the chapter
covering the <code>ext_tools</code> module, and distutils is covered
by a completely separate document xxx.
</p>
<h2>Passing Variables in/out of the C/C++ code</h2>
<em>
Note: Passing variables into the C code is pretty straight forward, but
there are subtlties to how variable modifications in C are returned to
Python. see <a href="#Returning%20Values">Returning Values</a> for a
more thorough discussion of this issue.
</em> <a name="Converting Types"></a>
<h2>Type Conversions</h2>
<em>
Note: Maybe <code>xxx_converter</code> instead of <code>xxx_specification</code>
is a more descriptive name. Might change in future version?
</em>
<p>By default, <code>inline()</code> makes the following type
conversions between
Python and C++ types.
</p>
<p></p>
<center>
<table border="1" width="100%">
<tbody>
<tr>
<td colspan="2" width="100%">
<p align="center">Default Data Type Conversions</p>
</td>
</tr>
<tr>
<td>
<p align="center">Python</p>
</td>
<td>
<p align="center">C++</p>
</td>
</tr>
<tr>
<td>&nbsp;&nbsp; int</td>
<td>&nbsp;&nbsp; int</td>
</tr>
<tr>
<td>&nbsp;&nbsp; float</td>
<td>&nbsp;&nbsp; double</td>
</tr>
<tr>
<td>&nbsp;&nbsp; complex</td>
<td>&nbsp;&nbsp; std::complex<double></double></td>
</tr>
<tr>
<td>&nbsp;&nbsp; string</td>
<td>&nbsp;&nbsp; py::string</td>
</tr>
<tr>
<td>&nbsp;&nbsp; list</td>
<td>&nbsp;&nbsp; py::list</td>
</tr>
<tr>
<td>&nbsp;&nbsp; dict</td>
<td>&nbsp;&nbsp; py::dict</td>
</tr>
<tr>
<td>&nbsp;&nbsp; tuple</td>
<td>&nbsp;&nbsp; py::tuple</td>
</tr>
<tr>
<td>&nbsp;&nbsp; file</td>
<td>&nbsp;&nbsp; FILE*</td>
</tr>
<tr>
<td>&nbsp;&nbsp; callable</td>
<td>&nbsp;&nbsp; py::object</td>
</tr>
<tr>
<td>&nbsp;&nbsp; instance</td>
<td>&nbsp;&nbsp; py::object</td>
</tr>
<tr>
<td>&nbsp;&nbsp; numpy.ndarray</td>
<td>&nbsp;&nbsp; PyArrayObject*</td>
</tr>
<tr>
<td>&nbsp;&nbsp; wxXXX</td>
<td>&nbsp;&nbsp; wxXXX*</td>
</tr>
</tbody>
</table>
</center>
<p>
The <code>Py::</code> namespace is defined by the SCXX library which
has C++ class
equivalents for many Python types. <code>std::</code> is the namespace
of the
standard library in C++.
</p>
<p><em>
Note:
<ul>
<li>I haven't figured out how to handle <code>long int</code> yet (I
think they are currenlty converted to int - - check this). </li>
<li>Hopefully VTK will be added to the list soon</li>
</ul>
</em>
</p>
<p>Python to C++ conversions fill in code in several locations in the
generated
<code>inline</code> extension function. Below is the basic template for
the
function. This is actually the exact code that is generated by calling
<code>weave.inline("")</code>. </p>
<blockquote>
<pre><code>
</code></pre>
</blockquote>
The <code>/* inline code */</code> section is filled with the code
passed to
the <code>inline()</code> function call. The <code>/*argument
convserion code*/</code> and <code>/* cleanup code */</code>
sections are filled with code that handles conversion from Python to
C++
types and code that deallocates memory or manipulates reference counts
before
the function returns. The following sections demostrate how these two
areas
are filled in by the default conversion methods.
<em> Note: I'm not sure I have reference counting correct on a few of
these. The only thing I increase/decrease the ref count on is NumPy
arrays. If you
see an issue, please let me know.
</em><a name="inline_numeric_argument_conversion"></a>
<h3> NumPy Argument Conversion </h3>
Integer, floating point, and complex arguments are handled in a very
similar
fashion. Consider the following inline function that has a single
integer variable passed in:
<blockquote>
<pre><code>
&gt;&gt;&gt; a = 1
&gt;&gt;&gt; inline("",['a'])
</code></pre>
</blockquote>
The argument conversion code inserted for <code>a</code> is:
<blockquote>
<pre><code>
/* argument conversion code */
int a = py_to_int (get_variable("a",raw_locals,raw_globals),"a");
</code></pre>
</blockquote>
<code>get_variable()</code> reads the variable <code>a</code>
from the local and global namespaces. <code>py_to_int()</code> has the
following
form:
<blockquote>
<pre><code>
static int py_to_int(PyObject* py_obj,char* name)
{
if (!py_obj || !PyInt_Check(py_obj))
handle_bad_type(py_obj,"int", name);
return (int) PyInt_AsLong(py_obj);
}
</code></pre>
</blockquote>
Similarly, the float and complex conversion routines look like:
<blockquote>
<pre><code> <br> static double py_to_float(PyObject* py_obj,char* name)<br> {<br> if (!py_obj || !PyFloat_Check(py_obj))<br> handle_bad_type(py_obj,"float", name);<br> return PyFloat_AsDouble(py_obj);<br> }<br> <br> static std::complex<double> py_to_complex(PyObject* py_obj,char* name)<br> {<br> if (!py_obj || !PyComplex_Check(py_obj))<br> handle_bad_type(py_obj,"complex", name);<br> return std::complex<double>(PyComplex_RealAsDouble(py_obj),<br> PyComplex_ImagAsDouble(py_obj)); <br> }<br> </double></double></code></pre>
</blockquote>
NumPy conversions do not require any clean up code.
<a name="inline_python_argument_conversion"></a>
<h3> String, List, Tuple, and Dictionary Conversion </h3>
Strings, Lists, Tuples and Dictionary conversions are all converted to
SCXX types by default.
For the following code,
<blockquote>
<pre><code>
&gt;&gt;&gt; a = [1]
&gt;&gt;&gt; inline("",['a'])
</code></pre>
</blockquote>
The argument conversion code inserted for <code>a</code> is:
<blockquote>
<pre><code>
/* argument conversion code */
Py::List a = py_to_list (get_variable("a",raw_locals,raw_globals),"a");
</code></pre>
</blockquote>
<code>get_variable()</code> reads the variable <code>a</code>
from the local and global namespaces. <code>py_to_list()</code> and
its
friends has the following form:
<blockquote>
<pre><code> <br> static Py::List py_to_list(PyObject* py_obj,char* name)<br> {<br> if (!py_obj || !PyList_Check(py_obj))<br> handle_bad_type(py_obj,"list", name);<br> return Py::List(py_obj);<br> }<br> <br> static Py::String py_to_string(PyObject* py_obj,char* name)<br> {<br> if (!PyString_Check(py_obj))<br> handle_bad_type(py_obj,"string", name);<br> return Py::String(py_obj);<br> }<br><br> static Py::Dict py_to_dict(PyObject* py_obj,char* name)<br> {<br> if (!py_obj || !PyDict_Check(py_obj))<br> handle_bad_type(py_obj,"dict", name);<br> return Py::Dict(py_obj);<br> }<br> <br> static Py::Tuple py_to_tuple(PyObject* py_obj,char* name)<br> {<br> if (!py_obj || !PyTuple_Check(py_obj))<br> handle_bad_type(py_obj,"tuple", name);<br> return Py::Tuple(py_obj);<br> }<br> </code></pre>
</blockquote>
SCXX handles reference counts on for strings, lists, tuples, and
dictionaries,
so clean up code isn't necessary.
<a name="#inline_file_argument_conversion"></a>
<h3> File Conversion </h3>
For the following code,
<blockquote>
<pre><code>
&gt;&gt;&gt; a = open("bob",'w')
&gt;&gt;&gt; inline("",['a'])
</code></pre>
</blockquote>
The argument conversion code is:
<blockquote>
<pre><code>
/* argument conversion code */
PyObject* py_a = get_variable("a",raw_locals,raw_globals);
FILE* a = py_to_file(py_a,"a");
</code></pre>
</blockquote>
<code>get_variable()</code> reads the variable <code>a</code>
from the local and global namespaces. <code>py_to_file()</code>
converts
PyObject* to a FILE* and increments the reference count of the
PyObject*:
<blockquote>
<pre><code>
FILE* py_to_file(PyObject* py_obj, char* name)
{
if (!py_obj || !PyFile_Check(py_obj))
handle_bad_type(py_obj,"file", name);
Py_INCREF(py_obj);
return PyFile_AsFile(py_obj);
}
</code></pre>
</blockquote>
Because the PyObject* was incremented, the clean up code needs to
decrement
the counter
<blockquote>
<pre><code>
/* cleanup code */
Py_XDECREF(py_a);
</code></pre>
</blockquote>
Its important to understand that file conversion only works on actual
files --
i.e. ones created using the <code>open()</code> command in Python. It
does
not support converting arbitrary objects that support the file
interface into
C <code>FILE*</code> pointers. This can affect many things. For
example, in
initial <code>printf()</code> examples, one might be tempted to solve
the problem of C and Python IDE's (PythonWin, PyCrust, etc.) writing to
different
stdout and stderr by using <code>fprintf()</code> and passing in <code>sys.stdout</code>
and <code>sys.stderr</code>. For example, instead of
<blockquote>
<pre><code>
&gt;&gt;&gt; weave.inline('printf("hello\\n");')
</code></pre>
</blockquote>
You might try:
<blockquote>
<pre><code>
&gt;&gt;&gt; buf = sys.stdout
&gt;&gt;&gt; weave.inline('fprintf(buf,"hello\\n");',['buf'])
</code></pre>
</blockquote>
This will work as expected from a standard python interpreter, but in
PythonWin,
the following occurs:
<blockquote>
<pre><code>
&gt;&gt;&gt; buf = sys.stdout
&gt;&gt;&gt; weave.inline('fprintf(buf,"hello\\n");',['buf'])
Traceback (most recent call last):
File "<interactive
input="">", line 1, in ?<br> File "C:\Python21\weave\inline_tools.py", line 315, in inline<br> auto_downcast = auto_downcast,<br> File "C:\Python21\weave\inline_tools.py", line 386, in compile_function<br> type_factories = type_factories)<br> File "C:\Python21\weave\ext_tools.py", line 197, in __init__<br> auto_downcast, type_factories)<br> File "C:\Python21\weave\ext_tools.py", line 390, in assign_variable_types<br> raise TypeError, format_error_msg(errors)<br> TypeError: {'buf': "Unable to convert variable 'buf' to a C++ type."}<br> </interactive></code></pre>
</blockquote>
The traceback tells us that <code>inline()</code> was unable to
convert 'buf' to a
C++ type (If instance conversion was implemented, the error would have
occurred at runtime instead). Why is this? Let's look at what the <code>buf</code>
object really is:
<blockquote>
<pre><code>
&gt;&gt;&gt; buf
pywin.framework.interact.InteractiveView instance at 00EAD014
</code></pre>
</blockquote>
PythonWin has reassigned <code>sys.stdout</code> to a special object
that implements the Python file interface. This works great in Python,
but since the special object doesn't have a FILE* pointer underlying
it, fprintf doesn't know what to do with it (well this will be the
problem when instance conversion is implemented...).
<a name="#inline_callable_argument_conversion"></a>
<h3> Callable, Instance, and Module Conversion </h3>
<em>Note: Need to look into how ref counts should be handled. Also,
Instance and Module conversion are not currently implemented.
</em>
<blockquote>
<pre><code>
&gt;&gt;&gt; def a():
pass
&gt;&gt;&gt; inline("",['a'])
</code></pre>
</blockquote>
Callable and instance variables are converted to PyObject*. Nothing is
done
to there reference counts.
<blockquote>
<pre><code>
/* argument conversion code */
PyObject* a = py_to_callable(get_variable("a",raw_locals,raw_globals),"a");
</code></pre>
</blockquote>
<code>get_variable()</code> reads the variable <code>a</code>
from the local and global namespaces. The <code>py_to_callable()</code>
and
<code>py_to_instance()</code> don't currently increment the ref count.
<blockquote>
<pre><code> <br> PyObject* py_to_callable(PyObject* py_obj, char* name)<br> {<br> if (!py_obj || !PyCallable_Check(py_obj))<br> handle_bad_type(py_obj,"callable", name); <br> return py_obj;<br> }<br><br> PyObject* py_to_instance(PyObject* py_obj, char* name)<br> {<br> if (!py_obj || !PyFile_Check(py_obj))<br> handle_bad_type(py_obj,"instance", name); <br> return py_obj;<br> }<br> </code></pre>
</blockquote>
There is no cleanup code for callables, modules, or instances.
<a name="#Customizing Conversions"></a>
<h3> Customizing Conversions </h3>
<p>
Converting from Python to C++ types is handled by xxx_specification
classes. A type specification class actually serve in two related but
different roles. The first is in determining whether a Python variable
that needs to be converted should be represented by the given class.
The second is as a code generator that generate C++ code needed to
convert from Python to C++ types for a specific variable.
</p>
<p>When </p>
<blockquote>
<pre><code>
&gt;&gt;&gt; a = 1
&gt;&gt;&gt; weave.inline('printf("%d",a);',['a'])
</code></pre>
</blockquote>
is called for the first time, the code snippet has to be compiled. In
this process, the variable 'a' is tested against a list of type
specifications (the default list is stored in weave/ext_tools.py). The <em>first</em>
specification in the list is used to represent the variable.
<p>Examples of <code>xxx_specification</code> are scattered throughout
numerous "xxx_spec.py" files in the <code>weave</code> package.
Closely related to the <code>xxx_specification</code> classes are <code>yyy_info</code>
classes. These classes contain compiler, header, and support code
information necessary for including a certain set of capabilities (such
as blitz++ or CXX support)
in a compiled module. <code>xxx_specification</code> classes have one
or more
<code>yyy_info</code> classes associated with them.
If you'd like to define your own set of type specifications, the
current best route
is to examine some of the existing spec and info files. Maybe looking
over
sequence_spec.py and cxx_info.py are a good place to start. After
defining specification classes, you'll need to pass them into <code>inline</code>
using the <code>type_factories</code> argument. A lot of times you may
just want to change how a specific variable type is represented. Say
you'd rather have Python strings converted to <code>std::string</code>
or maybe <code>char*</code> instead of using the CXX string object,
but would like all other type conversions to have default behavior.
This requires that a new specification class that handles strings
is written and then prepended to a list of the default type
specifications. Since
it is closer to the front of the list, it effectively overrides the
default
string specification.
The following code demonstrates how this is done:
...
<a name="The Catalog"></a></p>
<h2> The Catalog </h2>
<p>
<code>catalog.py</code> has a class called <code>catalog</code> that
helps keep track of previously compiled functions. This prevents <code>inline()</code>
and related functions from having to compile functions everytime they
are called. Instead, catalog will check an in memory cache to see if
the function has already been loaded into python. If it hasn't, then it
starts searching through persisent catalogs on disk to see if it finds
an entry for the given function. By saving information about compiled
functions to disk, it isn't
necessary to re-compile functions everytime you stop and restart the
interpreter.
Functions are compiled once and stored for future use.
</p>
<p>When <code>inline(cpp_code)</code> is called the following things
happen:
</p>
<ol>
<li> A fast local cache of functions is checked for the last function
called for <code>cpp_code</code>. If an entry for <code>cpp_code</code>
doesn't exist in the cache or the cached function call fails (perhaps
because the function doesn't have compatible types) then the next step
is to check the catalog. </li>
<li> The catalog class also keeps an in-memory cache with a list of
all the functions compiled for <code>cpp_code</code>. If <code>cpp_code</code>
has ever been called, then this cache will be present (loaded from
disk). If the cache isn't present, then it is loaded from disk.
<p> If the cache is present, each function in the cache is called
until one is found that was compiled for the correct argument types. If
none of the functions work, a new function is compiled with the given
argument types. This function is written to the on-disk catalog as well
as into the in-memory cache.</p>
</li>
<li> When a lookup for <code>cpp_code</code> fails, the catalog
looks through the on-disk function catalogs for the entries. The
PYTHONCOMPILED variable determines where to search for these catalogs
and in what order. If PYTHONCOMPILED is not present several platform
dependent locations are searched. All functions found for <code>cpp_code</code>
in the path are loaded into the in-memory cache with functions found
earlier in the search path closer to the front of the call list.
<p> If the function isn't found in the on-disk catalog, then the
function is compiled, written to the first writable directory in the
PYTHONCOMPILED path, and also loaded into the in-memory cache.</p>
</li>
</ol>
<a name="function storage"></a>
<h3> Function Storage: How functions are stored in caches and on disk </h3>
<p>
Function caches are stored as dictionaries where the key is the entire
C++
code string and the value is either a single function (as in the "level
1"
cache) or a list of functions (as in the main catalog cache). On disk
catalogs are stored in the same manor using standard Python shelves.
</p>
<p>Early on, there was a question as to whether md5 check sums of the
C++
code strings should be used instead of the actual code strings. I think
this
is the route inline Perl took. Some (admittedly quick) tests of the md5
vs.
the entire string showed that using the entire string was at least a
factor of 3 or 4 faster for Python. I think this is because it is more
time consuming to compute the md5 value than it is to do look-ups of
long
strings in the dictionary. Look at the examples/md5_speed.py file for
the
test run. <a name="PYTHONCOMPILED"></a>
</p>
<h3> Catalog search paths and the PYTHONCOMPILED variable</h3>
<p>
The default location for catalog files on Unix is is
~/.pythonXX_compiled where XX is version of Python being used. If this
directory doesn't exist, it is created the first time a catalog is
used. The directory must be writable. If, for any reason it isn't, then
the catalog attempts to create a directory based on your user id in the
/tmp directory. The directory permissions are set so that only you have
access to the directory. If this fails, I think you're out of luck. I
don't think either of these should ever fail though. On Windows, a
directory called pythonXX_compiled is created in the user's temporary
directory. </p>
<p>The actual catalog file that lives in this directory is a Python
shelve with
a platform specific name such as "nt21compiled_catalog" so that
multiple OSes
can share the same file systems without trampling on each other. Along
with
the catalog file, the .cpp and .so or .pyd files created by inline will
live
in this directory. The catalog file simply contains keys which are the
C++
code strings with values that are lists of functions. The function
lists point
at functions within these compiled modules. Each function in the lists
executes the same C++ code string, but compiled for different input
variables.
</p>
<p>You can use the PYTHONCOMPILED environment variable to specify
alternative
locations for compiled functions. On Unix this is a colon (':')
separated
list of directories. On windows, it is a (';') separated list of
directories.
These directories will be searched prior to the default directory for a
compiled function catalog. Also, the first writable directory in the
list
is where all new compiled function catalogs, .cpp and .so or .pyd files
are
written. Relative directory paths ('.' and '..') should work fine in
the
PYTHONCOMPILED variable as should environement variables.
</p>
<p>There is a "special" path variable called MODULE that can be placed
in the PYTHONCOMPILED variable. It specifies that the compiled catalog
should
reside in the same directory as the module that called it. This is
useful
if an admin wants to build a lot of compiled functions during the build
of a package and then install them in site-packages along with the
package.
User's who specify MODULE in their PYTHONCOMPILED variable will have
access
to these compiled functions. Note, however, that if they call the
function
with a set of argument types that it hasn't previously been built for,
the
new function will be stored in their default directory (or some other
writable
directory in the PYTHONCOMPILED path) because the user will not have
write
access to the site-packages directory.
</p>
<p>An example of using the PYTHONCOMPILED path on bash follows: </p>
<blockquote>
<pre><code>
PYTHONCOMPILED=MODULE:/some/path;export PYTHONCOMPILED;
</code></pre>
</blockquote>
If you are using python21 on linux, and the module bob.py in
site-packages
has a compiled function in it, then the catalog search order when
calling that
function for the first time in a python session would be:
<blockquote>
<pre><code>
/usr/lib/python21/site-packages/linuxpython_compiled
/some/path/linuxpython_compiled
~/.python21_compiled/linuxpython_compiled
</code></pre>
</blockquote>
The default location is always included in the search path.
<p><em> Note: hmmm. see a possible problem here. I should probably make
a sub-
directory such as /usr/lib/python21/site-
packages/python21_compiled/linuxpython_compiled so that library files
compiled with python21 are tried to link with python22 files in some
strange scenarios. Need to check this.
</em></p>
<p>The in-module cache (in <code>weave.inline_tools</code> reduces the
overhead of calling inline functions by about a factor of 2. It can be
reduced a little more for type loop calls where the same function is
called over and over again if the cache was a single value instead of a
dictionary, but the benefit is very small (less than 5%) and the
utility is quite a bit less. So, we'll stick with a dictionary as the
cache.
</p>
<p></p>
<a name="Blitz"></a>
<h1>Blitz</h1>
<em> Note: most of this section is lifted from old documentation. It
should be
pretty accurate, but there may be a few discrepancies.</em>
<p><code>weave.blitz()</code> compiles NumPy Python expressions for
fast execution. For most applications, compiled expressions should
provide a factor of 2-10 speed-up over NumPy arrays. Using compiled
expressions is meant to be as unobtrusive as possible and works much
like pythons exec statement. As an example, the following code fragment
takes a 5 point average of the 512x512 2d image, b, and stores it in
array, a: </p>
<blockquote>
<pre><code>
from scipy import * # or from NumPy import *
a = ones((512,512), Float64)
b = ones((512,512), Float64)
# ...do some stuff to fill in b...
# now average
a[1:-1,1:-1] = (b[1:-1,1:-1] + b[2:,1:-1] + b[:-2,1:-1] \
+ b[1:-1,2:] + b[1:-1,:-2]) / 5.
</code></pre>
</blockquote>
To compile the expression, convert the expression to a string by
putting
quotes around it and then use <code>weave.blitz</code>:
<blockquote>
<pre><code>
import weave
expr = "a[1:-1,1:-1] = (b[1:-1,1:-1] + b[2:,1:-1] + b[:-2,1:-1]" \
"+ b[1:-1,2:] + b[1:-1,:-2]) / 5."
weave.blitz(expr)
</code></pre>
</blockquote>
The first time <code>weave.blitz</code> is run for a given expression
and set of arguements, C++ code that accomplishes the exact same task
as the Python expression is generated and compiled to an extension
module. This can take up to a couple of minutes depending on the
complexity of the function. Subsequent calls to the function are very
fast. Futher, the generated module is saved between program executions
so that the compilation is only done once for a given expression and
associated set of array types. If the given expression
is executed with a new set of array types, the code most be compiled
again. This
does not overwrite the previously compiled function -- both of them are
saved and
available for exectution.
<p>The following table compares the run times for standard NumPy code
and compiled code for the 5 point averaging.
</p>
<p></p>
<center>
<table border="1">
<tbody>
<tr>
<td>Method</td>
<td>Run Time (seconds)</td>
</tr>
<tr>
<td>Standard NumPy</td>
<td>0.46349</td>
</tr>
<tr>
<td>blitz (1st time compiling)</td>
<td> 78.95526</td>
</tr>
<tr>
<td>blitz (subsequent calls)</td>
<td>0.05843 (factor of 8 speedup)</td>
</tr>
</tbody>
</table>
</center>
<p>
These numbers are for a 512x512 double precision image run on a 400 MHz
Celeron processor under RedHat Linux 6.2.
</p>
<p>Because of the slow compile times, its probably most effective to
develop algorithms as you usually do using the capabilities of scipy or
the NumPy module. Once the algorithm is perfected, put quotes around it
and execute it using <code>weave.blitz</code>. This provides the
standard rapid prototyping strengths of Python and results in
algorithms that run close to that of hand coded C or Fortran.
<a name="blitz_requirements"></a></p>
<h2>Requirements</h2>
Currently, the <code>weave.blitz</code> has only been tested under
Linux with gcc-2.95-3 and on Windows with Mingw32 (2.95.2). Its
compiler requirements are pretty heavy duty (see the <a
href="http://www.oonumerics.org/blitz/">blitz++ home page</a>), so it
won't work with just any compiler. Particularly MSVC++ isn't up to
snuff. A number of other compilers such as KAI++ will also work, but my
suspicions are that gcc will get the most use.
<a name="blitz_limitations"></a>
<h2>Limitations</h2>
<ol>
<li>Currently, <code>weave.blitz</code> handles all standard
mathematic
operators except for the ** power operator. The built-in
trigonmetric, log, floor/ceil, and fabs functions might work (but
haven't been tested). It also handles all types of array indexing
supported by the NumPy module. numarray's NumPy compatible array
indexing modes are likewise supported, but numarray's enhanced
(array based) indexing modes are not supported.
<p><code>weave.blitz</code> does not currently support operations
that use array broadcasting, nor have any of the special purpose
functions in NumPy such as take, compress, etc. been implemented. Note
that there are no obvious reasons why most of this functionality cannot
be added to scipy.weave, so it will likely trickle into future
versions. Using <code>slice()</code> objects directly instead of <code>start:stop:step</code>
is also not supported. </p>
</li>
<li>Currently Python only works on expressions that include
assignment such as
<blockquote>
<pre><code>
&gt;&gt;&gt; result = b + c + d
</code></pre>
</blockquote>
This means that the result array must exist before calling <code>weave.blitz</code>.
Future versions will allow the following:
<blockquote>
<pre><code>
&gt;&gt;&gt; result = weave.blitz_eval("b + c + d")
</code></pre>
</blockquote>
</li>
<li> <code>weave.blitz</code> works best when algorithms can be
expressed in a "vectorized" form. Algorithms that have a large number
of if/thens and other conditions are better hand written in C or
Fortran. Further, the restrictions imposed by requiring vectorized
expressions sometimes preclude the use of more efficient data
structures or algorithms. For maximum speed in these cases, hand-coded
C or Fortran code is the only way to go. </li>
<li> <code>weave.blitz</code> can produce different results than
NumPy
in certain situations. It can happen when the array receiving the
results of a calculation is also used during the calculation. The NumPy
behavior is to carry out the entire calculation on the right hand side
of an equation and store it in a temporary array. This temprorary array
is assigned to the array on the left hand side of the equation. blitz,
on the other hand, does a "running" calculation of the array elements
assigning values from the right hand
side to the elements on the left hand side immediately after they are
calculated.
Here is an example, provided by Prabhu Ramachandran, where this
happens:
<blockquote>
<pre><code>
# 4 point average.
&gt;&gt;&gt; expr = "u[1:-1, 1:-1] = (u[0:-2, 1:-1] + u[2:, 1:-1] + "\
... "u[1:-1,0:-2] + u[1:-1, 2:])*0.25"
&gt;&gt;&gt; u = zeros((5, 5), 'd'); u[0,:] = 100
&gt;&gt;&gt; exec (expr)
&gt;&gt;&gt; u
array([[ 100., 100., 100., 100., 100.],
[ 0., 25., 25., 25., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.]])
&gt;&gt;&gt; u = zeros((5, 5), 'd'); u[0,:] = 100
&gt;&gt;&gt; weave.blitz (expr)
&gt;&gt;&gt; u
array([[ 100. , 100. , 100. , 100. , 100. ],
[ 0. , 25. , 31.25 , 32.8125 , 0. ],
[ 0. , 6.25 , 9.375 , 10.546875 , 0. ],
[ 0. , 1.5625 , 2.734375 , 3.3203125, 0. ],
[ 0. , 0. , 0. , 0. , 0. ]])
</code></pre>
</blockquote>
You can prevent this behavior by using a temporary array.
<blockquote>
<pre><code>
&gt;&gt;&gt; u = zeros((5, 5), 'd'); u[0,:] = 100
&gt;&gt;&gt; temp = zeros((4, 4), 'd');
&gt;&gt;&gt; expr = "temp = (u[0:-2, 1:-1] + u[2:, 1:-1] + "\
... "u[1:-1,0:-2] + u[1:-1, 2:])*0.25;"\
... "u[1:-1,1:-1] = temp"
&gt;&gt;&gt; weave.blitz (expr)
&gt;&gt;&gt; u
array([[ 100., 100., 100., 100., 100.],
[ 0., 25., 25., 25., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.]])
</code></pre>
</blockquote>
</li>
<li>One other point deserves mention lest people be confused. <code>weave.blitz</code>
is not a general purpose Python-&gt;C compiler. It only works for
expressions that contain NumPy arrays and/or Python scalar values. This
focused scope concentrates effort on the compuationally intensive
regions of the program and sidesteps the difficult issues associated
with a general purpose Python-&gt;C compiler. </li>
</ol>
<a name="Numeric Efficiency"></a>
<h2>NumPy efficiency issues: What compilation buys you</h2>
Some might wonder why compiling NumPy expressions to C++ is beneficial
since operations on NumPy array operations are already executed within
C loops. The problem is that anything other than the simplest
expression are executed in less than optimal fashion. Consider the
following NumPy expression:
<blockquote>
<pre><code>
a = 1.2 * b + c * d
</code></pre>
</blockquote>
When NumPy calculates the value for the 2d array, <code>a</code>, it
does the following steps:
<blockquote>
<pre><code>
temp1 = 1.2 * b
temp2 = c * d
a = temp1 + temp2
</code></pre>
</blockquote>
Two things to note. Since <code>c</code> is an (perhaps large) array,
a large temporary array must be created to store the results of <code>1.2
* b</code>. The same is true for <code>temp2</code>. Allocation is
slow. The second thing is that we have 3 loops executing, one to
calculate <code>temp1</code>, one for <code>temp2</code> and one for
adding them up. A C loop for the same problem might look like:
<blockquote>
<pre><code>
for(int i = 0; i &lt; M; i++)
for(int j = 0; j &lt; N; j++)
a[i,j] = 1.2 * b[i,j] + c[i,j] * d[i,j]
</code></pre>
</blockquote>
Here, the 3 loops have been fused into a single loop and there is no
longer
a need for a temporary array. This provides a significant speed
improvement
over the above example (write me and tell me what you get).
<p>So, converting NumPy expressions into C/C++ loops that fuse the
loops and eliminate temporary arrays can provide big gains. The goal
then,is to convert NumPy expression to C/C++ loops, compile them in an
extension module, and then call the compiled extension function. The
good news is that there is an obvious correspondence between the NumPy
expression above and the C loop. The bad news is that NumPy is
generally much more powerful than this simple example illustrates and
handling all possible indexing possibilities results in loops that are
less than straight forward to write. (take a peak in NumPy for
confirmation). Luckily, there are several available tools that simplify
the process.
<a name="blitz_tools"></a></p>
<h2>The Tools</h2>
<code>weave.blitz</code> relies heavily on several remarkable tools. On
the Python side, the main facilitators are Jermey Hylton's parser
module and Travis Oliphant's NumPy module. On the compiled language
side,
Todd Veldhuizen's blitz++ array library, written in C++ (shhhh. don't
tell David Beazley), does the heavy lifting. Don't assume that, because
it's C++, it's much slower than C or Fortran. Blitz++ uses a jaw
dropping array of template techniques (metaprogramming, template
expression, etc) to convert innocent looking and readable C++
expressions into to code that usually executes within a few percentage
points of Fortran code for the same problem. This is good.
Unfortunately all the template raz-ma-taz is very expensive to compile,
so the 200 line extension modules often take 2 or more minutes to
compile. This isn't so good. <code>weave.blitz</code> works to
minimize this issue by remembering where compiled modules live and
reusing them instead of re-compiling every time a program is re-run.
<a name="blitz_parser"></a>
<h3>Parser</h3>
Tearing NumPy expressions apart, examining the pieces, and then
rebuilding them as C++ (blitz) expressions requires a parser of some
sort. I can imagine someone attacking this problem with regular
expressions, but it'd likely be ugly and fragile. Amazingly, Python
solves this problem for us. It actually exposes its parsing engine to
the world through the <code>parser</code> module. The following
fragment creates an Abstract Syntax Tree (AST) object for the
expression and then converts to a (rather unpleasant looking) deeply
nested list representation of the tree.
<blockquote>
<pre><code>
&gt;&gt;&gt; import parser
&gt;&gt;&gt; import scipy.weave.misc
&gt;&gt;&gt; ast = parser.suite("a = b * c + d")
&gt;&gt;&gt; ast_list = ast.tolist()
&gt;&gt;&gt; sym_list = scipy.weave.misc.translate_symbols(ast_list)
&gt;&gt;&gt; pprint.pprint(sym_list)
['file_input',
['stmt',
['simple_stmt',
['small_stmt',
['expr_stmt',
['testlist',
['test',
['and_test',
['not_test',
['comparison',
['expr',
['xor_expr',
['and_expr',
['shift_expr',
['arith_expr',
['term',
['factor', ['power', ['atom', ['NAME', 'a']]]]]]]]]]]]]]],
['EQUAL', '='],
['testlist',
['test',
['and_test',
['not_test',
['comparison',
['expr',
['xor_expr',
['and_expr',
['shift_expr',
['arith_expr',
['term',
['factor', ['power', ['atom', ['NAME', 'b']]]],
['STAR', '*'],
['factor', ['power', ['atom', ['NAME', 'c']]]]],
['PLUS', '+'],
['term',
['factor', ['power', ['atom', ['NAME', 'd']]]]]]]]]]]]]]]]],
['NEWLINE', '']]],
['ENDMARKER', '']]
</code></pre>
</blockquote>
Despite its looks, with some tools developed by Jermey H., its possible
to search these trees for specific patterns (sub-trees), extract the
sub-tree, manipulate them converting python specific code fragments
to blitz code fragments, and then re-insert it in the parse tree. The
parser
module documentation has some details on how to do this. Traversing the
new blitzified tree, writing out the terminal symbols as you go,
creates
our new blitz++ expression string.
<a name="blitz_blitz"></a>
<h3> Blitz and NumPy </h3>
The other nice discovery in the project is that the data structure used
for NumPy arrays and blitz arrays is nearly identical. NumPy stores
"strides" as byte offsets and blitz stores them as element offsets, but
other than that, they are the same. Further, most of the concept and
capabilities of the two libraries are remarkably similar. It is
satisfying that two completely different implementations solved the
problem with similar basic architectures. It is also fortuitous. The
work involved in converting NumPy expressions to blitz expressions was
greatly diminished.
As an example, consider the code for slicing an array in Python with a
stride:
<blockquote>
<pre><code>
&gt;&gt;&gt; a = b[0:4:2] + c
&gt;&gt;&gt; a
[0,2,4]
</code></pre>
</blockquote>
In Blitz it is as follows:
<blockquote>
<pre><code>
Array&lt;2,int&gt; b(10);
Array&lt;2,int&gt; c(3);
// ...
Array&lt;2,int&gt; a = b(Range(0,3,2)) + c;
</code></pre>
</blockquote>
Here the range object works exactly like Python slice objects with the
exception
that the top index (3) is inclusive where as Python's (4) is exclusive.
Other differences include the type declaraions in C++ and parentheses
instead of brackets for indexing arrays. Currently, <code>weave.blitz</code>
handles the inclusive/exclusive issue by subtracting one from upper
indices during the
translation. An alternative that is likely more robust/maintainable in
the long run, is to write a PyRange class that behaves like Python's
range. This is likely very easy.
<p>The stock blitz also doesn't handle negative indices in ranges. The
current implementation of the <code>blitz()</code> has a partial
solution to this problem. It calculates and index that starts with a
'-' sign by subtracting it from the maximum index in the array so that:
</p>
<blockquote>
<pre><code>
upper index limit
/-----\
b[:-1] -&gt; b(Range(0,Nb[0]-1-1))
</code></pre>
</blockquote>
This approach fails, however, when the top index is calculated from
other values. In the following scenario, if <code>i+j</code> evaluates
to a negative value, the compiled code will produce incorrect results
and could even core-
dump. Right now, all calculated indices are assumed to be positive.
<blockquote>
<pre><code>
b[:i-j] -&gt; b(Range(0,i+j))
</code></pre>
</blockquote>
A solution is to calculate all indices up front using if/then to handle
the
+/- cases. This is a little work and results in more code, so it hasn't
been
done. I'm holding out to see if blitz++ can be modified to handle
negative
indexing, but haven't looked into how much effort is involved yet.
While it needs fixin', I don't think there is a ton of code where this
is an issue.
<p>The actual translation of the Python expressions to blitz
expressions is currently a two part process. First, all x:y:z slicing
expression are removed
from the AST, converted to slice(x,y,z) and re-inserted into the tree.
Any
math needed on these expressions (subtracting from the maximum index,
etc.) are also preformed here. _beg and _end are used as special
variables that are defined as blitz::fromBegin and blitz::toEnd. </p>
<blockquote>
<pre><code>
a[i+j:i+j+1,:] = b[2:3,:]
</code></pre>
</blockquote>
becomes a more verbose:
<blockquote>
<pre><code>
a[slice(i+j,i+j+1),slice(_beg,_end)] = b[slice(2,3),slice(_beg,_end)]
</code></pre>
</blockquote>
The second part does a simple string search/replace to convert to a
blitz expression with the following translations:
<blockquote>
<pre><code>
slice(_beg,_end) -&gt; _all # not strictly needed, but cuts down on code.
slice -&gt; blitz::Range
[ -&gt; (
] -&gt; )
_stp -&gt; 1
</code></pre>
</blockquote>
<code>_all</code> is defined in the compiled function as <code>blitz::Range.all()</code>.
These translations could of course happen directly in the syntax tree.
But the string replacement is slightly easier. Note that name spaces
are maintained in the C++ code to lessen the likelyhood of name
clashes. Currently no effort is made to detect name clashes. A good
rule of thumb is don't use values that start with '_' or 'py_' in
compiled expressions and you'll be fine.
<a name="blitz_type_conversions"></a>
<h2>Type definitions and coersion</h2>
So far we've glossed over the dynamic vs. static typing issue between
Python and C++. In Python, the type of value that a variable holds can
change
through the course of program execution. C/C++, on the other hand,
forces you
to declare the type of value a variables will hold prior at compile
time.
<code>weave.blitz</code> handles this issue by examining the types of
the
variables in the expression being executed, and compiling a function
for those
explicit types. For example:
<blockquote>
<pre><code>
a = ones((5,5),Float32)
b = ones((5,5),Float32)
weave.blitz("a = a + b")
</code></pre>
</blockquote>
When compiling this expression to C++, <code>weave.blitz</code> sees
that the
values for a and b in the local scope have type <code>Float32</code>,
or 'float'
on a 32 bit architecture. As a result, it compiles the function using
the float type (no attempt has been made to deal with 64 bit issues).
It also goes one step further. If all arrays have the same type, a
templated
version of the function is made and instantiated for float, double,
complex<float>, and complex<double> arrays. <em> Note: This feature
has been removed from the current version of the code. Each version
will be compiled
separately </em>
</double></float>
<p>What happens if you call a compiled function with array types that
are different than the ones for which it was originally compiled? No
biggie, you'll just have to wait on it to compile a new version for
your new types. This doesn't overwrite the old functions, as they are
still accessible. See the catalog section in the inline() documentation
to see how this is handled. Suffice to say, the mechanism is
transparent to the user and behaves like dynamic typing with the
occasional wait for compiling newly typed functions.
</p>
<p>When working with combined scalar/array operations, the type of the
array is <em>always</em> used. This is similar to the savespace flag
that was recently added to NumPy. This prevents issues with the
following expression perhaps unexpectedly being calculated at a higher
(more expensive) precision that can occur in Python: </p>
<blockquote>
<pre><code>
&gt;&gt;&gt; a = array((1,2,3),typecode = Float32)
&gt;&gt;&gt; b = a * 2.1 # results in b being a Float64 array.
</code></pre>
</blockquote>
In this example,
<blockquote>
<pre><code>
&gt;&gt;&gt; a = ones((5,5),Float32)
&gt;&gt;&gt; b = ones((5,5),Float32)
&gt;&gt;&gt; weave.blitz("b = a * 2.1")
</code></pre>
</blockquote>
the <code>2.1</code> is cast down to a <code>float</code> before
carrying out the operation. If you really want to force the calculation
to be a <code>double</code>, define <code>a</code> and <code>b</code>
as <code>double</code> arrays.
<p>One other point of note. Currently, you must include both the right
hand side and left hand side (assignment side) of your equation in the
compiled expression. Also, the array being assigned to must be created
prior to calling <code>weave.blitz</code>. I'm pretty sure this is
easily changed so that a compiled_eval expression can be defined, but
no effort has been made to allocate new arrays (and decern their type)
on the fly.
<a name="blitz_catalog"></a> </p>
<h2>Cataloging Compiled Functions</h2>
See the <a href="#The%20Catalog">Cataloging functions</a> section in
the <code>weave.inline()</code> documentation.
<a name="blitz_array_sizes"></a>
<h2>Checking Array Sizes</h2>
Surprisingly, one of the big initial problems with compiled code was
making
sure all the arrays in an operation were of compatible type. The
following
case is trivially easy:
<blockquote>
<pre><code>
a = b + c
</code></pre>
</blockquote>
It only requires that arrays <code>a</code>, <code>b</code>, and <code>c</code>
have the same shape. However, expressions like:
<blockquote>
<pre><code>
a[i+j:i+j+1,:] = b[2:3,:] + c
</code></pre>
</blockquote>
are not so trivial. Since slicing is involved, the size of the slices,
not the input arrays must be checked. Broadcasting complicates things
further because arrays and slices with different dimensions and shapes
may be compatible for math operations (broadcasting isn't yet supported
by <code>weave.blitz</code>). Reductions have a similar effect as
their results are different shapes than their input operand. The binary
operators in NumPy compare the shapes of their two operands just before
they operate on them. This is possible because NumPy treats each
operation independently. The intermediate (temporary) arrays created
during sub-operations in an expression are tested for the correct shape
before they are combined by another operation. Because <code>weave.blitz</code>
fuses all operations into a single loop, this isn't possible. The shape
comparisons must be done and guaranteed compatible before evaluating
the expression.
<p>The solution chosen converts input arrays to "dummy arrays" that
only represent the dimensions of the arrays, not the data. Binary
operations on dummy arrays check that input array sizes are comptible
and return a dummy array with the size correct size. Evaluating an
expression of dummy arrays traces the changing array sizes through all
operations and fails if incompatible array sizes are ever found. </p>
<p>The machinery for this is housed in <code>weave.size_check</code>.
It basically involves writing a new class (dummy array) and overloading
it math operators to calculate the new sizes correctly. All the code is
in Python and there is a fair amount of logic (mainly to handle
indexing and slicing) so the operation does impose some overhead. For
large arrays (ie. 50x50x50), the overhead is negligible compared to
evaluating the actual expression. For small arrays (ie. 16x16), the
overhead imposed for checking the shapes with this method can cause the
<code>weave.blitz</code> to be slower than evaluating the expression in
Python. </p>
<p>What can be done to reduce the overhead? (1) The size checking code
could be moved into C. This would likely remove most of the overhead
penalty compared to NumPy (although there is also some calling
overhead), but no effort has been made to do this. (2) You can also
call <code>weave.blitz</code> with
<code>check_size=0</code> and the size checking isn't done. However, if
the sizes aren't compatible, it can cause a core-dump. So, foregoing
size_checking
isn't advisable until your code is well debugged.
<a name="blitz_extension_module"></a> </p>
<h2>Creating the Extension Module</h2>
<code>weave.blitz</code> uses the same machinery as <code>weave.inline</code>
to build the extension module. The only difference
is the code included in the function is automatically generated from
the NumPy array expression instead of supplied by the user.
<a name="#Extension Modules"></a>
<h1>Extension Modules</h1>
<code>weave.inline</code> and <code>weave.blitz</code> are high level
tools
that generate extension modules automatically. Under the covers, they
use several
classes from <code>weave.ext_tools</code> to help generate the
extension module.
The main two classes are <code>ext_module</code> and <code>ext_function</code>
(I'd
like to add <code>ext_class</code> and <code>ext_method</code> also).
These classes
simplify the process of generating extension modules by handling most
of the "boiler
plate" code automatically.
<em>Note: <code>inline</code> actually sub-classes <code>weave.ext_tools.ext_function</code>
to generate slightly different code than the standard <code>ext_function</code>.
The main difference is that the standard class converts function
arguments to
C types, while inline always has two arguments, the local and global
dicts, and
the grabs the variables that need to be convereted to C from these.
</em><a name="A Simple Example"></a>
<h2> A Simple Example </h2>
The following simple example demonstrates how to build an extension
module within
a Python function:
<blockquote>
<pre><code>
# examples/increment_example.py
from weave import ext_tools
def build_increment_ext():
""" Build a simple extension with functions that increment numbers.
The extension will be built in the local directory.
"""
mod = ext_tools.ext_module('increment_ext')
a = 1 # effectively a type declaration for 'a' in the
# following functions.
ext_code = "return_val = Py::new_reference_to(Py::Int(a+1));"
func = ext_tools.ext_function('increment',ext_code,['a'])
mod.add_function(func)
ext_code = "return_val = Py::new_reference_to(Py::Int(a+2));"
func = ext_tools.ext_function('increment_by_2',ext_code,['a'])
mod.add_function(func)
mod.compile()
</code></pre>
</blockquote>
The function <code>build_increment_ext()</code> creates an extension
module named <code>increment_ext</code> and compiles it to a shared
library (.so or .pyd) that can be loaded into Python.. <code>increment_ext</code>
contains two functions, <code>increment</code> and <code>increment_by_2</code>.
The first line of <code>build_increment_ext()</code>,
<blockquote>
<pre><code>
mod = ext_tools.ext_module('increment_ext')
</code></pre>
</blockquote>
creates an <code>ext_module</code> instance that is ready to have <code>ext_function</code>
instances added to it. <code>ext_function</code> instances are created
much with a calling convention similar to <code>weave.inline()</code>.
The most common call includes a C/C++ code snippet and a list of the
arguments for the function. The following
<blockquote>
<pre><code>
ext_code = "return_val = Py::new_reference_to(Py::Int(a+1));"
func = ext_tools.ext_function('increment',ext_code,['a'])
</code></pre>
</blockquote>
creates a C/C++ extension function that is equivalent to the following
Python
function:
<blockquote>
<pre><code>
def increment(a):
return a + 1
</code></pre>
</blockquote>
A second method is also added to the module and then,
<blockquote>
<pre><code>
mod.compile()
</code></pre>
</blockquote>
is called to build the extension module. By default, the module is
created
in the current working directory.
This example is available in the <code>examples/increment_example.py</code>
file
found in the <code>weave</code> directory. At the bottom of the file
in the
module's "main" program, an attempt to import <code>increment_ext</code>
without
building it is made. If this fails (the module doesn't exist in the
PYTHONPATH), the module is built by calling <code>build_increment_ext()</code>.
This approach
only takes the time consuming ( a few seconds for this example) process
of building
the module if it hasn't been built before.
<blockquote>
<pre><code>
if __name__ == "__main__":
try:
import increment_ext
except ImportError:
build_increment_ext()
import increment_ext
a = 1
print 'a, a+1:', a, increment_ext.increment(a)
print 'a, a+2:', a, increment_ext.increment_by_2(a)
</code></pre>
</blockquote>
<em>Note: If we were willing to always pay the penalty of building the
C++ code for a module, we could store the md5 checksum of the C++ code
along with some information about the compiler, platform, etc. Then, <code>ext_module.compile()</code>
could try importing the module before it actually
compiles it, check the md5 checksum and other meta-data in the imported
module
with the meta-data of the code it just produced and only compile the
code if
the module didn't exist or the meta-data didn't match. This would
reduce the
above code to:
</em>
<blockquote>
<pre><code>
if __name__ == "__main__":
build_increment_ext()
a = 1
print 'a, a+1:', a, increment_ext.increment(a)
print 'a, a+2:', a, increment_ext.increment_by_2(a)
</code></pre>
</blockquote>
<em>Note: There would always be the overhead of building the C++ code,
but it would only actually compile the code once. You pay a little in
overhead and get cleaner
"import" code. Needs some thought.
</em>
<p>If you run <code>increment_example.py</code> from the command line,
you get
the following: </p>
<blockquote>
<pre><code>
[eric@n0]$ python increment_example.py
a, a+1: 1 2
a, a+2: 1 3
</code></pre>
</blockquote>
If the module didn't exist before it was run, the module is created. If
it did
exist, it is just imported and used.
<a name="Fibonacci Example"></a>
<h2> Fibonacci Example </h2>
<code>examples/fibonacci.py</code> provides a little more complex
example of how to use <code>ext_tools</code>. Fibonacci numbers are a
series of numbers where each number in the series is the sum of the
previous two: 1, 1, 2, 3, 5, 8, etc. Here, the first two numbers in the
series are taken to be 1. One approach to calculating Fibonacci numbers
uses recursive function calls. In Python, it might be written as:
<blockquote>
<pre><code>
def fib(a):
if a &lt;= 2:
return 1
else:
return fib(a-2) + fib(a-1)
</code></pre>
</blockquote>
In C, the same function would look something like this:
<blockquote>
<pre><code>
int fib(int a)
{
if(a &lt;= 2)
return 1;
else
return fib(a-2) + fib(a-1);
}
</code></pre>
</blockquote>
Recursion is much faster in C than in Python, so it would be beneficial
to use the C version for fibonacci number calculations instead of the
Python version. We need an extension function that calls this C
function
to do this. This is possible by including the above code snippet as
"support code" and then calling it from the extension function. Support
code snippets (usually structure definitions, helper functions and the
like)
are inserted into the extension module C/C++ file before the extension
function code. Here is how to build the C version of the fibonacci
number
generator:
<blockquote>
<pre><code>
def build_fibonacci():
""" Builds an extension module with fibonacci calculators.
"""
mod = ext_tools.ext_module('fibonacci_ext')
a = 1 # this is effectively a type declaration
# recursive fibonacci in C
fib_code = """
int fib1(int a)
{
if(a &lt;= 2)
return 1;
else
return fib1(a-2) + fib1(a-1);
}
"""
ext_code = """
int val = fib1(a);
return_val = Py::new_reference_to(Py::Int(val));
"""
fib = ext_tools.ext_function('fib',ext_code,['a'])
fib.customize.add_support_code(fib_code)
mod.add_function(fib)
mod.compile()
</code></pre>
</blockquote>
XXX More about custom_info, and what xxx_info instances are good for.
<p><em>Note: recursion is not the fastest way to calculate fibonacci
numbers,
but this approach serves nicely for this example.
</em></p>
<p><a name="#Type Factories"></a>
</p>
<h1>Customizing Type Conversions -- Type Factories</h1>
not written
<h1>Things I wish <code>weave</code> did</h1>
It is possible to get name clashes if you uses a variable name that is
already defined
in a header automatically included (such as <code>stdio.h</code>) For
instance, if you
try to pass in a variable named <code>stdout</code>, you'll get a
cryptic error report
due to the fact that <code>stdio.h</code> also defines the name. <code>weave</code>
should probably try and handle this in some way.
Other things...
</body>
</html>
Jump to Line
Something went wrong with that request. Please try again.