Weave Documentation

By Eric Jones eric@enthought.com

Outline

Introduction

Requirements

Installation

Testing

Benchmarks

Inline

More with printf

More examples

Binary search

Dictionary sort

Numeric -- cast/copy/transpose

wxPython

Keyword options

Returning values

The issue with locals()

A quick look at the code

Technical Details

Converting Types

Numeric Argument Conversion

String, List, Tuple, and Dictionary Conversion

File Conversion

Callable, Instance, and Module Conversion

Customizing Conversions

Compiling Code

"Cataloging" functions

Function Storage

The PYTHONCOMPILED evnironment variable

Blitz

Requirements

Limitations

Numeric Efficiency Issues

The Tools

Parser

Blitz and Numeric

Type defintions and coersion

Cataloging Compiled Functions

Checking Array Sizes

Creating the Extension Module

Extension Modules

A Simple Example

Fibonacci Example

Customizing Type Conversions -- Type Factories (not written)

Type Specifications

Type Information

The Conversion Process

Introduction

The weave 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.

There are three basic ways to use weave. The weave.inline() function executes C code directly within Python, and weave.blitz() translates Python Numeric expressions to C++ for fast execution. blitz() was the original reason weave was built. For those interested in building extension libraries, the ext_tools module provides classes for building extension modules within Python.

Most of weave's functionality should work on Windows and Unix, although some of its functionality requires gcc 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).

The inline and blitz provide new functionality to Python (although I've recently learned about the PyInline project which may offer similar functionality to inline). 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 weave 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. weave is set up so that you can customize how Python types are converted to C types in weave. This is great for inline(), 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 weave does not do. I guess it would be possible to build such a tool on top of weave, 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 weave 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, weave.ext_tools() might be the tool for you. If you're wrapping legacy code, stick with SWIG.

The next several sections give the basics of how to use weave. 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. weave does quite a bit for you, but for anything complex, you'll need to do some conversions, reference counting, etc.

Note: weave is actually part of the SciPy package. However, it works fine as a standalone package. 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 from scipy import weave and the examples will work identically.

Requirements

• Python

  I use 2.1.1. Probably 2.0 or higher should work.

• C++ compiler

  weave uses distutils to actually build extension modules, so it uses whatever compiler was originally used to build Python. weave itself requires a C++ compiler. If you used a C++ compiler to build Python, your probably fine.

  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 inline and ext_tools. The one issue I'm not sure about is that I've hard coded things so that compilations are linked with the stdc++ library. Is this standard across Unix compilers, or is this a gcc-ism?

  For blitz(), 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.

  On Windows, either MSVC or gcc ( mingw32) should work. Again, you'll need gcc for blitz() as the MSVC compiler doesn't handle templates well.

  I have not tried Cygwin, so please report success if it works for you.

• Numeric (optional)

  The python Numeric module from here. is required for blitz() to work. Be sure and get NumPy, not NumArray which is the "next generation" implementation. This is not required for using inline() or ext_tools.

• scipy_distutils and scipy_test (packaged with weave)

  These two modules are packaged with weave in both the windows installer and the source distributions. If you are using CVS, however, you'll need to download these separately (also available through CVS at SciPy).

Installation

There are currently two ways to get weave. Fist, weave is part of SciPy and installed automatically (as a sub- package) whenever SciPy is installed (although the latest version isn't in SciPy yet, so use this one for now). Second, since weave is useful outside of the scientific community, it has been setup so that it can be used as a stand-alone module.

The stand-alone version can be downloaded from here. Unix users should grab the tar ball (.tgz file) and install it using the following commands.

    tar -xzvf weave-0.2.tar.gz
    cd weave-0.2
    python setup.py install
    

For Windows users, it's even easier. You can download the click-install .exe file and run it for automatic installation. There is also a .zip file of the source for those interested. It also includes a setup.py file to simplify installation.

If you're using the CVS version, you'll need to install scipy_distutils and scipy_test packages (also available from CVS) on your own.

Note: The dependency issue here is a little sticky. I hate to make people download more than one file (and so I haven't), but distutils doesn't have a way to do conditional installation -- at least that I know about. This can lead to undesired clobbering of the scipy_test and scipy_distutils modules. What to do, what to do... Right now it is a very minor issue.

Testing

    >>> import weave
    >>> weave.test()
    runs long time... spews tons of output and a few warnings
    .
    .
    .
    ..............................................................
    ................................................................
    ..................................................
    ----------------------------------------------------------------------
    Ran 184 tests in 158.418s

    OK
    
    >>> 
    

If you only want to test a single module of the package, you can do this by running test() for that specific module.

    >>> import weave.scalar_spec
    >>> weave.scalar_spec.test()
    .......
    ----------------------------------------------------------------------
    Ran 7 tests in 23.284s
    

• Windows 1

  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.

• Windows 2

  If you run the tests from PythonWin or some other GUI tool, you'll get a ton of DOS windows popping up periodically as weave spawns the compiler multiple times. Very annoying. Anyone know how to fix this?

• wxPython

  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?

Benchmarks

The benchmarks shown blitz in the best possible light. Numeric (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.

Inline

inline() 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)

Here's a trivial printf example using inline():

    >>> import weave    
    >>> a  = 1
    >>> weave.inline('printf("%d\\n",a);',['a'])
    1
    

In this, its most basic form, inline(c_code, var_list) requires two arguments. c_code is a string of valid C/C++ code. var_list is a list of variable names that are passed from Python into C/C++. Here we have a simple printf statement that writes the Python variable a 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 weave 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.

Note: If you try the printf 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.

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 printf example is actually slower by 30% or so than using Python print 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 Numeric 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.

More with printf

MSVC users will actually see a bit of compiler output that distutils does not supress the first time the code executes:

    
    >>> weave.inline(r'printf("%d\n",a);',['a'])
    sc_e013937dbc8c647ac62438874e5795131.cpp
       Creating library C:\DOCUME~1\eric\LOCALS~1\Temp\python21_compiled\temp
       \Release\sc_e013937dbc8c647ac62438874e5795131.lib and object C:\DOCUME
       ~1\eric\LOCALS~1\Temp\python21_compiled\temp\Release\sc_e013937dbc8c64
       7ac62438874e5795131.exp
    1
    

Nothing bad is happening, its just a bit annoying. Anyone know how to turn this off?

This example also demonstrates using 'raw strings'. The r 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 printf 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.

The printf statement in these examples is formatted to print out integers. What happens if a is a string? inline will happily, compile a new version of the code to accept strings as input, and execute the code. The result?

    
    >>> a = 'string'
    >>> weave.inline(r'printf("%d\n",a);',['a'])
    32956972
    

In this case, the result is non-sensical, but also non-fatal. In other situations, it might produce a compile time error because a 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 inline arguments of the wrong data type by using asserts in Python.

    
    >>> a = 'string'
    >>> def protected_printf(a):    
    ...     assert(type(a) == type(1))
    ...     weave.inline(r'printf("%d\n",a);',['a'])
    >>> protected_printf(1)
     1
    >>> protected_printf('string')
    AssertError...
    

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.

    
    >>> a = 'string'    
    >>> weave.inline(r'printf("%s\n",std::string(a).c_str());',['a'])
    string
    

This is a little convoluted. Perhaps strings should convert to std::string objects instead of CXX objects. Or maybe to char*.

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 std::.

    
    >>> weave.inline('std::cout << a << std::endl;',['a'])
    1    
    >>> a = 'string'
    >>> weave.inline('std::cout << a << std::endl;',['a'])
    string
    

Examples using printf and cout are included in examples/print_example.py.

More examples

Binary search

    def binary_search(seq, t):
        min = 0; max = len(seq) - 1
        while 1:
            if max < min:
                return -1
            m = (min  + max)  / 2
            if seq[m] < t: 
                min = m  + 1 
            elif seq[m] > t: 
                max = m  - 1 
            else:
                return m    
    

    
    def c_int_binary_search(seq,t):
        # do a little type checking in Python
        assert(type(t) == type(1))
        assert(type(seq) == type([]))
        
        # now the C code
        code = """
               #line 29 "binary_search.py"
               int val, m, min = 0;  
               int max = seq.length() - 1;
               PyObject *py_val; 
               for(;;)
               {
                   if (max < min  ) 
                   { 
                       return_val =  Py::new_reference_to(Py::Int(-1)); 
                       break;
                   } 
                   m =  (min + max) /2;
                   val =    py_to_int(PyList_GetItem(seq.ptr(),m),"val"); 
                   if (val  < t) 
                       min = m  + 1;
                   else if (val >  t)
                       max = m - 1;
                   else
                   {
                       return_val = Py::new_reference_to(Py::Int(m));
                       break;
                   }
               }
               """
        return inline(code,['seq','t'])
    

We have two variables seq and t passed in. t is guaranteed (by the assert) to be an integer. Python integers are converted to C int types in the transition from Python to C. seq 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 website. 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, seq.length() returns the length of the list. A little more about CXX and its class methods, etc. is in the ** type conversions ** section.

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.

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 return_val instead of directly returning from the C code with a return statement. return_val is an automatically defined variable of type PyObject* 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 Py::. The following code converts the integer m to a CXX Int() object and then to a PyObject* with an incremented reference count using Py::new_reference_to().

   
    return_val = Py::new_reference_to(Py::Int(m));
    

The second big differences shows up in the retrieval of integer values from the Python list. The simple Python seq[i] call balloons into a C Python API call to grab the value out of the list and then a separate call to py_to_int() that converts the PyObject* to an integer. py_to_int() includes both a NULL cheack and a PyInt_Check() 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 try/catch block that handles exceptions much like Python does. This removes the need for most error checking code.

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 seq.ptr() which is the underlying PyObject* of the List object.

The #line 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).

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 bisect 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:

   
    C:\home\ej\wrk\scipy\weave\examples> python binary_search.py
    Binary search for 3000 items in 1000000 length list of integers:
     speed in python: 0.159999966621
     speed of bisect: 0.121000051498
     speed up: 1.32
     speed in c: 0.110000014305
     speed up: 1.45
     speed in c(no asserts): 0.0900000333786
     speed up: 1.78
    

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 Numeric 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.

Dictionary Sort

The demo in examples/dict_sort.py is another example from the Python CookBook. This submission, by Alex Martelli, demonstrates how to return the values from a dictionary sorted by their keys:

       
    def sortedDictValues3(adict):
        keys = adict.keys()
        keys.sort()
        return map(adict.get, keys)
    

Alex provides 3 algorithms and this is the 3rd and fastest of the set. The C version of this same algorithm follows:

       
    def c_sort(adict):
        assert(type(adict) == type({}))
        code = """     
        #line 21 "dict_sort.py"  
        Py::List keys = adict.keys();
        Py::List items(keys.length()); keys.sort();     
        PyObject* item = NULL; 
        for(int i = 0;  i < keys.length();i++)
        {
            item = PyList_GET_ITEM(keys.ptr(),i);
            item = PyDict_GetItem(adict.ptr(),item);
            Py_XINCREF(item);
            PyList_SetItem(items.ptr(),i,item);              
        }           
        return_val = Py::new_reference_to(items);
        """   
        return inline_tools.inline(code,['adict'],verbose=1)
    

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.

       
    C:\home\ej\wrk\scipy\weave\examples> python dict_sort.py
    Dict sort of 1000 items for 300 iterations:
     speed in python: 0.319999933243
    [0, 1, 2, 3, 4]
     speed in c: 0.151000022888
     speed up: 2.12
    [0, 1, 2, 3, 4]
    

Numeric -- cast/copy/transpose

       
    def _castCopyAndTranspose(type, array):
        if a.typecode() == type:
            cast_array = copy.copy(Numeric.transpose(a))
        else:
            cast_array = copy.copy(Numeric.transpose(a).astype(type))
        return cast_array
    

    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)
        numeric_type = scalar_spec.numeric_to_blitz_type_mapping[type]
        code = \
        """  
        for(int i = 0;i < _Na_2d[0]; i++)  
            for(int j = 0;  j < _Na_2d[1]; j++)
                new_array(i,j) = (%s) a_2d(j,i);
        """ % numeric_type
        inline(code,['new_array','a_2d'],
               type_factories = blitz_type_factories,compiler='gcc')
        return new_array
    

     #C:\home\ej\wrk\scipy\weave\examples> 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
    

Keyword Options

The basic definition of the inline() function has a slew of optional variables. It also takes keyword arguments that are passed to distutils as compiler options. The following is a formatted cut/paste of the argument section of inline's doc-string. It explains all of the variables. Some examples using various options will follow.

       
    def inline(code,arg_names,local_dict = None, global_dict = None, 
               force = 0, 
               compiler='',
               verbose = 0, 
               support_code = None,
               customize=None, 
               type_factories = None, 
               auto_downcast=1,
               **kw):
    

inline Arguments:

code

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.

arg_names

list of strings. A list of Python variable names that should be transferred from Python into the C/C++ code.

local_dict

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.

global_dict

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.

force

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.

compiler

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).

On windows, the compiler defaults to the Microsoft C++ compiler. If this isn't available, it looks for mingw32 (the gcc compiler).

On Unix, it'll probably use the same compiler that was used when compiling Python. Cygwin's behavior should be similar.

verbose

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.

support_code

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.

customize

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).

type_factories

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.

auto_downcast

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 Numeric arrays are of type float. If even one of the arrays has type double or double complex, all variables maintain there standard types.

Distutils keywords:

inline() also accepts a number of distutils keywords for controlling how the code is compiled. The following descriptions have been copied from Greg Ward's distutils.extension.Extension class doc- strings for convenience:

sources

[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

include_dirs

[string] list of directories to search for C/C++ header files (in Unix form for portability)

define_macros

[(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)

undef_macros

[string] list of macros to undefine explicitly

library_dirs

[string] list of directories to search for C/C++ libraries at link time

libraries

[string] list of library names (not filenames or paths) to link against

runtime_library_dirs

[string] list of directories to search for C/C++ libraries at run time (for shared extensions, this is when the extension is loaded)

extra_objects

[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.)

extra_compile_args

[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.

extra_link_args

[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'.

export_symbols

[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.

Keyword Option Examples

    >>> from weave import inline
    >>> a = 'string'
    >>> code = """
    ...        int l = a.length();
    ...        return_val = Py::new_reference_to(Py::Int(l));
    ...        """
    >>> 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
    >>> inline(code,['a'])
    6
    

    >>> from weave import inline
    >>> a = 'a longer string' 
    >>> code = """ 
    ...        int l = a.length();
    ...        return_val = Py::new_reference_to(Py::Int(l));  
    ...        """
    >>> inline(code,['a'])
    15
    

Notice this time, inline() 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.

You can specify the local and global dictionaries if you'd like (much like exec or eval() in Python), but if they aren't specified, the "expected" ones are used -- i.e. the ones from the function that called inline() . 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:

    >>> a = 'a longer string'
    >>> b = 'an even  longer string' 
    >>> my_dict = {'a':b}
    >>> inline(code,['a'])
    15
    >>> inline(code,['a'],my_dict)
    21
    

Everytime, the code is changed, inline does a recompile. However, changing any of the other options in inline does not force a recompile. The force 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 by a single character (like adding a space some place) to force the recompile. Note: It also might be nice to add some methods for purging the cache and on disk catalogs.

I use verbose 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 verbose is set to 0.

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 force=1 flag is needed to make inline() ignore the previously compiled version and recompile using gcc. The verbose flag is added to show what is printed out:

    >>>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
    

    >>>inline(code,['a'],compiler='gcc',verbose=1,force=1)
    Compiling code...
    finished compiling (sec):  6.00800001621
    15
    

Note: I've only used the compiler option for switching between 'msvc' and 'gcc' on windows. It may have use on Unix also, but I don't know yet.

The support_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 string. Note that changes to support_code do not force a recompile. The catalog only relies on 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 in some way or use the force argument to get the code to recompile. I usually just add some inocuous whitespace to the end of one of the lines in code somewhere. Here's an example of defining a separate method for calculating the string length:

    >>> from weave import inline
    >>> a = 'a longer string'
    >>> support_code = """
    ...                PyObject* length(Py::String a)
    ...                {
    ...                    int l = a.length();  
    ...                    return Py::new_reference_to(Py::Int(l)); 
    ...                }
    ...                """        
    >>> inline("return_val = length(a);",['a'],
    ...        support_code = support_code)
    15
    

customize is a left over from a previous way of specifying compiler options. It is a custom_info object that can specify quite a bit of information about how a file is compiled. These info 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 support_code option, I think customize may be obsolete. We'll see if anyone cares to use it. If not, it'll get axed in the next version.

The type_factories 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.

auto_downcast handles one of the big type conversion issues that is common when using Numeric 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. auto_downcast 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 auto_downcast to 0.

Returning Values

    # THIS DOES NOT WORK
    >>> a = 1
    >>> weave.inline("a++;",['a'])
    >>> a
    2
    

    >>> a = 1
    >>> weave.inline("a++;",['a'])
    >>> a
    1
    

    >>> a= [1,2]
    >>> weave.inline("PyList_SetItem(a.ptr(),0,PyInt_FromLong(3));",['a'])
    >>> print a
    [3, 2]
    

    >>> a = 1
    >>> a = weave.inline("return_val = Py::new_reference_to(Py::Int(a+1));",['a'])  
    >>> a
    2
    

       
    # python version
    def multi_return():
        return 1, '2nd'
    
    # C version.
    def c_multi_return():    
        code =  """
     	        Py::Tuple results(2);
     	        results[0] = Py::Int(1);
     	        results[1] = Py::String("2nd");
     	        return_val = Py::new_reference_to(results); 	        
                """
        return inline_tools.inline(code)
    

The example is available in examples/tuple_return.py. It also has the dubious honor of demonstrating how much inline() can slow things down. The C version here is about 10 times slower than the Python version. Of course, something so trivial has no reason to be written in C anyway.

The issue with locals()

inline passes the locals() and globals() 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 locals() and globals() dictionaries so that the modified values were reflected in Python. Unfortunately, as pointed out by the Python manual, the locals() dictionary is not writable.

I suspect locals() 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 inline 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.

A quick look at the code

    >>> import weave    
    >>> a = 1
    >>> weave.inline('printf("%d\\n",a);',['a'])
    1
    

    static PyObject* compiled_func(PyObject*self, PyObject* args)
    {
        // The Py_None needs an incref before returning
        PyObject *return_val = NULL;
        int exception_occured = 0;
        PyObject *py__locals = NULL;
        PyObject *py__globals = NULL;
        PyObject *py_a;
        py_a = NULL;
        
        if(!PyArg_ParseTuple(args,"OO:compiled_func",&py__locals,&py__globals))
            return NULL;
        try                              
        {                                
            PyObject* raw_locals = py_to_raw_dict(py__locals,"_locals");
            PyObject* raw_globals = py_to_raw_dict(py__globals,"_globals");
            int a = py_to_int (get_variable("a",raw_locals,raw_globals),"a");
            /* Here is the inline code */            
            printf("%d\n",a);
            /* I would like to fill in changed locals and globals here... */
        }                                       
        catch( Py::Exception& e)           
        {                                
            return_val =  Py::Null();    
            exception_occured = 1;       
        }                                 
        if(!return_val && !exception_occured)
        {
                                      
            Py_INCREF(Py_None);              
            return_val = Py_None;            
        }
        /* clean up code */
        
        /* return */                              
        return return_val;           
    }                                
    

    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);
    }
    

    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) && PyInstance_Check(py_obj)) return "callable";
        if(PyInstance_Check(py_obj)) return "instance"; 
        if(PyCallable_Check(py_obj)) return "callable";
        return "unkown type";
    }
    

Technical Details

There are several main steps to using C/C++ code withing Python:

1. Type conversion
2. Generating C/C++ code
3. Compile the code to an extension module
4. Catalog (and cache) the function for future use

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 inline. Generating the C/C++ code is handled by ext_function and ext_module 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 ext_tools module, and distutils is covered by a completely separate document xxx.

Passing Variables in/out of the C/C++ code

Type Conversions

By default, inline() makes the following type conversions between Python and C++ types.

The Py:: namespace is defined by the CXX library which has C++ class equivalents for many Python types. std:: is the namespace of the standard library in C++.

Note:

• I haven't figured out how to handle long int yet (I think they are currenlty converted to int - - check this).
• Hopefully VTK will be added to the list soon

Python to C++ conversions fill in code in several locations in the generated inline extension function. Below is the basic template for the function. This is actually the exact code that is generated by calling weave.inline("").

    static PyObject* compiled_func(PyObject*self, PyObject* args)
    {
        PyObject *return_val = NULL;
        int exception_occured = 0;
        PyObject *py__locals = NULL;
        PyObject *py__globals = NULL;
        PyObject *py_a;
        py_a = NULL;
    
        if(!PyArg_ParseTuple(args,"OO:compiled_func",&py__locals,&py__globals))
            return NULL;
        try
        {
            PyObject* raw_locals = py_to_raw_dict(py__locals,"_locals");
            PyObject* raw_globals = py_to_raw_dict(py__globals,"_globals");
            /* argument conversion code */
            /* inline code */
            /*I would like to fill in changed locals and globals here...*/
    
        }
        catch( Py::Exception& e)
        {
            return_val =  Py::Null();
            exception_occured = 1;
        }
        /* cleanup code */
        if(!return_val && !exception_occured)
        {
    
            Py_INCREF(Py_None);
            return_val = Py_None;
        }
    
        return return_val;
    }
    

Numeric Argument Conversion

    >>> a = 1
    >>> inline("",['a'])
    

    /* argument conversion code */
    int a = py_to_int (get_variable("a",raw_locals,raw_globals),"a");
    

    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);
    }
    

    
    static double py_to_float(PyObject* py_obj,char* name)
    {
        if (!py_obj || !PyFloat_Check(py_obj))
            handle_bad_type(py_obj,"float", name);
        return PyFloat_AsDouble(py_obj);
    }
    
    static std::complex py_to_complex(PyObject* py_obj,char* name)
    {
        if (!py_obj || !PyComplex_Check(py_obj))
            handle_bad_type(py_obj,"complex", name);
        return std::complex(PyComplex_RealAsDouble(py_obj),
                                    PyComplex_ImagAsDouble(py_obj));    
    }
    

String, List, Tuple, and Dictionary Conversion

    >>> a = [1]
    >>> inline("",['a'])
    

    /* argument conversion code */
    Py::List a = py_to_list (get_variable("a",raw_locals,raw_globals),"a");
    

    
    static Py::List py_to_list(PyObject* py_obj,char* name)
    {
        if (!py_obj || !PyList_Check(py_obj))
            handle_bad_type(py_obj,"list", name);
        return Py::List(py_obj);
    }
    
    static Py::String py_to_string(PyObject* py_obj,char* name)
    {
        if (!PyString_Check(py_obj))
            handle_bad_type(py_obj,"string", name);
        return Py::String(py_obj);
    }

    static Py::Dict py_to_dict(PyObject* py_obj,char* name)
    {
        if (!py_obj || !PyDict_Check(py_obj))
            handle_bad_type(py_obj,"dict", name);
        return Py::Dict(py_obj);
    }
    
    static Py::Tuple py_to_tuple(PyObject* py_obj,char* name)
    {
        if (!py_obj || !PyTuple_Check(py_obj))
            handle_bad_type(py_obj,"tuple", name);
        return Py::Tuple(py_obj);
    }
    

File Conversion

    >>> a = open("bob",'w')  
    >>> inline("",['a'])
    

    /* argument conversion code */
    PyObject* py_a = get_variable("a",raw_locals,raw_globals);
    FILE* a = py_to_file(py_a,"a");
    

    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);
    }
    

    /* cleanup code */
    Py_XDECREF(py_a);
    

    >>> weave.inline('printf("hello\\n");')
    

    >>> buf = sys.stdout
    >>> weave.inline('fprintf(buf,"hello\\n");',['buf'])
    

    >>> buf = sys.stdout
    >>> weave.inline('fprintf(buf,"hello\\n");',['buf'])
    Traceback (most recent call last):
        File "", line 1, in ?
        File "C:\Python21\weave\inline_tools.py", line 315, in inline
            auto_downcast = auto_downcast,
        File "C:\Python21\weave\inline_tools.py", line 386, in compile_function
            type_factories = type_factories)
        File "C:\Python21\weave\ext_tools.py", line 197, in __init__
            auto_downcast, type_factories)
        File "C:\Python21\weave\ext_tools.py", line 390, in assign_variable_types
            raise TypeError, format_error_msg(errors)
        TypeError: {'buf': "Unable to convert variable 'buf' to a C++ type."}
    

    >>> buf
    pywin.framework.interact.InteractiveView instance at 00EAD014
    

Callable, Instance, and Module Conversion

    >>> def a(): 
        pass
    >>> inline("",['a'])
    

    /* argument conversion code */
    PyObject* a = py_to_callable(get_variable("a",raw_locals,raw_globals),"a");
    

    
    PyObject* py_to_callable(PyObject* py_obj, char* name)
    {
        if (!py_obj || !PyCallable_Check(py_obj))
            handle_bad_type(py_obj,"callable", name);    
        return py_obj;
    }

    PyObject* py_to_instance(PyObject* py_obj, char* name)
    {
        if (!py_obj || !PyFile_Check(py_obj))
            handle_bad_type(py_obj,"instance", name);    
        return py_obj;
    }
    

Customizing Conversions

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.

When

    >>> a = 1
    >>> weave.inline('printf("%d",a);',['a'])
    

Examples of xxx_specification are scattered throughout numerous "xxx_spec.py" files in the weave package. Closely related to the xxx_specification classes are yyy_info 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. xxx_specification classes have one or more yyy_info 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 inline using the type_factories 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 std::string or maybe char* 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: ...

The Catalog

catalog.py has a class called catalog that helps keep track of previously compiled functions. This prevents inline() 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.

When inline(cpp_code) is called the following things happen:

1. A fast local cache of functions is checked for the last function called for cpp_code. If an entry for cpp_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.
2. The catalog class also keeps an in-memory cache with a list of all the functions compiled for cpp_code. If cpp_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.

   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.

3. When a lookup for cpp_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 cpp_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.

   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.

Function Storage: How functions are stored in caches and on disk

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.

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.

Catalog search paths and the PYTHONCOMPILED variable

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.

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.

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.

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.

An example of using the PYTHONCOMPILED path on bash follows:

    PYTHONCOMPILED=MODULE:/some/path;export PYTHONCOMPILED;
    

    /usr/lib/python21/site-packages/linuxpython_compiled
    /some/path/linuxpython_compiled
    ~/.python21_compiled/linuxpython_compiled
    

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.

The in-module cache (in weave.inline_tools 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.

Blitz

weave.blitz() compiles Numeric Python expressions for fast execution. For most applications, compiled expressions should provide a factor of 2-10 speed-up over Numeric 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:

    from scipy import *  # or from Numeric 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.
    

    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)
    

The following table compares the run times for standard Numeric code and compiled code for the 5 point averaging.

These numbers are for a 512x512 double precision image run on a 400 MHz Celeron processor under RedHat Linux 6.2.

Because of the slow compile times, its probably most effective to develop algorithms as you usually do using the capabilities of scipy or the Numeric module. Once the algorithm is perfected, put quotes around it and execute it using weave.blitz. This provides the standard rapid prototyping strengths of Python and results in algorithms that run close to that of hand coded C or Fortran.

Requirements

Limitations

1. Currently, weave.blitz 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 Numeric module.

   weave.blitz does not currently support operations that use array broadcasting, nor have any of the special purpose functions in Numeric 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 slice() objects directly instead of start:stop:step is also not supported.

2. Currently Python only works on expressions that include assignment such as

   
       >>> result = b + c + d
       

   This means that the result array must exist before calling weave.blitz. Future versions will allow the following:

   
       >>> result = weave.blitz_eval("b + c + d")
       

3. weave.blitz 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.
4. weave.blitz can produce different results than Numeric in certain situations. It can happen when the array receiving the results of a calculation is also used during the calculation. The Numeric 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:

   
           # 4 point average.
           >>> 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"
           >>> u = zeros((5, 5), 'd'); u[0,:] = 100
           >>> exec (expr)
           >>> 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.]])
           
           >>> u = zeros((5, 5), 'd'); u[0,:] = 100
           >>> weave.blitz (expr)
           >>> 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. ]])    
           

   You can prevent this behavior by using a temporary array.

   
           >>> u = zeros((5, 5), 'd'); u[0,:] = 100
           >>> temp = zeros((4, 4), 'd');
           >>> 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"
           >>> weave.blitz (expr)
           >>> 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.]])
           

5. One other point deserves mention lest people be confused. weave.blitz is not a general purpose Python->C compiler. It only works for expressions that contain Numeric 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->C compiler.

Numeric efficiency issues: What compilation buys you

    a = 1.2 * b + c * d
    

    temp1 = 1.2 * b
    temp2 = c * d
    a = temp1 + temp2
    

    for(int i = 0; i < M; i++)
        for(int j = 0; j < N; j++)
            a[i,j] = 1.2 * b[i,j] + c[i,j] * d[i,j]
    

So, converting Numeric expressions into C/C++ loops that fuse the loops and eliminate temporary arrays can provide big gains. The goal then,is to convert Numeric 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 Numeric expression above and the C loop. The bad news is that Numeric 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 Numeric for confirmation). Luckily, there are several available tools that simplify the process.

The Tools

Parser

    >>> import parser
    >>> import scipy.weave.misc
    >>> ast = parser.suite("a = b * c + d")
    >>> ast_list = ast.tolist()
    >>> sym_list = scipy.weave.misc.translate_symbols(ast_list)
    >>> 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', '']]
    

Blitz and Numeric

    >>> a = b[0:4:2] + c
    >>> a
    [0,2,4]
    

    Array<2,int> b(10);
    Array<2,int> c(3);
    // ...
    Array<2,int> a = b(Range(0,3,2)) + c;
    

The stock blitz also doesn't handle negative indices in ranges. The current implementation of the blitz() 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:

                    upper index limit
                        /-----\
    b[:-1] -> b(Range(0,Nb[0]-1-1))
    

    b[:i-j] -> b(Range(0,i+j))
    

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.

    a[i+j:i+j+1,:] = b[2:3,:] 
    

    a[slice(i+j,i+j+1),slice(_beg,_end)] = b[slice(2,3),slice(_beg,_end)]
    

    slice(_beg,_end) -> _all  # not strictly needed, but cuts down on code.
    slice            -> blitz::Range
    [                -> (
    ]                -> )
    _stp             -> 1
    

Type definitions and coersion

    a = ones((5,5),Float32)
    b = ones((5,5),Float32)
    weave.blitz("a = a + b")