An Optimization Anecdote about PyInline

Mitchell N Charity mcharity at vendian.org
Sat Mar 9 14:44:02 EST 2002


Guido's "Python Patterns - An Optimization Anecdote"
(http://www.python.org/doc/essays/list2str.html) explores the
performance of various ways of converting a list of integers
into a string of characters.  Such as the straightforward

   def f1(list):
       string = ""
       for item in list:
           string = string + chr(item)
       return string

and the fastest version

   import array
   def f7(list):
       return array.array('B', list).tostring()

   # Use 'B' with Python2.2, 'b' with Python1.5

Guido writes
  [...] I had wanted to try one more approach: write the whole
  function in C.  [...]  Given the effort of writing and testing an
  extension (compared to whipping up those Python one-liners), as well
  as the dependency on a non-standard Python extension, I decided not
  to pursue this option...

I wanted to revisit this, because with PyInline, the effort has
diminished, and a non-standard extension is not required (just PyInline).
And, as Guido points out,

  If you feel the need for speed, [...] - you can't beat a loop
  written in C.

So, here then is a new version, f8(), which uses PyInline.

from PyInline import C
import __main__
C.Builder(code="""
PyObject* f8(PyObject *list)
{
    PyObject* string = NULL;
    unsigned char* buffer = NULL;
    int size, index;

#define RAISE(errtype,msg) { PyErr_SetString(errtype,msg); RE_RAISE; }
#define RE_RAISE           { Py_XDECREF(string); return NULL; }

    if(!PyList_Check(list))
       RAISE(PyExc_TypeError,"a list is required");

    size = PyList_Size(list);
    if(size < 0) RE_RAISE;

    string = PyString_FromStringAndSize(NULL, size);
    if(string == NULL) RE_RAISE;

    buffer = PyString_AsString(string);
    if(buffer == NULL) RE_RAISE;

    for(index = 0; index < size; index++) {
        PyObject *item = NULL;
        long number;

        item = PyList_GetItem(list,index);
        if(item == NULL) RE_RAISE;

        number = PyInt_AsLong(item);
        if(PyErr_Occurred() != NULL) RE_RAISE;

        if(number < 0 || number > 255)
            RAISE(PyExc_TypeError,"an integer was out of range");

        buffer[index] = (unsigned char) number;
    }

    return string;
}
""",targetmodule=__main__).build()

The test jig requires a slight change
   #for f in testfuncs: print f.func_name, f(testdata)
   for f in testfuncs: print f(testdata)
because PyInline currently generates glue which does not understand
f.func_name.

So, what is the result?

First, the downsides.  f8() is a C extension, and so only works with
C-based Pythons.  PyInline invokes an external C compiler, and
cruftily caches the result in the filesystem.  And, as PyInline is
alpha code, there are little misfeatures like not being able to use
func_name.  Oh yes, and being written in C, f8() is perhaps two orders
of magnitude more complex than the Python versions.  Though, this last
is perhaps a worst case, as f8() is almost all interface code, with
maybe just three lines of algorithm.  In more typical use, one might
see a one order of magnitude cost.  And with objects which provide a
nice C api, maybe even less.

The upside is f8() runs 5 times faster than f7(), the best alternative.
And uses half the memory.  That's 50 times faster than the naive versions.
And again, when python interface code is less dominant, the speedup
can be dramatically larger.

So, restoring the phrase I clipped out of the quote above, "If you
feel the need for speed, go for built-in functions - you can't beat a
loop written in C.".  And if there isn't one already, you might
consider creating one with PyInline.

Mitchell Charity

(I'm working on a web page http://www.vendian.org/mncharity/dir3/inline/
 which explores using PyInline, and Perl's Inline, with objects which
 provide C api's.)




More information about the Python-list mailing list