Erlkönig: Meeting Python

First Pythonic Impressions

I started using Python on 2007-11-23, following the tutorial on the Python website, generating the usual collection of small scripts to test various bits of the language. It quickly became clear that Python has an interesting and pretty clean style of expression, along with notably different performance tradeoffs than some other languages I've used.

Whitespace as Control, and Quirky Control Over Whitespace

Once past generating the obligatory Hello World and various code snippets to test parts of the tutorial, I wrote a small version of Hilo to test flow constructs, input from stdin, and exception handling with a project where I'd be fitting the language to the problem rather than the other way around. Some of my initial reactions were:

• It's funny that the character encoding directive is based on Emacs file variables.
• I expected to hate using whitespace for flow control, but it does reduce visual clutter a bit and removes the control-vs-indentation synchronization issue seen in many other languages.
• On the other hand, Bash can't save paragraphic Python one-liners in the history file without splitting them, even with with shopt -s lithist cmdhist, due to using newline as a inter-command history delimiter.
• Exception handling in Python has a clean feel.
• The inability to fully control the generation of additional spaces in print output is truly annoying. It has the comma to suppress newlines - why in the world isn't there an obvious to override its idea of whether you should have extra spaces injected? Should I really have to resort to sys.stdout.write for something so minor?
• The % operator is cute, in an anti-C++-STL kind of way (a contrast to the STL's dropping of the form() method from streams).
• I understand why Python doesn't want to assume the character encoding for output to files (or possibly to terminals), and the use of c.encode('utf-8') to choose an encoding on output is straightforward, albeit cumbersome. The unicodedata.name function's great, too - I'd love to see this kind of character identification in the Emacs modeline.
• BUT: While working with a piece of Unicode text (the classic Japanese 「いろは」 poem), I encountered some unpleasantness. The fact that one can output Unicode from a Python program to a terminal, then be completely unable to redirect that output to a file without the program crashing (much less cat the results from said file back your tty), or spontaneously encoding your output in some other character set, seems to violate a fundamental expection of datastream transitivity for the Unix realm. It seems like a perfect case for an overriding environment variable that I'm simply not seeing anywhere. Being able to open files via the codecs.open call looks like it might help, but not being able to change the default except on a system-wide level grates.

From 7, to 100 Frames per Second

Next, I started writing yet another Life cellular automaton, since I wanted to see how Python felt with respect to a familiar project that I'd written in a C/C++ environment where function calls are relatively cheap, and which would force me to reconceive the design once I was faced with a working, but slow, program. I also wanted to try out classes this time, deriving TTY and curses-focussed subvariants from base classes Cell and Grid, find the terminal size with TIOCGWINSZ even outside of curses, use select to check for input, and do profiling. The profiling was done with 140x50 cell grids, allowed to settle to a collection of trivial spinners and fixed configurations. Some of the surprises at this stage included:

• Having sys.argv and the in keyword makes certain types of argument parsing really convenient.
• Code like this is much more eloquent than the C equivalent, and the leading doc line gives a warm, LISPy fuzzy feeling:
  import select def StdinPending(): 'Returns whether there is input pending on stdin.' irdy, ordy, erdy = select.select([sys.stdin],[],[],0) return sys.stdin in irdy
• On the other hand, the crippled lambdas, harsh distinction between statements and expressions, and the somewhat high expense of function calls, are clear proof that this isn't LISP.
• Having to explicitly call del on instances to get their __del__ destructors to fire before leaving the function in which they were instantiated, even with no other (obvious) references to them, was counterintuitive, and seems to have other issues.
• The Hotshot profiler's quite nice, although it takes a surprisingly long time to convert its output into a textual summary. It also seems to interfere with instance GC and triggering of __del__ destructors - enough to make me curious if I've misinterpreted their intended use.
• Having to pass a buffer into an ioctl of an appropriate length for the result, yet have the result returned along a different channel, is just strange (the docs allege a fourth parameter to change this, but it didn't seem to be work). This works, though:
  import struct, fcntl, termios def TiocgWinsz(): 'Return the width (in columns) and height (rows) of the terminal (stdout).' buf = struct.pack('HHHH', 0, 0, 0, 0) result = fcntl.ioctl(sys.stdout.fileno(), termios.TIOCGWINSZ, buf) rows, cols, pxwidth, pxheight = (struct.unpack('HHHH', result)) return cols, rows, pxwidth, pxheight
• Although a bit of a surprise, there are two competing types of classes in Python (currently), and the newer style derived from the object built-in class works quite nicely. Especially nice was using super to call parent classes' __init__ constructors without having to name the parent explicitly.

  life.pydebugrc

  # Use: PYTHONSTARTUP=life.pydebugrc python from life import * # a debugging convenience def gc(x,y): # grid cell return g._grid[x][y] def s(): # show status print 65 * '-' # side-by-side grid views (rows for normal, debug, neighbors) for r3 in zip(g.rows(), g.drows(), g.nrows()): print '|'.join(r3) print 65 * '-' def up(): g.Update() s() # interactive utilities acting on required grid `g' def al(): return filter(lambda c: c._alive, g._cells) def em(): return filter(lambda c: not c._alive, g._cells) def bi(): return filter(lambda c: 3 == c._neighbors , em()) def di(): return filter(lambda c: not 2 <= c._neighbors <= 3, al()) g = GridTty(21,7) # setup - small, 3 side-by-side views, and a glider [ g._grid[x][y].Born() for (x,y) in [(8,4),(9,4),(10,4),(9,2),(10,3)] ] s()

  Yielding an initial display of:

  $ PYTHONSTARTUP=life.pydebugrc python Python 2.5.1 (r251:54863, Oct 5 2007, 13:36:32) [GCC 4.1.3 20070929 (prerelease) (Ubuntu 4.1.2-16ubuntu2)] on linux2 Type "help", "copyright", "credits" or "license" for more information. ----------------------------------------------------------------- | |000000000000000000000 | |000000001110000000000 * | x |000000001121000000000 * | . * |000000013532000000000 *** | x** |000000011322000000000 | . |000000012321000000000 | |000000000000000000000 ----------------------------------------------------------------- >>>
• Running Python with a $PYTHONSTARTUP script specific to testing the life.py module, including scripting an initial testing scenario and adding some interactive-specific utility functions greatly eased the edit/test/debug cycle.
• Initial runs of the Python life automaton were pretty low on the performance scale, only yielding about 7 fps at 140x50 at 90% CPU load. In contrast, a more complex C version does 250 fps at the same grid size with 0.3 % CPU load. I'll call that about 10700 times more efficient for the moment. Tweaking various parts of the Python code, particularly those generating lots of function calls, has so far brought it up to about 10 20 36 47 53 100 fps, ~14 X faster, and cut the 10700 factor down to about 750, a huge improvement. Still, I'm not really happy with the unchanged 90% CPU load.
• Poking around for ways to write-protect class and instance attributes gives the impression that one can arrange, at least with the newer class approach, to cause an exception to be raised whenever code tries to violate the protection.
• Python function call performance is slow enough, especially with the number of function calls the life.py code had in loops, to notably interfere with attempts to structure code in certain ways for readability or data modelling purposes. Hopefully I'll run into or discover some alternative approaches to this issue later.
• Certain permutations encouraged by Python comprehensions are PERL-like in their inversive nature. Similar to situations in PERL where the flow control constructs can end up following the controlled parts, there are Python chunks like:
  [ stdscr.addstr(y+1, x+1, grid[x][y].str()) \ for y in self.yvals() \ for x in self.xvals() \ if grid[x][y]._alive ]
  That chunk is actually a bit naïve, considering it could generate a lot of function calls for cells whose screen representations hadn't changed. Keeping a list of cells which were recently subject to Born() and Dies() would be faster...
  ...and it is, with an increase from 47 to 53 fps, with GridCurses' __init__ making a dict of cell-to-location pairs (indexed by repr() of the objects, perhaps a bit flippant), the Update method now recording changed cells in _changed, and this replacement code in the display method:
  for c in self._changed: (x,y) = self._locations[repr(c)] stdscr.addstr(y+1, x+1, c.str())
• Maintaining sets is more efficient than repeatedly using filters to extract list of members of interest. Replacing filters along the line of:
  def Update(self): alive = filter(lambda c: c._alive , self._cells) empty = filter(lambda c: not c._alive , self._cells) [...]
  with code transferring cells between the Grid sets ._alive and ._empty (different from the Cell's boolean ._alive seen above), instead of extracting data from ._cells, brought the speed from 53 fps up to ~90 fps.
• The code would clearly be faster if the class separation between Cell and Grid was removed, but that seems like it wouldn't be true to the idea of the code being an emulation-in-the-small of a larger program.
• Running help(Cell) and pydoc life showed some areas where the embedded doc in my life.py could be improved - especially when the -*- coding: utf-8 -*- directive and all of the top-of-file comments were exposed at the alleged “DESCRIPTION” section of the man-page like output of pydoc
• Unfortunately, pydoc explodes if any of the doc strings contain Unicode characters outside of the ASCII range. Considering that there are cases where embedded doc strings, or automated tests embedded in them, might require non-ASCII to even make sense, this has to be considered a pydoc bug if it is, itself, the cause of the problem. Especially considering that the life.py file has a coding directive explicity demanding Unicode UTF-8, of which pydoc code have taken note.

Using SWIG to Integrate Python and a C varags Function
- or Oh, why did I have to pick a variadic function first?
- or Oh, should I have tried ctypes first?

Of course, the real vexer here was starting my SWIG experimentation with a variadic function like printf, except worse, since this one allows for format continuations in later parameters. This ends up interleaving format strings and data, making the determination of the number of args basically impossible without running the full parsing pass outright on the whole chain of format string(s). That didn't stop me from rewriting the variadic function in question to have an alternate form taking its following args in a vector, macrozing around the va_arg calls to allow the same code chunk to live in both functions via an include. Once SWIG started grumbling about the void** second parameter, though, I decided to take a second pass at this will be with something other than a worst-case-scenario,

The latter, of course, went better, reaching some use of vectors fairly quickly, allowing interaction as shown below, based on a gratuitously-recursive C function along the lines of int Sum(int count, Int *vec) with Int just being a struct wrapped around an int (swigc is just the name of the test module, not an official name). Trying to %extend int rather than Int resulted - unsurpringly - in errors, but with Int, the following works just fine:

The %typemap approach is looking like the way to go next, since it should allow the Vec class to be jettisoned and the simpler passing of native Python lists into function calls. However, I'm also reading (and hearing) some rumors that a completely different mechanism, ctypes, might be preferable to SWIG in situations without and existing base of SWIG code.

AJAX with Python

While experimenting with an initial AJAX webpage, I realized that by default I was about to promote my sh-based CGI backend to PERL, and that going to Python instead might be more interesting. It only took a couple of minutes to find a convenient Python library (via import cgi) for CGI data parsing as well as a instant debugger option of:

Between these two, I found myself with Python instead of PERL in under five minutes, and without the need to drag in the PERL CGI parsing functions I've been using for so long now.