Performance profiling Python code

[Andreas Røsdal]

Hi,

I'm using the Python profiler to optimize a pathfinding algorithm of a
game, and would like some help from someone who knows how to improve the
performance of Python code. The algorithm is A-Star, and it works
correctly. However, the interface that I made between the A-Star
pathfinding algorithm and the map data structure is what I've found to be
a bottle-neck so far.

As a start, I'd like to optimize the methods with the highest cumulative
time spent using CPU, which are the ones on the top of the profiling
out-put attached below (get_tile_adjac ent etc.)

So are you able to spot any solutions to improving the performance of the
methods which I've indentified as bottlenecks here?

Thanks in advance!
Andreas R.

The source code of the main faile is here:

http://svn.gna.org/viewcvs/openrts/trunk/openrts/common/map.py?rev=89&view=markup

Profiling output is here:

ncalls tottime percall cumtime percall filename:lineno (function)
/home/andreas/openrts/openrts/openrts/common/map.py:190(get_ tile_adjacent)
162002 0.680 0.000 0.680 0.000
/home/andreas/openrts/openrts/openrts/common/map.py:56(get_t ile)
129600 0.570 0.000 0.570 0.000
/home/andreas/openrts/openrts/openrts/common/map.py:217(wrap _map_pos)
129837 0.480 0.000 0.480 0.000 :0(append)
8100 0.190 0.000 0.430 0.000
/home/andreas/openrts/openrts/openrts/common/map.py:253(get_ tile_height_pos )
8100 0.080 0.000 0.270 0.000
/usr/lib/python2.4/random.py:213(r andint)
8100 0.160 0.000 0.190 0.000
/usr/lib/python2.4/random.py:149(r andrange)
1 0.000 0.000 0.170 0.170
/home/andreas/openrts/openrts/openrts/common/map.py:26(__ini t__)
1 0.070 0.070 0.170 0.170
/home/andreas/openrts/openrts/openrts/common/map.py:32(init_ map)
8100 0.080 0.000 0.110 0.000
/home/andreas/openrts/openrts/openrts/common/map.py:234(get_ tile_height)

[Jack Diederich]

Re: Performance profiling Python code

On Fri, Mar 24, 2006 at 09:33:55PM +0100, Andreas R?sdal wrote:[color=blue]
> Hi,
>
> I'm using the Python profiler to optimize a pathfinding algorithm of a
> game, and would like some help from someone who knows how to improve the
> performance of Python code. The algorithm is A-Star, and it works
> correctly. However, the interface that I made between the A-Star
> pathfinding algorithm and the map data structure is what I've found to be
> a bottle-neck so far.
>
> So are you able to spot any solutions to improving the performance of the
> methods which I've indentified as bottlenecks here?
>[/color]

You have methods for accessing members, just access the members directly.
Differently worded: don't write try to write Java/C++ in python!
This will save you lots of function call overhead.

For the adjacent tiles you could compute the list once and store it in
the tiles (assuming the tiles don't move around).

As a last resort you can write the sensitive bits in C and add a python
wrapper. For the ICFP contest[1] each year I write the few primitives
in C and all the logic in python and it works quite well.

-Jack

[1] http://icfpcontest.org/

[Felipe Almeida Lessa]

Re: Performance profiling Python code

Em Sex, 2006-03-24 às 15:58 -0500, Jack Diederich escreveu:[color=blue]
> As a last resort you can write the sensitive bits in C and add a python
> wrapper. For the ICFP contest[1] each year I write the few primitives
> in C and all the logic in python and it works quite well.[/color]

Before that, try Psyco. It helps a lot sometimes.

HTH,

--
Felipe.