Recursion Performance Question

[B]

Hey I found some (VERY) old C++ code of mine that recursively built a
tree of the desktop window handles (on windows) using: (they are stored
in an STL vector)

void FWL(HWND hwnd, int nFlag) // Recursive Function
{
hwnd = GetWindow(hwnd, nFlag);
if(hwnd == NULL)
return;
AddWnd(hwnd);
nLevel++;
FWL(hwnd, GW_CHILD);
nLevel--;
FWL(hwnd, GW_HWNDNEXT);
return;
}

int FillWindowList( bool bReset) // Build Window List
{
WLI wli;

if(bReset)
ResetWindowList ();

nLevel = 0;
FWL(ui.hWnd, GW_HWNDFIRST);

return nCount;
}

Now the interface on this program is really ugly (i hate UI coding), so
I was thinking about re-writing it in python to use PyQt for an easy
GUI. So far I have (they are stored in a 'tree' class):

# pass in window handle and parent node
def gwl(node, hwnd):
if hwnd:
yield node, hwnd
for nd, wnd in Wnd.gwl(node.ch ildren[-1], GetWindow(hwnd,
GW_CHILD)):
yield nd, wnd
for nd, wnd in Wnd.gwl(node, GetWindow(hwnd, GW_HWNDNEXT)):
yield nd, wnd


def generateTree(se lf):
t = time.clock()
if self is not None:
self.children = []

for nd, wnd in Wnd.gwl(self, GetWindow(self. hwnd, GW_CHILD)):
nd.addChild(wnd )


Now it works, but it runs quite slow (compared to the c++ app). I
changed gwl from strait recursion to use a generator and that helped,
but it still takes 0.5-1.0 seconds to populate the tree. What I'm
wondering is am I doing it in a really inefficient way, or is it just
python?

The second problem is reseting the list. In C++ I would use the STL
Vector's clear() method. In python, I can't think of a good way to free
all the nodes, so there is a large memory leak.

I can post more of the code if it's unclear, I just didn't want to write
a really long post.

[Tim Golden]

Re: Recursion Performance Question

B wrote:

Now it works, but it runs quite slow (compared to the c++ app). I
changed gwl from strait recursion to use a generator and that helped,
but it still takes 0.5-1.0 seconds to populate the tree. What I'm
wondering is am I doing it in a really inefficient way, or is it just
python?

Well I did just enough to get your code to work and ran
it through timeit on my respectable but hardly bleeding-edge
desktop PC.

<dump>
C:\temp>python -m timeit "import get_windows; get_windows.run ()"
100 loops, best of 3: 17.1 msec per loop
</dump>

So it's not that slow. Full code is posted below; uncomment
the "print_tree " bit to see the results to confirm that it's
doing what you think. I did this really quickly so it's
possibly I've misunderstood what your code's up to.

I'm not saying there aren't other ways to do this, but
your code (at least inside my guessed-at wrapper) seems
to do an adequate job in a reasonable time.

<get_windows.py >
import time
from win32gui import GetWindow, GetWindowText, GetDesktopWindo w
from win32con import GW_CHILD, GW_HWNDNEXT

class Node (object):
def __init__ (self, hwnd):
self.hwnd = hwnd
self.children = []
def addChild (self, hwnd):
self.children.a ppend (Node (hwnd))

def gwl(node, hwnd):
if hwnd:
yield node, hwnd
for nd, wnd in gwl(node.childr en[-1], GetWindow(hwnd, GW_CHILD)):
yield nd, wnd
for nd, wnd in gwl(node, GetWindow(hwnd, GW_HWNDNEXT)):
yield nd, wnd


def generateTree(se lf):
t = time.clock()
if self is not None:
self.children = []

for nd, wnd in gwl(self, GetWindow(self. hwnd, GW_CHILD)):
nd.addChild(wnd )

def print_tree (root, level=0):
print level * " ", GetWindowText (root.hwnd) or hex (root.hwnd)
for child in root.children:
print_tree (child, level + 1)

def run ():
desktop = Node (GetDesktopWind ow ())
generateTree (desktop)
#~ print_tree (desktop)

</get_windows.py>

The second problem is reseting the list. In C++ I would use the STL
Vector's clear() method. In python, I can't think of a good way to free
all the nodes, so there is a large memory leak.

I'm not clear here what you're getting at. Memory handling in Python
is usually best left to Python unless you've got a very specific
case -- and they do crop up occasionally. Just to be clear on something:
Python does its own internal memory management, alloc-ing and free-ing
with a variety of policies broadly managed by an internal pool. You're
unlikely to see much memory released back to the system while Python's
running. But this isn't a memory leak as such. You don't need to (and,
in effect, can't) release the memory explicitly.

But I'd be surprised if you had enough windows open for this to be
even noticeable.

TJG

[Anders J. Munch]

Re: Recursion Performance Question

B wrote:

>
# pass in window handle and parent node
def gwl(node, hwnd):
if hwnd:
yield node, hwnd
for nd, wnd in Wnd.gwl(node.ch ildren[-1], GetWindow(hwnd,
GW_CHILD)):
yield nd, wnd
for nd, wnd in Wnd.gwl(node, GetWindow(hwnd, GW_HWNDNEXT)):
yield nd, wnd

[...]

>
Now it works, but it runs quite slow (compared to the c++ app). I
changed gwl from strait recursion to use a generator and that helped,
but it still takes 0.5-1.0 seconds to populate the tree.

Actually the generator could well be the problem here, because of time
spent on yield propagation: gwl has worst-case quadratic
performance, the worst case being if the tree is unbalanced and deep,
because every yielded value must pass through a chain of propagation
for-loops.

Straight recursion should be faster; I don't know what you did to make
it slow, but try something along these lines:

def gwl_helper(node , hwnd, collect):
if hwnd:
collect((node,h wnd))
gwl_helper(node .children[-1], GetWindow(hwnd, GW_CHILD), collect)
gwl_helper(node , GetWindow(hwnd, GW_HWNDNEXT), collect)

def gwl(node, hwnd):
result = []
gwl_helper(node , hwnd, result.append)
return result

- Anders