ocean wrote:

...

I've had a report from a user that Plex runs about half as fast in 2.5 as it did in 2.4. In particular, the NFA-to-DFA conversion phase, which does a lot of messing about with dicts representing mappings between sets of states.

That was me.

...

Does anyone in the Ministry for Making Python Blazingly fast happen to know of some change that might have pessimised things in this area?

Hello, I investigated. On my environment, consumed time is

E:\Plex-1.1.5>py24 plex_test2.py 0.710999965668

E:\Plex-1.1.5>py25 plex_test2.py 0.921999931335

And after I applied this patch to Plex/Machines, (make `Node' new style class)

62c62 < class Node: ---

...

class Node(object):

E:\Plex-1.1.5>py24 plex_test2.py 0.401000022888

E:\Plex-1.1.5>py25 plex_test2.py 0.350999832153

Nice!. Meanwhile I tried to replace the parsing I did with Plex by re.Scanner. And again there is a remarkable speed difference. Again python2.5 is slower: try: from re import Scanner except: from sre import Scanner pars = {} order = [] count = 0 def par(scanner,name): global count, order, pars if name in ['caller','e','pi']: return name if name not in pars.keys(): pars[name] = ('ns', count) order.append(name) ret = 'a[%d]'%count count += 1 else: ret = 'a[%d]'%(order.index(name)) return ret scanner = Scanner([ (r"x", lambda y,x: x), (r"[a-zA-Z]+\.", lambda y,x: x), (r"[a-z]+\(", lambda y,x: x), (r"[a-zA-Z_]\w*", par), (r"\d+\.\d*", lambda y,x: x), (r"\d+", lambda y,x: x), (r"\+|-|\*|/", lambda y,x: x), (r"\s+", None), (r"\)+", lambda y,x: x), (r"\(+", lambda y,x: x), (r",", lambda y,x: x), ]) import profile import pstats def run(): arg = '+amp*exp(-(x-pos)/fwhm)' for i in range(100): scanner.scan(arg) profile.run('run()','profscanner') p = pstats.Stats('profscanner') p.strip_dirs() p.sort_stats('cumulative') p.print_stats() Christian

On 6/13/07, ocean <ocean@m2.ccsnet.ne.jp> wrote:

...

Meanwhile I tried to replace the parsing I did with Plex by re.Scanner. And again there is a remarkable speed difference. Again python2.5 is slower:

try: from re import Scanner except: from sre import Scanner

pars = {} order = [] count = 0

def par(scanner,name): global count, order, pars

if name in ['caller','e','pi']: return name if name not in pars.keys(): pars[name] = ('ns', count) order.append(name) ret = 'a[%d]'%count count += 1 else: ret = 'a[%d]'%(order.index(name)) return ret

scanner = Scanner([ (r"x", lambda y,x: x), (r"[a-zA-Z]+\.", lambda y,x: x), (r"[a-z]+\(", lambda y,x: x), (r"[a-zA-Z_]\w*", par), (r"\d+\.\d*", lambda y,x: x), (r"\d+", lambda y,x: x), (r"\+|-|\*|/", lambda y,x: x), (r"\s+", None), (r"\)+", lambda y,x: x), (r"\(+", lambda y,x: x), (r",", lambda y,x: x), ])

import profile import pstats

def run(): arg = '+amp*exp(-(x-pos)/fwhm)' for i in range(100): scanner.scan(arg)

profile.run('run()','profscanner') p = pstats.Stats('profscanner') p.strip_dirs() p.sort_stats('cumulative') p.print_stats()

Well, I tried this script, there was no big difference. Python2.4 0.772sec Python2.5 0.816sec

Probably I found one reason comparation for classic style class is slower on Python2.5. Comparation function instance_compare() calls PyErr_GivenExceptionMatches(), and it was just flag operation on 2.4. But on 2.5, probably related to introduction of BaseException, it checks inherited type tuple. (ie: PyExceptionInstance_Check)

I'm curious about the speed of 2.6 (trunk). I think this should have become faster due to the introduction of fast subtype checks (he says without looking at the code). n