[Bug-gnubg] Random problem with CLI version in 0ply analysis

[Massimiliano Maini]

Hi all,
before week-end starts, I've created everything needed to check
this annoying problem (under Win).

1. Install the latest setup archive (possibly not under Program
Files, otherwise Vista may cause troubles in next steps):

http://gnubg.org/media/windows/gnubg-MAIN-20090904-setup.exe

2. Download the following archive and unzip it into your gnubg
directory. It contains various exes, a sample 7pt match and a
python script (saotest.py).

http://gnubg.org/media/windows/random-problem.zip

3. Invoke the python script (you'll need Python 2.6.2 installed).

In saotest.py you can specify which exe(s) you want to run and
with how many threads (0 for nomt exes).
Each exe/numthread is run to analyze a 7pt match at 0ply and save
the result into a sgf file (which is kept). A md5 checksum of the sgf
file is generated and printed to the screeen 9and in a result text file).

If the exe is a no-gui one, the output of "show match stats" is also
saved to a text file (which is kept) and generates another md5 checksum.


What happens for me:

- using a win thread / glib thread exe with >1 thread generates (sometimes)
a "wrong" md5 for the sgf and for the stats output too. This ahppens randomly
(you may have to insist a bit, open close the dos shell etc), and with -O0 as
well as with -O3.

- -O0 cli exes have a diferent md5 of the sgf file, but same md5 of
the "show match
output" (minor numerical diferences ? tbc)

- gui exes do not seem to have this problem.


MaX.