Page 1 of 1
Xiphos 0.2.2
Posted: Fri Apr 20, 2018 9:52 pm
by mt
I'm sorry for flooding the forum with an another Xiphos update.
I decided to make an official release this early because I found a terrible bug when the engine is tested using the opening books in FEN format (for example, .epd opening books). This affects all the previous releases.
The difference is over 100 elo:
Code: Select all
tc=10+0.1, hash=128MB
Score of Xiphos 0.2.2 SSE vs Xiphos 0.2 SSE: 447 - 86 - 467 [0.680] 1000
Elo difference: 131.34 +/- 15.70
In this version, there is an additional SMP stability patch (thanks TCEC organizers for reporting the issue). This patch also adds ~15 elo when tested on 4 threads.
Otherwise, on 1 thread and when the other opening book formats are used, the version 0.2.2 is exactly the same as 0.2.
Binaries:
https://github.com/milostatarevic/xiphos/releases
Sorry again for the trouble.
Re: Xiphos 0.2.2
Posted: Sat Apr 21, 2018 3:13 am
by AndrewGrant
Curious, which version is playing in TCEC right now?
0.2, 0.2.1, or 0.2.2?
Re: Xiphos 0.2.2
Posted: Sat Apr 21, 2018 3:28 am
by mt
AndrewGrant wrote:Curious, which version is playing in TCEC right now?
0.2, 0.2.1, or 0.2.2?
0.2.1. the organizer was having a trouble with 0.2. I have to submit a quick fix so I didn't update the version number at the time (it is still labeled 0.2 there).
in 0.2.2 i fixed a bug where after each `position fen` hash was cleared. that caused a bad performance at fastgm.de. doesn't seem to be an issue at TCEC (hopefully).
btw, great performance of Ethereal, congrats!
Re: Xiphos 0.2.2
Posted: Sat Apr 21, 2018 3:54 am
by AndrewGrant
Happy to hear that the SMP stuff made it into TCEC.
Did you measure the speedup for the changes you made to node counting?
It looked (at just a glance, so no offsense If I am off here) that you were using the
same counter for all threads. That would be a slowdown.
For Xiphos' sake, I hope your bug IS in TCEC at the moment. You would make a slaughter of D3 with +100 elo from hashing :)
Re: Xiphos 0.2.2
Posted: Sat Apr 21, 2018 7:06 am
by mt
AndrewGrant wrote:
Did you measure the speedup for the changes you made to node counting?
It looked (at just a glance, so no offsense If I am off here) that you were using the
same counter for all threads. That would be a slowdown.
i didn't test that one separately, but that was definitely a bad choice.
For Xiphos' sake, I hope your bug IS in TCEC at the moment. You would make a slaughter of D3 with +100 elo from hashing
Hah, thanks. I'm happy with the progress so far, no crashes on 43 cores, that's a big deal.
Re: Xiphos 0.2.2
Posted: Sat Apr 21, 2018 9:40 am
by Damir
Re: Xiphos 0.2.2
Posted: Thu Apr 26, 2018 7:15 am
by Gary Internet
Milos - I'm watching Xiphos vs Ethereal in the 3rd game of Division 3 of TCEC. Xiphos speed is fluctuating wildly. In Division 4 it used to stay above 90 mN/s for the whole game but now it's up at 90 for one move and down to 40 the next. I have no idea whether that's a problem or if it's some kind of gui reporting bug.
Re: Xiphos 0.2.2
Posted: Thu Apr 26, 2018 7:09 pm
by mt
Gary Internet wrote:Milos - I'm watching Xiphos vs Ethereal in the 3rd game of Division 3 of TCEC. Xiphos speed is fluctuating wildly. In Division 4 it used to stay above 90 mN/s for the whole game but now it's up at 90 for one move and down to 40 the next. I have no idea whether that's a problem or if it's some kind of gui reporting bug.
thanks for the report, it's all good.
it looks like the gui is now calculating nps based on the latest node count reported by the engine and the actual move time. xiphos is not reporting node count periodically, but only when the pv/score updates.
Re: Xiphos 0.2.2
Posted: Thu Apr 26, 2018 7:16 pm
by arunsoorya1309
Yours is the most popular engine in div3 so far, but your time management is pretty bad. You need to work on that to be more competitive. Engine makes bad moves under time pressure
Re: Xiphos 0.2.2
Posted: Tue May 01, 2018 6:25 am
by arunsoorya1309
any plans to update for div 2 in case it qualifies?
