Scorpio 3.0
Moderators: hgm, Rebel, chrisw
-
- Posts: 2283
- Joined: Sat Jun 02, 2012 2:13 am
-
- Posts: 2283
- Joined: Sat Jun 02, 2012 2:13 am
Re: Scorpio 3.0
OK, thanks Daniel.Does it work with non-cuda GPUs? I have a Radeon GPU. All the files are in place, but it doesn't want to run...
Top
Re: Scorpio 3.0
Post by Daniel Shawul » Mon Aug 19, 2019 11:49 pm
Unfortunately not right now. But it should be possible to support OpenCL via tensorflow backend.
I will consider this for the next version.
Daniel
-
- Posts: 2439
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
Re: Scorpio 3.0
I did all steps and all works (LittleBlitzerGUI accepts the .bat-file). BUT: Overnight I had 6 timelosses in 220 games (that is not acceptable, so I aborted the testrun). It seems, the wb2uci-adpater has a time-overhead. I need to set a large move overhead in Scorpio, otherwise a testrun is not possible. But I dont know how to do this (which Scorpio-option to set and how to set in wb2uci.eng file?). Need help!Daniel Shawul wrote: ↑Sat Aug 17, 2019 6:33 pmHi Torsten,Dokterchen wrote: ↑Sat Aug 17, 2019 12:09 pm Hi Daniel,
thanks for your new release!
I have tested it on my WIN10 4 core machine with a GTX 2070. The install.bat works very well, all paths and parameters in the ini file are fine and everythinbg works as expected. Also the change to INT8 works well:
I did my tests before your answer to Stefan. In your answer to Stefan this step is missing.:Code: Select all
HALF and delay=0 # nodes = 2003616 <0% qnodes> time = 10641ms nps = 188292 eps = 0 nneps = 19990 INT8 and delay=0 # nodes = 2587105 <0% qnodes> time = 9172ms nps = 282065 eps = 0 nneps = 28796
Is this not necessary anymore?Once that is done change the "float_type" to "INT8" in scorpio.ini and run scorpio one more time:
scorpio.bat float_type INT8 go quit
KR
Torsten
Indeed that is necessary, I forgot about it!
I have now put up a document for the installation steps with an example on a Linux GPU machine. If someone could follow similar instruction on a Windows machine from the command line, I would be happy to include the example run in the document.
Btw, you get really good speedup even using HALF precision while is like 2x slower for reasons I don't understand yet.
I suspect it is in the way TensorRT optimizes the network because i sometimes get a 2x speedup with FP16.
reards,
Daniel
Stefan (SPCC)
-
- Posts: 4185
- Joined: Tue Mar 14, 2006 11:34 am
- Location: Ethiopia
Re: Scorpio 3.0
Unfortunately there is no option to set move overhead in scorpio itself, so i guess it can only be done from the GUI or adaptor.
I am not an expert in Wb2UCI but you can try and set a larger initialization time. It takes about 9 sec to load the network on my machine
so put a 15 sec init time under [OPTIONS] in Wb2Uci.eng
I have played thousands of games in cutechess-cli using game in 1min time controls and no move overhead.
I have a few questions: Is Scorpio loaded once and then plays all games ? If the 6 losses you observed are when it being loaded from scratch
then the InitTime could solve the problem.
What time control are you using ?
Also what nps are you getting on your machine?
Thanks
Daniel
I am not an expert in Wb2UCI but you can try and set a larger initialization time. It takes about 9 sec to load the network on my machine
so put a 15 sec init time under [OPTIONS] in Wb2Uci.eng
Code: Select all
[OPTIONS]
InitTime=15
I have a few questions: Is Scorpio loaded once and then plays all games ? If the 6 losses you observed are when it being loaded from scratch
then the InitTime could solve the problem.
What time control are you using ?
Also what nps are you getting on your machine?
Thanks
Daniel
-
- Posts: 2439
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
Re: Scorpio 3.0
I use the LittleblitzerGUI. With 50“+500ms thinking-time.
The 6 timelosses happened randomly overnight. The first 35 games (before I went to bed) were all OK.
And as far as I know, the LBG loads engines fresh into memory with each game. But I am not sure about that.
The nps displayed by LBG were around 26000 n/s. Except the timelosses, Scorpio worked. The score out of the 220 games was around 41% (vs. 6 engines with average Elo of 3363) which means a performance of around 3300 Elo (Fire 7.1 has 3290 Elo in my ratinglist). So, not bad at all.
But without a move overhead parameter or (better) rewritten for UCI-protocol, Scorpio is untestable for me.
Winboard is totally outdated. And does not run in Fritz/Chessbase. The most important GUI for chessplayers. And not in the LittleBlitzerGUI, the most stable GUI for testers.And no other modern top-engine uses winboard anymore. So, please, use UCI !!!
The 6 timelosses happened randomly overnight. The first 35 games (before I went to bed) were all OK.
And as far as I know, the LBG loads engines fresh into memory with each game. But I am not sure about that.
The nps displayed by LBG were around 26000 n/s. Except the timelosses, Scorpio worked. The score out of the 220 games was around 41% (vs. 6 engines with average Elo of 3363) which means a performance of around 3300 Elo (Fire 7.1 has 3290 Elo in my ratinglist). So, not bad at all.
But without a move overhead parameter or (better) rewritten for UCI-protocol, Scorpio is untestable for me.
Winboard is totally outdated. And does not run in Fritz/Chessbase. The most important GUI for chessplayers. And not in the LittleBlitzerGUI, the most stable GUI for testers.And no other modern top-engine uses winboard anymore. So, please, use UCI !!!
-
- Posts: 247
- Joined: Tue Apr 13, 2010 10:41 am
Re: Scorpio 3.0
+1!
Dear Daniel, it would be highly appreciated if Scorpio supports the UCI protocol (in addition to wb). Please give it a serious consideration if your time allows.
In fact, the same hopeful request goes to the great authors of Crafty, Phalanx, ExChess ...
Hope we're not just the biological boot loader for digital super intelligence. Unfortunately, that is increasingly probable - Elon Musk
-
- Posts: 1296
- Joined: Sun Mar 12, 2006 6:46 pm
- Location: Kelowna
- Full name: Tony Mokonen
Re: Scorpio 3.0
None of Daniel's engines over the last 14 years have ever been UCI, so it seems not to be a priority for him. You will never in a million years see an official UCI version of Crafty. Bob has expressed his distaste for the UCI protocol on numerous occasions. Somehow I think Phalanx and ExChess will always be Winboard, too.
-
- Posts: 2439
- Joined: Sat Sep 03, 2011 7:25 am
- Location: Berlin, Germany
- Full name: Stefan Pohl
Re: Scorpio 3.0
That has changed !!!tmokonen wrote: ↑Wed Aug 21, 2019 9:08 pmNone of Daniel's engines over the last 14 years have ever been UCI, so it seems not to be a priority for him. You will never in a million years see an official UCI version of Crafty. Bob has expressed his distaste for the UCI protocol on numerous occasions. Somehow I think Phalanx and ExChess will always be Winboard, too.
“There is now a uci version of Scorpio that you may want to try. Please install again from scratch since there are changes in the ini file.“
https://github.com/dshawul/Scorpio/releases
I will restart the testrun tomorrow. Hopefully, it will work properly in the LittleBlitzerGUI, now. I will report here in 2 days...fingers crossed!
-
- Posts: 1296
- Joined: Sun Mar 12, 2006 6:46 pm
- Location: Kelowna
- Full name: Tony Mokonen
Re: Scorpio 3.0
Well, that's a surprise for sure. Nice of Daniel to add that UCI support.
-
- Posts: 247
- Joined: Tue Apr 13, 2010 10:41 am
Re: Scorpio 3.0
Thanks a lot Daniel, what a nice surprise!
While Scorpio version 3.0 compiles fine on a mac, there are issues compiling the new uci version 3.0.1 from Scorpio-master.
While Scorpio version 3.0 compiles fine on a mac, there are issues compiling the new uci version 3.0.1 from Scorpio-master.
Code: Select all
$ ./build.sh clang
rm -rf attack.o scorpio.o eval.o hash.o moves.o parallel.o probe.o search.o mcts.o see.o magics.o util.o ../bin/scorpio core.* cluster.* *.gcda
/Library/Developer/CommandLineTools/usr/bin/make ../bin/scorpio strip
clang++ -Wall -fstrict-aliasing -fno-exceptions -fno-rtti -Wno-unused-variable -Wno-unused-result -Wno-unused-result -std=c++11 -msse -Ofast -fomit-frame-pointer -DPARALLEL -DARC_64BIT -DHAS_POPCNT -DHAS_PREFETCH -DUSE_SPINLOCK -DTHREAD_POLLING -DLOG_FILE -DBOOK_PROBE -DBOOK_CREATE -DEGBB -c -o attack.o attack.cpp
clang++ -Wall -fstrict-aliasing -fno-exceptions -fno-rtti -Wno-unused-variable -Wno-unused-result -Wno-unused-result -std=c++11 -msse -Ofast -fomit-frame-pointer -DPARALLEL -DARC_64BIT -DHAS_POPCNT -DHAS_PREFETCH -DUSE_SPINLOCK -DTHREAD_POLLING -DLOG_FILE -DBOOK_PROBE -DBOOK_CREATE -DEGBB -c -o scorpio.o scorpio.cpp
scorpio.cpp:249:29: error: implicit instantiation of undefined template
'std::__1::basic_string<char, std::__1::char_traits<char>,
std::__1::allocator<char> >'
std::string s(buffer);
^
/Library/Developer/CommandLineTools/usr/include/c++/v1/iosfwd:193:32: note:
template is declared here
class _LIBCPP_TEMPLATE_VIS basic_string;
^
scorpio.cpp:250:48: error: implicit instantiation of undefined template
'std::__1::basic_string<char, std::__1::char_traits<char>,
std::__1::allocator<char> >'
while (s.find("value") != std::string::npos)
^
/Library/Developer/CommandLineTools/usr/include/c++/v1/iosfwd:193:32: note:
template is declared here
class _LIBCPP_TEMPLATE_VIS basic_string;
^
2 errors generated.
make[1]: *** [scorpio.o] Error 1
make: *** [default] Error 2
Hope we're not just the biological boot loader for digital super intelligence. Unfortunately, that is increasingly probable - Elon Musk