Halogen 8 time management problem

[Angle]

I downloaded Halogen compiles from the Chess Engines Diary:
https://chessengines.blogspot.com/2020/ ... -nnue.html
and now have no time losses using avx2 version. The problem of allocating too little time for the last few moves at ТС 40/2 still remains but there are no time losses in this case either.

I am now running a big tournament with various versions of Beef, Cheese, Counter, Francesca, Fruit Reloaded, Koivisto, Mr Bob, Orion, Rodent, Seer, Tucano, and Weiss. I will add Halogen to this torrney and report you the result by the end of this week.

Ouch, do you realize that this site just recycles otherwise free downloads and opensource projects just for making click money,
by re-uploading them to file hosters?

Actually (as expected) the binary is an identical copy of the avx2 compilation as from Halogens github release site! ( I did a file comparison)
What compilation did you use before? May be you used the nopopcount one?
https://github.com/KierenP/Halogen/releases

Code: Select all

Halogen8-x32.exe
848 KB
Halogen8-x64-AVX2.exe
961 KB
Halogen8-x64-nopopcnt.exe
966 KB
Halogen8-x64.exe
963 KB

I was surprised myself, but the fact is that avx2 compile from the Halogen github page losses on time systematically on my Laptop at 20s+3s TC while the similar compile from CED does not. I believe that the binary should be the same (though didn't make any identity tests) but the fact remaiins the same: it doesn't lose on time on my compiter. I have no idea how to explain it.

[Kieren Pearson]

I downloaded Halogen compiles from the Chess Engines Diary:
https://chessengines.blogspot.com/2020/ ... -nnue.html
and now have no time losses using avx2 version. The problem of allocating too little time for the last few moves at ТС 40/2 still remains but there are no time losses in this case either.

I am now running a big tournament with various versions of Beef, Cheese, Counter, Francesca, Fruit Reloaded, Koivisto, Mr Bob, Orion, Rodent, Seer, Tucano, and Weiss. I will add Halogen to this torrney and report you the result by the end of this week.

Well good news is the dev version more evenly allocates its time and is 27 elo stronger in self play with repeating time controls so the issue of allocating too little time for the last few moves is now fixed

[Kieren Pearson]

I downloaded Halogen compiles from the Chess Engines Diary:
https://chessengines.blogspot.com/2020/ ... -nnue.html
and now have no time losses using avx2 version. The problem of allocating too little time for the last few moves at ТС 40/2 still remains but there are no time losses in this case either.

I am now running a big tournament with various versions of Beef, Cheese, Counter, Francesca, Fruit Reloaded, Koivisto, Mr Bob, Orion, Rodent, Seer, Tucano, and Weiss. I will add Halogen to this torrney and report you the result by the end of this week.

Ouch, do you realize that this site just recycles otherwise free downloads and opensource projects just for making click money,
by re-uploading them to file hosters?

Actually (as expected) the binary is an identical copy of the avx2 compilation as from Halogens github release site! ( I did a file comparison)
What compilation did you use before? May be you used the nopopcount one?
https://github.com/KierenP/Halogen/releases

Code: Select all

Halogen8-x32.exe
848 KB
Halogen8-x64-AVX2.exe
961 KB
Halogen8-x64-nopopcnt.exe
966 KB
Halogen8-x64.exe
963 KB

Not sure where to post this but that site has Halogen as a NNUE engine which may or may not be appropriate. Halogen DOES use a efficiently updated neural network, but no NNUE code from SF and doesn't use SF nets or network trainer so I think it might be misleading.

[Guenther]

I downloaded Halogen compiles from the Chess Engines Diary:
https://chessengines.blogspot.com/2020/ ... -nnue.html
and now have no time losses using avx2 version. The problem of allocating too little time for the last few moves at ТС 40/2 still remains but there are no time losses in this case either.

I am now running a big tournament with various versions of Beef, Cheese, Counter, Francesca, Fruit Reloaded, Koivisto, Mr Bob, Orion, Rodent, Seer, Tucano, and Weiss. I will add Halogen to this torrney and report you the result by the end of this week.

Ouch, do you realize that this site just recycles otherwise free downloads and opensource projects just for making click money,
by re-uploading them to file hosters?

Actually (as expected) the binary is an identical copy of the avx2 compilation as from Halogens github release site! ( I did a file comparison)
What compilation did you use before? May be you used the nopopcount one?
https://github.com/KierenP/Halogen/releases

I was surprised myself, but the fact is that avx2 compile from the Halogen github page losses on time systematically on my Laptop at 20s+3s TC while the similar compile from CED does not. I believe that the binary should be the same (though didn't make any identity tests) but the fact remaiins the same: it doesn't lose on time on my compiter. I have no idea how to explain it.

May be something was running in the background or similar problems during the first test? Did it run from the same path originally BTW?
You could do a little retest or replacing the binaries again to be sure, if you like.
I also could provide you an avx2 compilation of current dev too, if you want, before Kieren releases a new version :)

[Angle]

I downloaded Halogen compiles from the Chess Engines Diary:
https://chessengines.blogspot.com/2020/ ... -nnue.html
and now have no time losses using avx2 version. The problem of allocating too little time for the last few moves at ТС 40/2 still remains but there are no time losses in this case either.

I am now running a big tournament with various versions of Beef, Cheese, Counter, Francesca, Fruit Reloaded, Koivisto, Mr Bob, Orion, Rodent, Seer, Tucano, and Weiss. I will add Halogen to this torrney and report you the result by the end of this week.

Ouch, do you realize that this site just recycles otherwise free downloads and opensource projects just for making click money,
by re-uploading them to file hosters?

Actually (as expected) the binary is an identical copy of the avx2 compilation as from Halogens github release site! ( I did a file comparison)
What compilation did you use before? May be you used the nopopcount one?
https://github.com/KierenP/Halogen/releases

I was surprised myself, but the fact is that avx2 compile from the Halogen github page losses on time systematically on my Laptop at 20s+3s TC while the similar compile from CED does not. I believe that the binary should be the same (though didn't make any identity tests) but the fact remaiins the same: it doesn't lose on time on my compiter. I have no idea how to explain it.

May be something was running in the background or similar problems during the first test? Did it run from the same path originally BTW?
You could do a little retest or replacing the binaries again to be sure, if you like.
I also could provide you an avx2 compilation of current dev too, if you want, before Kieren releases a new version

I tried both compiles several times (last time with the identical path). Every time the result was as I reported. It would be nice if you share the dev version with me because I have problems with compiling it myself. In this case I would put both the dev version and the original Halogen 8 into the same tournament to compare the results.

[Guenther]

I downloaded Halogen compiles from the Chess Engines Diary:
https://chessengines.blogspot.com/2020/ ... -nnue.html
and now have no time losses using avx2 version. The problem of allocating too little time for the last few moves at ТС 40/2 still remains but there are no time losses in this case either.

I am now running a big tournament with various versions of Beef, Cheese, Counter, Francesca, Fruit Reloaded, Koivisto, Mr Bob, Orion, Rodent, Seer, Tucano, and Weiss. I will add Halogen to this torrney and report you the result by the end of this week.

Ouch, do you realize that this site just recycles otherwise free downloads and opensource projects just for making click money,
by re-uploading them to file hosters?

Actually (as expected) the binary is an identical copy of the avx2 compilation as from Halogens github release site! ( I did a file comparison)
What compilation did you use before? May be you used the nopopcount one?
https://github.com/KierenP/Halogen/releases

I was surprised myself, but the fact is that avx2 compile from the Halogen github page losses on time systematically on my Laptop at 20s+3s TC while the similar compile from CED does not. I believe that the binary should be the same (though didn't make any identity tests) but the fact remaiins the same: it doesn't lose on time on my compiter. I have no idea how to explain it.

May be something was running in the background or similar problems during the first test? Did it run from the same path originally BTW?
You could do a little retest or replacing the binaries again to be sure, if you like.
I also could provide you an avx2 compilation of current dev too, if you want, before Kieren releases a new version :)

I tried both compiles several times (last time with the identical path). Every time the result was as I reported. It would be nice if you share the dev version with me because I have problems with compiling it myself. In this case I would put both the dev version and the original Halogen 8 into the same tournament to compare the results.

Well, I still cannot imagine a plausible scenario (except one) how this could happen.
The one exception would be that the Halogen 8 avx2 binary had been replaced silently shortly after its release
and you had the older one and later you had downloaded the one from the 'other site', which now was
the 'newer' one.
You could easily check this by a file comparison. I doubt though this had happened.

Here are the hash signatures for Halogen 8 avx2 from the github release folder and the 'recycled' from the 'other site' via rapidgator

Code: Select all

   Datei: Halogen8-x64-AVX2_github-release.exe
  CRC-32: fb3e2cda
   SHA-1: 8849488efb0a9277a5bafc78c82553e97af8b17c
 SHA-256: 8f522f3e60fdb9da0f7a65a0e2559f28ddbd17c09016caa968f6772cfa8f3e13
 SHA-512: f1bb6fda90e6a737e920539e11e749df3fad791e3062f667ccce74cfcfd63f6feadd6f5525611ebf42dfc2828e0d7cf3c8581d7e7341d1f4933f025c9ab33c01

---------

   Datei: Halogen8-x64-AVX2_CED-rapidgator.exe
  CRC-32: fb3e2cda
   SHA-1: 8849488efb0a9277a5bafc78c82553e97af8b17c
 SHA-256: 8f522f3e60fdb9da0f7a65a0e2559f28ddbd17c09016caa968f6772cfa8f3e13
 SHA-512: f1bb6fda90e6a737e920539e11e749df3fad791e3062f667ccce74cfcfd63f6feadd6f5525611ebf42dfc2828e0d7cf3c8581d7e7341d1f4933f025c9ab33c01

Anyhow here is a link for two avx2 compilations from current Halogen dev (only changed the internal version name for avoiding confusion):
One of both should at least run on your hardware, may be both, take the faster one then. They are untested because my hardware
cannot run them, but I made a non popcount one for me too again and it is running fine.
https://rwbc-chess.de/Downloads/Engines ... ev-avx2.7z

I have one wish though:
Please exchange it immediately with an official compilation, if Kieren releases one

[Guenther]

...

Anyhow here is a link for two avx2 compilations from current Halogen dev (only changed the internal version name for avoiding confusion):
One of both should at least run on your hardware, may be both, take the faster one then. They are untested because my hardware
cannot run them, but I made a non popcount one for me too again and it is running fine.
https://rwbc-chess.de/Downloads/Engines ... ev-avx2.7z

I have one wish though:
Please exchange it immediately with an official compilation, if Kieren releases one

Ha, just a few hours later Kieren commited an update including a new nn file.
https://github.com/KierenP/Halogen/pull/138

updated the avx2 dev compilations and readme.txt too
https://rwbc-chess.de/Downloads/Engines ... v2-avx2.7z

[Angle]

...

Anyhow here is a link for two avx2 compilations from current Halogen dev (only changed the internal version name for avoiding confusion):
One of both should at least run on your hardware, may be both, take the faster one then. They are untested because my hardware
cannot run them, but I made a non popcount one for me too again and it is running fine.
https://rwbc-chess.de/Downloads/Engines ... ev-avx2.7z

I have one wish though:
Please exchange it immediately with an official compilation, if Kieren releases one

Ha, just a few hours later Kieren commited an update including a new nn file.
https://github.com/KierenP/Halogen/pull/138

updated the avx2 dev compilations and readme.txt too
https://rwbc-chess.de/Downloads/Engines ... v2-avx2.7z

Thank you very much for this upload! Both compiles run smoothly without time consuming problems for both incremental TC and tournament TC. Popcnt-avx2 version is a bit faster, so I will use it in my tests.