IvanHoe 999953 beta
Re: IvanHoe 999953 beta
Comparing 9.53b x64 with H1.03a at longer time controls on two i7 machines I would say that whilst there is very little in it, so far Ivanhoe has the edge in that it finds the move both engines agree on, faster than Houdini does. Clearly the less time you give the engines, the more this becomes an important fact.
Also would like to say that on i7 machines the above two engines are very stables - I have not experience any failures yet.
Also would like to say that on i7 machines the above two engines are very stables - I have not experience any failures yet.
Re: IvanHoe 999953 beta
I get it as about 1-2% faster (AMD X6 Phenom II). Maybe it has to do with cache size, or 32/64-bit (64-bit mult is faster than 32-bit mult I suspect). It's also a compile time flag, so it can be turned off.Significant change is introduction of magic bit-boards; unfortunately, it's a 2% speed hit.
Does Hash need to be a power-of-2 in your build? The CHANGE_LOG says you can use other numbers.Just like Hash, must be power of 2.
Code: Select all
999957 HYPER_HASH to allow 3 bits in hash size, example 56 (0.7% more slow?)
- kingliveson
- Posts: 1388
- Joined: Thu Jun 10, 2010 1:22 am
- Real Name: Franklin Titus
- Location: 28°32'1"N 81°22'33"W
Re: IvanHoe 999953 beta
You're right about Hash not having to be power of 2 -- lazy post on my part. As for speed hit with new magic-bitboard, I initially disabled it, and then re-enabled it to compare. More investigating needed...BB+ wrote:I get it as about 1-2% faster (AMD X6 Phenom II). Maybe it has to do with cache size, or 32/64-bit (64-bit mult is faster than 32-bit mult I suspect). It's also a compile time flag, so it can be turned off.Significant change is introduction of magic bit-boards; unfortunately, it's a 2% speed hit.
Does Hash need to be a power-of-2 in your build? The CHANGE_LOG says you can use other numbers.Just like Hash, must be power of 2.Code: Select all
999957 HYPER_HASH to allow 3 bits in hash size, example 56 (0.7% more slow?)
PAWN : Knight >> Bishop >> Rook >>Queen
- kingliveson
- Posts: 1388
- Joined: Thu Jun 10, 2010 1:22 am
- Real Name: Franklin Titus
- Location: 28°32'1"N 81°22'33"W
Re: IvanHoe 999953 beta
IvanHoe 999953 beta has been re-uploaded yet again. The files has been reloaded about 5 times since initial upload and last upload was about 15 minutes ago. If it is yet re-uploaded again, it will probably be the last one.
PAWN : Knight >> Bishop >> Rook >>Queen
Re: IvanHoe 999953 beta
Does this "Pawn Hash set to 1/8th or 1/4th of Hash used" apply to all IvanHoes or just to those using magic boards?
If latter case, how do we tell when magic boards are used?
Thanks
If latter case, how do we tell when magic boards are used?
Thanks
- kingliveson
- Posts: 1388
- Joined: Thu Jun 10, 2010 1:22 am
- Real Name: Franklin Titus
- Location: 28°32'1"N 81°22'33"W
Re: IvanHoe 999953 beta
Magic bitboard introduction was with 999953 and it's enabled by default in the source. Pawn Hash set to either 1/16th, 1/8th, or 1/4th of Hash applies to all IvanHoe versions.Canto wrote:Does this "Pawn Hash set to 1/8th or 1/4th of Hash used" apply to all IvanHoes or just to those using magic boards?
If latter case, how do we tell when magic boards are used?
Thanks
Furthermore, default Pawn Hash is set to 2 MB in the source code which is 1/16th of default (32 MB) Hash. The developers recommend for 1024 MB hash, set Pawn Hash to 64 MB. 1/16th is a little too conservative -- almost communist. Kranium recommends 1/4th for Fire (FireBird).
PAWN : Knight >> Bishop >> Rook >>Queen
- kingliveson
- Posts: 1388
- Joined: Thu Jun 10, 2010 1:22 am
- Real Name: Franklin Titus
- Location: 28°32'1"N 81°22'33"W
Re: IvanHoe 999953 beta
999953 was re-uploaded a few minute ago and will not be touched again. Everything is now back to default source after playing around. Normally I try not to mess with parts of the code that affect engine play.
Interesting statement posted on Ippolit website that there are secret modification in my builds. This is not quite the case. It is true that the builds do have slight modification, but in most cases, these are code already in the source -- disabled/enabled by default.
Search, Eval, Material Weight... are not tampered with -- many others already do that. The goal has always been to have something as close as possible to the original source, and as intended by the developers. Here are things I do and have done with the builds:
- In the earlier versions, Smooth Scaling (introduced to Stockfish by Dann Corbit) was added as UCI option up to version 999963. Although UCI option for 999963 is there, I actually forgot to enable the code -- later discovered when playing around with the source. It was a copy/paste job from RobboLito -- I believe Sentinel and Kranium added it.
- Auto CPU cores detection is in the source code but disabled by default. So it is enabled and called in main. TitanicCPUsSizings renamed to Threads.
- Pawn Hash is set to 2 MB by default in the source code. Left it at default in some builds while others use 4 MB or 8 MB.
- Send currmove is off by default in the source code, but on by compile.
- Added POPCNT code for 32-bit Windows builds. But of course you should be running a 64-bit OS if your CPU supports POPCNT.
- AllowInstantMoveFromHash default is off by compile as of 999954 if am not mistaken.
- After running 3000 games, 2 additional time management adjustments disabled by default in the source code are enabled in 999953 (this is the file that was just re-uploaded).
- 999953 should also now display cpuload correctly -- copy/pasted code from RobboLito.
- DoHashFull default is off by compile as of 999953. It has no use in game play.
All in all, I think more credit is being given me than is deserved.
Thanks,
Franklin
Interesting statement posted on Ippolit website that there are secret modification in my builds. This is not quite the case. It is true that the builds do have slight modification, but in most cases, these are code already in the source -- disabled/enabled by default.
Search, Eval, Material Weight... are not tampered with -- many others already do that. The goal has always been to have something as close as possible to the original source, and as intended by the developers. Here are things I do and have done with the builds:
- In the earlier versions, Smooth Scaling (introduced to Stockfish by Dann Corbit) was added as UCI option up to version 999963. Although UCI option for 999963 is there, I actually forgot to enable the code -- later discovered when playing around with the source. It was a copy/paste job from RobboLito -- I believe Sentinel and Kranium added it.
- Auto CPU cores detection is in the source code but disabled by default. So it is enabled and called in main. TitanicCPUsSizings renamed to Threads.
- Pawn Hash is set to 2 MB by default in the source code. Left it at default in some builds while others use 4 MB or 8 MB.
- Send currmove is off by default in the source code, but on by compile.
- Added POPCNT code for 32-bit Windows builds. But of course you should be running a 64-bit OS if your CPU supports POPCNT.
- AllowInstantMoveFromHash default is off by compile as of 999954 if am not mistaken.
- After running 3000 games, 2 additional time management adjustments disabled by default in the source code are enabled in 999953 (this is the file that was just re-uploaded).
- 999953 should also now display cpuload correctly -- copy/pasted code from RobboLito.
- DoHashFull default is off by compile as of 999953. It has no use in game play.
All in all, I think more credit is being given me than is deserved.
Thanks,
Franklin
PAWN : Knight >> Bishop >> Rook >>Queen