Looking for 32-bit dll to probe tablebases

[Roger Zuehlsdorf]

Hello,

I want to add tablebase support for my 32-bit Visual Basic chess engine ChessbrainVB.

Are there any 32 bit Dlls to access tablebase files? May be for some older bitbases like Scorpio?

My current solution is using Fathom.exe output for root moves - so very slow and not usable for search.

Thank you, Roger

[hgm]

For Scorpio bitbases that must exist. I remember interfacing WinBoard to it (and I always compiled WinBoard for 32 bit).

[Daniel Shawul]

Yes, that is correct.
New exciting feature: You can now load and probe neural networks with it as well, both lc0 & scorpio format.
It uses Tensorflow and TensorRT backends with automatic multi-threaded batching and gets the same nps
as lc0. It is also very easy to add new network formats as the network structure is saved with the weights unlike lc0's.
All it cares about are the input planes and outputs (value+policy) match, you can use any operations within the body.

[chrisw]

Yes, that is correct.
New exciting feature: You can now load and probe neural networks with it as well, both lc0 & scorpio format.
It uses Tensorflow and TensorRT backends with automatic multi-threaded batching and gets the same nps
as lc0. It is also very easy to add new network formats as the network structure is saved with the weights unlike lc0's.
All it cares about are the input planes and outputs (value+policy) match, you can use any operations within the body.

Can a user define number of planes, number of auxiliary inputs? Can she define policy output space and value category space?

[Daniel Shawul]

It is not yet configurable like that but maybe i should do that if there is enough interest in it.
At the moment, you have to add code manually to make your network work with egbbdll, but it is not a lot of work.
Currently it supports the following input and output planes.

Inputs:

Code: Select all

Simple format: 12 piece locations 
Scorpio format: 12 piece + 12 attack planes
Lc0 format: 8 history planes of 13 planes, and other planes for castling, fifty etc

Outputs:

Value:

Code: Select all

Single value (lc0 30x nets use this)
WDL value (Scorpio uses this)

Policy:

Code: Select all

A0 format:   73 convolutions for the policy head (~4000 policy outputs) -- scorpio uses this too
Lc0 format:  1858 policy outputs

Adding other move to index mappings should be possible with a little bit of work but probably no need.

If you tell me your input/output planes, i can quickly implement them for you.

I tested scorpio+using leela ID-32930 using V100 against stockfish-dev-single thread at 40 / 1min. The score was about 52% after 1000 games
for scorpio+leela. So it is pretty competitive implementation that can be used with any network format with much ease.

Daniel

[chrisw]

It is not yet configurable like that but maybe i should do that if there is enough interest in it.
At the moment, you have to add code manually to make your network work with egbbdll, but it is not a lot of work.
Currently it supports the following input and output planes.

Inputs:

Code: Select all

Simple format: 12 piece locations 
Scorpio format: 12 piece + 12 attack planes
Lc0 format: 8 history planes of 13 planes, and other planes for castling, fifty etc

Outputs:

Value:

Code: Select all

Single value (lc0 30x nets use this)
WDL value (Scorpio uses this)

Policy:

Code: Select all

A0 format:   73 convolutions for the policy head (~4000 policy outputs) -- scorpio uses this too
Lc0 format:  1858 policy outputs

Adding other move to index mappings should be possible with a little bit of work but probably no need.

If you tell me your input/output planes, i can quickly implement them for you.

I tested scorpio+using leela ID-32930 using V100 against stockfish-dev-single thread at 40 / 1min. The score was about 52% after 1000 games
for scorpio+leela. So it is pretty competitive implementation that can be used with any network format with much ease.

Daniel

Oh, thanks! I’ve not yet tied down any of the final numbers, input planes are about 100, but that’s targeted to increase. Enhanced numeric inputs fed into the value head are 12, I think, but targeted to increase when I get round to it. Policy out is a wasteful 4096, and value out is 3, policy is targeted for efficiency reduction, and value is open to experiments, but none are at top of todo list. Right now everything is stable and learning gradual and consistent, so really, the next step upwards is search. The basics of UCT are straightforward, but I admit to being a bit unclear on how to parallelise, wait and find other useful things to do.