Included are a tablebase generator and probing code for adding tablebase probing to a chess engine. The tablebase generator is able to generate all tablebases for up to 6 pieces.
Requirements for the generator:
- 16 GB of RAM for 6-piece tables (much less for 5-piece tables).
- x86-64 CPU.
- 64-bit OS.
- Sufficiently recent gcc (producing 64-bit executables).
File names encode the type of tablebase: K+R+P vs K+R becomes KRPvKR. Each tablebase corresponds to two files: KRPvKR.rtbw and KRPvKR.rtbz. Note that KRPvKR also covers K+R vs K+R+P.
The .rtbw files store win/draw/loss information including, where applicable, information on the 50-move rule. During the search only the .rtbw files are accessed. These files are "two-sided": they store information for both white to move and black to move.
The .rtbz files store the distance-to-zero: the number of moves to the next capture or pawn move. These files only need to be accessed when the root position has 6 pieces or less. They are "single-sided".
The 6-piece WDL tables are 68.2 GB in total. The DTZ tables take up 81.9 GB. For up to 5 pieces, the numbers are 378 MB and 561 MB. Ideally, the WDL tables are stored on an SSD.
The directory src/ contains the tablebase generator code. It should be
easy to build on x86-64 Linux system and on 64-bit Windows with MinGW
("make all"). It might be necessary to edit src/Makefile. In particular,
if your CPU does not support the popcnt instruction, the line FLAGS += -DUSE_POPCNT should be commented out.
There are five programs:
- rtbgen for generating pawnless tablebases.
- rtbgenp for generating pawnful tablebases.
- rtbver for verifying pawnless tablebases.
- rtbverp for verifying pawnful tablebases.
- tbcheck for verifying integrity of tablebase files based on an embedded checksum.
Usage: rtbgen KQRvKR (or rtbgenp KRPvKR)
Produces two compressed files: KQRvKR.rtbw and KQRvKR.rtbz. Both files
contain an embedded checksum.
Options:
--threads n (or -t n)
Use n threads. If no other tasks are running, it is recommended to
set n to the number of CPU cores, or if the CPU supports hyperthreading,
to the number of CPU hyperthreads.
--wdl (or -w)
Only compress and save the WDL file (with .rtbw suffix).
--dtz (or -z)
Only compress and save the DTZ file (with .rtbz suffix).
-g
Generate the table but do not compress and save.
--stats (or -s)
Save statistics. Statistics are written to $RTBSTATSDIR/KQRvKR.txt
or to ./KQRvKR.txt if $RTBSTATSDIR is not set.
--disk (or -d)
Reduce RAM usage during compression. This takes a bit more time because
tables are temporarily saved to disk. This option is necessary to
generate 6-piece tables on systems with 16 GB RAM. This option is
not needed on system with 24 GB RAM or more.
Usage: rtbver KQRvKR (or rtbverp KRPvKR)
Verifies consistency of KQRvKR.rtbw and KQRvKR.rtbz. This should detect
(hardware) errors during generation and compression. For technical reasons
pawnful tables with symmetric material such as KPvKP and KRPvKRP cannot
(at least currently) be verified.
Options:
--threads n (or -t n)
See above.
--log (or -l)
Log verification results to rtblog.txt.
-d
Look for the WDL file in directory $RTBWDIR and look for the DTZ file in
directory $RTBZDIR. Without this option, both files should be present in
the current working directory.
Usage: tbcheck KQRvKR.rtbw KRPvKR.rtbz
Recalculates a checksum for each specified tablebase file and compares with
the embedded checksums. This should detect disk errors and transmission
errors.
Options:
--threads n (or -t n)
See above.
--print (or -p)
Print embedded checksums. Do not check correctness.
Note: The programs rtbgen, rtbgenp, rtbver and rtbverp require access to WDL tablebase files for "subtables". These should be present in the directory $RTBWDIR.
The somewhat primitive perl script src/run.pl can be used for generating and verifying all or part of the tables. Make sure the location of rtbgen, rtbgenp, rtbver and rtbverp is in your $PATH variable.
Usage: run.pl --generate
Options:
--threads n (or -t n)
See above.
--generate
Generate tablebases. Tablebases that already have been generated and are
found in the current working directory are skipped.
--verify
Verify tablebases.
--min n
Only treat tablebases with at least n pieces.
--max n
Only treat tablebases with at most n pieces.
--disk
Use this option to generate 6-piece tables on a system with 16 GB of RAM.
The directory interface/ contains probing code. It does not come in the form of a shared library, and requires some work to integrate into an engine. The main reason for this is efficiency. There are four files: tbcore.c, tbcore.h, tbprobe.cpp, tbprobe.h.
The files tbcore.c and tbcore.h should not require much changes, although engine authors might want to replace some printf()s with suitable logging statements. The files tbprobe.cpp and tbprobe.h do require some changes but these should be fairly straightforward when following the comments. The only reason for tbprobe.cpp having the .cpp extension is that I have used Stockfish as example. The probing code expects WDL files in $RTBWDIR and DTZ files in $RTBZDIR.
The files main.cpp, search.cpp and types.h are from Stockfish with calls to the probing code added (see // TB comments). The change in types.h is necessary in order to make room for "tablebase win in n" values distinct from "mate in n" values. Please note that the integration of probing code into Stockfish is merely intended as a proof of concept. It is far from perfect and might have some bugs.
The files lz4.c and lz4.h in src/ are copyrighted by Yann Collet and were released under the BSD 2-Clause License. The files city-c.c, city-c.h and citycrc.h in src/ (ported by me from C++ to C) are copyrighted by Google, Inc. and were released under an even more liberal license. Both licenses are compatible with the GPL. All other files in src/ are released under the GNU Public License, version 2.
The files main.cpp, search.cpp and types.h in interface/ obviously are copyrighted by the Stockfish authors and covered by the Stockfish GPL.
The files tbcore.c, tbcore.h, tbprobe.cpp and tbprobe.h in interface/ may be freely modified and redistributed in source and/or binary format.
Ronald de Man
[email protected]