HexWiki - User contributions [en]

Building KataHex

2026-01-19T02:14:35Z

4Dummies: Info on downloads

There is more than one version of KataHex out there, so these instructions have variants

----
== Ubuntu Linux ==
You can get Selinger's version of KataHex with the following command, which connects you to the Git history, and enables you to contribute edits:
<pre>
git clone https://github.com/selinger/katahex.git
</pre>

This creates a katahex directory wherever you execute that command. If you want to build multiple variants, you will have to do a separate load for each one
(trying to reuse a build directory creates chaos). Just rename the one you just created to reflect the variant name, and do another git clone.

You may have to install additional software to work with KataHex, using <code>apt-get</code> or <code>synaptic</code> to install packages.
For example, the TCMALLOC option protects from problems related to memory management when you run more than one instance at at time. You may have to install one or more packages
* <code>libtcmalloc-minimal4t64</code>
* <code>librust-tcmalloc-dev</code>
* <code>librust-tcmalloc-sys-dev</code>

The code works with at least 4 backends (software where the real work is done), EIGEN, OPENCL, CUDA, AND TENSORRT

==== EIGEN (CPU only) ====

The following BASH commands then may do the trick, but will produce code that may only work on the exact CPU model you use for the build. See comments in the KataHex source code, as well as the KataGo source code (on with KataHex is based, and which contains more details).

<pre>
cmake . -DUSE_BACKEND=EIGEN -DUSE_AVX2=1 -DUSE_TCMALLOC=1 -DCMAKE_CXX_FLAGS='-march=native' -DMAX_BOARD_LEN=19
make -j 4
</pre>

==== OPENCL ====
You may be able to use almost the same commands as for EIGEN, just by changing to -DUSE_BACKEND=OPENCL

==== CUDA ====
I have not yet been able to build with the CUDA backend

==== TENSORRT ====
I have not yet been able to build with the TENSORRT backend

Building KataHex

2026-01-19T02:10:30Z

4Dummies: Clarify

There is more than one version of KataHex out there, so these instructions have variants

----
== Ubuntu Linux ==
You can get Selinger's version of KataHex with the following command, which connects you to the Git history, and enables you to contribute edits:
<pre>
git clone https://github.com/selinger/katahex.git
</pre>

You may have to install additional software to work with KataHex, using <code>apt-get</code> or <code>synaptic</code> to install packages.

For example, the TCMALLOC option protects from problems related to memory management when you run more than one instance at at time. You may have to install one or more packages
* <code>libtcmalloc-minimal4t64</code>
* <code>librust-tcmalloc-dev</code>
* <code>librust-tcmalloc-sys-dev</code>

The code works with at least 4 backends (software where the real work is done), EIGEN, OPENCL, CUDA, AND TENSORRT

==== EIGEN (CPU only) ====

The following BASH commands then may do the trick, but will produce code that may only work on the exact CPU model you use for the build. See comments in the KataHex source code, as well as the KataGo source code (on with KataHex is based, and which contains more details).

<pre>
cmake . -DUSE_BACKEND=EIGEN -DUSE_AVX2=1 -DUSE_TCMALLOC=1 -DCMAKE_CXX_FLAGS='-march=native' -DMAX_BOARD_LEN=19
make -j 4
</pre>

==== OPENCL ====
You may be able to use almost the same commands as for EIGEN, just by changing to -DUSE_BACKEND=OPENCL

==== CUDA ====
I have not yet been able to build with the CUDA backend

==== TENSORRT ====
I have not yet been able to build with the TENSORRT backend

Building KataHex

2026-01-19T01:53:39Z

4Dummies: Add: how to get and build the code

There is more than one version of KataHex out there, so these instructions have variants

----
== Ubuntu Linux ==
You can get Selinger's version of KataHex with the following command, which connects you to the Git history, and enables you to contribute edits:
<pre>
git clone https://github.com/selinger/katahex.git
</pre>

You may have to install additional software to work with KataHex, using <code>apt-get</code> or <code>synaptic</code> to install packages.

For example, the TCMALLOC option protects from problems related to memory management when you run more than one instance at at time. You may have to install one or more packages
* <code>libtcmalloc-minimal4t64</code>
* <code>librust-tcmalloc-dev</code>
* <code>librust-tcmalloc-sys-dev</code>

The code works with at least 4 backends (software where the real work is done), EIGEN, CUDA, OPENCL AND TENSORRT.

==== EIGEN (CPU only) ====

The following BASH commands then may do the trick, but will produce code that may only work on the exact CPU model you use for the build. See comments in the KataHex source code, as well as the KataGo source code (on with KataHex is based, and which contains more details).

<pre>
cmake . -DUSE_BACKEND=EIGEN -DUSE_AVX2=1 -DUSE_TCMALLOC=1 -DCMAKE_CXX_FLAGS='-march=native'
make -j 4
</pre>

==== CUDA ====
I have not yet been able to build with the CUDA backend

==== TENSORRT ====
I have not yet been able to build with the TENSORRT backend

==== OPENCL ====
I have not yet been able to build with the OPENCL backend

Building KataHex

2026-01-18T20:18:08Z

4Dummies:

There is more than one version of KataHex out there, so these instructions have variants

----
== Ubuntu Linux ==
You may have to use <code>apt-get</code> or <code>synaptic</code> to install packages
the TCMALLOC option protects from problems related to memory management when you run more than one instance at at time. You may have to install one or more packages
* <code>libtcmalloc-minimal4t64</code>
* <code>librust-tcmalloc-dev</code>
* <code>librust-tcmalloc-sys-dev</code>

==== EIGEN (CPU only) ====

The following BASH commands then may do the trick, but will produce code that may only work on the exact CPU model you use for the build. See comments in the KataHex source code, as well as the KataGo source code (on with KataHex is based, and which contains more details).

<pre>
cmake . -DUSE_BACKEND=EIGEN -DUSE_AVX2=1 -DUSE_TCMALLOC=1 -DCMAKE_CXX_FLAGS='-march=native'
make -j 4
</pre>

==== CUDA ====
I have not yet been able to build with the CUDA backend

==== TENSORRT ====
I have not yet been able to build with the TENSORRT backend

==== OPENCL ====
I have not yet been able to build with the OPENCL backend

Building KataHex

2026-01-18T19:40:55Z

4Dummies: creation

There is more than one version of KataHex out there, so these instructions have variants

----
== Ubuntu Linux ==
You may have to use <code>apt-get</code> or <code>synaptic</code> to install packages

==== EIGEN (CPU only) ====

You may have to install packages <code>libtcmalloc-minimal4t64</code>,
<code>librust-tcmalloc-dev</code>, and
<code>librust-tcmalloc-sys-dev</code>

The following BASH commands then may do the trick, but will produce code that may only work on the exact CPU model you use for the build. See comments in the KataHex source code, as well as the KataGo source code (on with KataHex is based, and which contains more details).

<pre>
cmake . -DUSE_BACKEND=EIGEN -DUSE_AVX2=1 -DUSE_TCMALLOC=1 -DCMAKE_CXX_FLAGS='-march=native'
make -j 4
</pre>

==== CUDA ====
I have not yet been able to build with the CUDA backend

==== TENSORRT ====
I have not yet been able to build with the TENSORRT backend

==== OPENCL ====
I have not yet been able to build with the OPENCL backend

Building KataHex

2026-01-18T19:35:29Z

4Dummies: created

There is more than one version of KataHex out there, so these instructions have variants

=== Ubuntu Linux ===
You may have to use <code>apt-get</code> or <code>synaptic</code> to install packages

==== EIGEN (CPU only) ====

You may have to install packages <code>libtcmalloc-minimal4t64</code>,
<code>librust-tcmalloc-dev</code>, and
<code>librust-tcmalloc-sys-dev</code>

The following BASH commands then may do the trick, but will produce code that may only work on the exact CPU model you use for the build. See comments in the KataHex source code for details.

<pre>
cmake . -DUSE_BACKEND=EIGEN -DUSE_AVX2=1 -DUSE_TCMALLOC=1 -DCMAKE_CXX_FLAGS='-march=native'
make -j 4
</pre>

==== CUDA ====
I have not yet been able to build with the CUDA backend

Talk:Swap rule

2024-02-11T02:13:41Z

4Dummies: Answered

Does anybody know when the swap rule was introduced?

Apparently near the beginning of the 20th century in a game called mancala. [https://en.wikipedia.org/wiki/Pie_rule]

== Swapping the A column? ==

It has occurred to me to doubt the wisdom of swapping most of the left- or right-most columns. I'm just thinking about the templates and how they can block the opening move.

For example, on an 11x11 board, A4 is marked as a probable winner, but the bridge template played at B5 seems to be a good block and when I work it out, A4 seems to lose.
There are also template moves at C5 (ziggurat), and some fourth and fifth-row templates that seem promising.

Comments? [[User:4Dummies|4Dummies]] ([[User_talk:4Dummies|talk]]) 13:49, 8 February 2024‎

The problem is that B5 is not actually a good response to A4. Although B5 is a template and a ladder escape for White, A4 still gives Black a [[switchback]] for 2nd row ladders. Moreover, after White plays B5, it's Black's turn again, so White has lost the initiative. After A4 and B5, if Black plays anywhere near the center, Black is almost definitely winning (97-99% win rate according to [[KataHex]]). For example, H4, G5, D8 are all excellent choices for Black's next move. [[User:Selinger|Selinger]] ([[User talk:Selinger|talk]]) 19:16, 8 February 2024 (UTC)

Thanks for that. A bunch. I had just realized that B5 corresponds to an opening (swapped) at E2, which is noted as losing. So it's not so good on the face of it. However, there are taller templates that fit between A4 and the bottom of an 11x11 board, and they might be a bunch better.

Even more important in your reply was that I had never before heard of KataHex. I'm real interested in training an AI on my data. I have about 1.6 million games (no kidding) of 11x11 hex, gathered from real games and much amplified by six with a little guidance from HexHex. I don't know how to train HexHex either, but it sounds like KataHex is probably stronger. I just don't know how to do the training. My data is in HSGF format, that I process using the hexrec program that originally came from gamerecorder.

Talk:Swap rule

2024-02-08T19:47:13Z

4Dummies: followup

Does anybody know when the swap rule was introduced?

== Swapping the A column? ==

It has occurred to me to doubt the wisdom of swapping most of the left- or right-most columns. I'm just thinking about the templates and how they can block the opening move.

For example, on an 11x11 board, A4 is marked as a probable winner, but the bridge template played at B5 seems to be a good block and when I work it out, A4 seems to lose.
There are also template moves at C5 (ziggurat), and some fourth and fifth-row templates that seem promising.

Comments? [[User:4Dummies|4Dummies]] ([[User_talk:4Dummies|talk]]) 13:49, 8 February 2024‎

The problem is that B5 is not actually a good response to A4. Although B5 is a template and a ladder escape for White, A4 still gives Black a [[switchback]] for 2nd row ladders. Moreover, after White plays B5, it's Black's turn again, so White has lost the initiative. After A4 and B5, if Black plays anywhere near the center, Black is almost definitely winning (97-99% win rate according to [[KataHex]]). For example, H4, G5, D8 are all excellent choices for Black's next move. [[User:Selinger|Selinger]] ([[User talk:Selinger|talk]]) 19:16, 8 February 2024 (UTC)

Thanks for that. A bunch. I had just realized that B5 corresponds to an opening (swapped) at E2, which is noted as losing. So it's not so good on the face of it. However, there are taller templates that fit between A4 and the bottom of an 11x11 board, and they might be a bunch better.

Even more important in your reply was that I had never before heard of KataHex. I'm real interested in training an AI on my data. I have about 1.6 million games (no kidding) of 11x11 hex, gathered from real games and much amplified by six with a little guidance from HexHex. I don't know how to train HexHex either, but it sounds like KataHex is probably stronger. I just don't know how to do the training. My data is in HSGF format, that I process using the hexrec program that originally came from gamerecorder.

Talk:Swap rule

2024-02-08T16:49:07Z

4Dummies: /* Swapping the A column? */ new section