syzygy users (and Ronald)

[Evert]

I like simple, particularly when there is no down-side at all...

There is one: you're now dependent on the availability of the intrinsic, which isn't portable (works on gcc and clang, will/may not on other compilers).

[basil00]

Personally I have always preferred integrating TB probing code as tightly as possible with the engine for maximum performance. My probing code needs to generate and play out captures and what better functions to use for that than the engine's native move generation. (Back in the days of the Edwards tables I noticed that removing the multiple conversions of board representation on every probe doubled probing speed.)

It is possible to integrate the engine's "attacks" functions if compatible, which results in a closer integration. However for move generation and "do move" Fathom uses its own code mainly because it was too difficult to abstract out. In the case of Gull however, the Fathom code is probably faster, since Gull's move generator assumes evaluate() is called first.

[bob]

I like simple, particularly when there is no down-side at all...

There is one: you're now dependent on the availability of the intrinsic, which isn't portable (works on gcc and clang, will/may not on other compilers).

Also works on MSVC and Intel. Also IBM's C compiler...

[jdart]

I don't know about clang, but last I checked the GCC popcnt builtin was very slow, on systems w/o hardware popcnt. On these systems the software popcnt implementation in the Fathom library is much faster.

--Jon

[bob]

I don't know about clang, but last I checked the GCC popcnt builtin was very slow, on systems w/o hardware popcnt. On these systems the software popcnt implementation in the Fathom library is much faster.

--Jon

I looked at the code on my MacBook. It uses the folding idea that is in the syzygy code. I'll try to dig it out and publish it here, although I think I found it on the net with google also...