syzygy wrote: ↑Thu Jul 09, 2020 12:35 pm
Unfortunately the node counts of longer benches are still off
For example bench 512 1 24: 207806902 for Cfish, 209367314 for SF.
512 is just a number I picked. For other hash sizes the depth from where node counts start to differ may be lower or higher than 24.
It seems I missed the patch that decided that Stockfish should not evaluate this position as a draw:
[d]8/8/5k1p/8/7p/7K/8/8 b - - 1 11
Good catch Ronald, I had just tracked it down the simplification patch made on March 14 and you had already had posted a comment about the same patch on Github.
syzygy wrote: ↑Thu Jul 09, 2020 12:35 pm
I also noticed that, on my laptop, Cfish and SF are about equal in speed when I set the hash size to 1024 or higher. I haven't tried on my desktop yet, but this surprised me a bit.
syzygy wrote: ↑Thu Jul 09, 2020 12:35 pm
I also noticed that, on my laptop, Cfish and SF are about equal in speed when I set the hash size to 1024 or higher. I haven't tried on my desktop yet, but this surprised me a bit.
This makes a difference if /sys/kernel/mm/transparent_hugepage/enabled is set to "madvise".
This might explain why my system is set to "always". It was "madvise" and I recall changing it to "always". But I can't remember why. Let me revert it and check that Cfish still uses anonymous hugepages.
{EDIT} Yes, all is good.
$ cat /sys/kernel/mm/transparent_hugepage/enabled
always [madvise] never
syzygy wrote: ↑Thu Jul 09, 2020 12:35 pm
I also noticed that, on my laptop, Cfish and SF are about equal in speed when I set the hash size to 1024 or higher. I haven't tried on my desktop yet, but this surprised me a bit.