85th Amateur Series Division 4

Discussion of computer chess matches and engine tournaments.

Moderators: hgm, Rebel, chrisw

abulmo2
Posts: 433
Joined: Fri Dec 16, 2016 11:04 am
Location: France
Full name: Richard Delorme

Re: 85th Amateur Series Division 4

Post by abulmo2 »

Graham Banks wrote: Wed Apr 21, 2021 5:09 am You'd be best doing what others do - a non-popcount exe, a popcount exe and a bmi2 exe.
Well Stockfish 13, the reference, is published in six exe: bmi2, avx2, sse4.1+popcount, sse3, 64bit and 32bit.
The problem, for a rating list, is that all these versions should have a different rating as they do not perform as fast.
For example, on my computer amoeba 3.3 popcount is 28% faster than the non popcount version which could translate into ~15 Elo on a rating list. Mixing them will make the rating somewhat noisy I guess.
Richard Delorme
User avatar
xr_a_y
Posts: 1871
Joined: Sat Nov 25, 2017 2:28 pm
Location: France

Re: 85th Amateur Series Division 4

Post by xr_a_y »

For next release I'll have

core2 = SSE3, NO POPCNT
nehalem = SSE4.2, POPCNT
sandy-bridge = AVX, POPCNT
skylake = AVX2, POPCNT, BMI2

POPCNT affects many bitboard stuff of course, BMI2 affects "magics", and vectorisation affects NNUE.

I once posted that (http://talkchess.com/forum3/viewtopic.p ... on#p882163) showing that there is 100Elo between SSE3 and AVX2 build when NNUE is used (this means Minic with noisynotch using SSE3 is probably weaker than old Minic 2.46 using HCE).