bug-gnubg
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Bug-gnubg] Wrong evaluations with gcc 7 and AVX


From: Philippe Michel
Subject: [Bug-gnubg] Wrong evaluations with gcc 7 and AVX
Date: Tue, 12 Dec 2017 22:27:59 +0100
User-agent: Mutt/1.9.1 (2017-09-22)

In a post in the "Gnubg on MacOS High Sierra", I mentionned that I had 
noticed a problem with evaluations when built with gcc 7 and using SIMD 
instructions but only on FreeBSD.

My recollection was not too accurate. In fact the issue is only with 
AVX and probably on all operating systems.

A quick way to check is to run this from the CLI :

set evaluation chequerplay eval plies 0
set gnubgid 4HPwATDgc/ABMA:cIklAAAAAAAE
hint

I get the following results. Even if it is not obvious that a 0.2 equity 
is a bit high, it is clear that 24/20 cannot be that close to best.

FreeBSD 11, gcc 7.2.0, AVX
CentOS 6, gcc 7.2.1, AVX

    1. Cubeful 0-ply    8/5 6/5                      Eq.: +0.204
       0.602 0.186 0.010 - 0.398 0.137 0.034
        0-ply cubeful prune [expert]
    2. Cubeful 0-ply    24/20                        Eq.: +0.198 (-0.006)
       0.599 0.152 0.004 - 0.401 0.128 0.027
        0-ply cubeful prune [expert]

FreeBSD 11, clang 5.0.0, AVX
FreeBSD 11, gcc 7.2.0, SSE2
FreeBSD 11, gcc 7.2.0, no SIMD
CentOS 6, gcc 7.2.1, SSE2
FreeBSD 11, gcc 6.4.0, AVX
CentOS 6, gcc 6.3.1, AVX
CentOS 6, gcc 4.4.7, AVX

    1. Cubeful 0-ply    8/5 6/5                      Eq.: +0.102
       0.551 0.174 0.013 - 0.449 0.124 0.005
        0-ply cubeful prune [expert]
    2. Cubeful 0-ply    24/23 13/10                  Eq.: -0.006 (-0.108)
       0.497 0.137 0.008 - 0.503 0.140 0.007
        0-ply cubeful prune [expert]




reply via email to

[Prev in Thread] Current Thread [Next in Thread]