On Sun, Dec 10, 2017 at 06:49:17PM +1100, Marcell Fulop wrote:

> Much more interestingly, I found significant differences in match analyses.
> I've saved hundreds of those thousands matches played previously along with
> analyses. Now, when I open such a match, then open the analysis (Analyse ->
> Match or session statistics), then I open the same match in another
> instance but here first I clear then analyse again (Analyse -> Clear
> analysis -> Match or session; Analyse -> Analyse match or session), the
> results are not even remotely similar.
> 
> Both versions are 1.04.000. The Player and Analysis settings used are
> always the predefined Grandmaster with no modifications. I probably didn't
> build this version under High Sierra "right" because of the stupid
> behaviour, and because these differences are consistently showing much
> lower ratings in case of the latter for the same match.
> 
> Any ideas?

You wrote earlier that you built it with gcc 7. A few months ago, I 
tried to build gnubg on FreeBSD with its gcc 7 port (that was tagged as 
a development version at the time) and the evaluations were broken when 
using SIMD instructions (SSE or AVX). 

Other cases (gcc 7 when not using SIMD instructions, clang or earlier 
gcc versions, gcc 7 with SIMD instructions on Linux) worked fine. I 
didn't pursue the matter (the compiler was a development version...) and 
gave up on gcc 7 on FreeBSD.

I think gcc creates the include files it uses from the system files. 
Since MacOS and FreeBSD both use clang as system compiler, could it be 
possible that its SIMD intrinsics include files are somehow broken when 
generated from clang's ones ? Or that we use them in a subtly wrong way 
and don't get away with it any more in this configuration ?

You could try to build gnubg configured with --enable-simd=no with gcc 7 
(it will be 2-3 times slower), or in the default configuration with 
another compiler and see if it makes a difference.

In any case, the evaluations from the 1.x versions with various 
compilers and options should be identical or very close in terms 
of equity. The selected move may occasionally change in positions where 
some candidates are tied or almost tied but averaged data like analysis 
error rates shouldn't differ by more than a few tenth of a percent.

_______________________________________________
Bug-gnubg mailing list
Bug-gnubg@gnu.org
https://lists.gnu.org/mailman/listinfo/bug-gnubg

Reply via email to