Re: [gnugo-devel] new cache for readconnect and semeai

[Arend Bayer]

On Sun, 11 Apr 2004, Inge Wallin wrote:

> > - complete conversion to new cache
>
> Did you find out what I did wrong and why it didn't work?  Was the missing
>
>   str = origin(str);
>
> the only thing?

I think this did cause the increase in owl nodes that Gunnar observed when
owl was converted to new cache. I never got the huge breakage Gunnar
reported, but you used
        if (tt_get(...))
instead of
        if (tt_get(...) == 2)
in semeai reading which caused a couple fails in seki.tst if I remember
correctly.

> > Index: engine/cache.h
> > ===================================================================
> > RCS file: /cvsroot/gnugo/gnugo/engine/cache.h,v
> > retrieving revision 1.44
> > diff -u -p -r1.44 cache.h
> > --- engine/cache.h  5 Feb 2004 22:54:42 -0000       1.44
> > +++ engine/cache.h  11 Apr 2004 05:49:02 -0000
> > @@ -50,11 +50,13 @@
> >   * fields:
> >   *
> >   *   RESERVED       :  5 bits
> > - *   remaining_depth:  5 bits (depth - stackp)  NOTE: 
> > HN_MAX_REMAINING_DEPTH
> >   *   value1         :  4 bits
> >   *   value2         :  4 bits
> >   *   move           : 10 bits
> > - *   flags          :  4 bits
> > + *   cost           :  4 bits
> > + *   remaining_depth:  5 bits (depth - stackp)  NOTE: 
> > HN_MAX_REMAINING_DEPTH
> > + *
> > + *   The last 9 bits together give an index for the total costs.
> >   */
>
> Two things here:
>  - I think that maybe 4 bits for the cost is too much.  We don't need 16
>    levels.  2 bits should be enough.

Well I don't care much. We might decide otherwise later, and we have 5
unused bits in RESERVED anyway.

>  - If you add 4 bits for cost, you should reduce the RESERVED part with
>    the same amount.

Nope. Note that flags is gone. (Which was unused.)

Arend