Re: [gnugo-devel] paul_3_13.6

[Dave Denholm]

Paul Pogonyshev <address@hidden> writes:

> this patch fixes tree-based pattern matcher and speeds the
> matching up a bit.
> 
> - tree-based pattern matcher fixed
> - pattern matchers sped up
> 
> here is the full list of speedups:
> 

> - the order of indices in transformation[][] and spiral[][]
>   arays swapped. improves caching.

I'm confused...

>        
>      /* go to next state */
> -    delta = pdfa->states[state].next[dfa_p[dfa_pos + spiral[l][row]]];
> +    delta = pdfa->states[state].next[dfa_p[dfa_pos + spiral[row][l]]];
>      state += delta;
>      row++;
>    } while (delta != 0); /* while not on error state */
>  

I see the inner loop being over row, not 'l'. So swapping the
indices hurts the cache, rather than helping it ?

If there are enough registers, compiler should be able to put
spiral[row] into a register.

x86 can scale memory accesses by small powers of 2, but 8*sizeof(int)
is probably too big.


Or am I being particularly dense ?

dd
-- 
address@hidden          http://www.insignia.com