Author: Sune Fischer
Date: 00:42:49 08/18/02
Go up one level in this thread
On August 17, 2002 at 23:25:25, Vincent Diepeveen wrote:
>On August 17, 2002 at 14:12:03, Gerd Isenberg wrote:
>
>>Hi Vincent,
>>
>>actually it's faster on my AthlonXP with USE_LOOKUP_XYZ defined!?
>>Strange, i also would bet that the one without lookup is faster,
>>which only needs 4 instructions by your definition. I don't understand this
>>CISC-processors.
>
>Of course the instructions only outperform a table lookup when
>the processors L1 and L2 caches are overloaded busy with storing
>hashtable entries, doing a big evaluation and other stuff and not having
>any time to just put this lookup table in L1 cache.
Do you really think it is faster?
You need aditional off-set calculation (not visual to the C-code) when doing
this lookup, right?
I have a hard time believing a table lookup can beat a naked 4 clock operation,
even fetching things from the L2 catch is 5 clocks (IIRC that was the number
Eugene mentioned once), so it has to be in L1 cache to even stand a chance.
>Only testing the code versus the table is not a good idea obviously.
Right.
If that thing has to be in L1, it must mean that something else has to get out.
>DIEP doesn't fit in L2 cache at all. I don't need to mention what is
>faster for me :)
>
>>Tested in wrong bishop endings and KBN-K (without ETBs), where these inlines are
>>used quite often in eval and recognizers. In KBN-K 660KNodes versus 656KNodes.
>>But i don't played with optimizations so far (MSC++ minimize size optimization).
>
>see above. this is a typical case where the thing can put many relevant
>things in L1 cache.
>
>>May be it's because m_sUDR[a][b] is accessed frequently with quite equal "a" or
>>"b" and therefore is mostly already in first level cache - or a lack of
>>registers. But the code is definitely shorter with lookup.
>>
>>#ifdef USE_LOOKUP_XYZ
>>inline BOOL sameSquareColor(int a, int b) {return (m_sUDR[a][b]^1) & 1;}
>>inline BOOL oppoSquareColor(int a, int b) {return m_sUDR[a][b] & 1;}
>>#else
>>inline BOOL sameSquareColor(int a, int b) {return (((a^b)>>3)^(a^b)^1) & 1;}
>>inline BOOL oppoSquareColor(int a, int b) {return (((a^b)>>3)^(a^b)) & 1;}
>>#endif
>>
>>(((a^b)>>3)^(a^b)) & 1;
>>2 xors because only one (a^b) is necessary
>>1 shift
>>1 and
>>------
>>4 instructions
>
>ah great, you found a faster way :)
Well sort of, he bugfixed my (untested) method ;)
-S.
>>see you,
>>Gerd
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.