Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Avoiding if ( expr ) and branch misprediction ?

Author: Gerd Isenberg

Date: 11:41:07 01/19/06

Go up one level in this thread


On January 19, 2006 at 13:44:01, Dann Corbit wrote:

>On January 19, 2006 at 12:45:10, Chan Rasjid wrote:
>
>>
>>When using bitboards, often we have arrays that require only indices 0 / 1.
>>eg if we have no underpromotions, then there is either 1/2 knights per color.
>>Sometimes ,in such situations, codes may be re-written without using the if ()
>>statement. But whether such simple tricks can avoid the cost of branch
>>misprediction ( I only have a vague idea about this ).
>>
>>X[i] where i = 0 / 1;
>>
>>1) if ( a & b ){
>>     j = X[1];
>>   }else{
>>     j = X[0];
>>   }
>>   //etc.. codes dependent on j.
>>
>>2) Now without if() :-
>>
>>   j = X[(a & b)!= 0];
>>   //etc.. codes dependent on j.
>>
>>How (if ? ) can such a simple re-coding cause an improvement. Can there
>>be some situations related to the above with significanr misprediction
>>overheads ?
>
>If you have only 0 and 1 outcomes:
>
>Use an array of 2 function pointers and do this:
>
>assert(expression == 0 || expression == 1);
>foo[expression]();

No - avoid that for simple expressions!

>
>Not any great shakes, but it might be worthwhile.

Branch prediction is very poor with indirect branches.
If expression toggles each time 100% misspredictions for amd64 (and P4), while
it is perfectly predictable with conditional jumps. Also there is additional
call/ret overhead - the ret needs branch prediction ressources as well.

>
>Most of the time, this sort of stuff is a complete waste of time.
>
>The best way to speed up your program is to improve the algorithms.
>The best goal would be to reduce the branching factor while maintaining the good
>decisions.

Yes - of course assumed perfectly done already ;-)

Gerd

>
>IMO-YMMV.



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.