Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Population of disjoint Attacksets

Author: Dieter Buerssner

Date: 12:50:27 05/29/04

Go up one level in this thread


On May 29, 2004 at 06:27:47, Gerd Isenberg wrote:


Hi Gerd,

popcount inlined

>time determineAttackCounts = 73.345 sec/10**9 runs
>time 8,4,2,1-popCount      = 99.924 sec/10**9 runs
>time Dieter's pc_10bb      = 85.292 sec/10**9 runs

>determineAttackCounts, popCount and pc_10bb inlined:
>time 8,4,2,1-popCount      = 54.028 sec/10**9 runs
>time Dieter's pc_10bb      = 85.122 sec/10**9 runs

>Hmm, those dump loop tests are really not reliable!

strange indeed. Actually I would have guessed, that more inlining hurts here. Do
you remember the "Omid abs" (or was it max?), that was faster, than the dummy
loop?

I got

time determineAttackCounts = 84.351 sec/10**9 runs
time 8,4,2,1-popCount      = 134.213 sec/10**9 runs
time Dieter's pc_10bb      = 99.543 sec/10**9 runs

Without any inlines (besides the ones that have inline in the source). Now I
want to play the author's tourney at ICC, so I cannot test more.

>Should i try performance counter and cpuid?

Perhaps, I will try it later.

>Anyway determineAttackCounts(10) takes 34 (6*5 + 2*2) bitwise 64-bit
>instructions that are 68 32-bit instructions! And on x86-32 there are a lot of
>store/loads with local vars. That code cries for 64-bit and a few more registers
>as well.

Sure. I did not look at the assembly yet, but I guess my 10bb will also have
masses of loads/stores.

Cheers,
Dieter




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.