Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Matt Taylor's magic de Bruijn Constant

Author: Robert Hyatt

Date: 10:27:28 07/15/03

Go up one level in this thread


On July 15, 2003 at 12:06:30, Gerd Isenberg wrote:

>On >You can answer _all_ of the above by running lm-bench.  It will tell
>>you each one of those numbers, plus others.
>>
>
>oups - lm-bench is pure unix/linux tool, not yet installed.
>Is there something similar for Windows?

So far as I know, it will compile on windows, although I am not sure since
I haven't tried it.  I run it on our unix boxes (linux mostly) of course.



>
>What i miss, is a well defined algorithm, to messure latency correctly.
>The several implementations may vary a lot dependent on choosing random
>addresses. Is there always a real total cache miss?

Yes.  There has to be, otherwise you measure something besides memory
latency.  If you look up lm-bench on the net, you should find a reference
to a couple of online-papers that you can download and read.  They describe
the things lm-bench does in detail.



>
>At least i got an idea of this huge latencies -
>and it may be dangerous to hash with huge tables in qsearch.
>
>But i guess it is possible to hide this latencies in some way.
>What about PREFETCHING the hashaddress and doing few hundred cycles pure
>register processing eg. with mmx, sse2 and gp-registers (or some read/writes to
>often used and cached source and target blocks), before really reading the
>hashaddress?



That would work.  We used to call that "instruction lifting" in the Cray.  There
you would do something like s1   ,memory-location which says load
memory-location contents into register s1.  You "lift" that instruction up as
far as possible so that when you later refer to s1, there is no wait, the value
is already loaded.  Sort of like issuing a read before you need the data, then
checking to see if it has finished when you really need it.

There is a pre-fetch facility built in to new processors, but the compilers I
have looked at don't seem to use it and I haven't studied it very much myself
as a result.


>
>Cheers,
>Gerd



This page took 0.01 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.