Subject: re: Issue with large memory systems, and PPC overhead
To: matthew green <mrg@eterna.com.au>
From: Matt Thomas <matt@3am-software.com>
List: tech-kern
Date: 11/10/2002 12:44:14
At 05:52 PM 11/8/2002, matthew green wrote:
>
>    It's interesting to see that the Altivec version of pmap_copy_page
>    (vcopypage) is so much faster than the non-Altivec version.
>    Given that pmap_remove is 32us/call on the G4/400 and 63us on the
>    604ev/180 (roughly scaling with CPU speed), the difference between
>    19us for vcopypage and 136us for pmap_copy_page is amazing.  That's
>    over 3 times as fast as you'd except pmap_copy_page to be on the G4.
>
>
>do you use any of the ds* altivec insructions?

Actually, upon further reflection, I don't want to.  I use the LRU
vector instructions in vzeropage and vcopypage so they won't pollute
the L1/L2/L3 caches.  Using the data stream touch instructions would
pollute the cache too much for me.


-- 
Matt Thomas               Internet:   matt@3am-software.com
3am Software Foundry      WWW URL:    http://www.3am-software.com/bio/matt/
Cupertino, CA             Disclaimer: I avow all knowledge of this message