Am Donnerstag, 9. Januar 2003 14:05 schrieb Petr Sebor:
> Hello,
>
> somehow something got ommited, probably on my side.
> [as I can tell from the end of the patch, since there are references
> to removed code...]
>
> I'd like to send more cleanups if you are interested ;-)

Come along with it.

> Btw: Unfortunatelly, I have noticed that the current SSE code tends to
> be a little slower than the x86 path, at least on my XP2000+... not very
> much, but still.

I second that.

> Maybe this is given by the AMD SSE implementation

Could be. As I understand the AMD SSE implementation right, it isn't an extra 
CPU core but it use the 3DNow! infrastructure.

> or more likely, because we don't smartly precache data while we can.

Precaching is NOT by all means the fastest patch.
See AMD's developer notes/optimization guides.

There was discussion on LKM about a new memcopy routine.
http://marc.theaimsgroup.com/?l=linux-kernel&m=103548024914815&w=2
[-]
AMD recommends to perform memory copies with backward read operations 
instead of prefetch.

http://208.15.46.63/events/gdc2002.htm
[-]

Some numbers:
nuetzel/Entwicklung> ./athlon-DN
1600.156 MHz
clear_page by 'normal_clear_page'        took 12400 cycles (504.1 MB/s)
clear_page by 'slow_zero_page'           took 12412 cycles (503.6 MB/s)
clear_page by 'fast_clear_page'          took 9811 cycles (637.1 MB/s)
clear_page by 'faster_clear_page'        took 4239 cycles (1474.3 MB/s)

copy_page by 'normal_copy_page'  took 9231 cycles (677.1 MB/s)
copy_page by 'slow_copy_page'    took 9243 cycles (676.3 MB/s)
copy_page by 'fast_copy_page'    took 8318 cycles (751.4 MB/s)
copy_page by 'faster_copy'       took 5481 cycles (1140.3 MB/s)
copy_page by 'even_faster'       took 5545 cycles (1127.2 MB/s)
copy_page by 'no_prefetch'       took 4388 cycles (1424.3 MB/s)

But maybe you should stop the CPU scanning on AMD after 3DNow! 
(normal/ext/prof) detection, so that AMD's run 3DNow! and not SSE.

-Dieter


-------------------------------------------------------
This SF.NET email is sponsored by:
SourceForge Enterprise Edition + IBM + LinuxWorld = Something 2 See!
http://www.vasoftware.com
_______________________________________________
Dri-devel mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/dri-devel

Reply via email to