Re: AMD Geode GX/LX Support (Refreshed)

Re: AMD Geode GX/LX Support (Refreshed)

[John Zulauf]

Alan,

I've done the benchmarking of the 3DNOW and non-3DNOW configured Geode
GX/LX.  This was done by extracting the kernel memcpy code into a
userland test application.  Time measurement were made using rdtsc.

The results are averaged over 4 test runs on a Geode LX. The tot_memcpy
is the total CPU clocks to run every memcpy from len = 2
through 4095 four times.  For the 3DNOW version this even includes the 

if (len <512){  __memcpy ... } 

escape found in the kernel __memcpy3d mmx_memcpy wrapper function. 

               No 3DNOW w/ 3DNOW  improvement
tot_memcpy      4635754 4360040  5.9%
tot_copy_page      2324    1853 20.3%
tot_clear_page     4398    1427 67.5%

These are meaningful improvement and the 3DNOW config should be enabled
for Geode GX and LX.

Best Regards,

John Zulauf

On Tue, 2005-11-01 at 23:53 +0000, Alan Cox wrote:
> On Maw, 2005-11-01 at 16:10 -0700, Jordan Crouse wrote:
> > in my app, I think we should leave X86_USE_PPRO_CHECKSUM enabled for Geode
> > GX/LX.
> 
> Definitely.
> 
> 
> 
> 
> ---
> You are currently subscribed to info-linux@geode.amd.com
> as: john.zulauf@amd.com
> To unsubscribe send a blank email to:
> leave-info-linux-4271N@whitestar.amd.com
-- 
John Zulauf
Senior MTS, Embedded Core Software
Advanced Micro Devices, Inc.
1351 So Sunset St
Longmont, CO 80501

303 774 5166 office
303 774 5801 fax

Re: AMD Geode GX/LX Support (Refreshed)

[Alan Cox]

On Maw, 2005-11-01 at 16:10 -0700, Jordan Crouse wrote:
> in my app, I think we should leave X86_USE_PPRO_CHECKSUM enabled for Geode
> GX/LX.

Definitely.

Re: AMD Geode GX/LX Support (Refreshed)

[Jordan Crouse]

> >  config X86_USE_PPRO_CHECKSUM
> Does this mean you've now done actual performance analysis on whether
> this is a good idea for Geode GX/LX ?

Ok - here is some data that should put your mind at ease.  I pulled
the checksum functions from the kernel, stuck them in userland, and
ran each one N times at a variety of byte sizes (where N=10,000), took the
start and stop times with gettimeofday(), and added it all up.  The results
for the Geode LX platfrom are below.

I think the data shows that the ppro checksum is indeed useful on the
Geode LX platform (and the GX should have similar results, since the
pipelines are pretty close).  Unless I made a really boneheaded mistake
in my app, I think we should leave X86_USE_PPRO_CHECKSUM enabled for Geode
GX/LX.

Jordan

Starting non-ppro test....
Bytes   Time    Avg (usec/run)
----------------------------
16      2029    0.202900
32      1666    0.166600
64      2548    0.254800
128     3429    0.342900
256     5268    0.526800
512     8781    0.878100
1024    16031   1.603100
2048    30548   3.054800
4096    59265   5.926500
Total time: 129565

Starting ppro test....
Bytes   Time    Avg (usec/run)
----------------------------
16      1579    0.157900
32      1655    0.165500
64      1841    0.184100
128     2721    0.272100
256     3582    0.358200
512     5264    0.526400
1024    8900    0.890000
2048    15919   1.591900
4096    29369   2.936900
Total time: 70830

Re: AMD Geode GX/LX Support (Refreshed)

[Jordan Crouse]

> Does this mean you've now done actual performance analysis on whether
> this is a good idea for Geode GX/LX ?


I'm fairly sure that the ppro checksum is faster, but you're exactly right,
we should publish some numbers.  Mia culpa.

Jordan

-- 
Jordan Crouse
Senior Linux Engineer
AMD - Personal Connectivity Solutions Group
<www.amd.com/embeddedprocessors>