Re: [PATCH] [2.4] [2.5] [i386] Add support for GCC 3.1

Re: [PATCH] [2.4] [2.5] [i386] Add support for GCC 3.1

[rwhron]

Another processor config could be CONFIG_K62.
gcc-3.1 -march=k6-2 benchmarks a little better 
than -march=k6.  Adding CONFIG_XF86_USE_3DNOW=y 
seems to help a little too.

Based on grepping gcc-3.1 src, it appears:
k6-3 == k6-2
athlon-tbird == athlon
athlon-xp == athlon-4 == athlon-mp

-- 
Randy Hron

Re: [PATCH] [2.4] [2.5] [i386] Add support for GCC 3.1

[Alan Cox]

On Sun, 2002-05-26 at 21:02, rwhron@earthlink.net wrote:
> Another processor config could be CONFIG_K62.
> gcc-3.1 -march=k6-2 benchmarks a little better 
> than -march=k6.  Adding CONFIG_XF86_USE_3DNOW=y 
> seems to help a little too.

Make sure you benchmark FPU intensive use when testing USE_3DNOW. I've
seen several situations where it looks marginally better until you
actually measure two things - memory bandwidth available to user
programs, and also FPU performance, especially of FPU heavy apps that
schedule a lot (eg Quake)

Re: [PATCH] [2.4] [2.5] [i386] Add support for GCC 3.1

[Dave Jones]

On Sun, May 26, 2002 at 04:02:17PM -0400, rwhron@earthlink.net wrote:

 > athlon-xp == athlon-4 == athlon-mp

Not quite...

>From gcc/config/i386/i386.c ..

   if (!ix86_arch_string)
                   ix86_arch_string = TARGET_64BIT ? "athlon-4" : "i386";

Discussion with Jan Hubicka reveals that this should be checking for
"opteron", but as this was the closest match, it's not equivalent
to athlon-xp or athlon-mp in current gcc.

-- 
| Dave Jones.        http://www.codemonkey.org.uk
| SuSE Labs

Re: [PATCH] [2.4] [2.5] [i386] Add support for GCC 3.1

[Christopher E. Brown]

On Sun, 26 May 2002 rwhron@earthlink.net wrote:

> Another processor config could be CONFIG_K62.
> gcc-3.1 -march=k6-2 benchmarks a little better
> than -march=k6.  Adding CONFIG_XF86_USE_3DNOW=y
> seems to help a little too.
>
> Based on grepping gcc-3.1 src, it appears:
> k6-3 == k6-2


IIRC (not looking at the datasheets) there were some pretty nice
improvements made to the core in the K6 to K6-2 update (not just and
upclock and downsize, real changes(isn't this when they added MMX
too?)).


However, the K6-3 is simply a K6-2 with the addition of a 256K L2
cache on die at full cpu clock.  It was a great improvement
performance wise for many uses (think 256K L2 @ 450Mhz, and then a 2M
L3 on the mainboard) though.


-- 
I route, therefore you are.

Re: [PATCH] [2.4] [2.5] [i386] Add support for GCC 3.1

[rwhron]

>> Adding CONFIG_XF86_USE_3DNOW=y
>> seems to help a little too.

Mmm, maybe not.

> Make sure you benchmark FPU intensive use when testing USE_3DNOW. I've
> seen several situations where it looks marginally better until you
> actually measure two things - memory bandwidth available to user
> programs, and also FPU performance, especially of FPU heavy apps that
> schedule a lot (eg Quake)

Thanks for helping me narrow the search.  On an allocate and write to 
pages of memory test, USE_3DNOW increased test time by 15% (which is
bad) on k6-2.  

Both kernels compiled with gcc-3.1.0 -march=k6-2.

mtest01 -w 50 executed 100 times 

CONFIG_X86_USE_3DNOW=n	745 seconds
CONFIG_X86_USE_3DNOW=y	860 seconds

-- 
Randy Hron

Re: [PATCH] [2.4] [2.5] [i386] Add support for GCC 3.1

[Dave Jones]

On Mon, May 27, 2002 at 05:44:52AM -0600, Christopher E. Brown wrote:
 > > k6-3 == k6-2
 > IIRC (not looking at the datasheets) there were some pretty nice
 > improvements made to the core in the K6 to K6-2 update (not just and
 > upclock and downsize, real changes(isn't this when they added MMX
 > too?)).

original k6 had mmx. k6-2 brought 3dnow.

-- 
| Dave Jones.        http://www.codemonkey.org.uk
| SuSE Labs