gmp-4.2 performance on x86_64

On my 2.2GHz x86_64-unknown-linux-gnu machine
I get the following gmpbench results:

  gmp-4.1.4-gaudry  7863.2
  gmp-4.2           7375.2
  gmp-4.1.4         4645.6

gmp-4.1.4-gaudry is using the assembly code by
Patrick Gaudry at

  http://www.lor
ia.fr/~gaudry/mpn_AMD64/

All were compiled with the same compiler, gcc-4.1.0.

While I want to use gmp-4.2 to pick up bug fixes, I do not
want to lose the speed of gmp-4.1.4-gaudry.

Doing the obvious to use Gaudry's assembler code in gmp-4.2
causes a failure to compile (for me).

Comments?  Suggestions?

Kate Minola
Engineering Department
University of Maryland, College Park
_______________________________________________
gmp-discuss mailing list
gmp-discussswox.com
http://s
wox.com/mailman/listinfo/gmp-discuss

On Wed, Mar 29, 2006 at 11:07:00AM -0500, Kate Minola wrote:
> On my 2.2GHz x86_64-unknown-linux-gnu machine
> I get the following gmpbench results:
> 
>   gmp-4.1.4-gaudry  7863.2
>   gmp-4.2           7375.2
>   gmp-4.1.4         4645.6
> 
> gmp-4.1.4-gaudry is using the assembly code by
> Patrick Gaudry at
> 
>   http://www.lor
ia.fr/~gaudry/mpn_AMD64/
> 
> All were compiled with the same compiler, gcc-4.1.0.
> 
> While I want to use gmp-4.2 to pick up bug fixes, I do
not
> want to lose the speed of gmp-4.1.4-gaudry.
> 
> Doing the obvious to use Gaudry's assembler code in
gmp-4.2
> causes a failure to compile (for me).
> 
> Comments?  Suggestions?
> 
> Kate Minola
> Engineering Department
> University of Maryland, College Park

A version of the patch for GMP-4.2 is available at:
  http://www.loria.fr/~gaudry/mpn_AMD64/mpn_amd64.42.tgz

The webpage is not yet updated, but the install procedure is
as for the
patch of 4.1.4:
  tar zxf mpn_amd64.42.tgz
  cd mpn_amd64.42
  ./install /path_to_gmp/gmp-4.2
  cd /path_to_gmp/gmp-4.2
  ./configure && make && make install
  
I get a gmpbench result around 10000 on a 2.4 GHz Opteron
(8200 with
plain 4.2)

Pierrick.
_______________________________________________
gmp-discuss mailing list
gmp-discussswox.com
http://s
wox.com/mailman/listinfo/gmp-discuss

Hi!

Does that include my code as well? I added a few modules for
some basic 
operations to a previous patch for x86_64 (this was a while
back, 
searching the mail-list archives should be useful). May be
somebody 
improved on them, may be somebody wrote a better (read:
faster) version. 
But they did improve the benchmarks for me (compared to the
original + 
Pierrick's patch).

Regards,
Ashod Nakashian

Pierrick Gaudry wrote:
> On Wed, Mar 29, 2006 at 11:07:00AM -0500, Kate Minola
wrote:
> 
>>On my 2.2GHz x86_64-unknown-linux-gnu machine
>>I get the following gmpbench results:
>>
>>  gmp-4.1.4-gaudry  7863.2
>>  gmp-4.2           7375.2
>>  gmp-4.1.4         4645.6
>>
>>gmp-4.1.4-gaudry is using the assembly code by
>>Patrick Gaudry at
>>
>>  http://www.lor
ia.fr/~gaudry/mpn_AMD64/
>>
>>All were compiled with the same compiler, gcc-4.1.0.
>>
>>While I want to use gmp-4.2 to pick up bug fixes, I
do not
>>want to lose the speed of gmp-4.1.4-gaudry.
>>
>>Doing the obvious to use Gaudry's assembler code in
gmp-4.2
>>causes a failure to compile (for me).
>>
>>Comments?  Suggestions?
>>
>>Kate Minola
>>Engineering Department
>>University of Maryland, College Park
> 
> 
> A version of the patch for GMP-4.2 is available at:
>   http://www.loria.fr/~gaudry/mpn_AMD64/mpn_amd64.42.tgz
> 
> The webpage is not yet updated, but the install
procedure is as for the
> patch of 4.1.4:
>   tar zxf mpn_amd64.42.tgz
>   cd mpn_amd64.42
>   ./install /path_to_gmp/gmp-4.2
>   cd /path_to_gmp/gmp-4.2
>   ./configure && make && make install
>   
> I get a gmpbench result around 10000 on a 2.4 GHz
Opteron (8200 with
> plain 4.2)
> 
> Pierrick.
> _______________________________________________
> gmp-discuss mailing list
> gmp-discussswox.com
> http://s
wox.com/mailman/listinfo/gmp-discuss
> 
> 

_______________________________________________
gmp-discuss mailing list
gmp-discussswox.com
http://s
wox.com/mailman/listinfo/gmp-discuss

On Thu, Mar 30, 2006 at 12:37:27AM +0500, Ashod Nakashian
wrote:
> Hi!
> 
> Does that include my code as well? I added a few
modules for some basic 
> operations to a previous patch for x86_64 (this was a
while back, 
> searching the mail-list archives should be useful). May
be somebody 
> improved on them, may be somebody wrote a better (read:
faster) version. 
> But they did improve the benchmarks for me (compared to
the original + 
> Pierrick's patch).
> 
> Regards,
> Ashod Nakashian

I have compared your mul_1 and the one which is shipped with
4.2, and the
latter is faster, so I didn't include yours. But if I
remember well you
had also coded the popcount or some function of this sort. I
should
include it. 
I am also considering the inclusion of divrem_[12].s
functions written by
Timothy Jacobson, but for the moment I still have a few
problems to make
them work properly on my system. Once these are solved, this
should give a
nice speed-up to gmpbench, I guess.

All the best,
Pierrick.

_______________________________________________
gmp-discuss mailing list
gmp-discussswox.com
http://s
wox.com/mailman/listinfo/gmp-discuss