Is there a platform optimized version of gmp for INTEL?

[Michael Abshoff]

Patrick Pelissier wrote:

Hello,

>> I am wondering if there is a platform optimized version of gmp for the
>> Intel
>> architecture? I am using MS Visual Studio 2005. Has anybody tried to
>> compile
>> GMP using Intel compiler?
>  Yes

What exactly do you use as a target? x86, EMT64 or Itanic? In the unlikely
event that you are running your code on Itanium you might want to check
out the gmp-like implementation provided by Intel in the Intel Math Kernel
library. I am not aware of any performace numbers (and I did at least
search :) and I cannot tell you if it is 100% compatible, either.

If anybody has any experience with the gmp implementation in the MKL I
would be interested in the results.

>
>> Will the performance improve a lot than using MS
>> Visual Studio 2005?
>  No
>

It is all assembly anyway, so the compiler has little impact unless you
compile the generic C version. Obviously that code will be slower than the
assembly code.

>> I desparately need to reduce the run time of my program.

So what is the problem exactly? Chances are that you need algorithmic
improvements since "buying a bigger box" only gets you so far unless it is
something trivial like your working set not fitting into the RAM, etc.

>> Are there faster implementation of the GMP library or multiple precision
>> library?
>  No
>

Check out the "More Digits friendly Competition 2006" at
http://rnc7.loria.fr/comp_result.html and you will see that the mpfr
(which uses the gmp) looks very good. The Mathematica people "won" because
they solved more of the problems, not because they were faster compared to
the mpfr.

Cheers,

Michael