At 08:24 PM 1/21/2004, you wrote:
>Hi,
>
>On Thu, 22 Jan 2004, 10:14am +1100, [log in to unmask] wrote:
>
> > For our codes (numerical weather and climate models) the Intel
> > compiler's P4 specific optimizations have a large benefit.
> >
> > Time per simulated day on Intel P4 processor
> >
> > Options Time (s)
> > ifc -O2 319 Intel compiler
> > ifc -O2 -tpp6 -xK 299 P3 specific options
> > ifc -O2 -tpp7 -xW 213 P4 specific options
>
>(P3-P4)/P3 = 29% => impressive.
>
>
> > Another model
> >
> > ifc -O2 667
> > ifc -O2 -tpp6 -xK 504
> > ifc -O2 -tpp7 -xW 452
>
>(P3-P4)/P3 = 10% => uninteresting.
>
>Do you have a sense for what makes the first model so much
>more susceptible to performance tuning?
>
>-P.
>
>--
> Peter S. Shenkin Schrodinger, Inc.
>
The first group of timings would be typical of an application with
significant vectorizable double precision content. Likewise, the second
group likely has vectorizable single precision code.
You may have noticed that Intel is phasing out support for P-III optimized
code generation.
Tim Prince
|