Re: [eigen] CDASH-Speedup: "Homer" how did you do it?

[FMDSPAM <fmdspam@xxxxxxxxx>]

Hauke Heibel schrieb:
> On Sat, Aug 29, 2009 at 3:37 PM, FMDSPAM<fmdspam@xxxxxxxxx> wrote:
>   
> > Is it possible to change the compiler settings at the outermost level, the
> > calling batch?
> >     
>
> Probably, but see my comments below.
>
>   
> > All my experiments (even with EI_OFLAGS=...) fail.
> >     
>
> What do you mean? Are you saying that since my last commit all your
> experiments are failing?
>
>   
No, 2 or 3 experimtes several weeks ago.
> > A preferable result could be something in the (batch-)lines of e.g.:
> >
> > set CXX_VERSION=msvc-2008(9.0.30729.1)
> > set CXX_FLAGS=/Ox
> > set BUILD_STRING=%OS_VERSION%-%ARCH%-%CXX_VERSION%-%CXX_FLAGS%
> >
> >
> > what do you think?
> >     
>
> I am not sure whether this is necessary. O2 should definitely suffice.
>   
Hmm..., probably. But in theory maybe other flags...
> Btw, my test could also be so fast because quite a few are actually
> failing. :)
>   
That is not the case. E.g.:
product_large.exe: yours   5.25s, my   52.73s
product_trmm.exe: yours 24.76s, my 163.86s

By the way:  The timings could vary simply caused by the random numbers, 
right? Or is there a fixed random seed?
What is the expected bandwidth of variation by this?

Frank