Hi - we always see some numerical differences between different
platforms even sticking within gcc so some variation doesn't surprise
me. The error tolerances are set in FEEDS according to the typical
size of these differences so if the intel-compiler build is slightly
outside these tolerances that doesn't surprise me too much. Hopefully
the results won't be _very_ different though.
Cheers.
On 3 Sep 2007, at 06:12, Neil Killeen wrote:
> Hi
>
> I am having some further Intel problems. I succeeded in getting the
> entire system to build.
>
> 1) However,to do so, I built the Extras packages with gnu (I had
> too many
> problems with these third-party Makefiles) and the rest with
> Intel. My
> understanding is that the Intel compiler is GNU compatible.
> However, I would be interested in whether any problems
> might be expected in doing this or whether others have tried
> mixing compilers like this. I don't know what the dependencies
> are on the Extras package so I don't have any insight into
> what I might expect.
>
>
>
> 2) When I ran the feeds test suite I had two failures.
>
> 1. mcflirt failed with a segv
> 2. melodic failed with a glibc run-time error
>
> Now, regarding mcflirt alone at this point, I backed the optimization
> level off from O3 to O0 and rebuild just the mcflirt package.
> It then ran ok. However, I notice that the results are different
> from the
> GNU run. For example. Following the mcflirt step in the feat test
> the GNU test suite delivered:
>
>> cd results/fmri.feat/mc
>> more *rms
> ::::::::::::::
> prefiltered_func_data_mcf_abs_mean.rms
> ::::::::::::::
> 0.139814
> ::::::::::::::
> prefiltered_func_data_mcf_abs.rms
> ::::::::::::::
> 0.192819
> 0.191941
> 0.175656
>
>
> whereas the Intel delivered:
>
> prefiltered_func_data_mcf_abs_mean.rms
> ::::::::::::::
> 0.139704
> ::::::::::::::
> prefiltered_func_data_mcf_abs.rms
> ::::::::::::::
> 0.195441
> 0.194181
> 0.185148
>
> Other results (e.g. prefiltered_func_data_mcf.par) differ too.
>
>
>
> So my question is whether I should expect differences from run to
> run and
> between compilers, or whether it should all repeat precisely (round
> off aside). I am assuming that the results should repeat since they
> involve the same data.
>
>
> regards
> Neil
------------------------------------------------------------------------
---
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
------------------------------------------------------------------------
---
|