Re: [AMBER-Developers] Third release candidate for AmberTools 1.5

From: Ben Roberts <ben.roberts.geek.nz>
Date: Thu, 31 Mar 2011 15:01:08 -0400

On 31/3/2011, at 9:48 a.m., David A Case wrote:

> On Tue, Mar 29, 2011, Ben Roberts wrote:
>>>> nab/asp.out
>>>
>> Things pass with no MKL.
>
> OK...thanks for the update. I don't know quite what to do here. Hopefully,
> people who try the Mac/Intel/MKL combination will see that the tests are not
> passing. I'm hesitant to refuse to accept this combo at the configure stage,
> since the results *might* reflect a particular combination of versions, and
> that other versions would work. But comments would be appreciated.
>
> One idea, if you are willing to try one more test: put MKL back in, but export
> OMP_NUM_THREADS=1 in your environment. That will force MKL to use just a
> single thread.

Further updates:

* Intel with no MKL and macAccelerate passes those tests
* GNU compilers (MacPorts 4.4) with MKL and no macAccelerate do not pass (and are not the same numbers as with Intel compilers)
* Intel with MKL and OMP_NUM_THREADS=1 does not pass

Based on which, I think it's safe to say that MKL in particular does...interesting things. I don't think it's compiler-version specific, given that changing the compiler entirely (i.e., using GNU) produces similarly bad output. It may be MKL-version specific, though.

Shall we recommend that people not use the Intel MKL on the Mac for AT 1.5? I can put something up at the Amber On Mac blog to that effect.

Ben

-- 
For greater security, I support S/MIME encryption.





_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers


Received on Thu Mar 31 2011 - 12:30:03 PDT
Custom Search