[Pw_forum] compiling for Intel Xeon Phi

[Fabio Affinito]

Hi,
 
> It wasn't that simple (since I stumbled upon numerous complains about
> incompatible format), so I had to manually set up all compilers to
> use intel compilers. But OK.

This is because everything, i.e. code+libraries, must be cross-compiled with the -mmic flag.

> Another concern is if pw.x can benefit from setup. Specifically
> - can QE perform dynamic load balancing and utilize heterogeneous
> nodes efficiently?
> - can QU utilize phi cores efficiently (i.e. if pure native execution
> would be more efficient than offload or pure CPU modes)?

In my opinion, it's pretty unlikely that native mode can outperform the execution on the standard Xeon cpu. I strongly suggest to use the Xeon Phi in offload mode, for now.

Fabio


> 
> 
> On 24 March 2015 at 18:19, Fabio Affinito < f.affinito at cineca.it >
> wrote:
> 
> 
> Hi,
> 
> You can take advantage of the offload mode using the libxphi library.
> This library offloads the BLAS/MKL functions on the Xeon Phi
> platform hiding the latency times due to the communication. You just
> need to compile this library and then to link it dynamically. The
> library works with any version of QE.
> Libxphi library is available from https://github.com/cdahnken/libxphi
> Some documentation is available therein.
> 
> Instead, if you want to compile a native version of pw, you just need
> to add the -mmic flag and cross compile.
> If you want to use the symmetric mode, you need to compile twice:
> with and without the -mmic flag.
> 
> Best,
> 
> Fabio
> 
> 
> 
> ----- Messaggio originale -----
> > Da: "Евгений Пермяков" < permeakra at gmail.com >
> > A: "PWSCF Forum" < Pw_forum at pwscf.org >
> > Inviato: Martedì, 24 marzo 2015 15:54:58
> > Oggetto: [Pw_forum] compiling for Intel Xeon Phi
> > 
> > 
> > 
> > I have access to cluster with Xeon Phi coprocessor on nodes. (it
> > also
> > has Scalapack and mkl) The documentaion mentions several way to
> > compile programs:
> > 1) offload mode, executed on main CPU and offloaded onto
> > coprocessor
> > "automagically" (compiled with "mpicc" (I assume, in case of
> > fortran
> > applications it is "mpif90")
> > 2) native mode, executed completely on coprocessor, compiled with
> > "mpicc -mmic" ("mpif90 -mmic"
> > 
> > 3) symmetric mode, requiring creation of both binaries
> > 
> > 
> > Questions: does pw.x benefit from any of the modes? How should I
> > create 'native' binaries ? Is it possible to produce both "mmic"
> > and
> > "normal" binaries in same tree or I have to compile to source trees
> > separately ?
> > _______________________________________________
> > Pw_forum mailing list
> > Pw_forum at pwscf.org
> > http://pwscf.org/mailman/listinfo/pw_forum
> 
> --
> "When you are solving a problem, don’t worry. Now, after you have
> solved the problem, then that’s the time to worry." Richard Feynman
> 
> Fabio Affinito, PhD
> SuperComputing Applications and Innovation Department
> CINECA - via Magnanelli, 6/3, 40033 Casalecchio di Reno (Bologna) -
> ITALY
> Tel: +39 051 6171794 Fax: +39 051 6132198
> 
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://pwscf.org/mailman/listinfo/pw_forum
> 
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://pwscf.org/mailman/listinfo/pw_forum

-- 
"When you are solving a problem, don’t worry. Now, after you have solved the problem, then that’s the time to worry." Richard Feynman

Fabio Affinito, PhD
SuperComputing Applications and Innovation Department
CINECA - via Magnanelli, 6/3, 40033 Casalecchio di Reno (Bologna) - ITALY
Tel: +39 051 6171794  Fax: +39 051 6132198