[Trilinos-Users] Trilinos-Users Digest, Vol 60, Issue 1

[Bartlett, Roscoe A]

Expression templates will not work when the data is on a GPU.

My stop-gap measure for linear algebra operations is given here:

    http://trilinos.sandia.gov/packages/docs/dev/packages/thyra/doc/html/LinearAlgebraFunctionConvention.pdf

with examples at:

    http://trilinos.sandia.gov/packages/docs/dev/packages/thyra/doc/html/classThyra_1_1VectorBase.html

Therefore, for:

   y = alpha * x1 + beta * x2

you would have:

  V_StVpStV( y, alpha, x1, beta, x2 );

This is not the most beautiful thing you have ever seen but at least you can go back and forth between the linear algebra notation and the function declaration and call sequence.  You can also make this run very fast without fancy expression templates on the GPU in a fairly transparent way (at least transparent to the user).

Cheers,

- Ross
   

> -----Original Message-----
> From: trilinos-users-bounces at software.sandia.gov [mailto:trilinos-
> users-bounces at software.sandia.gov] On Behalf Of Mark Hoemmen
> Sent: Monday, August 02, 2010 12:25 PM
> To: trilinos-users at software.sandia.gov
> Subject: Re: [Trilinos-Users] Trilinos-Users Digest, Vol 60, Issue 1
> 
> On Aug 1, 2010, at 10:08 PM, trilinos-users-request at software.sandia.gov
> wrote:
> > Message: 4
> > Date: Mon, 02 Aug 2010 02:46:31 +0200
> > From: Nico Schl?mer <nico.schloemer at ua.ac.be>
> > Subject: Re: [Trilinos-Users] Epetra_SerialDenseVector vs.
> >        Teuchos::SerialDenseVector
> > To: "Bartlett, Roscoe A" <rabartl at sandia.gov>
> > Cc: "Heroux, Michael A" <maherou at sandia.gov>, "Thornquist,      Heidi
> K"
> >        <hkthorn at sandia.gov>, trilinos-users at software.sandia.gov
> > Message-ID: <5a5040fba02ffca0981ca1e5c125a8d0 at ua.ac.be>
> > Content-Type: text/plain; charset=utf-8
> >
> >> It is very hard to make operator+, operator-, operator* efficient.
> >
> > I see.
> > Would you have any recommendation for me what to do with expression
> like
> >
> >   y = alpha * x1 + beta * x2 + ...
> >
> > ?
> 
> "Efficient" here means "not creating temporary vectors."  Each of those
> "alpha*x1" terms will have to create a new temporary vector, and each
> operator+ invocation will have to do so as well.  The temporaries will
> go away, but you'll still have to spend time in memory allocation.
> 
> If you're not worried about that, then you could define non-member
> operator+, operator*, ... methods.  They don't need to belong to or be
> friends of Teuchos::SerialDenseVector, because SerialDenseVector has an
> operator[].  Each of these methods should create and return a new
> SerialDenseVector.
> 
> If you are worried about the cost of creating temporary vectors in such
> expressions, but still want to write expressions in this way, you may
> wish to investigate the uBLAS, which is in the Boost collection of C++
> libraries.  See e.g., the overview at
> 
> http://www.boost.org/doc/libs/1_43_0/libs/numeric/ublas/doc/overview.ht
> m
> 
> mfh
> 
> 
> 
> 
> _______________________________________________
> Trilinos-Users mailing list
> Trilinos-Users at software.sandia.gov
> http://software.sandia.gov/mailman/listinfo/trilinos-users