[petsc-dev] MATCOMPOSITE as fallback for matrix-matrix multiplication

[Matthew Knepley]

On Mon, Mar 11, 2019 at 6:59 PM Jed Brown via petsc-dev <
petsc-dev at mcs.anl.gov> wrote:

> One could imagine PETSc written so that all Mat operations like
> transpose and products return lazy representations, with a method that
> makes it explicit.  Would eliminate all the Transpose variants in the
> matrix products, for example, and enable an implementation to choose
> order of operations for chained products.
>

Didn't Liz Jessup (and others) do a BLAS subset this way in order to fuse
operations? Did these
not catch on because it is a maintenance burden, or just that you don't get
a big win?

   Matt


> "Smith, Barry F. via petsc-dev" <petsc-dev at mcs.anl.gov> writes:
>
> >> On Mar 11, 2019, at 1:28 PM, Isaac, Tobin G <tisaac at cc.gatech.edu>
> wrote:
> >>
> >>
> >> Okay, but do you have an opinion on the original issue: should we
> >> always return a matrix from various matrix-matrix routines, even if it
> >> has limited functionality?
> >
> >    I'm fine with it always returning the composite matrix unless you
> have a specific use case that would require MAT_INITIAL_MATRIX_GET_VALUES
> >
> >    Barry
> >
> >>
> >> On Mon, Mar 11, 2019 at 06:09:55PM +0000, Smith, Barry F. wrote:
> >>>
> >>>   Sounds like overkill on the complexity meter.
> >>>
> >>>   Barry
> >>>
> >>>
> >>>> On Mar 11, 2019, at 6:15 AM, Isaac, Tobin G via petsc-dev <
> petsc-dev at mcs.anl.gov> wrote:
> >>>>
> >>>> One thought: we could extend MatReuse with
> MAT_INITIAL_MATRIX_GET_VALUES to indicate that a matrix that implements
> MatGetValues() is desired, to allow some user intent.
> >>>>
> >>>>
> >>>> On March 9, 2019 7:44:06 PM EST, Matthew Knepley <knepley at gmail.com>
> wrote:
> >>>>> On Sat, Mar 9, 2019 at 4:42 PM Isaac, Tobin G via petsc-dev <
> >>>>> petsc-dev at mcs.anl.gov> wrote:
> >>>>>
> >>>>>>
> >>>>>> Let's say I want to do Newton-Krylov on some DAG of computations.
> >>>>>> Building the Jacobian is programmatic from the chain rule, requiring
> >>>>>> just a lot of matrix-matrix multiplications in the correct sequence.
> >>>>>>
> >>>>>> If it really is more efficient for two Jacobians to be contracted
> >>>>> into
> >>>>>> a matrix with random access, that's great, but if it's not, or if
> >>>>>> the appropriate implementation of MatMatMult_X_Y does not exist,
> >>>>>> having a MATCOMPOSITE is better than nothing.
> >>>>>>
> >>>>>> So why not make MATCOMPOSITE something that our matrix-matrix
> >>>>>> multiplication routines return when a specialization isn't found?
> >>>>>> It moves the "Not Implemented" warning from MatMatMult() to
> >>>>>> whenever I try to do something other than multiply with the output,
> >>>>>> but in some situations that's desirable.
> >>>>>>
> >>>>>> I'd be interested to hear what people think.
> >>>>>>
> >>>>>
> >>>>> I think this is a good idea. The most important thing is to maintain
> >>>>> transparency, so that the output
> >>>>> makes this clear in -ksp_view and we are easily able to see what
> >>>>> someone
> >>>>> has produced.
> >>>>>
> >>>>> Matt
> >>>>>
> >>>>>
> >>>>>> Thanks,
> >>>>>> Toby
> >>>>>>
> >>>>>>
> >>>>>>
> >>>
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener

https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20190311/0b91da29/attachment-0001.html>