Path: utzoo!mnetor!uunet!yale!lisper-bjorn
From: lisper-bjorn@CS.YALE.EDU (Bjorn Lisper)
Newsgroups: comp.arch
Subject: Re: FORTRAN Horror
Message-ID: <25462@yale-celray.yale.UUCP>
Date: 21 Mar 88 03:17:00 GMT
References: <1135@pembina.UUCP> <28200126@ccvaxa>
Sender: root@yale.UUCP
Reply-To: lisper-bjorn@CS.YALE.EDU (Bjorn Lisper)
Organization: Yale University Computer Science Dept, New Haven CT  06520-2158
Lines: 22

In article <28200126@ccvaxa> aglew@ccvaxa.UUCP writes:
>
>..> Chris Shaw on SPARSPAK
....
>(2) Do you have any ideas as to what a (vector) processor
>    can do to make sparse matrix codes run faster?

If the sparse matrix code is such that the vector instructions can be
utilized, yes. Multifrontal methods, for instance, allows this to some
degree. They are direct methods that work by breaking down the total, sparse
system to a tree of dense subsystems where each can be solved with the use
of vector instructions. When all sons to a node are solved the solutions of
the sons are combined with the equation system of the node itself which in
turn is solved, until all equations are solved.

Another possibility is if the machine has instructions for gathering and
scattering vector elements, then these can be used to increase the possible
vectorization. A paper on this is Lewis, Simon: "The Impact of Hardware
Gather/Scatter on Sparse Gaussian Elimination", SIAM J. Stat. Comput., vol.
9, no. 2, March 1988.

Bjorn Lisper