From: "Tim Davis" <davis@cise.ufl.edu>
Path: news.mathworks.com!newsfeed-00.mathworks.com!webcrossing
Newsgroups: comp.soft-sys.matlab
Subject: Re: Mac Intel OpenGL error
Message-ID: <ef42ae2.10@webcrossing.raydaftYaTP>
Date: Sat, 10 Mar 2007 19:54:03 -0500
References: <ef42ae2.5@webcrossing.raydaftYaTP> <45F1DE0A.7060709@mathworks.com> <ef42ae2.7@webcrossing.raydaftYaTP> <45F1F4B2.1070000@mathworks.com> <ef42ae2.9@webcrossing.raydaftYaTP>
Lines: 36
NNTP-Posting-Host: 70.171.50.145
MIME-Version: 1.0
Content-Type: text/plain; charset="ISO-8859-1"
Content-Transfer-Encoding: 8bit
Xref: news.mathworks.com comp.soft-sys.matlab:396711



Brian Powell wrote:
>> We are investigating the relatively poor result with SPARSE in
>> bench on
>> Intel-based Mac (this was reported by 7a pre-release customers
> and
>> by
>> internal users).

Is this real? I assume you're not comparing the old bench results
with the new ones. If so, take a close look at the old bench. The
new bench has a much larger matrix and does perhaps 40 times the work
as the old one. This is because the new sparse Cholesky is many
times faster than the old one. See page 6 of my paper: <http://www.cise.ufl.edu/~davis/techreports/cholmod/tr06-005.pdf>
where more details are given.

To get a better look at the performance, do

n = 300 ;
A=delsq(numgrid(’L’,n)) ;
p = amd (A) ;
fl = sum (symbfact (A (p,p)).^2) ;
b = rand (size(A,1),1) ;
tic
x = A\b ;
t = toc
mflops = 1e-6 * fl / t

You can do this in any version of MATLAB. The results in 7a should
be better than MATLAB 7.1 (R200whatever, I can't recall).

If the performance degradation is real, then I would guess it's a
BLAS problem.

Thanks,
Tim Davis (author of sparse chol in MATLAB).