Hi Jonathan,

Thanks very much for the benchmarks and detailed comments.  Your  
conclusions also ring true for me, especially that matrix  
construction is currently slow and that we need better documentation.

I have only one question for now.  Could you please explain the comment

 >  dok_matrix.setdiag() appears to change the matrix shape if you're not
 >  careful, rendering it pretty useless for building the matrices for  
this
 >  problem

in more detail?  The code for setdiag() is:

     def setdiag(self, values, k=0):
         N = len(values)
         for n in range(N):
             self[n, n+k] = values[n]
         return

which grows the matrix only if it's currently smaller than len 
(values) x len(values).  What's the problem this behaviour caused?   
Would you prefer to fix the size of a dok_matrix in advance and have  
a run-time check on the len(values)?  Or was the problem with the  
second argument?

-- Ed



On 22/11/2005, at 8:05 PM, Jonathan Guyer wrote:

> 
>  On Oct 7, 2005, at 4:00 PM, I wrote:
> 
> >  regardless, pysparse is BSD
> > licensed, so it's perfectly legal to use his code to improve
> > scipy.sparse (assuming that rigorous benchmarking determines that  
> > there
> > are, in fact, improvements to be made). We'll do some tests and, if a
> > merge is warranted, we'll run it by Roman out of courtesy.
> 
>  I've finally done some benchmarking of scipy.sparse and PySparse and
>  posted my results and comments on plone[*]:
> 
>       <http://www.scipy.org/wikis/featurerequests/SparseSolvers>
> 
>  Bottom line is that, with a couple of exceptions, PySparse is both
>  faster and less memory intensive than SciPy. I don't know whether
>  anything can be lifted from PySparse to improve SciPy's  
>  implementation.
> 
>  [*] Not that this comes as a surprise to anybody here, but the SciPy
>  plone wiki is scary slow. I also couldn't get it to accept reST, even
>  though I've seen other people using it there.
>  -- 
>  Jonathan E. Guyer, PhD
>  Metallurgy Division
>  National Institute of Standards and Technology
>  <http://www.metallurgy.nist.gov/>
> 
>  _______________________________________________
>  Scipy-dev mailing list
>  Scipy-dev@[...].net
>  http://www.scipy.net/mailman/listinfo/scipy-dev