Fwd: uBLAS vs MTL for sparse matrices

From: Paul C. Leopardi (leopardi_at_[hidden])
Date: 2002-10-24 19:34:41

Hi,
I posted the following message to boost and have so far had no response. Maybe 
boost-users is the correct mailing list?

Hi,
Thanks to Joerg Walter's original conversion of GluCat ( http://glucat.sf.net 
) to uBLAS, I have fairly painlessly been able to get GluCat going with the 
version of uBLAS now promoted to Boost. I'm using sparse matrices:

boost::numeric::ublas::sparse_matrix< Scalar_T, 
boost::numeric::ublas::row_major>

and comparing performance to MTL:

mtl::matrix< Scalar_T, mtl::rectangle<>, mtl::compressed<>,
mtl::row_major >::type

The operation I'm performing is quite complicated. Basically, I am generating 
a large number of sparse matrices, and computing the inner product of a given 
matrix with each of these matrices in turn.
( If you have a copy of GluCat, see basis_element() in matrix_multi_imp.h )

Because these sparse matrices have only one non-zero per row, operations on an 
n x n matrix are typically O(n), and seem to be so using MTL:

Size vs speed for MTL:
Size Time (ms)
  32    10
  64    20
 128    80
 256   160
 512   700

But with uBLAS, essentially the same operations seem to be O(n^2):

Size Time (ms) Slowdown
  32    30      3
  64    70      3.5
 128   830     10.4
 256  1940     12.1
 512 31770     45.4

o Is there a more appropriate data type in uBLAS? 
o Do you have any tips on where to tweak uBLAS to get better (ie. O(n) )   
performance? 
o What is BOOST_UBLAS_USE_CANONICAL_ITERATOR and what effect should it have on 
performance?
Thanks