
Ublas : 
From: Gunter Winkler (guwi17_at_[hidden])
Date: 20070303 07:28:32
Sourabh schrieb:
> Hi Gunter,
> Thanks for your kind reply.
> The structure of my matrices are as follows :
>
> X is a lower triangular matrix of size (100 x 100) (in addtion, all
> diagonal elements are also
> zero), with at most 3 or 4 nonzero entries per column.
>
So, I suggest using a compressed_matrix which will hold (lembda)
> What is your opinion, is construction of these matrices take much of the
> time ?
>
construction is usually expensive, because you have to allocate storage
and (sometimes) initialize some values. Thus I would move all
constructions out of the loop. A call of the clear() member function is
very cheap (only a few integer assigements). When you fill the matrix in
order (row by row for a row_major matrix) using push_back(i,j,value) you
gain optimal speed.
> I need to solve the equation (Ilembda) slopes = omegas
>
then you should initialize the matrix with (lemda) and use the unit
lower triangular solver ("unit" means the diagonal elements are treated
as they were 1.0)
> Same question as above. The delays vector creation or assignment to delays
> in each loop?
>
the creation need a call to new() which is always expensive. The
assignement has to occure anyway. So construct the vector before. (You
additionally improve the locality because the memory of this vector is
already inside the cache).
> Thank you very much for your help.
You could write a simple program with similar data and post it here. So
we can add it to the examples and benchmark collection.
mfg
Gunter