Re: [ublas] Solving linear equations in a loop (Optimization issues)

From: Sourabh (sourabh_at_[hidden])
Date: 2007-03-03 08:39:58

On Sat, 3 Mar 2007, Gunter Winkler wrote:

> Sourabh schrieb:
> > On Fri, 2 Mar 2007, Gunter Winkler wrote:
> >   
> > The program (see below), after some modifications is still taking around 
> > 8 minutes for the arguments 1 200000 100. 
> >
> > Even if I remove solve, and addition methods, the statement
> >  lembda1.assign (ident-lembda) is causing the program to take 1 minute.
> >   
> 1) You can avoid the assign if you fill lembda with the negative values, 
> such that you have to solve with (I+lemda).

Do you mean that I fill lembda with negative values and use this 
statement:

solve (lembda, omega, unit_lower_tag ()) ?

will it solve the equation treating lemda as (I+lembda)
Will it be equivalent to the following statement ?

lembda = I - lembda
solve (lembda, omega, lower_tag ())

> 2) you should try the current CVS head of boost.
> 3) use inplace_solve_lower<true>(A, x, row_major_tag()); from my examples.

Please see the below program.

int main ( int argc, char* argv[])
{
    int iterations = atoi (argv[1]);
    int gatesperpath = atoi (argv[2]);
    boost::timer t;
    t.restart();
    unsigned int i = 0;
    double j;
    identity_matrix<double> ident (gatesperpath);
    compressed_matrix<double> 	lembda1 (gatesperpath, gatesperpath);
    lembda1 (3, 1) = -0.1;
    vector <double> omegas(gatesperpath);
    vector<double> slopes;
    
    while (iterations--) {
        slopes = solve (lembda1, omegas, lower_tag ());	
    
    }
    std::cout << t.elapsed () << std::endl;
    return 0;
}

When I run this with arguments 200000 100 , the program takes 1.5 seconds 
on my machine. If I change solve to 
inplace_solve_lower<true>(lembda1, omegas, row_major_tag());
it takes 35 seconds. 

What is wrong in my way of using ? I think I not getting your point 
properly.