How to retrieve final_res and numiter?

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)

How to retrieve final_res and numiter?

Postby Klausb » Tue Oct 10, 2017 4:20 pm

Hello,

I need the final residual value and number of iterations needed at the end of the solver run.

I read that "final_res" and "numiter" are the related variables or functions, I assume from a monitoring class.

How can I retrieve these values?

Klaus
Klausb
 
Posts: 18
Joined: Thu Oct 05, 2017 3:04 pm

Retrieving final_res and numiter - I always get the same res

Postby Klausb » Sat Oct 14, 2017 9:02 am

In the meantime, I tried to figure it out based on the example related to struct magma_z_solver_par which mentions the variables final_res and numiter.

I added them to the magma example that comes with the install source code called "example_sparse":

Code: Select all
        printf("%i\n", opts.solver_par.numiter);
        printf("%.4f\n", opts.solver_par.final_res);


But however I modify "example_sparse" I get, numiter 1, and final_res 0.

What's wrong?

Klaus
Klausb
 
Posts: 18
Joined: Thu Oct 05, 2017 3:04 pm

Re: How to retrieve final_res and numiter?

Postby hartwig anzt » Sat Oct 14, 2017 11:40 am

Klaus,
please post the complete code. From that part, I can not deduce what is happening.
Thanks, Hartwig
hartwig anzt
 
Posts: 79
Joined: Tue Sep 02, 2014 5:44 pm

extended example_sparse

Postby Klausb » Sun Oct 15, 2017 7:14 am

Hello Hartwig,

there's the code:

Code: Select all
// This is a simple standalone example. See README.txt

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>

#include "magma_v2.h"
#include "magmasparse.h"


// ------------------------------------------------------------
// This is an example how magma can be integrated into another software.
int main( int argc, char** argv )
{
    // The software does e.g. discretization of a PDE,
    // ends up with a sparse linear system in CSR format and a RHS.
    // Let's assume this system is a diagonal system of size m.
   
    int i, m=1333, n=1;
    int *col, *row;
    double *val, *rhs, *sol;
   
    row = (int*) calloc(m+1, sizeof(int));
    col = (int*) calloc(m,   sizeof(int));
   
    val = (double*) calloc(m, sizeof(double));
    rhs = (double*) calloc(m, sizeof(double));
    sol = (double*) calloc(m, sizeof(double));
   
    for (i = 0; i < m; ++i) {
        col[i] = i;
        row[i] = i;
        val[i] = 55.0;
        rhs[i] = 3.0;
        sol[i] = 0.3;
    }
    row[m] = m;
   
    // Initialize MAGMA and create some LA structures.
    magma_init();
    magma_dopts opts;
    magma_queue_t queue;
    magma_queue_create( 0, &queue );
   
    magma_d_matrix A={Magma_CSR}, dA={Magma_CSR};
    magma_d_matrix b={Magma_CSR}, db={Magma_CSR};
    magma_d_matrix x={Magma_CSR}, dx={Magma_CSR};
   
    // Pass the system to MAGMA.
    magma_dcsrset( m, m, row, col, val, &A, queue );
    magma_dvset( m, 1, rhs, &b, queue );
    magma_dvset( m, 1, sol, &x, queue );

   
    // Choose a solver, preconditioner, etc. - see documentation for options.
    opts.solver_par.solver     = Magma_PIDRMERGE;
    opts.solver_par.restart    = 8;
    opts.solver_par.maxiter    = 1000;
    opts.solver_par.rtol       = 1e-10;
    opts.solver_par.maxiter    = 1000;
    opts.precond_par.solver    = Magma_ILU;
    opts.precond_par.levels    = 0;
    opts.precond_par.trisolver = Magma_CUSOLVE;
   
    // Initialize the solver.
    magma_dsolverinfo_init( &opts.solver_par, &opts.precond_par, queue );
   
    // Copy the system to the device (optional, only necessary if using the GPU)
    magma_dmtransfer( A, &dA, Magma_CPU, Magma_DEV, queue );
    magma_dmtransfer( b, &db, Magma_CPU, Magma_DEV, queue );
    magma_dmtransfer( x, &dx, Magma_CPU, Magma_DEV, queue );
   
    // Generate the preconditioner.
    magma_d_precondsetup( dA, db, &opts.solver_par, &opts.precond_par, queue );
   
    // In case we only wanted to generate a preconditioner, we are done.
    // The preconditioner in the opts.precond_par structure - in this case an ILU.
    // The lower ILU(0) factor is in opts.precond_par.L and
    // the upper ILU(0) factor is in opts.precond_par.U (in this case on the device).
   
    // If we want to solve the problem, run:
    magma_d_solver( dA, db, &dx, &opts, queue );

    // Then copy the solution back to the host...
    magma_dmfree( &x, queue );
    magma_dmtransfer( dx, &x, Magma_CPU, Magma_DEV, queue );
   
    // and back to the application code
    magma_dvget( x, &m, &n, &sol, queue );
   
    // CHECK SYNTAX TO RETRIEVE NUMBER OF ITERATIONS NEEDED AND THE FINAL RESIDUAL, NEEDED BY THE APPLICATION
    printf("%i\n", opts.solver_par.numiter);
    printf("%.4f\n\n", opts.solver_par.final_res);
   
    // Free the allocated memory...
    magma_dmfree( &dx, queue );
    magma_dmfree( &db, queue );
    magma_dmfree( &dA, queue );
   
    // and finalize MAGMA.
    magma_queue_destroy( queue );
    magma_finalize();
   
    // From here on, the application code may continue with the solution in sol...
    for (i = 0; i < 20; ++i) {
        printf("%.4f\n", sol[i]);
    }
   
    return 0;
}


Klaus
Klausb
 
Posts: 18
Joined: Thu Oct 05, 2017 3:04 pm

Re: How to retrieve final_res and numiter?

Postby hartwig anzt » Mon Oct 16, 2017 4:36 am

Klaus,
this is a diagonal matrix. You have the solution after one iteration.
Hartwig
hartwig anzt
 
Posts: 79
Joined: Tue Sep 02, 2014 5:44 pm


Return to User discussion

Who is online

Users browsing this forum: Bing [Bot] and 4 guests