tiled version with task groups

lu_tiled_taskgrp_i.cpp

/**** Tiled LU decomposition *****/
void lu(double **a, int range, int B)
{
    int i,j,k;
    double ** l_inv, ** u_inv;
    tbb::task_group g,f;
    
    for (k=0;k<range-1;k++) {

        /****Compute LU decomposition on upper left tile*****/
        lu_kernel(a,k*B,k*B,B,B);

        /****Compute inverted L and U matrices of upper left tile*****/
        l_inv=get_inv_l(a,k*B,k*B,B,B);
        u_inv=get_inv_u(a,k*B,k*B,B,B);

        /*****Compute LU decomposition on upper horizontal frame and left vertical frame*****/
        for (i=k+1;i<range;i++) {
            g.run([=]{
                mm_lower(l_inv,0,0,a,k*B,i*B,a,k*B,i*B,B,B,B);
                mm_upper(a,i*B,k*B,u_inv,0,0,a,i*B,k*B,B,B,B);});
        }
        g.wait();

        /*****Update trailing blocks*****/
        for (i=k+1;i<range;i++) {
            f.run([=,&j]{
                for (j=k+1;j<range;j++) {
                    mm_update(a,i*B,k*B,a,k*B,j*B,a,i*B,j*B,B,B,B);   
                }
            });
        } 
        f.wait();

    }

    /***** Compute LU on final diagonal block *****/
    lu_kernel(a,(range-1)*B,(range-1)*B,B,B);

}