LIS_INT lis_debug_trace_func(LIS_INT flag, char *func)
{
static int lis_tr_func = 0;
char buf[1024];
LIS_INT argc=0;
#ifdef USE_MPI
MPI_Initialized(&lis_mpi_initialized);
if (!lis_mpi_initialized) MPI_Init(&argc, NULL);
#endif
if( flag )
{
sprintf(buf, "%%%ds : %%s\n",lis_tr_func+3);
lis_printf(lis_debug_comm, buf, "IN ", func);
lis_tr_func++;
}
else
{
lis_tr_func--;
sprintf(buf, "%%%ds : %%s\n",lis_tr_func+3);
lis_printf(lis_debug_comm, buf, "OUT", func);
}
#ifdef USE_MPI
if( strcmp(func,"main")==0 && !flag )
{
MPI_Finalize();
}
#endif
return LIS_SUCCESS;
}
void lis_version(void)
{
LIS_DEBUG_FUNC_IN;
lis_printf(LIS_COMM_WORLD,"Lis Version %s\n",LIS_VERSION);
LIS_DEBUG_FUNC_OUT;
}
LIS_INT lis_error(const char *file, const char *func, const LIS_INT line, const LIS_INT code, const char *mess, ...)
{
va_list vvlist;
va_start( vvlist, mess );
lis_printf(lis_debug_comm,"%s(%d) : %s : error %s :",file,line,func,LIS_ERR_MESS[code-LIS_ERR_ILL_ARG]);
if( mess ) vprintf(mess,vvlist);
va_end(vvlist);
return LIS_SUCCESS;
}
LIS_INT lis_solve_kernel(LIS_MATRIX A, LIS_VECTOR b, LIS_VECTOR x, LIS_SOLVER solver, LIS_PRECON precon)
{
LIS_INT nsolver, precon_type, maxiter;
LIS_INT err;
LIS_SCALAR *residual;
LIS_VECTOR xx;
LIS_INT output;
LIS_INT scale;
LIS_INT conv_cond;
LIS_INT precision,is_use_at,storage,block;
LIS_INT i,n,np;
double p_c_times, p_i_times,itimes;
LIS_SCALAR nrm2,tol,tol_w;
LIS_VECTOR t;
LIS_VECTOR bb;
LIS_MATRIX AA,B;
LIS_MATRIX At;
char buf[64];
LIS_DEBUG_FUNC_IN;
nsolver = solver->options[LIS_OPTIONS_SOLVER];
precon_type = solver->options[LIS_OPTIONS_PRECON];
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
scale = solver->options[LIS_OPTIONS_SCALE];
precision = solver->options[LIS_OPTIONS_PRECISION];
is_use_at = solver->options[LIS_OPTIONS_USE_AT];
storage = solver->options[LIS_OPTIONS_STORAGE];
block = solver->options[LIS_OPTIONS_STORAGE_BLOCK];
conv_cond = solver->options[LIS_OPTIONS_CONV_COND];
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
tol_w = solver->params[LIS_PARAMS_RESID_WEIGHT-LIS_OPTIONS_LEN];
solver->precision = precision;
if( nsolver < 1 || nsolver > LIS_SOLVERS_LEN )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_OPTIONS_SOLVER is %d (Set between 1 to %d)\n",nsolver, LIS_SOLVERS_LEN);
return LIS_ERR_ILL_ARG;
}
if( precon_type < 0 || precon_type > precon_register_type )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_OPTIONS_PRECON is %d (Set between 0 to %d)\n",precon_type, precon_register_type-1);
return LIS_ERR_ILL_ARG;
}
if( maxiter<0 )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_OPTIONS_MAXITER(=%d) is less than 0\n",maxiter);
return LIS_ERR_ILL_ARG;
}
#ifdef USE_MPI
if( precon_type == LIS_PRECON_TYPE_SAAMG && solver->A->nprocs < 2)
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter A->nprocs (=%d) is less than 2 (Set more than 1 when using parallel version of SAAMG)\n",solver->A->nprocs);
return LIS_ERR_ILL_ARG;
}
#endif
#ifdef USE_QUAD_PRECISION
if( precision==LIS_PRECISION_QUAD && lis_solver_execute_quad[nsolver]==NULL )
{
LIS_SETERR1(LIS_ERR_NOT_IMPLEMENTED,"Quad precision solver %s is not implemented\n",lis_solvername[nsolver]);
return LIS_ERR_NOT_IMPLEMENTED;
}
else if( precision==LIS_PRECISION_SWITCH && lis_solver_execute_switch[nsolver]==NULL )
{
LIS_SETERR1(LIS_ERR_NOT_IMPLEMENTED,"Switch solver %s is not implemented\n",lis_solvername[nsolver]);
return LIS_ERR_NOT_IMPLEMENTED;
}
if( solver->options[LIS_OPTIONS_SWITCH_MAXITER]==-1 )
{
solver->options[LIS_OPTIONS_SWITCH_MAXITER] = maxiter;
}
#endif
err = lis_solver_check_params[nsolver](solver);
if( err )
{
solver->retcode = err;
return err;
}
/* end parameter check */
solver->A = A;
solver->b = b;
/* create initial vector */
#ifndef USE_QUAD_PRECISION
err = lis_vector_duplicate(A,&xx);
#else
if( precision==LIS_PRECISION_DOUBLE )
{
err = lis_vector_duplicate(A,&xx);
}
else
{
err = lis_vector_duplicateex(LIS_PRECISION_QUAD,A,&xx);
}
#endif
if( err )
{
solver->retcode = err;
return err;
}
if( solver->options[LIS_OPTIONS_INITGUESS_ZEROS] )
{
if( output ) lis_printf(A->comm,"initial vector x = 0\n");
#ifndef USE_QUAD_PRECISION
lis_vector_set_all(0.0,xx);
#else
if( precision==LIS_PRECISION_DOUBLE )
{
lis_vector_set_all(0.0,xx);
}
else
{
lis_vector_set_allex_nm(0.0,xx);
}
#endif
}
else
{
if( output ) lis_printf(A->comm,"initial vector x = user defined\n");
#ifndef USE_QUAD_PRECISION
lis_vector_copy(x,xx);
#else
if( precision==LIS_PRECISION_DOUBLE )
{
lis_vector_copy(x,xx);
}
else
{
lis_vector_copyex_nm(x,xx);
}
#endif
}
/* create residual history vector */
if( solver->residual ) lis_free(solver->residual);
residual = (LIS_SCALAR *)lis_malloc((maxiter+2)*sizeof(LIS_SCALAR),"lis_solve::residual");
if( residual==NULL )
{
LIS_SETERR_MEM((maxiter+2)*sizeof(LIS_SCALAR));
lis_vector_destroy(xx);
solver->retcode = err;
return err;
}
residual[0] = 1.0;
n = A->n;
np = A->np;
t = NULL;
At = NULL;
p_c_times = lis_wtime();
if( precon_type==LIS_PRECON_TYPE_IS )
{
if( solver->d==NULL )
{
err = lis_vector_duplicate(A,&solver->d);
if( err )
{
return err;
}
}
if( !A->is_scaled )
{
lis_matrix_scaling(A,b,solver->d,LIS_SCALE_JACOBI);
}
else if( !b->is_scaled )
{
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(i=0;i<n;i++)
{
b->value[i] = b->value[i]*solver->d->value[i];
}
}
if( nsolver >= LIS_SOLVER_JACOBI && nsolver <= LIS_SOLVER_SOR )
{
solver->options[LIS_OPTIONS_ISLEVEL] = 0;
}
}
else if( nsolver >= LIS_SOLVER_JACOBI && nsolver <= LIS_SOLVER_SOR && precon_type!=LIS_PRECON_TYPE_NONE )
{
if( solver->d==NULL )
{
err = lis_vector_duplicate(A,&solver->d);
if( err )
{
return err;
}
}
if( !A->is_scaled )
{
lis_matrix_scaling(A,b,solver->d,LIS_SCALE_JACOBI);
}
}
else if( scale )
{
if( storage==LIS_MATRIX_BSR && scale==LIS_SCALE_JACOBI )
{
if( A->matrix_type!=LIS_MATRIX_BSR )
{
err = lis_matrix_duplicate(A,&B);
if( err ) return err;
lis_matrix_set_blocksize(B,block,block,NULL,NULL);
lis_matrix_set_type(B,storage);
err = lis_matrix_convert(A,B);
if( err ) return err;
lis_matrix_storage_destroy(A);
lis_matrix_DLU_destroy(A);
lis_matrix_diag_destroy(A->WD);
if( A->l2g_map ) lis_free( A->l2g_map );
if( A->commtable ) lis_commtable_destroy( A->commtable );
if( A->ranges ) lis_free( A->ranges );
err = lis_matrix_copy_struct(B,A);
if( err ) return err;
lis_free(B);
}
err = lis_matrix_split(A);
if( err ) return err;
err = lis_matrix_diag_duplicate(A->D,&solver->WD);
if( err ) return err;
lis_matrix_diag_copy(A->D,solver->WD);
lis_matrix_diag_inverse(solver->WD);
lis_matrix_bscaling_bsr(A,solver->WD);
lis_vector_duplicate(A,&t);
lis_matrix_diag_matvec(solver->WD,b,t);
lis_vector_copy(t,b);
lis_vector_destroy(t);
t = NULL;
}
else
{
if( solver->d==NULL )
{
err = lis_vector_duplicate(A,&solver->d);
if( err )
{
return err;
}
}
if( scale==LIS_SCALE_JACOBI && nsolver==LIS_SOLVER_CG )
{
scale = LIS_SCALE_SYMM_DIAG;
}
if( !A->is_scaled )
{
lis_matrix_scaling(A,b,solver->d,scale);
}
else if( !b->is_scaled )
{
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(i=0;i<n;i++)
{
b->value[i] = b->value[i]*solver->d->value[i];
}
}
}
}
/* precon_type = precon->precon_type;*/
if( precon_type==LIS_PRECON_TYPE_IS )
{
if( nsolver < LIS_SOLVER_JACOBI || nsolver > LIS_SOLVER_SOR )
{
AA = solver->A;
bb = solver->b;
}
else
{
AA = precon->A;
bb = precon->Pb;
}
}
else
{
AA = A;
bb = b;
}
p_c_times = lis_wtime() - p_c_times;
itimes = lis_wtime();
/* Matrix Convert */
solver->A = AA;
solver->b = bb;
err = lis_matrix_convert_self(solver);
if( err )
{
lis_vector_destroy(xx);
lis_solver_work_destroy(solver);
lis_free(residual);
solver->retcode = err;
return err;
}
block = solver->A->bnr;
if( A->my_rank==0 )
{
if( output ) printf("precision : %s\n", lis_precisionname[precision]);
if( output ) printf("solver : %s %d\n", lis_solvername[nsolver],nsolver);
switch( precon_type )
{
case LIS_PRECON_TYPE_ILU:
i = solver->options[LIS_OPTIONS_FILL];
if( A->matrix_type==LIS_MATRIX_BSR || A->matrix_type==LIS_MATRIX_VBR )
{
if( output ) sprintf(buf,"Block %s(%d)",lis_preconname[precon_type],i);
}
else
{
if( output ) sprintf(buf,"%s(%d)",lis_preconname[precon_type],i);
}
break;
default:
if( output ) sprintf(buf,"%s",lis_preconname[precon_type]);
break;
}
if( solver->options[LIS_OPTIONS_ADDS] && precon_type )
{
if( output ) printf("precon : %s + additive schwarz\n", buf);
}
else
{
if( output ) printf("precon : %s\n", buf);
}
}
switch(conv_cond)
{
case LIS_CONV_COND_NRM2_R:
case LIS_CONV_COND_NRM2_B:
if( A->my_rank==0 )
{
if( output ) ("CONV_COND : ||r||_2 <= %6.1e*||r_0||_2\n", tol);
}
break;
case LIS_CONV_COND_NRM1_B:
lis_vector_nrm1(b,&nrm2);
nrm2 = nrm2*tol_w + tol;
if( A->my_rank==0 )
{
if( output ) printf("conv_cond : ||r||_1 <= %6.1e*||b||_1 + %6.1e = %6.1e\n", tol_w,tol,nrm2);
}
break;
}
if( A->my_rank==0 )
{
if( AA->matrix_type==LIS_MATRIX_BSR || AA->matrix_type==LIS_MATRIX_BSC )
{
if( output ) printf("storage : %s(%d x %d)\n", lis_storagename[AA->matrix_type-1],block,block);
}
else
{
if( output ) printf("storage : %s\n", lis_storagename[AA->matrix_type-1]);
}
}
/* create work vector */
err = lis_solver_malloc_work[nsolver](solver);
if( err )
{
lis_vector_destroy(xx);
lis_precon_destroy(precon);
solver->retcode = err;
return err;
}
if( nsolver==LIS_SOLVER_BICG && is_use_at )
{
if( output ) lis_printf(A->comm,"Use At\n");
lis_matrix_duplicate(AA,&At);
lis_matrix_set_type(At,LIS_USE_AT_TYPE[AA->matrix_type]);
lis_matrix_convert(AA,At);
solver->At = At;
}
solver->x = xx;
solver->xx = x;
solver->precon = precon;
solver->residual = residual;
/* execute solver */
#ifndef USE_QUAD_PRECISION
err = lis_solver_execute[nsolver](solver);
#else
if( precision==LIS_PRECISION_DOUBLE )
{
err = lis_solver_execute[nsolver](solver);
}
else if( precision==LIS_PRECISION_QUAD )
{
err = lis_solver_execute_quad[nsolver](solver);
}
else if( precision==LIS_PRECISION_SWITCH )
{
err = lis_solver_execute_switch[nsolver](solver);
}
#endif
solver->retcode = err;
if( scale==LIS_SCALE_SYMM_DIAG && precon_type!=LIS_PRECON_TYPE_IS)
{
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(i=0;i<n;i++)
{
x->value[i] = xx->value[i]*solver->d->value[i];
}
}
else
{
#ifndef USE_QUAD_PRECISION
lis_vector_copy(xx,x);
#else
if( precision==LIS_PRECISION_DOUBLE )
{
lis_vector_copy(xx,x);
}
else
{
lis_vector_copyex_mn(xx,x);
}
#endif
}
itimes = lis_wtime() - itimes - solver->ptimes;
p_i_times = solver->ptimes;
solver->ptimes = p_c_times + p_i_times;
solver->p_c_times = p_c_times;
solver->p_i_times = p_i_times;
solver->times = solver->ptimes + itimes;
solver->itimes = itimes;
lis_solver_work_destroy(solver);
lis_vector_duplicate(A,&t);
xx->precision = LIS_PRECISION_DEFAULT;
lis_matvec(A,xx,t);
lis_vector_xpay(b,-1.0,t);
if( scale==LIS_SCALE_SYMM_DIAG && precon_type!=LIS_PRECON_TYPE_IS)
{
#ifdef _OPENMP
#pragma omp parallel for
#endif
for(i=0;i<n;i++)
{
t->value[i] = t->value[i]/solver->d->value[i];
}
}
lis_vector_nrm2(t,&nrm2);
/*
solver->resid = nrm2;
*/
if( A->my_rank==0 )
{
if( err )
{
if( output ) printf("lis_solve : %s(code=%d)\n\n",lis_returncode[err],err);
}
else
{
if( output ) printf("lis_solve : normal end\n\n");
}
}
if( precision==LIS_PRECISION_DOUBLE )
{
solver->iter2 = solver->iter;
}
else if( precision==LIS_PRECISION_QUAD )
{
solver->iter2 = 0;
}
lis_vector_destroy(t);
/* lis_vector_destroy(d);*/
lis_vector_destroy(xx);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT main(int argc, char* argv[])
{
LIS_Comm comm;
LIS_INT err;
int nprocs,my_rank;
LIS_INT nesol;
LIS_MATRIX A,B;
LIS_VECTOR x;
LIS_SCALAR evalue0;
LIS_ESOLVER esolver;
LIS_REAL residual;
LIS_INT iter;
double time;
double itime,ptime,p_c_time,p_i_time;
char esolvername[128];
LIS_DEBUG_FUNC_IN;
lis_initialize(&argc, &argv);
comm = LIS_COMM_WORLD;
#ifdef USE_MPI
MPI_Comm_size(comm,&nprocs);
MPI_Comm_rank(comm,&my_rank);
#else
nprocs = 1;
my_rank = 0;
#endif
if( argc < 5 )
{
lis_printf(comm,"Usage: %s matrix_a_filename matrix_b_filename evector_filename rhistory_filename [options]\n", argv[0]);
CHKERR(1);
}
lis_printf(comm,"\n");
lis_printf(comm,"number of processes = %d\n",nprocs);
#ifdef _OPENMP
lis_printf(comm,"max number of threads = %d\n",omp_get_num_procs());
lis_printf(comm,"number of threads = %d\n",omp_get_max_threads());
#endif
/* create matrix and vectors */
lis_matrix_create(comm,&A);
lis_matrix_create(comm,&B);
lis_printf(comm,"\nmatrix A:\n");
lis_input_matrix(A,argv[1]);
lis_printf(comm,"matrix B:\n");
lis_input_matrix(B,argv[2]);
lis_vector_duplicate(A,&x);
lis_esolver_create(&esolver);
lis_esolver_set_option("-e gii -eprint mem",esolver);
err = lis_esolver_set_optionC(esolver);
CHKERR(err);
err = lis_gesolve(A,B,x,&evalue0,esolver);
CHKERR(err);
lis_esolver_get_esolver(esolver,&nesol);
lis_esolver_get_esolvername(nesol,esolvername);
lis_esolver_get_residualnorm(esolver,&residual);
lis_esolver_get_iter(esolver,&iter);
lis_esolver_get_timeex(esolver,&time,&itime,&ptime,&p_c_time,&p_i_time);
lis_printf(comm,"%s: mode number = %d\n", esolvername, 0);
#ifdef _COMPLEX
lis_printf(comm,"%s: eigenvalue = (%e, %e)\n", esolvername, (double)creal(evalue0), (double)cimag(evalue0));
#else
lis_printf(comm,"%s: eigenvalue = %e\n", esolvername, (double)evalue0);
#endif
lis_printf(comm,"%s: number of iterations = %D\n",esolvername, iter);
lis_printf(comm,"%s: elapsed time = %e sec.\n", esolvername, time);
lis_printf(comm,"%s: preconditioner = %e sec.\n", esolvername, ptime);
lis_printf(comm,"%s: matrix creation = %e sec.\n", esolvername, p_c_time);
lis_printf(comm,"%s: linear solver = %e sec.\n", esolvername, itime);
lis_printf(comm,"%s: relative residual = %e\n\n",esolvername, (double)residual);
/* write eigenvector */
lis_output_vector(x,LIS_FMT_MM,argv[3]);
/* write residual history */
lis_esolver_output_rhistory(esolver,argv[4]);
lis_esolver_destroy(esolver);
lis_matrix_destroy(A);
lis_matrix_destroy(B);
lis_vector_destroy(x);
lis_finalize();
LIS_DEBUG_FUNC_OUT;
return 0;
}
LIS_INT lis_egpi(LIS_ESOLVER esolver)
{
LIS_Comm comm;
LIS_MATRIX A,B;
LIS_VECTOR w,v,y,q;
LIS_SCALAR eta,theta;
LIS_INT emaxiter;
LIS_REAL tol;
LIS_INT iter,iter2,output;
LIS_SCALAR oshift,ishift;
LIS_REAL nrm2,resid;
LIS_SOLVER solver;
double time,itime,ptime,p_c_time,p_i_time;
LIS_INT err;
LIS_PRECON precon;
LIS_INT nsol, precon_type;
char solvername[128], preconname[128];
LIS_DEBUG_FUNC_IN;
comm = LIS_COMM_WORLD;
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
tol = esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN];
output = esolver->options[LIS_EOPTIONS_OUTPUT];
#ifdef _COMPLEX
oshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN] + esolver->params[LIS_EPARAMS_SHIFT_IM - LIS_EOPTIONS_LEN] * _Complex_I;
#else
oshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN];
#endif
A = esolver->A;
B = esolver->B;
v = esolver->x;
if (esolver->options[LIS_EOPTIONS_INITGUESS_ONES] )
{
lis_vector_set_all(1.0,v);
}
w = esolver->work[0];
y = esolver->work[1];
q = esolver->work[2];
if ( esolver->ishift != 0.0 ) oshift = ishift;
if ( oshift != 0.0 ) lis_matrix_shift_matrix(A, B, oshift);
if( output )
{
#ifdef _COMPLEX
lis_printf(comm,"shift : (%e, %e)\n", (double)creal(oshift), (double)cimag(oshift));
#else
lis_printf(comm,"shift : %e\n", (double)oshift);
#endif
}
lis_solver_create(&solver);
lis_solver_set_option("-i bicg -p none",solver);
err = lis_solver_set_optionC(solver);
CHKERR(err);
lis_solver_get_solver(solver, &nsol);
lis_solver_get_precon(solver, &precon_type);
lis_solver_get_solvername(nsol, solvername);
lis_solver_get_preconname(precon_type, preconname);
if( output )
{
lis_printf(comm,"linear solver : %s\n", solvername);
lis_printf(comm,"preconditioner : %s\n", preconname);
}
/* create preconditioner */
solver->A = B;
err = lis_precon_create(solver, &precon);
if( err )
{
lis_solver_work_destroy(solver);
solver->retcode = err;
return err;
}
iter=0;
while (iter<emaxiter)
{
iter = iter+1;
/* v = v / ||v||_2 */
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
/* w = A * v */
lis_matvec(A, v, w);
/* v = v / <v,w>^1/2, w = w / <v,w>^1/2 */
lis_vector_dot(v, w, &eta);
eta = sqrt(eta);
lis_vector_scale(1.0/eta, v);
lis_vector_scale(1.0/eta, w);
/* y = B^-1 * w */
err = lis_solve_kernel(B, w, y, solver, precon);
if( err )
{
lis_solver_work_destroy(solver);
solver->retcode = err;
return err;
}
lis_solver_get_iter(solver, &iter2);
/* theta = <w,y> */
lis_vector_dot(w, y, &theta);
/* resid = ||y - theta * v||_2 / |theta| */
lis_vector_axpyz(-theta, v, y, q);
lis_vector_nrm2(q, &resid);
resid = resid / fabs(theta);
/* v = y */
lis_vector_copy(y, v);
/* convergence check */
lis_solver_get_timeex(solver,&time,&itime,&ptime,&p_c_time,&p_i_time);
esolver->ptime += solver->ptime;
esolver->itime += solver->itime;
esolver->p_c_time += solver->p_c_time;
esolver->p_i_time += solver->p_i_time;
if( output )
{
if( output & LIS_EPRINT_MEM ) esolver->rhistory[iter] = resid;
if( output & LIS_EPRINT_OUT ) lis_print_rhistory(comm,iter,resid);
}
if( tol >= resid )
{
esolver->retcode = LIS_SUCCESS;
esolver->iter[0] = iter;
esolver->resid[0] = resid;
esolver->evalue[0] = theta + oshift;
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
if ( oshift != 0.0 ) lis_matrix_shift_matrix(A, B, -oshift);
lis_precon_destroy(precon);
lis_solver_destroy(solver);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
}
lis_precon_destroy(precon);
esolver->retcode = LIS_MAXITER;
esolver->iter[0] = iter;
esolver->resid[0] = resid;
esolver->evalue[0] = theta + oshift;
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
if ( oshift != 0.0 ) lis_matrix_shift_matrix(A, B, -oshift);
lis_solver_destroy(solver);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_epi(LIS_ESOLVER esolver)
{
LIS_Comm comm;
LIS_MATRIX A;
LIS_VECTOR v,y,q;
LIS_SCALAR theta;
LIS_INT emaxiter;
LIS_REAL tol;
LIS_INT iter,output;
LIS_SCALAR oshift,ishift;
LIS_REAL nrm2,resid;
LIS_DEBUG_FUNC_IN;
comm = LIS_COMM_WORLD;
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
tol = esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN];
output = esolver->options[LIS_EOPTIONS_OUTPUT];
#ifdef _COMPLEX
oshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN] + esolver->params[LIS_EPARAMS_SHIFT_IM - LIS_EOPTIONS_LEN] * _Complex_I;
#else
oshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN];
#endif
A = esolver->A;
v = esolver->x;
if (esolver->options[LIS_EOPTIONS_INITGUESS_ONES] )
{
lis_vector_set_all(1.0,v);
}
y = esolver->work[0];
q = esolver->work[1];
if ( esolver->ishift != 0.0 ) oshift = ishift;
if ( oshift != 0.0 ) lis_matrix_shift_diagonal(A, oshift);
if( output )
{
#ifdef _COMPLEX
lis_printf(comm,"shift : (%e, %e)\n", (double)creal(oshift), (double)cimag(oshift));
#else
lis_printf(comm,"shift : %e\n", (double)oshift);
#endif
}
iter=0;
while (iter<emaxiter)
{
iter = iter+1;
/* v = v / ||v||_2 */
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
/* y = A * v */
lis_matvec(A,v,y);
/* theta = <v,y> */
lis_vector_dot(v, y, &theta);
/* resid = ||y - theta * v||_2 / |theta| */
lis_vector_axpyz(-theta, v, y, q);
lis_vector_nrm2(q, &resid);
resid = resid / fabs(theta);
/* v = y */
lis_vector_copy(y, v);
/* convergence check */
if( output )
{
if( output & LIS_EPRINT_MEM ) esolver->rhistory[iter] = resid;
if( output & LIS_EPRINT_OUT ) lis_print_rhistory(comm,iter,resid);
}
if( tol >= resid )
{
esolver->retcode = LIS_SUCCESS;
esolver->iter[0] = iter;
esolver->resid[0] = resid;
esolver->evalue[0] = theta + oshift;
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
if ( oshift != 0.0 ) lis_matrix_shift_diagonal(A, -oshift);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
}
esolver->retcode = LIS_MAXITER;
esolver->iter[0] = iter;
esolver->resid[0] = resid;
esolver->evalue[0] = theta + oshift;
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
if ( oshift != 0.0 ) lis_matrix_shift_diagonal(A, -oshift);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_esolve(LIS_MATRIX A, LIS_VECTOR x, LIS_SCALAR *evalue0, LIS_ESOLVER esolver)
{
LIS_INT nesolver,niesolver,emaxiter;
LIS_SCALAR *evalue;
LIS_VECTOR *evector;
LIS_SCALAR *resid;
LIS_SCALAR *rhistory;
LIS_INT *iter,*iter2;
LIS_INT err;
LIS_INT output;
LIS_INT ss, mode;
double time;
double gshift;
LIS_INT estorage,eblock;
LIS_MATRIX B;
LIS_INT eprecision;
LIS_VECTOR xx;
LIS_DEBUG_FUNC_IN;
/* begin parameter check */
err = lis_matrix_check(A,LIS_MATRIX_CHECK_ALL);
if( err ) return err;
if( x==NULL )
{
LIS_SETERR(LIS_ERR_ILL_ARG,"vector x is undefined\n");
return LIS_ERR_ILL_ARG;
}
if( A->n!=x->n )
{
return LIS_ERR_ILL_ARG;
}
if( A->gn<=0 )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Size n(=%d) of matrix A is less than 0\n",A->gn);
return LIS_ERR_ILL_ARG;
}
nesolver = esolver->options[LIS_EOPTIONS_ESOLVER];
niesolver = esolver->options[LIS_EOPTIONS_INNER_ESOLVER];
ss = esolver->options[LIS_EOPTIONS_SUBSPACE];
mode = esolver->options[LIS_EOPTIONS_MODE];
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
gshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN];
output = esolver->options[LIS_EOPTIONS_OUTPUT];
estorage = esolver->options[LIS_EOPTIONS_STORAGE];
eblock = esolver->options[LIS_EOPTIONS_STORAGE_BLOCK];
eprecision = esolver->options[LIS_EOPTIONS_PRECISION];
esolver->eprecision = eprecision;
if( nesolver < 1 || nesolver > LIS_ESOLVER_LEN )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_ESOLVER is %d (Set between 1 to %d)\n",nesolver, LIS_ESOLVER_LEN);
return LIS_ERR_ILL_ARG;
}
if( niesolver < 1 || niesolver > 7 )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 1 to 7)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_SI && niesolver > 4 )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 1 to 4 for Subspace)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_LI && niesolver == LIS_ESOLVER_PI )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 2 to 7 for Lanczos)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_AI && (( niesolver == LIS_ESOLVER_PI ) || ( niesolver == LIS_ESOLVER_CG) || ( niesolver == LIS_ESOLVER_JD)) )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 2 to 4 or 6 for Arnoldi)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_SI && ss > A->gn )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_SUBSPACE is %d (Set less than or equal to matrix size %d for Subspace)\n", ss, A->gn);
return LIS_ERR_ILL_ARG;
}
if (( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_LI || esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_AI ) && ss > A->gn )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_SUBSPACE is %d (Set less than or equal to matrix size %d for Lanczos and Arnoldi)\n", ss, A->gn);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_SI && mode >= ss )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_MODE is %d (Set less than subspace size %d for Subspace)\n", mode, ss);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == ( LIS_ESOLVER_LI || LIS_ESOLVER_AI ) && mode >= ss )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_MODE is %d (Set less than subspace size %d for Lanczos or Arnoldi)\n", mode, ss);
return LIS_ERR_ILL_ARG;
}
#ifdef USE_QUAD_PRECISION
if( eprecision==LIS_PRECISION_QUAD && lis_esolver_execute_quad[nesolver]==NULL )
{
LIS_SETERR1(LIS_ERR_NOT_IMPLEMENTED,"Quad precision eigensolver %s is not implemented\n",lis_esolvername[nesolver]);
return LIS_ERR_NOT_IMPLEMENTED;
}
else if( eprecision==LIS_PRECISION_SWITCH && lis_esolver_execute_switch[nesolver]==NULL )
{
LIS_SETERR1(LIS_ERR_NOT_IMPLEMENTED,"Switch esolver %s is not implemented\n",lis_esolvername[nesolver]);
return LIS_ERR_NOT_IMPLEMENTED;
}
if( esolver->options[LIS_EOPTIONS_SWITCH_MAXITER]==-1 )
{
esolver->options[LIS_EOPTIONS_SWITCH_MAXITER] = emaxiter;
}
#endif
/* create eigenvalue array */
if( esolver->evalue ) lis_free(esolver->evalue);
evalue = (LIS_SCALAR *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::evalue");
if( evalue==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
evalue[0] = 1.0;
evalue[ss-1] = 1.0;
/* create residual norm array */
if( esolver->resid ) lis_free(esolver->resid);
resid = (LIS_SCALAR *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::resid");
if( resid==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
/* create number of iterations array */
if( esolver->iter ) lis_free(esolver->iter);
iter = (LIS_INT *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::iter");
if( iter==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
/* create quad precision number of iterations array */
if( esolver->iter2 ) lis_free(esolver->iter2);
iter2 = (LIS_INT *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::iter2");
if( iter2==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
/* create initial vector */
#ifndef USE_QUAD_PRECISION
err = lis_vector_duplicate(A,&xx);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
err = lis_vector_duplicate(A,&xx);
}
else
{
err = lis_vector_duplicateex(LIS_PRECISION_QUAD,A,&xx);
}
#endif
if( err )
{
esolver->retcode = err;
return err;
}
if( esolver->options[LIS_EOPTIONS_INITGUESS_ONES] )
{
if( output ) lis_printf(A->comm,"initial vector x : 1\n");
#ifndef USE_QUAD_PRECISION
lis_vector_set_all(1.0,xx);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
lis_vector_set_all(1.0,xx);
}
else
{
lis_vector_set_allex_nm(1.0,xx);
}
#endif
}
else
{
if( output ) lis_printf(A->comm,"initial vector x : user defined\n");
#ifndef USE_QUAD_PRECISION
lis_vector_copy(x,xx);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
lis_vector_copy(x,xx);
}
else
{
lis_vector_copyex_nm(x,xx);
}
#endif
}
/* global shift */
if ( output ) if( A->my_rank==0 ) printf("shift : %e\n", gshift);
/* create eigenvector array */
if( esolver->evector ) lis_free(esolver->evector);
evector = (LIS_VECTOR *)lis_malloc((ss+2)*sizeof(LIS_VECTOR),"lis_esolve::evector");
if( evector==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_VECTOR));
esolver->retcode = err;
return err;
}
/* create residual history array */
if( esolver->rhistory ) lis_free(esolver->rhistory);
rhistory = (LIS_SCALAR *)lis_malloc((emaxiter+2)*sizeof(LIS_SCALAR),"lis_esolve::rhistory");
if( rhistory==NULL )
{
LIS_SETERR_MEM((emaxiter+2)*sizeof(LIS_SCALAR));
lis_vector_destroy(xx);
esolver->retcode = err;
return err;
}
/* convert matrix */
if( estorage>0 && A->matrix_type!=estorage )
{
err = lis_matrix_duplicate(A,&B);
if( err ) return err;
lis_matrix_set_blocksize(B,eblock,eblock,NULL,NULL);
lis_matrix_set_type(B,estorage);
err = lis_matrix_convert(A,B);
if( err ) return err;
lis_matrix_storage_destroy(A);
lis_matrix_DLU_destroy(A);
lis_matrix_diag_destroy(A->WD);
if( A->l2g_map ) lis_free( A->l2g_map );
if( A->commtable ) lis_commtable_destroy( A->commtable );
if( A->ranges ) lis_free( A->ranges );
err = lis_matrix_copy_struct(B,A);
if( err ) return err;
lis_free(B);
}
esolver->A = A;
esolver->evalue = evalue;
esolver->x = x;
esolver->evector = evector;
rhistory[0] = 1.0;
esolver->rhistory = rhistory;
esolver->resid = resid;
esolver->iter = iter;
esolver->iter2 = iter2;
if( A->my_rank==0 )
{
#ifdef _LONG__DOUBLE
if ( output ) printf("precision : long double\n");
#else
if ( output ) printf("precision : %s\n", lis_eprecisionname[eprecision]);
#endif
#ifdef _LONG__LONG
if ( output ) printf("eigensolver : %s\n", lis_esolvername[nesolver]);
#else
if ( output ) printf("eigensolver : %s\n", lis_esolvername[nesolver]);
#endif
}
if( A->my_rank==0 )
{
#ifdef _LONG__DOUBLE
if ( output ) printf("convergence condition : ||lx-Ax||_2 <= %6.1Le * ||lx||_2\n", esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN]);
#else
if ( output ) printf("convergence condition : ||lx-Ax||_2 <= %6.1e * ||lx||_2\n", esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN]);
#endif
}
if( A->my_rank==0 )
{
if( A->matrix_type==LIS_MATRIX_BSR || A->matrix_type==LIS_MATRIX_BSC )
{
#ifdef _LONG__LONG
if ( output ) printf("matrix storage format : %s(%lld x %lld)\n", lis_estoragename[A->matrix_type-1],eblock,eblock);
#else
if ( output ) printf("matrix storage format : %s(%d x %d)\n", lis_estoragename[A->matrix_type-1],eblock,eblock);
#endif
}
else
{
if ( output ) printf("matrix storage format : %s\n", lis_estoragename[A->matrix_type-1]);
}
}
time = lis_wtime();
esolver->ptime = 0;
esolver->itime = 0;
esolver->p_c_time = 0;
esolver->p_i_time = 0;
if (gshift != 0.0) lis_matrix_shift_diagonal(A, gshift);
/* create work vector */
err = lis_esolver_malloc_work[nesolver](esolver);
if( err )
{
lis_vector_destroy(xx);
esolver->retcode = err;
return err;
}
esolver->x = xx;
esolver->xx = x;
/* execute esolver */
#ifndef USE_QUAD_PRECISION
err = lis_esolver_execute[nesolver](esolver);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
err = lis_esolver_execute[nesolver](esolver);
}
else if( eprecision==LIS_PRECISION_QUAD )
{
err = lis_esolver_execute_quad[nesolver](esolver);
}
else if( eprecision==LIS_PRECISION_SWITCH )
{
err = lis_esolver_execute_switch[nesolver](esolver);
}
#endif
esolver->retcode = err;
*evalue0 = esolver->evalue[0];
lis_vector_copy(esolver->x, x);
esolver->time = lis_wtime() - time;
lis_matrix_shift_diagonal(A, -gshift);
if( A->my_rank==0 )
{
if( err )
{
#ifdef _LONG__LONG
if ( output ) printf("eigensolver status : %s(code=%lld)\n\n",lis_ereturncode[err],err);
#else
if ( output ) printf("eigensolver status : %s(code=%d)\n\n",lis_ereturncode[err],err);
#endif
}
else
{
if ( output ) printf("eigensolver status : normal end\n\n");
}
}
if( eprecision==LIS_PRECISION_DOUBLE )
{
esolver->iter2[mode] = esolver->iter[mode];
}
else if( eprecision==LIS_PRECISION_QUAD )
{
esolver->iter2[mode] = 0;
}
lis_vector_destroy(xx);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
