void LisMatrix::multiply(const LisVector &x, LisVector &y) const
{
if (!_is_assembled)
throw std::logic_error("LisMatrix::matvec(): matrix not assembled.");
int ierr = lis_matvec(_AA, const_cast<LisVector*>(&x)->getRawVector(), y.getRawVector());
checkLisError(ierr);
}
void lis_matvec_f(LIS_MATRIX_F *A, LIS_VECTOR_F *x, LIS_VECTOR_F *y, LIS_INT *ierr)
{
LIS_DEBUG_FUNC_IN;
*ierr = lis_matvec((LIS_MATRIX)LIS_V2P(A),(LIS_VECTOR)LIS_V2P(x),(LIS_VECTOR)LIS_V2P(y));
LIS_DEBUG_FUNC_OUT;
return;
}
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 lis_epi_quad(LIS_ESOLVER esolver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_SCALAR evalue;
LIS_INT emaxiter;
LIS_REAL tol;
LIS_INT iter,output;
LIS_INT nprocs,my_rank;
LIS_REAL nrm2,resid;
LIS_QUAD_PTR qdot_xz;
LIS_VECTOR z,q;
double times, ptimes;
LIS_DEBUG_FUNC_IN;
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
tol = esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN];
output = esolver->options[LIS_EOPTIONS_OUTPUT];
A = esolver->A;
x = esolver->x;
if (esolver->options[LIS_EOPTIONS_INITGUESS_ONES] )
{
lis_vector_set_all(1.0,x);
}
z = esolver->work[0];
q = esolver->work[1];
LIS_QUAD_SCALAR_MALLOC(qdot_xz,0,1);
iter=0;
while (iter<emaxiter)
{
iter = iter+1;
/* x = x / ||x||_2 */
lis_vector_nrm2(x, &nrm2);
lis_vector_scale(1/nrm2, x);
/* z = A * x */
lis_matvec(A,x,z);
/* evalue = <x,z> */
lis_vector_dotex_mmm(x, z, &qdot_xz);
lis_quad_minus((LIS_QUAD *)qdot_xz.hi);
/* resid = ||z - evalue * x||_2 / |evalue| */
lis_vector_axpyzex_mmmm(qdot_xz,x,z,q);
lis_quad_minus((LIS_QUAD *)qdot_xz.hi);
lis_vector_nrm2(q, &resid);
evalue = qdot_xz.hi[0];
resid = fabs(resid / evalue);
/* x = z */
lis_vector_copy(z, x);
/* convergence check */
if( output )
{
if( output & LIS_EPRINT_MEM ) esolver->residual[iter] = resid;
if( output & LIS_EPRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,resid);
}
if( tol >= resid )
{
esolver->retcode = LIS_SUCCESS;
esolver->iter = iter;
esolver->resid = resid;
esolver->evalue[0] = evalue;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
}
esolver->retcode = LIS_MAXITER;
esolver->iter = iter;
esolver->resid = resid;
esolver->evalue[0] = evalue;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT main(LIS_INT argc, char* argv[])
{
LIS_MATRIX A;
LIS_VECTOR x,b,u;
LIS_SOLVER solver;
LIS_INT k,n,gn,ii,jj;
LIS_INT is,ie;
LIS_INT nprocs,my_rank;
int int_nprocs,int_my_rank;
LIS_INT nsol;
LIS_INT err,iter,iter_double,iter_quad;
double time,itime,ptime,p_c_time,p_i_time;
LIS_REAL resid;
char solvername[128];
LIS_INT *ptr,*index;
LIS_SCALAR *value,gamma;
LIS_DEBUG_FUNC_IN;
lis_initialize(&argc, &argv);
#ifdef USE_MPI
MPI_Comm_size(MPI_COMM_WORLD,&int_nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&int_my_rank);
nprocs = int_nprocs;
my_rank = int_my_rank;
#else
nprocs = 1;
my_rank = 0;
#endif
if( argc < 3 )
{
if( my_rank==0 )
{
printf("Usage: %s n gamma [options]\n", argv[0]);
}
CHKERR(1);
}
gn = atoi(argv[1]);
gamma = atof(argv[2]);
if( gn<=0 )
{
#ifdef _LONGLONG
if( my_rank==0 ) printf("n=%lld <=0 \n",gn);
#else
if( my_rank==0 ) printf("n=%d <=0 \n",gn);
#endif
CHKERR(1);
}
if( my_rank==0 )
{
printf("\n");
#ifdef _LONGLONG
printf("number of processes = %lld\n",nprocs);
#else
printf("number of processes = %d\n",nprocs);
#endif
}
#ifdef _OPENMP
if( my_rank==0 )
{
#ifdef _LONGLONG
printf("max number of threads = %lld\n",omp_get_num_procs());
printf("number of threads = %lld\n",omp_get_max_threads());
#else
printf("max number of threads = %d\n",omp_get_num_procs());
printf("number of threads = %d\n",omp_get_max_threads());
#endif
}
#endif
if(my_rank==0)
{
#ifdef _LONGLONG
printf("n = %lld, gamma = %f\n\n",gn,gamma);
#else
printf("n = %d, gamma = %f\n\n",gn,gamma);
#endif
}
/* create matrix and vectors */
err = lis_matrix_create(LIS_COMM_WORLD,&A); CHKERR(err);
err = lis_matrix_set_size(A,0,gn); CHKERR(err);
err = lis_matrix_get_size(A,&n,&gn); CHKERR(err);
err = lis_matrix_malloc_csr(n,3*n,&ptr,&index,&value); CHKERR(err);
err = lis_matrix_get_range(A,&is,&ie); CHKERR(err);
k = 0;
ptr[0] = 0;
for(ii=is;ii<ie;ii++)
{
if( ii>1 ) { jj = ii - 2; index[k] = jj; value[k++] = gamma;}
if( ii<gn-1 ) { jj = ii + 1; index[k] = jj; value[k++] = 1.0;}
index[k] = ii; value[k++] = 2.0;
ptr[ii-is+1] = k;
}
err = lis_matrix_set_csr(ptr[ie-is],ptr,index,value,A); CHKERR(err);
err = lis_matrix_assemble(A); CHKERR(err);
err = lis_vector_duplicate(A,&u); CHKERR(err);
err = lis_vector_duplicate(u,&b); CHKERR(err);
err = lis_vector_duplicate(u,&x); CHKERR(err);
err = lis_vector_set_all(1.0,u);
lis_matvec(A,u,b);
err = lis_solver_create(&solver); CHKERR(err);
lis_solver_set_option("-print mem",solver);
lis_solver_set_optionC(solver);
err = lis_solve(A,b,x,solver); CHKERR(err);
lis_solver_get_iterex(solver,&iter,&iter_double,&iter_quad);
lis_solver_get_timeex(solver,&time,&itime,&ptime,&p_c_time,&p_i_time);
lis_solver_get_residualnorm(solver,&resid);
lis_solver_get_solver(solver,&nsol);
lis_solver_get_solvername(nsol,solvername);
if( my_rank==0 )
{
#ifdef _LONGLONG
#ifdef _LONG__DOUBLE
printf("%s: number of iterations = %lld \n",solvername, iter);
#else
printf("%s: number of iterations = %lld (double = %lld, quad = %lld)\n",solvername,iter, iter_double, iter_quad);
#endif
#else
#ifdef _LONG__DOUBLE
printf("%s: number of iterations = %d \n",solvername, iter);
#else
printf("%s: number of iterations = %d (double = %d, quad = %d)\n",solvername,iter, iter_double, iter_quad);
#endif
#endif
printf("%s: elapsed time = %e sec.\n",solvername,time);
printf("%s: preconditioner = %e sec.\n",solvername, ptime);
printf("%s: matrix creation = %e sec.\n",solvername, p_c_time);
printf("%s: linear solver = %e sec.\n",solvername, itime);
#ifdef _LONG__DOUBLE
printf("%s: relative residual = %Le\n\n",solvername,resid);
#else
printf("%s: relative residual = %e\n\n",solvername,resid);
#endif
}
lis_solver_destroy(solver);
lis_matrix_destroy(A);
lis_vector_destroy(b);
lis_vector_destroy(x);
lis_vector_destroy(u);
lis_finalize();
LIS_DEBUG_FUNC_OUT;
return 0;
}
LIS_INT lis_bicrsafe_quad(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r, rtld, artld, mr, amr, p, ap, map;
LIS_VECTOR y, my, u, au, z;
LIS_QUAD_PTR alpha, beta, rho, rho_old;
LIS_QUAD_PTR qsi, eta, one;
LIS_QUAD_PTR tmp, tmpdot[5];
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
rtld = solver->work[0];
r = solver->work[1];
mr = solver->work[2];
amr = solver->work[3];
p = solver->work[4];
ap = solver->work[5];
map = solver->work[6];
my = solver->work[7];
y = solver->work[8];
u = solver->work[9];
z = solver->work[10];
au = solver->work[11];
artld = solver->work[12];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(qsi,4,1);
LIS_QUAD_SCALAR_MALLOC(eta,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[0],7,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[1],8,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[2],9,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[3],10,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[4],11,1);
LIS_QUAD_SCALAR_MALLOC(one,13,1);
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_matvect(A,rtld,artld);
time = lis_wtime();
lis_psolve(solver, r, mr);
ptime += lis_wtime()-time;
lis_matvec(A,mr,amr);
lis_vector_dotex_mmm(rtld,amr,&rho_old);
lis_vector_copyex_mm(amr,ap);
lis_vector_copyex_mm(mr,p);
one.hi[0] = -1.0;
one.lo[0] = 0.0;
for( iter=1; iter<=maxiter; iter++ )
{
/* map = M^-1 * ap */
time = lis_wtime();
lis_psolve(solver, ap, map);
ptime += lis_wtime()-time;
/* tmpdot[0] = <artld,map> */
/* alpha = rho_old / tmpdot[0] */
lis_vector_dotex_mmm(artld,map,&tmpdot[0]);
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho_old.hi,(LIS_QUAD *)tmpdot[0].hi);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <amr,r> */
/* tmpdot[2] = <y,r> */
/* tmpdot[3] = <amr,y> */
/* tmpdot[4] = <amr,amr> */
lis_vector_dotex_mmm(y,y,&tmpdot[0]);
lis_vector_dotex_mmm(amr,r,&tmpdot[1]);
lis_vector_dotex_mmm(y,r,&tmpdot[2]);
lis_vector_dotex_mmm(amr,y,&tmpdot[3]);
lis_vector_dotex_mmm(amr,amr,&tmpdot[4]);
if(iter==1)
{
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[1].hi,(LIS_QUAD *)tmpdot[4].hi);
eta.hi[0] = 0.0;
eta.lo[0] = 0.0;
}
else
{
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_sqr((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[2].hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)eta.hi);
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[2].hi);
lis_quad_mul((LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[3].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_sub((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_div((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmp.hi);
}
/* u = qsi*map + eta*my + eta*beta*u */
/* au = A * u */
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)beta.hi);
lis_vector_scaleex_mm(tmp,u);
lis_vector_axpyex_mmm(qsi,map,u);
lis_vector_axpyex_mmm(eta,my,u);
lis_matvec(A,u,au);
/* z = qsi*mr + eta*z - alpha*u */
lis_vector_scaleex_mm(eta,z);
lis_vector_axpyex_mmm(qsi,mr,z);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,u,z);
/* y = qsi*amr + eta*y - alpha*au */
/* my = M^-1 * y */
lis_vector_scaleex_mm(eta,y);
lis_vector_axpyex_mmm(qsi,amr,y);
lis_vector_axpyex_mmm(alpha,au,y);
time = lis_wtime();
lis_psolve(solver, y, my);
ptime += lis_wtime()-time;
/* x = x + alpha*p + z */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,p,x);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(one,z,x);
/* r = r - alpha*ap - y */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(alpha,ap,r);
lis_vector_axpyex_mmm(one,y,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* mr = mr - alpha*map - my */
/* amr = A * mr */
/* rho = <rtld,amr> */
lis_vector_axpyex_mmm(alpha,map,mr);
lis_vector_axpyex_mmm(one,my,mr);
lis_matvec(A,mr,amr);
lis_vector_dotex_mmm(rtld,amr,&rho);
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)alpha.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,(LIS_QUAD *)tmp.hi);
/* p = mr + beta*(p - u) */
/* ap = amr + beta*(ap - au) */
lis_vector_axpyex_mmm(one,u,p);
lis_vector_xpayex_mmm(mr,beta,p);
lis_vector_axpyex_mmm(one,au,ap);
lis_vector_xpayex_mmm(amr,beta,ap);
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_orthomin_quad(LIS_SOLVER solver)
{
LIS_Comm comm;
LIS_MATRIX A;
LIS_PRECON M;
LIS_VECTOR x;
LIS_VECTOR r, rtld, *p, *ap, *aptld;
LIS_QUAD *dotsave;
LIS_QUAD_PTR alpha, beta, tmp, one;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_INT m,l,lmax,ip,ip0;
LIS_DEBUG_FUNC_IN;
comm = LIS_COMM_WORLD;
A = solver->A;
M = solver->precon;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
m = solver->options[LIS_OPTIONS_RESTART];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(tmp,3,1);
LIS_QUAD_SCALAR_MALLOC(one,4,1);
r = solver->work[0];
rtld = solver->work[1];
p = &solver->work[2];
ap = &solver->work[ (m+1)+2];
aptld = &solver->work[2*(m+1)+2];
one.hi[0] = 1.0;
one.lo[0] = 0.0;
dotsave = (LIS_QUAD *)lis_malloc( sizeof(LIS_QUAD) * (m+1),"lis_orthomin_quad::dotsave" );
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,M,r,rtld,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
iter=1;
while( iter<=maxiter )
{
ip = (iter-1) % (m+1);
/* p[ip] = rtld */
lis_vector_copyex_mm(rtld,p[ip]);
/* ap[ip] = A*p[ip] */
/* aptld[ip] = M^-1 ap[ip] */
lis_matvec(A,p[ip],ap[ip]);
time = lis_wtime();
lis_psolve(solver, ap[ip], aptld[ip]);
ptime += lis_wtime()-time;
lmax = _min(m,iter-1);
for(l=1;l<=lmax;l++)
{
ip0 = (ip+m+1-l) % (m+1);
/* beta = -<Ar[ip],Ap[ip0]> / <Ap[ip0],Ap[ip0]> */
lis_vector_dotex_mmm(aptld[ip],aptld[ip0],&beta);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,&dotsave[l-1]);
lis_quad_minus((LIS_QUAD *)beta.hi);
lis_vector_axpyex_mmm(beta,p[ip0] ,p[ip]);
lis_vector_axpyex_mmm(beta,ap[ip0] ,ap[ip]);
lis_vector_axpyex_mmm(beta,aptld[ip0],aptld[ip]);
}
for(l=m-1;l>0;l--)
{
dotsave[l] = dotsave[l-1];
}
lis_vector_dotex_mmm(aptld[ip],aptld[ip],&tmp);
dotsave[0].hi = tmp.hi[0];
dotsave[0].lo = tmp.lo[0];
/* test breakdown */
if( tmp.hi[0]==0.0 && tmp.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
lis_free(dotsave);
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
lis_quad_div(&dotsave[0],(LIS_QUAD *)one.hi,&dotsave[0]);
/* alpha = <rtld,Aptld[ip]> */
lis_vector_dotex_mmm(rtld,aptld[ip],&alpha);
lis_quad_mul((LIS_QUAD *)alpha.hi,(LIS_QUAD *)alpha.hi,&dotsave[0]);
lis_vector_axpyex_mmm( alpha,p[ip],x);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,ap[ip],r);
lis_vector_axpyex_mmm(alpha,aptld[ip],rtld);
lis_quad_minus((LIS_QUAD *)alpha.hi);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT ) lis_print_rhistory(comm,iter,nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
lis_free(dotsave);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
iter++;
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
lis_free(dotsave);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_fgmres_quad(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR b,x;
LIS_VECTOR r,s,*z,*v;
LIS_QUAD *h;
LIS_QUAD_PTR aa,bb,rr,a2,b2,t,one,tmp;
LIS_REAL bnrm2,nrm2,tol;
LIS_INT iter,maxiter,n,output;
double time,ptime;
LIS_REAL rnorm;
LIS_INT i,j,k,m;
LIS_INT ii,i1,iiv,i1v,iih,jj;
LIS_INT h_dim;
LIS_INT cs,sn;
LIS_DEBUG_FUNC_IN;
A = solver->A;
b = solver->b;
x = solver->x;
n = A->n;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
m = solver->options[LIS_OPTIONS_RESTART];
h_dim = m+1;
ptime = 0.0;
s = solver->work[0];
r = solver->work[1];
z = &solver->work[2];
v = &solver->work[m+2];
h = (LIS_QUAD *)lis_malloc( sizeof(LIS_QUAD)*(h_dim+1)*(h_dim+2),"lis_fgmres_quad::h" );
cs = (m+1)*h_dim;
sn = (m+2)*h_dim;
LIS_QUAD_SCALAR_MALLOC(aa,0,1);
LIS_QUAD_SCALAR_MALLOC(bb,1,1);
LIS_QUAD_SCALAR_MALLOC(rr,2,1);
LIS_QUAD_SCALAR_MALLOC(a2,3,1);
LIS_QUAD_SCALAR_MALLOC(b2,4,1);
LIS_QUAD_SCALAR_MALLOC(t,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(one,7,1);
one.hi[0] = 1.0;
one.lo[0] = 0.0;
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,v[0],&bnrm2) )
{
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
rnorm = 1.0/bnrm2;
iter=0;
while( iter<maxiter )
{
/* first column of V */
/* v = r / ||r||_2 */
lis_vector_scaleex_nm(bnrm2,v[0]);
/* s = ||r||_2 e_1 */
lis_vector_set_allex_nm(0.0,s);
s->value[0] = rnorm;
s->value_lo[0] = 0.0;
i = 0;
do
{
iter++;
i++;
ii = i-1;
i1 = i;
iiv = i-1;
i1v = i;
iih = (i-1)*h_dim;
/* z = M^-1 * v */
time = lis_wtime();
lis_psolve(solver,v[iiv],z[iiv]);
ptime += lis_wtime()-time;
/* w = A * z */
lis_matvec(A,z[iiv], v[i1v]);
for(k=0;k<i;k++)
{
/* h[k,i] = <w,v[k]> */
/* w = w - h[k,i] * v[k] */
lis_vector_dotex_mmm(v[i1v],v[k],&t);
h[k+iih].hi = t.hi[0];
h[k+iih].lo = t.lo[0];
lis_quad_minus((LIS_QUAD *)t.hi);
lis_vector_axpyex_mmm(t,v[k],v[i1v]);
}
/* h[i+1,i] = ||w|| */
/* v[i+1] = w / h[i+1,i] */
lis_vector_nrm2ex_mm(v[i1v],&t);
h[i1+iih].hi = t.hi[0];
h[i1+iih].lo = t.lo[0];
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)one.hi,(LIS_QUAD *)t.hi);
lis_vector_scaleex_mm(tmp,v[i1v]);
for(k=1;k<=ii;k++)
{
jj = k-1;
t.hi[0] = h[jj+iih].hi;
t.lo[0] = h[jj+iih].lo;
lis_quad_mul((LIS_QUAD *)aa.hi,(LIS_QUAD *)&h[jj+cs],(LIS_QUAD *)t.hi);
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj+sn],(LIS_QUAD *)&h[k+iih]);
lis_quad_add((LIS_QUAD *)aa.hi,(LIS_QUAD *)aa.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)bb.hi,(LIS_QUAD *)&h[jj+sn],(LIS_QUAD *)t.hi);
lis_quad_minus((LIS_QUAD *)bb.hi);
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj+cs],(LIS_QUAD *)&h[k+iih]);
lis_quad_add((LIS_QUAD *)bb.hi,(LIS_QUAD *)bb.hi,(LIS_QUAD *)tmp.hi);
h[jj+iih].hi = aa.hi[0];
h[jj+iih].lo = aa.lo[0];
h[k+iih].hi = bb.hi[0];
h[k+iih].lo = bb.lo[0];
}
aa.hi[0] = h[ii+iih].hi;
aa.lo[0] = h[ii+iih].lo;
bb.hi[0] = h[i1+iih].hi;
bb.lo[0] = h[i1+iih].lo;
lis_quad_sqr((LIS_QUAD *)a2.hi,(LIS_QUAD *)aa.hi);
lis_quad_sqr((LIS_QUAD *)b2.hi,(LIS_QUAD *)bb.hi);
lis_quad_add((LIS_QUAD *)rr.hi,(LIS_QUAD *)a2.hi,(LIS_QUAD *)b2.hi);
lis_quad_sqrt((LIS_QUAD *)rr.hi,(LIS_QUAD *)rr.hi);
if( rr.hi[0]==0.0 )
{
rr.hi[0]=1.0e-17;
rr.lo[0]=0.0;
}
lis_quad_div((LIS_QUAD *)&h[ii+cs],(LIS_QUAD *)aa.hi,(LIS_QUAD *)rr.hi);
lis_quad_div((LIS_QUAD *)&h[ii+sn],(LIS_QUAD *)bb.hi,(LIS_QUAD *)rr.hi);
tmp.hi[0] = s->value[ii];
tmp.lo[0] = s->value_lo[ii];
lis_quad_mul((LIS_QUAD *)aa.hi,(LIS_QUAD *)&h[ii+sn],(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)bb.hi,(LIS_QUAD *)&h[ii+cs],(LIS_QUAD *)tmp.hi);
lis_quad_minus((LIS_QUAD *)aa.hi);
s->value[i1] = aa.hi[0];
s->value_lo[i1] = aa.lo[0];
s->value[ii] = bb.hi[0];
s->value_lo[ii] = bb.lo[0];
lis_quad_mul((LIS_QUAD *)aa.hi,(LIS_QUAD *)&h[ii+cs],(LIS_QUAD *)&h[ii+iih]);
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[ii+sn],(LIS_QUAD *)&h[i1+iih]);
lis_quad_add((LIS_QUAD *)aa.hi,(LIS_QUAD *)aa.hi,(LIS_QUAD *)tmp.hi);
h[ii+iih].hi = aa.hi[0];
h[ii+iih].lo = aa.lo[0];
/* convergence check */
nrm2 = fabs(s->value[i1]);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 ) break;
} while( i<m && iter <maxiter );
/* Solve H * Y = S for upper Hessenberg matrix H */
tmp.hi[0] = s->value[ii];
tmp.lo[0] = s->value_lo[ii];
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[ii+iih]);
s->value[ii] = tmp.hi[0];
s->value_lo[ii] = tmp.lo[0];
for(k=1;k<=ii;k++)
{
jj = ii-k;
t.hi[0] = s->value[jj];
t.lo[0] = s->value_lo[jj];
for(j=jj+1;j<=ii;j++)
{
tmp.hi[0] = s->value[j];
tmp.lo[0] = s->value_lo[j];
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj+j*h_dim]);
lis_quad_sub((LIS_QUAD *)t.hi,(LIS_QUAD *)t.hi,(LIS_QUAD *)tmp.hi);
}
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)t.hi,(LIS_QUAD *)&h[jj+jj*h_dim]);
s->value[jj] = tmp.hi[0];
s->value_lo[jj] = tmp.lo[0];
}
/* x = x + y * z */
for(j=0;j<=ii;j++)
{
aa.hi[0] = s->value[j];
aa.lo[0] = s->value_lo[j];
lis_vector_axpyex_mmm(aa,z[j],x);
}
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
lis_matvec(A,x,v[0]);
lis_vector_xpay(b,-1.0,v[0]);
memset(v[0]->value_lo,0,n*sizeof(LIS_SCALAR));
lis_vector_nrm2(v[0],&rnorm);
bnrm2 = 1.0/rnorm;
}
solver->retcode = LIS_MAXITER;
solver->iter = iter+1;
solver->resid = nrm2;
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_gmres_switch(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR b,x;
LIS_VECTOR r,s,z,*v;
LIS_QUAD *h;
LIS_SCALAR *hd;
LIS_QUAD_PTR aa,bb,rr,a2,b2,t,one,tmp;
LIS_QUAD_PTR rnorm;
LIS_REAL bnrm2,nrm2,tol,tol2;
LIS_INT iter,maxiter,n,output;
LIS_INT iter2,maxiter2;
double time,ptime;
LIS_INT i,j,k,m;
LIS_INT ii,i1,iiv,i1v,iih,jj;
LIS_INT h_dim;
LIS_INT cs,sn;
LIS_DEBUG_FUNC_IN;
A = solver->A;
b = solver->b;
x = solver->x;
n = A->n;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
maxiter2 = solver->options[LIS_OPTIONS_SWITCH_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
tol2 = solver->params[LIS_PARAMS_SWITCH_RESID-LIS_OPTIONS_LEN];
m = solver->options[LIS_OPTIONS_RESTART];
h_dim = m+1;
ptime = 0.0;
s = solver->work[0];
r = solver->work[1];
z = solver->work[2];
v = &solver->work[3];
LIS_QUAD_SCALAR_MALLOC(aa,0,1);
LIS_QUAD_SCALAR_MALLOC(bb,1,1);
LIS_QUAD_SCALAR_MALLOC(rr,2,1);
LIS_QUAD_SCALAR_MALLOC(a2,3,1);
LIS_QUAD_SCALAR_MALLOC(b2,4,1);
LIS_QUAD_SCALAR_MALLOC(t,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(one,7,1);
LIS_QUAD_SCALAR_MALLOC(rnorm,8,1);
h = (LIS_QUAD *)lis_malloc( sizeof(LIS_QUAD)*(h_dim+1)*(h_dim+2),"lis_gmres_switch::h" );
hd = (LIS_SCALAR *)h;
cs = (m+1)*h_dim;
sn = (m+2)*h_dim;
one.hi[0] = 1.0;
one.lo[0] = 0.0;
z->precision = LIS_PRECISION_DEFAULT;
/* r = M^-1 * (b - A * x) */
lis_matvec(A,x,z);
lis_vector_xpay(b,-1.0,z);
lis_psolve(solver,z,v[0]);
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,v[0],&bnrm2) )
{
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol2 = solver->tol_switch;
iter=0;
while( iter<maxiter2 )
{
/* first column of V */
/* v = r / ||r||_2 */
lis_vector_nrm2(v[0],&rnorm.hi[0]);
lis_vector_scale(1.0/rnorm.hi[0],v[0]);
/* s = ||r||_2 e_1 */
lis_vector_set_all(0,s);
s->value[0] = rnorm.hi[0];
i = 0;
do
{
iter++;
i++;
ii = i-1;
i1 = i;
iiv = i-1;
i1v = i;
iih = (i-1)*h_dim;
/* z = M^-1 * v */
time = lis_wtime();
lis_psolve(solver,v[iiv],z);
ptime += lis_wtime()-time;
/* w = A * z */
lis_matvec(A,z, v[i1v]);
for(k=0;k<i;k++)
{
/* h[k,i] = <w,v[k]> */
/* w = w - h[k,i] * v[k] */
lis_vector_dot(v[i1v],v[k],&t.hi[0]);
hd[k+iih] = t.hi[0];
lis_vector_axpy(-t.hi[0],v[k],v[i1v]);
}
/* h[i+1,i] = ||w|| */
/* v[i+1] = w / h[i+1,i] */
lis_vector_nrm2(v[i1v],&t.hi[0]);
hd[i1+iih] = t.hi[0];
lis_vector_scale(1.0/t.hi[0],v[i1v]);
for(k=1;k<=ii;k++)
{
jj = k-1;
t.hi[0] = hd[jj+iih];
aa.hi[0] = hd[jj+cs]*t.hi[0];
aa.hi[0] += hd[jj+sn]*hd[k+iih];
bb.hi[0] = -hd[jj+sn]*t.hi[0];
bb.hi[0] += hd[jj+cs]*hd[k+iih];
hd[jj+iih] = aa.hi[0];
hd[k+iih] = bb.hi[0];
}
aa.hi[0] = hd[ii+iih];
bb.hi[0] = hd[i1+iih];
a2.hi[0] = aa.hi[0]*aa.hi[0];
b2.hi[0] = bb.hi[0]*bb.hi[0];
rr.hi[0] = sqrt(a2.hi[0]+b2.hi[0]);
if( rr.hi[0]==0.0 ) rr.hi[0]=1.0e-17;
hd[ii+cs] = aa.hi[0]/rr.hi[0];
hd[ii+sn] = bb.hi[0]/rr.hi[0];
s->value[i1] = -hd[ii+sn]*s->value[ii];
s->value[ii] = hd[ii+cs]*s->value[ii];
aa.hi[0] = hd[ii+cs]*hd[ii+iih];
aa.hi[0] += hd[ii+sn]*hd[i1+iih];
hd[ii+iih] = aa.hi[0];
/* convergence check */
nrm2 = fabs(s->value[i1])*bnrm2;
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol2 >= nrm2 ) break;
} while( i<m && iter <maxiter2 );
/* Solve H * Y = S for upper Hessenberg matrix H */
s->value[ii] = s->value[ii]/hd[ii+iih];
for(k=1;k<=ii;k++)
{
jj = ii-k;
t.hi[0] = s->value[jj];
for(j=jj+1;j<=ii;j++)
{
t.hi[0] -= hd[jj+j*h_dim]*s->value[j];
}
s->value[jj] = t.hi[0]/hd[jj+jj*h_dim];
}
/* z = z + y * v */
for(k=0;k<n;k++)
{
z->value[k] = s->value[0]*v[0]->value[k];
}
for(j=1;j<=ii;j++)
{
lis_vector_axpy(s->value[j],v[j],z);
}
/* r = M^-1 * z */
time = lis_wtime();
lis_psolve(solver,z,r);
ptime += lis_wtime()-time;
/* x = x + r */
lis_vector_axpy(1,r,x);
if( tol2 >= nrm2 )
{
solver->iter = iter;
solver->iter2 = iter;
solver->ptime = ptime;
break;
}
for(j=1;j<=i;j++)
{
jj = i1-j+1;
s->value[jj-1] = -hd[jj-1+sn]*s->value[jj];
s->value[jj] = hd[jj-1+cs]*s->value[jj];
}
for(j=0;j<=i1;j++)
{
t.hi[0] = s->value[j];
if( j==0 ) t.hi[0] = t.hi[0]-1.0;
lis_vector_axpy(t.hi[0],v[j],v[0]);
}
}
/* Initial Residual */
z->precision = LIS_PRECISION_QUAD;
solver->options[LIS_OPTIONS_INITGUESS_ZEROS] = LIS_FALSE;
lis_vector_copyex_mn(x,solver->xx);
lis_solver_get_initial_residual(solver,NULL,NULL,v[0],&bnrm2);
tol = solver->tol;
iter2=iter;
while( iter2<maxiter )
{
/* first column of V */
/* v = r / ||r||_2 */
lis_vector_nrm2ex_mm(v[0],&rnorm);
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)one.hi,(LIS_QUAD *)rnorm.hi);
lis_vector_scaleex_mm(tmp,v[0]);
/* s = ||r||_2 e_1 */
lis_vector_set_allex_nm(0.0,s);
s->value[0] = rnorm.hi[0];
s->value_lo[0] = rnorm.lo[0];
i = 0;
do
{
iter2++;
i++;
ii = i-1;
i1 = i;
iiv = i-1;
i1v = i;
iih = (i-1)*h_dim;
/* z = M^-1 * v */
time = lis_wtime();
lis_psolve(solver,v[iiv],z);
ptime += lis_wtime()-time;
/* w = A * z */
lis_matvec(A,z, v[i1v]);
for(k=0;k<i;k++)
{
/* h[k,i] = <w,v[k]> */
/* w = w - h[k,i] * v[k] */
lis_vector_dotex_mmm(v[i1v],v[k],&t);
h[k+iih].hi = t.hi[0];
h[k+iih].lo = t.lo[0];
lis_quad_minus((LIS_QUAD *)t.hi);
lis_vector_axpyex_mmm(t,v[k],v[i1v]);
}
/* h[i+1,i] = ||w|| */
/* v[i+1] = w / h[i+1,i] */
lis_vector_nrm2ex_mm(v[i1v],&t);
h[i1+iih].hi = t.hi[0];
h[i1+iih].lo = t.lo[0];
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)one.hi,(LIS_QUAD *)t.hi);
lis_vector_scaleex_mm(tmp,v[i1v]);
for(k=1;k<=ii;k++)
{
jj = k-1;
t.hi[0] = h[jj+iih].hi;
t.lo[0] = h[jj+iih].lo;
lis_quad_mul((LIS_QUAD *)aa.hi,(LIS_QUAD *)&h[jj+cs],(LIS_QUAD *)t.hi);
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj+sn],(LIS_QUAD *)&h[k+iih]);
lis_quad_add((LIS_QUAD *)aa.hi,(LIS_QUAD *)aa.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)bb.hi,(LIS_QUAD *)&h[jj+sn],(LIS_QUAD *)t.hi);
lis_quad_minus((LIS_QUAD *)bb.hi);
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj+cs],(LIS_QUAD *)&h[k+iih]);
lis_quad_add((LIS_QUAD *)bb.hi,(LIS_QUAD *)bb.hi,(LIS_QUAD *)tmp.hi);
h[jj+iih].hi = aa.hi[0];
h[jj+iih].lo = aa.lo[0];
h[k+iih].hi = bb.hi[0];
h[k+iih].lo = bb.lo[0];
}
aa.hi[0] = h[ii+iih].hi;
aa.lo[0] = h[ii+iih].lo;
bb.hi[0] = h[i1+iih].hi;
bb.lo[0] = h[i1+iih].lo;
lis_quad_sqr((LIS_QUAD *)a2.hi,(LIS_QUAD *)aa.hi);
lis_quad_sqr((LIS_QUAD *)b2.hi,(LIS_QUAD *)bb.hi);
lis_quad_add((LIS_QUAD *)rr.hi,(LIS_QUAD *)a2.hi,(LIS_QUAD *)b2.hi);
lis_quad_sqrt((LIS_QUAD *)rr.hi,(LIS_QUAD *)rr.hi);
lis_quad_div((LIS_QUAD *)&h[ii+cs],(LIS_QUAD *)aa.hi,(LIS_QUAD *)rr.hi);
lis_quad_div((LIS_QUAD *)&h[ii+sn],(LIS_QUAD *)bb.hi,(LIS_QUAD *)rr.hi);
tmp.hi[0] = s->value[ii];
tmp.lo[0] = s->value_lo[ii];
lis_quad_mul((LIS_QUAD *)aa.hi,(LIS_QUAD *)&h[ii+sn],(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)bb.hi,(LIS_QUAD *)&h[ii+cs],(LIS_QUAD *)tmp.hi);
lis_quad_minus((LIS_QUAD *)aa.hi);
s->value[i1] = aa.hi[0];
s->value_lo[i1] = aa.lo[0];
s->value[ii] = bb.hi[0];
s->value_lo[ii] = bb.lo[0];
lis_quad_mul((LIS_QUAD *)aa.hi,(LIS_QUAD *)&h[ii+cs],(LIS_QUAD *)&h[ii+iih]);
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[ii+sn],(LIS_QUAD *)&h[i1+iih]);
lis_quad_add((LIS_QUAD *)aa.hi,(LIS_QUAD *)aa.hi,(LIS_QUAD *)tmp.hi);
h[ii+iih].hi = aa.hi[0];
h[ii+iih].lo = aa.lo[0];
/* convergence check */
nrm2 = fabs(s->value[i1])*bnrm2;
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter2] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 ) break;
} while( i<m && iter2 <maxiter );
/* Solve H * Y = S for upper Hessenberg matrix H */
tmp.hi[0] = s->value[ii];
tmp.lo[0] = s->value_lo[ii];
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[ii+iih]);
s->value[ii] = tmp.hi[0];
s->value_lo[ii] = tmp.lo[0];
for(k=1;k<=ii;k++)
{
jj = ii-k;
t.hi[0] = s->value[jj];
t.lo[0] = s->value_lo[jj];
for(j=jj+1;j<=ii;j++)
{
tmp.hi[0] = s->value[j];
tmp.lo[0] = s->value_lo[j];
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj+j*h_dim]);
lis_quad_sub((LIS_QUAD *)t.hi,(LIS_QUAD *)t.hi,(LIS_QUAD *)tmp.hi);
}
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)t.hi,(LIS_QUAD *)&h[jj+jj*h_dim]);
s->value[jj] = tmp.hi[0];
s->value_lo[jj] = tmp.lo[0];
}
/* z = z + y * v */
for(k=0;k<n;k++)
{
aa.hi[0] = s->value[0];
aa.lo[0] = s->value_lo[0];
bb.hi[0] = v[0]->value[k];
bb.lo[0] = v[0]->value_lo[k];
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)aa.hi,(LIS_QUAD *)bb.hi);
z->value[k] = tmp.hi[0];
z->value_lo[k] = tmp.lo[0];
}
for(j=1;j<=ii;j++)
{
aa.hi[0] = s->value[j];
aa.lo[0] = s->value_lo[j];
lis_vector_axpyex_mmm(aa,v[j],z);
}
/* r = M^-1 * z */
time = lis_wtime();
lis_psolve(solver,z,r);
ptime += lis_wtime()-time;
/* x = x + r */
lis_vector_axpyex_mmm(one,r,x);
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
solver->ptime = ptime;
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
for(j=1;j<=i;j++)
{
jj = i1-j+1;
tmp.hi[0] = s->value[jj];
tmp.lo[0] = s->value_lo[jj];
lis_quad_mul((LIS_QUAD *)aa.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj-1+sn]);
lis_quad_mul((LIS_QUAD *)bb.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)&h[jj-1+cs]);
lis_quad_minus((LIS_QUAD *)aa.hi);
s->value[jj-1] = aa.hi[0];
s->value_lo[jj-1] = aa.lo[0];
s->value[jj] = bb.hi[0];
s->value_lo[jj] = bb.lo[0];
}
for(j=0;j<=i1;j++)
{
t.hi[0] = s->value[j];
t.lo[0] = s->value_lo[j];
if( j==0 )
{
lis_quad_sub((LIS_QUAD *)t.hi,(LIS_QUAD *)t.hi,(LIS_QUAD *)one.hi);
}
lis_vector_axpyex_mmm(t,v[j],v[0]);
}
}
solver->retcode = LIS_MAXITER;
solver->iter = iter2+1;
solver->iter2 = iter;
solver->resid = nrm2;
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT main(LIS_INT argc, char* argv[])
{
LIS_MATRIX A,A0;
LIS_VECTOR b,x;
LIS_INT nprocs,my_rank;
int int_nprocs,int_my_rank;
LIS_INT nthreads, maxthreads;
LIS_INT nnz;
LIS_INT i,n,np;
LIS_INT block;
LIS_INT is,ie;
LIS_INT err,iter,matrix_type;
double time,time2,nnzs,nnzap,nnzt;
LIS_SCALAR val;
double commtime,comptime,flops;
char path[1024];
FILE *file;
LIS_DEBUG_FUNC_IN;
lis_initialize(&argc, &argv);
#ifdef USE_MPI
MPI_Comm_size(MPI_COMM_WORLD,&int_nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&int_my_rank);
nprocs = int_nprocs;
my_rank = int_my_rank;
#else
nprocs = 1;
my_rank = 0;
#endif
if( argc < 3 )
{
if( my_rank==0 )
{
printf("Usage: %s matrix_filename_list iter [block] \n", argv[0]);
}
lis_finalize();
exit(0);
}
file = fopen(argv[1], "r");
if( file==NULL ) CHKERR(1);
iter = atoi(argv[2]);
if (argv[3] == NULL) {
block = 2;
}
else {
block = atoi(argv[3]);
}
if( iter<=0 )
{
#ifdef _LONG__LONG
printf("iter=%lld <= 0\n",iter);
#else
printf("iter=%d <= 0\n",iter);
#endif
CHKERR(1);
}
if( my_rank==0 )
{
printf("\n");
#ifdef _LONG__LONG
printf("number of processes = %lld\n",nprocs);
#else
printf("number of processes = %d\n",nprocs);
#endif
}
#ifdef _OPENMP
if( my_rank==0 )
{
nthreads = omp_get_num_procs();
maxthreads = omp_get_max_threads();
#ifdef _LONG__LONG
printf("max number of threads = %lld\n", nthreads);
printf("number of threads = %lld\n", maxthreads);
#else
printf("max number of threads = %d\n", nthreads);
printf("number of threads = %d\n", maxthreads);
#endif
}
#else
nthreads = 1;
maxthreads = 1;
#endif
/* create matrix and vectors */
while( fscanf(file, "%s\n", path)==1 )
{
if( my_rank==0 )
{
printf("matrix_filename = %s\n", path);
}
lis_matrix_create(LIS_COMM_WORLD,&A0);
err = lis_input(A0,NULL,NULL,path);
if( err ) CHKERR(err);
n = A0->n;
nnz = A0->nnz;
np = A0->np-n;
#ifdef USE_MPI
MPI_Allreduce(&nnz,&i,1,LIS_MPI_INT,MPI_SUM,A0->comm);
nnzap = (double)i / (double)nprocs;
nnzt = ((double)nnz -nnzap)*((double)nnz -nnzap);
nnz = i;
MPI_Allreduce(&nnzt,&nnzs,1,MPI_DOUBLE,MPI_SUM,A0->comm);
nnzs = (nnzs / (double)nprocs)/nnzap;
MPI_Allreduce(&np,&i,1,LIS_MPI_INT,MPI_SUM,A0->comm);
np = i;
#endif
if( my_rank==0 )
{
#ifdef _LONG__LONG
printf("block size of BSR and BSC = %lld x %lld\n",block,block);
printf("number of iterations = %lld\n\n",iter);
#else
printf("block size of BSR and BSC = %d x %d\n",block,block);
printf("number of iterations = %d\n\n",iter);
#endif
}
err = lis_vector_duplicate(A0,&x);
if( err ) CHKERR(err);
err = lis_vector_duplicate(A0,&b);
if( err ) CHKERR(err);
lis_matrix_get_range(A0,&is,&ie);
for(i=0;i<n;i++)
{
err = lis_vector_set_value(LIS_INS_VALUE,i+is,1.0,x);
}
/*
MPI version of VBR is not implemented.
DNS is also excluded to reduce memory usage.
*/
for (matrix_type=1;matrix_type<11;matrix_type++)
{
if ( nprocs>1 && matrix_type==9 ) continue;
lis_matrix_duplicate(A0,&A);
lis_matrix_set_type(A,matrix_type);
err = lis_matrix_convert(A0,A);
if( err ) CHKERR(err);
if( my_rank==0 )
{
if( A->matrix_type==LIS_MATRIX_BSR || A->matrix_type==LIS_MATRIX_BSC )
{
A->bnr = block;
A->bnc = block;
}
}
comptime = 0.0;
commtime = 0.0;
for(i=0;i<iter;i++)
{
#ifdef USE_MPI
MPI_Barrier(A->comm);
time = lis_wtime();
lis_send_recv(A->commtable,x->value);
commtime += lis_wtime() - time;
#endif
time2 = lis_wtime();
lis_matvec(A,x,b);
comptime += lis_wtime() - time2;
}
lis_vector_nrm2(b,&val);
if( my_rank==0 )
{
flops = 2.0*nnz*iter*1.0e-6 / comptime;
#ifdef USE_MPI
#ifdef _LONG__DOUBLE
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#endif
#else
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#endif
#endif
#else
#ifdef _LONG__DOUBLE
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#endif
#else
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#endif
#endif
#endif
}
lis_matrix_destroy(A);
}
lis_matrix_destroy(A0);
lis_vector_destroy(b);
lis_vector_destroy(x);
}
fclose(file);
lis_finalize();
LIS_DEBUG_FUNC_OUT;
return 0;
}
LIS_INT lis_ecr(LIS_ESOLVER esolver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_SCALAR evalue;
LIS_INT emaxiter;
LIS_REAL tol;
LIS_INT iter,i,j,output;
LIS_INT nprocs,my_rank;
LIS_REAL nrm2,resid;
LIS_SCALAR lshift;
LIS_VECTOR r,p,Ax,Ar,Ap;
LIS_SCALAR alpha, beta;
LIS_SCALAR rAp, rp, ApAp, pAp, pp, ArAp, pAr;
double times,itimes,ptimes,p_c_times,p_i_times;
A = esolver->A;
x = esolver->x;
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
tol = esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN];
output = esolver->options[LIS_EOPTIONS_OUTPUT];
lshift = esolver->lshift;
if( A->my_rank==0 ) printf("local shift = %e\n", lshift);
if (lshift != 0) lis_matrix_shift_diagonal(A, lshift);
r = esolver->work[0];
p = esolver->work[1];
Ax = esolver->work[2];
Ar = esolver->work[3];
Ap = esolver->work[4];
lis_vector_set_all(1.0,x);
lis_vector_nrm2(x,&nrm2);
lis_vector_scale(1/nrm2,x);
lis_matvec(A,x,Ax);
lis_vector_dot(x,Ax,&evalue);
lis_vector_axpyz(-evalue,x,Ax,r);
lis_vector_scale(-1.0,r);
lis_vector_copy(r,p);
lis_matvec(A,p,Ap);
iter=0;
while (iter<emaxiter)
{
iter = iter + 1;
lis_vector_dot(r,Ap,&rAp);
lis_vector_dot(r,p,&rp);
lis_vector_dot(Ap,Ap,&ApAp);
lis_vector_dot(p,Ap,&pAp);
lis_vector_dot(p,p,&pp);
alpha = (rAp - evalue * rp) / (ApAp - evalue * (2.0 * pAp - evalue * pp));
lis_vector_axpy(alpha,p,x);
lis_matvec(A,x,Ax);
lis_vector_dot(x,Ax,&evalue);
lis_vector_nrm2(x, &nrm2);
evalue = evalue / (nrm2 * nrm2);
lis_vector_axpyz(-evalue,x,Ax,r);
lis_vector_scale(-1.0,r);
lis_matvec(A,r,Ar);
lis_vector_dot(Ar,Ap,&ArAp);
lis_vector_dot(p,Ar,&pAr);
lis_vector_dot(r,Ap,&rAp);
lis_vector_dot(r,p,&rp);
beta = - (ArAp - evalue * ((pAr + rAp) - evalue * rp))/ (ApAp - evalue * (2.0 * pAp - evalue * pp));
lis_vector_xpay(r,beta,p);
lis_vector_nrm2(r,&nrm2);
resid = fabs(nrm2 / (evalue));
if( output )
{
if( output & LIS_EPRINT_MEM ) esolver->residual[iter] = resid;
if( output & LIS_EPRINT_OUT && A->my_rank==0 ) printf("iter: %5d residual = %e\n", iter, resid);
}
if (resid<tol) break;
}
esolver->iter = iter;
esolver->resid = resid;
esolver->evalue[0] = evalue;
if (lshift != 0) lis_matrix_shift_diagonal(A, -lshift);
if (resid<tol)
{
esolver->retcode = LIS_SUCCESS;
return LIS_SUCCESS;
}
else
{
esolver->retcode = LIS_MAXITER;
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_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_bicgsafe_switch(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r, rtld, rhat, p, ptld, phat;
LIS_VECTOR t, ttld, that, t0, t0hat;
LIS_VECTOR y, w, u, z;
LIS_QUAD_PTR alpha, beta, rho, rho_old;
LIS_QUAD_PTR qsi, eta, one;
LIS_QUAD_PTR tmp, tmpdot[5];
LIS_REAL bnrm2, nrm2, tol, tol2;
LIS_INT iter,maxiter,output,conv;
LIS_INT iter2,maxiter2;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
maxiter2 = solver->options[LIS_OPTIONS_SWITCH_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
tol2 = solver->params[LIS_PARAMS_SWITCH_RESID-LIS_OPTIONS_LEN];
ptime = 0.0;
rtld = solver->work[0];
r = solver->work[1];
rhat = solver->work[2];
p = solver->work[3];
ptld = solver->work[4];
phat = solver->work[5];
t = solver->work[6];
ttld = solver->work[7];
that = solver->work[8];
t0 = solver->work[9];
t0hat = solver->work[10];
y = solver->work[11];
w = solver->work[12];
u = solver->work[13];
z = solver->work[14];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(qsi,4,1);
LIS_QUAD_SCALAR_MALLOC(eta,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[0],7,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[1],8,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[2],9,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[3],10,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[4],11,1);
LIS_QUAD_SCALAR_MALLOC(one,13,1);
rho_old.hi[0] = 1.0;
rho_old.lo[0] = 0.0;
alpha.hi[0] = 1.0;
alpha.lo[0] = 0.0;
qsi.hi[0] = 1.0;
qsi.lo[0] = 0.0;
one.hi[0] = -1.0;
one.lo[0] = 0.0;
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol2 = solver->tol_switch;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0, ttld);
lis_vector_set_allex_nm(0.0, ptld);
lis_vector_set_allex_nm(0.0, p);
lis_vector_set_allex_nm(0.0, u);
lis_vector_set_allex_nm(0.0, t);
lis_vector_set_allex_nm(0.0, t0);
for( iter=1; iter<=maxiter2; iter++ )
{
/* rho = <rtld,r> */
lis_vector_dot(rtld,r,&rho.hi[0]);
/* test breakdown */
if( rho.hi[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
beta.hi[0] = (rho.hi[0] / rho_old.hi[0]) * (alpha.hi[0] / qsi.hi[0]);
/* w = ttld + beta*ptld */
lis_vector_axpyz(beta.hi[0],ptld,ttld,w);
/* rhat = M^-1 * r */
time = lis_wtime();
lis_psolve(solver, r, rhat);
ptime += lis_wtime()-time;
/* p = rhat + beta*(p - u) */
lis_vector_axpy(-1,u,p);
lis_vector_xpay(rhat,beta.hi[0],p);
/* ptld = A * p */
lis_matvec(A,p,ptld);
/* tmpdot[0] = <rtld,ptld> */
lis_vector_dot(rtld,ptld,&tmpdot[0].hi[0]);
/* test breakdown */
/* */
/* alpha = rho / tmpdot[0] */
alpha.hi[0] = rho.hi[0] / tmpdot[0].hi[0];
/* y = t - r + alpha*(-w + ptld) */
lis_vector_axpyz(-1,w,ptld,y);
lis_vector_xpay(t,alpha.hi[0],y);
lis_vector_axpy(-1,r,y);
/* t = r - alpha*ptld */
lis_vector_axpyz(-alpha.hi[0],ptld,r,t);
/* that = M^-1 * t */
/* phat = M^-1 * ptld */
/* t0hat = M^-1 * t0 */
time = lis_wtime();
lis_psolve(solver, t, that);
lis_psolve(solver, ptld, phat);
lis_psolve(solver, t0, t0hat);
ptime += lis_wtime()-time;
/* ttld = A * that */
lis_matvec(A,that,ttld);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <ttld,t> */
/* tmpdot[2] = <y,t> */
/* tmpdot[3] = <ttld,y> */
/* tmpdot[4] = <ttld,ttld> */
lis_vector_dot(y,y,&tmpdot[0].hi[0]);
lis_vector_dot(ttld,t,&tmpdot[1].hi[0]);
lis_vector_dot(y,t,&tmpdot[2].hi[0]);
lis_vector_dot(ttld,y,&tmpdot[3].hi[0]);
lis_vector_dot(ttld,ttld,&tmpdot[4].hi[0]);
if(iter==1)
{
qsi.hi[0] = tmpdot[1].hi[0] / tmpdot[4].hi[0];
eta.hi[0] = 0.0;
}
else
{
tmp.hi[0] = tmpdot[4].hi[0]*tmpdot[0].hi[0] - tmpdot[3].hi[0]*tmpdot[3].hi[0];
qsi.hi[0] = (tmpdot[0].hi[0]*tmpdot[1].hi[0] - tmpdot[2].hi[0]*tmpdot[3].hi[0]) / tmp.hi[0];
eta.hi[0] = (tmpdot[4].hi[0]*tmpdot[2].hi[0] - tmpdot[3].hi[0]*tmpdot[1].hi[0]) / tmp.hi[0];
}
/* u = qsi*phat + eta*(t0hat - rhat + beta*u) */
lis_vector_xpay(t0hat,beta.hi[0],u);
lis_vector_axpy(-1,rhat,u);
lis_vector_scale(eta.hi[0],u);
lis_vector_axpy(qsi.hi[0],phat,u);
/* z = qsi*rhat + eta*z - alpha*u */
lis_vector_scale(eta.hi[0],z);
lis_vector_axpy(qsi.hi[0],rhat,z);
lis_vector_axpy(-alpha.hi[0],u,z);
/* x = x + alpha*p + z */
lis_vector_axpy(alpha.hi[0],p,x);
lis_vector_axpy(1,z,x);
/* r = t - eta*y - qsi*ttld */
lis_vector_axpyz(-eta.hi[0],y,t,r);
lis_vector_axpy(-qsi.hi[0],ttld,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol2 >= nrm2 )
{
solver->iter = iter;
solver->iter2 = iter;
solver->ptime = ptime;
break;
}
lis_vector_copy(t,t0);
rho_old.hi[0] = rho.hi[0];
}
r->precision = LIS_PRECISION_QUAD;
p->precision = LIS_PRECISION_QUAD;
t->precision = LIS_PRECISION_QUAD;
t0->precision = LIS_PRECISION_QUAD;
ptld->precision = LIS_PRECISION_QUAD;
that->precision = LIS_PRECISION_QUAD;
solver->options[LIS_OPTIONS_INITGUESS_ZEROS] = LIS_FALSE;
lis_vector_copyex_mn(x,solver->xx);
rho_old.hi[0] = 1.0;
alpha.hi[0] = 1.0;
qsi.hi[0] = 1.0;
one.hi[0] = -1.0;
/* Initial Residual */
lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2);
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0, ttld);
lis_vector_set_allex_nm(0.0, ptld);
lis_vector_set_allex_nm(0.0, p);
lis_vector_set_allex_nm(0.0, u);
lis_vector_set_allex_nm(0.0, t);
lis_vector_set_allex_nm(0.0, t0);
for( iter2=iter+1; iter2<=maxiter; iter2++ )
{
/* rho = <rtld,r> */
lis_vector_dotex_mmm(rtld,r,&rho);
/* test breakdown */
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)alpha.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,(LIS_QUAD *)tmp.hi);
/* w = ttld + beta*ptld */
lis_vector_axpyzex_mmmm(beta,ptld,ttld,w);
/* rhat = M^-1 * r */
time = lis_wtime();
lis_psolve(solver, r, rhat);
ptime += lis_wtime()-time;
/* p = rhat + beta*(p - u) */
lis_vector_axpyex_mmm(one,u,p);
lis_vector_xpayex_mmm(rhat,beta,p);
/* ptld = A * p */
lis_matvec(A,p,ptld);
/* tmpdot[0] = <rtld,ptld> */
lis_vector_dotex_mmm(rtld,ptld,&tmpdot[0]);
/* test breakdown */
/* */
/* alpha = rho / tmpdot[0] */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)tmpdot[0].hi);
/* y = t - r + alpha*(-w + ptld) */
lis_vector_axpyzex_mmmm(one,w,ptld,y);
lis_vector_xpayex_mmm(t,alpha,y);
lis_vector_axpyex_mmm(one,r,y);
/* t = r - alpha*ptld */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyzex_mmmm(alpha,ptld,r,t);
/* that = M^-1 * t */
/* phat = M^-1 * ptld */
/* t0hat = M^-1 * t0 */
time = lis_wtime();
lis_psolve(solver, t, that);
lis_psolve(solver, ptld, phat);
lis_psolve(solver, t0, t0hat);
ptime += lis_wtime()-time;
/* ttld = A * that */
lis_matvec(A,that,ttld);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <ttld,t> */
/* tmpdot[2] = <y,t> */
/* tmpdot[3] = <ttld,y> */
/* tmpdot[4] = <ttld,ttld> */
lis_vector_dotex_mmm(y,y,&tmpdot[0]);
lis_vector_dotex_mmm(ttld,t,&tmpdot[1]);
lis_vector_dotex_mmm(y,t,&tmpdot[2]);
lis_vector_dotex_mmm(ttld,y,&tmpdot[3]);
lis_vector_dotex_mmm(ttld,ttld,&tmpdot[4]);
if(iter==1)
{
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[1].hi,(LIS_QUAD *)tmpdot[4].hi);
eta.hi[0] = 0.0;
eta.lo[0] = 0.0;
}
else
{
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_sqr((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[2].hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)eta.hi);
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[2].hi);
lis_quad_mul((LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[3].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_sub((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_div((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmp.hi);
}
/* u = qsi*phat + eta*(t0hat - rhat + beta*u) */
lis_vector_xpayex_mmm(t0hat,beta,u);
lis_vector_axpyex_mmm(one,rhat,u);
lis_vector_scaleex_mm(eta,u);
lis_vector_axpyex_mmm(qsi,phat,u);
/* z = qsi*rhat + eta*z - alpha*u */
lis_vector_scaleex_mm(eta,z);
lis_vector_axpyex_mmm(qsi,rhat,z);
lis_vector_axpyex_mmm(alpha,u,z);
/* x = x + alpha*p + z */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(alpha,p,x);
lis_vector_axpyex_mmm(one,z,x);
lis_quad_minus((LIS_QUAD *)one.hi);
/* r = t - eta*y - qsi*ttld */
lis_quad_minus((LIS_QUAD *)eta.hi);
lis_quad_minus((LIS_QUAD *)qsi.hi);
lis_vector_axpyzex_mmmm(eta,y,t,r);
lis_vector_axpyex_mmm(qsi,ttld,r);
lis_quad_minus((LIS_QUAD *)eta.hi);
lis_quad_minus((LIS_QUAD *)qsi.hi);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter2] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
lis_vector_copyex_mm(t,t0);
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->iter2 = iter2;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_bicgsafe_quad(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r, rtld, rhat, p, ptld;
LIS_VECTOR t, ttld;
LIS_VECTOR y, v, u, utld, z;
LIS_QUAD_PTR alpha, beta, rho, rho_old;
LIS_QUAD_PTR qsi, eta;
LIS_QUAD_PTR tmp, tmpdot[5],one;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
rtld = solver->work[0];
r = solver->work[1];
rhat = solver->work[2];
p = solver->work[3];
ptld = solver->work[4];
t = solver->work[5];
ttld = solver->work[6];
y = solver->work[7];
v = solver->work[8];
u = solver->work[9];
z = solver->work[10];
utld = solver->work[11];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(qsi,4,1);
LIS_QUAD_SCALAR_MALLOC(eta,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[0],7,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[1],8,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[2],9,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[3],10,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[4],11,1);
LIS_QUAD_SCALAR_MALLOC(one,13,1);
rho_old.hi[0] = 1.0;
rho_old.lo[0] = 0.0;
alpha.hi[0] = 1.0;
alpha.lo[0] = 0.0;
qsi.hi[0] = 1.0;
qsi.lo[0] = 0.0;
one.hi[0] = -1.0;
one.lo[0] = 0.0;
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0,p);
lis_vector_set_allex_nm(0.0,u);
lis_vector_set_allex_nm(0.0,ptld);
lis_vector_set_allex_nm(0.0,utld);
for( iter=1; iter<=maxiter; iter++ )
{
/* rho = <rtld,r> */
lis_vector_dotex_mmm(rtld,r,&rho);
/* test breakdown */
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)alpha.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,(LIS_QUAD *)tmp.hi);
/* rhat = M^-1 * r */
/* v = A * rhat */
time = lis_wtime();
lis_psolve(solver, r, rhat);
ptime += lis_wtime()-time;
lis_matvec(A,rhat,v);
/* p = rhat + beta*(p - u) */
lis_vector_axpyex_mmm(one,u,p);
lis_vector_xpayex_mmm(rhat,beta,p);
/* ptld = v + beta*(ptld - utld) */
lis_vector_axpyex_mmm(one,utld,ptld);
lis_vector_xpayex_mmm(v,beta,ptld);
/* tmpdot[0] = <rtld,ptld> */
lis_vector_dotex_mmm(rtld,ptld,&tmpdot[0]);
/* test breakdown */
/* */
/* alpha = rho / tmpdot[0] */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)tmpdot[0].hi);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <v,r> */
/* tmpdot[2] = <y,r> */
/* tmpdot[3] = <v,y> */
/* tmpdot[4] = <v,v> */
lis_vector_dotex_mmm(y,y,&tmpdot[0]);
lis_vector_dotex_mmm(v,r,&tmpdot[1]);
lis_vector_dotex_mmm(y,r,&tmpdot[2]);
lis_vector_dotex_mmm(v,y,&tmpdot[3]);
lis_vector_dotex_mmm(v,v,&tmpdot[4]);
if(iter==1)
{
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[1].hi,(LIS_QUAD *)tmpdot[4].hi);
eta.hi[0] = 0.0;
eta.lo[0] = 0.0;
}
else
{
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_sqr((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[2].hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)eta.hi);
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[2].hi);
lis_quad_mul((LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[3].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_sub((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_div((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmp.hi);
}
/* t = qsi*ptld + eta*y */
lis_vector_copyex_mm(y,t);
lis_vector_scaleex_mm(eta,t);
lis_vector_axpyex_mmm(qsi,ptld,t);
/* ttld = M^-1 * t */
time = lis_wtime();
lis_psolve(solver, t, ttld);
ptime += lis_wtime()-time;
/* u = ttld + eta*beta*u */
/* utld = A * u */
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)beta.hi);
lis_vector_xpayex_mmm(ttld,tmp,u);
lis_matvec(A,u,utld);
/* z = qsi*rhat + eta*z - alpha*u */
lis_vector_scaleex_mm(eta,z);
lis_vector_axpyex_mmm(qsi,rhat,z);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,u,z);
/* y = qsi*v + eta*y - alpha*utld */
lis_vector_scaleex_mm(eta,y);
lis_vector_axpyex_mmm(qsi,v,y);
lis_vector_axpyex_mmm(alpha,utld,y);
lis_quad_minus((LIS_QUAD *)alpha.hi);
/* x = x + alpha*p + z */
lis_vector_axpyex_mmm(alpha,p,x);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(one,z,x);
lis_quad_minus((LIS_QUAD *)one.hi);
/* r = r - alpha*ptld - y */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,ptld,r);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(one,y,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_bicgsafe(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r, rtld, mr, amr, t, mt, p, ap;
LIS_VECTOR y, u, au, z;
LIS_SCALAR alpha, beta;
LIS_REAL rho, rho_old;
LIS_SCALAR qsi, eta;
LIS_SCALAR tmp, tmpdot[5];
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
rtld = solver->work[0];
r = solver->work[1];
mr = solver->work[2];
amr = solver->work[3];
p = solver->work[4];
ap = solver->work[5];
t = solver->work[6];
mt = solver->work[7];
y = solver->work[8];
u = solver->work[9];
z = solver->work[10];
au = solver->work[11];
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
time = lis_wtime();
lis_psolve(solver, r, mr);
ptime += lis_wtime()-time;
lis_matvec(A,mr,amr);
lis_vector_dot(rtld,r,&rho_old);
lis_vector_copy(amr,ap);
lis_vector_copy(mr,p);
beta = 0.0;
for( iter=1; iter<=maxiter; iter++ )
{
/* tmpdot[0] = <rtld,ap> */
/* alpha = rho_old / tmpdot[0] */
lis_vector_dot(rtld,ap,&tmpdot[0]);
alpha = rho_old / tmpdot[0];
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <amr,r> */
/* tmpdot[2] = <y,r> */
/* tmpdot[3] = <amr,y> */
/* tmpdot[4] = <amr,amr> */
lis_vector_dot(y,y,&tmpdot[0]);
lis_vector_dot(amr,r,&tmpdot[1]);
lis_vector_dot(y,r,&tmpdot[2]);
lis_vector_dot(amr,y,&tmpdot[3]);
lis_vector_dot(amr,amr,&tmpdot[4]);
if(iter==1)
{
qsi = tmpdot[1] / tmpdot[4];
eta = 0.0;
}
else
{
tmp = tmpdot[4]*tmpdot[0] - tmpdot[3]*tmpdot[3];
qsi = (tmpdot[0]*tmpdot[1] - tmpdot[2]*tmpdot[3]) / tmp;
eta = (tmpdot[4]*tmpdot[2] - tmpdot[3]*tmpdot[1]) / tmp;
}
/* t = qsi*ap + eta*y */
lis_vector_copy(y,t);
lis_vector_scale(eta,t);
lis_vector_axpy(qsi,ap,t);
/* mt = M^-1 * t */
time = lis_wtime();
lis_psolve(solver, t, mt);
ptime += lis_wtime()-time;
/* u = mt + eta*beta*u */
/* au = A * u */
lis_vector_xpay(mt,eta*beta,u);
lis_matvec(A,u,au);
/* z = qsi*mr + eta*z - alpha*u */
lis_vector_scale(eta,z);
lis_vector_axpy(qsi,mr,z);
lis_vector_axpy(-alpha,u,z);
/* y = qsi*amr + eta*y - alpha*au */
lis_vector_scale(eta,y);
lis_vector_axpy(qsi,amr,y);
lis_vector_axpy(-alpha,au,y);
/* x = x + alpha*p + z */
lis_vector_axpy(alpha,p,x);
lis_vector_axpy(1.0,z,x);
/* r = r - alpha*ap - y */
lis_vector_axpy(-alpha,ap,r);
lis_vector_axpy(-1.0,y,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* rho = <rtld,r> */
lis_vector_dot(rtld,r,&rho);
if( rho==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
beta = (rho / rho_old) * (alpha / qsi);
/* mr = M^-1 * r */
/* amr = A * mr */
time = lis_wtime();
lis_psolve(solver, r, mr);
ptime += lis_wtime()-time;
lis_matvec(A,mr,amr);
/* p = mr + beta*(p - u) */
/* ap = amr + beta*(ap - au) */
lis_vector_axpy(-1.0,u,p);
lis_vector_xpay(mr,beta,p);
lis_vector_axpy(-1.0,au,ap);
lis_vector_xpay(amr,beta,ap);
rho_old = rho;
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_solver_get_initial_residual(LIS_SOLVER solver, LIS_PRECON M, LIS_VECTOR t, LIS_VECTOR r, LIS_SCALAR *bnrm2)
{
LIS_INT output,conv;
#ifdef USE_QUAD_PRECISION
LIS_INT i;
#endif
LIS_MATRIX A;
LIS_VECTOR x,b,p,xx;
LIS_SCALAR nrm2;
LIS_REAL tol,tol_w,tol_switch;
LIS_DEBUG_FUNC_IN;
A = solver->A;
b = solver->b;
x = solver->x;
xx = solver->xx;
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = 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];
tol_switch = solver->params[LIS_PARAMS_SWITCH_RESID-LIS_OPTIONS_LEN];
/* Initial Residual */
if( M==NULL )
{
p = r;
}
else
{
p = t;
}
if( !solver->options[LIS_OPTIONS_INITGUESS_ZEROS] )
{
#ifndef USE_QUAD_PRECISION
lis_matvec(A,x,p); /* p = Ax */
lis_vector_xpay(b,-1,p); /* p = b - p */
#else
if( solver->precision==LIS_PRECISION_DOUBLE )
{
lis_matvec(A,x,p); /* p = Ax */
lis_vector_xpay(b,-1,p); /* p = b - p */
}
else
{
lis_matvec(A,xx,p); /* p = Ax */
lis_vector_xpay(b,-1,p); /* p = b - p */
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for(i=0;i<A->n;i++)
{
p->value_lo[i] = 0.0;
}
}
#endif
}
else
{
#ifndef USE_QUAD_PRECISION
lis_vector_copy(b,p);
#else
if( solver->precision==LIS_PRECISION_DOUBLE )
{
lis_vector_copy(b,p);
}
else
{
lis_vector_copyex_nm(b,p);
}
#endif
}
switch(conv)
{
case LIS_CONV_COND_NRM2_R:
lis_vector_nrm2(p,&nrm2);
*bnrm2 = nrm2;
solver->tol = tol;
solver->tol_switch = tol_switch;
break;
case LIS_CONV_COND_NRM2_B:
lis_vector_nrm2(p,&nrm2);
lis_vector_nrm2(b,bnrm2);
solver->tol = tol;
solver->tol_switch = tol_switch;
break;
case LIS_CONV_COND_NRM1_B:
lis_vector_nrm1(p,&nrm2);
lis_vector_nrm1(b,bnrm2);
solver->tol = *bnrm2*tol_w + tol;
solver->tol_switch = *bnrm2*tol_w + tol_switch;
break;
}
if( *bnrm2 == 0.0 )
{
*bnrm2 = 1.0;
}
else
{
*bnrm2 = 1.0 / *bnrm2;
}
solver->bnrm = *bnrm2;
nrm2 = nrm2 * *bnrm2;
if( output && (r->precision==LIS_PRECISION_QUAD && solver->precision!=LIS_PRECISION_SWITCH) )
{
if( output & LIS_PRINT_MEM ) solver->residual[0] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) printf("iter: %5d residual = %e\n", 0, nrm2);
}
if( nrm2 <= solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN] )
{
solver->retcode = LIS_SUCCESS;
solver->iter = 1;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_FAILS;
}
if( M!=NULL )
{
/* r = M^-1 * p */
lis_psolve(solver, p, r);
}
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_jacobi(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR b,x;
LIS_VECTOR d,r,t,s;
LIS_REAL bnrm2,nrm2,tol;
LIS_INT iter,maxiter,output;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
b = solver->b;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
ptime = 0.0;
r = solver->work[0];
t = solver->work[1];
s = solver->work[2];
d = solver->work[3];
lis_vector_nrm2(b,&bnrm2);
bnrm2 = 1.0 / bnrm2;
lis_matrix_get_diagonal(A,d);
lis_vector_reciprocal(d);
for( iter=1; iter<=maxiter; iter++ )
{
/* x += D^{-1}(b - Ax) */
time = lis_wtime();
lis_psolve(solver,x,s);
ptime += lis_wtime() - time;
lis_matvec(A,s,t);
/* lis_matvec(A,x,t);*/
lis_vector_axpyz(-1,t,b,r);
lis_vector_nrm2(r,&nrm2);
lis_vector_pmul(r,d,r);
lis_vector_axpy(1,r,x);
/* convergence check */
nrm2 = nrm2 * bnrm2;
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 )
{
time = lis_wtime();
lis_psolve(solver,x,s);
ptime += lis_wtime() - time;
lis_vector_copy(s,x);
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
}
lis_psolve(solver,x,s);
lis_vector_copy(s,x);
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT main(LIS_INT argc, char* argv[])
{
LIS_MATRIX A0,A;
LIS_VECTOR x,b,u;
LIS_SOLVER solver;
LIS_INT m,n,nn,nnz;
LIS_INT i,j,ii,jj,ctr;
LIS_INT is,ie;
LIS_INT nprocs,my_rank;
int int_nprocs,int_my_rank;
LIS_INT nsol;
LIS_INT err,iter,mtype,iter_double,iter_quad;
double time,itime,ptime,p_c_time,p_i_time;
LIS_REAL resid;
char solvername[128];
LIS_INT *ptr,*index;
LIS_SCALAR *value;
LIS_DEBUG_FUNC_IN;
lis_initialize(&argc, &argv);
#ifdef USE_MPI
MPI_Comm_size(MPI_COMM_WORLD,&int_nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&int_my_rank);
nprocs = int_nprocs;
my_rank = int_my_rank;
#else
nprocs = 1;
my_rank = 0;
#endif
if( argc < 6 )
{
if( my_rank==0 )
{
printf("Usage: %s m n matrix_type solution_filename rhistory_filename [options]\n", argv[0]);
}
CHKERR(1);
}
m = atoi(argv[1]);
n = atoi(argv[2]);
mtype = atoi(argv[3]);
if( m<=0 || n<=0 )
{
#ifdef _LONGLONG
if( my_rank==0 ) printf("m=%lld <=0 or n=%lld <=0\n",m,n);
#else
if( my_rank==0 ) printf("m=%d <=0 or n=%d <=0\n",m,n);
#endif
CHKERR(1);
}
if( my_rank==0 )
{
printf("\n");
#ifdef _LONGLONG
printf("number of processes = %lld\n",nprocs);
#else
printf("number of processes = %d\n",nprocs);
#endif
}
#ifdef _OPENMP
if( my_rank==0 )
{
#ifdef _LONGLONG
printf("max number of threads = %lld\n",omp_get_num_procs());
printf("number of threads = %lld\n",omp_get_max_threads());
#else
printf("max number of threads = %d\n",omp_get_num_procs());
printf("number of threads = %d\n",omp_get_max_threads());
#endif
}
#endif
/* create matrix and vectors */
nn = m*n;
err = lis_matrix_create(LIS_COMM_WORLD,&A);
err = lis_matrix_set_size(A,0,nn);
CHKERR(err);
ptr = (LIS_INT *)malloc((A->n+1)*sizeof(LIS_INT));
if( ptr==NULL ) CHKERR(1);
index = (LIS_INT *)malloc(5*A->n*sizeof(LIS_INT));
if( index==NULL ) CHKERR(1);
value = (LIS_SCALAR *)malloc(5*A->n*sizeof(LIS_SCALAR));
if( value==NULL ) CHKERR(1);
lis_matrix_get_range(A,&is,&ie);
ctr = 0;
for(ii=is;ii<ie;ii++)
{
i = ii/m;
j = ii - i*m;
if( i>0 ) { jj = ii - m; index[ctr] = jj; value[ctr++] = -1.0;}
if( i<n-1 ) { jj = ii + m; index[ctr] = jj; value[ctr++] = -1.0;}
if( j>0 ) { jj = ii - 1; index[ctr] = jj; value[ctr++] = -1.0;}
if( j<m-1 ) { jj = ii + 1; index[ctr] = jj; value[ctr++] = -1.0;}
index[ctr] = ii; value[ctr++] = 4.0;
ptr[ii-is+1] = ctr;
}
ptr[0] = 0;
err = lis_matrix_set_csr(ptr[ie-is],ptr,index,value,A);
CHKERR(err);
err = lis_matrix_assemble(A);
CHKERR(err);
nnz = A->nnz;
#ifdef USE_MPI
MPI_Allreduce(&nnz,&i,1,LIS_MPI_INT,MPI_SUM,A->comm);
nnz = i;
#endif
#ifdef _LONGLONG
if( my_rank==0 ) printf("matrix size = %lld x %lld (%lld nonzero entries)\n\n",nn,nn,nnz);
#else
if( my_rank==0 ) printf("matrix size = %d x %d (%d nonzero entries)\n\n",nn,nn,nnz);
#endif
err = lis_matrix_duplicate(A,&A0);
CHKERR(err);
lis_matrix_set_type(A0,mtype);
err = lis_matrix_convert(A,A0);
CHKERR(err);
lis_matrix_destroy(A);
A = A0;
err = lis_vector_duplicate(A,&u);
CHKERR(err);
err = lis_vector_duplicate(A,&b);
CHKERR(err);
err = lis_vector_duplicate(A,&x);
CHKERR(err);
err = lis_vector_set_all(1.0,u);
lis_matvec(A,u,b);
err = lis_solver_create(&solver); CHKERR(err);
lis_solver_set_option("-print mem",solver);
lis_solver_set_optionC(solver);
err = lis_solve(A,b,x,solver);
CHKERR(err);
lis_solver_get_iterex(solver,&iter,&iter_double,&iter_quad);
lis_solver_get_timeex(solver,&time,&itime,&ptime,&p_c_time,&p_i_time);
lis_solver_get_residualnorm(solver,&resid);
lis_solver_get_solver(solver,&nsol);
lis_solver_get_solvername(nsol,solvername);
if( my_rank==0 )
{
#ifdef _LONGLONG
#ifdef _LONG__DOUBLE
printf("%s: number of iterations = %lld \n",solvername, iter);
#else
printf("%s: number of iterations = %lld (double = %lld, quad = %lld)\n",solvername,iter, iter_double, iter_quad);
#endif
#else
#ifdef _LONG__DOUBLE
printf("%s: number of iterations = %d \n",solvername, iter);
#else
printf("%s: number of iterations = %d (double = %d, quad = %d)\n",solvername,iter, iter_double, iter_quad);
#endif
#endif
printf("%s: elapsed time = %e sec.\n",solvername,time);
printf("%s: preconditioner = %e sec.\n",solvername, ptime);
printf("%s: matrix creation = %e sec.\n",solvername, p_c_time);
printf("%s: linear solver = %e sec.\n",solvername, itime);
#ifdef _LONG__DOUBLE
printf("%s: relative residual = %Le\n\n",solvername,resid);
#else
printf("%s: relative residual = %e\n\n",solvername,resid);
#endif
}
/* write solution */
lis_output_vector(x,LIS_FMT_MM,argv[4]);
/* write residual history */
lis_solver_output_rhistory(solver, argv[5]);
lis_solver_destroy(solver);
lis_matrix_destroy(A);
lis_vector_destroy(b);
lis_vector_destroy(x);
lis_vector_destroy(u);
lis_finalize();
LIS_DEBUG_FUNC_OUT;
return 0;
}
LIS_INT lis_cgs(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r,rtld, p,phat, q, qhat, u, uhat, vhat;
LIS_SCALAR alpha, beta, rho, rho_old, tmpdot1;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
r = solver->work[0];
rtld = solver->work[1];
p = solver->work[2];
phat = solver->work[3];
q = solver->work[4];
qhat = solver->work[5];
u = solver->work[5];
uhat = solver->work[6];
vhat = solver->work[6];
alpha = (LIS_SCALAR)1.0;
rho_old = (LIS_SCALAR)1.0;
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_all(0,q);
lis_vector_set_all(0,p);
for( iter=1; iter<=maxiter; iter++ )
{
/* rho = <rtld,r> */
lis_vector_dot(rtld,r,&rho);
/* test breakdown */
if( rho==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) */
beta = (rho / rho_old);
/* u = r + beta*q */
lis_vector_axpyz(beta,q,r,u);
/* p = u + beta*(q + beta*p) */
lis_vector_xpay(q,beta,p);
lis_vector_xpay(u,beta,p);
/* phat = M^-1 * p */
time = lis_wtime();
lis_psolve(solver, p, phat);
ptime += lis_wtime()-time;
/* v = A * phat */
lis_matvec(A,phat,vhat);
/* tmpdot1 = <rtld,vhat> */
lis_vector_dot(rtld,vhat,&tmpdot1);
/* test breakdown */
if( tmpdot1==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* alpha = rho / tmpdot1 */
alpha = rho / tmpdot1;
/* q = u - alpha*vhat */
lis_vector_axpyz(-alpha,vhat,u,q);
/* phat = u + q */
/* uhat = M^-1 * (u + q) */
lis_vector_axpyz(1,u,q,phat);
time = lis_wtime();
lis_psolve(solver, phat, uhat);
ptime += lis_wtime()-time;
/* x = x + alpha*uhat */
lis_vector_axpy(alpha,uhat,x);
/* qhat = A * uhat */
lis_matvec(A,uhat,qhat);
/* r = r - alpha*qhat */
lis_vector_axpy(-alpha,qhat,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
rho_old = rho;
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_gmres(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR b,x;
LIS_VECTOR r,s,z,*v;
LIS_SCALAR *h;
LIS_SCALAR aa,bb,rr,a2,b2,t;
LIS_REAL tnrm2;
LIS_REAL bnrm2,nrm2,tol;
LIS_INT iter,maxiter,n,output;
double time,ptime;
LIS_REAL rnorm;
LIS_INT i,j,k,m;
LIS_INT ii,i1,iiv,i1v,iih,jj;
LIS_INT h_dim;
LIS_INT cs,sn;
LIS_DEBUG_FUNC_IN;
A = solver->A;
b = solver->b;
x = solver->x;
n = A->n;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
m = solver->options[LIS_OPTIONS_RESTART];
h_dim = m+1;
ptime = 0.0;
s = solver->work[0];
r = solver->work[1];
z = solver->work[2];
v = &solver->work[3];
h = (LIS_SCALAR *)lis_malloc( sizeof(LIS_SCALAR)*(h_dim+1)*(h_dim+2),"lis_gmres::h" );
cs = (m+1)*h_dim;
sn = (m+2)*h_dim;
/* r = M^-1 * (b - A * x) */
lis_matvec(A,x,z);
lis_vector_xpay(b,-1.0,z);
lis_psolve(solver,z,v[0]);
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,v[0],&bnrm2) )
{
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
iter=0;
while( iter<maxiter )
{
/* first column of V */
/* v = r / ||r||_2 */
lis_vector_nrm2(v[0],&rnorm);
lis_vector_scale(1.0/rnorm,v[0]);
/* s = ||r||_2 e_1 */
lis_vector_set_all(0,s);
s->value[0] = rnorm;
i = 0;
do
{
iter++;
i++;
ii = i-1;
i1 = i;
iiv = i-1;
i1v = i;
iih = (i-1)*h_dim;
/* z = M^-1 * v */
time = lis_wtime();
lis_psolve(solver,v[iiv],z);
ptime += lis_wtime()-time;
/* w = A * z */
lis_matvec(A,z,v[i1v]);
for(k=0;k<i;k++)
{
/* h[k,i] = <w,v[k]> */
/* w = w - h[k,i] * v[k] */
lis_vector_dot(v[i1v],v[k],&t);
h[k+iih] = t;
lis_vector_axpy(-t,v[k],v[i1v]);
}
/* h[i+1,i] = ||w|| */
/* v[i+1] = w / h[i+1,i] */
lis_vector_nrm2(v[i1v],&tnrm2);
h[i1+iih] = tnrm2;
lis_vector_scale(1.0/tnrm2,v[i1v]);
for(k=1;k<=ii;k++)
{
jj = k-1;
t = h[jj+iih];
aa = h[jj+cs]*t;
aa += h[jj+sn]*h[k+iih];
bb = -h[jj+sn]*t;
bb += h[jj+cs]*h[k+iih];
h[jj+iih] = aa;
h[k+iih] = bb;
}
aa = h[ii+iih];
bb = h[i1+iih];
a2 = aa*aa;
b2 = bb*bb;
rr = sqrt(a2+b2);
if( rr==0.0 ) rr=1.0e-17;
h[ii+cs] = aa/rr;
h[ii+sn] = bb/rr;
s->value[i1] = -h[ii+sn]*s->value[ii];
s->value[ii] = h[ii+cs]*s->value[ii];
aa = h[ii+cs]*h[ii+iih];
aa += h[ii+sn]*h[i1+iih];
h[ii+iih] = aa;
/* convergence check */
nrm2 = sabs(s->value[i1])*bnrm2;
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 ) break;
} while( i<m && iter <maxiter );
/* Solve H * Y = S for upper Hessenberg matrix H */
s->value[ii] = s->value[ii]/h[ii+iih];
for(k=1;k<=ii;k++)
{
jj = ii-k;
t = s->value[jj];
for(j=jj+1;j<=ii;j++)
{
t -= h[jj+j*h_dim]*s->value[j];
}
s->value[jj] = t/h[jj+jj*h_dim];
}
/* z = z + y * v */
#ifdef _OPENMP
#pragma omp parallel for private(k)
#endif
for(k=0;k<n;k++)
{
z->value[k] = s->value[0]*v[0]->value[k];
}
for(j=1;j<=ii;j++)
{
lis_vector_axpy(s->value[j],v[j],z);
}
/* r = M^-1 * z */
time = lis_wtime();
lis_psolve(solver,z,r);
ptime += lis_wtime()-time;
/* x = x + r */
lis_vector_axpy(1,r,x);
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
for(j=1;j<=i;j++)
{
jj = i1-j+1;
s->value[jj-1] = -h[jj-1+sn]*s->value[jj];
s->value[jj] = h[jj-1+cs]*s->value[jj];
}
for(j=0;j<=i1;j++)
{
t = s->value[j];
if( j==0 ) t = t-1.0;
lis_vector_axpy(t,v[j],v[0]);
}
}
solver->retcode = LIS_MAXITER;
solver->iter = iter+1;
solver->resid = nrm2;
lis_free(h);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_crs_quad(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r,rtld, p, q, u, z, ap, map, uq, auq;
LIS_QUAD_PTR alpha, beta, rho, rho_old, tmpdot1, one;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
r = solver->work[0];
rtld = solver->work[1];
p = solver->work[2];
z = solver->work[3];
u = solver->work[3];
uq = solver->work[3];
q = solver->work[4];
ap = solver->work[4];
map = solver->work[5];
auq = solver->work[5];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot1,4,1);
LIS_QUAD_SCALAR_MALLOC(one,6,1);
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,p);
lis_matvect(A,p,rtld);
lis_vector_set_allex_nm(0.0,q);
lis_vector_set_allex_nm(0.0,p);
rho_old.hi[0] = 1.0;
rho_old.lo[0] = 0.0;
one.hi[0] = 1.0;
one.lo[0] = 0.0;
for( iter=1; iter<=maxiter; iter++ )
{
/* z = M^-1 * r */
/* rho = <rtld,z> */
time = lis_wtime();
lis_psolve(solver, r, z);
ptime += lis_wtime()-time;
lis_vector_dotex_mmm(rtld,z,&rho);
/* test breakdown */
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = rho / rho_old */
/* u = z + beta*q */
/* p = u + beta*(q + beta*p) */
/* ap = A * p */
/* map = M^-1 * ap */
/* tmpdot1 = <rtld,map> */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_vector_axpyzex_mmmm(beta,q,z,u);
lis_vector_xpayex_mmm(q,beta,p);
lis_vector_xpayex_mmm(u,beta,p);
lis_matvec(A,p,ap);
time = lis_wtime();
lis_psolve(solver, ap, map);
ptime += lis_wtime()-time;
lis_vector_dotex_mmm(rtld,map,&tmpdot1);
/* test breakdown */
if( tmpdot1.hi[0]==0.0 && tmpdot1.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* alpha = rho / tmpdot1 */
/* q = u - alpha*map */
/* uq = u + q */
/* auq = A * uq */
/* x = x + alpha*uq */
/* r = r - alpha*auq */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)tmpdot1.hi);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyzex_mmmm(alpha,map,u,q);
lis_vector_axpyzex_mmmm(one,u,q,uq);
lis_matvec(A,uq,auq);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,uq,x);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,auq,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_esi_quad(LIS_ESOLVER esolver)
{
LIS_MATRIX A;
LIS_VECTOR x, Ax;
LIS_SCALAR xAx, xx, mu, lshift;
LIS_INT ss;
LIS_INT emaxiter;
LIS_REAL tol;
LIS_INT i,j,k;
LIS_SCALAR evalue,dotvr;
LIS_INT iter,giter,output,niesolver;
LIS_INT nprocs,my_rank;
LIS_REAL nrm2,dot,resid,resid0;
LIS_QUAD_PTR qdot_vv, qdot_vr;
LIS_VECTOR *v,r,q;
LIS_SOLVER solver;
LIS_PRECON precon;
double times,itimes,ptimes,p_c_times,p_i_times;
LIS_INT err;
LIS_INT nsol, precon_type;
char solvername[128], preconname[128];
LIS_DEBUG_FUNC_IN;
A = esolver->A;
x = esolver->x;
ss = esolver->options[LIS_EOPTIONS_SUBSPACE];
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
tol = esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN];
lshift = esolver->lshift;
output = esolver->options[LIS_EOPTIONS_OUTPUT];
niesolver = esolver->options[LIS_EOPTIONS_INNER_ESOLVER];
r = esolver->work[0];
q = esolver->work[1];
v = &esolver->work[2];
Ax = esolver->work[3];
LIS_QUAD_SCALAR_MALLOC(qdot_vv,0,1);
LIS_QUAD_SCALAR_MALLOC(qdot_vr,1,1);
lis_vector_set_all(1.0,r);
lis_vector_nrm2(r, &nrm2);
lis_vector_scale(1/nrm2,r);
switch ( niesolver )
{
case LIS_ESOLVER_II:
lis_solver_create(&solver);
lis_solver_set_option("-i bicg -p ilu -precision quad",solver);
lis_solver_set_optionC(solver);
lis_solver_get_solver(solver, &nsol);
lis_solver_get_precon(solver, &precon_type);
lis_get_solvername(nsol, solvername);
lis_get_preconname(precon_type, preconname);
printf("solver : %s %d\n", solvername, nsol);
printf("precon : %s %d\n", preconname, precon_type);
if( A->my_rank==0 ) printf("local shift = %e\n", lshift);
if (lshift != 0) lis_matrix_shift_diagonal(A, lshift);
break;
case LIS_ESOLVER_AII:
lis_solver_create(&solver);
lis_solver_set_option("-i bicg -p ilu -precision quad",solver);
lis_solver_set_optionC(solver);
lis_solver_get_solver(solver, &nsol);
lis_solver_get_precon(solver, &precon_type);
lis_get_solvername(nsol, solvername);
lis_get_preconname(precon_type, preconname);
printf("solver : %s %d\n", solvername, nsol);
printf("precon : %s %d\n", preconname, precon_type);
if( A->my_rank==0 ) printf("local shift = %e\n", lshift);
if (lshift != 0) lis_matrix_shift_diagonal(A, lshift);
lis_vector_set_all(1.0,q);
lis_solve(A, q, x, solver);
lis_precon_create(solver, &precon);
solver->precon = precon;
break;
case LIS_ESOLVER_RQI:
lis_solver_create(&solver);
lis_solver_set_option("-p ilu -precision quad -maxiter 10",solver);
lis_solver_set_optionC(solver);
lis_solver_get_solver(solver, &nsol);
lis_solver_get_precon(solver, &precon_type);
lis_get_solvername(nsol, solvername);
lis_get_preconname(precon_type, preconname);
printf("solver : %s %d\n", solvername, nsol);
printf("precon : %s %d\n", preconname, precon_type);
if( A->my_rank==0 ) printf("local shift = %e\n", lshift);
if (lshift != 0) lis_matrix_shift_diagonal(A, lshift);
break;
}
giter=0;
j=0;
while (j<ss)
{
lis_vector_duplicate(A,&esolver->evector[j]);
j = j+1;
lis_vector_copy(r, v[j]);
if (niesolver==LIS_ESOLVER_II || niesolver==LIS_ESOLVER_RQI)
{
/* create preconditioner */
solver->A = A;
err = lis_precon_create(solver, &precon);
if( err )
{
lis_solver_work_destroy(solver);
solver->retcode = err;
return err;
}
}
if (niesolver==LIS_ESOLVER_RQI)
{
lis_vector_nrm2(x, &nrm2);
lis_vector_scale(1/nrm2, x);
lis_matvec(A, x, Ax);
lis_vector_dot(x, Ax, &xAx);
lis_vector_dot(x, x, &xx);
mu = xAx / xx;
}
iter = 0;
while (iter<emaxiter)
{
/* diagonalization */
iter = iter+1;
giter = giter+1;
for (k=1;k<j;k++)
{
lis_vector_dotex_mmm(v[j], v[k], &qdot_vv);
lis_quad_minus((LIS_QUAD *)qdot_vv.hi);
lis_vector_axpyex_mmm(qdot_vv,v[k],v[j]);
}
switch( niesolver )
{
case LIS_ESOLVER_PI:
lis_matvec(A,v[j],r);
break;
case LIS_ESOLVER_II:
lis_solve_kernel(A, v[j], r, solver, precon);
break;
case LIS_ESOLVER_AII:
lis_psolve(solver, v[j], r);
break;
case LIS_ESOLVER_RQI:
lis_vector_nrm2(v[j], &nrm2);
lis_vector_scale(1/nrm2, v[j]);
lis_matrix_shift_diagonal(A, -mu);
lis_solve_kernel(A, v[j], r, solver, precon);
lis_matrix_shift_diagonal(A, mu);
break;
}
if ( j==1 && ( niesolver==LIS_ESOLVER_II || niesolver==LIS_ESOLVER_AII || niesolver==LIS_ESOLVER_RQI ))
{
lis_solver_get_timeex(solver,×,&itimes,&ptimes,&p_c_times,&p_i_times);
esolver->ptimes += solver->ptimes;
esolver->itimes += solver->itimes;
esolver->p_c_times += solver->p_c_times;
esolver->p_i_times += solver->p_i_times;
}
lis_vector_nrm2(r, &nrm2);
lis_vector_dotex_mmm(v[j], r, &qdot_vr);
lis_quad_minus((LIS_QUAD *)qdot_vr.hi);
lis_vector_axpyzex_mmmm(qdot_vr,v[j],r,q);
lis_quad_minus((LIS_QUAD *)qdot_vr.hi);
dotvr = qdot_vr.hi[0];
mu = mu + 1/dotvr;
lis_vector_nrm2(q, &resid);
resid = fabs(resid / dotvr);
lis_vector_scale(1/nrm2,r);
lis_vector_copy(r, v[j]);
if ( j==1 )
{
if( output & LIS_PRINT_MEM ) esolver->residual[iter] = resid;
if( output & LIS_PRINT_OUT ) printf("iter: %5d residual = %e\n", iter, resid);
esolver->iter = iter;
esolver->resid = resid;
}
if (tol>resid) break;
}
if (niesolver==LIS_ESOLVER_II || niesolver==LIS_ESOLVER_RQI)
{
lis_precon_destroy(precon);
}
switch ( niesolver )
{
case LIS_ESOLVER_PI:
esolver->evalue[j-1] = dotvr;
break;
case LIS_ESOLVER_II:
esolver->evalue[j-1] = 1/dotvr;
break;
case LIS_ESOLVER_AII:
esolver->evalue[j-1] = 1/dotvr;
break;
case LIS_ESOLVER_RQI:
esolver->evalue[j-1] = mu;
break;
}
lis_vector_copy(v[j], esolver->evector[j-1]);
if (A->my_rank==0 && ss>1)
{
#ifdef _LONGLONG
printf("Subspace: mode number = %lld\n", j-1);
#else
printf("Subspace: mode number = %d\n", j-1);
#endif
printf("Subspace: eigenvalue = %e\n", esolver->evalue[j-1]);
#ifdef _LONGLONG
printf("Subspace: number of iterations = %lld\n",iter);
#else
printf("Subspace: number of iterations = %d\n",iter);
#endif
printf("Subspace: relative residual 2-norm = %e\n",resid);
}
}
lis_vector_copy(esolver->evector[esolver->options[LIS_EOPTIONS_MODE]], esolver->x);
switch ( niesolver )
{
case LIS_ESOLVER_II:
if (lshift != 0) lis_matrix_shift_diagonal(A, -lshift);
lis_solver_destroy(solver);
break;
case LIS_ESOLVER_AII:
if (lshift != 0) lis_matrix_shift_diagonal(A, -lshift);
lis_precon_destroy(precon);
lis_solver_destroy(solver);
break;
case LIS_ESOLVER_RQI:
if (lshift != 0) lis_matrix_shift_diagonal(A, -lshift);
lis_solver_destroy(solver);
break;
}
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_gs(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR b,x;
LIS_VECTOR r,t,s;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output;
double time,ptime;
LIS_INT err;
LIS_DEBUG_FUNC_IN;
A = solver->A;
b = solver->b;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
ptime = 0.0;
r = solver->work[0];
t = solver->work[1];
s = solver->work[2];
lis_vector_nrm2(b,&bnrm2);
bnrm2 = 1.0 / bnrm2;
err = lis_matrix_split(A);
if( err ) return err;
if( A->use_wd!=LIS_SOLVER_GS )
{
if( !A->WD )
{
err = lis_matrix_diag_duplicate(A->D,&A->WD);
if( err ) return err;
}
lis_matrix_diag_copy(A->D,A->WD);
lis_matrix_diag_inverse(A->WD);
A->use_wd = LIS_SOLVER_GS;
}
for( iter=1; iter<=maxiter; iter++ )
{
/* x += (D-L)^{-1}(b - Ax) */
time = lis_wtime();
lis_psolve(solver,x,s);
ptime += lis_wtime() - time;
lis_matvec(A,s,t);
/* lis_matvec(A,x,t);*/
lis_vector_axpyz(-1,t,b,r);
lis_vector_nrm2(r,&nrm2);
lis_matrix_solve(A,r,t,LIS_MATRIX_LOWER);
lis_vector_axpy(1,t,x);
/* convergence check */
nrm2 = nrm2 * bnrm2;
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 )
{
time = lis_wtime();
lis_psolve(solver,x,s);
ptime += lis_wtime() - time;
lis_vector_copy(s,x);
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
}
lis_psolve(solver,x,s);
lis_vector_copy(s,x);
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_cgs_switch(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r,rtld, p,phat, q, qhat, u, uhat, vhat;
LIS_QUAD_PTR alpha, beta, rho, rho_old, tmpdot1, one;
LIS_REAL bnrm2, nrm2, tol, tol2;
LIS_INT iter,maxiter,output,conv;
LIS_INT iter2,maxiter2;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
maxiter2 = solver->options[LIS_OPTIONS_SWITCH_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
tol2 = solver->params[LIS_PARAMS_SWITCH_RESID-LIS_OPTIONS_LEN];
ptime = 0.0;
r = solver->work[0];
rtld = solver->work[1];
p = solver->work[2];
phat = solver->work[3];
q = solver->work[4];
qhat = solver->work[5];
u = solver->work[5];
uhat = solver->work[6];
vhat = solver->work[6];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot1,4,1);
LIS_QUAD_SCALAR_MALLOC(one,6,1);
rho_old.hi[0] = 1.0;
rho_old.lo[0] = 0.0;
alpha.hi[0] = 1.0;
alpha.lo[0] = 0.0;
one.hi[0] = 1.0;
one.lo[0] = 0.0;
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol2 = solver->tol_switch;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0, q);
lis_vector_set_allex_nm(0.0, p);
uhat->precision = LIS_PRECISION_DEFAULT;
p->precision = LIS_PRECISION_DEFAULT;
phat->precision = LIS_PRECISION_DEFAULT;
for( iter=1; iter<=maxiter2; iter++ )
{
/* rho = <rtld,r> */
lis_vector_dot(rtld,r,&rho.hi[0]);
/* test breakdown */
if( rho.hi[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) */
beta.hi[0] = (rho.hi[0] / rho_old.hi[0]);
/* u = r + beta*q */
lis_vector_axpyz(beta.hi[0],q,r,u);
/* p = u + beta*(q + beta*p) */
lis_vector_xpay(q,beta.hi[0],p);
lis_vector_xpay(u,beta.hi[0],p);
/* phat = M^-1 * p */
time = lis_wtime();
lis_psolve(solver, p, phat);
ptime += lis_wtime()-time;
/* v = A * phat */
lis_matvec(A,phat,vhat);
/* tmpdot1 = <rtld,vhat> */
lis_vector_dot(rtld,vhat,&tmpdot1.hi[0]);
/* test breakdown */
if( tmpdot1.hi[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* alpha = rho / tmpdot1 */
alpha.hi[0] = rho.hi[0] / tmpdot1.hi[0];
/* q = u - alpha*vhat */
lis_vector_axpyz(-alpha.hi[0],vhat,u,q);
/* phat = u + q */
/* uhat = M^-1 * (u + q) */
lis_vector_axpyz(1.0,u,q,phat);
time = lis_wtime();
lis_psolve(solver, phat, uhat);
ptime += lis_wtime()-time;
/* x = x + alpha*uhat */
lis_vector_axpy(alpha.hi[0],uhat,x);
/* qhat = A * uhat */
lis_matvec(A,uhat,qhat);
/* r = r - alpha*qhat */
lis_vector_axpy(-alpha.hi[0],qhat,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( nrm2 <= tol2 )
{
solver->iter = iter;
solver->iter2 = iter;
solver->ptime = ptime;
break;
}
rho_old.hi[0] = rho.hi[0];
}
uhat->precision = LIS_PRECISION_QUAD;
p->precision = LIS_PRECISION_QUAD;
phat->precision = LIS_PRECISION_QUAD;
solver->options[LIS_OPTIONS_INITGUESS_ZEROS] = LIS_FALSE;
lis_vector_copyex_mn(x,solver->xx);
rho_old.hi[0] = 1.0;
lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2);
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0, q);
lis_vector_set_allex_nm(0.0, p);
for( iter2=iter+1; iter2<=maxiter; iter2++ )
{
/* rho = <rtld,r> */
lis_vector_dotex_mmm(rtld,r,&rho);
/* test breakdown */
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
/* u = r + beta*q */
lis_vector_axpyzex_mmmm(beta,q,r,u);
/* p = u + beta*(q + beta*p) */
lis_vector_xpayex_mmm(q,beta,p);
lis_vector_xpayex_mmm(u,beta,p);
/* phat = M^-1 * p */
time = lis_wtime();
lis_psolve(solver, p, phat);
ptime += lis_wtime()-time;
/* v = A * phat */
lis_matvec(A,phat,vhat);
/* tmpdot1 = <rtld,vhat> */
lis_vector_dotex_mmm(rtld,vhat,&tmpdot1);
/* test breakdown */
if( tmpdot1.hi[0]==0.0 && tmpdot1.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* alpha = rho / tmpdot1 */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)tmpdot1.hi);
/* q = u - alpha*vhat */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyzex_mmmm(alpha,vhat,u,q);
/* phat = u + q */
/* uhat = M^-1 * (u + q) */
lis_vector_axpyzex_mmmm(one,u,q,phat);
time = lis_wtime();
lis_psolve(solver, phat, uhat);
ptime += lis_wtime()-time;
/* x = x + alpha*uhat */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,uhat,x);
/* qhat = A * uhat */
lis_matvec(A,uhat,qhat);
/* r = r - alpha*qhat */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,qhat,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter2] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT main(LIS_INT argc, char* argv[])
{
LIS_MATRIX A;
LIS_VECTOR b,x,u;
LIS_SOLVER solver;
LIS_INT my_rank;
#ifdef USE_MPI
int int_nprocs,int_my_rank;
#endif
LIS_INT err,i,n,gn,is,ie,iter;
n = 12;
lis_initialize(&argc, &argv);
#ifdef USE_MPI
MPI_Comm_size(MPI_COMM_WORLD,&int_nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&int_my_rank);
my_rank = int_my_rank;
#else
my_rank = 0;
#endif
lis_matrix_create(LIS_COMM_WORLD,&A);
err = lis_matrix_set_size(A,0,n);
CHKERR(err);
lis_matrix_get_size(A,&n,&gn);
lis_matrix_get_range(A,&is,&ie);
for(i=is;i<ie;i++)
{
if( i>0 ) lis_matrix_set_value(LIS_INS_VALUE,i,i-1,-1.0,A);
if( i<gn-1 ) lis_matrix_set_value(LIS_INS_VALUE,i,i+1,-1.0,A);
lis_matrix_set_value(LIS_INS_VALUE,i,i,2.0,A);
}
lis_matrix_set_type(A,LIS_MATRIX_CSR);
lis_matrix_assemble(A);
lis_vector_duplicate(A,&u);
lis_vector_duplicate(A,&b);
lis_vector_duplicate(A,&x);
lis_vector_set_all(1.0,u);
lis_matvec(A,u,b);
lis_solver_create(&solver);
lis_solver_set_option("-print mem",solver);
lis_solver_set_optionC(solver);
lis_solve(A,b,x,solver);
lis_solver_get_iter(solver,&iter);
if (my_rank==0)
{
#ifdef _LONG__LONG
printf("number of iterations = %lld\n",iter);
#else
printf("number of iterations = %d\n",iter);
#endif
printf("\n");
}
lis_vector_print(x);
lis_matrix_destroy(A);
lis_vector_destroy(b);
lis_vector_destroy(x);
lis_vector_destroy(u);
lis_solver_destroy(solver);
lis_finalize();
return 0;
}
LIS_INT main(LIS_INT argc, char* argv[])
{
LIS_MATRIX A0,A,B;
LIS_VECTOR x,b,u;
LIS_SOLVER solver;
LIS_INT nprocs,my_rank;
int int_nprocs,int_my_rank;
LIS_INT nsol,rhs,len,i,j,k,jj,kk,p,nrow_p,l,n;
LIS_INT err,iter,iter_double,iter_quad;
LIS_INT *iw,*nrow,*index,*ptr;
LIS_SCALAR *s,t,*value;
double times,itimes,ptimes,p_c_times,p_i_times;
LIS_REAL resid;
char solvername[128];
LIS_DEBUG_FUNC_IN;
lis_initialize(&argc, &argv);
#ifdef USE_MPI
MPI_Comm_size(MPI_COMM_WORLD,&int_nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&int_my_rank);
nprocs = int_nprocs;
my_rank = int_my_rank;
#else
nprocs = 1;
my_rank = 0;
#endif
if( argc < 5 )
{
if( my_rank==0 )
{
printf("Usage: %s matrix_filename rhs_setting solution_filename residual_filename [options]\n", argv[0]);
}
CHKERR(1);
}
len = (LIS_INT)strlen(argv[2]);
if( len==1 )
{
if( argv[2][0]=='0' || argv[2][0]=='1' || argv[2][0]=='2' )
{
rhs = atoi(argv[2]);
}
else
{
rhs = -1;
}
}
else
{
rhs = -1;
}
if( my_rank==0 )
{
printf("\n");
#ifdef _LONGLONG
printf("number of processes = %lld\n",nprocs);
#else
printf("number of processes = %d\n",nprocs);
#endif
}
#ifdef _OPENMP
if( my_rank==0 )
{
#ifdef _LONGLONG
printf("max number of threads = %lld\n",omp_get_num_procs());
printf("number of threads = %lld\n",omp_get_max_threads());
#else
printf("max number of threads = %d\n",omp_get_num_procs());
printf("number of threads = %d\n",omp_get_max_threads());
#endif
}
#endif
/* read matrix and vectors from file */
err = lis_matrix_create(LIS_COMM_WORLD,&A); CHKERR(err);
err = lis_vector_create(LIS_COMM_WORLD,&b); CHKERR(err);
err = lis_vector_create(LIS_COMM_WORLD,&x); CHKERR(err);
err = lis_input(A,b,x,argv[1]);
CHKERR(err);
err = lis_matrix_duplicate(A,&A0);
CHKERR(err);
lis_matrix_set_type(A0,LIS_MATRIX_CSR);
err = lis_matrix_convert(A,A0);
CHKERR(err);
lis_matrix_destroy(A);
A = A0;
err = lis_vector_duplicate(A,&u);
CHKERR(err);
if( lis_vector_is_null(b) )
{
lis_vector_destroy(b);
lis_vector_duplicate(A,&b);
CHKERR(err);
if( rhs==0 )
{
CHKERR(1);
}
else if( rhs==1 )
{
err = lis_vector_set_all(1.0,b);
}
else
{
err = lis_vector_set_all(1.0,u);
lis_matvec(A,u,b);
}
}
if( rhs==-1 )
{
lis_input_vector(b,argv[2]);
}
if( lis_vector_is_null(x) )
{
lis_vector_destroy(x);
err = lis_vector_duplicate(A,&x);
CHKERR(err);
}
err = lis_solver_create(&solver); CHKERR(err);
lis_solver_set_option("-print mem",solver);
lis_solver_set_optionC(solver);
err = lis_solve(A,b,x,solver);
CHKERR(err);
lis_solver_get_itersex(solver,&iter,&iter_double,&iter_quad);
lis_solver_get_timeex(solver,×,&itimes,&ptimes,&p_c_times,&p_i_times);
lis_solver_get_residualnorm(solver,&resid);
lis_solver_get_solver(solver,&nsol);
lis_solver_get_solvername(nsol,solvername);
/* write results */
if( my_rank==0 )
{
#ifdef _LONGLONG
#ifdef _LONG__DOUBLE
printf("%s: number of iterations = %lld \n",solvername, iter);
#else
printf("%s: number of iterations = %lld (double = %lld, quad = %lld)\n",solvername,iter, iter_double, iter_quad);
#endif
#else
#ifdef _LONG__DOUBLE
printf("%s: number of iterations = %d \n",solvername, iter);
#else
printf("%s: number of iterations = %d (double = %d, quad = %d)\n",solvername,iter, iter_double, iter_quad);
#endif
#endif
printf("%s: elapsed time = %e sec.\n",solvername,times);
printf("%s: preconditioner = %e sec.\n",solvername, ptimes);
printf("%s: matrix creation = %e sec.\n",solvername, p_c_times);
printf("%s: linear solver = %e sec.\n",solvername, itimes);
#ifdef _LONG__DOUBLE
printf("%s: relative residual 2-norm = %Le\n\n",solvername,resid);
#else
printf("%s: relative residual 2-norm = %e\n\n",solvername,resid);
#endif
}
/* write solution */
lis_output_vector(x,LIS_FMT_MM,argv[3]);
/* write residual */
lis_solver_output_rhistory(solver, argv[4]);
lis_solver_destroy(solver);
lis_vector_destroy(x);
lis_vector_destroy(u);
lis_vector_destroy(b);
lis_matrix_destroy(A);
lis_finalize();
LIS_DEBUG_FUNC_OUT;
return 0;
}
LIS_INT lis_psolve_adds(LIS_SOLVER solver, LIS_VECTOR B, LIS_VECTOR X)
{
LIS_INT i,k,n,np,iter,ptype;
LIS_SCALAR *b,*x,*w,*r,*rl;
LIS_VECTOR W,R;
LIS_PRECON precon;
LIS_QUAD_DECLAR;
LIS_DEBUG_FUNC_IN;
precon = solver->precon;
n = precon->A->n;
np = precon->A->np;
W = precon->work[0];
R = precon->work[1];
b = B->value;
x = X->value;
w = W->value;
r = R->value;
rl = R->value_lo;
iter = solver->options[LIS_OPTIONS_ADDS_ITER];
ptype = solver->options[LIS_OPTIONS_PRECON];
#ifdef USE_QUAD_PRECISION
if( solver->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_set_all(0.0,X);
lis_vector_copy(B,R);
for(k=0;k<iter+1;k++)
{
for(i=n;i<np;i++)
{
r[i] = 0.0;
}
lis_psolve_xxx[ptype](solver,R,W);
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for(i=0;i<n;i++)
{
x[i] += w[i];
}
if(k!=iter)
{
lis_matvec(precon->A,X,R);
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for(i=0;i<n;i++)
{
r[i] = b[i] - r[i];
}
}
}
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_set_allex_nm(0.0,X);
lis_vector_copyex_mm(B,R);
for(k=0;k<iter+1;k++)
{
for(i=n;i<np;i++)
{
r[i] = 0.0;
rl[i] = 0.0;
}
lis_psolve_xxx[ptype](solver,R,W);
for(i=0;i<n;i++)
{
#ifndef USE_SSE2
LIS_QUAD_ADD(X->value[i],X->value_lo[i],X->value[i],X->value_lo[i],W->value[i],W->value_lo[i]);
#else
LIS_QUAD_ADD_SSE2(X->value[i],X->value_lo[i],X->value[i],X->value_lo[i],W->value[i],W->value_lo[i]);
#endif
/* x[i] += w[i];*/
}
if(k==iter) break;
lis_matvec(precon->A,X,R);
for(i=0;i<n;i++)
{
#ifndef USE_SSE2
LIS_QUAD_ADD(R->value[i],R->value_lo[i],B->value[i],B->value_lo[i],-R->value[i],-R->value_lo[i]);
#else
LIS_QUAD_ADD_SSE2(R->value[i],R->value_lo[i],B->value[i],B->value_lo[i],-R->value[i],-R->value_lo[i]);
#endif
/* r[i] = b[i] - r[i];*/
}
}
}
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_minres(LIS_SOLVER solver)
{
LIS_Comm comm;
LIS_MATRIX A;
LIS_VECTOR b,x;
LIS_VECTOR v1,v2,v3,v4,w0,w1,w2;
LIS_REAL nrm2,tol;
LIS_SCALAR alpha;
LIS_REAL beta2,beta3;
LIS_SCALAR gamma1,gamma2,gamma3;
LIS_SCALAR delta,eta;
LIS_SCALAR sigma1,sigma2,sigma3;
LIS_SCALAR rho1,rho2,rho3;
LIS_REAL r0_euc,r_euc;
LIS_INT iter,maxiter,output;
double time,ptime;
LIS_DEBUG_FUNC_IN;
comm = LIS_COMM_WORLD;
A = solver->A;
b = solver->b;
x = solver->x;
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
ptime = 0.0;
v1 = solver->work[0];
v2 = solver->work[1];
v3 = solver->work[2];
v4 = solver->work[3];
w0 = solver->work[4];
w1 = solver->work[5];
w2 = solver->work[6];
/* Lanczos algorithm */
lis_matvec(A,x,v2);
lis_vector_xpay(b,-1.0,v2);
time = lis_wtime();
lis_psolve(solver,v2,v3);
ptime += lis_wtime()-time;
lis_vector_copy(v3,v2);
/* Compute elements of Hermitian tridiagonal matrix */
lis_vector_nrm2(v2,&r_euc);
eta = beta2 = r0_euc = r_euc;
gamma2 = gamma1 = 1.0;
sigma2 = sigma1 = 0.0;
lis_vector_set_all(0.0,v1);
lis_vector_set_all(0.0,w0);
lis_vector_set_all(0.0,w1);
nrm2 = r_euc / r0_euc;
for(iter=1;iter<=maxiter;iter++)
{
/* Lanczos algorithm */
lis_vector_scale(1.0 / beta2,v2);
lis_matvec(A,v2,v3);
time = lis_wtime();
lis_psolve(solver,v3,v4);
ptime += lis_wtime()-time;
lis_vector_dot(v2,v4,&alpha);
lis_vector_axpy(-alpha,v2,v4);
lis_vector_axpy(-beta2,v1,v4);
lis_vector_nrm2(v4,&beta3);
/* Compute elements of Hermitian tridiagonal matrix */
delta = gamma2 * alpha - gamma1 * sigma2 * beta2;
rho1 = sqrt(delta * delta + beta3 * beta3);
rho2 = sigma2 * alpha + gamma1 * gamma2 * beta2;
rho3 = sigma1 * beta2;
gamma3 = delta / rho1;
sigma3 = beta3 / rho1;
lis_vector_axpyz(-rho3,w0,v2,w2);
lis_vector_axpy(-rho2,w1,w2);
lis_vector_scale(1.0 / rho1,w2);
lis_vector_axpy(gamma3 * eta,w2,x);
/* convergence check */
r_euc *= fabs(sigma3);
nrm2 = r_euc / r0_euc;
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT ) lis_print_rhistory(comm,iter,nrm2);
}
if( nrm2 <= tol )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
eta *= -sigma3;
lis_vector_copy(v2,v1);
lis_vector_copy(v4,v2);
lis_vector_copy(w1,w0);
lis_vector_copy(w2,w1);
beta2 = beta3;
gamma1 = gamma2;
gamma2 = gamma3;
sigma1 = sigma2;
sigma2 = sigma3;
}
lis_vector_destroy(v1);
lis_vector_destroy(v2);
lis_vector_destroy(v3);
lis_vector_destroy(v4);
lis_vector_destroy(w0);
lis_vector_destroy(w1);
lis_vector_destroy(w2);
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT main(LIS_INT argc, char* argv[])
{
LIS_MATRIX A,A0;
LIS_VECTOR b,x;
LIS_SCALAR *value;
LIS_INT nprocs,my_rank;
int int_nprocs,int_my_rank;
LIS_INT nthreads,maxthreads;
LIS_INT gn,nnz,np;
LIS_INT i,j,k,si,sj,sk,ii,jj,ctr;
LIS_INT l,m,n,nn;
LIS_INT is,ie;
LIS_INT err,iter,matrix_type,storage,ss,se;
LIS_INT *ptr,*index;
double time,time2,nnzs,nnzap,nnzt;
LIS_SCALAR val;
double commtime,comptime,flops;
LIS_DEBUG_FUNC_IN;
lis_initialize(&argc, &argv);
#ifdef USE_MPI
MPI_Comm_size(MPI_COMM_WORLD,&int_nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&int_my_rank);
nprocs = int_nprocs;
my_rank = int_my_rank;
#else
nprocs = 1;
my_rank = 0;
#endif
if( argc < 5 )
{
if( my_rank==0 )
{
printf("Usage: %s l m n iter [matrix_type]\n", argv[0]);
}
CHKERR(1);
}
l = atoi(argv[1]);
m = atoi(argv[2]);
n = atoi(argv[3]);
iter = atoi(argv[4]);
if (argv[5] == NULL) {
storage = 0;
}
else {
storage = atoi(argv[5]);
}
if( iter<=0 )
{
#ifdef _LONG__LONG
if( my_rank==0 ) printf("iter=%lld <= 0\n",iter);
#else
if( my_rank==0 ) printf("iter=%d <= 0\n",iter);
#endif
CHKERR(1);
}
if( l<=0 || m<=0 || n<=0 )
{
#ifdef _LONG__LONG
if( my_rank==0 ) printf("l=%lld <=0, m=%lld <=0 or n=%lld <=0\n",l,m,n);
#else
if( my_rank==0 ) printf("l=%d <=0, m=%d <=0 or n=%d <=0\n",l,m,n);
#endif
CHKERR(1);
}
if( storage<0 || storage>11 )
{
#ifdef _LONG__LONG
if( my_rank==0 ) printf("matrix_type=%lld < 0 or matrix_type=%lld > 11\n",storage,storage);
#else
if( my_rank==0 ) printf("matrix_type=%d < 0 or matrix_type=%d > 11\n",storage,storage);
#endif
CHKERR(1);
}
if( my_rank==0 )
{
printf("\n");
#ifdef _LONG__LONG
printf("number of processes = %lld\n",nprocs);
#else
printf("number of processes = %d\n",nprocs);
#endif
}
#ifdef _OPENMP
nthreads = omp_get_num_procs();
maxthreads = omp_get_max_threads();
if( my_rank==0 )
{
#ifdef _LONG__LONG
printf("max number of threads = %lld\n", nthreads);
printf("number of threads = %lld\n", maxthreads);
#else
printf("max number of threads = %d\n", nthreads);
printf("number of threads = %d\n", maxthreads);
#endif
}
#else
nthreads = 1;
maxthreads = 1;
#endif
/* create matrix and vectors */
nn = l*m*n;
err = lis_matrix_create(LIS_COMM_WORLD,&A0);
err = lis_matrix_set_size(A0,0,nn);
CHKERR(err);
ptr = (LIS_INT *)malloc((A0->n+1)*sizeof(LIS_INT));
if( ptr==NULL ) CHKERR(1);
index = (LIS_INT *)malloc(27*A0->n*sizeof(LIS_INT));
if( index==NULL ) CHKERR(1);
value = (LIS_SCALAR *)malloc(27*A0->n*sizeof(LIS_SCALAR));
if( value==NULL ) CHKERR(1);
lis_matrix_get_range(A0,&is,&ie);
ctr = 0;
for(ii=is;ii<ie;ii++)
{
i = ii/(m*n);
j = (ii - i*m*n)/n;
k = ii - i*m*n - j*n;
for(si=-1;si<=1;si++) {
if( i+si>-1 && i+si<l ) {
for(sj=-1;sj<=1;sj++) {
if( j+sj>-1 && j+sj<m ) {
for(sk=-1;sk<=1;sk++) {
if( k+sk>-1 && k+sk<n ) {
jj = ii + si*m*n + sj*n + sk;
index[ctr] = jj;
if( jj==ii ) { value[ctr++] = 26.0;}
else { value[ctr++] = -1.0;}
}
}
}
}
}
}
ptr[ii-is+1] = ctr;
}
ptr[0] = 0;
err = lis_matrix_set_csr(ptr[ie-is],ptr,index,value,A0);
CHKERR(err);
err = lis_matrix_assemble(A0);
CHKERR(err);
n = A0->n;
gn = A0->gn;
nnz = A0->nnz;
np = A0->np-n;
#ifdef USE_MPI
MPI_Allreduce(&nnz,&i,1,LIS_MPI_INT,MPI_SUM,A0->comm);
nnzap = (double)i / (double)nprocs;
nnzt = ((double)nnz -nnzap)*((double)nnz -nnzap);
nnz = i;
MPI_Allreduce(&nnzt,&nnzs,1,MPI_DOUBLE,MPI_SUM,A0->comm);
nnzs = (nnzs / (double)nprocs)/nnzap;
MPI_Allreduce(&np,&i,1,LIS_MPI_INT,MPI_SUM,A0->comm);
np = i;
#endif
if( my_rank==0 )
{
#ifdef _LONG__LONG
printf("matrix size = %lld x %lld (%lld nonzero entries)\n",gn,gn,nnz);
printf("number of iterations = %lld\n\n",iter);
#else
printf("matrix size = %d x %d (%d nonzero entries)\n",gn,gn,nnz);
printf("number of iterations = %d\n\n",iter);
#endif
}
err = lis_vector_duplicate(A0,&x);
if( err ) CHKERR(err);
err = lis_vector_duplicate(A0,&b);
if( err ) CHKERR(err);
lis_matrix_get_range(A0,&is,&ie);
for(i=0;i<n;i++)
{
err = lis_vector_set_value(LIS_INS_VALUE,i+is,1.0,x);
}
for(i=0;i<n;i++)
{
lis_sort_id(A0->ptr[i],A0->ptr[i+1]-1,A0->index,A0->value);
}
/*
MPI version of VBR is not implemented.
DNS is also excluded to reduce memory usage.
*/
if (storage==0)
{
ss = 1;
se = 11;
}
else
{
ss = storage;
se = storage+1;
}
for (matrix_type=ss;matrix_type<se;matrix_type++)
{
if ( nprocs>1 && matrix_type==9 ) continue;
lis_matrix_duplicate(A0,&A);
lis_matrix_set_type(A,matrix_type);
err = lis_matrix_convert(A0,A);
if( err ) CHKERR(err);
comptime = 0.0;
commtime = 0.0;
for(i=0;i<iter;i++)
{
#ifdef USE_MPI
MPI_Barrier(A->comm);
time = lis_wtime();
lis_send_recv(A->commtable,x->value);
commtime += lis_wtime() - time;
#endif
time2 = lis_wtime();
lis_matvec(A,x,b);
comptime += lis_wtime() - time2;
}
lis_vector_nrm2(b,&val);
if( my_rank==0 )
{
flops = 2.0*nnz*iter*1.0e-6 / comptime;
#ifdef USE_MPI
#ifdef _LONG__DOUBLE
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#endif
#else
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, communication = %e sec, communication/computation = %3.3f %%, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,commtime,commtime/comptime*100,val);
#endif
#endif
#else
#ifdef _LONG__DOUBLE
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %Le\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#endif
#else
#ifdef _LONG__LONG
printf("matrix_type = %2lld (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#else
printf("matrix_type = %2d (%s), computation = %e sec, %8.3f MFLOPS, 2-norm = %e\n",matrix_type,lis_storagename2[matrix_type-1],comptime,flops,val);
#endif
#endif
#endif
}
lis_matrix_destroy(A);
}
lis_matrix_destroy(A0);
lis_vector_destroy(b);
lis_vector_destroy(x);
lis_finalize();
LIS_DEBUG_FUNC_OUT;
return 0;
}
