예제 #1
0
파일: lis_solver_gmres.c 프로젝트: rwl/lis
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;
}
예제 #2
0
파일: lis_solver_gmres.c 프로젝트: rwl/lis
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 lis_gmres_quad(LIS_SOLVER solver)
{
	LIS_MATRIX A;
	LIS_PRECON M;
	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_QUAD_PTR rnorm;

	LIS_REAL   bnrm2, nrm2, tol;
	LIS_INT iter,maxiter,n,output,conv;
	double times,ptimes;

	LIS_INT i,j,k,m;
	LIS_INT ii,i1,iiv,i1v,iih,i1h,jj;
	LIS_INT h_dim;
	LIS_INT cs,sn;

	LIS_DEBUG_FUNC_IN;

	A       = solver->A;
	M       = solver->precon;
	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];
	conv    = solver->options[LIS_OPTIONS_CONV_COND];
	h_dim   = m+1;
	ptimes  = 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_quad::h" );
	cs      = (m+1)*h_dim;
	sn      = (m+2)*h_dim;
	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;


	iter=0;
	while( iter<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
		{
			iter++;
			i++;
			ii  = i-1;
			i1  = i;
			iiv = i-1;
			i1v = i;
			iih = (i-1)*h_dim;
			i1h = i*h_dim;


			/* z = M^-1 v */
			times = lis_wtime();
			lis_psolve(solver, v[iiv], z);
			ptimes += lis_wtime()-times;

			/* v = Az */
			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);
			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]) * bnrm2;

			if( output )
			{
				if( output & LIS_PRINT_MEM ) solver->residual[iter] = nrm2;
				if( output & LIS_PRINT_OUT && A->my_rank==0 ) printf("iter: %5d  residual = %e\n", iter, nrm2);
			}

			if( tol >= nrm2 ) break;
		} while( i<m && iter <maxiter );

		/* Solve H*Y =S for upper triangular 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 + yv */
		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 */
		times = lis_wtime();
		lis_psolve(solver, z, r);
		ptimes += lis_wtime()-times;

		/* x = x + r */
		lis_vector_axpyex_mmm(one,r,x);

		if( tol >= nrm2 )
		{
			solver->retcode    = LIS_SUCCESS;
			solver->iter       = iter;
			solver->iter2      = 0;
			solver->resid      = nrm2;
			solver->ptimes     = ptimes;
			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      = iter+1;
	solver->iter2     = 0;
	solver->resid     = nrm2;
	lis_free(h);
	LIS_DEBUG_FUNC_OUT;
	return LIS_MAXITER;
}