LIS_INT lis_psolvet_saamg(LIS_SOLVER solver, LIS_VECTOR b, LIS_VECTOR x)
{
#if defined(USE_SAAMG)
LIS_INT n;
LIS_PRECON precon;
LIS_MATRIX A;
#ifdef USE_MPI
LIS_MPI_Fint comm;
#endif
LIS_DEBUG_FUNC_IN;
A = solver->A;
precon = solver->precon;
#ifdef USE_MPI
comm = MPI_Comm_c2f(A->comm);
n = b->np;
(*(void (*)())f_v_cycle_ptr)(b->value,x->value,precon->temp->value,&precon->level_num,
&comm,A->commtable->ws,A->commtable->wr,&n,&precon->wsize);
#else
n = b->n;
(*(void (*)())f_v_cycle_ptr)(&n,b->value,x->value,&precon->level_num,precon->temp->value);
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
#else
LIS_DEBUG_FUNC_IN;
lis_vector_copy(b,x);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
#endif
}
LIS_INT lis_matrix_solvet_msr(LIS_MATRIX A, LIS_VECTOR B, LIS_VECTOR X, LIS_INT flag)
{
LIS_INT i,j,n;
LIS_SCALAR t;
LIS_SCALAR *b,*x;
LIS_DEBUG_FUNC_IN;
n = A->n;
b = B->value;
x = X->value;
lis_vector_copy(B,X);
switch(flag)
{
case LIS_MATRIX_LOWER:
for(i=n-1;i>=0;i--)
{
x[i] = x[i] * A->WD->value[i];
for(j=A->U->index[i];j<A->U->index[i+1];j++)
{
x[A->U->index[j]] -= A->U->value[j] * x[i];
}
}
break;
case LIS_MATRIX_UPPER:
for(i=0;i<n;i++)
{
x[i] = x[i] * A->WD->value[i];
for(j=A->L->index[i];j<A->L->index[i+1];j++)
{
x[A->L->index[j]] -= A->L->value[j] * x[i];
}
}
break;
case LIS_MATRIX_SSOR:
for(i=0;i<n;i++)
{
t = x[i] * A->WD->value[i];
for(j=A->U->index[i];j<A->U->index[i+1];j++)
{
x[A->U->index[j]] -= A->U->value[j] * t;
}
}
for(i=n-1;i>=0;i--)
{
t = x[i] * A->WD->value[i];
x[i] = t;
for(j=A->L->index[i];j<A->L->index[i+1];j++)
{
x[A->L->index[j]] -= A->L->value[j] * t;
}
}
break;
}
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_solver_set_shadowresidual(LIS_SOLVER solver, LIS_VECTOR r0, LIS_VECTOR rs0)
{
unsigned long init[4]={0x123, 0x234, 0x345, 0x456}, length=4;
LIS_INT i,n,resid;
LIS_DEBUG_FUNC_IN;
resid = solver->options[LIS_OPTIONS_INIT_SHADOW_RESID];
if( resid==LIS_RANDOM )
{
n = solver->A->n;
init_by_array(init, length);
#ifdef USE_QUAD_PRECISION
if( solver->precision==LIS_PRECISION_DEFAULT )
#endif
{
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for(i=0;i<n;i++)
{
rs0->value[i] = genrand_real1();
}
}
#ifdef USE_QUAD_PRECISION
else
{
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for(i=0;i<n;i++)
{
rs0->value[i] = genrand_real1();
rs0->value_lo[i] = 0.0;
}
}
#endif
}
else
{
#ifdef USE_QUAD_PRECISION
if( solver->precision==LIS_PRECISION_DEFAULT )
#endif
{
lis_vector_copy(r0,rs0);
}
#ifdef USE_QUAD_PRECISION
else
{
lis_vector_copyex_mm(r0,rs0);
}
#endif
}
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
void lis_vector_copy_f(LIS_VECTOR_F *x, LIS_VECTOR_F *y, LIS_INT *ierr)
{
LIS_DEBUG_FUNC_IN;
*ierr = lis_vector_copy((LIS_VECTOR)LIS_V2P(x),(LIS_VECTOR)LIS_V2P(y));
if( *ierr ) return;
LIS_DEBUG_FUNC_OUT;
return;
}
LIS_INT lis_psolvet_none(LIS_SOLVER solver, LIS_VECTOR b, LIS_VECTOR x)
{
LIS_DEBUG_FUNC_IN;
#ifndef USE_QUAD_PRECISION
lis_vector_copy(b,x);
#else
if( solver->precision==LIS_PRECISION_DOUBLE )
{
lis_vector_copy(b,x);
}
else
{
lis_vector_copyex_mm(b,x);
}
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_psolve_ilut_bsr(LIS_SOLVER solver, LIS_VECTOR B, LIS_VECTOR X)
{
LIS_INT i,j,jj,nr,bnr,bs;
LIS_SCALAR w[9];
LIS_SCALAR *b,*x;
LIS_MATRIX_ILU L,U;
LIS_MATRIX_DIAG D;
LIS_PRECON precon;
/*
* LUx = b
* LU = (D + L*A) * (I + D^-1 * U*A)
*/
LIS_DEBUG_FUNC_IN;
precon = solver->precon;
L = precon->L;
U = precon->U;
D = precon->WD;
b = B->value;
x = X->value;
nr = solver->A->nr;
bnr = solver->A->bnr;
bs = bnr*bnr;
lis_vector_copy(B,X);
for(i=0; i<nr; i++)
{
for(j=0;j<L->nnz[i];j++)
{
jj = L->index[i][j];
lis_array_matvec(bnr,&L->value[i][bs*j],&x[bnr*jj],&x[bnr*i],LIS_SUB_VALUE);
}
}
for(i=nr-1; i>=0; i--)
{
for(j=0;j<U->nnz[i];j++)
{
jj = U->index[i][j];
lis_array_matvec(bnr,&U->value[i][bs*j],&x[bnr*jj],&x[bnr*i],LIS_SUB_VALUE);
}
/* lis_array_matvec(bnr,&D->value[bs*i],&x[bnr*i],w,LIS_INS_VALUE);*/
lis_array_invvec(bnr,&D->value[bs*i],&x[bnr*i],w);
memcpy(&x[bnr*i],w,bnr*sizeof(LIS_SCALAR));
}
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
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_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_psolvet_adds(LIS_SOLVER solver, LIS_VECTOR B, LIS_VECTOR X)
{
LIS_INT i,k,n,np,iter,ptype;
LIS_SCALAR *b,*x,*w,*r;
LIS_VECTOR W,R;
LIS_PRECON precon;
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;
iter = solver->options[LIS_OPTIONS_ADDS_ITER];
ptype = solver->options[LIS_OPTIONS_PRECON];
if( solver->precision==LIS_PRECISION_DEFAULT )
{
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_psolvet_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_matvect(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];
}
}
}
}
else
{
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_psolvet_xxx[ptype](solver,R,W);
for(i=0;i<n;i++)
{
x[i] += w[i];
}
if(k==iter) break;
X->precision = LIS_PRECISION_DEFAULT;
lis_matvect(precon->A,X,R);
X->precision = LIS_PRECISION_QUAD;
for(i=0;i<n;i++)
{
r[i] = b[i] - r[i];
}
}
}
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_orthomin(LIS_SOLVER solver)
{
LIS_Comm comm;
LIS_MATRIX A;
LIS_PRECON M;
LIS_VECTOR x;
LIS_VECTOR r,rtld,*p,*ap,*aptld;
LIS_SCALAR *dotsave;
LIS_SCALAR alpha, beta;
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;
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];
dotsave = (LIS_SCALAR *)lis_malloc( sizeof(LIS_SCALAR) * (m+1),"lis_orthomin::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_copy(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_dot(aptld[ip],aptld[ip0],&beta);
beta = -beta * dotsave[l-1];
lis_vector_axpy(beta,p[ip0] ,p[ip]);
lis_vector_axpy(beta,ap[ip0] ,ap[ip]);
lis_vector_axpy(beta,aptld[ip0],aptld[ip]);
}
for(l=m-1;l>0;l--)
{
dotsave[l] = dotsave[l-1];
}
lis_vector_dot(aptld[ip],aptld[ip],&dotsave[0]);
/* test breakdown */
if( dotsave[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
lis_free(dotsave);
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
dotsave[0] = 1.0/dotsave[0];
/* alpha = <rtld,Aptld[ip]> */
lis_vector_dot(rtld,aptld[ip],&alpha);
alpha = alpha * dotsave[0];
lis_vector_axpy( alpha,p[ip],x);
lis_vector_axpy(-alpha,ap[ip],r);
lis_vector_axpy(-alpha,aptld[ip],rtld);
/* 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_bicrstab(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_PRECON M;
LIS_VECTOR b,x;
LIS_VECTOR r,rtld, p, s, ap, ms, map, ams, z;
LIS_SCALAR alpha, beta, omega, rho, rho_old, tmpdot1, tmpdot2;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,n,output,conv;
double times,ptimes;
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];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptimes = 0.0;
rtld = solver->work[0];
r = solver->work[1];
s = solver->work[2];
ms = solver->work[3];
ams = solver->work[4];
p = solver->work[5];
ap = solver->work[6];
map = solver->work[7];
z = solver->work[8];
/* 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);
times = lis_wtime();
lis_psolve(solver, r, z);
ptimes += lis_wtime()-times;
lis_vector_copy(z,p);
lis_vector_dot(rtld,z,&rho_old);
for( iter=1; iter<=maxiter; iter++ )
{
/* ap = A * p */
/* map = M^-1 * ap */
/* tmpdot1 = <rtld,map> */
/* alpha = rho_old / tmpdot1 */
/* s = r - alpha*ap */
LIS_MATVEC(A,p,ap);
times = lis_wtime();
lis_psolve(solver, ap, map);
ptimes += lis_wtime()-times;
lis_vector_dot(rtld,map,&tmpdot1);
alpha = rho_old / tmpdot1;
lis_vector_axpyz(-alpha,ap,r,s);
/* Early check for tolerance */
lis_solver_get_residual[conv](s,solver,&nrm2);
if( nrm2 <= tol )
{
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);
}
lis_vector_axpy(alpha,p,x);
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptimes = ptimes;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* ms = z - alpha*map */
/* ams = A * ms */
/* tmpdot1 = <ams,s> */
/* tmpdot2 = <ams,ams> */
/* omega = tmpdot1 / tmpdot2 */
lis_vector_axpyz(-alpha,map,z,ms);
LIS_MATVEC(A,ms,ams);
lis_vector_dot(ams,s,&tmpdot1);
lis_vector_dot(ams,ams,&tmpdot2);
omega = tmpdot1 / tmpdot2;
/* x = x + alpha*p + omega*ms */
/* r = s - omega*ams */
lis_vector_axpy(alpha,p,x);
lis_vector_axpy(omega,ms,x);
lis_vector_axpyz(-omega,ams,s,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
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 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptimes = ptimes;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* z = M^-1 * r */
/* rho = <rtld,z> */
times = lis_wtime();
lis_psolve(solver, r, z);
ptimes += lis_wtime()-times;
lis_vector_dot(rtld,z,&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 / omega) */
/* p = z + beta*(p - omega*map) */
beta = (rho / rho_old) * (alpha / omega);
lis_vector_axpy(-omega,map,p);
lis_vector_xpay(z,beta,p);
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_bicr(LIS_SOLVER solver)
{
LIS_MATRIX A,At;
LIS_PRECON M;
LIS_VECTOR b,x;
LIS_VECTOR r,rtld, z,ztld,p, ptld, ap, map, az, aptld;
LIS_SCALAR alpha, beta, rho, rho_old, tmpdot1;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,n,output,conv;
double times,ptimes;
LIS_DEBUG_FUNC_IN;
A = solver->A;
At = 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];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptimes = 0.0;
r = solver->work[0];
rtld = solver->work[1];
z = solver->work[2];
ztld = solver->work[3];
p = solver->work[4];
ptld = solver->work[5];
ap = solver->work[6];
az = solver->work[7];
map = solver->work[8];
aptld = solver->work[9];
/* 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_psolve(solver, r, z);
lis_psolvet(solver, rtld, ztld);
lis_vector_copy(z,p);
lis_vector_copy(ztld,ptld);
LIS_MATVEC(A,z,ap);
lis_vector_dot(ap,ztld,&rho_old);
for( iter=1; iter<=maxiter; iter++ )
{
/* aptld = A^T * ptld */
/* map = M^-1 * ap */
LIS_MATVECT(A,ptld,aptld);
times = lis_wtime();
lis_psolve(solver, ap, map);
ptimes += lis_wtime()-times;
/* tmpdot1 = <map,aptld> */
lis_vector_dot(map,aptld,&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_old / tmpdot1 */
/* x = x + alpha*p */
/* r = r - alpha*ap */
alpha = rho_old / tmpdot1;
lis_vector_axpy(alpha,p,x);
lis_vector_axpy(-alpha,ap,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->residual[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->ptimes = ptimes;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* rtld = rtld - alpha*aptld */
/* z = z - alpha*map */
/* ztld = M^-T * rtld */
/* az = A * z */
/* rho = <az,ztld> */
lis_vector_axpy(-alpha,aptld,rtld);
lis_vector_axpy(-alpha,map,z);
times = lis_wtime();
lis_psolvet(solver, rtld, ztld);
ptimes += lis_wtime()-times;
LIS_MATVEC(A,z,az);
lis_vector_dot(az,ztld,&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 */
/* p = z + beta*p */
/* ptld = ztld + beta*ptld */
/* ap = az + beta*ap */
beta = rho / rho_old;
lis_vector_xpay(z,beta,p);
lis_vector_xpay(ztld,beta,ptld);
lis_vector_xpay(az,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_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_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 lis_psolvet_hybrid(LIS_SOLVER solver, LIS_VECTOR B, LIS_VECTOR X)
{
LIS_VECTOR xx;
LIS_SOLVER solver2;
LIS_INT nsolver;
LIS_PRECON precon;
/*
* Mx = b
* M = A
*/
LIS_DEBUG_FUNC_IN;
precon = solver->precon;
solver2 = precon->solver;
xx = precon->solver->x;
nsolver = solver2->options[LIS_OPTIONS_SOLVER];
solver2->b = B;
LIS_MATVEC = lis_matvect;
LIS_MATVECT = lis_matvec;
if( solver2->options[LIS_OPTIONS_INITGUESS_ZEROS] )
{
#ifdef USE_QUAD_PRECISION
if( solver->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_set_all(0,xx);
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_set_allex_nm(0,xx);
}
#endif
}
else
{
#ifdef USE_QUAD_PRECISION
if( solver->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_copy(B,xx);
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_copyex_mm(B,xx);
}
#endif
}
/* execute solver */
lis_solver_execute[nsolver](solver2);
#ifdef USE_QUAD_PRECISION
if( solver->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_copy(solver2->x,X);
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_copyex_mm(solver2->x,X);
}
#endif
LIS_MATVEC = lis_matvec;
LIS_MATVECT = lis_matvect;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
LIS_INT lis_ecg(LIS_ESOLVER esolver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_SCALAR evalue;
LIS_INT emaxiter;
LIS_REAL tol;
LIS_INT iter,iter3,nsolver,i,j,output;
LIS_INT nprocs,my_rank;
LIS_REAL nrm2,resid,resid3;
LIS_SCALAR lshift;
LIS_VECTOR b,D,r,w,p,Aw,Ax,Ap,ones,Ds;
LIS_SCALAR *SA, *SB, *SW, *v3, *SAv3, *SBv3, *z3, *q3, *SBz3, evalue3, ievalue3;
LIS_SOLVER solver;
LIS_PRECON precon;
LIS_MATRIX A0;
LIS_VECTOR x0,z,q;
double times,itimes,ptimes,p_c_times,p_i_times;
LIS_INT nsol, precon_type;
char solvername[128], preconname[128];
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);
SA = (LIS_SCALAR *)lis_malloc(3*3*sizeof(LIS_SCALAR), "lis_ecg::SA");
SB = (LIS_SCALAR *)lis_malloc(3*3*sizeof(LIS_SCALAR), "lis_ecg::SB");
SW = (LIS_SCALAR *)lis_malloc(3*3*sizeof(LIS_SCALAR), "lis_ecg::SW");
v3 = (LIS_SCALAR *)lis_malloc(3*sizeof(LIS_SCALAR), "lis_ecg::v3");
SAv3 = (LIS_SCALAR *)lis_malloc(3*sizeof(LIS_SCALAR), "lis_ecg::SAv3");
SBv3 = (LIS_SCALAR *)lis_malloc(3*sizeof(LIS_SCALAR), "lis_ecg::SBv3");
SBz3 = (LIS_SCALAR *)lis_malloc(3*sizeof(LIS_SCALAR), "lis_ecg::SBz3");
z3 = (LIS_SCALAR *)lis_malloc(3*sizeof(LIS_SCALAR), "lis_ecg::z3");
q3 = (LIS_SCALAR *)lis_malloc(3*sizeof(LIS_SCALAR), "lis_ecg::q3");
b = esolver->work[0];
D = esolver->work[1];
Ds = esolver->work[2];
r = esolver->work[3];
w = esolver->work[4];
p = esolver->work[5];
Aw = esolver->work[6];
Ax = esolver->work[7];
Ap = esolver->work[8];
lis_vector_set_all(1.0,b);
lis_vector_nrm2(b, &nrm2);
lis_vector_scale(1/nrm2, b);
lis_solver_create(&solver);
lis_solver_set_option("-i bicg -p ilu",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);
lis_solve(A, b, x, solver);
lis_vector_copy(b,Ax);
lis_vector_nrm2(x, &nrm2);
lis_vector_set_all(0.0,p);
lis_vector_set_all(0.0,Ap);
lis_precon_create(solver, &precon);
solver->precon = precon;
iter=0;
while (iter<emaxiter)
{
iter = iter + 1;
lis_vector_dot(x,Ax,&evalue);
lis_vector_axpyz(-(evalue),x,Ax,r);
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;
lis_psolve(solver, x, w);
lis_vector_copy(x,Aw);
lis_vector_nrm2(w, &nrm2);
lis_vector_dot(w,Aw,&SA[0]);
lis_vector_dot(x,Aw,&SA[3]);
lis_vector_dot(p,Aw,&SA[6]);
SA[1] = SA[3];
lis_vector_dot(x,Ax,&SA[4]);
lis_vector_dot(p,Ax,&SA[7]);
SA[2] = SA[6];
SA[5] = SA[7];
lis_vector_dot(p,Ap,&SA[8]);
lis_vector_dot(w,w,&SB[0]);
lis_vector_dot(x,w,&SB[3]);
lis_vector_dot(p,w,&SB[6]);
SB[1] = SB[3];
lis_vector_dot(x,x,&SB[4]);
lis_vector_dot(p,x,&SB[7]);
SB[2] = SB[6];
SB[5] = SB[7];
lis_vector_dot(p,p,&SB[8]);
lis_array_set_all(3, 1.0, v3);
iter3=0;
while (iter3<emaxiter)
{
iter3 = iter3 + 1;
lis_array_nrm2(3, v3, &nrm2);
lis_array_scale(3, 1/nrm2, v3);
lis_array_matvec(3, SB, v3, SBv3, LIS_INS_VALUE);
lis_array_invvec(3, SA, SBv3, z3);
lis_array_dot2(3, SBv3, z3, &ievalue3);
if (ievalue3==0)
{
printf("ievalue3 is zero\n");
lis_precon_destroy(precon);
lis_solver_destroy(solver);
esolver->iter = iter;
esolver->resid = resid;
esolver->evalue[0] = evalue;
if (lshift != 0) lis_matrix_shift_diagonal(A, -lshift);
lis_free(SA);
lis_free(SB);
lis_free(SW);
lis_free(v3);
lis_free(SAv3);
lis_free(SBv3);
lis_free(SBz3);
lis_free(z3);
lis_free(q3);
return LIS_BREAKDOWN;
}
lis_array_axpyz(3, -ievalue3, SBv3, z3, q3);
lis_array_nrm2(3, q3, &resid3);
resid3 = fabs(resid3 / ievalue3);
if (resid3<1e-12) break;
lis_array_copy(3,z3,v3);
}
evalue3 = 1 / ievalue3;
lis_vector_scale(v3[0],w);
lis_vector_axpy(v3[2],p,w);
lis_vector_xpay(w,v3[1],x);
lis_vector_copy(w,p);
lis_vector_scale(v3[0],Aw);
lis_vector_axpy(v3[2],Ap,Aw);
lis_vector_xpay(Aw,v3[1],Ax);
lis_vector_copy(Aw,Ap);
lis_vector_nrm2(x,&nrm2);
lis_vector_scale(1/nrm2,x);
lis_vector_scale(1/nrm2,Ax);
lis_vector_nrm2(p,&nrm2);
lis_vector_scale(1/nrm2,p);
lis_vector_scale(1/nrm2,Ap);
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_precon_destroy(precon);
lis_solver_destroy(solver);
esolver->iter = iter;
esolver->resid = resid;
esolver->evalue[0] = evalue;
if (lshift != 0) lis_matrix_shift_diagonal(A, -lshift);
lis_free(SA);
lis_free(SB);
lis_free(SW);
lis_free(v3);
lis_free(SAv3);
lis_free(SBv3);
lis_free(SBz3);
lis_free(z3);
lis_free(q3);
if (resid<tol)
{
esolver->retcode = LIS_SUCCESS;
return LIS_SUCCESS;
}
else
{
esolver->retcode = LIS_MAXITER;
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_epi(LIS_ESOLVER esolver)
{
LIS_Comm comm;
LIS_MATRIX A;
LIS_VECTOR v,y,q;
LIS_SCALAR theta;
LIS_INT emaxiter;
LIS_REAL tol;
LIS_INT iter,output;
LIS_SCALAR oshift,ishift;
LIS_REAL nrm2,resid;
LIS_DEBUG_FUNC_IN;
comm = LIS_COMM_WORLD;
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
tol = esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN];
output = esolver->options[LIS_EOPTIONS_OUTPUT];
#ifdef _COMPLEX
oshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN] + esolver->params[LIS_EPARAMS_SHIFT_IM - LIS_EOPTIONS_LEN] * _Complex_I;
#else
oshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN];
#endif
A = esolver->A;
v = esolver->x;
if (esolver->options[LIS_EOPTIONS_INITGUESS_ONES] )
{
lis_vector_set_all(1.0,v);
}
y = esolver->work[0];
q = esolver->work[1];
if ( esolver->ishift != 0.0 ) oshift = ishift;
if ( oshift != 0.0 ) lis_matrix_shift_diagonal(A, oshift);
if( output )
{
#ifdef _COMPLEX
lis_printf(comm,"shift : (%e, %e)\n", (double)creal(oshift), (double)cimag(oshift));
#else
lis_printf(comm,"shift : %e\n", (double)oshift);
#endif
}
iter=0;
while (iter<emaxiter)
{
iter = iter+1;
/* v = v / ||v||_2 */
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
/* y = A * v */
lis_matvec(A,v,y);
/* theta = <v,y> */
lis_vector_dot(v, y, &theta);
/* resid = ||y - theta * v||_2 / |theta| */
lis_vector_axpyz(-theta, v, y, q);
lis_vector_nrm2(q, &resid);
resid = resid / fabs(theta);
/* v = y */
lis_vector_copy(y, v);
/* convergence check */
if( output )
{
if( output & LIS_EPRINT_MEM ) esolver->rhistory[iter] = resid;
if( output & LIS_EPRINT_OUT ) lis_print_rhistory(comm,iter,resid);
}
if( tol >= resid )
{
esolver->retcode = LIS_SUCCESS;
esolver->iter[0] = iter;
esolver->resid[0] = resid;
esolver->evalue[0] = theta + oshift;
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
if ( oshift != 0.0 ) lis_matrix_shift_diagonal(A, -oshift);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
}
esolver->retcode = LIS_MAXITER;
esolver->iter[0] = iter;
esolver->resid[0] = resid;
esolver->evalue[0] = theta + oshift;
lis_vector_nrm2(v, &nrm2);
lis_vector_scale(1.0/nrm2, v);
if ( oshift != 0.0 ) lis_matrix_shift_diagonal(A, -oshift);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_esolve(LIS_MATRIX A, LIS_VECTOR x, LIS_SCALAR *evalue0, LIS_ESOLVER esolver)
{
LIS_INT nesolver,niesolver,emaxiter;
LIS_SCALAR *evalue;
LIS_VECTOR *evector;
LIS_SCALAR *resid;
LIS_SCALAR *rhistory;
LIS_INT *iter,*iter2;
LIS_INT err;
LIS_INT output;
LIS_INT ss, mode;
double time;
double gshift;
LIS_INT estorage,eblock;
LIS_MATRIX B;
LIS_INT eprecision;
LIS_VECTOR xx;
LIS_DEBUG_FUNC_IN;
/* begin parameter check */
err = lis_matrix_check(A,LIS_MATRIX_CHECK_ALL);
if( err ) return err;
if( x==NULL )
{
LIS_SETERR(LIS_ERR_ILL_ARG,"vector x is undefined\n");
return LIS_ERR_ILL_ARG;
}
if( A->n!=x->n )
{
return LIS_ERR_ILL_ARG;
}
if( A->gn<=0 )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Size n(=%d) of matrix A is less than 0\n",A->gn);
return LIS_ERR_ILL_ARG;
}
nesolver = esolver->options[LIS_EOPTIONS_ESOLVER];
niesolver = esolver->options[LIS_EOPTIONS_INNER_ESOLVER];
ss = esolver->options[LIS_EOPTIONS_SUBSPACE];
mode = esolver->options[LIS_EOPTIONS_MODE];
emaxiter = esolver->options[LIS_EOPTIONS_MAXITER];
gshift = esolver->params[LIS_EPARAMS_SHIFT - LIS_EOPTIONS_LEN];
output = esolver->options[LIS_EOPTIONS_OUTPUT];
estorage = esolver->options[LIS_EOPTIONS_STORAGE];
eblock = esolver->options[LIS_EOPTIONS_STORAGE_BLOCK];
eprecision = esolver->options[LIS_EOPTIONS_PRECISION];
esolver->eprecision = eprecision;
if( nesolver < 1 || nesolver > LIS_ESOLVER_LEN )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_ESOLVER is %d (Set between 1 to %d)\n",nesolver, LIS_ESOLVER_LEN);
return LIS_ERR_ILL_ARG;
}
if( niesolver < 1 || niesolver > 7 )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 1 to 7)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_SI && niesolver > 4 )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 1 to 4 for Subspace)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_LI && niesolver == LIS_ESOLVER_PI )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 2 to 7 for Lanczos)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_AI && (( niesolver == LIS_ESOLVER_PI ) || ( niesolver == LIS_ESOLVER_CG) || ( niesolver == LIS_ESOLVER_JD)) )
{
LIS_SETERR1(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_INNER_ESOLVER is %d (Set between 2 to 4 or 6 for Arnoldi)\n", niesolver);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_SI && ss > A->gn )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_SUBSPACE is %d (Set less than or equal to matrix size %d for Subspace)\n", ss, A->gn);
return LIS_ERR_ILL_ARG;
}
if (( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_LI || esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_AI ) && ss > A->gn )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_SUBSPACE is %d (Set less than or equal to matrix size %d for Lanczos and Arnoldi)\n", ss, A->gn);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == LIS_ESOLVER_SI && mode >= ss )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_MODE is %d (Set less than subspace size %d for Subspace)\n", mode, ss);
return LIS_ERR_ILL_ARG;
}
if ( esolver->options[LIS_EOPTIONS_ESOLVER] == ( LIS_ESOLVER_LI || LIS_ESOLVER_AI ) && mode >= ss )
{
LIS_SETERR2(LIS_ERR_ILL_ARG,"Parameter LIS_EOPTIONS_MODE is %d (Set less than subspace size %d for Lanczos or Arnoldi)\n", mode, ss);
return LIS_ERR_ILL_ARG;
}
#ifdef USE_QUAD_PRECISION
if( eprecision==LIS_PRECISION_QUAD && lis_esolver_execute_quad[nesolver]==NULL )
{
LIS_SETERR1(LIS_ERR_NOT_IMPLEMENTED,"Quad precision eigensolver %s is not implemented\n",lis_esolvername[nesolver]);
return LIS_ERR_NOT_IMPLEMENTED;
}
else if( eprecision==LIS_PRECISION_SWITCH && lis_esolver_execute_switch[nesolver]==NULL )
{
LIS_SETERR1(LIS_ERR_NOT_IMPLEMENTED,"Switch esolver %s is not implemented\n",lis_esolvername[nesolver]);
return LIS_ERR_NOT_IMPLEMENTED;
}
if( esolver->options[LIS_EOPTIONS_SWITCH_MAXITER]==-1 )
{
esolver->options[LIS_EOPTIONS_SWITCH_MAXITER] = emaxiter;
}
#endif
/* create eigenvalue array */
if( esolver->evalue ) lis_free(esolver->evalue);
evalue = (LIS_SCALAR *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::evalue");
if( evalue==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
evalue[0] = 1.0;
evalue[ss-1] = 1.0;
/* create residual norm array */
if( esolver->resid ) lis_free(esolver->resid);
resid = (LIS_SCALAR *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::resid");
if( resid==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
/* create number of iterations array */
if( esolver->iter ) lis_free(esolver->iter);
iter = (LIS_INT *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::iter");
if( iter==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
/* create quad precision number of iterations array */
if( esolver->iter2 ) lis_free(esolver->iter2);
iter2 = (LIS_INT *)lis_malloc((ss+2)*sizeof(LIS_SCALAR),"lis_esolve::iter2");
if( iter2==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_SCALAR));
esolver->retcode = err;
return err;
}
/* create initial vector */
#ifndef USE_QUAD_PRECISION
err = lis_vector_duplicate(A,&xx);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
err = lis_vector_duplicate(A,&xx);
}
else
{
err = lis_vector_duplicateex(LIS_PRECISION_QUAD,A,&xx);
}
#endif
if( err )
{
esolver->retcode = err;
return err;
}
if( esolver->options[LIS_EOPTIONS_INITGUESS_ONES] )
{
if( output ) lis_printf(A->comm,"initial vector x : 1\n");
#ifndef USE_QUAD_PRECISION
lis_vector_set_all(1.0,xx);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
lis_vector_set_all(1.0,xx);
}
else
{
lis_vector_set_allex_nm(1.0,xx);
}
#endif
}
else
{
if( output ) lis_printf(A->comm,"initial vector x : user defined\n");
#ifndef USE_QUAD_PRECISION
lis_vector_copy(x,xx);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
lis_vector_copy(x,xx);
}
else
{
lis_vector_copyex_nm(x,xx);
}
#endif
}
/* global shift */
if ( output ) if( A->my_rank==0 ) printf("shift : %e\n", gshift);
/* create eigenvector array */
if( esolver->evector ) lis_free(esolver->evector);
evector = (LIS_VECTOR *)lis_malloc((ss+2)*sizeof(LIS_VECTOR),"lis_esolve::evector");
if( evector==NULL )
{
LIS_SETERR_MEM((ss+2)*sizeof(LIS_VECTOR));
esolver->retcode = err;
return err;
}
/* create residual history array */
if( esolver->rhistory ) lis_free(esolver->rhistory);
rhistory = (LIS_SCALAR *)lis_malloc((emaxiter+2)*sizeof(LIS_SCALAR),"lis_esolve::rhistory");
if( rhistory==NULL )
{
LIS_SETERR_MEM((emaxiter+2)*sizeof(LIS_SCALAR));
lis_vector_destroy(xx);
esolver->retcode = err;
return err;
}
/* convert matrix */
if( estorage>0 && A->matrix_type!=estorage )
{
err = lis_matrix_duplicate(A,&B);
if( err ) return err;
lis_matrix_set_blocksize(B,eblock,eblock,NULL,NULL);
lis_matrix_set_type(B,estorage);
err = lis_matrix_convert(A,B);
if( err ) return err;
lis_matrix_storage_destroy(A);
lis_matrix_DLU_destroy(A);
lis_matrix_diag_destroy(A->WD);
if( A->l2g_map ) lis_free( A->l2g_map );
if( A->commtable ) lis_commtable_destroy( A->commtable );
if( A->ranges ) lis_free( A->ranges );
err = lis_matrix_copy_struct(B,A);
if( err ) return err;
lis_free(B);
}
esolver->A = A;
esolver->evalue = evalue;
esolver->x = x;
esolver->evector = evector;
rhistory[0] = 1.0;
esolver->rhistory = rhistory;
esolver->resid = resid;
esolver->iter = iter;
esolver->iter2 = iter2;
if( A->my_rank==0 )
{
#ifdef _LONG__DOUBLE
if ( output ) printf("precision : long double\n");
#else
if ( output ) printf("precision : %s\n", lis_eprecisionname[eprecision]);
#endif
#ifdef _LONG__LONG
if ( output ) printf("eigensolver : %s\n", lis_esolvername[nesolver]);
#else
if ( output ) printf("eigensolver : %s\n", lis_esolvername[nesolver]);
#endif
}
if( A->my_rank==0 )
{
#ifdef _LONG__DOUBLE
if ( output ) printf("convergence condition : ||lx-Ax||_2 <= %6.1Le * ||lx||_2\n", esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN]);
#else
if ( output ) printf("convergence condition : ||lx-Ax||_2 <= %6.1e * ||lx||_2\n", esolver->params[LIS_EPARAMS_RESID - LIS_EOPTIONS_LEN]);
#endif
}
if( A->my_rank==0 )
{
if( A->matrix_type==LIS_MATRIX_BSR || A->matrix_type==LIS_MATRIX_BSC )
{
#ifdef _LONG__LONG
if ( output ) printf("matrix storage format : %s(%lld x %lld)\n", lis_estoragename[A->matrix_type-1],eblock,eblock);
#else
if ( output ) printf("matrix storage format : %s(%d x %d)\n", lis_estoragename[A->matrix_type-1],eblock,eblock);
#endif
}
else
{
if ( output ) printf("matrix storage format : %s\n", lis_estoragename[A->matrix_type-1]);
}
}
time = lis_wtime();
esolver->ptime = 0;
esolver->itime = 0;
esolver->p_c_time = 0;
esolver->p_i_time = 0;
if (gshift != 0.0) lis_matrix_shift_diagonal(A, gshift);
/* create work vector */
err = lis_esolver_malloc_work[nesolver](esolver);
if( err )
{
lis_vector_destroy(xx);
esolver->retcode = err;
return err;
}
esolver->x = xx;
esolver->xx = x;
/* execute esolver */
#ifndef USE_QUAD_PRECISION
err = lis_esolver_execute[nesolver](esolver);
#else
if( eprecision==LIS_PRECISION_DOUBLE )
{
err = lis_esolver_execute[nesolver](esolver);
}
else if( eprecision==LIS_PRECISION_QUAD )
{
err = lis_esolver_execute_quad[nesolver](esolver);
}
else if( eprecision==LIS_PRECISION_SWITCH )
{
err = lis_esolver_execute_switch[nesolver](esolver);
}
#endif
esolver->retcode = err;
*evalue0 = esolver->evalue[0];
lis_vector_copy(esolver->x, x);
esolver->time = lis_wtime() - time;
lis_matrix_shift_diagonal(A, -gshift);
if( A->my_rank==0 )
{
if( err )
{
#ifdef _LONG__LONG
if ( output ) printf("eigensolver status : %s(code=%lld)\n\n",lis_ereturncode[err],err);
#else
if ( output ) printf("eigensolver status : %s(code=%d)\n\n",lis_ereturncode[err],err);
#endif
}
else
{
if ( output ) printf("eigensolver status : normal end\n\n");
}
}
if( eprecision==LIS_PRECISION_DOUBLE )
{
esolver->iter2[mode] = esolver->iter[mode];
}
else if( eprecision==LIS_PRECISION_QUAD )
{
esolver->iter2[mode] = 0;
}
lis_vector_destroy(xx);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
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_bicgstab(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_PRECON M;
LIS_VECTOR b,x;
LIS_VECTOR r,rtld, t,p,v, s, phat, shat;
LIS_SCALAR alpha, beta, omega, rho, rho_old, tmpdot1, tmpdot2;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,n,output,conv;
double times,ptimes;
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];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptimes = 0.0;
rtld = solver->work[0];
r = solver->work[1];
s = solver->work[1];
t = solver->work[2];
p = solver->work[3];
v = solver->work[4];
phat = solver->work[5];
shat = solver->work[6];
alpha = (LIS_SCALAR)1.0;
omega = (LIS_SCALAR)1.0;
rho_old = (LIS_SCALAR)1.0;
lis_vector_set_all(0.0,p);
lis_vector_set_all(0.0,phat);
lis_vector_set_all(0.0,s);
lis_vector_set_all(0.0,shat);
/* 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);
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;
}
if( iter==1 )
{
lis_vector_copy(r,p);
}
else
{
/* beta = (rho / rho_old) * (alpha / omega) */
beta = (rho / rho_old) * (alpha / omega);
/* p = r + beta*(p - omega*v) */
lis_vector_axpy(-omega,v,p);
lis_vector_xpay(r,beta,p);
}
/* phat = M^-1 * p */
times = lis_wtime();
lis_psolve(solver, p, phat);
ptimes += lis_wtime()-times;
/* v = A * phat */
LIS_MATVEC(A,phat,v);
/* tmpdot1 = <rtld,v> */
lis_vector_dot(rtld,v,&tmpdot1);
/* test breakdown */
/* */
/* alpha = rho / tmpdot1 */
alpha = rho / tmpdot1;
/* s = r - alpha*v */
lis_vector_axpy(-alpha,v,r);
/* Early check for tolerance */
lis_solver_get_residual[conv](s,solver,&nrm2);
/* lis_vector_nrm2(s,&nrm2);
nrm2 = nrm2 * bnrm2;*/
if( nrm2 <= tol )
{
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);
}
lis_vector_axpy(alpha,phat,x);
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptimes = ptimes;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* shat = M^-1 * s */
times = lis_wtime();
lis_psolve(solver, s, shat);
ptimes += lis_wtime()-times;
/* t = A * shat */
LIS_MATVEC(A,shat,t);
/* tmpdot1 = <t,s> */
/* tmpdot2 = <t,t> */
/* omega = tmpdot1 / tmpdot2 */
lis_vector_dot(t,s,&tmpdot1);
lis_vector_dot(t,t,&tmpdot2);
omega = tmpdot1 / tmpdot2;
/* x = x + alpha*phat + omega*shat */
lis_vector_axpy(alpha,phat,x);
lis_vector_axpy(omega,shat,x);
/* r = s - omega*t */
lis_vector_axpy(-omega,t,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
/* lis_vector_nrm2(r,&nrm2);
nrm2 = nrm2 * 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 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptimes = ptimes;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
if( omega==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
rho_old = rho;
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
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_psolvet_ilut_csr(LIS_SOLVER solver, LIS_VECTOR B, LIS_VECTOR X)
{
#ifdef _OPENMP
LIS_INT i,j,jj,n;
LIS_INT is,ie,my_rank,nprocs;
LIS_SCALAR *b,*x;
LIS_MATRIX_ILU L,U;
LIS_VECTOR D;
LIS_PRECON precon;
LIS_QUAD_DECLAR;
#ifdef USE_QUAD_PRECISION
LIS_SCALAR *xl;
#endif
LIS_DEBUG_FUNC_IN;
precon = solver->precon;
L = precon->L;
U = precon->U;
D = precon->D;
b = B->value;
x = X->value;
#ifdef USE_QUAD_PRECISION
xl = X->value_lo;
#endif
n = solver->A->n;
nprocs = omp_get_max_threads();
#ifdef USE_QUAD_PRECISION
if( B->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_copy(B,X);
#pragma omp parallel private(i,j,jj,is,ie,my_rank)
{
my_rank = omp_get_thread_num();
LIS_GET_ISIE(my_rank,nprocs,n,is,ie);
for(i=is;i<ie;i++)
{
x[i] = D->value[i]*x[i];
for(j=0;j<U->nnz[i];j++)
{
jj = U->index[i][j];
x[jj] -= U->value[i][j] * x[i];
}
}
for(i=ie-1;i>=is;i--)
{
for(j=0;j<L->nnz[i];j++)
{
jj = L->index[i][j];
x[jj] -= L->value[i][j] * x[i];
}
}
}
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_copyex_mm(B,X);
nprocs = omp_get_max_threads();
#ifndef USE_SSE2
#pragma omp parallel private(i,j,jj,is,ie,my_rank,p1,p2,tq,bhi,blo,chi,clo,sh,sl,th,tl,eh,el)
#else
#pragma omp parallel private(i,j,jj,is,ie,my_rank,bh,ch,sh,wh,th,bl,cl,sl,wl,tl,p1,p2,t0,t1,t2,eh)
#endif
{
my_rank = omp_get_thread_num();
LIS_GET_ISIE(my_rank,nprocs,n,is,ie);
for(i=is;i<ie;i++)
{
#ifndef USE_SSE2
LIS_QUAD_MULD(x[i],xl[i],x[i],xl[i],D->value[i]);
#else
LIS_QUAD_MULD_SSE2(x[i],xl[i],x[i],xl[i],D->value[i]);
#endif
/* x[i] = D->value[i]*x[i];*/
for(j=0;j<U->nnz[i];j++)
{
jj = U->index[i][j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-U->value[i][j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-U->value[i][j]);
#endif
/* x[jj] -= U->value[i][j] * x[i];*/
}
}
for(i=ie-1;i>=is;i--)
{
for(j=0;j<L->nnz[i];j++)
{
jj = L->index[i][j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-L->value[i][j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-L->value[i][j]);
#endif
/* x[jj] -= L->value[i][j] * x[i];*/
}
}
}
}
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
#else
LIS_INT i,j,jj,n;
LIS_SCALAR *b,*x;
LIS_MATRIX_ILU L,U;
LIS_VECTOR D;
LIS_PRECON precon;
LIS_QUAD_DECLAR;
#ifdef USE_QUAD_PRECISION
LIS_SCALAR *xl;
#endif
LIS_DEBUG_FUNC_IN;
precon = solver->precon;
L = precon->L;
U = precon->U;
D = precon->D;
b = B->value;
x = X->value;
#ifdef USE_QUAD_PRECISION
xl = X->value_lo;
#endif
n = solver->A->n;
#ifdef USE_QUAD_PRECISION
if( B->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_copy(B,X);
for(i=0; i<n; i++)
{
x[i] = D->value[i]*x[i];
for(j=0;j<U->nnz[i];j++)
{
jj = U->index[i][j];
x[jj] -= U->value[i][j] * x[i];
}
}
for(i=n-1; i>=0; i--)
{
for(j=0;j<L->nnz[i];j++)
{
jj = L->index[i][j];
x[jj] -= L->value[i][j] * x[i];
}
}
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_copy(B,X);
for(i=0; i<n; i++)
{
#ifndef USE_SSE2
LIS_QUAD_MULD(x[i],xl[i],x[i],xl[i],D->value[i]);
#else
LIS_QUAD_MULD_SSE2(x[i],xl[i],x[i],xl[i],D->value[i]);
#endif
/* x[i] = D->value[i]*x[i];*/
for(j=0;j<U->nnz[i];j++)
{
jj = U->index[i][j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-U->value[i][j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-U->value[i][j]);
#endif
/* x[jj] -= U->value[i][j] * x[i];*/
}
}
for(i=n-1; i>=0; i--)
{
for(j=0;j<L->nnz[i];j++)
{
jj = L->index[i][j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-L->value[i][j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],x[i],xl[i],-L->value[i][j]);
#endif
/* x[jj] -= L->value[i][j] * x[i];*/
}
}
}
#endif
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
#endif
}
LIS_INT lis_matrix_solvet_csr(LIS_MATRIX A, LIS_VECTOR B, LIS_VECTOR X, LIS_INT flag)
{
LIS_INT i,j,jj,n;
LIS_SCALAR t;
LIS_SCALAR *x;
#ifdef _OPENMP
LIS_INT is,ie,my_rank,nprocs;
#endif
#ifdef USE_QUAD_PRECISION
LIS_QUAD w1,w2;
LIS_SCALAR *xl;
#endif
LIS_QUAD_DECLAR;
LIS_DEBUG_FUNC_IN;
n = A->n;
x = X->value;
#ifdef USE_QUAD_PRECISION
xl = X->value_lo;
#endif
#ifdef USE_QUAD_PRECISION
if( B->precision==LIS_PRECISION_DEFAULT )
{
#endif
lis_vector_copy(B,X);
#ifdef USE_QUAD_PRECISION
}
else
{
lis_vector_copyex_mm(B,X);
}
#endif
switch(flag)
{
case LIS_MATRIX_LOWER:
for(i=0;i<n;i++)
{
x[i] = x[i] * A->WD->value[i];
for(j=A->U->ptr[i];j<A->U->ptr[i+1];j++)
{
x[A->U->index[j]] -= A->U->value[j] * x[i];
}
}
break;
case LIS_MATRIX_UPPER:
for(i=n-1;i>=0;i--)
{
x[i] = x[i] * A->WD->value[i];
for(j=A->L->ptr[i];j<A->L->ptr[i+1];j++)
{
x[A->L->index[j]] -= A->L->value[j] * x[i];
}
}
break;
case LIS_MATRIX_SSOR:
#ifdef USE_QUAD_PRECISION
if( B->precision==LIS_PRECISION_DEFAULT )
{
#endif
#ifdef _OPENMP
nprocs = omp_get_max_threads();
#pragma omp parallel private(i,j,jj,t,is,ie,my_rank)
{
my_rank = omp_get_thread_num();
LIS_GET_ISIE(my_rank,nprocs,n,is,ie);
for(i=is;i<ie;i++)
{
t = x[i] * A->WD->value[i];
for(j=A->U->ptr[i];j<A->U->ptr[i+1];j++)
{
jj = A->U->index[j];
if( jj<is || jj>=ie ) continue;
x[jj] -= A->U->value[j] * t;
}
}
for(i=ie-1;i>=is;i--)
{
t = x[i] * A->WD->value[i];
x[i] = t;
for(j=A->L->ptr[i];j<A->L->ptr[i+1];j++)
{
jj = A->L->index[j];
if( jj<is ) continue;
x[jj] -= A->L->value[j] * t;
}
}
}
#else
for(i=0;i<n;i++)
{
t = x[i] * A->WD->value[i];
for(j=A->U->ptr[i];j<A->U->ptr[i+1];j++)
{
x[A->U->index[j]] -= A->U->value[j] * t;
}
}
for(i=n-1;i>=0;i--)
{
t = x[i] * A->WD->value[i];
x[i] = t;
for(j=A->L->ptr[i];j<A->L->ptr[i+1];j++)
{
x[A->L->index[j]] -= A->L->value[j] * t;
}
}
#endif
#ifdef USE_QUAD_PRECISION
}
else
{
#ifdef _OPENMP
nprocs = omp_get_max_threads();
#ifndef USE_SSE2
#pragma omp parallel private(i,j,jj,is,ie,w1,my_rank,p1,p2,tq,bhi,blo,chi,clo,sh,sl,th,tl,eh,el)
#else
#pragma omp parallel private(i,j,jj,is,ie,w1,my_rank,bh,ch,sh,wh,th,bl,cl,sl,wl,tl,p1,p2,t0,t1,t2,eh)
#endif
{
my_rank = omp_get_thread_num();
LIS_GET_ISIE(my_rank,nprocs,n,is,ie);
for(i=is;i<ie;i++)
{
#ifndef USE_SSE2
LIS_QUAD_MULD(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#else
LIS_QUAD_MULD_SSE2(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#endif
/* t = x[i] * A->WD->value[i]; */
for(j=A->U->ptr[i];j<A->U->ptr[i+1];j++)
{
jj = A->U->index[j];
if( jj<is || jj>=ie ) continue;
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->U->value[j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->U->value[j]);
#endif
/* x[A->U->index[j]] -= A->U->value[j] * t; */
}
}
for(i=ie-1;i>=is;i--)
{
#ifndef USE_SSE2
LIS_QUAD_MULD(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#else
LIS_QUAD_MULD_SSE2(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#endif
x[i] = w1.hi;
xl[i] = w1.lo;
/* t = x[i] * A->WD->value[i]; */
/* x[i] = t; */
for(j=A->L->ptr[i];j<A->L->ptr[i+1];j++)
{
jj = A->L->index[j];
if( jj<is || jj>=ie ) continue;
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->L->value[j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->L->value[j]);
#endif
/* x[A->L->index[j]] -= A->L->value[j] * t; */
}
}
}
#else
for(i=0;i<n;i++)
{
#ifndef USE_SSE2
LIS_QUAD_MULD(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#else
LIS_QUAD_MULD_SSE2(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#endif
/* t = x[i] * A->WD->value[i]; */
for(j=A->U->ptr[i];j<A->U->ptr[i+1];j++)
{
jj = A->U->index[j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->U->value[j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->U->value[j]);
#endif
/* x[A->U->index[j]] -= A->U->value[j] * t; */
}
}
for(i=n-1;i>=0;i--)
{
#ifndef USE_SSE2
LIS_QUAD_MULD(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#else
LIS_QUAD_MULD_SSE2(w1.hi,w1.lo,x[i],xl[i],A->WD->value[i]);
#endif
x[i] = w1.hi;
xl[i] = w1.lo;
/* t = x[i] * A->WD->value[i]; */
/* x[i] = t; */
for(j=A->L->ptr[i];j<A->L->ptr[i+1];j++)
{
jj = A->L->index[j];
#ifndef USE_SSE2
LIS_QUAD_FMAD(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->L->value[j]);
#else
LIS_QUAD_FMAD_SSE2(x[jj],xl[jj],x[jj],xl[jj],w1.hi,w1.lo,-A->L->value[j]);
#endif
/* x[A->L->index[j]] -= A->L->value[j] * t; */
}
}
#endif
}
#endif
break;
}
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
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;
}
LisVector::LisVector(LisVector const& src)
{
lis_vector_duplicate(src._vec, &_vec);
lis_vector_copy(src._vec, _vec);
}
LisVector& LisVector::operator= (const LisVector &src)
{
lis_vector_copy(src._vec, _vec);
return *this;
}
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 lis_bicgstab_switch(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_PRECON M;
LIS_VECTOR b,x;
LIS_VECTOR r,rtld, t,p,v, s, phat, shat;
LIS_QUAD_PTR alpha, beta, omega, rho, rho_old, tmpdot1, tmpdot2;
LIS_REAL bnrm2, nrm2, tol, tol2;
LIS_INT iter,maxiter,n,output,conv;
LIS_INT iter2,maxiter2;
double times,ptimes;
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];
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];
ptimes = 0.0;
rtld = solver->work[0];
r = solver->work[1];
s = solver->work[1];
t = solver->work[2];
p = solver->work[3];
v = solver->work[4];
phat = solver->work[5];
shat = 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(omega,6,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot2,7,1);
rho_old.hi[0] = 1.0;
rho_old.lo[0] = 0.0;
alpha.hi[0] = 1.0;
alpha.lo[0] = 0.0;
omega.hi[0] = 1.0;
omega.lo[0] = 0.0;
lis_vector_set_allex_nm(0.0, p);
lis_vector_set_allex_nm(0.0, phat);
lis_vector_set_allex_nm(0.0, s);
lis_vector_set_allex_nm(0.0, shat);
/* 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);
s->precision = LIS_PRECISION_DEFAULT;
shat->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;
}
if( iter==1 )
{
lis_vector_copy(r,p);
}
else
{
/* beta = (rho / rho_old) * (alpha / omega) */
beta.hi[0] = (rho.hi[0] / rho_old.hi[0]) * (alpha.hi[0] / omega.hi[0]);
/* p = r + beta*(p - omega*v) */
lis_vector_axpy(-omega.hi[0],v,p);
lis_vector_xpay(r,beta.hi[0],p);
}
/* phat = M^-1 * p */
times = lis_wtime();
lis_psolve(solver, p, phat);
ptimes += lis_wtime()-times;
/* v = A * phat */
LIS_MATVEC(A,phat,v);
/* tmpdot1 = <rtld,v> */
lis_vector_dot(rtld,v,&tmpdot1.hi[0]);
/* test breakdown */
/* */
/* alpha = rho / tmpdot1 */
alpha.hi[0] = rho.hi[0] / tmpdot1.hi[0];
/* s = r - alpha*v */
lis_vector_axpy(-alpha.hi[0],v,r);
/* Early check for tolerance */
lis_solver_get_residual[conv](s,solver,&nrm2);
if( nrm2 <= tol2 )
{
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);
}
lis_vector_axpy(alpha.hi[0],phat,x);
solver->iter = iter;
solver->iter2 = iter;
solver->ptimes = ptimes;
break;
}
/* shat = M^-1 * s */
times = lis_wtime();
lis_psolve(solver, s, shat);
ptimes += lis_wtime()-times;
/* t = A * shat */
LIS_MATVEC(A,shat,t);
/* tmpdot1 = <t,s> */
/* tmpdot2 = <t,t> */
/* omega = tmpdot1 / tmpdot2 */
lis_vector_dot(t,s,&tmpdot1.hi[0]);
lis_vector_dot(t,t,&tmpdot2.hi[0]);
omega.hi[0] = tmpdot1.hi[0] / tmpdot2.hi[0];
/* x = x + alpha*phat + omega*shat */
lis_vector_axpy(alpha.hi[0],phat,x);
lis_vector_axpy(omega.hi[0],shat,x);
/* r = s - omega*t */
lis_vector_axpy(-omega.hi[0],t,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
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( nrm2 <= tol2 )
{
solver->iter = iter;
solver->iter2 = iter;
solver->ptimes = ptimes;
break;
}
if( omega.hi[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
rho_old.hi[0] = rho.hi[0];
}
s->precision = LIS_PRECISION_QUAD;
shat->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;
alpha.hi[0] = 1.0;
omega.hi[0] = 1.0;
lis_vector_set_allex_nm(0.0, p);
lis_vector_set_allex_nm(0.0, phat);
lis_vector_set_allex_nm(0.0, s);
lis_vector_set_allex_nm(0.0, shat);
/* Initial Residual */
lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2);
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
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;
}
if( iter2==1 )
{
lis_vector_copyex_mm(r,p);
}
else
{
/* beta = (rho / rho_old) * (alpha / omega) */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_quad_div((LIS_QUAD *)tmpdot1.hi,(LIS_QUAD *)alpha.hi,(LIS_QUAD *)omega.hi);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,(LIS_QUAD *)tmpdot1.hi);
/* p = r + beta*(p - omega*v) */
lis_quad_minus((LIS_QUAD *)omega.hi);
lis_vector_axpyex_mmm(omega,v,p);
lis_vector_xpayex_mmm(r,beta,p);
}
/* phat = M^-1 * p */
times = lis_wtime();
lis_psolve(solver, p, phat);
ptimes += lis_wtime()-times;
/* v = A * phat */
LIS_MATVEC(A,phat,v);
/* tmpdot1 = <rtld,v> */
lis_vector_dotex_mmm(rtld,v,&tmpdot1);
/* test breakdown */
/* */
/* alpha = rho / tmpdot1 */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)tmpdot1.hi);
/* s = r - alpha*v */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,v,r);
/* Early check for tolerance */
lis_solver_get_residual[conv](s,solver,&nrm2);
if( tol > nrm2 )
{
if( output )
{
if( output & LIS_PRINT_MEM ) solver->residual[iter2] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) printf("iter: %5d residual = %e\n", iter2, nrm2);
}
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,phat,x);
solver->retcode = LIS_SUCCESS;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
solver->ptimes = ptimes;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* shat = M^-1 * s */
times = lis_wtime();
lis_psolve(solver, s, shat);
ptimes += lis_wtime()-times;
/* t = A * shat */
LIS_MATVEC(A,shat,t);
/* tmpdot1 = <t,s> */
/* tmpdot2 = <t,t> */
/* omega = tmpdot1 / tmpdot2 */
lis_vector_dotex_mmm(t,s,&tmpdot1);
lis_vector_dotex_mmm(t,t,&tmpdot2);
lis_quad_div((LIS_QUAD *)omega.hi,(LIS_QUAD *)tmpdot1.hi,(LIS_QUAD *)tmpdot2.hi);
/* x = x + alpha*phat + omega*shat */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,phat,x);
lis_vector_axpyex_mmm(omega,shat,x);
/* r = s - omega*t */
lis_quad_minus((LIS_QUAD *)omega.hi);
lis_vector_axpyex_mmm(omega,t,r);
lis_quad_minus((LIS_QUAD *)omega.hi);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->residual[iter2] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) printf("iter: %5d residual = %e\n", iter2, nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
solver->ptimes = ptimes;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
if( omega.hi[0]==0.0 && omega.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
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;
}
