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;
}
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_orthomin_quad(LIS_SOLVER solver)
{
LIS_Comm comm;
LIS_MATRIX A;
LIS_PRECON M;
LIS_VECTOR x;
LIS_VECTOR r, rtld, *p, *ap, *aptld;
LIS_QUAD *dotsave;
LIS_QUAD_PTR alpha, beta, tmp, one;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_INT m,l,lmax,ip,ip0;
LIS_DEBUG_FUNC_IN;
comm = LIS_COMM_WORLD;
A = solver->A;
M = solver->precon;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
m = solver->options[LIS_OPTIONS_RESTART];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(tmp,3,1);
LIS_QUAD_SCALAR_MALLOC(one,4,1);
r = solver->work[0];
rtld = solver->work[1];
p = &solver->work[2];
ap = &solver->work[ (m+1)+2];
aptld = &solver->work[2*(m+1)+2];
one.hi[0] = 1.0;
one.lo[0] = 0.0;
dotsave = (LIS_QUAD *)lis_malloc( sizeof(LIS_QUAD) * (m+1),"lis_orthomin_quad::dotsave" );
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,M,r,rtld,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
iter=1;
while( iter<=maxiter )
{
ip = (iter-1) % (m+1);
/* p[ip] = rtld */
lis_vector_copyex_mm(rtld,p[ip]);
/* ap[ip] = A*p[ip] */
/* aptld[ip] = M^-1 ap[ip] */
lis_matvec(A,p[ip],ap[ip]);
time = lis_wtime();
lis_psolve(solver, ap[ip], aptld[ip]);
ptime += lis_wtime()-time;
lmax = _min(m,iter-1);
for(l=1;l<=lmax;l++)
{
ip0 = (ip+m+1-l) % (m+1);
/* beta = -<Ar[ip],Ap[ip0]> / <Ap[ip0],Ap[ip0]> */
lis_vector_dotex_mmm(aptld[ip],aptld[ip0],&beta);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,&dotsave[l-1]);
lis_quad_minus((LIS_QUAD *)beta.hi);
lis_vector_axpyex_mmm(beta,p[ip0] ,p[ip]);
lis_vector_axpyex_mmm(beta,ap[ip0] ,ap[ip]);
lis_vector_axpyex_mmm(beta,aptld[ip0],aptld[ip]);
}
for(l=m-1;l>0;l--)
{
dotsave[l] = dotsave[l-1];
}
lis_vector_dotex_mmm(aptld[ip],aptld[ip],&tmp);
dotsave[0].hi = tmp.hi[0];
dotsave[0].lo = tmp.lo[0];
/* test breakdown */
if( tmp.hi[0]==0.0 && tmp.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
lis_free(dotsave);
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
lis_quad_div(&dotsave[0],(LIS_QUAD *)one.hi,&dotsave[0]);
/* alpha = <rtld,Aptld[ip]> */
lis_vector_dotex_mmm(rtld,aptld[ip],&alpha);
lis_quad_mul((LIS_QUAD *)alpha.hi,(LIS_QUAD *)alpha.hi,&dotsave[0]);
lis_vector_axpyex_mmm( alpha,p[ip],x);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,ap[ip],r);
lis_vector_axpyex_mmm(alpha,aptld[ip],rtld);
lis_quad_minus((LIS_QUAD *)alpha.hi);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT ) lis_print_rhistory(comm,iter,nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
lis_free(dotsave);
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
iter++;
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
lis_free(dotsave);
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_bicr_quad(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_QUAD_PTR 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];
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);
/* 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_copyex_mm(z,p);
lis_vector_copyex_mm(ztld,ptld);
LIS_MATVEC(A,z,ap);
lis_vector_dotex_mmm(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_dotex_mmm(map,aptld,&tmpdot1);
/* test breakdown */
if( tmpdot1.hi[0]==0.0 && tmpdot1.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* alpha = rho_old / tmpdot1 */
/* x = x + alpha*p */
/* r = r - alpha*ap */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho_old.hi,(LIS_QUAD *)tmpdot1.hi);
lis_vector_axpyex_mmm(alpha,p,x);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(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_axpyex_mmm(alpha,aptld,rtld);
lis_vector_axpyex_mmm(alpha,map,z);
times = lis_wtime();
lis_psolvet(solver, rtld, ztld);
ptimes += lis_wtime()-times;
LIS_MATVEC(A,z,az);
lis_vector_dotex_mmm(az,ztld,&rho);
/* test breakdown */
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = rho / rho_old */
/* p = z + beta*p */
/* ptld = ztld + beta*ptld */
/* ap = az + beta*ap */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_vector_xpayex_mmm(z,beta,p);
lis_vector_xpayex_mmm(ztld,beta,ptld);
lis_vector_xpayex_mmm(az,beta,ap);
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_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_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_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_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_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;
}
LIS_INT lis_bicrstab_quad(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_QUAD_PTR 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];
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);
/* 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_copyex_mm(z,p);
lis_vector_dotex_mmm(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_dotex_mmm(rtld,map,&tmpdot1);
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho_old.hi,(LIS_QUAD *)tmpdot1.hi);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyzex_mmmm(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_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(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_axpyzex_mmmm(alpha,map,z,ms);
LIS_MATVEC(A,ms,ams);
lis_vector_dotex_mmm(ams,s,&tmpdot1);
lis_vector_dotex_mmm(ams,ams,&tmpdot2);
lis_quad_div((LIS_QUAD *)omega.hi,(LIS_QUAD *)tmpdot1.hi,(LIS_QUAD *)tmpdot2.hi);
/* x = x + alpha*p + omega*ms */
/* r = s - omega*ams */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,p,x);
lis_vector_axpyex_mmm(omega,ms,x);
lis_quad_minus((LIS_QUAD *)omega.hi);
lis_vector_axpyzex_mmmm(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_dotex_mmm(rtld,z,&rho);
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / omega) */
/* p = z + beta*(p - omega*map) */
lis_quad_minus((LIS_QUAD *)omega.hi);
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);
lis_quad_minus((LIS_QUAD *)omega.hi);
lis_vector_axpyex_mmm(omega,map,p);
lis_vector_xpayex_mmm(z,beta,p);
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_bicgsafe_switch(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r, rtld, rhat, p, ptld, phat;
LIS_VECTOR t, ttld, that, t0, t0hat;
LIS_VECTOR y, w, u, z;
LIS_QUAD_PTR alpha, beta, rho, rho_old;
LIS_QUAD_PTR qsi, eta, one;
LIS_QUAD_PTR tmp, tmpdot[5];
LIS_REAL bnrm2, nrm2, tol, tol2;
LIS_INT iter,maxiter,output,conv;
LIS_INT iter2,maxiter2;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
maxiter2 = solver->options[LIS_OPTIONS_SWITCH_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
tol = solver->params[LIS_PARAMS_RESID-LIS_OPTIONS_LEN];
tol2 = solver->params[LIS_PARAMS_SWITCH_RESID-LIS_OPTIONS_LEN];
ptime = 0.0;
rtld = solver->work[0];
r = solver->work[1];
rhat = solver->work[2];
p = solver->work[3];
ptld = solver->work[4];
phat = solver->work[5];
t = solver->work[6];
ttld = solver->work[7];
that = solver->work[8];
t0 = solver->work[9];
t0hat = solver->work[10];
y = solver->work[11];
w = solver->work[12];
u = solver->work[13];
z = solver->work[14];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(qsi,4,1);
LIS_QUAD_SCALAR_MALLOC(eta,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[0],7,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[1],8,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[2],9,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[3],10,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[4],11,1);
LIS_QUAD_SCALAR_MALLOC(one,13,1);
rho_old.hi[0] = 1.0;
rho_old.lo[0] = 0.0;
alpha.hi[0] = 1.0;
alpha.lo[0] = 0.0;
qsi.hi[0] = 1.0;
qsi.lo[0] = 0.0;
one.hi[0] = -1.0;
one.lo[0] = 0.0;
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol2 = solver->tol_switch;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0, ttld);
lis_vector_set_allex_nm(0.0, ptld);
lis_vector_set_allex_nm(0.0, p);
lis_vector_set_allex_nm(0.0, u);
lis_vector_set_allex_nm(0.0, t);
lis_vector_set_allex_nm(0.0, t0);
for( iter=1; iter<=maxiter2; iter++ )
{
/* rho = <rtld,r> */
lis_vector_dot(rtld,r,&rho.hi[0]);
/* test breakdown */
if( rho.hi[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
beta.hi[0] = (rho.hi[0] / rho_old.hi[0]) * (alpha.hi[0] / qsi.hi[0]);
/* w = ttld + beta*ptld */
lis_vector_axpyz(beta.hi[0],ptld,ttld,w);
/* rhat = M^-1 * r */
time = lis_wtime();
lis_psolve(solver, r, rhat);
ptime += lis_wtime()-time;
/* p = rhat + beta*(p - u) */
lis_vector_axpy(-1,u,p);
lis_vector_xpay(rhat,beta.hi[0],p);
/* ptld = A * p */
lis_matvec(A,p,ptld);
/* tmpdot[0] = <rtld,ptld> */
lis_vector_dot(rtld,ptld,&tmpdot[0].hi[0]);
/* test breakdown */
/* */
/* alpha = rho / tmpdot[0] */
alpha.hi[0] = rho.hi[0] / tmpdot[0].hi[0];
/* y = t - r + alpha*(-w + ptld) */
lis_vector_axpyz(-1,w,ptld,y);
lis_vector_xpay(t,alpha.hi[0],y);
lis_vector_axpy(-1,r,y);
/* t = r - alpha*ptld */
lis_vector_axpyz(-alpha.hi[0],ptld,r,t);
/* that = M^-1 * t */
/* phat = M^-1 * ptld */
/* t0hat = M^-1 * t0 */
time = lis_wtime();
lis_psolve(solver, t, that);
lis_psolve(solver, ptld, phat);
lis_psolve(solver, t0, t0hat);
ptime += lis_wtime()-time;
/* ttld = A * that */
lis_matvec(A,that,ttld);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <ttld,t> */
/* tmpdot[2] = <y,t> */
/* tmpdot[3] = <ttld,y> */
/* tmpdot[4] = <ttld,ttld> */
lis_vector_dot(y,y,&tmpdot[0].hi[0]);
lis_vector_dot(ttld,t,&tmpdot[1].hi[0]);
lis_vector_dot(y,t,&tmpdot[2].hi[0]);
lis_vector_dot(ttld,y,&tmpdot[3].hi[0]);
lis_vector_dot(ttld,ttld,&tmpdot[4].hi[0]);
if(iter==1)
{
qsi.hi[0] = tmpdot[1].hi[0] / tmpdot[4].hi[0];
eta.hi[0] = 0.0;
}
else
{
tmp.hi[0] = tmpdot[4].hi[0]*tmpdot[0].hi[0] - tmpdot[3].hi[0]*tmpdot[3].hi[0];
qsi.hi[0] = (tmpdot[0].hi[0]*tmpdot[1].hi[0] - tmpdot[2].hi[0]*tmpdot[3].hi[0]) / tmp.hi[0];
eta.hi[0] = (tmpdot[4].hi[0]*tmpdot[2].hi[0] - tmpdot[3].hi[0]*tmpdot[1].hi[0]) / tmp.hi[0];
}
/* u = qsi*phat + eta*(t0hat - rhat + beta*u) */
lis_vector_xpay(t0hat,beta.hi[0],u);
lis_vector_axpy(-1,rhat,u);
lis_vector_scale(eta.hi[0],u);
lis_vector_axpy(qsi.hi[0],phat,u);
/* z = qsi*rhat + eta*z - alpha*u */
lis_vector_scale(eta.hi[0],z);
lis_vector_axpy(qsi.hi[0],rhat,z);
lis_vector_axpy(-alpha.hi[0],u,z);
/* x = x + alpha*p + z */
lis_vector_axpy(alpha.hi[0],p,x);
lis_vector_axpy(1,z,x);
/* r = t - eta*y - qsi*ttld */
lis_vector_axpyz(-eta.hi[0],y,t,r);
lis_vector_axpy(-qsi.hi[0],ttld,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol2 >= nrm2 )
{
solver->iter = iter;
solver->iter2 = iter;
solver->ptime = ptime;
break;
}
lis_vector_copy(t,t0);
rho_old.hi[0] = rho.hi[0];
}
r->precision = LIS_PRECISION_QUAD;
p->precision = LIS_PRECISION_QUAD;
t->precision = LIS_PRECISION_QUAD;
t0->precision = LIS_PRECISION_QUAD;
ptld->precision = LIS_PRECISION_QUAD;
that->precision = LIS_PRECISION_QUAD;
solver->options[LIS_OPTIONS_INITGUESS_ZEROS] = LIS_FALSE;
lis_vector_copyex_mn(x,solver->xx);
rho_old.hi[0] = 1.0;
alpha.hi[0] = 1.0;
qsi.hi[0] = 1.0;
one.hi[0] = -1.0;
/* Initial Residual */
lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2);
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0, ttld);
lis_vector_set_allex_nm(0.0, ptld);
lis_vector_set_allex_nm(0.0, p);
lis_vector_set_allex_nm(0.0, u);
lis_vector_set_allex_nm(0.0, t);
lis_vector_set_allex_nm(0.0, t0);
for( iter2=iter+1; iter2<=maxiter; iter2++ )
{
/* rho = <rtld,r> */
lis_vector_dotex_mmm(rtld,r,&rho);
/* test breakdown */
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)alpha.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,(LIS_QUAD *)tmp.hi);
/* w = ttld + beta*ptld */
lis_vector_axpyzex_mmmm(beta,ptld,ttld,w);
/* rhat = M^-1 * r */
time = lis_wtime();
lis_psolve(solver, r, rhat);
ptime += lis_wtime()-time;
/* p = rhat + beta*(p - u) */
lis_vector_axpyex_mmm(one,u,p);
lis_vector_xpayex_mmm(rhat,beta,p);
/* ptld = A * p */
lis_matvec(A,p,ptld);
/* tmpdot[0] = <rtld,ptld> */
lis_vector_dotex_mmm(rtld,ptld,&tmpdot[0]);
/* test breakdown */
/* */
/* alpha = rho / tmpdot[0] */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)tmpdot[0].hi);
/* y = t - r + alpha*(-w + ptld) */
lis_vector_axpyzex_mmmm(one,w,ptld,y);
lis_vector_xpayex_mmm(t,alpha,y);
lis_vector_axpyex_mmm(one,r,y);
/* t = r - alpha*ptld */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyzex_mmmm(alpha,ptld,r,t);
/* that = M^-1 * t */
/* phat = M^-1 * ptld */
/* t0hat = M^-1 * t0 */
time = lis_wtime();
lis_psolve(solver, t, that);
lis_psolve(solver, ptld, phat);
lis_psolve(solver, t0, t0hat);
ptime += lis_wtime()-time;
/* ttld = A * that */
lis_matvec(A,that,ttld);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <ttld,t> */
/* tmpdot[2] = <y,t> */
/* tmpdot[3] = <ttld,y> */
/* tmpdot[4] = <ttld,ttld> */
lis_vector_dotex_mmm(y,y,&tmpdot[0]);
lis_vector_dotex_mmm(ttld,t,&tmpdot[1]);
lis_vector_dotex_mmm(y,t,&tmpdot[2]);
lis_vector_dotex_mmm(ttld,y,&tmpdot[3]);
lis_vector_dotex_mmm(ttld,ttld,&tmpdot[4]);
if(iter==1)
{
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[1].hi,(LIS_QUAD *)tmpdot[4].hi);
eta.hi[0] = 0.0;
eta.lo[0] = 0.0;
}
else
{
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_sqr((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[2].hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)eta.hi);
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[2].hi);
lis_quad_mul((LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[3].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_sub((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_div((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmp.hi);
}
/* u = qsi*phat + eta*(t0hat - rhat + beta*u) */
lis_vector_xpayex_mmm(t0hat,beta,u);
lis_vector_axpyex_mmm(one,rhat,u);
lis_vector_scaleex_mm(eta,u);
lis_vector_axpyex_mmm(qsi,phat,u);
/* z = qsi*rhat + eta*z - alpha*u */
lis_vector_scaleex_mm(eta,z);
lis_vector_axpyex_mmm(qsi,rhat,z);
lis_vector_axpyex_mmm(alpha,u,z);
/* x = x + alpha*p + z */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(alpha,p,x);
lis_vector_axpyex_mmm(one,z,x);
lis_quad_minus((LIS_QUAD *)one.hi);
/* r = t - eta*y - qsi*ttld */
lis_quad_minus((LIS_QUAD *)eta.hi);
lis_quad_minus((LIS_QUAD *)qsi.hi);
lis_vector_axpyzex_mmmm(eta,y,t,r);
lis_vector_axpyex_mmm(qsi,ttld,r);
lis_quad_minus((LIS_QUAD *)eta.hi);
lis_quad_minus((LIS_QUAD *)qsi.hi);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter2] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter2;
solver->iter2 = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
lis_vector_copyex_mm(t,t0);
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->iter2 = iter2;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_bicgsafe_quad(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r, rtld, rhat, p, ptld;
LIS_VECTOR t, ttld;
LIS_VECTOR y, v, u, utld, z;
LIS_QUAD_PTR alpha, beta, rho, rho_old;
LIS_QUAD_PTR qsi, eta;
LIS_QUAD_PTR tmp, tmpdot[5],one;
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
rtld = solver->work[0];
r = solver->work[1];
rhat = solver->work[2];
p = solver->work[3];
ptld = solver->work[4];
t = solver->work[5];
ttld = solver->work[6];
y = solver->work[7];
v = solver->work[8];
u = solver->work[9];
z = solver->work[10];
utld = solver->work[11];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(qsi,4,1);
LIS_QUAD_SCALAR_MALLOC(eta,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[0],7,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[1],8,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[2],9,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[3],10,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[4],11,1);
LIS_QUAD_SCALAR_MALLOC(one,13,1);
rho_old.hi[0] = 1.0;
rho_old.lo[0] = 0.0;
alpha.hi[0] = 1.0;
alpha.lo[0] = 0.0;
qsi.hi[0] = 1.0;
qsi.lo[0] = 0.0;
one.hi[0] = -1.0;
one.lo[0] = 0.0;
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_vector_set_allex_nm(0.0,p);
lis_vector_set_allex_nm(0.0,u);
lis_vector_set_allex_nm(0.0,ptld);
lis_vector_set_allex_nm(0.0,utld);
for( iter=1; iter<=maxiter; iter++ )
{
/* rho = <rtld,r> */
lis_vector_dotex_mmm(rtld,r,&rho);
/* test breakdown */
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)alpha.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,(LIS_QUAD *)tmp.hi);
/* rhat = M^-1 * r */
/* v = A * rhat */
time = lis_wtime();
lis_psolve(solver, r, rhat);
ptime += lis_wtime()-time;
lis_matvec(A,rhat,v);
/* p = rhat + beta*(p - u) */
lis_vector_axpyex_mmm(one,u,p);
lis_vector_xpayex_mmm(rhat,beta,p);
/* ptld = v + beta*(ptld - utld) */
lis_vector_axpyex_mmm(one,utld,ptld);
lis_vector_xpayex_mmm(v,beta,ptld);
/* tmpdot[0] = <rtld,ptld> */
lis_vector_dotex_mmm(rtld,ptld,&tmpdot[0]);
/* test breakdown */
/* */
/* alpha = rho / tmpdot[0] */
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)tmpdot[0].hi);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <v,r> */
/* tmpdot[2] = <y,r> */
/* tmpdot[3] = <v,y> */
/* tmpdot[4] = <v,v> */
lis_vector_dotex_mmm(y,y,&tmpdot[0]);
lis_vector_dotex_mmm(v,r,&tmpdot[1]);
lis_vector_dotex_mmm(y,r,&tmpdot[2]);
lis_vector_dotex_mmm(v,y,&tmpdot[3]);
lis_vector_dotex_mmm(v,v,&tmpdot[4]);
if(iter==1)
{
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[1].hi,(LIS_QUAD *)tmpdot[4].hi);
eta.hi[0] = 0.0;
eta.lo[0] = 0.0;
}
else
{
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_sqr((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[2].hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)eta.hi);
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[2].hi);
lis_quad_mul((LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[3].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_sub((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_div((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmp.hi);
}
/* t = qsi*ptld + eta*y */
lis_vector_copyex_mm(y,t);
lis_vector_scaleex_mm(eta,t);
lis_vector_axpyex_mmm(qsi,ptld,t);
/* ttld = M^-1 * t */
time = lis_wtime();
lis_psolve(solver, t, ttld);
ptime += lis_wtime()-time;
/* u = ttld + eta*beta*u */
/* utld = A * u */
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)beta.hi);
lis_vector_xpayex_mmm(ttld,tmp,u);
lis_matvec(A,u,utld);
/* z = qsi*rhat + eta*z - alpha*u */
lis_vector_scaleex_mm(eta,z);
lis_vector_axpyex_mmm(qsi,rhat,z);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,u,z);
/* y = qsi*v + eta*y - alpha*utld */
lis_vector_scaleex_mm(eta,y);
lis_vector_axpyex_mmm(qsi,v,y);
lis_vector_axpyex_mmm(alpha,utld,y);
lis_quad_minus((LIS_QUAD *)alpha.hi);
/* x = x + alpha*p + z */
lis_vector_axpyex_mmm(alpha,p,x);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(one,z,x);
lis_quad_minus((LIS_QUAD *)one.hi);
/* r = r - alpha*ptld - y */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,ptld,r);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(one,y,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol > nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
LIS_INT lis_bicrsafe_quad(LIS_SOLVER solver)
{
LIS_MATRIX A;
LIS_VECTOR x;
LIS_VECTOR r, rtld, artld, mr, amr, p, ap, map;
LIS_VECTOR y, my, u, au, z;
LIS_QUAD_PTR alpha, beta, rho, rho_old;
LIS_QUAD_PTR qsi, eta, one;
LIS_QUAD_PTR tmp, tmpdot[5];
LIS_REAL bnrm2, nrm2, tol;
LIS_INT iter,maxiter,output,conv;
double time,ptime;
LIS_DEBUG_FUNC_IN;
A = solver->A;
x = solver->x;
maxiter = solver->options[LIS_OPTIONS_MAXITER];
output = solver->options[LIS_OPTIONS_OUTPUT];
conv = solver->options[LIS_OPTIONS_CONV_COND];
ptime = 0.0;
rtld = solver->work[0];
r = solver->work[1];
mr = solver->work[2];
amr = solver->work[3];
p = solver->work[4];
ap = solver->work[5];
map = solver->work[6];
my = solver->work[7];
y = solver->work[8];
u = solver->work[9];
z = solver->work[10];
au = solver->work[11];
artld = solver->work[12];
LIS_QUAD_SCALAR_MALLOC(alpha,0,1);
LIS_QUAD_SCALAR_MALLOC(beta,1,1);
LIS_QUAD_SCALAR_MALLOC(rho,2,1);
LIS_QUAD_SCALAR_MALLOC(rho_old,3,1);
LIS_QUAD_SCALAR_MALLOC(qsi,4,1);
LIS_QUAD_SCALAR_MALLOC(eta,5,1);
LIS_QUAD_SCALAR_MALLOC(tmp,6,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[0],7,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[1],8,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[2],9,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[3],10,1);
LIS_QUAD_SCALAR_MALLOC(tmpdot[4],11,1);
LIS_QUAD_SCALAR_MALLOC(one,13,1);
/* Initial Residual */
if( lis_solver_get_initial_residual(solver,NULL,NULL,r,&bnrm2) )
{
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
tol = solver->tol;
lis_solver_set_shadowresidual(solver,r,rtld);
lis_matvect(A,rtld,artld);
time = lis_wtime();
lis_psolve(solver, r, mr);
ptime += lis_wtime()-time;
lis_matvec(A,mr,amr);
lis_vector_dotex_mmm(rtld,amr,&rho_old);
lis_vector_copyex_mm(amr,ap);
lis_vector_copyex_mm(mr,p);
one.hi[0] = -1.0;
one.lo[0] = 0.0;
for( iter=1; iter<=maxiter; iter++ )
{
/* map = M^-1 * ap */
time = lis_wtime();
lis_psolve(solver, ap, map);
ptime += lis_wtime()-time;
/* tmpdot[0] = <artld,map> */
/* alpha = rho_old / tmpdot[0] */
lis_vector_dotex_mmm(artld,map,&tmpdot[0]);
lis_quad_div((LIS_QUAD *)alpha.hi,(LIS_QUAD *)rho_old.hi,(LIS_QUAD *)tmpdot[0].hi);
/* tmpdot[0] = <y,y> */
/* tmpdot[1] = <amr,r> */
/* tmpdot[2] = <y,r> */
/* tmpdot[3] = <amr,y> */
/* tmpdot[4] = <amr,amr> */
lis_vector_dotex_mmm(y,y,&tmpdot[0]);
lis_vector_dotex_mmm(amr,r,&tmpdot[1]);
lis_vector_dotex_mmm(y,r,&tmpdot[2]);
lis_vector_dotex_mmm(amr,y,&tmpdot[3]);
lis_vector_dotex_mmm(amr,amr,&tmpdot[4]);
if(iter==1)
{
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[1].hi,(LIS_QUAD *)tmpdot[4].hi);
eta.hi[0] = 0.0;
eta.lo[0] = 0.0;
}
else
{
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_sqr((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)tmp.hi,(LIS_QUAD *)tmp.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[2].hi,(LIS_QUAD *)tmpdot[3].hi);
lis_quad_sub((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)eta.hi);
lis_quad_div((LIS_QUAD *)qsi.hi,(LIS_QUAD *)qsi.hi,(LIS_QUAD *)tmp.hi);
lis_quad_mul((LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[4].hi,(LIS_QUAD *)tmpdot[2].hi);
lis_quad_mul((LIS_QUAD *)tmpdot[0].hi,(LIS_QUAD *)tmpdot[3].hi,(LIS_QUAD *)tmpdot[1].hi);
lis_quad_sub((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmpdot[0].hi);
lis_quad_div((LIS_QUAD *)eta.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)tmp.hi);
}
/* u = qsi*map + eta*my + eta*beta*u */
/* au = A * u */
lis_quad_mul((LIS_QUAD *)tmp.hi,(LIS_QUAD *)eta.hi,(LIS_QUAD *)beta.hi);
lis_vector_scaleex_mm(tmp,u);
lis_vector_axpyex_mmm(qsi,map,u);
lis_vector_axpyex_mmm(eta,my,u);
lis_matvec(A,u,au);
/* z = qsi*mr + eta*z - alpha*u */
lis_vector_scaleex_mm(eta,z);
lis_vector_axpyex_mmm(qsi,mr,z);
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,u,z);
/* y = qsi*amr + eta*y - alpha*au */
/* my = M^-1 * y */
lis_vector_scaleex_mm(eta,y);
lis_vector_axpyex_mmm(qsi,amr,y);
lis_vector_axpyex_mmm(alpha,au,y);
time = lis_wtime();
lis_psolve(solver, y, my);
ptime += lis_wtime()-time;
/* x = x + alpha*p + z */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_vector_axpyex_mmm(alpha,p,x);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(one,z,x);
/* r = r - alpha*ap - y */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_quad_minus((LIS_QUAD *)one.hi);
lis_vector_axpyex_mmm(alpha,ap,r);
lis_vector_axpyex_mmm(one,y,r);
/* convergence check */
lis_solver_get_residual[conv](r,solver,&nrm2);
if( output )
{
if( output & LIS_PRINT_MEM ) solver->rhistory[iter] = nrm2;
if( output & LIS_PRINT_OUT && A->my_rank==0 ) lis_print_rhistory(iter,nrm2);
}
if( tol >= nrm2 )
{
solver->retcode = LIS_SUCCESS;
solver->iter = iter;
solver->resid = nrm2;
solver->ptime = ptime;
LIS_DEBUG_FUNC_OUT;
return LIS_SUCCESS;
}
/* mr = mr - alpha*map - my */
/* amr = A * mr */
/* rho = <rtld,amr> */
lis_vector_axpyex_mmm(alpha,map,mr);
lis_vector_axpyex_mmm(one,my,mr);
lis_matvec(A,mr,amr);
lis_vector_dotex_mmm(rtld,amr,&rho);
if( rho.hi[0]==0.0 && rho.lo[0]==0.0 )
{
solver->retcode = LIS_BREAKDOWN;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_BREAKDOWN;
}
/* beta = (rho / rho_old) * (alpha / qsi) */
lis_quad_minus((LIS_QUAD *)alpha.hi);
lis_quad_div((LIS_QUAD *)beta.hi,(LIS_QUAD *)rho.hi,(LIS_QUAD *)rho_old.hi);
lis_quad_div((LIS_QUAD *)tmp.hi,(LIS_QUAD *)alpha.hi,(LIS_QUAD *)qsi.hi);
lis_quad_mul((LIS_QUAD *)beta.hi,(LIS_QUAD *)beta.hi,(LIS_QUAD *)tmp.hi);
/* p = mr + beta*(p - u) */
/* ap = amr + beta*(ap - au) */
lis_vector_axpyex_mmm(one,u,p);
lis_vector_xpayex_mmm(mr,beta,p);
lis_vector_axpyex_mmm(one,au,ap);
lis_vector_xpayex_mmm(amr,beta,ap);
rho_old.hi[0] = rho.hi[0];
rho_old.lo[0] = rho.lo[0];
}
solver->retcode = LIS_MAXITER;
solver->iter = iter;
solver->resid = nrm2;
LIS_DEBUG_FUNC_OUT;
return LIS_MAXITER;
}
