void
Nyx::strang_second_step (Real time, Real dt, MultiFab& S_new, MultiFab& D_new)
{
BL_PROFILE("Nyx::strang_second_step()");
Real half_dt = 0.5*dt;
int min_iter = 100000;
int max_iter = 0;
int min_iter_grid;
int max_iter_grid;
// Set a at the half of the time step in the second strang
const Real a = get_comoving_a(time-half_dt);
MultiFab reset_e_src(S_new.boxArray(), S_new.DistributionMap(), 1, NUM_GROW);
reset_e_src.setVal(0.0);
reset_internal_energy(S_new,D_new,reset_e_src);
compute_new_temp (S_new,D_new);
#ifndef FORCING
{
const Real z = 1.0/a - 1.0;
fort_interp_to_this_z(&z);
}
#endif
#ifdef _OPENMP
#pragma omp parallel private(min_iter_grid,max_iter_grid) reduction(min:min_iter) reduction(max:max_iter)
#endif
for (MFIter mfi(S_new,true); mfi.isValid(); ++mfi)
{
// Here bx is just the valid region
const Box& bx = mfi.tilebox();
min_iter_grid = 100000;
max_iter_grid = 0;
integrate_state
(bx.loVect(), bx.hiVect(),
BL_TO_FORTRAN(S_new[mfi]),
BL_TO_FORTRAN(D_new[mfi]),
&a, &half_dt, &min_iter_grid, &max_iter_grid);
if (S_new[mfi].contains_nan(bx,0,S_new.nComp()))
{
std::cout << "NANS IN THIS GRID " << bx << std::endl;
}
min_iter = std::min(min_iter,min_iter_grid);
max_iter = std::max(max_iter,max_iter_grid);
}
ParallelDescriptor::ReduceIntMax(max_iter);
ParallelDescriptor::ReduceIntMin(min_iter);
if (heat_cool_type == 1)
if (ParallelDescriptor::IOProcessor())
std::cout << "Min/Max Number of Iterations in Second Strang: " << min_iter << " " << max_iter << std::endl;
}
void
MCLinOp::applyBC (MultiFab& inout,
int level,
MCBC_Mode bc_mode)
{
//
// The inout MultiFab must have at least MCLinOp_grow ghost cells
// for applyBC()
//
BL_ASSERT(inout.nGrow() >= MCLinOp_grow);
//
// The inout MultiFab must have at least Periodic_BC_grow cells for the
// algorithms taking care of periodic boundary conditions.
//
BL_ASSERT(inout.nGrow() >= MCLinOp_grow);
//
// No coarsened boundary values, cannot apply inhomog at lev>0.
//
BL_ASSERT(!(level>0 && bc_mode == MCInhomogeneous_BC));
int flagden = 1; // fill in the bndry data and undrrelxr
int flagbc = 1; // with values
if (bc_mode == MCHomogeneous_BC)
flagbc = 0; // nodata if homog
int nc = inout.nComp();
BL_ASSERT(nc == numcomp );
inout.setBndry(-1.e30);
inout.FillBoundary();
prepareForLevel(level);
geomarray[level].FillPeriodicBoundary(inout,0,nc);
//
// Fill boundary cells.
//
#ifdef _OPENMP
#pragma omp parallel
#endif
for (MFIter mfi(inout); mfi.isValid(); ++mfi)
{
const int gn = mfi.index();
BL_ASSERT(gbox[level][gn] == inout.box(gn));
const BndryData::RealTuple& bdl = bgb.bndryLocs(gn);
const Array< Array<BoundCond> >& bdc = bgb.bndryConds(gn);
const MaskTuple& msk = maskvals[level][gn];
for (OrientationIter oitr; oitr; ++oitr)
{
const Orientation face = oitr();
FabSet& f = (*undrrelxr[level])[face];
FabSet& td = (*tangderiv[level])[face];
int cdr(face);
const FabSet& fs = bgb.bndryValues(face);
Real bcl = bdl[face];
const Array<BoundCond>& bc = bdc[face];
const int *bct = (const int*) bc.dataPtr();
const FArrayBox& fsfab = fs[gn];
const Real* bcvalptr = fsfab.dataPtr();
//
// Way external derivs stored.
//
const Real* exttdptr = fsfab.dataPtr(numcomp);
const int* fslo = fsfab.loVect();
const int* fshi = fsfab.hiVect();
FArrayBox& inoutfab = inout[gn];
FArrayBox& denfab = f[gn];
FArrayBox& tdfab = td[gn];
#if BL_SPACEDIM==2
int cdir = face.coordDir(), perpdir = -1;
if (cdir == 0)
perpdir = 1;
else if (cdir == 1)
perpdir = 0;
else
BoxLib::Abort("MCLinOp::applyBC(): bad logic");
const Mask& m = *msk[face];
const Mask& mphi = *msk[Orientation(perpdir,Orientation::high)];
const Mask& mplo = *msk[Orientation(perpdir,Orientation::low)];
FORT_APPLYBC(
&flagden, &flagbc, &maxorder,
inoutfab.dataPtr(),
ARLIM(inoutfab.loVect()), ARLIM(inoutfab.hiVect()),
&cdr, bct, &bcl,
bcvalptr, ARLIM(fslo), ARLIM(fshi),
m.dataPtr(), ARLIM(m.loVect()), ARLIM(m.hiVect()),
mphi.dataPtr(), ARLIM(mphi.loVect()), ARLIM(mphi.hiVect()),
mplo.dataPtr(), ARLIM(mplo.loVect()), ARLIM(mplo.hiVect()),
denfab.dataPtr(),
ARLIM(denfab.loVect()), ARLIM(denfab.hiVect()),
exttdptr, ARLIM(fslo), ARLIM(fshi),
tdfab.dataPtr(),ARLIM(tdfab.loVect()),ARLIM(tdfab.hiVect()),
inout.box(gn).loVect(), inout.box(gn).hiVect(),
&nc, h[level]);
#elif BL_SPACEDIM==3
const Mask& mn = *msk[Orientation(1,Orientation::high)];
const Mask& me = *msk[Orientation(0,Orientation::high)];
const Mask& mw = *msk[Orientation(0,Orientation::low)];
const Mask& ms = *msk[Orientation(1,Orientation::low)];
const Mask& mt = *msk[Orientation(2,Orientation::high)];
const Mask& mb = *msk[Orientation(2,Orientation::low)];
FORT_APPLYBC(
&flagden, &flagbc, &maxorder,
inoutfab.dataPtr(),
ARLIM(inoutfab.loVect()), ARLIM(inoutfab.hiVect()),
&cdr, bct, &bcl,
bcvalptr, ARLIM(fslo), ARLIM(fshi),
mn.dataPtr(),ARLIM(mn.loVect()),ARLIM(mn.hiVect()),
me.dataPtr(),ARLIM(me.loVect()),ARLIM(me.hiVect()),
mw.dataPtr(),ARLIM(mw.loVect()),ARLIM(mw.hiVect()),
ms.dataPtr(),ARLIM(ms.loVect()),ARLIM(ms.hiVect()),
mt.dataPtr(),ARLIM(mt.loVect()),ARLIM(mt.hiVect()),
mb.dataPtr(),ARLIM(mb.loVect()),ARLIM(mb.hiVect()),
denfab.dataPtr(),
ARLIM(denfab.loVect()), ARLIM(denfab.hiVect()),
exttdptr, ARLIM(fslo), ARLIM(fshi),
tdfab.dataPtr(),ARLIM(tdfab.loVect()),ARLIM(tdfab.hiVect()),
inout.box(gn).loVect(), inout.box(gn).hiVect(),
&nc, h[level]);
#endif
}
}
#if 0
// This "probably" works, but is not strictly needed just because of the way Bill
// coded up the tangential derivative stuff. It's handy code though, so I want to
// keep it around/
// Clean up corners:
// The problem here is that APPLYBC fills only grow cells normal to the boundary.
// As a result, any corner cell on the boundary (either coarse-fine or fine-fine)
// is not filled. For coarse-fine, the operator adjusts itself, sliding away from
// the box edge to avoid referencing that corner point. On the physical boundary
// though, the corner point is needed. Particularly if a fine-fine boundary intersects
// the physical boundary, since we want the stencil to be independent of the box
// blocking. FillBoundary operations wont fix the problem because the "good"
// data we need is living in the grow region of adjacent fabs. So, here we play
// the usual games to treat the newly filled grow cells as "valid" data.
// Note that we only need to do something where the grids touch the physical boundary.
const Geometry& geomlev = geomarray[level];
const BoxArray& grids = inout.boxArray();
const Box& domain = geomlev.Domain();
int nGrow = 1;
int src_comp = 0;
int num_comp = BL_SPACEDIM;
// Lets do a quick check to see if we need to do anything at all here
BoxArray BIGba = BoxArray(grids).grow(nGrow);
if (! (domain.contains(BIGba.minimalBox())) ) {
BoxArray boundary_pieces;
Array<int> proc_idxs;
Array<Array<int> > old_to_new(grids.size());
const DistributionMapping& dmap=inout.DistributionMap();
for (int d=0; d<BL_SPACEDIM; ++d) {
if (! (geomlev.isPeriodic(d)) ) {
BoxArray gba = BoxArray(grids).grow(d,nGrow);
for (int i=0; i<gba.size(); ++i) {
BoxArray new_pieces = BoxLib::boxComplement(gba[i],domain);
int size_new = new_pieces.size();
if (size_new>0) {
int size_old = boundary_pieces.size();
boundary_pieces.resize(size_old+size_new);
proc_idxs.resize(boundary_pieces.size());
for (int j=0; j<size_new; ++j) {
boundary_pieces.set(size_old+j,new_pieces[j]);
proc_idxs[size_old+j] = dmap[i];
old_to_new[i].push_back(size_old+j);
}
}
}
}
}
proc_idxs.push_back(ParallelDescriptor::MyProc());
MultiFab boundary_data(boundary_pieces,num_comp,nGrow,
DistributionMapping(proc_idxs));
for (MFIter mfi(inout); mfi.isValid(); ++mfi) {
const FArrayBox& src_fab = inout[mfi];
for (int j=0; j<old_to_new[mfi.index()].size(); ++j) {
int new_box_idx = old_to_new[mfi.index()][j];
boundary_data[new_box_idx].copy(src_fab,src_comp,0,num_comp);
}
}
boundary_data.FillBoundary();
// Use a hacked Geometry object to handle the periodic intersections for us.
// Here, the "domain" is the plane of cells on non-periodic boundary faces.
// and there may be cells over the periodic boundary in the remaining directions.
// We do a Geometry::PFB on each non-periodic face to sync these up.
if (geomlev.isAnyPeriodic()) {
Array<int> is_per(BL_SPACEDIM,0);
for (int d=0; d<BL_SPACEDIM; ++d) {
is_per[d] = geomlev.isPeriodic(d);
}
for (int d=0; d<BL_SPACEDIM; ++d) {
if (! is_per[d]) {
Box tmpLo = BoxLib::adjCellLo(geomlev.Domain(),d,1);
Geometry tmpGeomLo(tmpLo,&(geomlev.ProbDomain()),(int)geomlev.Coord(),is_per.dataPtr());
tmpGeomLo.FillPeriodicBoundary(boundary_data);
Box tmpHi = BoxLib::adjCellHi(geomlev.Domain(),d,1);
Geometry tmpGeomHi(tmpHi,&(geomlev.ProbDomain()),(int)geomlev.Coord(),is_per.dataPtr());
tmpGeomHi.FillPeriodicBoundary(boundary_data);
}
}
}
for (MFIter mfi(inout); mfi.isValid(); ++mfi) {
int idx = mfi.index();
FArrayBox& dst_fab = inout[mfi];
for (int j=0; j<old_to_new[idx].size(); ++j) {
int new_box_idx = old_to_new[mfi.index()][j];
const FArrayBox& src_fab = boundary_data[new_box_idx];
const Box& src_box = src_fab.box();
BoxArray pieces_outside_domain = BoxLib::boxComplement(src_box,domain);
for (int k=0; k<pieces_outside_domain.size(); ++k) {
const Box& outside = pieces_outside_domain[k] & dst_fab.box();
if (outside.ok()) {
dst_fab.copy(src_fab,outside,0,outside,src_comp,num_comp);
}
}
}
}
}
#endif
}
void
Nyx::sdc_reactions (MultiFab& S_old, MultiFab& S_new, MultiFab& D_new,
MultiFab& hydro_src, MultiFab& IR,
Real delta_time, Real a_old, Real a_new, int sdc_iter)
{
BL_PROFILE("Nyx::sdc_reactions()");
const Real* dx = geom.CellSize();
// First reset internal energy before call to compute_temp
MultiFab reset_e_src(S_new.boxArray(), S_new.DistributionMap(), 1, NUM_GROW);
reset_e_src.setVal(0.0);
reset_internal_energy(S_new,D_new,reset_e_src);
compute_new_temp (S_new,D_new);
#ifndef FORCING
{
const Real z = 1.0/a_old - 1.0;
fort_interp_to_this_z(&z);
}
#endif
int min_iter = 100000;
int max_iter = 0;
int min_iter_grid, max_iter_grid;
/////////////////////Consider adding ifdefs for whether CVODE is compiled in for these statements
if(heat_cool_type == 3)
{
#ifdef _OPENMP
#pragma omp parallel
#endif
for (MFIter mfi(S_old,true); mfi.isValid(); ++mfi)
{
// Note that this "bx" is only the valid region (unlike for Strang)
const Box& bx = mfi.tilebox();
min_iter_grid = 100000;
max_iter_grid = 0;
integrate_state_with_source
(bx.loVect(), bx.hiVect(),
BL_TO_FORTRAN(S_old[mfi]),
BL_TO_FORTRAN(S_new[mfi]),
BL_TO_FORTRAN(D_new[mfi]),
BL_TO_FORTRAN(hydro_src[mfi]),
BL_TO_FORTRAN(reset_e_src[mfi]),
BL_TO_FORTRAN(IR[mfi]),
&a_old, &delta_time, &min_iter_grid, &max_iter_grid);
min_iter = std::min(min_iter,min_iter_grid);
max_iter = std::max(max_iter,max_iter_grid);
}
}
else if(heat_cool_type == 5)
{
#ifdef _OPENMP
#pragma omp parallel
#endif
for (MFIter mfi(S_old,true); mfi.isValid(); ++mfi)
{
// Note that this "bx" is only the valid region (unlike for Strang)
const Box& bx = mfi.tilebox();
min_iter_grid = 100000;
max_iter_grid = 0;
integrate_state_fcvode_with_source
(bx.loVect(), bx.hiVect(),
BL_TO_FORTRAN(S_old[mfi]),
BL_TO_FORTRAN(S_new[mfi]),
BL_TO_FORTRAN(D_new[mfi]),
BL_TO_FORTRAN(hydro_src[mfi]),
BL_TO_FORTRAN(reset_e_src[mfi]),
BL_TO_FORTRAN(IR[mfi]),
&a_old, &delta_time, &min_iter_grid, &max_iter_grid);
min_iter = std::min(min_iter,min_iter_grid);
max_iter = std::max(max_iter,max_iter_grid);
}
}
ParallelDescriptor::ReduceIntMax(max_iter);
ParallelDescriptor::ReduceIntMin(min_iter);
amrex::Print() << "Min/Max Number of Iterations in SDC: " << min_iter << " " << max_iter << std::endl;
}
int
CGSolver::solve_bicgstab (MultiFab& sol,
const MultiFab& rhs,
Real eps_rel,
Real eps_abs,
LinOp::BC_Mode bc_mode)
{
BL_PROFILE("CGSolver::solve_bicgstab()");
const int nghost = sol.nGrow(), ncomp = 1;
const BoxArray& ba = sol.boxArray();
const DistributionMapping& dm = sol.DistributionMap();
BL_ASSERT(sol.nComp() == ncomp);
BL_ASSERT(sol.boxArray() == Lp.boxArray(lev));
BL_ASSERT(rhs.boxArray() == Lp.boxArray(lev));
MultiFab ph(ba, ncomp, nghost, dm);
MultiFab sh(ba, ncomp, nghost, dm);
MultiFab sorig(ba, ncomp, 0, dm);
MultiFab p (ba, ncomp, 0, dm);
MultiFab r (ba, ncomp, 0, dm);
MultiFab s (ba, ncomp, 0, dm);
MultiFab rh (ba, ncomp, 0, dm);
MultiFab v (ba, ncomp, 0, dm);
MultiFab t (ba, ncomp, 0, dm);
Lp.residual(r, rhs, sol, lev, bc_mode);
MultiFab::Copy(sorig,sol,0,0,1,0);
MultiFab::Copy(rh, r, 0,0,1,0);
sol.setVal(0);
const LinOp::BC_Mode temp_bc_mode = LinOp::Homogeneous_BC;
#ifdef CG_USE_OLD_CONVERGENCE_CRITERIA
Real rnorm = norm_inf(r);
#else
//
// Calculate the local values of these norms & reduce their values together.
//
Real vals[2] = { norm_inf(r, true), Lp.norm(0, lev, true) };
ParallelDescriptor::ReduceRealMax(vals,2,color());
Real rnorm = vals[0];
const Real Lp_norm = vals[1];
Real sol_norm = 0;
#endif
const Real rnorm0 = rnorm;
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_BiCGStab: Initial error (error0) = " << rnorm0 << '\n';
}
int ret = 0, nit = 1;
Real rho_1 = 0, alpha = 0, omega = 0;
if ( rnorm0 == 0 || rnorm0 < eps_abs )
{
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_BiCGStab: niter = 0,"
<< ", rnorm = " << rnorm
<< ", eps_abs = " << eps_abs << std::endl;
}
return ret;
}
for (; nit <= maxiter; ++nit)
{
const Real rho = dotxy(rh,r);
if ( rho == 0 )
{
ret = 1; break;
}
if ( nit == 1 )
{
MultiFab::Copy(p,r,0,0,1,0);
}
else
{
const Real beta = (rho/rho_1)*(alpha/omega);
sxay(p, p, -omega, v);
sxay(p, r, beta, p);
}
if ( use_mg_precond )
{
ph.setVal(0);
mg_precond->solve(ph, p, eps_rel, eps_abs, temp_bc_mode);
}
else if ( use_jacobi_precond )
{
ph.setVal(0);
Lp.jacobi_smooth(ph, p, lev, temp_bc_mode);
}
else
{
MultiFab::Copy(ph,p,0,0,1,0);
}
Lp.apply(v, ph, lev, temp_bc_mode);
if ( Real rhTv = dotxy(rh,v) )
{
alpha = rho/rhTv;
}
else
{
ret = 2; break;
}
sxay(sol, sol, alpha, ph);
sxay(s, r, -alpha, v);
rnorm = norm_inf(s);
if ( verbose > 2 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_BiCGStab: Half Iter "
<< std::setw(11) << nit
<< " rel. err. "
<< rnorm/(rnorm0) << '\n';
}
#ifdef CG_USE_OLD_CONVERGENCE_CRITERIA
if ( rnorm < eps_rel*rnorm0 || rnorm < eps_abs ) break;
#else
sol_norm = norm_inf(sol);
if ( rnorm < eps_rel*(Lp_norm*sol_norm + rnorm0 ) || rnorm < eps_abs ) break;
#endif
if ( use_mg_precond )
{
sh.setVal(0);
mg_precond->solve(sh, s, eps_rel, eps_abs, temp_bc_mode);
}
else if ( use_jacobi_precond )
{
sh.setVal(0);
Lp.jacobi_smooth(sh, s, lev, temp_bc_mode);
}
else
{
MultiFab::Copy(sh,s,0,0,1,0);
}
Lp.apply(t, sh, lev, temp_bc_mode);
//
// This is a little funky. I want to elide one of the reductions
// in the following two dotxy()s. We do that by calculating the "local"
// values and then reducing the two local values at the same time.
//
Real vals[2] = { dotxy(t,t,true), dotxy(t,s,true) };
ParallelDescriptor::ReduceRealSum(vals,2,color());
if ( vals[0] )
{
omega = vals[1]/vals[0];
}
else
{
ret = 3; break;
}
sxay(sol, sol, omega, sh);
sxay(r, s, -omega, t);
rnorm = norm_inf(r);
if ( verbose > 2 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_BiCGStab: Iteration "
<< std::setw(11) << nit
<< " rel. err. "
<< rnorm/(rnorm0) << '\n';
}
#ifdef CG_USE_OLD_CONVERGENCE_CRITERIA
if ( rnorm < eps_rel*rnorm0 || rnorm < eps_abs ) break;
#else
sol_norm = norm_inf(sol);
if ( rnorm < eps_rel*(Lp_norm*sol_norm + rnorm0 ) || rnorm < eps_abs ) break;
#endif
if ( omega == 0 )
{
ret = 4; break;
}
rho_1 = rho;
}
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_BiCGStab: Final: Iteration "
<< std::setw(4) << nit
<< " rel. err. "
<< rnorm/(rnorm0) << '\n';
}
#ifdef CG_USE_OLD_CONVERGENCE_CRITERIA
if ( ret == 0 && rnorm > eps_rel*rnorm0 && rnorm > eps_abs)
#else
if ( ret == 0 && rnorm > eps_rel*(Lp_norm*sol_norm + rnorm0 ) && rnorm > eps_abs )
#endif
{
if ( ParallelDescriptor::IOProcessor(color()) )
BoxLib::Warning("CGSolver_BiCGStab:: failed to converge!");
ret = 8;
}
if ( ( ret == 0 || ret == 8 ) && (rnorm < rnorm0) )
{
sol.plus(sorig, 0, 1, 0);
}
else
{
sol.setVal(0);
sol.plus(sorig, 0, 1, 0);
}
return ret;
}
int
CGSolver::solve_cabicgstab (MultiFab& sol,
const MultiFab& rhs,
Real eps_rel,
Real eps_abs,
LinOp::BC_Mode bc_mode)
{
BL_PROFILE("CGSolver::solve_cabicgstab()");
BL_ASSERT(sol.nComp() == 1);
BL_ASSERT(sol.boxArray() == Lp.boxArray(lev));
BL_ASSERT(rhs.boxArray() == Lp.boxArray(lev));
Real temp1[4*SSS_MAX+1];
Real temp2[4*SSS_MAX+1];
Real temp3[4*SSS_MAX+1];
Real Tp[4*SSS_MAX+1][4*SSS_MAX+1];
Real Tpp[4*SSS_MAX+1][4*SSS_MAX+1];
Real aj[4*SSS_MAX+1];
Real cj[4*SSS_MAX+1];
Real ej[4*SSS_MAX+1];
Real Tpaj[4*SSS_MAX+1];
Real Tpcj[4*SSS_MAX+1];
Real Tppaj[4*SSS_MAX+1];
Real G[4*SSS_MAX+1][4*SSS_MAX+1]; // Extracted from first 4*SSS+1 columns of Gg[][]. indexed as [row][col]
Real g[4*SSS_MAX+1]; // Extracted from last [4*SSS+1] column of Gg[][].
Real Gg[(4*SSS_MAX+1)*(4*SSS_MAX+2)]; // Buffer to hold the Gram-like matrix produced by matmul(). indexed as [row*(4*SSS+2) + col]
//
// If variable_SSS we "telescope" SSS.
// We start with 1 and increase it up to SSS_MAX on the outer iterations.
//
if (variable_SSS) SSS = 1;
zero( aj, 4*SSS_MAX+1);
zero( cj, 4*SSS_MAX+1);
zero( ej, 4*SSS_MAX+1);
zero( Tpaj, 4*SSS_MAX+1);
zero( Tpcj, 4*SSS_MAX+1);
zero(Tppaj, 4*SSS_MAX+1);
zero(temp1, 4*SSS_MAX+1);
zero(temp2, 4*SSS_MAX+1);
zero(temp3, 4*SSS_MAX+1);
SetMonomialBasis(Tp,Tpp,SSS);
const int ncomp = 1, nghost = sol.nGrow();
//
// Contains the matrix powers of p[] and r[].
//
// First 2*SSS+1 components are powers of p[].
// Next 2*SSS components are powers of r[].
//
const BoxArray& ba = sol.boxArray();
const DistributionMapping& dm = sol.DistributionMap();
MultiFab PR(ba, 4*SSS_MAX+1, 0, dm);
MultiFab p(ba, ncomp, 0, dm);
MultiFab r(ba, ncomp, 0, dm);
MultiFab rt(ba, ncomp, 0, dm);
MultiFab tmp(ba, 4, nghost, dm);
Lp.residual(r, rhs, sol, lev, bc_mode);
BL_ASSERT(!r.contains_nan());
MultiFab::Copy(rt,r,0,0,1,0);
MultiFab::Copy( p,r,0,0,1,0);
const Real rnorm0 = norm_inf(r);
Real delta = dotxy(r,rt);
const Real L2_norm_of_rt = sqrt(delta);
const LinOp::BC_Mode temp_bc_mode = LinOp::Homogeneous_BC;
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_CABiCGStab: Initial error (error0) = " << rnorm0 << '\n';
}
if ( rnorm0 == 0 || delta == 0 || rnorm0 < eps_abs )
{
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_CABiCGStab: niter = 0,"
<< ", rnorm = " << rnorm0
<< ", delta = " << delta
<< ", eps_abs = " << eps_abs << '\n';
}
return 0;
}
int niters = 0, ret = 0;
Real L2_norm_of_resid = 0, atime = 0, gtime = 0;
bool BiCGStabFailed = false, BiCGStabConverged = false;
for (int m = 0; m < maxiter && !BiCGStabFailed && !BiCGStabConverged; )
{
const Real time1 = ParallelDescriptor::second();
//
// Compute the matrix powers on p[] & r[] (monomial basis).
// The 2*SSS+1 powers of p[] followed by the 2*SSS powers of r[].
//
MultiFab::Copy(PR,p,0,0,1,0);
MultiFab::Copy(PR,r,0,2*SSS+1,1,0);
BL_ASSERT(!PR.contains_nan(0, 1));
BL_ASSERT(!PR.contains_nan(2*SSS+1,1));
//
// We use "tmp" to minimize the number of Lp.apply()s.
// We do this by doing p & r together in a single call.
//
MultiFab::Copy(tmp,p,0,0,1,0);
MultiFab::Copy(tmp,r,0,1,1,0);
for (int n = 1; n < 2*SSS; n++)
{
Lp.apply(tmp, tmp, lev, temp_bc_mode, false, 0, 2, 2);
MultiFab::Copy(tmp,tmp,2,0,2,0);
MultiFab::Copy(PR,tmp,0, n,1,0);
MultiFab::Copy(PR,tmp,1,2*SSS+n+1,1,0);
BL_ASSERT(!PR.contains_nan(n, 1));
BL_ASSERT(!PR.contains_nan(2*SSS+n+1,1));
}
MultiFab::Copy(tmp,PR,2*SSS-1,0,1,0);
Lp.apply(tmp, tmp, lev, temp_bc_mode, false, 0, 1, 1);
MultiFab::Copy(PR,tmp,1,2*SSS,1,0);
BL_ASSERT(!PR.contains_nan(2*SSS-1,1));
BL_ASSERT(!PR.contains_nan(2*SSS, 1));
Real time2 = ParallelDescriptor::second();
atime += (time2-time1);
BuildGramMatrix(Gg, PR, rt, SSS);
const Real time3 = ParallelDescriptor::second();
gtime += (time3-time2);
//
// Form G[][] and g[] from Gg.
//
for (int i = 0, k = 0; i < 4*SSS+1; i++)
{
for (int j = 0; j < 4*SSS+1; j++)
//
// First 4*SSS+1 elements in each row go to G[][].
//
G[i][j] = Gg[k++];
//
// Last element in row goes to g[].
//
g[i] = Gg[k++];
}
zero(aj, 4*SSS+1); aj[0] = 1;
zero(cj, 4*SSS+1); cj[2*SSS+1] = 1;
zero(ej, 4*SSS+1);
for (int nit = 0; nit < SSS; nit++)
{
gemv( Tpaj, Tp, aj, 4*SSS+1, 4*SSS+1);
gemv( Tpcj, Tp, cj, 4*SSS+1, 4*SSS+1);
gemv(Tppaj, Tpp, aj, 4*SSS+1, 4*SSS+1);
const Real g_dot_Tpaj = dot(g, Tpaj, 4*SSS+1);
if ( g_dot_Tpaj == 0 )
{
if ( verbose > 1 && ParallelDescriptor::IOProcessor(color()) )
std::cout << "CGSolver_CABiCGStab: g_dot_Tpaj == 0, nit = " << nit << '\n';
BiCGStabFailed = true; ret = 1; break;
}
const Real alpha = delta / g_dot_Tpaj;
if ( std::isinf(alpha) )
{
if ( verbose > 1 && ParallelDescriptor::IOProcessor(color()) )
std::cout << "CGSolver_CABiCGStab: alpha == inf, nit = " << nit << '\n';
BiCGStabFailed = true; ret = 2; break;
}
axpy(temp1, Tpcj, -alpha, Tppaj, 4*SSS+1);
gemv(temp2, G, temp1, 4*SSS+1, 4*SSS+1);
axpy(temp3, cj, -alpha, Tpaj, 4*SSS+1);
const Real omega_numerator = dot(temp3, temp2, 4*SSS+1);
const Real omega_denominator = dot(temp1, temp2, 4*SSS+1);
//
// NOTE: omega_numerator/omega_denominator can be 0/x or 0/0, but should never be x/0.
//
// If omega_numerator==0, and ||s||==0, then convergence, x=x+alpha*aj.
// If omega_numerator==0, and ||s||!=0, then stabilization breakdown.
//
// Partial update of ej must happen before the check on omega to ensure forward progress !!!
//
axpy(ej, ej, alpha, aj, 4*SSS+1);
//
// ej has been updated so consider that we've done an iteration since
// even if we break out of the loop we'll be able to update both sol.
//
niters++;
//
// Calculate the norm of Saad's vector 's' to check intra s-step convergence.
//
axpy(temp1, cj,-alpha, Tpaj, 4*SSS+1);
gemv(temp2, G, temp1, 4*SSS+1, 4*SSS+1);
const Real L2_norm_of_s = dot(temp1,temp2,4*SSS+1);
L2_norm_of_resid = (L2_norm_of_s < 0 ? 0 : sqrt(L2_norm_of_s));
if ( L2_norm_of_resid < eps_rel*L2_norm_of_rt )
{
if ( verbose > 1 && L2_norm_of_resid == 0 && ParallelDescriptor::IOProcessor(color()) )
std::cout << "CGSolver_CABiCGStab: L2 norm of s: " << L2_norm_of_s << '\n';
BiCGStabConverged = true; break;
}
if ( omega_denominator == 0 )
{
if ( verbose > 1 && ParallelDescriptor::IOProcessor(color()) )
std::cout << "CGSolver_CABiCGStab: omega_denominator == 0, nit = " << nit << '\n';
BiCGStabFailed = true; ret = 3; break;
}
const Real omega = omega_numerator / omega_denominator;
if ( verbose > 1 && ParallelDescriptor::IOProcessor(color()) )
{
if ( omega == 0 ) std::cout << "CGSolver_CABiCGStab: omega == 0, nit = " << nit << '\n';
if ( std::isinf(omega) ) std::cout << "CGSolver_CABiCGStab: omega == inf, nit = " << nit << '\n';
}
if ( omega == 0 ) { BiCGStabFailed = true; ret = 4; break; }
if ( std::isinf(omega) ) { BiCGStabFailed = true; ret = 4; break; }
//
// Complete the update of ej & cj now that omega is known to be ok.
//
axpy(ej, ej, omega, cj, 4*SSS+1);
axpy(ej, ej,-omega*alpha, Tpaj, 4*SSS+1);
axpy(cj, cj, -omega, Tpcj, 4*SSS+1);
axpy(cj, cj, -alpha, Tpaj, 4*SSS+1);
axpy(cj, cj, omega*alpha, Tppaj, 4*SSS+1);
//
// Do an early check of the residual to determine convergence.
//
gemv(temp1, G, cj, 4*SSS+1, 4*SSS+1);
//
// sqrt( (cj,Gcj) ) == L2 norm of the intermediate residual in exact arithmetic.
// However, finite precision can lead to the norm^2 being < 0 (Jim Demmel).
// If cj_dot_Gcj < 0 we flush to zero and consider ourselves converged.
//
const Real L2_norm_of_r = dot(cj, temp1, 4*SSS+1);
L2_norm_of_resid = (L2_norm_of_r > 0 ? sqrt(L2_norm_of_r) : 0);
if ( L2_norm_of_resid < eps_rel*L2_norm_of_rt )
{
if ( verbose > 1 && L2_norm_of_resid == 0 && ParallelDescriptor::IOProcessor(color()) )
std::cout << "CGSolver_CABiCGStab: L2_norm_of_r: " << L2_norm_of_r << '\n';
BiCGStabConverged = true; break;
}
const Real delta_next = dot(g, cj, 4*SSS+1);
if ( verbose > 1 && ParallelDescriptor::IOProcessor(color()) )
{
if ( delta_next == 0 ) std::cout << "CGSolver_CABiCGStab: delta == 0, nit = " << nit << '\n';
if ( std::isinf(delta_next) ) std::cout << "CGSolver_CABiCGStab: delta == inf, nit = " << nit << '\n';
}
if ( std::isinf(delta_next) ) { BiCGStabFailed = true; ret = 5; break; } // delta = inf?
if ( delta_next == 0 ) { BiCGStabFailed = true; ret = 5; break; } // Lanczos breakdown...
const Real beta = (delta_next/delta)*(alpha/omega);
if ( verbose > 1 && ParallelDescriptor::IOProcessor(color()) )
{
if ( beta == 0 ) std::cout << "CGSolver_CABiCGStab: beta == 0, nit = " << nit << '\n';
if ( std::isinf(beta) ) std::cout << "CGSolver_CABiCGStab: beta == inf, nit = " << nit << '\n';
}
if ( std::isinf(beta) ) { BiCGStabFailed = true; ret = 6; break; } // beta = inf?
if ( beta == 0 ) { BiCGStabFailed = true; ret = 6; break; } // beta = 0? can't make further progress(?)
axpy(aj, cj, beta, aj, 4*SSS+1);
axpy(aj, aj, -omega*beta, Tpaj, 4*SSS+1);
delta = delta_next;
}
//
// Update iterates.
//
for (int i = 0; i < 4*SSS+1; i++)
sxay(sol,sol,ej[i],PR,i);
MultiFab::Copy(p,PR,0,0,1,0);
p.mult(aj[0],0,1);
for (int i = 1; i < 4*SSS+1; i++)
sxay(p,p,aj[i],PR,i);
MultiFab::Copy(r,PR,0,0,1,0);
r.mult(cj[0],0,1);
for (int i = 1; i < 4*SSS+1; i++)
sxay(r,r,cj[i],PR,i);
if ( !BiCGStabFailed && !BiCGStabConverged )
{
m += SSS;
if ( variable_SSS && SSS < SSS_MAX ) { SSS++; SetMonomialBasis(Tp,Tpp,SSS); }
}
}
if ( verbose > 0 )
{
if ( ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_CABiCGStab: Final: Iteration "
<< std::setw(4) << niters
<< " rel. err. "
<< L2_norm_of_resid << '\n';
}
if ( verbose > 1 )
{
Real tmp[2] = { atime, gtime };
ParallelDescriptor::ReduceRealMax(tmp,2,color());
if ( ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_CABiCGStab apply time: " << tmp[0] << ", gram time: " << tmp[1] << '\n';
}
}
}
if ( niters >= maxiter && !BiCGStabFailed && !BiCGStabConverged)
{
if ( L2_norm_of_resid > L2_norm_of_rt )
{
if ( ParallelDescriptor::IOProcessor(color()) )
BoxLib::Warning("CGSolver_CABiCGStab: failed to converge!");
//
// Return code 8 tells the MultiGrid driver to zero out the solution!
//
ret = 8;
}
else
{
//
// Return codes 1-7 tells the MultiGrid driver to smooth the solution!
//
ret = 7;
}
}
return ret;
}
int
CGSolver::jbb_precond (MultiFab& sol,
const MultiFab& rhs,
int lev,
LinOp& Lp)
{
//
// This is a local routine. No parallel is allowed to happen here.
//
int lev_loc = lev;
const Real eps_rel = 1.e-2;
const Real eps_abs = 1.e-16;
const int nghost = sol.nGrow();
const int ncomp = sol.nComp();
const bool local = true;
const LinOp::BC_Mode bc_mode = LinOp::Homogeneous_BC;
BL_ASSERT(ncomp == 1 );
BL_ASSERT(sol.boxArray() == Lp.boxArray(lev_loc));
BL_ASSERT(rhs.boxArray() == Lp.boxArray(lev_loc));
const BoxArray& ba = sol.boxArray();
const DistributionMapping& dm = sol.DistributionMap();
MultiFab sorig(ba, ncomp, nghost, dm);
MultiFab r(ba, ncomp, nghost, dm);
MultiFab z(ba, ncomp, nghost, dm);
MultiFab q(ba, ncomp, nghost, dm);
MultiFab p(ba, ncomp, nghost, dm);
sorig.copy(sol);
Lp.residual(r, rhs, sorig, lev_loc, LinOp::Homogeneous_BC, local);
sol.setVal(0);
Real rnorm = norm_inf(r,local);
const Real rnorm0 = rnorm;
Real minrnorm = rnorm;
if ( verbose > 2 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev_loc);
std::cout << " jbb_precond: Initial error : " << rnorm0 << '\n';
}
const Real Lp_norm = Lp.norm(0, lev_loc, local);
Real sol_norm = 0;
int ret = 0; // will return this value if all goes well
Real rho_1 = 0;
int nit = 1;
if ( rnorm0 == 0 || rnorm0 < eps_abs )
{
if ( verbose > 2 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev_loc);
std::cout << "jbb_precond: niter = 0,"
<< ", rnorm = " << rnorm
<< ", eps_abs = " << eps_abs << std::endl;
}
return 0;
}
for (; nit <= maxiter; ++nit)
{
z.copy(r);
Real rho = dotxy(z,r,local);
if (nit == 1)
{
p.copy(z);
}
else
{
Real beta = rho/rho_1;
sxay(p, z, beta, p);
}
Lp.apply(q, p, lev_loc, bc_mode, local);
Real alpha;
if ( Real pw = dotxy(p,q,local) )
{
alpha = rho/pw;
}
else
{
ret = 1; break;
}
if ( verbose > 3 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev_loc);
std::cout << "jbb_precond:" << " nit " << nit
<< " rho " << rho << " alpha " << alpha << '\n';
}
sxay(sol, sol, alpha, p);
sxay( r, r,-alpha, q);
rnorm = norm_inf(r, local);
sol_norm = norm_inf(sol, local);
if ( verbose > 2 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev_loc);
std::cout << "jbb_precond: Iteration"
<< std::setw(4) << nit
<< " rel. err. "
<< rnorm/(rnorm0) << '\n';
}
if ( rnorm < eps_rel*(Lp_norm*sol_norm + rnorm0) || rnorm < eps_abs )
{
break;
}
if ( rnorm > def_unstable_criterion*minrnorm )
{
ret = 2; break;
}
else if ( rnorm < minrnorm )
{
minrnorm = rnorm;
}
rho_1 = rho;
}
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev_loc);
std::cout << "jbb_precond: Final Iteration"
<< std::setw(4) << nit
<< " rel. err. "
<< rnorm/(rnorm0) << '\n';
}
if ( ret == 0 && rnorm > eps_rel*(Lp_norm*sol_norm + rnorm0) && rnorm > eps_abs )
{
if ( ParallelDescriptor::IOProcessor(color()) )
{
BoxLib::Warning("jbb_precond:: failed to converge!");
}
ret = 8;
}
if ( ( ret == 0 || ret == 8 ) && (rnorm < rnorm0) )
{
sol.plus(sorig, 0, 1, 0);
}
else
{
sol.setVal(0);
sol.plus(sorig, 0, 1, 0);
}
return ret;
}
int
CGSolver::solve_cg (MultiFab& sol,
const MultiFab& rhs,
Real eps_rel,
Real eps_abs,
LinOp::BC_Mode bc_mode)
{
BL_PROFILE("CGSolver::solve_cg()");
const int nghost = sol.nGrow(), ncomp = 1;
const BoxArray& ba = sol.boxArray();
const DistributionMapping& dm = sol.DistributionMap();
BL_ASSERT(sol.nComp() == ncomp);
BL_ASSERT(sol.boxArray() == Lp.boxArray(lev));
BL_ASSERT(rhs.boxArray() == Lp.boxArray(lev));
MultiFab sorig(ba, ncomp, nghost, dm);
MultiFab r(ba, ncomp, nghost, dm);
MultiFab z(ba, ncomp, nghost, dm);
MultiFab q(ba, ncomp, nghost, dm);
MultiFab p(ba, ncomp, nghost, dm);
MultiFab r1(ba, ncomp, nghost, dm);
MultiFab z1(ba, ncomp, nghost, dm);
MultiFab r2(ba, ncomp, nghost, dm);
MultiFab z2(ba, ncomp, nghost, dm);
MultiFab::Copy(sorig,sol,0,0,1,0);
Lp.residual(r, rhs, sorig, lev, bc_mode);
sol.setVal(0);
const LinOp::BC_Mode temp_bc_mode=LinOp::Homogeneous_BC;
Real rnorm = norm_inf(r);
const Real rnorm0 = rnorm;
Real minrnorm = rnorm;
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << " CG: Initial error : " << rnorm0 << '\n';
}
const Real Lp_norm = Lp.norm(0, lev);
Real sol_norm = 0;
Real rho_1 = 0;
int ret = 0;
int nit = 1;
if ( rnorm == 0 || rnorm < eps_abs )
{
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << " CG: niter = 0,"
<< ", rnorm = " << rnorm
<< ", eps_rel*(Lp_norm*sol_norm + rnorm0 )" << eps_rel*(Lp_norm*sol_norm + rnorm0 )
<< ", eps_abs = " << eps_abs << std::endl;
}
return 0;
}
for (; nit <= maxiter; ++nit)
{
if (use_jbb_precond && ParallelDescriptor::NProcs(color()) > 1)
{
z.setVal(0);
jbb_precond(z,r,lev,Lp);
}
else
{
MultiFab::Copy(z,r,0,0,1,0);
}
Real rho = dotxy(z,r);
if (nit == 1)
{
MultiFab::Copy(p,z,0,0,1,0);
}
else
{
Real beta = rho/rho_1;
sxay(p, z, beta, p);
}
Lp.apply(q, p, lev, temp_bc_mode);
Real alpha;
if ( Real pw = dotxy(p,q) )
{
alpha = rho/pw;
}
else
{
ret = 1; break;
}
if ( verbose > 2 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << "CGSolver_cg:"
<< " nit " << nit
<< " rho " << rho
<< " alpha " << alpha << '\n';
}
sxay(sol, sol, alpha, p);
sxay( r, r,-alpha, q);
rnorm = norm_inf(r);
sol_norm = norm_inf(sol);
if ( verbose > 2 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << " CG: Iteration"
<< std::setw(4) << nit
<< " rel. err. "
<< rnorm/(rnorm0) << '\n';
}
#ifdef CG_USE_OLD_CONVERGENCE_CRITERIA
if ( rnorm < eps_rel*rnorm0 || rnorm < eps_abs ) break;
#else
if ( rnorm < eps_rel*(Lp_norm*sol_norm + rnorm0) || rnorm < eps_abs ) break;
#endif
if ( rnorm > def_unstable_criterion*minrnorm )
{
ret = 2; break;
}
else if ( rnorm < minrnorm )
{
minrnorm = rnorm;
}
rho_1 = rho;
}
if ( verbose > 0 && ParallelDescriptor::IOProcessor(color()) )
{
Spacer(std::cout, lev);
std::cout << " CG: Final Iteration"
<< std::setw(4) << nit
<< " rel. err. "
<< rnorm/(rnorm0) << '\n';
}
#ifdef CG_USE_OLD_CONVERGENCE_CRITERIA
if ( ret == 0 && rnorm > eps_rel*rnorm0 && rnorm > eps_abs )
#else
if ( ret == 0 && rnorm > eps_rel*(Lp_norm*sol_norm + rnorm0) && rnorm > eps_abs )
#endif
{
if ( ParallelDescriptor::IOProcessor(color()) )
BoxLib::Warning("CGSolver_cg: failed to converge!");
ret = 8;
}
if ( ( ret == 0 || ret == 8 ) && (rnorm < rnorm0) )
{
sol.plus(sorig, 0, 1, 0);
}
else
{
sol.setVal(0);
sol.plus(sorig, 0, 1, 0);
}
return ret;
}