virtual void linearIn(EBCellFAB& arg, void* buf, const Box& R,
const Interval& comps) const
{
EBCellFAB tmp;
tmp.clone(arg);
tmp.linearIn(buf, R, comps);
arg.plus(tmp, R, comps.begin(), comps.begin(), comps.size());
}
void
EBMGAverage::averageFAB(EBCellFAB& a_coar,
const Box& a_boxCoar,
const EBCellFAB& a_refCoar,
const DataIndex& a_datInd,
const Interval& a_variables) const
{
CH_TIMERS("EBMGAverage::average");
CH_TIMER("regular_average", t1);
CH_TIMER("irregular_average", t2);
CH_assert(isDefined());
const Box& coarBox = a_boxCoar;
//do all cells as if they were regular
Box refBox(IntVect::Zero, IntVect::Zero);
refBox.refine(m_refRat);
int numFinePerCoar = refBox.numPts();
BaseFab<Real>& coarRegFAB = a_coar.getSingleValuedFAB();
const BaseFab<Real>& refCoarRegFAB = a_refCoar.getSingleValuedFAB();
//set to zero because the fortran is a bit simpleminded
//and does stuff additively
a_coar.setVal(0.);
CH_START(t1);
for (int comp = a_variables.begin(); comp <= a_variables.end(); comp++)
{
FORT_REGAVERAGE(CHF_FRA1(coarRegFAB,comp),
CHF_CONST_FRA1(refCoarRegFAB,comp),
CHF_BOX(coarBox),
CHF_BOX(refBox),
CHF_CONST_INT(numFinePerCoar),
CHF_CONST_INT(m_refRat));
}
CH_STOP(t1);
//this is really volume-weighted averaging even though it does
//not look that way.
//so (in the traditional sense) we want to preserve
//rhoc * volc = sum(rhof * volf)
//this translates to
//volfrac_C * rhoC = (1/numFinePerCoar)(sum(volFrac_F * rhoF))
//but the data input to this routine is all kappa weigthed so
//the volumefractions have already been multiplied
//which means
// rhoC = (1/numFinePerCoar)(sum(rhoF))
//which is what this does
CH_START(t2);
for (int comp = a_variables.begin(); comp <= a_variables.end(); comp++)
{
m_averageEBStencil[a_datInd]->apply(a_coar, a_refCoar, false, comp);
}
CH_STOP(t2);
}
void op(EBCellFAB& dest,
const Box& RegionFrom,
const Interval& Cdest,
const Box& RegionTo,
const EBCellFAB& src,
const Interval& Csrc) const
{
dest.plus(src, RegionFrom, Csrc.begin(), Cdest.begin(), Cdest.size());
}
void
MappedLevelFluxRegister::incrementCoarse(const FArrayBox& a_coarseFlux,
Real a_scale,
const DataIndex& a_coarseDataIndex,
const Interval& a_srcInterval,
const Interval& a_dstInterval,
int a_dir,
Side::LoHiSide a_sd)
{
CH_assert(isDefined());
if (!(m_isDefined & FluxRegCoarseDefined)) return;
CH_TIME("MappedLevelFluxRegister::incrementCoarse");
const Vector<Box>& intersect =
m_coarseLocations[a_dir + a_sd * CH_SPACEDIM][a_coarseDataIndex];
FArrayBox& coarse = m_coarFlux[a_coarseDataIndex];
// We cast away the constness in a_coarseFlux for the scope of this function. This
// should be acceptable, since at the end of the day there is no change to it. -JNJ
FArrayBox& coarseFlux = const_cast<FArrayBox&>(a_coarseFlux); // Muhahaha.
coarseFlux.shiftHalf(a_dir, sign(a_sd));
Real scale = -sign(a_sd) * a_scale;
int s = a_srcInterval.begin();
int d = a_dstInterval.begin();
int size = a_srcInterval.size();
for (int b = 0; b < intersect.size(); ++b) {
const Box& box = intersect[b];
Vector<Real> regbefore(coarse.nComp());
Vector<Real> regafter(coarse.nComp());
if (s_verbose && (a_dir == debugdir) && box.contains(ivdebnoeb)) {
for (int ivar = 0; ivar < coarse.nComp(); ivar++) {
regbefore[ivar] = coarse(ivdebnoeb, ivar);
}
}
coarse.plus(coarseFlux, box, box, scale, s, d, size);
if (s_verbose && (a_dir == debugdir) && box.contains(ivdebnoeb)) {
for (int ivar = 0; ivar < coarse.nComp(); ivar++) {
regafter[ivar] = coarse(ivdebnoeb, ivar);
}
pout() << "levelfluxreg::incrementCoar: scale = " << scale << ", ";
for (int ivar = 0; ivar < coarse.nComp(); ivar++) {
pout() << " input flux = " << coarseFlux(ivdebnoeb, ivar) << ", ";
pout() << " reg before = " << regbefore[ivar] << ", ";
pout() << " reg after = " << regafter[ivar] << ", ";
}
pout() << endl;
}
}
coarseFlux.shiftHalf(a_dir, - sign(a_sd));
}
void
EBCoarsen::coarsenIrreg(EBCellFAB& a_coar,
const EBCellFAB& a_fine,
const DataIndex& a_dit,
const Interval& a_variables)
{
const BaseIVFAB<VoFStencil>& stenBaseIVFAB = m_coarsenStencil[a_dit];
VoFIterator& vofit = m_vofIt[a_dit];
for (vofit.reset(); vofit.ok(); ++vofit)
{
const VolIndex& vofCoar = vofit();
const VoFStencil& stencil = stenBaseIVFAB(vofCoar,0);
for (int icomp=a_variables.begin();icomp ==a_variables.end();icomp++)
{
//coarsen irreg fine vofs to coarse
// compute the coarsening by using fine data
Real phi = 0.0;
for (int i = 0; i < stencil.size(); ++i )
{
const Real& weight = stencil.weight(i);
const VolIndex& vofFine = stencil.vof(i);
const Real& phiF = a_fine(vofFine,icomp);
phi += weight * phiF;
}
//set the coarse value
a_coar(vofCoar,icomp) = phi;
}
}
}
// ---------------------------------------------------------
// 28 March 2003:
// This is called by other functions, and should not be called directly.
Real
integral(const BoxLayoutData<NodeFArrayBox>& a_layout,
const Real a_dx,
const Interval& a_interval,
bool a_verbose)
{
Real integralTotal = 0.;
for (DataIterator it = a_layout.dataIterator(); it.ok(); ++it)
{
const NodeFArrayBox& thisNfab = a_layout[it()];
const Box& thisBox(a_layout.box(it())); // CELL-centered
Real thisNfabIntegral =
integral(thisNfab, a_dx, thisBox,
a_interval.begin(), a_interval.size());
integralTotal += thisNfabIntegral;
}
# ifdef CH_MPI
Real recv;
// add up
int result = MPI_Allreduce(&integralTotal, &recv, 1, MPI_CH_REAL,
MPI_SUM, Chombo_MPI::comm);
if (result != MPI_SUCCESS)
{ //bark!!!
MayDay::Error("sorry, but I had a communication error on integral");
}
integralTotal = recv;
# endif
return integralTotal;
}
// ---------------------------------------------------------
// 27 March 2003:
// This is called by other functions, and should not be called directly.
Real
maxnorm(const BoxLayoutData<NodeFArrayBox>& a_layout,
const Interval& a_interval,
bool a_verbose)
{
Real normTotal = 0.;
// a_p == 0: max norm
for (DataIterator it = a_layout.dataIterator(); it.ok(); ++it)
{
const Box& thisBox(a_layout.box(it())); // CELL-centered
const NodeFArrayBox& thisNfab = a_layout[it()];
Real thisNfabNorm =
maxnorm(thisNfab, thisBox, a_interval.begin(), a_interval.size());
if (a_verbose)
cout << "maxnorm(" << thisBox << ") = " << thisNfabNorm << endl;
normTotal = Max(normTotal, thisNfabNorm);
}
# ifdef CH_MPI
Real recv;
int result = MPI_Allreduce(&normTotal, &recv, 1, MPI_CH_REAL,
MPI_MAX, Chombo_MPI::comm);
if (result != MPI_SUCCESS)
{ //bark!!!
MayDay::Error("sorry, but I had a communication error on maxnorm");
}
normTotal = recv;
# endif
return normTotal;
}
void
LevelFluxRegisterEdge::incrementFine(
FArrayBox& a_fineFlux,
Real a_scale,
const DataIndex& a_fineDataIndex,
const Interval& a_srcInterval,
const Interval& a_dstInterval)
{
CH_assert(isDefined());
CH_assert(!a_fineFlux.box().isEmpty());
CH_assert(a_srcInterval.size() == a_dstInterval.size());
CH_assert(a_srcInterval.begin() >= 0);
CH_assert(a_srcInterval.end() < a_fineFlux.nComp());
CH_assert(a_dstInterval.begin() >= 0);
CH_assert(a_dstInterval.end() < m_nComp);
int edgeDir = -1;
for (int sideDir = 0; sideDir<SpaceDim; sideDir++)
{
if (a_fineFlux.box().type(sideDir) == IndexType::CELL)
{
edgeDir = sideDir;
}
}
CH_assert(edgeDir >= 0);
CH_assert(edgeDir < SpaceDim);
for (int faceDir=0; faceDir<SpaceDim; faceDir++)
{
if (faceDir != edgeDir)
{
SideIterator sit;
for (sit.begin(); sit.ok(); ++sit)
{
incrementFine(a_fineFlux,
a_scale,
a_fineDataIndex,
a_srcInterval,
a_dstInterval,
faceDir,
sit());
}
}
}
}
// ---------------------------------------------------------
// 7 Dec 2005
Real
maxnorm(const BoxLayoutData<NodeFArrayBox>& a_layout,
const LevelData<NodeFArrayBox>& a_mask,
const ProblemDomain& a_domain,
const Interval& a_interval,
bool a_verbose)
{
Real normTotal = 0.;
// a_p == 0: max norm
int ncomp = a_interval.size();
for (DataIterator it = a_layout.dataIterator(); it.ok(); ++it)
{
const NodeFArrayBox& thisNfab = a_layout[it()];
const FArrayBox& dataFab = thisNfab.getFab();
const FArrayBox& maskFab = a_mask[it()].getFab();
const Box& thisBox(a_layout.box(it())); // CELL-centered
NodeFArrayBox dataMasked(thisBox, ncomp);
FArrayBox& dataMaskedFab = dataMasked.getFab();
dataMaskedFab.copy(dataFab);
// dataMaskedFab *= maskFab;
for (int comp = a_interval.begin(); comp <= a_interval.end(); comp++)
{
// Set dataMaskedFab[comp] *= maskFab[0].
dataMaskedFab.mult(maskFab, 0, comp);
}
Real thisNfabNorm =
maxnorm(dataMasked, thisBox, a_interval.begin(), a_interval.size());
if (a_verbose)
cout << "maxnorm(" << thisBox << ") = " << thisNfabNorm << endl;
normTotal = Max(normTotal, thisNfabNorm);
}
# ifdef CH_MPI
Real recv;
// add up (a_p is not 0)
int result = MPI_Allreduce(&normTotal, &recv, 1, MPI_CH_REAL,
MPI_MAX, Chombo_MPI::comm);
if (result != MPI_SUCCESS)
{ //bark!!!
MayDay::Error("sorry, but I had a communication error on norm");
}
normTotal = recv;
# endif
return normTotal;
}
void
LevelFluxRegisterEdge::incrementCoarse(FArrayBox& a_coarseFlux,
Real a_scale,
const DataIndex& a_coarseDataIndex,
const Interval& a_srcInterval,
const Interval& a_dstInterval)
{
CH_assert(isDefined());
CH_assert(!a_coarseFlux.box().isEmpty());
CH_assert(a_srcInterval.size() == a_dstInterval.size());
CH_assert(a_srcInterval.begin() >= 0);
CH_assert(a_srcInterval.end() < a_coarseFlux.nComp());
CH_assert(a_dstInterval.begin() >= 0);
CH_assert(a_dstInterval.end() < m_nComp);
// get edge-centering of coarseFlux
const Box& edgeBox = a_coarseFlux.box();
int edgeDir = -1;
for (int dir=0; dir<SpaceDim; dir++)
{
if (edgeBox.type(dir) == IndexType::CELL)
{
if (edgeDir == -1)
{
edgeDir = dir;
}
else
{
// already found a cell-centered direction (should only be
// one for edge-centering)
MayDay::Error("LevelFluxRegisterEdge::incrementCoarse -- e-field not edge-centered");
}
}
} // end loop over directions
CH_assert(edgeDir != -1);
FArrayBox& thisCrseReg = m_regCoarse[a_coarseDataIndex][edgeDir];
thisCrseReg.plus(a_coarseFlux, -a_scale, a_srcInterval.begin(),
a_dstInterval.begin(), a_srcInterval.size());
}
// ---------------------------------------------------------
// 27 March 2003:
// This is called by other functions, and should not be called directly.
Real
norm(const BoxLayoutData<NodeFArrayBox>& a_layout,
const Real a_dx,
const int a_p,
const Interval& a_interval,
bool a_verbose)
{
if (a_p == 0)
return maxnorm(a_layout, a_interval, a_verbose);
Real normTotal = 0.;
for (DataIterator it = a_layout.dataIterator(); it.ok(); ++it)
{
const NodeFArrayBox& thisNfab = a_layout[it()];
const Box& thisBox(a_layout.box(it())); // CELL-centered
Real thisNfabNorm = norm(thisNfab, a_dx, thisBox, a_p,
a_interval.begin(), a_interval.size());
if (a_verbose)
cout << a_p << "norm(" << thisBox << ") = " << thisNfabNorm << endl;
if (a_p == 1)
{
normTotal += thisNfabNorm;
}
else if (a_p == 2)
{
normTotal += thisNfabNorm * thisNfabNorm;
}
else
{
normTotal += pow(thisNfabNorm, Real(a_p));
}
}
# ifdef CH_MPI
Real recv;
// add up (a_p is not 0)
int result = MPI_Allreduce(&normTotal, &recv, 1, MPI_CH_REAL,
MPI_SUM, Chombo_MPI::comm);
if (result != MPI_SUCCESS)
{ //bark!!!
MayDay::Error("sorry, but I had a communication error on norm");
}
normTotal = recv;
# endif
// now do sqrt, etc
if (a_p == 2)
normTotal = sqrt(normTotal);
else
if ((a_p != 0) && (a_p != 1))
normTotal = pow(normTotal, (Real)1.0/Real(a_p));
return normTotal;
}
void
EBFineToCoarRedist::
redistribute(LevelData<EBCellFAB>& a_coarSolution,
const Interval& a_variables)
{
CH_TIME("EBFineToCoarRedist::redistribute");
Real nrefD = 1.0;
for (int idir = 0; idir < SpaceDim; idir++)
nrefD *= m_refRat;
//copy the buffer to the coarse layout
m_regsFine.copyTo(a_variables, m_regsRefCoar, a_variables);
//redistribute the refined coarse registers to the coarse solution
int ibox = 0;
for (DataIterator dit = m_gridsCoar.dataIterator(); dit.ok(); ++dit)
{
const BaseIVFAB<Real>& regRefCoar = m_regsRefCoar[dit()];
const IntVectSet& ivsRefCoar = m_setsRefCoar[dit()];
const EBISBox& ebisBoxRefCoar = m_ebislRefCoar[dit()];
const EBISBox& ebisBoxCoar = m_ebislCoar[dit()];
const BaseIVFAB<VoFStencil>& stenFAB = m_stenRefCoar[dit()];
EBCellFAB& solFAB = a_coarSolution[dit()];
for (VoFIterator vofit(ivsRefCoar, ebisBoxRefCoar.getEBGraph());
vofit.ok(); ++vofit)
{
const VolIndex& srcVoFFine = vofit();
const VoFStencil& vofsten = stenFAB(srcVoFFine, 0);
for (int isten = 0; isten < vofsten.size(); isten++)
{
const Real& weight = vofsten.weight(isten);
const VolIndex& dstVoFFine = vofsten.vof(isten);
VolIndex dstVoFCoar =
m_ebislRefCoar.coarsen(dstVoFFine,m_refRat, dit());
//by construction...
CH_assert(m_gridsCoar.get(dit()).contains(dstVoFCoar.gridIndex()));
Real dstVolFracFine = ebisBoxRefCoar.volFrac(dstVoFFine);
Real dstVolFracCoar = ebisBoxCoar.volFrac(dstVoFCoar);
Real denom = dstVolFracCoar*nrefD;
for (int ivar = a_variables.begin();
ivar <= a_variables.end(); ivar++)
{
Real dmFine = regRefCoar(srcVoFFine, ivar);
//ucoar+= massfine/volcoar, ie.
//ucoar+= (wcoar*dmCoar*volFracfine/volfraccoar)=massfine/volcoar
Real dUCoar = dmFine*weight*dstVolFracFine/denom;
solFAB(dstVoFCoar, ivar) += dUCoar;
}
}
}
ibox++;
}
}
void
EBCoarToFineRedist::
redistribute(LevelData<EBCellFAB>& a_fineSolution,
const Interval& a_variables)
{
CH_TIME("EBCoarToFineRedist::redistribute");
//copy the buffer to the fine layout
m_regsCoar.copyTo(a_variables, m_regsCedFine, a_variables);
//redistribute the coarsened fine registers to the fine solution
for (DataIterator dit = m_gridsFine.dataIterator(); dit.ok(); ++dit)
{
const BaseIVFAB<Real>& regCoar = m_regsCedFine[dit()];
const IntVectSet& ivsCoar = m_setsCedFine[dit()];
const EBISBox& ebisBoxCoar = m_ebislCedFine[dit()];
const BaseIVFAB<VoFStencil>& stenFAB = m_stenCedFine[dit()];
EBCellFAB& solFAB = a_fineSolution[dit()];
for (VoFIterator vofitCoar(ivsCoar, ebisBoxCoar.getEBGraph());
vofitCoar.ok(); ++vofitCoar)
{
const VolIndex& srcVoFCoar = vofitCoar();
const VoFStencil& vofsten = stenFAB(srcVoFCoar, 0);
for (int isten = 0; isten < vofsten.size(); isten++)
{
const Real& weight = vofsten.weight(isten);
const VolIndex& dstVoFCoar = vofsten.vof(isten);
Vector<VolIndex> vofsFine =
m_ebislCedFine.refine(dstVoFCoar,m_refRat, dit());
for (int ivar = a_variables.begin();
ivar <= a_variables.end(); ivar++)
{
Real dmCoar = regCoar(srcVoFCoar, ivar);
for (int ifine = 0; ifine < vofsFine.size(); ifine++)
{
const VolIndex& dstVoFFine = vofsFine[ifine];
//ufine += (wcoar*dmCoar) (piecewise constant density diff)
Real dUFine = dmCoar*weight;
solFAB(dstVoFFine, ivar) += dUFine;
}
}
}
}
}
}
void
EBCoarseAverage::averageFAB(BaseIVFAB<Real>& a_coar,
const BaseIVFAB<Real>& a_fine,
const DataIndex& a_datInd,
const Interval& a_variables) const
{
CH_assert(isDefined());
//recall that datInd is from the fine layout.
const EBISBox& ebisBoxCoar = m_eblgCoFi.getEBISL()[a_datInd];
const EBISBox& ebisBoxFine = m_eblgFine.getEBISL()[a_datInd];
const IntVectSet& coarIrregIVS = a_coar.getIVS();
const IntVectSet& fineIrregIVS = a_fine.getIVS();
for (VoFIterator vofitCoar(coarIrregIVS, ebisBoxCoar.getEBGraph());
vofitCoar.ok(); ++vofitCoar)
{
const VolIndex& coarVoF = vofitCoar();
Vector<VolIndex> fineVoFs =
m_eblgCoFi.getEBISL().refine(coarVoF, m_refRat, a_datInd);
for (int ivar = a_variables.begin(); ivar <= a_variables.end(); ivar++)
{
int numVoFs = 0;
Real areaTot = 0;
Real dataVal = 0;
for (int ifine = 0; ifine < fineVoFs.size(); ifine++)
{
const VolIndex& fineVoF = fineVoFs[ifine];
if (fineIrregIVS.contains(fineVoF.gridIndex()))
{
Real bndryArea = ebisBoxFine.bndryArea(fineVoF);
if (bndryArea > 0)
{
areaTot += bndryArea;
numVoFs++;
dataVal += a_fine(fineVoF, ivar);
}
}
}
if (numVoFs > 1)
{
dataVal /= Real(numVoFs);
}
a_coar(coarVoF, ivar) = dataVal;
}
}
}
void BlockTab::findIntervals( void )
{
vector< bool > visited( curBlockNum + 1, false );
list< BasicBlock * > left;
left.push_back( blockList[ 1 ] ); //entry node
visited[ blockList[ 1 ] -> no ] = true;
while ( !left.empty( ) ) {
Interval current;
current.Head( *left.begin( ) );
bool added = true;
while ( added == true ) {
added = false;
for ( list< BasicBlock * >::iterator i = left.begin( );
i != left.end( ); ) {
//includes:check if first set includes the second set
if ( current.Head( ) == *i ||
includes( current.begin( ), current.end( ),
( *i ) -> predecessors.begin( ),
(*i) -> predecessors.end())) {
current.insert(*i);
(*i) -> head = current.Head();
if ((*i) -> getTakenPtr() &&
!visited[(*i) -> getTakenPtr() -> no]) {
visited[(*i) -> getTakenPtr() -> no] = true;
left.push_back((*i) -> getTakenPtr());
added = true;
}
if ((*i) -> getNTakenPtr() &&
!visited[(*i)-> getNTakenPtr() -> no]) {
visited[(*i) -> getNTakenPtr() -> no] = true;
left.push_back((*i) -> getNTakenPtr());
added = true;
}
left.erase(i++);
}
else
i++;
}
}
all.push_back(current);
}
}
void TimeInterpolatorRK4::interpolate(/// interpolated solution on this level coarsened
LevelData<FArrayBox>& a_U,
/// time interpolation coefficient in range [0:1]
const Real& a_timeInterpCoeff,
/// interval of a_U to fill in
const Interval& a_intvl)
{
CH_assert(m_defined);
CH_assert(m_gotFullTaylorPoly);
CH_assert(a_U.nComp() == m_numStates);
LevelData<FArrayBox> UComp;
aliasLevelData(UComp, &a_U, a_intvl);
// For i in 0:m_numCoeffs-1,
// coeffFirst[i] is index of first component of m_taylorCoeffs
// that corresponds to a coefficient of t^i.
Vector<int> coeffFirst(m_numCoeffs);
int intervalLength = a_intvl.size();
for (int i = 0; i < m_numCoeffs; i++)
{
coeffFirst[i] = a_intvl.begin() + i * m_numStates;
}
DataIterator dit = UComp.dataIterator();
for (dit.begin(); dit.ok(); ++dit)
{
FArrayBox& UFab = UComp[dit];
const FArrayBox& taylorFab = m_taylorCoeffs[dit];
// Evaluate a0 + a1*t + a2*t^2 + a3*t^3
// as a0 + t * (a1 + t * (a2 + t * a3)).
UFab.copy(taylorFab, coeffFirst[m_numCoeffs-1], 0, intervalLength);
for (int ind = m_numCoeffs - 2; ind >=0; ind--)
{
UFab *= a_timeInterpCoeff;
UFab.plus(taylorFab, coeffFirst[ind], 0, intervalLength);
}
}
// dummy statement in order to get around gdb bug
int dummy_unused = 0; dummy_unused = 0;
}
void
EBMGInterp::pwcInterpFAB(EBCellFAB& a_refCoar,
const Box& a_coarBox,
const EBCellFAB& a_coar,
const DataIndex& a_datInd,
const Interval& a_variables) const
{
CH_TIMERS("EBMGInterp::interp");
CH_TIMER("regular_interp", t1);
CH_TIMER("irregular_interp", t2);
CH_assert(isDefined());
const Box& coarBox = a_coarBox;
for (int ivar = a_variables.begin(); ivar <= a_variables.end(); ivar++)
{
m_interpEBStencil[a_datInd]->cache(a_refCoar, ivar);
//do all cells as if they were regular
Box refBox(IntVect::Zero, IntVect::Zero);
refBox.refine(m_refRat);
const BaseFab<Real>& coarRegFAB = a_coar.getSingleValuedFAB();
BaseFab<Real>& refCoarRegFAB = a_refCoar.getSingleValuedFAB();
CH_START(t1);
FORT_REGPROLONG(CHF_FRA1(refCoarRegFAB,ivar),
CHF_CONST_FRA1(coarRegFAB,ivar),
CHF_BOX(coarBox),
CHF_BOX(refBox),
CHF_CONST_INT(m_refRat));
CH_STOP(t1);
m_interpEBStencil[a_datInd]->uncache(a_refCoar, ivar);
CH_START(t2);
m_interpEBStencil[a_datInd]->apply(a_refCoar, a_coar, true, ivar);
CH_STOP(t2);
}
}
void
EBCoarToFineRedist::increment(const BaseIVFAB<Real>& a_coarseMass,
const DataIndex& a_coarseDataIndex,
const Interval& a_variables)
{
BaseIVFAB<Real>& coarBuf = m_regsCoar[a_coarseDataIndex];
const EBISBox& ebisBox = m_ebislCoar[a_coarseDataIndex];
const IntVectSet& ivsLoc = m_setsCoar[a_coarseDataIndex];
CH_assert(a_coarseMass.getIVS().contains(ivsLoc));
CH_assert(coarBuf.getIVS().contains(ivsLoc));
for (VoFIterator vofit(ivsLoc, ebisBox.getEBGraph()); vofit.ok(); ++vofit)
{
const VolIndex& vof = vofit();
for (int ivar = a_variables.begin();
ivar <= a_variables.end(); ivar++)
{
coarBuf(vof, ivar) += a_coarseMass(vof, ivar);
}
}
}
void
MappedLevelFluxRegister::incrementFine(const FArrayBox& a_fineFlux,
Real a_scale,
const DataIndex& a_fineDataIndex,
const Interval& a_srcInterval,
const Interval& a_dstInterval,
int a_dir,
Side::LoHiSide a_sd)
{
CH_assert(isDefined());
if (!(m_isDefined & FluxRegFineDefined)) return;
CH_assert(a_srcInterval.size() == a_dstInterval.size());
CH_TIME("MappedLevelFluxRegister::incrementFine");
// We cast away the constness in a_coarseFlux for the scope of this function. This
// should be acceptable, since at the end of the day there is no change to it. -JNJ
FArrayBox& fineFlux = const_cast<FArrayBox&>(a_fineFlux); // Muhahaha.
fineFlux.shiftHalf(a_dir, sign(a_sd));
Real denom = 1.0;
if (m_scaleFineFluxes) {
denom = m_nRefine.product() / m_nRefine[a_dir];
}
Real scale = sign(a_sd) * a_scale / denom;
FArrayBox& cFine = m_fineFlux[a_fineDataIndex];
// FArrayBox cFineFortran(cFine.box(), cFine.nComp());
// cFineFortran.copy(cFine);
Box clipBox = m_fineFlux.box(a_fineDataIndex);
clipBox.refine(m_nRefine);
Box fineBox;
if (a_sd == Side::Lo) {
fineBox = adjCellLo(clipBox, a_dir, 1);
fineBox &= fineFlux.box();
} else {
fineBox = adjCellHi(clipBox, a_dir, 1);
fineBox &= fineFlux.box();
}
#if 0
for (BoxIterator b(fineBox); b.ok(); ++b) {
int s = a_srcInterval.begin();
int d = a_dstInterval.begin();
for (; s <= a_srcInterval.end(); ++s, ++d) {
cFine(coarsen(b(), m_nRefine), d) += scale * fineFlux(b(), s);
}
}
#else
// shifting to ensure fineBox is in the positive quadrant, so IntVect coarening
// is just integer division.
const Box& box = coarsen(fineBox, m_nRefine);
Vector<Real> regbefore(cFine.nComp());
Vector<Real> regafter(cFine.nComp());
if (s_verbose && (a_dir == debugdir) && box.contains(ivdebnoeb)) {
for (int ivar = 0; ivar < cFine.nComp(); ivar++) {
regbefore[ivar] = cFine(ivdebnoeb, ivar);
}
}
const IntVect& iv = fineBox.smallEnd();
IntVect civ = coarsen(iv, m_nRefine);
int srcComp = a_srcInterval.begin();
int destComp = a_dstInterval.begin();
int ncomp = a_srcInterval.size();
FORT_MAPPEDINCREMENTFINE(CHF_CONST_FRA_SHIFT(fineFlux, iv),
CHF_FRA_SHIFT(cFine, civ),
CHF_BOX_SHIFT(fineBox, iv),
CHF_CONST_INTVECT(m_nRefine),
CHF_CONST_REAL(scale),
CHF_CONST_INT(srcComp),
CHF_CONST_INT(destComp),
CHF_CONST_INT(ncomp));
if (s_verbose && (a_dir == debugdir) && box.contains(ivdebnoeb)) {
for (int ivar = 0; ivar < cFine.nComp(); ivar++) {
regafter[ivar] = cFine(ivdebnoeb, ivar);
}
}
if (s_verbose && (a_dir == debugdir) && box.contains(ivdebnoeb)) {
pout() << "levelfluxreg::incrementFine: scale = " << scale << endl;
Box refbox(ivdebnoeb, ivdebnoeb);
refbox.refine(m_nRefine);
refbox &= fineBox;
if (!refbox.isEmpty()) {
pout() << "fine fluxes = " << endl;
for (BoxIterator bit(refbox); bit.ok(); ++bit) {
for (int ivar = 0; ivar < cFine.nComp(); ivar++) {
pout() << "iv = " << bit() << "(";
for (int ivar = 0; ivar < cFine.nComp(); ivar++) {
pout() << fineFlux(bit(), ivar);
}
pout() << ")" << endl;
}
}
}
for (int ivar = 0; ivar < cFine.nComp(); ivar++) {
pout() << " reg before = " << regbefore[ivar] << ", ";
pout() << " reg after = " << regafter[ivar] << ", ";
}
pout() << endl;
}
//fineBox.shift(-shift);
// now, check that cFineFortran and cFine are the same
//fineBox.coarsen(m_nRefine);
// for (BoxIterator b(fineBox); b.ok(); ++b)
// {
// if (cFineFortran(b(),0) != cFine(b(),0))
// {
// MayDay::Error("Fortran doesn't match C++");
// }
// }
// need to shift boxes back to where they were on entry.
// cFine.shift(-shift/m_nRefine);
#endif
fineFlux.shiftHalf(a_dir, - sign(a_sd));
}
void
EBCoarsen::coarsenFAB(EBCellFAB& a_coar,
const EBCellFAB& a_fine,
const DataIndex& a_datInd,
const Interval& a_variables)
{
CH_assert(isDefined());
//do all cells as if they were regular
BaseFab<Real>& coarRegFAB = a_coar.getSingleValuedFAB();
const BaseFab<Real>& fineRegFAB = a_fine.getSingleValuedFAB();
//this is how much we need to grow a coarse box to get a fine box that
// only has the fine iv's that are neighbors of the coarse iv
const int fac = 1 - m_refRat/2;//NOTE: fac = 0 for refRat==2...
Box refbox(IntVect::Zero,
(m_refRat-1)*IntVect::Unit);
refbox.grow(fac);
Box coarBox = m_coarsenedFineGrids.get(a_datInd);
Box fineBox = refine(coarBox, m_refRat);
CH_assert(coarRegFAB.box().contains(coarBox));
CH_assert(fineRegFAB.box().contains(fineBox));
for (int ivar = a_variables.begin();
ivar <= a_variables.end(); ivar++)
{
BaseFab<Real> laplFine(fineBox, 1);
laplFine.setVal(0.);
for (int idir = 0; idir < SpaceDim; idir++)
{
Box loBox, hiBox, centerBox;
int hasLo, hasHi;
EBArith::loHiCenter(loBox, hasLo,
hiBox, hasHi,
centerBox, m_domainFine,
fineBox, idir, &((*m_cfivsPtr)[a_datInd]));
FORT_H2LAPL1DADDITIVE(CHF_FRA1(laplFine, 0),
CHF_FRA1(fineRegFAB, ivar),
CHF_CONST_INT(idir),
CHF_BOX(loBox),
CHF_CONST_INT(hasLo),
CHF_BOX(hiBox),
CHF_CONST_INT(hasHi),
CHF_BOX(centerBox));
}
FORT_EBCOARSEN(CHF_FRA1(coarRegFAB,ivar),
CHF_CONST_FRA1(fineRegFAB,ivar),
CHF_CONST_FRA1(laplFine, 0),
CHF_BOX(coarBox),
CHF_CONST_INT(m_refRat),
CHF_BOX(refbox));
}
//overwrite irregular vofs and coarse vofs next to the cfivs if refRat<4,
// for these vofs we coarsen based on
// taylor expansions from fine vofs to the coarse cell center
coarsenIrreg(a_coar,a_fine,a_datInd,a_variables);
}
int UniformGridLayoutData<Dim>::touchesAllocLocal(const OtherDomain &d, OutIter o,
const ConstructTag &ctag) const
{
// If there are no internal guard cells, then this calculation is the
// same as the normal touches calculation.
if (!this->hasInternalGuards_m) return touches(d,o,ctag);
int i, count = 0;
// Make sure we have a valid touching domain.
PAssert(this->initialized());
PAssert(contains(this->domain_m, d));
// Find the starting and ending grid positions, and store as an
// Interval.
Interval<Dim> box = Pooma::NoInit();
for (i = 0; i < Dim; ++i)
{
int a, b;
// This part is unchanged from touches. What we're going to do
// here is simply extend the range in each direction by one block
// and then let the intersection calculation below sort out if
// there is actually an intersection. Otherwise we'd have to do
// comparisons on all blocks up here and then still do the
// intersection on the remaining blocks below. On average, this
// should be slightly faster I think.
if (!this->hasExternalGuards_m)
{
a = (d[i].min() - this->firsti_m[i]) / blocksizes_m[i];
b = (d[i].max() - this->firsti_m[i]) / blocksizes_m[i];
}
else
{
// If we're in the lower guards, this will fall
// through to give zero.
a = b = 0;
int pos = d[i].min();
int last = this->innerdomain_m[i].last();
int del = pos - this->firsti_m[i];
if (del >= 0)
if (pos <= last)
a = del / blocksizes_m[i];
else
a = allDomain_m[i].last();
pos = d[i].max();
del = pos - this->firsti_m[i];
if (del >= 0)
if (pos <= last)
b = del / blocksizes_m[i];
else
b = allDomain_m[i].last();
}
// Now we check that we're not at the ends of the domain for the
// brick of blocks, and extend the region accordingly.
if (a > 0) --a;
if (b < allDomain_m[i].last()) ++b;
box[i] = Interval<1>(a, b);
}
// Figure the type of the domain resulting from the intersection.
typedef typename
IntersectReturnType<Domain_t,OtherDomain>::Type_t OutDomain_t;
OutDomain_t outDomain = Pooma::NoInit();
// Generate the type of the node pushed on the output iterator.
typedef Node<OutDomain_t,Domain_t> OutNode_t;
// Iterate through the Interval grid positions.
typename Interval<Dim>::const_iterator boxiter = box.begin();
while (boxiter != box.end())
{
// Calculate the linear position of the current node.
int indx = (*boxiter)[0].first();
for (i = 1; i < Dim; ++i)
indx += blockstride_m[i] * (*boxiter)[i].first();
// Get that node, intersect the *allocated* domain with the
// requested one, and write the result out to the output iterator.
PAssert(indx >= 0 && indx < this->local_m.size());
outDomain = intersect(d, this->local_m[indx]->allocated());
// We can no longer assert that outDomain is not empty since
// we extended the search box without checking. Thus we now
// have an if tests around the output.
if (!outDomain.empty())
{
*o = touchesConstruct(outDomain,
this->local_m[indx]->allocated(),
this->local_m[indx]->affinity(),
this->local_m[indx]->context(),
this->local_m[indx]->globalID(),
this->local_m[indx]->localID(),
ctag);
}
// Increment output iterator, count, and grid iterator.
++o;
++count;
++boxiter;
}
// Return the number of non-empty domains we found.
return count;
}
void
LevelFluxRegisterEdge::incrementFine(
FArrayBox& a_fineFlux,
Real a_scale,
const DataIndex& a_fineDataIndex,
const Interval& a_srcInterval,
const Interval& a_dstInterval,
int a_dir,
Side::LoHiSide a_sd)
{
CH_assert(isDefined());
CH_assert(!a_fineFlux.box().isEmpty());
CH_assert(a_srcInterval.size() == a_dstInterval.size());
CH_assert(a_srcInterval.begin() >= 0);
CH_assert(a_srcInterval.end() < a_fineFlux.nComp());
CH_assert(a_dstInterval.begin() >= 0);
CH_assert(a_dstInterval.end() < m_nComp);
CH_assert(a_dir >= 0);
CH_assert(a_dir < SpaceDim);
CH_assert((a_sd == Side::Lo)||(a_sd == Side::Hi));
//
//
//denom is the number of fine faces per coarse face
//this is intrinsically dimension-dependent
#if (CH_SPACEDIM == 2)
Real denom = 1;
#elif (CH_SPACEDIM == 3)
Real denom = m_nRefine;
#else
// This code doesn't make any sense in 1D, and hasn't been implemented
// for DIM > 3
Real denom = -1.0;
MayDay::Error("LevelFluxRegisterEdge -- bad SpaceDim");
#endif
Real scale = a_scale/denom;
// need which fluxbox face we're doing this for
Box thisBox = a_fineFlux.box();
int fluxComp = -1;
for (int sideDir=0; sideDir<SpaceDim; sideDir++)
{
// we do nothing in the direction normal to face
if (sideDir != a_dir)
{
if (thisBox.type(sideDir) == IndexType::CELL)
{
fluxComp = sideDir;
}
}
}
CH_assert (fluxComp >= 0);
int regcomp = getRegComp(a_dir, fluxComp);
FluxBox& thisReg = m_fabFine[index(a_dir, a_sd)][a_fineDataIndex];
FArrayBox& reg = thisReg[regcomp];
a_fineFlux.shiftHalf(a_dir, sign(a_sd));
// this is a way of geting a face-centered domain
// box which we can then use to intersect with things
// to screen out cells outside the physical domain
// (nothing is screened out in periodic case)
Box shiftedValidDomain = m_domainCoarse.domainBox();
shiftedValidDomain.grow(2);
shiftedValidDomain &= m_domainCoarse;
shiftedValidDomain.surroundingNodes(regcomp);
BoxIterator regIt(reg.box() & shiftedValidDomain);
for (regIt.begin(); regIt.ok(); ++regIt)
{
const IntVect& coarseIndex = regIt();
// create a cell-centered box, then shift back to face-centered
Box box(coarseIndex, coarseIndex);
box.shiftHalf(regcomp,-1);
// to avoid adding in edges which do not overlie coarse-grid
// edges, will refine only in non-fluxComp directions to
// determine box from which to grab fluxes.
IntVect refineVect(m_nRefine*IntVect::Unit);
//refineVect.setVal(fluxComp,1);
box.refine(refineVect);
if (a_sd == Side::Lo) box.growLo(a_dir,-(m_nRefine-1));
else box.growHi(a_dir,-(m_nRefine-1));
BoxIterator fluxIt(box);
for (fluxIt.begin(); fluxIt.ok(); ++fluxIt)
{
int src = a_srcInterval.begin();
int dest = a_dstInterval.begin();
for ( ; src <=a_srcInterval.end(); ++src,++dest)
reg(coarseIndex, dest) += scale*a_fineFlux(fluxIt(), src);
}
}
a_fineFlux.shiftHalf(a_dir, -sign(a_sd));
}
void
EBFluxRegister::
incrementRedistRegister(EBCoarToCoarRedist& a_register,
const Interval& a_variables,
const Real& a_scale)
{
if (m_hasEBCF)
{
LevelData<BaseIVFAB<Real> >& registerMass = a_register.m_regsCoar;
LayoutData<IntVectSet>& registerSets = a_register.m_setsCoar;
//reflux into an empty LevelData<EBCellFAB>
//Multiply this by (kappa)(1-kappa)
//add result into register mass
EBCellFactory ebcfCoar(m_eblgCoar.getEBISL());
LevelData<EBCellFAB> increment(m_eblgCoar.getDBL(), m_nComp, m_saveCoar.ghostVect(), ebcfCoar);
EBLevelDataOps::clone (increment, m_saveCoar);
EBLevelDataOps::setVal(increment, 0.0);
reflux(increment, a_variables, a_scale, true);
for (DataIterator dit = m_eblgCoar.getDBL().dataIterator(); dit.ok(); ++dit)
{
for (int idir = 0; idir < SpaceDim; idir++)
{
for (SideIterator sit; sit.ok(); ++sit)
{
int iindex = index(idir, sit());
Vector<IntVectSet> setsCoar = (m_setsCoar[iindex])[dit()];
for (int iset = 0; iset < setsCoar.size(); iset++)
{
const IntVectSet& setCoa = registerSets[dit()];
const IntVectSet& setReg = setsCoar[iset];
IntVectSet set = setCoa;
set &= setReg;
const EBISBox& ebisBox =m_eblgCoar.getEBISL()[dit()];
for (VoFIterator vofit(set, ebisBox.getEBGraph()); vofit.ok(); ++vofit)
{
const VolIndex& vof = vofit();
int ibleck = 0;
if ((vof.gridIndex() == ivdebugfr) && EBFastFR::s_verbose)
{
ibleck = 1;
pout() << setprecision(10)
<< setiosflags(ios::showpoint)
<< setiosflags(ios::scientific);
pout() << "incrcotoco:" << endl;
}
for (int icomp = a_variables.begin(); icomp <= a_variables.end(); icomp++)
{
Real extraMass = increment[dit()](vofit(), icomp);
Real oldMass = registerMass[dit()](vofit(), icomp);
Real newMass = oldMass + extraMass;
Real diff = extraMass;
if (ibleck == 1)
{
pout() << "( " << oldMass << ", " << newMass << ", " << diff << ")";
}
registerMass[dit()](vofit(), icomp) += extraMass;
//set increment to zero in case it gets
//hit twice (more than one direction or
//whatever
increment[dit()](vof, icomp) = 0;
} //loop over comps
if (ibleck == 1)
{
ibleck = 0;
pout() << endl;
}
}//loop over vofs in the set
}//loop over sets in this coarse box
} //loop over sides
} //loop over directions
} //dataiterator loop
} //You are using Bonetti's defense against me, uh?
} //I thought it fitting, considering the rocky terrain.
void
EBFluxRegister::
incrementRedistRegister(EBCoarToFineRedist& a_register,
const Interval& a_variables,
const Real& a_scale)
{
if (m_hasEBCF)
{
LevelData<BaseIVFAB<Real> >& registerMass = a_register.m_regsCoar;
LayoutData<IntVectSet>& registerSets = a_register.m_setsCoar;
//reflux into an empty LevelData<EBCellFAB>
//Multiply this by (kappa)(1-kappa)
//add result into register mass
EBCellFactory ebcfCoar(m_eblgCoar.getEBISL());
LevelData<EBCellFAB> increment(m_eblgCoar.getDBL(), m_nComp, m_saveCoar.ghostVect(), ebcfCoar);
EBLevelDataOps::clone(increment, m_saveCoar);
EBLevelDataOps::setVal(increment, 0.0);
reflux(increment, a_variables, a_scale, true);
for (DataIterator dit = m_eblgCoar.getDBL().dataIterator(); dit.ok(); ++dit)
{
for (int idir = 0; idir < SpaceDim; idir++)
{
for (SideIterator sit; sit.ok(); ++sit)
{
int iindex = index(idir, sit());
Vector<IntVectSet> setsCoar = (m_setsCoar[iindex])[dit()];
for (int iset = 0; iset < setsCoar.size(); iset++)
{
const IntVectSet& setCoa = registerSets[dit()];
const IntVectSet& setReg = setsCoar[iset];
IntVectSet set = setCoa;
set &= setReg;
const EBISBox& ebisBox =m_eblgCoar.getEBISL()[dit()];
for (VoFIterator vofit(set, ebisBox.getEBGraph()); vofit.ok(); ++vofit)
{
const VolIndex vof = vofit();
int ibleck = 0;
if ((vof.gridIndex() == ivdebugfr) && EBFastFR::s_verbose)
{
ibleck = 1;
pout() << setprecision(10)
<< setiosflags(ios::showpoint)
<< setiosflags(ios::scientific);
pout() << "incrcotofi:" << endl;
}
for (int icomp = a_variables.begin(); icomp <= a_variables.end(); icomp++)
{
Real extraMass = increment[dit()](vofit(), icomp);
if ((ibleck == 1) && icomp == 0)
{
//incrredist
Real soluOld = registerMass[dit()](vof, icomp);
Real soluNew = soluOld + extraMass;
Real fluxDif = -extraMass;
pout() << "(" << extraMass << ", " << soluOld << ", " << soluNew << ", " << fluxDif << ") " ;
}
registerMass[dit()](vof, icomp) += extraMass;
//set increment to zero in case it gets
//hit twice (more than one direction or
//whatever
increment[dit()](vof, icomp) = 0;
}
if (ibleck == 1)
{
pout() << endl;
ibleck = 0;
}
}
}
}
}
}
}
}
void TimeInterpolatorRK4::intermediate(/// intermediate RK4 solution on this level coarsened
LevelData<FArrayBox>& a_U,
/// time interpolation coefficient in range [0:1]
const Real& a_timeInterpCoeff,
/// which RK4 stage: 0, 1, 2, 3
const int& a_stage,
/// interval of a_U to fill in
const Interval& a_intvl) const
{
CH_assert(m_defined);
CH_assert(m_gotFullTaylorPoly);
CH_assert(a_U.nComp() == m_numStates);
CH_assert(a_stage >= 0);
CH_assert(a_stage < 4);
Real rinv = 1. / Real(m_refineCoarse);
Vector<Real> intermCoeffs(4);
// 0 is coefficient of m_taylorCoeffs[0] = Ucoarse(0)
// 1 is coefficient of m_taylorCoeffs[1] = K1
// 2 is coefficient of m_taylorCoeffs[2] = 1/2 * (-3*K1 + 2*K2 + 2*K3 - K4)
// 3 is coefficient of m_taylorCoeffs[3] = 2/3 * ( K1 - K2 - K3 + K4)
Real diff12Coeffs;
// coefficient of m_diff12 = - K2 + K3
switch (a_stage)
{
case 0:
intermCoeffs[0] = 1.;
intermCoeffs[1] = a_timeInterpCoeff;
intermCoeffs[2] = a_timeInterpCoeff * a_timeInterpCoeff;
intermCoeffs[3] = a_timeInterpCoeff * a_timeInterpCoeff * a_timeInterpCoeff;
diff12Coeffs = 0.;
break;
case 1:
intermCoeffs[0] = 1.;
intermCoeffs[1] = 0.5*rinv + a_timeInterpCoeff;
intermCoeffs[2] = a_timeInterpCoeff * (rinv + a_timeInterpCoeff);
intermCoeffs[3] = a_timeInterpCoeff * a_timeInterpCoeff * (1.5*rinv + a_timeInterpCoeff);
diff12Coeffs = 0.;
break;
case 2:
intermCoeffs[0] = 1.;
intermCoeffs[1] = 0.5*rinv + a_timeInterpCoeff;
intermCoeffs[2] = 0.5*rinv*rinv + a_timeInterpCoeff * (rinv + a_timeInterpCoeff);
intermCoeffs[3] = 0.375*rinv*rinv*rinv + a_timeInterpCoeff * (1.5*rinv*rinv + a_timeInterpCoeff * (1.5*rinv + a_timeInterpCoeff));
diff12Coeffs = -0.25 * rinv * rinv;
break;
case 3:
intermCoeffs[0] = 1.;
intermCoeffs[1] = rinv + a_timeInterpCoeff;
intermCoeffs[2] = rinv*rinv + a_timeInterpCoeff * (2.*rinv + a_timeInterpCoeff);
intermCoeffs[3] = 0.75*rinv*rinv*rinv + a_timeInterpCoeff * (3.*rinv*rinv + a_timeInterpCoeff * (3.*rinv + a_timeInterpCoeff));
diff12Coeffs = 0.5 * rinv * rinv;
break;
default:
MayDay::Error("TimeInterpolatorRK4::intermediate must have a_stage in range 0:3");
}
LevelData<FArrayBox> UComp;
aliasLevelData(UComp, &a_U, a_intvl);
// For i in 0:m_numCoeffs-1,
// coeffFirst[i] is index of first component of m_taylorCoeffs
// that corresponds to a coefficient of t^i.
Vector<int> coeffFirst(m_numCoeffs);
int intervalLength = a_intvl.size();
for (int i = 0; i < m_numCoeffs; i++)
{
coeffFirst[i] = a_intvl.begin() + i * m_numStates;
}
DataIterator dit = UComp.dataIterator();
for (dit.begin(); dit.ok(); ++dit)
{
FArrayBox& UFab = UComp[dit];
const FArrayBox& taylorFab = m_taylorCoeffs[dit];
// WAS: Evaluate a0 + a1*t + a2*t^2 + a3*t^3
// as a0 + t * (a1 + t * (a2 + t * a3)):
// that is, set UFab to
// a3, t*a3 + a2, t*(t*a3 + a2) + a1, t*(t*(t*a3 + a2) + a1) * a0.
// NEW: Evaluate a0*c0 + a1*c1 + a2*c2 + a3*c3,
// where c0, c1, c2, c3 are scalars,
// and c0 = intermCoeffs[0] = 1.
UFab.copy(taylorFab, coeffFirst[0], 0, intervalLength);
for (int ind = 1; ind < 4; ind++)
{
UFab.plus(taylorFab, intermCoeffs[ind],
coeffFirst[ind], 0, intervalLength);
}
const FArrayBox& diff12Fab = m_diff12[dit];
UFab.plus(diff12Fab, diff12Coeffs, 0, 0, intervalLength);
}
// dummy statement in order to get around gdb bug
int dummy_unused = 0; dummy_unused = 0;
}
//----------------------------------------------------------------------------
void
EBNormalizeByVolumeFraction::
operator()(LevelData<EBCellFAB>& a_Q,
const Interval& a_compInterval) const
{
CH_TIME("EBNormalizer::operator()");
// Endpoints of the given interval.
int begin = a_compInterval.begin(), end = a_compInterval.end(),
length = a_compInterval.size();
// Loop over the EBISBoxes within our grid. The EB data structures are
// indexed in the same manner as the non-EB data structures, so we piggy-
// back the former on the latter.
a_Q.exchange();
EBISLayout ebisLayout = m_levelGrid.getEBISL();
DisjointBoxLayout layout = m_levelGrid.getDBL();
for (DataIterator dit = layout.dataIterator(); dit.ok(); ++dit)
{
const EBISBox& box = ebisLayout[dit()];
EBCellFAB& QFAB = a_Q[dit()];
// Go over the irregular cells in this box.
const IntVectSet& irregCells = box.getIrregIVS(layout[dit()]);
// The average has to be computed from the uncorrected data from all
// the neighbors, so we can't apply the corrections in place. For now,
// we stash them in a map.
map<VolIndex, vector<Real> > correctedValues;
for (VoFIterator vit(irregCells, box.getEBGraph()); vit.ok(); ++vit)
{
Real kappajSum = 0.0;
vector<Real> kappajQjSum(length, 0.0);
// Get all of the indices of the VoFs within a monotone path
// radius of 1.
VolIndex vofi = vit();
Vector<VolIndex> vofjs;
EBArith::getAllVoFsInMonotonePath(vofjs, vofi, box, 1);
// Accumulate the contributions from the neighboring cells.
for (unsigned int j = 0; j < vofjs.size(); ++j)
{
VolIndex vofj = vofjs[j];
Real kappaj = box.volFrac(vofj);
for (int icomp = begin; icomp <= end; ++icomp)
{
kappajQjSum[icomp] += QFAB(vofj, icomp);
}
// Add this volume fraction to the sum.
kappajSum += kappaj;
}
if (kappajSum > 0.)
{
// Normalize the quantity and stow it.
vector<Real> correctedValue(length);
// Real kappai = box.volFrac(vofi); //unused dtg
for (int icomp = begin; icomp <= end; ++icomp)
{
// correctedValue[icomp - begin] =
// QFAB(vofi, icomp) + (1.0 - kappai) * kappajQjSum[icomp] / kappajSum;
correctedValue[icomp - begin] = kappajQjSum[icomp] / kappajSum;
}
correctedValues[vofi] = correctedValue;
}
}
// Apply the corrections.
for (map<VolIndex, vector<Real> >::const_iterator
cit = correctedValues.begin(); cit != correctedValues.end(); ++cit)
{
for (int icomp = begin; icomp <= end; ++icomp)
{
QFAB(cit->first, icomp) = cit->second[icomp-begin];
}
}
}
}
void
EBMGInterp::pwlInterpFAB(EBCellFAB& a_refCoar,
const Box& a_coarBox,
const EBCellFAB& a_coar,
const DataIndex& a_datInd,
const Interval& a_variables) const
{
CH_TIMERS("EBMGInterp::pwlinterpfab");
CH_TIMER("regular_interp", t1);
CH_TIMER("irregular_interp", t2);
CH_assert(isDefined());
//first interpolate piecewise constant.
pwcInterpFAB(a_refCoar,
a_coarBox,
a_coar,
a_datInd,
a_variables);
//then add in slope*distance.
const Box& coarBox = a_coarBox;
for (int ivar = a_variables.begin(); ivar <= a_variables.end(); ivar++)
{
//save stuff at irreg cells because fortran result will be garbage
//and this is an incremental process
m_linearEBStencil[a_datInd]->cache(a_refCoar, ivar);
//do all cells as if they were regular
Box refBox(IntVect::Zero, IntVect::Zero);
refBox.refine(m_refRat);
const BaseFab<Real>& coarRegFAB = a_coar.getSingleValuedFAB();
BaseFab<Real>& refCoarRegFAB = a_refCoar.getSingleValuedFAB();
CH_START(t1);
Real dxf = 1.0/m_coarDomain.size(0);
Real dxc = 2.0*dxf;
//do every cell as regular
for (int idir = 0; idir < SpaceDim; idir++)
{
FORT_PROLONGADDSLOPE(CHF_FRA1(refCoarRegFAB,ivar),
CHF_CONST_FRA1(coarRegFAB,ivar),
CHF_BOX(coarBox),
CHF_BOX(refBox),
CHF_INT(idir),
CHF_REAL(dxf),
CHF_REAL(dxc),
CHF_CONST_INT(m_refRat));
}
CH_STOP(t1);
//replace garbage fortran put in with original values (at irregular cells)
m_linearEBStencil[a_datInd]->uncache(a_refCoar, ivar);
CH_START(t2);
//do what fortran should have done at irregular cells.
m_linearEBStencil[a_datInd]->apply(a_refCoar, a_coar, true, ivar);
CH_STOP(t2);
}
}
Real DotProduct(const BoxLayoutData<FArrayBox>& a_dataOne,
const BoxLayoutData<FArrayBox>& a_dataTwo,
const BoxLayout& a_dblIn,
const Interval& a_comps)
{
Vector<Real> rhodot;
rhodot.reserve(10);
const int startcomp = a_comps.begin();
const int endcomp = a_comps.end();
//calculate the single-processor dot product
DataIterator dit = a_dataOne.dataIterator();
for (dit.reset(); dit.ok(); ++dit)
{
Box fabbox = a_dblIn.get(dit());
const FArrayBox& onefab = a_dataOne[dit()];
const FArrayBox& twofab = a_dataTwo[dit()];
CH_assert(onefab.box().contains(fabbox));
CH_assert(twofab.box().contains(fabbox));
Real dotgrid = 0;
FORT_DOTPRODUCT(CHF_REAL(dotgrid),
CHF_CONST_FRA(onefab),
CHF_CONST_FRA(twofab),
CHF_BOX(fabbox),
CHF_CONST_INT(startcomp),
CHF_CONST_INT(endcomp));
rhodot.push_back(dotgrid);
}
// now for the multi-processor fandango
//gather all the rhodots onto a vector and add them up
int baseProc = 0;
Vector<Vector<Real> >dotVec;
gather(dotVec, rhodot, baseProc);
Real rhodotTot = 0.0;
if (procID() == baseProc)
{
CH_assert(dotVec.size() == numProc());
rhodot.resize(a_dblIn.size());
int index = 0;
for (int p = 0; p < dotVec.size(); p++)
{
Vector<Real>& v = dotVec[p];
for (int i = 0; i < v.size(); i++, index++)
{
rhodot[index] = v[i];
}
}
rhodot.sort();
for (int ivec = 0; ivec < rhodot.size(); ivec++)
{
rhodotTot += rhodot[ivec];
}
}
//broadcast the sum to all processors.
broadcast(rhodotTot, baseProc);
//return the total
return rhodotTot;
}
void TraceState(/// state at time t+dt/2 on edges in direction dir
FArrayBox& a_stateHalf,
/// cell-centered state at time t
const FArrayBox& a_state,
/// cell-centered velocity at time t
const FArrayBox& a_cellVel,
/// edge-centered advection velocity at time t+dt/2
const FluxBox& a_advectionVel,
/// cell-centered source
const FArrayBox& a_source,
/// Physical domain
const ProblemDomain& a_dProblem,
/// interior of grid patch
const Box& a_gridBox,
/// timeStep
const Real a_dt,
/// cell-spacing
const Real a_dx,
/// direction in which to perform tracing
const int a_dir,
/// which components to trace
const Interval& a_srcComps,
/// where to put traced components in a_stateHalf,
const Interval& a_destComps)
{
int ncomp = a_srcComps.size();
CH_assert (ncomp == a_destComps.size());
int offset = a_destComps.begin() - a_srcComps.begin();
Box edgeBox = a_gridBox;
edgeBox.surroundingNodes(a_dir);
#ifdef SIMPLEUPWIND
// simple cell-to-edge averaging may be causing us problems --
// instead do simple upwinding
for (int comp=a_srcComps.begin(); comp <= a_srcComps.end(); comp++)
{
int destcomp = comp+offset;
FORT_UPWINDCELLTOEDGE(CHF_FRA1(a_stateHalf, destComp),
CHF_CONST_FRA1(a_state,comp),
CHF_CONST_FRA(a_advectionVel[a_dir]),
CHF_BOX(edgeBox),
CHF_CONST_INT(a_dir));
}
#else
// first compute slopes
Box slopesBox = grow(a_gridBox,1);
// these are debugging changes to sync with old code
#ifdef MATCH_OLDCODE
FluxBox tempEdgeVel(slopesBox,1);
tempEdgeVel.setVal(0.0);
FArrayBox tempCellVel(slopesBox,SpaceDim);
tempCellVel.setVal(0.0);
CellToEdge(a_cellVel, tempEdgeVel);
EdgeToCell(tempEdgeVel, tempCellVel);
tempEdgeVel.clear(); // done with this, so reclaim memory
EdgeToCell(a_advectionVel, tempCellVel);
#endif
FArrayBox delS(slopesBox,1);
FArrayBox sHat(slopesBox,1);
FArrayBox sTilde(a_state.box(),1);
for (int comp=a_srcComps.begin(); comp<=a_srcComps.end(); comp++)
{
int destComp = comp+offset;
#ifdef SET_BOGUS_VALUES
delS.setVal(BOGUS_VALUE);
sHat.setVal(BOGUS_VALUE);
sTilde.setVal(BOGUS_VALUE);
#endif
// compute van leer limited slopes in the normal direction
// for now, always limit slopes, although might want to make
// this a parameter
int limitSlopes = 1;
FORT_SLOPES(CHF_FRA1(delS,0),
CHF_CONST_FRA1(a_state,comp),
CHF_BOX(slopesBox),
CHF_CONST_INT(a_dir),
CHF_CONST_INT(limitSlopes));
// this is a way to incorporate the minion correction --
// stateTilde = state + (dt/2)*source
sTilde.copy(a_source,comp,0,1);
Real sourceFactor = a_dt/2.0;
sTilde *= sourceFactor;
sTilde.plus(a_state, comp,0,1);
sHat.copy(sTilde,slopesBox);
// now loop over directions, adding transverse components
for (int localDir=0; localDir < SpaceDim; localDir++)
{
if (localDir != a_dir)
{
// add simple transverse components
FORT_TRANSVERSE(CHF_FRA1(sHat,0),
CHF_CONST_FRA1(sTilde,0),
#ifdef MATCH_OLDCODE
CHF_CONST_FRA1(tempCellVel, localDir),
#else
CHF_CONST_FRA1(a_cellVel,localDir),
#endif
CHF_BOX(slopesBox),
CHF_CONST_REAL(a_dt),
CHF_CONST_REAL(a_dx),
CHF_CONST_INT(localDir));
//CHF_FRA1(temp,0)); // not used in function
}
else if (SpaceDim == 3)
{
// only add cross derivative transverse piece if we're in 3D
// note that we need both components of velocity for this one
FORT_TRANSVERSECROSS(CHF_FRA1(sHat,0),
CHF_CONST_FRA1(sTilde,0),
CHF_CONST_FRA(a_cellVel),
CHF_BOX(slopesBox),
CHF_CONST_REAL(a_dt),
CHF_CONST_REAL(a_dx),
CHF_CONST_INT(localDir));
}
} // end loop over directions
// now compute left and right states and resolve the Reimann
// problem to get single set of edge-centered values
FORT_PREDICT(CHF_FRA1(a_stateHalf,destComp),
CHF_CONST_FRA1(sHat,0),
CHF_CONST_FRA1(delS,0),
CHF_CONST_FRA1(a_cellVel,a_dir),
CHF_CONST_FRA1(a_advectionVel[a_dir],0),
CHF_BOX(edgeBox),
CHF_CONST_REAL(a_dt),
CHF_CONST_REAL(a_dx),
CHF_CONST_INT(a_dir));
} // end loop over components
#endif
}
int main(int argc, char *argv[])
{
Pooma::initialize(argc, argv);
Pooma::Tester tester(argc, argv);
// To declare a field, you first need to set up a layout. This requires
// knowing the physical vertex-domain and the number of external guard
// cell layers. Vertex domains contain enough points to hold all of the
// rectilinear centerings that POOMA is likely to support for quite
// awhile. Also, it means that the same layout can be used for all
// fields, regardless of centering.
Interval<2> physicalVertexDomain(14, 14);
Loc<2> blocks(3, 3);
GridLayout<2> layout1(physicalVertexDomain, blocks, GuardLayers<2>(1),
LayoutTag_t());
GridLayout<2> layout0(physicalVertexDomain, blocks, GuardLayers<2>(0),
LayoutTag_t());
Centering<2> cell = canonicalCentering<2>(CellType, Continuous, AllDim);
Centering<2> vert = canonicalCentering<2>(VertexType, Continuous, AllDim);
Centering<2> yedge = canonicalCentering<2>(EdgeType, Continuous, YDim);
Vector<2> origin(0.0);
Vector<2> spacings(1.0, 2.0);
// First basic test verifies that we're assigning to the correct areas
// on a brick.
typedef
Field<UniformRectilinearMesh<2>, double,
MultiPatch<GridTag, BrickTag_t> > Field_t;
Field_t b0(cell, layout1, origin, spacings);
Field_t b1(vert, layout1, origin, spacings);
Field_t b2(yedge, layout1, origin, spacings);
Field_t b3(yedge, layout1, origin, spacings);
Field_t bb0(cell, layout0, origin, spacings);
Field_t bb1(vert, layout0, origin, spacings);
Field_t bb2(yedge, layout0, origin, spacings);
b0.all() = 0.0;
b1.all() = 0.0;
b2.all() = 0.0;
b0 = 1.0;
b1 = 1.0;
b2 = 1.0;
bb0.all() = 0.0;
bb1.all() = 0.0;
bb2.all() = 0.0;
bb0 = 1.0;
bb1 = 1.0;
bb2 = 1.0;
// SPMD code follows.
// Note, SPMD code will work with the evaluator if you are careful
// to perform assignment on all the relevant contexts. The patchLocal
// function creates a brick on the local context, so you can just perform
// the assignment on that context.
int i;
for (i = 0; i < b0.numPatchesLocal(); ++i)
{
Patch<Field_t>::Type_t patch = b0.patchLocal(i);
// tester.out() << "context " << Pooma::context() << ": assigning to patch " << i
// << " with domain " << patch.domain() << std::endl;
patch += 1.5;
}
// This is safe to do since b1 and b2 are built with the same layout.
for (i = 0; i < b1.numPatchesLocal(); ++i)
{
b1.patchLocal(i) += 1.5;
b2.patchLocal(i) += 1.5;
}
for (i = 0; i < bb0.numPatchesLocal(); ++i)
{
Patch<Field_t>::Type_t patch = bb0.patchLocal(i);
// tester.out() << "context " << Pooma::context() << ": assigning to patch on bb0 " << i
// << " with domain " << patch.domain() << std::endl;
patch += 1.5;
}
// This is safe to do since bb1 and bb2 are built with the same layout.
for (i = 0; i < bb1.numPatchesLocal(); ++i)
{
bb1.patchLocal(i) += 1.5;
bb2.patchLocal(i) += 1.5;
}
tester.check("cell centered field is 2.5", all(b0 == 2.5));
tester.check("vert centered field is 2.5", all(b1 == 2.5));
tester.check("edge centered field is 2.5", all(b2 == 2.5));
tester.out() << "b0.all():" << std::endl << b0.all() << std::endl;
tester.out() << "b1.all():" << std::endl << b1.all() << std::endl;
tester.out() << "b2.all():" << std::endl << b2.all() << std::endl;
tester.check("didn't write into b0 boundary",
sum(b0.all()) == 2.5 * b0.physicalDomain().size());
tester.check("didn't write into b1 boundary",
sum(b1.all()) == 2.5 * b1.physicalDomain().size());
tester.check("didn't write into b2 boundary",
sum(b2.all()) == 2.5 * b2.physicalDomain().size());
tester.check("cell centered field is 2.5", all(bb0 == 2.5));
tester.check("vert centered field is 2.5", all(bb1 == 2.5));
tester.check("edge centered field is 2.5", all(bb2 == 2.5));
tester.out() << "bb0:" << std::endl << bb0 << std::endl;
tester.out() << "bb1:" << std::endl << bb1 << std::endl;
tester.out() << "bb2:" << std::endl << bb2 << std::endl;
typedef
Field<UniformRectilinearMesh<2>, double,
MultiPatch<GridTag, CompressibleBrickTag_t> > CField_t;
CField_t c0(cell, layout1, origin, spacings);
CField_t c1(vert, layout1, origin, spacings);
CField_t c2(yedge, layout1, origin, spacings);
CField_t cb0(cell, layout0, origin, spacings);
CField_t cb1(vert, layout0, origin, spacings);
CField_t cb2(yedge, layout0, origin, spacings);
c0.all() = 0.0;
c1.all() = 0.0;
c2.all() = 0.0;
c0 = 1.0;
c1 = 1.0;
c2 = 1.0;
cb0.all() = 0.0;
cb1.all() = 0.0;
cb2.all() = 0.0;
cb0 = 1.0;
cb1 = 1.0;
cb2 = 1.0;
// SPMD code follows.
// Note, SPMD code will work with the evaluator if you are careful
// to perform assignment on all the relevant contexts. The patchLocal
// function creates a brick on the local context, so you can just perform
// the assignment on that context.
for (i = 0; i < c0.numPatchesLocal(); ++i)
{
Patch<CField_t>::Type_t patch = c0.patchLocal(i);
tester.out() << "context " << Pooma::context() << ": assigning to patch " << i
<< " with domain " << patch.domain() << std::endl;
patch += 1.5;
}
// This is safe to do since c1 and c2 are built with the same layout.
for (i = 0; i < c1.numPatchesLocal(); ++i)
{
c1.patchLocal(i) += 1.5;
c2.patchLocal(i) += 1.5;
}
for (i = 0; i < cb0.numPatchesLocal(); ++i)
{
Patch<CField_t>::Type_t patch = cb0.patchLocal(i);
tester.out() << "context " << Pooma::context() << ": assigning to patch on cb0 " << i
<< " with domain " << patch.domain() << std::endl;
patch += 1.5;
}
// This is safe to do since cb1 and cb2 are cuilt with the same layout.
for (i = 0; i < cb1.numPatchesLocal(); ++i)
{
cb1.patchLocal(i) += 1.5;
cb2.patchLocal(i) += 1.5;
}
tester.check("cell centered field is 2.5", all(c0 == 2.5));
tester.check("vert centered field is 2.5", all(c1 == 2.5));
tester.check("edge centered field is 2.5", all(c2 == 2.5));
tester.out() << "c0.all():" << std::endl << c0.all() << std::endl;
tester.out() << "c1.all():" << std::endl << c1.all() << std::endl;
tester.out() << "c2.all():" << std::endl << c2.all() << std::endl;
tester.check("didn't write into c0 boundary",
sum(c0.all()) == 2.5 * c0.physicalDomain().size());
tester.check("didn't write into c1 boundary",
sum(c1.all()) == 2.5 * c1.physicalDomain().size());
tester.check("didn't write into c2 boundary",
sum(c2.all()) == 2.5 * c2.physicalDomain().size());
tester.check("cell centered field is 2.5", all(cb0 == 2.5));
tester.check("vert centered field is 2.5", all(cb1 == 2.5));
tester.check("edge centered field is 2.5", all(cb2 == 2.5));
tester.out() << "cb0:" << std::endl << cb0 << std::endl;
tester.out() << "cb1:" << std::endl << cb1 << std::endl;
tester.out() << "cb2:" << std::endl << cb2 << std::endl;
//------------------------------------------------------------------
// Scalar code example:
//
c0 = iota(c0.domain()).comp(0);
c1 = iota(c1.domain()).comp(1);
// Make sure all the data-parallel are done:
Pooma::blockAndEvaluate();
for (i = 0; i < c0.numPatchesLocal(); ++i)
{
Patch<CField_t>::Type_t local0 = c0.patchLocal(i);
Patch<CField_t>::Type_t local1 = c1.patchLocal(i);
Patch<CField_t>::Type_t local2 = c2.patchLocal(i);
Interval<2> domain = local2.domain(); // physical domain of local y-edges
// --------------------------------------------------------------
// I believe the following is probably the most efficient approach
// for sparse computations. For data-parallel computations, the
// evaluator will uncompress the patches and take brick views, which
// provide the most efficient access. If you are only performing
// the computation on a small portion of cells, then the gains would
// be outweighed by the act of copying the compressed value to all the
// cells.
//
// The read function is used on the right hand side, because
// operator() is forced to uncompress the patch just in case you want
// to write to it.
for(Interval<2>::iterator pos = domain.begin(); pos != domain.end(); ++pos)
{
Loc<2> edge = *pos;
Loc<2> rightCell = edge; // cell to right is same cell
Loc<2> leftCell = edge - Loc<2>(1,0);
Loc<2> topVert = edge + Loc<2>(0, 1);
Loc<2> bottomVert = edge;
local2(edge) =
local0.read(rightCell) + local0.read(leftCell) +
local1.read(topVert) + local1.read(bottomVert);
}
// This statement is optional, it tries to compress the patch after
// we're done computing on it. Since I used .read() for the local0 and 1
// they remained in their original state. compress() can be expensive, so
// it may not be worth trying unless space is really important.
compress(local2);
}
tester.out() << "c0" << std::endl << c0 << std::endl;
tester.out() << "c1" << std::endl << c1 << std::endl;
tester.out() << "c2" << std::endl << c2 << std::endl;
//------------------------------------------------------------------
// Interfacing with a c-function:
//
// This example handles the corner cases, where the patches from a
// cell centered field with no guard layers actually contain some
// extra data.
Pooma::blockAndEvaluate();
for (i = 0; i < cb0.numPatchesLocal(); ++i)
{
Patch<CField_t>::Type_t local0 = cb0.patchLocal(i);
Interval<2> physicalDomain = local0.physicalDomain();
double *data;
int size = physicalDomain.size();
if (physicalDomain == local0.totalDomain())
{
uncompress(local0);
data = &local0(physicalDomain.firsts());
nonsense(data, size);
}
else
{
// In this case, the engine has extra storage even though the
// field has the right domain. We copy it to a brick engine,
// call the function and copy it back. No uncompress is required,
// since the assignment will copy the compressed value into the
// brick.
// arrayView is a work-around. Array = Field doesn't work at
// the moment.
Array<2, double, Brick> brick(physicalDomain);
Array<2, double, CompressibleBrick> arrayView(local0.engine());
brick = arrayView(physicalDomain);
Pooma::blockAndEvaluate();
data = &brick(Loc<2>(0));
nonsense(data, size);
arrayView(physicalDomain) = brick;
// Note that we don't need a blockAndEvaluate here, since an iterate has
// been spawned to perform the copy.
}
// If you want to try compress(local0) here, you should do blockAndEvaluate
// first in case the local0 = brick hasn't been executed yet.
}
tester.out() << "cb0.all()" << std::endl << cb0 << std::endl;
b2 = positions(b2).comp(0);
RefCountedBlockPtr<double> block = pack(b2);
// The following functions give you access to the raw data from pack.
// Note that the lifetime of the data is managed by the RefCountedBlockPtr,
// so when "block" goes out of scope, the data goes away. (i.e. Don't write
// a function where you return block.beginPointer().)
double *start = block.beginPointer(); // start of the data
double *end = block.endPointer(); // one past the end
int size = block.size(); // size of the data
tester.out() << Pooma::context() << ":" << block.size() << std::endl;
unpack(b3, block);
tester.out() << "b2" << std::endl << b2 << std::endl;
tester.out() << "b3" << std::endl << b3 << std::endl;
tester.check("pack, unpack", all(b2 == b3));
int ret = tester.results("LocalPatch");
Pooma::finalize();
return ret;
}