static void gen_vsldoi(DisasContext *ctx)
{
TCGv_ptr ra, rb, rd;
TCGv_i32 sh;
if (unlikely(!ctx->altivec_enabled)) {
gen_exception(ctx, POWERPC_EXCP_VPU);
return;
}
ra = gen_avr_ptr(rA(ctx->opcode));
rb = gen_avr_ptr(rB(ctx->opcode));
rd = gen_avr_ptr(rD(ctx->opcode));
sh = tcg_const_i32(VSH(ctx->opcode));
gen_helper_vsldoi (rd, ra, rb, sh);
tcg_temp_free_ptr(ra);
tcg_temp_free_ptr(rb);
tcg_temp_free_ptr(rd);
tcg_temp_free_i32(sh);
}
Foam::coupledSolverPerformance Foam::coupledCgSolver::solve
(
FieldField<Field, scalar>& x,
const FieldField<Field, scalar>& b,
const direction cmpt
) const
{
// Prepare solver performance
coupledSolverPerformance solverPerf(typeName, fieldName());
FieldField<Field, scalar> wA(x.size());
FieldField<Field, scalar> rA(x.size());
forAll (x, rowI)
{
wA.set(rowI, new scalarField(x[rowI].size(), 0));
rA.set(rowI, new scalarField(x[rowI].size(), 0));
}
static void gen_vmrgew(DisasContext *ctx)
{
TCGv_i64 tmp;
int VT, VA, VB;
if (unlikely(!ctx->altivec_enabled)) {
gen_exception(ctx, POWERPC_EXCP_VPU);
return;
}
VT = rD(ctx->opcode);
VA = rA(ctx->opcode);
VB = rB(ctx->opcode);
tmp = tcg_temp_new_i64();
tcg_gen_shri_i64(tmp, cpu_avrh[VB], 32);
tcg_gen_deposit_i64(cpu_avrh[VT], cpu_avrh[VA], tmp, 0, 32);
tcg_gen_shri_i64(tmp, cpu_avrl[VB], 32);
tcg_gen_deposit_i64(cpu_avrl[VT], cpu_avrl[VA], tmp, 0, 32);
tcg_temp_free_i64(tmp);
}
Foam::lduMatrix::solverPerformance Foam::PCG::solve
(
scalarField& psi,
const scalarField& source,
const direction cmpt
) const
{
// --- Setup class containing solver performance data
lduMatrix::solverPerformance solverPerf
(
lduMatrix::preconditioner::getName(controlDict_) + typeName,
fieldName_
);
register label nCells = psi.size();
scalar* __restrict__ psiPtr = psi.begin();
scalarField pA(nCells);
scalar* __restrict__ pAPtr = pA.begin();
scalarField wA(nCells);
scalar* __restrict__ wAPtr = wA.begin();
scalar wArA = matrix_.great_;
scalar wArAold = wArA;
// --- Calculate A.psi
matrix_.Amul(wA, psi, interfaceBouCoeffs_, interfaces_, cmpt);
// --- Calculate initial residual field
scalarField rA(source - wA);
scalar* __restrict__ rAPtr = rA.begin();
// --- Calculate normalisation factor
scalar normFactor = this->normFactor(psi, source, wA, pA);
if (lduMatrix::debug >= 2)
{
Info<< " Normalisation factor = " << normFactor << endl;
}
// --- Calculate normalised residual norm
solverPerf.initialResidual() = gSumMag(rA)/normFactor;
solverPerf.finalResidual() = solverPerf.initialResidual();
// --- Check convergence, solve if not converged
if (!solverPerf.checkConvergence(tolerance_, relTol_))
{
// --- Select and construct the preconditioner
autoPtr<lduMatrix::preconditioner> preconPtr =
lduMatrix::preconditioner::New
(
*this,
controlDict_
);
// --- Solver iteration
do
{
// --- Store previous wArA
wArAold = wArA;
// --- Precondition residual
preconPtr->precondition(wA, rA, cmpt);
// --- Update search directions:
wArA = gSumProd(wA, rA);
if (solverPerf.nIterations() == 0)
{
for (register label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell];
}
}
else
{
scalar beta = wArA/wArAold;
for (register label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell] + beta*pAPtr[cell];
}
}
// --- Update preconditioned residual
matrix_.Amul(wA, pA, interfaceBouCoeffs_, interfaces_, cmpt);
scalar wApA = gSumProd(wA, pA);
// --- Test for singularity
if (solverPerf.checkSingularity(mag(wApA)/normFactor)) break;
// --- Update solution and residual:
scalar alpha = wArA/wApA;
for (register label cell=0; cell<nCells; cell++)
{
psiPtr[cell] += alpha*pAPtr[cell];
rAPtr[cell] -= alpha*wAPtr[cell];
}
solverPerf.finalResidual() = gSumMag(rA)/normFactor;
} while
(
solverPerf.nIterations()++ < maxIter_
&& !(solverPerf.checkConvergence(tolerance_, relTol_))
);
}
return solverPerf;
}
typename Foam::SolverPerformance<Type>
Foam::PCICG<Type, DType, LUType>::solve(Field<Type>& psi) const
{
word preconditionerName(this->controlDict_.lookup("preconditioner"));
// --- Setup class containing solver performance data
SolverPerformance<Type> solverPerf
(
preconditionerName + typeName,
this->fieldName_
);
label nCells = psi.size();
Type* __restrict__ psiPtr = psi.begin();
Field<Type> pA(nCells);
Type* __restrict__ pAPtr = pA.begin();
Field<Type> wA(nCells);
Type* __restrict__ wAPtr = wA.begin();
Type wArA = solverPerf.great_*pTraits<Type>::one;
Type wArAold = wArA;
// --- Calculate A.psi
this->matrix_.Amul(wA, psi);
// --- Calculate initial residual field
Field<Type> rA(this->matrix_.source() - wA);
Type* __restrict__ rAPtr = rA.begin();
// --- Calculate normalisation factor
Type normFactor = this->normFactor(psi, wA, pA);
if (LduMatrix<Type, DType, LUType>::debug >= 2)
{
Info<< " Normalisation factor = " << normFactor << endl;
}
// --- Calculate normalised residual norm
solverPerf.initialResidual() = cmptDivide(gSumCmptMag(rA), normFactor);
solverPerf.finalResidual() = solverPerf.initialResidual();
// --- Check convergence, solve if not converged
if
(
this->minIter_ > 0
|| !solverPerf.checkConvergence(this->tolerance_, this->relTol_)
)
{
// --- Select and construct the preconditioner
autoPtr<typename LduMatrix<Type, DType, LUType>::preconditioner>
preconPtr = LduMatrix<Type, DType, LUType>::preconditioner::New
(
*this,
this->controlDict_
);
// --- Solver iteration
do
{
// --- Store previous wArA
wArAold = wArA;
// --- Precondition residual
preconPtr->precondition(wA, rA);
// --- Update search directions:
wArA = gSumCmptProd(wA, rA);
if (solverPerf.nIterations() == 0)
{
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell];
}
}
else
{
Type beta = cmptDivide
(
wArA,
stabilise(wArAold, solverPerf.vsmall_)
);
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell] + cmptMultiply(beta, pAPtr[cell]);
}
}
// --- Update preconditioned residual
this->matrix_.Amul(wA, pA);
Type wApA = gSumCmptProd(wA, pA);
// --- Test for singularity
if
(
solverPerf.checkSingularity
(
cmptDivide(cmptMag(wApA), normFactor)
)
)
{
break;
}
// --- Update solution and residual:
Type alpha = cmptDivide
(
wArA,
stabilise(wApA, solverPerf.vsmall_)
);
for (label cell=0; cell<nCells; cell++)
{
psiPtr[cell] += cmptMultiply(alpha, pAPtr[cell]);
rAPtr[cell] -= cmptMultiply(alpha, wAPtr[cell]);
}
solverPerf.finalResidual() =
cmptDivide(gSumCmptMag(rA), normFactor);
} while
(
(
solverPerf.nIterations()++ < this->maxIter_
&& !solverPerf.checkConvergence(this->tolerance_, this->relTol_)
)
|| solverPerf.nIterations() < this->minIter_
);
}
return solverPerf;
}
Foam::lduMatrix::solverPerformance Foam::paralution_PFGMRES::solve
(
scalarField& psi,
const scalarField& source,
const direction cmpt
) const
{
word precond_name = lduMatrix::preconditioner::getName(controlDict_);
double div = controlDict_.lookupOrDefault<double>("div", 1e+08);
int basis = controlDict_.lookupOrDefault<int>("BasisSize", 30);
bool accel = controlDict_.lookupOrDefault<bool>("useAccelerator", true);
word mformat = controlDict_.lookupOrDefault<word>("MatrixFormat", "CSR");
word pformat = controlDict_.lookupOrDefault<word>("PrecondFormat", "CSR");
int ILUp = controlDict_.lookupOrDefault<int>("ILUp", 0);
int ILUq = controlDict_.lookupOrDefault<int>("ILUq", 1);
int MEp = controlDict_.lookupOrDefault<int>("MEp", 1);
word LBPre = controlDict_.lookupOrDefault<word>("LastBlockPrecond", "paralution_Jacobi");
lduMatrix::solverPerformance solverPerf(typeName + '(' + precond_name + ')', fieldName_);
register label nCells = psi.size();
scalarField pA(nCells);
scalarField wA(nCells);
// --- Calculate A.psi
matrix_.Amul(wA, psi, interfaceBouCoeffs_, interfaces_, cmpt);
// --- Calculate initial residual field
scalarField rA(source - wA);
// --- Calculate normalisation factor
scalar normFactor = this->normFactor(psi, source, wA, pA);
// --- Calculate normalised residual norm
solverPerf.initialResidual() = gSumMag(rA)/normFactor;
solverPerf.finalResidual() = solverPerf.initialResidual();
// TODO check why we cannot skip 1 iteration when initial residual < relTol_ or why initial residual actually
// does not drop below relTol_
if (!solverPerf.checkConvergence(tolerance_, relTol_)) {
paralution::_matrix_format mf = paralution::CSR;
if (mformat == "CSR") mf = paralution::CSR;
else if (mformat == "DIA") mf = paralution::DIA;
else if (mformat == "HYB") mf = paralution::HYB;
else if (mformat == "ELL") mf = paralution::ELL;
else if (mformat == "MCSR") mf = paralution::MCSR;
else if (mformat == "BCSR") mf = paralution::BCSR;
else if (mformat == "COO") mf = paralution::COO;
else if (mformat == "DENSE") mf = paralution::DENSE;
paralution::init_paralution();
paralution::LocalVector<double> x;
paralution::LocalVector<double> rhs;
paralution::LocalMatrix<double> mat;
paralution::FGMRES<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double> ls;
import_openfoam_matrix(matrix(), &mat);
import_openfoam_vector(source, &rhs);
import_openfoam_vector(psi, &x);
ls.Clear();
if (accel) {
mat.MoveToAccelerator();
rhs.MoveToAccelerator();
x.MoveToAccelerator();
}
paralution::Preconditioner<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double > *precond = NULL;
precond = GetPreconditioner<double>(precond_name, LBPre, pformat, ILUp, ILUq, MEp);
if (precond != NULL) ls.SetPreconditioner(*precond);
ls.SetOperator(mat);
ls.SetBasisSize(basis);
ls.Verbose(0);
ls.Init(tolerance_*normFactor, // abs
relTol_, // rel
div, // div
maxIter_); // max iter
ls.Build();
switch(mf) {
case paralution::DENSE:
mat.ConvertToDENSE();
break;
case paralution::CSR:
mat.ConvertToCSR();
break;
case paralution::MCSR:
mat.ConvertToMCSR();
break;
case paralution::BCSR:
mat.ConvertToBCSR();
break;
case paralution::COO:
mat.ConvertToCOO();
break;
case paralution::DIA:
mat.ConvertToDIA();
break;
case paralution::ELL:
mat.ConvertToELL();
break;
case paralution::HYB:
mat.ConvertToHYB();
break;
}
// mat.info();
ls.Solve(rhs, &x);
export_openfoam_vector(x, &psi);
solverPerf.finalResidual() = ls.GetCurrentResidual() / normFactor; // divide by normFactor, see lduMatrixSolver.C
solverPerf.nIterations() = ls.GetIterationCount();
solverPerf.checkConvergence(tolerance_, relTol_);
ls.Clear();
if (precond != NULL) {
precond->Clear();
delete precond;
}
paralution::stop_paralution();
}
return solverPerf;
}
void Foam::DICSmoother::smooth
(
scalarField& psi,
const scalarField& source,
const direction cmpt,
const label nSweeps
) const
{
//add by Xiaow:begin
Foam::Time::enterSec("DICSmoother");
//add by Xiaow:end
const scalar* const __restrict__ rDPtr = rD_.begin();
const scalar* const __restrict__ upperPtr = matrix_.upper().begin();
const label* const __restrict__ uPtr =
matrix_.lduAddr().upperAddr().begin();
const label* const __restrict__ lPtr =
matrix_.lduAddr().lowerAddr().begin();
// Temporary storage for the residual
scalarField rA(rD_.size());
scalar* __restrict__ rAPtr = rA.begin();
//add by Xiaow:begin
Foam::label interid =Foam::Time::commProfiler_.enterIterSec();
//add by Xiaow:end
for (label sweep=0; sweep<nSweeps; sweep++)
{
matrix_.residual
(
rA,
psi,
source,
interfaceBouCoeffs_,
interfaces_,
cmpt
);
rA *= rD_;
register label nFaces = matrix_.upper().size();
for (register label face=0; face<nFaces; face++)
{
register label u = uPtr[face];
rAPtr[u] -= rDPtr[u]*upperPtr[face]*rAPtr[lPtr[face]];
}
register label nFacesM1 = nFaces - 1;
for (register label face=nFacesM1; face>=0; face--)
{
register label l = lPtr[face];
rAPtr[l] -= rDPtr[l]*upperPtr[face]*rAPtr[uPtr[face]];
}
psi += rA;
//add by Xiaow:begin
Foam::Time::commProfiler_.endSingleIter();
//add by Xiaow:end
}
//add by Xiaow:begin
Foam::Time::leaveSec();
//add by Xiaow:end
}
Foam::solverPerformance Foam::PBiCGStab::solve
(
scalarField& psi,
const scalarField& source,
const direction cmpt
) const
{
// --- Setup class containing solver performance data
solverPerformance solverPerf
(
lduMatrix::preconditioner::getName(controlDict_) + typeName,
fieldName_
);
const label nCells = psi.size();
scalar* __restrict__ psiPtr = psi.begin();
scalarField pA(nCells);
scalar* __restrict__ pAPtr = pA.begin();
scalarField yA(nCells);
scalar* __restrict__ yAPtr = yA.begin();
// --- Calculate A.psi
matrix_.Amul(yA, psi, interfaceBouCoeffs_, interfaces_, cmpt);
// --- Calculate initial residual field
scalarField rA(source - yA);
scalar* __restrict__ rAPtr = rA.begin();
// --- Calculate normalisation factor
const scalar normFactor = this->normFactor(psi, source, yA, pA);
if (lduMatrix::debug >= 2)
{
Info<< " Normalisation factor = " << normFactor << endl;
}
// --- Calculate normalised residual norm
solverPerf.initialResidual() =
gSumMag(rA, matrix().mesh().comm())
/normFactor;
solverPerf.finalResidual() = solverPerf.initialResidual();
// --- Check convergence, solve if not converged
if
(
minIter_ > 0
|| !solverPerf.checkConvergence(tolerance_, relTol_)
)
{
scalarField AyA(nCells);
scalar* __restrict__ AyAPtr = AyA.begin();
scalarField sA(nCells);
scalar* __restrict__ sAPtr = sA.begin();
scalarField zA(nCells);
scalar* __restrict__ zAPtr = zA.begin();
scalarField tA(nCells);
scalar* __restrict__ tAPtr = tA.begin();
// --- Store initial residual
const scalarField rA0(rA);
// --- Initial values not used
scalar rA0rA = 0;
scalar alpha = 0;
scalar omega = 0;
// --- Select and construct the preconditioner
autoPtr<lduMatrix::preconditioner> preconPtr =
lduMatrix::preconditioner::New
(
*this,
controlDict_
);
// --- Solver iteration
do
{
// --- Store previous rA0rA
const scalar rA0rAold = rA0rA;
rA0rA = gSumProd(rA0, rA, matrix().mesh().comm());
// --- Test for singularity
if (solverPerf.checkSingularity(mag(rA0rA)))
{
break;
}
// --- Update pA
if (solverPerf.nIterations() == 0)
{
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = rAPtr[cell];
}
}
else
{
// --- Test for singularity
if (solverPerf.checkSingularity(mag(omega)))
{
break;
}
const scalar beta = (rA0rA/rA0rAold)*(alpha/omega);
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] =
rAPtr[cell] + beta*(pAPtr[cell] - omega*AyAPtr[cell]);
}
}
// --- Precondition pA
preconPtr->precondition(yA, pA, cmpt);
// --- Calculate AyA
matrix_.Amul(AyA, yA, interfaceBouCoeffs_, interfaces_, cmpt);
const scalar rA0AyA = gSumProd(rA0, AyA, matrix().mesh().comm());
alpha = rA0rA/rA0AyA;
// --- Calculate sA
for (label cell=0; cell<nCells; cell++)
{
sAPtr[cell] = rAPtr[cell] - alpha*AyAPtr[cell];
}
// --- Test sA for convergence
solverPerf.finalResidual() =
gSumMag(sA, matrix().mesh().comm())/normFactor;
if (solverPerf.checkConvergence(tolerance_, relTol_))
{
for (label cell=0; cell<nCells; cell++)
{
psiPtr[cell] += alpha*yAPtr[cell];
}
solverPerf.nIterations()++;
return solverPerf;
}
// --- Precondition sA
preconPtr->precondition(zA, sA, cmpt);
// --- Calculate tA
matrix_.Amul(tA, zA, interfaceBouCoeffs_, interfaces_, cmpt);
const scalar tAtA = gSumSqr(tA, matrix().mesh().comm());
// --- Calculate omega from tA and sA
// (cheaper than using zA with preconditioned tA)
omega = gSumProd(tA, sA, matrix().mesh().comm())/tAtA;
// --- Update solution and residual
for (label cell=0; cell<nCells; cell++)
{
psiPtr[cell] += alpha*yAPtr[cell] + omega*zAPtr[cell];
rAPtr[cell] = sAPtr[cell] - omega*tAPtr[cell];
}
solverPerf.finalResidual() =
gSumMag(rA, matrix().mesh().comm())
/normFactor;
} while
(
(
solverPerf.nIterations()++ < maxIter_
&& !solverPerf.checkConvergence(tolerance_, relTol_)
)
|| solverPerf.nIterations() < minIter_
);
}
return solverPerf;
}
Foam::SolverPerformance<Type>
Foam::PBiCCCG<Type, DType, LUType>::solve
(
Field<Type>& psi
) const
{
word preconditionerName(this->controlDict_.lookup("preconditioner"));
// --- Setup class containing solver performance data
SolverPerformance<Type> solverPerf
(
preconditionerName + typeName,
this->fieldName_
);
label nCells = psi.size();
Type* __restrict__ psiPtr = psi.begin();
Field<Type> pA(nCells);
Type* __restrict__ pAPtr = pA.begin();
Field<Type> pT(nCells, Zero);
Type* __restrict__ pTPtr = pT.begin();
Field<Type> wA(nCells);
Type* __restrict__ wAPtr = wA.begin();
Field<Type> wT(nCells);
Type* __restrict__ wTPtr = wT.begin();
scalar wArT = 1e15; //this->matrix_.great_;
scalar wArTold = wArT;
// --- Calculate A.psi and T.psi
this->matrix_.Amul(wA, psi);
this->matrix_.Tmul(wT, psi);
// --- Calculate initial residual and transpose residual fields
Field<Type> rA(this->matrix_.source() - wA);
Field<Type> rT(this->matrix_.source() - wT);
Type* __restrict__ rAPtr = rA.begin();
Type* __restrict__ rTPtr = rT.begin();
// --- Calculate normalisation factor
Type normFactor = this->normFactor(psi, wA, pA);
if (LduMatrix<Type, DType, LUType>::debug >= 2)
{
Info<< " Normalisation factor = " << normFactor << endl;
}
// --- Calculate normalised residual norm
solverPerf.initialResidual() = cmptDivide(gSumCmptMag(rA), normFactor);
solverPerf.finalResidual() = solverPerf.initialResidual();
// --- Check convergence, solve if not converged
if
(
this->minIter_ > 0
|| !solverPerf.checkConvergence(this->tolerance_, this->relTol_)
)
{
// --- Select and construct the preconditioner
autoPtr<typename LduMatrix<Type, DType, LUType>::preconditioner>
preconPtr = LduMatrix<Type, DType, LUType>::preconditioner::New
(
*this,
this->controlDict_
);
// --- Solver iteration
do
{
// --- Store previous wArT
wArTold = wArT;
// --- Precondition residuals
preconPtr->precondition(wA, rA);
preconPtr->preconditionT(wT, rT);
// --- Update search directions:
wArT = gSumProd(wA, rT);
if (solverPerf.nIterations() == 0)
{
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell];
pTPtr[cell] = wTPtr[cell];
}
}
else
{
scalar beta = wArT/wArTold;
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell] + (beta* pAPtr[cell]);
pTPtr[cell] = wTPtr[cell] + (beta* pTPtr[cell]);
}
}
// --- Update preconditioned residuals
this->matrix_.Amul(wA, pA);
this->matrix_.Tmul(wT, pT);
scalar wApT = gSumProd(wA, pT);
// --- Test for singularity
if
(
solverPerf.checkSingularity
(
cmptDivide(pTraits<Type>::one*mag(wApT), normFactor)
)
)
{
break;
}
// --- Update solution and residual:
scalar alpha = wArT/wApT;
for (label cell=0; cell<nCells; cell++)
{
psiPtr[cell] += (alpha* pAPtr[cell]);
rAPtr[cell] -= (alpha* wAPtr[cell]);
rTPtr[cell] -= (alpha* wTPtr[cell]);
}
solverPerf.finalResidual() =
cmptDivide(gSumCmptMag(rA), normFactor);
} while
(
(
solverPerf.nIterations()++ < this->maxIter_
&& !solverPerf.checkConvergence(this->tolerance_, this->relTol_)
)
|| solverPerf.nIterations() < this->minIter_
);
}
return solverPerf;
}
Foam::lduSolverPerformance Foam::PBiCG::solve
(
scalarField& x,
const scalarField& b,
const direction cmpt
) const
{
// --- Setup class containing solver performance data
lduSolverPerformance solverPerf
(
lduMatrix::preconditioner::getName(dict()) + typeName,
fieldName()
);
register label nCells = x.size();
scalar* __restrict__ xPtr = x.begin();
scalarField pA(nCells);
scalar* __restrict__ pAPtr = pA.begin();
scalarField pT(nCells, 0.0);
scalar* __restrict__ pTPtr = pT.begin();
scalarField wA(nCells);
scalar* __restrict__ wAPtr = wA.begin();
scalarField wT(nCells);
scalar* __restrict__ wTPtr = wT.begin();
scalar wArT = matrix_.great_;
scalar wArTold = wArT;
// Calculate A.x and T.x
matrix_.Amul(wA, x, coupleBouCoeffs_, interfaces_, cmpt);
matrix_.Tmul(wT, x, coupleIntCoeffs_, interfaces_, cmpt);
// Calculate initial residual and transpose residual fields
scalarField rA(b - wA);
scalarField rT(b - wT);
scalar* __restrict__ rAPtr = rA.begin();
scalar* __restrict__ rTPtr = rT.begin();
// Calculate normalisation factor
scalar normFactor = this->normFactor(x, b, wA, pA, cmpt);
if (lduMatrix::debug >= 2)
{
Info<< " Normalisation factor = " << normFactor << endl;
}
// Calculate normalised residual norm
solverPerf.initialResidual() = gSumMag(rA)/normFactor;
solverPerf.finalResidual() = solverPerf.initialResidual();
// Check convergence, solve if not converged
if (!stop(solverPerf))
{
// Select and construct the preconditioner
autoPtr<lduPreconditioner> preconPtr;
preconPtr =
lduPreconditioner::New
(
matrix_,
coupleBouCoeffs_,
coupleIntCoeffs_,
interfaces_,
dict()
);
// Solver iteration
do
{
// Store previous wArT
wArTold = wArT;
// Precondition residuals
preconPtr->precondition(wA, rA, cmpt);
preconPtr->preconditionT(wT, rT, cmpt);
// Update search directions:
wArT = gSumProd(wA, rT);
if (solverPerf.nIterations() == 0)
{
for (register label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell];
pTPtr[cell] = wTPtr[cell];
}
}
else
{
scalar beta = wArT/wArTold;
for (register label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell] + beta*pAPtr[cell];
pTPtr[cell] = wTPtr[cell] + beta*pTPtr[cell];
}
}
// Update preconditioned residuals
matrix_.Amul(wA, pA, coupleBouCoeffs_, interfaces_, cmpt);
matrix_.Tmul(wT, pT, coupleIntCoeffs_, interfaces_, cmpt);
scalar wApT = gSumProd(wA, pT);
// Test for singularity
if (solverPerf.checkSingularity(mag(wApT)/normFactor)) break;
// Update solution and residual:
scalar alpha = wArT/wApT;
for (register label cell=0; cell<nCells; cell++)
{
xPtr[cell] += alpha*pAPtr[cell];
rAPtr[cell] -= alpha*wAPtr[cell];
rTPtr[cell] -= alpha*wTPtr[cell];
}
solverPerf.finalResidual() = gSumMag(rA)/normFactor;
solverPerf.nIterations()++;
} while (!stop(solverPerf));
}
return solverPerf;
}
Foam::solverPerformance Foam::PBiCG::solve
(
scalarField& psi,
const scalarField& source,
const direction cmpt
) const
{
// --- Setup class containing solver performance data
solverPerformance solverPerf
(
lduMatrix::preconditioner::getName(controlDict_) + typeName,
fieldName_
);
label nCells = psi.size();
scalar* __restrict__ psiPtr = psi.begin();
scalarField pA(nCells);
scalar* __restrict__ pAPtr = pA.begin();
scalarField pT(nCells, 0.0);
scalar* __restrict__ pTPtr = pT.begin();
scalarField wA(nCells);
scalar* __restrict__ wAPtr = wA.begin();
scalarField wT(nCells);
scalar* __restrict__ wTPtr = wT.begin();
scalar wArT = solverPerf.great_;
scalar wArTold = wArT;
// --- Calculate A.psi and T.psi
matrix_.Amul(wA, psi, interfaceBouCoeffs_, interfaces_, cmpt);
matrix_.Tmul(wT, psi, interfaceIntCoeffs_, interfaces_, cmpt);
// --- Calculate initial residual and transpose residual fields
scalarField rA(source - wA);
scalarField rT(source - wT);
scalar* __restrict__ rAPtr = rA.begin();
scalar* __restrict__ rTPtr = rT.begin();
// --- Calculate normalisation factor
scalar normFactor = this->normFactor(psi, source, wA, pA);
if (lduMatrix::debug >= 2)
{
Info<< " Normalisation factor = " << normFactor << endl;
}
// --- Calculate normalised residual norm
solverPerf.initialResidual() =
gSumMag(rA, matrix().mesh().comm())
/normFactor;
solverPerf.finalResidual() = solverPerf.initialResidual();
// --- Check convergence, solve if not converged
if
(
minIter_ > 0
|| !solverPerf.checkConvergence(tolerance_, relTol_)
)
{
// --- Select and construct the preconditioner
autoPtr<lduMatrix::preconditioner> preconPtr =
lduMatrix::preconditioner::New
(
*this,
controlDict_
);
// --- Solver iteration
do
{
// --- Store previous wArT
wArTold = wArT;
// --- Precondition residuals
preconPtr->precondition(wA, rA, cmpt);
preconPtr->preconditionT(wT, rT, cmpt);
// --- Update search directions:
wArT = gSumProd(wA, rT, matrix().mesh().comm());
if (solverPerf.nIterations() == 0)
{
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell];
pTPtr[cell] = wTPtr[cell];
}
}
else
{
scalar beta = wArT/wArTold;
for (label cell=0; cell<nCells; cell++)
{
pAPtr[cell] = wAPtr[cell] + beta*pAPtr[cell];
pTPtr[cell] = wTPtr[cell] + beta*pTPtr[cell];
}
}
// --- Update preconditioned residuals
matrix_.Amul(wA, pA, interfaceBouCoeffs_, interfaces_, cmpt);
matrix_.Tmul(wT, pT, interfaceIntCoeffs_, interfaces_, cmpt);
scalar wApT = gSumProd(wA, pT, matrix().mesh().comm());
// --- Test for singularity
if (solverPerf.checkSingularity(mag(wApT)/normFactor))
{
break;
}
// --- Update solution and residual:
scalar alpha = wArT/wApT;
for (label cell=0; cell<nCells; cell++)
{
psiPtr[cell] += alpha*pAPtr[cell];
rAPtr[cell] -= alpha*wAPtr[cell];
rTPtr[cell] -= alpha*wTPtr[cell];
}
solverPerf.finalResidual() =
gSumMag(rA, matrix().mesh().comm())
/normFactor;
} while
(
(
solverPerf.nIterations()++ < maxIter_
&& !solverPerf.checkConvergence(tolerance_, relTol_)
)
|| solverPerf.nIterations() < minIter_
);
}
return solverPerf;
}
Foam::solverPerformance Foam::paralution_AMG::solve
(
scalarField& psi,
const scalarField& source,
const direction cmpt
) const
{
word precond_name = lduMatrix::preconditioner::getName(controlDict_);
double div = controlDict_.lookupOrDefault<double>("div", 1e+08);
bool accel = controlDict_.lookupOrDefault<bool>("useAccelerator", true);
word mformat = controlDict_.lookupOrDefault<word>("MatrixFormat", "CSR");
word pformat = controlDict_.lookupOrDefault<word>("PrecondFormat", "CSR");
word sformat = controlDict_.lookupOrDefault<word>("SmootherFormat", "CSR");
word solver_name = controlDict_.lookupOrDefault<word>("CoarseGridSolver", "CG");
word smoother_name = controlDict_.lookupOrDefault<word>("smoother", "paralution_MultiColoredGS");
int MEp = controlDict_.lookupOrDefault<int>("MEp", 1);
word LBPre = controlDict_.lookupOrDefault<word>("LastBlockPrecond", "paralution_Jacobi");
int iterPreSmooth = controlDict_.lookupOrDefault<int>("nPreSweeps", 1);
int iterPostSmooth = controlDict_.lookupOrDefault<int>("nPostSweeps", 2);
double epsCoupling = controlDict_.lookupOrDefault<double>("couplingStrength", 0.01);
int coarsestCells = controlDict_.lookupOrDefault<int>("nCellsInCoarsestLevel", 300);
int ILUp = controlDict_.lookupOrDefault<int>("ILUp", 0);
int ILUq = controlDict_.lookupOrDefault<int>("ILUq", 1);
double relax = controlDict_.lookupOrDefault<double>("Relaxation", 1.0);
double aggrrelax = controlDict_.lookupOrDefault<double>("AggrRelax", 2./3.);
bool scaling = controlDict_.lookupOrDefault<bool>("scaleCorrection", true);
word interp_name = controlDict_.lookupOrDefault<word>("InterpolationType", "SmoothedAggregation");
solverPerformance solverPerf(typeName + '(' + precond_name + ')', fieldName_);
register label nCells = psi.size();
scalarField pA(nCells);
scalarField wA(nCells);
// --- Calculate A.psi
matrix_.Amul(wA, psi, interfaceBouCoeffs_, interfaces_, cmpt);
// --- Calculate initial residual field
scalarField rA(source - wA);
// --- Calculate normalisation factor
scalar normFactor = this->normFactor(psi, source, wA, pA);
// --- Calculate normalised residual norm
solverPerf.initialResidual() = gSumMag(rA)/normFactor;
solverPerf.finalResidual() = solverPerf.initialResidual();
if ( !solverPerf.checkConvergence(tolerance_, relTol_) ) {
paralution::_matrix_format mf = paralution::CSR;
if (mformat == "CSR") mf = paralution::CSR;
else if (mformat == "DIA") mf = paralution::DIA;
else if (mformat == "HYB") mf = paralution::HYB;
else if (mformat == "ELL") mf = paralution::ELL;
else if (mformat == "MCSR") mf = paralution::MCSR;
else if (mformat == "BCSR") mf = paralution::BCSR;
else if (mformat == "COO") mf = paralution::COO;
else if (mformat == "DENSE") mf = paralution::DENSE;
paralution::_interp ip = paralution::SmoothedAggregation;
if (interp_name == "SmoothedAggregation") ip = paralution::SmoothedAggregation;
else if (interp_name == "Aggregation") ip = paralution::Aggregation;
paralution::LocalVector<double> x;
paralution::LocalVector<double> rhs;
paralution::LocalMatrix<double> mat;
paralution::AMG<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double> ls;
paralution::import_openfoam_matrix(matrix(), &mat);
paralution::import_openfoam_vector(source, &rhs);
paralution::import_openfoam_vector(psi, &x);
ls.SetOperator(mat);
// coupling strength
ls.SetCouplingStrength(epsCoupling);
// number of unknowns on coarsest level
ls.SetCoarsestLevel(coarsestCells);
// interpolation type for grid transfer operators
ls.SetInterpolation(ip);
// Relaxation parameter for smoothed interpolation aggregation
ls.SetInterpRelax(aggrrelax);
// Manual smoothers
ls.SetManualSmoothers(true);
// Manual course grid solver
ls.SetManualSolver(true);
// grid transfer scaling
ls.SetScaling(scaling);
// operator format
ls.SetOperatorFormat(mf);
ls.SetSmootherPreIter(iterPreSmooth);
ls.SetSmootherPostIter(iterPostSmooth);
ls.BuildHierarchy();
int levels = ls.GetNumLevels();
// Smoother via preconditioned FixedPoint iteration
paralution::IterativeLinearSolver<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double > **fp = NULL;
fp = new paralution::IterativeLinearSolver<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double >*[levels-1];
paralution::Preconditioner<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double > **sm = NULL;
sm = new paralution::Preconditioner<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double >*[levels-1];
for (int i=0; i<levels-1; ++i) {
fp[i] = paralution::GetIterativeLinearSolver<double>("paralution_FixedPoint", relax);
sm[i] = paralution::GetPreconditioner<double>(smoother_name, LBPre, sformat, ILUp, ILUq, MEp);
fp[i]->SetPreconditioner(*sm[i]);
fp[i]->Verbose(0);
}
// Coarse Grid Solver and its Preconditioner
paralution::IterativeLinearSolver<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double > *cgs = NULL;
cgs = paralution::GetIterativeLinearSolver<double>(solver_name, relax);
cgs->Verbose(0);
paralution::Preconditioner<paralution::LocalMatrix<double>,
paralution::LocalVector<double>,
double > *cgp = NULL;
cgp = paralution::GetPreconditioner<double>(precond_name, LBPre, pformat, ILUp, ILUq, MEp);
if (cgp != NULL) cgs->SetPreconditioner(*cgp);
ls.SetSmoother(fp);
ls.SetSolver(*cgs);
// Switch to L1 norm to be consistent with OpenFOAM solvers
ls.SetResidualNorm(1);
ls.Init(tolerance_*normFactor, // abs
relTol_, // rel
div, // div
maxIter_); // max iter
ls.Build();
if (accel) {
mat.MoveToAccelerator();
rhs.MoveToAccelerator();
x.MoveToAccelerator();
ls.MoveToAccelerator();
}
switch(mf) {
case paralution::DENSE:
mat.ConvertToDENSE();
break;
case paralution::CSR:
mat.ConvertToCSR();
break;
case paralution::MCSR:
mat.ConvertToMCSR();
break;
case paralution::BCSR:
mat.ConvertToBCSR();
break;
case paralution::COO:
mat.ConvertToCOO();
break;
case paralution::DIA:
mat.ConvertToDIA();
break;
case paralution::ELL:
mat.ConvertToELL();
break;
case paralution::HYB:
mat.ConvertToHYB();
break;
}
ls.Verbose(0);
// Solve linear system
ls.Solve(rhs, &x);
paralution::export_openfoam_vector(x, &psi);
solverPerf.finalResidual() = ls.GetCurrentResidual() / normFactor; // divide by normFactor, see lduMatrixSolver.C
solverPerf.nIterations() = ls.GetIterationCount();
solverPerf.checkConvergence(tolerance_, relTol_);
// Clear MultiGrid object
ls.Clear();
// Free all structures
for (int i=0; i<levels-1; ++i) {
delete fp[i];
delete sm[i];
}
cgs->Clear();
if (cgp != NULL)
delete cgp;
delete[] fp;
delete[] sm;
delete cgs;
}
return solverPerf;
}
// Solve linear system using Gaussian elimination
// This method changes the content of the matrix mpA
Vector LinearSystem::Solve()
{
Vector m(mSize); //See description in Appendix A
Vector solution(mSize);
// We introduce references to make the syntax readable
Matrix& rA = *mpA; //NB not reference in published version
Vector& rb = *mpb; //NB not reference in published version
// forward sweep of Gaussian elimination
for (int k=0; k<mSize-1; k++)
{
// see if pivoting is necessary
double max = 0.0;
int row = -1;
for (int i=k; i<mSize; i++)
{
if (fabs(rA(i+1,k+1)) > max)
{
row = i;
max=fabs(rA(i+1,k+1)); //NB bug in published version
}
}
assert(row >= 0); //NB bug in published version
// pivot if necessary
if (row != k)
{
// swap matrix rows k+1 with row+1
for (int i=0; i<mSize; i++)
{
double temp = rA(k+1,i+1);
rA(k+1,i+1) = rA(row+1,i+1);
rA(row+1,i+1) = temp;
}
// swap vector entries k+1 with row+1
double temp = rb(k+1);
rb(k+1) = rb(row+1);
rb(row+1) = temp;
}
// create zeros in lower part of column k
for (int i=k+1; i<mSize; i++)
{
m(i+1) = rA(i+1,k+1)/rA(k+1,k+1);
for (int j=k; j<mSize; j++)
{
rA(i+1,j+1) -= rA(k+1,j+1)*m(i+1);
}
rb(i+1) -= rb(k+1)*m(i+1);
}
}
// back substitution
for (int i=mSize-1; i>-1; i--)
{
solution(i+1) = rb(i+1);
for (int j=i+1; j<mSize; j++)
{
solution(i+1) -= rA(i+1,j+1)*solution(j+1);
}
solution(i+1) /= rA(i+1,i+1);
}
return solution;
}
Foam::SolverPerformance<Type>
Foam::PBiCCCG<Type, DType, LUType>::solve
(
gpuField<Type>& psi
) const
{
word preconditionerName(this->controlDict_.lookup("preconditioner"));
// --- Setup class containing solver performance data
SolverPerformance<Type> solverPerf
(
preconditionerName + typeName,
this->fieldName_
);
register label nCells = psi.size();
gpuField<Type> pA(nCells);
gpuField<Type> pT(nCells, pTraits<Type>::zero);
gpuField<Type> wA(nCells);
gpuField<Type> wT(nCells);
scalar wArT = 1e15; //this->matrix_.great_;
scalar wArTold = wArT;
// --- Calculate A.psi and T.psi
this->matrix_.Amul(wA, psi);
this->matrix_.Tmul(wT, psi);
// --- Calculate initial residual and transpose residual fields
gpuField<Type> rA(this->matrix_.source() - wA);
gpuField<Type> rT(this->matrix_.source() - wT);
// --- Calculate normalisation factor
Type normFactor = this->normFactor(psi, wA, pA);
if (LduMatrix<Type, DType, LUType>::debug >= 2)
{
Info<< " Normalisation factor = " << normFactor << endl;
}
// --- Calculate normalised residual norm
solverPerf.initialResidual() = cmptDivide(gSumCmptMag(rA), normFactor);
solverPerf.finalResidual() = solverPerf.initialResidual();
// --- Check convergence, solve if not converged
if
(
this->minIter_ > 0
|| !solverPerf.checkConvergence(this->tolerance_, this->relTol_)
)
{
// --- Select and construct the preconditioner
autoPtr<typename LduMatrix<Type, DType, LUType>::preconditioner>
preconPtr = LduMatrix<Type, DType, LUType>::preconditioner::New
(
*this,
this->controlDict_
);
// --- Solver iteration
do
{
// --- Store previous wArT
wArTold = wArT;
// --- Precondition residuals
preconPtr->precondition(wA, rA);
preconPtr->preconditionT(wT, rT);
// --- Update search directions:
wArT = gSumProd(wA, rT);
if (solverPerf.nIterations() == 0)
{
thrust::copy(wA.begin(),wA.end(),pA.begin());
thrust::copy(wT.begin(),wT.end(),pT.begin());
}
else
{
scalar beta = wArT/wArTold;
thrust::transform
(
wA.begin(),
wA.end(),
thrust::make_transform_iterator
(
pA.begin(),
multiplyOperatorSFFunctor<scalar,Type,Type>(beta)
),
pA.begin(),
addOperatorFunctor<Type,Type,Type>()
);
thrust::transform
(
wT.begin(),
wT.end(),
thrust::make_transform_iterator
(
pT.begin(),
multiplyOperatorSFFunctor<scalar,Type,Type>(beta)
),
pT.begin(),
addOperatorFunctor<Type,Type,Type>()
);
}
// --- Update preconditioned residuals
this->matrix_.Amul(wA, pA);
this->matrix_.Tmul(wT, pT);
scalar wApT = gSumProd(wA, pT);
// --- Test for singularity
if
(
solverPerf.checkSingularity
(
cmptDivide(pTraits<Type>::one*mag(wApT), normFactor)
)
)
{
break;
}
// --- Update solution and residual:
scalar alpha = wArT/wApT;
thrust::transform
(
psi.begin(),
psi.end(),
thrust::make_transform_iterator
(
pA.begin(),
multiplyOperatorSFFunctor<scalar,Type,Type>(alpha)
),
psi.begin(),
addOperatorFunctor<Type,Type,Type>()
);
thrust::transform
(
rA.begin(),
rA.end(),
thrust::make_transform_iterator
(
wA.begin(),
multiplyOperatorSFFunctor<scalar,Type,Type>(alpha)
),
rA.begin(),
subtractOperatorFunctor<Type,Type,Type>()
);
thrust::transform
(
rT.begin(),
rT.end(),
thrust::make_transform_iterator
(
wT.begin(),
multiplyOperatorSFFunctor<scalar,Type,Type>(alpha)
),
rT.begin(),
subtractOperatorFunctor<Type,Type,Type>()
);
solverPerf.finalResidual() =
cmptDivide(gSumCmptMag(rA), normFactor);
} while
(
(
solverPerf.nIterations()++ < this->maxIter_
&& !solverPerf.checkConvergence(this->tolerance_, this->relTol_)
)
|| solverPerf.nIterations() < this->minIter_
);
}
return solverPerf;
}
