#include <PreconditionerBlock.h>

Inheritance diagram for Nektar::MultiRegions::PreconditionerBlock:

Public Member Functions
	PreconditionerBlock (const std::shared_ptr< GlobalLinSys > &plinsys, const AssemblyMapSharedPtr &pLocToGloMap)

	~PreconditionerBlock () override

Public Member Functions inherited from Nektar::MultiRegions::Preconditioner
	Preconditioner (const std::shared_ptr< GlobalLinSys > &plinsys, const AssemblyMapSharedPtr &pLocToGloMap)

virtual	~Preconditioner ()

void	DoPreconditioner (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput, const bool &IsLocal=false)

void	DoAssembleLoc (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput, const bool &ZeroDir)
	Apply an assembly and scatter back to lcoal array. More...

void	DoPreconditionerWithNonVertOutput (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput, const Array< OneD, NekDouble > &pNonVertOutput, Array< OneD, NekDouble > &pVertForce=NullNekDouble1DArray)

void	DoTransformBasisToLowEnergy (Array< OneD, NekDouble > &pInOut)

void	DoTransformCoeffsFromLowEnergy (Array< OneD, NekDouble > &pInOut)

void	DoTransformCoeffsToLowEnergy (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput)

void	DoTransformBasisFromLowEnergy (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput)

void	BuildPreconditioner ()

void	InitObject ()

Array< OneD, NekDouble >	AssembleStaticCondGlobalDiagonals ()
	Performs global assembly of diagonal entries to global Schur complement matrix. More...

const DNekScalBlkMatSharedPtr &	GetBlockTransformedSchurCompl () const

const DNekScalBlkMatSharedPtr &	GetBlockCMatrix () const

const DNekScalBlkMatSharedPtr &	GetBlockInvDMatrix () const

const DNekScalBlkMatSharedPtr &	GetBlockSchurCompl () const

const DNekScalBlkMatSharedPtr &	GetBlockTransformationMatrix () const

const DNekScalBlkMatSharedPtr &	GetBlockTransposedTransformationMatrix () const

DNekScalMatSharedPtr	TransformedSchurCompl (int offset, int bndoffset, const std::shared_ptr< DNekScalMat > &loc_mat)

Static Public Member Functions
static PreconditionerSharedPtr	create (const std::shared_ptr< GlobalLinSys > &plinsys, const std::shared_ptr< AssemblyMap > &pLocToGloMap)
	Creates an instance of this class. More...

Static Public Attributes
static std::string	className
	Name of class. More...

Protected Member Functions
void	v_InitObject () override

void	v_DoPreconditioner (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput, const bool &isLocal=false) override
	Apply preconditioner to `pInput` and store the result in `pOutput`. More...

void	v_BuildPreconditioner () override

Protected Member Functions inherited from Nektar::MultiRegions::Preconditioner
virtual DNekScalMatSharedPtr	v_TransformedSchurCompl (int offset, int bndoffset, const std::shared_ptr< DNekScalMat > &loc_mat)
	Get block elemental transposed transformation matrix \(\mathbf{R}^{T}\). More...

virtual void	v_InitObject ()

virtual void	v_DoPreconditioner (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput, const bool &isLocal=false)
	Apply a preconditioner to the conjugate gradient method. More...

virtual void	v_DoPreconditionerWithNonVertOutput (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput, const Array< OneD, NekDouble > &pNonVertOutput, Array< OneD, NekDouble > &pVertForce)
	Apply a preconditioner to the conjugate gradient method with an output for non-vertex degrees of freedom. More...

virtual void	v_DoTransformBasisToLowEnergy (Array< OneD, NekDouble > &pInOut)
	Transform from original basis to low energy basis. More...

virtual void	v_DoTransformCoeffsFromLowEnergy (Array< OneD, NekDouble > &pInOut)
	Transform from low energy coeffs to orignal basis. More...

virtual void	v_DoTransformCoeffsToLowEnergy (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput)
	Multiply by the block inverse transformation matrix. More...

virtual void	v_DoTransformBasisFromLowEnergy (const Array< OneD, NekDouble > &pInput, Array< OneD, NekDouble > &pOutput)
	Multiply by the block transposed inverse transformation matrix. More...

virtual void	v_BuildPreconditioner ()

Protected Attributes
bool	m_isFull

DNekBlkMatSharedPtr	m_blkMat

Gs::gs_data *	m_gs_mapping

Protected Attributes inherited from Nektar::MultiRegions::Preconditioner
const std::weak_ptr< GlobalLinSys >	m_linsys

std::string	m_preconType

DNekMatSharedPtr	m_preconditioner

std::weak_ptr< AssemblyMap >	m_locToGloMap

Private Member Functions
void	BlockPreconditionerCG (void)
	Construct a block preconditioner from \(\mathbf{S}_{1}\) for the continuous Galerkin system. More...

void	BlockPreconditionerHDG (void)
	Construct a block preconditioner for the hybridized discontinuous Galerkin system. More...

Detailed Description

Definition at line 48 of file PreconditionerBlock.h.

Constructor & Destructor Documentation

◆ PreconditionerBlock()

Nektar::MultiRegions::PreconditionerBlock::PreconditionerBlock	(	const std::shared_ptr< GlobalLinSys > &	plinsys,
		const AssemblyMapSharedPtr &	pLocToGloMap
	)

Definition at line 61 of file PreconditionerBlock.cpp.

    : Preconditioner(plinsys, pLocToGloMap)
{
}

◆ ~PreconditionerBlock()

Nektar::MultiRegions::PreconditionerBlock::~PreconditionerBlock ( )

inlineoverride

Definition at line 71 of file PreconditionerBlock.h.

    {
        Gs::Free(m_gs_mapping);
    }

References Gs::Free(), and m_gs_mapping.

Member Function Documentation

◆ BlockPreconditionerCG()

void Nektar::MultiRegions::PreconditionerBlock::BlockPreconditionerCG ( void )

private

Construct a block preconditioner from \(\mathbf{S}_{1}\) for the continuous Galerkin system.

The preconditioner for the statically-condensed system is defined as:

\[\mathbf{M}^{-1}=\left[\begin{array}{ccc} \mathrm{Diag}[(\mathbf{S_{1}})_{vv}] & & \\ & (\mathbf{S}_{1})_{eb} & \\ & & (\mathbf{S}_{1})_{fb} \end{array}\right] \]

where \(\mathbf{S}_{1}\) is the local Schur complement matrix for each element and the subscript denotes the portion of the Schur complement associated with the vertex (vv), edge (eb) and face blocks (fb) respectively.

In case of the full system \(\mathbf{A}\), the preconditioner includes the interior blocks as well. The full block preconditioner is defined as:

\[\mathbf{M}^{-1}=\left[\begin{array}{cccc} \mathrm{Diag}[(\mathbf{A})_{vv}] & & & \\ & (\mathbf{A})_{eb} & & \\ & & (\mathbf{A})_{fb} & \\ & & & (\mathbf{A})_{int} \end{array}\right] \]

where the interior portion is added at the end following the ordering inside data structures.

Process boundary matrices

Definition at line 123 of file PreconditionerBlock.cpp.

{
    ExpListSharedPtr expList = m_linsys.lock()->GetLocMat().lock();
    LocalRegions::ExpansionSharedPtr exp;
    DNekScalBlkMatSharedPtr loc_mat;
    DNekScalMatSharedPtr bnd_mat;
 
    int i, j, k, n, cnt, gId;
    int meshVertId, meshEdgeId, meshFaceId;
 
    auto asmMap = m_locToGloMap.lock();
 
    const int nExp    = expList->GetExpSize();
    const int nDirBnd = asmMap->GetNumGlobalDirBndCoeffs();
 
    // Grab periodic geometry information.
    PeriodicMap periodicVerts, periodicEdges, periodicFaces;
    expList->GetPeriodicEntities(periodicVerts, periodicEdges, periodicFaces);
 
    // The vectors below are of size 3(+1) to have separate storage for
    // vertices, edges and faces.
    // For the full matrix approach additional storage for the
    // interior matrices is required
    int nStorage = m_isFull ? 4 : 3;
 
    // Maps from geometry ID to the matrix representing the extracted
    // portion of S_1. For example idMats[2] folds the S_1 face blocks.
    vector<map<int, vector<NekDouble>>> idMats(nStorage);
 
    // Maps from the global ID, as obtained from AssemblyMapCG's
    // localToGlobalMap, to the geometry ID.
    vector<map<int, int>> gidMeshIds(3);
 
    // Maps from the global ID to the number of degrees of freedom for
    // this geometry object.
    vector<map<int, int>> gidDofs(nStorage);
 
    // Array containing maximum information needed for the universal
    // numbering later. For i = 0,1,2 for each geometry dimension:
    //
    // maxVertIds[2*i]   = maximum geometry ID at dimension i
    // maxVertIds[2*i+1] = maximum number of degrees of freedom for all
    //                     elements of dimension i.
    Array<OneD, int> maxVertIds(6, -1);
 
    // Figure out mapping from each elemental contribution to offset in
    // (vert,edge,face) triples.
    for (cnt = n = 0; n < nExp; ++n)
    {
        exp = expList->GetExp(n);
 
        // Grab reference to
        // StaticCond: Local Schur complement matrix
        // Full:       Boundary and interior matrix
        DNekScalMatSharedPtr schurMat;
 
        if (m_isFull)
        {
            // Get local matrix
            auto tmpMat = m_linsys.lock()->GetBlock(n);
 
            // Get number of boundary dofs
            int nBndDofs = exp->NumBndryCoeffs();
            int nIntDofs = exp->GetNcoeffs() - nBndDofs;
 
            /// Process boundary matrices
            // Get boundary map
            Array<OneD, unsigned int> bndMap(nBndDofs);
            exp->GetBoundaryMap(bndMap);
 
            // Create temporary matrices
            DNekMatSharedPtr bndryMat =
                MemoryManager<DNekMat>::AllocateSharedPtr(nBndDofs, nBndDofs);
 
            // Extract boundary and send to StaticCond (Schur)
            // framework
            for (int i = 0; i < nBndDofs; ++i)
            {
                for (int j = 0; j < nBndDofs; ++j)
                {
                    (*bndryMat)(i, j) = (*tmpMat)(bndMap[i], bndMap[j]);
                }
            }
 
            schurMat =
                MemoryManager<DNekScalMat>::AllocateSharedPtr(1.0, bndryMat);
 
            // Extract interior matrix and save for block matrix setting
            if (nIntDofs)
            {
                Array<OneD, unsigned int> intMap(nIntDofs);
                exp->GetInteriorMap(intMap);
 
                vector<NekDouble> tmpStore(nIntDofs * nIntDofs);
                for (int i = 0; i < nIntDofs; ++i)
                {
                    for (int j = 0; j < nIntDofs; ++j)
                    {
                        tmpStore[j + i * nIntDofs] =
                            (*tmpMat)(intMap[i], intMap[j]);
                    }
                }
 
                // Save interior matrices and nDofs
                idMats[3][n]  = tmpStore;
                gidDofs[3][n] = nIntDofs;
                // then intMat gets added to a block diagonal matrix that
                // handles the extra interior dofs no need to do any
                // communication for that because all interior dofs are
                // local to each element
            }
        }
        else
        {
            schurMat = m_linsys.lock()->GetStaticCondBlock(n)->GetBlock(0, 0);
        }
 
        // Process vertices to extract relevant portion of the Schur
        // complement matrix.
        for (i = 0; i < exp->GetNverts(); ++i)
        {
            meshVertId = exp->GetGeom()->GetVid(i);
            int locId  = exp->GetVertexMap(i);
 
            // Get the global ID of this vertex.
            gId = asmMap->GetLocalToGlobalMap(cnt + locId) - nDirBnd;
 
            // Ignore all Dirichlet vertices.
            if (gId < 0)
            {
                continue;
            }
 
            gidDofs[0][gId] = 1;
 
            // Extract vertex value from Schur complement matrix.
            NekDouble vertVal = (*schurMat)(locId, locId);
 
            // See if we have processed this vertex from another
            // element.
            auto gIt = idMats[0].find(gId);
 
            if (gIt == idMats[0].end())
            {
                // If not then put our 'matrix' inside idMats.
                idMats[0][gId] = vector<NekDouble>(1, vertVal);
            }
            else
            {
                // Otherwise combine with the value that is already
                // there (i.e. do assembly on this degree of freedom).
                gIt->second[0] += vertVal;
            }
 
            // Now check to see if the vertex is periodic. If it is,
            // then we change meshVertId to be the minimum of all the
            // other periodic vertices, so that we don't end up
            // duplicating the matrix in our final block matrix.
            auto pIt = periodicVerts.find(meshVertId);
            if (pIt != periodicVerts.end())
            {
                for (j = 0; j < pIt->second.size(); ++j)
                {
                    meshVertId = min(meshVertId, pIt->second[j].id);
                }
            }
 
            // Finally record the other information we need into the
            // other matrices.
            gidMeshIds[0][gId] = meshVertId;
            maxVertIds[0]      = max(maxVertIds[0], meshVertId);
            maxVertIds[1]      = 1;
        }
 
        // Process edges. This logic is mostly the same as the previous
        // block.
        for (i = 0; i < exp->GetGeom()->GetNumEdges(); ++i)
        {
            meshEdgeId = exp->GetGeom()->GetEid(i);
 
            Array<OneD, unsigned int> bmap, bmap2;
            Array<OneD, int> sign;
            StdRegions::Orientation edgeOrient = exp->GetGeom()->GetEorient(i);
 
            // Check if this edge is periodic. We may need to flip
            // orientation if it is.
            auto pIt = periodicEdges.find(meshEdgeId);
            if (pIt != periodicEdges.end())
            {
                pair<int, StdRegions::Orientation> idOrient =
                    DeterminePeriodicEdgeOrientId(meshEdgeId, edgeOrient,
                                                  pIt->second);
                meshEdgeId = idOrient.first;
                edgeOrient = idOrient.second;
            }
 
            // Grab edge interior map, and the edge inverse boundary
            // map, so that we can extract this edge from the Schur
            // complement matrix.
            if (exp->GetGeom()->GetNumFaces()) // 3D Element calls
            {
                exp->as<LocalRegions::Expansion3D>()
                    ->GetEdgeInteriorToElementMap(i, bmap, sign, edgeOrient);
                bmap2 = exp->as<LocalRegions::Expansion3D>()
                            ->GetEdgeInverseBoundaryMap(i);
            }
            else
            {
                exp->GetTraceInteriorToElementMap(i, bmap, sign, edgeOrient);
                bmap2 = exp->as<LocalRegions::Expansion2D>()
                            ->GetTraceInverseBoundaryMap(i);
            }
 
            // Check that edge has edge-interior dofs
            const int nEdgeCoeffs = bmap.size();
            if (nEdgeCoeffs == 0)
            {
                continue;
            }
 
            // Allocate temporary storage for the extracted edge matrix.
            vector<NekDouble> tmpStore(nEdgeCoeffs * nEdgeCoeffs);
 
            gId = asmMap->GetLocalToGlobalMap(cnt + bmap[0]);
 
            for (j = 0; j < nEdgeCoeffs; ++j)
            {
                // We record the minimum ID from the edge for our
                // maps. This follows the logic that the assembly map
                // ordering will always give us a contiguous ordering of
                // global degrees of freedom for edge interior
                // coefficients.
                gId = min(gId,
                          asmMap->GetLocalToGlobalMap(cnt + bmap[j]) - nDirBnd);
 
                // Ignore Dirichlet edges.
                if (gId < 0)
                {
                    continue;
                }
 
                const NekDouble sign1 = sign[j];
 
                // Extract this edge, along with sign array for assembly
                // later.
                for (k = 0; k < nEdgeCoeffs; ++k)
                {
                    tmpStore[k + j * nEdgeCoeffs] =
                        sign1 * sign[k] * (*schurMat)(bmap2[j], bmap2[k]);
                }
            }
 
            if (gId < 0)
            {
                continue;
            }
 
            gidDofs[1][gId] = nEdgeCoeffs;
 
            // Assemble this edge matrix with another one, if it exists.
            auto gIt = idMats[1].find(gId);
            if (gIt == idMats[1].end())
            {
                idMats[1][gId] = tmpStore;
            }
            else
            {
                ASSERTL1(tmpStore.size() == gIt->second.size(),
                         "Number of modes mismatch");
                Vmath::Vadd(nEdgeCoeffs * nEdgeCoeffs, &gIt->second[0], 1,
                            &tmpStore[0], 1, &gIt->second[0], 1);
            }
 
            gidMeshIds[1][gId] = meshEdgeId;
            maxVertIds[2]      = max(maxVertIds[2], meshEdgeId);
            maxVertIds[3]      = max(maxVertIds[3], nEdgeCoeffs);
        }
 
        // Process faces. This logic is mostly the same as the previous
        // block.
        for (i = 0; i < exp->GetGeom()->GetNumFaces(); ++i)
        {
            meshFaceId = exp->GetGeom()->GetFid(i);
 
            Array<OneD, unsigned int> bmap, bmap2;
            Array<OneD, int> sign;
            StdRegions::Orientation faceOrient = exp->GetGeom()->GetForient(i);
 
            // Check if this face is periodic. We may need to flip
            // orientation if it is.
            auto pIt = periodicFaces.find(meshFaceId);
            if (pIt != periodicFaces.end())
            {
                meshFaceId = min(meshFaceId, pIt->second[0].id);
                faceOrient = DeterminePeriodicFaceOrient(faceOrient,
                                                         pIt->second[0].orient);
            }
 
            // Trace Interior to Element Map
            exp->GetTraceInteriorToElementMap(i, bmap, sign, faceOrient);
            bmap2 = exp->as<LocalRegions::Expansion3D>()
                        ->GetTraceInverseBoundaryMap(i);
 
            // Check that face has face-interior dofs
            const int nFaceCoeffs = bmap.size();
            if (nFaceCoeffs == 0)
            {
                continue;
            }
 
            // Allocate temporary storage for the extracted face matrix.
            vector<NekDouble> tmpStore(nFaceCoeffs * nFaceCoeffs);
 
            gId = asmMap->GetLocalToGlobalMap(cnt + bmap[0]);
 
            for (j = 0; j < nFaceCoeffs; ++j)
            {
                gId = min(gId,
                          asmMap->GetLocalToGlobalMap(cnt + bmap[j]) - nDirBnd);
 
                // Ignore Dirichlet faces.
                if (gId < 0)
                {
                    continue;
                }
 
                const NekDouble sign1 = sign[j];
 
                // Extract this face, along with sign array for assembly
                // later.
                for (k = 0; k < nFaceCoeffs; ++k)
                {
                    tmpStore[k + j * nFaceCoeffs] =
                        sign1 * sign[k] * (*schurMat)(bmap2[j], bmap2[k]);
                }
            }
 
            if (gId < 0)
            {
                continue;
            }
 
            gidDofs[2][gId] = nFaceCoeffs;
 
            // Assemble this face matrix with another one, if it exists.
            auto gIt = idMats[2].find(gId);
            if (gIt == idMats[2].end())
            {
                idMats[2][gId] = tmpStore;
            }
            else
            {
                ASSERTL1(tmpStore.size() == gIt->second.size(),
                         "Number of modes mismatch");
                Vmath::Vadd(nFaceCoeffs * nFaceCoeffs, &gIt->second[0], 1,
                            &tmpStore[0], 1, &gIt->second[0], 1);
            }
 
            gidMeshIds[2][gId] = meshFaceId;
            maxVertIds[4]      = max(maxVertIds[4], meshFaceId);
            maxVertIds[5]      = max(maxVertIds[5], nFaceCoeffs);
        }
 
        cnt += exp->GetNcoeffs();
    }
 
    // Perform a reduction to find maximum vertex, edge and face
    // geometry IDs.
    LibUtilities::CommSharedPtr vComm =
        expList->GetSession()->GetComm()->GetRowComm();
    vComm->AllReduce(maxVertIds, LibUtilities::ReduceMax);
 
    // Concatenate all matrices into contiguous storage and figure out
    // universal ID numbering.
    vector<NekDouble> storageBuf;
    vector<long> globalToUniversal;
 
    for (i = 0, cnt = 1; i < 3; ++i)
    {
        const int maxDofs = maxVertIds[2 * i + 1];
 
        // Note that iterating over the map uses the property that keys
        // are accessed in order of ascending order, putting everything
        // in the right order for the global system.
        for (auto &gIt : idMats[i])
        {
            // Copy matrix into storage.
            storageBuf.insert(storageBuf.end(), gIt.second.begin(),
                              gIt.second.end());
 
            // Get mesh ID from global ID number.
            ASSERTL1(gidMeshIds[i].count(gIt.first) > 0,
                     "Unable to find global ID " +
                         boost::lexical_cast<string>(gIt.first) +
                         " inside map");
            meshVertId = gidMeshIds[i][gIt.first];
 
            for (j = 0; j < gIt.second.size(); ++j)
            {
                globalToUniversal.push_back(cnt +
                                            meshVertId * maxDofs * maxDofs + j);
            }
 
            // Free up the temporary storage.
            gIt.second.clear();
        }
 
        cnt += (maxVertIds[2 * i] + 1) * maxDofs * maxDofs;
    }
 
    ASSERTL1(storageBuf.size() == globalToUniversal.size(),
             "Storage buffer and global to universal map size does "
             "not match");
 
    Array<OneD, NekDouble> storageData(storageBuf.size(), &storageBuf[0]);
    Array<OneD, long> globalToUniversalMap(globalToUniversal.size(),
                                           &globalToUniversal[0]);
 
    // Initialise Gs mapping
    m_gs_mapping =
        Gs::Init(globalToUniversalMap, vComm,
                 expList->GetSession()->DefinesCmdLineArgument("verbose"));
 
    // Use GSlib to assemble data between processors.
    Gs::Gather(storageData, Gs::gs_add, m_gs_mapping);
 
    // Figure out what storage we need in the block matrix.
    int nblksSize = 1 + idMats[1].size() + idMats[2].size();
    if (m_isFull)
    {
        nblksSize += idMats[3].size();
    }
    Array<OneD, unsigned int> n_blks(nblksSize);
 
    // Vertex block is a diagonal matrix.
    n_blks[0] = idMats[0].size();
 
    // Now extract number of rows in each edge and face block from the
    // gidDofs map.
    cnt = 1;
    for (i = 1; i < nStorage; ++i)
    {
        for (auto &gIt : idMats[i])
        {
            ASSERTL1(gidDofs[i].count(gIt.first) > 0,
                     "Unable to find number of degrees of freedom for "
                     "global ID " +
                         boost::lexical_cast<string>(gIt.first));
 
            // Check for non-zero DoF count
            if (gidDofs[i][gIt.first])
            {
                n_blks[cnt++] = gidDofs[i][gIt.first];
            }
        }
    }
 
    // Allocate storage for the block matrix.
    m_blkMat =
        MemoryManager<DNekBlkMat>::AllocateSharedPtr(n_blks, n_blks, eDIAGONAL);
 
    // We deal with the vertex matrix separately since all vertices can
    // be condensed into a single, block-diagonal matrix.
    DNekMatSharedPtr vertMat = MemoryManager<DNekMat>::AllocateSharedPtr(
        n_blks[0], n_blks[0], 0.0, eDIAGONAL);
 
    // Fill the vertex matrix with the inverse of each vertex value.
    cnt = 0;
    for (auto gIt = idMats[0].begin(); gIt != idMats[0].end(); ++gIt, ++cnt)
    {
        (*vertMat)(cnt, cnt) = 1.0 / storageData[cnt];
    }
 
    // Put the vertex matrix in the block matrix.
    m_blkMat->SetBlock(0, 0, vertMat);
 
    // Finally, grab the matrices from the block storage, invert them
    // and place them in the correct position inside the block matrix.
    int cnt2 = 1;
    for (i = 1; i < nStorage; ++i)
    {
        for (auto &gIt : idMats[i])
        {
            int nDofs = gidDofs[i][gIt.first];
 
            // Skip, if zero DoFs
            if (nDofs == 0)
            {
                continue;
            }
 
            DNekMatSharedPtr tmp =
                MemoryManager<DNekMat>::AllocateSharedPtr(nDofs, nDofs);
 
            for (j = 0; j < nDofs; ++j)
            {
                for (k = 0; k < nDofs; ++k)
                {
                    // Interior matrices do not need mapping
                    // and are stored in idMats[3]
                    if (m_isFull && i == 3)
                    {
                        (*tmp)(j, k) = gIt.second[k + j * nDofs];
                    }
                    // Boundary matrices
                    else
                    {
                        (*tmp)(j, k) = storageData[k + j * nDofs + cnt];
                    }
                }
            }
 
            cnt += nDofs * nDofs;
 
            tmp->Invert();
            m_blkMat->SetBlock(cnt2, cnt2, tmp);
            ++cnt2;
        }
    }
}

References Nektar::MemoryManager< DataType >::AllocateSharedPtr(), ASSERTL1, Nektar::MultiRegions::DeterminePeriodicEdgeOrientId(), Nektar::MultiRegions::DeterminePeriodicFaceOrient(), Nektar::eDIAGONAL, Gs::Gather(), Gs::gs_add, Gs::Init(), m_blkMat, m_gs_mapping, m_isFull, Nektar::MultiRegions::Preconditioner::m_linsys, Nektar::MultiRegions::Preconditioner::m_locToGloMap, Nektar::LibUtilities::ReduceMax, sign, and Vmath::Vadd().

Referenced by v_BuildPreconditioner().

◆ BlockPreconditionerHDG()

void Nektar::MultiRegions::PreconditionerBlock::BlockPreconditionerHDG ( void )

private

Construct a block preconditioner for the hybridized discontinuous Galerkin system.

This system is built in a similar fashion to its continuous variant found in PreconditionerBlock::BlockPreconditionerCG. In this setting however, the matrix is constructed as:

\[ M^{-1} = \mathrm{Diag}[ (\mathbf{S_{1}})_{f}^{-1} ] \]

where each matrix is the Schur complement system restricted to a single face of the trace system.

Definition at line 657 of file PreconditionerBlock.cpp.

{
    std::shared_ptr<MultiRegions::ExpList> expList =
        ((m_linsys.lock())->GetLocMat()).lock();
    std::shared_ptr<MultiRegions::ExpList> trace = expList->GetTrace();
    LocalRegions::ExpansionSharedPtr locExpansion;
    DNekScalBlkMatSharedPtr loc_mat;
    DNekScalMatSharedPtr bnd_mat;
 
    AssemblyMapDGSharedPtr asmMap =
        std::dynamic_pointer_cast<AssemblyMapDG>(m_locToGloMap.lock());
 
    int i, j, k, n, cnt, cnt2;
 
    // Figure out number of Dirichlet trace elements
    int nTrace = expList->GetTrace()->GetExpSize();
    int nDir   = asmMap->GetNumGlobalDirBndCoeffs();
 
    for (cnt = n = 0; n < nTrace; ++n)
    {
        if (cnt >= nDir)
        {
            break;
        }
 
        cnt += trace->GetExp(n)->GetNcoeffs();
    }
 
    nDir = n;
 
    // Allocate storage for block matrix. Need number of unique faces in
    // trace space.
    int maxTraceSize = -1;
    map<int, int> arrayOffsets;
 
    for (cnt = 0, n = nDir; n < nTrace; ++n)
    {
        int nCoeffs  = trace->GetExp(n)->GetNcoeffs();
        int nCoeffs2 = nCoeffs * nCoeffs;
 
        arrayOffsets[n] = cnt;
        cnt += nCoeffs2;
 
        if (maxTraceSize < nCoeffs2)
        {
            maxTraceSize = nCoeffs2;
        }
    }
 
    // Find maximum trace size.
    LibUtilities::CommSharedPtr vComm =
        expList->GetSession()->GetComm()->GetRowComm();
    vComm->AllReduce(maxTraceSize, LibUtilities::ReduceMax);
 
    // Zero matrix storage.
    Array<OneD, NekDouble> tmpStore(cnt, 0.0);
 
    // Assemble block matrices for each trace element.
    for (cnt = n = 0; n < expList->GetExpSize(); ++n)
    {
        int elmt     = n;
        locExpansion = expList->GetExp(elmt);
 
        Array<OneD, LocalRegions::ExpansionSharedPtr> &elmtToTraceMap =
            asmMap->GetElmtToTrace()[elmt];
 
        // Block matrix (lambda)
        loc_mat = (m_linsys.lock())->GetStaticCondBlock(n);
        bnd_mat = loc_mat->GetBlock(0, 0);
 
        int nFacets = locExpansion->GetNtraces();
 
        for (cnt2 = i = 0; i < nFacets; ++i)
        {
            int nCoeffs = elmtToTraceMap[i]->GetNcoeffs();
            int elmtId  = elmtToTraceMap[i]->GetElmtId();
 
            // Ignore Dirichlet edges.
            if (elmtId < nDir)
            {
                cnt += nCoeffs;
                cnt2 += nCoeffs;
                continue;
            }
 
            NekDouble *off = &tmpStore[arrayOffsets[elmtId]];
 
            for (j = 0; j < nCoeffs; ++j)
            {
                NekDouble sign1 = asmMap->GetLocalToGlobalBndSign(cnt + j);
                for (k = 0; k < nCoeffs; ++k)
                {
                    NekDouble sign2 = asmMap->GetLocalToGlobalBndSign(cnt + k);
                    off[j * nCoeffs + k] +=
                        (*bnd_mat)(cnt2 + j, cnt2 + k) * sign1 * sign2;
                }
            }
 
            cnt += nCoeffs;
            cnt2 += nCoeffs;
        }
    }
 
    // Set up IDs for universal numbering.
    Array<OneD, long> uniIds(tmpStore.size());
    for (cnt = 0, n = nDir; n < nTrace; ++n)
    {
        LocalRegions::ExpansionSharedPtr traceExp = trace->GetExp(n);
        int nCoeffs                               = traceExp->GetNcoeffs();
        int geomId = traceExp->GetGeom()->GetGlobalID();
 
        for (i = 0; i < nCoeffs * nCoeffs; ++i)
        {
            uniIds[cnt++] = geomId * maxTraceSize + i + 1;
        }
    }
 
    // Assemble matrices across partitions.
    Gs::gs_data *gsh =
        Gs::Init(uniIds, vComm,
                 expList->GetSession()->DefinesCmdLineArgument("verbose"));
    Gs::Gather(tmpStore, Gs::gs_add, gsh);
 
    // Set up diagonal block matrix
    Array<OneD, unsigned int> n_blks(nTrace - nDir);
    for (n = 0; n < nTrace - nDir; ++n)
    {
        n_blks[n] = trace->GetExp(n + nDir)->GetNcoeffs();
    }
 
    m_blkMat =
        MemoryManager<DNekBlkMat>::AllocateSharedPtr(n_blks, n_blks, eDIAGONAL);
 
    for (n = 0; n < nTrace - nDir; ++n)
    {
        int nCoeffs = trace->GetExp(n + nDir)->GetNcoeffs();
        DNekMatSharedPtr tmp =
            MemoryManager<DNekMat>::AllocateSharedPtr(nCoeffs, nCoeffs);
        NekDouble *off = &tmpStore[arrayOffsets[n + nDir]];
 
        for (i = 0; i < nCoeffs; ++i)
        {
            for (j = 0; j < nCoeffs; ++j)
            {
                (*tmp)(i, j) = off[i * nCoeffs + j];
            }
        }
 
        tmp->Invert();
        m_blkMat->SetBlock(n, n, tmp);
    }
}

References Nektar::MemoryManager< DataType >::AllocateSharedPtr(), Nektar::eDIAGONAL, Gs::Gather(), Gs::gs_add, Gs::Init(), m_blkMat, Nektar::MultiRegions::Preconditioner::m_linsys, Nektar::MultiRegions::Preconditioner::m_locToGloMap, and Nektar::LibUtilities::ReduceMax.

Referenced by v_BuildPreconditioner().

◆ create()

static PreconditionerSharedPtr Nektar::MultiRegions::PreconditionerBlock::create	(	const std::shared_ptr< GlobalLinSys > &	plinsys,
		const std::shared_ptr< AssemblyMap > &	pLocToGloMap
	)

inlinestatic

Creates an instance of this class.

Definition at line 52 of file PreconditionerBlock.h.

    {
        PreconditionerSharedPtr p =
            MemoryManager<PreconditionerBlock>::AllocateSharedPtr(plinsys,
                                                                  pLocToGloMap);
        p->InitObject();
        return p;
    }

References Nektar::MemoryManager< DataType >::AllocateSharedPtr(), and CellMLToNektar.cellml_metadata::p.

◆ v_BuildPreconditioner()

void Nektar::MultiRegions::PreconditionerBlock::v_BuildPreconditioner ( )

overrideprotectedvirtual

Reimplemented from Nektar::MultiRegions::Preconditioner.

Definition at line 79 of file PreconditionerBlock.cpp.

{
    GlobalLinSysKey key = m_linsys.lock()->GetKey();
 
    // Different setup for HDG and CG.
    if (key.GetMatrixType() == StdRegions::eHybridDGHelmBndLam)
    {
        BlockPreconditionerHDG();
    }
    else
    {
        BlockPreconditionerCG();
    }
}

References BlockPreconditionerCG(), BlockPreconditionerHDG(), Nektar::StdRegions::eHybridDGHelmBndLam, Nektar::MultiRegions::GlobalMatrixKey::GetMatrixType(), and Nektar::MultiRegions::Preconditioner::m_linsys.

◆ v_DoPreconditioner()

void Nektar::MultiRegions::PreconditionerBlock::v_DoPreconditioner	(	const Array< OneD, NekDouble > &	pInput,
		Array< OneD, NekDouble > &	pOutput,
		const bool &	isLocal = `false`
	)

overrideprotectedvirtual

Apply preconditioner to pInput and store the result in pOutput.

Reimplemented from Nektar::MultiRegions::Preconditioner.

Definition at line 814 of file PreconditionerBlock.cpp.

{
    // Get assembly map and solver type
    auto asmMap = m_locToGloMap.lock();
 
    // Determine matrix size
    int nGlobal = m_isFull ? asmMap->GetNumGlobalCoeffs()
                           : asmMap->GetNumGlobalBndCoeffs();
    int nDir    = asmMap->GetNumGlobalDirBndCoeffs();
    int nNonDir = nGlobal - nDir;
 
    // Apply preconditioner
    DNekBlkMat &M = (*m_blkMat);
 
    if (isLocal)
    {
        Array<OneD, NekDouble> wk(nGlobal);
        Array<OneD, NekDouble> wk1(nGlobal);
        asmMap->Assemble(pInput, wk);
 
        NekVector<NekDouble> r(nNonDir, wk + nDir, eWrapper);
        NekVector<NekDouble> z(nNonDir, wk1 + nDir, eWrapper);
 
        z = M * r;
        Vmath::Zero(nDir, wk1, 1);
 
        asmMap->GlobalToLocal(wk1, pOutput);
    }
    else
    {
        NekVector<NekDouble> r(nNonDir, pInput, eWrapper);
        NekVector<NekDouble> z(nNonDir, pOutput, eWrapper);
        z = M * r;
    }
}

References Nektar::eWrapper, m_isFull, Nektar::MultiRegions::Preconditioner::m_locToGloMap, Nektar::UnitTests::z(), and Vmath::Zero().

◆ v_InitObject()

void Nektar::MultiRegions::PreconditionerBlock::v_InitObject ( )

overrideprotectedvirtual

Reimplemented from Nektar::MultiRegions::Preconditioner.

Definition at line 68 of file PreconditionerBlock.cpp.

{
    GlobalSysSolnType solverType = m_locToGloMap.lock()->GetGlobalSysSolnType();
    m_isFull =
        (solverType == eIterativeFull) || (solverType == ePETScFullMatrix);
 
    ASSERTL0(solverType == MultiRegions::eIterativeStaticCond ||
                 solverType == MultiRegions::ePETScStaticCond || m_isFull,
             "Solver type not valid");
}

References ASSERTL0, Nektar::MultiRegions::eIterativeFull, Nektar::MultiRegions::eIterativeStaticCond, Nektar::MultiRegions::ePETScFullMatrix, Nektar::MultiRegions::ePETScStaticCond, m_isFull, and Nektar::MultiRegions::Preconditioner::m_locToGloMap.

Member Data Documentation

◆ className

string Nektar::MultiRegions::PreconditionerBlock::className

static

Initial value:

=
    GetPreconFactory().RegisterCreatorFunction(
        "Block", PreconditionerBlock::create, "Block Diagonal Preconditioning")

Name of class.

Registers the class with the Factory.

Definition at line 64 of file PreconditionerBlock.h.

◆ m_blkMat

DNekBlkMatSharedPtr Nektar::MultiRegions::PreconditionerBlock::m_blkMat

protected

Definition at line 78 of file PreconditionerBlock.h.

Referenced by BlockPreconditionerCG(), and BlockPreconditionerHDG().

◆ m_gs_mapping

Gs::gs_data* Nektar::MultiRegions::PreconditionerBlock::m_gs_mapping

protected

Definition at line 81 of file PreconditionerBlock.h.

Referenced by BlockPreconditionerCG(), and ~PreconditionerBlock().

◆ m_isFull

bool Nektar::MultiRegions::PreconditionerBlock::m_isFull

protected

Definition at line 77 of file PreconditionerBlock.h.

Referenced by BlockPreconditionerCG(), v_DoPreconditioner(), and v_InitObject().

Public Member Functions

Static Public Member Functions

Static Public Attributes

Protected Member Functions

Protected Attributes

Private Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ PreconditionerBlock()

◆ ~PreconditionerBlock()

Member Function Documentation

◆ BlockPreconditionerCG()

◆ BlockPreconditionerHDG()

◆ create()

◆ v_BuildPreconditioner()

◆ v_DoPreconditioner()

◆ v_InitObject()

Member Data Documentation

◆ className

◆ m_blkMat

◆ m_gs_mapping

◆ m_isFull