// prepares the file catalogue and searches for given text
void Executive::PrepareCatalogue()
{
bool IsRecursive = false;
DataStore ds;
FileManager Fm(path_, ds);
Fm.addAllPatterns(Patterns_);
Arguments::iterator argIter = std::find(Options_.begin(), Options_.end(), "/s");
if (argIter != Options_.end())
{
IsRecursive = true;
}
Fm.search(IsRecursive);
Display dp;
dp.DisplayCatalogue(ds);
argIter = std::find(Options_.begin(), Options_.end(), "/d");
if (argIter != Options_.end())
{
dp.DisplayDuplicates(ds);
}
argIter = std::find(Options_.begin(), Options_.end(), "/f");
if (argIter != Options_.end())
{
Fm.TextSearch(ds, searchText);
dp.DisplaySearchFiles(ds, searchText);
}
DoSearch(Fm, ds, dp);
}
// Right-hand side of the Lax-Wendroff discretization:
//
// ( q(t+dt) - q(t) )/dt = -F_{x} - G_{y}.
//
// This routine constructs the 'time-integrated' flux functions F and G using the
// Cauchy-Kowalewski procedure.
//
// First, consider time derivatives of q
//
// q^{n+1} = q^n + dt * (q^n_t + dt/2 * q^n_tt + dt^3/6 q^n_ttt ).
//
// Formally, these are given by
//
// q_{t} = ( -f(q) )_x + ( -g(q) )_y,
// q_{tt} = ( f'(q) * ( f_x + g_y )_x + ( g'(q) * ( f_x + g_y )_y
// q_{ttt} = ...
//
// Now, considering Taylor expansions of f and g, centered about t = t^n
//
// F = f^n + (t-t^n) \dot{f^n} + \cdots
// G = g^n + (t-t^n) \dot{g^n} + \cdots
//
// We have the following form, after integrating in time
//
// F: = ( f - dt/2 * ( f'(q)*( f_x+g_y )
// + dt^2/6 ( \pd2{f}{q} \cdot (f_x+g_y,f_x+g_y) +
// \pd{f}{q} ( f_x + g_y )_t ) + \cdots
//
// G: = ( g - dt/2 * ( g'(q)*( f_x+g_y )
// + dt^2/6 ( \pd2{g}{q} \cdot (f_x+g_y,f_x+g_y) +
// \pd{g}{q} ( f_x + g_y )_t ) + \cdots
//
// where the final ingredient is
//
// (f_x+g_y)_t = \pd2{f}{q} \cdot (q_x, f_x+g_y ) + \pd{f}{q} ( f_xx + g_xy ) +
// \pd2{g}{q} \cdot (q_y, f_x+g_y ) + \pd{g}{q} ( f_xy + g_yy ).
//
// At the end of the day, we set
//
// L(q) := -F_x - G_y.
//
// See also: ConstructL.
void LaxWendroff(double dt,
const dTensorBC4& aux, const dTensorBC4& q, // set bndy values modifies these
dTensorBC4& Lstar, dTensorBC3& smax)
{
printf("This call hasn't been tested \n");
if ( !dogParams.get_flux_term() )
{ return; }
const edge_data& edgeData = Legendre2d::get_edgeData();
const int space_order = dogParams.get_space_order();
const int mx = q.getsize(1);
const int my = q.getsize(2);
const int meqn = q.getsize(3);
const int kmax = q.getsize(4);
const int mbc = q.getmbc();
const int maux = aux.getsize(3);
// Flux values
//
// Space-order = number of quadrature points needed for 1D integration
// along cell edges.
//
dTensorBC4 Fm(mx, my, meqn, space_order, mbc);
dTensorBC4 Fp(mx, my, meqn, space_order, mbc);
dTensorBC4 Gm(mx, my, meqn, space_order, mbc);
dTensorBC4 Gp(mx, my, meqn, space_order, mbc);
// Flux function
void FluxFunc(const dTensor2& xpts,
const dTensor2& Q,
const dTensor2& Aux,
dTensor3& flux);
// Jacobian of the flux function:
void DFluxFunc(const dTensor2& xpts,
const dTensor2& Q,
const dTensor2& Aux,
dTensor4& Dflux );
// Hessian of the flux function:
void D2FluxFunc(const dTensor2& xpts,
const dTensor2& Q,
const dTensor2& Aux,
dTensor5& D2flux );
// Riemann solver that relies on the fact that we already have
// f(ql) and f(qr) already computed:
double RiemannSolveLxW(const dTensor1& nvec,
const dTensor1& xedge,
const dTensor1& Ql, const dTensor1& Qr,
const dTensor1& Auxl, const dTensor1& Auxr,
const dTensor1& ffl, const dTensor1& ffr,
dTensor1& Fl, dTensor1& Fr);
void LstarExtra(const dTensorBC4*,
const dTensorBC4*,
dTensorBC4*);
void ArtificialViscosity(const dTensorBC4* aux,
const dTensorBC4* q,
dTensorBC4* Lstar);
// Grid information
const double xlower = dogParamsCart2.get_xlow();
const double ylower = dogParamsCart2.get_ylow();
const double dx = dogParamsCart2.get_dx();
const double dy = dogParamsCart2.get_dy();
// --------------------------------------------------------------------- //
// Boundary data:
// --------------------------------------------------------------------- //
// TODO - call this routine before calling this function.
// void SetBndValues(dTensorBC4& q, dTensorBC4& aux);
// SetBndValues(q, aux);
// ---------------------------------------------------------
// --------------------------------------------------------------------- //
// Part 0: Compute the Lax-Wendroff "flux" function:
//
// Here, we include the extra information about time derivatives.
// --------------------------------------------------------------------- //
dTensorBC4 F(mx, my, meqn, kmax, mbc); F.setall(0.);
dTensorBC4 G(mx, my, meqn, kmax, mbc); G.setall(0.);
void L2ProjectLxW( const int mterms,
const double alpha, const double beta_dt, const double charlie_dt,
const int istart, const int iend,
const int jstart, const int jend,
const int QuadOrder,
const int BasisOrder_auxin,
const int BasisOrder_fout,
const dTensorBC4* qin,
const dTensorBC4* auxin,
dTensorBC4* F,
dTensorBC4* G,
void FluxFunc (const dTensor2& xpts,
const dTensor2& Q, const dTensor2& Aux, dTensor3& flux),
void DFluxFunc (const dTensor2& xpts,
const dTensor2& Q, const dTensor2& aux, dTensor4& Dflux),
void D2FluxFunc (const dTensor2& xpts,
const dTensor2& Q, const dTensor2& aux, dTensor5& D2flux) );
printf("hello\n");
L2ProjectLxW( 3, 1.0, 0.5*dt, dt*dt/6.0,
1-mbc, mx+mbc, 1-mbc, my+mbc,
space_order, space_order, space_order,
&q, &aux, &F, &G, &FluxFunc, &DFluxFunc, D2FluxFunc );
// ---------------------------------------------------------
// Part I: compute source term
// ---------------------------------------------------------
if( dogParams.get_source_term() > 0 )
{
// eprintf("error: have not implemented source term for LxW solver.");
printf("Source term has not been implemented for LxW solver. Terminating program.");
exit(1);
}
Lstar.setall( 0. );
// ---------------------------------------------------------
// Part II: compute inter-element interaction fluxes
//
// N = int( F(q,x,t) * phi_x, x ) / dA
//
// ---------------------------------------------------------
// 1-direction: loop over interior edges and solve Riemann problems
dTensor1 nvec(2);
nvec.set(1, 1.0e0 );
nvec.set(2, 0.0e0 );
#pragma omp parallel for
for (int i=(2-mbc); i<=(mx+mbc); i++)
{
dTensor1 Ql(meqn), Qr(meqn);
dTensor1 ffl(meqn), ffr(meqn);
dTensor1 Fl(meqn), Fr(meqn);
dTensor1 DFl(meqn), DFr(meqn);
dTensor1 Auxl(maux), Auxr(maux);
dTensor1 xedge(2);
for (int j=(2-mbc); j<=(my+mbc-1); j++)
{
// ell indexes Riemann point along the edge
for (int ell=1; ell<=space_order; ell++)
{
// Riemann data - q and f (from basis functions/q)
for (int m=1; m<=meqn; m++)
{
Ql.set (m, 0.0 );
Qr.set (m, 0.0 );
ffl.set(m, 0.0 );
ffr.set(m, 0.0 );
for (int k=1; k<=kmax; k++)
{
// phi_xl( xi=1.0, eta ), phi_xr( xi=-1.0, eta )
Ql.fetch(m) += edgeData.phi_xl->get(ell,k)*q.get(i-1, j, m, k );
Qr.fetch(m) += edgeData.phi_xr->get(ell,k)*q.get(i, j, m, k );
ffl.fetch(m) += edgeData.phi_xl->get(ell,k)*F.get(i-1, j, m, k );
ffr.fetch(m) += edgeData.phi_xr->get(ell,k)*F.get(i, j, m, k );
}
}
// Riemann data - aux
for (int m=1; m<=maux; m++)
{
Auxl.set(m, 0.0 );
Auxr.set(m, 0.0 );
for (int k=1; k<=kmax; k++)
{
Auxl.fetch(m) += edgeData.phi_xl->get(ell,k)*aux.get(i-1, j, m, k);
Auxr.fetch(m) += edgeData.phi_xr->get(ell,k)*aux.get(i, j, m, k);
}
}
// Solve Riemann problem
xedge.set(1, xlower + (double(i)-1.0)*dx );
xedge.set(2, ylower + (double(j)-0.5)*dy );
const double smax_edge = RiemannSolveLxW(
nvec, xedge, Ql, Qr, Auxl, Auxr, ffl, ffr, Fl, Fr);
smax.set(i-1, j, 1, Max(dy*smax_edge,smax.get(i-1, j, 1)) );
smax.set(i, j, 1, Max(dy*smax_edge,smax.get(i, j, 1)) );
// Construct fluxes
for (int m=1; m<=meqn; m++)
{
Fm.set(i , j, m, ell, Fr.get(m) );
Fp.set(i-1, j, m, ell, Fl.get(m) );
}
}
}
}
// 2-direction: loop over interior edges and solve Riemann problems
nvec.set(1, 0.0e0 );
nvec.set(2, 1.0e0 );
#pragma omp parallel for
for (int i=(2-mbc); i<=(mx+mbc-1); i++)
{
dTensor1 Ql(meqn), Qr(meqn);
dTensor1 Fl(meqn), Fr(meqn);
dTensor1 ffl(meqn), ffr(meqn);
dTensor1 Auxl(maux),Auxr(maux);
dTensor1 xedge(2);
for (int j=(2-mbc); j<=(my+mbc); j++)
for (int ell=1; ell<=space_order; ell++)
{
// Riemann data - q
for (int m=1; m<=meqn; m++)
{
Ql.set (m, 0.0 );
Qr.set (m, 0.0 );
ffl.set (m, 0.0 );
ffr.set (m, 0.0 );
for (int k=1; k<=kmax; k++)
{
Ql.fetch(m) += edgeData.phi_yl->get(ell, k)*q.get(i, j-1, m, k );
Qr.fetch(m) += edgeData.phi_yr->get(ell, k)*q.get(i, j, m, k );
ffl.fetch(m) += edgeData.phi_yl->get(ell, k)*G.get(i, j-1, m, k );
ffr.fetch(m) += edgeData.phi_yr->get(ell, k)*G.get(i, j, m, k );
}
}
// Riemann data - aux
for (int m=1; m<=maux; m++)
{
Auxl.set(m, 0.0 );
Auxr.set(m, 0.0 );
for (int k=1; k<=kmax; k++)
{
Auxl.fetch(m) += edgeData.phi_yl->get(ell,k)*aux.get(i,j-1,m,k);
Auxr.fetch(m) += edgeData.phi_yr->get(ell,k)*aux.get(i,j,m,k);
}
}
// Solve Riemann problem
xedge.set(1, xlower + (double(i)-0.5)*dx );
xedge.set(2, ylower + (double(j)-1.0)*dy );
const double smax_edge = RiemannSolveLxW(
nvec, xedge, Ql, Qr, Auxl, Auxr, ffl, ffr, Fl, Fr);
smax.set(i, j-1, 2, Max(dx*smax_edge, smax.get(i, j-1, 2)) );
smax.set(i, j, 2, Max(dx*smax_edge, smax.get(i, j, 2)) );
// Construct fluxes
for (int m=1; m<=meqn; m++)
{
Gm.set(i, j, m, ell, Fr.get(m) );
Gp.set(i, j-1, m, ell, Fl.get(m) );
}
}
}
// Compute ``flux differences'' dF and dG
const double half_dx_inv = 0.5/dx;
const double half_dy_inv = 0.5/dy;
const int mlength = Lstar.getsize(3); assert_eq( meqn, mlength );
// Use the four values, Gm, Gp, Fm, Fp to construct the boundary integral:
#pragma omp parallel for
for (int i=(2-mbc); i<=(mx+mbc-1); i++)
for (int j=(2-mbc); j<=(my+mbc-1); j++)
for (int m=1; m<=mlength; m++)
for (int k=1; k<=kmax; k++)
{
// 1-direction: dF
double F1 = 0.0;
double F2 = 0.0;
for (int ell=1; ell<=space_order; ell++)
{
F1 = F1 + edgeData.wght_phi_xr->get(ell,k)*Fm.get(i,j,m,ell);
F2 = F2 + edgeData.wght_phi_xl->get(ell,k)*Fp.get(i,j,m,ell);
}
// 2-direction: dG
double G1 = 0.0;
double G2 = 0.0;
for (int ell=1; ell<=space_order; ell++)
{
G1 = G1 + edgeData.wght_phi_yr->get(ell,k)*Gm.get(i,j,m,ell);
G2 = G2 + edgeData.wght_phi_yl->get(ell,k)*Gp.get(i,j,m,ell);
}
Lstar.fetch(i,j,m,k) -= (half_dx_inv*(F2-F1) + half_dy_inv*(G2-G1));
}
// ---------------------------------------------------------
// ---------------------------------------------------------
// Part III: compute intra-element contributions
// ---------------------------------------------------------
// No need to call this if first-order in space
if(dogParams.get_space_order()>1)
{
dTensorBC4 Ltmp( mx, my, meqn, kmax, mbc );
void L2ProjectGradAddLegendre(const int istart,
const int iend,
const int jstart,
const int jend,
const int QuadOrder,
const dTensorBC4* F,
const dTensorBC4* G,
dTensorBC4* fout );
L2ProjectGradAddLegendre( 1-mbc, mx+mbc, 1-mbc, my+mbc,
space_order, &F, &G, &Lstar );
}
// ---------------------------------------------------------
// ---------------------------------------------------------
// Part IV: add extra contributions to Lstar
// ---------------------------------------------------------
// Call LstarExtra
LstarExtra(&q,&aux,&Lstar);
// ---------------------------------------------------------
// ---------------------------------------------------------
// Part V: artificial viscosity limiter
// ---------------------------------------------------------
if (dogParams.get_space_order()>1 &&
dogParams.using_viscosity_limiter())
{ ArtificialViscosity(&aux,&q,&Lstar); }
// ---------------------------------------------------------
}
//---------------------------------------------------------
DVec& NDG2D::PoissonIPDGbc2D
(DVec& ubc, //[in]
DVec& qbc //[in]
)
//---------------------------------------------------------
{
// function [OP] = PoissonIPDGbc2D()
// Purpose: Set up the discrete Poisson matrix directly
// using LDG. The operator is set up in the weak form
// build DG derivative matrices
int max_OP = (K*Np*Np*(1+Nfaces));
// initialize parameters
DVec faceR("faceR"), faceS("faceS");
DMat V1D("V1D"), Dx("Dx"),Dy("Dy"), Dn1("Dn1"), mmE_Fm1("mmE(:,Fm1)");
IVec Fm("Fm"), Fm1("Fm1"), fidM("fidM");
double lnx=0.0,lny=0.0,lsJ=0.0,hinv=0.0,gtau=0.0;
int i=0,k1=0,f1=0,id=0;
IVec i1_Nfp = Range(1,Nfp);
double N1N1 = double((N+1)*(N+1));
// build local face matrices
DMat massEdge[4]; // = zeros(Np,Np,Nfaces);
for (i=1; i<=Nfaces; ++i) {
massEdge[i].resize(Np,Np);
}
// face mass matrix 1
Fm = Fmask(All,1); faceR = r(Fm);
V1D = Vandermonde1D(N, faceR);
massEdge[1](Fm,Fm) = inv(V1D*trans(V1D));
// face mass matrix 2
Fm = Fmask(All,2); faceR = r(Fm);
V1D = Vandermonde1D(N, faceR);
massEdge[2](Fm,Fm) = inv(V1D*trans(V1D));
// face mass matrix 3
Fm = Fmask(All,3); faceS = s(Fm);
V1D = Vandermonde1D(N, faceS);
massEdge[3](Fm,Fm) = inv(V1D*trans(V1D));
// build DG right hand side
DVec* pBC = new DVec(Np*K, "bc", OBJ_temp);
DVec& bc = (*pBC); // reference, for syntax
////////////////////////////////////////////////////////////////
umMSG(1, "\n ==> {OP} assembly [bc]: ");
for (k1=1; k1<=K; ++k1)
{
if (! (k1%100)) { umMSG(1, "%d, ",k1); }
// rows1 = outer(Range((k1-1)*Np+1,k1*Np), Ones(NGauss));
// Build element-to-element parts of operator
for (f1=1; f1<=Nfaces; ++f1)
{
if (BCType(k1,f1))
{
////////////////////////added by Kevin ///////////////////////////////
Fm1 = Fmask(All,f1);
fidM = (k1-1)*Nfp*Nfaces + (f1-1)*Nfp + i1_Nfp;
id = 1+(f1-1)*Nfp + (k1-1)*Nfp*Nfaces;
lnx = nx(id); lny = ny(id);
lsJ = sJ(id); hinv = Fscale(id);
Dx = rx(1,k1)*Dr + sx(1,k1)*Ds;
Dy = ry(1,k1)*Dr + sy(1,k1)*Ds;
Dn1 = lnx*Dx + lny*Dy;
//mmE = lsJ*massEdge(:,:,f1);
//bc(All,k1) += (gtau*mmE(All,Fm1) - Dn1'*mmE(All,Fm1))*ubc(fidM);
mmE_Fm1 = massEdge[f1](All,Fm1); mmE_Fm1 *= lsJ;
gtau = 10*N1N1*hinv; // set penalty scaling
//bc(All,k1) += (gtau*mmE_Fm1 - trans(Dn1)*mmE_Fm1) * ubc(fidM);
switch(BCType(k1,f1)){
case BC_Dirichlet:
bc(Np*(k1-1)+Range(1,Np)) += (gtau*mmE_Fm1 - trans(Dn1)*mmE_Fm1)*ubc(fidM);
break;
case BC_Neuman:
bc(Np*(k1-1)+Range(1,Np)) += mmE_Fm1*qbc(fidM);
break;
default:
std::cout<<"warning: boundary condition is incorrect"<<std::endl;
}
}
}
}
return bc;
}
//---------------------------------------------------------
void NDG3D::PoissonIPDG3D(CSd& spOP, CSd& spMM)
//---------------------------------------------------------
{
// function [OP,MM] = PoissonIPDG3D()
//
// Purpose: Set up the discrete Poisson matrix directly
// using LDG. The operator is set up in the weak form
DVec faceR("faceR"), faceS("faceS"), faceT("faceT");
DMat V2D; IVec Fm("Fm"); IVec i1_Nfp = Range(1,Nfp);
double opti1=0.0, opti2=0.0; int i=0;
umLOG(1, "\n ==> {OP,MM} assembly: ");
opti1 = timer.read(); // time assembly
// build local face matrices
DMat massEdge[5]; // = zeros(Np,Np,Nfaces);
for (i=1; i<=Nfaces; ++i) {
massEdge[i].resize(Np,Np);
}
// face mass matrix 1
Fm = Fmask(All,1); faceR=r(Fm); faceS=s(Fm);
V2D = Vandermonde2D(N, faceR, faceS);
massEdge[1](Fm,Fm) = inv(V2D*trans(V2D));
// face mass matrix 2
Fm = Fmask(All,2); faceR = r(Fm); faceT = t(Fm);
V2D = Vandermonde2D(N, faceR, faceT);
massEdge[2](Fm,Fm) = inv(V2D*trans(V2D));
// face mass matrix 3
Fm = Fmask(All,3); faceS = s(Fm); faceT = t(Fm);
V2D = Vandermonde2D(N, faceS, faceT);
massEdge[3](Fm,Fm) = inv(V2D*trans(V2D));
// face mass matrix 4
Fm = Fmask(All,4); faceS = s(Fm); faceT = t(Fm);
V2D = Vandermonde2D(N, faceS, faceT);
massEdge[4](Fm,Fm) = inv(V2D*trans(V2D));
// build local volume mass matrix
MassMatrix = trans(invV)*invV;
DMat Dx("Dx"),Dy("Dy"),Dz("Dz"), Dx2("Dx2"),Dy2("Dy2"),Dz2("Dz2");
DMat Dn1("Dn1"),Dn2("Dn2"), mmE("mmE"), OP11("OP11"), OP12("OP12");
DMat mmE_All_Fm1, mmE_Fm1_Fm1, Dn2_Fm2_All;
IMat rows1,cols1,rows2,cols2; int k1=0,f1=0,k2=0,f2=0,id=0;
Index1D entries, entriesMM, idsM; IVec fidM,vidM,Fm1,vidP,Fm2;
double lnx=0.0,lny=0.0,lnz=0.0,lsJ=0.0,hinv=0.0,gtau=0.0;
double N1N1 = double((N+1)*(N+1)); int NpNp = Np*Np;
// build DG derivative matrices
int max_OP = (K*Np*Np*(1+Nfaces));
int max_MM = (K*Np*Np);
// "OP" triplets (i,j,x), extracted to {Ai,Aj,Ax}
IVec OPi(max_OP), OPj(max_OP), Ai,Aj; DVec OPx(max_OP), Ax;
// "MM" triplets (i,j,x)
IVec MMi(max_MM), MMj(max_MM); DVec MMx(max_MM);
IVec OnesNp = Ones(Np);
// global node numbering
entries.reset(1,NpNp); entriesMM.reset(1,NpNp);
OP12.resize(Np,Np);
for (k1=1; k1<=K; ++k1)
{
if (! (k1%250)) { umLOG(1, "%d, ",k1); }
rows1 = outer( Range((k1-1)*Np+1,k1*Np), OnesNp );
cols1 = trans(rows1);
// Build local operators
Dx = rx(1,k1)*Dr + sx(1,k1)*Ds + tx(1,k1)*Dt;
Dy = ry(1,k1)*Dr + sy(1,k1)*Ds + ty(1,k1)*Dt;
Dz = rz(1,k1)*Dr + sz(1,k1)*Ds + tz(1,k1)*Dt;
OP11 = J(1,k1)*(trans(Dx)*MassMatrix*Dx +
trans(Dy)*MassMatrix*Dy +
trans(Dz)*MassMatrix*Dz);
// Build element-to-element parts of operator
for (f1=1; f1<=Nfaces; ++f1) {
k2 = EToE(k1,f1); f2 = EToF(k1,f1);
rows2 = outer( Range((k2-1)*Np+1, k2*Np), OnesNp );
cols2 = trans(rows2);
fidM = (k1-1)*Nfp*Nfaces + (f1-1)*Nfp + i1_Nfp;
vidM = vmapM(fidM); Fm1 = mod(vidM-1,Np)+1;
vidP = vmapP(fidM); Fm2 = mod(vidP-1,Np)+1;
id = 1+(f1-1)*Nfp + (k1-1)*Nfp*Nfaces;
lnx = nx(id); lny = ny(id); lnz = nz(id); lsJ = sJ(id);
hinv = std::max(Fscale(id), Fscale(1+(f2-1)*Nfp, k2));
Dx2 = rx(1,k2)*Dr + sx(1,k2)*Ds + tx(1,k2)*Dt;
Dy2 = ry(1,k2)*Dr + sy(1,k2)*Ds + ty(1,k2)*Dt;
Dz2 = rz(1,k2)*Dr + sz(1,k2)*Ds + tz(1,k2)*Dt;
Dn1 = lnx*Dx + lny*Dy + lnz*Dz;
Dn2 = lnx*Dx2 + lny*Dy2 + lnz*Dz2;
mmE = lsJ*massEdge[f1];
gtau = 2.0 * N1N1 * hinv; // set penalty scaling
if (EToE(k1,f1)==k1) {
OP11 += ( gtau*mmE - mmE*Dn1 - trans(Dn1)*mmE ); // ok
}
else
{
// interior face variational terms
OP11 += 0.5*( gtau*mmE - mmE*Dn1 - trans(Dn1)*mmE );
// extract mapped regions:
mmE_All_Fm1 = mmE(All,Fm1);
mmE_Fm1_Fm1 = mmE(Fm1,Fm1);
Dn2_Fm2_All = Dn2(Fm2,All);
OP12 = 0.0; // reset to zero
OP12(All,Fm2) = -0.5*( gtau*mmE_All_Fm1 );
OP12(Fm1,All) -= 0.5*( mmE_Fm1_Fm1*Dn2_Fm2_All );
//OP12(All,Fm2) -= 0.5*(-trans(Dn1)*mmE_All_Fm1 );
OP12(All,Fm2) += 0.5*( trans(Dn1)*mmE_All_Fm1 );
// load this set of triplets
#if (1)
OPi(entries)=rows1; OPj(entries)=cols2, OPx(entries)=OP12;
entries += (NpNp);
#else
//###########################################################
// load only the lower triangle (after droptol test?)
sk=0; start=entries(1);
for (int i=1; i<=NpNp; ++i) {
eid = start+i;
id=entries(eid); rid=rows1(i); cid=cols2(i);
if (rows1(rid) >= cid) { // take lower triangle
if ( fabs(OP12(id)) > 1e-15) { // drop small entries
++sk; OPi(id)=rid; OPj(id)=cid, OPx(id)=OP12(id);
}
}
}
entries += sk;
//###########################################################
#endif
}
}
OPi(entries )=rows1; OPj(entries )=cols1, OPx(entries )=OP11;
MMi(entriesMM)=rows1; MMj(entriesMM)=cols1; MMx(entriesMM)=J(1,k1)*MassMatrix;
entries += (NpNp); entriesMM += (NpNp);
}
umLOG(1, "\n ==> {OP,MM} to sparse\n");
entries.reset(1, entries.hi()-Np*Np);
// Extract triplets from the large buffers. Note: this
// requires copying each array, and since these arrays
// can be HUGE(!), we force immediate deallocation:
Ai=OPi(entries); OPi.Free();
Aj=OPj(entries); OPj.Free();
Ax=OPx(entries); OPx.Free();
umLOG(1, " ==> triplets ready (OP) nnz = %10d\n", entries.hi());
// adjust triplet indices for 0-based sparse operators
Ai -= 1; Aj -= 1; MMi -= 1; MMj -= 1; int npk=Np*K;
#if defined(NDG_USE_CHOLMOD) || defined(NDG_New_CHOLINC)
// load only the lower triangle tril(OP) free args?
spOP.load(npk,npk, Ai,Aj,Ax, sp_LT, false,1e-15, true); // {LT, false} -> TriL
#else
// select {upper,lower,both} triangles
//spOP.load(npk,npk, Ai,Aj,Ax, sp_LT, true,1e-15,true); // LT -> enforce symmetry
//spOP.load(npk,npk, Ai,Aj,Ax, sp_All,true,1e-15,true); // All-> includes "noise"
//spOP.load(npk,npk, Ai,Aj,Ax, sp_UT, false,1e-15,true); // UT -> triu(OP) only
#endif
Ai.Free(); Aj.Free(); Ax.Free();
umLOG(1, " ==> triplets ready (MM) nnz = %10d\n", entriesMM.hi());
//-------------------------------------------------------
// The mass matrix operator will NOT be factorised,
// Load ALL elements (both upper and lower triangles):
//-------------------------------------------------------
spMM.load(npk,npk, MMi,MMj,MMx, sp_All,false,1.00e-15,true);
MMi.Free(); MMj.Free(); MMx.Free();
opti2 = timer.read(); // time assembly
umLOG(1, " ==> {OP,MM} converted to csc. (%g secs)\n", opti2-opti1);
}
void NDG2D::PoissonIPDGbc2D(
CSd& spOP //[out] sparse operator
)
{
// function [OP] = PoissonIPDGbc2D()
// Purpose: Set up the discrete Poisson matrix directly
// using LDG. The operator is set up in the weak form
// build DG derivative matrices
int max_OP = (K*Np*Np*(1+Nfaces));
//initialize parameters
DVec faceR("faceR"), faceS("faceS");
IVec Fm("Fm"), Fm1("Fm1"), fidM("fidM");
DMat V1D("V1D"); int i=0;
// build local face matrices
DMat massEdge[4]; // = zeros(Np,Np,Nfaces);
for (i=1; i<=Nfaces; ++i) {
massEdge[i].resize(Np,Np);
}
// face mass matrix 1
Fm = Fmask(All,1); faceR = r(Fm);
V1D = Vandermonde1D(N, faceR);
massEdge[1](Fm,Fm) = inv(V1D*trans(V1D));
// face mass matrix 2
Fm = Fmask(All,2); faceR = r(Fm);
V1D = Vandermonde1D(N, faceR);
massEdge[2](Fm,Fm) = inv(V1D*trans(V1D));
// face mass matrix 3
Fm = Fmask(All,3); faceS = s(Fm);
V1D = Vandermonde1D(N, faceS);
massEdge[3](Fm,Fm) = inv(V1D*trans(V1D));
//continue initialize parameters
DMat Dx("Dx"),Dy("Dy"), Dn1("Dn1"), mmE_Fm1("mmE(:,Fm1)");
double lnx=0.0,lny=0.0,lsJ=0.0,hinv=0.0,gtau=0.0;
int k1=0,f1=0,id=0;
IVec i1_Nfp = Range(1,Nfp);
double N1N1 = double((N+1)*(N+1));
// "OP" triplets (i,j,x), extracted to {Ai,Aj,Ax}
IVec OPi(max_OP),OPj(max_OP), Ai,Aj; DVec OPx(max_OP), Ax;
IMat rows1, cols1; Index1D entries; DMat OP11(Np,Nfp, 0.0);
// global node numbering
entries.reset(1,Np*Nfp);
cols1 = outer(Ones(Np), Range(1,Nfp));
umMSG(1, "\n ==> {OP} assembly [bc]: ");
for (k1=1; k1<=K; ++k1)
{
if (! (k1%100)) { umMSG(1, "%d, ",k1); }
rows1 = outer(Range((k1-1)*Np+1,k1*Np), Ones(Nfp));
// Build element-to-element parts of operator
for (f1=1; f1<=Nfaces; ++f1)
{
if (BCType(k1,f1))
{
////////////////////////added by Kevin ///////////////////////////////
Fm1 = Fmask(All,f1);
fidM = (k1-1)*Nfp*Nfaces + (f1-1)*Nfp + i1_Nfp;
id = 1+(f1-1)*Nfp + (k1-1)*Nfp*Nfaces;
lnx = nx(id); lny = ny(id);
lsJ = sJ(id); hinv = Fscale(id);
Dx = rx(1,k1)*Dr + sx(1,k1)*Ds;
Dy = ry(1,k1)*Dr + sy(1,k1)*Ds;
Dn1 = lnx*Dx + lny*Dy;
//mmE = lsJ*massEdge(:,:,f1);
//bc(All,k1) += (gtau*mmE(All,Fm1) - Dn1'*mmE(All,Fm1))*ubc(fidM);
mmE_Fm1 = massEdge[f1](All,Fm1); mmE_Fm1 *= lsJ;
gtau = 10*N1N1*hinv; // set penalty scaling
//bc(All,k1) += (gtau*mmE_Fm1 - trans(Dn1)*mmE_Fm1) * ubc(fidM);
switch(BCType(k1,f1)){
case BC_Dirichlet:
OP11 = gtau*mmE_Fm1 - trans(Dn1)*mmE_Fm1;
break;
case BC_Neuman:
OP11 = mmE_Fm1;
break;
default:
std::cout<<"warning: boundary condition is incorrect"<<std::endl;
}
OPi(entries)=rows1; OPj(entries)=cols1; OPx(entries)=OP11;
entries += (Np*Nfp);
}
cols1 += Nfp;
}
}
umMSG(1, "\n ==> {OPbc} to sparse\n");
entries.reset(1, entries.hi()-(Np*Nfp));
// extract triplets from large buffers
Ai=OPi(entries); Aj=OPj(entries); Ax=OPx(entries);
// These arrays can be HUGE, so force deallocation
OPi.Free(); OPj.Free(); OPx.Free();
// return 0-based sparse result
Ai -= 1; Aj -= 1;
//-------------------------------------------------------
// This operator is not symmetric, and will NOT be
// factorised, only used to create reference RHS's:
//
// refrhsbcPR = spOP1 * bcPR;
// refrhsbcUx = spOP2 * bcUx;
// refrhsbcUy = spOP2 * bcUy;
//
// Load ALL elements (both upper and lower triangles):
//-------------------------------------------------------
spOP.load(Np*K, Nfp*Nfaces*K, Ai,Aj,Ax, sp_All,false, 1e-15,true);
Ai.Free(); Aj.Free(); Ax.Free();
umMSG(1, " ==> {OPbc} ready.\n");
#if (1)
// check on original estimates for nnx
umMSG(1, " ==> max_OP: %12d\n", max_OP);
umMSG(1, " ==> nnz_OP: %12d\n", entries.hi());
#endif
}
