// updated
PetscErrorCode BSSCR_MatStokesMVBlockDefaultBuildScaling( MatStokesBlockScaling BA, Mat A, Vec b, Vec x, PetscTruth sym )
{
Mat K,G,D,C;
Vec rG;
PetscScalar rg2, rg, ra;
PetscInt N;
Vec rA, rC;
Vec L1,L2, R1,R2;
Mat S;
VecNestGetSubVec( BA->Lz, 0, &L1 );
VecNestGetSubVec( BA->Lz, 1, &L2 );
VecNestGetSubVec( BA->Rz, 0, &R1 );
VecNestGetSubVec( BA->Rz, 1, &R2 );
rA = L1;
rC = L2;
MatNestGetSubMat( A, 0,0, &K );
MatNestGetSubMat( A, 0,1, &G );
MatNestGetSubMat( A, 1,0, &D );
MatNestGetSubMat( A, 1,1, &C );
VecDuplicate( rA, &rG );
/* Get magnitude of K */
//px_MatGetAbsRowSum( K, rA );
//MatGetRowMax( K, rA, PETSC_NULL );
MatGetDiagonal( K, rA);
VecSqrt( rA );
VecReciprocal( rA );
/* VecDot( rA,rA, &ra ); */
/* VecGetSize( rA, &N ); */
/* ra = PetscSqrtScalar( ra/N ); */
/* Get magnitude of G */
//px_MatGetAbsRowSum( G, rG );
//MatGetRowMax( G, rG, PETSC_NULL );
Mat A21_cpy;
Mat Shat;
Vec diag; /* same as rA*rA */
MatGetVecs( K, &diag, PETSC_NULL );
MatGetDiagonal( K, diag );
VecReciprocal( diag );
if( sym )
#if( PETSC_VERSION_MAJOR <= 2 )
MatTranspose( G, &A21_cpy );
#else
MatTranspose( G, MAT_INITIAL_MATRIX, &A21_cpy );
#endif
else {
/*
RHSMatrixLaplacian - User-provided routine to compute the right-hand-side
matrix for the Laplacian operator
Input Parameters:
ts - the TS context
t - current time (ignored)
X - current solution (ignored)
dummy - optional user-defined context, as set by TSetRHSJacobian()
Output Parameters:
AA - Jacobian matrix
BB - optionally different matrix from which the preconditioner is built
str - flag indicating matrix structure
*/
PetscErrorCode RHSMatrixLaplaciangllDM(TS ts,PetscReal t,Vec X,Mat A,Mat BB,void *ctx)
{
PetscReal **temp;
PetscReal vv;
AppCtx *appctx = (AppCtx*)ctx; /* user-defined application context */
PetscErrorCode ierr;
PetscInt i,xs,xn,l,j;
PetscInt *rowsDM;
PetscBool flg = PETSC_FALSE;
ierr = PetscOptionsGetBool(NULL,NULL,"-gll_mf",&flg,NULL);CHKERRQ(ierr);
if (!flg) {
/*
Creates the element stiffness matrix for the given gll
*/
ierr = PetscGLLElementLaplacianCreate(&appctx->SEMop.gll,&temp);CHKERRQ(ierr);
/* workarround for clang analyzer warning: Division by zero */
if (appctx->param.N <= 1) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_ARG_WRONG,"Spectral element order should be > 1");
/* scale by the size of the element */
for (i=0; i<appctx->param.N; i++) {
vv=-appctx->param.mu*2.0/appctx->param.Le;
for (j=0; j<appctx->param.N; j++) temp[i][j]=temp[i][j]*vv;
}
ierr = MatSetOption(A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);
ierr = DMDAGetCorners(appctx->da,&xs,NULL,NULL,&xn,NULL,NULL);CHKERRQ(ierr);
xs = xs/(appctx->param.N-1);
xn = xn/(appctx->param.N-1);
ierr = PetscMalloc1(appctx->param.N,&rowsDM);CHKERRQ(ierr);
/*
loop over local elements
*/
for (j=xs; j<xs+xn; j++) {
for (l=0; l<appctx->param.N; l++) {
rowsDM[l] = 1+(j-xs)*(appctx->param.N-1)+l;
}
ierr = MatSetValuesLocal(A,appctx->param.N,rowsDM,appctx->param.N,rowsDM,&temp[0][0],ADD_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree(rowsDM);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = VecReciprocal(appctx->SEMop.mass);CHKERRQ(ierr);
ierr = MatDiagonalScale(A,appctx->SEMop.mass,0);CHKERRQ(ierr);
ierr = VecReciprocal(appctx->SEMop.mass);CHKERRQ(ierr);
ierr = PetscGLLElementLaplacianDestroy(&appctx->SEMop.gll,&temp);CHKERRQ(ierr);
} else {
ierr = MatSetType(A,MATSHELL);CHKERRQ(ierr);
ierr = MatSetUp(A);CHKERRQ(ierr);
ierr = MatShellSetContext(A,appctx);CHKERRQ(ierr);
ierr = MatShellSetOperation(A,MATOP_MULT,(void (*)(void))MatMult_Laplacian);CHKERRQ(ierr);
}
return 0;
}
PetscErrorCode StokesSetupApproxSchur(Stokes *s)
{
Vec diag;
PetscErrorCode ierr;
PetscFunctionBeginUser;
/* Schur complement approximation: myS = A11 - A10 diag(A00)^(-1) A01 */
/* note: A11 is zero */
/* note: in real life this matrix would be build directly, */
/* i.e. without MatMatMult */
/* inverse of diagonal of A00 */
ierr = VecCreate(PETSC_COMM_WORLD,&diag);CHKERRQ(ierr);
ierr = VecSetSizes(diag,PETSC_DECIDE,2*s->nx*s->ny);CHKERRQ(ierr);
ierr = VecSetType(diag,VECMPI);CHKERRQ(ierr);
ierr = MatGetDiagonal(s->subA[0],diag);
ierr = VecReciprocal(diag);
/* compute: - A10 diag(A00)^(-1) A01 */
ierr = MatDiagonalScale(s->subA[1],diag,NULL); /* (*warning* overwrites subA[1]) */
ierr = MatMatMult(s->subA[2],s->subA[1],MAT_INITIAL_MATRIX,PETSC_DEFAULT,&s->myS);CHKERRQ(ierr);
ierr = MatScale(s->myS,-1.0);CHKERRQ(ierr);
/* restore A10 */
ierr = MatGetDiagonal(s->subA[0],diag);
ierr = MatDiagonalScale(s->subA[1],diag,NULL);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode bsscr_GMiGt( Mat *_K2, Mat K, Mat G, Mat M){
Mat K2;
Vec Mdiag;
Mat MinvGt;
Mat Gtrans;
PetscErrorCode ierr;
PetscFunctionBegin;
MatGetVecs( M, &Mdiag, PETSC_NULL );
MatGetDiagonal( M, Mdiag );
VecReciprocal(Mdiag);
#if( PETSC_VERSION_MAJOR <= 2 )
ierr=MatTranspose(G, &Gtrans);CHKERRQ(ierr);
#else
ierr=MatTranspose(G, MAT_INITIAL_MATRIX,&Gtrans);CHKERRQ(ierr);
#endif
ierr=MatConvert(Gtrans, MATSAME, MAT_INITIAL_MATRIX, &MinvGt);CHKERRQ(ierr);/* copy Gtrans -> MinvGt */
MatDiagonalScale(MinvGt, Mdiag, PETSC_NULL);/* Minv*Gtrans */
/* MAT_INITIAL_MATRIX -> creates K2 matrix : PETSC_DEFAULT for fill ratio: run with -info to find what it should be*/
ierr=MatMatMult( G, MinvGt, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &K2);CHKERRQ(ierr);/* K2 = G*Minv*Gtrans */
Stg_MatDestroy(&Gtrans);
Stg_MatDestroy(&MinvGt);
Stg_VecDestroy(&Mdiag);
*_K2=K2;
PetscFunctionReturn(0);
}
void FETI_Operations::set_jacobi_precond_vector()
{
homemade_assert_msg(m_bC_RR_MatrixSet,"Preconditioner matrix not set yet!");
// Create and set the vector
VecCreate(m_comm.get(),&m_coupling_jacobi_precond_vec);
VecSetSizes(m_coupling_jacobi_precond_vec,m_C_RR_M_local,m_C_RR_M);
VecSetFromOptions(m_coupling_jacobi_precond_vec);
// Get the diagonal
MatGetDiagonal(m_C_RR,m_coupling_jacobi_precond_vec);
// Calculate the reciprocal
VecReciprocal(m_coupling_jacobi_precond_vec);
// Export it
write_PETSC_vector(m_coupling_jacobi_precond_vec,m_scratch_folder_path + "/precond_Jacobi_vector.petscvec",m_comm.rank(),m_comm.get());
// Print MatLab debugging output? Variable defined at "carl_headers.h"
#ifdef PRINT_MATLAB_DEBUG
write_PETSC_vector_MATLAB(m_coupling_jacobi_precond_vec,m_scratch_folder_path + "/precond_Jacobi_vector.m",m_comm.get());
#endif
// Set flag
m_bCreatedPrecondJacobiVec = true;
}
/*@
PCMGGetRScale - Gets the pointwise scaling for the restriction operator from level l to l-1.
Collective on PC
Input Parameters:
+ pc - the multigrid context
. rscale - the scaling
- l - the level (0 is coarsest) to supply [Do not supply 0]
Level: advanced
Notes:
When evaluating a function on a coarse level one does not want to do F(R * x) one does F(rscale * R * x) where rscale is 1 over the row sums of R.
.keywords: MG, set, multigrid, restriction, level
.seealso: PCMGSetInterpolation(), PCMGGetRestriction()
@*/
PetscErrorCode PCMGGetRScale(PC pc,PetscInt l,Vec *rscale)
{
PetscErrorCode ierr;
PC_MG *mg = (PC_MG*)pc->data;
PC_MG_Levels **mglevels = mg->levels;
PetscFunctionBegin;
PetscValidHeaderSpecific(pc,PC_CLASSID,1);
if (!mglevels) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_WRONGSTATE,"Must set MG levels before calling");
if (l <= 0 || mg->nlevels <= l) SETERRQ2(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_OUTOFRANGE,"Level %D must be in range {1,...,%D}",l,mg->nlevels-1);
if (!mglevels[l]->rscale) {
Mat R;
Vec X,Y,coarse,fine;
PetscInt M,N;
ierr = PCMGGetRestriction(pc,l,&R);CHKERRQ(ierr);
ierr = MatCreateVecs(R,&X,&Y);CHKERRQ(ierr);
ierr = MatGetSize(R,&M,&N);CHKERRQ(ierr);
if (M < N) {
fine = X;
coarse = Y;
} else if (N < M) {
fine = Y; coarse = X;
} else SETERRQ(PetscObjectComm((PetscObject)R),PETSC_ERR_SUP,"Restriction matrix is square, cannot determine which Vec is coarser");
ierr = VecSet(fine,1.);CHKERRQ(ierr);
ierr = MatRestrict(R,fine,coarse);CHKERRQ(ierr);
ierr = VecDestroy(&fine);CHKERRQ(ierr);
ierr = VecReciprocal(coarse);CHKERRQ(ierr);
mglevels[l]->rscale = coarse;
}
*rscale = mglevels[l]->rscale;
PetscFunctionReturn(0);
}
static PetscErrorCode PCSetUp_LSC(PC pc)
{
PC_LSC *lsc = (PC_LSC*)pc->data;
Mat L,Lp,B,C;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PCLSCAllocate_Private(pc);CHKERRQ(ierr);
ierr = PetscObjectQuery((PetscObject)pc->mat,"LSC_L",(PetscObject*)&L);CHKERRQ(ierr);
if (!L) {ierr = PetscObjectQuery((PetscObject)pc->pmat,"LSC_L",(PetscObject*)&L);CHKERRQ(ierr);}
ierr = PetscObjectQuery((PetscObject)pc->pmat,"LSC_Lp",(PetscObject*)&Lp);CHKERRQ(ierr);
if (!Lp) {ierr = PetscObjectQuery((PetscObject)pc->mat,"LSC_Lp",(PetscObject*)&Lp);CHKERRQ(ierr);}
if (!L) {
ierr = MatSchurComplementGetSubMatrices(pc->mat,NULL,NULL,&B,&C,NULL);CHKERRQ(ierr);
if (!lsc->L) {
ierr = MatMatMult(C,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&lsc->L);CHKERRQ(ierr);
} else {
ierr = MatMatMult(C,B,MAT_REUSE_MATRIX,PETSC_DEFAULT,&lsc->L);CHKERRQ(ierr);
}
Lp = L = lsc->L;
}
if (lsc->scale) {
Mat Ap;
ierr = MatSchurComplementGetSubMatrices(pc->mat,NULL,&Ap,NULL,NULL,NULL);CHKERRQ(ierr);
ierr = MatGetDiagonal(Ap,lsc->scale);CHKERRQ(ierr); /* Should be the mass matrix, but we don't have plumbing for that yet */
ierr = VecReciprocal(lsc->scale);CHKERRQ(ierr);
}
ierr = KSPSetOperators(lsc->kspL,L,Lp);CHKERRQ(ierr);
ierr = KSPSetFromOptions(lsc->kspL);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
void PetscVector<T>::reciprocal()
{
PetscErrorCode ierr = 0;
// VecReciprocal has been in PETSc since at least 2.3.3 days
ierr = VecReciprocal(_vec);
CHKERRABORT(libMesh::COMM_WORLD,ierr);
}
PetscErrorCode BSSCR_PCScGtKGUseStandardScaling( PC pc )
{
PC_SC_GtKG ctx = (PC_SC_GtKG)pc->data;
Mat K,G,D,C;
Vec rG;
PetscScalar rg2, rg, ra;
PetscInt N;
Vec rA, rC;
Vec L1,L2, R1,R2;
BSSCR_BSSCR_pc_error_ScGtKG( pc, __func__ );
L1 = ctx->X1;
L2 = ctx->X2;
R1 = ctx->Y1;
R2 = ctx->Y2;
rA = L1;
rC = L2;
K = ctx->F;
G = ctx->Bt;
D = ctx->B;
C = ctx->C;
VecDuplicate( rA, &rG );
/* Get magnitude of K */
MatGetRowMax( K, rA, PETSC_NULL );
VecSqrt( rA );
VecReciprocal( rA );
VecDot( rA,rA, &ra );
VecGetSize( rA, &N );
ra = PetscSqrtScalar( ra/N );
/* Get magnitude of G */
MatGetRowMax( G, rG, PETSC_NULL );
VecDot( rG, rG, &rg2 );
VecGetSize( rG, &N );
rg = PetscSqrtScalar(rg2/N);
// printf("rg = %f \n", rg );
VecSet( rC, 1.0/(rg*ra) );
Stg_VecDestroy(&rG );
VecCopy( L1, R1 );
VecCopy( L2, R2 );
PetscFunctionReturn(0);
}
/*
RHSMatrixAdvection - User-provided routine to compute the right-hand-side
matrix for the Advection (gradient) operator.
Input Parameters:
ts - the TS context
t - current time
global_in - global input vector
dummy - optional user-defined context, as set by TSetRHSJacobian()
Output Parameters:
AA - Jacobian matrix
BB - optionally different preconditioning matrix
str - flag indicating matrix structure
*/
PetscErrorCode RHSMatrixAdvectiongllDM(TS ts,PetscReal t,Vec X,Mat A,Mat BB,void *ctx)
{
PetscReal **temp;
AppCtx *appctx = (AppCtx*)ctx; /* user-defined application context */
PetscErrorCode ierr;
PetscInt xs,xn,l,j;
PetscInt *rowsDM;
PetscBool flg = PETSC_FALSE;
ierr = PetscOptionsGetBool(NULL,NULL,"-gll_mf",&flg,NULL);CHKERRQ(ierr);
if (!flg) {
/*
Creates the advection matrix for the given gll
*/
ierr = PetscGLLElementAdvectionCreate(&appctx->SEMop.gll,&temp);CHKERRQ(ierr);
ierr = MatSetOption(A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr);
ierr = DMDAGetCorners(appctx->da,&xs,NULL,NULL,&xn,NULL,NULL);CHKERRQ(ierr);
xs = xs/(appctx->param.N-1);
xn = xn/(appctx->param.N-1);
ierr = PetscMalloc1(appctx->param.N,&rowsDM);CHKERRQ(ierr);
for (j=xs; j<xs+xn; j++) {
for (l=0; l<appctx->param.N; l++) {
rowsDM[l] = 1+(j-xs)*(appctx->param.N-1)+l;
}
ierr = MatSetValuesLocal(A,appctx->param.N,rowsDM,appctx->param.N,rowsDM,&temp[0][0],ADD_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree(rowsDM);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = VecReciprocal(appctx->SEMop.mass);CHKERRQ(ierr);
ierr = MatDiagonalScale(A,appctx->SEMop.mass,0);CHKERRQ(ierr);
ierr = VecReciprocal(appctx->SEMop.mass);CHKERRQ(ierr);
ierr = PetscGLLElementAdvectionDestroy(&appctx->SEMop.gll,&temp);CHKERRQ(ierr);
} else {
ierr = MatSetType(A,MATSHELL);CHKERRQ(ierr);
ierr = MatSetUp(A);CHKERRQ(ierr);
ierr = MatShellSetContext(A,appctx);CHKERRQ(ierr);
ierr = MatShellSetOperation(A,MATOP_MULT,(void (*)(void))MatMult_Advection);CHKERRQ(ierr);
}
return 0;
}
PetscErrorCode PCGAMGOptProl_Classical_Jacobi(PC pc,const Mat A,Mat *P)
{
PetscErrorCode ierr;
PetscInt f,s,n,cf,cs,i,idx;
PetscInt *coarserows;
PetscInt ncols;
const PetscInt *pcols;
const PetscScalar *pvals;
Mat Pnew;
Vec diag;
PC_MG *mg = (PC_MG*)pc->data;
PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx;
PC_GAMG_Classical *cls = (PC_GAMG_Classical*)pc_gamg->subctx;
PetscFunctionBegin;
if (cls->nsmooths == 0) {
ierr = PCGAMGTruncateProlongator_Private(pc,P);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ierr = MatGetOwnershipRange(*P,&s,&f);CHKERRQ(ierr);
n = f-s;
ierr = MatGetOwnershipRangeColumn(*P,&cs,&cf);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscInt)*n,&coarserows);CHKERRQ(ierr);
/* identify the rows corresponding to coarse unknowns */
idx = 0;
for (i=s;i<f;i++) {
ierr = MatGetRow(*P,i,&ncols,&pcols,&pvals);CHKERRQ(ierr);
/* assume, for now, that it's a coarse unknown if it has a single unit entry */
if (ncols == 1) {
if (pvals[0] == 1.) {
coarserows[idx] = i;
idx++;
}
}
ierr = MatRestoreRow(*P,i,&ncols,&pcols,&pvals);CHKERRQ(ierr);
}
ierr = MatGetVecs(A,&diag,0);CHKERRQ(ierr);
ierr = MatGetDiagonal(A,diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
for (i=0;i<cls->nsmooths;i++) {
ierr = MatMatMult(A,*P,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&Pnew);CHKERRQ(ierr);
ierr = MatZeroRows(Pnew,idx,coarserows,0.,NULL,NULL);CHKERRQ(ierr);
ierr = MatDiagonalScale(Pnew,diag,0);CHKERRQ(ierr);
ierr = MatAYPX(Pnew,-1.0,*P,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
ierr = MatDestroy(P);CHKERRQ(ierr);
*P = Pnew;
Pnew = NULL;
}
ierr = VecDestroy(&diag);CHKERRQ(ierr);
ierr = PetscFree(coarserows);CHKERRQ(ierr);
ierr = PCGAMGTruncateProlongator_Private(pc,P);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
void N_VInv_Petsc(N_Vector x, N_Vector z)
{
Vec *xv = NV_PVEC_PTC(x);
Vec *zv = NV_PVEC_PTC(z);
if(z != x)
VecCopy(*xv, *zv); /* copy x~>z */
VecReciprocal(*zv);
return;
}
// updated
PetscErrorCode BSSCR_mat_mvblock_invert_scalings( MatStokesBlockScaling BA )
{
Vec L1,L2, R1,R2;
VecNestGetSubVec( BA->Lz, 0, &L1 );
VecNestGetSubVec( BA->Lz, 1, &L2 );
VecNestGetSubVec( BA->Rz, 0, &R1 );
VecNestGetSubVec( BA->Rz, 1, &R2 );
VecReciprocal(L1);
VecReciprocal(L2);
VecReciprocal(R1);
VecReciprocal(R2);
/* toggle inversion flag */
if( BA->scalings_have_been_inverted == PETSC_TRUE ) { BA->scalings_have_been_inverted = PETSC_FALSE; }
if( BA->scalings_have_been_inverted == PETSC_FALSE ) { BA->scalings_have_been_inverted = PETSC_TRUE; }
PetscFunctionReturn(0);
}
static PetscErrorCode VecReciprocal_Nest(Vec x)
{
Vec_Nest *bx = (Vec_Nest*)x->data;
PetscInt i,nr;
PetscErrorCode ierr;
PetscFunctionBegin;
nr = bx->nb;
for (i=0; i<nr; i++) {
ierr = VecReciprocal(bx->v[i]);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
/*@
MatCreateSchurComplementPmat - create a preconditioning matrix for the Schur complement by assembling Sp = A11 - A10 inv(diag(A00)) A01
Collective on Mat
Input Parameters:
+ A00,A01,A10,A11 - the four parts of the original matrix A = [A00 A01; A10 A11] (A01,A10, and A11 are optional, implying zero matrices)
. ainvtype - type of approximation for inv(A00) used when forming Sp = A11 - A10 inv(A00) A01
- preuse - MAT_INITIAL_MATRIX for a new Sp, or MAT_REUSE_MATRIX to reuse an existing Sp, or MAT_IGNORE_MATRIX to put nothing in Sp
Output Parameter:
- Spmat - approximate Schur complement suitable for preconditioning S = A11 - A10 inv(diag(A00)) A01
Note:
Since the real Schur complement is usually dense, providing a good approximation to newpmat usually requires
application-specific information. The default for assembled matrices is to use the inverse of the diagonal of
the (0,0) block A00 in place of A00^{-1}. This rarely produce a scalable algorithm. Optionally, A00 can be lumped
before forming inv(diag(A00)).
Level: advanced
Concepts: matrices^submatrices
.seealso: MatCreateSchurComplement(), MatGetSchurComplement(), MatSchurComplementGetPmat(), MatSchurComplementAinvType
@*/
PetscErrorCode MatCreateSchurComplementPmat(Mat A00,Mat A01,Mat A10,Mat A11,MatSchurComplementAinvType ainvtype,MatReuse preuse,Mat *Spmat)
{
PetscErrorCode ierr;
PetscInt N00;
PetscFunctionBegin;
/* Use an appropriate approximate inverse of A00 to form A11 - A10 inv(diag(A00)) A01; a NULL A01, A10 or A11 indicates a zero matrix. */
/* TODO: Perhaps should create an appropriately-sized zero matrix of the same type as A00? */
if ((!A01 || !A10) & !A11) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot assemble Spmat: A01, A10 and A11 are all NULL.");
if (preuse == MAT_IGNORE_MATRIX) PetscFunctionReturn(0);
/* A zero size A00 or empty A01 or A10 imply S = A11. */
ierr = MatGetSize(A00,&N00,NULL);CHKERRQ(ierr);
if (!A01 || !A10 || !N00) {
if (preuse == MAT_INITIAL_MATRIX) {
ierr = MatDuplicate(A11,MAT_COPY_VALUES,Spmat);CHKERRQ(ierr);
} else { /* MAT_REUSE_MATRIX */
/* TODO: when can we pass SAME_NONZERO_PATTERN? */
ierr = MatCopy(A11,*Spmat,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
}
} else {
Mat AdB;
Vec diag;
ierr = MatCreateVecs(A00,&diag,NULL);CHKERRQ(ierr);
if (ainvtype == MAT_SCHUR_COMPLEMENT_AINV_LUMP) {
ierr = MatGetRowSum(A00,diag);CHKERRQ(ierr);
} else if (ainvtype == MAT_SCHUR_COMPLEMENT_AINV_DIAG) {
ierr = MatGetDiagonal(A00,diag);CHKERRQ(ierr);
} else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown MatSchurComplementAinvType: %D", ainvtype);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = MatDuplicate(A01,MAT_COPY_VALUES,&AdB);CHKERRQ(ierr);
ierr = MatDiagonalScale(AdB,diag,NULL);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
/* Cannot really reuse Spmat in MatMatMult() because of MatAYPX() -->
MatAXPY() --> MatHeaderReplace() --> MatDestroy_XXX_MatMatMult() */
ierr = MatDestroy(Spmat);CHKERRQ(ierr);
ierr = MatMatMult(A10,AdB,MAT_INITIAL_MATRIX,PETSC_DEFAULT,Spmat);CHKERRQ(ierr);
if (!A11) {
ierr = MatScale(*Spmat,-1.0);CHKERRQ(ierr);
} else {
/* TODO: when can we pass SAME_NONZERO_PATTERN? */
ierr = MatAYPX(*Spmat,-1,A11,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
}
ierr = MatDestroy(&AdB);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
/*
SampleShellPCSetUp - This routine sets up a user-defined
preconditioner context.
Input Parameters:
. pc - preconditioner object
. pmat - preconditioner matrix
. x - vector
Output Parameter:
. shell - fully set up user-defined preconditioner context
Notes:
In this example, we define the shell preconditioner to be Jacobi's
method. Thus, here we create a work vector for storing the reciprocal
of the diagonal of the preconditioner matrix; this vector is then
used within the routine SampleShellPCApply().
*/
PetscErrorCode SampleShellPCSetUp(PC pc,Mat pmat,Vec x)
{
SampleShellPC *shell;
Vec diag;
PetscErrorCode ierr;
ierr = PCShellGetContext(pc,(void**)&shell);CHKERRQ(ierr);
ierr = VecDuplicate(x,&diag);CHKERRQ(ierr);
ierr = MatGetDiagonal(pmat,diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
shell->diag = diag;
return 0;
}
/*
I should not modify setup called!!
This is handled via petsc.
*/
PetscErrorCode BSSCR_PCSetUp_GtKG( PC pc )
{
PC_GtKG ctx = (PC_GtKG)pc->data;
PetscReal fill;
Mat Ident;
Vec diag;
PetscInt M,N, m,n;
MPI_Comm comm;
PetscInt nnz_I, nnz_G;
const MatType mtype;
const char *prefix;
PetscTruth wasSetup;
if( ctx->K == PETSC_NULL ) {
Stg_SETERRQ( PETSC_ERR_SUP, "gtkg: K not set" );
}
if( ctx->G == PETSC_NULL ) {
Stg_SETERRQ( PETSC_ERR_SUP, "gtkg: G not set" );
}
PetscObjectGetComm( (PetscObject)ctx->K, &comm );
/* Check for existence of objects and trash any which exist */
if( ctx->form_GtG == PETSC_TRUE && ctx->GtG != PETSC_NULL ) {
Stg_MatDestroy(&ctx->GtG );
ctx->GtG = PETSC_NULL;
}
if( ctx->s != PETSC_NULL ) {
Stg_VecDestroy(&ctx->s );
ctx->s = PETSC_NULL;
}
if( ctx->X != PETSC_NULL ) {
Stg_VecDestroy(&ctx->X );
ctx->X = PETSC_NULL;
}
if( ctx->t != PETSC_NULL ) {
Stg_VecDestroy(&ctx->t );
ctx->t = PETSC_NULL;
}
if( ctx->inv_diag_M != PETSC_NULL ) {
Stg_VecDestroy(&ctx->inv_diag_M );
ctx->inv_diag_M = PETSC_NULL;
}
/* Create vectors */
MatGetVecs( ctx->K, &ctx->s, &ctx->X );
MatGetVecs( ctx->G, &ctx->t, PETSC_NULL );
if( ctx->M != PETSC_NULL ) {
MatGetVecs( ctx->K, &ctx->inv_diag_M, PETSC_NULL );
MatGetDiagonal( ctx->M, ctx->inv_diag_M );
VecReciprocal( ctx->inv_diag_M );
/* change the pc_apply routines */
pc->ops->apply = BSSCR_BSSCR_PCApply_GtKG_diagonal_scaling;
pc->ops->applytranspose = BSSCR_BSSCR_PCApplyTranspose_GtKG_diagonal_scaling;
}
/* Assemble GtG */
MatGetSize( ctx->G, &M, &N );
MatGetLocalSize( ctx->G, &m, &n );
MatGetVecs( ctx->G, PETSC_NULL, &diag );
VecSet( diag, 1.0 );
MatCreate( comm, &Ident );
MatSetSizes( Ident, m,m , M, M );
#if (((PETSC_VERSION_MAJOR==3) && (PETSC_VERSION_MINOR>=3)) || (PETSC_VERSION_MAJOR>3) )
MatSetUp(Ident);
#endif
MatGetType( ctx->G, &mtype );
MatSetType( Ident, mtype );
if( ctx->M == PETSC_NULL ) {
MatDiagonalSet( Ident, diag, INSERT_VALUES );
}
else {
MatDiagonalSet( Ident, ctx->inv_diag_M, INSERT_VALUES );
}
BSSCR_get_number_nonzeros_AIJ( Ident, &nnz_I );
BSSCR_get_number_nonzeros_AIJ( ctx->G, &nnz_G );
//fill = 1.0;
/*
Not sure the best way to estimate the fill factor.
GtG is a laplacian on the pressure space.
This might tell us something useful...
*/
fill = (PetscReal)(nnz_G)/(PetscReal)( nnz_I );
MatPtAP( Ident, ctx->G, MAT_INITIAL_MATRIX, fill, &ctx->GtG );
Stg_MatDestroy(&Ident);
Stg_VecDestroy(&diag );
Stg_KSPSetOperators( ctx->ksp, ctx->GtG, ctx->GtG, SAME_NONZERO_PATTERN );
if (!pc->setupcalled) {
wasSetup = PETSC_FALSE;
PCGetOptionsPrefix( pc,&prefix );
KSPSetOptionsPrefix( ctx->ksp, prefix );
KSPAppendOptionsPrefix( ctx->ksp, "pc_gtkg_" ); /* -pc_GtKG_ksp_type <type>, -ksp_GtKG_pc_type <type> */
}
else {
wasSetup = PETSC_TRUE;
}
// if (!wasSetup && pc->setfromoptionscalled) {
if (!wasSetup) {
KSPSetFromOptions(ctx->ksp);
}
PetscFunctionReturn(0);
}
/* Developer Notes: This should be implemented with a MatCreate_SchurComplement() as that is the standard design for new Mat classes. */
PetscErrorCode MatGetSchurComplement_Basic(Mat mat,IS isrow0,IS iscol0,IS isrow1,IS iscol1,MatReuse mreuse,Mat *newmat,MatReuse preuse,Mat *newpmat)
{
PetscErrorCode ierr;
Mat A=0,Ap=0,B=0,C=0,D=0;
PetscFunctionBegin;
PetscValidHeaderSpecific(mat,MAT_CLASSID,1);
PetscValidHeaderSpecific(isrow0,IS_CLASSID,2);
PetscValidHeaderSpecific(iscol0,IS_CLASSID,3);
PetscValidHeaderSpecific(isrow1,IS_CLASSID,4);
PetscValidHeaderSpecific(iscol1,IS_CLASSID,5);
if (mreuse == MAT_REUSE_MATRIX) PetscValidHeaderSpecific(*newmat,MAT_CLASSID,7);
if (preuse == MAT_REUSE_MATRIX) PetscValidHeaderSpecific(*newpmat,MAT_CLASSID,9);
PetscValidType(mat,1);
if (mat->factortype) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
if (mreuse != MAT_IGNORE_MATRIX) {
/* Use MatSchurComplement */
if (mreuse == MAT_REUSE_MATRIX) {
ierr = MatSchurComplementGetSubmatrices(*newmat,&A,&Ap,&B,&C,&D);CHKERRQ(ierr);
if (!A || !Ap || !B || !C) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Attempting to reuse matrix but Schur complement matrices unset");
if (A != Ap) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Preconditioning matrix does not match operator");
ierr = MatDestroy(&Ap);CHKERRQ(ierr); /* get rid of extra reference */
}
ierr = MatGetSubMatrix(mat,isrow0,iscol0,mreuse,&A);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow0,iscol1,mreuse,&B);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow1,iscol0,mreuse,&C);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow1,iscol1,mreuse,&D);CHKERRQ(ierr);
switch (mreuse) {
case MAT_INITIAL_MATRIX:
ierr = MatCreateSchurComplement(A,A,B,C,D,newmat);CHKERRQ(ierr);
break;
case MAT_REUSE_MATRIX:
ierr = MatSchurComplementUpdate(*newmat,A,A,B,C,D,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
break;
default:
SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_SUP,"Unrecognized value of mreuse");
}
}
if (preuse != MAT_IGNORE_MATRIX) {
/* Use the diagonal part of A to form D - C inv(diag(A)) B */
Mat Ad,AdB,S;
Vec diag;
PetscInt i,m,n,mstart,mend;
PetscScalar *x;
/* We could compose these with newpmat so that the matrices can be reused. */
if (!A) {ierr = MatGetSubMatrix(mat,isrow0,iscol0,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr);}
if (!B) {ierr = MatGetSubMatrix(mat,isrow0,iscol1,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr);}
if (!C) {ierr = MatGetSubMatrix(mat,isrow1,iscol0,MAT_INITIAL_MATRIX,&C);CHKERRQ(ierr);}
if (!D) {ierr = MatGetSubMatrix(mat,isrow1,iscol1,MAT_INITIAL_MATRIX,&D);CHKERRQ(ierr);}
ierr = MatGetVecs(A,&diag,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetDiagonal(A,diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr);
/* We need to compute S = D - C inv(diag(A)) B. For row-oriented formats, it is easy to scale the rows of B and
* for column-oriented formats the columns of C can be scaled. Would skip creating a silly diagonal matrix. */
ierr = MatCreate(((PetscObject)A)->comm,&Ad);CHKERRQ(ierr);
ierr = MatSetSizes(Ad,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetOptionsPrefix(Ad,((PetscObject)mat)->prefix);CHKERRQ(ierr);
ierr = MatAppendOptionsPrefix(Ad,"diag_");CHKERRQ(ierr);
ierr = MatSetFromOptions(Ad);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(Ad,1,PETSC_NULL);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(Ad,1,PETSC_NULL,0,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Ad,&mstart,&mend);CHKERRQ(ierr);
ierr = VecGetArray(diag,&x);CHKERRQ(ierr);
for (i=mstart; i<mend; i++) {
ierr = MatSetValue(Ad,i,i,x[i-mstart],INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecRestoreArray(diag,&x);CHKERRQ(ierr);
ierr = MatAssemblyBegin(Ad,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Ad,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
ierr = MatMatMult(Ad,B,MAT_INITIAL_MATRIX,1,&AdB);CHKERRQ(ierr);
S = (preuse == MAT_REUSE_MATRIX) ? *newpmat : (Mat)0;
ierr = MatMatMult(C,AdB,preuse,PETSC_DEFAULT,&S);CHKERRQ(ierr);
ierr = MatAYPX(S,-1,D,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
*newpmat = S;
ierr = MatDestroy(&Ad);CHKERRQ(ierr);
ierr = MatDestroy(&AdB);CHKERRQ(ierr);
}
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&D);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode BSSCR_DRIVER_auglag( KSP ksp, Mat stokes_A, Vec stokes_x, Vec stokes_b, Mat approxS,
MatStokesBlockScaling BA, PetscTruth sym, KSP_BSSCR * bsscrp_self )
{
AugLagStokes_SLE * stokesSLE = (AugLagStokes_SLE*)bsscrp_self->solver->st_sle;
PetscTruth uzawastyle, KisJustK=PETSC_TRUE, restorek, change_A11rhspresolve;
PetscTruth usePreviousGuess, useNormInfStoppingConditions, useNormInfMonitor, found, forcecorrection;
PetscTruth change_backsolve, mg_active;
PetscErrorCode ierr;
//PetscInt monitor_index;
PetscInt max_it,min_it;
KSP ksp_inner, ksp_S, ksp_new_inner;
PC pc_S, pcInner;
Mat K,G,D,C, S, K2;// Korig;
Vec u,p,f,f2=0,f3=0,h, h_hat,t;
MGContext mgCtx;
double mgSetupTime, scrSolveTime, a11SingleSolveTime, penaltyNumber;// hFactor;
static int been_here = 0; /* Ha Ha Ha !! */
char name[PETSC_MAX_PATH_LEN];
char matname[PETSC_MAX_PATH_LEN];
char suffix[PETSC_MAX_PATH_LEN];
char str[PETSC_MAX_PATH_LEN];
PetscTruth flg, extractMats;
PetscLogDouble flopsA,flopsB;
/***************************************************************************************************************/
/***************************************************************************************************************/
//if( bsscrp_self->solver->st_sle->context->loadFromCheckPoint ){
// been_here=1;
//}
/* get sub matrix / vector objects */
/* note that here, the matrix D should always exist. It is set up in _StokesBlockKSPInterface_Solve in StokesBlockKSPInterface.c */
/* now extract K,G etc from a MatNest object */
MatNestGetSubMat( stokes_A, 0,0, &K );
MatNestGetSubMat( stokes_A, 0,1, &G );
MatNestGetSubMat( stokes_A, 1,0, &D );if(!D){ PetscPrintf( PETSC_COMM_WORLD, "D does not exist but should!!\n"); exit(1); }
MatNestGetSubMat( stokes_A, 1,1, &C );
VecNestGetSubVec( stokes_x, 0, &u );
VecNestGetSubVec( stokes_x, 1, &p );
VecNestGetSubVec( stokes_b, 0, &f );
VecNestGetSubVec( stokes_b, 1, &h );
PetscPrintf( PETSC_COMM_WORLD, "\n\n---------- AUGMENTED LAGRANGIAN K2 METHOD ---------\n\n" );
PetscPrintf( PETSC_COMM_WORLD, "----------- Penalty = %f\n\n", bsscrp_self->solver->penaltyNumber );
sprintf(suffix,"%s","x");
PetscOptionsGetString( PETSC_NULL, "-matsuffix", suffix, PETSC_MAX_PATH_LEN-1, &extractMats );
flg=0;
PetscOptionsGetString( PETSC_NULL, "-matdumpdir", name, PETSC_MAX_PATH_LEN-1, &flg );
if(flg){
sprintf(str,"%s/",name); sprintf(matname,"K%s",suffix);
bsscr_dirwriteMat( K, matname,str, "Writing K matrix in al Solver");
sprintf(str,"%s/",name); sprintf(matname,"G%s",suffix);
bsscr_dirwriteMat( G, matname,str, "Writing G matrix in al Solver");
sprintf(str,"%s/",name); sprintf(matname,"D%s",suffix);
bsscr_dirwriteMat( D, matname,str, "Writing D matrix in al Solver");
sprintf(str,"%s/",name); sprintf(matname,"f%s",suffix);
bsscr_dirwriteVec( f, matname,str, "Writing f vector in al Solver");
sprintf(str,"%s/",name); sprintf(matname,"h%s",suffix);
bsscr_dirwriteVec( h, matname,str, "Writing h vector in al Solver");
sprintf(str,"%s/",name); sprintf(matname,"Shat%s",suffix);
bsscr_dirwriteMat( approxS, matname,str, "Writing Shat matrix in al Solver");
if(C){
sprintf(str,"%s/",name); sprintf(matname,"C%s",suffix);
bsscr_dirwriteMat( C, matname,str, "Writing C matrix in al Solver");
}
}
mg_active=PETSC_TRUE;
PetscOptionsGetTruth( PETSC_NULL ,"-A11_mg_active", &mg_active, &found );
bsscrp_self->solver->mg_active=mg_active;
penaltyNumber = bsscrp_self->solver->penaltyNumber;
//hFactor = stokesSLE->hFactor;
/***************************************************************************************************************/
/***************************************************************************************************************/
/****** GET K2 ****************************************************************************************/
if(penaltyNumber > 1e-10 && bsscrp_self->k2type){
flg=0;
PetscOptionsGetString( PETSC_NULL, "-matdumpdir", name, PETSC_MAX_PATH_LEN-1, &flg );
if(flg){
sprintf(str,"%s/",name); sprintf(matname,"K2%s",suffix);
bsscr_dirwriteMat( bsscrp_self->K2, matname,str, "Writing K2 matrix in al Solver");
}
K2=bsscrp_self->K2;
scrSolveTime = MPI_Wtime();
ierr=MatAXPY(K,penaltyNumber,K2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);/* Computes K = penaltyNumber*K2 + K */
scrSolveTime = MPI_Wtime() - scrSolveTime;
PetscPrintf( PETSC_COMM_WORLD, "\n\t* K+p*K2 in time: %lf seconds\n\n", scrSolveTime);
KisJustK=PETSC_FALSE;
forcecorrection=PETSC_TRUE;
PetscOptionsGetTruth( PETSC_NULL ,"-force_correction", &forcecorrection, &found );
if(forcecorrection){
if(bsscrp_self->f2 && forcecorrection){
f2=bsscrp_self->f2;
ierr=VecAXPY(f,penaltyNumber,f2);/* f <- f +a*f2 */
}else{
switch (bsscrp_self->k2type) {
case (K2_GG):
{
VecDuplicate( u, &f3 );
MatMult( G, h, f3); ierr=VecAXPY(f,penaltyNumber,f3);/* f <- f +a*f2 */
}
break;
case (K2_GMG):
{
Mat M;
Vec Mdiag;
VecDuplicate( u, &f3 );
M = bsscrp_self->solver->mStiffMat->matrix;
MatGetVecs( M, &Mdiag, PETSC_NULL );
MatGetDiagonal( M, Mdiag );
VecReciprocal(Mdiag);
VecPointwiseMult(Mdiag,Mdiag,h);
MatMult( G, Mdiag, f3); ierr=VecAXPY(f,penaltyNumber,f3);/* f <- f +a*f2 */
Stg_VecDestroy(&Mdiag);
}
break;
case (K2_DGMGD):
{
Mat M;
Vec Mdiag;
VecDuplicate( u, &f3 );
M = bsscrp_self->solver->mStiffMat->matrix;
MatGetVecs( M, &Mdiag, PETSC_NULL );
MatGetDiagonal( M, Mdiag );
VecReciprocal(Mdiag);
VecPointwiseMult(Mdiag,Mdiag,h);
MatMult( G, Mdiag, f3); ierr=VecAXPY(f,penaltyNumber,f3);/* f <- f +a*f2 */
Stg_VecDestroy(&Mdiag);
}
break;
case (K2_NULL):
{
;
}
break;
case (K2_SLE):
{
;
}
break;
}
}
}
}
/* Create Schur complement matrix */
MatCreateSchurComplement(K,K,G,D,C, &S);
MatSchurComplementGetKSP( S, &ksp_inner);
KSPGetPC( ksp_inner, &pcInner );
/***************************************************************************************************************/
/***************************************************************************************************************/
/********* SET PREFIX FOR INNER/VELOCITY KSP *************************************************************/
KSPSetOptionsPrefix( ksp_inner, "A11_" );
KSPSetFromOptions( ksp_inner );
Stg_KSPSetOperators(ksp_inner, K, K, DIFFERENT_NONZERO_PATTERN);
useNormInfStoppingConditions = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL ,"-A11_use_norm_inf_stopping_condition", &useNormInfStoppingConditions, &found );
if(useNormInfStoppingConditions)
BSSCR_KSPSetNormInfConvergenceTest( ksp_inner );
useNormInfMonitor = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-A11_ksp_norm_inf_monitor", &useNormInfMonitor, &found );
if(useNormInfMonitor) KSPMonitorSet( ksp_inner, BSSCR_KSPNormInfMonitor, PETSC_NULL, PETSC_NULL );
useNormInfMonitor = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-A11_ksp_norm_inf_to_norm_2_monitor", &useNormInfMonitor, &found );
if(useNormInfMonitor) KSPMonitorSet( ksp_inner, BSSCR_KSPNormInfToNorm2Monitor, PETSC_NULL, PETSC_NULL );
/***************************************************************************************************************/
/***************************************************************************************************************/
/* If multigrid is enabled, set it now. */
change_A11rhspresolve = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-change_A11rhspresolve", &change_A11rhspresolve, &found );
if(bsscrp_self->solver->mg_active && !change_A11rhspresolve) { mgSetupTime=setupMG( bsscrp_self, ksp_inner, pcInner, K, &mgCtx ); }
/***************************************************************************************************************/
/***************************************************************************************************************/
/* create right hand side */
if(change_A11rhspresolve){
//Stg_KSPDestroy(&ksp_inner );
KSPCreate(PETSC_COMM_WORLD, &ksp_new_inner);
Stg_KSPSetOperators(ksp_new_inner, K, K, DIFFERENT_NONZERO_PATTERN);
KSPSetOptionsPrefix(ksp_new_inner, "rhsA11_");
MatSchurComplementSetKSP( S, ksp_new_inner );/* this call destroys the ksp_inner that is already set on S */
ksp_inner=ksp_new_inner;
KSPGetPC( ksp_inner, &pcInner );
KSPSetFromOptions(ksp_inner); /* make sure we are setting up our solver how we want it */
}
MatGetVecs( S, PETSC_NULL, &h_hat );
Vec f_tmp;
/* It may be the case that the current velocity solution might not be bad guess for f_tmp? <-- maybe not */
MatGetVecs( K, PETSC_NULL, &f_tmp );
scrSolveTime = MPI_Wtime();
KSPSolve(ksp_inner, f, f_tmp);
scrSolveTime = MPI_Wtime() - scrSolveTime;
PetscPrintf( PETSC_COMM_WORLD, "\n\t* KSPSolve for RHS setup Finished in time: %lf seconds\n\n", scrSolveTime);
MatMult(D, f_tmp, h_hat);
VecAYPX(h_hat, -1.0, h); /* Computes y = x + alpha y. h_hat -> h - Gt*K^(-1)*f*/
Stg_VecDestroy(&f_tmp);
if(bsscrp_self->solver->mg_active && change_A11rhspresolve) {
//Stg_KSPDestroy(&ksp_inner );
KSPCreate(PETSC_COMM_WORLD, &ksp_new_inner);
Stg_KSPSetOperators(ksp_new_inner, K, K, DIFFERENT_NONZERO_PATTERN);
KSPSetOptionsPrefix( ksp_new_inner, "A11_" );
MatSchurComplementSetKSP( S, ksp_new_inner );
ksp_inner=ksp_new_inner;
//MatSchurSetKSP( S, ksp_inner );
KSPGetPC( ksp_inner, &pcInner );
KSPSetFromOptions( ksp_inner );
mgSetupTime=setupMG( bsscrp_self, ksp_inner, pcInner, K, &mgCtx );
}
/* create solver for S p = h_hat */
KSPCreate( PETSC_COMM_WORLD, &ksp_S );
KSPSetOptionsPrefix( ksp_S, "scr_");
/* By default use the UW approxS Schur preconditioner -- same as the one used by the Uzawa solver */
/* Note that if scaling is activated then the approxS matrix has been scaled already */
/* so no need to rebuild in the case of scaling as we have been doing */
if(!approxS){ PetscPrintf( PETSC_COMM_WORLD, "WARNING approxS is NULL\n"); }
Stg_KSPSetOperators( ksp_S, S, S, SAME_NONZERO_PATTERN );
KSPSetType( ksp_S, "cg" );
KSPGetPC( ksp_S, &pc_S );
BSSCR_BSSCR_StokesCreatePCSchur2( K,G,D,C,approxS, pc_S, sym, bsscrp_self );
flg=0;
PetscOptionsGetString( PETSC_NULL, "-NN", name, PETSC_MAX_PATH_LEN-1, &flg );
if(flg){
Mat Smat, Pmat;
//MatStructure mstruct;
Stg_PCGetOperators( pc_S, &Smat, &Pmat, NULL );
sprintf(str,"%s/",name); sprintf(matname,"Pmat%s",suffix);
bsscr_dirwriteMat( Pmat, matname,str, "Writing Pmat matrix in al Solver");
}
uzawastyle=PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-uzawa_style", &uzawastyle, &found );
if(uzawastyle){
/* now want to set up the ksp_S->pc to be of type ksp (gmres) by default to match Uzawa */
KSP pc_ksp;
KSPGetPC( ksp_S, &pc_S );
PCSetType(pc_S,PCKSP);
PCKSPGetKSP( pc_S, &pc_ksp);
KSPSetType(pc_ksp, "gmres" );
KSPSetOptionsPrefix( pc_ksp, "scrPCKSP_");
KSPSetFromOptions( pc_ksp );
}
KSPSetFromOptions( ksp_S );
/* Set specific monitor test */
KSPGetTolerances( ksp_S, PETSC_NULL, PETSC_NULL, PETSC_NULL, &max_it );
// Weirdness with petsc 3.2 here...look at it later
//BSSCR_KSPLogSetMonitor( ksp_S, max_it, &monitor_index );
/***************************************************************************************************************/
/* Pressure / Velocity Solve */
/***************************************************************************************************************/
PetscPrintf( PETSC_COMM_WORLD, "\t* Pressure / Velocity Solve \n");
/***************************************************************************************************************/
/***************************************************************************************************************/
usePreviousGuess = PETSC_FALSE;
if(been_here)
PetscOptionsGetTruth( PETSC_NULL, "-scr_use_previous_guess", &usePreviousGuess, &found );
if(usePreviousGuess) { /* Note this should actually look at checkpoint information */
KSPSetInitialGuessNonzero( ksp_S, PETSC_TRUE ); }
else {
KSPSetInitialGuessNonzero( ksp_S, PETSC_FALSE ); }
/***************************************************************************************************************/
/***************************************************************************************************************/
/******* SET CONVERGENCE TESTS *************************************************************************/
useNormInfStoppingConditions = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL ,"-scr_use_norm_inf_stopping_condition", &useNormInfStoppingConditions, &found );
if(useNormInfStoppingConditions)
BSSCR_KSPSetNormInfConvergenceTest(ksp_S);
useNormInfMonitor = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-scr_ksp_norm_inf_monitor", &useNormInfMonitor, &found );
if(useNormInfMonitor)
KSPMonitorSet( ksp_S, BSSCR_KSPNormInfToNorm2Monitor, PETSC_NULL, PETSC_NULL );
/***************************************************************************************************************/
/***************************************************************************************************************/
/******* PRESSURE SOLVE ************************************************************************************/
PetscPrintf( PETSC_COMM_WORLD, "\t* KSPSolve( ksp_S, h_hat, p )\n");
/* if h_hat needs to be fixed up ..take out any nullspace vectors here */
/* we want to check that there is no "noise" in the null-space in the h vector */
/* this causes problems when we are trying to solve a Jacobian system when the Residual is almost converged */
if(bsscrp_self->check_pressureNS){
bsscrp_self->buildPNS(ksp);/* build and set nullspace vectors on bsscr - which is on ksp (function pointer is set in KSPSetUp_BSSCR) */
}
PetscScalar hnorm, gnorm;
MatNorm(G,NORM_INFINITY,&gnorm);
VecNorm(h_hat, NORM_2, &hnorm);
hnorm=hnorm/gnorm;
if((hnorm < 1e-6) && (hnorm > 1e-20)){
VecScale(h_hat,1.0/hnorm);
}
/* test to see if v or t are in nullspace of G and orthogonalize wrt h_hat if needed */
KSPRemovePressureNullspace_BSSCR(ksp, h_hat);
/* set convergence test to use min_it */
found = PETSC_FALSE;
min_it = 0;
PetscOptionsGetInt( PETSC_NULL,"-scr_ksp_set_min_it_converge", &min_it, &found);
if(found && min_it > 0){
BSSCR_KSPSetConvergenceMinIts(ksp_S, min_it, bsscrp_self);
}
/** Pressure Solve **/
PetscGetFlops(&flopsA);
scrSolveTime = MPI_Wtime();
KSPSolve( ksp_S, h_hat, p );
scrSolveTime = MPI_Wtime() - scrSolveTime;
PetscGetFlops(&flopsB);
PetscPrintf( PETSC_COMM_WORLD, "\n\t* KSPSolve( ksp_S, h_hat, p ) Solve Finished in time: %lf seconds\n\n", scrSolveTime);
bsscrp_self->solver->stats.pressure_time=scrSolveTime;
bsscrp_self->solver->stats.pressure_flops=(double)(flopsB-flopsA);
/***************************************/
if((hnorm < 1e-6) && (hnorm > 1e-20)){
VecScale(h_hat,hnorm);
VecScale(p,hnorm);
}
KSPRemovePressureNullspace_BSSCR(ksp, p);
/***************************************************************************************************************/
/***************************************************************************************************************/
/* restore K and f for the Velocity back solve */
found = PETSC_FALSE;
restorek = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-restore_K", &restorek, &found);
//PetscOptionsGetString( PETSC_NULL, "-restore_K", name, PETSC_MAX_PATH_LEN-1, &flg );
if(penaltyNumber > 1e-10 && bsscrp_self->k2type){
if(restorek){
penaltyNumber = -penaltyNumber;
if(f2) { ierr=VecAXPY(f,penaltyNumber,f2); }/* f <- f +a*f2 */
if(f3) { ierr=VecAXPY(f,penaltyNumber,f3); }/* f <- f +a*f3 */
ierr=MatAXPY(K,penaltyNumber,K2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);/* Computes K = penaltyNumber*K2 + K */
//K=Korig;
Stg_KSPSetOperators(ksp_inner, K, K, DIFFERENT_NONZERO_PATTERN);
KisJustK=PETSC_TRUE;
}
}
if(f3){ Stg_VecDestroy(&f3 ); } /* always destroy this local vector if was created */
/* obtain solution for u */
VecDuplicate( u, &t ); MatMult( G, p, t); VecAYPX( t, -1.0, f ); /*** t <- -t + f = f - G*p ***/
MatSchurComplementGetKSP( S, &ksp_inner );
a11SingleSolveTime = MPI_Wtime(); /* ---------------------------------- Final V Solve */
if(usePreviousGuess) KSPSetInitialGuessNonzero( ksp_inner, PETSC_TRUE );
/***************************************************************************************************************/
/***************************************************************************************************************/
/******* VELOCITY SOLVE ************************************************************************************/
/** Easier to just create a new KSP here if we want to do backsolve diffferently. (getting petsc errors now when switching from fgmres) */
change_backsolve=PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-change_backsolve", &change_backsolve, &found );
if(change_backsolve){
//Stg_KSPDestroy(&ksp_inner );
KSPCreate(PETSC_COMM_WORLD, &ksp_new_inner);
Stg_KSPSetOperators(ksp_new_inner, K, K, DIFFERENT_NONZERO_PATTERN);
KSPSetOptionsPrefix(ksp_new_inner, "backsolveA11_");
KSPSetFromOptions(ksp_new_inner); /* make sure we are setting up our solver how we want it */
MatSchurComplementSetKSP( S, ksp_new_inner );/* need to give the Schur it's inner ksp back for when we destroy it at end */
ksp_inner=ksp_new_inner;
}
PetscGetFlops(&flopsA);
KSPSolve(ksp_inner, t, u); /* Solve, then restore default tolerance and initial guess */
PetscGetFlops(&flopsB);
bsscrp_self->solver->stats.velocity_backsolve_flops=(double)(flopsB-flopsA);
a11SingleSolveTime = MPI_Wtime() - a11SingleSolveTime; /* ------------------ Final V Solve */
bsscrp_self->solver->stats.velocity_backsolve_time=a11SingleSolveTime;
flg=0;
PetscOptionsGetString( PETSC_NULL, "-solutiondumpdir", name, PETSC_MAX_PATH_LEN-1, &flg );
if(flg){
sprintf(str,"%s/",name); sprintf(matname,"p%s",suffix);
bsscr_dirwriteVec( p, matname,str, "Writing p vector in al Solver");
sprintf(str,"%s/",name); sprintf(matname,"u%s",suffix);
bsscr_dirwriteVec( u, matname,str, "Writing u vector in al Solver");
sprintf(str,"%s/",name); sprintf(matname,"h_hat%s",suffix);
bsscr_dirwriteVec( h_hat, matname,str, "Writing h_hat vector in al Solver");
}
found = PETSC_FALSE;
restorek = PETSC_FALSE;
PetscOptionsGetTruth( PETSC_NULL, "-restore_K_after_solve", &restorek, &found);
if(penaltyNumber > 1e-10 && bsscrp_self->k2type){
if(restorek){
penaltyNumber = -penaltyNumber;
ierr=MatAXPY(K,penaltyNumber,K2,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);/* Computes K = penaltyNumber*K2 + K */
KisJustK=PETSC_TRUE;
}
}
/***************************************************************************************************************/
/***************************************************************************************************************/
/****** SOLUTION SUMMARY ******************************************************************************/
bsscr_summary(bsscrp_self,ksp_S,ksp_inner,K,K2,D,G,C,u,p,f,h,t,penaltyNumber,KisJustK, mgSetupTime, scrSolveTime, a11SingleSolveTime);
//bsscr_summary(bsscrp_self,ksp_S,ksp_inner,K,Korig,K2,D,G,C,u,p,f,h,t,penaltyNumber,KisJustK, mgSetupTime, scrSolveTime, a11SingleSolveTime);
/***************************************************************************************************************/
/***************************************************************************************************************/
Stg_VecDestroy(&t );
Stg_KSPDestroy(&ksp_S );
Stg_VecDestroy(&h_hat );
Stg_MatDestroy(&S );//This will destroy ksp_inner: also.. pcInner == pc_MG and is destroyed when ksp_inner is
been_here = 1;
PetscFunctionReturn(0);
}
extern PetscErrorCode MatLMVMUpdate(Mat M, Vec x, Vec g)
{
MatLMVMCtx *ctx;
PetscReal rhotemp, rhotol;
PetscReal y0temp, s0temp;
PetscReal yDy, yDs, sDs;
PetscReal sigmanew, denom;
PetscErrorCode ierr;
PetscInt i;
PetscBool same;
PetscReal yy_sum=0.0, ys_sum=0.0, ss_sum=0.0;
PetscFunctionBegin;
PetscValidHeaderSpecific(x,VEC_CLASSID,2);
PetscValidHeaderSpecific(g,VEC_CLASSID,3);
ierr = PetscObjectTypeCompare((PetscObject)M,MATSHELL,&same);CHKERRQ(ierr);
if (!same) SETERRQ(PETSC_COMM_SELF,1,"Matrix M is not type MatLMVM");
ierr = MatShellGetContext(M,(void**)&ctx);CHKERRQ(ierr);
if (!ctx->allocated) {
ierr = MatLMVMAllocateVectors(M, x); CHKERRQ(ierr);
}
if (0 == ctx->iter) {
ierr = MatLMVMReset(M);CHKERRQ(ierr);
} else {
ierr = VecAYPX(ctx->Gprev,-1.0,g);CHKERRQ(ierr);
ierr = VecAYPX(ctx->Xprev,-1.0,x);CHKERRQ(ierr);
ierr = VecDot(ctx->Gprev,ctx->Xprev,&rhotemp);CHKERRQ(ierr);
ierr = VecDot(ctx->Gprev,ctx->Gprev,&y0temp);CHKERRQ(ierr);
rhotol = ctx->eps * y0temp;
if (rhotemp > rhotol) {
++ctx->nupdates;
ctx->lmnow = PetscMin(ctx->lmnow+1, ctx->lm);
ierr=PetscObjectDereference((PetscObject)ctx->S[ctx->lm]);CHKERRQ(ierr);
ierr=PetscObjectDereference((PetscObject)ctx->Y[ctx->lm]);CHKERRQ(ierr);
for (i = ctx->lm-1; i >= 0; --i) {
ctx->S[i+1] = ctx->S[i];
ctx->Y[i+1] = ctx->Y[i];
ctx->rho[i+1] = ctx->rho[i];
}
ctx->S[0] = ctx->Xprev;
ctx->Y[0] = ctx->Gprev;
PetscObjectReference((PetscObject)ctx->S[0]);
PetscObjectReference((PetscObject)ctx->Y[0]);
ctx->rho[0] = 1.0 / rhotemp;
/* Compute the scaling */
switch(ctx->scaleType) {
case MatLMVM_Scale_None:
break;
case MatLMVM_Scale_Scalar:
/* Compute s^T s */
ierr = VecDot(ctx->Xprev,ctx->Xprev,&s0temp);CHKERRQ(ierr);
/* Scalar is positive; safeguards are not required. */
/* Save information for scalar scaling */
ctx->yy_history[(ctx->nupdates - 1) % ctx->scalar_history] = y0temp;
ctx->ys_history[(ctx->nupdates - 1) % ctx->scalar_history] = rhotemp;
ctx->ss_history[(ctx->nupdates - 1) % ctx->scalar_history] = s0temp;
/* Compute summations for scalar scaling */
yy_sum = 0; /* No safeguard required; y^T y > 0 */
ys_sum = 0; /* No safeguard required; y^T s > 0 */
ss_sum = 0; /* No safeguard required; s^T s > 0 */
for (i = 0; i < PetscMin(ctx->nupdates, ctx->scalar_history); ++i) {
yy_sum += ctx->yy_history[i];
ys_sum += ctx->ys_history[i];
ss_sum += ctx->ss_history[i];
}
if (0.0 == ctx->s_alpha) {
/* Safeguard ys_sum */
if (0.0 == ys_sum) {
ys_sum = TAO_ZERO_SAFEGUARD;
}
sigmanew = ss_sum / ys_sum;
} else if (1.0 == ctx->s_alpha) {
/* Safeguard yy_sum */
if (0.0 == yy_sum) {
yy_sum = TAO_ZERO_SAFEGUARD;
}
sigmanew = ys_sum / yy_sum;
} else {
denom = 2*ctx->s_alpha*yy_sum;
/* Safeguard denom */
if (0.0 == denom) {
denom = TAO_ZERO_SAFEGUARD;
}
sigmanew = ((2*ctx->s_alpha-1)*ys_sum + PetscSqrtScalar((2*ctx->s_alpha-1)*(2*ctx->s_alpha-1)*ys_sum*ys_sum - 4*(ctx->s_alpha)*(ctx->s_alpha-1)*yy_sum*ss_sum)) / denom;
}
switch(ctx->limitType) {
case MatLMVM_Limit_Average:
if (1.0 == ctx->mu) {
ctx->sigma = sigmanew;
} else if (ctx->mu) {
ctx->sigma = ctx->mu * sigmanew + (1.0 - ctx->mu) * ctx->sigma;
}
break;
case MatLMVM_Limit_Relative:
if (ctx->mu) {
ctx->sigma = TaoMid((1.0 - ctx->mu) * ctx->sigma, sigmanew, (1.0 + ctx->mu) * ctx->sigma);
}
break;
case MatLMVM_Limit_Absolute:
if (ctx->nu) {
ctx->sigma = TaoMid(ctx->sigma - ctx->nu, sigmanew, ctx->sigma + ctx->nu);
}
break;
default:
ctx->sigma = sigmanew;
break;
}
break;
case MatLMVM_Scale_Broyden:
/* Original version */
/* Combine DFP and BFGS */
/* This code appears to be numerically unstable. We use the */
/* original version because this was used to generate all of */
/* the data and because it may be the least unstable of the */
/* bunch. */
/* P = Q = inv(D); */
ierr = VecCopy(ctx->D,ctx->P);CHKERRQ(ierr);
ierr = VecReciprocal(ctx->P);CHKERRQ(ierr);
ierr = VecCopy(ctx->P,ctx->Q);CHKERRQ(ierr);
/* V = y*y */
ierr = VecPointwiseMult(ctx->V,ctx->Gprev,ctx->Gprev);CHKERRQ(ierr);
/* W = inv(D)*s */
ierr = VecPointwiseMult(ctx->W,ctx->Xprev,ctx->P);CHKERRQ(ierr);
ierr = VecDot(ctx->W,ctx->Xprev,&sDs);CHKERRQ(ierr);
/* Safeguard rhotemp and sDs */
if (0.0 == rhotemp) {
rhotemp = TAO_ZERO_SAFEGUARD;
}
if (0.0 == sDs) {
sDs = TAO_ZERO_SAFEGUARD;
}
if (1.0 != ctx->phi) {
/* BFGS portion of the update */
/* U = (inv(D)*s)*(inv(D)*s) */
ierr = VecPointwiseMult(ctx->U,ctx->W,ctx->W);CHKERRQ(ierr);
/* Assemble */
ierr = VecAXPY(ctx->P,1.0/rhotemp,ctx->V);CHKERRQ(ierr);
ierr = VecAXPY(ctx->P,-1.0/sDs,ctx->U);CHKERRQ(ierr);
}
if (0.0 != ctx->phi) {
/* DFP portion of the update */
/* U = inv(D)*s*y */
ierr = VecPointwiseMult(ctx->U, ctx->W, ctx->Gprev);CHKERRQ(ierr);
/* Assemble */
ierr = VecAXPY(ctx->Q,1.0/rhotemp + sDs/(rhotemp*rhotemp), ctx->V);CHKERRQ(ierr);
ierr = VecAXPY(ctx->Q,-2.0/rhotemp,ctx->U);CHKERRQ(ierr);
}
if (0.0 == ctx->phi) {
ierr = VecCopy(ctx->P,ctx->U);CHKERRQ(ierr);
} else if (1.0 == ctx->phi) {
ierr = VecCopy(ctx->Q,ctx->U);CHKERRQ(ierr);
} else {
/* Broyden update U=(1-phi)*P + phi*Q */
ierr = VecCopy(ctx->Q,ctx->U);CHKERRQ(ierr);
ierr = VecAXPBY(ctx->U,1.0-ctx->phi, ctx->phi, ctx->P);CHKERRQ(ierr);
}
/* Obtain inverse and ensure positive definite */
ierr = VecReciprocal(ctx->U);CHKERRQ(ierr);
ierr = VecAbs(ctx->U);CHKERRQ(ierr);
switch(ctx->rScaleType) {
case MatLMVM_Rescale_None:
break;
case MatLMVM_Rescale_Scalar:
case MatLMVM_Rescale_GL:
if (ctx->rScaleType == MatLMVM_Rescale_GL) {
/* Gilbert and Lemarachal use the old diagonal */
ierr = VecCopy(ctx->D,ctx->P);CHKERRQ(ierr);
} else {
/* The default version uses the current diagonal */
ierr = VecCopy(ctx->U,ctx->P);CHKERRQ(ierr);
}
/* Compute s^T s */
ierr = VecDot(ctx->Xprev,ctx->Xprev,&s0temp);CHKERRQ(ierr);
/* Save information for special cases of scalar rescaling */
ctx->yy_rhistory[(ctx->nupdates - 1) % ctx->rescale_history] = y0temp;
ctx->ys_rhistory[(ctx->nupdates - 1) % ctx->rescale_history] = rhotemp;
ctx->ss_rhistory[(ctx->nupdates - 1) % ctx->rescale_history] = s0temp;
if (0.5 == ctx->r_beta) {
if (1 == PetscMin(ctx->nupdates, ctx->rescale_history)) {
ierr = VecPointwiseMult(ctx->V,ctx->Y[0],ctx->P);CHKERRQ(ierr);
ierr = VecDot(ctx->V,ctx->Y[0],&yy_sum);CHKERRQ(ierr);
ierr = VecPointwiseDivide(ctx->W,ctx->S[0],ctx->P);CHKERRQ(ierr);
ierr = VecDot(ctx->W,ctx->S[0],&ss_sum);CHKERRQ(ierr);
ys_sum = ctx->ys_rhistory[0];
} else {
ierr = VecCopy(ctx->P,ctx->Q);CHKERRQ(ierr);
ierr = VecReciprocal(ctx->Q);CHKERRQ(ierr);
/* Compute summations for scalar scaling */
yy_sum = 0; /* No safeguard required */
ys_sum = 0; /* No safeguard required */
ss_sum = 0; /* No safeguard required */
for (i = 0; i < PetscMin(ctx->nupdates, ctx->rescale_history); ++i) {
ierr = VecPointwiseMult(ctx->V,ctx->Y[i],ctx->P);CHKERRQ(ierr);
ierr = VecDot(ctx->V,ctx->Y[i],&yDy);CHKERRQ(ierr);
yy_sum += yDy;
ierr = VecPointwiseMult(ctx->W,ctx->S[i],ctx->Q);CHKERRQ(ierr);
ierr = VecDot(ctx->W,ctx->S[i],&sDs);CHKERRQ(ierr);
ss_sum += sDs;
ys_sum += ctx->ys_rhistory[i];
}
}
} else if (0.0 == ctx->r_beta) {
if (1 == PetscMin(ctx->nupdates, ctx->rescale_history)) {
/* Compute summations for scalar scaling */
ierr = VecPointwiseDivide(ctx->W,ctx->S[0],ctx->P);CHKERRQ(ierr);
ierr = VecDot(ctx->W, ctx->Y[0], &ys_sum);CHKERRQ(ierr);
ierr = VecDot(ctx->W, ctx->W, &ss_sum);CHKERRQ(ierr);
yy_sum += ctx->yy_rhistory[0];
} else {
ierr = VecCopy(ctx->Q, ctx->P);CHKERRQ(ierr);
ierr = VecReciprocal(ctx->Q);CHKERRQ(ierr);
/* Compute summations for scalar scaling */
yy_sum = 0; /* No safeguard required */
ys_sum = 0; /* No safeguard required */
ss_sum = 0; /* No safeguard required */
for (i = 0; i < PetscMin(ctx->nupdates, ctx->rescale_history); ++i) {
ierr = VecPointwiseMult(ctx->W, ctx->S[i], ctx->Q);CHKERRQ(ierr);
ierr = VecDot(ctx->W, ctx->Y[i], &yDs);CHKERRQ(ierr);
ys_sum += yDs;
ierr = VecDot(ctx->W, ctx->W, &sDs);CHKERRQ(ierr);
ss_sum += sDs;
yy_sum += ctx->yy_rhistory[i];
}
}
} else if (1.0 == ctx->r_beta) {
/* Compute summations for scalar scaling */
yy_sum = 0; /* No safeguard required */
ys_sum = 0; /* No safeguard required */
ss_sum = 0; /* No safeguard required */
for (i = 0; i < PetscMin(ctx->nupdates, ctx->rescale_history); ++i) {
ierr = VecPointwiseMult(ctx->V, ctx->Y[i], ctx->P);CHKERRQ(ierr);
ierr = VecDot(ctx->V, ctx->S[i], &yDs);CHKERRQ(ierr);
ys_sum += yDs;
ierr = VecDot(ctx->V, ctx->V, &yDy);CHKERRQ(ierr);
yy_sum += yDy;
ss_sum += ctx->ss_rhistory[i];
}
} else {
ierr = VecCopy(ctx->Q, ctx->P);CHKERRQ(ierr);
ierr = VecPow(ctx->P, ctx->r_beta);CHKERRQ(ierr);
ierr = VecPointwiseDivide(ctx->Q, ctx->P, ctx->Q);CHKERRQ(ierr);
/* Compute summations for scalar scaling */
yy_sum = 0; /* No safeguard required */
ys_sum = 0; /* No safeguard required */
ss_sum = 0; /* No safeguard required */
for (i = 0; i < PetscMin(ctx->nupdates, ctx->rescale_history); ++i) {
ierr = VecPointwiseMult(ctx->V, ctx->P, ctx->Y[i]);CHKERRQ(ierr);
ierr = VecPointwiseMult(ctx->W, ctx->Q, ctx->S[i]);CHKERRQ(ierr);
ierr = VecDot(ctx->V, ctx->V, &yDy);CHKERRQ(ierr);
ierr = VecDot(ctx->V, ctx->W, &yDs);CHKERRQ(ierr);
ierr = VecDot(ctx->W, ctx->W, &sDs);CHKERRQ(ierr);
yy_sum += yDy;
ys_sum += yDs;
ss_sum += sDs;
}
}
if (0.0 == ctx->r_alpha) {
/* Safeguard ys_sum */
if (0.0 == ys_sum) {
ys_sum = TAO_ZERO_SAFEGUARD;
}
sigmanew = ss_sum / ys_sum;
} else if (1.0 == ctx->r_alpha) {
/* Safeguard yy_sum */
if (0.0 == yy_sum) {
ys_sum = TAO_ZERO_SAFEGUARD;
}
sigmanew = ys_sum / yy_sum;
} else {
denom = 2*ctx->r_alpha*yy_sum;
/* Safeguard denom */
if (0.0 == denom) {
denom = TAO_ZERO_SAFEGUARD;
}
sigmanew = ((2*ctx->r_alpha-1)*ys_sum + PetscSqrtScalar((2*ctx->r_alpha-1)*(2*ctx->r_alpha-1)*ys_sum*ys_sum - 4*ctx->r_alpha*(ctx->r_alpha-1)*yy_sum*ss_sum)) / denom;
}
/* If Q has small values, then Q^(r_beta - 1) */
/* can have very large values. Hence, ys_sum */
/* and ss_sum can be infinity. In this case, */
/* sigmanew can either be not-a-number or infinity. */
if (PetscIsInfOrNanReal(sigmanew)) {
/* sigmanew is not-a-number; skip rescaling */
} else if (!sigmanew) {
/* sigmanew is zero; this is a bad case; skip rescaling */
} else {
/* sigmanew is positive */
ierr = VecScale(ctx->U, sigmanew);CHKERRQ(ierr);
}
break;
}
/* Modify for previous information */
switch(ctx->limitType) {
case MatLMVM_Limit_Average:
if (1.0 == ctx->mu) {
ierr = VecCopy(ctx->D, ctx->U);CHKERRQ(ierr);
} else if (ctx->mu) {
ierr = VecAXPBY(ctx->D,ctx->mu, 1.0-ctx->mu,ctx->U);CHKERRQ(ierr);
}
break;
case MatLMVM_Limit_Relative:
if (ctx->mu) {
/* P = (1-mu) * D */
ierr = VecAXPBY(ctx->P, 1.0-ctx->mu, 0.0, ctx->D);CHKERRQ(ierr);
/* Q = (1+mu) * D */
ierr = VecAXPBY(ctx->Q, 1.0+ctx->mu, 0.0, ctx->D);CHKERRQ(ierr);
ierr = VecMedian(ctx->P, ctx->U, ctx->Q, ctx->D);CHKERRQ(ierr);
}
break;
case MatLMVM_Limit_Absolute:
if (ctx->nu) {
ierr = VecCopy(ctx->P, ctx->D);CHKERRQ(ierr);
ierr = VecShift(ctx->P, -ctx->nu);CHKERRQ(ierr);
ierr = VecCopy(ctx->D, ctx->Q);CHKERRQ(ierr);
ierr = VecShift(ctx->Q, ctx->nu);CHKERRQ(ierr);
ierr = VecMedian(ctx->P, ctx->U, ctx->Q, ctx->P);CHKERRQ(ierr);
}
break;
default:
ierr = VecCopy(ctx->U, ctx->D);CHKERRQ(ierr);
break;
}
break;
}
ierr = PetscObjectDereference((PetscObject)ctx->Xprev);CHKERRQ(ierr);
ierr = PetscObjectDereference((PetscObject)ctx->Gprev);CHKERRQ(ierr);
ctx->Xprev = ctx->S[ctx->lm];
ctx->Gprev = ctx->Y[ctx->lm];
ierr = PetscObjectReference((PetscObject)ctx->S[ctx->lm]);CHKERRQ(ierr);
ierr = PetscObjectReference((PetscObject)ctx->Y[ctx->lm]);CHKERRQ(ierr);
} else {
++ctx->nrejects;
}
}
++ctx->iter;
ierr = VecCopy(x, ctx->Xprev);CHKERRQ(ierr);
ierr = VecCopy(g, ctx->Gprev);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@C
PCGAMGFilterGraph - filter (remove zero and possibly small values from the) graph and make it symmetric if requested
Collective on Mat
Input Parameter:
+ a_Gmat - the graph
. vfilter - threshold paramter [0,1)
- symm - make the result symmetric
Level: developer
Notes:
This is called before graph coarsers are called.
.seealso: PCGAMGSetThreshold()
@*/
PetscErrorCode PCGAMGFilterGraph(Mat *a_Gmat,PetscReal vfilter,PetscBool symm)
{
PetscErrorCode ierr;
PetscInt Istart,Iend,Ii,jj,ncols,nnz0,nnz1, NN, MM, nloc;
PetscMPIInt rank;
Mat Gmat = *a_Gmat, tGmat, matTrans;
MPI_Comm comm;
const PetscScalar *vals;
const PetscInt *idx;
PetscInt *d_nnz, *o_nnz;
Vec diag;
MatType mtype;
PetscFunctionBegin;
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
/* scale Gmat for all values between -1 and 1 */
ierr = MatCreateVecs(Gmat, &diag, 0);CHKERRQ(ierr);
ierr = MatGetDiagonal(Gmat, diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecSqrtAbs(diag);CHKERRQ(ierr);
ierr = MatDiagonalScale(Gmat, diag, diag);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
if (vfilter < 0.0 && !symm) {
/* Just use the provided matrix as the graph but make all values positive */
MatInfo info;
PetscScalar *avals;
PetscBool isaij,ismpiaij;
ierr = PetscObjectBaseTypeCompare((PetscObject)Gmat,MATSEQAIJ,&isaij);CHKERRQ(ierr);
ierr = PetscObjectBaseTypeCompare((PetscObject)Gmat,MATMPIAIJ,&ismpiaij);CHKERRQ(ierr);
if (!isaij && !ismpiaij) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_USER,"Require (MPI)AIJ matrix type");
if (isaij) {
ierr = MatGetInfo(Gmat,MAT_LOCAL,&info);CHKERRQ(ierr);
ierr = MatSeqAIJGetArray(Gmat,&avals);CHKERRQ(ierr);
for (jj = 0; jj<info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
ierr = MatSeqAIJRestoreArray(Gmat,&avals);CHKERRQ(ierr);
} else {
Mat_MPIAIJ *aij = (Mat_MPIAIJ*)Gmat->data;
ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr);
ierr = MatSeqAIJGetArray(aij->A,&avals);CHKERRQ(ierr);
for (jj = 0; jj<info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
ierr = MatSeqAIJRestoreArray(aij->A,&avals);CHKERRQ(ierr);
ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr);
ierr = MatSeqAIJGetArray(aij->B,&avals);CHKERRQ(ierr);
for (jj = 0; jj<info.nz_used; jj++) avals[jj] = PetscAbsScalar(avals[jj]);
ierr = MatSeqAIJRestoreArray(aij->B,&avals);CHKERRQ(ierr);
}
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
PetscFunctionReturn(0);
}
ierr = PetscObjectGetComm((PetscObject)Gmat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Gmat, &Istart, &Iend);CHKERRQ(ierr);
nloc = Iend - Istart;
ierr = MatGetSize(Gmat, &MM, &NN);CHKERRQ(ierr);
if (symm) {
ierr = MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans);CHKERRQ(ierr);
}
/* Determine upper bound on nonzeros needed in new filtered matrix */
ierr = PetscMalloc2(nloc, &d_nnz,nloc, &o_nnz);CHKERRQ(ierr);
for (Ii = Istart, jj = 0; Ii < Iend; Ii++, jj++) {
ierr = MatGetRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] = ncols;
o_nnz[jj] = ncols;
ierr = MatRestoreRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
if (symm) {
ierr = MatGetRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] += ncols;
o_nnz[jj] += ncols;
ierr = MatRestoreRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
}
if (d_nnz[jj] > nloc) d_nnz[jj] = nloc;
if (o_nnz[jj] > (MM-nloc)) o_nnz[jj] = MM - nloc;
}
ierr = MatGetType(Gmat,&mtype);CHKERRQ(ierr);
ierr = MatCreate(comm, &tGmat);CHKERRQ(ierr);
ierr = MatSetSizes(tGmat,nloc,nloc,MM,MM);CHKERRQ(ierr);
ierr = MatSetBlockSizes(tGmat, 1, 1);CHKERRQ(ierr);
ierr = MatSetType(tGmat, mtype);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(tGmat,0,d_nnz);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(tGmat,0,d_nnz,0,o_nnz);CHKERRQ(ierr);
ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr);
if (symm) {
ierr = MatDestroy(&matTrans);CHKERRQ(ierr);
} else {
/* all entries are generated locally so MatAssembly will be slightly faster for large process counts */
ierr = MatSetOption(tGmat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr);
}
for (Ii = Istart, nnz0 = nnz1 = 0; Ii < Iend; Ii++) {
ierr = MatGetRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
for (jj=0; jj<ncols; jj++,nnz0++) {
PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
if (PetscRealPart(sv) > vfilter) {
nnz1++;
if (symm) {
sv *= 0.5;
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(tGmat,1,&idx[jj],1,&Ii,&sv,ADD_VALUES);CHKERRQ(ierr);
} else {
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatRestoreRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
#if defined(PETSC_USE_INFO)
{
double t1 = (!nnz0) ? 1. : 100.*(double)nnz1/(double)nnz0, t2 = (!nloc) ? 1. : (double)nnz0/(double)nloc;
ierr = PetscInfo4(*a_Gmat,"\t %g%% nnz after filtering, with threshold %g, %g nnz ave. (N=%D)\n",t1,vfilter,t2,MM);CHKERRQ(ierr);
}
#endif
ierr = MatDestroy(&Gmat);CHKERRQ(ierr);
*a_Gmat = tGmat;
PetscFunctionReturn(0);
}
EXTERN_C_END
/* -------------------------------------------------------------------------- */
/*
PCSetUp_Jacobi - Prepares for the use of the Jacobi preconditioner
by setting data structures and options.
Input Parameter:
. pc - the preconditioner context
Application Interface Routine: PCSetUp()
Notes:
The interface routine PCSetUp() is not usually called directly by
the user, but instead is called by PCApply() if necessary.
*/
#undef __FUNCT__
#define __FUNCT__ "PCSetUp_Jacobi"
static PetscErrorCode PCSetUp_Jacobi(PC pc)
{
PC_Jacobi *jac = (PC_Jacobi*)pc->data;
Vec diag,diagsqrt;
PetscErrorCode ierr;
PetscInt n,i;
PetscScalar *x;
PetscBool zeroflag = PETSC_FALSE;
PetscFunctionBegin;
/*
For most preconditioners the code would begin here something like
if (pc->setupcalled == 0) { allocate space the first time this is ever called
ierr = MatGetVecs(pc->mat,&jac->diag);CHKERRQ(ierr);
PetscLogObjectParent(pc,jac->diag);
}
But for this preconditioner we want to support use of both the matrix' diagonal
elements (for left or right preconditioning) and square root of diagonal elements
(for symmetric preconditioning). Hence we do not allocate space here, since we
don't know at this point which will be needed (diag and/or diagsqrt) until the user
applies the preconditioner, and we don't want to allocate BOTH unless we need
them both. Thus, the diag and diagsqrt are allocated in PCSetUp_Jacobi_NonSymmetric()
and PCSetUp_Jacobi_Symmetric(), respectively.
*/
/*
Here we set up the preconditioner; that is, we copy the diagonal values from
the matrix and put them into a format to make them quick to apply as a preconditioner.
*/
diag = jac->diag;
diagsqrt = jac->diagsqrt;
if (diag) {
if (jac->userowmax) {
ierr = MatGetRowMaxAbs(pc->pmat,diag,PETSC_NULL);CHKERRQ(ierr);
} else if (jac->userowsum) {
ierr = MatGetRowSum(pc->pmat,diag);CHKERRQ(ierr);
} else {
ierr = MatGetDiagonal(pc->pmat,diag);CHKERRQ(ierr);
}
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecGetLocalSize(diag,&n);CHKERRQ(ierr);
ierr = VecGetArray(diag,&x);CHKERRQ(ierr);
if (jac->useabs) {
for (i=0; i<n; i++) {
x[i] = PetscAbsScalar(x[i]);
}
}
for (i=0; i<n; i++) {
if (x[i] == 0.0) {
x[i] = 1.0;
zeroflag = PETSC_TRUE;
}
}
ierr = VecRestoreArray(diag,&x);CHKERRQ(ierr);
}
if (diagsqrt) {
if (jac->userowmax) {
ierr = MatGetRowMaxAbs(pc->pmat,diagsqrt,PETSC_NULL);CHKERRQ(ierr);
} else if (jac->userowsum) {
ierr = MatGetRowSum(pc->pmat,diagsqrt);CHKERRQ(ierr);
} else {
ierr = MatGetDiagonal(pc->pmat,diagsqrt);CHKERRQ(ierr);
}
ierr = VecGetLocalSize(diagsqrt,&n);CHKERRQ(ierr);
ierr = VecGetArray(diagsqrt,&x);CHKERRQ(ierr);
for (i=0; i<n; i++) {
if (x[i] != 0.0) x[i] = 1.0/PetscSqrtReal(PetscAbsScalar(x[i]));
else {
x[i] = 1.0;
zeroflag = PETSC_TRUE;
}
}
ierr = VecRestoreArray(diagsqrt,&x);CHKERRQ(ierr);
}
if (zeroflag) {
ierr = PetscInfo(pc,"Zero detected in diagonal of matrix, using 1 at those locations\n");CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
PetscErrorCode PCGAMGFilterGraph(Mat *a_Gmat,const PetscReal vfilter,const PetscBool symm,const PetscInt verbose)
{
PetscErrorCode ierr;
PetscInt Istart,Iend,Ii,jj,ncols,nnz0,nnz1, NN, MM, nloc;
PetscMPIInt rank, size;
Mat Gmat = *a_Gmat, tGmat, matTrans;
MPI_Comm comm;
const PetscScalar *vals;
const PetscInt *idx;
PetscInt *d_nnz, *o_nnz;
Vec diag;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Gmat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Gmat, &Istart, &Iend);CHKERRQ(ierr);
nloc = Iend - Istart;
ierr = MatGetSize(Gmat, &MM, &NN);CHKERRQ(ierr);
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
/* scale Gmat so filter works */
ierr = MatGetVecs(Gmat, &diag, 0);CHKERRQ(ierr);
ierr = MatGetDiagonal(Gmat, diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecSqrtAbs(diag);CHKERRQ(ierr);
ierr = MatDiagonalScale(Gmat, diag, diag);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
if (symm) {
ierr = MatTranspose(Gmat, MAT_INITIAL_MATRIX, &matTrans);CHKERRQ(ierr);
}
/* filter - dup zeros out matrix */
ierr = PetscMalloc1(nloc, &d_nnz);CHKERRQ(ierr);
ierr = PetscMalloc1(nloc, &o_nnz);CHKERRQ(ierr);
for (Ii = Istart, jj = 0; Ii < Iend; Ii++, jj++) {
ierr = MatGetRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] = ncols;
o_nnz[jj] = ncols;
ierr = MatRestoreRow(Gmat,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
if (symm) {
ierr = MatGetRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
d_nnz[jj] += ncols;
o_nnz[jj] += ncols;
ierr = MatRestoreRow(matTrans,Ii,&ncols,NULL,NULL);CHKERRQ(ierr);
}
if (d_nnz[jj] > nloc) d_nnz[jj] = nloc;
if (o_nnz[jj] > (MM-nloc)) o_nnz[jj] = MM - nloc;
}
ierr = MatCreateAIJ(comm, nloc, nloc, MM, MM, 0, d_nnz, 0, o_nnz, &tGmat);CHKERRQ(ierr);
ierr = PetscFree(d_nnz);CHKERRQ(ierr);
ierr = PetscFree(o_nnz);CHKERRQ(ierr);
if (symm) {
ierr = MatDestroy(&matTrans);CHKERRQ(ierr);
}
for (Ii = Istart, nnz0 = nnz1 = 0; Ii < Iend; Ii++) {
ierr = MatGetRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
for (jj=0; jj<ncols; jj++,nnz0++) {
PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
if (PetscRealPart(sv) > vfilter) {
nnz1++;
if (symm) {
sv *= 0.5;
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(tGmat,1,&idx[jj],1,&Ii,&sv,ADD_VALUES);CHKERRQ(ierr);
} else {
ierr = MatSetValues(tGmat,1,&Ii,1,&idx[jj],&sv,ADD_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatRestoreRow(Gmat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(tGmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
if (verbose) {
if (verbose == 1) {
ierr = PetscPrintf(comm,"\t[%d]%s %g%% nnz after filtering, with threshold %g, %g nnz ave. (N=%d)\n",rank,__FUNCT__,
100.*(double)nnz1/(double)nnz0,vfilter,(double)nnz0/(double)nloc,MM);CHKERRQ(ierr);
} else {
PetscInt nnz[2],out[2];
nnz[0] = nnz0; nnz[1] = nnz1;
ierr = MPI_Allreduce(nnz, out, 2, MPIU_INT, MPI_SUM, comm);CHKERRQ(ierr);
ierr = PetscPrintf(comm,"\t[%d]%s %g%% nnz after filtering, with threshold %g, %g nnz ave. (N=%d)\n",rank,__FUNCT__,
100.*(double)out[1]/(double)out[0],vfilter,(double)out[0]/(double)MM,MM);CHKERRQ(ierr);
}
}
ierr = MatDestroy(&Gmat);CHKERRQ(ierr);
*a_Gmat = tGmat;
PetscFunctionReturn(0);
}
/*
Forms the interpolation (and restriction) operator from
coarse grid to fine.
*/
PetscErrorCode FormInterpolation(AppCtx *user)
{
PetscErrorCode ierr;
PetscInt i,j,i_start,m_fine,j_start,m,n,*idx;
PetscInt m_ghost,n_ghost,*idx_c,m_ghost_c,n_ghost_c,m_coarse;
PetscInt row,i_start_ghost,j_start_ghost,cols[4], m_c;
PetscInt nc,ratio = user->ratio,m_c_local,m_fine_local;
PetscInt i_c,j_c,i_start_c,j_start_c,n_c,i_start_ghost_c,j_start_ghost_c,col;
PetscScalar v[4],x,y, one = 1.0;
Mat mat;
Vec Rscale;
ierr = DMDAGetCorners(user->fine.da,&i_start,&j_start,0,&m,&n,0);CHKERRQ(ierr);
ierr = DMDAGetGhostCorners(user->fine.da,&i_start_ghost,&j_start_ghost,0,&m_ghost,&n_ghost,0);CHKERRQ(ierr);
ierr = DMDAGetGlobalIndices(user->fine.da,PETSC_NULL,&idx);CHKERRQ(ierr);
ierr = DMDAGetCorners(user->coarse.da,&i_start_c,&j_start_c,0,&m_c,&n_c,0);CHKERRQ(ierr);
ierr = DMDAGetGhostCorners(user->coarse.da,&i_start_ghost_c,&j_start_ghost_c,0,&m_ghost_c,&n_ghost_c,0);CHKERRQ(ierr);
ierr = DMDAGetGlobalIndices(user->coarse.da,PETSC_NULL,&idx_c);CHKERRQ(ierr);
/* create interpolation matrix */
ierr = VecGetLocalSize(user->fine.x,&m_fine_local);CHKERRQ(ierr);
ierr = VecGetLocalSize(user->coarse.x,&m_c_local);CHKERRQ(ierr);
ierr = VecGetSize(user->fine.x,&m_fine);CHKERRQ(ierr);
ierr = VecGetSize(user->coarse.x,&m_coarse);CHKERRQ(ierr);
ierr = MatCreateAIJ(PETSC_COMM_WORLD,m_fine_local,m_c_local,m_fine,m_coarse,
5,0,3,0,&mat);CHKERRQ(ierr);
/* loop over local fine grid nodes setting interpolation for those*/
for ( j=j_start; j<j_start+n; j++ ) {
for ( i=i_start; i<i_start+m; i++ ) {
/* convert to local "natural" numbering and then to PETSc global numbering */
row = idx[m_ghost*(j-j_start_ghost) + (i-i_start_ghost)];
i_c = (i/ratio); /* coarse grid node to left of fine grid node */
j_c = (j/ratio); /* coarse grid node below fine grid node */
/*
Only include those interpolation points that are truly
nonzero. Note this is very important for final grid lines
in x and y directions; since they have no right/top neighbors
*/
x = ((double)(i - i_c*ratio))/((double)ratio);
y = ((double)(j - j_c*ratio))/((double)ratio);
nc = 0;
/* one left and below; or we are right on it */
if (j_c < j_start_ghost_c || j_c > j_start_ghost_c+n_ghost_c) {
SETERRQ3(PETSC_COMM_SELF,1,"Sorry j %D %D %D",j_c,j_start_ghost_c,j_start_ghost_c+n_ghost_c);
}
if (i_c < i_start_ghost_c || i_c > i_start_ghost_c+m_ghost_c) {
SETERRQ3(PETSC_COMM_SELF,1,"Sorry i %D %D %D",i_c,i_start_ghost_c,i_start_ghost_c+m_ghost_c);
}
col = m_ghost_c*(j_c-j_start_ghost_c) + (i_c-i_start_ghost_c);
cols[nc] = idx_c[col];
v[nc++] = x*y - x - y + 1.0;
/* one right and below */
if (i_c*ratio != i) {
cols[nc] = idx_c[col+1];
v[nc++] = -x*y + x;
}
/* one left and above */
if (j_c*ratio != j) {
cols[nc] = idx_c[col+m_ghost_c];
v[nc++] = -x*y + y;
}
/* one right and above */
if (j_c*ratio != j && i_c*ratio != i) {
cols[nc] = idx_c[col+m_ghost_c+1];
v[nc++] = x*y;
}
ierr = MatSetValues(mat,1,&row,nc,cols,v,INSERT_VALUES);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = VecDuplicate(user->coarse.x,&Rscale);CHKERRQ(ierr);
ierr = VecSet(user->fine.x,one);CHKERRQ(ierr);
ierr = MatMultTranspose(mat,user->fine.x,Rscale);CHKERRQ(ierr);
ierr = VecReciprocal(Rscale);CHKERRQ(ierr);
user->Rscale = Rscale;
user->R = mat;
return 0;
}
void PETSC_STDCALL vecreciprocal_(Vec vec, int *__ierr ){
*__ierr = VecReciprocal(
(Vec)PetscToPointer((vec) ));
}
void _Stokes_SLE_UzawaSolver_Solve( void* solver, void* stokesSLE ) {
Stokes_SLE_UzawaSolver* self = (Stokes_SLE_UzawaSolver*)solver;
Stokes_SLE* sle = (Stokes_SLE*)stokesSLE;
/* Create shortcuts to stuff needed on sle */
Mat K_Mat = sle->kStiffMat->matrix;
Mat G_Mat = sle->gStiffMat->matrix;
Mat D_Mat = NULL;
Mat M_Mat = NULL;
Vec uVec = sle->uSolnVec->vector;
Vec qVec = sle->pSolnVec->vector;
Vec fVec = sle->fForceVec->vector;
Vec hVec = sle->hForceVec->vector;
/* Create shortcuts to solver related stuff */
Vec qTempVec = self->pTempVec;
Vec rVec = self->rVec;
Vec sVec = self->sVec;
Vec fTempVec = self->fTempVec;
Vec vStarVec = self->vStarVec;
KSP velSolver = self->velSolver; /* Inner velocity solver */
KSP pcSolver = self->pcSolver; /* Preconditioner */
Iteration_Index maxIterations = self->maxUzawaIterations;
Iteration_Index minIterations = self->minUzawaIterations;
Iteration_Index iteration_I = 0;
Iteration_Index outputInterval = 1;
double zdotr_current = 0.0;
double zdotr_previous = 1.0;
double sdotGTrans_v;
double alpha, beta;
double absResidual;
double relResidual;
double* chosenResidual; /* We can opt to use either the absolute or relative residual in termination condition */
double uzawaRhsScale;
double divU;
double weightedResidual;
double weightedVelocityScale;
double momentumEquationResidual;
Iteration_Index innerLoopIterations;
Stream* errorStream = Journal_Register( Error_Type, (Name)Stokes_SLE_UzawaSolver_Type );
PetscInt fVecSize, qTempVecSize, uVecSize, qVecSize;
PetscScalar fVecNorm, qTempVecNorm, uVecNorm, rVecNorm, fTempVecNorm, uVecNormInf, qVecNorm, qVecNormInf;
double qGlobalProblemScale;
double qReciprocalGlobalProblemScale;
int init_info_stream_rank;
PetscScalar p_sum;
/* Bool nullsp_present; */
Bool uzawa_summary;
double time,t0,rnorm0;
PetscTruth flg;
double ksptime;
VecGetSize( qTempVec, &qTempVecSize );
qGlobalProblemScale = sqrt( (double) qTempVecSize );
qReciprocalGlobalProblemScale = 1.0 / qGlobalProblemScale;
init_info_stream_rank = Stream_GetPrintingRank( self->info );
Stream_SetPrintingRank( self->info, 0 );
/* DEFINITIONS:
See accompanying documentation
u - the displacement / velocity solution (to which constraints are applied)
q - the pressure-like variable which constrains the divergence displacement / velocity (= pressure for incompressible)
F - standard FE force vector
Fhat - Uzawa RHS = K^{-1} G F - h
K - standard FE stiffness matrix
Khat - Uzawa transformed stiffness matrix = G^T K^{-1} G
G matrix - discrete gradient operator
D matrix - discrete divergence operator = G^T for this particular algorithm
C matrix - Mass matrix (M) for compressibility
LM & DAM
*/
/* CHOICE OF RESIDUAL:
we may opt to converge on the absolute value (self->useAbsoluteTolerance == True ... default)
or the relative value of the residual (self->useAbsoluteTolerance == False)
(another possibility would be always to improve the residual by a given tolerance)
The Moresi & Solomatov (Phys Fluids, 1995) approach is to use the relative tolerance
*/
VecNorm( fVec, NORM_2, &fVecNorm );
VecGetSize( fVec, &fVecSize );
if ( fVecNorm / sqrt( (double)fVecSize ) <= 1e-99 ) {
Journal_Printf( errorStream,
"Error in func %s: The momentum force vector \"%s\" is zero. "
"The force vector should be non-zero either because of your chosen boundary "
"conditions, or because of the element force vector assembly. You have %d "
"element force vectors attached.\n",
__func__, sle->fForceVec->name, sle->fForceVec->assembleForceVector->hooks->count );
if ( sle->fForceVec->assembleForceVector->hooks->count > 0 ) {
Journal_Printf( errorStream, "You used the following force vector assembly terms:\n" );
EntryPoint_PrintConcise( sle->fForceVec->assembleForceVector, errorStream );
/* TODO : need to print the elementForceVector assembly, not the global guy!! */
}
Journal_Printf( errorStream,
"Please check values for building the force vector.\n" );
Journal_Firewall( 0, errorStream, "Exiting.\n" );
}
Journal_DPrintf( self->debug, "In %s:\n", __func__ );
Journal_RPrintfL( self->debug, 2, "Conjugate Gradient Uzawa solver with:\n");
Stream_IndentBranch( StgFEM_Debug );
Journal_RPrintfL( self->debug, 2, "Compressibility %s\n", (sle->cStiffMat)? "on" : "off");
Journal_RPrintfL( self->debug, 2, "Preconditioning %s\n", (pcSolver)? "on" : "off" );
if ( sle->cStiffMat ) {
Journal_DPrintfL( self->debug, 2, "(compressibility active)\n" );
M_Mat = sle->cStiffMat->matrix;
}
else {
Journal_DPrintfL( self->debug, 2, "(compressibility inactive)\n" );
}
if ( sle->dStiffMat ) {
Journal_DPrintfL( self->debug, 2, "(asymmetric geometry: handling D Matrix [incorrectly - will be ignored])\n" );
D_Mat = sle->dStiffMat->matrix;
}
else {
Journal_DPrintfL( self->debug, 2, "(No D -> symmetric geometry: D = Gt)\n" );
}
#if DEBUG
if ( Stream_IsPrintableLevel( self->debug, 3 ) ) {
Journal_DPrintf( self->debug, "Matrices and Vectors to solve are:\n" );
Journal_DPrintf( self->debug, "K Matrix:\n" );
/* No nice way of viewing Matrices, so commented out as incompatible with
* new 3D decomp at present --Kathleen Humble 30-04-07
* Matrix_View( sle->kStiffMat->matrix, self->debug ); */
Journal_DPrintf( self->debug, "G Matrix:\n" );
if ( D_Mat ) {
Journal_DPrintf( self->debug, "D Matrix:\n" );
}
if ( M_Mat ) {
Journal_DPrintf( self->debug, "M Matrix:\n" );
}
Journal_DPrintf( self->debug, "Z (preconditioner) Matrix:\n" );
Journal_DPrintf( self->debug, "f Vector:\n" );
_SLE_VectorView( fVec, self->debug );
Journal_DPrintf( self->debug, "h Vector:\n" );
_SLE_VectorView( hVec, self->debug );
}
#endif
/* STEP 1: Estimate the magnitude of the RHS for the transformed problem
we compute (usually to lower accuracy than elsewhere) the RHS (Fhat - h)
and store the result in qTempVec.
LM & DAM
*/
Journal_DPrintfL( self->debug, 2, "Building Fhat - h.\n" );
PetscOptionsHasName(PETSC_NULL,"-uzawa_printksptimes",&flg);
KSPSetTolerances( velSolver, self->tolerance, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT );
if (flg) {
ksptime = MPI_Wtime();
}
KSPSolve( velSolver, fVec, vStarVec );
if (flg) {
ksptime = MPI_Wtime() - ksptime;
PetscPrintf( PETSC_COMM_WORLD, "KSP on velSolver took %lf seconds in Building Fhat step\n", ksptime);
}
KSPGetIterationNumber( velSolver, &innerLoopIterations );
Journal_DPrintfL( self->debug, 2, "Fhat inner solution: Number of iterations: %d\n", innerLoopIterations );
if ( D_Mat ) {
MatMult( D_Mat, vStarVec, qTempVec );
}
else {
MatMultTranspose( G_Mat, vStarVec, qTempVec );
}
VecAXPY( qTempVec, -1.0, hVec );
/* WARNING:
If D != G^T then the resulting \hat{K} is not likely to be symmetric, positive definite as
required by this implementation of the Uzawa iteration. This next piece of code
is VERY unlikely to work properly so it's in the sin bin for the time being - LM.
if ( D_Mat ) {
MatrixMultiply( D_Mat, vStarVec, qTempVec );
}
else {
MatrixTransposeMultiply( G_Mat, vStarVec, qTempVec );
}
LM & DAM
*/
/* STEP 2: The problem scaling - optionally normalize the uzawa residual by the magnitude of the RHS (use a relative tolerance)
For the inner velocity solver, Citcom uses a relative tolerance equal to that used for the Uzawa iteration as a whole
LM & DAM
*/
if (self->useAbsoluteTolerance) {
chosenResidual = &absResidual;
Journal_PrintfL( self->info, 2, "Absolute residual < %g for Uzawa stopping condition\n", self->tolerance);
/* We should calculate the effective relative tolerance and insert that here !! */
KSPSetTolerances( velSolver, 0.1 * self->tolerance, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT );
}
else { /* The CITCOM compatible choice */
chosenResidual = &relResidual;
Journal_PrintfL( self->info, 2, "Relative residual < %g for Uzawa stopping condition\n", self->tolerance);
KSPSetTolerances( velSolver, 0.1 * self->tolerance, PETSC_DEFAULT, PETSC_DEFAULT, PETSC_DEFAULT );
}
Journal_DPrintfL( self->debug, 2, "Determining scaling factor for residual:\n" );
VecNorm( qTempVec, NORM_2, &qTempVecNorm );
uzawaRhsScale = ((double)qTempVecNorm) * qReciprocalGlobalProblemScale;
Journal_DPrintfL( self->debug, 2, "uzawaRhsScale = %f\n", uzawaRhsScale );
Journal_Firewall( isGoodNumber( uzawaRhsScale ), errorStream,
"Error in func '%s' for %s '%s' - uzawaRhsScale has illegal value '%g'.\n", __func__, self->type, self->name, uzawaRhsScale );
/* STEP 3: Calculate initial residual for transformed equation (\hat{F} - h - \hat{K} q_0)
Compute the solution to K u_0 = F - G q_0 (u_0 unknown)
Then G^T u* = \hat{F} - \hat{K} q_0
u_0 is also the initial velocity solution to which the constraint is applied by the subsequent iteration
LM & DAM
*/
Journal_DPrintfL( self->debug, 2, "Solving for transformed Uzawa RHS.\n" );
VecCopy( fVec, fTempVec );
VecScale( fTempVec, -1.0 );
MatMultAdd( G_Mat, qVec, fTempVec, fTempVec );
VecScale( fTempVec, -1.0 );
KSPSolve( velSolver, fTempVec, uVec );
/* Handling for NON-SYMMETRIC: relegated to sin bin (see comment above) LM & DAM */
if ( D_Mat ) {
MatMult( D_Mat, uVec, rVec );
}
else {
MatMultTranspose( G_Mat, uVec, rVec );
}
VecNorm( rVec, NORM_2, &rnorm0 );
VecNorm( uVec, NORM_2, &uVecNorm );
divU = rnorm0 / uVecNorm;
Journal_PrintfL( self->info, 2, "Initial l2Norm( Div u ) / l2Norm( u ) = %f \n", divU);
Journal_Firewall( isGoodNumber( divU ), errorStream,
"Error in func '%s' for %s '%s' - l2Norm( Div u ) has illegal value '%g'.\n",
__func__, self->type, self->name, divU );
Journal_DPrintfL( self->debug, 2, "Adding compressibility and prescribed divergence terms.\n" );
if ( M_Mat ) {
MatMultAdd( M_Mat, qVec, rVec, rVec );
}
VecAXPY( rVec, -1.0, hVec );
/* Check for existence of constant null space */
#if 0
nullsp_present = _check_if_constant_nullsp_present( self, K_Mat,G_Mat,M_Mat, fTempVec,vStarVec,qTempVec,sVec, velSolver );
#endif
/* STEP 4: Preconditioned conjugate gradient iteration loop */
Journal_DPrintfL( self->debug, 1, "Beginning main Uzawa conjugate gradient loop:\n" );
iteration_I = 0;
/* outer_it, residual, time */
uzawa_summary = self->monitor;
time = 0.0;
t0 = MPI_Wtime();
// Journal_PrintfL( self->info, 1, " |r0| = %.8e \n", rnorm0 );
do{
/* reset initial time and end time for inner its back to 0 - probs don't need to do this but just in case */
self->outeritsinitialtime = 0;
self->outeritsendtime = 0;
//BEGINNING OF OUTER ITERATIONS!!!!!
/*get wall time for start of outer loop*/
self->outeritsinitialtime = MPI_Wtime();
Journal_DPrintfL( self->debug, 2, "Beginning solve '%u'.\n", iteration_I );
Stream_IndentBranch( StgFEM_Debug );
/* STEP 4.1: Preconditioner
Solve:
Q_\hat{K} z_1 = r_1
Q_\hat{K} is an approximation to \hat{K} which is simple / trivial / quick to invert
LM & DAM
*/
if ( pcSolver ) {
PetscOptionsHasName(PETSC_NULL,"-uzawa_printksptimes",&flg);
if (flg) {
ksptime = MPI_Wtime();
}
KSPSolve( pcSolver, rVec, qTempVec );
if (flg) {
ksptime = MPI_Wtime() - ksptime;
PetscPrintf( PETSC_COMM_WORLD, "KSP on pcSolver took %lf seconds\n", ksptime);
}
}
else {
VecCopy( rVec, qTempVec );
}
/* Remove the constant null space, but only if NOT compressible */
#if 0
if( nullsp_present == True ) {
_remove_constant_nullsp( qTempVec );
}
#endif
/* STEP 4.2: Calculate s_I, the pressure search direction
z_{I-1} . r_{I-1}
\beta = (z_{I-1} . r_{I-1}) / (z_{I-2} . r_{I-2})
\beta = 0 for the first iteration
s_I = z_(I-1) + \beta * s_(I-1)
LM & DAM
*/
VecDot( qTempVec, rVec, &zdotr_current );
VecNorm( qTempVec, NORM_2, &qTempVecNorm );
VecNorm( rVec, NORM_2, &rVecNorm );
Journal_DPrintfL( self->debug, 2, "l2Norm (qTempVec) %g; (rVec) %g \n",
qTempVecNorm * qReciprocalGlobalProblemScale,
rVecNorm * qReciprocalGlobalProblemScale );
if ( iteration_I == 0 ) {
VecCopy( qTempVec, sVec );
}
else {
beta = zdotr_current/zdotr_previous;
VecAYPX( sVec, beta, qTempVec );
}
/* STEP 4.3: Velocity search direction corresponding to s_I is found by solving
K u* = G s_I
LM & DAM
*/
MatMult( G_Mat, sVec, fTempVec );
Journal_DPrintfL( self->debug, 2, "Uzawa inner iteration step\n");
//START OF INNER ITERATIONS!!!!
PetscOptionsHasName(PETSC_NULL,"-uzawa_printksptimes",&flg);
/*get initial wall time for inner loop*/
self->inneritsinitialtime = MPI_Wtime();
if (flg) {
ksptime = MPI_Wtime();
}
KSPSolve( velSolver, fTempVec, vStarVec );
if (flg) {
ksptime = MPI_Wtime() - ksptime;
PetscPrintf( PETSC_COMM_WORLD, "KSP on velSolver took %lf seconds in Uzawa inner iteration step\n", ksptime);
}
/*get end wall time for inner loop*/
self->inneritsendtime = MPI_Wtime();
/* add time to total time inner its: */
self->totalinneritstime = self->totalinneritstime + (-self->inneritsinitialtime + self->inneritsendtime);
/* reset initial time and end time for inner its back to 0 - probs don't need to do this but just in case */
self->inneritsinitialtime = 0;
self->inneritsendtime = 0;
KSPGetIterationNumber( velSolver, &innerLoopIterations );
/* add the inner loop iterations to the total inner iterations */
self->totalnuminnerits = self->totalnuminnerits + innerLoopIterations;
Journal_DPrintfL( self->debug, 2, "Completed Uzawa inner iteration in '%u' iterations \n", innerLoopIterations );
/* STEP 4.4: Calculate the step size ( \alpha = z_{I-1} . r_{I-1} / (s_I . \hat{K} s_I) )
\hat{K} s_I = G^T u* - M s_I (u* from step 4.3)
LM & DAM
*/
if ( D_Mat ) {
MatMult( D_Mat, vStarVec, qTempVec );
}
else {
MatMultTranspose( G_Mat, vStarVec, qTempVec );
}
/* Handling for NON-SYMMETRIC: relegated to sin bin (see comment above)
if ( D_Mat ) {
MatrixMultiply( D_Mat, vStarVec, qTempVec );
}
else {
MatrixTransposeMultiply( G_Mat, vStarVec, qTempVec );
}
LM & DAM
*/
if ( M_Mat ) {
Journal_DPrintfL( self->debug, 2, "Correcting for Compressibility\n" );
VecScale( qTempVec, -1.0 );
MatMultAdd( M_Mat, sVec, qTempVec, qTempVec );
VecScale( qTempVec, -1.0 );
}
VecDot( sVec, qTempVec, &sdotGTrans_v );
alpha = zdotr_current/sdotGTrans_v;
/* STEP 4.5: Update pressure, velocity and value of residual
by \alpha times corresponding search direction
LM & DAM
*/
Journal_DPrintfL( self->debug, 2, "zdotr_current = %g \n", zdotr_current);
Journal_DPrintfL( self->debug, 2, "sdotGTrans_v = %g \n", sdotGTrans_v);
Journal_DPrintfL( self->debug, 2, "alpha = %g \n", alpha);
Journal_Firewall(
isGoodNumber( zdotr_current ) && isGoodNumber( sdotGTrans_v ) && isGoodNumber( alpha ),
errorStream,
"Error in func '%s' for %s '%s' - zdotr_current, sdotGTrans_v or alpha has an illegal value: '%g','%g' or '%g'\n",
__func__, self->type, self->name, zdotr_current, sdotGTrans_v, alpha );
VecAXPY( qVec, alpha, sVec );
VecAXPY( uVec, -alpha, vStarVec );
VecAXPY( rVec, -alpha, qTempVec );
/* STEP 4.6: store the value of z_{I-1} . r_{I-1} for the next iteration
LM & DAM
*/
zdotr_previous = zdotr_current;
VecNorm( rVec, NORM_2, &rVecNorm );
absResidual = rVecNorm * qReciprocalGlobalProblemScale;
relResidual = absResidual / uzawaRhsScale;
Stream_UnIndentBranch( StgFEM_Debug );
if( iteration_I % outputInterval == 0 ) {
Journal_PrintfL( self->info, 2, "\tLoop = %u, absResidual = %.8e, relResidual = %.8e\n",
iteration_I, absResidual, relResidual );
}
Journal_Firewall( isGoodNumber( absResidual ), errorStream,
"Error in func '%s' for %s '%s' - absResidual has an illegal value: '%g'\n",
__func__, self->type, self->name, absResidual );
Journal_Firewall( iteration_I < maxIterations,
errorStream, "In func %s: Reached maximum number of iterations %u without converging; absResidual = %.5g, relResidual = %.5g \n",
__func__, iteration_I, absResidual, relResidual );
/* TODO: test for small change in 10 iterations and if so restart? */
time = MPI_Wtime()-t0;
if (uzawa_summary) {
Journal_PrintfL( self->info, 1, " %1.4d uzawa residual norm %12.13e, cpu time %5.5e\n", iteration_I+1,*chosenResidual,time );
}
iteration_I++;
//END OF OUTER ITERATION LOOP!!!
/*get wall time for end of outer loop*/
self->outeritsendtime = MPI_Wtime();
/* add time to total time inner its: */
self->totalouteritstime = self->totalouteritstime + (-self->outeritsinitialtime + self->outeritsendtime);
/* reset initial time and end time for inner its back to 0 - probs don't need to do this but just in case */
self->outeritsinitialtime = 0;
self->outeritsendtime = 0;
/* add the outer loop iterations to the total outer iterations */
self->totalnumouterits++;
} while ( (*chosenResidual > self->tolerance) || (iteration_I<minIterations) );
// } while ( *chosenResidual > self->tolerance );
Journal_DPrintfL( self->debug, 1, "Pressure solution converged. Exiting uzawa \n ");
/* STEP 5: Check all the relevant residuals and report back */
if (Stream_IsEnable( self->info ) ) {
/* This information should be in an info stream */
Journal_PrintfL( self->info, 1, "Summary:\n");
Journal_PrintfL( self->info, 1, " Uzawa its. = %04d , Uzawa residual = %12.13e\n", iteration_I, relResidual );
MatMultTranspose( G_Mat, uVec, rVec );
VecNorm( rVec, NORM_2, &rVecNorm );
VecNorm( uVec, NORM_2, &uVecNorm );
divU = rVecNorm / uVecNorm;
Journal_PrintfL( self->info, 1, " |G^T u|/|u| = %.8e\n", divU);
/* Residual for the momentum equation
Compute r = || F - Ku - Gp || / || F ||
*/
MatMult( G_Mat, qVec, vStarVec );
MatMultAdd( K_Mat, uVec, vStarVec, fTempVec );
VecAYPX( fTempVec, -1.0, fVec );
VecNorm( fTempVec, NORM_2, &fTempVecNorm );
VecNorm( fVec, NORM_2, &fVecNorm );
momentumEquationResidual = fTempVecNorm / fVecNorm;
Journal_PrintfL( self->info, 1, " |f - K u - G p|/|f| = %.8e\n", momentumEquationResidual );
Journal_Firewall( isGoodNumber( momentumEquationResidual ), errorStream,
"Bad residual for the momentum equation (|| F - Ku - Gp || / || F || = %g):\n"
"\tCheck to see if forcing term is zero or nan - \n\t|| F - Ku - Gp || = %g \n\t|| F || = %g.\n",
momentumEquationResidual,
fTempVecNorm, fVecNorm );
/* "Preconditioned" residual for the momentum equation
r_{w} = || Q_{K}(r) || / || Q_{K}(F)
fTempVec contains the residual but is overwritten once used
vStarVec is used to hold the diagonal preconditioner Q_{K}
*/
MatGetDiagonal( K_Mat, vStarVec );
VecReciprocal( vStarVec );
VecPointwiseMult( vStarVec, fTempVec, fTempVec );
VecNorm( fTempVec, NORM_2, &weightedResidual );
VecPointwiseMult( vStarVec, fVec, fTempVec );
VecNorm( fTempVec, NORM_2, &weightedVelocityScale );
Journal_PrintfL( self->info, 1, " |f - K u - G p|_w/|f|_w = %.8e\n", weightedResidual / weightedVelocityScale );
/* Report back on the solution - velocity and pressure
Note - correction for dof in Vrms ??
*/
VecNorm( uVec, NORM_INFINITY, &uVecNormInf );
VecNorm( uVec, NORM_2, &uVecNorm );
VecGetSize( uVec, &uVecSize );
VecNorm( qVec, NORM_INFINITY, &qVecNormInf );
VecNorm( qVec, NORM_2, &qVecNorm );
VecGetSize( qVec, &qVecSize );
Journal_PrintfL( self->info, 1, " |u|_{\\infty} = %.8e , u_rms = %.8e\n",
uVecNormInf, uVecNorm / sqrt( (double)uVecSize ) );
Journal_PrintfL( self->info, 1, " |p|_{\\infty} = %.8e , p_rms = %.8e\n",
qVecNormInf, qVecNorm / sqrt( (double)qVecSize ) );
{ PetscInt lmin,lmax;
PetscReal min,max;
VecMax( uVec, &lmax, &max );
VecMin( uVec, &lmin, &min );
Journal_PrintfL( self->info, 1, " min/max(u) = %.8e [%d] / %.8e [%d]\n",min,lmin,max,lmax);
VecMax( qVec, &lmax, &max );
VecMin( qVec, &lmin, &min );
Journal_PrintfL( self->info, 1, " min/max(p) = %.8e [%d] / %.8e [%d]\n",min,lmin,max,lmax);
}
VecSum( qVec, &p_sum );
Journal_PrintfL( self->info, 1, " \\sum_i p_i = %.8e \n", p_sum );
} /* journal stream enabled */
#if DEBUG
if ( Stream_IsPrintableLevel( self->debug, 3 ) ) {
Journal_DPrintf( self->debug, "Velocity solution:\n" );
_SLE_VectorView( uVec, self->debug );
Journal_DPrintf( self->debug, "Pressure solution:\n" );
_SLE_VectorView( qVec, self->debug );
}
#endif
Stream_UnIndentBranch( StgFEM_Debug );
Stream_SetPrintingRank( self->info, init_info_stream_rank );
/* Now gather up data for printing out to FrequentOutput file: */
/*!!! if non-linear need to divide by number of nonlinear iterations and we do this in SystemLinearEquations */
if((sle->isNonLinear != True)){
self->avgnuminnerits = self->totalnuminnerits/self->totalnumouterits;
self->avgnumouterits = self->totalnumouterits;
self->avgtimeouterits = (self->totalouteritstime - self->totalinneritstime)/self->totalnumouterits;
self->avgtimeinnerits = self->totalinneritstime/self->totalnuminnerits;
}
}
static PetscErrorCode PCSetUp_Jacobi(PC pc)
{
PC_Jacobi *jac = (PC_Jacobi*)pc->data;
Vec diag,diagsqrt;
PetscErrorCode ierr;
PetscInt n,i;
PetscScalar *x;
PetscBool zeroflag = PETSC_FALSE;
PetscFunctionBegin;
/*
For most preconditioners the code would begin here something like
if (pc->setupcalled == 0) { allocate space the first time this is ever called
ierr = MatCreateVecs(pc->mat,&jac->diag);CHKERRQ(ierr);
PetscLogObjectParent((PetscObject)pc,(PetscObject)jac->diag);
}
But for this preconditioner we want to support use of both the matrix' diagonal
elements (for left or right preconditioning) and square root of diagonal elements
(for symmetric preconditioning). Hence we do not allocate space here, since we
don't know at this point which will be needed (diag and/or diagsqrt) until the user
applies the preconditioner, and we don't want to allocate BOTH unless we need
them both. Thus, the diag and diagsqrt are allocated in PCSetUp_Jacobi_NonSymmetric()
and PCSetUp_Jacobi_Symmetric(), respectively.
*/
/*
Here we set up the preconditioner; that is, we copy the diagonal values from
the matrix and put them into a format to make them quick to apply as a preconditioner.
*/
diag = jac->diag;
diagsqrt = jac->diagsqrt;
if (diag) {
if (jac->userowmax) {
ierr = MatGetRowMaxAbs(pc->pmat,diag,NULL);CHKERRQ(ierr);
} else if (jac->userowsum) {
ierr = MatGetRowSum(pc->pmat,diag);CHKERRQ(ierr);
} else {
ierr = MatGetDiagonal(pc->pmat,diag);CHKERRQ(ierr);
}
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecGetLocalSize(diag,&n);CHKERRQ(ierr);
ierr = VecGetArray(diag,&x);CHKERRQ(ierr);
if (jac->useabs) {
for (i=0; i<n; i++) x[i] = PetscAbsScalar(x[i]);
}
for (i=0; i<n; i++) {
if (x[i] == 0.0) {
x[i] = 1.0;
zeroflag = PETSC_TRUE;
}
}
ierr = VecRestoreArray(diag,&x);CHKERRQ(ierr);
}
if (diagsqrt) {
if (jac->userowmax) {
ierr = MatGetRowMaxAbs(pc->pmat,diagsqrt,NULL);CHKERRQ(ierr);
} else if (jac->userowsum) {
ierr = MatGetRowSum(pc->pmat,diagsqrt);CHKERRQ(ierr);
} else {
ierr = MatGetDiagonal(pc->pmat,diagsqrt);CHKERRQ(ierr);
}
ierr = VecGetLocalSize(diagsqrt,&n);CHKERRQ(ierr);
ierr = VecGetArray(diagsqrt,&x);CHKERRQ(ierr);
for (i=0; i<n; i++) {
if (x[i] != 0.0) x[i] = 1.0/PetscSqrtReal(PetscAbsScalar(x[i]));
else {
x[i] = 1.0;
zeroflag = PETSC_TRUE;
}
}
ierr = VecRestoreArray(diagsqrt,&x);CHKERRQ(ierr);
}
if (zeroflag) {
ierr = PetscInfo(pc,"Zero detected in diagonal of matrix, using 1 at those locations\n");CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
void _Stokes_SLE_PenaltySolver_Solve( void* solver,void* stokesSLE ) {
Stokes_SLE_PenaltySolver* self = (Stokes_SLE_PenaltySolver*)solver;
Stokes_SLE* sle = (Stokes_SLE*)stokesSLE;
/* Create shortcuts to stuff needed on sle */
Mat kMatrix = sle->kStiffMat->matrix;
Mat gradMat = sle->gStiffMat->matrix;
Mat divMat = NULL;
Mat C_Mat = sle->cStiffMat->matrix;
Vec uVec = sle->uSolnVec->vector;
Vec pVec = sle->pSolnVec->vector;
Vec fVec = sle->fForceVec->vector;
Vec hVec = sle->hForceVec->vector;
Vec hTempVec;
Vec fTempVec;
Vec penalty;
Mat GTrans, kHat;
KSP ksp_v;
double negOne=-1.0;
double one=1.0;
Mat C_InvMat;
Vec diagC;
PC pc;
int rank;
MPI_Comm_rank( MPI_COMM_WORLD, &rank );
Journal_DPrintf( self->debug, "In %s():\n", __func__ );
VecDuplicate( hVec, &hTempVec );
VecDuplicate( fVec, &fTempVec );
VecDuplicate( pVec, &diagC );
if( sle->dStiffMat == NULL ) {
Journal_DPrintf( self->debug, "Div matrix == NULL : Problem is assumed to be symmetric. ie Div = GTrans \n");
#if( PETSC_VERSION_MAJOR <= 2 )
MatTranspose( gradMat, >rans );
#else
MatTranspose( gradMat, MAT_INITIAL_MATRIX, >rans );
#endif
divMat = GTrans;
}
else {
MatType type;
PetscInt size[2];
MatGetType( sle->dStiffMat->matrix, &type );
MatGetLocalSize( sle->dStiffMat->matrix, size + 0, size + 1 );
/* make a copy we can play with */
MatCreate( sle->comm, >rans );
MatSetSizes( GTrans, size[0], size[1], PETSC_DECIDE, PETSC_DECIDE );
MatSetType( GTrans, type );
#if (((PETSC_VERSION_MAJOR==3) && (PETSC_VERSION_MINOR>=3)) || (PETSC_VERSION_MAJOR>3) )
MatSetUp(GTrans);
#endif
MatCopy( sle->dStiffMat->matrix, GTrans, DIFFERENT_NONZERO_PATTERN );
divMat = GTrans;
}
Stokes_SLE_PenaltySolver_MakePenalty( self, sle, &penalty );
/* Create CInv */
MatGetDiagonal( C_Mat, diagC );
VecReciprocal( diagC );
VecPointwiseMult( diagC, penalty, diagC );
{ /* Print the maximum and minimum penalties in my system. */
PetscInt idx;
PetscReal min, max;
VecMin( diagC, &idx, &min );
VecMax( diagC, &idx, &max );
if( rank == 0 ) {
printf( "PENALTY RANGE:\n" );
printf( " MIN: %e\n", min );
printf( " MAX: %e\n", max );
}
}
MatDiagonalSet( C_Mat, diagC, INSERT_VALUES );
C_InvMat = C_Mat; /* Use pointer CInv since C has been inverted */
/* Build RHS : rhs = f - GCInv h */
MatMult( C_InvMat, hVec, hTempVec ); /* hTempVec = C_InvMat * hVec */
VecScale( hTempVec, -1.0 );
MatMult( gradMat, hTempVec, fTempVec );
#if 0
VecPointwiseMult( fTempVec, penalty, fTempVec );
{ /* Print the maximum and minimum penalties in my system. */
PetscInt idx;
PetscReal min, max;
VecMin( fTempVec, &idx, &min );
VecMax( fTempVec, &idx, &max );
printf( "PENALTY RANGE:\n" );
printf( " MIN: %e\n", min );
printf( " MAX: %e\n", max );
}
#endif
VecAXPY( fTempVec, 1.0, fVec );
/*MatMultAdd( gradMat, hTempVec, fVec, fTempVec );*/
/* Build G CInv GTrans */
/* MatTranspose( gradMat, >rans ); */
/* since CInv is diagonal we can just scale mat entries by the diag vector */
MatDiagonalScale( divMat, diagC, PETSC_NULL ); /* Div = CInve Div */
MatMatMult( gradMat, divMat, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &kHat );
/*MatDiagonalScale( kHat, penalty, PETSC_NULL );*/
MatScale( kHat, -1.0 );
MatAXPY( kMatrix, 1.0, kHat, SAME_NONZERO_PATTERN );
/* Setup solver context and make sure that it uses a direct solver */
KSPCreate( sle->comm, &ksp_v );
Stg_KSPSetOperators( ksp_v, kMatrix, kMatrix, DIFFERENT_NONZERO_PATTERN );
KSPSetType( ksp_v, KSPPREONLY );
KSPGetPC( ksp_v, &pc );
PCSetType( pc, PCLU );
KSPSetFromOptions( ksp_v );
KSPSolve( ksp_v, fTempVec, uVec );
/* Recover p */
if( sle->dStiffMat == NULL ) {
/* since Div was modified when C is diagonal, re build the transpose */
if( GTrans != PETSC_NULL )
Stg_MatDestroy(>rans );
#if( PETSC_VERSION_MAJOR <= 2 )
MatTranspose( gradMat, >rans );
#else
MatTranspose( gradMat, MAT_INITIAL_MATRIX, >rans );
#endif
divMat = GTrans;
}
else {
/* never modified Div_null so set divMat to point back to it */
divMat = sle->dStiffMat->matrix;
}
MatMult( divMat, uVec, hTempVec ); /* hTemp = Div v */
VecAYPX( hTempVec, negOne, hVec ); /* hTemp = H - hTemp : hTemp = H - Div v */
MatMult( C_InvMat, hTempVec, pVec ); /* p = CInv hTemp : p = CInv ( H - Div v ) */
Stg_MatDestroy(&kHat );
if( fTempVec != PETSC_NULL ) Stg_VecDestroy(&fTempVec );
if( hTempVec != PETSC_NULL ) Stg_VecDestroy(&hTempVec );
if( diagC != PETSC_NULL ) Stg_VecDestroy(&diagC );
if( ksp_v != PETSC_NULL ) Stg_KSPDestroy(&ksp_v );
if( GTrans != PETSC_NULL ) Stg_MatDestroy(>rans );
}
static PetscErrorCode QPIPSetInitialPoint(TAO_BQPIP *qp, Tao tao)
{
PetscErrorCode ierr;
PetscReal two=2.0,p01=1;
PetscReal gap1,gap2,fff,mu;
PetscFunctionBegin;
/* Compute function, Gradient R=Hx+b, and Hessian */
ierr = TaoComputeVariableBounds(tao);CHKERRQ(ierr);
ierr = VecMedian(qp->XL, tao->solution, qp->XU, tao->solution);CHKERRQ(ierr);
ierr = MatMult(tao->hessian, tao->solution, tao->gradient);CHKERRQ(ierr);
ierr = VecCopy(qp->C0, qp->Work);CHKERRQ(ierr);
ierr = VecAXPY(qp->Work, 0.5, tao->gradient);CHKERRQ(ierr);
ierr = VecAXPY(tao->gradient, 1.0, qp->C0);CHKERRQ(ierr);
ierr = VecDot(tao->solution, qp->Work, &fff);CHKERRQ(ierr);
qp->pobj = fff + qp->c;
/* Initialize Primal Vectors */
/* T = XU - X; G = X - XL */
ierr = VecCopy(qp->XU, qp->T);CHKERRQ(ierr);
ierr = VecAXPY(qp->T, -1.0, tao->solution);CHKERRQ(ierr);
ierr = VecCopy(tao->solution, qp->G);CHKERRQ(ierr);
ierr = VecAXPY(qp->G, -1.0, qp->XL);CHKERRQ(ierr);
ierr = VecSet(qp->GZwork, p01);CHKERRQ(ierr);
ierr = VecSet(qp->TSwork, p01);CHKERRQ(ierr);
ierr = VecPointwiseMax(qp->G, qp->G, qp->GZwork);CHKERRQ(ierr);
ierr = VecPointwiseMax(qp->T, qp->T, qp->TSwork);CHKERRQ(ierr);
/* Initialize Dual Variable Vectors */
ierr = VecCopy(qp->G, qp->Z);CHKERRQ(ierr);
ierr = VecReciprocal(qp->Z);CHKERRQ(ierr);
ierr = VecCopy(qp->T, qp->S);CHKERRQ(ierr);
ierr = VecReciprocal(qp->S);CHKERRQ(ierr);
ierr = MatMult(tao->hessian, qp->Work, qp->RHS);CHKERRQ(ierr);
ierr = VecAbs(qp->RHS);CHKERRQ(ierr);
ierr = VecSet(qp->Work, p01);CHKERRQ(ierr);
ierr = VecPointwiseMax(qp->RHS, qp->RHS, qp->Work);CHKERRQ(ierr);
ierr = VecPointwiseDivide(qp->RHS, tao->gradient, qp->RHS);CHKERRQ(ierr);
ierr = VecNorm(qp->RHS, NORM_1, &gap1);CHKERRQ(ierr);
mu = PetscMin(10.0,(gap1+10.0)/qp->m);
ierr = VecScale(qp->S, mu);CHKERRQ(ierr);
ierr = VecScale(qp->Z, mu);CHKERRQ(ierr);
ierr = VecSet(qp->TSwork, p01);CHKERRQ(ierr);
ierr = VecSet(qp->GZwork, p01);CHKERRQ(ierr);
ierr = VecPointwiseMax(qp->S, qp->S, qp->TSwork);CHKERRQ(ierr);
ierr = VecPointwiseMax(qp->Z, qp->Z, qp->GZwork);CHKERRQ(ierr);
qp->mu=0;qp->dinfeas=1.0;qp->pinfeas=1.0;
while ( (qp->dinfeas+qp->pinfeas)/(qp->m+qp->n) >= qp->mu ){
ierr = VecScale(qp->G, two);CHKERRQ(ierr);
ierr = VecScale(qp->Z, two);CHKERRQ(ierr);
ierr = VecScale(qp->S, two);CHKERRQ(ierr);
ierr = VecScale(qp->T, two);CHKERRQ(ierr);
ierr = QPIPComputeResidual(qp,tao);CHKERRQ(ierr);
ierr = VecCopy(tao->solution, qp->R3);CHKERRQ(ierr);
ierr = VecAXPY(qp->R3, -1.0, qp->G);CHKERRQ(ierr);
ierr = VecAXPY(qp->R3, -1.0, qp->XL);CHKERRQ(ierr);
ierr = VecCopy(tao->solution, qp->R5);CHKERRQ(ierr);
ierr = VecAXPY(qp->R5, 1.0, qp->T);CHKERRQ(ierr);
ierr = VecAXPY(qp->R5, -1.0, qp->XU);CHKERRQ(ierr);
ierr = VecNorm(qp->R3, NORM_INFINITY, &gap1);CHKERRQ(ierr);
ierr = VecNorm(qp->R5, NORM_INFINITY, &gap2);CHKERRQ(ierr);
qp->pinfeas=PetscMax(gap1,gap2);
/* Compute the duality gap */
ierr = VecDot(qp->G, qp->Z, &gap1);CHKERRQ(ierr);
ierr = VecDot(qp->T, qp->S, &gap2);CHKERRQ(ierr);
qp->gap = (gap1+gap2);
qp->dobj = qp->pobj - qp->gap;
if (qp->m>0) qp->mu=qp->gap/(qp->m); else qp->mu=0.0;
qp->rgap=qp->gap/( PetscAbsReal(qp->dobj) + PetscAbsReal(qp->pobj) + 1.0 );
}
PetscFunctionReturn(0);
}
static PetscErrorCode TaoSolve_NTR(Tao tao)
{
TAO_NTR *tr = (TAO_NTR *)tao->data;
PC pc;
KSPConvergedReason ksp_reason;
TaoConvergedReason reason;
PetscReal fmin, ftrial, prered, actred, kappa, sigma, beta;
PetscReal tau, tau_1, tau_2, tau_max, tau_min, max_radius;
PetscReal f, gnorm;
PetscReal delta;
PetscReal norm_d;
PetscErrorCode ierr;
PetscInt iter = 0;
PetscInt bfgsUpdates = 0;
PetscInt needH;
PetscInt i_max = 5;
PetscInt j_max = 1;
PetscInt i, j, N, n, its;
PetscFunctionBegin;
if (tao->XL || tao->XU || tao->ops->computebounds) {
ierr = PetscPrintf(((PetscObject)tao)->comm,"WARNING: Variable bounds have been set but will be ignored by ntr algorithm\n");CHKERRQ(ierr);
}
tao->trust = tao->trust0;
/* Modify the radius if it is too large or small */
tao->trust = PetscMax(tao->trust, tr->min_radius);
tao->trust = PetscMin(tao->trust, tr->max_radius);
if (NTR_PC_BFGS == tr->pc_type && !tr->M) {
ierr = VecGetLocalSize(tao->solution,&n);CHKERRQ(ierr);
ierr = VecGetSize(tao->solution,&N);CHKERRQ(ierr);
ierr = MatCreateLMVM(((PetscObject)tao)->comm,n,N,&tr->M);CHKERRQ(ierr);
ierr = MatLMVMAllocateVectors(tr->M,tao->solution);CHKERRQ(ierr);
}
/* Check convergence criteria */
ierr = TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient);CHKERRQ(ierr);
ierr = VecNorm(tao->gradient,NORM_2,&gnorm);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
needH = 1;
ierr = TaoMonitor(tao, iter, f, gnorm, 0.0, 1.0, &reason);CHKERRQ(ierr);
if (reason != TAO_CONTINUE_ITERATING) PetscFunctionReturn(0);
/* Create vectors for the limited memory preconditioner */
if ((NTR_PC_BFGS == tr->pc_type) &&
(BFGS_SCALE_BFGS != tr->bfgs_scale_type)) {
if (!tr->Diag) {
ierr = VecDuplicate(tao->solution, &tr->Diag);CHKERRQ(ierr);
}
}
switch(tr->ksp_type) {
case NTR_KSP_NASH:
ierr = KSPSetType(tao->ksp, KSPNASH);CHKERRQ(ierr);
if (tao->ksp->ops->setfromoptions) {
(*tao->ksp->ops->setfromoptions)(tao->ksp);
}
break;
case NTR_KSP_STCG:
ierr = KSPSetType(tao->ksp, KSPSTCG);CHKERRQ(ierr);
if (tao->ksp->ops->setfromoptions) {
(*tao->ksp->ops->setfromoptions)(tao->ksp);
}
break;
default:
ierr = KSPSetType(tao->ksp, KSPGLTR);CHKERRQ(ierr);
if (tao->ksp->ops->setfromoptions) {
(*tao->ksp->ops->setfromoptions)(tao->ksp);
}
break;
}
/* Modify the preconditioner to use the bfgs approximation */
ierr = KSPGetPC(tao->ksp, &pc);CHKERRQ(ierr);
switch(tr->pc_type) {
case NTR_PC_NONE:
ierr = PCSetType(pc, PCNONE);CHKERRQ(ierr);
if (pc->ops->setfromoptions) {
(*pc->ops->setfromoptions)(pc);
}
break;
case NTR_PC_AHESS:
ierr = PCSetType(pc, PCJACOBI);CHKERRQ(ierr);
if (pc->ops->setfromoptions) {
(*pc->ops->setfromoptions)(pc);
}
ierr = PCJacobiSetUseAbs(pc);CHKERRQ(ierr);
break;
case NTR_PC_BFGS:
ierr = PCSetType(pc, PCSHELL);CHKERRQ(ierr);
if (pc->ops->setfromoptions) {
(*pc->ops->setfromoptions)(pc);
}
ierr = PCShellSetName(pc, "bfgs");CHKERRQ(ierr);
ierr = PCShellSetContext(pc, tr->M);CHKERRQ(ierr);
ierr = PCShellSetApply(pc, MatLMVMSolveShell);CHKERRQ(ierr);
break;
default:
/* Use the pc method set by pc_type */
break;
}
/* Initialize trust-region radius */
switch(tr->init_type) {
case NTR_INIT_CONSTANT:
/* Use the initial radius specified */
break;
case NTR_INIT_INTERPOLATION:
/* Use the initial radius specified */
max_radius = 0.0;
for (j = 0; j < j_max; ++j) {
fmin = f;
sigma = 0.0;
if (needH) {
ierr = TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr);
needH = 0;
}
for (i = 0; i < i_max; ++i) {
ierr = VecCopy(tao->solution, tr->W);CHKERRQ(ierr);
ierr = VecAXPY(tr->W, -tao->trust/gnorm, tao->gradient);CHKERRQ(ierr);
ierr = TaoComputeObjective(tao, tr->W, &ftrial);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(ftrial)) {
tau = tr->gamma1_i;
}
else {
if (ftrial < fmin) {
fmin = ftrial;
sigma = -tao->trust / gnorm;
}
ierr = MatMult(tao->hessian, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
ierr = VecDot(tao->gradient, tao->stepdirection, &prered);CHKERRQ(ierr);
prered = tao->trust * (gnorm - 0.5 * tao->trust * prered / (gnorm * gnorm));
actred = f - ftrial;
if ((PetscAbsScalar(actred) <= tr->epsilon) &&
(PetscAbsScalar(prered) <= tr->epsilon)) {
kappa = 1.0;
}
else {
kappa = actred / prered;
}
tau_1 = tr->theta_i * gnorm * tao->trust / (tr->theta_i * gnorm * tao->trust + (1.0 - tr->theta_i) * prered - actred);
tau_2 = tr->theta_i * gnorm * tao->trust / (tr->theta_i * gnorm * tao->trust - (1.0 + tr->theta_i) * prered + actred);
tau_min = PetscMin(tau_1, tau_2);
tau_max = PetscMax(tau_1, tau_2);
if (PetscAbsScalar(kappa - 1.0) <= tr->mu1_i) {
/* Great agreement */
max_radius = PetscMax(max_radius, tao->trust);
if (tau_max < 1.0) {
tau = tr->gamma3_i;
}
else if (tau_max > tr->gamma4_i) {
tau = tr->gamma4_i;
}
else {
tau = tau_max;
}
}
else if (PetscAbsScalar(kappa - 1.0) <= tr->mu2_i) {
/* Good agreement */
max_radius = PetscMax(max_radius, tao->trust);
if (tau_max < tr->gamma2_i) {
tau = tr->gamma2_i;
}
else if (tau_max > tr->gamma3_i) {
tau = tr->gamma3_i;
}
else {
tau = tau_max;
}
}
else {
/* Not good agreement */
if (tau_min > 1.0) {
tau = tr->gamma2_i;
}
else if (tau_max < tr->gamma1_i) {
tau = tr->gamma1_i;
}
else if ((tau_min < tr->gamma1_i) && (tau_max >= 1.0)) {
tau = tr->gamma1_i;
}
else if ((tau_1 >= tr->gamma1_i) && (tau_1 < 1.0) &&
((tau_2 < tr->gamma1_i) || (tau_2 >= 1.0))) {
tau = tau_1;
}
else if ((tau_2 >= tr->gamma1_i) && (tau_2 < 1.0) &&
((tau_1 < tr->gamma1_i) || (tau_2 >= 1.0))) {
tau = tau_2;
}
else {
tau = tau_max;
}
}
}
tao->trust = tau * tao->trust;
}
if (fmin < f) {
f = fmin;
ierr = VecAXPY(tao->solution, sigma, tao->gradient);CHKERRQ(ierr);
ierr = TaoComputeGradient(tao,tao->solution, tao->gradient);CHKERRQ(ierr);
ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
needH = 1;
ierr = TaoMonitor(tao, iter, f, gnorm, 0.0, 1.0, &reason);CHKERRQ(ierr);
if (reason != TAO_CONTINUE_ITERATING) {
PetscFunctionReturn(0);
}
}
}
tao->trust = PetscMax(tao->trust, max_radius);
/* Modify the radius if it is too large or small */
tao->trust = PetscMax(tao->trust, tr->min_radius);
tao->trust = PetscMin(tao->trust, tr->max_radius);
break;
default:
/* Norm of the first direction will initialize radius */
tao->trust = 0.0;
break;
}
/* Set initial scaling for the BFGS preconditioner
This step is done after computing the initial trust-region radius
since the function value may have decreased */
if (NTR_PC_BFGS == tr->pc_type) {
if (f != 0.0) {
delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
}
else {
delta = 2.0 / (gnorm*gnorm);
}
ierr = MatLMVMSetDelta(tr->M,delta);CHKERRQ(ierr);
}
/* Have not converged; continue with Newton method */
while (reason == TAO_CONTINUE_ITERATING) {
++iter;
tao->ksp_its=0;
/* Compute the Hessian */
if (needH) {
ierr = TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);CHKERRQ(ierr);
needH = 0;
}
if (NTR_PC_BFGS == tr->pc_type) {
if (BFGS_SCALE_AHESS == tr->bfgs_scale_type) {
/* Obtain diagonal for the bfgs preconditioner */
ierr = MatGetDiagonal(tao->hessian, tr->Diag);CHKERRQ(ierr);
ierr = VecAbs(tr->Diag);CHKERRQ(ierr);
ierr = VecReciprocal(tr->Diag);CHKERRQ(ierr);
ierr = MatLMVMSetScale(tr->M,tr->Diag);CHKERRQ(ierr);
}
/* Update the limited memory preconditioner */
ierr = MatLMVMUpdate(tr->M, tao->solution, tao->gradient);CHKERRQ(ierr);
++bfgsUpdates;
}
while (reason == TAO_CONTINUE_ITERATING) {
ierr = KSPSetOperators(tao->ksp, tao->hessian, tao->hessian_pre);CHKERRQ(ierr);
/* Solve the trust region subproblem */
if (NTR_KSP_NASH == tr->ksp_type) {
ierr = KSPNASHSetRadius(tao->ksp,tao->trust);CHKERRQ(ierr);
ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
tao->ksp_its+=its;
tao->ksp_tot_its+=its;
ierr = KSPNASHGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
} else if (NTR_KSP_STCG == tr->ksp_type) {
ierr = KSPSTCGSetRadius(tao->ksp,tao->trust);CHKERRQ(ierr);
ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
tao->ksp_its+=its;
tao->ksp_tot_its+=its;
ierr = KSPSTCGGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
} else { /* NTR_KSP_GLTR */
ierr = KSPGLTRSetRadius(tao->ksp,tao->trust);CHKERRQ(ierr);
ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
tao->ksp_its+=its;
tao->ksp_tot_its+=its;
ierr = KSPGLTRGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
}
if (0.0 == tao->trust) {
/* Radius was uninitialized; use the norm of the direction */
if (norm_d > 0.0) {
tao->trust = norm_d;
/* Modify the radius if it is too large or small */
tao->trust = PetscMax(tao->trust, tr->min_radius);
tao->trust = PetscMin(tao->trust, tr->max_radius);
}
else {
/* The direction was bad; set radius to default value and re-solve
the trust-region subproblem to get a direction */
tao->trust = tao->trust0;
/* Modify the radius if it is too large or small */
tao->trust = PetscMax(tao->trust, tr->min_radius);
tao->trust = PetscMin(tao->trust, tr->max_radius);
if (NTR_KSP_NASH == tr->ksp_type) {
ierr = KSPNASHSetRadius(tao->ksp,tao->trust);CHKERRQ(ierr);
ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
tao->ksp_its+=its;
tao->ksp_tot_its+=its;
ierr = KSPNASHGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
} else if (NTR_KSP_STCG == tr->ksp_type) {
ierr = KSPSTCGSetRadius(tao->ksp,tao->trust);CHKERRQ(ierr);
ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
tao->ksp_its+=its;
tao->ksp_tot_its+=its;
ierr = KSPSTCGGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
} else { /* NTR_KSP_GLTR */
ierr = KSPGLTRSetRadius(tao->ksp,tao->trust);CHKERRQ(ierr);
ierr = KSPSolve(tao->ksp, tao->gradient, tao->stepdirection);CHKERRQ(ierr);
ierr = KSPGetIterationNumber(tao->ksp,&its);CHKERRQ(ierr);
tao->ksp_its+=its;
tao->ksp_tot_its+=its;
ierr = KSPGLTRGetNormD(tao->ksp, &norm_d);CHKERRQ(ierr);
}
if (norm_d == 0.0) SETERRQ(PETSC_COMM_SELF,1, "Initial direction zero");
}
}
ierr = VecScale(tao->stepdirection, -1.0);CHKERRQ(ierr);
ierr = KSPGetConvergedReason(tao->ksp, &ksp_reason);CHKERRQ(ierr);
if ((KSP_DIVERGED_INDEFINITE_PC == ksp_reason) &&
(NTR_PC_BFGS == tr->pc_type) && (bfgsUpdates > 1)) {
/* Preconditioner is numerically indefinite; reset the
approximate if using BFGS preconditioning. */
if (f != 0.0) {
delta = 2.0 * PetscAbsScalar(f) / (gnorm*gnorm);
}
else {
delta = 2.0 / (gnorm*gnorm);
}
ierr = MatLMVMSetDelta(tr->M, delta);CHKERRQ(ierr);
ierr = MatLMVMReset(tr->M);CHKERRQ(ierr);
ierr = MatLMVMUpdate(tr->M, tao->solution, tao->gradient);CHKERRQ(ierr);
bfgsUpdates = 1;
}
if (NTR_UPDATE_REDUCTION == tr->update_type) {
/* Get predicted reduction */
if (NTR_KSP_NASH == tr->ksp_type) {
ierr = KSPNASHGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
} else if (NTR_KSP_STCG == tr->ksp_type) {
ierr = KSPSTCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
} else { /* gltr */
ierr = KSPGLTRGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
}
if (prered >= 0.0) {
/* The predicted reduction has the wrong sign. This cannot
happen in infinite precision arithmetic. Step should
be rejected! */
tao->trust = tr->alpha1 * PetscMin(tao->trust, norm_d);
}
else {
/* Compute trial step and function value */
ierr = VecCopy(tao->solution,tr->W);CHKERRQ(ierr);
ierr = VecAXPY(tr->W, 1.0, tao->stepdirection);CHKERRQ(ierr);
ierr = TaoComputeObjective(tao, tr->W, &ftrial);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(ftrial)) {
tao->trust = tr->alpha1 * PetscMin(tao->trust, norm_d);
} else {
/* Compute and actual reduction */
actred = f - ftrial;
prered = -prered;
if ((PetscAbsScalar(actred) <= tr->epsilon) &&
(PetscAbsScalar(prered) <= tr->epsilon)) {
kappa = 1.0;
}
else {
kappa = actred / prered;
}
/* Accept or reject the step and update radius */
if (kappa < tr->eta1) {
/* Reject the step */
tao->trust = tr->alpha1 * PetscMin(tao->trust, norm_d);
}
else {
/* Accept the step */
if (kappa < tr->eta2) {
/* Marginal bad step */
tao->trust = tr->alpha2 * PetscMin(tao->trust, norm_d);
}
else if (kappa < tr->eta3) {
/* Reasonable step */
tao->trust = tr->alpha3 * tao->trust;
}
else if (kappa < tr->eta4) {
/* Good step */
tao->trust = PetscMax(tr->alpha4 * norm_d, tao->trust);
}
else {
/* Very good step */
tao->trust = PetscMax(tr->alpha5 * norm_d, tao->trust);
}
break;
}
}
}
}
else {
/* Get predicted reduction */
if (NTR_KSP_NASH == tr->ksp_type) {
ierr = KSPNASHGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
} else if (NTR_KSP_STCG == tr->ksp_type) {
ierr = KSPSTCGGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
} else { /* gltr */
ierr = KSPGLTRGetObjFcn(tao->ksp,&prered);CHKERRQ(ierr);
}
if (prered >= 0.0) {
/* The predicted reduction has the wrong sign. This cannot
happen in infinite precision arithmetic. Step should
be rejected! */
tao->trust = tr->gamma1 * PetscMin(tao->trust, norm_d);
}
else {
ierr = VecCopy(tao->solution, tr->W);CHKERRQ(ierr);
ierr = VecAXPY(tr->W, 1.0, tao->stepdirection);CHKERRQ(ierr);
ierr = TaoComputeObjective(tao, tr->W, &ftrial);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(ftrial)) {
tao->trust = tr->gamma1 * PetscMin(tao->trust, norm_d);
}
else {
ierr = VecDot(tao->gradient, tao->stepdirection, &beta);CHKERRQ(ierr);
actred = f - ftrial;
prered = -prered;
if ((PetscAbsScalar(actred) <= tr->epsilon) &&
(PetscAbsScalar(prered) <= tr->epsilon)) {
kappa = 1.0;
}
else {
kappa = actred / prered;
}
tau_1 = tr->theta * beta / (tr->theta * beta - (1.0 - tr->theta) * prered + actred);
tau_2 = tr->theta * beta / (tr->theta * beta + (1.0 + tr->theta) * prered - actred);
tau_min = PetscMin(tau_1, tau_2);
tau_max = PetscMax(tau_1, tau_2);
if (kappa >= 1.0 - tr->mu1) {
/* Great agreement; accept step and update radius */
if (tau_max < 1.0) {
tao->trust = PetscMax(tao->trust, tr->gamma3 * norm_d);
}
else if (tau_max > tr->gamma4) {
tao->trust = PetscMax(tao->trust, tr->gamma4 * norm_d);
}
else {
tao->trust = PetscMax(tao->trust, tau_max * norm_d);
}
break;
}
else if (kappa >= 1.0 - tr->mu2) {
/* Good agreement */
if (tau_max < tr->gamma2) {
tao->trust = tr->gamma2 * PetscMin(tao->trust, norm_d);
}
else if (tau_max > tr->gamma3) {
tao->trust = PetscMax(tao->trust, tr->gamma3 * norm_d);
}
else if (tau_max < 1.0) {
tao->trust = tau_max * PetscMin(tao->trust, norm_d);
}
else {
tao->trust = PetscMax(tao->trust, tau_max * norm_d);
}
break;
}
else {
/* Not good agreement */
if (tau_min > 1.0) {
tao->trust = tr->gamma2 * PetscMin(tao->trust, norm_d);
}
else if (tau_max < tr->gamma1) {
tao->trust = tr->gamma1 * PetscMin(tao->trust, norm_d);
}
else if ((tau_min < tr->gamma1) && (tau_max >= 1.0)) {
tao->trust = tr->gamma1 * PetscMin(tao->trust, norm_d);
}
else if ((tau_1 >= tr->gamma1) && (tau_1 < 1.0) &&
((tau_2 < tr->gamma1) || (tau_2 >= 1.0))) {
tao->trust = tau_1 * PetscMin(tao->trust, norm_d);
}
else if ((tau_2 >= tr->gamma1) && (tau_2 < 1.0) &&
((tau_1 < tr->gamma1) || (tau_2 >= 1.0))) {
tao->trust = tau_2 * PetscMin(tao->trust, norm_d);
}
else {
tao->trust = tau_max * PetscMin(tao->trust, norm_d);
}
}
}
}
}
/* The step computed was not good and the radius was decreased.
Monitor the radius to terminate. */
ierr = TaoMonitor(tao, iter, f, gnorm, 0.0, tao->trust, &reason);CHKERRQ(ierr);
}
/* The radius may have been increased; modify if it is too large */
tao->trust = PetscMin(tao->trust, tr->max_radius);
if (reason == TAO_CONTINUE_ITERATING) {
ierr = VecCopy(tr->W, tao->solution);CHKERRQ(ierr);
f = ftrial;
ierr = TaoComputeGradient(tao, tao->solution, tao->gradient);
ierr = VecNorm(tao->gradient, NORM_2, &gnorm);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(f) || PetscIsInfOrNanReal(gnorm)) SETERRQ(PETSC_COMM_SELF,1, "User provided compute function generated Inf or NaN");
needH = 1;
ierr = TaoMonitor(tao, iter, f, gnorm, 0.0, tao->trust, &reason);CHKERRQ(ierr);
}
}
PetscFunctionReturn(0);
}
