PetscErrorCode SNESQNApply_BadBroyden(SNES snes,PetscInt it,Vec Y,Vec X,Vec Xold,Vec D,Vec Dold)
{
PetscErrorCode ierr;
SNES_QN *qn = (SNES_QN*)snes->data;
Vec W = snes->work[3];
Vec *U = qn->U;
Vec *T = qn->V;
/* ksp thing for Jacobian scaling */
PetscInt h,k,j,i,lits;
PetscInt m = qn->m;
PetscScalar gdot,udot;
PetscInt l = m;
PetscFunctionBegin;
if (it < m) l = it;
ierr = VecCopy(D,Y);CHKERRQ(ierr);
if (l > 0) {
k = (it-1)%l;
ierr = SNESLineSearchGetLambda(snes->linesearch,&qn->lambda[k]);CHKERRQ(ierr);
ierr = VecCopy(Dold,U[k]);CHKERRQ(ierr);
ierr = VecAXPY(U[k],-1.0,D);CHKERRQ(ierr);
ierr = VecCopy(Xold,T[k]);CHKERRQ(ierr);
ierr = VecAXPY(T[k],-1.0,X);CHKERRQ(ierr);
}
if (qn->scale_type == SNES_QN_SCALE_JACOBIAN) {
ierr = KSPSolve(snes->ksp,Y,W);CHKERRQ(ierr);
SNESCheckKSPSolve(snes);
ierr = KSPGetIterationNumber(snes->ksp,&lits);CHKERRQ(ierr);
snes->linear_its += lits;
ierr = VecCopy(W,Y);CHKERRQ(ierr);
} else {
ierr = VecScale(Y,qn->scaling);CHKERRQ(ierr);
}
/* inward recursion starting at the first update and working forward */
if (l > 0) {
for (i = 0; i < l-1; i++) {
j = (it+i-l)%l;
k = (it+i-l+1)%l;
h = (it-1)%l;
ierr = VecDotBegin(U[j],U[h],&gdot);CHKERRQ(ierr);
ierr = VecDotBegin(U[j],U[j],&udot);CHKERRQ(ierr);
ierr = VecDotEnd(U[j],U[h],&gdot);CHKERRQ(ierr);
ierr = VecDotEnd(U[j],U[j],&udot);CHKERRQ(ierr);
ierr = VecAXPY(Y,PetscRealPart(gdot)/PetscRealPart(udot),T[k]);CHKERRQ(ierr);
ierr = VecAXPY(Y,-(1.-qn->lambda[k])*PetscRealPart(gdot)/PetscRealPart(udot),T[j]);CHKERRQ(ierr);
if (qn->monitor) {
ierr = PetscViewerASCIIAddTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(qn->monitor, "it: %d k: %d gdot: %14.12e\n", it, k, PetscRealPart(gdot));CHKERRQ(ierr);
ierr = PetscViewerASCIISubtractTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
}
}
}
PetscFunctionReturn(0);
}
/*
MatMFFDCompute_DS - Standard PETSc code for computing the
differencing paramter (h) for use with matrix-free finite differences.
Input Parameters:
+ ctx - the matrix free context
. U - the location at which you want the Jacobian
- a - the direction you want the derivative
Output Parameter:
. h - the scale computed
*/
static PetscErrorCode MatMFFDCompute_DS(MatMFFD ctx,Vec U,Vec a,PetscScalar *h,PetscBool *zeroa)
{
MatMFFD_DS *hctx = (MatMFFD_DS*)ctx->hctx;
PetscReal nrm,sum,umin = hctx->umin;
PetscScalar dot;
PetscErrorCode ierr;
PetscFunctionBegin;
if (!(ctx->count % ctx->recomputeperiod)) {
/*
This algorithm requires 2 norms and 1 inner product. Rather than
use directly the VecNorm() and VecDot() routines (and thus have
three separate collective operations, we use the VecxxxBegin/End() routines
*/
ierr = VecDotBegin(U,a,&dot);CHKERRQ(ierr);
ierr = VecNormBegin(a,NORM_1,&sum);CHKERRQ(ierr);
ierr = VecNormBegin(a,NORM_2,&nrm);CHKERRQ(ierr);
ierr = VecDotEnd(U,a,&dot);CHKERRQ(ierr);
ierr = VecNormEnd(a,NORM_1,&sum);CHKERRQ(ierr);
ierr = VecNormEnd(a,NORM_2,&nrm);CHKERRQ(ierr);
if (nrm == 0.0) {
*zeroa = PETSC_TRUE;
PetscFunctionReturn(0);
}
*zeroa = PETSC_FALSE;
/*
Safeguard for step sizes that are "too small"
*/
if (PetscAbsScalar(dot) < umin*sum && PetscRealPart(dot) >= 0.0) dot = umin*sum;
else if (PetscAbsScalar(dot) < 0.0 && PetscRealPart(dot) > -umin*sum) dot = -umin*sum;
*h = ctx->error_rel*dot/(nrm*nrm);
} else {
*h = ctx->currenth;
}
if (*h != *h) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Differencing parameter is not a number sum = %g dot = %g norm = %g",(double)sum,(double)PetscRealPart(dot),(double)nrm);
ctx->count++;
PetscFunctionReturn(0);
}
PetscErrorCode SNESSolve_NCG(SNES snes)
{
SNES_NCG *ncg = (SNES_NCG*)snes->data;
Vec X,dX,lX,F,dXold;
PetscReal fnorm, ynorm, xnorm, beta = 0.0;
PetscScalar dXdotdX, dXolddotdXold, dXdotdXold, lXdotdX, lXdotdXold;
PetscInt maxits, i;
PetscErrorCode ierr;
PetscBool lsSuccess = PETSC_TRUE;
SNESLineSearch linesearch;
SNESConvergedReason reason;
PetscFunctionBegin;
ierr = PetscCitationsRegister(SNESCitation,&SNEScite);CHKERRQ(ierr);
snes->reason = SNES_CONVERGED_ITERATING;
maxits = snes->max_its; /* maximum number of iterations */
X = snes->vec_sol; /* X^n */
dXold = snes->work[0]; /* The previous iterate of X */
dX = snes->work[1]; /* the preconditioned direction */
lX = snes->vec_sol_update; /* the conjugate direction */
F = snes->vec_func; /* residual vector */
ierr = SNESGetLineSearch(snes, &linesearch);CHKERRQ(ierr);
ierr = PetscObjectSAWsTakeAccess((PetscObject)snes);CHKERRQ(ierr);
snes->iter = 0;
snes->norm = 0.;
ierr = PetscObjectSAWsGrantAccess((PetscObject)snes);CHKERRQ(ierr);
/* compute the initial function and preconditioned update dX */
if (snes->pc && snes->functype == SNES_FUNCTION_PRECONDITIONED) {
ierr = SNESApplyNPC(snes,X,NULL,dX);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = VecCopy(dX,F);CHKERRQ(ierr);
ierr = VecNorm(F,NORM_2,&fnorm);CHKERRQ(ierr);
} else {
if (!snes->vec_func_init_set) {
ierr = SNESComputeFunction(snes,X,F);CHKERRQ(ierr);
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
PetscFunctionReturn(0);
}
} else snes->vec_func_init_set = PETSC_FALSE;
/* convergence test */
ierr = VecNorm(F,NORM_2,&fnorm);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(fnorm)) {
snes->reason = SNES_DIVERGED_FNORM_NAN;
PetscFunctionReturn(0);
}
ierr = VecCopy(F,dX);CHKERRQ(ierr);
}
if (snes->pc) {
if (snes->functype == SNES_FUNCTION_UNPRECONDITIONED) {
ierr = SNESApplyNPC(snes,X,F,dX);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
}
}
ierr = VecCopy(dX,lX);CHKERRQ(ierr);
ierr = VecDot(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = PetscObjectSAWsTakeAccess((PetscObject)snes);CHKERRQ(ierr);
snes->norm = fnorm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)snes);CHKERRQ(ierr);
ierr = SNESLogConvergenceHistory(snes,fnorm,0);CHKERRQ(ierr);
ierr = SNESMonitor(snes,0,fnorm);CHKERRQ(ierr);
/* test convergence */
ierr = (*snes->ops->converged)(snes,0,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
/* Call general purpose update function */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
}
/* first update -- just use the (preconditioned) residual direction for the initial conjugate direction */
for (i = 1; i < maxits + 1; i++) {
lsSuccess = PETSC_TRUE;
/* some update types require the old update direction or conjugate direction */
if (ncg->type != SNES_NCG_FR) {
ierr = VecCopy(dX, dXold);CHKERRQ(ierr);
}
ierr = SNESLineSearchApply(linesearch,X,F,&fnorm,lX);CHKERRQ(ierr);
ierr = SNESLineSearchGetSuccess(linesearch, &lsSuccess);CHKERRQ(ierr);
if (!lsSuccess) {
if (++snes->numFailures >= snes->maxFailures) {
snes->reason = SNES_DIVERGED_LINE_SEARCH;
PetscFunctionReturn(0);
}
}
if (snes->nfuncs >= snes->max_funcs) {
snes->reason = SNES_DIVERGED_FUNCTION_COUNT;
PetscFunctionReturn(0);
}
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
PetscFunctionReturn(0);
}
ierr = SNESLineSearchGetNorms(linesearch, &xnorm, &fnorm, &ynorm);CHKERRQ(ierr);
/* Monitor convergence */
ierr = PetscObjectSAWsTakeAccess((PetscObject)snes);CHKERRQ(ierr);
snes->iter = i;
snes->norm = fnorm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)snes);CHKERRQ(ierr);
ierr = SNESLogConvergenceHistory(snes,snes->norm,0);CHKERRQ(ierr);
ierr = SNESMonitor(snes,snes->iter,snes->norm);CHKERRQ(ierr);
/* Test for convergence */
ierr = (*snes->ops->converged)(snes,snes->iter,xnorm,ynorm,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
/* Call general purpose update function */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
}
if (snes->pc) {
if (snes->functype == SNES_FUNCTION_PRECONDITIONED) {
ierr = SNESApplyNPC(snes,X,NULL,dX);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = VecCopy(dX,F);CHKERRQ(ierr);
} else {
ierr = SNESApplyNPC(snes,X,F,dX);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
}
} else {
ierr = VecCopy(F,dX);CHKERRQ(ierr);
}
/* compute the conjugate direction lX = dX + beta*lX with beta = ((dX, dX) / (dX_old, dX_old) (Fletcher-Reeves update)*/
switch (ncg->type) {
case SNES_NCG_FR: /* Fletcher-Reeves */
dXolddotdXold= dXdotdX;
ierr = VecDot(dX, dX, &dXdotdX);CHKERRQ(ierr);
beta = PetscRealPart(dXdotdX / dXolddotdXold);
break;
case SNES_NCG_PRP: /* Polak-Ribiere-Poylak */
dXolddotdXold= dXdotdX;
ierr = VecDotBegin(dX, dX, &dXdotdXold);CHKERRQ(ierr);
ierr = VecDotBegin(dXold, dX, &dXdotdXold);CHKERRQ(ierr);
ierr = VecDotEnd(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = VecDotEnd(dXold, dX, &dXdotdXold);CHKERRQ(ierr);
beta = PetscRealPart(((dXdotdX - dXdotdXold) / dXolddotdXold));
if (beta < 0.0) beta = 0.0; /* restart */
break;
case SNES_NCG_HS: /* Hestenes-Stiefel */
ierr = VecDotBegin(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = VecDotBegin(dX, dXold, &dXdotdXold);CHKERRQ(ierr);
ierr = VecDotBegin(lX, dX, &lXdotdX);CHKERRQ(ierr);
ierr = VecDotBegin(lX, dXold, &lXdotdXold);CHKERRQ(ierr);
ierr = VecDotEnd(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = VecDotEnd(dX, dXold, &dXdotdXold);CHKERRQ(ierr);
ierr = VecDotEnd(lX, dX, &lXdotdX);CHKERRQ(ierr);
ierr = VecDotEnd(lX, dXold, &lXdotdXold);CHKERRQ(ierr);
beta = PetscRealPart((dXdotdX - dXdotdXold) / (lXdotdX - lXdotdXold));
break;
case SNES_NCG_DY: /* Dai-Yuan */
ierr = VecDotBegin(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = VecDotBegin(lX, dX, &lXdotdX);CHKERRQ(ierr);
ierr = VecDotBegin(lX, dXold, &lXdotdXold);CHKERRQ(ierr);
ierr = VecDotEnd(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = VecDotEnd(lX, dX, &lXdotdX);CHKERRQ(ierr);
ierr = VecDotEnd(lX, dXold, &lXdotdXold);CHKERRQ(ierr);
beta = PetscRealPart(dXdotdX / (lXdotdXold - lXdotdX));CHKERRQ(ierr);
break;
case SNES_NCG_CD: /* Conjugate Descent */
ierr = VecDotBegin(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = VecDotBegin(lX, dXold, &lXdotdXold);CHKERRQ(ierr);
ierr = VecDotEnd(dX, dX, &dXdotdX);CHKERRQ(ierr);
ierr = VecDotEnd(lX, dXold, &lXdotdXold);CHKERRQ(ierr);
beta = PetscRealPart(dXdotdX / lXdotdXold);CHKERRQ(ierr);
break;
}
if (ncg->monitor) {
ierr = PetscViewerASCIIPrintf(ncg->monitor, "beta = %e\n", (double)beta);CHKERRQ(ierr);
}
ierr = VecAYPX(lX, beta, dX);CHKERRQ(ierr);
}
ierr = PetscInfo1(snes, "Maximum number of iterations has been reached: %D\n", maxits);CHKERRQ(ierr);
if (!snes->reason) snes->reason = SNES_DIVERGED_MAX_IT;
PetscFunctionReturn(0);
}
/* approximately solve the overdetermined system:
2*F(x_i)\cdot F(\x_j)\alpha_i = 0
\alpha_i = 1
Which minimizes the L2 norm of the linearization of:
||F(\sum_i \alpha_i*x_i)||^2
With the constraint that \sum_i\alpha_i = 1
Where x_i is the solution from the ith subsolver.
*/
static PetscErrorCode SNESCompositeApply_AdditiveOptimal(SNES snes,Vec X,Vec B,Vec F,PetscReal *fnorm)
{
PetscErrorCode ierr;
SNES_Composite *jac = (SNES_Composite*)snes->data;
SNES_CompositeLink next = jac->head;
Vec *Xes = jac->Xes,*Fes = jac->Fes;
PetscInt i,j;
PetscScalar tot,total,ftf;
PetscReal min_fnorm;
PetscInt min_i;
SNESConvergedReason reason;
PetscFunctionBegin;
if (!next) SETERRQ(PetscObjectComm((PetscObject)snes),PETSC_ERR_ARG_WRONGSTATE,"No composite SNESes supplied via SNESCompositeAddSNES() or -snes_composite_sneses");
if (snes->normschedule == SNES_NORM_ALWAYS) {
next = jac->head;
ierr = SNESSetInitialFunction(next->snes,F);CHKERRQ(ierr);
while (next->next) {
next = next->next;
ierr = SNESSetInitialFunction(next->snes,F);CHKERRQ(ierr);
}
}
next = jac->head;
i = 0;
ierr = VecCopy(X,Xes[i]);CHKERRQ(ierr);
ierr = SNESSolve(next->snes,B,Xes[i]);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(next->snes,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
jac->innerFailures++;
if (jac->innerFailures >= snes->maxFailures) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
}
while (next->next) {
i++;
next = next->next;
ierr = VecCopy(X,Xes[i]);CHKERRQ(ierr);
ierr = SNESSolve(next->snes,B,Xes[i]);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(next->snes,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
jac->innerFailures++;
if (jac->innerFailures >= snes->maxFailures) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
}
}
/* all the solutions are collected; combine optimally */
for (i=0;i<jac->n;i++) {
for (j=0;j<i+1;j++) {
ierr = VecDotBegin(Fes[i],Fes[j],&jac->h[i + j*jac->n]);CHKERRQ(ierr);
}
ierr = VecDotBegin(Fes[i],F,&jac->g[i]);CHKERRQ(ierr);
}
for (i=0;i<jac->n;i++) {
for (j=0;j<i+1;j++) {
ierr = VecDotEnd(Fes[i],Fes[j],&jac->h[i + j*jac->n]);CHKERRQ(ierr);
if (i == j) jac->fnorms[i] = PetscSqrtReal(PetscRealPart(jac->h[i + j*jac->n]));
}
ierr = VecDotEnd(Fes[i],F,&jac->g[i]);CHKERRQ(ierr);
}
ftf = (*fnorm)*(*fnorm);
for (i=0; i<jac->n; i++) {
for (j=i+1;j<jac->n;j++) {
jac->h[i + j*jac->n] = jac->h[j + i*jac->n];
}
}
for (i=0; i<jac->n; i++) {
for (j=0; j<jac->n; j++) {
jac->h[i + j*jac->n] = jac->h[i + j*jac->n] - jac->g[j] - jac->g[i] + ftf;
}
jac->beta[i] = ftf - jac->g[i];
}
#if defined(PETSC_MISSING_LAPACK_GELSS)
SETERRQ(PetscObjectComm((PetscObject)snes),PETSC_ERR_SUP,"SNESCOMPOSITE with ADDITIVEOPTIMAL requires the LAPACK GELSS routine.");
#else
jac->info = 0;
jac->rcond = -1.;
ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
#if defined(PETSC_USE_COMPLEX)
PetscStackCall("LAPACKgelss",LAPACKgelss_(&jac->n,&jac->n,&jac->nrhs,jac->h,&jac->lda,jac->beta,&jac->lda,jac->s,&jac->rcond,&jac->rank,jac->work,&jac->lwork,jac->rwork,&jac->info));
#else
PetscStackCall("LAPACKgelss",LAPACKgelss_(&jac->n,&jac->n,&jac->nrhs,jac->h,&jac->lda,jac->beta,&jac->lda,jac->s,&jac->rcond,&jac->rank,jac->work,&jac->lwork,&jac->info));
#endif
ierr = PetscFPTrapPop();CHKERRQ(ierr);
if (jac->info < 0) SETERRQ(PetscObjectComm((PetscObject)snes),PETSC_ERR_LIB,"Bad argument to GELSS");
if (jac->info > 0) SETERRQ(PetscObjectComm((PetscObject)snes),PETSC_ERR_LIB,"SVD failed to converge");
#endif
tot = 0.;
total = 0.;
for (i=0; i<jac->n; i++) {
if (snes->errorifnotconverged && PetscIsInfOrNanScalar(jac->beta[i])) SETERRQ(PetscObjectComm((PetscObject)snes),PETSC_ERR_LIB,"SVD generated inconsistent output");
ierr = PetscInfo2(snes,"%D: %g\n",i,(double)PetscRealPart(jac->beta[i]));CHKERRQ(ierr);
tot += jac->beta[i];
total += PetscAbsScalar(jac->beta[i]);
}
ierr = VecScale(X,(1. - tot));CHKERRQ(ierr);
ierr = VecMAXPY(X,jac->n,jac->beta,Xes);CHKERRQ(ierr);
ierr = SNESComputeFunction(snes,X,F);CHKERRQ(ierr);
if (snes->xl && snes->xu) {
ierr = SNESVIComputeInactiveSetFnorm(snes, F, X, fnorm);CHKERRQ(ierr);
} else {
ierr = VecNorm(F, NORM_2, fnorm);CHKERRQ(ierr);
}
/* take the minimum-normed candidate if it beats the combination by a factor of rtol or the combination has stagnated */
min_fnorm = jac->fnorms[0];
min_i = 0;
for (i=0; i<jac->n; i++) {
if (jac->fnorms[i] < min_fnorm) {
min_fnorm = jac->fnorms[i];
min_i = i;
}
}
/* stagnation or divergence restart to the solution of the solver that failed the least */
if (PetscRealPart(total) < jac->stol || min_fnorm*jac->rtol < *fnorm) {
ierr = VecCopy(jac->Xes[min_i],X);CHKERRQ(ierr);
ierr = VecCopy(jac->Fes[min_i],F);CHKERRQ(ierr);
*fnorm = min_fnorm;
}
PetscFunctionReturn(0);
}
EXTERN_C_END
/*
SNESSolve_NCG - Solves a nonlinear system with the Nonlinear Conjugate Gradient method.
Input Parameters:
. snes - the SNES context
Output Parameter:
. outits - number of iterations until termination
Application Interface Routine: SNESSolve()
*/
#undef __FUNCT__
#define __FUNCT__ "SNESSolve_NCG"
PetscErrorCode SNESSolve_NCG(SNES snes)
{
SNES_NCG *ncg = (SNES_NCG *)snes->data;
Vec X, dX, lX, F, B, Fold;
PetscReal fnorm, ynorm, xnorm, beta = 0.0;
PetscScalar dXdotF, dXolddotFold, dXdotFold, lXdotF, lXdotFold;
PetscInt maxits, i;
PetscErrorCode ierr;
SNESConvergedReason reason;
PetscBool lsSuccess = PETSC_TRUE;
SNESLineSearch linesearch;
PetscFunctionBegin;
snes->reason = SNES_CONVERGED_ITERATING;
maxits = snes->max_its; /* maximum number of iterations */
X = snes->vec_sol; /* X^n */
Fold = snes->work[0]; /* The previous iterate of X */
dX = snes->work[1]; /* the preconditioned direction */
lX = snes->vec_sol_update; /* the conjugate direction */
F = snes->vec_func; /* residual vector */
B = snes->vec_rhs; /* the right hand side */
ierr = SNESGetSNESLineSearch(snes, &linesearch);
CHKERRQ(ierr);
ierr = PetscObjectTakeAccess(snes);
CHKERRQ(ierr);
snes->iter = 0;
snes->norm = 0.;
ierr = PetscObjectGrantAccess(snes);
CHKERRQ(ierr);
/* compute the initial function and preconditioned update dX */
if (!snes->vec_func_init_set) {
ierr = SNESComputeFunction(snes,X,F);
CHKERRQ(ierr);
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
PetscFunctionReturn(0);
}
} else {
snes->vec_func_init_set = PETSC_FALSE;
}
if (!snes->norm_init_set) {
/* convergence test */
ierr = VecNorm(F, NORM_2, &fnorm);
CHKERRQ(ierr); /* fnorm <- ||F|| */
if (PetscIsInfOrNanReal(fnorm)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FP,"Infinite or not-a-number generated in norm");
} else {
fnorm = snes->norm_init;
snes->norm_init_set = PETSC_FALSE;
}
ierr = PetscObjectTakeAccess(snes);
CHKERRQ(ierr);
snes->norm = fnorm;
ierr = PetscObjectGrantAccess(snes);
CHKERRQ(ierr);
SNESLogConvHistory(snes,fnorm,0);
ierr = SNESMonitor(snes,0,fnorm);
CHKERRQ(ierr);
/* set parameter for default relative tolerance convergence test */
snes->ttol = fnorm*snes->rtol;
/* test convergence */
ierr = (*snes->ops->converged)(snes,0,0.0,0.0,fnorm,&snes->reason,snes->cnvP);
CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
/* Call general purpose update function */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);
CHKERRQ(ierr);
}
/* first update -- just use the (preconditioned) residual direction for the initial conjugate direction */
if (snes->pc && snes->pcside == PC_RIGHT) {
ierr = VecCopy(X, dX);
CHKERRQ(ierr);
ierr = SNESSetInitialFunction(snes->pc, F);
CHKERRQ(ierr);
ierr = SNESSetInitialFunctionNorm(snes->pc, fnorm);
CHKERRQ(ierr);
ierr = SNESSolve(snes->pc, B, dX);
CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);
CHKERRQ(ierr);
if (reason < 0 && (reason != SNES_DIVERGED_MAX_IT)) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = VecAYPX(dX,-1.0,X);
CHKERRQ(ierr);
} else {
ierr = VecCopy(F, dX);
CHKERRQ(ierr);
}
ierr = VecCopy(dX, lX);
CHKERRQ(ierr);
ierr = VecDot(F, dX, &dXdotF);
CHKERRQ(ierr);
/*
} else {
ierr = SNESNCGComputeYtJtF_Private(snes, X, F, dX, W, G, &dXdotF);CHKERRQ(ierr);
}
*/
for (i = 1; i < maxits + 1; i++) {
lsSuccess = PETSC_TRUE;
/* some update types require the old update direction or conjugate direction */
if (ncg->type != SNES_NCG_FR) {
ierr = VecCopy(F, Fold);
CHKERRQ(ierr);
}
ierr = SNESLineSearchApply(linesearch, X, F, &fnorm, lX);
CHKERRQ(ierr);
ierr = SNESLineSearchGetSuccess(linesearch, &lsSuccess);
CHKERRQ(ierr);
if (!lsSuccess) {
if (++snes->numFailures >= snes->maxFailures) {
snes->reason = SNES_DIVERGED_LINE_SEARCH;
PetscFunctionReturn(0);
}
}
if (snes->nfuncs >= snes->max_funcs) {
snes->reason = SNES_DIVERGED_FUNCTION_COUNT;
PetscFunctionReturn(0);
}
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
PetscFunctionReturn(0);
}
ierr = SNESLineSearchGetNorms(linesearch, &xnorm, &fnorm, &ynorm);
CHKERRQ(ierr);
/* Monitor convergence */
ierr = PetscObjectTakeAccess(snes);
CHKERRQ(ierr);
snes->iter = i;
snes->norm = fnorm;
ierr = PetscObjectGrantAccess(snes);
CHKERRQ(ierr);
SNESLogConvHistory(snes,snes->norm,0);
ierr = SNESMonitor(snes,snes->iter,snes->norm);
CHKERRQ(ierr);
/* Test for convergence */
ierr = (*snes->ops->converged)(snes,snes->iter,xnorm,ynorm,fnorm,&snes->reason,snes->cnvP);
CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
/* Call general purpose update function */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);
CHKERRQ(ierr);
}
if (snes->pc && snes->pcside == PC_RIGHT) {
ierr = VecCopy(X,dX);
CHKERRQ(ierr);
ierr = SNESSetInitialFunction(snes->pc, F);
CHKERRQ(ierr);
ierr = SNESSetInitialFunctionNorm(snes->pc, fnorm);
CHKERRQ(ierr);
ierr = SNESSolve(snes->pc, B, dX);
CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);
CHKERRQ(ierr);
if (reason < 0 && (reason != SNES_DIVERGED_MAX_IT)) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = VecAYPX(dX,-1.0,X);
CHKERRQ(ierr);
} else {
ierr = VecCopy(F, dX);
CHKERRQ(ierr);
}
/* compute the conjugate direction lX = dX + beta*lX with beta = ((dX, dX) / (dX_old, dX_old) (Fletcher-Reeves update)*/
switch(ncg->type) {
case SNES_NCG_FR: /* Fletcher-Reeves */
dXolddotFold = dXdotF;
ierr = VecDot(dX, dX, &dXdotF);
CHKERRQ(ierr);
beta = PetscRealPart(dXdotF / dXolddotFold);
break;
case SNES_NCG_PRP: /* Polak-Ribiere-Poylak */
dXolddotFold = dXdotF;
ierr = VecDotBegin(F, dX, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotBegin(Fold, dX, &dXdotFold);
CHKERRQ(ierr);
ierr = VecDotEnd(F, dX, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotEnd(Fold, dX, &dXdotFold);
CHKERRQ(ierr);
beta = PetscRealPart(((dXdotF - dXdotFold) / dXolddotFold));
if (beta < 0.0) beta = 0.0; /* restart */
break;
case SNES_NCG_HS: /* Hestenes-Stiefel */
ierr = VecDotBegin(dX, F, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotBegin(dX, Fold, &dXdotFold);
CHKERRQ(ierr);
ierr = VecDotBegin(lX, F, &lXdotF);
CHKERRQ(ierr);
ierr = VecDotBegin(lX, Fold, &lXdotFold);
CHKERRQ(ierr);
ierr = VecDotEnd(dX, F, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotEnd(dX, Fold, &dXdotFold);
CHKERRQ(ierr);
ierr = VecDotEnd(lX, F, &lXdotF);
CHKERRQ(ierr);
ierr = VecDotEnd(lX, Fold, &lXdotFold);
CHKERRQ(ierr);
beta = PetscRealPart((dXdotF - dXdotFold) / (lXdotF - lXdotFold));
break;
case SNES_NCG_DY: /* Dai-Yuan */
ierr = VecDotBegin(dX, F, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotBegin(lX, F, &lXdotF);
CHKERRQ(ierr);
ierr = VecDotBegin(lX, Fold, &lXdotFold);
CHKERRQ(ierr);
ierr = VecDotEnd(dX, F, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotEnd(lX, F, &lXdotF);
CHKERRQ(ierr);
ierr = VecDotEnd(lX, Fold, &lXdotFold);
CHKERRQ(ierr);
beta = PetscRealPart(dXdotF / (lXdotFold - lXdotF));
CHKERRQ(ierr);
break;
case SNES_NCG_CD: /* Conjugate Descent */
ierr = VecDotBegin(dX, F, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotBegin(lX, Fold, &lXdotFold);
CHKERRQ(ierr);
ierr = VecDotEnd(dX, F, &dXdotF);
CHKERRQ(ierr);
ierr = VecDotEnd(lX, Fold, &lXdotFold);
CHKERRQ(ierr);
beta = PetscRealPart(dXdotF / lXdotFold);
CHKERRQ(ierr);
break;
}
if (ncg->monitor) {
ierr = PetscViewerASCIIPrintf(ncg->monitor, "beta = %e\n", beta);
CHKERRQ(ierr);
}
ierr = VecAYPX(lX, beta, dX);
CHKERRQ(ierr);
}
ierr = PetscInfo1(snes, "Maximum number of iterations has been reached: %D\n", maxits);
CHKERRQ(ierr);
if (!snes->reason) snes->reason = SNES_DIVERGED_MAX_IT;
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_GROPPCG(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i;
PetscScalar alpha,beta = 0.0,gamma,gammaNew,t;
PetscReal dp = 0.0;
Vec x,b,r,p,s,S,z,Z;
Mat Amat,Pmat;
PetscBool diagonalscale;
PetscFunctionBegin;
ierr = PCGetDiagonalScale(ksp->pc,&diagonalscale);CHKERRQ(ierr);
if (diagonalscale) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Krylov method %s does not support diagonal scaling",((PetscObject)ksp)->type_name);
x = ksp->vec_sol;
b = ksp->vec_rhs;
r = ksp->work[0];
p = ksp->work[1];
s = ksp->work[2];
S = ksp->work[3];
z = ksp->work[4];
Z = ksp->work[5];
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
ksp->its = 0;
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,x,r);CHKERRQ(ierr); /* r <- b - Ax */
ierr = VecAYPX(r,-1.0,b);CHKERRQ(ierr);
} else {
ierr = VecCopy(b,r);CHKERRQ(ierr); /* r <- b (x is 0) */
}
ierr = KSP_PCApply(ksp,r,z);CHKERRQ(ierr); /* z <- Br */
ierr = VecCopy(z,p);CHKERRQ(ierr); /* p <- z */
ierr = VecDotBegin(r,z,&gamma);CHKERRQ(ierr); /* gamma <- z'*r */
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)r));CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,p,s);CHKERRQ(ierr); /* s <- Ap */
ierr = VecDotEnd(r,z,&gamma);CHKERRQ(ierr); /* gamma <- z'*r */
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
/* This could be merged with the computation of gamma above */
ierr = VecNorm(z,NORM_2,&dp);CHKERRQ(ierr); /* dp <- z'*z = e'*A'*B'*B*A'*e' */
break;
case KSP_NORM_UNPRECONDITIONED:
/* This could be merged with the computation of gamma above */
ierr = VecNorm(r,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r = e'*A'*A*e */
break;
case KSP_NORM_NATURAL:
if (PetscIsInfOrNanScalar(gamma)) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_FP,"Infinite or not-a-number generated in dot product");
dp = PetscSqrtReal(PetscAbsScalar(gamma)); /* dp <- r'*z = r'*B*r = e'*A'*B*A*e */
break;
case KSP_NORM_NONE:
dp = 0.0;
break;
default: SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"%s",KSPNormTypes[ksp->normtype]);
}
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,dp);CHKERRQ(ierr);
ksp->rnorm = dp;
ierr = (*ksp->converged)(ksp,0,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); /* test for convergence */
if (ksp->reason) PetscFunctionReturn(0);
i = 0;
do {
ksp->its = i+1;
i++;
ierr = VecDotBegin(p,s,&t);CHKERRQ(ierr);
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)p));CHKERRQ(ierr);
ierr = KSP_PCApply(ksp,s,S);CHKERRQ(ierr); /* S <- Bs */
ierr = VecDotEnd(p,s,&t);CHKERRQ(ierr);
alpha = gamma / t;
ierr = VecAXPY(x, alpha,p);CHKERRQ(ierr); /* x <- x + alpha * p */
ierr = VecAXPY(r,-alpha,s);CHKERRQ(ierr); /* r <- r - alpha * s */
ierr = VecAXPY(z,-alpha,S);CHKERRQ(ierr); /* z <- z - alpha * S */
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNormBegin(r,NORM_2,&dp);CHKERRQ(ierr);
} else if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNormBegin(z,NORM_2,&dp);CHKERRQ(ierr);
}
ierr = VecDotBegin(r,z,&gammaNew);CHKERRQ(ierr);
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)r));CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,z,Z);CHKERRQ(ierr); /* Z <- Az */
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNormEnd(r,NORM_2,&dp);CHKERRQ(ierr);
} else if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNormEnd(z,NORM_2,&dp);CHKERRQ(ierr);
}
ierr = VecDotEnd(r,z,&gammaNew);CHKERRQ(ierr);
if (ksp->normtype == KSP_NORM_NATURAL) {
if (PetscIsInfOrNanScalar(gammaNew)) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_FP,"Infinite or not-a-number generated in dot product");
dp = PetscSqrtReal(PetscAbsScalar(gammaNew)); /* dp <- r'*z = r'*B*r = e'*A'*B*A*e */
} else if (ksp->normtype == KSP_NORM_NONE) {
dp = 0.0;
}
ksp->rnorm = dp;
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i,dp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
beta = gammaNew / gamma;
gamma = gammaNew;
ierr = VecAYPX(p,beta,z);CHKERRQ(ierr); /* p <- z + beta * p */
ierr = VecAYPX(s,beta,Z);CHKERRQ(ierr); /* s <- Z + beta * s */
} while (i<ksp->max_it);
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
/*
EPSDelayedArnoldi - This function is equivalent to EPSBasicArnoldi but
performs the computation in a different way. The main idea is that
reorthogonalization is delayed to the next Arnoldi step. This version is
more scalable but in some cases convergence may stagnate.
*/
PetscErrorCode EPSDelayedArnoldi(EPS eps,PetscScalar *H,PetscInt ldh,Vec *V,PetscInt k,PetscInt *M,Vec f,PetscReal *beta,PetscBool *breakdown)
{
PetscErrorCode ierr;
PetscInt i,j,m=*M;
Vec u,t;
PetscScalar shh[100],*lhh,dot,dot2;
PetscReal norm1=0.0,norm2;
PetscFunctionBegin;
if (m<=100) lhh = shh;
else {
ierr = PetscMalloc1(m,&lhh);CHKERRQ(ierr);
}
ierr = VecDuplicate(f,&u);CHKERRQ(ierr);
ierr = VecDuplicate(f,&t);CHKERRQ(ierr);
for (j=k;j<m;j++) {
ierr = STApply(eps->st,V[j],f);CHKERRQ(ierr);
ierr = IPOrthogonalize(eps->ip,0,NULL,eps->nds,NULL,eps->defl,f,NULL,NULL,NULL);CHKERRQ(ierr);
ierr = IPMInnerProductBegin(eps->ip,f,j+1,V,H+ldh*j);CHKERRQ(ierr);
if (j>k) {
ierr = IPMInnerProductBegin(eps->ip,V[j],j,V,lhh);CHKERRQ(ierr);
ierr = IPInnerProductBegin(eps->ip,V[j],V[j],&dot);CHKERRQ(ierr);
}
if (j>k+1) {
ierr = IPNormBegin(eps->ip,u,&norm2);CHKERRQ(ierr);
ierr = VecDotBegin(u,V[j-2],&dot2);CHKERRQ(ierr);
}
ierr = IPMInnerProductEnd(eps->ip,f,j+1,V,H+ldh*j);CHKERRQ(ierr);
if (j>k) {
ierr = IPMInnerProductEnd(eps->ip,V[j],j,V,lhh);CHKERRQ(ierr);
ierr = IPInnerProductEnd(eps->ip,V[j],V[j],&dot);CHKERRQ(ierr);
}
if (j>k+1) {
ierr = IPNormEnd(eps->ip,u,&norm2);CHKERRQ(ierr);
ierr = VecDotEnd(u,V[j-2],&dot2);CHKERRQ(ierr);
if (PetscAbsScalar(dot2/norm2) > PETSC_MACHINE_EPSILON) {
*breakdown = PETSC_TRUE;
*M = j-1;
*beta = norm2;
if (m>100) { ierr = PetscFree(lhh);CHKERRQ(ierr); }
ierr = VecDestroy(&u);CHKERRQ(ierr);
ierr = VecDestroy(&t);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
}
if (j>k) {
norm1 = PetscSqrtReal(PetscRealPart(dot));
for (i=0;i<j;i++)
H[ldh*j+i] = H[ldh*j+i]/norm1;
H[ldh*j+j] = H[ldh*j+j]/dot;
ierr = VecCopy(V[j],t);CHKERRQ(ierr);
ierr = VecScale(V[j],1.0/norm1);CHKERRQ(ierr);
ierr = VecScale(f,1.0/norm1);CHKERRQ(ierr);
}
ierr = SlepcVecMAXPBY(f,1.0,-1.0,j+1,H+ldh*j,V);CHKERRQ(ierr);
if (j>k) {
ierr = SlepcVecMAXPBY(t,1.0,-1.0,j,lhh,V);CHKERRQ(ierr);
for (i=0;i<j;i++)
H[ldh*(j-1)+i] += lhh[i];
}
if (j>k+1) {
ierr = VecCopy(u,V[j-1]);CHKERRQ(ierr);
ierr = VecScale(V[j-1],1.0/norm2);CHKERRQ(ierr);
H[ldh*(j-2)+j-1] = norm2;
}
if (j<m-1) {
ierr = VecCopy(f,V[j+1]);CHKERRQ(ierr);
ierr = VecCopy(t,u);CHKERRQ(ierr);
}
}
ierr = IPNorm(eps->ip,t,&norm2);CHKERRQ(ierr);
ierr = VecScale(t,1.0/norm2);CHKERRQ(ierr);
ierr = VecCopy(t,V[m-1]);CHKERRQ(ierr);
H[ldh*(m-2)+m-1] = norm2;
ierr = IPMInnerProduct(eps->ip,f,m,V,lhh);CHKERRQ(ierr);
ierr = SlepcVecMAXPBY(f,1.0,-1.0,m,lhh,V);CHKERRQ(ierr);
for (i=0;i<m;i++)
H[ldh*(m-1)+i] += lhh[i];
ierr = IPNorm(eps->ip,f,beta);CHKERRQ(ierr);
ierr = VecScale(f,1.0 / *beta);CHKERRQ(ierr);
*breakdown = PETSC_FALSE;
if (m>100) { ierr = PetscFree(lhh);CHKERRQ(ierr); }
ierr = VecDestroy(&u);CHKERRQ(ierr);
ierr = VecDestroy(&t);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_PIPECG(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i;
PetscScalar alpha = 0.0,beta = 0.0,gamma = 0.0,gammaold = 0.0,delta = 0.0;
PetscReal dp = 0.0;
Vec X,B,Z,P,W,Q,U,M,N,R,S;
Mat Amat,Pmat;
PetscBool diagonalscale;
PetscFunctionBegin;
ierr = PCGetDiagonalScale(ksp->pc,&diagonalscale);CHKERRQ(ierr);
if (diagonalscale) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Krylov method %s does not support diagonal scaling",((PetscObject)ksp)->type_name);
X = ksp->vec_sol;
B = ksp->vec_rhs;
M = ksp->work[0];
Z = ksp->work[1];
P = ksp->work[2];
N = ksp->work[3];
W = ksp->work[4];
Q = ksp->work[5];
U = ksp->work[6];
R = ksp->work[7];
S = ksp->work[8];
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
ksp->its = 0;
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr); /* r <- b - Ax */
ierr = VecAYPX(R,-1.0,B);CHKERRQ(ierr);
} else {
ierr = VecCopy(B,R);CHKERRQ(ierr); /* r <- b (x is 0) */
}
ierr = KSP_PCApply(ksp,R,U);CHKERRQ(ierr); /* u <- Br */
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
ierr = VecNormBegin(U,NORM_2,&dp);CHKERRQ(ierr); /* dp <- u'*u = e'*A'*B'*B*A'*e' */
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)U));CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,U,W);CHKERRQ(ierr); /* w <- Au */
ierr = VecNormEnd(U,NORM_2,&dp);CHKERRQ(ierr);
break;
case KSP_NORM_UNPRECONDITIONED:
ierr = VecNormBegin(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r = e'*A'*A*e */
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)R));CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,U,W);CHKERRQ(ierr); /* w <- Au */
ierr = VecNormEnd(R,NORM_2,&dp);CHKERRQ(ierr);
break;
case KSP_NORM_NATURAL:
ierr = VecDotBegin(R,U,&gamma);CHKERRQ(ierr); /* gamma <- u'*r */
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)R));CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,U,W);CHKERRQ(ierr); /* w <- Au */
ierr = VecDotEnd(R,U,&gamma);CHKERRQ(ierr);
KSPCheckDot(ksp,gamma);
dp = PetscSqrtReal(PetscAbsScalar(gamma)); /* dp <- r'*u = r'*B*r = e'*A'*B*A*e */
break;
case KSP_NORM_NONE:
ierr = KSP_MatMult(ksp,Amat,U,W);CHKERRQ(ierr);
dp = 0.0;
break;
default: SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"%s",KSPNormTypes[ksp->normtype]);
}
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,dp);CHKERRQ(ierr);
ksp->rnorm = dp;
ierr = (*ksp->converged)(ksp,0,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); /* test for convergence */
if (ksp->reason) PetscFunctionReturn(0);
i = 0;
do {
if (i > 0 && ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNormBegin(R,NORM_2,&dp);CHKERRQ(ierr);
} else if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNormBegin(U,NORM_2,&dp);CHKERRQ(ierr);
}
if (!(i == 0 && ksp->normtype == KSP_NORM_NATURAL)) {
ierr = VecDotBegin(R,U,&gamma);CHKERRQ(ierr);
}
ierr = VecDotBegin(W,U,&delta);CHKERRQ(ierr);
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)R));CHKERRQ(ierr);
ierr = KSP_PCApply(ksp,W,M);CHKERRQ(ierr); /* m <- Bw */
ierr = KSP_MatMult(ksp,Amat,M,N);CHKERRQ(ierr); /* n <- Am */
if (i > 0 && ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNormEnd(R,NORM_2,&dp);CHKERRQ(ierr);
} else if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNormEnd(U,NORM_2,&dp);CHKERRQ(ierr);
}
if (!(i == 0 && ksp->normtype == KSP_NORM_NATURAL)) {
ierr = VecDotEnd(R,U,&gamma);CHKERRQ(ierr);
}
ierr = VecDotEnd(W,U,&delta);CHKERRQ(ierr);
if (i > 0) {
if (ksp->normtype == KSP_NORM_NATURAL) dp = PetscSqrtReal(PetscAbsScalar(gamma));
else if (ksp->normtype == KSP_NORM_NONE) dp = 0.0;
ksp->rnorm = dp;
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i,dp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
}
if (i == 0) {
alpha = gamma / delta;
ierr = VecCopy(N,Z);CHKERRQ(ierr); /* z <- n */
ierr = VecCopy(M,Q);CHKERRQ(ierr); /* q <- m */
ierr = VecCopy(U,P);CHKERRQ(ierr); /* p <- u */
ierr = VecCopy(W,S);CHKERRQ(ierr); /* s <- w */
} else {
beta = gamma / gammaold;
alpha = gamma / (delta - beta / alpha * gamma);
ierr = VecAYPX(Z,beta,N);CHKERRQ(ierr); /* z <- n + beta * z */
ierr = VecAYPX(Q,beta,M);CHKERRQ(ierr); /* q <- m + beta * q */
ierr = VecAYPX(P,beta,U);CHKERRQ(ierr); /* p <- u + beta * p */
ierr = VecAYPX(S,beta,W);CHKERRQ(ierr); /* s <- w + beta * s */
}
ierr = VecAXPY(X, alpha,P);CHKERRQ(ierr); /* x <- x + alpha * p */
ierr = VecAXPY(U,-alpha,Q);CHKERRQ(ierr); /* u <- u - alpha * q */
ierr = VecAXPY(W,-alpha,Z);CHKERRQ(ierr); /* w <- w - alpha * z */
ierr = VecAXPY(R,-alpha,S);CHKERRQ(ierr); /* r <- r - alpha * s */
gammaold = gamma;
i++;
ksp->its = i;
/* if (i%50 == 0) { */
/* ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr); /\* w <- b - Ax *\/ */
/* ierr = VecAYPX(R,-1.0,B);CHKERRQ(ierr); */
/* ierr = KSP_PCApply(ksp,R,U);CHKERRQ(ierr); */
/* ierr = KSP_MatMult(ksp,Amat,U,W);CHKERRQ(ierr); */
/* } */
} while (i<ksp->max_it);
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_CR(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i = 0;
MatStructure pflag;
PetscReal dp;
PetscScalar ai, bi;
PetscScalar apq,btop, bbot;
Vec X,B,R,RT,P,AP,ART,Q;
Mat Amat, Pmat;
PetscFunctionBegin;
X = ksp->vec_sol;
B = ksp->vec_rhs;
R = ksp->work[0];
RT = ksp->work[1];
P = ksp->work[2];
AP = ksp->work[3];
ART = ksp->work[4];
Q = ksp->work[5];
/* R is the true residual norm, RT is the preconditioned residual norm */
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat,&pflag);CHKERRQ(ierr);
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr); /* R <- A*X */
ierr = VecAYPX(R,-1.0,B);CHKERRQ(ierr); /* R <- B-R == B-A*X */
} else {
ierr = VecCopy(B,R);CHKERRQ(ierr); /* R <- B (X is 0) */
}
ierr = KSP_PCApply(ksp,R,P);CHKERRQ(ierr); /* P <- B*R */
ierr = KSP_MatMult(ksp,Amat,P,AP);CHKERRQ(ierr); /* AP <- A*P */
ierr = VecCopy(P,RT);CHKERRQ(ierr); /* RT <- P */
ierr = VecCopy(AP,ART);CHKERRQ(ierr); /* ART <- AP */
ierr = VecDotBegin(RT,ART,&btop);CHKERRQ(ierr); /* (RT,ART) */
if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNormBegin(RT,NORM_2,&dp);CHKERRQ(ierr); /* dp <- RT'*RT */
ierr = VecDotEnd (RT,ART,&btop);CHKERRQ(ierr); /* (RT,ART) */
ierr = VecNormEnd (RT,NORM_2,&dp);CHKERRQ(ierr); /* dp <- RT'*RT */
} else if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNormBegin(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- R'*R */
ierr = VecDotEnd (RT,ART,&btop);CHKERRQ(ierr); /* (RT,ART) */
ierr = VecNormEnd (R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- RT'*RT */
} else if (ksp->normtype == KSP_NORM_NATURAL) {
ierr = VecDotEnd (RT,ART,&btop);CHKERRQ(ierr); /* (RT,ART) */
dp = PetscSqrtReal(PetscAbsScalar(btop)); /* dp = sqrt(R,AR) */
}
if (PetscAbsScalar(btop) < 0.0) {
ksp->reason = KSP_DIVERGED_INDEFINITE_MAT;
ierr = PetscInfo(ksp,"diverging due to indefinite or negative definite matrix\n");CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ksp->its = 0;
ierr = KSPMonitor(ksp,0,dp);CHKERRQ(ierr);
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->rnorm = dp;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,0,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
i = 0;
do {
ierr = KSP_PCApply(ksp,AP,Q);CHKERRQ(ierr); /* Q <- B* AP */
ierr = VecDot(AP,Q,&apq);CHKERRQ(ierr);
if (PetscRealPart(apq) <= 0.0) {
ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
ierr = PetscInfo(ksp,"KSPSolve_CR:diverging due to indefinite or negative definite PC\n");CHKERRQ(ierr);
break;
}
ai = btop/apq; /* ai = (RT,ART)/(AP,Q) */
ierr = VecAXPY(X,ai,P);CHKERRQ(ierr); /* X <- X + ai*P */
ierr = VecAXPY(RT,-ai,Q);CHKERRQ(ierr); /* RT <- RT - ai*Q */
ierr = KSP_MatMult(ksp,Amat,RT,ART);CHKERRQ(ierr); /* ART <- A*RT */
bbot = btop;
ierr = VecDotBegin(RT,ART,&btop);CHKERRQ(ierr);
if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNormBegin(RT,NORM_2,&dp);CHKERRQ(ierr); /* dp <- || RT || */
ierr = VecDotEnd (RT,ART,&btop);CHKERRQ(ierr);
ierr = VecNormEnd (RT,NORM_2,&dp);CHKERRQ(ierr); /* dp <- || RT || */
} else if (ksp->normtype == KSP_NORM_NATURAL) {
ierr = VecDotEnd(RT,ART,&btop);CHKERRQ(ierr);
dp = PetscSqrtReal(PetscAbsScalar(btop)); /* dp = sqrt(R,AR) */
} else if (ksp->normtype == KSP_NORM_NONE) {
ierr = VecDotEnd(RT,ART,&btop);CHKERRQ(ierr);
dp = 0.0;
} else if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecAXPY(R,ai,AP);CHKERRQ(ierr); /* R <- R - ai*AP */
ierr = VecNormBegin(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- R'*R */
ierr = VecDotEnd (RT,ART,&btop);CHKERRQ(ierr);
ierr = VecNormEnd (R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- R'*R */
} else SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"KSPNormType of %d not supported",(int)ksp->normtype);
if (PetscAbsScalar(btop) < 0.0) {
ksp->reason = KSP_DIVERGED_INDEFINITE_MAT;
ierr = PetscInfo(ksp,"diverging due to indefinite or negative definite PC\n");CHKERRQ(ierr);
break;
}
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = dp;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i+1,dp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i+1,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
bi = btop/bbot;
ierr = VecAYPX(P,bi,RT);CHKERRQ(ierr); /* P <- RT + Bi P */
ierr = VecAYPX(AP,bi,ART);CHKERRQ(ierr); /* AP <- ART + Bi AP */
i++;
} while (i<ksp->max_it);
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
PetscErrorCode ierr;
PetscInt n = 25,i,row0 = 0;
PetscScalar one = 1.0,two = 2.0,result1,result2,results[40],value,ten = 10.0;
PetscScalar result1a,result2a;
PetscReal result3,result4,result[2],result3a,result4a,resulta[2];
Vec x,y,vecs[40];
ierr = PetscInitialize(&argc,&argv,(char*)0,help);CHKERRQ(ierr);
/* create vector */
ierr = VecCreate(PETSC_COMM_WORLD,&x);CHKERRQ(ierr);
ierr = VecSetSizes(x,n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetFromOptions(x);CHKERRQ(ierr);
ierr = VecDuplicate(x,&y);CHKERRQ(ierr);
ierr = VecSet(x,one);CHKERRQ(ierr);
ierr = VecSet(y,two);CHKERRQ(ierr);
/*
Test mixing dot products and norms that require sums
*/
result1 = result2 = 0.0;
result3 = result4 = 0.0;
ierr = VecDotBegin(x,y,&result1);CHKERRQ(ierr);
ierr = VecDotBegin(y,x,&result2);CHKERRQ(ierr);
ierr = VecNormBegin(y,NORM_2,&result3);CHKERRQ(ierr);
ierr = VecNormBegin(x,NORM_1,&result4);CHKERRQ(ierr);
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)x));CHKERRQ(ierr);
ierr = VecDotEnd(x,y,&result1);CHKERRQ(ierr);
ierr = VecDotEnd(y,x,&result2);CHKERRQ(ierr);
ierr = VecNormEnd(y,NORM_2,&result3);CHKERRQ(ierr);
ierr = VecNormEnd(x,NORM_1,&result4);CHKERRQ(ierr);
ierr = VecDot(x,y,&result1a);CHKERRQ(ierr);
ierr = VecDot(y,x,&result2a);CHKERRQ(ierr);
ierr = VecNorm(y,NORM_2,&result3a);CHKERRQ(ierr);
ierr = VecNorm(x,NORM_1,&result4a);CHKERRQ(ierr);
if (result1 != result1a || result2 != result2a) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"Error dot: result1 %g result2 %g\n",(double)PetscRealPart(result1),(double)PetscRealPart(result2));CHKERRQ(ierr);
}
if (result3 != result3a || result4 != result4a) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"Error 1,2 norms: result3 %g result4 %g\n",(double)result3,(double)result4);CHKERRQ(ierr);
}
/*
Test norms that only require abs
*/
result1 = result2 = 0.0;
result3 = result4 = 0.0;
ierr = VecNormBegin(y,NORM_MAX,&result3);CHKERRQ(ierr);
ierr = VecNormBegin(x,NORM_MAX,&result4);CHKERRQ(ierr);
ierr = VecNormEnd(y,NORM_MAX,&result3);CHKERRQ(ierr);
ierr = VecNormEnd(x,NORM_MAX,&result4);CHKERRQ(ierr);
ierr = VecNorm(x,NORM_MAX,&result4a);CHKERRQ(ierr);
ierr = VecNorm(y,NORM_MAX,&result3a);CHKERRQ(ierr);
if (result3 != result3a || result4 != result4a) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"Error max norm: result3 %g result4 %g\n",(double)result3,(double)result4);CHKERRQ(ierr);
}
/*
Tests dot, max, 1, norm
*/
result1 = result2 = 0.0;
result3 = result4 = 0.0;
ierr = VecSetValues(x,1,&row0,&ten,INSERT_VALUES);CHKERRQ(ierr);
ierr = VecAssemblyBegin(x);CHKERRQ(ierr);
ierr = VecAssemblyEnd(x);CHKERRQ(ierr);
ierr = VecDotBegin(x,y,&result1);CHKERRQ(ierr);
ierr = VecDotBegin(y,x,&result2);CHKERRQ(ierr);
ierr = VecNormBegin(x,NORM_MAX,&result3);CHKERRQ(ierr);
ierr = VecNormBegin(x,NORM_1,&result4);CHKERRQ(ierr);
ierr = VecDotEnd(x,y,&result1);CHKERRQ(ierr);
ierr = VecDotEnd(y,x,&result2);CHKERRQ(ierr);
ierr = VecNormEnd(x,NORM_MAX,&result3);CHKERRQ(ierr);
ierr = VecNormEnd(x,NORM_1,&result4);CHKERRQ(ierr);
ierr = VecDot(x,y,&result1a);CHKERRQ(ierr);
ierr = VecDot(y,x,&result2a);CHKERRQ(ierr);
ierr = VecNorm(x,NORM_MAX,&result3a);CHKERRQ(ierr);
ierr = VecNorm(x,NORM_1,&result4a);CHKERRQ(ierr);
if (result1 != result1a || result2 != result2a) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"Error dot: result1 %g result2 %g\n",(double)PetscRealPart(result1),(double)PetscRealPart(result2));CHKERRQ(ierr);
}
if (result3 != result3a || result4 != result4a) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"Error max 1 norms: result3 %g result4 %g\n",(double)result3,(double)result4);CHKERRQ(ierr);
}
/*
tests 1_and_2 norm
*/
ierr = VecNormBegin(x,NORM_MAX,&result3);CHKERRQ(ierr);
ierr = VecNormBegin(x,NORM_1_AND_2,result);CHKERRQ(ierr);
ierr = VecNormBegin(y,NORM_MAX,&result4);CHKERRQ(ierr);
ierr = VecNormEnd(x,NORM_MAX,&result3);CHKERRQ(ierr);
ierr = VecNormEnd(x,NORM_1_AND_2,result);CHKERRQ(ierr);
ierr = VecNormEnd(y,NORM_MAX,&result4);CHKERRQ(ierr);
ierr = VecNorm(x,NORM_MAX,&result3a);CHKERRQ(ierr);
ierr = VecNorm(x,NORM_1_AND_2,resulta);CHKERRQ(ierr);
ierr = VecNorm(y,NORM_MAX,&result4a);CHKERRQ(ierr);
if (result3 != result3a || result4 != result4a) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"Error max: result1 %g result2 %g\n",(double)result3,(double)result4);CHKERRQ(ierr);
}
if (PetscAbsReal(result[0]-resulta[0]) > .01 || PetscAbsReal(result[1]-resulta[1]) > .01) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"Error 1 and 2 norms: result[0] %g result[1] %g\n",(double)result[0],(double)result[1]);CHKERRQ(ierr);
}
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&y);CHKERRQ(ierr);
/*
Tests computing a large number of operations that require
allocating a larger data structure internally
*/
for (i=0; i<40; i++) {
ierr = VecCreate(PETSC_COMM_WORLD,vecs+i);CHKERRQ(ierr);
ierr = VecSetSizes(vecs[i],PETSC_DECIDE,n);CHKERRQ(ierr);
ierr = VecSetFromOptions(vecs[i]);CHKERRQ(ierr);
value = (PetscReal)i;
ierr = VecSet(vecs[i],value);CHKERRQ(ierr);
}
for (i=0; i<39; i++) {
ierr = VecDotBegin(vecs[i],vecs[i+1],results+i);CHKERRQ(ierr);
}
for (i=0; i<39; i++) {
ierr = VecDotEnd(vecs[i],vecs[i+1],results+i);CHKERRQ(ierr);
if (results[i] != 25.0*i*(i+1)) {
ierr = PetscPrintf(PETSC_COMM_WORLD,"i %D expected %g got %g\n",i,25.0*i*(i+1),(double)PetscRealPart(results[i]));CHKERRQ(ierr);
}
}
for (i=0; i<40; i++) {
ierr = VecDestroy(&vecs[i]);CHKERRQ(ierr);
}
ierr = PetscFinalize();
return 0;
}
static PetscErrorCode SNESSolve_QN(SNES snes)
{
PetscErrorCode ierr;
SNES_QN *qn = (SNES_QN*) snes->data;
Vec X,Xold;
Vec F,W;
Vec Y,D,Dold;
PetscInt i, i_r;
PetscReal fnorm,xnorm,ynorm,gnorm;
SNESLineSearchReason lssucceed;
PetscBool powell,periodic;
PetscScalar DolddotD,DolddotDold;
SNESConvergedReason reason;
/* basically just a regular newton's method except for the application of the Jacobian */
PetscFunctionBegin;
if (snes->xl || snes->xu || snes->ops->computevariablebounds) {
SETERRQ1(PetscObjectComm((PetscObject)snes),PETSC_ERR_ARG_WRONGSTATE, "SNES solver %s does not support bounds", ((PetscObject)snes)->type_name);
}
ierr = PetscCitationsRegister(SNESCitation,&SNEScite);CHKERRQ(ierr);
F = snes->vec_func; /* residual vector */
Y = snes->vec_sol_update; /* search direction generated by J^-1D*/
W = snes->work[3];
X = snes->vec_sol; /* solution vector */
Xold = snes->work[0];
/* directions generated by the preconditioned problem with F_pre = F or x - M(x, b) */
D = snes->work[1];
Dold = snes->work[2];
snes->reason = SNES_CONVERGED_ITERATING;
ierr = PetscObjectSAWsTakeAccess((PetscObject)snes);CHKERRQ(ierr);
snes->iter = 0;
snes->norm = 0.;
ierr = PetscObjectSAWsGrantAccess((PetscObject)snes);CHKERRQ(ierr);
if (snes->pc && snes->pcside == PC_LEFT && snes->functype == SNES_FUNCTION_PRECONDITIONED) {
ierr = SNESApplyNPC(snes,X,NULL,F);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = VecNorm(F,NORM_2,&fnorm);CHKERRQ(ierr);
} else {
if (!snes->vec_func_init_set) {
ierr = SNESComputeFunction(snes,X,F);CHKERRQ(ierr);
} else snes->vec_func_init_set = PETSC_FALSE;
ierr = VecNorm(F,NORM_2,&fnorm);CHKERRQ(ierr);
SNESCheckFunctionNorm(snes,fnorm);
}
if (snes->pc && snes->pcside == PC_LEFT && snes->functype == SNES_FUNCTION_UNPRECONDITIONED) {
ierr = SNESApplyNPC(snes,X,F,D);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
} else {
ierr = VecCopy(F,D);CHKERRQ(ierr);
}
ierr = PetscObjectSAWsTakeAccess((PetscObject)snes);CHKERRQ(ierr);
snes->norm = fnorm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)snes);CHKERRQ(ierr);
ierr = SNESLogConvergenceHistory(snes,fnorm,0);CHKERRQ(ierr);
ierr = SNESMonitor(snes,0,fnorm);CHKERRQ(ierr);
/* test convergence */
ierr = (*snes->ops->converged)(snes,0,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
if (snes->pc && snes->pcside == PC_RIGHT) {
ierr = PetscLogEventBegin(SNES_NPCSolve,snes->pc,X,0,0);CHKERRQ(ierr);
ierr = SNESSolve(snes->pc,snes->vec_rhs,X);CHKERRQ(ierr);
ierr = PetscLogEventEnd(SNES_NPCSolve,snes->pc,X,0,0);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = SNESGetNPCFunction(snes,F,&fnorm);CHKERRQ(ierr);
ierr = VecCopy(F,D);CHKERRQ(ierr);
}
/* scale the initial update */
if (qn->scale_type == SNES_QN_SCALE_JACOBIAN) {
ierr = SNESComputeJacobian(snes,X,snes->jacobian,snes->jacobian_pre);CHKERRQ(ierr);
ierr = KSPSetOperators(snes->ksp,snes->jacobian,snes->jacobian_pre);CHKERRQ(ierr);
}
for (i = 0, i_r = 0; i < snes->max_its; i++, i_r++) {
if (qn->scale_type == SNES_QN_SCALE_SHANNO && i_r > 0) {
PetscScalar ff,xf;
ierr = VecCopy(Dold,Y);CHKERRQ(ierr);
ierr = VecCopy(Xold,W);CHKERRQ(ierr);
ierr = VecAXPY(Y,-1.0,D);CHKERRQ(ierr);
ierr = VecAXPY(W,-1.0,X);CHKERRQ(ierr);
ierr = VecDotBegin(Y,Y,&ff);CHKERRQ(ierr);
ierr = VecDotBegin(W,Y,&xf);CHKERRQ(ierr);
ierr = VecDotEnd(Y,Y,&ff);CHKERRQ(ierr);
ierr = VecDotEnd(W,Y,&xf);CHKERRQ(ierr);
qn->scaling = PetscRealPart(xf)/PetscRealPart(ff);
}
switch (qn->type) {
case SNES_QN_BADBROYDEN:
ierr = SNESQNApply_BadBroyden(snes,i_r,Y,X,Xold,D,Dold);CHKERRQ(ierr);
break;
case SNES_QN_BROYDEN:
ierr = SNESQNApply_Broyden(snes,i_r,Y,X,Xold,D);CHKERRQ(ierr);
break;
case SNES_QN_LBFGS:
SNESQNApply_LBFGS(snes,i_r,Y,X,Xold,D,Dold);CHKERRQ(ierr);
break;
}
/* line search for lambda */
ynorm = 1; gnorm = fnorm;
ierr = VecCopy(D, Dold);CHKERRQ(ierr);
ierr = VecCopy(X, Xold);CHKERRQ(ierr);
ierr = SNESLineSearchApply(snes->linesearch, X, F, &fnorm, Y);CHKERRQ(ierr);
if (snes->reason == SNES_DIVERGED_FUNCTION_COUNT) break;
ierr = SNESLineSearchGetReason(snes->linesearch, &lssucceed);CHKERRQ(ierr);
ierr = SNESLineSearchGetNorms(snes->linesearch, &xnorm, &fnorm, &ynorm);CHKERRQ(ierr);
if (lssucceed) {
if (++snes->numFailures >= snes->maxFailures) {
snes->reason = SNES_DIVERGED_LINE_SEARCH;
break;
}
}
if (qn->scale_type == SNES_QN_SCALE_LINESEARCH) {
ierr = SNESLineSearchGetLambda(snes->linesearch, &qn->scaling);CHKERRQ(ierr);
}
/* convergence monitoring */
ierr = PetscInfo4(snes,"fnorm=%18.16e, gnorm=%18.16e, ynorm=%18.16e, lssucceed=%d\n",(double)fnorm,(double)gnorm,(double)ynorm,(int)lssucceed);CHKERRQ(ierr);
if (snes->pc && snes->pcside == PC_RIGHT) {
ierr = PetscLogEventBegin(SNES_NPCSolve,snes->pc,X,0,0);CHKERRQ(ierr);
ierr = SNESSolve(snes->pc,snes->vec_rhs,X);CHKERRQ(ierr);
ierr = PetscLogEventEnd(SNES_NPCSolve,snes->pc,X,0,0);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = SNESGetNPCFunction(snes,F,&fnorm);CHKERRQ(ierr);
}
ierr = SNESSetIterationNumber(snes, i+1);CHKERRQ(ierr);
snes->norm = fnorm;
ierr = SNESLogConvergenceHistory(snes,snes->norm,snes->iter);CHKERRQ(ierr);
ierr = SNESMonitor(snes,snes->iter,snes->norm);CHKERRQ(ierr);
/* set parameter for default relative tolerance convergence test */
ierr = (*snes->ops->converged)(snes,snes->iter,xnorm,ynorm,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
if (snes->pc && snes->pcside == PC_LEFT && snes->functype == SNES_FUNCTION_UNPRECONDITIONED) {
ierr = SNESApplyNPC(snes,X,F,D);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && reason != SNES_DIVERGED_MAX_IT) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
} else {
ierr = VecCopy(F, D);CHKERRQ(ierr);
}
powell = PETSC_FALSE;
if (qn->restart_type == SNES_QN_RESTART_POWELL) {
/* check restart by Powell's Criterion: |F^T H_0 Fold| > 0.2 * |Fold^T H_0 Fold| */
if (qn->scale_type == SNES_QN_SCALE_JACOBIAN) {
ierr = MatMult(snes->jacobian_pre,Dold,W);CHKERRQ(ierr);
} else {
ierr = VecCopy(Dold,W);CHKERRQ(ierr);
}
ierr = VecDotBegin(W, Dold, &DolddotDold);CHKERRQ(ierr);
ierr = VecDotBegin(W, D, &DolddotD);CHKERRQ(ierr);
ierr = VecDotEnd(W, Dold, &DolddotDold);CHKERRQ(ierr);
ierr = VecDotEnd(W, D, &DolddotD);CHKERRQ(ierr);
if (PetscAbs(PetscRealPart(DolddotD)) > qn->powell_gamma*PetscAbs(PetscRealPart(DolddotDold))) powell = PETSC_TRUE;
}
periodic = PETSC_FALSE;
if (qn->restart_type == SNES_QN_RESTART_PERIODIC) {
if (i_r>qn->m-1) periodic = PETSC_TRUE;
}
/* restart if either powell or periodic restart is satisfied. */
if (powell || periodic) {
if (qn->monitor) {
ierr = PetscViewerASCIIAddTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(qn->monitor, "restart! |%14.12e| > %4.2f*|%14.12e| or i_r = %d\n", PetscRealPart(DolddotD), qn->powell_gamma, PetscRealPart(DolddotDold), i_r);CHKERRQ(ierr);
ierr = PetscViewerASCIISubtractTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
}
i_r = -1;
/* general purpose update */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
}
if (qn->scale_type == SNES_QN_SCALE_JACOBIAN) {
ierr = SNESComputeJacobian(snes,X,snes->jacobian,snes->jacobian_pre);CHKERRQ(ierr);
}
}
/* general purpose update */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
}
}
if (i == snes->max_its) {
ierr = PetscInfo1(snes, "Maximum number of iterations has been reached: %D\n", snes->max_its);CHKERRQ(ierr);
if (!snes->reason) snes->reason = SNES_DIVERGED_MAX_IT;
}
PetscFunctionReturn(0);
}
/*
SNESMatrixFreeMult2_Private - Default matrix-free form for Jacobian-vector
product, y = F'(u)*a:
y = (F(u + ha) - F(u)) /h,
where F = nonlinear function, as set by SNESSetFunction()
u = current iterate
h = difference interval
*/
PetscErrorCode SNESMatrixFreeMult2_Private(Mat mat,Vec a,Vec y)
{
MFCtx_Private *ctx;
SNES snes;
PetscReal h,norm,sum,umin,noise;
PetscScalar hs,dot;
Vec w,U,F;
PetscErrorCode ierr,(*eval_fct)(SNES,Vec,Vec);
MPI_Comm comm;
PetscInt iter;
PetscFunctionBegin;
/* We log matrix-free matrix-vector products separately, so that we can
separate the performance monitoring from the cases that use conventional
storage. We may eventually modify event logging to associate events
with particular objects, hence alleviating the more general problem. */
ierr = PetscLogEventBegin(MATMFFD_Mult,a,y,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr);
ierr = MatShellGetContext(mat,(void**)&ctx);CHKERRQ(ierr);
snes = ctx->snes;
w = ctx->w;
umin = ctx->umin;
ierr = SNESGetSolution(snes,&U);CHKERRQ(ierr);
eval_fct = SNESComputeFunction;
ierr = SNESGetFunction(snes,&F,NULL,NULL);CHKERRQ(ierr);
/* Determine a "good" step size, h */
if (ctx->need_h) {
/* Use Jorge's method to compute h */
if (ctx->jorge) {
ierr = SNESDiffParameterCompute_More(snes,ctx->data,U,a,&noise,&h);CHKERRQ(ierr);
/* Use the Brown/Saad method to compute h */
} else {
/* Compute error if desired */
ierr = SNESGetIterationNumber(snes,&iter);CHKERRQ(ierr);
if ((ctx->need_err) || ((ctx->compute_err_freq) && (ctx->compute_err_iter != iter) && (!((iter-1)%ctx->compute_err_freq)))) {
/* Use Jorge's method to compute noise */
ierr = SNESDiffParameterCompute_More(snes,ctx->data,U,a,&noise,&h);CHKERRQ(ierr);
ctx->error_rel = PetscSqrtReal(noise);
ierr = PetscInfo3(snes,"Using Jorge's noise: noise=%g, sqrt(noise)=%g, h_more=%g\n",(double)noise,(double)ctx->error_rel,(double)h);CHKERRQ(ierr);
ctx->compute_err_iter = iter;
ctx->need_err = PETSC_FALSE;
}
ierr = VecDotBegin(U,a,&dot);CHKERRQ(ierr);
ierr = VecNormBegin(a,NORM_1,&sum);CHKERRQ(ierr);
ierr = VecNormBegin(a,NORM_2,&norm);CHKERRQ(ierr);
ierr = VecDotEnd(U,a,&dot);CHKERRQ(ierr);
ierr = VecNormEnd(a,NORM_1,&sum);CHKERRQ(ierr);
ierr = VecNormEnd(a,NORM_2,&norm);CHKERRQ(ierr);
/* Safeguard for step sizes too small */
if (sum == 0.0) {
dot = 1.0;
norm = 1.0;
} else if (PetscAbsScalar(dot) < umin*sum && PetscRealPart(dot) >= 0.0) dot = umin*sum;
else if (PetscAbsScalar(dot) < 0.0 && PetscRealPart(dot) > -umin*sum) dot = -umin*sum;
h = PetscRealPart(ctx->error_rel*dot/(norm*norm));
}
} else h = ctx->h;
if (!ctx->jorge || !ctx->need_h) {ierr = PetscInfo1(snes,"h = %g\n",(double)h);CHKERRQ(ierr);}
/* Evaluate function at F(u + ha) */
hs = h;
ierr = VecWAXPY(w,hs,a,U);CHKERRQ(ierr);
ierr = eval_fct(snes,w,y);CHKERRQ(ierr);
ierr = VecAXPY(y,-1.0,F);CHKERRQ(ierr);
ierr = VecScale(y,1.0/hs);CHKERRQ(ierr);
if (ctx->sp) {ierr = MatNullSpaceRemove(ctx->sp,y);CHKERRQ(ierr);}
ierr = PetscLogEventEnd(MATMFFD_Mult,a,y,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode SNESQNApply_LBFGS(SNES snes,PetscInt it,Vec Y,Vec X,Vec Xold,Vec D,Vec Dold)
{
PetscErrorCode ierr;
SNES_QN *qn = (SNES_QN*)snes->data;
Vec W = snes->work[3];
Vec *dX = qn->U;
Vec *dF = qn->V;
PetscScalar *alpha = qn->alpha;
PetscScalar *beta = qn->beta;
PetscScalar *dXtdF = qn->dXtdF;
PetscScalar *dFtdX = qn->dFtdX;
PetscScalar *YtdX = qn->YtdX;
/* ksp thing for jacobian scaling */
KSPConvergedReason kspreason;
PetscInt k,i,j,g,lits;
PetscInt m = qn->m;
PetscScalar t;
PetscInt l = m;
PetscFunctionBegin;
if (it < m) l = it;
ierr = VecCopy(D,Y);CHKERRQ(ierr);
if (it > 0) {
k = (it - 1) % l;
ierr = VecCopy(D,dF[k]);CHKERRQ(ierr);
ierr = VecAXPY(dF[k], -1.0, Dold);CHKERRQ(ierr);
ierr = VecCopy(X, dX[k]);CHKERRQ(ierr);
ierr = VecAXPY(dX[k], -1.0, Xold);CHKERRQ(ierr);
if (qn->singlereduction) {
PetscScalar dFtdF;
ierr = VecMDotBegin(dF[k],l,dX,dXtdF);CHKERRQ(ierr);
ierr = VecMDotBegin(dX[k],l,dF,dFtdX);CHKERRQ(ierr);
ierr = VecMDotBegin(Y,l,dX,YtdX);CHKERRQ(ierr);
if (qn->scale_type == SNES_QN_SCALE_SHANNO) {ierr = VecDotBegin(dF[k],dF[k],&dFtdF);CHKERRQ(ierr);}
ierr = VecMDotEnd(dF[k],l,dX,dXtdF);CHKERRQ(ierr);
ierr = VecMDotEnd(dX[k],l,dF,dFtdX);CHKERRQ(ierr);
ierr = VecMDotEnd(Y,l,dX,YtdX);CHKERRQ(ierr);
if (qn->scale_type == SNES_QN_SCALE_SHANNO) {
ierr = VecDotEnd(dF[k],dF[k],&dFtdF);CHKERRQ(ierr);
qn->scaling = PetscRealPart(dXtdF[k]) / PetscRealPart(dFtdF);
}
for (j = 0; j < l; j++) {
H(k, j) = dFtdX[j];
H(j, k) = dXtdF[j];
}
/* copy back over to make the computation of alpha and beta easier */
for (j = 0; j < l; j++) dXtdF[j] = H(j, j);
} else {
ierr = VecDot(dX[k], dF[k], &dXtdF[k]);CHKERRQ(ierr);
if (qn->scale_type == SNES_QN_SCALE_SHANNO) {
PetscReal dFtdF;
ierr = VecDotRealPart(dF[k],dF[k],&dFtdF);CHKERRQ(ierr);
qn->scaling = PetscRealPart(dXtdF[k])/dFtdF;
}
}
if (qn->scale_type == SNES_QN_SCALE_LINESEARCH) {
ierr = SNESLineSearchGetLambda(snes->linesearch,&qn->scaling);CHKERRQ(ierr);
}
}
/* outward recursion starting at iteration k's update and working back */
for (i=0; i<l; i++) {
k = (it-i-1)%l;
if (qn->singlereduction) {
/* construct t = dX[k] dot Y as Y_0 dot dX[k] + sum(-alpha[j]dX[k]dF[j]) */
t = YtdX[k];
for (j=0; j<i; j++) {
g = (it-j-1)%l;
t -= alpha[g]*H(k, g);
}
alpha[k] = t/H(k,k);
} else {
ierr = VecDot(dX[k],Y,&t);CHKERRQ(ierr);
alpha[k] = t/dXtdF[k];
}
if (qn->monitor) {
ierr = PetscViewerASCIIAddTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(qn->monitor, "it: %d k: %d alpha: %14.12e\n", it, k, PetscRealPart(alpha[k]));CHKERRQ(ierr);
ierr = PetscViewerASCIISubtractTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
}
ierr = VecAXPY(Y,-alpha[k],dF[k]);CHKERRQ(ierr);
}
if (qn->scale_type == SNES_QN_SCALE_JACOBIAN) {
ierr = KSPSolve(snes->ksp,Y,W);CHKERRQ(ierr);
ierr = KSPGetConvergedReason(snes->ksp,&kspreason);CHKERRQ(ierr);
if (kspreason < 0) {
if (++snes->numLinearSolveFailures >= snes->maxLinearSolveFailures) {
ierr = PetscInfo2(snes,"iter=%D, number linear solve failures %D greater than current SNES allowed, stopping solve\n",snes->iter,snes->numLinearSolveFailures);CHKERRQ(ierr);
snes->reason = SNES_DIVERGED_LINEAR_SOLVE;
PetscFunctionReturn(0);
}
}
ierr = KSPGetIterationNumber(snes->ksp,&lits);CHKERRQ(ierr);
snes->linear_its += lits;
ierr = VecCopy(W, Y);CHKERRQ(ierr);
} else {
ierr = VecScale(Y, qn->scaling);CHKERRQ(ierr);
}
if (qn->singlereduction) {
ierr = VecMDot(Y,l,dF,YtdX);CHKERRQ(ierr);
}
/* inward recursion starting at the first update and working forward */
for (i = 0; i < l; i++) {
k = (it + i - l) % l;
if (qn->singlereduction) {
t = YtdX[k];
for (j = 0; j < i; j++) {
g = (it + j - l) % l;
t += (alpha[g] - beta[g])*H(g, k);
}
beta[k] = t / H(k, k);
} else {
ierr = VecDot(dF[k], Y, &t);CHKERRQ(ierr);
beta[k] = t / dXtdF[k];
}
ierr = VecAXPY(Y, (alpha[k] - beta[k]), dX[k]);CHKERRQ(ierr);
if (qn->monitor) {
ierr = PetscViewerASCIIAddTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(qn->monitor, "it: %d k: %d alpha - beta: %14.12e\n", it, k, PetscRealPart(alpha[k] - beta[k]));CHKERRQ(ierr);
ierr = PetscViewerASCIISubtractTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
}
}
PetscFunctionReturn(0);
}
static PetscErrorCode SNESSolve_QN(SNES snes)
{
PetscErrorCode ierr;
SNES_QN *qn = (SNES_QN*) snes->data;
Vec X,Xold;
Vec F,B;
Vec Y,FPC,D,Dold;
SNESConvergedReason reason;
PetscInt i, i_r;
PetscReal fnorm,xnorm,ynorm,gnorm;
PetscBool lssucceed,powell,periodic;
PetscScalar DolddotD,DolddotDold,DdotD,YdotD;
MatStructure flg = DIFFERENT_NONZERO_PATTERN;
/* basically just a regular newton's method except for the application of the jacobian */
PetscFunctionBegin;
F = snes->vec_func; /* residual vector */
Y = snes->vec_sol_update; /* search direction generated by J^-1D*/
B = snes->vec_rhs;
X = snes->vec_sol; /* solution vector */
Xold = snes->work[0];
/* directions generated by the preconditioned problem with F_pre = F or x - M(x, b) */
D = snes->work[1];
Dold = snes->work[2];
snes->reason = SNES_CONVERGED_ITERATING;
ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
snes->iter = 0;
snes->norm = 0.;
ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
if (!snes->vec_func_init_set){
ierr = SNESComputeFunction(snes,X,F);CHKERRQ(ierr);
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
PetscFunctionReturn(0);
}
} else {
snes->vec_func_init_set = PETSC_FALSE;
}
if (!snes->norm_init_set) {
ierr = VecNorm(F, NORM_2, &fnorm);CHKERRQ(ierr); /* fnorm <- ||F|| */
if (PetscIsInfOrNanReal(fnorm)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FP,"Infinite or not-a-number generated in norm");
} else {
fnorm = snes->norm_init;
snes->norm_init_set = PETSC_FALSE;
}
ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
snes->norm = fnorm;
ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
SNESLogConvHistory(snes,fnorm,0);
ierr = SNESMonitor(snes,0,fnorm);CHKERRQ(ierr);
/* set parameter for default relative tolerance convergence test */
snes->ttol = fnorm*snes->rtol;
/* test convergence */
ierr = (*snes->ops->converged)(snes,0,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
/* composed solve */
if (snes->pc && snes->pcside == PC_RIGHT) {
ierr = SNESSetInitialFunction(snes->pc, F);CHKERRQ(ierr);
ierr = SNESSetInitialFunctionNorm(snes->pc, fnorm);CHKERRQ(ierr);
ierr = SNESSolve(snes->pc, B, X);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && (reason != SNES_DIVERGED_MAX_IT)) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = SNESGetFunction(snes->pc, &FPC, PETSC_NULL, PETSC_NULL);CHKERRQ(ierr);
ierr = VecCopy(FPC, F);CHKERRQ(ierr);
ierr = SNESGetFunctionNorm(snes->pc, &fnorm);CHKERRQ(ierr);
ierr = VecCopy(F, Y);CHKERRQ(ierr);
} else {
ierr = VecCopy(F, Y);CHKERRQ(ierr);
}
ierr = VecCopy(Y, D);CHKERRQ(ierr);
/* scale the initial update */
if (qn->scale_type == SNES_QN_SCALE_JACOBIAN) {
ierr = SNESComputeJacobian(snes,X,&snes->jacobian,&snes->jacobian_pre,&flg);CHKERRQ(ierr);
}
for (i = 0, i_r = 0; i < snes->max_its; i++, i_r++) {
switch(qn->type) {
case SNES_QN_BADBROYDEN:
ierr = SNESQNApply_BadBroyden(snes,i_r,Y,X,Xold,D,Dold);CHKERRQ(ierr);
break;
case SNES_QN_BROYDEN:
ierr = SNESQNApply_Broyden(snes,i_r,Y,X,Xold,D,Dold);CHKERRQ(ierr);
break;
case SNES_QN_LBFGS:
SNESQNApply_LBFGS(snes,i_r,Y,X,Xold,D,Dold);CHKERRQ(ierr);
break;
}
/* line search for lambda */
ynorm = 1; gnorm = fnorm;
ierr = VecCopy(D, Dold);CHKERRQ(ierr);
ierr = VecCopy(X, Xold);CHKERRQ(ierr);
ierr = SNESLineSearchApply(snes->linesearch, X, F, &fnorm, Y);CHKERRQ(ierr);
if (snes->reason == SNES_DIVERGED_FUNCTION_COUNT) break;
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
PetscFunctionReturn(0);
}
ierr = SNESLineSearchGetSuccess(snes->linesearch, &lssucceed);CHKERRQ(ierr);
if (!lssucceed) {
if (++snes->numFailures >= snes->maxFailures) {
snes->reason = SNES_DIVERGED_LINE_SEARCH;
break;
}
}
ierr = SNESLineSearchGetNorms(snes->linesearch, &xnorm, &fnorm, &ynorm);CHKERRQ(ierr);
if (qn->scale_type == SNES_QN_SCALE_LINESEARCH) {
ierr = SNESLineSearchGetLambda(snes->linesearch, &qn->scaling);CHKERRQ(ierr);
}
/* convergence monitoring */
ierr = PetscInfo4(snes,"fnorm=%18.16e, gnorm=%18.16e, ynorm=%18.16e, lssucceed=%d\n",(double)fnorm,(double)gnorm,(double)ynorm,(int)lssucceed);CHKERRQ(ierr);
ierr = SNESSetIterationNumber(snes, i+1);CHKERRQ(ierr);
ierr = SNESSetFunctionNorm(snes, fnorm);CHKERRQ(ierr);
SNESLogConvHistory(snes,snes->norm,snes->iter);
ierr = SNESMonitor(snes,snes->iter,snes->norm);CHKERRQ(ierr);
/* set parameter for default relative tolerance convergence test */
ierr = (*snes->ops->converged)(snes,snes->iter,xnorm,ynorm,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
if (snes->pc && snes->pcside == PC_RIGHT) {
ierr = SNESSetInitialFunction(snes->pc, F);CHKERRQ(ierr);
ierr = SNESSetInitialFunctionNorm(snes->pc, fnorm);CHKERRQ(ierr);
ierr = SNESSolve(snes->pc, B, X);CHKERRQ(ierr);
ierr = SNESGetConvergedReason(snes->pc,&reason);CHKERRQ(ierr);
if (reason < 0 && (reason != SNES_DIVERGED_MAX_IT)) {
snes->reason = SNES_DIVERGED_INNER;
PetscFunctionReturn(0);
}
ierr = SNESGetFunction(snes->pc, &FPC, PETSC_NULL, PETSC_NULL);CHKERRQ(ierr);
ierr = VecCopy(FPC, F);CHKERRQ(ierr);
ierr = SNESGetFunctionNorm(snes->pc, &fnorm);CHKERRQ(ierr);
ierr = VecCopy(F, D);CHKERRQ(ierr);
} else {
ierr = VecCopy(F, D);CHKERRQ(ierr);
}
powell = PETSC_FALSE;
if (qn->restart_type == SNES_QN_RESTART_POWELL) {
/* check restart by Powell's Criterion: |F^T H_0 Fold| > 0.2 * |Fold^T H_0 Fold| */
ierr = VecDotBegin(Dold, Dold, &DolddotDold);CHKERRQ(ierr);
ierr = VecDotBegin(Dold, D, &DolddotD);CHKERRQ(ierr);
ierr = VecDotBegin(D, D, &DdotD);CHKERRQ(ierr);
ierr = VecDotBegin(Y, D, &YdotD);CHKERRQ(ierr);
ierr = VecDotEnd(Dold, Dold, &DolddotDold);CHKERRQ(ierr);
ierr = VecDotEnd(Dold, D, &DolddotD);CHKERRQ(ierr);
ierr = VecDotEnd(D, D, &DdotD);CHKERRQ(ierr);
ierr = VecDotEnd(Y, D, &YdotD);CHKERRQ(ierr);
if (PetscAbs(PetscRealPart(DolddotD)) > qn->powell_gamma*PetscAbs(PetscRealPart(DolddotDold))) powell = PETSC_TRUE;
}
periodic = PETSC_FALSE;
if (qn->restart_type == SNES_QN_RESTART_PERIODIC) {
if (i_r>qn->m-1) periodic = PETSC_TRUE;
}
/* restart if either powell or periodic restart is satisfied. */
if (powell || periodic) {
if (qn->monitor) {
ierr = PetscViewerASCIIAddTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
ierr = PetscViewerASCIIPrintf(qn->monitor, "restart! |%14.12e| > %4.2f*|%14.12e| or i_r = %d\n", PetscRealPart(DolddotD), qn->powell_gamma, PetscRealPart(DolddotDold), i_r);CHKERRQ(ierr);
ierr = PetscViewerASCIISubtractTab(qn->monitor,((PetscObject)snes)->tablevel+2);CHKERRQ(ierr);
}
i_r = -1;
/* general purpose update */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
}
if (qn->scale_type == SNES_QN_SCALE_JACOBIAN) {
ierr = SNESComputeJacobian(snes,X,&snes->jacobian,&snes->jacobian_pre,&flg);CHKERRQ(ierr);
}
}
/* general purpose update */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
}
}
if (i == snes->max_its) {
ierr = PetscInfo1(snes, "Maximum number of iterations has been reached: %D\n", snes->max_its);CHKERRQ(ierr);
if (!snes->reason) snes->reason = SNES_DIVERGED_MAX_IT;
}
PetscFunctionReturn(0);
}
