static PetscErrorCode KSPSolve_GCR(KSP ksp)
{
KSP_GCR *ctx = (KSP_GCR*)ksp->data;
PetscErrorCode ierr;
Mat A, B;
Vec r,b,x;
PetscReal norm_r;
PetscFunctionBegin;
ierr = KSPGetOperators(ksp, &A, &B);CHKERRQ(ierr);
x = ksp->vec_sol;
b = ksp->vec_rhs;
r = ctx->R;
/* compute initial residual */
ierr = KSP_MatMult(ksp,A, x, r);CHKERRQ(ierr);
ierr = VecAYPX(r, -1.0, b);CHKERRQ(ierr); /* r = b - A x */
ierr = VecNorm(r, NORM_2, &norm_r);CHKERRQ(ierr);
KSPCheckNorm(ksp,norm_r);
ksp->its = 0;
ksp->rnorm0 = norm_r;
ierr = KSPLogResidualHistory(ksp,ksp->rnorm0);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,ksp->rnorm0);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,ksp->its,ksp->rnorm0,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
do {
ierr = KSPSolve_GCR_cycle(ksp);CHKERRQ(ierr);
if (ksp->reason) break; /* catch case when convergence occurs inside the cycle */
} while (ksp->its < ksp->max_it);CHKERRQ(ierr);
if (ksp->its >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_CGLS(KSP ksp)
{
PetscErrorCode ierr;
KSP_CGLS *cgls = (KSP_CGLS*)ksp->data;
Mat A;
Vec x,b,r,p,q,ss;
PetscScalar beta;
PetscReal alpha,gamma,oldgamma;
PetscInt maxiter_ls = 15;
PetscFunctionBegin;
ierr = KSPGetOperators(ksp,&A,NULL);CHKERRQ(ierr); /* Matrix of the system */
/* vectors of length n, where system size is mxn */
x = ksp->vec_sol; /* Solution vector */
p = cgls->vwork_n[0];
ss = cgls->vwork_n[1];
/* vectors of length m, where system size is mxn */
b = ksp->vec_rhs; /* Right-hand side vector */
r = cgls->vwork_m[0];
q = cgls->vwork_m[1];
/* Minimization with the CGLS method */
ksp->its = 0;
ierr = MatMult(A,x,r);CHKERRQ(ierr);
ierr = VecAYPX(r,-1,b);CHKERRQ(ierr); /* r_0 = b - A * x_0 */
ierr = MatMultTranspose(A,r,p);CHKERRQ(ierr); /* p_0 = A^T * r_0 */
ierr = VecCopy(p,ss);CHKERRQ(ierr); /* s_0 = p_0 */
ierr = VecNorm(ss,NORM_2,&gamma);CHKERRQ(ierr);
KSPCheckNorm(ksp,gamma);
ksp->rnorm = gamma;
ierr = (*ksp->converged)(ksp,ksp->its,ksp->rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
gamma = gamma*gamma; /* gamma = norm2(s)^2 */
do {
ierr = MatMult(A,p,q);CHKERRQ(ierr); /* q = A * p */
ierr = VecNorm(q,NORM_2,&alpha);CHKERRQ(ierr);
KSPCheckNorm(ksp,alpha);
alpha = alpha*alpha; /* alpha = norm2(q)^2 */
alpha = gamma / alpha; /* alpha = gamma / alpha */
ierr = VecAXPY(x,alpha,p);CHKERRQ(ierr); /* x += alpha * p */
ierr = VecAXPY(r,-alpha,q);CHKERRQ(ierr); /* r -= alpha * q */
ierr = MatMultTranspose(A,r,ss);CHKERRQ(ierr); /* ss = A^T * r */
oldgamma = gamma; /* oldgamma = gamma */
ierr = VecNorm(ss,NORM_2,&gamma);CHKERRQ(ierr);
KSPCheckNorm(ksp,gamma);
ksp->its++;
ksp->rnorm = gamma;
ierr = KSPMonitor(ksp,ksp->its,ksp->rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,ksp->its,ksp->rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
gamma = gamma*gamma; /* gamma = norm2(s)^2 */
beta = gamma/oldgamma; /* beta = gamma / oldgamma */
ierr = VecAYPX(p,beta,ss);CHKERRQ(ierr); /* p = s + beta * p */
} while (ksp->its<ksp->max_it && !ksp->reason);
if (ksp->its>=maxiter_ls && !ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPAGMRESCycle(PetscInt *itcount,KSP ksp)
{
KSP_AGMRES *agmres = (KSP_AGMRES*)(ksp->data);
PetscReal res;
PetscErrorCode ierr;
PetscInt KspSize = KSPSIZE;
PetscFunctionBegin;
/* check for the convergence */
res = ksp->rnorm; /* Norm of the initial residual vector */
if (!res) {
if (itcount) *itcount = 0;
ksp->reason = KSP_CONVERGED_ATOL;
ierr = PetscInfo(ksp,"Converged due to zero residual norm on entry\n");CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
/* Build the Krylov basis with Newton polynomials */
ierr = KSPAGMRESBuildBasis(ksp);CHKERRQ(ierr);
/* QR Factorize the basis with RODDEC */
ierr = KSPAGMRESRoddec(ksp, KspSize+1);CHKERRQ(ierr);
/* Recover a (partial) Hessenberg matrix for the Arnoldi-like relation */
ierr = KSPAGMRESBuildHessenberg(ksp);CHKERRQ(ierr);
/* Solve the least square problem and unwind the preconditioner */
ierr = KSPAGMRESBuildSoln(ksp, KspSize);CHKERRQ(ierr);
res = ksp->rnorm;
ksp->its += KspSize;
agmres->it = KspSize-1;
/* Test for the convergence */
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
*itcount = KspSize;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_PIPEFCG_cycle(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,j,k,idx,kdx,mi;
KSP_PIPEFCG *pipefcg;
PetscScalar alpha=0.0,gamma,*betas,*dots;
PetscReal dp=0.0, delta,*eta,*etas;
Vec B,R,Z,X,Qcurr,W,ZETAcurr,M,N,Pcurr,Scurr,*redux;
Mat Amat,Pmat;
PetscFunctionBegin;
/* We have not checked these routines for use with complex numbers. The inner products
are likely not defined correctly for that case */
#if (defined(PETSC_USE_COMPLEX) && !defined(PETSC_SKIP_COMPLEX))
SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_SUP,"PIPEFGMRES has not been implemented for use with complex scalars");
#endif
#define VecXDot(x,y,a) (((pipefcg->type) == (KSP_CG_HERMITIAN)) ? VecDot (x,y,a) : VecTDot (x,y,a))
#define VecXDotBegin(x,y,a) (((pipefcg->type) == (KSP_CG_HERMITIAN)) ? VecDotBegin (x,y,a) : VecTDotBegin (x,y,a))
#define VecXDotEnd(x,y,a) (((pipefcg->type) == (KSP_CG_HERMITIAN)) ? VecDotEnd (x,y,a) : VecTDotEnd (x,y,a))
#define VecMXDot(x,n,y,a) (((pipefcg->type) == (KSP_CG_HERMITIAN)) ? VecMDot (x,n,y,a) : VecMTDot (x,n,y,a))
#define VecMXDotBegin(x,n,y,a) (((pipefcg->type) == (KSP_CG_HERMITIAN)) ? VecMDotBegin (x,n,y,a) : VecMTDotBegin (x,n,y,a))
#define VecMXDotEnd(x,n,y,a) (((pipefcg->type) == (KSP_CG_HERMITIAN)) ? VecMDotEnd (x,n,y,a) : VecMTDotEnd (x,n,y,a))
pipefcg = (KSP_PIPEFCG*)ksp->data;
X = ksp->vec_sol;
B = ksp->vec_rhs;
R = ksp->work[0];
Z = ksp->work[1];
W = ksp->work[2];
M = ksp->work[3];
N = ksp->work[4];
redux = pipefcg->redux;
dots = pipefcg->dots;
etas = pipefcg->etas;
betas = dots; /* dots takes the result of all dot products of which the betas are a subset */
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
/* Compute cycle initial residual */
ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr);
ierr = VecAYPX(R,-1.0,B);CHKERRQ(ierr); /* r <- b - Ax */
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
Pcurr = pipefcg->Pvecs[0];
Scurr = pipefcg->Svecs[0];
Qcurr = pipefcg->Qvecs[0];
ZETAcurr = pipefcg->ZETAvecs[0];
ierr = VecCopy(Z,Pcurr);CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,Pcurr,Scurr);CHKERRQ(ierr); /* S = Ap */
ierr = VecCopy(Scurr,W);CHKERRQ(ierr); /* w = s = Az */
/* Initial state of pipelining intermediates */
redux[0] = R;
redux[1] = W;
ierr = VecMXDotBegin(Z,2,redux,dots);CHKERRQ(ierr);
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)Z));CHKERRQ(ierr); /* perform asynchronous reduction */
ierr = KSP_PCApply(ksp,W,M);CHKERRQ(ierr); /* m = B(w) */
ierr = KSP_MatMult(ksp,Amat,M,N);CHKERRQ(ierr); /* n = Am */
ierr = VecCopy(M,Qcurr);CHKERRQ(ierr); /* q = m */
ierr = VecCopy(N,ZETAcurr);CHKERRQ(ierr); /* zeta = n */
ierr = VecMXDotEnd(Z,2,redux,dots);CHKERRQ(ierr);
gamma = dots[0];
delta = PetscRealPart(dots[1]);
etas[0] = delta;
alpha = gamma/delta;
i = 0;
do {
ksp->its++;
/* Update X, R, Z, W */
ierr = VecAXPY(X,+alpha,Pcurr);CHKERRQ(ierr); /* x <- x + alpha * pi */
ierr = VecAXPY(R,-alpha,Scurr);CHKERRQ(ierr); /* r <- r - alpha * si */
ierr = VecAXPY(Z,-alpha,Qcurr);CHKERRQ(ierr); /* z <- z - alpha * qi */
ierr = VecAXPY(W,-alpha,ZETAcurr);CHKERRQ(ierr); /* w <- w - alpha * zetai */
/* Compute norm for convergence check */
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
ierr = VecNorm(Z,NORM_2,&dp);CHKERRQ(ierr); /* dp <- sqrt(z'*z) = sqrt(e'*A'*B'*B*A*e) */
break;
case KSP_NORM_UNPRECONDITIONED:
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- sqrt(r'*r) = sqrt(e'*A'*A*e) */
break;
case KSP_NORM_NATURAL:
dp = PetscSqrtReal(PetscAbsScalar(gamma)); /* dp <- sqrt(r'*z) = sqrt(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]);
}
/* Check for convergence */
ksp->rnorm = dp;
KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,dp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,ksp->its+1,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
/* Computations of current iteration done */
++i;
/* If needbe, allocate a new chunk of vectors in P and C */
ierr = KSPAllocateVectors_PIPEFCG(ksp,i+1,pipefcg->vecb);CHKERRQ(ierr);
/* Note that we wrap around and start clobbering old vectors */
idx = i % (pipefcg->mmax+1);
Pcurr = pipefcg->Pvecs[idx];
Scurr = pipefcg->Svecs[idx];
Qcurr = pipefcg->Qvecs[idx];
ZETAcurr = pipefcg->ZETAvecs[idx];
eta = pipefcg->etas+idx;
/* number of old directions to orthogonalize against */
switch(pipefcg->truncstrat){
case KSP_FCD_TRUNC_TYPE_STANDARD:
mi = pipefcg->mmax;
break;
case KSP_FCD_TRUNC_TYPE_NOTAY:
mi = ((i-1) % pipefcg->mmax)+1;
break;
default:
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unrecognized Truncation Strategy");
}
/* Pick old p,s,q,zeta in a way suitable for VecMDot */
ierr = VecCopy(Z,Pcurr);CHKERRQ(ierr);
for(k=PetscMax(0,i-mi),j=0;k<i;++j,++k){
kdx = k % (pipefcg->mmax+1);
pipefcg->Pold[j] = pipefcg->Pvecs[kdx];
pipefcg->Sold[j] = pipefcg->Svecs[kdx];
pipefcg->Qold[j] = pipefcg->Qvecs[kdx];
pipefcg->ZETAold[j] = pipefcg->ZETAvecs[kdx];
redux[j] = pipefcg->Svecs[kdx];
}
redux[j] = R; /* If the above loop is not executed redux contains only R => all beta_k = 0, only gamma, delta != 0 */
redux[j+1] = W;
ierr = VecMXDotBegin(Z,j+2,redux,betas);CHKERRQ(ierr); /* Start split reductions for beta_k = (z,s_k), gamma = (z,r), delta = (z,w) */
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)Z));CHKERRQ(ierr); /* perform asynchronous reduction */
ierr = VecWAXPY(N,-1.0,R,W);CHKERRQ(ierr); /* m = u + B(w-r): (a) ntmp = w-r */
ierr = KSP_PCApply(ksp,N,M);CHKERRQ(ierr); /* m = u + B(w-r): (b) mtmp = B(ntmp) = B(w-r) */
ierr = VecAXPY(M,1.0,Z);CHKERRQ(ierr); /* m = u + B(w-r): (c) m = z + mtmp */
ierr = KSP_MatMult(ksp,Amat,M,N);CHKERRQ(ierr); /* n = Am */
ierr = VecMXDotEnd(Z,j+2,redux,betas);CHKERRQ(ierr); /* Finish split reductions */
gamma = betas[j];
delta = PetscRealPart(betas[j+1]);
*eta = 0.;
for(k=PetscMax(0,i-mi),j=0;k<i;++j,++k){
kdx = k % (pipefcg->mmax+1);
betas[j] /= -etas[kdx]; /* betak /= etak */
*eta -= ((PetscReal)(PetscAbsScalar(betas[j])*PetscAbsScalar(betas[j]))) * etas[kdx];
/* etaitmp = -betaik^2 * etak */
}
*eta += delta; /* etai = delta -betaik^2 * etak */
if(*eta < 0.) {
pipefcg->norm_breakdown = PETSC_TRUE;
ierr = PetscInfo1(ksp,"Restart due to square root breakdown at it = \n",ksp->its);CHKERRQ(ierr);
break;
} else {
alpha= gamma/(*eta); /* alpha = gamma/etai */
}
/* project out stored search directions using classical G-S */
ierr = VecCopy(Z,Pcurr);CHKERRQ(ierr);
ierr = VecCopy(W,Scurr);CHKERRQ(ierr);
ierr = VecCopy(M,Qcurr);CHKERRQ(ierr);
ierr = VecCopy(N,ZETAcurr);CHKERRQ(ierr);
ierr = VecMAXPY(Pcurr ,j,betas,pipefcg->Pold);CHKERRQ(ierr); /* pi <- ui - sum_k beta_k p_k */
ierr = VecMAXPY(Scurr ,j,betas,pipefcg->Sold);CHKERRQ(ierr); /* si <- wi - sum_k beta_k s_k */
ierr = VecMAXPY(Qcurr ,j,betas,pipefcg->Qold);CHKERRQ(ierr); /* qi <- m - sum_k beta_k q_k */
ierr = VecMAXPY(ZETAcurr,j,betas,pipefcg->ZETAold);CHKERRQ(ierr); /* zetai <- n - sum_k beta_k zeta_k */
} while (ksp->its < ksp->max_it);
PetscFunctionReturn(0);
}
PetscErrorCode KSPFGMRESCycle(PetscInt *itcount,KSP ksp)
{
KSP_FGMRES *fgmres = (KSP_FGMRES*)(ksp->data);
PetscReal res_norm;
PetscReal hapbnd,tt;
PetscBool hapend = PETSC_FALSE; /* indicates happy breakdown ending */
PetscErrorCode ierr;
PetscInt loc_it; /* local count of # of dir. in Krylov space */
PetscInt max_k = fgmres->max_k; /* max # of directions Krylov space */
Mat Amat,Pmat;
MatStructure pflag;
PetscFunctionBegin;
/* Number of pseudo iterations since last restart is the number
of prestart directions */
loc_it = 0;
/* note: (fgmres->it) is always set one less than (loc_it) It is used in
KSPBUILDSolution_FGMRES, where it is passed to KSPFGMRESBuildSoln.
Note that when KSPFGMRESBuildSoln is called from this function,
(loc_it -1) is passed, so the two are equivalent */
fgmres->it = (loc_it - 1);
/* initial residual is in VEC_VV(0) - compute its norm*/
ierr = VecNorm(VEC_VV(0),NORM_2,&res_norm);CHKERRQ(ierr);
/* first entry in right-hand-side of hessenberg system is just
the initial residual norm */
*RS(0) = res_norm;
ksp->rnorm = res_norm;
ierr = KSPLogResidualHistory(ksp,res_norm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res_norm);CHKERRQ(ierr);
/* check for the convergence - maybe the current guess is good enough */
ierr = (*ksp->converged)(ksp,ksp->its,res_norm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) {
if (itcount) *itcount = 0;
PetscFunctionReturn(0);
}
/* scale VEC_VV (the initial residual) */
ierr = VecScale(VEC_VV(0),1.0/res_norm);CHKERRQ(ierr);
/* MAIN ITERATION LOOP BEGINNING*/
/* keep iterating until we have converged OR generated the max number
of directions OR reached the max number of iterations for the method */
while (!ksp->reason && loc_it < max_k && ksp->its < ksp->max_it) {
if (loc_it) {
ierr = KSPLogResidualHistory(ksp,res_norm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res_norm);CHKERRQ(ierr);
}
fgmres->it = (loc_it - 1);
/* see if more space is needed for work vectors */
if (fgmres->vv_allocated <= loc_it + VEC_OFFSET + 1) {
ierr = KSPFGMRESGetNewVectors(ksp,loc_it+1);CHKERRQ(ierr);
/* (loc_it+1) is passed in as number of the first vector that should
be allocated */
}
/* CHANGE THE PRECONDITIONER? */
/* ModifyPC is the callback function that can be used to
change the PC or its attributes before its applied */
(*fgmres->modifypc)(ksp,ksp->its,loc_it,res_norm,fgmres->modifyctx);
/* apply PRECONDITIONER to direction vector and store with
preconditioned vectors in prevec */
ierr = KSP_PCApply(ksp,VEC_VV(loc_it),PREVEC(loc_it));CHKERRQ(ierr);
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat,&pflag);CHKERRQ(ierr);
/* Multiply preconditioned vector by operator - put in VEC_VV(loc_it+1) */
ierr = MatMult(Amat,PREVEC(loc_it),VEC_VV(1+loc_it));CHKERRQ(ierr);
/* update hessenberg matrix and do Gram-Schmidt - new direction is in
VEC_VV(1+loc_it)*/
ierr = (*fgmres->orthog)(ksp,loc_it);CHKERRQ(ierr);
/* new entry in hessenburg is the 2-norm of our new direction */
ierr = VecNorm(VEC_VV(loc_it+1),NORM_2,&tt);CHKERRQ(ierr);
*HH(loc_it+1,loc_it) = tt;
*HES(loc_it+1,loc_it) = tt;
/* Happy Breakdown Check */
hapbnd = PetscAbsScalar((tt) / *RS(loc_it));
/* RS(loc_it) contains the res_norm from the last iteration */
hapbnd = PetscMin(fgmres->haptol,hapbnd);
if (tt > hapbnd) {
/* scale new direction by its norm */
ierr = VecScale(VEC_VV(loc_it+1),1.0/tt);CHKERRQ(ierr);
} else {
/* This happens when the solution is exactly reached. */
/* So there is no new direction... */
ierr = VecSet(VEC_TEMP,0.0);CHKERRQ(ierr); /* set VEC_TEMP to 0 */
hapend = PETSC_TRUE;
}
/* note that for FGMRES we could get HES(loc_it+1, loc_it) = 0 and the
current solution would not be exact if HES was singular. Note that
HH non-singular implies that HES is no singular, and HES is guaranteed
to be nonsingular when PREVECS are linearly independent and A is
nonsingular (in GMRES, the nonsingularity of A implies the nonsingularity
of HES). So we should really add a check to verify that HES is nonsingular.*/
/* Now apply rotations to new col of hessenberg (and right side of system),
calculate new rotation, and get new residual norm at the same time*/
ierr = KSPFGMRESUpdateHessenberg(ksp,loc_it,hapend,&res_norm);CHKERRQ(ierr);
if (ksp->reason) break;
loc_it++;
fgmres->it = (loc_it-1); /* Add this here in case it has converged */
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = res_norm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,ksp->its,res_norm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
/* Catch error in happy breakdown and signal convergence and break from loop */
if (hapend) {
if (!ksp->reason) {
if (ksp->errorifnotconverged) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"You reached the happy break down, but convergence was not indicated. Residual norm = %G",res_norm);
else {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
}
}
}
/* END OF ITERATION LOOP */
ierr = KSPLogResidualHistory(ksp,res_norm);CHKERRQ(ierr);
/*
Monitor if we know that we will not return for a restart */
if (loc_it && (ksp->reason || ksp->its >= ksp->max_it)) {
ierr = KSPMonitor(ksp,ksp->its,res_norm);CHKERRQ(ierr);
}
if (itcount) *itcount = loc_it;
/*
Down here we have to solve for the "best" coefficients of the Krylov
columns, add the solution values together, and possibly unwind the
preconditioning from the solution
*/
/* Form the solution (or the solution so far) */
/* Note: must pass in (loc_it-1) for iteration count so that KSPFGMRESBuildSoln
properly navigates */
ierr = KSPFGMRESBuildSoln(RS(0),ksp->vec_sol,ksp->vec_sol,ksp,loc_it-1);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_QCG(KSP ksp)
{
/*
Correpondence with documentation above:
B = g = gradient,
X = s = step
Note: This is not coded correctly for complex arithmetic!
*/
KSP_QCG *pcgP = (KSP_QCG*)ksp->data;
Mat Amat,Pmat;
Vec W,WA,WA2,R,P,ASP,BS,X,B;
PetscScalar scal,beta,rntrn,step;
PetscReal q1,q2,xnorm,step1,step2,rnrm,btx,xtax;
PetscReal ptasp,rtr,wtasp,bstp;
PetscReal dzero = 0.0,bsnrm;
PetscErrorCode ierr;
PetscInt i,maxit;
PC pc = ksp->pc;
PCSide side;
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);
if (ksp->transpose_solve) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Currently does not support transpose solve");
ksp->its = 0;
maxit = ksp->max_it;
WA = ksp->work[0];
R = ksp->work[1];
P = ksp->work[2];
ASP = ksp->work[3];
BS = ksp->work[4];
W = ksp->work[5];
WA2 = ksp->work[6];
X = ksp->vec_sol;
B = ksp->vec_rhs;
if (pcgP->delta <= dzero) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_ARG_OUTOFRANGE,"Input error: delta <= 0");
ierr = KSPGetPCSide(ksp,&side);CHKERRQ(ierr);
if (side != PC_SYMMETRIC) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_ARG_OUTOFRANGE,"Requires symmetric preconditioner!");
/* Initialize variables */
ierr = VecSet(W,0.0);CHKERRQ(ierr); /* W = 0 */
ierr = VecSet(X,0.0);CHKERRQ(ierr); /* X = 0 */
ierr = PCGetOperators(pc,&Amat,&Pmat);CHKERRQ(ierr);
/* Compute: BS = D^{-1} B */
ierr = PCApplySymmetricLeft(pc,B,BS);CHKERRQ(ierr);
ierr = VecNorm(BS,NORM_2,&bsnrm);CHKERRQ(ierr);
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = bsnrm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,bsnrm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,bsnrm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,0,bsnrm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
/* Compute the initial scaled direction and scaled residual */
ierr = VecCopy(BS,R);CHKERRQ(ierr);
ierr = VecScale(R,-1.0);CHKERRQ(ierr);
ierr = VecCopy(R,P);CHKERRQ(ierr);
ierr = VecDotRealPart(R,R,&rtr);CHKERRQ(ierr);
for (i=0; i<=maxit; i++) {
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
/* Compute: asp = D^{-T}*A*D^{-1}*p */
ierr = PCApplySymmetricRight(pc,P,WA);CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,WA,WA2);CHKERRQ(ierr);
ierr = PCApplySymmetricLeft(pc,WA2,ASP);CHKERRQ(ierr);
/* Check for negative curvature */
ierr = VecDotRealPart(P,ASP,&ptasp);CHKERRQ(ierr);
if (ptasp <= dzero) {
/* Scaled negative curvature direction: Compute a step so that
||w + step*p|| = delta and QS(w + step*p) is least */
if (!i) {
ierr = VecCopy(P,X);CHKERRQ(ierr);
ierr = VecNorm(X,NORM_2,&xnorm);CHKERRQ(ierr);
scal = pcgP->delta / xnorm;
ierr = VecScale(X,scal);CHKERRQ(ierr);
} else {
/* Compute roots of quadratic */
ierr = KSPQCGQuadraticRoots(W,P,pcgP->delta,&step1,&step2);CHKERRQ(ierr);
ierr = VecDotRealPart(W,ASP,&wtasp);CHKERRQ(ierr);
ierr = VecDotRealPart(BS,P,&bstp);CHKERRQ(ierr);
ierr = VecCopy(W,X);CHKERRQ(ierr);
q1 = step1*(bstp + wtasp + .5*step1*ptasp);
q2 = step2*(bstp + wtasp + .5*step2*ptasp);
if (q1 <= q2) {
ierr = VecAXPY(X,step1,P);CHKERRQ(ierr);
} else {
ierr = VecAXPY(X,step2,P);CHKERRQ(ierr);
}
}
pcgP->ltsnrm = pcgP->delta; /* convergence in direction of */
ksp->reason = KSP_CONVERGED_CG_NEG_CURVE; /* negative curvature */
if (!i) {
ierr = PetscInfo1(ksp,"negative curvature: delta=%g\n",(double)pcgP->delta);CHKERRQ(ierr);
} else {
ierr = PetscInfo3(ksp,"negative curvature: step1=%g, step2=%g, delta=%g\n",(double)step1,(double)step2,(double)pcgP->delta);CHKERRQ(ierr);
}
} else {
/* Compute step along p */
step = rtr/ptasp;
ierr = VecCopy(W,X);CHKERRQ(ierr); /* x = w */
ierr = VecAXPY(X,step,P);CHKERRQ(ierr); /* x <- step*p + x */
ierr = VecNorm(X,NORM_2,&pcgP->ltsnrm);CHKERRQ(ierr);
if (pcgP->ltsnrm > pcgP->delta) {
/* Since the trial iterate is outside the trust region,
evaluate a constrained step along p so that
||w + step*p|| = delta
The positive step is always better in this case. */
if (!i) {
scal = pcgP->delta / pcgP->ltsnrm;
ierr = VecScale(X,scal);CHKERRQ(ierr);
} else {
/* Compute roots of quadratic */
ierr = KSPQCGQuadraticRoots(W,P,pcgP->delta,&step1,&step2);CHKERRQ(ierr);
ierr = VecCopy(W,X);CHKERRQ(ierr);
ierr = VecAXPY(X,step1,P);CHKERRQ(ierr); /* x <- step1*p + x */
}
pcgP->ltsnrm = pcgP->delta;
ksp->reason = KSP_CONVERGED_CG_CONSTRAINED; /* convergence along constrained step */
if (!i) {
ierr = PetscInfo1(ksp,"constrained step: delta=%g\n",(double)pcgP->delta);CHKERRQ(ierr);
} else {
ierr = PetscInfo3(ksp,"constrained step: step1=%g, step2=%g, delta=%g\n",(double)step1,(double)step2,(double)pcgP->delta);CHKERRQ(ierr);
}
} else {
/* Evaluate the current step */
ierr = VecCopy(X,W);CHKERRQ(ierr); /* update interior iterate */
ierr = VecAXPY(R,-step,ASP);CHKERRQ(ierr); /* r <- -step*asp + r */
ierr = VecNorm(R,NORM_2,&rnrm);CHKERRQ(ierr);
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->rnorm = rnrm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rnrm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i+1,rnrm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i+1,rnrm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) { /* convergence for */
ierr = PetscInfo3(ksp,"truncated step: step=%g, rnrm=%g, delta=%g\n",(double)PetscRealPart(step),(double)rnrm,(double)pcgP->delta);CHKERRQ(ierr);
}
}
}
if (ksp->reason) break; /* Convergence has been attained */
else { /* Compute a new AS-orthogonal direction */
ierr = VecDot(R,R,&rntrn);CHKERRQ(ierr);
beta = rntrn/rtr;
ierr = VecAYPX(P,beta,R);CHKERRQ(ierr); /* p <- r + beta*p */
rtr = PetscRealPart(rntrn);
}
}
if (!ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
/* Unscale x */
ierr = VecCopy(X,WA2);CHKERRQ(ierr);
ierr = PCApplySymmetricRight(pc,WA2,X);CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,X,WA);CHKERRQ(ierr);
ierr = VecDotRealPart(B,X,&btx);CHKERRQ(ierr);
ierr = VecDotRealPart(X,WA,&xtax);CHKERRQ(ierr);
pcgP->quadratic = btx + .5*xtax;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPPGMRESCycle(PetscInt *itcount,KSP ksp)
{
KSP_PGMRES *pgmres = (KSP_PGMRES*)(ksp->data);
PetscReal res_norm,res,newnorm;
PetscErrorCode ierr;
PetscInt it = 0,j,k;
PetscBool hapend = PETSC_FALSE;
PetscFunctionBegin;
if (itcount) *itcount = 0;
ierr = VecNormalize(VEC_VV(0),&res_norm);CHKERRQ(ierr);
res = res_norm;
*RS(0) = res_norm;
/* check for the convergence */
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->rnorm = res;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
pgmres->it = it-2;
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
if (!res) {
ksp->reason = KSP_CONVERGED_ATOL;
ierr = PetscInfo(ksp,"Converged due to zero residual norm on entry\n");CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
for (; !ksp->reason; it++) {
Vec Zcur,Znext;
if (pgmres->vv_allocated <= it + VEC_OFFSET + 1) {
ierr = KSPGMRESGetNewVectors(ksp,it+1);CHKERRQ(ierr);
}
/* VEC_VV(it-1) is orthogonal, it will be normalized once the VecNorm arrives. */
Zcur = VEC_VV(it); /* Zcur is not yet orthogonal, but the VecMDot to orthogonalize it has been started. */
Znext = VEC_VV(it+1); /* This iteration will compute Znext, update with a deferred correction once we know how
* Zcur relates to the previous vectors, and start the reduction to orthogonalize it. */
if (it < pgmres->max_k+1 && ksp->its+1 < PetscMax(2,ksp->max_it)) { /* We don't know whether what we have computed is enough, so apply the matrix. */
ierr = KSP_PCApplyBAorAB(ksp,Zcur,Znext,VEC_TEMP_MATOP);CHKERRQ(ierr);
}
if (it > 1) { /* Complete the pending reduction */
ierr = VecNormEnd(VEC_VV(it-1),NORM_2,&newnorm);CHKERRQ(ierr);
*HH(it-1,it-2) = newnorm;
}
if (it > 0) { /* Finish the reduction computing the latest column of H */
ierr = VecMDotEnd(Zcur,it,&(VEC_VV(0)),HH(0,it-1));CHKERRQ(ierr);
}
if (it > 1) {
/* normalize the base vector from two iterations ago, basis is complete up to here */
ierr = VecScale(VEC_VV(it-1),1./ *HH(it-1,it-2));CHKERRQ(ierr);
ierr = KSPPGMRESUpdateHessenberg(ksp,it-2,&hapend,&res);CHKERRQ(ierr);
pgmres->it = it-2;
ksp->its++;
ksp->rnorm = res;
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (it < pgmres->max_k+1 || ksp->reason || ksp->its == ksp->max_it) { /* Monitor if we are done or still iterating, but not before a restart. */
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
}
if (ksp->reason) break;
/* Catch error in happy breakdown and signal convergence and break from loop */
if (hapend) {
if (ksp->errorifnotconverged) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"You reached the happy break down, but convergence was not indicated. Residual norm = %G",res);
else {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
}
if (!(it < pgmres->max_k+1 && ksp->its < ksp->max_it)) break;
/* The it-2 column of H was not scaled when we computed Zcur, apply correction */
ierr = VecScale(Zcur,1./ *HH(it-1,it-2));CHKERRQ(ierr);
/* And Znext computed in this iteration was computed using the under-scaled Zcur */
ierr = VecScale(Znext,1./ *HH(it-1,it-2));CHKERRQ(ierr);
/* In the previous iteration, we projected an unnormalized Zcur against the Krylov basis, so we need to fix the column of H resulting from that projection. */
for (k=0; k<it; k++) *HH(k,it-1) /= *HH(it-1,it-2);
/* When Zcur was projected against the Krylov basis, VV(it-1) was still not normalized, so fix that too. This
* column is complete except for HH(it,it-1) which we won't know until the next iteration. */
*HH(it-1,it-1) /= *HH(it-1,it-2);
}
if (it > 0) {
PetscScalar *work;
if (!pgmres->orthogwork) {ierr = PetscMalloc((pgmres->max_k + 2)*sizeof(PetscScalar),&pgmres->orthogwork);CHKERRQ(ierr);}
work = pgmres->orthogwork;
/* Apply correction computed by the VecMDot in the last iteration to Znext. The original form is
*
* Znext -= sum_{j=0}^{i-1} Z[j+1] * H[j,i-1]
*
* where
*
* Z[j] = sum_{k=0}^j V[k] * H[k,j-1]
*
* substituting
*
* Znext -= sum_{j=0}^{i-1} sum_{k=0}^{j+1} V[k] * H[k,j] * H[j,i-1]
*
* rearranging the iteration space from row-column to column-row
*
* Znext -= sum_{k=0}^i sum_{j=k-1}^{i-1} V[k] * H[k,j] * H[j,i-1]
*
* Note that column it-1 of HH is correct. For all previous columns, we must look at HES because HH has already
* been transformed to upper triangular form.
*/
for (k=0; k<it+1; k++) {
work[k] = 0;
for (j=PetscMax(0,k-1); j<it-1; j++) work[k] -= *HES(k,j) * *HH(j,it-1);
}
ierr = VecMAXPY(Znext,it+1,work,&VEC_VV(0));CHKERRQ(ierr);
ierr = VecAXPY(Znext,-*HH(it-1,it-1),Zcur);CHKERRQ(ierr);
/* Orthogonalize Zcur against existing basis vectors. */
for (k=0; k<it; k++) work[k] = -*HH(k,it-1);
ierr = VecMAXPY(Zcur,it,work,&VEC_VV(0));CHKERRQ(ierr);
/* Zcur is now orthogonal, and will be referred to as VEC_VV(it) again, though it is still not normalized. */
/* Begin computing the norm of the new vector, will be normalized after the MatMult in the next iteration. */
ierr = VecNormBegin(VEC_VV(it),NORM_2,&newnorm);CHKERRQ(ierr);
}
/* Compute column of H (to the diagonal, but not the subdiagonal) to be able to orthogonalize the newest vector. */
ierr = VecMDotBegin(Znext,it+1,&VEC_VV(0),HH(0,it));CHKERRQ(ierr);
/* Start an asynchronous split-mode reduction, the result of the MDot and Norm will be collected on the next iteration. */
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)Znext));CHKERRQ(ierr);
}
if (itcount) *itcount = it-1; /* Number of iterations actually completed. */
/*
Down here we have to solve for the "best" coefficients of the Krylov
columns, add the solution values together, and possibly unwind the
preconditioning from the solution
*/
/* Form the solution (or the solution so far) */
ierr = KSPPGMRESBuildSoln(RS(0),ksp->vec_sol,ksp->vec_sol,ksp,it-2);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_CGS(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i;
PetscScalar rho,rhoold,a,s,b;
Vec X,B,V,P,R,RP,T,Q,U,AUQ;
PetscReal dp = 0.0;
PetscBool diagonalscale;
PetscFunctionBegin;
/* not sure what residual norm it does use, should use for right preconditioning */
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];
RP = ksp->work[1];
V = ksp->work[2];
T = ksp->work[3];
Q = ksp->work[4];
P = ksp->work[5];
U = ksp->work[6];
AUQ = V;
/* Compute initial preconditioned residual */
ierr = KSPInitialResidual(ksp,X,V,T,R,B);CHKERRQ(ierr);
/* Test for nothing to do */
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr);
if (ksp->normtype == KSP_NORM_NATURAL) dp *= dp;
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = dp;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,dp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,0,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
/* Make the initial Rp == R */
ierr = VecCopy(R,RP);CHKERRQ(ierr);
/* added for Fidap */
/* Penalize Startup - Isaac Hasbani Trick for CGS
Since most initial conditions result in a mostly 0 residual,
we change all the 0 values in the vector RP to the maximum.
*/
if (ksp->normtype == KSP_NORM_NATURAL) {
PetscReal vr0max;
PetscScalar *tmp_RP=0;
PetscInt numnp =0, *max_pos=0;
ierr = VecMax(RP, max_pos, &vr0max);CHKERRQ(ierr);
ierr = VecGetArray(RP, &tmp_RP);CHKERRQ(ierr);
ierr = VecGetLocalSize(RP, &numnp);CHKERRQ(ierr);
for (i=0; i<numnp; i++) {
if (tmp_RP[i] == 0.0) tmp_RP[i] = vr0max;
}
ierr = VecRestoreArray(RP, &tmp_RP);CHKERRQ(ierr);
}
/* end of addition for Fidap */
/* Set the initial conditions */
ierr = VecDot(R,RP,&rhoold);CHKERRQ(ierr); /* rhoold = (r,rp) */
ierr = VecCopy(R,U);CHKERRQ(ierr);
ierr = VecCopy(R,P);CHKERRQ(ierr);
ierr = KSP_PCApplyBAorAB(ksp,P,V,T);CHKERRQ(ierr);
i = 0;
do {
ierr = VecDot(V,RP,&s);CHKERRQ(ierr); /* s <- (v,rp) */
a = rhoold / s; /* a <- rho / s */
ierr = VecWAXPY(Q,-a,V,U);CHKERRQ(ierr); /* q <- u - a v */
ierr = VecWAXPY(T,1.0,U,Q);CHKERRQ(ierr); /* t <- u + q */
ierr = VecAXPY(X,a,T);CHKERRQ(ierr); /* x <- x + a (u + q) */
ierr = KSP_PCApplyBAorAB(ksp,T,AUQ,U);CHKERRQ(ierr);
ierr = VecAXPY(R,-a,AUQ);CHKERRQ(ierr); /* r <- r - a K (u + q) */
ierr = VecDot(R,RP,&rho);CHKERRQ(ierr); /* rho <- (r,rp) */
if (ksp->normtype == KSP_NORM_NATURAL) {
dp = PetscAbsScalar(rho);
} else {
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr);
}
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = dp;
ierr = PetscObjectSAWsGrantAccess((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;
b = rho / rhoold; /* b <- rho / rhoold */
ierr = VecWAXPY(U,b,Q,R);CHKERRQ(ierr); /* u <- r + b q */
ierr = VecAXPY(Q,b,P);CHKERRQ(ierr);
ierr = VecWAXPY(P,b,Q,U);CHKERRQ(ierr); /* p <- u + b(q + b p) */
ierr = KSP_PCApplyBAorAB(ksp,P,V,Q);CHKERRQ(ierr); /* v <- K p */
rhoold = rho;
i++;
} while (i<ksp->max_it);
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
ierr = KSPUnwindPreconditioner(ksp,X,T);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_Chebyshev(KSP ksp)
{
KSP_Chebyshev *cheb = (KSP_Chebyshev*)ksp->data;
PetscErrorCode ierr;
PetscInt k,kp1,km1,maxit,ktmp,i;
PetscScalar alpha,omegaprod,mu,omega,Gamma,c[3],scale;
PetscReal rnorm = 0.0;
Vec sol_orig,b,p[3],r;
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);
if (cheb->kspest && !cheb->estimate_current) {
PetscReal max=0.0,min=0.0;
Vec X,B;
X = ksp->work[0];
if (cheb->random) {
B = ksp->work[1];
ierr = VecSetRandom(B,cheb->random);CHKERRQ(ierr);
} else {
B = ksp->vec_rhs;
}
ierr = KSPSolve(cheb->kspest,B,X);CHKERRQ(ierr);
if (ksp->guess_zero) {
ierr = VecZeroEntries(X);CHKERRQ(ierr);
}
ierr = KSPChebyshevComputeExtremeEigenvalues_Private(cheb->kspest,&min,&max);CHKERRQ(ierr);
cheb->emin = cheb->tform[0]*min + cheb->tform[1]*max;
cheb->emax = cheb->tform[2]*min + cheb->tform[3]*max;
cheb->estimate_current = PETSC_TRUE;
}
ksp->its = 0;
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
maxit = ksp->max_it;
/* These three point to the three active solutions, we
rotate these three at each solution update */
km1 = 0; k = 1; kp1 = 2;
sol_orig = ksp->vec_sol; /* ksp->vec_sol will be asigned to rotating vector p[k], thus save its address */
b = ksp->vec_rhs;
p[km1] = sol_orig;
p[k] = ksp->work[0];
p[kp1] = ksp->work[1];
r = ksp->work[2];
/* use scale*B as our preconditioner */
scale = 2.0/(cheb->emax + cheb->emin);
/* -alpha <= scale*lambda(B^{-1}A) <= alpha */
alpha = 1.0 - scale*(cheb->emin);
Gamma = 1.0;
mu = 1.0/alpha;
omegaprod = 2.0/alpha;
c[km1] = 1.0;
c[k] = mu;
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,p[km1],r);CHKERRQ(ierr); /* r = b - A*p[km1] */
ierr = VecAYPX(r,-1.0,b);CHKERRQ(ierr);
} else {
ierr = VecCopy(b,r);CHKERRQ(ierr);
}
ierr = KSP_PCApply(ksp,r,p[k]);CHKERRQ(ierr); /* p[k] = scale B^{-1}r + p[km1] */
ierr = VecAYPX(p[k],scale,p[km1]);CHKERRQ(ierr);
for (i=0; i<maxit; i++) {
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
c[kp1] = 2.0*mu*c[k] - c[km1];
omega = omegaprod*c[k]/c[kp1];
ierr = KSP_MatMult(ksp,Amat,p[k],r);CHKERRQ(ierr); /* r = b - Ap[k] */
ierr = VecAYPX(r,-1.0,b);CHKERRQ(ierr);
ierr = KSP_PCApply(ksp,r,p[kp1]);CHKERRQ(ierr); /* p[kp1] = B^{-1}r */
ksp->vec_sol = p[k];
/* calculate residual norm if requested */
if (ksp->normtype != KSP_NORM_NONE || ksp->numbermonitors) {
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNorm(r,NORM_2,&rnorm);CHKERRQ(ierr);
} else {
ierr = VecNorm(p[kp1],NORM_2,&rnorm);CHKERRQ(ierr);
}
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->rnorm = rnorm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
}
/* y^{k+1} = omega(y^{k} - y^{k-1} + Gamma*r^{k}) + y^{k-1} */
ierr = VecAXPBYPCZ(p[kp1],1.0-omega,omega,omega*Gamma*scale,p[km1],p[k]);CHKERRQ(ierr);
ktmp = km1;
km1 = k;
k = kp1;
kp1 = ktmp;
}
if (!ksp->reason) {
if (ksp->normtype != KSP_NORM_NONE) {
ierr = KSP_MatMult(ksp,Amat,p[k],r);CHKERRQ(ierr); /* r = b - Ap[k] */
ierr = VecAYPX(r,-1.0,b);CHKERRQ(ierr);
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNorm(r,NORM_2,&rnorm);CHKERRQ(ierr);
} else {
ierr = KSP_PCApply(ksp,r,p[kp1]);CHKERRQ(ierr); /* p[kp1] = B^{-1}r */
ierr = VecNorm(p[kp1],NORM_2,&rnorm);CHKERRQ(ierr);
}
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->rnorm = rnorm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->vec_sol = p[k];
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i,rnorm);CHKERRQ(ierr);
}
if (ksp->its >= ksp->max_it) {
if (ksp->normtype != KSP_NORM_NONE) {
ierr = (*ksp->converged)(ksp,i,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (!ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
} else ksp->reason = KSP_CONVERGED_ITS;
}
}
/* make sure solution is in vector x */
ksp->vec_sol = sol_orig;
if (k) {
ierr = VecCopy(p[k],sol_orig);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_IBCGS(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,N;
PetscReal rnorm,rnormin = 0.0;
#if defined(PETSC_HAVE_MPI_LONG_DOUBLE) && !defined(PETSC_USE_COMPLEX) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
/* Because of possible instabilities in the algorithm (as indicated by different residual histories for the same problem
on the same number of processes with different runs) we support computing the inner products using Intel's 80 bit arithematic
rather than just 64 bit. Thus we copy our double precision values into long doubles (hoping this keeps the 16 extra bits)
and tell MPI to do its ALlreduces with MPI_LONG_DOUBLE.
Note for developers that does not effect the code. Intel's long double is implemented by storing the 80 bits of extended double
precision into a 16 byte space (the rest of the space is ignored) */
long double insums[7],outsums[7];
#else
PetscScalar insums[7],outsums[7];
#endif
PetscScalar sigman_2, sigman_1, sigman, pin_1, pin, phin_1, phin,tmp1,tmp2;
PetscScalar taun_1, taun, rhon, alphan_1, alphan, omegan_1, omegan;
const PetscScalar *PETSC_RESTRICT r0, *PETSC_RESTRICT f0, *PETSC_RESTRICT qn, *PETSC_RESTRICT b, *PETSC_RESTRICT un;
PetscScalar *PETSC_RESTRICT rn, *PETSC_RESTRICT xn, *PETSC_RESTRICT vn, *PETSC_RESTRICT zn;
/* the rest do not have to keep n_1 values */
PetscScalar kappan, thetan, etan, gamman, betan, deltan;
const PetscScalar *PETSC_RESTRICT tn;
PetscScalar *PETSC_RESTRICT sn;
Vec R0,Rn,Xn,F0,Vn,Zn,Qn,Tn,Sn,B,Un;
Mat A;
PetscFunctionBegin;
if (!ksp->vec_rhs->petscnative) SETERRQ(((PetscObject)ksp)->comm,PETSC_ERR_SUP,"Only coded for PETSc vectors");
ierr = PCGetOperators(ksp->pc,&A,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr);
ierr = VecGetLocalSize(ksp->vec_sol,&N);CHKERRQ(ierr);
Xn = ksp->vec_sol;ierr = VecGetArray(Xn_1,(PetscScalar**)&xn_1);CHKERRQ(ierr);ierr = VecRestoreArray(Xn_1,PETSC_NULL);CHKERRQ(ierr);
B = ksp->vec_rhs;ierr = VecGetArrayRead(B,(const PetscScalar**)&b);ierr = VecRestoreArrayRead(B,PETSC_NULL);CHKERRQ(ierr);
R0 = ksp->work[0];ierr = VecGetArrayRead(R0,(const PetscScalar**)&r0);CHKERRQ(ierr);ierr = VecRestoreArrayRead(R0,PETSC_NULL);CHKERRQ(ierr);
Rn = ksp->work[1];ierr = VecGetArray(Rn_1,(PetscScalar**)&rn_1);CHKERRQ(ierr);ierr = VecRestoreArray(Rn_1,PETSC_NULL);CHKERRQ(ierr);
Un = ksp->work[2];ierr = VecGetArrayRead(Un_1,(const PetscScalar**)&un_1);CHKERRQ(ierr);ierr = VecRestoreArrayRead(Un_1,PETSC_NULL);CHKERRQ(ierr);
F0 = ksp->work[3];ierr = VecGetArrayRead(F0,(const PetscScalar**)&f0);CHKERRQ(ierr);ierr = VecRestoreArrayRead(F0,PETSC_NULL);CHKERRQ(ierr);
Vn = ksp->work[4];ierr = VecGetArray(Vn_1,(PetscScalar**)&vn_1);CHKERRQ(ierr);ierr = VecRestoreArray(Vn_1,PETSC_NULL);CHKERRQ(ierr);
Zn = ksp->work[5];ierr = VecGetArray(Zn_1,(PetscScalar**)&zn_1);CHKERRQ(ierr);ierr = VecRestoreArray(Zn_1,PETSC_NULL);CHKERRQ(ierr);
Qn = ksp->work[6];ierr = VecGetArrayRead(Qn_1,(const PetscScalar**)&qn_1);CHKERRQ(ierr);ierr = VecRestoreArrayRead(Qn_1,PETSC_NULL);CHKERRQ(ierr);
Tn = ksp->work[7];ierr = VecGetArrayRead(Tn,(const PetscScalar**)&tn);CHKERRQ(ierr);ierr = VecRestoreArrayRead(Tn,PETSC_NULL);CHKERRQ(ierr);
Sn = ksp->work[8];ierr = VecGetArrayRead(Sn,(const PetscScalar**)&sn);CHKERRQ(ierr);ierr = VecRestoreArrayRead(Sn,PETSC_NULL);CHKERRQ(ierr);
/* r0 = rn_1 = b - A*xn_1; */
/* ierr = KSP_PCApplyBAorAB(ksp,Xn_1,Rn_1,Tn);CHKERRQ(ierr);
ierr = VecAYPX(Rn_1,-1.0,B);CHKERRQ(ierr); */
ierr = KSPInitialResidual(ksp,Xn_1,Tn,Sn,Rn_1,B);CHKERRQ(ierr);
ierr = VecNorm(Rn_1,NORM_2,&rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,0,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
ierr = VecCopy(Rn_1,R0);CHKERRQ(ierr);
/* un_1 = A*rn_1; */
ierr = KSP_PCApplyBAorAB(ksp,Rn_1,Un_1,Tn);CHKERRQ(ierr);
/* f0 = A'*rn_1; */
if (ksp->pc_side == PC_RIGHT) { /* B' A' */
ierr = MatMultTranspose(A,R0,Tn);CHKERRQ(ierr);
ierr = PCApplyTranspose(ksp->pc,Tn,F0);CHKERRQ(ierr);
} else if (ksp->pc_side == PC_LEFT) { /* A' B' */
ierr = PCApplyTranspose(ksp->pc,R0,Tn);CHKERRQ(ierr);
ierr = MatMultTranspose(A,Tn,F0);CHKERRQ(ierr);
}
/*qn_1 = vn_1 = zn_1 = 0.0; */
ierr = VecSet(Qn_1,0.0);CHKERRQ(ierr);
ierr = VecSet(Vn_1,0.0);CHKERRQ(ierr);
ierr = VecSet(Zn_1,0.0);CHKERRQ(ierr);
sigman_2 = pin_1 = taun_1 = 0.0;
/* the paper says phin_1 should be initialized to zero, it is actually R0'R0 */
ierr = VecDot(R0,R0,&phin_1);CHKERRQ(ierr);
/* sigman_1 = rn_1'un_1 */
ierr = VecDot(R0,Un_1,&sigman_1);CHKERRQ(ierr);
alphan_1 = omegan_1 = 1.0;
for (ksp->its = 1; ksp->its<ksp->max_it+1; ksp->its++) {
rhon = phin_1 - omegan_1*sigman_2 + omegan_1*alphan_1*pin_1;
/* if (rhon == 0.0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"rhon is zero, iteration %D",n); */
if (ksp->its == 1) deltan = rhon;
else deltan = rhon/taun_1;
betan = deltan/omegan_1;
taun = sigman_1 + betan*taun_1 - deltan*pin_1;
if (taun == 0.0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"taun is zero, iteration %D",ksp->its);
alphan = rhon/taun;
ierr = PetscLogFlops(15.0);
/*
zn = alphan*rn_1 + (alphan/alphan_1)betan*zn_1 - alphan*deltan*vn_1
vn = un_1 + betan*vn_1 - deltan*qn_1
sn = rn_1 - alphan*vn
The algorithm in the paper is missing the alphan/alphan_1 term in the zn update
*/
ierr = PetscLogEventBegin(VEC_Ops,0,0,0,0);CHKERRQ(ierr);
tmp1 = (alphan/alphan_1)*betan;
tmp2 = alphan*deltan;
for (i=0; i<N; i++) {
zn[i] = alphan*rn_1[i] + tmp1*zn_1[i] - tmp2*vn_1[i];
vn[i] = un_1[i] + betan*vn_1[i] - deltan*qn_1[i];
sn[i] = rn_1[i] - alphan*vn[i];
}
ierr = PetscLogFlops(3.0+11.0*N);
ierr = PetscLogEventEnd(VEC_Ops,0,0,0,0);CHKERRQ(ierr);
/*
qn = A*vn
*/
ierr = KSP_PCApplyBAorAB(ksp,Vn,Qn,Tn);CHKERRQ(ierr);
/*
tn = un_1 - alphan*qn
*/
ierr = VecWAXPY(Tn,-alphan,Qn,Un_1);CHKERRQ(ierr);
/*
phin = r0'sn
pin = r0'qn
gamman = f0'sn
etan = f0'tn
thetan = sn'tn
kappan = tn'tn
*/
ierr = PetscLogEventBegin(VEC_ReduceArithmetic,0,0,0,0);CHKERRQ(ierr);
phin = pin = gamman = etan = thetan = kappan = 0.0;
for (i=0; i<N; i++) {
phin += r0[i]*sn[i];
pin += r0[i]*qn[i];
gamman += f0[i]*sn[i];
etan += f0[i]*tn[i];
thetan += sn[i]*tn[i];
kappan += tn[i]*tn[i];
}
ierr = PetscLogFlops(12.0*N);
ierr = PetscLogEventEnd(VEC_ReduceArithmetic,0,0,0,0);CHKERRQ(ierr);
insums[0] = phin;
insums[1] = pin;
insums[2] = gamman;
insums[3] = etan;
insums[4] = thetan;
insums[5] = kappan;
insums[6] = rnormin;
ierr = PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr);
#if defined(PETSC_HAVE_MPI_LONG_DOUBLE) && !defined(PETSC_USE_COMPLEX) && (defined(PETSC_USE_REAL_SINGLE) || defined(PETSC_USE_REAL_DOUBLE))
if (ksp->lagnorm && ksp->its > 1) {
ierr = MPI_Allreduce(insums,outsums,7,MPI_LONG_DOUBLE,MPI_SUM,((PetscObject)ksp)->comm);CHKERRQ(ierr);
} else {
ierr = MPI_Allreduce(insums,outsums,6,MPI_LONG_DOUBLE,MPI_SUM,((PetscObject)ksp)->comm);CHKERRQ(ierr);
}
#else
if (ksp->lagnorm && ksp->its > 1) {
ierr = MPI_Allreduce(insums,outsums,7,MPIU_SCALAR,MPIU_SUM,((PetscObject)ksp)->comm);CHKERRQ(ierr);
} else {
ierr = MPI_Allreduce(insums,outsums,6,MPIU_SCALAR,MPIU_SUM,((PetscObject)ksp)->comm);CHKERRQ(ierr);
}
#endif
ierr = PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr);
phin = outsums[0];
pin = outsums[1];
gamman = outsums[2];
etan = outsums[3];
thetan = outsums[4];
kappan = outsums[5];
if (ksp->lagnorm && ksp->its > 1) rnorm = PetscSqrtReal(PetscRealPart(outsums[6]));
if (kappan == 0.0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"kappan is zero, iteration %D",ksp->its);
if (thetan == 0.0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"thetan is zero, iteration %D",ksp->its);
omegan = thetan/kappan;
sigman = gamman - omegan*etan;
/*
rn = sn - omegan*tn
xn = xn_1 + zn + omegan*sn
*/
ierr = PetscLogEventBegin(VEC_Ops,0,0,0,0);CHKERRQ(ierr);
rnormin = 0.0;
for (i=0; i<N; i++) {
rn[i] = sn[i] - omegan*tn[i];
rnormin += PetscRealPart(PetscConj(rn[i])*rn[i]);
xn[i] += zn[i] + omegan*sn[i];
}
ierr = PetscObjectStateIncrease((PetscObject)Xn);CHKERRQ(ierr);
ierr = PetscLogFlops(7.0*N);
ierr = PetscLogEventEnd(VEC_Ops,0,0,0,0);CHKERRQ(ierr);
if (!ksp->lagnorm && ksp->chknorm < ksp->its) {
ierr = PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr);
ierr = MPI_Allreduce(&rnormin,&rnorm,1,MPIU_REAL,MPIU_SUM,((PetscObject)ksp)->comm);CHKERRQ(ierr);
ierr = PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr);
rnorm = PetscSqrtReal(rnorm);
}
/* Test for convergence */
ierr = KSPMonitor(ksp,ksp->its,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,ksp->its,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
/* un = A*rn */
ierr = KSP_PCApplyBAorAB(ksp,Rn,Un,Tn);CHKERRQ(ierr);
/* Update n-1 locations with n locations */
sigman_2 = sigman_1;
sigman_1 = sigman;
pin_1 = pin;
phin_1 = phin;
alphan_1 = alphan;
taun_1 = taun;
omegan_1 = omegan;
}
if (ksp->its >= ksp->max_it) {
ksp->reason = KSP_DIVERGED_ITS;
}
ierr = KSPUnwindPreconditioner(ksp,Xn,Tn);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_BiCG(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i;
PetscBool diagonalscale;
PetscScalar dpi,a=1.0,beta,betaold=1.0,b,ma;
PetscReal dp;
Vec X,B,Zl,Zr,Rl,Rr,Pl,Pr;
Mat Amat,Pmat;
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;
Rl = ksp->work[0];
Zl = ksp->work[1];
Pl = ksp->work[2];
Rr = ksp->work[3];
Zr = ksp->work[4];
Pr = ksp->work[5];
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,Rr);CHKERRQ(ierr); /* r <- b - Ax */
ierr = VecAYPX(Rr,-1.0,B);CHKERRQ(ierr);
} else {
ierr = VecCopy(B,Rr);CHKERRQ(ierr); /* r <- b (x is 0) */
}
ierr = VecCopy(Rr,Rl);CHKERRQ(ierr);
ierr = KSP_PCApply(ksp,Rr,Zr);CHKERRQ(ierr); /* z <- Br */
ierr = VecConjugate(Rl);CHKERRQ(ierr);
ierr = KSP_PCApplyTranspose(ksp,Rl,Zl);CHKERRQ(ierr);
ierr = VecConjugate(Rl);CHKERRQ(ierr);
ierr = VecConjugate(Zl);CHKERRQ(ierr);
if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNorm(Zr,NORM_2,&dp);CHKERRQ(ierr); /* dp <- z'*z */
} else {
ierr = VecNorm(Rr,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r */
}
ierr = KSPMonitor(ksp,0,dp);CHKERRQ(ierr);
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = dp;
ierr = PetscObjectSAWsGrantAccess((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 = VecDot(Zr,Rl,&beta);CHKERRQ(ierr); /* beta <- r'z */
if (!i) {
if (beta == 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN_BICG;
PetscFunctionReturn(0);
}
ierr = VecCopy(Zr,Pr);CHKERRQ(ierr); /* p <- z */
ierr = VecCopy(Zl,Pl);CHKERRQ(ierr);
} else {
b = beta/betaold;
ierr = VecAYPX(Pr,b,Zr);CHKERRQ(ierr); /* p <- z + b* p */
b = PetscConj(b);
ierr = VecAYPX(Pl,b,Zl);CHKERRQ(ierr);
}
betaold = beta;
ierr = KSP_MatMult(ksp,Amat,Pr,Zr);CHKERRQ(ierr); /* z <- Kp */
ierr = VecConjugate(Pl);CHKERRQ(ierr);
ierr = KSP_MatMultTranspose(ksp,Amat,Pl,Zl);CHKERRQ(ierr);
ierr = VecConjugate(Pl);CHKERRQ(ierr);
ierr = VecConjugate(Zl);CHKERRQ(ierr);
ierr = VecDot(Zr,Pl,&dpi);CHKERRQ(ierr); /* dpi <- z'p */
a = beta/dpi; /* a = beta/p'z */
ierr = VecAXPY(X,a,Pr);CHKERRQ(ierr); /* x <- x + ap */
ma = -a;
ierr = VecAXPY(Rr,ma,Zr);CHKERRQ(ierr);
ma = PetscConj(ma);
ierr = VecAXPY(Rl,ma,Zl);CHKERRQ(ierr);
if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = KSP_PCApply(ksp,Rr,Zr);CHKERRQ(ierr); /* z <- Br */
ierr = VecConjugate(Rl);CHKERRQ(ierr);
ierr = KSP_PCApplyTranspose(ksp,Rl,Zl);CHKERRQ(ierr);
ierr = VecConjugate(Rl);CHKERRQ(ierr);
ierr = VecConjugate(Zl);CHKERRQ(ierr);
ierr = VecNorm(Zr,NORM_2,&dp);CHKERRQ(ierr); /* dp <- z'*z */
} else {
ierr = VecNorm(Rr,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r */
}
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = i+1;
ksp->rnorm = dp;
ierr = PetscObjectSAWsGrantAccess((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;
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = KSP_PCApply(ksp,Rr,Zr);CHKERRQ(ierr); /* z <- Br */
ierr = VecConjugate(Rl);CHKERRQ(ierr);
ierr = KSP_PCApplyTranspose(ksp,Rl,Zl);CHKERRQ(ierr);
ierr = VecConjugate(Rl);CHKERRQ(ierr);
ierr = VecConjugate(Zl);CHKERRQ(ierr);
}
i++;
} while (i<ksp->max_it);
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPPIPEFGMRESCycle(PetscInt *itcount,KSP ksp)
{
KSP_PIPEFGMRES *pipefgmres = (KSP_PIPEFGMRES*)(ksp->data);
PetscReal res_norm;
PetscReal hapbnd,tt;
PetscScalar *hh,*hes,*lhh,shift = pipefgmres->shift;
PetscBool hapend = PETSC_FALSE; /* indicates happy breakdown ending */
PetscErrorCode ierr;
PetscInt loc_it; /* local count of # of dir. in Krylov space */
PetscInt max_k = pipefgmres->max_k; /* max # of directions Krylov space */
PetscInt i,j,k;
Mat Amat,Pmat;
Vec Q,W; /* Pipelining vectors */
Vec *redux = pipefgmres->redux; /* workspace for single reduction */
PetscFunctionBegin;
if (itcount) *itcount = 0;
/* Assign simpler names to these vectors, allocated as pipelining workspace */
Q = VEC_Q;
W = VEC_W;
/* Allocate memory for orthogonalization work (freed in the GMRES Destroy routine)*/
/* Note that we add an extra value here to allow for a single reduction */
if (!pipefgmres->orthogwork) { ierr = PetscMalloc1(pipefgmres->max_k + 2 ,&pipefgmres->orthogwork);CHKERRQ(ierr);
}
lhh = pipefgmres->orthogwork;
/* Number of pseudo iterations since last restart is the number
of prestart directions */
loc_it = 0;
/* note: (pipefgmres->it) is always set one less than (loc_it) It is used in
KSPBUILDSolution_PIPEFGMRES, where it is passed to KSPPIPEFGMRESBuildSoln.
Note that when KSPPIPEFGMRESBuildSoln is called from this function,
(loc_it -1) is passed, so the two are equivalent */
pipefgmres->it = (loc_it - 1);
/* initial residual is in VEC_VV(0) - compute its norm*/
ierr = VecNorm(VEC_VV(0),NORM_2,&res_norm);CHKERRQ(ierr);
/* first entry in right-hand-side of hessenberg system is just
the initial residual norm */
*RS(0) = res_norm;
ksp->rnorm = res_norm;
ierr = KSPLogResidualHistory(ksp,res_norm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res_norm);CHKERRQ(ierr);
/* check for the convergence - maybe the current guess is good enough */
ierr = (*ksp->converged)(ksp,ksp->its,res_norm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) {
if (itcount) *itcount = 0;
PetscFunctionReturn(0);
}
/* scale VEC_VV (the initial residual) */
ierr = VecScale(VEC_VV(0),1.0/res_norm);CHKERRQ(ierr);
/* Fill the pipeline */
ierr = KSP_PCApply(ksp,VEC_VV(loc_it),PREVEC(loc_it));CHKERRQ(ierr);
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,PREVEC(loc_it),ZVEC(loc_it));CHKERRQ(ierr);
ierr = VecAXPY(ZVEC(loc_it),-shift,VEC_VV(loc_it));CHKERRQ(ierr); /* Note shift */
/* MAIN ITERATION LOOP BEGINNING*/
/* keep iterating until we have converged OR generated the max number
of directions OR reached the max number of iterations for the method */
while (!ksp->reason && loc_it < max_k && ksp->its < ksp->max_it) {
if (loc_it) {
ierr = KSPLogResidualHistory(ksp,res_norm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res_norm);CHKERRQ(ierr);
}
pipefgmres->it = (loc_it - 1);
/* see if more space is needed for work vectors */
if (pipefgmres->vv_allocated <= loc_it + VEC_OFFSET + 1) {
ierr = KSPPIPEFGMRESGetNewVectors(ksp,loc_it+1);CHKERRQ(ierr);
/* (loc_it+1) is passed in as number of the first vector that should
be allocated */
}
/* Note that these inner products are with "Z" now, so
in particular, lhh[loc_it] is the 'barred' or 'shifted' value,
not the value from the equivalent FGMRES run (even in exact arithmetic)
That is, the H we need for the Arnoldi relation is different from the
coefficients we use in the orthogonalization process,because of the shift */
/* Do some local twiddling to allow for a single reduction */
for(i=0;i<loc_it+1;i++){
redux[i] = VEC_VV(i);
}
redux[loc_it+1] = ZVEC(loc_it);
/* note the extra dot product which ends up in lh[loc_it+1], which computes ||z||^2 */
ierr = VecMDotBegin(ZVEC(loc_it),loc_it+2,redux,lhh);CHKERRQ(ierr);
/* Start the split reduction (This actually calls the MPI_Iallreduce, otherwise, the reduction is simply delayed until the "end" call)*/
ierr = PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)ZVEC(loc_it)));CHKERRQ(ierr);
/* The work to be overlapped with the inner products follows.
This is application of the preconditioner and the operator
to compute intermediate quantites which will be combined (locally)
with the results of the inner products.
*/
ierr = KSP_PCApply(ksp,ZVEC(loc_it),Q);CHKERRQ(ierr);
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
ierr = KSP_MatMult(ksp,Amat,Q,W);CHKERRQ(ierr);
/* Compute inner products of the new direction with previous directions,
and the norm of the to-be-orthogonalized direction "Z".
This information is enough to build the required entries
of H. The inner product with VEC_VV(it_loc) is
*different* than in the standard FGMRES and need to be dealt with specially.
That is, for standard FGMRES the orthogonalization coefficients are the same
as the coefficients used in the Arnoldi relation to reconstruct, but here this
is not true (albeit only for the one entry of H which we "unshift" below. */
/* Finish the dot product, retrieving the extra entry */
ierr = VecMDotEnd(ZVEC(loc_it),loc_it+2,redux,lhh);CHKERRQ(ierr);
tt = PetscRealPart(lhh[loc_it+1]);
/* Hessenberg entries, and entries for (naive) classical Graham-Schmidt
Note that the Hessenberg entries require a shift, as these are for the
relation AU = VH, which is wrt unshifted basis vectors */
hh = HH(0,loc_it); hes=HES(0,loc_it);
for (j=0; j<loc_it; j++) {
hh[j] = lhh[j];
hes[j] = lhh[j];
}
hh[loc_it] = lhh[loc_it] + shift;
hes[loc_it] = lhh[loc_it] + shift;
/* we delay applying the shift here */
for (j=0; j<=loc_it; j++) {
lhh[j] = -lhh[j]; /* flip sign */
}
/* Compute the norm of the un-normalized new direction using the rearranged formula
Note that these are shifted ("barred") quantities */
for(k=0;k<=loc_it;k++) tt -= ((PetscReal)(PetscAbsScalar(lhh[k]) * PetscAbsScalar(lhh[k])));
/* On AVX512 this is accumulating roundoff errors for eg: tt=-2.22045e-16 */
if ((tt < 0.0) && tt > -PETSC_SMALL) tt = 0.0 ;
if (tt < 0.0) {
/* If we detect square root breakdown in the norm, we must restart the algorithm.
Here this means we simply break the current loop and reconstruct the solution
using the basis we have computed thus far. Note that by breaking immediately,
we do not update the iteration count, so computation done in this iteration
should be disregarded.
*/
ierr = PetscInfo2(ksp,"Restart due to square root breakdown at it = %D, tt=%g\n",ksp->its,(double)tt);CHKERRQ(ierr);
break;
} else {
tt = PetscSqrtReal(tt);
}
/* new entry in hessenburg is the 2-norm of our new direction */
hh[loc_it+1] = tt;
hes[loc_it+1] = tt;
/* The recurred computation for the new direction
The division by tt is delayed to the happy breakdown check later
Note placement BEFORE the unshift
*/
ierr = VecCopy(ZVEC(loc_it),VEC_VV(loc_it+1));CHKERRQ(ierr);
ierr = VecMAXPY(VEC_VV(loc_it+1),loc_it+1,lhh,&VEC_VV(0));CHKERRQ(ierr);
/* (VEC_VV(loc_it+1) is not normalized yet) */
/* The recurred computation for the preconditioned vector (u) */
ierr = VecCopy(Q,PREVEC(loc_it+1));CHKERRQ(ierr);
ierr = VecMAXPY(PREVEC(loc_it+1),loc_it+1,lhh,&PREVEC(0));CHKERRQ(ierr);
ierr = VecScale(PREVEC(loc_it+1),1.0/tt);CHKERRQ(ierr);
/* Unshift an entry in the GS coefficients ("removing the bar") */
lhh[loc_it] -= shift;
/* The recurred computation for z (Au)
Note placement AFTER the "unshift" */
ierr = VecCopy(W,ZVEC(loc_it+1));CHKERRQ(ierr);
ierr = VecMAXPY(ZVEC(loc_it+1),loc_it+1,lhh,&ZVEC(0));CHKERRQ(ierr);
ierr = VecScale(ZVEC(loc_it+1),1.0/tt);CHKERRQ(ierr);
/* Happy Breakdown Check */
hapbnd = PetscAbsScalar((tt) / *RS(loc_it));
/* RS(loc_it) contains the res_norm from the last iteration */
hapbnd = PetscMin(pipefgmres->haptol,hapbnd);
if (tt > hapbnd) {
/* scale new direction by its norm */
ierr = VecScale(VEC_VV(loc_it+1),1.0/tt);CHKERRQ(ierr);
} else {
/* This happens when the solution is exactly reached. */
/* So there is no new direction... */
ierr = VecSet(VEC_TEMP,0.0);CHKERRQ(ierr); /* set VEC_TEMP to 0 */
hapend = PETSC_TRUE;
}
/* note that for pipefgmres we could get HES(loc_it+1, loc_it) = 0 and the
current solution would not be exact if HES was singular. Note that
HH non-singular implies that HES is not singular, and HES is guaranteed
to be nonsingular when PREVECS are linearly independent and A is
nonsingular (in GMRES, the nonsingularity of A implies the nonsingularity
of HES). So we should really add a check to verify that HES is nonsingular.*/
/* Note that to be thorough, in debug mode, one could call a LAPACK routine
here to check that the hessenberg matrix is indeed non-singular (since
FGMRES does not guarantee this) */
/* Now apply rotations to new col of hessenberg (and right side of system),
calculate new rotation, and get new residual norm at the same time*/
ierr = KSPPIPEFGMRESUpdateHessenberg(ksp,loc_it,&hapend,&res_norm);CHKERRQ(ierr);
if (ksp->reason) break;
loc_it++;
pipefgmres->it = (loc_it-1); /* Add this here in case it has converged */
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = res_norm;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,ksp->its,res_norm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
/* Catch error in happy breakdown and signal convergence and break from loop */
if (hapend) {
if (!ksp->reason) {
if (ksp->errorifnotconverged) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"You reached the happy break down, but convergence was not indicated. Residual norm = %g",(double)res_norm);
else {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
}
}
}
/* END OF ITERATION LOOP */
ierr = KSPLogResidualHistory(ksp,res_norm);CHKERRQ(ierr);
/*
Monitor if we know that we will not return for a restart */
if (loc_it && (ksp->reason || ksp->its >= ksp->max_it)) {
ierr = KSPMonitor(ksp,ksp->its,res_norm);CHKERRQ(ierr);
}
if (itcount) *itcount = loc_it;
/*
Down here we have to solve for the "best" coefficients of the Krylov
columns, add the solution values together, and possibly unwind the
preconditioning from the solution
*/
/* Form the solution (or the solution so far) */
/* Note: must pass in (loc_it-1) for iteration count so that KSPPIPEGMRESIIBuildSoln
properly navigates */
ierr = KSPPIPEFGMRESBuildSoln(RS(0),ksp->vec_sol,ksp->vec_sol,ksp,loc_it-1);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_TSIRM(KSP ksp)
{
PetscErrorCode ierr;
KSP_TSIRM *tsirm = (KSP_TSIRM*)ksp->data;
KSP sub_ksp;
PC pc;
Mat AS;
Vec x,b;
PetscScalar *array;
PetscReal norm = 20;
PetscInt i,*ind_row,first_iteration = 1,its = 0,total = 0,col = 0;
PetscInt restart = 30;
KSP ksp_min; /* KSP for minimization */
PC pc_min; /* PC for minimization */
PetscFunctionBegin;
x = ksp->vec_sol; /* Solution vector */
b = ksp->vec_rhs; /* Right-hand side vector */
/* Row indexes (these indexes are global) */
ierr = PetscMalloc1(tsirm->Iend-tsirm->Istart,&ind_row);
CHKERRQ(ierr);
for (i=0; i<tsirm->Iend-tsirm->Istart; i++) ind_row[i] = i+tsirm->Istart;
/* Inner solver */
ierr = KSPGetPC(ksp,&pc);
CHKERRQ(ierr);
ierr = PCKSPGetKSP(pc,&sub_ksp);
CHKERRQ(ierr);
ierr = KSPSetTolerances(sub_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,restart);
CHKERRQ(ierr);
/* previously it seemed good but with SNES it seems not good... */
ierr = KSP_MatMult(sub_ksp,tsirm->A,x,tsirm->r);
CHKERRQ(ierr);
ierr = VecAXPY(tsirm->r,-1,b);
CHKERRQ(ierr);
ierr = VecNorm(tsirm->r,NORM_2,&norm);
CHKERRQ(ierr);
ksp->its = 0;
ierr = KSPConvergedDefault(ksp,ksp->its,norm,&ksp->reason,ksp->cnvP);
CHKERRQ(ierr);
ierr = KSPSetInitialGuessNonzero(sub_ksp,PETSC_TRUE);
CHKERRQ(ierr);
do {
for (col=0; col<tsirm->size_ls && ksp->reason==0; col++) {
/* Solve (inner iteration) */
ierr = KSPSolve(sub_ksp,b,x);
CHKERRQ(ierr);
ierr = KSPGetIterationNumber(sub_ksp,&its);
CHKERRQ(ierr);
total += its;
/* Build S^T */
ierr = VecGetArray(x,&array);
CHKERRQ(ierr);
ierr = MatSetValues(tsirm->S,tsirm->Iend-tsirm->Istart,ind_row,1,&col,array,INSERT_VALUES);
CHKERRQ(ierr);
ierr = VecRestoreArray(x,&array);
CHKERRQ(ierr);
ierr = KSPGetResidualNorm(sub_ksp,&norm);
CHKERRQ(ierr);
ksp->rnorm = norm;
ksp->its ++;
ierr = KSPConvergedDefault(ksp,ksp->its,norm,&ksp->reason,ksp->cnvP);
CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,norm);
CHKERRQ(ierr);
}
/* Minimization step */
if (!ksp->reason) {
ierr = MatAssemblyBegin(tsirm->S,MAT_FINAL_ASSEMBLY);
CHKERRQ(ierr);
ierr = MatAssemblyEnd(tsirm->S,MAT_FINAL_ASSEMBLY);
CHKERRQ(ierr);
if (first_iteration) {
ierr = MatMatMult(tsirm->A,tsirm->S,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&AS);
CHKERRQ(ierr);
first_iteration = 0;
} else {
ierr = MatMatMult(tsirm->A,tsirm->S,MAT_REUSE_MATRIX,PETSC_DEFAULT,&AS);
CHKERRQ(ierr);
}
/* CGLS or LSQR method to minimize the residuals*/
ierr = KSPCreate(PETSC_COMM_WORLD,&ksp_min);
CHKERRQ(ierr);
if (tsirm->cgls) {
ierr = KSPSetType(ksp_min,KSPCGLS);
CHKERRQ(ierr);
} else {
ierr = KSPSetType(ksp_min,KSPLSQR);
CHKERRQ(ierr);
}
ierr = KSPSetOperators(ksp_min,AS,AS);
CHKERRQ(ierr);
ierr = KSPSetTolerances(ksp_min,tsirm->tol_ls,PETSC_DEFAULT,PETSC_DEFAULT,tsirm->maxiter_ls);
CHKERRQ(ierr);
ierr = KSPGetPC(ksp_min,&pc_min);
CHKERRQ(ierr);
ierr = PCSetType(pc_min,PCNONE);
CHKERRQ(ierr);
ierr = KSPSolve(ksp_min,b,tsirm->Alpha);
CHKERRQ(ierr); /* Find Alpha such that ||AS Alpha = b|| */
ierr = KSPDestroy(&ksp_min);
CHKERRQ(ierr);
/* Apply minimization */
ierr = MatMult(tsirm->S,tsirm->Alpha,x);
CHKERRQ(ierr); /* x = S * Alpha */
}
} while (ksp->its<ksp->max_it && !ksp->reason);
ierr = MatDestroy(&AS);
CHKERRQ(ierr);
ierr = PetscFree(ind_row);
CHKERRQ(ierr);
ksp->its = total;
PetscFunctionReturn(0);
}
PetscErrorCode KSPLGMRESCycle(PetscInt *itcount,KSP ksp)
{
KSP_LGMRES *lgmres = (KSP_LGMRES*)(ksp->data);
PetscReal res_norm, res;
PetscReal hapbnd, tt;
PetscScalar tmp;
PetscBool hapend = PETSC_FALSE; /* indicates happy breakdown ending */
PetscErrorCode ierr;
PetscInt loc_it; /* local count of # of dir. in Krylov space */
PetscInt max_k = lgmres->max_k; /* max approx space size */
PetscInt max_it = ksp->max_it; /* max # of overall iterations for the method */
/* LGMRES_MOD - new variables*/
PetscInt aug_dim = lgmres->aug_dim;
PetscInt spot = 0;
PetscInt order = 0;
PetscInt it_arnoldi; /* number of arnoldi steps to take */
PetscInt it_total; /* total number of its to take (=approx space size)*/
PetscInt ii, jj;
PetscReal tmp_norm;
PetscScalar inv_tmp_norm;
PetscScalar *avec;
PetscFunctionBegin;
/* Number of pseudo iterations since last restart is the number
of prestart directions */
loc_it = 0;
/* LGMRES_MOD: determine number of arnoldi steps to take */
/* if approx_constant then we keep the space the same size even if
we don't have the full number of aug vectors yet*/
if (lgmres->approx_constant) it_arnoldi = max_k - lgmres->aug_ct;
else it_arnoldi = max_k - aug_dim;
it_total = it_arnoldi + lgmres->aug_ct;
/* initial residual is in VEC_VV(0) - compute its norm*/
ierr = VecNorm(VEC_VV(0),NORM_2,&res_norm);CHKERRQ(ierr);
res = res_norm;
/* first entry in right-hand-side of hessenberg system is just
the initial residual norm */
*GRS(0) = res_norm;
/* check for the convergence */
if (!res) {
if (itcount) *itcount = 0;
ksp->reason = KSP_CONVERGED_ATOL;
ierr = PetscInfo(ksp,"Converged due to zero residual norm on entry\n");CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* scale VEC_VV (the initial residual) */
tmp = 1.0/res_norm; ierr = VecScale(VEC_VV(0),tmp);CHKERRQ(ierr);
ksp->rnorm = res;
/* note: (lgmres->it) is always set one less than (loc_it) It is used in
KSPBUILDSolution_LGMRES, where it is passed to KSPLGMRESBuildSoln.
Note that when KSPLGMRESBuildSoln is called from this function,
(loc_it -1) is passed, so the two are equivalent */
lgmres->it = (loc_it - 1);
/* MAIN ITERATION LOOP BEGINNING*/
/* keep iterating until we have converged OR generated the max number
of directions OR reached the max number of iterations for the method */
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
while (!ksp->reason && loc_it < it_total && ksp->its < max_it) { /* LGMRES_MOD: changed to it_total */
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
lgmres->it = (loc_it - 1);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
/* see if more space is needed for work vectors */
if (lgmres->vv_allocated <= loc_it + VEC_OFFSET + 1) {
ierr = KSPLGMRESGetNewVectors(ksp,loc_it+1);CHKERRQ(ierr);
/* (loc_it+1) is passed in as number of the first vector that should
be allocated */
}
/*LGMRES_MOD: decide whether this is an arnoldi step or an aug step */
if (loc_it < it_arnoldi) { /* Arnoldi */
ierr = KSP_PCApplyBAorAB(ksp,VEC_VV(loc_it),VEC_VV(1+loc_it),VEC_TEMP_MATOP);CHKERRQ(ierr);
} else { /*aug step */
order = loc_it - it_arnoldi + 1; /* which aug step */
for (ii=0; ii<aug_dim; ii++) {
if (lgmres->aug_order[ii] == order) {
spot = ii;
break; /* must have this because there will be duplicates before aug_ct = aug_dim */
}
}
ierr = VecCopy(A_AUGVEC(spot), VEC_VV(1+loc_it));CHKERRQ(ierr);
/*note: an alternate implementation choice would be to only save the AUGVECS and
not A_AUGVEC and then apply the PC here to the augvec */
}
/* update hessenberg matrix and do Gram-Schmidt - new direction is in
VEC_VV(1+loc_it)*/
ierr = (*lgmres->orthog)(ksp,loc_it);CHKERRQ(ierr);
/* new entry in hessenburg is the 2-norm of our new direction */
ierr = VecNorm(VEC_VV(loc_it+1),NORM_2,&tt);CHKERRQ(ierr);
*HH(loc_it+1,loc_it) = tt;
*HES(loc_it+1,loc_it) = tt;
/* check for the happy breakdown */
hapbnd = PetscAbsScalar(tt / *GRS(loc_it)); /* GRS(loc_it) contains the res_norm from the last iteration */
if (hapbnd > lgmres->haptol) hapbnd = lgmres->haptol;
if (tt > hapbnd) {
tmp = 1.0/tt;
ierr = VecScale(VEC_VV(loc_it+1),tmp);CHKERRQ(ierr); /* scale new direction by its norm */
} else {
ierr = PetscInfo2(ksp,"Detected happy breakdown, current hapbnd = %G tt = %G\n",hapbnd,tt);CHKERRQ(ierr);
hapend = PETSC_TRUE;
}
/* Now apply rotations to new col of hessenberg (and right side of system),
calculate new rotation, and get new residual norm at the same time*/
ierr = KSPLGMRESUpdateHessenberg(ksp,loc_it,hapend,&res);CHKERRQ(ierr);
if (ksp->reason) break;
loc_it++;
lgmres->it = (loc_it-1); /* Add this here in case it has converged */
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = res;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
/* Catch error in happy breakdown and signal convergence and break from loop */
if (hapend) {
if (!ksp->reason) {
if (ksp->errorifnotconverged) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"You reached the happy break down, but convergence was not indicated. Residual norm = %G",res);
else {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
}
}
}
/* END OF ITERATION LOOP */
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
/* Monitor if we know that we will not return for a restart */
if (ksp->reason || ksp->its >= max_it) {
ierr = KSPMonitor(ksp, ksp->its, res);CHKERRQ(ierr);
}
if (itcount) *itcount = loc_it;
/*
Down here we have to solve for the "best" coefficients of the Krylov
columns, add the solution values together, and possibly unwind the
preconditioning from the solution
*/
/* Form the solution (or the solution so far) */
/* Note: must pass in (loc_it-1) for iteration count so that KSPLGMRESBuildSoln
properly navigates */
ierr = KSPLGMRESBuildSoln(GRS(0),ksp->vec_sol,ksp->vec_sol,ksp,loc_it-1);CHKERRQ(ierr);
/* LGMRES_MOD collect aug vector and A*augvector for future restarts -
only if we will be restarting (i.e. this cycle performed it_total
iterations) */
if (!ksp->reason && ksp->its < max_it && aug_dim > 0) {
/*AUG_TEMP contains the new augmentation vector (assigned in KSPLGMRESBuildSoln) */
if (!lgmres->aug_ct) {
spot = 0;
lgmres->aug_ct++;
} else if (lgmres->aug_ct < aug_dim) {
spot = lgmres->aug_ct;
lgmres->aug_ct++;
} else { /* truncate */
for (ii=0; ii<aug_dim; ii++) {
if (lgmres->aug_order[ii] == aug_dim) spot = ii;
}
}
ierr = VecCopy(AUG_TEMP, AUGVEC(spot));CHKERRQ(ierr);
/*need to normalize */
ierr = VecNorm(AUGVEC(spot), NORM_2, &tmp_norm);CHKERRQ(ierr);
inv_tmp_norm = 1.0/tmp_norm;
ierr = VecScale(AUGVEC(spot),inv_tmp_norm);CHKERRQ(ierr);
/*set new aug vector to order 1 - move all others back one */
for (ii=0; ii < aug_dim; ii++) AUG_ORDER(ii)++;
AUG_ORDER(spot) = 1;
/*now add the A*aug vector to A_AUGVEC(spot) - this is independ. of preconditioning type*/
/* want V*H*y - y is in GRS, V is in VEC_VV and H is in HES */
/* first do H+*y */
avec = lgmres->hwork;
ierr = PetscMemzero(avec,(it_total+1)*sizeof(*avec));CHKERRQ(ierr);
for (ii=0; ii < it_total + 1; ii++) {
for (jj=0; jj <= ii+1 && jj < it_total+1; jj++) {
avec[jj] += *HES(jj ,ii) * *GRS(ii);
}
}
/*now multiply result by V+ */
ierr = VecSet(VEC_TEMP,0.0);CHKERRQ(ierr);
ierr = VecMAXPY(VEC_TEMP, it_total+1, avec, &VEC_VV(0));CHKERRQ(ierr); /*answer is in VEC_TEMP*/
/*copy answer to aug location and scale*/
ierr = VecCopy(VEC_TEMP, A_AUGVEC(spot));CHKERRQ(ierr);
ierr = VecScale(A_AUGVEC(spot),inv_tmp_norm);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_MINRES(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i;
PetscScalar alpha,beta,ibeta,betaold,eta,c=1.0,ceta,cold=1.0,coold,s=0.0,sold=0.0,soold;
PetscScalar rho0,rho1,irho1,rho2,mrho2,rho3,mrho3,dp = 0.0;
PetscReal np;
Vec X,B,R,Z,U,V,W,UOLD,VOLD,WOLD,WOOLD;
Mat Amat,Pmat;
KSP_MINRES *minres = (KSP_MINRES*)ksp->data;
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];
Z = ksp->work[1];
U = ksp->work[2];
V = ksp->work[3];
W = ksp->work[4];
UOLD = ksp->work[5];
VOLD = ksp->work[6];
WOLD = ksp->work[7];
WOOLD = ksp->work[8];
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
ksp->its = 0;
ierr = VecSet(UOLD,0.0);CHKERRQ(ierr); /* u_old <- 0 */
ierr = VecCopy(UOLD,VOLD);CHKERRQ(ierr); /* v_old <- 0 */
ierr = VecCopy(UOLD,W);CHKERRQ(ierr); /* w <- 0 */
ierr = VecCopy(UOLD,WOLD);CHKERRQ(ierr); /* w_old <- 0 */
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr); /* r <- b - A*x */
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 <- B*r */
ierr = VecDot(R,Z,&dp);CHKERRQ(ierr);
if (PetscRealPart(dp) < minres->haptol) {
ierr = PetscInfo2(ksp,"Detected indefinite operator %g tolerance %g\n",(double)PetscRealPart(dp),(double)minres->haptol);CHKERRQ(ierr);
ksp->reason = KSP_DIVERGED_INDEFINITE_MAT;
PetscFunctionReturn(0);
}
dp = PetscAbsScalar(dp);
dp = PetscSqrtScalar(dp);
beta = dp; /* beta <- sqrt(r'*z */
eta = beta;
ierr = VecCopy(R,V);CHKERRQ(ierr);
ierr = VecCopy(Z,U);CHKERRQ(ierr);
ibeta = 1.0 / beta;
ierr = VecScale(V,ibeta);CHKERRQ(ierr); /* v <- r / beta */
ierr = VecScale(U,ibeta);CHKERRQ(ierr); /* u <- z / beta */
ierr = VecNorm(Z,NORM_2,&np);CHKERRQ(ierr); /* np <- ||z|| */
ierr = KSPLogResidualHistory(ksp,np);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,np);CHKERRQ(ierr);
ksp->rnorm = np;
ierr = (*ksp->converged)(ksp,0,np,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); /* test for convergence */
if (ksp->reason) PetscFunctionReturn(0);
i = 0;
do {
ksp->its = i+1;
/* Lanczos */
ierr = KSP_MatMult(ksp,Amat,U,R);CHKERRQ(ierr); /* r <- A*u */
ierr = VecDot(U,R,&alpha);CHKERRQ(ierr); /* alpha <- r'*u */
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- B*r */
ierr = VecAXPY(R,-alpha,V);CHKERRQ(ierr); /* r <- r - alpha v */
ierr = VecAXPY(Z,-alpha,U);CHKERRQ(ierr); /* z <- z - alpha u */
ierr = VecAXPY(R,-beta,VOLD);CHKERRQ(ierr); /* r <- r - beta v_old */
ierr = VecAXPY(Z,-beta,UOLD);CHKERRQ(ierr); /* z <- z - beta u_old */
betaold = beta;
ierr = VecDot(R,Z,&dp);CHKERRQ(ierr);
if ( PetscRealPart(dp) < minres->haptol) {
ierr = PetscInfo2(ksp,"Detected indefinite operator %g tolerance %g\n",(double)PetscRealPart(dp),(double)minres->haptol);CHKERRQ(ierr);
ksp->reason = KSP_DIVERGED_INDEFINITE_MAT;
break;
}
dp = PetscAbsScalar(dp);
beta = PetscSqrtScalar(dp); /* beta <- sqrt(r'*z) */
/* QR factorisation */
coold = cold; cold = c; soold = sold; sold = s;
rho0 = cold * alpha - coold * sold * betaold;
rho1 = PetscSqrtScalar(rho0*rho0 + beta*beta);
rho2 = sold * alpha + coold * cold * betaold;
rho3 = soold * betaold;
/* Givens rotation */
c = rho0 / rho1;
s = beta / rho1;
/* Update */
ierr = VecCopy(WOLD,WOOLD);CHKERRQ(ierr); /* w_oold <- w_old */
ierr = VecCopy(W,WOLD);CHKERRQ(ierr); /* w_old <- w */
ierr = VecCopy(U,W);CHKERRQ(ierr); /* w <- u */
mrho2 = -rho2;
ierr = VecAXPY(W,mrho2,WOLD);CHKERRQ(ierr); /* w <- w - rho2 w_old */
mrho3 = -rho3;
ierr = VecAXPY(W,mrho3,WOOLD);CHKERRQ(ierr); /* w <- w - rho3 w_oold */
irho1 = 1.0 / rho1;
ierr = VecScale(W,irho1);CHKERRQ(ierr); /* w <- w / rho1 */
ceta = c * eta;
ierr = VecAXPY(X,ceta,W);CHKERRQ(ierr); /* x <- x + c eta w */
eta = -s * eta;
ierr = VecCopy(V,VOLD);CHKERRQ(ierr);
ierr = VecCopy(U,UOLD);CHKERRQ(ierr);
ierr = VecCopy(R,V);CHKERRQ(ierr);
ierr = VecCopy(Z,U);CHKERRQ(ierr);
ibeta = 1.0 / beta;
ierr = VecScale(V,ibeta);CHKERRQ(ierr); /* v <- r / beta */
ierr = VecScale(U,ibeta);CHKERRQ(ierr); /* u <- z / beta */
np = ksp->rnorm * PetscAbsScalar(s);
ksp->rnorm = np;
ierr = KSPLogResidualHistory(ksp,np);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i+1,np);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i+1,np,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); /* test for convergence */
if (ksp->reason) break;
i++;
} while (i<ksp->max_it);
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
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);
}
static PetscErrorCode KSPSolve_BCGSL(KSP ksp)
{
KSP_BCGSL *bcgsl = (KSP_BCGSL*) ksp->data;
PetscScalar alpha, beta, omega, sigma;
PetscScalar rho0, rho1;
PetscReal kappa0, kappaA, kappa1;
PetscReal ghat;
PetscReal zeta, zeta0, rnmax_computed, rnmax_true, nrm0;
PetscBool bUpdateX;
PetscInt maxit;
PetscInt h, i, j, k, vi, ell;
PetscBLASInt ldMZ,bierr;
PetscScalar utb;
PetscReal max_s, pinv_tol;
PetscErrorCode ierr;
PetscFunctionBegin;
/* set up temporary vectors */
vi = 0;
ell = bcgsl->ell;
bcgsl->vB = ksp->work[vi]; vi++;
bcgsl->vRt = ksp->work[vi]; vi++;
bcgsl->vTm = ksp->work[vi]; vi++;
bcgsl->vvR = ksp->work+vi; vi += ell+1;
bcgsl->vvU = ksp->work+vi; vi += ell+1;
bcgsl->vXr = ksp->work[vi]; vi++;
ierr = PetscBLASIntCast(ell+1,&ldMZ);CHKERRQ(ierr);
/* Prime the iterative solver */
ierr = KSPInitialResidual(ksp, VX, VTM, VB, VVR[0], ksp->vec_rhs);CHKERRQ(ierr);
ierr = VecNorm(VVR[0], NORM_2, &zeta0);CHKERRQ(ierr);
rnmax_computed = zeta0;
rnmax_true = zeta0;
ierr = (*ksp->converged)(ksp, 0, zeta0, &ksp->reason, ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) {
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = zeta0;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ierr = VecSet(VVU[0],0.0);CHKERRQ(ierr);
alpha = 0.;
rho0 = omega = 1;
if (bcgsl->delta>0.0) {
ierr = VecCopy(VX, VXR);CHKERRQ(ierr);
ierr = VecSet(VX,0.0);CHKERRQ(ierr);
ierr = VecCopy(VVR[0], VB);CHKERRQ(ierr);
} else {
ierr = VecCopy(ksp->vec_rhs, VB);CHKERRQ(ierr);
}
/* Life goes on */
ierr = VecCopy(VVR[0], VRT);CHKERRQ(ierr);
zeta = zeta0;
ierr = KSPGetTolerances(ksp, NULL, NULL, NULL, &maxit);CHKERRQ(ierr);
for (k=0; k<maxit; k += bcgsl->ell) {
ksp->its = k;
ksp->rnorm = zeta;
ierr = KSPLogResidualHistory(ksp, zeta);CHKERRQ(ierr);
ierr = KSPMonitor(ksp, ksp->its, zeta);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp, k, zeta, &ksp->reason, ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason < 0) PetscFunctionReturn(0);
else if (ksp->reason) break;
/* BiCG part */
rho0 = -omega*rho0;
nrm0 = zeta;
for (j=0; j<bcgsl->ell; j++) {
/* rho1 <- r_j' * r_tilde */
ierr = VecDot(VVR[j], VRT, &rho1);CHKERRQ(ierr);
if (rho1 == 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN_BICG;
PetscFunctionReturn(0);
}
beta = alpha*(rho1/rho0);
rho0 = rho1;
for (i=0; i<=j; i++) {
/* u_i <- r_i - beta*u_i */
ierr = VecAYPX(VVU[i], -beta, VVR[i]);CHKERRQ(ierr);
}
/* u_{j+1} <- inv(K)*A*u_j */
ierr = KSP_PCApplyBAorAB(ksp, VVU[j], VVU[j+1], VTM);CHKERRQ(ierr);
ierr = VecDot(VVU[j+1], VRT, &sigma);CHKERRQ(ierr);
if (sigma == 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN_BICG;
PetscFunctionReturn(0);
}
alpha = rho1/sigma;
/* x <- x + alpha*u_0 */
ierr = VecAXPY(VX, alpha, VVU[0]);CHKERRQ(ierr);
for (i=0; i<=j; i++) {
/* r_i <- r_i - alpha*u_{i+1} */
ierr = VecAXPY(VVR[i], -alpha, VVU[i+1]);CHKERRQ(ierr);
}
/* r_{j+1} <- inv(K)*A*r_j */
ierr = KSP_PCApplyBAorAB(ksp, VVR[j], VVR[j+1], VTM);CHKERRQ(ierr);
ierr = VecNorm(VVR[0], NORM_2, &nrm0);CHKERRQ(ierr);
if (bcgsl->delta>0.0) {
if (rnmax_computed<nrm0) rnmax_computed = nrm0;
if (rnmax_true<nrm0) rnmax_true = nrm0;
}
/* NEW: check for early exit */
ierr = (*ksp->converged)(ksp, k+j, nrm0, &ksp->reason, ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) {
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = k+j;
ksp->rnorm = nrm0;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
if (ksp->reason < 0) PetscFunctionReturn(0);
}
}
/* Polynomial part */
for (i = 0; i <= bcgsl->ell; ++i) {
ierr = VecMDot(VVR[i], i+1, VVR, &MZa[i*ldMZ]);CHKERRQ(ierr);
}
/* Symmetrize MZa */
for (i = 0; i <= bcgsl->ell; ++i) {
for (j = i+1; j <= bcgsl->ell; ++j) {
MZa[i*ldMZ+j] = MZa[j*ldMZ+i] = PetscConj(MZa[j*ldMZ+i]);
}
}
/* Copy MZa to MZb */
ierr = PetscMemcpy(MZb,MZa,ldMZ*ldMZ*sizeof(PetscScalar));CHKERRQ(ierr);
if (!bcgsl->bConvex || bcgsl->ell==1) {
PetscBLASInt ione = 1,bell;
ierr = PetscBLASIntCast(bcgsl->ell,&bell);CHKERRQ(ierr);
AY0c[0] = -1;
if (bcgsl->pinv) {
#if defined(PETSC_MISSING_LAPACK_GESVD)
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"GESVD - Lapack routine is unavailable.");
#else
# if defined(PETSC_USE_COMPLEX)
PetscStackCall("LAPACKgesvd",LAPACKgesvd_("A","A",&bell,&bell,&MZa[1+ldMZ],&ldMZ,bcgsl->s,bcgsl->u,&bell,bcgsl->v,&bell,bcgsl->work,&bcgsl->lwork,bcgsl->realwork,&bierr));
# else
PetscStackCall("LAPACKgesvd",LAPACKgesvd_("A","A",&bell,&bell,&MZa[1+ldMZ],&ldMZ,bcgsl->s,bcgsl->u,&bell,bcgsl->v,&bell,bcgsl->work,&bcgsl->lwork,&bierr));
# endif
#endif
if (bierr!=0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
PetscFunctionReturn(0);
}
/* Apply pseudo-inverse */
max_s = bcgsl->s[0];
for (i=1; i<bell; i++) {
if (bcgsl->s[i] > max_s) {
max_s = bcgsl->s[i];
}
}
/* tolerance is hardwired to bell*max(s)*PETSC_MACHINE_EPSILON */
pinv_tol = bell*max_s*PETSC_MACHINE_EPSILON;
ierr = PetscMemzero(&AY0c[1],bell*sizeof(PetscScalar));CHKERRQ(ierr);
for (i=0; i<bell; i++) {
if (bcgsl->s[i] >= pinv_tol) {
utb=0.;
for (j=0; j<bell; j++) {
utb += MZb[1+j]*bcgsl->u[i*bell+j];
}
for (j=0; j<bell; j++) {
AY0c[1+j] += utb/bcgsl->s[i]*bcgsl->v[j*bell+i];
}
}
}
} else {
#if defined(PETSC_MISSING_LAPACK_POTRF)
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"POTRF - Lapack routine is unavailable.");
#else
PetscStackCall("LAPACKpotrf",LAPACKpotrf_("Lower", &bell, &MZa[1+ldMZ], &ldMZ, &bierr));
#endif
if (bierr!=0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
PetscFunctionReturn(0);
}
ierr = PetscMemcpy(&AY0c[1],&MZb[1],bcgsl->ell*sizeof(PetscScalar));CHKERRQ(ierr);
PetscStackCall("LAPACKpotrs",LAPACKpotrs_("Lower", &bell, &ione, &MZa[1+ldMZ], &ldMZ, &AY0c[1], &ldMZ, &bierr));
}
} else {
PetscBLASInt ione = 1;
PetscScalar aone = 1.0, azero = 0.0;
PetscBLASInt neqs;
ierr = PetscBLASIntCast(bcgsl->ell-1,&neqs);CHKERRQ(ierr);
#if defined(PETSC_MISSING_LAPACK_POTRF)
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"POTRF - Lapack routine is unavailable.");
#else
PetscStackCall("LAPACKpotrf",LAPACKpotrf_("Lower", &neqs, &MZa[1+ldMZ], &ldMZ, &bierr));
#endif
if (bierr!=0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
PetscFunctionReturn(0);
}
ierr = PetscMemcpy(&AY0c[1],&MZb[1],(bcgsl->ell-1)*sizeof(PetscScalar));CHKERRQ(ierr);
PetscStackCall("LAPACKpotrs",LAPACKpotrs_("Lower", &neqs, &ione, &MZa[1+ldMZ], &ldMZ, &AY0c[1], &ldMZ, &bierr));
AY0c[0] = -1;
AY0c[bcgsl->ell] = 0.;
ierr = PetscMemcpy(&AYlc[1],&MZb[1+ldMZ*(bcgsl->ell)],(bcgsl->ell-1)*sizeof(PetscScalar));CHKERRQ(ierr);
PetscStackCall("LAPACKpotrs",LAPACKpotrs_("Lower", &neqs, &ione, &MZa[1+ldMZ], &ldMZ, &AYlc[1], &ldMZ, &bierr));
AYlc[0] = 0.;
AYlc[bcgsl->ell] = -1;
PetscStackCall("BLASgemv",BLASgemv_("NoTr", &ldMZ, &ldMZ, &aone, MZb, &ldMZ, AY0c, &ione, &azero, AYtc, &ione));
kappa0 = PetscRealPart(BLASdot_(&ldMZ, AY0c, &ione, AYtc, &ione));
/* round-off can cause negative kappa's */
if (kappa0<0) kappa0 = -kappa0;
kappa0 = PetscSqrtReal(kappa0);
kappaA = PetscRealPart(BLASdot_(&ldMZ, AYlc, &ione, AYtc, &ione));
PetscStackCall("BLASgemv",BLASgemv_("noTr", &ldMZ, &ldMZ, &aone, MZb, &ldMZ, AYlc, &ione, &azero, AYtc, &ione));
kappa1 = PetscRealPart(BLASdot_(&ldMZ, AYlc, &ione, AYtc, &ione));
if (kappa1<0) kappa1 = -kappa1;
kappa1 = PetscSqrtReal(kappa1);
if (kappa0!=0.0 && kappa1!=0.0) {
if (kappaA<0.7*kappa0*kappa1) {
ghat = (kappaA<0.0) ? -0.7*kappa0/kappa1 : 0.7*kappa0/kappa1;
} else {
ghat = kappaA/(kappa1*kappa1);
}
for (i=0; i<=bcgsl->ell; i++) {
AY0c[i] = AY0c[i] - ghat* AYlc[i];
}
}
}
omega = AY0c[bcgsl->ell];
for (h=bcgsl->ell; h>0 && omega==0.0; h--) omega = AY0c[h];
if (omega==0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
PetscFunctionReturn(0);
}
ierr = VecMAXPY(VX, bcgsl->ell,AY0c+1, VVR);CHKERRQ(ierr);
for (i=1; i<=bcgsl->ell; i++) AY0c[i] *= -1.0;
ierr = VecMAXPY(VVU[0], bcgsl->ell,AY0c+1, VVU+1);CHKERRQ(ierr);
ierr = VecMAXPY(VVR[0], bcgsl->ell,AY0c+1, VVR+1);CHKERRQ(ierr);
for (i=1; i<=bcgsl->ell; i++) AY0c[i] *= -1.0;
ierr = VecNorm(VVR[0], NORM_2, &zeta);CHKERRQ(ierr);
/* Accurate Update */
if (bcgsl->delta>0.0) {
if (rnmax_computed<zeta) rnmax_computed = zeta;
if (rnmax_true<zeta) rnmax_true = zeta;
bUpdateX = (PetscBool) (zeta<bcgsl->delta*zeta0 && zeta0<=rnmax_computed);
if ((zeta<bcgsl->delta*rnmax_true && zeta0<=rnmax_true) || bUpdateX) {
/* r0 <- b-inv(K)*A*X */
ierr = KSP_PCApplyBAorAB(ksp, VX, VVR[0], VTM);CHKERRQ(ierr);
ierr = VecAYPX(VVR[0], -1.0, VB);CHKERRQ(ierr);
rnmax_true = zeta;
if (bUpdateX) {
ierr = VecAXPY(VXR,1.0,VX);CHKERRQ(ierr);
ierr = VecSet(VX,0.0);CHKERRQ(ierr);
ierr = VecCopy(VVR[0], VB);CHKERRQ(ierr);
rnmax_computed = zeta;
}
}
}
}
if (bcgsl->delta>0.0) {
ierr = VecAXPY(VX,1.0,VXR);CHKERRQ(ierr);
}
ierr = (*ksp->converged)(ksp, k, zeta, &ksp->reason, ksp->cnvP);CHKERRQ(ierr);
if (!ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
PetscErrorCode KSPGMRESCycle(PetscInt *itcount,KSP ksp)
{
KSP_GMRES *gmres = (KSP_GMRES*)(ksp->data);
PetscReal res_norm,res,hapbnd,tt;
PetscErrorCode ierr;
PetscInt it = 0, max_k = gmres->max_k;
PetscBool hapend = PETSC_FALSE;
PetscFunctionBegin;
ierr = VecNormalize(VEC_VV(0),&res_norm);CHKERRQ(ierr);
res = res_norm;
*GRS(0) = res_norm;
/* check for the convergence */
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->rnorm = res;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
gmres->it = (it - 1);
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
if (!res) {
if (itcount) *itcount = 0;
ksp->reason = KSP_CONVERGED_ATOL;
ierr = PetscInfo(ksp,"Converged due to zero residual norm on entry\n");CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
while (!ksp->reason && it < max_k && ksp->its < ksp->max_it) {
if (it) {
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
}
gmres->it = (it - 1);
if (gmres->vv_allocated <= it + VEC_OFFSET + 1) {
ierr = KSPGMRESGetNewVectors(ksp,it+1);CHKERRQ(ierr);
}
ierr = KSP_PCApplyBAorAB(ksp,VEC_VV(it),VEC_VV(1+it),VEC_TEMP_MATOP);CHKERRQ(ierr);
/* update hessenberg matrix and do Gram-Schmidt */
ierr = (*gmres->orthog)(ksp,it);CHKERRQ(ierr);
/* vv(i+1) . vv(i+1) */
ierr = VecNormalize(VEC_VV(it+1),&tt);CHKERRQ(ierr);
/* save the magnitude */
*HH(it+1,it) = tt;
*HES(it+1,it) = tt;
/* check for the happy breakdown */
hapbnd = PetscAbsScalar(tt / *GRS(it));
if (hapbnd > gmres->haptol) hapbnd = gmres->haptol;
if (tt < hapbnd) {
ierr = PetscInfo2(ksp,"Detected happy breakdown, current hapbnd = %14.12e tt = %14.12e\n",(double)hapbnd,(double)tt);CHKERRQ(ierr);
hapend = PETSC_TRUE;
}
ierr = KSPGMRESUpdateHessenberg(ksp,it,hapend,&res);CHKERRQ(ierr);
it++;
gmres->it = (it-1); /* For converged */
ksp->its++;
ksp->rnorm = res;
if (ksp->reason) break;
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
/* Catch error in happy breakdown and signal convergence and break from loop */
if (hapend) {
if (!ksp->reason) {
if (ksp->errorifnotconverged) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"You reached the happy break down, but convergence was not indicated. Residual norm = %G",res);
else {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
}
}
}
/* Monitor if we know that we will not return for a restart */
if (it && (ksp->reason || ksp->its >= ksp->max_it)) {
ierr = KSPLogResidualHistory(ksp,res);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
}
if (itcount) *itcount = it;
/*
Down here we have to solve for the "best" coefficients of the Krylov
columns, add the solution values together, and possibly unwind the
preconditioning from the solution
*/
/* Form the solution (or the solution so far) */
ierr = KSPGMRESBuildSoln(GRS(0),ksp->vec_sol,ksp->vec_sol,ksp,it-1);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_SYMMLQ(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i;
PetscScalar alpha,beta,ibeta,betaold,beta1,ceta = 0,ceta_oold = 0.0, ceta_old = 0.0,ceta_bar;
PetscScalar c = 1.0,cold=1.0,s=0.0,sold=0.0,coold,soold,rho0,rho1,rho2,rho3;
PetscScalar dp = 0.0;
PetscReal np,s_prod;
Vec X,B,R,Z,U,V,W,UOLD,VOLD,Wbar;
Mat Amat,Pmat;
MatStructure pflag;
KSP_SYMMLQ *symmlq = (KSP_SYMMLQ*)ksp->data;
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];
Z = ksp->work[1];
U = ksp->work[2];
V = ksp->work[3];
W = ksp->work[4];
UOLD = ksp->work[5];
VOLD = ksp->work[6];
Wbar = ksp->work[7];
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat,&pflag);CHKERRQ(ierr);
ksp->its = 0;
ierr = VecSet(UOLD,0.0);CHKERRQ(ierr); /* u_old <- zeros; */
ierr = VecCopy(UOLD,VOLD);CHKERRQ(ierr); /* v_old <- u_old; */
ierr = VecCopy(UOLD,W);CHKERRQ(ierr); /* w <- u_old; */
ierr = VecCopy(UOLD,Wbar);CHKERRQ(ierr); /* w_bar <- u_old; */
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr); /* r <- b - A*x */
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 <- B*r */
ierr = VecDot(R,Z,&dp);CHKERRQ(ierr); /* dp = r'*z; */
if (PetscAbsScalar(dp) < symmlq->haptol) {
ierr = PetscInfo2(ksp,"Detected happy breakdown %G tolerance %G\n",PetscAbsScalar(dp),symmlq->haptol);CHKERRQ(ierr);
ksp->rnorm = 0.0; /* what should we really put here? */
ksp->reason = KSP_CONVERGED_HAPPY_BREAKDOWN; /* bugfix proposed by Lourens ([email protected]) */
PetscFunctionReturn(0);
}
#if !defined(PETSC_USE_COMPLEX)
if (dp < 0.0) {
ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
PetscFunctionReturn(0);
}
#endif
dp = PetscSqrtScalar(dp);
beta = dp; /* beta <- sqrt(r'*z) */
beta1 = beta;
s_prod = PetscAbsScalar(beta1);
ierr = VecCopy(R,V);CHKERRQ(ierr); /* v <- r; */
ierr = VecCopy(Z,U);CHKERRQ(ierr); /* u <- z; */
ibeta = 1.0 / beta;
ierr = VecScale(V,ibeta);CHKERRQ(ierr); /* v <- ibeta*v; */
ierr = VecScale(U,ibeta);CHKERRQ(ierr); /* u <- ibeta*u; */
ierr = VecCopy(U,Wbar);CHKERRQ(ierr); /* w_bar <- u; */
ierr = VecNorm(Z,NORM_2,&np);CHKERRQ(ierr); /* np <- ||z|| */
ierr = KSPLogResidualHistory(ksp,np);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,np);CHKERRQ(ierr);
ksp->rnorm = np;
ierr = (*ksp->converged)(ksp,0,np,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); /* test for convergence */
if (ksp->reason) PetscFunctionReturn(0);
i = 0; ceta = 0.;
do {
ksp->its = i+1;
/* Update */
if (ksp->its > 1) {
ierr = VecCopy(V,VOLD);CHKERRQ(ierr); /* v_old <- v; */
ierr = VecCopy(U,UOLD);CHKERRQ(ierr); /* u_old <- u; */
ierr = VecCopy(R,V);CHKERRQ(ierr);
ierr = VecScale(V,1.0/beta);CHKERRQ(ierr); /* v <- ibeta*r; */
ierr = VecCopy(Z,U);CHKERRQ(ierr);
ierr = VecScale(U,1.0/beta);CHKERRQ(ierr); /* u <- ibeta*z; */
ierr = VecCopy(Wbar,W);CHKERRQ(ierr);
ierr = VecScale(W,c);CHKERRQ(ierr);
ierr = VecAXPY(W,s,U);CHKERRQ(ierr); /* w <- c*w_bar + s*u; (w_k) */
ierr = VecScale(Wbar,-s);CHKERRQ(ierr);
ierr = VecAXPY(Wbar,c,U);CHKERRQ(ierr); /* w_bar <- -s*w_bar + c*u; (w_bar_(k+1)) */
ierr = VecAXPY(X,ceta,W);CHKERRQ(ierr); /* x <- x + ceta * w; (xL_k) */
ceta_oold = ceta_old;
ceta_old = ceta;
}
/* Lanczos */
ierr = KSP_MatMult(ksp,Amat,U,R);CHKERRQ(ierr); /* r <- Amat*u; */
ierr = VecDot(U,R,&alpha);CHKERRQ(ierr); /* alpha <- u'*r; */
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- B*r; */
ierr = VecAXPY(R,-alpha,V);CHKERRQ(ierr); /* r <- r - alpha* v; */
ierr = VecAXPY(Z,-alpha,U);CHKERRQ(ierr); /* z <- z - alpha* u; */
ierr = VecAXPY(R,-beta,VOLD);CHKERRQ(ierr); /* r <- r - beta * v_old; */
ierr = VecAXPY(Z,-beta,UOLD);CHKERRQ(ierr); /* z <- z - beta * u_old; */
betaold = beta; /* beta_k */
ierr = VecDot(R,Z,&dp);CHKERRQ(ierr); /* dp <- r'*z; */
if (PetscAbsScalar(dp) < symmlq->haptol) {
ierr = PetscInfo2(ksp,"Detected happy breakdown %G tolerance %G\n",PetscAbsScalar(dp),symmlq->haptol);CHKERRQ(ierr);
dp = 0.0;
}
#if !defined(PETSC_USE_COMPLEX)
if (dp < 0.0) {
ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
break;
}
#endif
beta = PetscSqrtScalar(dp); /* beta = sqrt(dp); */
/* QR factorization */
coold = cold; cold = c; soold = sold; sold = s;
rho0 = cold * alpha - coold * sold * betaold; /* gamma_bar */
rho1 = PetscSqrtScalar(rho0*rho0 + beta*beta); /* gamma */
rho2 = sold * alpha + coold * cold * betaold; /* delta */
rho3 = soold * betaold; /* epsilon */
/* Givens rotation: [c -s; s c] (different from the Reference!) */
c = rho0 / rho1; s = beta / rho1;
if (ksp->its==1) ceta = beta1/rho1;
else ceta = -(rho2*ceta_old + rho3*ceta_oold)/rho1;
s_prod = s_prod*PetscAbsScalar(s);
if (c == 0.0) np = s_prod*1.e16;
else np = s_prod/PetscAbsScalar(c); /* residual norm for xc_k (CGNORM) */
ksp->rnorm = np;
ierr = KSPLogResidualHistory(ksp,np);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i+1,np);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i+1,np,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); /* test for convergence */
if (ksp->reason) break;
i++;
} while (i<ksp->max_it);
/* move to the CG point: xc_(k+1) */
if (c == 0.0) ceta_bar = ceta*1.e15;
else ceta_bar = ceta/c;
ierr = VecAXPY(X,ceta_bar,Wbar);CHKERRQ(ierr); /* x <- x + ceta_bar*w_bar */
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_Richardson(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,maxit;
PetscReal rnorm = 0.0,abr;
PetscScalar scale,rdot;
Vec x,b,r,z,w = NULL,y = NULL;
PetscInt xs, ws;
Mat Amat,Pmat;
KSP_Richardson *richardsonP = (KSP_Richardson*)ksp->data;
PetscBool exists,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);
ksp->its = 0;
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
x = ksp->vec_sol;
b = ksp->vec_rhs;
ierr = VecGetSize(x,&xs);CHKERRQ(ierr);
ierr = VecGetSize(ksp->work[0],&ws);CHKERRQ(ierr);
if (xs != ws) {
if (richardsonP->selfscale) {
ierr = KSPSetWorkVecs(ksp,4);CHKERRQ(ierr);
} else {
ierr = KSPSetWorkVecs(ksp,2);CHKERRQ(ierr);
}
}
r = ksp->work[0];
z = ksp->work[1];
if (richardsonP->selfscale) {
w = ksp->work[2];
y = ksp->work[3];
}
maxit = ksp->max_it;
/* if user has provided fast Richardson code use that */
ierr = PCApplyRichardsonExists(ksp->pc,&exists);CHKERRQ(ierr);
if (exists && !ksp->numbermonitors && !ksp->transpose_solve & !ksp->nullsp) {
PCRichardsonConvergedReason reason;
ierr = PCApplyRichardson(ksp->pc,b,x,r,ksp->rtol,ksp->abstol,ksp->divtol,maxit,ksp->guess_zero,&ksp->its,&reason);CHKERRQ(ierr);
ksp->reason = (KSPConvergedReason)reason;
PetscFunctionReturn(0);
}
scale = richardsonP->scale;
if (!ksp->guess_zero) { /* r <- b - A x */
ierr = KSP_MatMult(ksp,Amat,x,r);CHKERRQ(ierr);
ierr = VecAYPX(r,-1.0,b);CHKERRQ(ierr);
} else {
ierr = VecCopy(b,r);CHKERRQ(ierr);
}
ksp->its = 0;
if (richardsonP->selfscale) {
ierr = KSP_PCApply(ksp,r,z);CHKERRQ(ierr); /* z <- B r */
for (i=0; i<maxit; i++) {
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNorm(r,NORM_2,&rnorm);CHKERRQ(ierr); /* rnorm <- r'*r */
ierr = KSPMonitor(ksp,i,rnorm);CHKERRQ(ierr);
ksp->rnorm = rnorm;
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
} else if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNorm(z,NORM_2,&rnorm);CHKERRQ(ierr); /* rnorm <- z'*z */
ierr = KSPMonitor(ksp,i,rnorm);CHKERRQ(ierr);
ksp->rnorm = rnorm;
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
}
ierr = KSP_PCApplyBAorAB(ksp,z,y,w);CHKERRQ(ierr); /* y = BAz = BABr */
ierr = VecDotNorm2(z,y,&rdot,&abr);CHKERRQ(ierr); /* rdot = (Br)^T(BABR); abr = (BABr)^T (BABr) */
scale = rdot/abr;
ierr = PetscInfo1(ksp,"Self-scale factor %g\n",(double)PetscRealPart(scale));CHKERRQ(ierr);
ierr = VecAXPY(x,scale,z);CHKERRQ(ierr); /* x <- x + scale z */
ierr = VecAXPY(r,-scale,w);CHKERRQ(ierr); /* r <- r - scale*Az */
ierr = VecAXPY(z,-scale,y);CHKERRQ(ierr); /* z <- z - scale*y */
ksp->its++;
}
} else {
for (i=0; i<maxit; i++) {
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNorm(r,NORM_2,&rnorm);CHKERRQ(ierr); /* rnorm <- r'*r */
ierr = KSPMonitor(ksp,i,rnorm);CHKERRQ(ierr);
ksp->rnorm = rnorm;
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
}
ierr = KSP_PCApply(ksp,r,z);CHKERRQ(ierr); /* z <- B r */
if (ksp->normtype == KSP_NORM_PRECONDITIONED) {
ierr = VecNorm(z,NORM_2,&rnorm);CHKERRQ(ierr); /* rnorm <- z'*z */
ierr = KSPMonitor(ksp,i,rnorm);CHKERRQ(ierr);
ksp->rnorm = rnorm;
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
}
ierr = VecAXPY(x,scale,z);CHKERRQ(ierr); /* x <- x + scale z */
ksp->its++;
if (i+1 < maxit || ksp->normtype != KSP_NORM_NONE) {
ierr = KSP_MatMult(ksp,Amat,x,r);CHKERRQ(ierr); /* r <- b - Ax */
ierr = VecAYPX(r,-1.0,b);CHKERRQ(ierr);
}
}
}
if (!ksp->reason) {
if (ksp->normtype != KSP_NORM_NONE) {
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
ierr = VecNorm(r,NORM_2,&rnorm);CHKERRQ(ierr); /* rnorm <- r'*r */
} else {
ierr = KSP_PCApply(ksp,r,z);CHKERRQ(ierr); /* z <- B r */
ierr = VecNorm(z,NORM_2,&rnorm);CHKERRQ(ierr); /* rnorm <- z'*z */
}
ksp->rnorm = rnorm;
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i,rnorm);CHKERRQ(ierr);
}
if (ksp->its >= ksp->max_it) {
if (ksp->normtype != KSP_NORM_NONE) {
ierr = (*ksp->converged)(ksp,i,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (!ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
} else {
ksp->reason = KSP_CONVERGED_ITS;
}
}
}
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_LSQR(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,size1,size2;
PetscScalar rho,rhobar,phi,phibar,theta,c,s,tmp,tau;
PetscReal beta,alpha,rnorm;
Vec X,B,V,V1,U,U1,TMP,W,W2,SE,Z = NULL;
Mat Amat,Pmat;
MatStructure pflag;
KSP_LSQR *lsqr = (KSP_LSQR*)ksp->data;
PetscBool diagonalscale,nopreconditioner;
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);
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat,&pflag);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)ksp->pc,PCNONE,&nopreconditioner);CHKERRQ(ierr);
/* nopreconditioner =PETSC_FALSE; */
/* Calculate norm of right hand side */
ierr = VecNorm(ksp->vec_rhs,NORM_2,&lsqr->rhs_norm);CHKERRQ(ierr);
/* mark norm of matrix with negative number to indicate it has not yet been computed */
lsqr->anorm = -1.0;
/* vectors of length m, where system size is mxn */
B = ksp->vec_rhs;
U = lsqr->vwork_m[0];
U1 = lsqr->vwork_m[1];
/* vectors of length n */
X = ksp->vec_sol;
W = lsqr->vwork_n[0];
V = lsqr->vwork_n[1];
V1 = lsqr->vwork_n[2];
W2 = lsqr->vwork_n[3];
if (!nopreconditioner) Z = lsqr->vwork_n[4];
/* standard error vector */
SE = lsqr->se;
if (SE) {
ierr = VecGetSize(SE,&size1);CHKERRQ(ierr);
ierr = VecGetSize(X,&size2);CHKERRQ(ierr);
if (size1 != size2) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Standard error vector (size %d) does not match solution vector (size %d)",size1,size2);
ierr = VecSet(SE,0.0);CHKERRQ(ierr);
}
/* Compute initial residual, temporarily use work vector u */
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,U);CHKERRQ(ierr); /* u <- b - Ax */
ierr = VecAYPX(U,-1.0,B);CHKERRQ(ierr);
} else {
ierr = VecCopy(B,U);CHKERRQ(ierr); /* u <- b (x is 0) */
}
/* Test for nothing to do */
ierr = VecNorm(U,NORM_2,&rnorm);CHKERRQ(ierr);
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = rnorm;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,0,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
beta = rnorm;
ierr = VecScale(U,1.0/beta);CHKERRQ(ierr);
ierr = KSP_MatMultTranspose(ksp,Amat,U,V);CHKERRQ(ierr);
if (nopreconditioner) {
ierr = VecNorm(V,NORM_2,&alpha);CHKERRQ(ierr);
} else {
ierr = PCApply(ksp->pc,V,Z);CHKERRQ(ierr);
ierr = VecDotRealPart(V,Z,&alpha);CHKERRQ(ierr);
if (alpha <= 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
PetscFunctionReturn(0);
}
alpha = PetscSqrtReal(alpha);
ierr = VecScale(Z,1.0/alpha);CHKERRQ(ierr);
}
ierr = VecScale(V,1.0/alpha);CHKERRQ(ierr);
if (nopreconditioner) {
ierr = VecCopy(V,W);CHKERRQ(ierr);
} else {
ierr = VecCopy(Z,W);CHKERRQ(ierr);
}
lsqr->arnorm = alpha * beta;
phibar = beta;
rhobar = alpha;
i = 0;
do {
if (nopreconditioner) {
ierr = KSP_MatMult(ksp,Amat,V,U1);CHKERRQ(ierr);
} else {
ierr = KSP_MatMult(ksp,Amat,Z,U1);CHKERRQ(ierr);
}
ierr = VecAXPY(U1,-alpha,U);CHKERRQ(ierr);
ierr = VecNorm(U1,NORM_2,&beta);CHKERRQ(ierr);
if (beta == 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
ierr = VecScale(U1,1.0/beta);CHKERRQ(ierr);
ierr = KSP_MatMultTranspose(ksp,Amat,U1,V1);CHKERRQ(ierr);
ierr = VecAXPY(V1,-beta,V);CHKERRQ(ierr);
if (nopreconditioner) {
ierr = VecNorm(V1,NORM_2,&alpha);CHKERRQ(ierr);
} else {
ierr = PCApply(ksp->pc,V1,Z);CHKERRQ(ierr);
ierr = VecDotRealPart(V1,Z,&alpha);CHKERRQ(ierr);
if (alpha <= 0.0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
break;
}
alpha = PetscSqrtReal(alpha);
ierr = VecScale(Z,1.0/alpha);CHKERRQ(ierr);
}
ierr = VecScale(V1,1.0/alpha);CHKERRQ(ierr);
rho = PetscSqrtScalar(rhobar*rhobar + beta*beta);
c = rhobar / rho;
s = beta / rho;
theta = s * alpha;
rhobar = -c * alpha;
phi = c * phibar;
phibar = s * phibar;
tau = s * phi;
ierr = VecAXPY(X,phi/rho,W);CHKERRQ(ierr); /* x <- x + (phi/rho) w */
if (SE) {
ierr = VecCopy(W,W2);CHKERRQ(ierr);
ierr = VecSquare(W2);CHKERRQ(ierr);
ierr = VecScale(W2,1.0/(rho*rho));CHKERRQ(ierr);
ierr = VecAXPY(SE, 1.0, W2);CHKERRQ(ierr); /* SE <- SE + (w^2/rho^2) */
}
if (nopreconditioner) {
ierr = VecAYPX(W,-theta/rho,V1);CHKERRQ(ierr); /* w <- v - (theta/rho) w */
} else {
ierr = VecAYPX(W,-theta/rho,Z);CHKERRQ(ierr); /* w <- z - (theta/rho) w */
}
lsqr->arnorm = alpha*PetscAbsScalar(tau);
rnorm = PetscRealPart(phibar);
ierr = PetscObjectAMSTakeAccess((PetscObject)ksp);CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = rnorm;
ierr = PetscObjectAMSGrantAccess((PetscObject)ksp);CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rnorm);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i+1,rnorm);CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i+1,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
SWAP(U1,U,TMP);
SWAP(V1,V,TMP);
i++;
} while (i<ksp->max_it);
if (i >= ksp->max_it && !ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
/* Finish off the standard error estimates */
if (SE) {
tmp = 1.0;
ierr = MatGetSize(Amat,&size1,&size2);CHKERRQ(ierr);
if (size1 > size2) tmp = size1 - size2;
tmp = rnorm / PetscSqrtScalar(tmp);
ierr = VecSqrtAbs(SE);CHKERRQ(ierr);
ierr = VecScale(SE,tmp);CHKERRQ(ierr);
}
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);
}
static PetscErrorCode KSPSolve_TCQMR(KSP ksp)
{
PetscReal rnorm0,rnorm,dp1,Gamma;
PetscScalar theta,ep,cl1,sl1,cl,sl,sprod,tau_n1,f;
PetscScalar deltmp,rho,beta,eptmp,ta,s,c,tau_n,delta;
PetscScalar dp11,dp2,rhom1,alpha,tmp;
PetscErrorCode ierr;
PetscFunctionBegin;
ksp->its = 0;
ierr = KSPInitialResidual(ksp,x,u,v,r,b);CHKERRQ(ierr);
ierr = VecNorm(r,NORM_2,&rnorm0);CHKERRQ(ierr); /* rnorm0 = ||r|| */
KSPCheckNorm(ksp,rnorm0);
ierr = (*ksp->converged)(ksp,0,rnorm0,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
ierr = VecSet(um1,0.0);CHKERRQ(ierr);
ierr = VecCopy(r,u);CHKERRQ(ierr);
rnorm = rnorm0;
tmp = 1.0/rnorm; ierr = VecScale(u,tmp);CHKERRQ(ierr);
ierr = VecSet(vm1,0.0);CHKERRQ(ierr);
ierr = VecCopy(u,v);CHKERRQ(ierr);
ierr = VecCopy(u,v0);CHKERRQ(ierr);
ierr = VecSet(pvec1,0.0);CHKERRQ(ierr);
ierr = VecSet(pvec2,0.0);CHKERRQ(ierr);
ierr = VecSet(p,0.0);CHKERRQ(ierr);
theta = 0.0;
ep = 0.0;
cl1 = 0.0;
sl1 = 0.0;
cl = 0.0;
sl = 0.0;
sprod = 1.0;
tau_n1= rnorm0;
f = 1.0;
Gamma = 1.0;
rhom1 = 1.0;
/*
CALCULATE SQUARED LANCZOS vectors
*/
ierr = (*ksp->converged)(ksp,ksp->its,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
while (!ksp->reason) {
ierr = KSPMonitor(ksp,ksp->its,rnorm);CHKERRQ(ierr);
ksp->its++;
ierr = KSP_PCApplyBAorAB(ksp,u,y,vtmp);CHKERRQ(ierr); /* y = A*u */
ierr = VecDot(y,v0,&dp11);CHKERRQ(ierr);
KSPCheckDot(ksp,dp11);
ierr = VecDot(u,v0,&dp2);CHKERRQ(ierr);
alpha = dp11 / dp2; /* alpha = v0'*y/v0'*u */
deltmp = alpha;
ierr = VecCopy(y,z);CHKERRQ(ierr);
ierr = VecAXPY(z,-alpha,u);CHKERRQ(ierr); /* z = y - alpha u */
ierr = VecDot(u,v0,&rho);CHKERRQ(ierr);
beta = rho / (f*rhom1);
rhom1 = rho;
ierr = VecCopy(z,utmp);CHKERRQ(ierr); /* up1 = (A-alpha*I)*
(z-2*beta*p) + f*beta*
beta*um1 */
ierr = VecAXPY(utmp,-2.0*beta,p);CHKERRQ(ierr);
ierr = KSP_PCApplyBAorAB(ksp,utmp,up1,vtmp);CHKERRQ(ierr);
ierr = VecAXPY(up1,-alpha,utmp);CHKERRQ(ierr);
ierr = VecAXPY(up1,f*beta*beta,um1);CHKERRQ(ierr);
ierr = VecNorm(up1,NORM_2,&dp1);CHKERRQ(ierr);
KSPCheckNorm(ksp,dp1);
f = 1.0 / dp1;
ierr = VecScale(up1,f);CHKERRQ(ierr);
ierr = VecAYPX(p,-beta,z);CHKERRQ(ierr); /* p = f*(z-beta*p) */
ierr = VecScale(p,f);CHKERRQ(ierr);
ierr = VecCopy(u,um1);CHKERRQ(ierr);
ierr = VecCopy(up1,u);CHKERRQ(ierr);
beta = beta/Gamma;
eptmp = beta;
ierr = KSP_PCApplyBAorAB(ksp,v,vp1,vtmp);CHKERRQ(ierr);
ierr = VecAXPY(vp1,-alpha,v);CHKERRQ(ierr);
ierr = VecAXPY(vp1,-beta,vm1);CHKERRQ(ierr);
ierr = VecNorm(vp1,NORM_2,&Gamma);CHKERRQ(ierr);
KSPCheckNorm(ksp,Gamma);
ierr = VecScale(vp1,1.0/Gamma);CHKERRQ(ierr);
ierr = VecCopy(v,vm1);CHKERRQ(ierr);
ierr = VecCopy(vp1,v);CHKERRQ(ierr);
/*
SOLVE Ax = b
*/
/* Apply last two Given's (Gl-1 and Gl) rotations to (beta,alpha,Gamma) */
if (ksp->its > 2) {
theta = sl1*beta;
eptmp = -cl1*beta;
}
if (ksp->its > 1) {
ep = -cl*eptmp + sl*alpha;
deltmp = -sl*eptmp - cl*alpha;
}
if (PetscAbsReal(Gamma) > PetscAbsScalar(deltmp)) {
ta = -deltmp / Gamma;
s = 1.0 / PetscSqrtScalar(1.0 + ta*ta);
c = s*ta;
} else {
ta = -Gamma/deltmp;
c = 1.0 / PetscSqrtScalar(1.0 + ta*ta);
s = c*ta;
}
delta = -c*deltmp + s*Gamma;
tau_n = -c*tau_n1; tau_n1 = -s*tau_n1;
ierr = VecCopy(vm1,pvec);CHKERRQ(ierr);
ierr = VecAXPY(pvec,-theta,pvec2);CHKERRQ(ierr);
ierr = VecAXPY(pvec,-ep,pvec1);CHKERRQ(ierr);
ierr = VecScale(pvec,1.0/delta);CHKERRQ(ierr);
ierr = VecAXPY(x,tau_n,pvec);CHKERRQ(ierr);
cl1 = cl; sl1 = sl; cl = c; sl = s;
ierr = VecCopy(pvec1,pvec2);CHKERRQ(ierr);
ierr = VecCopy(pvec,pvec1);CHKERRQ(ierr);
/* Compute the upper bound on the residual norm r (See QMR paper p. 13) */
sprod = sprod*PetscAbsScalar(s);
rnorm = rnorm0 * PetscSqrtReal((PetscReal)ksp->its+2.0) * PetscRealPart(sprod);
ierr = (*ksp->converged)(ksp,ksp->its,rnorm,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->its >= ksp->max_it) {
if (!ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
break;
}
}
ierr = KSPMonitor(ksp,ksp->its,rnorm);CHKERRQ(ierr);
ierr = KSPUnwindPreconditioner(ksp,x,vtmp);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);
}
PetscErrorCode KSPSolve_CG(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,stored_max_it,eigs;
PetscScalar dpi = 0.0,a = 1.0,beta,betaold = 1.0,b = 0,*e = 0,*d = 0,delta,dpiold;
PetscReal dp = 0.0;
Vec X,B,Z,R,P,S,W;
KSP_CG *cg;
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);
cg = (KSP_CG*)ksp->data;
eigs = ksp->calc_sings;
stored_max_it = ksp->max_it;
X = ksp->vec_sol;
B = ksp->vec_rhs;
R = ksp->work[0];
Z = ksp->work[1];
P = ksp->work[2];
if (cg->singlereduction) {
S = ksp->work[3];
W = ksp->work[4];
} else {
S = 0; /* unused */
W = Z;
}
#define VecXDot(x,y,a) (((cg->type) == (KSP_CG_HERMITIAN)) ? VecDot(x,y,a) : VecTDot(x,y,a))
if (eigs) {e = cg->e; d = cg->d; e[0] = 0.0; }
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) */
}
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
ierr = VecNorm(Z,NORM_2,&dp);CHKERRQ(ierr); /* dp <- z'*z = e'*A'*B'*B*A'*e' */
break;
case KSP_NORM_UNPRECONDITIONED:
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r = e'*A'*A*e */
break;
case KSP_NORM_NATURAL:
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
ierr = VecXDot(Z,S,&delta);CHKERRQ(ierr);
}
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- z'*r */
if (PetscIsInfOrNanScalar(beta)) {
if (ksp->errorifnotconverged) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"KSPSolve has not converged due to Nan or Inf inner product");
else {
ksp->reason = KSP_DIVERGED_NANORINF;
PetscFunctionReturn(0);
}
}
dp = PetscSqrtReal(PetscAbsScalar(beta)); /* 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);
if (ksp->normtype != KSP_NORM_PRECONDITIONED && (ksp->normtype != KSP_NORM_NATURAL)) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
}
if (ksp->normtype != KSP_NORM_NATURAL) {
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
ierr = VecXDot(Z,S,&delta);CHKERRQ(ierr);
}
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- z'*r */
if (PetscIsInfOrNanScalar(beta)) {
if (ksp->errorifnotconverged) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"KSPSolve has not converged due to Nan or Inf inner product");
else {
ksp->reason = KSP_DIVERGED_NANORINF;
PetscFunctionReturn(0);
}
}
}
i = 0;
do {
ksp->its = i+1;
if (beta == 0.0) {
ksp->reason = KSP_CONVERGED_ATOL;
ierr = PetscInfo(ksp,"converged due to beta = 0\n");CHKERRQ(ierr);
break;
#if !defined(PETSC_USE_COMPLEX)
} else if ((i > 0) && (beta*betaold < 0.0)) {
ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
ierr = PetscInfo(ksp,"diverging due to indefinite preconditioner\n");CHKERRQ(ierr);
break;
#endif
}
if (!i) {
ierr = VecCopy(Z,P);CHKERRQ(ierr); /* p <- z */
b = 0.0;
} else {
b = beta/betaold;
if (eigs) {
if (ksp->max_it != stored_max_it) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Can not change maxit AND calculate eigenvalues");
e[i] = PetscSqrtReal(PetscAbsScalar(b))/a;
}
ierr = VecAYPX(P,b,Z);CHKERRQ(ierr); /* p <- z + b* p */
}
dpiold = dpi;
if (!cg->singlereduction || !i) {
ierr = KSP_MatMult(ksp,Amat,P,W);CHKERRQ(ierr); /* w <- Ap */
ierr = VecXDot(P,W,&dpi);CHKERRQ(ierr); /* dpi <- p'w */
} else {
ierr = VecAYPX(W,beta/betaold,S);CHKERRQ(ierr); /* w <- Ap */
dpi = delta - beta*beta*dpiold/(betaold*betaold); /* dpi <- p'w */
}
betaold = beta;
if (PetscIsInfOrNanScalar(dpi)) {
if (ksp->errorifnotconverged) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"KSPSolve has not converged due to Nan or Inf inner product");
else {
ksp->reason = KSP_DIVERGED_NANORINF;
PetscFunctionReturn(0);
}
}
if ((dpi == 0.0) || ((i > 0) && (PetscRealPart(dpi*dpiold) <= 0.0))) {
ksp->reason = KSP_DIVERGED_INDEFINITE_MAT;
ierr = PetscInfo(ksp,"diverging due to indefinite or negative definite matrix\n");CHKERRQ(ierr);
break;
}
a = beta/dpi; /* a = beta/p'w */
if (eigs) d[i] = PetscSqrtReal(PetscAbsScalar(b))*e[i] + 1.0/a;
ierr = VecAXPY(X,a,P);CHKERRQ(ierr); /* x <- x + ap */
ierr = VecAXPY(R,-a,W);CHKERRQ(ierr); /* r <- r - aw */
if (ksp->normtype == KSP_NORM_PRECONDITIONED && ksp->chknorm < i+2) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
}
ierr = VecNorm(Z,NORM_2,&dp);CHKERRQ(ierr); /* dp <- z'*z */
} else if (ksp->normtype == KSP_NORM_UNPRECONDITIONED && ksp->chknorm < i+2) {
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r */
} else if (ksp->normtype == KSP_NORM_NATURAL) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
if (cg->singlereduction) {
PetscScalar tmp[2];
Vec vecs[2];
vecs[0] = S; vecs[1] = R;
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
ierr = VecMDot(Z,2,vecs,tmp);CHKERRQ(ierr);
delta = tmp[0]; beta = tmp[1];
} else {
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- r'*z */
}
if (PetscIsInfOrNanScalar(beta)) {
if (ksp->errorifnotconverged) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"KSPSolve has not converged due to Nan or Inf inner product");
else {
ksp->reason = KSP_DIVERGED_NANORINF;
PetscFunctionReturn(0);
}
}
dp = PetscSqrtReal(PetscAbsScalar(beta));
} else {
dp = 0.0;
}
ksp->rnorm = dp;
CHKERRQ(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;
if ((ksp->normtype != KSP_NORM_PRECONDITIONED && (ksp->normtype != KSP_NORM_NATURAL)) || (ksp->chknorm >= i+2)) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
}
}
if ((ksp->normtype != KSP_NORM_NATURAL) || (ksp->chknorm >= i+2)) {
if (cg->singlereduction) {
PetscScalar tmp[2];
Vec vecs[2];
vecs[0] = S; vecs[1] = R;
ierr = VecMDot(Z,2,vecs,tmp);CHKERRQ(ierr);
delta = tmp[0]; beta = tmp[1];
} else {
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- z'*r */
}
if (PetscIsInfOrNanScalar(beta)) {
if (ksp->errorifnotconverged) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_NOT_CONVERGED,"KSPSolve has not converged due to Nan or Inf inner product");
else {
ksp->reason = KSP_DIVERGED_NANORINF;
PetscFunctionReturn(0);
}
}
}
i++;
} while (i<ksp->max_it);
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_FBCGSR(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,j,N;
PetscScalar tau,sigma,alpha,omega,beta;
PetscReal rho;
PetscScalar xi1,xi2,xi3,xi4;
Vec X,B,P,P2,RP,R,V,S,T,S2;
PetscScalar *PETSC_RESTRICT rp, *PETSC_RESTRICT r, *PETSC_RESTRICT p;
PetscScalar *PETSC_RESTRICT v, *PETSC_RESTRICT s, *PETSC_RESTRICT t, *PETSC_RESTRICT s2;
PetscScalar insums[4],outsums[4];
KSP_BCGS *bcgs = (KSP_BCGS*)ksp->data;
PC pc;
PetscFunctionBegin;
if (!ksp->vec_rhs->petscnative) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Only coded for PETSc vectors");
ierr = VecGetLocalSize(ksp->vec_sol,&N);
CHKERRQ(ierr);
X = ksp->vec_sol;
B = ksp->vec_rhs;
P2 = ksp->work[0];
/* The followings are involved in modified inner product calculations and vector updates */
RP = ksp->work[1];
ierr = VecGetArray(RP,(PetscScalar**)&rp);
CHKERRQ(ierr);
ierr = VecRestoreArray(RP,NULL);
CHKERRQ(ierr);
R = ksp->work[2];
ierr = VecGetArray(R,(PetscScalar**)&r);
CHKERRQ(ierr);
ierr = VecRestoreArray(R,NULL);
CHKERRQ(ierr);
P = ksp->work[3];
ierr = VecGetArray(P,(PetscScalar**)&p);
CHKERRQ(ierr);
ierr = VecRestoreArray(P,NULL);
CHKERRQ(ierr);
V = ksp->work[4];
ierr = VecGetArray(V,(PetscScalar**)&v);
CHKERRQ(ierr);
ierr = VecRestoreArray(V,NULL);
CHKERRQ(ierr);
S = ksp->work[5];
ierr = VecGetArray(S,(PetscScalar**)&s);
CHKERRQ(ierr);
ierr = VecRestoreArray(S,NULL);
CHKERRQ(ierr);
T = ksp->work[6];
ierr = VecGetArray(T,(PetscScalar**)&t);
CHKERRQ(ierr);
ierr = VecRestoreArray(T,NULL);
CHKERRQ(ierr);
S2 = ksp->work[7];
ierr = VecGetArray(S2,(PetscScalar**)&s2);
CHKERRQ(ierr);
ierr = VecRestoreArray(S2,NULL);
CHKERRQ(ierr);
/* Only supports right preconditioning */
if (ksp->pc_side != PC_RIGHT) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"KSP fbcgsr does not support %s",PCSides[ksp->pc_side]);
if (!ksp->guess_zero) {
if (!bcgs->guess) {
ierr = VecDuplicate(X,&bcgs->guess);
CHKERRQ(ierr);
}
ierr = VecCopy(X,bcgs->guess);
CHKERRQ(ierr);
} else {
ierr = VecSet(X,0.0);
CHKERRQ(ierr);
}
/* Compute initial residual */
ierr = KSPGetPC(ksp,&pc);
CHKERRQ(ierr);
ierr = PCSetUp(pc);
CHKERRQ(ierr);
if (!ksp->guess_zero) {
ierr = MatMult(pc->mat,X,P2);
CHKERRQ(ierr); /* P2 is used as temporary storage */
ierr = VecCopy(B,R);
CHKERRQ(ierr);
ierr = VecAXPY(R,-1.0,P2);
CHKERRQ(ierr);
} else {
ierr = VecCopy(B,R);
CHKERRQ(ierr);
}
/* Test for nothing to do */
if (ksp->normtype != KSP_NORM_NONE) {
ierr = VecNorm(R,NORM_2,&rho);
CHKERRQ(ierr);
}
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);
CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = rho;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);
CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rho);
CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,rho);
CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,0,rho,&ksp->reason,ksp->cnvP);
CHKERRQ(ierr);
if (ksp->reason) PetscFunctionReturn(0);
/* Initialize iterates */
ierr = VecCopy(R,RP);
CHKERRQ(ierr); /* rp <- r */
ierr = VecCopy(R,P);
CHKERRQ(ierr); /* p <- r */
/* Big loop */
for (i=0; i<ksp->max_it; i++) {
/* matmult and pc */
ierr = PCApply(pc,P,P2);
CHKERRQ(ierr); /* p2 <- K p */
ierr = MatMult(pc->mat,P2,V);
CHKERRQ(ierr); /* v <- A p2 */
/* inner prodcuts */
if (i==0) {
tau = rho*rho;
ierr = VecDot(V,RP,&sigma);
CHKERRQ(ierr); /* sigma <- (v,rp) */
} else {
ierr = PetscLogEventBegin(VEC_ReduceArithmetic,0,0,0,0);
CHKERRQ(ierr);
tau = sigma = 0.0;
for (j=0; j<N; j++) {
tau += r[j]*rp[j]; /* tau <- (r,rp) */
sigma += v[j]*rp[j]; /* sigma <- (v,rp) */
}
PetscLogFlops(4.0*N);
ierr = PetscLogEventEnd(VEC_ReduceArithmetic,0,0,0,0);
CHKERRQ(ierr);
insums[0] = tau;
insums[1] = sigma;
ierr = PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
CHKERRQ(ierr);
ierr = MPI_Allreduce(insums,outsums,2,MPIU_SCALAR,MPIU_SUM,PetscObjectComm((PetscObject)ksp));
CHKERRQ(ierr);
ierr = PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
CHKERRQ(ierr);
tau = outsums[0];
sigma = outsums[1];
}
/* scalar update */
alpha = tau / sigma;
/* vector update */
ierr = VecWAXPY(S,-alpha,V,R);
CHKERRQ(ierr); /* s <- r - alpha v */
/* matmult and pc */
ierr = PCApply(pc,S,S2);
CHKERRQ(ierr); /* s2 <- K s */
ierr = MatMult(pc->mat,S2,T);
CHKERRQ(ierr); /* t <- A s2 */
/* inner prodcuts */
ierr = PetscLogEventBegin(VEC_ReduceArithmetic,0,0,0,0);
CHKERRQ(ierr);
xi1 = xi2 = xi3 = xi4 = 0.0;
for (j=0; j<N; j++) {
xi1 += s[j]*s[j]; /* xi1 <- (s,s) */
xi2 += t[j]*s[j]; /* xi2 <- (t,s) */
xi3 += t[j]*t[j]; /* xi3 <- (t,t) */
xi4 += t[j]*rp[j]; /* xi4 <- (t,rp) */
}
PetscLogFlops(8.0*N);
ierr = PetscLogEventEnd(VEC_ReduceArithmetic,0,0,0,0);
CHKERRQ(ierr);
insums[0] = xi1;
insums[1] = xi2;
insums[2] = xi3;
insums[3] = xi4;
ierr = PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
CHKERRQ(ierr);
ierr = MPI_Allreduce(insums,outsums,4,MPIU_SCALAR,MPIU_SUM,PetscObjectComm((PetscObject)ksp));
CHKERRQ(ierr);
ierr = PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
CHKERRQ(ierr);
xi1 = outsums[0];
xi2 = outsums[1];
xi3 = outsums[2];
xi4 = outsums[3];
/* test denominator */
if (xi3 == 0.0) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_PLIB,"Divide by zero");
if (sigma == 0.0) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_PLIB,"Divide by zero");
/* scalar updates */
omega = xi2 / xi3;
beta = -xi4 / sigma;
rho = PetscSqrtReal(PetscAbsScalar(xi1 - omega * xi2)); /* residual norm */
/* vector updates */
ierr = VecAXPBYPCZ(X,alpha,omega,1.0,P2,S2);
CHKERRQ(ierr); /* x <- alpha * p2 + omega * s2 + x */
/* convergence test */
ierr = PetscObjectSAWsTakeAccess((PetscObject)ksp);
CHKERRQ(ierr);
ksp->its++;
ksp->rnorm = rho;
ierr = PetscObjectSAWsGrantAccess((PetscObject)ksp);
CHKERRQ(ierr);
ierr = KSPLogResidualHistory(ksp,rho);
CHKERRQ(ierr);
ierr = KSPMonitor(ksp,i+1,rho);
CHKERRQ(ierr);
ierr = (*ksp->converged)(ksp,i+1,rho,&ksp->reason,ksp->cnvP);
CHKERRQ(ierr);
if (ksp->reason) break;
/* vector updates */
ierr = PetscLogEventBegin(VEC_Ops,0,0,0,0);
CHKERRQ(ierr);
for (j=0; j<N; j++) {
r[j] = s[j] - omega * t[j]; /* r <- s - omega t */
p[j] = r[j] + beta * (p[j] - omega * v[j]); /* p <- r + beta * (p - omega v) */
}
PetscLogFlops(6.0*N);
ierr = PetscLogEventEnd(VEC_Ops,0,0,0,0);
CHKERRQ(ierr);
}
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
static PetscErrorCode KSPSolve_PIPEFCG(KSP ksp)
{
PetscErrorCode ierr;
KSP_PIPEFCG *pipefcg;
PetscScalar gamma;
PetscReal dp=0.0;
Vec B,R,Z,X;
Mat Amat,Pmat;
#define VecXDot(x,y,a) (((pipefcg->type) == (KSP_CG_HERMITIAN)) ? VecDot (x,y,a) : VecTDot (x,y,a))
PetscFunctionBegin;
ierr = PetscCitationsRegister(citation,&cited);CHKERRQ(ierr);
pipefcg = (KSP_PIPEFCG*)ksp->data;
X = ksp->vec_sol;
B = ksp->vec_rhs;
R = ksp->work[0];
Z = ksp->work[1];
ierr = PCGetOperators(ksp->pc,&Amat,&Pmat);CHKERRQ(ierr);
/* Compute initial residual needed for convergence check*/
ksp->its = 0;
if (!ksp->guess_zero) {
ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr);
ierr = VecAYPX(R,-1.0,B);CHKERRQ(ierr); /* r <- b - Ax */
} else {
ierr = VecCopy(B,R);CHKERRQ(ierr); /* r <- b (x is 0) */
}
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
ierr = VecNorm(Z,NORM_2,&dp);CHKERRQ(ierr); /* dp <- dqrt(z'*z) = sqrt(e'*A'*B'*B*A*e) */
break;
case KSP_NORM_UNPRECONDITIONED:
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- sqrt(r'*r) = sqrt(e'*A'*A*e) */
break;
case KSP_NORM_NATURAL:
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
ierr = VecXDot(Z,R,&gamma);CHKERRQ(ierr);
dp = PetscSqrtReal(PetscAbsScalar(gamma)); /* dp <- sqrt(r'*z) = sqrt(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]);
}
/* Initial Convergence Check */
ierr = KSPLogResidualHistory(ksp,dp);CHKERRQ(ierr);
ierr = KSPMonitor(ksp,0,dp);CHKERRQ(ierr);
ksp->rnorm = dp;
if (ksp->normtype == KSP_NORM_NONE) {
ierr = KSPConvergedSkip (ksp,0,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
} else {
ierr = (*ksp->converged)(ksp,0,dp,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
}
if (ksp->reason) PetscFunctionReturn(0);
do {
/* A cycle is broken only if a norm breakdown occurs. If not the entire solve happens in a single cycle.
This is coded this way to allow both truncation and truncation-restart strategy
(see KSPFCDGetNumOldDirections()) */
ierr = KSPSolve_PIPEFCG_cycle(ksp);CHKERRQ(ierr);
if (ksp->reason) break;
if (pipefcg->norm_breakdown) {
pipefcg->n_restarts++;
pipefcg->norm_breakdown = PETSC_FALSE;
}
} while (ksp->its < ksp->max_it);
if (ksp->its >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_STCG(KSP ksp)
{
#if defined(PETSC_USE_COMPLEX)
SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP, "STCG is not available for complex systems");
#else
KSP_STCG *cg = (KSP_STCG*)ksp->data;
PetscErrorCode ierr;
MatStructure pflag;
Mat Qmat, Mmat;
Vec r, z, p, d;
PC pc;
PetscReal norm_r, norm_d, norm_dp1, norm_p, dMp;
PetscReal alpha, beta, kappa, rz, rzm1;
PetscReal rr, r2, step;
PetscInt max_cg_its;
PetscBool diagonalscale;
/***************************************************************************/
/* Check the arguments and parameters. */
/***************************************************************************/
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);
if (cg->radius < 0.0) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_ARG_OUTOFRANGE, "Input error: radius < 0");
/***************************************************************************/
/* Get the workspace vectors and initialize variables */
/***************************************************************************/
r2 = cg->radius * cg->radius;
r = ksp->work[0];
z = ksp->work[1];
p = ksp->work[2];
d = ksp->vec_sol;
pc = ksp->pc;
ierr = PCGetOperators(pc, &Qmat, &Mmat, &pflag);CHKERRQ(ierr);
ierr = VecGetSize(d, &max_cg_its);CHKERRQ(ierr);
max_cg_its = PetscMin(max_cg_its, ksp->max_it);
ksp->its = 0;
/***************************************************************************/
/* Initialize objective function and direction. */
/***************************************************************************/
cg->o_fcn = 0.0;
ierr = VecSet(d, 0.0);CHKERRQ(ierr); /* d = 0 */
cg->norm_d = 0.0;
/***************************************************************************/
/* Begin the conjugate gradient method. Check the right-hand side for */
/* numerical problems. The check for not-a-number and infinite values */
/* need be performed only once. */
/***************************************************************************/
ierr = VecCopy(ksp->vec_rhs, r);CHKERRQ(ierr); /* r = -grad */
ierr = VecDot(r, r, &rr);CHKERRQ(ierr); /* rr = r^T r */
if (PetscIsInfOrNanScalar(rr)) {
/*************************************************************************/
/* The right-hand side contains not-a-number or an infinite value. */
/* The gradient step does not work; return a zero value for the step. */
/*************************************************************************/
ksp->reason = KSP_DIVERGED_NANORINF;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: bad right-hand side: rr=%g\n", rr);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/***************************************************************************/
/* Check the preconditioner for numerical problems and for positive */
/* definiteness. The check for not-a-number and infinite values need be */
/* performed only once. */
/***************************************************************************/
ierr = KSP_PCApply(ksp, r, z);CHKERRQ(ierr); /* z = inv(M) r */
ierr = VecDot(r, z, &rz);CHKERRQ(ierr); /* rz = r^T inv(M) r */
if (PetscIsInfOrNanScalar(rz)) {
/*************************************************************************/
/* The preconditioner contains not-a-number or an infinite value. */
/* Return the gradient direction intersected with the trust region. */
/*************************************************************************/
ksp->reason = KSP_DIVERGED_NANORINF;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: bad preconditioner: rz=%g\n", rz);CHKERRQ(ierr);
if (cg->radius != 0) {
if (r2 >= rr) {
alpha = 1.0;
cg->norm_d = PetscSqrtReal(rr);
} else {
alpha = PetscSqrtReal(r2 / rr);
cg->norm_d = cg->radius;
}
ierr = VecAXPY(d, alpha, r);CHKERRQ(ierr); /* d = d + alpha r */
/***********************************************************************/
/* Compute objective function. */
/***********************************************************************/
ierr = KSP_MatMult(ksp, Qmat, d, z);CHKERRQ(ierr);
ierr = VecAYPX(z, -0.5, ksp->vec_rhs);CHKERRQ(ierr);
ierr = VecDot(d, z, &cg->o_fcn);CHKERRQ(ierr);
cg->o_fcn = -cg->o_fcn;
++ksp->its;
}
PetscFunctionReturn(0);
}
if (rz < 0.0) {
/*************************************************************************/
/* The preconditioner is indefinite. Because this is the first */
/* and we do not have a direction yet, we use the gradient step. Note */
/* that we cannot use the preconditioned norm when computing the step */
/* because the matrix is indefinite. */
/*************************************************************************/
ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: indefinite preconditioner: rz=%g\n", rz);CHKERRQ(ierr);
if (cg->radius != 0.0) {
if (r2 >= rr) {
alpha = 1.0;
cg->norm_d = PetscSqrtReal(rr);
} else {
alpha = PetscSqrtReal(r2 / rr);
cg->norm_d = cg->radius;
}
ierr = VecAXPY(d, alpha, r);CHKERRQ(ierr); /* d = d + alpha r */
/***********************************************************************/
/* Compute objective function. */
/***********************************************************************/
ierr = KSP_MatMult(ksp, Qmat, d, z);CHKERRQ(ierr);
ierr = VecAYPX(z, -0.5, ksp->vec_rhs);CHKERRQ(ierr);
ierr = VecDot(d, z, &cg->o_fcn);CHKERRQ(ierr);
cg->o_fcn = -cg->o_fcn;
++ksp->its;
}
PetscFunctionReturn(0);
}
/***************************************************************************/
/* As far as we know, the preconditioner is positive semidefinite. */
/* Compute and log the residual. Check convergence because this */
/* initializes things, but do not terminate until at least one conjugate */
/* gradient iteration has been performed. */
/***************************************************************************/
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
ierr = VecNorm(z, NORM_2, &norm_r);CHKERRQ(ierr); /* norm_r = |z| */
break;
case KSP_NORM_UNPRECONDITIONED:
norm_r = PetscSqrtReal(rr); /* norm_r = |r| */
break;
case KSP_NORM_NATURAL:
norm_r = PetscSqrtReal(rz); /* norm_r = |r|_M */
break;
default:
norm_r = 0.0;
break;
}
ierr = KSPLogResidualHistory(ksp, norm_r);CHKERRQ(ierr);
ierr = KSPMonitor(ksp, ksp->its, norm_r);CHKERRQ(ierr);
ksp->rnorm = norm_r;
ierr = (*ksp->converged)(ksp, ksp->its, norm_r, &ksp->reason, ksp->cnvP);CHKERRQ(ierr);
/***************************************************************************/
/* Compute the first direction and update the iteration. */
/***************************************************************************/
ierr = VecCopy(z, p);CHKERRQ(ierr); /* p = z */
ierr = KSP_MatMult(ksp, Qmat, p, z);CHKERRQ(ierr); /* z = Q * p */
++ksp->its;
/***************************************************************************/
/* Check the matrix for numerical problems. */
/***************************************************************************/
ierr = VecDot(p, z, &kappa);CHKERRQ(ierr); /* kappa = p^T Q p */
if (PetscIsInfOrNanScalar(kappa)) {
/*************************************************************************/
/* The matrix produced not-a-number or an infinite value. In this case, */
/* we must stop and use the gradient direction. This condition need */
/* only be checked once. */
/*************************************************************************/
ksp->reason = KSP_DIVERGED_NANORINF;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: bad matrix: kappa=%g\n", kappa);CHKERRQ(ierr);
if (cg->radius) {
if (r2 >= rr) {
alpha = 1.0;
cg->norm_d = PetscSqrtReal(rr);
} else {
alpha = PetscSqrtReal(r2 / rr);
cg->norm_d = cg->radius;
}
ierr = VecAXPY(d, alpha, r);CHKERRQ(ierr); /* d = d + alpha r */
/***********************************************************************/
/* Compute objective function. */
/***********************************************************************/
ierr = KSP_MatMult(ksp, Qmat, d, z);CHKERRQ(ierr);
ierr = VecAYPX(z, -0.5, ksp->vec_rhs);CHKERRQ(ierr);
ierr = VecDot(d, z, &cg->o_fcn);CHKERRQ(ierr);
cg->o_fcn = -cg->o_fcn;
++ksp->its;
}
PetscFunctionReturn(0);
}
/***************************************************************************/
/* Initialize variables for calculating the norm of the direction. */
/***************************************************************************/
dMp = 0.0;
norm_d = 0.0;
switch (cg->dtype) {
case STCG_PRECONDITIONED_DIRECTION:
norm_p = rz;
break;
default:
ierr = VecDot(p, p, &norm_p);CHKERRQ(ierr);
break;
}
/***************************************************************************/
/* Check for negative curvature. */
/***************************************************************************/
if (kappa <= 0.0) {
/*************************************************************************/
/* In this case, the matrix is indefinite and we have encountered a */
/* direction of negative curvature. Because negative curvature occurs */
/* during the first step, we must follow a direction. */
/*************************************************************************/
ksp->reason = KSP_CONVERGED_CG_NEG_CURVE;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: negative curvature: kappa=%g\n", kappa);CHKERRQ(ierr);
if (cg->radius != 0.0 && norm_p > 0.0) {
/***********************************************************************/
/* Follow direction of negative curvature to the boundary of the */
/* trust region. */
/***********************************************************************/
step = PetscSqrtReal(r2 / norm_p);
cg->norm_d = cg->radius;
ierr = VecAXPY(d, step, p);CHKERRQ(ierr); /* d = d + step p */
/***********************************************************************/
/* Update objective function. */
/***********************************************************************/
cg->o_fcn += step * (0.5 * step * kappa - rz);
} else if (cg->radius != 0.0) {
/***********************************************************************/
/* The norm of the preconditioned direction is zero; use the gradient */
/* step. */
/***********************************************************************/
if (r2 >= rr) {
alpha = 1.0;
cg->norm_d = PetscSqrtReal(rr);
} else {
alpha = PetscSqrtReal(r2 / rr);
cg->norm_d = cg->radius;
}
ierr = VecAXPY(d, alpha, r);CHKERRQ(ierr); /* d = d + alpha r */
/***********************************************************************/
/* Compute objective function. */
/***********************************************************************/
ierr = KSP_MatMult(ksp, Qmat, d, z);CHKERRQ(ierr);
ierr = VecAYPX(z, -0.5, ksp->vec_rhs);CHKERRQ(ierr);
ierr = VecDot(d, z, &cg->o_fcn);CHKERRQ(ierr);
cg->o_fcn = -cg->o_fcn;
++ksp->its;
}
PetscFunctionReturn(0);
}
/***************************************************************************/
/* Run the conjugate gradient method until either the problem is solved, */
/* we encounter the boundary of the trust region, or the conjugate */
/* gradient method breaks down. */
/***************************************************************************/
while (1) {
/*************************************************************************/
/* Know that kappa is nonzero, because we have not broken down, so we */
/* can compute the steplength. */
/*************************************************************************/
alpha = rz / kappa;
/*************************************************************************/
/* Compute the steplength and check for intersection with the trust */
/* region. */
/*************************************************************************/
norm_dp1 = norm_d + alpha*(2.0*dMp + alpha*norm_p);
if (cg->radius != 0.0 && norm_dp1 >= r2) {
/***********************************************************************/
/* In this case, the matrix is positive definite as far as we know. */
/* However, the full step goes beyond the trust region. */
/***********************************************************************/
ksp->reason = KSP_CONVERGED_CG_CONSTRAINED;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: constrained step: radius=%g\n", cg->radius);CHKERRQ(ierr);
if (norm_p > 0.0) {
/*********************************************************************/
/* Follow the direction to the boundary of the trust region. */
/*********************************************************************/
step = (PetscSqrtReal(dMp*dMp+norm_p*(r2-norm_d))-dMp)/norm_p;
cg->norm_d = cg->radius;
ierr = VecAXPY(d, step, p);CHKERRQ(ierr); /* d = d + step p */
/*********************************************************************/
/* Update objective function. */
/*********************************************************************/
cg->o_fcn += step * (0.5 * step * kappa - rz);
} else {
/*********************************************************************/
/* The norm of the direction is zero; there is nothing to follow. */
/*********************************************************************/
}
break;
}
/*************************************************************************/
/* Now we can update the direction and residual. */
/*************************************************************************/
ierr = VecAXPY(d, alpha, p);CHKERRQ(ierr); /* d = d + alpha p */
ierr = VecAXPY(r, -alpha, z);CHKERRQ(ierr); /* r = r - alpha Q p */
ierr = KSP_PCApply(ksp, r, z);CHKERRQ(ierr); /* z = inv(M) r */
switch (cg->dtype) {
case STCG_PRECONDITIONED_DIRECTION:
norm_d = norm_dp1;
break;
default:
ierr = VecDot(d, d, &norm_d);CHKERRQ(ierr);
break;
}
cg->norm_d = PetscSqrtReal(norm_d);
/*************************************************************************/
/* Update objective function. */
/*************************************************************************/
cg->o_fcn -= 0.5 * alpha * rz;
/*************************************************************************/
/* Check that the preconditioner appears positive semidefinite. */
/*************************************************************************/
rzm1 = rz;
ierr = VecDot(r, z, &rz);CHKERRQ(ierr); /* rz = r^T z */
if (rz < 0.0) {
/***********************************************************************/
/* The preconditioner is indefinite. */
/***********************************************************************/
ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: cg indefinite preconditioner: rz=%g\n", rz);CHKERRQ(ierr);
break;
}
/*************************************************************************/
/* As far as we know, the preconditioner is positive semidefinite. */
/* Compute the residual and check for convergence. */
/*************************************************************************/
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
ierr = VecNorm(z, NORM_2, &norm_r);CHKERRQ(ierr); /* norm_r = |z| */
break;
case KSP_NORM_UNPRECONDITIONED:
ierr = VecNorm(r, NORM_2, &norm_r);CHKERRQ(ierr); /* norm_r = |r| */
break;
case KSP_NORM_NATURAL:
norm_r = PetscSqrtReal(rz); /* norm_r = |r|_M */
break;
default:
norm_r = 0.0;
break;
}
ierr = KSPLogResidualHistory(ksp, norm_r);CHKERRQ(ierr);
ierr = KSPMonitor(ksp, ksp->its, norm_r);CHKERRQ(ierr);
ksp->rnorm = norm_r;
ierr = (*ksp->converged)(ksp, ksp->its, norm_r, &ksp->reason, ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) {
/***********************************************************************/
/* The method has converged. */
/***********************************************************************/
ierr = PetscInfo2(ksp, "KSPSolve_STCG: truncated step: rnorm=%g, radius=%g\n", norm_r, cg->radius);CHKERRQ(ierr);
break;
}
/*************************************************************************/
/* We have not converged yet. Check for breakdown. */
/*************************************************************************/
beta = rz / rzm1;
if (fabs(beta) <= 0.0) {
/***********************************************************************/
/* Conjugate gradients has broken down. */
/***********************************************************************/
ksp->reason = KSP_DIVERGED_BREAKDOWN;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: breakdown: beta=%g\n", beta);CHKERRQ(ierr);
break;
}
/*************************************************************************/
/* Check iteration limit. */
/*************************************************************************/
if (ksp->its >= max_cg_its) {
ksp->reason = KSP_DIVERGED_ITS;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: iterlim: its=%d\n", ksp->its);CHKERRQ(ierr);
break;
}
/*************************************************************************/
/* Update p and the norms. */
/*************************************************************************/
ierr = VecAYPX(p, beta, z);CHKERRQ(ierr); /* p = z + beta p */
switch (cg->dtype) {
case STCG_PRECONDITIONED_DIRECTION:
dMp = beta*(dMp + alpha*norm_p);
norm_p = beta*(rzm1 + beta*norm_p);
break;
default:
ierr = VecDot(d, p, &dMp);CHKERRQ(ierr);
ierr = VecDot(p, p, &norm_p);CHKERRQ(ierr);
break;
}
/*************************************************************************/
/* Compute the new direction and update the iteration. */
/*************************************************************************/
ierr = KSP_MatMult(ksp, Qmat, p, z);CHKERRQ(ierr); /* z = Q * p */
ierr = VecDot(p, z, &kappa);CHKERRQ(ierr); /* kappa = p^T Q p */
++ksp->its;
/*************************************************************************/
/* Check for negative curvature. */
/*************************************************************************/
if (kappa <= 0.0) {
/***********************************************************************/
/* In this case, the matrix is indefinite and we have encountered */
/* a direction of negative curvature. Follow the direction to the */
/* boundary of the trust region. */
/***********************************************************************/
ksp->reason = KSP_CONVERGED_CG_NEG_CURVE;
ierr = PetscInfo1(ksp, "KSPSolve_STCG: negative curvature: kappa=%g\n", kappa);CHKERRQ(ierr);
if (cg->radius != 0.0 && norm_p > 0.0) {
/*********************************************************************/
/* Follow direction of negative curvature to boundary. */
/*********************************************************************/
step = (PetscSqrtReal(dMp*dMp+norm_p*(r2-norm_d))-dMp)/norm_p;
cg->norm_d = cg->radius;
ierr = VecAXPY(d, step, p);CHKERRQ(ierr); /* d = d + step p */
/*********************************************************************/
/* Update objective function. */
/*********************************************************************/
cg->o_fcn += step * (0.5 * step * kappa - rz);
} else if (cg->radius != 0.0) {
/*********************************************************************/
/* The norm of the direction is zero; there is nothing to follow. */
/*********************************************************************/
}
break;
}
}
PetscFunctionReturn(0);
#endif
}
PetscErrorCode KSPSolve_GCR_cycle( KSP ksp )
{
KSP_GCR *ctx = (KSP_GCR*)ksp->data;
PetscErrorCode ierr;
PetscScalar r_dot_v;
Mat A, B;
PC pc;
Vec s,v,r;
PetscReal norm_r,nrm;
PetscInt k, i, restart;
Vec x;
PetscReal res;
PetscFunctionBegin;
restart = ctx->restart;
ierr = KSPGetPC( ksp, &pc );CHKERRQ(ierr);
ierr = KSPGetOperators( ksp, &A, &B, 0 );CHKERRQ(ierr);
x = ksp->vec_sol;
r = ctx->R;
for ( k=0; k<restart; k++ ) {
v = ctx->VV[k];
s = ctx->SS[k];
if (ctx->modifypc) {
ierr = (*ctx->modifypc)(ksp,ksp->its,ksp->rnorm,ctx->modifypc_ctx);CHKERRQ(ierr);
}
ierr = PCApply( pc, r, s );CHKERRQ(ierr); /* s = B^{-1} r */
ierr = MatMult( A, s, v );CHKERRQ(ierr); /* v = A s */
ierr = VecMDot( v,k, ctx->VV, ctx->val );CHKERRQ(ierr);
for (i=0; i<k; i++) ctx->val[i] = -ctx->val[i];
ierr = VecMAXPY(v,k,ctx->val,ctx->VV);CHKERRQ(ierr); /* v = v - sum_{i=0}^{k-1} alpha_i v_i */
ierr = VecMAXPY(s,k,ctx->val,ctx->SS);CHKERRQ(ierr); /* s = s - sum_{i=0}^{k-1} alpha_i s_i */
ierr = VecDotNorm2(r,v,&r_dot_v,&nrm);CHKERRQ(ierr);
nrm = PetscSqrtReal(nrm);
r_dot_v = r_dot_v/nrm;
ierr = VecScale( v, 1.0/nrm );CHKERRQ(ierr);
ierr = VecScale( s, 1.0/nrm );CHKERRQ(ierr);
ierr = VecAXPY( x, r_dot_v, s );CHKERRQ(ierr);
ierr = VecAXPY( r, -r_dot_v, v );CHKERRQ(ierr);
if (ksp->its > ksp->chknorm ) {
ierr = VecNorm( r, NORM_2, &norm_r );CHKERRQ(ierr);
}
/* update the local counter and the global counter */
ksp->its++;
res = norm_r;
ksp->rnorm = res;
KSPLogResidualHistory(ksp,res);
ierr = KSPMonitor(ksp,ksp->its,res);CHKERRQ(ierr);
if ( ksp->its > ksp->chknorm ) {
ierr = (*ksp->converged)(ksp,ksp->its,res,&ksp->reason,ksp->cnvP);CHKERRQ(ierr);
if (ksp->reason) break;
}
if ( ksp->its >= ksp->max_it ) {
ksp->reason = KSP_CONVERGED_ITS;
break;
}
}
ctx->n_restarts++;
PetscFunctionReturn(0);
}
PetscErrorCode KSPSolve_CG(KSP ksp)
{
PetscErrorCode ierr;
PetscInt i,stored_max_it,eigs;
PetscScalar dpi = 0.0,a = 1.0,beta,betaold = 1.0,b = 0,*e = 0,*d = 0,delta,dpiold;
PetscReal dp = 0.0;
Vec X,B,Z,R,P,S,W;
KSP_CG *cg;
Mat Amat,Pmat;
PetscBool diagonalscale;
/* Dingwen */
PetscInt itv_d, itv_c;
PetscScalar CKSX1,CKSZ1,CKSR1,CKSP1,CKSS1,CKSW1;
PetscScalar CKSX2,CKSZ2,CKSR2,CKSP2,CKSS2,CKSW2;
Vec CKSAmat1;
Vec CKSAmat2;
Vec C1,C2;
PetscScalar d1,d2;
PetscScalar sumX1,sumR1;
PetscScalar sumX2,sumR2;
Vec CKPX,CKPP;
PetscScalar CKPbetaold;
PetscInt CKPi;
PetscBool flag1 = PETSC_TRUE, flag2 = PETSC_TRUE;
PetscInt pos;
PetscScalar v;
VecScatter ctx;
Vec W_SEQ;
PetscScalar *_W;
/* Dingwen */
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);
cg = (KSP_CG*)ksp->data;
eigs = ksp->calc_sings;
stored_max_it = ksp->max_it;
X = ksp->vec_sol;
B = ksp->vec_rhs;
R = ksp->work[0];
Z = ksp->work[1];
P = ksp->work[2];
/* Dingwen */
CKPX = ksp->work[3];
CKPP = ksp->work[4];
CKSAmat1 = ksp->work[5];
CKSAmat2 = ksp->work[6];
C1 = ksp->work[7];
C2 = ksp->work[8];
/* Dingwen */
/* Dingwen */
int rank; /* Get MPI variables */
MPI_Comm_rank (MPI_COMM_WORLD,&rank);
/* Dingwen */
#define VecXDot(x,y,a) (((cg->type) == (KSP_CG_HERMITIAN)) ? VecDot(x,y,a) : VecTDot(x,y,a))
if (cg->singlereduction) {
S = ksp->work[9];
W = ksp->work[10];
} else {
S = 0; /* unused */
W = Z;
}
if (eigs) {e = cg->e; d = cg->d; e[0] = 0.0; }
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) */
}
/* Dingwen */
/* checksum coefficients initialization */
PetscInt size;
ierr = VecGetSize(B,&size);
for (i=0; i<size; i++)
{
v = 1.0;
ierr = VecSetValues(C1,1,&i,&v,INSERT_VALUES);CHKERRQ(ierr);
v = i;
ierr = VecSetValues(C2,1,&i,&v,INSERT_VALUES);CHKERRQ(ierr);
}
d1 = 1.0;
d2 = 2.0;
/* Dingwen */
switch (ksp->normtype) {
case KSP_NORM_PRECONDITIONED:
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
ierr = VecNorm(Z,NORM_2,&dp);CHKERRQ(ierr); /* dp <- z'*z = e'*A'*B'*B*A'*e' */
break;
case KSP_NORM_UNPRECONDITIONED:
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r = e'*A'*A*e */
break;
case KSP_NORM_NATURAL:
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
ierr = VecXDot(Z,S,&delta);CHKERRQ(ierr);
/* Dingwen */
ierr = VecXDot(C1,S,&CKSS1);CHKERRQ(ierr); /* Compute the initial checksum1(S) */
ierr = VecXDot(C2,S,&CKSS2);CHKERRQ(ierr); /* Compute the initial checksum2(S) */
/* Dingwen */
}
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- z'*r */
KSPCheckDot(ksp,beta);
dp = PetscSqrtReal(PetscAbsScalar(beta)); /* 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);
if (ksp->normtype != KSP_NORM_PRECONDITIONED && (ksp->normtype != KSP_NORM_NATURAL)) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
}
if (ksp->normtype != KSP_NORM_NATURAL) {
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
ierr = VecXDot(Z,S,&delta);CHKERRQ(ierr);
}
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- z'*r */
KSPCheckDot(ksp,beta);
}
/* Dingwen */
/* Checksum Initialization */
ierr = VecXDot(C1,X,&CKSX1);CHKERRQ(ierr); /* Compute the initial checksum1(X) */
ierr = VecXDot(C1,W,&CKSW1);CHKERRQ(ierr); /* Compute the initial checksum1(W) */
ierr = VecXDot(C1,R,&CKSR1);CHKERRQ(ierr); /* Compute the initial checksum1(R) */
ierr = VecXDot(C1,Z,&CKSZ1);CHKERRQ(ierr); /* Compute the initial checksum1(Z) */
ierr = VecXDot(C2,X,&CKSX2);CHKERRQ(ierr); /* Compute the initial checksum2(X) */
ierr = VecXDot(C2,W,&CKSW2);CHKERRQ(ierr); /* Compute the initial checksum2(W) */
ierr = VecXDot(C2,R,&CKSR2);CHKERRQ(ierr); /* Compute the initial checksum2(R) */
ierr = VecXDot(C2,Z,&CKSZ2);CHKERRQ(ierr); /* Compute the initial checksum2(Z) */
ierr = KSP_MatMultTranspose(ksp,Amat,C1,CKSAmat1);CHKERRQ(ierr);
ierr = VecAXPY(CKSAmat1,-d1,C1);CHKERRQ(ierr);
ierr = VecAXPY(CKSAmat1,-d2,C2);CHKERRQ(ierr); /* Compute the initial checksum1(A) */
ierr = KSP_MatMultTranspose(ksp,Amat,C2,CKSAmat2);CHKERRQ(ierr);
ierr = VecAXPY(CKSAmat2,-d2,C1);CHKERRQ(ierr);
ierr = VecAXPY(CKSAmat2,-d1,C2);CHKERRQ(ierr); /* Compute the initial checksum2(A) */
itv_c = 2;
itv_d = 10;
/* Dingwen */
i = 0;
do {
/* Dingwen */
if ((i>0) && (i%itv_d == 0))
{
ierr = VecXDot(C1,X,&sumX1);CHKERRQ(ierr);
ierr = VecXDot(C1,R,&sumR1);CHKERRQ(ierr);
if ((PetscAbsScalar(sumX1-CKSX1) > 1.0e-6) || (PetscAbsScalar(sumR1-CKSR1) > 1.0e-6))
{
/* Rollback and Recovery */
if (rank==0) printf ("Recovery start...\n");
if (rank==0) printf ("Rollback from iteration-%d to iteration-%d\n",i,CKPi);
betaold = CKPbetaold; /* Recovery scalar betaold by checkpoint*/
i = CKPi; /* Recovery integer i by checkpoint */
ierr = VecCopy(CKPP,P);CHKERRQ(ierr); /* Recovery vector P from checkpoint */
ierr = VecXDot(C1,P,&CKSP1);CHKERRQ(ierr); /* Recovery checksum1(P) by P */
ierr = VecXDot(C2,P,&CKSP2);CHKERRQ(ierr); /* Recovery checksum2(P) by P */
ierr = KSP_MatMult(ksp,Amat,P,W);CHKERRQ(ierr); /* Recovery vector W by P */
ierr = VecXDot(P,W,&dpi);CHKERRQ(ierr); /* Recovery scalar dpi by P and W */
ierr = VecCopy(CKPX,X);CHKERRQ(ierr); /* Recovery vector X from checkpoint */
ierr = VecXDot(C1,X,&CKSX1);CHKERRQ(ierr); /* Recovery checksum1(X) by X */
ierr = VecXDot(C2,X,&CKSX2);CHKERRQ(ierr); /* Recovery checksum2(X) by X */
ierr = KSP_MatMult(ksp,Amat,X,R);CHKERRQ(ierr); /* Recovery vector R by X */
ierr = VecAYPX(R,-1.0,B);CHKERRQ(ierr);
ierr = VecXDot(C1,R,&CKSR1);CHKERRQ(ierr); /* Recovery checksum1(R) by R */
ierr = VecXDot(C2,R,&CKSR2);CHKERRQ(ierr); /* Recovery checksum2(R) by R */
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* Recovery vector Z by R */
ierr = VecXDot(C1,Z,&CKSZ1);CHKERRQ(ierr); /* Recovery checksum1(Z) by Z */
ierr = VecXDot(C2,Z,&CKSZ2);CHKERRQ(ierr); /* Recovery checksum2(Z) by Z */
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* Recovery scalar beta by Z and R */
if (rank==0) printf ("Recovery end.\n");
}
else if (i%(itv_c*itv_d) == 0)
{
if (rank==0) printf ("Checkpoint iteration-%d\n",i);
ierr = VecCopy(X,CKPX);CHKERRQ(ierr);
ierr = VecCopy(P,CKPP);CHKERRQ(ierr);
CKPbetaold = betaold;
CKPi = i;
}
}
ksp->its = i+1;
if (beta == 0.0) {
ksp->reason = KSP_CONVERGED_ATOL;
ierr = PetscInfo(ksp,"converged due to beta = 0\n");CHKERRQ(ierr);
break;
#if !defined(PETSC_USE_COMPLEX)
} else if ((i > 0) && (beta*betaold < 0.0)) {
ksp->reason = KSP_DIVERGED_INDEFINITE_PC;
ierr = PetscInfo(ksp,"diverging due to indefinite preconditioner\n");CHKERRQ(ierr);
break;
#endif
}
if (!i) {
ierr = VecCopy(Z,P);CHKERRQ(ierr); /* p <- z */
b = 0.0;
/* Dingwen */
ierr = VecXDot(C1,P, &CKSP1);CHKERRQ(ierr); /* Compute the initial checksum1(P) */
ierr = VecXDot(C2,P, &CKSP2);CHKERRQ(ierr); /* Compute the initial checksum2(P) */
/* Dingwen */
} else {
b = beta/betaold;
if (eigs) {
if (ksp->max_it != stored_max_it) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Can not change maxit AND calculate eigenvalues");
e[i] = PetscSqrtReal(PetscAbsScalar(b))/a;
}
ierr = VecAYPX(P,b,Z);CHKERRQ(ierr); /* p <- z + b* p */
/* Dingwen */
CKSP1 = CKSZ1 + b*CKSP1; /* Update checksum1(P) = checksum1(Z) + b*checksum1(P); */
CKSP2 = CKSZ2 + b*CKSP2; /* Update checksum2(P) = checksum2(Z) + b*checksum2(P); */
/* Dingwen */
}
dpiold = dpi;
if (!cg->singlereduction || !i) {
ierr = KSP_MatMult(ksp,Amat,P,W);CHKERRQ(ierr); /* w <- Ap */ /* MVM */
ierr = VecXDot(P,W,&dpi);CHKERRQ(ierr); /* dpi <- p'w */
/* Dingwen */
ierr = VecXDot(CKSAmat1, P, &CKSW1);CHKERRQ(ierr);
CKSW1 = CKSW1 + d1*CKSP1 + d2*CKSP2; /* Update checksum1(W) = checksum1(A)P + d1*checksum1(P) + d2*checksum2(P); */
ierr = VecXDot(CKSAmat2, P, &CKSW2);CHKERRQ(ierr);
CKSW2 = CKSW2 + d2*CKSP1 + d1*CKSP2; /* Update checksum2(W) = checksum2(A)P + d2*checksum1(P) + d1*checksum2(P); */
if((i==41)&&(flag2))
{
pos = 100;
v = 1000;
ierr = VecSetValue(W,pos,v,INSERT_VALUES);CHKERRQ(ierr);
VecAssemblyBegin(W);
VecAssemblyEnd(W);
if (rank==0) printf ("Inject an error in %d-th element of vector W after MVM W=AP at iteration-%d\n", pos,i);
flag2 = PETSC_FALSE;
}
PetscScalar delta1,delta2;
PetscScalar sumW1,sumW2;
ierr = VecXDot(C1,W,&sumW1);CHKERRQ(ierr);
ierr = VecXDot(C2,W,&sumW2);CHKERRQ(ierr);
delta1 = sumW1 - CKSW1;
delta2 = sumW2 - CKSW2;
if (PetscAbsScalar(delta1) > 1.0e-6)
{
VecScatterCreateToAll(W,&ctx,&W_SEQ);
VecScatterBegin(ctx,W,W_SEQ,INSERT_VALUES,SCATTER_FORWARD);
VecScatterEnd(ctx,W,W_SEQ,INSERT_VALUES,SCATTER_FORWARD);
VecGetArray(W_SEQ,&_W);
pos = rint(delta2/delta1);
v = _W[pos];
v = v - delta1;
ierr = VecSetValues(W,1,&pos,&v,INSERT_VALUES);CHKERRQ(ierr);
if (rank==0) printf ("Correct an error of %d-th elements of vector W after MVM W=AP at iteration-%d\n", pos, i);
}
} else {
ierr = VecAYPX(W,beta/betaold,S);CHKERRQ(ierr); /* w <- Ap */
dpi = delta - beta*beta*dpiold/(betaold*betaold); /* dpi <- p'w */
/* Dingwen */
CKSW1 = beta/betaold*CKSW1 + CKSS1; /* Update checksum1(W) = checksum1(S) + beta/betaold*checksum1(W); */
CKSW2 = beta/betaold*CKSW2 + CKSS2; /* Update checksum2(W) = checksum2(S) + beta/betaold*checksum2(W); */
/* Dingwen */
}
betaold = beta;
KSPCheckDot(ksp,beta);
if ((dpi == 0.0) || ((i > 0) && (PetscRealPart(dpi*dpiold) <= 0.0))) {
ksp->reason = KSP_DIVERGED_INDEFINITE_MAT;
ierr = PetscInfo(ksp,"diverging due to indefinite or negative definite matrix\n");CHKERRQ(ierr);
break;
}
a = beta/dpi; /* a = beta/p'w */
if (eigs) d[i] = PetscSqrtReal(PetscAbsScalar(b))*e[i] + 1.0/a;
ierr = VecAXPY(X,a,P);CHKERRQ(ierr); /* x <- x + ap */
/* Dingwen */
CKSX1 = CKSX1 + a*CKSP1; /* Update checksum1(X) = checksum1(X) + a*checksum1(P); */
CKSX2 = CKSX2 + a*CKSP2; /* Update checksum2(X) = checksum2(X) + a*checksum2(P); */
/* Dingwen */
ierr = VecAXPY(R,-a,W);CHKERRQ(ierr); /* r <- r - aw */
/* Dingwen */
CKSR1 = CKSR1 - a*CKSW1; /* Update checksum1(R) = checksum1(R) - a*checksum1(W); */
CKSR2 = CKSR2 - a*CKSW2; /* Update checksum2(R) = checksum2(R) - a*checksum2(W); */
/* Dingwen */
if (ksp->normtype == KSP_NORM_PRECONDITIONED && ksp->chknorm < i+2) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
/* Dingwen */
ierr = VecXDot(C1,Z, &CKSZ1);CHKERRQ(ierr); /* Update checksum1(Z) */
ierr = VecXDot(C2,Z, &CKSZ2);CHKERRQ(ierr); /* Update checksum2(Z) */
/* Dingwen */
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr); /* MVM */
/* Dingwen */
ierr = VecXDot(CKSAmat1, Z, &CKSS1);CHKERRQ(ierr);
CKSS1 = CKSS1 + d1*CKSZ1 + d2*CKSZ2; /* Update checksum1(S) = checksum1(A)Z + d1*chekcsum1(Z) + d2*checksum2(Z); */
ierr = VecXDot(CKSAmat2, Z, &CKSS2);CHKERRQ(ierr);
CKSS2 = CKSS2 + d2*CKSZ1 + d1*CKSZ2; /* Update checksum2(S) = checksum2(A)Z + d2*chekcsum1(Z) + d1*checksum2(Z); */
/* Dingwen */
}
ierr = VecNorm(Z,NORM_2,&dp);CHKERRQ(ierr); /* dp <- z'*z */
} else if (ksp->normtype == KSP_NORM_UNPRECONDITIONED && ksp->chknorm < i+2) {
ierr = VecNorm(R,NORM_2,&dp);CHKERRQ(ierr); /* dp <- r'*r */
} else if (ksp->normtype == KSP_NORM_NATURAL) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
/* Dingwen */
ierr = VecXDot(C1,Z, &CKSZ1);CHKERRQ(ierr); /* Update checksum1(Z) */
ierr = VecXDot(C2,Z, &CKSZ2);CHKERRQ(ierr); /* Update checksum2(Z) */
/* Dingwen */
if (cg->singlereduction) {
PetscScalar tmp[2];
Vec vecs[2];
vecs[0] = S; vecs[1] = R;
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
ierr = VecMDot(Z,2,vecs,tmp);CHKERRQ(ierr);
delta = tmp[0]; beta = tmp[1];
} else {
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- r'*z */
}
KSPCheckDot(ksp,beta);
dp = PetscSqrtReal(PetscAbsScalar(beta));
} else {
dp = 0.0;
}
ksp->rnorm = dp;
CHKERRQ(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;
if ((ksp->normtype != KSP_NORM_PRECONDITIONED && (ksp->normtype != KSP_NORM_NATURAL)) || (ksp->chknorm >= i+2)) {
ierr = KSP_PCApply(ksp,R,Z);CHKERRQ(ierr); /* z <- Br */
/* Dingwen */
ierr = VecXDot(C1,Z, &CKSZ1);CHKERRQ(ierr); /* Update checksum1(Z) */
ierr = VecXDot(C2,Z, &CKSZ2);CHKERRQ(ierr); /* Update checksum2(Z) */
/* Dingwen */
if (cg->singlereduction) {
ierr = KSP_MatMult(ksp,Amat,Z,S);CHKERRQ(ierr);
}
}
if ((ksp->normtype != KSP_NORM_NATURAL) || (ksp->chknorm >= i+2)) {
if (cg->singlereduction) {
PetscScalar tmp[2];
Vec vecs[2];
vecs[0] = S; vecs[1] = R;
ierr = VecMDot(Z,2,vecs,tmp);CHKERRQ(ierr);
delta = tmp[0]; beta = tmp[1];
} else {
ierr = VecXDot(Z,R,&beta);CHKERRQ(ierr); /* beta <- z'*r */
}
KSPCheckDot(ksp,beta);
}
i++;
/* Dingwen */
/* Inject error */
if ((i==50)&&(flag1))
{
pos = 1000;
v = -1;
ierr = VecSetValues(X,1,&pos,&v,INSERT_VALUES);CHKERRQ(ierr);
ierr = VecAssemblyBegin(X);CHKERRQ(ierr);
ierr = VecAssemblyEnd(X);CHKERRQ(ierr);
flag1 = PETSC_FALSE;
if (rank==0)printf ("Inject an error in vector X at the end of iteration-%d\n", i-1);
}
/* Dingwen */
} while (i<ksp->max_it);
/* Dingwen */
ierr = VecXDot(C1,X,&sumX1);CHKERRQ(ierr);
ierr = VecXDot(C1,R,&sumR1);CHKERRQ(ierr);
ierr = VecXDot(C2,X,&sumX2);CHKERRQ(ierr);
ierr = VecXDot(C2,R,&sumR2);CHKERRQ(ierr);
if (rank==0)
{
printf ("sum1 of X = %f\n", sumX1);
printf ("checksum1(X) = %f\n", CKSX1);
printf ("sum2 of X = %f\n", sumX2);
printf ("checksum2(X) = %f\n", CKSX2);
printf ("sum1 of R = %f\n", sumR1);
printf ("checksum1(R) = %f\n", CKSR1);
printf ("sum2 of R = %f\n", sumR2);
printf ("checksum2(R) = %f\n", CKSR2);
}
VecDestroy(&W_SEQ);
VecScatterDestroy(&ctx);
/* Dingwen */
if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
if (eigs) cg->ned = ksp->its;
PetscFunctionReturn(0);
}