PetscErrorCode DSVectors_NHEP_Refined_Some(DS ds,PetscInt *k,PetscReal *rnorm,PetscBool left)
{
#if defined(SLEPC_MISSING_LAPACK_GESVD)
PetscFunctionBegin;
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"GESVD - Lapack routine is unavailable");
#else
PetscErrorCode ierr;
PetscInt i,j;
PetscBLASInt info,ld,n,n1,lwork,inc=1;
PetscScalar sdummy,done=1.0,zero=0.0;
PetscReal *sigma;
PetscBool iscomplex = PETSC_FALSE;
PetscScalar *A = ds->mat[DS_MAT_A];
PetscScalar *Q = ds->mat[DS_MAT_Q];
PetscScalar *X = ds->mat[left?DS_MAT_Y:DS_MAT_X];
PetscScalar *W;
PetscFunctionBegin;
if (left) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented for left vectors");
ierr = PetscBLASIntCast(ds->n,&n);CHKERRQ(ierr);
ierr = PetscBLASIntCast(ds->ld,&ld);CHKERRQ(ierr);
n1 = n+1;
if ((*k)<n-1 && A[(*k)+1+(*k)*ld]!=0.0) iscomplex = PETSC_TRUE;
if (iscomplex) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented for complex eigenvalues yet");
ierr = DSAllocateWork_Private(ds,5*ld,6*ld,0);CHKERRQ(ierr);
ierr = DSAllocateMat_Private(ds,DS_MAT_W);CHKERRQ(ierr);
W = ds->mat[DS_MAT_W];
lwork = 5*ld;
sigma = ds->rwork+5*ld;
/* build A-w*I in W */
for (j=0;j<n;j++)
for (i=0;i<=n;i++)
W[i+j*ld] = A[i+j*ld];
for (i=0;i<n;i++)
W[i+i*ld] -= A[(*k)+(*k)*ld];
/* compute SVD of W */
#if !defined(PETSC_USE_COMPLEX)
PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("N","O",&n1,&n,W,&ld,sigma,&sdummy,&ld,&sdummy,&ld,ds->work,&lwork,&info));
#else
PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("N","O",&n1,&n,W,&ld,sigma,&sdummy,&ld,&sdummy,&ld,ds->work,&lwork,ds->rwork,&info));
#endif
if (info) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in Lapack xGESVD %d",info);
/* the smallest singular value is the new error estimate */
if (rnorm) *rnorm = sigma[n-1];
/* update vector with right singular vector associated to smallest singular value,
accumulating the transformation matrix Q */
PetscStackCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&done,Q,&ld,W+n-1,&ld,&zero,X+(*k)*ld,&inc));
PetscFunctionReturn(0);
#endif
}
/*
Matrix free operation of 1d Laplacian and Grad for GLL spectral elements
*/
PetscErrorCode MatMult_Laplacian(Mat A,Vec x,Vec y)
{
AppCtx *appctx;
PetscErrorCode ierr;
PetscReal **temp,vv;
PetscInt i,j,xs,xn;
Vec xlocal,ylocal;
const PetscScalar *xl;
PetscScalar *yl;
PetscBLASInt _One = 1,n;
PetscScalar _DOne = 1;
ierr = MatShellGetContext(A,&appctx);CHKERRQ(ierr);
ierr = DMGetLocalVector(appctx->da,&xlocal);CHKERRQ(ierr);
ierr = DMGlobalToLocalBegin(appctx->da,x,INSERT_VALUES,xlocal);CHKERRQ(ierr);
ierr = DMGlobalToLocalEnd(appctx->da,x,INSERT_VALUES,xlocal);CHKERRQ(ierr);
ierr = DMGetLocalVector(appctx->da,&ylocal);CHKERRQ(ierr);
ierr = VecSet(ylocal,0.0);CHKERRQ(ierr);
ierr = PetscGLLElementLaplacianCreate(&appctx->SEMop.gll,&temp);CHKERRQ(ierr);
for (i=0; i<appctx->param.N; i++) {
vv =-appctx->param.mu*2.0/appctx->param.Le;
for (j=0; j<appctx->param.N; j++) temp[i][j]=temp[i][j]*vv;
}
ierr = DMDAVecGetArrayRead(appctx->da,xlocal,(void*)&xl);CHKERRQ(ierr);
ierr = DMDAVecGetArray(appctx->da,ylocal,&yl);CHKERRQ(ierr);
ierr = DMDAGetCorners(appctx->da,&xs,NULL,NULL,&xn,NULL,NULL);CHKERRQ(ierr);
ierr = PetscBLASIntCast(appctx->param.N,&n);CHKERRQ(ierr);
for (j=xs; j<xs+xn; j += appctx->param.N-1) {
PetscStackCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&_DOne,&temp[0][0],&n,&xl[j],&_One,&_DOne,&yl[j],&_One));
}
ierr = DMDAVecRestoreArrayRead(appctx->da,xlocal,(void*)&xl);CHKERRQ(ierr);
ierr = DMDAVecRestoreArray(appctx->da,ylocal,&yl);CHKERRQ(ierr);
ierr = PetscGLLElementLaplacianDestroy(&appctx->SEMop.gll,&temp);CHKERRQ(ierr);
ierr = VecSet(y,0.0);CHKERRQ(ierr);
ierr = DMLocalToGlobalBegin(appctx->da,ylocal,ADD_VALUES,y);CHKERRQ(ierr);
ierr = DMLocalToGlobalEnd(appctx->da,ylocal,ADD_VALUES,y);CHKERRQ(ierr);
ierr = DMRestoreLocalVector(appctx->da,&xlocal);CHKERRQ(ierr);
ierr = DMRestoreLocalVector(appctx->da,&ylocal);CHKERRQ(ierr);
ierr = VecPointwiseDivide(y,y,appctx->SEMop.mass);CHKERRQ(ierr);
return 0;
}
PetscErrorCode DSUpdateExtraRow_NHEP(DS ds)
{
PetscErrorCode ierr;
PetscInt i;
PetscBLASInt n,ld,incx=1;
PetscScalar *A,*Q,*x,*y,one=1.0,zero=0.0;
PetscFunctionBegin;
ierr = PetscBLASIntCast(ds->n,&n);CHKERRQ(ierr);
ierr = PetscBLASIntCast(ds->ld,&ld);CHKERRQ(ierr);
A = ds->mat[DS_MAT_A];
Q = ds->mat[DS_MAT_Q];
ierr = DSAllocateWork_Private(ds,2*ld,0,0);CHKERRQ(ierr);
x = ds->work;
y = ds->work+ld;
for (i=0;i<n;i++) x[i] = PetscConj(A[n+i*ld]);
PetscStackCallBLAS("BLASgemv",BLASgemv_("C",&n,&n,&one,Q,&ld,x,&incx,&zero,y,&incx));
for (i=0;i<n;i++) A[n+i*ld] = PetscConj(y[i]);
ds->k = n;
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);
}
void DenseMatrix<T>::_matvec_blas(T alpha, T beta,
DenseVector<T>& dest,
const DenseVector<T>& arg,
bool trans) const
{
// Ensure that dest and arg sizes are compatible
if (!trans)
{
// dest ~ A * arg
// (mx1) (mxn) * (nx1)
if ((dest.size() != this->m()) || (arg.size() != this->n()))
{
libMesh::out << "Improper input argument sizes!" << std::endl;
libmesh_error();
}
}
else // trans == true
{
// Ensure that dest and arg are proper size
// dest ~ A^T * arg
// (nx1) (nxm) * (mx1)
if ((dest.size() != this->n()) || (arg.size() != this->m()))
{
libMesh::out << "Improper input argument sizes!" << std::endl;
libmesh_error();
}
}
// Calling sequence for dgemv:
//
// dgemv(TRANS,M,N,ALPHA,A,LDA,X,INCX,BETA,Y,INCY)
// TRANS - CHARACTER*1, 't' for transpose, 'n' for non-transpose multiply
// We store everything in row-major order, so pass the transpose flag for
// non-transposed matvecs and the 'n' flag for transposed matvecs
char TRANS[] = "t";
if (trans)
TRANS[0] = 'n';
// M - INTEGER.
// On entry, M specifies the number of rows of the matrix A.
// In C/C++, pass the number of *cols* of A
int M = this->n();
// N - INTEGER.
// On entry, N specifies the number of columns of the matrix A.
// In C/C++, pass the number of *rows* of A
int N = this->m();
// ALPHA - DOUBLE PRECISION.
// The scalar constant passed to this function
// A - DOUBLE PRECISION array of DIMENSION ( LDA, n ).
// Before entry, the leading m by n part of the array A must
// contain the matrix of coefficients.
// The matrix, *this. Note that _matvec_blas is called from
// a const function, vector_mult(), and so we have made this function const
// as well. Since BLAS knows nothing about const, we have to cast it away
// now.
DenseMatrix<T>& a_ref = const_cast< DenseMatrix<T>& > ( *this );
std::vector<T>& a = a_ref.get_values();
// LDA - INTEGER.
// On entry, LDA specifies the first dimension of A as declared
// in the calling (sub) program. LDA must be at least
// max( 1, m ).
int LDA = M;
// X - DOUBLE PRECISION array of DIMENSION at least
// ( 1 + ( n - 1 )*abs( INCX ) ) when TRANS = 'N' or 'n'
// and at least
// ( 1 + ( m - 1 )*abs( INCX ) ) otherwise.
// Before entry, the incremented array X must contain the
// vector x.
// Here, we must cast away the const-ness of "arg" since BLAS knows
// nothing about const
DenseVector<T>& x_ref = const_cast< DenseVector<T>& > ( arg );
std::vector<T>& x = x_ref.get_values();
// INCX - INTEGER.
// On entry, INCX specifies the increment for the elements of
// X. INCX must not be zero.
int INCX = 1;
// BETA - DOUBLE PRECISION.
// On entry, BETA specifies the scalar beta. When BETA is
// supplied as zero then Y need not be set on input.
// The second scalar constant passed to this function
// Y - DOUBLE PRECISION array of DIMENSION at least
// ( 1 + ( m - 1 )*abs( INCY ) ) when TRANS = 'N' or 'n'
// and at least
// ( 1 + ( n - 1 )*abs( INCY ) ) otherwise.
// Before entry with BETA non-zero, the incremented array Y
// must contain the vector y. On exit, Y is overwritten by the
// updated vector y.
// The input vector "dest"
std::vector<T>& y = dest.get_values();
// INCY - INTEGER.
// On entry, INCY specifies the increment for the elements of
// Y. INCY must not be zero.
int INCY = 1;
// Finally, ready to call the BLAS function
BLASgemv_(TRANS, &M, &N, &alpha, &(a[0]), &LDA, &(x[0]), &INCX, &beta, &(y[0]), &INCY);
}
PetscErrorCode DSTranslateHarmonic_NHEP(DS ds,PetscScalar tau,PetscReal beta,PetscBool recover,PetscScalar *gin,PetscReal *gamma)
{
#if defined(PETSC_MISSING_LAPACK_GETRF) || defined(PETSC_MISSING_LAPACK_GETRS)
PetscFunctionBegin;
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"GETRF/GETRS - Lapack routines are unavailable");
#else
PetscErrorCode ierr;
PetscInt i,j;
PetscBLASInt *ipiv,info,n,ld,one=1,ncol;
PetscScalar *A,*B,*Q,*g=gin,*ghat;
PetscScalar done=1.0,dmone=-1.0,dzero=0.0;
PetscReal gnorm;
PetscFunctionBegin;
ierr = PetscBLASIntCast(ds->n,&n);CHKERRQ(ierr);
ierr = PetscBLASIntCast(ds->ld,&ld);CHKERRQ(ierr);
A = ds->mat[DS_MAT_A];
if (!recover) {
ierr = DSAllocateWork_Private(ds,0,0,ld);CHKERRQ(ierr);
ipiv = ds->iwork;
if (!g) {
ierr = DSAllocateWork_Private(ds,ld,0,0);CHKERRQ(ierr);
g = ds->work;
}
/* use workspace matrix W to factor A-tau*eye(n) */
ierr = DSAllocateMat_Private(ds,DS_MAT_W);CHKERRQ(ierr);
B = ds->mat[DS_MAT_W];
ierr = PetscMemcpy(B,A,sizeof(PetscScalar)*ld*ld);CHKERRQ(ierr);
/* Vector g initialy stores b = beta*e_n^T */
ierr = PetscMemzero(g,n*sizeof(PetscScalar));CHKERRQ(ierr);
g[n-1] = beta;
/* g = (A-tau*eye(n))'\b */
for (i=0;i<n;i++)
B[i+i*ld] -= tau;
PetscStackCallBLAS("LAPACKgetrf",LAPACKgetrf_(&n,&n,B,&ld,ipiv,&info));
if (info<0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"Bad argument to LU factorization");
if (info>0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Bad LU factorization");
ierr = PetscLogFlops(2.0*n*n*n/3.0);CHKERRQ(ierr);
PetscStackCallBLAS("LAPACKgetrs",LAPACKgetrs_("C",&n,&one,B,&ld,ipiv,g,&ld,&info));
if (info) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"GETRS - Bad solve");
ierr = PetscLogFlops(2.0*n*n-n);CHKERRQ(ierr);
/* A = A + g*b' */
for (i=0;i<n;i++)
A[i+(n-1)*ld] += g[i]*beta;
} else { /* recover */
PetscValidPointer(g,6);
ierr = DSAllocateWork_Private(ds,ld,0,0);CHKERRQ(ierr);
ghat = ds->work;
Q = ds->mat[DS_MAT_Q];
/* g^ = -Q(:,idx)'*g */
ierr = PetscBLASIntCast(ds->l+ds->k,&ncol);CHKERRQ(ierr);
PetscStackCallBLAS("BLASgemv",BLASgemv_("C",&n,&ncol,&dmone,Q,&ld,g,&one,&dzero,ghat,&one));
/* A = A + g^*b' */
for (i=0;i<ds->l+ds->k;i++)
for (j=ds->l;j<ds->l+ds->k;j++)
A[i+j*ld] += ghat[i]*Q[n-1+j*ld]*beta;
/* g~ = (I-Q(:,idx)*Q(:,idx)')*g = g+Q(:,idx)*g^ */
PetscStackCallBLAS("BLASgemv",BLASgemv_("N",&n,&ncol,&done,Q,&ld,ghat,&one,&done,g,&one));
}
/* Compute gamma factor */
if (gamma) {
gnorm = 0.0;
for (i=0;i<n;i++)
gnorm = gnorm + PetscRealPart(g[i]*PetscConj(g[i]));
*gamma = PetscSqrtReal(1.0+gnorm);
}
PetscFunctionReturn(0);
#endif
}
PetscErrorCode DSVectors_NHEP_Eigen_Some(DS ds,PetscInt *k,PetscReal *rnorm,PetscBool left)
{
#if defined(SLEPC_MISSING_LAPACK_TREVC)
PetscFunctionBegin;
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"TREVC - Lapack routine is unavailable");
#else
PetscErrorCode ierr;
PetscInt i;
PetscBLASInt mm=1,mout,info,ld,n,inc = 1;
PetscScalar tmp,done=1.0,zero=0.0;
PetscReal norm;
PetscBool iscomplex = PETSC_FALSE;
PetscBLASInt *select;
PetscScalar *A = ds->mat[DS_MAT_A];
PetscScalar *Q = ds->mat[DS_MAT_Q];
PetscScalar *X = ds->mat[left?DS_MAT_Y:DS_MAT_X];
PetscScalar *Y;
PetscFunctionBegin;
ierr = PetscBLASIntCast(ds->n,&n);CHKERRQ(ierr);
ierr = PetscBLASIntCast(ds->ld,&ld);CHKERRQ(ierr);
ierr = DSAllocateWork_Private(ds,0,0,ld);CHKERRQ(ierr);
select = ds->iwork;
for (i=0;i<n;i++) select[i] = (PetscBLASInt)PETSC_FALSE;
/* Compute k-th eigenvector Y of A */
Y = X+(*k)*ld;
select[*k] = (PetscBLASInt)PETSC_TRUE;
#if !defined(PETSC_USE_COMPLEX)
if ((*k)<n-1 && A[(*k)+1+(*k)*ld]!=0.0) iscomplex = PETSC_TRUE;
mm = iscomplex? 2: 1;
if (iscomplex) select[(*k)+1] = (PetscBLASInt)PETSC_TRUE;
ierr = DSAllocateWork_Private(ds,3*ld,0,0);CHKERRQ(ierr);
PetscStackCallBLAS("LAPACKtrevc",LAPACKtrevc_(left?"L":"R","S",select,&n,A,&ld,Y,&ld,Y,&ld,&mm,&mout,ds->work,&info));
#else
ierr = DSAllocateWork_Private(ds,2*ld,ld,0);CHKERRQ(ierr);
PetscStackCallBLAS("LAPACKtrevc",LAPACKtrevc_(left?"L":"R","S",select,&n,A,&ld,Y,&ld,Y,&ld,&mm,&mout,ds->work,ds->rwork,&info));
#endif
if (info) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in Lapack xTREVC %d",info);
if (mout != mm) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Inconsistent arguments");
/* accumulate and normalize eigenvectors */
if (ds->state>=DS_STATE_CONDENSED) {
ierr = PetscMemcpy(ds->work,Y,mout*ld*sizeof(PetscScalar));CHKERRQ(ierr);
PetscStackCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&done,Q,&ld,ds->work,&inc,&zero,Y,&inc));
#if !defined(PETSC_USE_COMPLEX)
if (iscomplex) PetscStackCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&done,Q,&ld,ds->work+ld,&inc,&zero,Y+ld,&inc));
#endif
norm = BLASnrm2_(&n,Y,&inc);
#if !defined(PETSC_USE_COMPLEX)
if (iscomplex) {
tmp = BLASnrm2_(&n,Y+ld,&inc);
norm = SlepcAbsEigenvalue(norm,tmp);
}
#endif
tmp = 1.0 / norm;
PetscStackCallBLAS("BLASscal",BLASscal_(&n,&tmp,Y,&inc));
#if !defined(PETSC_USE_COMPLEX)
if (iscomplex) PetscStackCallBLAS("BLASscal",BLASscal_(&n,&tmp,Y+ld,&inc));
#endif
}
/* set output arguments */
if (iscomplex) (*k)++;
if (rnorm) {
if (iscomplex) *rnorm = SlepcAbsEigenvalue(Y[n-1],Y[n-1+ld]);
else *rnorm = PetscAbsScalar(Y[n-1]);
}
PetscFunctionReturn(0);
#endif
}
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, epsilon, abstol;
PetscReal zeta, zeta0, rnmax_computed, rnmax_true, nrm0;
PetscTruth bUpdateX;
PetscTruth bBombed = PETSC_FALSE;
PetscInt maxit;
PetscInt h, i, j, k, vi, ell;
PetscBLASInt ldMZ,bierr;
PetscErrorCode ierr;
PetscFunctionBegin;
if (ksp->normtype == KSP_NORM_NATURAL) SETERRQ(PETSC_ERR_SUP,"Cannot use natural norm with KSPBCGSL");
if (ksp->normtype == KSP_NORM_PRECONDITIONED && ksp->pc_side != PC_LEFT) SETERRQ(PETSC_ERR_SUP,"Use -ksp_norm_type unpreconditioned for right preconditioning and KSPBCGSL");
if (ksp->normtype == KSP_NORM_UNPRECONDITIONED && ksp->pc_side != PC_RIGHT) SETERRQ(PETSC_ERR_SUP,"Use -ksp_norm_type preconditioned for left preconditioning and KSPBCGSL");
/* 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++;
ldMZ = PetscBLASIntCast(ell+1);
/* 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 = PetscObjectTakeAccess(ksp);
CHKERRQ(ierr);
ksp->its = 0;
ksp->rnorm = zeta0;
ierr = PetscObjectGrantAccess(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, &epsilon, &abstol, PETSC_NULL, &maxit);
CHKERRQ(ierr);
for (k=0; k<maxit; k += bcgsl->ell) {
ksp->its = k;
ksp->rnorm = zeta;
KSPLogResidualHistory(ksp, zeta);
KSPMonitor(ksp, ksp->its, zeta);
ierr = (*ksp->converged)(ksp, k, zeta, &ksp->reason, ksp->cnvP);
CHKERRQ(ierr);
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;
bBombed = PETSC_TRUE;
break;
}
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;
bBombed = PETSC_TRUE;
break;
}
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 = PetscObjectTakeAccess(ksp);
CHKERRQ(ierr);
ksp->its = k+j;
ksp->rnorm = nrm0;
ierr = PetscObjectGrantAccess(ksp);
CHKERRQ(ierr);
break;
}
}
if (bBombed==PETSC_TRUE) break;
/* 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 = PetscBLASIntCast(bcgsl->ell);
AY0c[0] = -1;
LAPACKpotrf_("Lower", &bell, &MZa[1+ldMZ], &ldMZ, &bierr);
if (ierr!=0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
bBombed = PETSC_TRUE;
break;
}
ierr = PetscMemcpy(&AY0c[1],&MZb[1],bcgsl->ell*sizeof(PetscScalar));
CHKERRQ(ierr);
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 = PetscBLASIntCast(bcgsl->ell-1);
LAPACKpotrf_("Lower", &neqs, &MZa[1+ldMZ], &ldMZ, &bierr);
if (ierr!=0) {
ksp->reason = KSP_DIVERGED_BREAKDOWN;
bBombed = PETSC_TRUE;
break;
}
ierr = PetscMemcpy(&AY0c[1],&MZb[1],(bcgsl->ell-1)*sizeof(PetscScalar));
CHKERRQ(ierr);
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);
LAPACKpotrs_("Lower", &neqs, &ione, &MZa[1+ldMZ], &ldMZ, &AYlc[1], &ldMZ, &bierr);
AYlc[0] = 0.;
AYlc[bcgsl->ell] = -1;
BLASgemv_("NoTr", &ldMZ, &ldMZ, &aone, MZb, &ldMZ, AY0c, &ione, &azero, AYtc, &ione);
kappa0 = BLASdot_(&ldMZ, AY0c, &ione, AYtc, &ione);
/* round-off can cause negative kappa's */
if (kappa0<0) kappa0 = -kappa0;
kappa0 = sqrt(kappa0);
kappaA = BLASdot_(&ldMZ, AYlc, &ione, AYtc, &ione);
BLASgemv_("noTr", &ldMZ, &ldMZ, &aone, MZb, &ldMZ, AYlc, &ione, &azero, AYtc, &ione);
kappa1 = BLASdot_(&ldMZ, AYlc, &ione, AYtc, &ione);
if (kappa1<0) kappa1 = -kappa1;
kappa1 = sqrt(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;
break;
}
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 = (PetscTruth) (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);
}
