Actual source code: qcg.c

  1: #include <../src/ksp/ksp/impls/qcg/qcgimpl.h>

  3: /*
  4:   KSPQCGQuadraticRoots - Computes the roots of the quadratic,
  5:          ||s + step*p|| - delta = 0
  6:    such that step1 >= 0 >= step2.
  7:    where
  8:       delta:
  9:         On entry delta must contain scalar delta.
 10:         On exit delta is unchanged.
 11:       step1:
 12:         On entry step1 need not be specified.
 13:         On exit step1 contains the non-negative root.
 14:       step2:
 15:         On entry step2 need not be specified.
 16:         On exit step2 contains the non-positive root.
 17:    C code is translated from the Fortran version of the MINPACK-2 Project,
 18:    Argonne National Laboratory, Brett M. Averick and Richard G. Carter.
 19: */
 20: static PetscErrorCode KSPQCGQuadraticRoots(Vec s, Vec p, PetscReal delta, PetscReal *step1, PetscReal *step2)
 21: {
 22:   PetscReal dsq, ptp, pts, rad, sts;

 24:   PetscFunctionBegin;
 25:   PetscCall(VecDotRealPart(p, s, &pts));
 26:   PetscCall(VecDotRealPart(p, p, &ptp));
 27:   PetscCall(VecDotRealPart(s, s, &sts));
 28:   dsq = delta * delta;
 29:   rad = PetscSqrtReal((pts * pts) - ptp * (sts - dsq));
 30:   if (pts > 0.0) {
 31:     *step2 = -(pts + rad) / ptp;
 32:     *step1 = (sts - dsq) / (ptp * *step2);
 33:   } else {
 34:     *step1 = -(pts - rad) / ptp;
 35:     *step2 = (sts - dsq) / (ptp * *step1);
 36:   }
 37:   PetscFunctionReturn(PETSC_SUCCESS);
 38: }

 40: /*@
 41:   KSPQCGSetTrustRegionRadius - Sets the radius of the trust region for `KSPQCG`

 43:   Logically Collective

 45:   Input Parameters:
 46: + ksp   - the iterative context
 47: - delta - the trust region radius (Infinity is the default)

 49:   Options Database Key:
 50: . -ksp_qcg_trustregionradius <delta> - trust region radius

 52:   Level: advanced

 54:   Developer Note:
 55:   `KSPMINRESSetRadius()`, for example, does not have TrustRegion in the name

 57: .seealso: [](ch_ksp), `KSPQCG`, `KSPQCGGetTrialStepNorm()`
 58: @*/
 59: PetscErrorCode KSPQCGSetTrustRegionRadius(KSP ksp, PetscReal delta)
 60: {
 61:   PetscFunctionBegin;
 63:   PetscCheck(delta >= 0.0, PetscObjectComm((PetscObject)ksp), PETSC_ERR_ARG_OUTOFRANGE, "Tolerance must be non-negative");
 64:   PetscTryMethod(ksp, "KSPQCGSetTrustRegionRadius_C", (KSP, PetscReal), (ksp, delta));
 65:   PetscFunctionReturn(PETSC_SUCCESS);
 66: }

 68: /*@
 69:   KSPQCGGetTrialStepNorm - Gets the norm of a trial step vector in `KSPQCG`.  The WCG step may be
 70:   constrained, so this is not necessarily the length of the ultimate step taken in `KSPQCG`.

 72:   Not Collective

 74:   Input Parameter:
 75: . ksp - the iterative context

 77:   Output Parameter:
 78: . tsnorm - the norm

 80:   Level: advanced

 82: .seealso: [](ch_ksp), `KSPQCG`, `KSPQCGSetTrustRegionRadius()`
 83: @*/
 84: PetscErrorCode KSPQCGGetTrialStepNorm(KSP ksp, PetscReal *tsnorm)
 85: {
 86:   PetscFunctionBegin;
 88:   PetscUseMethod(ksp, "KSPQCGGetTrialStepNorm_C", (KSP, PetscReal *), (ksp, tsnorm));
 89:   PetscFunctionReturn(PETSC_SUCCESS);
 90: }

 92: /*@
 93:   KSPQCGGetQuadratic - Gets the value of the quadratic function, evaluated at the new iterate

 95:   Collective

 97:   Input Parameter:
 98: . ksp - the iterative context

100:   Output Parameter:
101: . quadratic - the quadratic function evaluated at the new iterate

103:   Level: advanced

105:   Note:
106:   The quadratic function is

108:   $$
109:   q(s) = g^T * s + 0.5 * s^T * H * s
110:   $$

112:   which satisfies the Euclidean Norm trust region constraint

114:   $$
115:   || D * s || \le delta,
116:   $$

118:   where
119: .vb
120:   delta is the trust region radius,
121:   g is the gradient vector, and
122:   H is Hessian matrix,
123:   D is a scaling matrix.
124: .ve

126: .seealso: [](ch_ksp), `KSPQCG`
127: @*/
128: PetscErrorCode KSPQCGGetQuadratic(KSP ksp, PetscReal *quadratic)
129: {
130:   PetscFunctionBegin;
132:   PetscUseMethod(ksp, "KSPQCGGetQuadratic_C", (KSP, PetscReal *), (ksp, quadratic));
133:   PetscFunctionReturn(PETSC_SUCCESS);
134: }

136: static PetscErrorCode KSPSolve_QCG(KSP ksp)
137: {
138:   /*
139:    Correspondence with documentation above:
140:       B = g = gradient,
141:       X = s = step
142:    Note:  This is not coded correctly for complex arithmetic!
143:  */

145:   KSP_QCG    *pcgP = (KSP_QCG *)ksp->data;
146:   Mat         Amat, Pmat;
147:   Vec         W, WA, WA2, R, P, ASP, BS, X, B;
148:   PetscScalar scal, beta, rntrn, step;
149:   PetscReal   q1, q2, xnorm, step1, step2, rnrm = 0.0, btx, xtax;
150:   PetscReal   ptasp, rtr, wtasp, bstp;
151:   PetscReal   dzero = 0.0, bsnrm = 0.0;
152:   PetscInt    i, maxit;
153:   PC          pc = ksp->pc;
154:   PetscBool   diagonalscale;

156:   PetscFunctionBegin;
157:   PetscCall(PCGetDiagonalScale(ksp->pc, &diagonalscale));
158:   PetscCheck(!diagonalscale, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Krylov method %s does not support diagonal scaling", ((PetscObject)ksp)->type_name);
159:   PetscCheck(!ksp->transpose_solve, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Currently does not support transpose solve");

161:   ksp->its = 0;
162:   maxit    = ksp->max_it;
163:   WA       = ksp->work[0];
164:   R        = ksp->work[1];
165:   P        = ksp->work[2];
166:   ASP      = ksp->work[3];
167:   BS       = ksp->work[4];
168:   W        = ksp->work[5];
169:   WA2      = ksp->work[6];
170:   X        = ksp->vec_sol;
171:   B        = ksp->vec_rhs;

173:   PetscCheck(pcgP->delta > dzero, PetscObjectComm((PetscObject)ksp), PETSC_ERR_ARG_OUTOFRANGE, "Input error: delta <= 0");

175:   /* Initialize variables */
176:   PetscCall(VecSet(W, 0.0)); /* W = 0 */
177:   PetscCall(VecSet(X, 0.0)); /* X = 0 */
178:   PetscCall(PCGetOperators(pc, &Amat, &Pmat));

180:   /* Compute:  BS = D^{-1} B */
181:   PetscCall(PCApplySymmetricLeft(pc, B, BS));

183:   if (ksp->normtype != KSP_NORM_NONE) {
184:     PetscCall(VecNorm(BS, NORM_2, &bsnrm));
185:     KSPCheckNorm(ksp, bsnrm);
186:   }
187:   PetscCall(PetscObjectSAWsTakeAccess((PetscObject)ksp));
188:   ksp->its   = 0;
189:   ksp->rnorm = bsnrm;
190:   PetscCall(PetscObjectSAWsGrantAccess((PetscObject)ksp));
191:   PetscCall(KSPLogResidualHistory(ksp, bsnrm));
192:   PetscCall(KSPMonitor(ksp, 0, bsnrm));
193:   PetscCall((*ksp->converged)(ksp, 0, bsnrm, &ksp->reason, ksp->cnvP));
194:   if (ksp->reason) PetscFunctionReturn(PETSC_SUCCESS);

196:   /* Compute the initial scaled direction and scaled residual */
197:   PetscCall(VecCopy(BS, R));
198:   PetscCall(VecScale(R, -1.0));
199:   PetscCall(VecCopy(R, P));
200:   PetscCall(VecDotRealPart(R, R, &rtr));

202:   for (i = 0; i <= maxit; i++) {
203:     PetscCall(PetscObjectSAWsTakeAccess((PetscObject)ksp));
204:     ksp->its++;
205:     PetscCall(PetscObjectSAWsGrantAccess((PetscObject)ksp));

207:     /* Compute:  asp = D^{-T}*A*D^{-1}*p  */
208:     PetscCall(PCApplySymmetricRight(pc, P, WA));
209:     PetscCall(KSP_MatMult(ksp, Amat, WA, WA2));
210:     PetscCall(PCApplySymmetricLeft(pc, WA2, ASP));

212:     /* Check for negative curvature */
213:     PetscCall(VecDotRealPart(P, ASP, &ptasp));
214:     if (ptasp <= dzero) {
215:       /* Scaled negative curvature direction:  Compute a step so that
216:         ||w + step*p|| = delta and QS(w + step*p) is least */

218:       if (!i) {
219:         PetscCall(VecCopy(P, X));
220:         PetscCall(VecNorm(X, NORM_2, &xnorm));
221:         KSPCheckNorm(ksp, xnorm);
222:         scal = pcgP->delta / xnorm;
223:         PetscCall(VecScale(X, scal));
224:       } else {
225:         /* Compute roots of quadratic */
226:         PetscCall(KSPQCGQuadraticRoots(W, P, pcgP->delta, &step1, &step2));
227:         PetscCall(VecDotRealPart(W, ASP, &wtasp));
228:         PetscCall(VecDotRealPart(BS, P, &bstp));
229:         PetscCall(VecCopy(W, X));
230:         q1 = step1 * (bstp + wtasp + .5 * step1 * ptasp);
231:         q2 = step2 * (bstp + wtasp + .5 * step2 * ptasp);
232:         if (q1 <= q2) {
233:           PetscCall(VecAXPY(X, step1, P));
234:         } else {
235:           PetscCall(VecAXPY(X, step2, P));
236:         }
237:       }
238:       pcgP->ltsnrm = pcgP->delta; /* convergence in direction of */
239:       ksp->reason  = ksp->converged_neg_curve ? KSP_CONVERGED_NEG_CURVE : KSP_DIVERGED_INDEFINITE_MAT;
240:       if (!i) {
241:         PetscCall(PetscInfo(ksp, "negative curvature: delta=%g\n", (double)pcgP->delta));
242:       } else {
243:         PetscCall(PetscInfo(ksp, "negative curvature: step1=%g, step2=%g, delta=%g\n", (double)step1, (double)step2, (double)pcgP->delta));
244:       }

246:     } else {
247:       /* Compute step along p */
248:       step = rtr / ptasp;
249:       PetscCall(VecCopy(W, X));       /*  x = w  */
250:       PetscCall(VecAXPY(X, step, P)); /*  x <- step*p + x  */
251:       PetscCall(VecNorm(X, NORM_2, &pcgP->ltsnrm));
252:       KSPCheckNorm(ksp, pcgP->ltsnrm);

254:       if (pcgP->ltsnrm > pcgP->delta) {
255:         /* Since the trial iterate is outside the trust region,
256:             evaluate a constrained step along p so that
257:                     ||w + step*p|| = delta
258:           The positive step is always better in this case. */
259:         if (!i) {
260:           scal = pcgP->delta / pcgP->ltsnrm;
261:           PetscCall(VecScale(X, scal));
262:         } else {
263:           /* Compute roots of quadratic */
264:           PetscCall(KSPQCGQuadraticRoots(W, P, pcgP->delta, &step1, &step2));
265:           PetscCall(VecCopy(W, X));
266:           PetscCall(VecAXPY(X, step1, P)); /*  x <- step1*p + x  */
267:         }
268:         pcgP->ltsnrm = pcgP->delta;
269:         ksp->reason  = KSP_CONVERGED_STEP_LENGTH; /* convergence along constrained step */
270:         if (!i) {
271:           PetscCall(PetscInfo(ksp, "constrained step: delta=%g\n", (double)pcgP->delta));
272:         } else {
273:           PetscCall(PetscInfo(ksp, "constrained step: step1=%g, step2=%g, delta=%g\n", (double)step1, (double)step2, (double)pcgP->delta));
274:         }

276:       } else {
277:         /* Evaluate the current step */
278:         PetscCall(VecCopy(X, W));          /* update interior iterate */
279:         PetscCall(VecAXPY(R, -step, ASP)); /* r <- -step*asp + r */
280:         if (ksp->normtype != KSP_NORM_NONE) {
281:           PetscCall(VecNorm(R, NORM_2, &rnrm));
282:           KSPCheckNorm(ksp, rnrm);
283:         }
284:         PetscCall(PetscObjectSAWsTakeAccess((PetscObject)ksp));
285:         ksp->rnorm = rnrm;
286:         PetscCall(PetscObjectSAWsGrantAccess((PetscObject)ksp));
287:         PetscCall(KSPLogResidualHistory(ksp, rnrm));
288:         PetscCall(KSPMonitor(ksp, i + 1, rnrm));
289:         PetscCall((*ksp->converged)(ksp, i + 1, rnrm, &ksp->reason, ksp->cnvP));
290:         if (ksp->reason) { /* convergence for */
291:           PetscCall(PetscInfo(ksp, "truncated step: step=%g, rnrm=%g, delta=%g\n", (double)PetscRealPart(step), (double)rnrm, (double)pcgP->delta));
292:         }
293:       }
294:     }
295:     if (ksp->reason) break; /* Convergence has been attained */
296:     else { /* Compute a new AS-orthogonal direction */ PetscCall(VecDot(R, R, &rntrn));
297:       beta = rntrn / rtr;
298:       PetscCall(VecAYPX(P, beta, R)); /*  p <- r + beta*p  */
299:       rtr = PetscRealPart(rntrn);
300:     }
301:   }
302:   if (!ksp->reason) ksp->reason = KSP_DIVERGED_ITS;

304:   /* Unscale x */
305:   PetscCall(VecCopy(X, WA2));
306:   PetscCall(PCApplySymmetricRight(pc, WA2, X));

308:   PetscCall(KSP_MatMult(ksp, Amat, X, WA));
309:   PetscCall(VecDotRealPart(B, X, &btx));
310:   PetscCall(VecDotRealPart(X, WA, &xtax));

312:   pcgP->quadratic = btx + .5 * xtax;
313:   PetscFunctionReturn(PETSC_SUCCESS);
314: }

316: static PetscErrorCode KSPSetUp_QCG(KSP ksp)
317: {
318:   PetscFunctionBegin;
319:   /* Get work vectors from user code */
320:   PetscCall(KSPSetWorkVecs(ksp, 7));
321:   PetscFunctionReturn(PETSC_SUCCESS);
322: }

324: static PetscErrorCode KSPDestroy_QCG(KSP ksp)
325: {
326:   PetscFunctionBegin;
327:   PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPQCGGetQuadratic_C", NULL));
328:   PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPQCGGetTrialStepNorm_C", NULL));
329:   PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPQCGSetTrustRegionRadius_C", NULL));
330:   PetscCall(KSPDestroyDefault(ksp));
331:   PetscFunctionReturn(PETSC_SUCCESS);
332: }

334: static PetscErrorCode KSPQCGSetTrustRegionRadius_QCG(KSP ksp, PetscReal delta)
335: {
336:   KSP_QCG *cgP = (KSP_QCG *)ksp->data;

338:   PetscFunctionBegin;
339:   cgP->delta = delta;
340:   PetscFunctionReturn(PETSC_SUCCESS);
341: }

343: static PetscErrorCode KSPQCGGetTrialStepNorm_QCG(KSP ksp, PetscReal *ltsnrm)
344: {
345:   KSP_QCG *cgP = (KSP_QCG *)ksp->data;

347:   PetscFunctionBegin;
348:   *ltsnrm = cgP->ltsnrm;
349:   PetscFunctionReturn(PETSC_SUCCESS);
350: }

352: static PetscErrorCode KSPQCGGetQuadratic_QCG(KSP ksp, PetscReal *quadratic)
353: {
354:   KSP_QCG *cgP = (KSP_QCG *)ksp->data;

356:   PetscFunctionBegin;
357:   *quadratic = cgP->quadratic;
358:   PetscFunctionReturn(PETSC_SUCCESS);
359: }

361: static PetscErrorCode KSPSetFromOptions_QCG(KSP ksp, PetscOptionItems *PetscOptionsObject)
362: {
363:   PetscReal delta;
364:   KSP_QCG  *cgP = (KSP_QCG *)ksp->data;
365:   PetscBool flg;

367:   PetscFunctionBegin;
368:   PetscOptionsHeadBegin(PetscOptionsObject, "KSP QCG Options");
369:   PetscCall(PetscOptionsReal("-ksp_qcg_trustregionradius", "Trust Region Radius", "KSPQCGSetTrustRegionRadius", cgP->delta, &delta, &flg));
370:   if (flg) PetscCall(KSPQCGSetTrustRegionRadius(ksp, delta));
371:   PetscOptionsHeadEnd();
372:   PetscFunctionReturn(PETSC_SUCCESS);
373: }

375: /*MC
376:    KSPQCG - Code to run conjugate gradient method subject to a constraint on the solution norm  {cite}`steihaug:83`.

378:    Options Database Key:
379: .      -ksp_qcg_trustregionradius <r> - Trust Region Radius

381:    Level: developer

383:    Notes:
384:    This is rarely used directly, ir is used in Trust Region methods for nonlinear equations, `SNESNEWTONTR`

386:    Uses preconditioned conjugate gradient to compute
387:    an approximate minimizer of the quadratic function $ q(s) = g^T * s + .5 * s^T * H * s $   subject to the Euclidean norm trust region constraint
388:    $ || D * s || \le delta$, where
389: .vb
390:      delta is the trust region radius,
391:      g is the gradient vector, and
392:      H is Hessian matrix,
393:      D is a scaling matrix.
394: .ve

396:    `KSPConvergedReason` may include
397: +  `KSP_CONVERGED_NEG_CURVE` - if convergence is reached along a negative curvature direction,
398: -  `KSP_CONVERGED_STEP_LENGTH` - if convergence is reached along a constrained step,

400:   Note:
401:   Allows symmetric preconditioning with the following scaling matrices:
402: .vb
403:       `PCNONE`:   D = Identity matrix
404:       `PCJACOBI`: D = diag [d_1, d_2, ...., d_n], where d_i = sqrt(H[i,i])
405:       `PCICC`:    D = L^T, implemented with forward and backward solves. Here L is an incomplete Cholesky factor of H.
406: .ve

408: .seealso: [](ch_ksp), `KSPNASH`, `KSPGLTR`, `KSPSTCG`, `KSPCreate()`, `KSPSetType()`, `KSPType`, `KSP`, `KSPQCGSetTrustRegionRadius()`
409:           `KSPQCGGetTrialStepNorm()`, `KSPQCGGetQuadratic()`
410: M*/

412: PETSC_EXTERN PetscErrorCode KSPCreate_QCG(KSP ksp)
413: {
414:   KSP_QCG *cgP;

416:   PetscFunctionBegin;
417:   PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_PRECONDITIONED, PC_SYMMETRIC, 3));
418:   PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_NONE, PC_SYMMETRIC, 1));
419:   PetscCall(KSPSetConvergedNegativeCurvature(ksp, PETSC_TRUE));
420:   PetscCall(PetscNew(&cgP));

422:   ksp->data                = (void *)cgP;
423:   ksp->ops->setup          = KSPSetUp_QCG;
424:   ksp->ops->setfromoptions = KSPSetFromOptions_QCG;
425:   ksp->ops->solve          = KSPSolve_QCG;
426:   ksp->ops->destroy        = KSPDestroy_QCG;
427:   ksp->ops->buildsolution  = KSPBuildSolutionDefault;
428:   ksp->ops->buildresidual  = KSPBuildResidualDefault;
429:   ksp->ops->view           = NULL;

431:   PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPQCGGetQuadratic_C", KSPQCGGetQuadratic_QCG));
432:   PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPQCGGetTrialStepNorm_C", KSPQCGGetTrialStepNorm_QCG));
433:   PetscCall(PetscObjectComposeFunction((PetscObject)ksp, "KSPQCGSetTrustRegionRadius_C", KSPQCGSetTrustRegionRadius_QCG));
434:   cgP->delta = PETSC_MAX_REAL; /* default trust region radius is infinite */
435:   PetscFunctionReturn(PETSC_SUCCESS);
436: }