xref: /petsc/src/tao/quadratic/impls/gpcg/gpcg.c (revision 03047865b8d8757cf1cf9cda45785c1537b01dc1)
1 #include <petscksp.h>
2 #include <../src/tao/quadratic/impls/gpcg/gpcg.h> /*I "gpcg.h" I*/
3 
4 static PetscErrorCode GPCGGradProjections(Tao tao);
5 static PetscErrorCode GPCGObjectiveAndGradient(TaoLineSearch, Vec, PetscReal *, Vec, void *);
6 
TaoDestroy_GPCG(Tao tao)7 static PetscErrorCode TaoDestroy_GPCG(Tao tao)
8 {
9   TAO_GPCG *gpcg = (TAO_GPCG *)tao->data;
10 
11   /* Free allocated memory in GPCG structure */
12   PetscFunctionBegin;
13   PetscCall(VecDestroy(&gpcg->B));
14   PetscCall(VecDestroy(&gpcg->Work));
15   PetscCall(VecDestroy(&gpcg->X_New));
16   PetscCall(VecDestroy(&gpcg->G_New));
17   PetscCall(VecDestroy(&gpcg->DXFree));
18   PetscCall(VecDestroy(&gpcg->R));
19   PetscCall(VecDestroy(&gpcg->PG));
20   PetscCall(MatDestroy(&gpcg->Hsub));
21   PetscCall(MatDestroy(&gpcg->Hsub_pre));
22   PetscCall(ISDestroy(&gpcg->Free_Local));
23   PetscCall(KSPDestroy(&tao->ksp));
24   PetscCall(PetscFree(tao->data));
25   PetscFunctionReturn(PETSC_SUCCESS);
26 }
27 
TaoSetFromOptions_GPCG(Tao tao,PetscOptionItems PetscOptionsObject)28 static PetscErrorCode TaoSetFromOptions_GPCG(Tao tao, PetscOptionItems PetscOptionsObject)
29 {
30   TAO_GPCG *gpcg = (TAO_GPCG *)tao->data;
31   PetscBool flg;
32 
33   PetscFunctionBegin;
34   PetscOptionsHeadBegin(PetscOptionsObject, "Gradient Projection, Conjugate Gradient method for bound constrained optimization");
35   PetscCall(PetscOptionsInt("-tao_gpcg_maxpgits", "maximum number of gradient projections per GPCG iterate", NULL, gpcg->maxgpits, &gpcg->maxgpits, &flg));
36   PetscOptionsHeadEnd();
37   PetscCall(KSPSetFromOptions(tao->ksp));
38   PetscCall(TaoLineSearchSetFromOptions(tao->linesearch));
39   PetscFunctionReturn(PETSC_SUCCESS);
40 }
41 
TaoView_GPCG(Tao tao,PetscViewer viewer)42 static PetscErrorCode TaoView_GPCG(Tao tao, PetscViewer viewer)
43 {
44   TAO_GPCG *gpcg = (TAO_GPCG *)tao->data;
45   PetscBool isascii;
46 
47   PetscFunctionBegin;
48   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
49   if (isascii) {
50     PetscCall(PetscViewerASCIIPrintf(viewer, "Total PG its: %" PetscInt_FMT ",", gpcg->total_gp_its));
51     PetscCall(PetscViewerASCIIPrintf(viewer, "PG tolerance: %g \n", (double)gpcg->pg_ftol));
52   }
53   PetscCall(TaoLineSearchView(tao->linesearch, viewer));
54   PetscFunctionReturn(PETSC_SUCCESS);
55 }
56 
57 /* GPCGObjectiveAndGradient()
58    Compute f=0.5 * x'Hx + b'x + c
59            g=Hx + b
60 */
GPCGObjectiveAndGradient(TaoLineSearch ls,Vec X,PetscReal * f,Vec G,void * tptr)61 static PetscErrorCode GPCGObjectiveAndGradient(TaoLineSearch ls, Vec X, PetscReal *f, Vec G, void *tptr)
62 {
63   Tao       tao  = (Tao)tptr;
64   TAO_GPCG *gpcg = (TAO_GPCG *)tao->data;
65   PetscReal f1, f2;
66 
67   PetscFunctionBegin;
68   PetscCall(MatMult(tao->hessian, X, G));
69   PetscCall(VecDot(G, X, &f1));
70   PetscCall(VecDot(gpcg->B, X, &f2));
71   PetscCall(VecAXPY(G, 1.0, gpcg->B));
72   *f = f1 / 2.0 + f2 + gpcg->c;
73   PetscFunctionReturn(PETSC_SUCCESS);
74 }
75 
TaoSetup_GPCG(Tao tao)76 static PetscErrorCode TaoSetup_GPCG(Tao tao)
77 {
78   TAO_GPCG *gpcg = (TAO_GPCG *)tao->data;
79 
80   PetscFunctionBegin;
81   /* Allocate some arrays */
82   if (!tao->gradient) PetscCall(VecDuplicate(tao->solution, &tao->gradient));
83   if (!tao->stepdirection) PetscCall(VecDuplicate(tao->solution, &tao->stepdirection));
84 
85   PetscCall(VecDuplicate(tao->solution, &gpcg->B));
86   PetscCall(VecDuplicate(tao->solution, &gpcg->Work));
87   PetscCall(VecDuplicate(tao->solution, &gpcg->X_New));
88   PetscCall(VecDuplicate(tao->solution, &gpcg->G_New));
89   PetscCall(VecDuplicate(tao->solution, &gpcg->DXFree));
90   PetscCall(VecDuplicate(tao->solution, &gpcg->R));
91   PetscCall(VecDuplicate(tao->solution, &gpcg->PG));
92   /*
93     if (gpcg->ksp_type == GPCG_KSP_NASH) {
94         PetscCall(KSPSetType(tao->ksp,KSPNASH));
95       } else if (gpcg->ksp_type == GPCG_KSP_STCG) {
96         PetscCall(KSPSetType(tao->ksp,KSPSTCG));
97       } else {
98         PetscCall(KSPSetType(tao->ksp,KSPGLTR));
99       }
100       if (tao->ksp->ops->setfromoptions) (*tao->ksp->ops->setfromoptions)(tao->ksp);
101 
102     }
103   */
104   PetscFunctionReturn(PETSC_SUCCESS);
105 }
106 
TaoSolve_GPCG(Tao tao)107 static PetscErrorCode TaoSolve_GPCG(Tao tao)
108 {
109   TAO_GPCG                    *gpcg = (TAO_GPCG *)tao->data;
110   PetscInt                     its;
111   PetscReal                    actred, f, f_new, gnorm, gdx, stepsize, xtb;
112   PetscReal                    xtHx;
113   TaoLineSearchConvergedReason ls_status = TAOLINESEARCH_CONTINUE_ITERATING;
114 
115   PetscFunctionBegin;
116   PetscCall(TaoComputeVariableBounds(tao));
117   PetscCall(VecMedian(tao->XL, tao->solution, tao->XU, tao->solution));
118   PetscCall(TaoLineSearchSetVariableBounds(tao->linesearch, tao->XL, tao->XU));
119 
120   /* Using f = .5*x'Hx + x'b + c and g=Hx + b,  compute b,c */
121   PetscCall(TaoComputeHessian(tao, tao->solution, tao->hessian, tao->hessian_pre));
122   PetscCall(TaoComputeObjectiveAndGradient(tao, tao->solution, &f, tao->gradient));
123   PetscCall(VecCopy(tao->gradient, gpcg->B));
124   PetscCall(MatMult(tao->hessian, tao->solution, gpcg->Work));
125   PetscCall(VecDot(gpcg->Work, tao->solution, &xtHx));
126   PetscCall(VecAXPY(gpcg->B, -1.0, gpcg->Work));
127   PetscCall(VecDot(gpcg->B, tao->solution, &xtb));
128   gpcg->c = f - xtHx / 2.0 - xtb;
129   if (gpcg->Free_Local) PetscCall(ISDestroy(&gpcg->Free_Local));
130   PetscCall(VecWhichInactive(tao->XL, tao->solution, tao->gradient, tao->XU, PETSC_TRUE, &gpcg->Free_Local));
131 
132   /* Project the gradient and calculate the norm */
133   PetscCall(VecCopy(tao->gradient, gpcg->G_New));
134   PetscCall(VecBoundGradientProjection(tao->gradient, tao->solution, tao->XL, tao->XU, gpcg->PG));
135   PetscCall(VecNorm(gpcg->PG, NORM_2, &gpcg->gnorm));
136   tao->step = 1.0;
137   gpcg->f   = f;
138 
139   /* Check Stopping Condition      */
140   tao->reason = TAO_CONTINUE_ITERATING;
141   PetscCall(TaoLogConvergenceHistory(tao, f, gpcg->gnorm, 0.0, tao->ksp_its));
142   PetscCall(TaoMonitor(tao, tao->niter, f, gpcg->gnorm, 0.0, tao->step));
143   PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
144 
145   while (tao->reason == TAO_CONTINUE_ITERATING) {
146     /* Call general purpose update function */
147     PetscTryTypeMethod(tao, update, tao->niter, tao->user_update);
148     tao->ksp_its = 0;
149 
150     PetscCall(GPCGGradProjections(tao));
151     PetscCall(ISGetSize(gpcg->Free_Local, &gpcg->n_free));
152 
153     f     = gpcg->f;
154     gnorm = gpcg->gnorm;
155 
156     PetscCall(KSPReset(tao->ksp));
157 
158     if (gpcg->n_free > 0) {
159       /* Create a reduced linear system */
160       PetscCall(VecDestroy(&gpcg->R));
161       PetscCall(VecDestroy(&gpcg->DXFree));
162       PetscCall(TaoVecGetSubVec(tao->gradient, gpcg->Free_Local, tao->subset_type, 0.0, &gpcg->R));
163       PetscCall(VecScale(gpcg->R, -1.0));
164       PetscCall(TaoVecGetSubVec(tao->stepdirection, gpcg->Free_Local, tao->subset_type, 0.0, &gpcg->DXFree));
165       PetscCall(VecSet(gpcg->DXFree, 0.0));
166 
167       PetscCall(TaoMatGetSubMat(tao->hessian, gpcg->Free_Local, gpcg->Work, tao->subset_type, &gpcg->Hsub));
168 
169       if (tao->hessian_pre == tao->hessian) {
170         PetscCall(MatDestroy(&gpcg->Hsub_pre));
171         PetscCall(PetscObjectReference((PetscObject)gpcg->Hsub));
172         gpcg->Hsub_pre = gpcg->Hsub;
173       } else {
174         PetscCall(TaoMatGetSubMat(tao->hessian, gpcg->Free_Local, gpcg->Work, tao->subset_type, &gpcg->Hsub_pre));
175       }
176 
177       PetscCall(KSPReset(tao->ksp));
178       PetscCall(KSPSetOperators(tao->ksp, gpcg->Hsub, gpcg->Hsub_pre));
179 
180       PetscCall(KSPSolve(tao->ksp, gpcg->R, gpcg->DXFree));
181       PetscCall(KSPGetIterationNumber(tao->ksp, &its));
182       tao->ksp_its += its;
183       tao->ksp_tot_its += its;
184       PetscCall(VecSet(tao->stepdirection, 0.0));
185       PetscCall(VecISAXPY(tao->stepdirection, gpcg->Free_Local, 1.0, gpcg->DXFree));
186 
187       PetscCall(VecDot(tao->stepdirection, tao->gradient, &gdx));
188       PetscCall(TaoLineSearchSetInitialStepLength(tao->linesearch, 1.0));
189       f_new = f;
190       PetscCall(TaoLineSearchApply(tao->linesearch, tao->solution, &f_new, tao->gradient, tao->stepdirection, &stepsize, &ls_status));
191 
192       actred = f_new - f;
193 
194       /* Evaluate the function and gradient at the new point */
195       PetscCall(VecBoundGradientProjection(tao->gradient, tao->solution, tao->XL, tao->XU, gpcg->PG));
196       PetscCall(VecNorm(gpcg->PG, NORM_2, &gnorm));
197       f = f_new;
198       PetscCall(ISDestroy(&gpcg->Free_Local));
199       PetscCall(VecWhichInactive(tao->XL, tao->solution, tao->gradient, tao->XU, PETSC_TRUE, &gpcg->Free_Local));
200     } else {
201       actred     = 0;
202       gpcg->step = 1.0;
203       /* if there were no free variables, no cg method */
204     }
205 
206     tao->niter++;
207     gpcg->f      = f;
208     gpcg->gnorm  = gnorm;
209     gpcg->actred = actred;
210     PetscCall(TaoLogConvergenceHistory(tao, f, gpcg->gnorm, 0.0, tao->ksp_its));
211     PetscCall(TaoMonitor(tao, tao->niter, f, gpcg->gnorm, 0.0, tao->step));
212     PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
213     if (tao->reason != TAO_CONTINUE_ITERATING) break;
214   } /* END MAIN LOOP  */
215   PetscFunctionReturn(PETSC_SUCCESS);
216 }
217 
GPCGGradProjections(Tao tao)218 static PetscErrorCode GPCGGradProjections(Tao tao)
219 {
220   TAO_GPCG                    *gpcg = (TAO_GPCG *)tao->data;
221   PetscInt                     i;
222   PetscReal                    actred = -1.0, actred_max = 0.0, gAg, gtg = gpcg->gnorm, alpha;
223   PetscReal                    f_new, gdx, stepsize;
224   Vec                          DX = tao->stepdirection, XL = tao->XL, XU = tao->XU, Work = gpcg->Work;
225   Vec                          X = tao->solution, G = tao->gradient;
226   TaoLineSearchConvergedReason lsflag = TAOLINESEARCH_CONTINUE_ITERATING;
227 
228   /*
229      The free, active, and binding variables should be already identified
230   */
231   PetscFunctionBegin;
232   for (i = 0; i < gpcg->maxgpits; i++) {
233     if (-actred <= (gpcg->pg_ftol) * actred_max) break;
234     PetscCall(VecBoundGradientProjection(G, X, XL, XU, DX));
235     PetscCall(VecScale(DX, -1.0));
236     PetscCall(VecDot(DX, G, &gdx));
237 
238     PetscCall(MatMult(tao->hessian, DX, Work));
239     PetscCall(VecDot(DX, Work, &gAg));
240 
241     gpcg->gp_iterates++;
242     gpcg->total_gp_its++;
243 
244     gtg = -gdx;
245     if (PetscAbsReal(gAg) == 0.0) {
246       alpha = 1.0;
247     } else {
248       alpha = PetscAbsReal(gtg / gAg);
249     }
250     PetscCall(TaoLineSearchSetInitialStepLength(tao->linesearch, alpha));
251     f_new = gpcg->f;
252     PetscCall(TaoLineSearchApply(tao->linesearch, X, &f_new, G, DX, &stepsize, &lsflag));
253 
254     /* Update the iterate */
255     actred     = f_new - gpcg->f;
256     actred_max = PetscMax(actred_max, -(f_new - gpcg->f));
257     gpcg->f    = f_new;
258     PetscCall(ISDestroy(&gpcg->Free_Local));
259     PetscCall(VecWhichInactive(XL, X, tao->gradient, XU, PETSC_TRUE, &gpcg->Free_Local));
260   }
261 
262   gpcg->gnorm = gtg;
263   PetscFunctionReturn(PETSC_SUCCESS);
264 } /* End gradient projections */
265 
TaoComputeDual_GPCG(Tao tao,Vec DXL,Vec DXU)266 static PetscErrorCode TaoComputeDual_GPCG(Tao tao, Vec DXL, Vec DXU)
267 {
268   TAO_GPCG *gpcg = (TAO_GPCG *)tao->data;
269 
270   PetscFunctionBegin;
271   PetscCall(VecBoundGradientProjection(tao->gradient, tao->solution, tao->XL, tao->XU, gpcg->Work));
272   PetscCall(VecCopy(gpcg->Work, DXL));
273   PetscCall(VecAXPY(DXL, -1.0, tao->gradient));
274   PetscCall(VecSet(DXU, 0.0));
275   PetscCall(VecPointwiseMax(DXL, DXL, DXU));
276 
277   PetscCall(VecCopy(tao->gradient, DXU));
278   PetscCall(VecAXPY(DXU, -1.0, gpcg->Work));
279   PetscCall(VecSet(gpcg->Work, 0.0));
280   PetscCall(VecPointwiseMin(DXU, gpcg->Work, DXU));
281   PetscFunctionReturn(PETSC_SUCCESS);
282 }
283 
284 /*MC
285   TAOGPCG - gradient projected conjugate gradient algorithm is an active-set
286         conjugate-gradient based method for bound-constrained minimization
287 
288   Options Database Keys:
289 + -tao_gpcg_maxpgits - maximum number of gradient projections for GPCG iterate
290 - -tao_subset_type - "subvec","mask","matrix-free", strategies for handling active-sets
291 
292   Level: beginner
293 M*/
TaoCreate_GPCG(Tao tao)294 PETSC_EXTERN PetscErrorCode TaoCreate_GPCG(Tao tao)
295 {
296   TAO_GPCG *gpcg;
297 
298   PetscFunctionBegin;
299   tao->ops->setup          = TaoSetup_GPCG;
300   tao->ops->solve          = TaoSolve_GPCG;
301   tao->ops->view           = TaoView_GPCG;
302   tao->ops->setfromoptions = TaoSetFromOptions_GPCG;
303   tao->ops->destroy        = TaoDestroy_GPCG;
304   tao->ops->computedual    = TaoComputeDual_GPCG;
305 
306   PetscCall(PetscNew(&gpcg));
307   tao->data = (void *)gpcg;
308 
309   /* Override default settings (unless already changed) */
310   PetscCall(TaoParametersInitialize(tao));
311   PetscObjectParameterSetDefault(tao, max_it, 500);
312   PetscObjectParameterSetDefault(tao, max_funcs, 100000);
313   PetscObjectParameterSetDefault(tao, gatol, PetscDefined(USE_REAL_SINGLE) ? 1e-6 : 1e-12);
314   PetscObjectParameterSetDefault(tao, grtol, PetscDefined(USE_REAL_SINGLE) ? 1e-6 : 1e-12);
315 
316   /* Initialize pointers and variables */
317   gpcg->n        = 0;
318   gpcg->maxgpits = 8;
319   gpcg->pg_ftol  = 0.1;
320 
321   gpcg->gp_iterates  = 0; /* Cumulative number */
322   gpcg->total_gp_its = 0;
323 
324   /* Initialize pointers and variables */
325   gpcg->n_bind      = 0;
326   gpcg->n_free      = 0;
327   gpcg->n_upper     = 0;
328   gpcg->n_lower     = 0;
329   gpcg->subset_type = TAO_SUBSET_MASK;
330   gpcg->Hsub        = NULL;
331   gpcg->Hsub_pre    = NULL;
332 
333   PetscCall(KSPCreate(((PetscObject)tao)->comm, &tao->ksp));
334   PetscCall(PetscObjectIncrementTabLevel((PetscObject)tao->ksp, (PetscObject)tao, 1));
335   PetscCall(KSPSetOptionsPrefix(tao->ksp, tao->hdr.prefix));
336   PetscCall(KSPSetType(tao->ksp, KSPNASH));
337 
338   PetscCall(TaoLineSearchCreate(((PetscObject)tao)->comm, &tao->linesearch));
339   PetscCall(PetscObjectIncrementTabLevel((PetscObject)tao->linesearch, (PetscObject)tao, 1));
340   PetscCall(TaoLineSearchSetType(tao->linesearch, TAOLINESEARCHGPCG));
341   PetscCall(TaoLineSearchSetObjectiveAndGradientRoutine(tao->linesearch, GPCGObjectiveAndGradient, tao));
342   PetscCall(TaoLineSearchSetOptionsPrefix(tao->linesearch, tao->hdr.prefix));
343   PetscFunctionReturn(PETSC_SUCCESS);
344 }
345