Actual source code: pdipm.c
petsc-3.13.0 2020-03-29
1: #include <petsctaolinesearch.h>
2: #include <../src/tao/constrained/impls/ipm/pdipm.h>
3: #include <petscsnes.h>
5: /*
6: TaoPDIPMEvaluateFunctionsAndJacobians - Evaluate the objective function f, gradient fx, constraints, and all the Jacobians at current vector
8: Collective on tao
10: Input Parameter:
11: + tao - solver context
12: - x - vector at which all objects to be evaluated
14: Level: beginner
16: .seealso: TaoPDIPMUpdateConstraints(), TaoPDIPMSetUpBounds()
17: */
18: PetscErrorCode TaoPDIPMEvaluateFunctionsAndJacobians(Tao tao,Vec x)
19: {
21: TAO_PDIPM *pdipm=(TAO_PDIPM*)tao->data;
24: /* Compute user objective function and gradient */
25: TaoComputeObjectiveAndGradient(tao,x,&pdipm->obj,tao->gradient);
27: /* Equality constraints and Jacobian */
28: if (pdipm->Ng) {
29: TaoComputeEqualityConstraints(tao,x,tao->constraints_equality);
30: TaoComputeJacobianEquality(tao,x,tao->jacobian_equality,tao->jacobian_equality_pre);
31: }
33: /* Inequality constraints and Jacobian */
34: if (pdipm->Nh) {
35: TaoComputeInequalityConstraints(tao,x,tao->constraints_inequality);
36: TaoComputeJacobianInequality(tao,x,tao->jacobian_inequality,tao->jacobian_inequality_pre);
37: }
38: return(0);
39: }
41: /*
42: TaoPDIPMUpdateConstraints - Update the vectors ce and ci at x
44: Collective
46: Input Parameter:
47: + tao - Tao context
48: - x - vector at which constraints to be evaluted
50: Level: beginner
52: .seealso: TaoPDIPMEvaluateFunctionsAndJacobians()
53: */
54: PetscErrorCode TaoPDIPMUpdateConstraints(Tao tao,Vec x)
55: {
56: PetscErrorCode ierr;
57: TAO_PDIPM *pdipm=(TAO_PDIPM*)tao->data;
58: PetscInt i,offset,offset1,k,xstart;
59: PetscScalar *carr;
60: const PetscInt *ubptr,*lbptr,*bxptr,*fxptr;
61: const PetscScalar *xarr,*xuarr,*xlarr,*garr,*harr;
64: VecGetOwnershipRange(x,&xstart,NULL);
66: VecGetArrayRead(x,&xarr);
67: VecGetArrayRead(tao->XU,&xuarr);
68: VecGetArrayRead(tao->XL,&xlarr);
70: /* (1) Update ce vector */
71: VecGetArray(pdipm->ce,&carr);
73: if(pdipm->Ng) {
74: /* (1.a) Inserting updated g(x) */
75: VecGetArrayRead(tao->constraints_equality,&garr);
76: PetscMemcpy(carr,garr,pdipm->ng*sizeof(PetscScalar));
77: VecRestoreArrayRead(tao->constraints_equality,&garr);
78: }
80: /* (1.b) Update xfixed */
81: if (pdipm->Nxfixed) {
82: offset = pdipm->ng;
83: ISGetIndices(pdipm->isxfixed,&fxptr); /* global indices in x */
84: for (k=0;k < pdipm->nxfixed;k++){
85: i = fxptr[k]-xstart;
86: carr[offset + k] = xarr[i] - xuarr[i];
87: }
88: }
89: VecRestoreArray(pdipm->ce,&carr);
91: /* (2) Update ci vector */
92: VecGetArray(pdipm->ci,&carr);
94: if(pdipm->Nh) {
95: /* (2.a) Inserting updated h(x) */
96: VecGetArrayRead(tao->constraints_inequality,&harr);
97: PetscMemcpy(carr,harr,pdipm->nh*sizeof(PetscScalar));
98: VecRestoreArrayRead(tao->constraints_inequality,&harr);
99: }
101: /* (2.b) Update xub */
102: offset = pdipm->nh;
103: if (pdipm->Nxub) {
104: ISGetIndices(pdipm->isxub,&ubptr);
105: for (k=0; k<pdipm->nxub; k++){
106: i = ubptr[k]-xstart;
107: carr[offset + k] = xuarr[i] - xarr[i];
108: }
109: }
111: if (pdipm->Nxlb) {
112: /* (2.c) Update xlb */
113: offset += pdipm->nxub;
114: ISGetIndices(pdipm->isxlb,&lbptr); /* global indices in x */
115: for (k=0; k<pdipm->nxlb; k++){
116: i = lbptr[k]-xstart;
117: carr[offset + k] = xarr[i] - xlarr[i];
118: }
119: }
121: if (pdipm->Nxbox) {
122: /* (2.d) Update xbox */
123: offset += pdipm->nxlb;
124: offset1 = offset + pdipm->nxbox;
125: ISGetIndices(pdipm->isxbox,&bxptr); /* global indices in x */
126: for (k=0; k<pdipm->nxbox; k++){
127: i = bxptr[k]-xstart; /* local indices in x */
128: carr[offset+k] = xuarr[i] - xarr[i];
129: carr[offset1+k] = xarr[i] - xlarr[i];
130: }
131: }
132: VecRestoreArray(pdipm->ci,&carr);
134: /* Restoring Vectors */
135: VecRestoreArrayRead(x,&xarr);
136: VecRestoreArrayRead(tao->XU,&xuarr);
137: VecRestoreArrayRead(tao->XL,&xlarr);
138: return(0);
139: }
141: /*
142: TaoPDIPMSetUpBounds - Create upper and lower bound vectors of x
144: Collective
146: Input Parameter:
147: . tao - holds pdipm and XL & XU
149: Level: beginner
151: .seealso: TaoPDIPMUpdateConstraints
152: */
153: PetscErrorCode TaoPDIPMSetUpBounds(Tao tao)
154: {
155: PetscErrorCode ierr;
156: TAO_PDIPM *pdipm=(TAO_PDIPM*)tao->data;
157: const PetscScalar *xl,*xu;
158: PetscInt n,*ixlb,*ixub,*ixfixed,*ixfree,*ixbox,i,low,high,idx;
159: MPI_Comm comm;
160: PetscInt sendbuf[5],recvbuf[5];
163: /* Creates upper and lower bounds vectors on x, if not created already */
164: TaoComputeVariableBounds(tao);
166: VecGetLocalSize(tao->XL,&n);
167: PetscMalloc5(n,&ixlb,n,&ixub,n,&ixfree,n,&ixfixed,n,&ixbox);
169: VecGetOwnershipRange(tao->XL,&low,&high);
170: VecGetArrayRead(tao->XL,&xl);
171: VecGetArrayRead(tao->XU,&xu);
172: for (i=0; i<n; i++) {
173: idx = low + i;
174: if((PetscRealPart(xl[i]) > PETSC_NINFINITY) && (PetscRealPart(xu[i]) < PETSC_INFINITY)) {
175: if (PetscRealPart(xl[i]) == PetscRealPart(xu[i])) {
176: ixfixed[pdipm->nxfixed++] = idx;
177: } else ixbox[pdipm->nxbox++] = idx;
178: } else {
179: if ((PetscRealPart(xl[i]) > PETSC_NINFINITY) && (PetscRealPart(xu[i]) >= PETSC_INFINITY)) {
180: ixlb[pdipm->nxlb++] = idx;
181: } else if ((PetscRealPart(xl[i]) <= PETSC_NINFINITY) && (PetscRealPart(xu[i]) < PETSC_INFINITY)) {
182: ixub[pdipm->nxlb++] = idx;
183: } else ixfree[pdipm->nxfree++] = idx;
184: }
185: }
186: VecRestoreArrayRead(tao->XL,&xl);
187: VecRestoreArrayRead(tao->XU,&xu);
189: PetscObjectGetComm((PetscObject)tao,&comm);
190: sendbuf[0] = pdipm->nxlb;
191: sendbuf[1] = pdipm->nxub;
192: sendbuf[2] = pdipm->nxfixed;
193: sendbuf[3] = pdipm->nxbox;
194: sendbuf[4] = pdipm->nxfree;
196: MPI_Allreduce(sendbuf,recvbuf,5,MPIU_INT,MPI_SUM,comm);
197: pdipm->Nxlb = recvbuf[0];
198: pdipm->Nxub = recvbuf[1];
199: pdipm->Nxfixed = recvbuf[2];
200: pdipm->Nxbox = recvbuf[3];
201: pdipm->Nxfree = recvbuf[4];
203: if (pdipm->Nxlb) {
204: ISCreateGeneral(comm,pdipm->nxlb,ixlb,PETSC_COPY_VALUES,&pdipm->isxlb);
205: }
206: if (pdipm->Nxub) {
207: ISCreateGeneral(comm,pdipm->nxub,ixub,PETSC_COPY_VALUES,&pdipm->isxub);
208: }
209: if (pdipm->Nxfixed) {
210: ISCreateGeneral(comm,pdipm->nxfixed,ixfixed,PETSC_COPY_VALUES,&pdipm->isxfixed);
211: }
212: if (pdipm->Nxbox) {
213: ISCreateGeneral(comm,pdipm->nxbox,ixbox,PETSC_COPY_VALUES,&pdipm->isxbox);
214: }
215: if (pdipm->Nxfree) {
216: ISCreateGeneral(comm,pdipm->nxfree,ixfree,PETSC_COPY_VALUES,&pdipm->isxfree);
217: }
218: PetscFree5(ixlb,ixub,ixfixed,ixbox,ixfree);
219: return(0);
220: }
222: /*
223: TaoPDIPMInitializeSolution - Initialize PDIPM solution X = [x; lambdae; lambdai; z].
224: X consists of four subvectors in the order [x; lambdae; lambdai; z]. These
225: four subvectors need to be initialized and its values copied over to X. Instead
226: of copying, we use VecPlace/ResetArray functions to share the memory locations for
227: X and the subvectors
229: Collective
231: Input Parameter:
232: . tao - Tao context
234: Level: beginner
235: */
236: PetscErrorCode TaoPDIPMInitializeSolution(Tao tao)
237: {
239: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
240: PetscScalar *Xarr,*z,*lambdai;
241: PetscInt i;
242: const PetscScalar *xarr,*h;
245: VecGetArray(pdipm->X,&Xarr);
247: /* Set Initialize X.x = tao->solution */
248: VecGetArrayRead(tao->solution,&xarr);
249: PetscMemcpy(Xarr,xarr,pdipm->nx*sizeof(PetscScalar));
250: VecRestoreArrayRead(tao->solution,&xarr);
252: /* Initialize X.lambdae = 0.0 */
253: VecSet(pdipm->lambdae,0.0);
255: /* Initialize X.lambdai = push_init_lambdai, X.z = push_init_slack */
256: VecSet(pdipm->lambdai,pdipm->push_init_lambdai);
257: VecSet(pdipm->z,pdipm->push_init_slack);
259: /* Additional modification for X.lambdai and X.z */
260: VecGetArray(pdipm->lambdai,&lambdai);
261: VecGetArray(pdipm->z,&z);
262: if(pdipm->Nh) {
263: VecGetArrayRead(tao->constraints_inequality,&h);
264: for (i=0; i < pdipm->nh; i++) {
265: if (h[i] < -pdipm->push_init_slack) z[i] = -h[i];
266: if (pdipm->mu/z[i] > pdipm->push_init_lambdai) lambdai[i] = pdipm->mu/z[i];
267: }
268: VecRestoreArrayRead(tao->constraints_inequality,&h);
269: }
270: VecRestoreArray(pdipm->lambdai,&lambdai);
271: VecRestoreArray(pdipm->z,&z);
273: VecRestoreArray(pdipm->X,&Xarr);
274: return(0);
275: }
277: /*
278: TaoSNESJacobian_PDIPM - Evaluate the Hessian matrix at X
280: Input Parameter:
281: snes - SNES context
282: X - KKT Vector
283: *ctx - pdipm context
285: Output Parameter:
286: J - Hessian matrix
287: Jpre - Preconditioner
288: */
289: PetscErrorCode TaoSNESJacobian_PDIPM(SNES snes,Vec X, Mat J, Mat Jpre, void *ctx)
290: {
291: PetscErrorCode ierr;
292: Tao tao=(Tao)ctx;
293: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
294: PetscInt i,row,cols[2],Jrstart,rjstart,nc,j;
295: const PetscInt *aj,*ranges,*Jranges,*rranges,*cranges;
296: const PetscScalar *Xarr,*aa;
297: PetscScalar vals[2];
298: PetscInt proc,nx_all,*nce_all=pdipm->nce_all;
299: MPI_Comm comm;
300: PetscMPIInt rank,size;
301: Mat jac_equality_trans=pdipm->jac_equality_trans,jac_inequality_trans=pdipm->jac_inequality_trans;
304: PetscObjectGetComm((PetscObject)snes,&comm);
305: MPI_Comm_rank(comm,&rank);
306: MPI_Comm_rank(comm,&size);
308: MatGetOwnershipRanges(Jpre,&Jranges);
309: MatGetOwnershipRange(Jpre,&Jrstart,NULL);
310: MatGetOwnershipRangesColumn(tao->hessian,&rranges);
311: MatGetOwnershipRangesColumn(tao->hessian,&cranges);
313: VecGetArrayRead(X,&Xarr);
315: /* (2) insert Z and Ci to Jpre -- overwrite existing values */
316: for (i=0; i < pdipm->nci; i++) {
317: row = Jrstart + pdipm->off_z + i;
318: cols[0] = Jrstart + pdipm->off_lambdai + i;
319: cols[1] = row;
320: vals[0] = Xarr[pdipm->off_z + i];
321: vals[1] = Xarr[pdipm->off_lambdai + i];
322: MatSetValues(Jpre,1,&row,2,cols,vals,INSERT_VALUES);
323: }
325: /* (3) insert 2nd row block of Jpre: [ grad g, 0, 0, 0] */
326: if(pdipm->Ng) {
327: MatGetOwnershipRange(tao->jacobian_equality,&rjstart,NULL);
328: for (i=0; i<pdipm->ng; i++){
329: row = Jrstart + pdipm->off_lambdae + i;
330:
331: MatGetRow(tao->jacobian_equality,i+rjstart,&nc,&aj,&aa);
332: proc = 0;
333: for (j=0; j < nc; j++) {
334: while (aj[j] >= cranges[proc+1]) proc++;
335: cols[0] = aj[j] - cranges[proc] + Jranges[proc];
336: MatSetValue(Jpre,row,cols[0],aa[j],INSERT_VALUES);
337: }
338: MatRestoreRow(tao->jacobian_equality,i+rjstart,&nc,&aj,&aa);
339: }
340: }
342: if(pdipm->Nh) {
343: /* (4) insert 3nd row block of Jpre: [ grad h, 0, 0, 0] */
344: MatGetOwnershipRange(tao->jacobian_inequality,&rjstart,NULL);
345: for (i=0; i < pdipm->nh; i++){
346: row = Jrstart + pdipm->off_lambdai + i;
347:
348: MatGetRow(tao->jacobian_inequality,i+rjstart,&nc,&aj,&aa);
349: proc = 0;
350: for (j=0; j < nc; j++) {
351: while (aj[j] >= cranges[proc+1]) proc++;
352: cols[0] = aj[j] - cranges[proc] + Jranges[proc];
353: MatSetValue(Jpre,row,cols[0],aa[j],INSERT_VALUES);
354: }
355: MatRestoreRow(tao->jacobian_inequality,i+rjstart,&nc,&aj,&aa);
356: }
357: }
359: /* (5) insert Wxx, grad g' and -grad h' to Jpre */
360: if(pdipm->Ng) {
361: MatTranspose(tao->jacobian_equality,MAT_REUSE_MATRIX,&jac_equality_trans);
362: }
363: if(pdipm->Nh) {
364: MatTranspose(tao->jacobian_inequality,MAT_REUSE_MATRIX,&jac_inequality_trans);
365: }
367: VecPlaceArray(pdipm->x,Xarr);
368: TaoComputeHessian(tao,pdipm->x,tao->hessian,tao->hessian_pre);
369: VecResetArray(pdipm->x);
371: MatGetOwnershipRange(tao->hessian,&rjstart,NULL);
372: for (i=0; i<pdipm->nx; i++){
373: row = Jrstart + i;
375: /* insert Wxx */
376: MatGetRow(tao->hessian,i+rjstart,&nc,&aj,&aa);
377: proc = 0;
378: for (j=0; j < nc; j++) {
379: while (aj[j] >= cranges[proc+1]) proc++;
380: cols[0] = aj[j] - cranges[proc] + Jranges[proc];
381: MatSetValue(Jpre,row,cols[0],aa[j],INSERT_VALUES);
382: }
383: MatRestoreRow(tao->hessian,i+rjstart,&nc,&aj,&aa);
385: if(pdipm->ng) {
386: /* insert grad g' */
387: MatGetRow(jac_equality_trans,i+rjstart,&nc,&aj,&aa);
388: MatGetOwnershipRanges(tao->jacobian_equality,&ranges);
389: proc = 0;
390: for (j=0; j < nc; j++) {
391: /* find row ownership of */
392: while (aj[j] >= ranges[proc+1]) proc++;
393: nx_all = rranges[proc+1] - rranges[proc];
394: cols[0] = aj[j] - ranges[proc] + Jranges[proc] + nx_all;
395: MatSetValue(Jpre,row,cols[0],aa[j],INSERT_VALUES);
396: }
397: MatRestoreRow(jac_equality_trans,i+rjstart,&nc,&aj,&aa);
398: }
400: if(pdipm->nh) {
401: /* insert -grad h' */
402: MatGetRow(jac_inequality_trans,i+rjstart,&nc,&aj,&aa);
403: MatGetOwnershipRanges(tao->jacobian_inequality,&ranges);
404: proc = 0;
405: for (j=0; j < nc; j++) {
406: /* find row ownership of */
407: while (aj[j] >= ranges[proc+1]) proc++;
408: nx_all = rranges[proc+1] - rranges[proc];
409: cols[0] = aj[j] - ranges[proc] + Jranges[proc] + nx_all + nce_all[proc];
410: MatSetValue(Jpre,row,cols[0],-aa[j],INSERT_VALUES);
411: }
412: MatRestoreRow(jac_inequality_trans,i+rjstart,&nc,&aj,&aa);
413: }
414: }
415: VecRestoreArrayRead(X,&Xarr);
417: /* (6) assemble Jpre and J */
418: MatAssemblyBegin(Jpre,MAT_FINAL_ASSEMBLY);
419: MatAssemblyEnd(Jpre,MAT_FINAL_ASSEMBLY);
421: if (J != Jpre) {
422: MatAssemblyBegin(J,MAT_FINAL_ASSEMBLY);
423: MatAssemblyEnd(J,MAT_FINAL_ASSEMBLY);
424: }
425: return(0);
426: }
428: /*
429: TaoSnesFunction_PDIPM - Evaluate KKT function at X
431: Input Parameter:
432: snes - SNES context
433: X - KKT Vector
434: *ctx - pdipm
436: Output Parameter:
437: F - Updated Lagrangian vector
438: */
439: PetscErrorCode TaoSNESFunction_PDIPM(SNES snes,Vec X,Vec F,void *ctx)
440: {
442: Tao tao=(Tao)ctx;
443: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
444: PetscScalar *Farr;
445: Vec x,L1;
446: PetscInt i;
447: PetscReal res[2],cnorm[2];
448: const PetscScalar *Xarr,*carr,*zarr,*larr;
451: VecSet(F,0.0);
453: VecGetArrayRead(X,&Xarr);
454: VecGetArray(F,&Farr);
456: /* (0) Evaluate f, fx, Gx, Hx at X.x Note: pdipm->x is not changed below */
457: x = pdipm->x;
458: VecPlaceArray(x,Xarr);
459: TaoPDIPMEvaluateFunctionsAndJacobians(tao,x);
461: /* Update ce, ci, and Jci at X.x */
462: TaoPDIPMUpdateConstraints(tao,x);
463: VecResetArray(x);
465: /* (1) L1 = fx + (gradG'*DE + Jce_xfixed'*lambdae_xfixed) - (gradH'*DI + Jci_xb'*lambdai_xb) */
466: L1 = pdipm->x;
467: VecPlaceArray(L1,Farr);
468: if (pdipm->Nci) {
469: if(pdipm->Nh) {
470: /* L1 += gradH'*DI. Note: tao->DI is not changed below */
471: VecPlaceArray(tao->DI,Xarr+pdipm->off_lambdai);
472: MatMultTransposeAdd(tao->jacobian_inequality,tao->DI,L1,L1);
473: VecResetArray(tao->DI);
474: }
476: /* L1 += Jci_xb'*lambdai_xb */
477: VecPlaceArray(pdipm->lambdai_xb,Xarr+pdipm->off_lambdai+pdipm->nh);
478: MatMultTransposeAdd(pdipm->Jci_xb,pdipm->lambdai_xb,L1,L1);
479: VecResetArray(pdipm->lambdai_xb);
481: /* (1.4) L1 = - (gradH'*DI + Jci_xb'*lambdai_xb) */
482: VecScale(L1,-1.0);
483: }
485: /* L1 += fx */
486: VecAXPY(L1,1.0,tao->gradient);
488: if (pdipm->Nce) {
489: if(pdipm->Ng) {
490: /* L1 += gradG'*DE. Note: tao->DE is not changed below */
491: VecPlaceArray(tao->DE,Xarr+pdipm->off_lambdae);
492: MatMultTransposeAdd(tao->jacobian_equality,tao->DE,L1,L1);
493: VecResetArray(tao->DE);
494: }
495: if (pdipm->Nxfixed) {
496: /* L1 += Jce_xfixed'*lambdae_xfixed */
497: VecPlaceArray(pdipm->lambdae_xfixed,Xarr+pdipm->off_lambdae+pdipm->ng);
498: MatMultTransposeAdd(pdipm->Jce_xfixed,pdipm->lambdae_xfixed,L1,L1);
499: VecResetArray(pdipm->lambdae_xfixed);
500: }
501: }
502: VecNorm(L1,NORM_2,&res[0]);
503: VecResetArray(L1);
505: /* (2) L2 = ce(x) */
506: if (pdipm->Nce) {
507: VecGetArrayRead(pdipm->ce,&carr);
508: for (i=0; i<pdipm->nce; i++) Farr[pdipm->off_lambdae + i] = carr[i];
509: VecRestoreArrayRead(pdipm->ce,&carr);
510: }
511: VecNorm(pdipm->ce,NORM_2,&cnorm[0]);
513: if (pdipm->Nci) {
514: /* (3) L3 = ci(x) - z;
515: (4) L4 = Z * Lambdai * e - mu * e
516: */
517: VecGetArrayRead(pdipm->ci,&carr);
518: larr = Xarr+pdipm->off_lambdai;
519: zarr = Xarr+pdipm->off_z;
520: for (i=0; i<pdipm->nci; i++) {
521: Farr[pdipm->off_lambdai + i] = carr[i] - zarr[i];
522: Farr[pdipm->off_z + i] = zarr[i]*larr[i] - pdipm->mu;
523: }
524: VecRestoreArrayRead(pdipm->ci,&carr);
525: }
527: VecPlaceArray(pdipm->ci,Farr+pdipm->off_lambdai);
528: VecNorm(pdipm->ci,NORM_2,&cnorm[1]);
529: VecResetArray(pdipm->ci);
531: /* note: pdipm->z is not changed below */
532: VecPlaceArray(pdipm->z,Farr+pdipm->off_z);
533: VecNorm(pdipm->z,NORM_2,&res[1]);
534: VecResetArray(pdipm->z);
536: tao->residual = PetscSqrtReal(res[0]*res[0] + res[1]*res[1]);
537: tao->cnorm = PetscSqrtReal(cnorm[0]*cnorm[0] + cnorm[1]*cnorm[1]);
539: VecRestoreArrayRead(X,&Xarr);
540: VecRestoreArray(F,&Farr);
541: return(0);
542: }
544: /*
545: PDIPMLineSearch - Custom line search used with PDIPM.
547: Collective on TAO
549: Notes:
550: PDIPMLineSearch employs a simple backtracking line-search to keep
551: the slack variables (z) and inequality constraints lagrange multipliers
552: (lambdai) positive, i.e., z,lambdai >=0. It does this by calculating scalars
553: alpha_p and alpha_d to keep z,lambdai non-negative. The decision (x), and the
554: slack variables are updated as X = X + alpha_d*dx. The constraint multipliers
555: are updated as Lambdai = Lambdai + alpha_p*dLambdai. The barrier parameter mu
556: is also updated as mu = mu + z'lambdai/Nci
557: */
558: PetscErrorCode PDIPMLineSearch(SNESLineSearch linesearch,void *ctx)
559: {
561: Tao tao=(Tao)ctx;
562: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
563: SNES snes;
564: Vec X,F,Y,W,G;
565: PetscInt i,iter;
566: PetscReal alpha_p=1.0,alpha_d=1.0,alpha[4];
567: PetscScalar *Xarr,*z,*lambdai,dot;
568: const PetscScalar *dXarr,*dz,*dlambdai;
569: PetscScalar *taosolarr;
572: SNESLineSearchGetSNES(linesearch,&snes);
573: SNESGetIterationNumber(snes,&iter);
575: SNESLineSearchSetReason(linesearch,SNES_LINESEARCH_SUCCEEDED);
576: SNESLineSearchGetVecs(linesearch,&X,&F,&Y,&W,&G);
578: VecGetArray(X,&Xarr);
579: VecGetArrayRead(Y,&dXarr);
580: z = Xarr + pdipm->off_z;
581: dz = dXarr + pdipm->off_z;
582: for (i=0; i < pdipm->nci; i++) {
583: if (z[i] - dz[i] < 0.0) {
584: alpha_p = PetscMin(alpha_p,0.9999*z[i]/dz[i]);
585: }
586: }
588: lambdai = Xarr + pdipm->off_lambdai;
589: dlambdai = dXarr + pdipm->off_lambdai;
591: for (i=0; i<pdipm->nci; i++) {
592: if (lambdai[i] - dlambdai[i] < 0.0) {
593: alpha_d = PetscMin(0.9999*lambdai[i]/dlambdai[i],alpha_d);
594: }
595: }
597: alpha[0] = alpha_p;
598: alpha[1] = alpha_d;
599: VecRestoreArrayRead(Y,&dXarr);
600: VecRestoreArray(X,&Xarr);
602: /* alpha = min(alpha) over all processes */
603: MPI_Allreduce(alpha,alpha+2,2,MPIU_REAL,MPIU_MIN,PetscObjectComm((PetscObject)tao));
605: alpha_p = alpha[2];
606: alpha_d = alpha[3];
608: VecGetArray(X,&Xarr);
609: VecGetArrayRead(Y,&dXarr);
610: for (i=0; i<pdipm->nx; i++) {
611: Xarr[i] = Xarr[i] - alpha_p * dXarr[i];
612: }
614: for (i=0; i<pdipm->nce; i++) {
615: Xarr[i+pdipm->off_lambdae] = Xarr[i+pdipm->off_lambdae] - alpha_d * dXarr[i+pdipm->off_lambdae];
616: }
618: for (i=0; i<pdipm->nci; i++) {
619: Xarr[i+pdipm->off_lambdai] = Xarr[i+pdipm->off_lambdai] - alpha_d * dXarr[i+pdipm->off_lambdai];
620: }
622: for (i=0; i<pdipm->nci; i++) {
623: Xarr[i+pdipm->off_z] = Xarr[i+pdipm->off_z] - alpha_p * dXarr[i+pdipm->off_z];
624: }
626: VecGetArray(tao->solution,&taosolarr);
627: PetscMemcpy(taosolarr,Xarr,pdipm->nx*sizeof(PetscScalar));
628: VecRestoreArray(tao->solution,&taosolarr);
631: VecRestoreArray(X,&Xarr);
632: VecRestoreArrayRead(Y,&dXarr);
634: /* Evaluate F at X */
635: SNESComputeFunction(snes,X,F);
636: SNESLineSearchComputeNorms(linesearch); /* must call this func, do not know why */
638: /* update mu = mu_update_factor * dot(z,lambdai)/pdipm->nci at updated X */
639: VecDot(pdipm->z,pdipm->lambdai,&dot);
641: /* if (PetscAbsReal(pdipm->gradL) < 0.9*pdipm->mu) */
642: pdipm->mu = pdipm->mu_update_factor * dot/pdipm->Nci;
644: /* Update F; get tao->residual and tao->cnorm */
645: TaoSNESFunction_PDIPM(snes,X,F,(void*)tao);
647: tao->niter++;
648: TaoLogConvergenceHistory(tao,pdipm->obj,tao->residual,tao->cnorm,tao->niter);
649: TaoMonitor(tao,tao->niter,pdipm->obj,tao->residual,tao->cnorm,pdipm->mu);
651: (*tao->ops->convergencetest)(tao,tao->cnvP);
652: if (tao->reason) {
653: SNESSetConvergedReason(snes,SNES_CONVERGED_FNORM_ABS);
654: }
655: return(0);
656: }
658: /*
659: TaoSolve_PDIPM
661: Input Parameter:
662: tao - TAO context
664: Output Parameter:
665: tao - TAO context
666: */
667: PetscErrorCode TaoSolve_PDIPM(Tao tao)
668: {
669: PetscErrorCode ierr;
670: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
671: SNESLineSearch linesearch; /* SNESLineSearch context */
672: Vec dummy;
675: /* Initialize all variables */
676: TaoPDIPMInitializeSolution(tao);
678: /* Set linesearch */
679: SNESGetLineSearch(pdipm->snes,&linesearch);
680: SNESLineSearchSetType(linesearch,SNESLINESEARCHSHELL);
681: SNESLineSearchShellSetUserFunc(linesearch,PDIPMLineSearch,tao);
682: SNESLineSearchSetFromOptions(linesearch);
684: tao->reason = TAO_CONTINUE_ITERATING;
686: /* -tao_monitor for iteration 0 and check convergence */
687: VecDuplicate(pdipm->X,&dummy);
688: TaoSNESFunction_PDIPM(pdipm->snes,pdipm->X,dummy,(void*)tao);
690: TaoLogConvergenceHistory(tao,pdipm->obj,tao->residual,tao->cnorm,tao->niter);
691: TaoMonitor(tao,tao->niter,pdipm->obj,tao->residual,tao->cnorm,pdipm->mu);
692: VecDestroy(&dummy);
693: (*tao->ops->convergencetest)(tao,tao->cnvP);
694: if (tao->reason) {
695: SNESSetConvergedReason(pdipm->snes,SNES_CONVERGED_FNORM_ABS);
696: }
698: while (tao->reason == TAO_CONTINUE_ITERATING) {
699: SNESConvergedReason reason;
700: SNESSolve(pdipm->snes,NULL,pdipm->X);
702: /* Check SNES convergence */
703: SNESGetConvergedReason(pdipm->snes,&reason);
704: if (reason < 0) {
705: PetscPrintf(PETSC_COMM_WORLD,"SNES solve did not converged due to reason %D\n",reason);
706: }
708: /* Check TAO convergence */
709: if (PetscIsInfOrNanReal(pdipm->obj)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"User-provided compute function generated Inf or NaN");
710: }
711: return(0);
712: }
714: /*
715: TaoSetup_PDIPM - Sets up tao and pdipm
717: Input Parameter:
718: tao - TAO object
720: Output: pdipm - initialized object
721: */
722: PetscErrorCode TaoSetup_PDIPM(Tao tao)
723: {
724: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
726: MPI_Comm comm;
727: PetscMPIInt rank,size;
728: PetscInt row,col,Jcrstart,Jcrend,k,tmp,nc,proc,*nh_all,*ng_all;
729: PetscInt offset,*xa,*xb,i,j,rstart,rend;
730: PetscScalar one=1.0,neg_one=-1.0,*Xarr;
731: const PetscInt *cols,*rranges,*cranges,*aj,*ranges;
732: const PetscScalar *aa;
733: Mat J,jac_equality_trans,jac_inequality_trans;
734: Mat Jce_xfixed_trans,Jci_xb_trans;
735: PetscInt *dnz,*onz,rjstart,nx_all,*nce_all,*Jranges,cols1[2];
738: PetscObjectGetComm((PetscObject)tao,&comm);
739: MPI_Comm_rank(comm,&rank);
740: MPI_Comm_size(comm,&size);
742: /* (1) Setup Bounds and create Tao vectors */
743: TaoPDIPMSetUpBounds(tao);
745: if (!tao->gradient) {
746: VecDuplicate(tao->solution,&tao->gradient);
747: VecDuplicate(tao->solution,&tao->stepdirection);
748: }
750: /* (2) Get sizes */
751: /* Size of vector x - This is set by TaoSetInitia√lVector */
752: VecGetSize(tao->solution,&pdipm->Nx);
753: VecGetLocalSize(tao->solution,&pdipm->nx);
755: /* Size of equality constraints and vectors */
756: if (tao->constraints_equality) {
757: VecGetSize(tao->constraints_equality,&pdipm->Ng);
758: VecGetLocalSize(tao->constraints_equality,&pdipm->ng);
759: } else {
760: pdipm->ng = pdipm->Ng = 0;
761: }
763: pdipm->nce = pdipm->ng + pdipm->nxfixed;
764: pdipm->Nce = pdipm->Ng + pdipm->Nxfixed;
766: /* Size of inequality constraints and vectors */
767: if (tao->constraints_inequality) {
768: VecGetSize(tao->constraints_inequality,&pdipm->Nh);
769: VecGetLocalSize(tao->constraints_inequality,&pdipm->nh);
770: } else {
771: pdipm->nh = pdipm->Nh = 0;
772: }
774: pdipm->nci = pdipm->nh + pdipm->nxlb + pdipm->nxub + 2*pdipm->nxbox;
775: pdipm->Nci = pdipm->Nh + pdipm->Nxlb + pdipm->Nxub + 2*pdipm->Nxbox;
777: /* Full size of the KKT system to be solved */
778: pdipm->n = pdipm->nx + pdipm->nce + 2*pdipm->nci;
779: pdipm->N = pdipm->Nx + pdipm->Nce + 2*pdipm->Nci;
781: /* list below to TaoView_PDIPM()? */
782: /* PetscPrintf(PETSC_COMM_SELF,"[%d] nce %d = ng %d + nxfixed %d\n",rank,pdipm->nce,pdipm->ng,pdipm->nxfixed); */
783: /* PetscPrintf(PETSC_COMM_SELF,"[%d] nci %d = nh %d + nxlb %d + nxub %d + 2*nxbox %d\n",rank,pdipm->nci,pdipm->nh,pdipm->nxlb,pdipm->nxub,pdipm->nxbox); */
784: /* PetscPrintf(PETSC_COMM_SELF,"[%d] n %d = nx %d + nce %d + 2*nci %d\n",rank,pdipm->n,pdipm->nx,pdipm->nce,pdipm->nci); */
786: /* (3) Offsets for subvectors */
787: pdipm->off_lambdae = pdipm->nx;
788: pdipm->off_lambdai = pdipm->off_lambdae + pdipm->nce;
789: pdipm->off_z = pdipm->off_lambdai + pdipm->nci;
791: /* (4) Create vectors and subvectors */
792: /* Ce and Ci vectors */
793: VecCreate(comm,&pdipm->ce);
794: VecSetSizes(pdipm->ce,pdipm->nce,pdipm->Nce);
795: VecSetFromOptions(pdipm->ce);
797: VecCreate(comm,&pdipm->ci);
798: VecSetSizes(pdipm->ci,pdipm->nci,pdipm->Nci);
799: VecSetFromOptions(pdipm->ci);
801: /* X=[x; lambdae; lambdai; z] for the big KKT system */
802: VecCreate(comm,&pdipm->X);
803: VecSetSizes(pdipm->X,pdipm->n,pdipm->N);
804: VecSetFromOptions(pdipm->X);
806: /* Subvectors; they share local arrays with X */
807: VecGetArray(pdipm->X,&Xarr);
808: /* x shares local array with X.x */
809: if (pdipm->Nx) {
810: VecCreateMPIWithArray(comm,1,pdipm->nx,pdipm->Nx,Xarr,&pdipm->x);
811: }
813: /* lambdae shares local array with X.lambdae */
814: if (pdipm->Nce) {
815: VecCreateMPIWithArray(comm,1,pdipm->nce,pdipm->Nce,Xarr+pdipm->off_lambdae,&pdipm->lambdae);
816: }
818: /* tao->DE shares local array with X.lambdae_g */
819: if (pdipm->Ng) {
820: VecCreateMPIWithArray(comm,1,pdipm->ng,pdipm->Ng,Xarr+pdipm->off_lambdae,&tao->DE);
822: VecCreate(comm,&pdipm->lambdae_xfixed);
823: VecSetSizes(pdipm->lambdae_xfixed,pdipm->nxfixed,PETSC_DECIDE);
824: VecSetFromOptions(pdipm->lambdae_xfixed);
825: }
827: if (pdipm->Nci) {
828: /* lambdai shares local array with X.lambdai */
829: VecCreateMPIWithArray(comm,1,pdipm->nci,pdipm->Nci,Xarr+pdipm->off_lambdai,&pdipm->lambdai);
831: /* z for slack variables; it shares local array with X.z */
832: VecCreateMPIWithArray(comm,1,pdipm->nci,pdipm->Nci,Xarr+pdipm->off_z,&pdipm->z);
833: }
835: /* tao->DI which shares local array with X.lambdai_h */
836: if (pdipm->Nh) {
837: VecCreateMPIWithArray(comm,1,pdipm->nh,pdipm->Nh,Xarr+pdipm->off_lambdai,&tao->DI);
838: }
840: VecCreate(comm,&pdipm->lambdai_xb);
841: VecSetSizes(pdipm->lambdai_xb,(pdipm->nci - pdipm->nh),PETSC_DECIDE);
842: VecSetFromOptions(pdipm->lambdai_xb);
844: VecRestoreArray(pdipm->X,&Xarr);
846: /* (5) Create Jacobians Jce_xfixed and Jci */
847: /* (5.1) PDIPM Jacobian of equality bounds cebound(x) = J_nxfixed */
848: if (pdipm->Nxfixed) {
849: /* Create Jce_xfixed */
850: MatCreate(comm,&pdipm->Jce_xfixed);
851: MatSetSizes(pdipm->Jce_xfixed,pdipm->nxfixed,pdipm->nx,PETSC_DECIDE,pdipm->Nx);
852: MatSetFromOptions(pdipm->Jce_xfixed);
853: MatSeqAIJSetPreallocation(pdipm->Jce_xfixed,1,NULL);
854: MatMPIAIJSetPreallocation(pdipm->Jce_xfixed,1,NULL,1,NULL);
856: MatGetOwnershipRange(pdipm->Jce_xfixed,&Jcrstart,&Jcrend);
857: ISGetIndices(pdipm->isxfixed,&cols);
858: k = 0;
859: for (row = Jcrstart; row < Jcrend; row++) {
860: MatSetValues(pdipm->Jce_xfixed,1,&row,1,cols+k,&one,INSERT_VALUES);
861: k++;
862: }
863: ISRestoreIndices(pdipm->isxfixed, &cols);
864: MatAssemblyBegin(pdipm->Jce_xfixed,MAT_FINAL_ASSEMBLY);
865: MatAssemblyEnd(pdipm->Jce_xfixed,MAT_FINAL_ASSEMBLY);
866: }
868: /* (5.2) PDIPM inequality Jacobian Jci = [tao->jacobian_inequality; ...] */
869: MatCreate(comm,&pdipm->Jci_xb);
870: MatSetSizes(pdipm->Jci_xb,pdipm->nci-pdipm->nh,pdipm->nx,PETSC_DECIDE,pdipm->Nx);
871: MatSetFromOptions(pdipm->Jci_xb);
872: MatSeqAIJSetPreallocation(pdipm->Jci_xb,1,NULL);
873: MatMPIAIJSetPreallocation(pdipm->Jci_xb,1,NULL,1,NULL);
875: MatGetOwnershipRange(pdipm->Jci_xb,&Jcrstart,&Jcrend);
876: offset = Jcrstart;
877: if (pdipm->Nxub) {
878: /* Add xub to Jci_xb */
879: ISGetIndices(pdipm->isxub,&cols);
880: k = 0;
881: for (row = offset; row < offset + pdipm->nxub; row++) {
882: MatSetValues(pdipm->Jci_xb,1,&row,1,cols+k,&neg_one,INSERT_VALUES);
883: k++;
884: }
885: ISRestoreIndices(pdipm->isxub, &cols);
886: }
888: if (pdipm->Nxlb) {
889: /* Add xlb to Jci_xb */
890: ISGetIndices(pdipm->isxlb,&cols);
891: k = 0;
892: offset += pdipm->nxub;
893: for (row = offset; row < offset + pdipm->nxlb; row++) {
894: MatSetValues(pdipm->Jci_xb,1,&row,1,cols+k,&one,INSERT_VALUES);
895: k++;
896: }
897: ISRestoreIndices(pdipm->isxlb, &cols);
898: }
900: /* Add xbox to Jci_xb */
901: if (pdipm->Nxbox) {
902: ISGetIndices(pdipm->isxbox,&cols);
903: k = 0;
904: offset += pdipm->nxlb;
905: for (row = offset; row < offset + pdipm->nxbox; row++) {
906: MatSetValues(pdipm->Jci_xb,1,&row,1,cols+k,&neg_one,INSERT_VALUES);
907: tmp = row + pdipm->nxbox;
908: MatSetValues(pdipm->Jci_xb,1,&tmp,1,cols+k,&one,INSERT_VALUES);
909: k++;
910: }
911: ISRestoreIndices(pdipm->isxbox, &cols);
912: }
914: MatAssemblyBegin(pdipm->Jci_xb,MAT_FINAL_ASSEMBLY);
915: MatAssemblyEnd(pdipm->Jci_xb,MAT_FINAL_ASSEMBLY);
916: /* MatView(pdipm->Jci_xb,PETSC_VIEWER_STDOUT_WORLD); */
918: /* (6) Set up ISs for PC Fieldsplit */
919: if (pdipm->solve_reduced_kkt) {
920: PetscMalloc2(pdipm->nx+pdipm->nce,&xa,2*pdipm->nci,&xb);
921: for(i=0; i < pdipm->nx + pdipm->nce; i++) xa[i] = i;
922: for(i=0; i < 2*pdipm->nci; i++) xb[i] = pdipm->off_lambdai + i;
924: ISCreateGeneral(comm,pdipm->nx+pdipm->nce,xa,PETSC_OWN_POINTER,&pdipm->is1);
925: ISCreateGeneral(comm,2*pdipm->nci,xb,PETSC_OWN_POINTER,&pdipm->is2);
926: }
928: /* (7) Gather offsets from all processes */
929: PetscMalloc1(size,&pdipm->nce_all);
931: /* Get rstart of KKT matrix */
932: MPI_Scan(&pdipm->n,&rstart,1,MPIU_INT,MPI_SUM,comm);
933: rstart -= pdipm->n;
935: MPI_Allgather(&pdipm->nce,1,MPIU_INT,pdipm->nce_all,1,MPIU_INT,comm);
937: PetscMalloc3(size,&ng_all,size,&nh_all,size,&Jranges);
938: MPI_Allgather(&rstart,1,MPIU_INT,Jranges,1,MPIU_INT,comm);
939: MPI_Allgather(&pdipm->nh,1,MPIU_INT,nh_all,1,MPIU_INT,comm);
940: MPI_Allgather(&pdipm->ng,1,MPIU_INT,ng_all,1,MPIU_INT,comm);
942: MatGetOwnershipRanges(tao->hessian,&rranges);
943: MatGetOwnershipRangesColumn(tao->hessian,&cranges);
945: if (pdipm->Ng) {
946: TaoComputeJacobianEquality(tao,tao->solution,tao->jacobian_equality,tao->jacobian_equality_pre);
947: MatTranspose(tao->jacobian_equality,MAT_INITIAL_MATRIX,&pdipm->jac_equality_trans);
948: }
949: if (pdipm->Nh) {
950: TaoComputeJacobianInequality(tao,tao->solution,tao->jacobian_inequality,tao->jacobian_inequality_pre);
951: MatTranspose(tao->jacobian_inequality,MAT_INITIAL_MATRIX,&pdipm->jac_inequality_trans);
952: }
954: /* Count dnz,onz for preallocation of KKT matrix */
955: jac_equality_trans = pdipm->jac_equality_trans;
956: jac_inequality_trans = pdipm->jac_inequality_trans;
957: nce_all = pdipm->nce_all;
959: if (pdipm->Nxfixed) {
960: MatTranspose(pdipm->Jce_xfixed,MAT_INITIAL_MATRIX,&Jce_xfixed_trans);
961: }
962: MatTranspose(pdipm->Jci_xb,MAT_INITIAL_MATRIX,&Jci_xb_trans);
964: MatPreallocateInitialize(comm,pdipm->n,pdipm->n,dnz,onz);
966: /* 1st row block of KKT matrix: [Wxx; gradCe'; -gradCi'; 0] */
967: TaoPDIPMEvaluateFunctionsAndJacobians(tao,pdipm->x);
968: TaoComputeHessian(tao,tao->solution,tao->hessian,tao->hessian_pre);
970: /* Insert tao->hessian */
971: MatGetOwnershipRange(tao->hessian,&rjstart,NULL);
972: for (i=0; i<pdipm->nx; i++){
973: row = rstart + i;
975: MatGetRow(tao->hessian,i+rjstart,&nc,&aj,NULL);
976: proc = 0;
977: for (j=0; j < nc; j++) {
978: while (aj[j] >= cranges[proc+1]) proc++;
979: col = aj[j] - cranges[proc] + Jranges[proc];
980: MatPreallocateSet(row,1,&col,dnz,onz);
981: }
982: MatRestoreRow(tao->hessian,i+rjstart,&nc,&aj,NULL);
984: if(pdipm->ng) {
985: /* Insert grad g' */
986: MatGetRow(jac_equality_trans,i+rjstart,&nc,&aj,NULL);
987: MatGetOwnershipRanges(tao->jacobian_equality,&ranges);
988: proc = 0;
989: for (j=0; j < nc; j++) {
990: /* find row ownership of */
991: while (aj[j] >= ranges[proc+1]) proc++;
992: nx_all = rranges[proc+1] - rranges[proc];
993: col = aj[j] - ranges[proc] + Jranges[proc] + nx_all;
994: MatPreallocateSet(row,1,&col,dnz,onz);
995: }
996: MatRestoreRow(jac_equality_trans,i+rjstart,&nc,&aj,NULL);
997: }
999: /* Insert Jce_xfixed^T' */
1000: if (pdipm->nxfixed) {
1001: MatGetRow(Jce_xfixed_trans,i+rjstart,&nc,&aj,NULL);
1002: MatGetOwnershipRanges(pdipm->Jce_xfixed,&ranges);
1003: proc = 0;
1004: for (j=0; j < nc; j++) {
1005: /* find row ownership of */
1006: while (aj[j] >= ranges[proc+1]) proc++;
1007: nx_all = rranges[proc+1] - rranges[proc];
1008: col = aj[j] - ranges[proc] + Jranges[proc] + nx_all + ng_all[proc];
1009: MatPreallocateSet(row,1,&col,dnz,onz);
1010: }
1011: MatRestoreRow(Jce_xfixed_trans,i+rjstart,&nc,&aj,NULL);
1012: }
1014: if(pdipm->nh) {
1015: /* Insert -grad h' */
1016: MatGetRow(jac_inequality_trans,i+rjstart,&nc,&aj,NULL);
1017: MatGetOwnershipRanges(tao->jacobian_inequality,&ranges);
1018: proc = 0;
1019: for (j=0; j < nc; j++) {
1020: /* find row ownership of */
1021: while (aj[j] >= ranges[proc+1]) proc++;
1022: nx_all = rranges[proc+1] - rranges[proc];
1023: col = aj[j] - ranges[proc] + Jranges[proc] + nx_all + nce_all[proc];
1024: MatPreallocateSet(row,1,&col,dnz,onz);
1025: }
1026: MatRestoreRow(jac_inequality_trans,i+rjstart,&nc,&aj,NULL);
1027: }
1029: /* Insert Jci_xb^T' */
1030: MatGetRow(Jci_xb_trans,i+rjstart,&nc,&aj,NULL);
1031: MatGetOwnershipRanges(pdipm->Jci_xb,&ranges);
1032: proc = 0;
1033: for (j=0; j < nc; j++) {
1034: /* find row ownership of */
1035: while (aj[j] >= ranges[proc+1]) proc++;
1036: nx_all = rranges[proc+1] - rranges[proc];
1037: col = aj[j] - ranges[proc] + Jranges[proc] + nx_all + nce_all[proc] + nh_all[proc];
1038: MatPreallocateSet(row,1,&col,dnz,onz);
1039: }
1040: MatRestoreRow(Jci_xb_trans,i+rjstart,&nc,&aj,NULL);
1041: }
1043: /* 2nd Row block of KKT matrix: [grad Ce, 0, 0, 0] */
1044: if(pdipm->Ng) {
1045: MatGetOwnershipRange(tao->jacobian_equality,&rjstart,NULL);
1046: for (i=0; i < pdipm->ng; i++){
1047: row = rstart + pdipm->off_lambdae + i;
1049: MatGetRow(tao->jacobian_equality,i+rjstart,&nc,&aj,NULL);
1050: proc = 0;
1051: for (j=0; j < nc; j++) {
1052: while (aj[j] >= cranges[proc+1]) proc++;
1053: col = aj[j] - cranges[proc] + Jranges[proc];
1054: MatPreallocateSet(row,1,&col,dnz,onz); /* grad g */
1055: }
1056: MatRestoreRow(tao->jacobian_equality,i+rjstart,&nc,&aj,NULL);
1057: }
1058: }
1059: /* Jce_xfixed */
1060: if (pdipm->Nxfixed) {
1061: MatGetOwnershipRange(pdipm->Jce_xfixed,&Jcrstart,NULL);
1062: for (i=0; i < (pdipm->nce - pdipm->ng); i++ ){
1063: row = rstart + pdipm->off_lambdae + pdipm->ng + i;
1065: MatGetRow(pdipm->Jce_xfixed,i+Jcrstart,&nc,&cols,NULL);
1066: if (nc != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"nc != 1");
1068: proc = 0;
1069: j = 0;
1070: while (cols[j] >= cranges[proc+1]) proc++;
1071: col = cols[j] - cranges[proc] + Jranges[proc];
1072: MatPreallocateSet(row,1,&col,dnz,onz);
1073: MatRestoreRow(pdipm->Jce_xfixed,i+Jcrstart,&nc,&cols,NULL);
1074: }
1075: }
1077: /* 3rd Row block of KKT matrix: [ gradCi, 0, 0, -I] */
1078: if(pdipm->Nh) {
1079: MatGetOwnershipRange(tao->jacobian_inequality,&rjstart,NULL);
1080: for (i=0; i < pdipm->nh; i++){
1081: row = rstart + pdipm->off_lambdai + i;
1083: MatGetRow(tao->jacobian_inequality,i+rjstart,&nc,&aj,NULL);
1084: proc = 0;
1085: for (j=0; j < nc; j++) {
1086: while (aj[j] >= cranges[proc+1]) proc++;
1087: col = aj[j] - cranges[proc] + Jranges[proc];
1088: MatPreallocateSet(row,1,&col,dnz,onz); /* grad h */
1089: }
1090: MatRestoreRow(tao->jacobian_inequality,i+rjstart,&nc,&aj,NULL);
1091: }
1092: /* -I */
1093: for (i=0; i < pdipm->nh; i++){
1094: row = rstart + pdipm->off_lambdai + i;
1095: col = rstart + pdipm->off_z + i;
1096: MatPreallocateSet(row,1,&col,dnz,onz);
1097: }
1098: }
1100: /* Jci_xb */
1101: MatGetOwnershipRange(pdipm->Jci_xb,&Jcrstart,NULL);
1102: for (i=0; i < (pdipm->nci - pdipm->nh); i++ ){
1103: row = rstart + pdipm->off_lambdai + pdipm->nh + i;
1105: MatGetRow(pdipm->Jci_xb,i+Jcrstart,&nc,&cols,NULL);
1106: if (nc != 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"nc != 1");
1107: proc = 0;
1108: for (j=0; j < nc; j++) {
1109: while (cols[j] >= cranges[proc+1]) proc++;
1110: col = cols[j] - cranges[proc] + Jranges[proc];
1111: MatPreallocateSet(row,1,&col,dnz,onz);
1112: }
1113: MatRestoreRow(pdipm->Jci_xb,i+Jcrstart,&nc,&cols,NULL);
1114: /* -I */
1115: col = rstart + pdipm->off_z + pdipm->nh + i;
1116: MatPreallocateSet(row,1,&col,dnz,onz);
1117: }
1119: /* 4-th Row block of KKT matrix: Z and Ci */
1120: for (i=0; i < pdipm->nci; i++) {
1121: row = rstart + pdipm->off_z + i;
1122: cols1[0] = rstart + pdipm->off_lambdai + i;
1123: cols1[1] = row;
1124: MatPreallocateSet(row,2,cols1,dnz,onz);
1125: }
1127: /* diagonal entry */
1128: for (i=0; i<pdipm->n; i++) dnz[i]++; /* diagonal entry */
1130: /* Create KKT matrix */
1131: MatCreate(comm,&J);
1132: MatSetSizes(J,pdipm->n,pdipm->n,PETSC_DECIDE,PETSC_DECIDE);
1133: MatSetFromOptions(J);
1134: MatSeqAIJSetPreallocation(J,0,dnz);
1135: MatMPIAIJSetPreallocation(J,0,dnz,0,onz);
1136: /* MatSetOption(J,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE); */
1137: MatPreallocateFinalize(dnz,onz);
1138: pdipm->K = J;
1140: /* (8) Set up nonlinear solver SNES */
1141: SNESSetFunction(pdipm->snes,NULL,TaoSNESFunction_PDIPM,(void*)tao);
1142: SNESSetJacobian(pdipm->snes,J,J,TaoSNESJacobian_PDIPM,(void*)tao);
1144: if (pdipm->solve_reduced_kkt) {
1145: PC pc;
1146: KSPGetPC(tao->ksp,&pc);
1147: PCSetType(pc,PCFIELDSPLIT);
1148: PCFieldSplitSetType(pc,PC_COMPOSITE_SCHUR);
1149: PCFieldSplitSetIS(pc,"2",pdipm->is2);
1150: PCFieldSplitSetIS(pc,"1",pdipm->is1);
1151: }
1152: SNESSetFromOptions(pdipm->snes);
1154: /* (9) Insert constant entries to K */
1155: /* Set 0.0 to diagonal of K, so that the solver does not complain *about missing diagonal value */
1156: MatGetOwnershipRange(J,&rstart,&rend);
1157: for (i=rstart; i<rend; i++){
1158: MatSetValue(J,i,i,0.0,INSERT_VALUES);
1159: }
1161: /* Row block of K: [ grad Ce, 0, 0, 0] */
1162: if (pdipm->Nxfixed) {
1163: MatGetOwnershipRange(pdipm->Jce_xfixed,&Jcrstart,NULL);
1164: for (i=0; i < (pdipm->nce - pdipm->ng); i++ ){
1165: row = rstart + pdipm->off_lambdae + pdipm->ng + i;
1167: MatGetRow(pdipm->Jce_xfixed,i+Jcrstart,&nc,&cols,&aa);
1168: proc = 0;
1169: for (j=0; j < nc; j++) {
1170: while (cols[j] >= cranges[proc+1]) proc++;
1171: col = cols[j] - cranges[proc] + Jranges[proc];
1172: MatSetValue(J,row,col,aa[j],INSERT_VALUES); /* grad Ce */
1173: MatSetValue(J,col,row,aa[j],INSERT_VALUES); /* grad Ce' */
1174: }
1175: MatRestoreRow(pdipm->Jce_xfixed,i+Jcrstart,&nc,&cols,&aa);
1176: }
1177: }
1179: /* Row block of K: [ grad Ci, 0, 0, -I] */
1180: MatGetOwnershipRange(pdipm->Jci_xb,&Jcrstart,NULL);
1181: for (i=0; i < (pdipm->nci - pdipm->nh); i++ ){
1182: row = rstart + pdipm->off_lambdai + pdipm->nh + i;
1184: MatGetRow(pdipm->Jci_xb,i+Jcrstart,&nc,&cols,&aa);
1185: proc = 0;
1186: for (j=0; j < nc; j++) {
1187: while (cols[j] >= cranges[proc+1]) proc++;
1188: col = cols[j] - cranges[proc] + Jranges[proc];
1189: MatSetValue(J,col,row,-aa[j],INSERT_VALUES);
1190: MatSetValue(J,row,col,aa[j],INSERT_VALUES);
1191: }
1192: MatRestoreRow(pdipm->Jci_xb,i+Jcrstart,&nc,&cols,&aa);
1194: col = rstart + pdipm->off_z + pdipm->nh + i;
1195: MatSetValue(J,row,col,-1,INSERT_VALUES);
1196: }
1198: for (i=0; i < pdipm->nh; i++){
1199: row = rstart + pdipm->off_lambdai + i;
1200: col = rstart + pdipm->off_z + i;
1201: MatSetValue(J,row,col,-1,INSERT_VALUES);
1202: }
1204: if (pdipm->Nxfixed) {
1205: MatDestroy(&Jce_xfixed_trans);
1206: }
1207: MatDestroy(&Jci_xb_trans);
1208: PetscFree3(ng_all,nh_all,Jranges);
1209: return(0);
1210: }
1212: /*
1213: TaoDestroy_PDIPM - Destroys the pdipm object
1215: Input:
1216: full pdipm
1218: Output:
1219: Destroyed pdipm
1220: */
1221: PetscErrorCode TaoDestroy_PDIPM(Tao tao)
1222: {
1223: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
1227: /* Freeing Vectors assocaiated with KKT (X) */
1228: VecDestroy(&pdipm->x); /* Solution x */
1229: VecDestroy(&pdipm->lambdae); /* Equality constraints lagrangian multiplier*/
1230: VecDestroy(&pdipm->lambdai); /* Inequality constraints lagrangian multiplier*/
1231: VecDestroy(&pdipm->z); /* Slack variables */
1232: VecDestroy(&pdipm->X); /* Big KKT system vector [x; lambdae; lambdai; z] */
1234: /* work vectors */
1235: VecDestroy(&pdipm->lambdae_xfixed);
1236: VecDestroy(&pdipm->lambdai_xb);
1238: /* Legrangian equality and inequality Vec */
1239: VecDestroy(&pdipm->ce); /* Vec of equality constraints */
1240: VecDestroy(&pdipm->ci); /* Vec of inequality constraints */
1242: /* Matrices */
1243: MatDestroy(&pdipm->Jce_xfixed);
1244: MatDestroy(&pdipm->Jci_xb); /* Jacobian of inequality constraints Jci = [tao->jacobian_inequality ; J(nxub); J(nxlb); J(nxbx)] */
1245: MatDestroy(&pdipm->K);
1247: /* Index Sets */
1248: if (pdipm->Nxub) {
1249: ISDestroy(&pdipm->isxub); /* Finite upper bound only -inf < x < ub */
1250: }
1252: if (pdipm->Nxlb) {
1253: ISDestroy(&pdipm->isxlb); /* Finite lower bound only lb <= x < inf */
1254: }
1256: if (pdipm->Nxfixed) {
1257: ISDestroy(&pdipm->isxfixed); /* Fixed variables lb = x = ub */
1258: }
1260: if (pdipm->Nxbox) {
1261: ISDestroy(&pdipm->isxbox); /* Boxed variables lb <= x <= ub */
1262: }
1264: if (pdipm->Nxfree) {
1265: ISDestroy(&pdipm->isxfree); /* Free variables -inf <= x <= inf */
1266: }
1268: if (pdipm->solve_reduced_kkt) {
1269: ISDestroy(&pdipm->is1);
1270: ISDestroy(&pdipm->is2);
1271: }
1273: /* SNES */
1274: SNESDestroy(&pdipm->snes); /* Nonlinear solver */
1275: PetscFree(pdipm->nce_all);
1276: MatDestroy(&pdipm->jac_equality_trans);
1277: MatDestroy(&pdipm->jac_inequality_trans);
1279: /* Destroy pdipm */
1280: PetscFree(tao->data); /* Holding locations of pdipm */
1282: /* Destroy Dual */
1283: VecDestroy(&tao->DE); /* equality dual */
1284: VecDestroy(&tao->DI); /* dinequality dual */
1285: return(0);
1286: }
1288: PetscErrorCode TaoSetFromOptions_PDIPM(PetscOptionItems *PetscOptionsObject,Tao tao)
1289: {
1290: TAO_PDIPM *pdipm = (TAO_PDIPM*)tao->data;
1294: PetscOptionsHead(PetscOptionsObject,"PDIPM method for constrained optimization");
1295: PetscOptionsReal("-tao_pdipm_push_init_slack","parameter to push initial slack variables away from bounds",NULL,pdipm->push_init_slack,&pdipm->push_init_slack,NULL);
1296: PetscOptionsReal("-tao_pdipm_push_init_lambdai","parameter to push initial (inequality) dual variables away from bounds",NULL,pdipm->push_init_lambdai,&pdipm->push_init_lambdai,NULL);
1297: PetscOptionsBool("-tao_pdipm_solve_reduced_kkt","Solve reduced KKT system using Schur-complement",NULL,pdipm->solve_reduced_kkt,&pdipm->solve_reduced_kkt,NULL);
1298: PetscOptionsReal("-tao_pdipm_mu_update_factor","Update scalar for barrier parameter (mu) update",NULL,pdipm->mu_update_factor,&pdipm->mu_update_factor,NULL);
1299: PetscOptionsTail();
1300: return(0);
1301: }
1303: /*MC
1304: TAOPDIPM - Barrier-based primal-dual interior point algorithm for generally constrained optimization.
1306: Option Database Keys:
1307: + -tao_pdipm_push_init_lambdai - parameter to push initial dual variables away from bounds (> 0)
1308: . -tao_pdipm_push_init_slack - parameter to push initial slack variables away from bounds (> 0)
1309: - -tao_pdipm_mu_update_factor - update scalar for barrier parameter (mu) update (> 0)
1311: Level: beginner
1312: M*/
1313: PETSC_EXTERN PetscErrorCode TaoCreate_PDIPM(Tao tao)
1314: {
1315: TAO_PDIPM *pdipm;
1319: tao->ops->setup = TaoSetup_PDIPM;
1320: tao->ops->solve = TaoSolve_PDIPM;
1321: tao->ops->setfromoptions = TaoSetFromOptions_PDIPM;
1322: tao->ops->destroy = TaoDestroy_PDIPM;
1324: PetscNewLog(tao,&pdipm);
1325: tao->data = (void*)pdipm;
1327: pdipm->nx = pdipm->Nx = 0;
1328: pdipm->nxfixed = pdipm->Nxfixed = 0;
1329: pdipm->nxlb = pdipm->Nxlb = 0;
1330: pdipm->nxub = pdipm->Nxub = 0;
1331: pdipm->nxbox = pdipm->Nxbox = 0;
1332: pdipm->nxfree = pdipm->Nxfree = 0;
1334: pdipm->ng = pdipm->Ng = pdipm->nce = pdipm->Nce = 0;
1335: pdipm->nh = pdipm->Nh = pdipm->nci = pdipm->Nci = 0;
1336: pdipm->n = pdipm->N = 0;
1337: pdipm->mu = 1.0;
1338: pdipm->mu_update_factor = 0.1;
1340: pdipm->push_init_slack = 1.0;
1341: pdipm->push_init_lambdai = 1.0;
1342: pdipm->solve_reduced_kkt = PETSC_FALSE;
1344: /* Override default settings (unless already changed) */
1345: if (!tao->max_it_changed) tao->max_it = 200;
1346: if (!tao->max_funcs_changed) tao->max_funcs = 500;
1348: SNESCreate(((PetscObject)tao)->comm,&pdipm->snes);
1349: SNESSetOptionsPrefix(pdipm->snes,tao->hdr.prefix);
1350: SNESGetKSP(pdipm->snes,&tao->ksp);
1351: PetscObjectReference((PetscObject)tao->ksp);
1352: return(0);
1353: }