Actual source code: bddcprivate.c
petsc-3.13.6 2020-09-29
1: #include <../src/mat/impls/aij/seq/aij.h>
2: #include <../src/ksp/pc/impls/bddc/bddc.h>
3: #include <../src/ksp/pc/impls/bddc/bddcprivate.h>
4: #include <../src/mat/impls/dense/seq/dense.h>
5: #include <petscdmplex.h>
6: #include <petscblaslapack.h>
7: #include <petsc/private/sfimpl.h>
8: #include <petsc/private/dmpleximpl.h>
9: #include <petscdmda.h>
11: static PetscErrorCode MatMPIAIJRestrict(Mat,MPI_Comm,Mat*);
13: /* if range is true, it returns B s.t. span{B} = range(A)
14: if range is false, it returns B s.t. range(B) _|_ range(A) */
15: PetscErrorCode MatDenseOrthogonalRangeOrComplement(Mat A, PetscBool range, PetscInt lw, PetscScalar *work, PetscReal *rwork, Mat *B)
16: {
17: #if !defined(PETSC_USE_COMPLEX)
18: PetscScalar *uwork,*data,*U, ds = 0.;
19: PetscReal *sing;
20: PetscBLASInt bM,bN,lwork,lierr,di = 1;
21: PetscInt ulw,i,nr,nc,n;
25: MatGetSize(A,&nr,&nc);
26: if (!nr || !nc) return(0);
28: /* workspace */
29: if (!work) {
30: ulw = PetscMax(PetscMax(1,5*PetscMin(nr,nc)),3*PetscMin(nr,nc)+PetscMax(nr,nc));
31: PetscMalloc1(ulw,&uwork);
32: } else {
33: ulw = lw;
34: uwork = work;
35: }
36: n = PetscMin(nr,nc);
37: if (!rwork) {
38: PetscMalloc1(n,&sing);
39: } else {
40: sing = rwork;
41: }
43: /* SVD */
44: PetscMalloc1(nr*nr,&U);
45: PetscBLASIntCast(nr,&bM);
46: PetscBLASIntCast(nc,&bN);
47: PetscBLASIntCast(ulw,&lwork);
48: MatDenseGetArray(A,&data);
49: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
50: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("A","N",&bM,&bN,data,&bM,sing,U,&bM,&ds,&di,uwork,&lwork,&lierr));
51: PetscFPTrapPop();
52: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr);
53: MatDenseRestoreArray(A,&data);
54: for (i=0;i<n;i++) if (sing[i] < PETSC_SMALL) break;
55: if (!rwork) {
56: PetscFree(sing);
57: }
58: if (!work) {
59: PetscFree(uwork);
60: }
61: /* create B */
62: if (!range) {
63: MatCreateSeqDense(PETSC_COMM_SELF,nr,nr-i,NULL,B);
64: MatDenseGetArray(*B,&data);
65: PetscArraycpy(data,U+nr*i,(nr-i)*nr);
66: } else {
67: MatCreateSeqDense(PETSC_COMM_SELF,nr,i,NULL,B);
68: MatDenseGetArray(*B,&data);
69: PetscArraycpy(data,U,i*nr);
70: }
71: MatDenseRestoreArray(*B,&data);
72: PetscFree(U);
73: #else /* PETSC_USE_COMPLEX */
75: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented for complexes");
76: #endif
77: return(0);
78: }
80: /* TODO REMOVE */
81: #if defined(PRINT_GDET)
82: static int inc = 0;
83: static int lev = 0;
84: #endif
86: PetscErrorCode PCBDDCComputeNedelecChangeEdge(Mat lG, IS edge, IS extrow, IS extcol, IS corners, Mat* Gins, Mat* GKins, PetscScalar cvals[2], PetscScalar *work, PetscReal *rwork)
87: {
89: Mat GE,GEd;
90: PetscInt rsize,csize,esize;
91: PetscScalar *ptr;
94: ISGetSize(edge,&esize);
95: if (!esize) return(0);
96: ISGetSize(extrow,&rsize);
97: ISGetSize(extcol,&csize);
99: /* gradients */
100: ptr = work + 5*esize;
101: MatCreateSubMatrix(lG,extrow,extcol,MAT_INITIAL_MATRIX,&GE);
102: MatCreateSeqDense(PETSC_COMM_SELF,rsize,csize,ptr,Gins);
103: MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,Gins);
104: MatDestroy(&GE);
106: /* constants */
107: ptr += rsize*csize;
108: MatCreateSeqDense(PETSC_COMM_SELF,esize,csize,ptr,&GEd);
109: MatCreateSubMatrix(lG,edge,extcol,MAT_INITIAL_MATRIX,&GE);
110: MatConvert(GE,MATSEQDENSE,MAT_REUSE_MATRIX,&GEd);
111: MatDestroy(&GE);
112: MatDenseOrthogonalRangeOrComplement(GEd,PETSC_FALSE,5*esize,work,rwork,GKins);
113: MatDestroy(&GEd);
115: if (corners) {
116: Mat GEc;
117: const PetscScalar *vals;
118: PetscScalar v;
120: MatCreateSubMatrix(lG,edge,corners,MAT_INITIAL_MATRIX,&GEc);
121: MatTransposeMatMult(GEc,*GKins,MAT_INITIAL_MATRIX,1.0,&GEd);
122: MatDenseGetArrayRead(GEd,&vals);
123: /* v = PetscAbsScalar(vals[0]) */;
124: v = 1.;
125: cvals[0] = vals[0]/v;
126: cvals[1] = vals[1]/v;
127: MatDenseRestoreArrayRead(GEd,&vals);
128: MatScale(*GKins,1./v);
129: #if defined(PRINT_GDET)
130: {
131: PetscViewer viewer;
132: char filename[256];
133: sprintf(filename,"Gdet_l%d_r%d_cc%d.m",lev,PetscGlobalRank,inc++);
134: PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);
135: PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);
136: PetscObjectSetName((PetscObject)GEc,"GEc");
137: MatView(GEc,viewer);
138: PetscObjectSetName((PetscObject)(*GKins),"GK");
139: MatView(*GKins,viewer);
140: PetscObjectSetName((PetscObject)GEd,"Gproj");
141: MatView(GEd,viewer);
142: PetscViewerDestroy(&viewer);
143: }
144: #endif
145: MatDestroy(&GEd);
146: MatDestroy(&GEc);
147: }
149: return(0);
150: }
152: PetscErrorCode PCBDDCNedelecSupport(PC pc)
153: {
154: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
155: Mat_IS *matis = (Mat_IS*)pc->pmat->data;
156: Mat G,T,conn,lG,lGt,lGis,lGall,lGe,lGinit;
157: Vec tvec;
158: PetscSF sfv;
159: ISLocalToGlobalMapping el2g,vl2g,fl2g,al2g;
160: MPI_Comm comm;
161: IS lned,primals,allprimals,nedfieldlocal;
162: IS *eedges,*extrows,*extcols,*alleedges;
163: PetscBT btv,bte,btvc,btb,btbd,btvcand,btvi,btee,bter;
164: PetscScalar *vals,*work;
165: PetscReal *rwork;
166: const PetscInt *idxs,*ii,*jj,*iit,*jjt;
167: PetscInt ne,nv,Lv,order,n,field;
168: PetscInt n_neigh,*neigh,*n_shared,**shared;
169: PetscInt i,j,extmem,cum,maxsize,nee;
170: PetscInt *extrow,*extrowcum,*marks,*vmarks,*gidxs;
171: PetscInt *sfvleaves,*sfvroots;
172: PetscInt *corners,*cedges;
173: PetscInt *ecount,**eneighs,*vcount,**vneighs;
174: #if defined(PETSC_USE_DEBUG)
175: PetscInt *emarks;
176: #endif
177: PetscBool print,eerr,done,lrc[2],conforming,global,singular,setprimal;
178: PetscErrorCode ierr;
181: /* If the discrete gradient is defined for a subset of dofs and global is true,
182: it assumes G is given in global ordering for all the dofs.
183: Otherwise, the ordering is global for the Nedelec field */
184: order = pcbddc->nedorder;
185: conforming = pcbddc->conforming;
186: field = pcbddc->nedfield;
187: global = pcbddc->nedglobal;
188: setprimal = PETSC_FALSE;
189: print = PETSC_FALSE;
190: singular = PETSC_FALSE;
192: /* Command line customization */
193: PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC Nedelec options","PC");
194: PetscOptionsBool("-pc_bddc_nedelec_field_primal","All edge dofs set as primals: Toselli's algorithm C",NULL,setprimal,&setprimal,NULL);
195: PetscOptionsBool("-pc_bddc_nedelec_singular","Infer nullspace from discrete gradient",NULL,singular,&singular,NULL);
196: PetscOptionsInt("-pc_bddc_nedelec_order","Test variable order code (to be removed)",NULL,order,&order,NULL);
197: /* print debug info TODO: to be removed */
198: PetscOptionsBool("-pc_bddc_nedelec_print","Print debug info",NULL,print,&print,NULL);
199: PetscOptionsEnd();
201: /* Return if there are no edges in the decomposition and the problem is not singular */
202: MatGetLocalToGlobalMapping(pc->pmat,&al2g,NULL);
203: ISLocalToGlobalMappingGetSize(al2g,&n);
204: PetscObjectGetComm((PetscObject)pc,&comm);
205: if (!singular) {
206: VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);
207: lrc[0] = PETSC_FALSE;
208: for (i=0;i<n;i++) {
209: if (PetscRealPart(vals[i]) > 2.) {
210: lrc[0] = PETSC_TRUE;
211: break;
212: }
213: }
214: VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);
215: MPIU_Allreduce(&lrc[0],&lrc[1],1,MPIU_BOOL,MPI_LOR,comm);
216: if (!lrc[1]) return(0);
217: }
219: /* Get Nedelec field */
220: if (pcbddc->n_ISForDofsLocal && field >= pcbddc->n_ISForDofsLocal) SETERRQ2(comm,PETSC_ERR_USER,"Invalid field for Nedelec %D: number of fields is %D",field,pcbddc->n_ISForDofsLocal);
221: if (pcbddc->n_ISForDofsLocal && field >= 0) {
222: PetscObjectReference((PetscObject)pcbddc->ISForDofsLocal[field]);
223: nedfieldlocal = pcbddc->ISForDofsLocal[field];
224: ISGetLocalSize(nedfieldlocal,&ne);
225: } else if (!pcbddc->n_ISForDofsLocal && field != PETSC_DECIDE) {
226: ne = n;
227: nedfieldlocal = NULL;
228: global = PETSC_TRUE;
229: } else if (field == PETSC_DECIDE) {
230: PetscInt rst,ren,*idx;
232: PetscArrayzero(matis->sf_leafdata,n);
233: PetscArrayzero(matis->sf_rootdata,pc->pmat->rmap->n);
234: MatGetOwnershipRange(pcbddc->discretegradient,&rst,&ren);
235: for (i=rst;i<ren;i++) {
236: PetscInt nc;
238: MatGetRow(pcbddc->discretegradient,i,&nc,NULL,NULL);
239: if (nc > 1) matis->sf_rootdata[i-rst] = 1;
240: MatRestoreRow(pcbddc->discretegradient,i,&nc,NULL,NULL);
241: }
242: PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);
243: PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);
244: PetscMalloc1(n,&idx);
245: for (i=0,ne=0;i<n;i++) if (matis->sf_leafdata[i]) idx[ne++] = i;
246: ISCreateGeneral(comm,ne,idx,PETSC_OWN_POINTER,&nedfieldlocal);
247: } else {
248: SETERRQ(comm,PETSC_ERR_USER,"When multiple fields are present, the Nedelec field has to be specified");
249: }
251: /* Sanity checks */
252: if (!order && !conforming) SETERRQ(comm,PETSC_ERR_SUP,"Variable order and non-conforming spaces are not supported at the same time");
253: if (pcbddc->user_ChangeOfBasisMatrix) SETERRQ(comm,PETSC_ERR_SUP,"Cannot generate Nedelec support with user defined change of basis");
254: if (order && ne%order) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"The number of local edge dofs %D it's not a multiple of the order %D",ne,order);
256: /* Just set primal dofs and return */
257: if (setprimal) {
258: IS enedfieldlocal;
259: PetscInt *eidxs;
261: PetscMalloc1(ne,&eidxs);
262: VecGetArrayRead(matis->counter,(const PetscScalar**)&vals);
263: if (nedfieldlocal) {
264: ISGetIndices(nedfieldlocal,&idxs);
265: for (i=0,cum=0;i<ne;i++) {
266: if (PetscRealPart(vals[idxs[i]]) > 2.) {
267: eidxs[cum++] = idxs[i];
268: }
269: }
270: ISRestoreIndices(nedfieldlocal,&idxs);
271: } else {
272: for (i=0,cum=0;i<ne;i++) {
273: if (PetscRealPart(vals[i]) > 2.) {
274: eidxs[cum++] = i;
275: }
276: }
277: }
278: VecRestoreArrayRead(matis->counter,(const PetscScalar**)&vals);
279: ISCreateGeneral(comm,cum,eidxs,PETSC_COPY_VALUES,&enedfieldlocal);
280: PCBDDCSetPrimalVerticesLocalIS(pc,enedfieldlocal);
281: PetscFree(eidxs);
282: ISDestroy(&nedfieldlocal);
283: ISDestroy(&enedfieldlocal);
284: return(0);
285: }
287: /* Compute some l2g maps */
288: if (nedfieldlocal) {
289: IS is;
291: /* need to map from the local Nedelec field to local numbering */
292: ISLocalToGlobalMappingCreateIS(nedfieldlocal,&fl2g);
293: /* need to map from the local Nedelec field to global numbering for the whole dofs*/
294: ISLocalToGlobalMappingApplyIS(al2g,nedfieldlocal,&is);
295: ISLocalToGlobalMappingCreateIS(is,&al2g);
296: /* need to map from the local Nedelec field to global numbering (for Nedelec only) */
297: if (global) {
298: PetscObjectReference((PetscObject)al2g);
299: el2g = al2g;
300: } else {
301: IS gis;
303: ISRenumber(is,NULL,NULL,&gis);
304: ISLocalToGlobalMappingCreateIS(gis,&el2g);
305: ISDestroy(&gis);
306: }
307: ISDestroy(&is);
308: } else {
309: /* restore default */
310: pcbddc->nedfield = -1;
311: /* one ref for the destruction of al2g, one for el2g */
312: PetscObjectReference((PetscObject)al2g);
313: PetscObjectReference((PetscObject)al2g);
314: el2g = al2g;
315: fl2g = NULL;
316: }
318: /* Start communication to drop connections for interior edges (for cc analysis only) */
319: PetscArrayzero(matis->sf_leafdata,n);
320: PetscArrayzero(matis->sf_rootdata,pc->pmat->rmap->n);
321: if (nedfieldlocal) {
322: ISGetIndices(nedfieldlocal,&idxs);
323: for (i=0;i<ne;i++) matis->sf_leafdata[idxs[i]] = 1;
324: ISRestoreIndices(nedfieldlocal,&idxs);
325: } else {
326: for (i=0;i<ne;i++) matis->sf_leafdata[i] = 1;
327: }
328: PetscSFReduceBegin(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);
329: PetscSFReduceEnd(matis->sf,MPIU_INT,matis->sf_leafdata,matis->sf_rootdata,MPI_SUM);
331: if (!singular) { /* drop connections with interior edges to avoid unneeded communications and memory movements */
332: MatDuplicate(pcbddc->discretegradient,MAT_COPY_VALUES,&G);
333: MatSetOption(G,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);
334: if (global) {
335: PetscInt rst;
337: MatGetOwnershipRange(G,&rst,NULL);
338: for (i=0,cum=0;i<pc->pmat->rmap->n;i++) {
339: if (matis->sf_rootdata[i] < 2) {
340: matis->sf_rootdata[cum++] = i + rst;
341: }
342: }
343: MatSetOption(G,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE);
344: MatZeroRows(G,cum,matis->sf_rootdata,0.,NULL,NULL);
345: } else {
346: PetscInt *tbz;
348: PetscMalloc1(ne,&tbz);
349: PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);
350: PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);
351: ISGetIndices(nedfieldlocal,&idxs);
352: for (i=0,cum=0;i<ne;i++)
353: if (matis->sf_leafdata[idxs[i]] == 1)
354: tbz[cum++] = i;
355: ISRestoreIndices(nedfieldlocal,&idxs);
356: ISLocalToGlobalMappingApply(el2g,cum,tbz,tbz);
357: MatZeroRows(G,cum,tbz,0.,NULL,NULL);
358: PetscFree(tbz);
359: }
360: } else { /* we need the entire G to infer the nullspace */
361: PetscObjectReference((PetscObject)pcbddc->discretegradient);
362: G = pcbddc->discretegradient;
363: }
365: /* Extract subdomain relevant rows of G */
366: ISLocalToGlobalMappingGetIndices(el2g,&idxs);
367: ISCreateGeneral(comm,ne,idxs,PETSC_USE_POINTER,&lned);
368: MatCreateSubMatrix(G,lned,NULL,MAT_INITIAL_MATRIX,&lGall);
369: ISLocalToGlobalMappingRestoreIndices(el2g,&idxs);
370: ISDestroy(&lned);
371: MatConvert(lGall,MATIS,MAT_INITIAL_MATRIX,&lGis);
372: MatDestroy(&lGall);
373: MatISGetLocalMat(lGis,&lG);
375: /* SF for nodal dofs communications */
376: MatGetLocalSize(G,NULL,&Lv);
377: MatGetLocalToGlobalMapping(lGis,NULL,&vl2g);
378: PetscObjectReference((PetscObject)vl2g);
379: ISLocalToGlobalMappingGetSize(vl2g,&nv);
380: PetscSFCreate(comm,&sfv);
381: ISLocalToGlobalMappingGetIndices(vl2g,&idxs);
382: PetscSFSetGraphLayout(sfv,lGis->cmap,nv,NULL,PETSC_OWN_POINTER,idxs);
383: ISLocalToGlobalMappingRestoreIndices(vl2g,&idxs);
384: i = singular ? 2 : 1;
385: PetscMalloc2(i*nv,&sfvleaves,i*Lv,&sfvroots);
387: /* Destroy temporary G created in MATIS format and modified G */
388: PetscObjectReference((PetscObject)lG);
389: MatDestroy(&lGis);
390: MatDestroy(&G);
392: if (print) {
393: PetscObjectSetName((PetscObject)lG,"initial_lG");
394: MatView(lG,NULL);
395: }
397: /* Save lG for values insertion in change of basis */
398: MatDuplicate(lG,MAT_COPY_VALUES,&lGinit);
400: /* Analyze the edge-nodes connections (duplicate lG) */
401: MatDuplicate(lG,MAT_COPY_VALUES,&lGe);
402: MatSetOption(lGe,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);
403: PetscBTCreate(nv,&btv);
404: PetscBTCreate(ne,&bte);
405: PetscBTCreate(ne,&btb);
406: PetscBTCreate(ne,&btbd);
407: PetscBTCreate(nv,&btvcand);
408: /* need to import the boundary specification to ensure the
409: proper detection of coarse edges' endpoints */
410: if (pcbddc->DirichletBoundariesLocal) {
411: IS is;
413: if (fl2g) {
414: ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->DirichletBoundariesLocal,&is);
415: } else {
416: is = pcbddc->DirichletBoundariesLocal;
417: }
418: ISGetLocalSize(is,&cum);
419: ISGetIndices(is,&idxs);
420: for (i=0;i<cum;i++) {
421: if (idxs[i] >= 0) {
422: PetscBTSet(btb,idxs[i]);
423: PetscBTSet(btbd,idxs[i]);
424: }
425: }
426: ISRestoreIndices(is,&idxs);
427: if (fl2g) {
428: ISDestroy(&is);
429: }
430: }
431: if (pcbddc->NeumannBoundariesLocal) {
432: IS is;
434: if (fl2g) {
435: ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_MASK,pcbddc->NeumannBoundariesLocal,&is);
436: } else {
437: is = pcbddc->NeumannBoundariesLocal;
438: }
439: ISGetLocalSize(is,&cum);
440: ISGetIndices(is,&idxs);
441: for (i=0;i<cum;i++) {
442: if (idxs[i] >= 0) {
443: PetscBTSet(btb,idxs[i]);
444: }
445: }
446: ISRestoreIndices(is,&idxs);
447: if (fl2g) {
448: ISDestroy(&is);
449: }
450: }
452: /* Count neighs per dof */
453: ISLocalToGlobalMappingGetNodeInfo(el2g,NULL,&ecount,&eneighs);
454: ISLocalToGlobalMappingGetNodeInfo(vl2g,NULL,&vcount,&vneighs);
456: /* need to remove coarse faces' dofs and coarse edges' dirichlet dofs
457: for proper detection of coarse edges' endpoints */
458: PetscBTCreate(ne,&btee);
459: for (i=0;i<ne;i++) {
460: if ((ecount[i] > 2 && !PetscBTLookup(btbd,i)) || (ecount[i] == 2 && PetscBTLookup(btb,i))) {
461: PetscBTSet(btee,i);
462: }
463: }
464: PetscMalloc1(ne,&marks);
465: if (!conforming) {
466: MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);
467: MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
468: }
469: MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
470: MatSeqAIJGetArray(lGe,&vals);
471: cum = 0;
472: for (i=0;i<ne;i++) {
473: /* eliminate rows corresponding to edge dofs belonging to coarse faces */
474: if (!PetscBTLookup(btee,i)) {
475: marks[cum++] = i;
476: continue;
477: }
478: /* set badly connected edge dofs as primal */
479: if (!conforming) {
480: if (ii[i+1]-ii[i] != order + 1) { /* every row of G on the coarse edge should list order+1 nodal dofs */
481: marks[cum++] = i;
482: PetscBTSet(bte,i);
483: for (j=ii[i];j<ii[i+1];j++) {
484: PetscBTSet(btv,jj[j]);
485: }
486: } else {
487: /* every edge dofs should be connected trough a certain number of nodal dofs
488: to other edge dofs belonging to coarse edges
489: - at most 2 endpoints
490: - order-1 interior nodal dofs
491: - no undefined nodal dofs (nconn < order)
492: */
493: PetscInt ends = 0,ints = 0, undef = 0;
494: for (j=ii[i];j<ii[i+1];j++) {
495: PetscInt v = jj[j],k;
496: PetscInt nconn = iit[v+1]-iit[v];
497: for (k=iit[v];k<iit[v+1];k++) if (!PetscBTLookup(btee,jjt[k])) nconn--;
498: if (nconn > order) ends++;
499: else if (nconn == order) ints++;
500: else undef++;
501: }
502: if (undef || ends > 2 || ints != order -1) {
503: marks[cum++] = i;
504: PetscBTSet(bte,i);
505: for (j=ii[i];j<ii[i+1];j++) {
506: PetscBTSet(btv,jj[j]);
507: }
508: }
509: }
510: }
511: /* We assume the order on the element edge is ii[i+1]-ii[i]-1 */
512: if (!order && ii[i+1] != ii[i]) {
513: PetscScalar val = 1./(ii[i+1]-ii[i]-1);
514: for (j=ii[i];j<ii[i+1];j++) vals[j] = val;
515: }
516: }
517: PetscBTDestroy(&btee);
518: MatSeqAIJRestoreArray(lGe,&vals);
519: MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
520: if (!conforming) {
521: MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
522: MatDestroy(&lGt);
523: }
524: MatZeroRows(lGe,cum,marks,0.,NULL,NULL);
526: /* identify splitpoints and corner candidates */
527: MatTranspose(lGe,MAT_INITIAL_MATRIX,&lGt);
528: if (print) {
529: PetscObjectSetName((PetscObject)lGe,"edgerestr_lG");
530: MatView(lGe,NULL);
531: PetscObjectSetName((PetscObject)lGt,"edgerestr_lGt");
532: MatView(lGt,NULL);
533: }
534: MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
535: MatSeqAIJGetArray(lGt,&vals);
536: for (i=0;i<nv;i++) {
537: PetscInt ord = order, test = ii[i+1]-ii[i], vc = vcount[i];
538: PetscBool sneighs = PETSC_TRUE, bdir = PETSC_FALSE;
539: if (!order) { /* variable order */
540: PetscReal vorder = 0.;
542: for (j=ii[i];j<ii[i+1];j++) vorder += PetscRealPart(vals[j]);
543: test = PetscFloorReal(vorder+10.*PETSC_SQRT_MACHINE_EPSILON);
544: if (vorder-test > PETSC_SQRT_MACHINE_EPSILON) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected value for vorder: %g (%D)",vorder,test);
545: ord = 1;
546: }
547: #if defined(PETSC_USE_DEBUG)
548: if (test%ord) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected number of edge dofs %D connected with nodal dof %D with order %D",test,i,ord);
549: #endif
550: for (j=ii[i];j<ii[i+1] && sneighs;j++) {
551: if (PetscBTLookup(btbd,jj[j])) {
552: bdir = PETSC_TRUE;
553: break;
554: }
555: if (vc != ecount[jj[j]]) {
556: sneighs = PETSC_FALSE;
557: } else {
558: PetscInt k,*vn = vneighs[i], *en = eneighs[jj[j]];
559: for (k=0;k<vc;k++) {
560: if (vn[k] != en[k]) {
561: sneighs = PETSC_FALSE;
562: break;
563: }
564: }
565: }
566: }
567: if (!sneighs || test >= 3*ord || bdir) { /* splitpoints */
568: if (print) PetscPrintf(PETSC_COMM_SELF,"SPLITPOINT %D (%D %D %D)\n",i,!sneighs,test >= 3*ord,bdir);
569: PetscBTSet(btv,i);
570: } else if (test == ord) {
571: if (order == 1 || (!order && ii[i+1]-ii[i] == 1)) {
572: if (print) PetscPrintf(PETSC_COMM_SELF,"ENDPOINT %D\n",i);
573: PetscBTSet(btv,i);
574: } else {
575: if (print) PetscPrintf(PETSC_COMM_SELF,"CORNER CANDIDATE %D\n",i);
576: PetscBTSet(btvcand,i);
577: }
578: }
579: }
580: ISLocalToGlobalMappingRestoreNodeInfo(el2g,NULL,&ecount,&eneighs);
581: ISLocalToGlobalMappingRestoreNodeInfo(vl2g,NULL,&vcount,&vneighs);
582: PetscBTDestroy(&btbd);
584: /* a candidate is valid if it is connected to another candidate via a non-primal edge dof */
585: if (order != 1) {
586: if (print) PetscPrintf(PETSC_COMM_SELF,"INSPECTING CANDIDATES\n");
587: MatGetRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
588: for (i=0;i<nv;i++) {
589: if (PetscBTLookup(btvcand,i)) {
590: PetscBool found = PETSC_FALSE;
591: for (j=ii[i];j<ii[i+1] && !found;j++) {
592: PetscInt k,e = jj[j];
593: if (PetscBTLookup(bte,e)) continue;
594: for (k=iit[e];k<iit[e+1];k++) {
595: PetscInt v = jjt[k];
596: if (v != i && PetscBTLookup(btvcand,v)) {
597: found = PETSC_TRUE;
598: break;
599: }
600: }
601: }
602: if (!found) {
603: if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %D CLEARED\n",i);
604: PetscBTClear(btvcand,i);
605: } else {
606: if (print) PetscPrintf(PETSC_COMM_SELF," CANDIDATE %D ACCEPTED\n",i);
607: }
608: }
609: }
610: MatRestoreRowIJ(lGe,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
611: }
612: MatSeqAIJRestoreArray(lGt,&vals);
613: MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
614: MatDestroy(&lGe);
616: /* Get the local G^T explicitly */
617: MatDestroy(&lGt);
618: MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);
619: MatSetOption(lGt,MAT_KEEP_NONZERO_PATTERN,PETSC_FALSE);
621: /* Mark interior nodal dofs */
622: ISLocalToGlobalMappingGetInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);
623: PetscBTCreate(nv,&btvi);
624: for (i=1;i<n_neigh;i++) {
625: for (j=0;j<n_shared[i];j++) {
626: PetscBTSet(btvi,shared[i][j]);
627: }
628: }
629: ISLocalToGlobalMappingRestoreInfo(vl2g,&n_neigh,&neigh,&n_shared,&shared);
631: /* communicate corners and splitpoints */
632: PetscMalloc1(nv,&vmarks);
633: PetscArrayzero(sfvleaves,nv);
634: PetscArrayzero(sfvroots,Lv);
635: for (i=0;i<nv;i++) if (PetscUnlikely(PetscBTLookup(btv,i))) sfvleaves[i] = 1;
637: if (print) {
638: IS tbz;
640: cum = 0;
641: for (i=0;i<nv;i++)
642: if (sfvleaves[i])
643: vmarks[cum++] = i;
645: ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);
646: PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_local");
647: ISView(tbz,NULL);
648: ISDestroy(&tbz);
649: }
651: PetscSFReduceBegin(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);
652: PetscSFReduceEnd(sfv,MPIU_INT,sfvleaves,sfvroots,MPI_SUM);
653: PetscSFBcastBegin(sfv,MPIU_INT,sfvroots,sfvleaves);
654: PetscSFBcastEnd(sfv,MPIU_INT,sfvroots,sfvleaves);
656: /* Zero rows of lGt corresponding to identified corners
657: and interior nodal dofs */
658: cum = 0;
659: for (i=0;i<nv;i++) {
660: if (sfvleaves[i]) {
661: vmarks[cum++] = i;
662: PetscBTSet(btv,i);
663: }
664: if (!PetscBTLookup(btvi,i)) vmarks[cum++] = i;
665: }
666: PetscBTDestroy(&btvi);
667: if (print) {
668: IS tbz;
670: ISCreateGeneral(PETSC_COMM_SELF,cum,vmarks,PETSC_COPY_VALUES,&tbz);
671: PetscObjectSetName((PetscObject)tbz,"corners_to_be_zeroed_with_interior");
672: ISView(tbz,NULL);
673: ISDestroy(&tbz);
674: }
675: MatZeroRows(lGt,cum,vmarks,0.,NULL,NULL);
676: PetscFree(vmarks);
677: PetscSFDestroy(&sfv);
678: PetscFree2(sfvleaves,sfvroots);
680: /* Recompute G */
681: MatDestroy(&lG);
682: MatTranspose(lGt,MAT_INITIAL_MATRIX,&lG);
683: if (print) {
684: PetscObjectSetName((PetscObject)lG,"used_lG");
685: MatView(lG,NULL);
686: PetscObjectSetName((PetscObject)lGt,"used_lGt");
687: MatView(lGt,NULL);
688: }
690: /* Get primal dofs (if any) */
691: cum = 0;
692: for (i=0;i<ne;i++) {
693: if (PetscUnlikely(PetscBTLookup(bte,i))) marks[cum++] = i;
694: }
695: if (fl2g) {
696: ISLocalToGlobalMappingApply(fl2g,cum,marks,marks);
697: }
698: ISCreateGeneral(comm,cum,marks,PETSC_COPY_VALUES,&primals);
699: if (print) {
700: PetscObjectSetName((PetscObject)primals,"prescribed_primal_dofs");
701: ISView(primals,NULL);
702: }
703: PetscBTDestroy(&bte);
704: /* TODO: what if the user passed in some of them ? */
705: PCBDDCSetPrimalVerticesLocalIS(pc,primals);
706: ISDestroy(&primals);
708: /* Compute edge connectivity */
709: PetscObjectSetOptionsPrefix((PetscObject)lG,"econn_");
711: /* Symbolic conn = lG*lGt */
712: MatProductCreate(lG,lGt,NULL,&conn);
713: MatProductSetType(conn,MATPRODUCT_AB);
714: MatProductSetAlgorithm(conn,"default");
715: MatProductSetFill(conn,PETSC_DEFAULT);
716: PetscObjectSetOptionsPrefix((PetscObject)conn,"econn_");
717: MatProductSetFromOptions(conn);
718: MatProductSymbolic(conn);
720: MatGetRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
721: if (fl2g) {
722: PetscBT btf;
723: PetscInt *iia,*jja,*iiu,*jju;
724: PetscBool rest = PETSC_FALSE,free = PETSC_FALSE;
726: /* create CSR for all local dofs */
727: PetscMalloc1(n+1,&iia);
728: if (pcbddc->mat_graph->nvtxs_csr) { /* the user has passed in a CSR graph */
729: if (pcbddc->mat_graph->nvtxs_csr != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid size of CSR graph %D. Should be %D",pcbddc->mat_graph->nvtxs_csr,n);
730: iiu = pcbddc->mat_graph->xadj;
731: jju = pcbddc->mat_graph->adjncy;
732: } else if (pcbddc->use_local_adj) {
733: rest = PETSC_TRUE;
734: MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);
735: } else {
736: free = PETSC_TRUE;
737: PetscMalloc2(n+1,&iiu,n,&jju);
738: iiu[0] = 0;
739: for (i=0;i<n;i++) {
740: iiu[i+1] = i+1;
741: jju[i] = -1;
742: }
743: }
745: /* import sizes of CSR */
746: iia[0] = 0;
747: for (i=0;i<n;i++) iia[i+1] = iiu[i+1]-iiu[i];
749: /* overwrite entries corresponding to the Nedelec field */
750: PetscBTCreate(n,&btf);
751: ISGetIndices(nedfieldlocal,&idxs);
752: for (i=0;i<ne;i++) {
753: PetscBTSet(btf,idxs[i]);
754: iia[idxs[i]+1] = ii[i+1]-ii[i];
755: }
757: /* iia in CSR */
758: for (i=0;i<n;i++) iia[i+1] += iia[i];
760: /* jja in CSR */
761: PetscMalloc1(iia[n],&jja);
762: for (i=0;i<n;i++)
763: if (!PetscBTLookup(btf,i))
764: for (j=0;j<iiu[i+1]-iiu[i];j++)
765: jja[iia[i]+j] = jju[iiu[i]+j];
767: /* map edge dofs connectivity */
768: if (jj) {
769: ISLocalToGlobalMappingApply(fl2g,ii[ne],jj,(PetscInt *)jj);
770: for (i=0;i<ne;i++) {
771: PetscInt e = idxs[i];
772: for (j=0;j<ii[i+1]-ii[i];j++) jja[iia[e]+j] = jj[ii[i]+j];
773: }
774: }
775: ISRestoreIndices(nedfieldlocal,&idxs);
776: PCBDDCSetLocalAdjacencyGraph(pc,n,iia,jja,PETSC_OWN_POINTER);
777: if (rest) {
778: MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&i,(const PetscInt**)&iiu,(const PetscInt**)&jju,&done);
779: }
780: if (free) {
781: PetscFree2(iiu,jju);
782: }
783: PetscBTDestroy(&btf);
784: } else {
785: PCBDDCSetLocalAdjacencyGraph(pc,n,ii,jj,PETSC_USE_POINTER);
786: }
788: /* Analyze interface for edge dofs */
789: PCBDDCAnalyzeInterface(pc);
790: pcbddc->mat_graph->twodim = PETSC_FALSE;
792: /* Get coarse edges in the edge space */
793: PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);
794: MatRestoreRowIJ(conn,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
796: if (fl2g) {
797: ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);
798: PetscMalloc1(nee,&eedges);
799: for (i=0;i<nee;i++) {
800: ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);
801: }
802: } else {
803: eedges = alleedges;
804: primals = allprimals;
805: }
807: /* Mark fine edge dofs with their coarse edge id */
808: PetscArrayzero(marks,ne);
809: ISGetLocalSize(primals,&cum);
810: ISGetIndices(primals,&idxs);
811: for (i=0;i<cum;i++) marks[idxs[i]] = nee+1;
812: ISRestoreIndices(primals,&idxs);
813: if (print) {
814: PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs");
815: ISView(primals,NULL);
816: }
818: maxsize = 0;
819: for (i=0;i<nee;i++) {
820: PetscInt size,mark = i+1;
822: ISGetLocalSize(eedges[i],&size);
823: ISGetIndices(eedges[i],&idxs);
824: for (j=0;j<size;j++) marks[idxs[j]] = mark;
825: ISRestoreIndices(eedges[i],&idxs);
826: maxsize = PetscMax(maxsize,size);
827: }
829: /* Find coarse edge endpoints */
830: MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
831: MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
832: for (i=0;i<nee;i++) {
833: PetscInt mark = i+1,size;
835: ISGetLocalSize(eedges[i],&size);
836: if (!size && nedfieldlocal) continue;
837: if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i);
838: ISGetIndices(eedges[i],&idxs);
839: if (print) {
840: PetscPrintf(PETSC_COMM_SELF,"ENDPOINTS ANALYSIS EDGE %D\n",i);
841: ISView(eedges[i],NULL);
842: }
843: for (j=0;j<size;j++) {
844: PetscInt k, ee = idxs[j];
845: if (print) PetscPrintf(PETSC_COMM_SELF," idx %D\n",ee);
846: for (k=ii[ee];k<ii[ee+1];k++) {
847: if (print) PetscPrintf(PETSC_COMM_SELF," inspect %D\n",jj[k]);
848: if (PetscBTLookup(btv,jj[k])) {
849: if (print) PetscPrintf(PETSC_COMM_SELF," corner found (already set) %D\n",jj[k]);
850: } else if (PetscBTLookup(btvcand,jj[k])) { /* is it ok? */
851: PetscInt k2;
852: PetscBool corner = PETSC_FALSE;
853: for (k2 = iit[jj[k]];k2 < iit[jj[k]+1];k2++) {
854: if (print) PetscPrintf(PETSC_COMM_SELF," INSPECTING %D: mark %D (ref mark %D), boundary %D\n",jjt[k2],marks[jjt[k2]],mark,!!PetscBTLookup(btb,jjt[k2]));
855: /* it's a corner if either is connected with an edge dof belonging to a different cc or
856: if the edge dof lie on the natural part of the boundary */
857: if ((marks[jjt[k2]] && marks[jjt[k2]] != mark) || (!marks[jjt[k2]] && PetscBTLookup(btb,jjt[k2]))) {
858: corner = PETSC_TRUE;
859: break;
860: }
861: }
862: if (corner) { /* found the nodal dof corresponding to the endpoint of the edge */
863: if (print) PetscPrintf(PETSC_COMM_SELF," corner found %D\n",jj[k]);
864: PetscBTSet(btv,jj[k]);
865: } else {
866: if (print) PetscPrintf(PETSC_COMM_SELF," no corners found\n");
867: }
868: }
869: }
870: }
871: ISRestoreIndices(eedges[i],&idxs);
872: }
873: MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
874: MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
875: PetscBTDestroy(&btb);
877: /* Reset marked primal dofs */
878: ISGetLocalSize(primals,&cum);
879: ISGetIndices(primals,&idxs);
880: for (i=0;i<cum;i++) marks[idxs[i]] = 0;
881: ISRestoreIndices(primals,&idxs);
883: /* Now use the initial lG */
884: MatDestroy(&lG);
885: MatDestroy(&lGt);
886: lG = lGinit;
887: MatTranspose(lG,MAT_INITIAL_MATRIX,&lGt);
889: /* Compute extended cols indices */
890: PetscBTCreate(nv,&btvc);
891: PetscBTCreate(nee,&bter);
892: MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
893: MatSeqAIJGetMaxRowNonzeros(lG,&i);
894: i *= maxsize;
895: PetscCalloc1(nee,&extcols);
896: PetscMalloc2(i,&extrow,i,&gidxs);
897: eerr = PETSC_FALSE;
898: for (i=0;i<nee;i++) {
899: PetscInt size,found = 0;
901: cum = 0;
902: ISGetLocalSize(eedges[i],&size);
903: if (!size && nedfieldlocal) continue;
904: if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i);
905: ISGetIndices(eedges[i],&idxs);
906: PetscBTMemzero(nv,btvc);
907: for (j=0;j<size;j++) {
908: PetscInt k,ee = idxs[j];
909: for (k=ii[ee];k<ii[ee+1];k++) {
910: PetscInt vv = jj[k];
911: if (!PetscBTLookup(btv,vv)) extrow[cum++] = vv;
912: else if (!PetscBTLookupSet(btvc,vv)) found++;
913: }
914: }
915: ISRestoreIndices(eedges[i],&idxs);
916: PetscSortRemoveDupsInt(&cum,extrow);
917: ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);
918: PetscSortIntWithArray(cum,gidxs,extrow);
919: ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);
920: /* it may happen that endpoints are not defined at this point
921: if it is the case, mark this edge for a second pass */
922: if (cum != size -1 || found != 2) {
923: PetscBTSet(bter,i);
924: if (print) {
925: PetscObjectSetName((PetscObject)eedges[i],"error_edge");
926: ISView(eedges[i],NULL);
927: PetscObjectSetName((PetscObject)extcols[i],"error_extcol");
928: ISView(extcols[i],NULL);
929: }
930: eerr = PETSC_TRUE;
931: }
932: }
933: /* if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL FIRST PASS"); */
934: MPIU_Allreduce(&eerr,&done,1,MPIU_BOOL,MPI_LOR,comm);
935: if (done) {
936: PetscInt *newprimals;
938: PetscMalloc1(ne,&newprimals);
939: ISGetLocalSize(primals,&cum);
940: ISGetIndices(primals,&idxs);
941: PetscArraycpy(newprimals,idxs,cum);
942: ISRestoreIndices(primals,&idxs);
943: MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
944: if (print) PetscPrintf(PETSC_COMM_SELF,"DOING SECOND PASS (eerr %D)\n",eerr);
945: for (i=0;i<nee;i++) {
946: PetscBool has_candidates = PETSC_FALSE;
947: if (PetscBTLookup(bter,i)) {
948: PetscInt size,mark = i+1;
950: ISGetLocalSize(eedges[i],&size);
951: ISGetIndices(eedges[i],&idxs);
952: /* for (j=0;j<size;j++) newprimals[cum++] = idxs[j]; */
953: for (j=0;j<size;j++) {
954: PetscInt k,ee = idxs[j];
955: if (print) PetscPrintf(PETSC_COMM_SELF,"Inspecting edge dof %D [%D %D)\n",ee,ii[ee],ii[ee+1]);
956: for (k=ii[ee];k<ii[ee+1];k++) {
957: /* set all candidates located on the edge as corners */
958: if (PetscBTLookup(btvcand,jj[k])) {
959: PetscInt k2,vv = jj[k];
960: has_candidates = PETSC_TRUE;
961: if (print) PetscPrintf(PETSC_COMM_SELF," Candidate set to vertex %D\n",vv);
962: PetscBTSet(btv,vv);
963: /* set all edge dofs connected to candidate as primals */
964: for (k2=iit[vv];k2<iit[vv+1];k2++) {
965: if (marks[jjt[k2]] == mark) {
966: PetscInt k3,ee2 = jjt[k2];
967: if (print) PetscPrintf(PETSC_COMM_SELF," Connected edge dof set to primal %D\n",ee2);
968: newprimals[cum++] = ee2;
969: /* finally set the new corners */
970: for (k3=ii[ee2];k3<ii[ee2+1];k3++) {
971: if (print) PetscPrintf(PETSC_COMM_SELF," Connected nodal dof set to vertex %D\n",jj[k3]);
972: PetscBTSet(btv,jj[k3]);
973: }
974: }
975: }
976: } else {
977: if (print) PetscPrintf(PETSC_COMM_SELF," Not a candidate vertex %D\n",jj[k]);
978: }
979: }
980: }
981: if (!has_candidates) { /* circular edge */
982: PetscInt k, ee = idxs[0],*tmarks;
984: PetscCalloc1(ne,&tmarks);
985: if (print) PetscPrintf(PETSC_COMM_SELF," Circular edge %D\n",i);
986: for (k=ii[ee];k<ii[ee+1];k++) {
987: PetscInt k2;
988: if (print) PetscPrintf(PETSC_COMM_SELF," Set to corner %D\n",jj[k]);
989: PetscBTSet(btv,jj[k]);
990: for (k2=iit[jj[k]];k2<iit[jj[k]+1];k2++) tmarks[jjt[k2]]++;
991: }
992: for (j=0;j<size;j++) {
993: if (tmarks[idxs[j]] > 1) {
994: if (print) PetscPrintf(PETSC_COMM_SELF," Edge dof set to primal %D\n",idxs[j]);
995: newprimals[cum++] = idxs[j];
996: }
997: }
998: PetscFree(tmarks);
999: }
1000: ISRestoreIndices(eedges[i],&idxs);
1001: }
1002: ISDestroy(&extcols[i]);
1003: }
1004: PetscFree(extcols);
1005: MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&iit,&jjt,&done);
1006: PetscSortRemoveDupsInt(&cum,newprimals);
1007: if (fl2g) {
1008: ISLocalToGlobalMappingApply(fl2g,cum,newprimals,newprimals);
1009: ISDestroy(&primals);
1010: for (i=0;i<nee;i++) {
1011: ISDestroy(&eedges[i]);
1012: }
1013: PetscFree(eedges);
1014: }
1015: PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);
1016: ISCreateGeneral(comm,cum,newprimals,PETSC_COPY_VALUES,&primals);
1017: PetscFree(newprimals);
1018: PCBDDCSetPrimalVerticesLocalIS(pc,primals);
1019: ISDestroy(&primals);
1020: PCBDDCAnalyzeInterface(pc);
1021: pcbddc->mat_graph->twodim = PETSC_FALSE;
1022: PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);
1023: if (fl2g) {
1024: ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,allprimals,&primals);
1025: PetscMalloc1(nee,&eedges);
1026: for (i=0;i<nee;i++) {
1027: ISGlobalToLocalMappingApplyIS(fl2g,IS_GTOLM_DROP,alleedges[i],&eedges[i]);
1028: }
1029: } else {
1030: eedges = alleedges;
1031: primals = allprimals;
1032: }
1033: PetscCalloc1(nee,&extcols);
1035: /* Mark again */
1036: PetscArrayzero(marks,ne);
1037: for (i=0;i<nee;i++) {
1038: PetscInt size,mark = i+1;
1040: ISGetLocalSize(eedges[i],&size);
1041: ISGetIndices(eedges[i],&idxs);
1042: for (j=0;j<size;j++) marks[idxs[j]] = mark;
1043: ISRestoreIndices(eedges[i],&idxs);
1044: }
1045: if (print) {
1046: PetscObjectSetName((PetscObject)primals,"obtained_primal_dofs_secondpass");
1047: ISView(primals,NULL);
1048: }
1050: /* Recompute extended cols */
1051: eerr = PETSC_FALSE;
1052: for (i=0;i<nee;i++) {
1053: PetscInt size;
1055: cum = 0;
1056: ISGetLocalSize(eedges[i],&size);
1057: if (!size && nedfieldlocal) continue;
1058: if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i);
1059: ISGetIndices(eedges[i],&idxs);
1060: for (j=0;j<size;j++) {
1061: PetscInt k,ee = idxs[j];
1062: for (k=ii[ee];k<ii[ee+1];k++) if (!PetscBTLookup(btv,jj[k])) extrow[cum++] = jj[k];
1063: }
1064: ISRestoreIndices(eedges[i],&idxs);
1065: PetscSortRemoveDupsInt(&cum,extrow);
1066: ISLocalToGlobalMappingApply(vl2g,cum,extrow,gidxs);
1067: PetscSortIntWithArray(cum,gidxs,extrow);
1068: ISCreateGeneral(PETSC_COMM_SELF,cum,extrow,PETSC_COPY_VALUES,&extcols[i]);
1069: if (cum != size -1) {
1070: if (print) {
1071: PetscObjectSetName((PetscObject)eedges[i],"error_edge_secondpass");
1072: ISView(eedges[i],NULL);
1073: PetscObjectSetName((PetscObject)extcols[i],"error_extcol_secondpass");
1074: ISView(extcols[i],NULL);
1075: }
1076: eerr = PETSC_TRUE;
1077: }
1078: }
1079: }
1080: MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
1081: PetscFree2(extrow,gidxs);
1082: PetscBTDestroy(&bter);
1083: if (print) { PCBDDCGraphASCIIView(pcbddc->mat_graph,5,PETSC_VIEWER_STDOUT_SELF); }
1084: /* an error should not occur at this point */
1085: if (eerr) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected SIZE OF EDGE > EXTCOL SECOND PASS");
1087: /* Check the number of endpoints */
1088: MatGetRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
1089: PetscMalloc1(2*nee,&corners);
1090: PetscMalloc1(nee,&cedges);
1091: for (i=0;i<nee;i++) {
1092: PetscInt size, found = 0, gc[2];
1094: /* init with defaults */
1095: cedges[i] = corners[i*2] = corners[i*2+1] = -1;
1096: ISGetLocalSize(eedges[i],&size);
1097: if (!size && nedfieldlocal) continue;
1098: if (!size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Unexpected zero sized edge %D",i);
1099: ISGetIndices(eedges[i],&idxs);
1100: PetscBTMemzero(nv,btvc);
1101: for (j=0;j<size;j++) {
1102: PetscInt k,ee = idxs[j];
1103: for (k=ii[ee];k<ii[ee+1];k++) {
1104: PetscInt vv = jj[k];
1105: if (PetscBTLookup(btv,vv) && !PetscBTLookupSet(btvc,vv)) {
1106: if (found == 2) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found more then two corners for edge %D",i);
1107: corners[i*2+found++] = vv;
1108: }
1109: }
1110: }
1111: if (found != 2) {
1112: PetscInt e;
1113: if (fl2g) {
1114: ISLocalToGlobalMappingApply(fl2g,1,idxs,&e);
1115: } else {
1116: e = idxs[0];
1117: }
1118: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Found %D corners for edge %D (astart %D, estart %D)",found,i,e,idxs[0]);
1119: }
1121: /* get primal dof index on this coarse edge */
1122: ISLocalToGlobalMappingApply(vl2g,2,corners+2*i,gc);
1123: if (gc[0] > gc[1]) {
1124: PetscInt swap = corners[2*i];
1125: corners[2*i] = corners[2*i+1];
1126: corners[2*i+1] = swap;
1127: }
1128: cedges[i] = idxs[size-1];
1129: ISRestoreIndices(eedges[i],&idxs);
1130: if (print) PetscPrintf(PETSC_COMM_SELF,"EDGE %D: ce %D, corners (%D,%D)\n",i,cedges[i],corners[2*i],corners[2*i+1]);
1131: }
1132: MatRestoreRowIJ(lG,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
1133: PetscBTDestroy(&btvc);
1135: #if defined(PETSC_USE_DEBUG)
1136: /* Inspects columns of lG (rows of lGt) and make sure the change of basis will
1137: not interfere with neighbouring coarse edges */
1138: PetscMalloc1(nee+1,&emarks);
1139: MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
1140: for (i=0;i<nv;i++) {
1141: PetscInt emax = 0,eemax = 0;
1143: if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue;
1144: PetscArrayzero(emarks,nee+1);
1145: for (j=ii[i];j<ii[i+1];j++) emarks[marks[jj[j]]]++;
1146: for (j=1;j<nee+1;j++) {
1147: if (emax < emarks[j]) {
1148: emax = emarks[j];
1149: eemax = j;
1150: }
1151: }
1152: /* not relevant for edges */
1153: if (!eemax) continue;
1155: for (j=ii[i];j<ii[i+1];j++) {
1156: if (marks[jj[j]] && marks[jj[j]] != eemax) {
1157: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_SUP,"Found 2 coarse edges (id %D and %D) connected through the %D nodal dof at edge dof %D",marks[jj[j]]-1,eemax,i,jj[j]);
1158: }
1159: }
1160: }
1161: PetscFree(emarks);
1162: MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
1163: #endif
1165: /* Compute extended rows indices for edge blocks of the change of basis */
1166: MatGetRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
1167: MatSeqAIJGetMaxRowNonzeros(lGt,&extmem);
1168: extmem *= maxsize;
1169: PetscMalloc1(extmem*nee,&extrow);
1170: PetscMalloc1(nee,&extrows);
1171: PetscCalloc1(nee,&extrowcum);
1172: for (i=0;i<nv;i++) {
1173: PetscInt mark = 0,size,start;
1175: if (ii[i+1]==ii[i] || PetscBTLookup(btv,i)) continue;
1176: for (j=ii[i];j<ii[i+1];j++)
1177: if (marks[jj[j]] && !mark)
1178: mark = marks[jj[j]];
1180: /* not relevant */
1181: if (!mark) continue;
1183: /* import extended row */
1184: mark--;
1185: start = mark*extmem+extrowcum[mark];
1186: size = ii[i+1]-ii[i];
1187: if (extrowcum[mark] + size > extmem) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Not enough memory allocated %D > %D",extrowcum[mark] + size,extmem);
1188: PetscArraycpy(extrow+start,jj+ii[i],size);
1189: extrowcum[mark] += size;
1190: }
1191: MatRestoreRowIJ(lGt,0,PETSC_FALSE,PETSC_FALSE,&i,&ii,&jj,&done);
1192: MatDestroy(&lGt);
1193: PetscFree(marks);
1195: /* Compress extrows */
1196: cum = 0;
1197: for (i=0;i<nee;i++) {
1198: PetscInt size = extrowcum[i],*start = extrow + i*extmem;
1199: PetscSortRemoveDupsInt(&size,start);
1200: ISCreateGeneral(PETSC_COMM_SELF,size,start,PETSC_USE_POINTER,&extrows[i]);
1201: cum = PetscMax(cum,size);
1202: }
1203: PetscFree(extrowcum);
1204: PetscBTDestroy(&btv);
1205: PetscBTDestroy(&btvcand);
1207: /* Workspace for lapack inner calls and VecSetValues */
1208: PetscMalloc2((5+cum+maxsize)*maxsize,&work,maxsize,&rwork);
1210: /* Create change of basis matrix (preallocation can be improved) */
1211: MatCreate(comm,&T);
1212: MatSetSizes(T,pc->pmat->rmap->n,pc->pmat->rmap->n,
1213: pc->pmat->rmap->N,pc->pmat->rmap->N);
1214: MatSetType(T,MATAIJ);
1215: MatSeqAIJSetPreallocation(T,10,NULL);
1216: MatMPIAIJSetPreallocation(T,10,NULL,10,NULL);
1217: MatSetLocalToGlobalMapping(T,al2g,al2g);
1218: MatSetOption(T,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
1219: MatSetOption(T,MAT_ROW_ORIENTED,PETSC_FALSE);
1220: ISLocalToGlobalMappingDestroy(&al2g);
1222: /* Defaults to identity */
1223: MatCreateVecs(pc->pmat,&tvec,NULL);
1224: VecSet(tvec,1.0);
1225: MatDiagonalSet(T,tvec,INSERT_VALUES);
1226: VecDestroy(&tvec);
1228: /* Create discrete gradient for the coarser level if needed */
1229: MatDestroy(&pcbddc->nedcG);
1230: ISDestroy(&pcbddc->nedclocal);
1231: if (pcbddc->current_level < pcbddc->max_levels) {
1232: ISLocalToGlobalMapping cel2g,cvl2g;
1233: IS wis,gwis;
1234: PetscInt cnv,cne;
1236: ISCreateGeneral(comm,nee,cedges,PETSC_COPY_VALUES,&wis);
1237: if (fl2g) {
1238: ISLocalToGlobalMappingApplyIS(fl2g,wis,&pcbddc->nedclocal);
1239: } else {
1240: PetscObjectReference((PetscObject)wis);
1241: pcbddc->nedclocal = wis;
1242: }
1243: ISLocalToGlobalMappingApplyIS(el2g,wis,&gwis);
1244: ISDestroy(&wis);
1245: ISRenumber(gwis,NULL,&cne,&wis);
1246: ISLocalToGlobalMappingCreateIS(wis,&cel2g);
1247: ISDestroy(&wis);
1248: ISDestroy(&gwis);
1250: ISCreateGeneral(comm,2*nee,corners,PETSC_USE_POINTER,&wis);
1251: ISLocalToGlobalMappingApplyIS(vl2g,wis,&gwis);
1252: ISDestroy(&wis);
1253: ISRenumber(gwis,NULL,&cnv,&wis);
1254: ISLocalToGlobalMappingCreateIS(wis,&cvl2g);
1255: ISDestroy(&wis);
1256: ISDestroy(&gwis);
1258: MatCreate(comm,&pcbddc->nedcG);
1259: MatSetSizes(pcbddc->nedcG,PETSC_DECIDE,PETSC_DECIDE,cne,cnv);
1260: MatSetType(pcbddc->nedcG,MATAIJ);
1261: MatSeqAIJSetPreallocation(pcbddc->nedcG,2,NULL);
1262: MatMPIAIJSetPreallocation(pcbddc->nedcG,2,NULL,2,NULL);
1263: MatSetLocalToGlobalMapping(pcbddc->nedcG,cel2g,cvl2g);
1264: ISLocalToGlobalMappingDestroy(&cel2g);
1265: ISLocalToGlobalMappingDestroy(&cvl2g);
1266: }
1267: ISLocalToGlobalMappingDestroy(&vl2g);
1269: #if defined(PRINT_GDET)
1270: inc = 0;
1271: lev = pcbddc->current_level;
1272: #endif
1274: /* Insert values in the change of basis matrix */
1275: for (i=0;i<nee;i++) {
1276: Mat Gins = NULL, GKins = NULL;
1277: IS cornersis = NULL;
1278: PetscScalar cvals[2];
1280: if (pcbddc->nedcG) {
1281: ISCreateGeneral(PETSC_COMM_SELF,2,corners+2*i,PETSC_USE_POINTER,&cornersis);
1282: }
1283: PCBDDCComputeNedelecChangeEdge(lG,eedges[i],extrows[i],extcols[i],cornersis,&Gins,&GKins,cvals,work,rwork);
1284: if (Gins && GKins) {
1285: const PetscScalar *data;
1286: const PetscInt *rows,*cols;
1287: PetscInt nrh,nch,nrc,ncc;
1289: ISGetIndices(eedges[i],&cols);
1290: /* H1 */
1291: ISGetIndices(extrows[i],&rows);
1292: MatGetSize(Gins,&nrh,&nch);
1293: MatDenseGetArrayRead(Gins,&data);
1294: MatSetValuesLocal(T,nrh,rows,nch,cols,data,INSERT_VALUES);
1295: MatDenseRestoreArrayRead(Gins,&data);
1296: ISRestoreIndices(extrows[i],&rows);
1297: /* complement */
1298: MatGetSize(GKins,&nrc,&ncc);
1299: if (!ncc) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Constant function has not been generated for coarse edge %D",i);
1300: if (ncc + nch != nrc) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB,"The sum of the number of columns of GKins %D and Gins %D does not match %D for coarse edge %D",ncc,nch,nrc,i);
1301: if (ncc != 1 && pcbddc->nedcG) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot generate the coarse discrete gradient for coarse edge %D with ncc %D",i,ncc);
1302: MatDenseGetArrayRead(GKins,&data);
1303: MatSetValuesLocal(T,nrc,cols,ncc,cols+nch,data,INSERT_VALUES);
1304: MatDenseRestoreArrayRead(GKins,&data);
1306: /* coarse discrete gradient */
1307: if (pcbddc->nedcG) {
1308: PetscInt cols[2];
1310: cols[0] = 2*i;
1311: cols[1] = 2*i+1;
1312: MatSetValuesLocal(pcbddc->nedcG,1,&i,2,cols,cvals,INSERT_VALUES);
1313: }
1314: ISRestoreIndices(eedges[i],&cols);
1315: }
1316: ISDestroy(&extrows[i]);
1317: ISDestroy(&extcols[i]);
1318: ISDestroy(&cornersis);
1319: MatDestroy(&Gins);
1320: MatDestroy(&GKins);
1321: }
1322: ISLocalToGlobalMappingDestroy(&el2g);
1324: /* Start assembling */
1325: MatAssemblyBegin(T,MAT_FINAL_ASSEMBLY);
1326: if (pcbddc->nedcG) {
1327: MatAssemblyBegin(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);
1328: }
1330: /* Free */
1331: if (fl2g) {
1332: ISDestroy(&primals);
1333: for (i=0;i<nee;i++) {
1334: ISDestroy(&eedges[i]);
1335: }
1336: PetscFree(eedges);
1337: }
1339: /* hack mat_graph with primal dofs on the coarse edges */
1340: {
1341: PCBDDCGraph graph = pcbddc->mat_graph;
1342: PetscInt *oqueue = graph->queue;
1343: PetscInt *ocptr = graph->cptr;
1344: PetscInt ncc,*idxs;
1346: /* find first primal edge */
1347: if (pcbddc->nedclocal) {
1348: ISGetIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);
1349: } else {
1350: if (fl2g) {
1351: ISLocalToGlobalMappingApply(fl2g,nee,cedges,cedges);
1352: }
1353: idxs = cedges;
1354: }
1355: cum = 0;
1356: while (cum < nee && cedges[cum] < 0) cum++;
1358: /* adapt connected components */
1359: PetscMalloc2(graph->nvtxs+1,&graph->cptr,ocptr[graph->ncc],&graph->queue);
1360: graph->cptr[0] = 0;
1361: for (i=0,ncc=0;i<graph->ncc;i++) {
1362: PetscInt lc = ocptr[i+1]-ocptr[i];
1363: if (cum != nee && oqueue[ocptr[i+1]-1] == cedges[cum]) { /* this cc has a primal dof */
1364: graph->cptr[ncc+1] = graph->cptr[ncc]+1;
1365: graph->queue[graph->cptr[ncc]] = cedges[cum];
1366: ncc++;
1367: lc--;
1368: cum++;
1369: while (cum < nee && cedges[cum] < 0) cum++;
1370: }
1371: graph->cptr[ncc+1] = graph->cptr[ncc] + lc;
1372: for (j=0;j<lc;j++) graph->queue[graph->cptr[ncc]+j] = oqueue[ocptr[i]+j];
1373: ncc++;
1374: }
1375: graph->ncc = ncc;
1376: if (pcbddc->nedclocal) {
1377: ISRestoreIndices(pcbddc->nedclocal,(const PetscInt**)&idxs);
1378: }
1379: PetscFree2(ocptr,oqueue);
1380: }
1381: ISLocalToGlobalMappingDestroy(&fl2g);
1382: PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,&nee,&alleedges,&allprimals);
1383: PCBDDCGraphResetCSR(pcbddc->mat_graph);
1384: MatDestroy(&conn);
1386: ISDestroy(&nedfieldlocal);
1387: PetscFree(extrow);
1388: PetscFree2(work,rwork);
1389: PetscFree(corners);
1390: PetscFree(cedges);
1391: PetscFree(extrows);
1392: PetscFree(extcols);
1393: MatDestroy(&lG);
1395: /* Complete assembling */
1396: MatAssemblyEnd(T,MAT_FINAL_ASSEMBLY);
1397: if (pcbddc->nedcG) {
1398: MatAssemblyEnd(pcbddc->nedcG,MAT_FINAL_ASSEMBLY);
1399: #if 0
1400: PetscObjectSetName((PetscObject)pcbddc->nedcG,"coarse_G");
1401: MatView(pcbddc->nedcG,NULL);
1402: #endif
1403: }
1405: /* set change of basis */
1406: PCBDDCSetChangeOfBasisMat(pc,T,singular);
1407: MatDestroy(&T);
1409: return(0);
1410: }
1412: /* the near-null space of BDDC carries information on quadrature weights,
1413: and these can be collinear -> so cheat with MatNullSpaceCreate
1414: and create a suitable set of basis vectors first */
1415: PetscErrorCode PCBDDCNullSpaceCreate(MPI_Comm comm, PetscBool has_const, PetscInt nvecs, Vec quad_vecs[], MatNullSpace *nnsp)
1416: {
1418: PetscInt i;
1421: for (i=0;i<nvecs;i++) {
1422: PetscInt first,last;
1424: VecGetOwnershipRange(quad_vecs[i],&first,&last);
1425: if (last-first < 2*nvecs && has_const) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not implemented");
1426: if (i>=first && i < last) {
1427: PetscScalar *data;
1428: VecGetArray(quad_vecs[i],&data);
1429: if (!has_const) {
1430: data[i-first] = 1.;
1431: } else {
1432: data[2*i-first] = 1./PetscSqrtReal(2.);
1433: data[2*i-first+1] = -1./PetscSqrtReal(2.);
1434: }
1435: VecRestoreArray(quad_vecs[i],&data);
1436: }
1437: PetscObjectStateIncrease((PetscObject)quad_vecs[i]);
1438: }
1439: MatNullSpaceCreate(comm,has_const,nvecs,quad_vecs,nnsp);
1440: for (i=0;i<nvecs;i++) { /* reset vectors */
1441: PetscInt first,last;
1442: VecLockReadPop(quad_vecs[i]);
1443: VecGetOwnershipRange(quad_vecs[i],&first,&last);
1444: if (i>=first && i < last) {
1445: PetscScalar *data;
1446: VecGetArray(quad_vecs[i],&data);
1447: if (!has_const) {
1448: data[i-first] = 0.;
1449: } else {
1450: data[2*i-first] = 0.;
1451: data[2*i-first+1] = 0.;
1452: }
1453: VecRestoreArray(quad_vecs[i],&data);
1454: }
1455: PetscObjectStateIncrease((PetscObject)quad_vecs[i]);
1456: VecLockReadPush(quad_vecs[i]);
1457: }
1458: return(0);
1459: }
1461: PetscErrorCode PCBDDCComputeNoNetFlux(Mat A, Mat divudotp, PetscBool transpose, IS vl2l, PCBDDCGraph graph, MatNullSpace *nnsp)
1462: {
1463: Mat loc_divudotp;
1464: Vec p,v,vins,quad_vec,*quad_vecs;
1465: ISLocalToGlobalMapping map;
1466: PetscScalar *vals;
1467: const PetscScalar *array;
1468: PetscInt i,maxneighs = 0,maxsize,*gidxs;
1469: PetscInt n_neigh,*neigh,*n_shared,**shared;
1470: PetscMPIInt rank;
1471: PetscErrorCode ierr;
1474: ISLocalToGlobalMappingGetInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);
1475: for (i=0;i<n_neigh;i++) maxneighs = PetscMax(graph->count[shared[i][0]]+1,maxneighs);
1476: MPIU_Allreduce(MPI_IN_PLACE,&maxneighs,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)A));
1477: if (!maxneighs) {
1478: ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);
1479: *nnsp = NULL;
1480: return(0);
1481: }
1482: maxsize = 0;
1483: for (i=0;i<n_neigh;i++) maxsize = PetscMax(n_shared[i],maxsize);
1484: PetscMalloc2(maxsize,&gidxs,maxsize,&vals);
1485: /* create vectors to hold quadrature weights */
1486: MatCreateVecs(A,&quad_vec,NULL);
1487: if (!transpose) {
1488: MatGetLocalToGlobalMapping(A,&map,NULL);
1489: } else {
1490: MatGetLocalToGlobalMapping(A,NULL,&map);
1491: }
1492: VecDuplicateVecs(quad_vec,maxneighs,&quad_vecs);
1493: VecDestroy(&quad_vec);
1494: PCBDDCNullSpaceCreate(PetscObjectComm((PetscObject)A),PETSC_FALSE,maxneighs,quad_vecs,nnsp);
1495: for (i=0;i<maxneighs;i++) {
1496: VecLockReadPop(quad_vecs[i]);
1497: }
1499: /* compute local quad vec */
1500: MatISGetLocalMat(divudotp,&loc_divudotp);
1501: if (!transpose) {
1502: MatCreateVecs(loc_divudotp,&v,&p);
1503: } else {
1504: MatCreateVecs(loc_divudotp,&p,&v);
1505: }
1506: VecSet(p,1.);
1507: if (!transpose) {
1508: MatMultTranspose(loc_divudotp,p,v);
1509: } else {
1510: MatMult(loc_divudotp,p,v);
1511: }
1512: if (vl2l) {
1513: Mat lA;
1514: VecScatter sc;
1516: MatISGetLocalMat(A,&lA);
1517: MatCreateVecs(lA,&vins,NULL);
1518: VecScatterCreate(v,NULL,vins,vl2l,&sc);
1519: VecScatterBegin(sc,v,vins,INSERT_VALUES,SCATTER_FORWARD);
1520: VecScatterEnd(sc,v,vins,INSERT_VALUES,SCATTER_FORWARD);
1521: VecScatterDestroy(&sc);
1522: } else {
1523: vins = v;
1524: }
1525: VecGetArrayRead(vins,&array);
1526: VecDestroy(&p);
1528: /* insert in global quadrature vecs */
1529: MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
1530: for (i=1;i<n_neigh;i++) {
1531: const PetscInt *idxs;
1532: PetscInt idx,nn,j;
1534: idxs = shared[i];
1535: nn = n_shared[i];
1536: for (j=0;j<nn;j++) vals[j] = array[idxs[j]];
1537: PetscFindInt(rank,graph->count[idxs[0]],graph->neighbours_set[idxs[0]],&idx);
1538: idx = -(idx+1);
1539: if (idx < 0 || idx >= maxneighs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid index %D not in [0,%D)",idx,maxneighs);
1540: ISLocalToGlobalMappingApply(map,nn,idxs,gidxs);
1541: VecSetValues(quad_vecs[idx],nn,gidxs,vals,INSERT_VALUES);
1542: }
1543: ISLocalToGlobalMappingRestoreInfo(graph->l2gmap,&n_neigh,&neigh,&n_shared,&shared);
1544: VecRestoreArrayRead(vins,&array);
1545: if (vl2l) {
1546: VecDestroy(&vins);
1547: }
1548: VecDestroy(&v);
1549: PetscFree2(gidxs,vals);
1551: /* assemble near null space */
1552: for (i=0;i<maxneighs;i++) {
1553: VecAssemblyBegin(quad_vecs[i]);
1554: }
1555: for (i=0;i<maxneighs;i++) {
1556: VecAssemblyEnd(quad_vecs[i]);
1557: VecViewFromOptions(quad_vecs[i],NULL,"-pc_bddc_quad_vecs_view");
1558: VecLockReadPush(quad_vecs[i]);
1559: }
1560: VecDestroyVecs(maxneighs,&quad_vecs);
1561: return(0);
1562: }
1564: PetscErrorCode PCBDDCAddPrimalVerticesLocalIS(PC pc, IS primalv)
1565: {
1566: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
1570: if (primalv) {
1571: if (pcbddc->user_primal_vertices_local) {
1572: IS list[2], newp;
1574: list[0] = primalv;
1575: list[1] = pcbddc->user_primal_vertices_local;
1576: ISConcatenate(PetscObjectComm((PetscObject)pc),2,list,&newp);
1577: ISSortRemoveDups(newp);
1578: ISDestroy(&list[1]);
1579: pcbddc->user_primal_vertices_local = newp;
1580: } else {
1581: PCBDDCSetPrimalVerticesLocalIS(pc,primalv);
1582: }
1583: }
1584: return(0);
1585: }
1587: static PetscErrorCode func_coords_private(PetscInt dim, PetscReal t, const PetscReal X[], PetscInt Nf, PetscScalar *out, void *ctx)
1588: {
1589: PetscInt f, *comp = (PetscInt *)ctx;
1592: for (f=0;f<Nf;f++) out[f] = X[*comp];
1593: return(0);
1594: }
1596: PetscErrorCode PCBDDCComputeLocalTopologyInfo(PC pc)
1597: {
1599: Vec local,global;
1600: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
1601: Mat_IS *matis = (Mat_IS*)pc->pmat->data;
1602: PetscBool monolithic = PETSC_FALSE;
1605: PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC topology options","PC");
1606: PetscOptionsBool("-pc_bddc_monolithic","Discard any information on dofs splitting",NULL,monolithic,&monolithic,NULL);
1607: PetscOptionsEnd();
1608: /* need to convert from global to local topology information and remove references to information in global ordering */
1609: MatCreateVecs(pc->pmat,&global,NULL);
1610: MatCreateVecs(matis->A,&local,NULL);
1611: VecBindToCPU(global,PETSC_TRUE);
1612: VecBindToCPU(local,PETSC_TRUE);
1613: if (monolithic) { /* just get block size to properly compute vertices */
1614: if (pcbddc->vertex_size == 1) {
1615: MatGetBlockSize(pc->pmat,&pcbddc->vertex_size);
1616: }
1617: goto boundary;
1618: }
1620: if (pcbddc->user_provided_isfordofs) {
1621: if (pcbddc->n_ISForDofs) {
1622: PetscInt i;
1624: PetscMalloc1(pcbddc->n_ISForDofs,&pcbddc->ISForDofsLocal);
1625: for (i=0;i<pcbddc->n_ISForDofs;i++) {
1626: PetscInt bs;
1628: PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->ISForDofs[i],&pcbddc->ISForDofsLocal[i]);
1629: ISGetBlockSize(pcbddc->ISForDofs[i],&bs);
1630: ISSetBlockSize(pcbddc->ISForDofsLocal[i],bs);
1631: ISDestroy(&pcbddc->ISForDofs[i]);
1632: }
1633: pcbddc->n_ISForDofsLocal = pcbddc->n_ISForDofs;
1634: pcbddc->n_ISForDofs = 0;
1635: PetscFree(pcbddc->ISForDofs);
1636: }
1637: } else {
1638: if (!pcbddc->n_ISForDofsLocal) { /* field split not present */
1639: DM dm;
1641: MatGetDM(pc->pmat, &dm);
1642: if (!dm) {
1643: PCGetDM(pc, &dm);
1644: }
1645: if (dm) {
1646: IS *fields;
1647: PetscInt nf,i;
1649: DMCreateFieldDecomposition(dm,&nf,NULL,&fields,NULL);
1650: PetscMalloc1(nf,&pcbddc->ISForDofsLocal);
1651: for (i=0;i<nf;i++) {
1652: PetscInt bs;
1654: PCBDDCGlobalToLocal(matis->rctx,global,local,fields[i],&pcbddc->ISForDofsLocal[i]);
1655: ISGetBlockSize(fields[i],&bs);
1656: ISSetBlockSize(pcbddc->ISForDofsLocal[i],bs);
1657: ISDestroy(&fields[i]);
1658: }
1659: PetscFree(fields);
1660: pcbddc->n_ISForDofsLocal = nf;
1661: } else { /* See if MATIS has fields attached by the conversion from MatNest */
1662: PetscContainer c;
1664: PetscObjectQuery((PetscObject)pc->pmat,"_convert_nest_lfields",(PetscObject*)&c);
1665: if (c) {
1666: MatISLocalFields lf;
1667: PetscContainerGetPointer(c,(void**)&lf);
1668: PCBDDCSetDofsSplittingLocal(pc,lf->nr,lf->rf);
1669: } else { /* fallback, create the default fields if bs > 1 */
1670: PetscInt i, n = matis->A->rmap->n;
1671: MatGetBlockSize(pc->pmat,&i);
1672: if (i > 1) {
1673: pcbddc->n_ISForDofsLocal = i;
1674: PetscMalloc1(pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal);
1675: for (i=0;i<pcbddc->n_ISForDofsLocal;i++) {
1676: ISCreateStride(PetscObjectComm((PetscObject)pc),n/pcbddc->n_ISForDofsLocal,i,pcbddc->n_ISForDofsLocal,&pcbddc->ISForDofsLocal[i]);
1677: }
1678: }
1679: }
1680: }
1681: } else {
1682: PetscInt i;
1683: for (i=0;i<pcbddc->n_ISForDofsLocal;i++) {
1684: PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->ISForDofsLocal[i]);
1685: }
1686: }
1687: }
1689: boundary:
1690: if (!pcbddc->DirichletBoundariesLocal && pcbddc->DirichletBoundaries) {
1691: PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->DirichletBoundaries,&pcbddc->DirichletBoundariesLocal);
1692: } else if (pcbddc->DirichletBoundariesLocal) {
1693: PCBDDCConsistencyCheckIS(pc,MPI_LAND,&pcbddc->DirichletBoundariesLocal);
1694: }
1695: if (!pcbddc->NeumannBoundariesLocal && pcbddc->NeumannBoundaries) {
1696: PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->NeumannBoundaries,&pcbddc->NeumannBoundariesLocal);
1697: } else if (pcbddc->NeumannBoundariesLocal) {
1698: PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->NeumannBoundariesLocal);
1699: }
1700: if (!pcbddc->user_primal_vertices_local && pcbddc->user_primal_vertices) {
1701: PCBDDCGlobalToLocal(matis->rctx,global,local,pcbddc->user_primal_vertices,&pcbddc->user_primal_vertices_local);
1702: }
1703: VecDestroy(&global);
1704: VecDestroy(&local);
1705: /* detect local disconnected subdomains if requested (use matis->A) */
1706: if (pcbddc->detect_disconnected) {
1707: IS primalv = NULL;
1708: PetscInt i;
1709: PetscBool filter = pcbddc->detect_disconnected_filter;
1711: for (i=0;i<pcbddc->n_local_subs;i++) {
1712: ISDestroy(&pcbddc->local_subs[i]);
1713: }
1714: PetscFree(pcbddc->local_subs);
1715: PCBDDCDetectDisconnectedComponents(pc,filter,&pcbddc->n_local_subs,&pcbddc->local_subs,&primalv);
1716: PCBDDCAddPrimalVerticesLocalIS(pc,primalv);
1717: ISDestroy(&primalv);
1718: }
1719: /* early stage corner detection */
1720: {
1721: DM dm;
1723: MatGetDM(pc->pmat,&dm);
1724: if (!dm) {
1725: PCGetDM(pc,&dm);
1726: }
1727: if (dm) {
1728: PetscBool isda;
1730: PetscObjectTypeCompare((PetscObject)dm,DMDA,&isda);
1731: if (isda) {
1732: ISLocalToGlobalMapping l2l;
1733: IS corners;
1734: Mat lA;
1735: PetscBool gl,lo;
1737: {
1738: Vec cvec;
1739: const PetscScalar *coords;
1740: PetscInt dof,n,cdim;
1741: PetscBool memc = PETSC_TRUE;
1743: DMDAGetInfo(dm,NULL,NULL,NULL,NULL,NULL,NULL,NULL,&dof,NULL,NULL,NULL,NULL,NULL);
1744: DMGetCoordinates(dm,&cvec);
1745: VecGetLocalSize(cvec,&n);
1746: VecGetBlockSize(cvec,&cdim);
1747: n /= cdim;
1748: PetscFree(pcbddc->mat_graph->coords);
1749: PetscMalloc1(dof*n*cdim,&pcbddc->mat_graph->coords);
1750: VecGetArrayRead(cvec,&coords);
1751: #if defined(PETSC_USE_COMPLEX)
1752: memc = PETSC_FALSE;
1753: #endif
1754: if (dof != 1) memc = PETSC_FALSE;
1755: if (memc) {
1756: PetscArraycpy(pcbddc->mat_graph->coords,coords,cdim*n*dof);
1757: } else { /* BDDC graph does not use any blocked information, we need to replicate the data */
1758: PetscReal *bcoords = pcbddc->mat_graph->coords;
1759: PetscInt i, b, d;
1761: for (i=0;i<n;i++) {
1762: for (b=0;b<dof;b++) {
1763: for (d=0;d<cdim;d++) {
1764: bcoords[i*dof*cdim + b*cdim + d] = PetscRealPart(coords[i*cdim+d]);
1765: }
1766: }
1767: }
1768: }
1769: VecRestoreArrayRead(cvec,&coords);
1770: pcbddc->mat_graph->cdim = cdim;
1771: pcbddc->mat_graph->cnloc = dof*n;
1772: pcbddc->mat_graph->cloc = PETSC_FALSE;
1773: }
1774: DMDAGetSubdomainCornersIS(dm,&corners);
1775: MatISGetLocalMat(pc->pmat,&lA);
1776: MatGetLocalToGlobalMapping(lA,&l2l,NULL);
1777: MatISRestoreLocalMat(pc->pmat,&lA);
1778: lo = (PetscBool)(l2l && corners);
1779: MPIU_Allreduce(&lo,&gl,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));
1780: if (gl) { /* From PETSc's DMDA */
1781: const PetscInt *idx;
1782: PetscInt dof,bs,*idxout,n;
1784: DMDAGetInfo(dm,NULL,NULL,NULL,NULL,NULL,NULL,NULL,&dof,NULL,NULL,NULL,NULL,NULL);
1785: ISLocalToGlobalMappingGetBlockSize(l2l,&bs);
1786: ISGetLocalSize(corners,&n);
1787: ISGetIndices(corners,&idx);
1788: if (bs == dof) {
1789: PetscMalloc1(n,&idxout);
1790: ISLocalToGlobalMappingApplyBlock(l2l,n,idx,idxout);
1791: } else { /* the original DMDA local-to-local map have been modified */
1792: PetscInt i,d;
1794: PetscMalloc1(dof*n,&idxout);
1795: for (i=0;i<n;i++) for (d=0;d<dof;d++) idxout[dof*i+d] = dof*idx[i]+d;
1796: ISLocalToGlobalMappingApply(l2l,dof*n,idxout,idxout);
1798: bs = 1;
1799: n *= dof;
1800: }
1801: ISRestoreIndices(corners,&idx);
1802: DMDARestoreSubdomainCornersIS(dm,&corners);
1803: ISCreateBlock(PetscObjectComm((PetscObject)pc),bs,n,idxout,PETSC_OWN_POINTER,&corners);
1804: PCBDDCAddPrimalVerticesLocalIS(pc,corners);
1805: ISDestroy(&corners);
1806: pcbddc->corner_selected = PETSC_TRUE;
1807: pcbddc->corner_selection = PETSC_TRUE;
1808: }
1809: if (corners) {
1810: DMDARestoreSubdomainCornersIS(dm,&corners);
1811: }
1812: }
1813: }
1814: }
1815: if (pcbddc->corner_selection && !pcbddc->mat_graph->cdim) {
1816: DM dm;
1818: MatGetDM(pc->pmat,&dm);
1819: if (!dm) {
1820: PCGetDM(pc,&dm);
1821: }
1822: if (dm) { /* this can get very expensive, I need to find a faster alternative */
1823: Vec vcoords;
1824: PetscSection section;
1825: PetscReal *coords;
1826: PetscInt d,cdim,nl,nf,**ctxs;
1827: PetscErrorCode (**funcs)(PetscInt, PetscReal, const PetscReal *, PetscInt, PetscScalar *, void *);
1829: DMGetCoordinateDim(dm,&cdim);
1830: DMGetLocalSection(dm,§ion);
1831: PetscSectionGetNumFields(section,&nf);
1832: DMCreateGlobalVector(dm,&vcoords);
1833: VecGetLocalSize(vcoords,&nl);
1834: PetscMalloc1(nl*cdim,&coords);
1835: PetscMalloc2(nf,&funcs,nf,&ctxs);
1836: PetscMalloc1(nf,&ctxs[0]);
1837: for (d=0;d<nf;d++) funcs[d] = func_coords_private;
1838: for (d=1;d<nf;d++) ctxs[d] = ctxs[d-1] + 1;
1839: for (d=0;d<cdim;d++) {
1840: PetscInt i;
1841: const PetscScalar *v;
1843: for (i=0;i<nf;i++) ctxs[i][0] = d;
1844: DMProjectFunction(dm,0.0,funcs,(void**)ctxs,INSERT_VALUES,vcoords);
1845: VecGetArrayRead(vcoords,&v);
1846: for (i=0;i<nl;i++) coords[i*cdim+d] = PetscRealPart(v[i]);
1847: VecRestoreArrayRead(vcoords,&v);
1848: }
1849: VecDestroy(&vcoords);
1850: PCSetCoordinates(pc,cdim,nl,coords);
1851: PetscFree(coords);
1852: PetscFree(ctxs[0]);
1853: PetscFree2(funcs,ctxs);
1854: }
1855: }
1856: return(0);
1857: }
1859: PetscErrorCode PCBDDCConsistencyCheckIS(PC pc, MPI_Op mop, IS *is)
1860: {
1861: Mat_IS *matis = (Mat_IS*)(pc->pmat->data);
1862: PetscErrorCode ierr;
1863: IS nis;
1864: const PetscInt *idxs;
1865: PetscInt i,nd,n = matis->A->rmap->n,*nidxs,nnd;
1866: PetscBool *ld;
1869: if (mop != MPI_LAND && mop != MPI_LOR) SETERRQ(PetscObjectComm((PetscObject)(pc)),PETSC_ERR_SUP,"Supported are MPI_LAND and MPI_LOR");
1870: if (mop == MPI_LAND) {
1871: /* init rootdata with true */
1872: ld = (PetscBool*) matis->sf_rootdata;
1873: for (i=0;i<pc->pmat->rmap->n;i++) ld[i] = PETSC_TRUE;
1874: } else {
1875: PetscArrayzero(matis->sf_rootdata,pc->pmat->rmap->n);
1876: }
1877: PetscArrayzero(matis->sf_leafdata,n);
1878: ISGetLocalSize(*is,&nd);
1879: ISGetIndices(*is,&idxs);
1880: ld = (PetscBool*) matis->sf_leafdata;
1881: for (i=0;i<nd;i++)
1882: if (-1 < idxs[i] && idxs[i] < n)
1883: ld[idxs[i]] = PETSC_TRUE;
1884: ISRestoreIndices(*is,&idxs);
1885: PetscSFReduceBegin(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);
1886: PetscSFReduceEnd(matis->sf,MPIU_BOOL,matis->sf_leafdata,matis->sf_rootdata,mop);
1887: PetscSFBcastBegin(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);
1888: PetscSFBcastEnd(matis->sf,MPIU_BOOL,matis->sf_rootdata,matis->sf_leafdata);
1889: if (mop == MPI_LAND) {
1890: PetscMalloc1(nd,&nidxs);
1891: } else {
1892: PetscMalloc1(n,&nidxs);
1893: }
1894: for (i=0,nnd=0;i<n;i++)
1895: if (ld[i])
1896: nidxs[nnd++] = i;
1897: ISCreateGeneral(PetscObjectComm((PetscObject)(*is)),nnd,nidxs,PETSC_OWN_POINTER,&nis);
1898: ISDestroy(is);
1899: *is = nis;
1900: return(0);
1901: }
1903: PetscErrorCode PCBDDCBenignRemoveInterior(PC pc,Vec r,Vec z)
1904: {
1905: PC_IS *pcis = (PC_IS*)(pc->data);
1906: PC_BDDC *pcbddc = (PC_BDDC*)(pc->data);
1907: PetscErrorCode ierr;
1910: if (!pcbddc->benign_have_null) {
1911: return(0);
1912: }
1913: if (pcbddc->ChangeOfBasisMatrix) {
1914: Vec swap;
1916: MatMultTranspose(pcbddc->ChangeOfBasisMatrix,r,pcbddc->work_change);
1917: swap = pcbddc->work_change;
1918: pcbddc->work_change = r;
1919: r = swap;
1920: }
1921: VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
1922: VecScatterEnd(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
1923: KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);
1924: KSPCheckSolve(pcbddc->ksp_D,pc,pcis->vec2_D);
1925: VecSet(z,0.);
1926: VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);
1927: VecScatterEnd(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);
1928: if (pcbddc->ChangeOfBasisMatrix) {
1929: pcbddc->work_change = r;
1930: VecCopy(z,pcbddc->work_change);
1931: MatMult(pcbddc->ChangeOfBasisMatrix,pcbddc->work_change,z);
1932: }
1933: return(0);
1934: }
1936: PetscErrorCode PCBDDCBenignMatMult_Private_Private(Mat A, Vec x, Vec y, PetscBool transpose)
1937: {
1938: PCBDDCBenignMatMult_ctx ctx;
1939: PetscErrorCode ierr;
1940: PetscBool apply_right,apply_left,reset_x;
1943: MatShellGetContext(A,&ctx);
1944: if (transpose) {
1945: apply_right = ctx->apply_left;
1946: apply_left = ctx->apply_right;
1947: } else {
1948: apply_right = ctx->apply_right;
1949: apply_left = ctx->apply_left;
1950: }
1951: reset_x = PETSC_FALSE;
1952: if (apply_right) {
1953: const PetscScalar *ax;
1954: PetscInt nl,i;
1956: VecGetLocalSize(x,&nl);
1957: VecGetArrayRead(x,&ax);
1958: PetscArraycpy(ctx->work,ax,nl);
1959: VecRestoreArrayRead(x,&ax);
1960: for (i=0;i<ctx->benign_n;i++) {
1961: PetscScalar sum,val;
1962: const PetscInt *idxs;
1963: PetscInt nz,j;
1964: ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);
1965: ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);
1966: sum = 0.;
1967: if (ctx->apply_p0) {
1968: val = ctx->work[idxs[nz-1]];
1969: for (j=0;j<nz-1;j++) {
1970: sum += ctx->work[idxs[j]];
1971: ctx->work[idxs[j]] += val;
1972: }
1973: } else {
1974: for (j=0;j<nz-1;j++) {
1975: sum += ctx->work[idxs[j]];
1976: }
1977: }
1978: ctx->work[idxs[nz-1]] -= sum;
1979: ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);
1980: }
1981: VecPlaceArray(x,ctx->work);
1982: reset_x = PETSC_TRUE;
1983: }
1984: if (transpose) {
1985: MatMultTranspose(ctx->A,x,y);
1986: } else {
1987: MatMult(ctx->A,x,y);
1988: }
1989: if (reset_x) {
1990: VecResetArray(x);
1991: }
1992: if (apply_left) {
1993: PetscScalar *ay;
1994: PetscInt i;
1996: VecGetArray(y,&ay);
1997: for (i=0;i<ctx->benign_n;i++) {
1998: PetscScalar sum,val;
1999: const PetscInt *idxs;
2000: PetscInt nz,j;
2001: ISGetLocalSize(ctx->benign_zerodiag_subs[i],&nz);
2002: ISGetIndices(ctx->benign_zerodiag_subs[i],&idxs);
2003: val = -ay[idxs[nz-1]];
2004: if (ctx->apply_p0) {
2005: sum = 0.;
2006: for (j=0;j<nz-1;j++) {
2007: sum += ay[idxs[j]];
2008: ay[idxs[j]] += val;
2009: }
2010: ay[idxs[nz-1]] += sum;
2011: } else {
2012: for (j=0;j<nz-1;j++) {
2013: ay[idxs[j]] += val;
2014: }
2015: ay[idxs[nz-1]] = 0.;
2016: }
2017: ISRestoreIndices(ctx->benign_zerodiag_subs[i],&idxs);
2018: }
2019: VecRestoreArray(y,&ay);
2020: }
2021: return(0);
2022: }
2024: PetscErrorCode PCBDDCBenignMatMultTranspose_Private(Mat A, Vec x, Vec y)
2025: {
2029: PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_TRUE);
2030: return(0);
2031: }
2033: PetscErrorCode PCBDDCBenignMatMult_Private(Mat A, Vec x, Vec y)
2034: {
2038: PCBDDCBenignMatMult_Private_Private(A,x,y,PETSC_FALSE);
2039: return(0);
2040: }
2042: PetscErrorCode PCBDDCBenignShellMat(PC pc, PetscBool restore)
2043: {
2044: PC_IS *pcis = (PC_IS*)pc->data;
2045: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
2046: PCBDDCBenignMatMult_ctx ctx;
2047: PetscErrorCode ierr;
2050: if (!restore) {
2051: Mat A_IB,A_BI;
2052: PetscScalar *work;
2053: PCBDDCReuseSolvers reuse = pcbddc->sub_schurs ? pcbddc->sub_schurs->reuse_solver : NULL;
2055: if (pcbddc->benign_original_mat) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Benign original mat has not been restored");
2056: if (!pcbddc->benign_change || !pcbddc->benign_n || pcbddc->benign_change_explicit) return(0);
2057: PetscMalloc1(pcis->n,&work);
2058: MatCreate(PETSC_COMM_SELF,&A_IB);
2059: MatSetSizes(A_IB,pcis->n-pcis->n_B,pcis->n_B,PETSC_DECIDE,PETSC_DECIDE);
2060: MatSetType(A_IB,MATSHELL);
2061: MatShellSetOperation(A_IB,MATOP_MULT,(void (*)(void))PCBDDCBenignMatMult_Private);
2062: MatShellSetOperation(A_IB,MATOP_MULT_TRANSPOSE,(void (*)(void))PCBDDCBenignMatMultTranspose_Private);
2063: PetscNew(&ctx);
2064: MatShellSetContext(A_IB,ctx);
2065: ctx->apply_left = PETSC_TRUE;
2066: ctx->apply_right = PETSC_FALSE;
2067: ctx->apply_p0 = PETSC_FALSE;
2068: ctx->benign_n = pcbddc->benign_n;
2069: if (reuse) {
2070: ctx->benign_zerodiag_subs = reuse->benign_zerodiag_subs;
2071: ctx->free = PETSC_FALSE;
2072: } else { /* TODO: could be optimized for successive solves */
2073: ISLocalToGlobalMapping N_to_D;
2074: PetscInt i;
2076: ISLocalToGlobalMappingCreateIS(pcis->is_I_local,&N_to_D);
2077: PetscMalloc1(pcbddc->benign_n,&ctx->benign_zerodiag_subs);
2078: for (i=0;i<pcbddc->benign_n;i++) {
2079: ISGlobalToLocalMappingApplyIS(N_to_D,IS_GTOLM_DROP,pcbddc->benign_zerodiag_subs[i],&ctx->benign_zerodiag_subs[i]);
2080: }
2081: ISLocalToGlobalMappingDestroy(&N_to_D);
2082: ctx->free = PETSC_TRUE;
2083: }
2084: ctx->A = pcis->A_IB;
2085: ctx->work = work;
2086: MatSetUp(A_IB);
2087: MatAssemblyBegin(A_IB,MAT_FINAL_ASSEMBLY);
2088: MatAssemblyEnd(A_IB,MAT_FINAL_ASSEMBLY);
2089: pcis->A_IB = A_IB;
2091: /* A_BI as A_IB^T */
2092: MatCreateTranspose(A_IB,&A_BI);
2093: pcbddc->benign_original_mat = pcis->A_BI;
2094: pcis->A_BI = A_BI;
2095: } else {
2096: if (!pcbddc->benign_original_mat) {
2097: return(0);
2098: }
2099: MatShellGetContext(pcis->A_IB,&ctx);
2100: MatDestroy(&pcis->A_IB);
2101: pcis->A_IB = ctx->A;
2102: ctx->A = NULL;
2103: MatDestroy(&pcis->A_BI);
2104: pcis->A_BI = pcbddc->benign_original_mat;
2105: pcbddc->benign_original_mat = NULL;
2106: if (ctx->free) {
2107: PetscInt i;
2108: for (i=0;i<ctx->benign_n;i++) {
2109: ISDestroy(&ctx->benign_zerodiag_subs[i]);
2110: }
2111: PetscFree(ctx->benign_zerodiag_subs);
2112: }
2113: PetscFree(ctx->work);
2114: PetscFree(ctx);
2115: }
2116: return(0);
2117: }
2119: /* used just in bddc debug mode */
2120: PetscErrorCode PCBDDCBenignProject(PC pc, IS is1, IS is2, Mat *B)
2121: {
2122: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
2123: Mat_IS *matis = (Mat_IS*)pc->pmat->data;
2124: Mat An;
2128: MatPtAP(matis->A,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&An);
2129: MatZeroRowsColumns(An,pcbddc->benign_n,pcbddc->benign_p0_lidx,1.0,NULL,NULL);
2130: if (is1) {
2131: MatCreateSubMatrix(An,is1,is2,MAT_INITIAL_MATRIX,B);
2132: MatDestroy(&An);
2133: } else {
2134: *B = An;
2135: }
2136: return(0);
2137: }
2139: /* TODO: add reuse flag */
2140: PetscErrorCode MatSeqAIJCompress(Mat A, Mat *B)
2141: {
2142: Mat Bt;
2143: PetscScalar *a,*bdata;
2144: const PetscInt *ii,*ij;
2145: PetscInt m,n,i,nnz,*bii,*bij;
2146: PetscBool flg_row;
2150: MatGetSize(A,&n,&m);
2151: MatGetRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);
2152: MatSeqAIJGetArray(A,&a);
2153: nnz = n;
2154: for (i=0;i<ii[n];i++) {
2155: if (PetscLikely(PetscAbsScalar(a[i]) > PETSC_SMALL)) nnz++;
2156: }
2157: PetscMalloc1(n+1,&bii);
2158: PetscMalloc1(nnz,&bij);
2159: PetscMalloc1(nnz,&bdata);
2160: nnz = 0;
2161: bii[0] = 0;
2162: for (i=0;i<n;i++) {
2163: PetscInt j;
2164: for (j=ii[i];j<ii[i+1];j++) {
2165: PetscScalar entry = a[j];
2166: if (PetscLikely(PetscAbsScalar(entry) > PETSC_SMALL) || (n == m && ij[j] == i)) {
2167: bij[nnz] = ij[j];
2168: bdata[nnz] = entry;
2169: nnz++;
2170: }
2171: }
2172: bii[i+1] = nnz;
2173: }
2174: MatSeqAIJRestoreArray(A,&a);
2175: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),n,m,bii,bij,bdata,&Bt);
2176: MatRestoreRowIJ(A,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&ij,&flg_row);
2177: {
2178: Mat_SeqAIJ *b = (Mat_SeqAIJ*)(Bt->data);
2179: b->free_a = PETSC_TRUE;
2180: b->free_ij = PETSC_TRUE;
2181: }
2182: if (*B == A) {
2183: MatDestroy(&A);
2184: }
2185: *B = Bt;
2186: return(0);
2187: }
2189: PetscErrorCode PCBDDCDetectDisconnectedComponents(PC pc, PetscBool filter, PetscInt *ncc, IS* cc[], IS* primalv)
2190: {
2191: Mat B = NULL;
2192: DM dm;
2193: IS is_dummy,*cc_n;
2194: ISLocalToGlobalMapping l2gmap_dummy;
2195: PCBDDCGraph graph;
2196: PetscInt *xadj_filtered = NULL,*adjncy_filtered = NULL;
2197: PetscInt i,n;
2198: PetscInt *xadj,*adjncy;
2199: PetscBool isplex = PETSC_FALSE;
2200: PetscErrorCode ierr;
2203: if (ncc) *ncc = 0;
2204: if (cc) *cc = NULL;
2205: if (primalv) *primalv = NULL;
2206: PCBDDCGraphCreate(&graph);
2207: MatGetDM(pc->pmat,&dm);
2208: if (!dm) {
2209: PCGetDM(pc,&dm);
2210: }
2211: if (dm) {
2212: PetscObjectTypeCompare((PetscObject)dm,DMPLEX,&isplex);
2213: }
2214: if (filter) isplex = PETSC_FALSE;
2216: if (isplex) { /* this code has been modified from plexpartition.c */
2217: PetscInt p, pStart, pEnd, a, adjSize, idx, size, nroots;
2218: PetscInt *adj = NULL;
2219: IS cellNumbering;
2220: const PetscInt *cellNum;
2221: PetscBool useCone, useClosure;
2222: PetscSection section;
2223: PetscSegBuffer adjBuffer;
2224: PetscSF sfPoint;
2228: DMPlexGetHeightStratum(dm, 0, &pStart, &pEnd);
2229: DMGetPointSF(dm, &sfPoint);
2230: PetscSFGetGraph(sfPoint, &nroots, NULL, NULL, NULL);
2231: /* Build adjacency graph via a section/segbuffer */
2232: PetscSectionCreate(PetscObjectComm((PetscObject) dm), §ion);
2233: PetscSectionSetChart(section, pStart, pEnd);
2234: PetscSegBufferCreate(sizeof(PetscInt),1000,&adjBuffer);
2235: /* Always use FVM adjacency to create partitioner graph */
2236: DMGetBasicAdjacency(dm, &useCone, &useClosure);
2237: DMSetBasicAdjacency(dm, PETSC_TRUE, PETSC_FALSE);
2238: DMPlexGetCellNumbering(dm, &cellNumbering);
2239: ISGetIndices(cellNumbering, &cellNum);
2240: for (n = 0, p = pStart; p < pEnd; p++) {
2241: /* Skip non-owned cells in parallel (ParMetis expects no overlap) */
2242: if (nroots > 0) {if (cellNum[p] < 0) continue;}
2243: adjSize = PETSC_DETERMINE;
2244: DMPlexGetAdjacency(dm, p, &adjSize, &adj);
2245: for (a = 0; a < adjSize; ++a) {
2246: const PetscInt point = adj[a];
2247: if (pStart <= point && point < pEnd) {
2248: PetscInt *PETSC_RESTRICT pBuf;
2249: PetscSectionAddDof(section, p, 1);
2250: PetscSegBufferGetInts(adjBuffer, 1, &pBuf);
2251: *pBuf = point;
2252: }
2253: }
2254: n++;
2255: }
2256: DMSetBasicAdjacency(dm, useCone, useClosure);
2257: /* Derive CSR graph from section/segbuffer */
2258: PetscSectionSetUp(section);
2259: PetscSectionGetStorageSize(section, &size);
2260: PetscMalloc1(n+1, &xadj);
2261: for (idx = 0, p = pStart; p < pEnd; p++) {
2262: if (nroots > 0) {if (cellNum[p] < 0) continue;}
2263: PetscSectionGetOffset(section, p, &(xadj[idx++]));
2264: }
2265: xadj[n] = size;
2266: PetscSegBufferExtractAlloc(adjBuffer, &adjncy);
2267: /* Clean up */
2268: PetscSegBufferDestroy(&adjBuffer);
2269: PetscSectionDestroy(§ion);
2270: PetscFree(adj);
2271: graph->xadj = xadj;
2272: graph->adjncy = adjncy;
2273: } else {
2274: Mat A;
2275: PetscBool isseqaij, flg_row;
2277: MatISGetLocalMat(pc->pmat,&A);
2278: if (!A->rmap->N || !A->cmap->N) {
2279: PCBDDCGraphDestroy(&graph);
2280: return(0);
2281: }
2282: PetscObjectBaseTypeCompare((PetscObject)A,MATSEQAIJ,&isseqaij);
2283: if (!isseqaij && filter) {
2284: PetscBool isseqdense;
2286: PetscObjectTypeCompare((PetscObject)A,MATSEQDENSE,&isseqdense);
2287: if (!isseqdense) {
2288: MatConvert(A,MATSEQAIJ,MAT_INITIAL_MATRIX,&B);
2289: } else { /* TODO: rectangular case and LDA */
2290: PetscScalar *array;
2291: PetscReal chop=1.e-6;
2293: MatDuplicate(A,MAT_COPY_VALUES,&B);
2294: MatDenseGetArray(B,&array);
2295: MatGetSize(B,&n,NULL);
2296: for (i=0;i<n;i++) {
2297: PetscInt j;
2298: for (j=i+1;j<n;j++) {
2299: PetscReal thresh = chop*(PetscAbsScalar(array[i*(n+1)])+PetscAbsScalar(array[j*(n+1)]));
2300: if (PetscAbsScalar(array[i*n+j]) < thresh) array[i*n+j] = 0.;
2301: if (PetscAbsScalar(array[j*n+i]) < thresh) array[j*n+i] = 0.;
2302: }
2303: }
2304: MatDenseRestoreArray(B,&array);
2305: MatConvert(B,MATSEQAIJ,MAT_INPLACE_MATRIX,&B);
2306: }
2307: } else {
2308: PetscObjectReference((PetscObject)A);
2309: B = A;
2310: }
2311: MatGetRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);
2313: /* if filter is true, then removes entries lower than PETSC_SMALL in magnitude */
2314: if (filter) {
2315: PetscScalar *data;
2316: PetscInt j,cum;
2318: PetscCalloc2(n+1,&xadj_filtered,xadj[n],&adjncy_filtered);
2319: MatSeqAIJGetArray(B,&data);
2320: cum = 0;
2321: for (i=0;i<n;i++) {
2322: PetscInt t;
2324: for (j=xadj[i];j<xadj[i+1];j++) {
2325: if (PetscUnlikely(PetscAbsScalar(data[j]) < PETSC_SMALL)) {
2326: continue;
2327: }
2328: adjncy_filtered[cum+xadj_filtered[i]++] = adjncy[j];
2329: }
2330: t = xadj_filtered[i];
2331: xadj_filtered[i] = cum;
2332: cum += t;
2333: }
2334: MatSeqAIJRestoreArray(B,&data);
2335: graph->xadj = xadj_filtered;
2336: graph->adjncy = adjncy_filtered;
2337: } else {
2338: graph->xadj = xadj;
2339: graph->adjncy = adjncy;
2340: }
2341: }
2342: /* compute local connected components using PCBDDCGraph */
2343: ISCreateStride(PETSC_COMM_SELF,n,0,1,&is_dummy);
2344: ISLocalToGlobalMappingCreateIS(is_dummy,&l2gmap_dummy);
2345: ISDestroy(&is_dummy);
2346: PCBDDCGraphInit(graph,l2gmap_dummy,n,PETSC_MAX_INT);
2347: ISLocalToGlobalMappingDestroy(&l2gmap_dummy);
2348: PCBDDCGraphSetUp(graph,1,NULL,NULL,0,NULL,NULL);
2349: PCBDDCGraphComputeConnectedComponents(graph);
2351: /* partial clean up */
2352: PetscFree2(xadj_filtered,adjncy_filtered);
2353: if (B) {
2354: PetscBool flg_row;
2355: MatRestoreRowIJ(B,0,PETSC_TRUE,PETSC_FALSE,&n,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);
2356: MatDestroy(&B);
2357: }
2358: if (isplex) {
2359: PetscFree(xadj);
2360: PetscFree(adjncy);
2361: }
2363: /* get back data */
2364: if (isplex) {
2365: if (ncc) *ncc = graph->ncc;
2366: if (cc || primalv) {
2367: Mat A;
2368: PetscBT btv,btvt;
2369: PetscSection subSection;
2370: PetscInt *ids,cum,cump,*cids,*pids;
2372: DMPlexGetSubdomainSection(dm,&subSection);
2373: MatISGetLocalMat(pc->pmat,&A);
2374: PetscMalloc3(A->rmap->n,&ids,graph->ncc+1,&cids,A->rmap->n,&pids);
2375: PetscBTCreate(A->rmap->n,&btv);
2376: PetscBTCreate(A->rmap->n,&btvt);
2378: cids[0] = 0;
2379: for (i = 0, cump = 0, cum = 0; i < graph->ncc; i++) {
2380: PetscInt j;
2382: PetscBTMemzero(A->rmap->n,btvt);
2383: for (j = graph->cptr[i]; j < graph->cptr[i+1]; j++) {
2384: PetscInt k, size, *closure = NULL, cell = graph->queue[j];
2386: DMPlexGetTransitiveClosure(dm,cell,PETSC_TRUE,&size,&closure);
2387: for (k = 0; k < 2*size; k += 2) {
2388: PetscInt s, pp, p = closure[k], off, dof, cdof;
2390: PetscSectionGetConstraintDof(subSection,p,&cdof);
2391: PetscSectionGetOffset(subSection,p,&off);
2392: PetscSectionGetDof(subSection,p,&dof);
2393: for (s = 0; s < dof-cdof; s++) {
2394: if (PetscBTLookupSet(btvt,off+s)) continue;
2395: if (!PetscBTLookup(btv,off+s)) {
2396: ids[cum++] = off+s;
2397: } else { /* cross-vertex */
2398: pids[cump++] = off+s;
2399: }
2400: }
2401: DMPlexGetTreeParent(dm,p,&pp,NULL);
2402: if (pp != p) {
2403: PetscSectionGetConstraintDof(subSection,pp,&cdof);
2404: PetscSectionGetOffset(subSection,pp,&off);
2405: PetscSectionGetDof(subSection,pp,&dof);
2406: for (s = 0; s < dof-cdof; s++) {
2407: if (PetscBTLookupSet(btvt,off+s)) continue;
2408: if (!PetscBTLookup(btv,off+s)) {
2409: ids[cum++] = off+s;
2410: } else { /* cross-vertex */
2411: pids[cump++] = off+s;
2412: }
2413: }
2414: }
2415: }
2416: DMPlexRestoreTransitiveClosure(dm,cell,PETSC_TRUE,&size,&closure);
2417: }
2418: cids[i+1] = cum;
2419: /* mark dofs as already assigned */
2420: for (j = cids[i]; j < cids[i+1]; j++) {
2421: PetscBTSet(btv,ids[j]);
2422: }
2423: }
2424: if (cc) {
2425: PetscMalloc1(graph->ncc,&cc_n);
2426: for (i = 0; i < graph->ncc; i++) {
2427: ISCreateGeneral(PETSC_COMM_SELF,cids[i+1]-cids[i],ids+cids[i],PETSC_COPY_VALUES,&cc_n[i]);
2428: }
2429: *cc = cc_n;
2430: }
2431: if (primalv) {
2432: ISCreateGeneral(PetscObjectComm((PetscObject)pc),cump,pids,PETSC_COPY_VALUES,primalv);
2433: }
2434: PetscFree3(ids,cids,pids);
2435: PetscBTDestroy(&btv);
2436: PetscBTDestroy(&btvt);
2437: }
2438: } else {
2439: if (ncc) *ncc = graph->ncc;
2440: if (cc) {
2441: PetscMalloc1(graph->ncc,&cc_n);
2442: for (i=0;i<graph->ncc;i++) {
2443: ISCreateGeneral(PETSC_COMM_SELF,graph->cptr[i+1]-graph->cptr[i],graph->queue+graph->cptr[i],PETSC_COPY_VALUES,&cc_n[i]);
2444: }
2445: *cc = cc_n;
2446: }
2447: }
2448: /* clean up graph */
2449: graph->xadj = 0;
2450: graph->adjncy = 0;
2451: PCBDDCGraphDestroy(&graph);
2452: return(0);
2453: }
2455: PetscErrorCode PCBDDCBenignCheck(PC pc, IS zerodiag)
2456: {
2457: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
2458: PC_IS* pcis = (PC_IS*)(pc->data);
2459: IS dirIS = NULL;
2460: PetscInt i;
2464: PCBDDCGraphGetDirichletDofs(pcbddc->mat_graph,&dirIS);
2465: if (zerodiag) {
2466: Mat A;
2467: Vec vec3_N;
2468: PetscScalar *vals;
2469: const PetscInt *idxs;
2470: PetscInt nz,*count;
2472: /* p0 */
2473: VecSet(pcis->vec1_N,0.);
2474: PetscMalloc1(pcis->n,&vals);
2475: ISGetLocalSize(zerodiag,&nz);
2476: ISGetIndices(zerodiag,&idxs);
2477: for (i=0;i<nz;i++) vals[i] = 1.;
2478: VecSetValues(pcis->vec1_N,nz,idxs,vals,INSERT_VALUES);
2479: VecAssemblyBegin(pcis->vec1_N);
2480: VecAssemblyEnd(pcis->vec1_N);
2481: /* v_I */
2482: VecSetRandom(pcis->vec2_N,NULL);
2483: for (i=0;i<nz;i++) vals[i] = 0.;
2484: VecSetValues(pcis->vec2_N,nz,idxs,vals,INSERT_VALUES);
2485: ISRestoreIndices(zerodiag,&idxs);
2486: ISGetIndices(pcis->is_B_local,&idxs);
2487: for (i=0;i<pcis->n_B;i++) vals[i] = 0.;
2488: VecSetValues(pcis->vec2_N,pcis->n_B,idxs,vals,INSERT_VALUES);
2489: ISRestoreIndices(pcis->is_B_local,&idxs);
2490: if (dirIS) {
2491: PetscInt n;
2493: ISGetLocalSize(dirIS,&n);
2494: ISGetIndices(dirIS,&idxs);
2495: for (i=0;i<n;i++) vals[i] = 0.;
2496: VecSetValues(pcis->vec2_N,n,idxs,vals,INSERT_VALUES);
2497: ISRestoreIndices(dirIS,&idxs);
2498: }
2499: VecAssemblyBegin(pcis->vec2_N);
2500: VecAssemblyEnd(pcis->vec2_N);
2501: VecDuplicate(pcis->vec1_N,&vec3_N);
2502: VecSet(vec3_N,0.);
2503: MatISGetLocalMat(pc->pmat,&A);
2504: MatMult(A,pcis->vec1_N,vec3_N);
2505: VecDot(vec3_N,pcis->vec2_N,&vals[0]);
2506: if (PetscAbsScalar(vals[0]) > 1.e-1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! b(v_I,p_0) = %1.6e (should be numerically 0.)",PetscAbsScalar(vals[0]));
2507: PetscFree(vals);
2508: VecDestroy(&vec3_N);
2510: /* there should not be any pressure dofs lying on the interface */
2511: PetscCalloc1(pcis->n,&count);
2512: ISGetIndices(pcis->is_B_local,&idxs);
2513: for (i=0;i<pcis->n_B;i++) count[idxs[i]]++;
2514: ISRestoreIndices(pcis->is_B_local,&idxs);
2515: ISGetIndices(zerodiag,&idxs);
2516: for (i=0;i<nz;i++) if (count[idxs[i]]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Benign trick can not be applied! pressure dof %D is an interface dof",idxs[i]);
2517: ISRestoreIndices(zerodiag,&idxs);
2518: PetscFree(count);
2519: }
2520: ISDestroy(&dirIS);
2522: /* check PCBDDCBenignGetOrSetP0 */
2523: VecSetRandom(pcis->vec1_global,NULL);
2524: for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = -PetscGlobalRank-i;
2525: PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_FALSE);
2526: for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = 1;
2527: PCBDDCBenignGetOrSetP0(pc,pcis->vec1_global,PETSC_TRUE);
2528: for (i=0;i<pcbddc->benign_n;i++) {
2529: PetscInt val = PetscRealPart(pcbddc->benign_p0[i]);
2530: if (val != -PetscGlobalRank-i) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error testing PCBDDCBenignGetOrSetP0! Found %g at %D instead of %g",PetscRealPart(pcbddc->benign_p0[i]),i,-PetscGlobalRank-i);
2531: }
2532: return(0);
2533: }
2535: PetscErrorCode PCBDDCBenignDetectSaddlePoint(PC pc, PetscBool reuse, IS *zerodiaglocal)
2536: {
2537: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
2538: IS pressures = NULL,zerodiag = NULL,*bzerodiag = NULL,zerodiag_save,*zerodiag_subs;
2539: PetscInt nz,n,benign_n,bsp = 1;
2540: PetscInt *interior_dofs,n_interior_dofs,nneu;
2541: PetscBool sorted,have_null,has_null_pressures,recompute_zerodiag,checkb;
2545: if (reuse) goto project_b0;
2546: PetscSFDestroy(&pcbddc->benign_sf);
2547: MatDestroy(&pcbddc->benign_B0);
2548: for (n=0;n<pcbddc->benign_n;n++) {
2549: ISDestroy(&pcbddc->benign_zerodiag_subs[n]);
2550: }
2551: PetscFree(pcbddc->benign_zerodiag_subs);
2552: has_null_pressures = PETSC_TRUE;
2553: have_null = PETSC_TRUE;
2554: /* if a local information on dofs is present, gets pressure dofs from command line (uses the last field is not provided)
2555: Without local information, it uses only the zerodiagonal dofs (ok if the pressure block is all zero and it is a scalar field)
2556: Checks if all the pressure dofs in each subdomain have a zero diagonal
2557: If not, a change of basis on pressures is not needed
2558: since the local Schur complements are already SPD
2559: */
2560: if (pcbddc->n_ISForDofsLocal) {
2561: IS iP = NULL;
2562: PetscInt p,*pp;
2563: PetscBool flg;
2565: PetscMalloc1(pcbddc->n_ISForDofsLocal,&pp);
2566: n = pcbddc->n_ISForDofsLocal;
2567: PetscOptionsBegin(PetscObjectComm((PetscObject)pc),((PetscObject)pc)->prefix,"BDDC benign options","PC");
2568: PetscOptionsIntArray("-pc_bddc_pressure_field","Field id for pressures",NULL,pp,&n,&flg);
2569: PetscOptionsEnd();
2570: if (!flg) {
2571: n = 1;
2572: pp[0] = pcbddc->n_ISForDofsLocal-1;
2573: }
2575: bsp = 0;
2576: for (p=0;p<n;p++) {
2577: PetscInt bs;
2579: if (pp[p] < 0 || pp[p] > pcbddc->n_ISForDofsLocal-1) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Invalid field id for pressures %D",pp[p]);
2580: ISGetBlockSize(pcbddc->ISForDofsLocal[pp[p]],&bs);
2581: bsp += bs;
2582: }
2583: PetscMalloc1(bsp,&bzerodiag);
2584: bsp = 0;
2585: for (p=0;p<n;p++) {
2586: const PetscInt *idxs;
2587: PetscInt b,bs,npl,*bidxs;
2589: ISGetBlockSize(pcbddc->ISForDofsLocal[pp[p]],&bs);
2590: ISGetLocalSize(pcbddc->ISForDofsLocal[pp[p]],&npl);
2591: ISGetIndices(pcbddc->ISForDofsLocal[pp[p]],&idxs);
2592: PetscMalloc1(npl/bs,&bidxs);
2593: for (b=0;b<bs;b++) {
2594: PetscInt i;
2596: for (i=0;i<npl/bs;i++) bidxs[i] = idxs[bs*i+b];
2597: ISCreateGeneral(PETSC_COMM_SELF,npl/bs,bidxs,PETSC_COPY_VALUES,&bzerodiag[bsp]);
2598: bsp++;
2599: }
2600: PetscFree(bidxs);
2601: ISRestoreIndices(pcbddc->ISForDofsLocal[pp[p]],&idxs);
2602: }
2603: ISConcatenate(PETSC_COMM_SELF,bsp,bzerodiag,&pressures);
2605: /* remove zeroed out pressures if we are setting up a BDDC solver for a saddle-point FETI-DP */
2606: PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lP",(PetscObject*)&iP);
2607: if (iP) {
2608: IS newpressures;
2610: ISDifference(pressures,iP,&newpressures);
2611: ISDestroy(&pressures);
2612: pressures = newpressures;
2613: }
2614: ISSorted(pressures,&sorted);
2615: if (!sorted) {
2616: ISSort(pressures);
2617: }
2618: PetscFree(pp);
2619: }
2621: /* pcis has not been setup yet, so get the local size from the subdomain matrix */
2622: MatGetLocalSize(pcbddc->local_mat,&n,NULL);
2623: if (!n) pcbddc->benign_change_explicit = PETSC_TRUE;
2624: MatFindZeroDiagonals(pcbddc->local_mat,&zerodiag);
2625: ISSorted(zerodiag,&sorted);
2626: if (!sorted) {
2627: ISSort(zerodiag);
2628: }
2629: PetscObjectReference((PetscObject)zerodiag);
2630: zerodiag_save = zerodiag;
2631: ISGetLocalSize(zerodiag,&nz);
2632: if (!nz) {
2633: if (n) have_null = PETSC_FALSE;
2634: has_null_pressures = PETSC_FALSE;
2635: ISDestroy(&zerodiag);
2636: }
2637: recompute_zerodiag = PETSC_FALSE;
2639: /* in case disconnected subdomains info is present, split the pressures accordingly (otherwise the benign trick could fail) */
2640: zerodiag_subs = NULL;
2641: benign_n = 0;
2642: n_interior_dofs = 0;
2643: interior_dofs = NULL;
2644: nneu = 0;
2645: if (pcbddc->NeumannBoundariesLocal) {
2646: ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&nneu);
2647: }
2648: checkb = (PetscBool)(!pcbddc->NeumannBoundariesLocal || pcbddc->current_level);
2649: if (checkb) { /* need to compute interior nodes */
2650: PetscInt n,i,j;
2651: PetscInt n_neigh,*neigh,*n_shared,**shared;
2652: PetscInt *iwork;
2654: ISLocalToGlobalMappingGetSize(pc->pmat->rmap->mapping,&n);
2655: ISLocalToGlobalMappingGetInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);
2656: PetscCalloc1(n,&iwork);
2657: PetscMalloc1(n,&interior_dofs);
2658: for (i=1;i<n_neigh;i++)
2659: for (j=0;j<n_shared[i];j++)
2660: iwork[shared[i][j]] += 1;
2661: for (i=0;i<n;i++)
2662: if (!iwork[i])
2663: interior_dofs[n_interior_dofs++] = i;
2664: PetscFree(iwork);
2665: ISLocalToGlobalMappingRestoreInfo(pc->pmat->rmap->mapping,&n_neigh,&neigh,&n_shared,&shared);
2666: }
2667: if (has_null_pressures) {
2668: IS *subs;
2669: PetscInt nsubs,i,j,nl;
2670: const PetscInt *idxs;
2671: PetscScalar *array;
2672: Vec *work;
2673: Mat_IS* matis = (Mat_IS*)(pc->pmat->data);
2675: subs = pcbddc->local_subs;
2676: nsubs = pcbddc->n_local_subs;
2677: /* these vectors are needed to check if the constant on pressures is in the kernel of the local operator B (i.e. B(v_I,p0) should be zero) */
2678: if (checkb) {
2679: VecDuplicateVecs(matis->y,2,&work);
2680: ISGetLocalSize(zerodiag,&nl);
2681: ISGetIndices(zerodiag,&idxs);
2682: /* work[0] = 1_p */
2683: VecSet(work[0],0.);
2684: VecGetArray(work[0],&array);
2685: for (j=0;j<nl;j++) array[idxs[j]] = 1.;
2686: VecRestoreArray(work[0],&array);
2687: /* work[0] = 1_v */
2688: VecSet(work[1],1.);
2689: VecGetArray(work[1],&array);
2690: for (j=0;j<nl;j++) array[idxs[j]] = 0.;
2691: VecRestoreArray(work[1],&array);
2692: ISRestoreIndices(zerodiag,&idxs);
2693: }
2695: if (nsubs > 1 || bsp > 1) {
2696: IS *is;
2697: PetscInt b,totb;
2699: totb = bsp;
2700: is = bsp > 1 ? bzerodiag : &zerodiag;
2701: nsubs = PetscMax(nsubs,1);
2702: PetscCalloc1(nsubs*totb,&zerodiag_subs);
2703: for (b=0;b<totb;b++) {
2704: for (i=0;i<nsubs;i++) {
2705: ISLocalToGlobalMapping l2g;
2706: IS t_zerodiag_subs;
2707: PetscInt nl;
2709: if (subs) {
2710: ISLocalToGlobalMappingCreateIS(subs[i],&l2g);
2711: } else {
2712: IS tis;
2714: MatGetLocalSize(pcbddc->local_mat,&nl,NULL);
2715: ISCreateStride(PETSC_COMM_SELF,nl,0,1,&tis);
2716: ISLocalToGlobalMappingCreateIS(tis,&l2g);
2717: ISDestroy(&tis);
2718: }
2719: ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,is[b],&t_zerodiag_subs);
2720: ISGetLocalSize(t_zerodiag_subs,&nl);
2721: if (nl) {
2722: PetscBool valid = PETSC_TRUE;
2724: if (checkb) {
2725: VecSet(matis->x,0);
2726: ISGetLocalSize(subs[i],&nl);
2727: ISGetIndices(subs[i],&idxs);
2728: VecGetArray(matis->x,&array);
2729: for (j=0;j<nl;j++) array[idxs[j]] = 1.;
2730: VecRestoreArray(matis->x,&array);
2731: ISRestoreIndices(subs[i],&idxs);
2732: VecPointwiseMult(matis->x,work[0],matis->x);
2733: MatMult(matis->A,matis->x,matis->y);
2734: VecPointwiseMult(matis->y,work[1],matis->y);
2735: VecGetArray(matis->y,&array);
2736: for (j=0;j<n_interior_dofs;j++) {
2737: if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) {
2738: valid = PETSC_FALSE;
2739: break;
2740: }
2741: }
2742: VecRestoreArray(matis->y,&array);
2743: }
2744: if (valid && nneu) {
2745: const PetscInt *idxs;
2746: PetscInt nzb;
2748: ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);
2749: ISGlobalToLocalMappingApply(l2g,IS_GTOLM_DROP,nneu,idxs,&nzb,NULL);
2750: ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);
2751: if (nzb) valid = PETSC_FALSE;
2752: }
2753: if (valid && pressures) {
2754: IS t_pressure_subs,tmp;
2755: PetscInt i1,i2;
2757: ISGlobalToLocalMappingApplyIS(l2g,IS_GTOLM_DROP,pressures,&t_pressure_subs);
2758: ISEmbed(t_zerodiag_subs,t_pressure_subs,PETSC_TRUE,&tmp);
2759: ISGetLocalSize(tmp,&i1);
2760: ISGetLocalSize(t_zerodiag_subs,&i2);
2761: if (i2 != i1) valid = PETSC_FALSE;
2762: ISDestroy(&t_pressure_subs);
2763: ISDestroy(&tmp);
2764: }
2765: if (valid) {
2766: ISLocalToGlobalMappingApplyIS(l2g,t_zerodiag_subs,&zerodiag_subs[benign_n]);
2767: benign_n++;
2768: } else recompute_zerodiag = PETSC_TRUE;
2769: }
2770: ISDestroy(&t_zerodiag_subs);
2771: ISLocalToGlobalMappingDestroy(&l2g);
2772: }
2773: }
2774: } else { /* there's just one subdomain (or zero if they have not been detected */
2775: PetscBool valid = PETSC_TRUE;
2777: if (nneu) valid = PETSC_FALSE;
2778: if (valid && pressures) {
2779: ISEqual(pressures,zerodiag,&valid);
2780: }
2781: if (valid && checkb) {
2782: MatMult(matis->A,work[0],matis->x);
2783: VecPointwiseMult(matis->x,work[1],matis->x);
2784: VecGetArray(matis->x,&array);
2785: for (j=0;j<n_interior_dofs;j++) {
2786: if (PetscAbsScalar(array[interior_dofs[j]]) > PETSC_SMALL) {
2787: valid = PETSC_FALSE;
2788: break;
2789: }
2790: }
2791: VecRestoreArray(matis->x,&array);
2792: }
2793: if (valid) {
2794: benign_n = 1;
2795: PetscMalloc1(benign_n,&zerodiag_subs);
2796: PetscObjectReference((PetscObject)zerodiag);
2797: zerodiag_subs[0] = zerodiag;
2798: }
2799: }
2800: if (checkb) {
2801: VecDestroyVecs(2,&work);
2802: }
2803: }
2804: PetscFree(interior_dofs);
2806: if (!benign_n) {
2807: PetscInt n;
2809: ISDestroy(&zerodiag);
2810: recompute_zerodiag = PETSC_FALSE;
2811: MatGetLocalSize(pcbddc->local_mat,&n,NULL);
2812: if (n) have_null = PETSC_FALSE;
2813: }
2815: /* final check for null pressures */
2816: if (zerodiag && pressures) {
2817: ISEqual(pressures,zerodiag,&have_null);
2818: }
2820: if (recompute_zerodiag) {
2821: ISDestroy(&zerodiag);
2822: if (benign_n == 1) {
2823: PetscObjectReference((PetscObject)zerodiag_subs[0]);
2824: zerodiag = zerodiag_subs[0];
2825: } else {
2826: PetscInt i,nzn,*new_idxs;
2828: nzn = 0;
2829: for (i=0;i<benign_n;i++) {
2830: PetscInt ns;
2831: ISGetLocalSize(zerodiag_subs[i],&ns);
2832: nzn += ns;
2833: }
2834: PetscMalloc1(nzn,&new_idxs);
2835: nzn = 0;
2836: for (i=0;i<benign_n;i++) {
2837: PetscInt ns,*idxs;
2838: ISGetLocalSize(zerodiag_subs[i],&ns);
2839: ISGetIndices(zerodiag_subs[i],(const PetscInt**)&idxs);
2840: PetscArraycpy(new_idxs+nzn,idxs,ns);
2841: ISRestoreIndices(zerodiag_subs[i],(const PetscInt**)&idxs);
2842: nzn += ns;
2843: }
2844: PetscSortInt(nzn,new_idxs);
2845: ISCreateGeneral(PETSC_COMM_SELF,nzn,new_idxs,PETSC_OWN_POINTER,&zerodiag);
2846: }
2847: have_null = PETSC_FALSE;
2848: }
2850: /* determines if the coarse solver will be singular or not */
2851: MPIU_Allreduce(&have_null,&pcbddc->benign_null,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)pc));
2853: /* Prepare matrix to compute no-net-flux */
2854: if (pcbddc->compute_nonetflux && !pcbddc->divudotp) {
2855: Mat A,loc_divudotp;
2856: ISLocalToGlobalMapping rl2g,cl2g,l2gmap;
2857: IS row,col,isused = NULL;
2858: PetscInt M,N,n,st,n_isused;
2860: if (pressures) {
2861: isused = pressures;
2862: } else {
2863: isused = zerodiag_save;
2864: }
2865: MatGetLocalToGlobalMapping(pc->pmat,&l2gmap,NULL);
2866: MatISGetLocalMat(pc->pmat,&A);
2867: MatGetLocalSize(A,&n,NULL);
2868: if (!isused && n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"Don't know how to extract div u dot p! Please provide the pressure field");
2869: n_isused = 0;
2870: if (isused) {
2871: ISGetLocalSize(isused,&n_isused);
2872: }
2873: MPI_Scan(&n_isused,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));
2874: st = st-n_isused;
2875: if (n) {
2876: const PetscInt *gidxs;
2878: MatCreateSubMatrix(A,isused,NULL,MAT_INITIAL_MATRIX,&loc_divudotp);
2879: ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);
2880: /* TODO: extend ISCreateStride with st = PETSC_DECIDE */
2881: ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);
2882: ISCreateGeneral(PetscObjectComm((PetscObject)pc),n,gidxs,PETSC_COPY_VALUES,&col);
2883: ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);
2884: } else {
2885: MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&loc_divudotp);
2886: ISCreateStride(PetscObjectComm((PetscObject)pc),n_isused,st,1,&row);
2887: ISCreateGeneral(PetscObjectComm((PetscObject)pc),0,NULL,PETSC_COPY_VALUES,&col);
2888: }
2889: MatGetSize(pc->pmat,NULL,&N);
2890: ISGetSize(row,&M);
2891: ISLocalToGlobalMappingCreateIS(row,&rl2g);
2892: ISLocalToGlobalMappingCreateIS(col,&cl2g);
2893: ISDestroy(&row);
2894: ISDestroy(&col);
2895: MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->divudotp);
2896: MatSetType(pcbddc->divudotp,MATIS);
2897: MatSetSizes(pcbddc->divudotp,PETSC_DECIDE,PETSC_DECIDE,M,N);
2898: MatSetLocalToGlobalMapping(pcbddc->divudotp,rl2g,cl2g);
2899: ISLocalToGlobalMappingDestroy(&rl2g);
2900: ISLocalToGlobalMappingDestroy(&cl2g);
2901: MatISSetLocalMat(pcbddc->divudotp,loc_divudotp);
2902: MatDestroy(&loc_divudotp);
2903: MatAssemblyBegin(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);
2904: MatAssemblyEnd(pcbddc->divudotp,MAT_FINAL_ASSEMBLY);
2905: }
2906: ISDestroy(&zerodiag_save);
2907: ISDestroy(&pressures);
2908: if (bzerodiag) {
2909: PetscInt i;
2911: for (i=0;i<bsp;i++) {
2912: ISDestroy(&bzerodiag[i]);
2913: }
2914: PetscFree(bzerodiag);
2915: }
2916: pcbddc->benign_n = benign_n;
2917: pcbddc->benign_zerodiag_subs = zerodiag_subs;
2919: /* determines if the problem has subdomains with 0 pressure block */
2920: have_null = (PetscBool)(!!pcbddc->benign_n);
2921: MPIU_Allreduce(&have_null,&pcbddc->benign_have_null,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));
2923: project_b0:
2924: MatGetLocalSize(pcbddc->local_mat,&n,NULL);
2925: /* change of basis and p0 dofs */
2926: if (pcbddc->benign_n) {
2927: PetscInt i,s,*nnz;
2929: /* local change of basis for pressures */
2930: MatDestroy(&pcbddc->benign_change);
2931: MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_change);
2932: MatSetType(pcbddc->benign_change,MATAIJ);
2933: MatSetSizes(pcbddc->benign_change,n,n,PETSC_DECIDE,PETSC_DECIDE);
2934: PetscMalloc1(n,&nnz);
2935: for (i=0;i<n;i++) nnz[i] = 1; /* defaults to identity */
2936: for (i=0;i<pcbddc->benign_n;i++) {
2937: const PetscInt *idxs;
2938: PetscInt nzs,j;
2940: ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nzs);
2941: ISGetIndices(pcbddc->benign_zerodiag_subs[i],&idxs);
2942: for (j=0;j<nzs-1;j++) nnz[idxs[j]] = 2; /* change on pressures */
2943: nnz[idxs[nzs-1]] = nzs; /* last local pressure dof in subdomain */
2944: ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],&idxs);
2945: }
2946: MatSeqAIJSetPreallocation(pcbddc->benign_change,0,nnz);
2947: MatSetOption(pcbddc->benign_change,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2948: PetscFree(nnz);
2949: /* set identity by default */
2950: for (i=0;i<n;i++) {
2951: MatSetValue(pcbddc->benign_change,i,i,1.,INSERT_VALUES);
2952: }
2953: PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);
2954: PetscMalloc3(pcbddc->benign_n,&pcbddc->benign_p0_lidx,pcbddc->benign_n,&pcbddc->benign_p0_gidx,pcbddc->benign_n,&pcbddc->benign_p0);
2955: /* set change on pressures */
2956: for (s=0;s<pcbddc->benign_n;s++) {
2957: PetscScalar *array;
2958: const PetscInt *idxs;
2959: PetscInt nzs;
2961: ISGetLocalSize(pcbddc->benign_zerodiag_subs[s],&nzs);
2962: ISGetIndices(pcbddc->benign_zerodiag_subs[s],&idxs);
2963: for (i=0;i<nzs-1;i++) {
2964: PetscScalar vals[2];
2965: PetscInt cols[2];
2967: cols[0] = idxs[i];
2968: cols[1] = idxs[nzs-1];
2969: vals[0] = 1.;
2970: vals[1] = 1.;
2971: MatSetValues(pcbddc->benign_change,1,cols,2,cols,vals,INSERT_VALUES);
2972: }
2973: PetscMalloc1(nzs,&array);
2974: for (i=0;i<nzs-1;i++) array[i] = -1.;
2975: array[nzs-1] = 1.;
2976: MatSetValues(pcbddc->benign_change,1,idxs+nzs-1,nzs,idxs,array,INSERT_VALUES);
2977: /* store local idxs for p0 */
2978: pcbddc->benign_p0_lidx[s] = idxs[nzs-1];
2979: ISRestoreIndices(pcbddc->benign_zerodiag_subs[s],&idxs);
2980: PetscFree(array);
2981: }
2982: MatAssemblyBegin(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);
2983: MatAssemblyEnd(pcbddc->benign_change,MAT_FINAL_ASSEMBLY);
2985: /* project if needed */
2986: if (pcbddc->benign_change_explicit) {
2987: Mat M;
2989: MatPtAP(pcbddc->local_mat,pcbddc->benign_change,MAT_INITIAL_MATRIX,2.0,&M);
2990: MatDestroy(&pcbddc->local_mat);
2991: MatSeqAIJCompress(M,&pcbddc->local_mat);
2992: MatDestroy(&M);
2993: }
2994: /* store global idxs for p0 */
2995: ISLocalToGlobalMappingApply(pc->pmat->rmap->mapping,pcbddc->benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx);
2996: }
2997: *zerodiaglocal = zerodiag;
2998: return(0);
2999: }
3001: PetscErrorCode PCBDDCBenignGetOrSetP0(PC pc, Vec v, PetscBool get)
3002: {
3003: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
3004: PetscScalar *array;
3008: if (!pcbddc->benign_sf) {
3009: PetscSFCreate(PetscObjectComm((PetscObject)pc),&pcbddc->benign_sf);
3010: PetscSFSetGraphLayout(pcbddc->benign_sf,pc->pmat->rmap,pcbddc->benign_n,NULL,PETSC_OWN_POINTER,pcbddc->benign_p0_gidx);
3011: }
3012: if (get) {
3013: VecGetArrayRead(v,(const PetscScalar**)&array);
3014: PetscSFBcastBegin(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);
3015: PetscSFBcastEnd(pcbddc->benign_sf,MPIU_SCALAR,array,pcbddc->benign_p0);
3016: VecRestoreArrayRead(v,(const PetscScalar**)&array);
3017: } else {
3018: VecGetArray(v,&array);
3019: PetscSFReduceBegin(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);
3020: PetscSFReduceEnd(pcbddc->benign_sf,MPIU_SCALAR,pcbddc->benign_p0,array,MPIU_REPLACE);
3021: VecRestoreArray(v,&array);
3022: }
3023: return(0);
3024: }
3026: PetscErrorCode PCBDDCBenignPopOrPushB0(PC pc, PetscBool pop)
3027: {
3028: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
3032: /* TODO: add error checking
3033: - avoid nested pop (or push) calls.
3034: - cannot push before pop.
3035: - cannot call this if pcbddc->local_mat is NULL
3036: */
3037: if (!pcbddc->benign_n) {
3038: return(0);
3039: }
3040: if (pop) {
3041: if (pcbddc->benign_change_explicit) {
3042: IS is_p0;
3043: MatReuse reuse;
3045: /* extract B_0 */
3046: reuse = MAT_INITIAL_MATRIX;
3047: if (pcbddc->benign_B0) {
3048: reuse = MAT_REUSE_MATRIX;
3049: }
3050: ISCreateGeneral(PETSC_COMM_SELF,pcbddc->benign_n,pcbddc->benign_p0_lidx,PETSC_COPY_VALUES,&is_p0);
3051: MatCreateSubMatrix(pcbddc->local_mat,is_p0,NULL,reuse,&pcbddc->benign_B0);
3052: /* remove rows and cols from local problem */
3053: MatSetOption(pcbddc->local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);
3054: MatSetOption(pcbddc->local_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
3055: MatZeroRowsColumnsIS(pcbddc->local_mat,is_p0,1.0,NULL,NULL);
3056: ISDestroy(&is_p0);
3057: } else {
3058: Mat_IS *matis = (Mat_IS*)pc->pmat->data;
3059: PetscScalar *vals;
3060: PetscInt i,n,*idxs_ins;
3062: VecGetLocalSize(matis->y,&n);
3063: PetscMalloc2(n,&idxs_ins,n,&vals);
3064: if (!pcbddc->benign_B0) {
3065: PetscInt *nnz;
3066: MatCreate(PetscObjectComm((PetscObject)pcbddc->local_mat),&pcbddc->benign_B0);
3067: MatSetType(pcbddc->benign_B0,MATAIJ);
3068: MatSetSizes(pcbddc->benign_B0,pcbddc->benign_n,n,PETSC_DECIDE,PETSC_DECIDE);
3069: PetscMalloc1(pcbddc->benign_n,&nnz);
3070: for (i=0;i<pcbddc->benign_n;i++) {
3071: ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nnz[i]);
3072: nnz[i] = n - nnz[i];
3073: }
3074: MatSeqAIJSetPreallocation(pcbddc->benign_B0,0,nnz);
3075: MatSetOption(pcbddc->benign_B0,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
3076: PetscFree(nnz);
3077: }
3079: for (i=0;i<pcbddc->benign_n;i++) {
3080: PetscScalar *array;
3081: PetscInt *idxs,j,nz,cum;
3083: VecSet(matis->x,0.);
3084: ISGetLocalSize(pcbddc->benign_zerodiag_subs[i],&nz);
3085: ISGetIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);
3086: for (j=0;j<nz;j++) vals[j] = 1.;
3087: VecSetValues(matis->x,nz,idxs,vals,INSERT_VALUES);
3088: VecAssemblyBegin(matis->x);
3089: VecAssemblyEnd(matis->x);
3090: VecSet(matis->y,0.);
3091: MatMult(matis->A,matis->x,matis->y);
3092: VecGetArray(matis->y,&array);
3093: cum = 0;
3094: for (j=0;j<n;j++) {
3095: if (PetscUnlikely(PetscAbsScalar(array[j]) > PETSC_SMALL)) {
3096: vals[cum] = array[j];
3097: idxs_ins[cum] = j;
3098: cum++;
3099: }
3100: }
3101: MatSetValues(pcbddc->benign_B0,1,&i,cum,idxs_ins,vals,INSERT_VALUES);
3102: VecRestoreArray(matis->y,&array);
3103: ISRestoreIndices(pcbddc->benign_zerodiag_subs[i],(const PetscInt**)&idxs);
3104: }
3105: MatAssemblyBegin(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);
3106: MatAssemblyEnd(pcbddc->benign_B0,MAT_FINAL_ASSEMBLY);
3107: PetscFree2(idxs_ins,vals);
3108: }
3109: } else { /* push */
3110: if (pcbddc->benign_change_explicit) {
3111: PetscInt i;
3113: for (i=0;i<pcbddc->benign_n;i++) {
3114: PetscScalar *B0_vals;
3115: PetscInt *B0_cols,B0_ncol;
3117: MatGetRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);
3118: MatSetValues(pcbddc->local_mat,1,pcbddc->benign_p0_lidx+i,B0_ncol,B0_cols,B0_vals,INSERT_VALUES);
3119: MatSetValues(pcbddc->local_mat,B0_ncol,B0_cols,1,pcbddc->benign_p0_lidx+i,B0_vals,INSERT_VALUES);
3120: MatSetValue(pcbddc->local_mat,pcbddc->benign_p0_lidx[i],pcbddc->benign_p0_lidx[i],0.0,INSERT_VALUES);
3121: MatRestoreRow(pcbddc->benign_B0,i,&B0_ncol,(const PetscInt**)&B0_cols,(const PetscScalar**)&B0_vals);
3122: }
3123: MatAssemblyBegin(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);
3124: MatAssemblyEnd(pcbddc->local_mat,MAT_FINAL_ASSEMBLY);
3125: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cannot push B0!");
3126: }
3127: return(0);
3128: }
3130: PetscErrorCode PCBDDCAdaptiveSelection(PC pc)
3131: {
3132: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
3133: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
3134: PetscBLASInt B_dummyint,B_neigs,B_ierr,B_lwork;
3135: PetscBLASInt *B_iwork,*B_ifail;
3136: PetscScalar *work,lwork;
3137: PetscScalar *St,*S,*eigv;
3138: PetscScalar *Sarray,*Starray;
3139: PetscReal *eigs,thresh,lthresh,uthresh;
3140: PetscInt i,nmax,nmin,nv,cum,mss,cum2,cumarray,maxneigs;
3141: PetscBool allocated_S_St;
3142: #if defined(PETSC_USE_COMPLEX)
3143: PetscReal *rwork;
3144: #endif
3145: PetscErrorCode ierr;
3148: if (!sub_schurs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Adaptive selection of constraints requires SubSchurs data");
3149: if (!sub_schurs->schur_explicit) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Adaptive selection of constraints requires MUMPS and/or MKL_CPARDISO");
3150: if (sub_schurs->n_subs && (!sub_schurs->is_symmetric)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_SUP,"Adaptive selection not yet implemented for this matrix pencil (herm %d, symm %d, posdef %d)",sub_schurs->is_hermitian,sub_schurs->is_symmetric,sub_schurs->is_posdef);
3151: PetscLogEventBegin(PC_BDDC_AdaptiveSetUp[pcbddc->current_level],pc,0,0,0);
3153: if (pcbddc->dbg_flag) {
3154: PetscViewerFlush(pcbddc->dbg_viewer);
3155: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");
3156: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check adaptive selection of constraints\n");
3157: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
3158: }
3160: if (pcbddc->dbg_flag) {
3161: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d cc %D (%d,%d).\n",PetscGlobalRank,sub_schurs->n_subs,sub_schurs->is_hermitian,sub_schurs->is_posdef);
3162: }
3164: /* max size of subsets */
3165: mss = 0;
3166: for (i=0;i<sub_schurs->n_subs;i++) {
3167: PetscInt subset_size;
3169: ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);
3170: mss = PetscMax(mss,subset_size);
3171: }
3173: /* min/max and threshold */
3174: nmax = pcbddc->adaptive_nmax > 0 ? pcbddc->adaptive_nmax : mss;
3175: nmin = pcbddc->adaptive_nmin > 0 ? pcbddc->adaptive_nmin : 0;
3176: nmax = PetscMax(nmin,nmax);
3177: allocated_S_St = PETSC_FALSE;
3178: if (nmin || !sub_schurs->is_posdef) { /* XXX */
3179: allocated_S_St = PETSC_TRUE;
3180: }
3182: /* allocate lapack workspace */
3183: cum = cum2 = 0;
3184: maxneigs = 0;
3185: for (i=0;i<sub_schurs->n_subs;i++) {
3186: PetscInt n,subset_size;
3188: ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);
3189: n = PetscMin(subset_size,nmax);
3190: cum += subset_size;
3191: cum2 += subset_size*n;
3192: maxneigs = PetscMax(maxneigs,n);
3193: }
3194: lwork = 0;
3195: if (mss) {
3196: if (sub_schurs->is_symmetric) {
3197: PetscScalar sdummy = 0.;
3198: PetscBLASInt B_itype = 1;
3199: PetscBLASInt B_N = mss, idummy = 0;
3200: PetscReal rdummy = 0.,zero = 0.0;
3201: PetscReal eps = 0.0; /* dlamch? */
3203: B_lwork = -1;
3204: /* some implementations may complain about NULL pointers, even if we are querying */
3205: S = &sdummy;
3206: St = &sdummy;
3207: eigs = &rdummy;
3208: eigv = &sdummy;
3209: B_iwork = &idummy;
3210: B_ifail = &idummy;
3211: #if defined(PETSC_USE_COMPLEX)
3212: rwork = &rdummy;
3213: #endif
3214: thresh = 1.0;
3215: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
3216: #if defined(PETSC_USE_COMPLEX)
3217: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3218: #else
3219: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&zero,&thresh,&B_dummyint,&B_dummyint,&eps,&B_neigs,eigs,eigv,&B_N,&lwork,&B_lwork,B_iwork,B_ifail,&B_ierr));
3220: #endif
3221: if (B_ierr != 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYGVX Lapack routine %d",(int)B_ierr);
3222: PetscFPTrapPop();
3223: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented");
3224: }
3226: nv = 0;
3227: if (sub_schurs->is_vertices && pcbddc->use_vertices) { /* complement set of active subsets, each entry is a vertex (boundary made by active subsets, vertices and dirichlet dofs) */
3228: ISGetLocalSize(sub_schurs->is_vertices,&nv);
3229: }
3230: PetscBLASIntCast((PetscInt)PetscRealPart(lwork),&B_lwork);
3231: if (allocated_S_St) {
3232: PetscMalloc2(mss*mss,&S,mss*mss,&St);
3233: }
3234: PetscMalloc5(mss*mss,&eigv,mss,&eigs,B_lwork,&work,5*mss,&B_iwork,mss,&B_ifail);
3235: #if defined(PETSC_USE_COMPLEX)
3236: PetscMalloc1(7*mss,&rwork);
3237: #endif
3238: PetscMalloc5(nv+sub_schurs->n_subs,&pcbddc->adaptive_constraints_n,
3239: nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_idxs_ptr,
3240: nv+sub_schurs->n_subs+1,&pcbddc->adaptive_constraints_data_ptr,
3241: nv+cum,&pcbddc->adaptive_constraints_idxs,
3242: nv+cum2,&pcbddc->adaptive_constraints_data);
3243: PetscArrayzero(pcbddc->adaptive_constraints_n,nv+sub_schurs->n_subs);
3245: maxneigs = 0;
3246: cum = cumarray = 0;
3247: pcbddc->adaptive_constraints_idxs_ptr[0] = 0;
3248: pcbddc->adaptive_constraints_data_ptr[0] = 0;
3249: if (sub_schurs->is_vertices && pcbddc->use_vertices) {
3250: const PetscInt *idxs;
3252: ISGetIndices(sub_schurs->is_vertices,&idxs);
3253: for (cum=0;cum<nv;cum++) {
3254: pcbddc->adaptive_constraints_n[cum] = 1;
3255: pcbddc->adaptive_constraints_idxs[cum] = idxs[cum];
3256: pcbddc->adaptive_constraints_data[cum] = 1.0;
3257: pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum]+1;
3258: pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum]+1;
3259: }
3260: ISRestoreIndices(sub_schurs->is_vertices,&idxs);
3261: }
3263: if (mss) { /* multilevel */
3264: MatSeqAIJGetArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);
3265: MatSeqAIJGetArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);
3266: }
3268: lthresh = pcbddc->adaptive_threshold[0];
3269: uthresh = pcbddc->adaptive_threshold[1];
3270: for (i=0;i<sub_schurs->n_subs;i++) {
3271: const PetscInt *idxs;
3272: PetscReal upper,lower;
3273: PetscInt j,subset_size,eigs_start = 0;
3274: PetscBLASInt B_N;
3275: PetscBool same_data = PETSC_FALSE;
3276: PetscBool scal = PETSC_FALSE;
3278: if (pcbddc->use_deluxe_scaling) {
3279: upper = PETSC_MAX_REAL;
3280: lower = uthresh;
3281: } else {
3282: if (!sub_schurs->is_posdef) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented without deluxe scaling");
3283: upper = 1./uthresh;
3284: lower = 0.;
3285: }
3286: ISGetLocalSize(sub_schurs->is_subs[i],&subset_size);
3287: ISGetIndices(sub_schurs->is_subs[i],&idxs);
3288: PetscBLASIntCast(subset_size,&B_N);
3289: /* this is experimental: we assume the dofs have been properly grouped to have
3290: the diagonal blocks Schur complements either positive or negative definite (true for Stokes) */
3291: if (!sub_schurs->is_posdef) {
3292: Mat T;
3294: for (j=0;j<subset_size;j++) {
3295: if (PetscRealPart(*(Sarray+cumarray+j*(subset_size+1))) < 0.0) {
3296: MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,Sarray+cumarray,&T);
3297: MatScale(T,-1.0);
3298: MatDestroy(&T);
3299: MatCreateSeqDense(PETSC_COMM_SELF,subset_size,subset_size,Starray+cumarray,&T);
3300: MatScale(T,-1.0);
3301: MatDestroy(&T);
3302: if (sub_schurs->change_primal_sub) {
3303: PetscInt nz,k;
3304: const PetscInt *idxs;
3306: ISGetLocalSize(sub_schurs->change_primal_sub[i],&nz);
3307: ISGetIndices(sub_schurs->change_primal_sub[i],&idxs);
3308: for (k=0;k<nz;k++) {
3309: *( Sarray + cumarray + idxs[k]*(subset_size+1)) *= -1.0;
3310: *(Starray + cumarray + idxs[k]*(subset_size+1)) = 0.0;
3311: }
3312: ISRestoreIndices(sub_schurs->change_primal_sub[i],&idxs);
3313: }
3314: scal = PETSC_TRUE;
3315: break;
3316: }
3317: }
3318: }
3320: if (allocated_S_St) { /* S and S_t should be copied since we could need them later */
3321: if (sub_schurs->is_symmetric) {
3322: PetscInt j,k;
3323: if (sub_schurs->n_subs == 1) { /* zeroing memory to use PetscArraycmp() later */
3324: PetscArrayzero(S,subset_size*subset_size);
3325: PetscArrayzero(St,subset_size*subset_size);
3326: }
3327: for (j=0;j<subset_size;j++) {
3328: for (k=j;k<subset_size;k++) {
3329: S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k];
3330: St[j*subset_size+k] = Starray[cumarray+j*subset_size+k];
3331: }
3332: }
3333: } else {
3334: PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);
3335: PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);
3336: }
3337: } else {
3338: S = Sarray + cumarray;
3339: St = Starray + cumarray;
3340: }
3341: /* see if we can save some work */
3342: if (sub_schurs->n_subs == 1 && pcbddc->use_deluxe_scaling) {
3343: PetscArraycmp(S,St,subset_size*subset_size,&same_data);
3344: }
3346: if (same_data && !sub_schurs->change) { /* there's no need of constraints here */
3347: B_neigs = 0;
3348: } else {
3349: if (sub_schurs->is_symmetric) {
3350: PetscBLASInt B_itype = 1;
3351: PetscBLASInt B_IL, B_IU;
3352: PetscReal eps = -1.0; /* dlamch? */
3353: PetscInt nmin_s;
3354: PetscBool compute_range;
3356: B_neigs = 0;
3357: compute_range = (PetscBool)!same_data;
3358: if (nmin >= subset_size) compute_range = PETSC_FALSE;
3360: if (pcbddc->dbg_flag) {
3361: PetscInt nc = 0;
3363: if (sub_schurs->change_primal_sub) {
3364: ISGetLocalSize(sub_schurs->change_primal_sub[i],&nc);
3365: }
3366: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Computing for sub %D/%D size %D count %D fid %D (range %d) (change %D).\n",i,sub_schurs->n_subs,subset_size,pcbddc->mat_graph->count[idxs[0]]+1,pcbddc->mat_graph->which_dof[idxs[0]],compute_range,nc);
3367: }
3369: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
3370: if (compute_range) {
3372: /* ask for eigenvalues larger than thresh */
3373: if (sub_schurs->is_posdef) {
3374: #if defined(PETSC_USE_COMPLEX)
3375: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3376: #else
3377: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3378: #endif
3379: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3380: } else { /* no theory so far, but it works nicely */
3381: PetscInt recipe = 0,recipe_m = 1;
3382: PetscReal bb[2];
3384: PetscOptionsGetInt(NULL,((PetscObject)pc)->prefix,"-pc_bddc_adaptive_recipe",&recipe,NULL);
3385: switch (recipe) {
3386: case 0:
3387: if (scal) { bb[0] = PETSC_MIN_REAL; bb[1] = lthresh; }
3388: else { bb[0] = uthresh; bb[1] = PETSC_MAX_REAL; }
3389: #if defined(PETSC_USE_COMPLEX)
3390: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3391: #else
3392: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3393: #endif
3394: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3395: break;
3396: case 1:
3397: bb[0] = PETSC_MIN_REAL; bb[1] = lthresh*lthresh;
3398: #if defined(PETSC_USE_COMPLEX)
3399: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3400: #else
3401: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3402: #endif
3403: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3404: if (!scal) {
3405: PetscBLASInt B_neigs2 = 0;
3407: bb[0] = PetscMax(lthresh*lthresh,uthresh); bb[1] = PETSC_MAX_REAL;
3408: PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);
3409: PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);
3410: #if defined(PETSC_USE_COMPLEX)
3411: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3412: #else
3413: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3414: #endif
3415: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3416: B_neigs += B_neigs2;
3417: }
3418: break;
3419: case 2:
3420: if (scal) {
3421: bb[0] = PETSC_MIN_REAL;
3422: bb[1] = 0;
3423: #if defined(PETSC_USE_COMPLEX)
3424: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3425: #else
3426: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3427: #endif
3428: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3429: } else {
3430: PetscBLASInt B_neigs2 = 0;
3431: PetscBool import = PETSC_FALSE;
3433: lthresh = PetscMax(lthresh,0.0);
3434: if (lthresh > 0.0) {
3435: bb[0] = PETSC_MIN_REAL;
3436: bb[1] = lthresh*lthresh;
3438: import = PETSC_TRUE;
3439: #if defined(PETSC_USE_COMPLEX)
3440: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3441: #else
3442: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3443: #endif
3444: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3445: }
3446: bb[0] = PetscMax(lthresh*lthresh,uthresh);
3447: bb[1] = PETSC_MAX_REAL;
3448: if (import) {
3449: PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);
3450: PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);
3451: }
3452: #if defined(PETSC_USE_COMPLEX)
3453: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3454: #else
3455: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3456: #endif
3457: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3458: B_neigs += B_neigs2;
3459: }
3460: break;
3461: case 3:
3462: if (scal) {
3463: PetscOptionsGetInt(NULL,((PetscObject)pc)->prefix,"-pc_bddc_adaptive_recipe3_min_scal",&recipe_m,NULL);
3464: } else {
3465: PetscOptionsGetInt(NULL,((PetscObject)pc)->prefix,"-pc_bddc_adaptive_recipe3_min",&recipe_m,NULL);
3466: }
3467: if (!scal) {
3468: bb[0] = uthresh;
3469: bb[1] = PETSC_MAX_REAL;
3470: #if defined(PETSC_USE_COMPLEX)
3471: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3472: #else
3473: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3474: #endif
3475: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3476: }
3477: if (recipe_m > 0 && B_N - B_neigs > 0) {
3478: PetscBLASInt B_neigs2 = 0;
3480: B_IL = 1;
3481: PetscBLASIntCast(PetscMin(recipe_m,B_N - B_neigs),&B_IU);
3482: PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);
3483: PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);
3484: #if defined(PETSC_USE_COMPLEX)
3485: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3486: #else
3487: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3488: #endif
3489: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3490: B_neigs += B_neigs2;
3491: }
3492: break;
3493: case 4:
3494: bb[0] = PETSC_MIN_REAL; bb[1] = lthresh;
3495: #if defined(PETSC_USE_COMPLEX)
3496: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3497: #else
3498: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3499: #endif
3500: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3501: {
3502: PetscBLASInt B_neigs2 = 0;
3504: bb[0] = PetscMax(lthresh+PETSC_SMALL,uthresh); bb[1] = PETSC_MAX_REAL;
3505: PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);
3506: PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);
3507: #if defined(PETSC_USE_COMPLEX)
3508: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3509: #else
3510: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","V","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*B_N,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3511: #endif
3512: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3513: B_neigs += B_neigs2;
3514: }
3515: break;
3516: case 5: /* same as before: first compute all eigenvalues, then filter */
3517: #if defined(PETSC_USE_COMPLEX)
3518: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","A","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3519: #else
3520: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","A","L",&B_N,St,&B_N,S,&B_N,&bb[0],&bb[1],&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3521: #endif
3522: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3523: {
3524: PetscInt e,k,ne;
3525: for (e=0,ne=0;e<B_neigs;e++) {
3526: if (eigs[e] < lthresh || eigs[e] > uthresh) {
3527: for (k=0;k<B_N;k++) S[ne*B_N+k] = eigv[e*B_N+k];
3528: eigs[ne] = eigs[e];
3529: ne++;
3530: }
3531: }
3532: PetscArraycpy(eigv,S,B_N*ne);
3533: B_neigs = ne;
3534: }
3535: break;
3536: default:
3537: SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Unknown recipe %D",recipe);
3538: break;
3539: }
3540: }
3541: } else if (!same_data) { /* this is just to see all the eigenvalues */
3542: B_IU = PetscMax(1,PetscMin(B_N,nmax));
3543: B_IL = 1;
3544: #if defined(PETSC_USE_COMPLEX)
3545: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3546: #else
3547: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs,eigs,eigv,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3548: #endif
3549: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3550: } else { /* same_data is true, so just get the adaptive functional requested by the user */
3551: PetscInt k;
3552: if (!sub_schurs->change_primal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");
3553: ISGetLocalSize(sub_schurs->change_primal_sub[i],&nmax);
3554: PetscBLASIntCast(nmax,&B_neigs);
3555: nmin = nmax;
3556: PetscArrayzero(eigv,subset_size*nmax);
3557: for (k=0;k<nmax;k++) {
3558: eigs[k] = 1./PETSC_SMALL;
3559: eigv[k*(subset_size+1)] = 1.0;
3560: }
3561: }
3562: PetscFPTrapPop();
3563: if (B_ierr) {
3564: if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr);
3565: else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr);
3566: else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1);
3567: }
3569: if (B_neigs > nmax) {
3570: if (pcbddc->dbg_flag) {
3571: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, more than maximum required %D.\n",B_neigs,nmax);
3572: }
3573: if (pcbddc->use_deluxe_scaling) eigs_start = scal ? 0 : B_neigs-nmax;
3574: B_neigs = nmax;
3575: }
3577: nmin_s = PetscMin(nmin,B_N);
3578: if (B_neigs < nmin_s) {
3579: PetscBLASInt B_neigs2 = 0;
3581: if (pcbddc->use_deluxe_scaling) {
3582: if (scal) {
3583: B_IU = nmin_s;
3584: B_IL = B_neigs + 1;
3585: } else {
3586: B_IL = B_N - nmin_s + 1;
3587: B_IU = B_N - B_neigs;
3588: }
3589: } else {
3590: B_IL = B_neigs + 1;
3591: B_IU = nmin_s;
3592: }
3593: if (pcbddc->dbg_flag) {
3594: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," found %d eigs, less than minimum required %D. Asking for %d to %d incl (fortran like)\n",B_neigs,nmin,B_IL,B_IU);
3595: }
3596: if (sub_schurs->is_symmetric) {
3597: PetscInt j,k;
3598: for (j=0;j<subset_size;j++) {
3599: for (k=j;k<subset_size;k++) {
3600: S [j*subset_size+k] = Sarray [cumarray+j*subset_size+k];
3601: St[j*subset_size+k] = Starray[cumarray+j*subset_size+k];
3602: }
3603: }
3604: } else {
3605: PetscArraycpy(S,Sarray+cumarray,subset_size*subset_size);
3606: PetscArraycpy(St,Starray+cumarray,subset_size*subset_size);
3607: }
3608: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
3609: #if defined(PETSC_USE_COMPLEX)
3610: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,rwork,B_iwork,B_ifail,&B_ierr));
3611: #else
3612: PetscStackCallBLAS("LAPACKsygvx",LAPACKsygvx_(&B_itype,"V","I","L",&B_N,St,&B_N,S,&B_N,&lower,&upper,&B_IL,&B_IU,&eps,&B_neigs2,eigs+B_neigs,eigv+B_neigs*subset_size,&B_N,work,&B_lwork,B_iwork,B_ifail,&B_ierr));
3613: #endif
3614: PetscLogFlops((4.0*subset_size*subset_size*subset_size)/3.0);
3615: PetscFPTrapPop();
3616: B_neigs += B_neigs2;
3617: }
3618: if (B_ierr) {
3619: if (B_ierr < 0 ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: illegal value for argument %d",-(int)B_ierr);
3620: else if (B_ierr <= B_N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: %d eigenvalues failed to converge",(int)B_ierr);
3621: else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYGVX Lapack routine: leading minor of order %d is not positive definite",(int)B_ierr-B_N-1);
3622: }
3623: if (pcbddc->dbg_flag) {
3624: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Got %d eigs\n",B_neigs);
3625: for (j=0;j<B_neigs;j++) {
3626: if (eigs[j] == 0.0) {
3627: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," Inf\n");
3628: } else {
3629: if (pcbddc->use_deluxe_scaling) {
3630: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",eigs[j+eigs_start]);
3631: } else {
3632: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.6e\n",1./eigs[j+eigs_start]);
3633: }
3634: }
3635: }
3636: }
3637: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented");
3638: }
3639: /* change the basis back to the original one */
3640: if (sub_schurs->change) {
3641: Mat change,phi,phit;
3643: if (pcbddc->dbg_flag > 2) {
3644: PetscInt ii;
3645: for (ii=0;ii<B_neigs;ii++) {
3646: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector (old basis) %d/%d (%d)\n",ii,B_neigs,B_N);
3647: for (j=0;j<B_N;j++) {
3648: #if defined(PETSC_USE_COMPLEX)
3649: PetscReal r = PetscRealPart(eigv[(ii+eigs_start)*subset_size+j]);
3650: PetscReal c = PetscImaginaryPart(eigv[(ii+eigs_start)*subset_size+j]);
3651: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);
3652: #else
3653: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",eigv[(ii+eigs_start)*subset_size+j]);
3654: #endif
3655: }
3656: }
3657: }
3658: KSPGetOperators(sub_schurs->change[i],&change,NULL);
3659: MatCreateSeqDense(PETSC_COMM_SELF,subset_size,B_neigs,eigv+eigs_start*subset_size,&phit);
3660: MatMatMult(change,phit,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&phi);
3661: MatCopy(phi,phit,SAME_NONZERO_PATTERN);
3662: MatDestroy(&phit);
3663: MatDestroy(&phi);
3664: }
3665: maxneigs = PetscMax(B_neigs,maxneigs);
3666: pcbddc->adaptive_constraints_n[i+nv] = B_neigs;
3667: if (B_neigs) {
3668: PetscArraycpy(pcbddc->adaptive_constraints_data+pcbddc->adaptive_constraints_data_ptr[cum],eigv+eigs_start*subset_size,B_neigs*subset_size);
3670: if (pcbddc->dbg_flag > 1) {
3671: PetscInt ii;
3672: for (ii=0;ii<B_neigs;ii++) {
3673: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," -> Eigenvector %d/%d (%d)\n",ii,B_neigs,B_N);
3674: for (j=0;j<B_N;j++) {
3675: #if defined(PETSC_USE_COMPLEX)
3676: PetscReal r = PetscRealPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);
3677: PetscReal c = PetscImaginaryPart(pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);
3678: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e + %1.4e i\n",r,c);
3679: #else
3680: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer," %1.4e\n",pcbddc->adaptive_constraints_data[ii*subset_size+j+pcbddc->adaptive_constraints_data_ptr[cum]]);
3681: #endif
3682: }
3683: }
3684: }
3685: PetscArraycpy(pcbddc->adaptive_constraints_idxs+pcbddc->adaptive_constraints_idxs_ptr[cum],idxs,subset_size);
3686: pcbddc->adaptive_constraints_idxs_ptr[cum+1] = pcbddc->adaptive_constraints_idxs_ptr[cum] + subset_size;
3687: pcbddc->adaptive_constraints_data_ptr[cum+1] = pcbddc->adaptive_constraints_data_ptr[cum] + subset_size*B_neigs;
3688: cum++;
3689: }
3690: ISRestoreIndices(sub_schurs->is_subs[i],&idxs);
3691: /* shift for next computation */
3692: cumarray += subset_size*subset_size;
3693: }
3694: if (pcbddc->dbg_flag) {
3695: PetscViewerFlush(pcbddc->dbg_viewer);
3696: }
3698: if (mss) {
3699: MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_inv_all,&Sarray);
3700: MatSeqAIJRestoreArray(sub_schurs->sum_S_Ej_tilda_all,&Starray);
3701: /* destroy matrices (junk) */
3702: MatDestroy(&sub_schurs->sum_S_Ej_inv_all);
3703: MatDestroy(&sub_schurs->sum_S_Ej_tilda_all);
3704: }
3705: if (allocated_S_St) {
3706: PetscFree2(S,St);
3707: }
3708: PetscFree5(eigv,eigs,work,B_iwork,B_ifail);
3709: #if defined(PETSC_USE_COMPLEX)
3710: PetscFree(rwork);
3711: #endif
3712: if (pcbddc->dbg_flag) {
3713: PetscInt maxneigs_r;
3714: MPIU_Allreduce(&maxneigs,&maxneigs_r,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));
3715: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of constraints per cc %D\n",maxneigs_r);
3716: }
3717: PetscLogEventEnd(PC_BDDC_AdaptiveSetUp[pcbddc->current_level],pc,0,0,0);
3718: return(0);
3719: }
3721: PetscErrorCode PCBDDCSetUpSolvers(PC pc)
3722: {
3723: PetscScalar *coarse_submat_vals;
3727: /* Setup local scatters R_to_B and (optionally) R_to_D */
3728: /* PCBDDCSetUpLocalWorkVectors should be called first! */
3729: PCBDDCSetUpLocalScatters(pc);
3731: /* Setup local neumann solver ksp_R */
3732: /* PCBDDCSetUpLocalScatters should be called first! */
3733: PCBDDCSetUpLocalSolvers(pc,PETSC_FALSE,PETSC_TRUE);
3735: /*
3736: Setup local correction and local part of coarse basis.
3737: Gives back the dense local part of the coarse matrix in column major ordering
3738: */
3739: PCBDDCSetUpCorrection(pc,&coarse_submat_vals);
3741: /* Compute total number of coarse nodes and setup coarse solver */
3742: PCBDDCSetUpCoarseSolver(pc,coarse_submat_vals);
3744: /* free */
3745: PetscFree(coarse_submat_vals);
3746: return(0);
3747: }
3749: PetscErrorCode PCBDDCResetCustomization(PC pc)
3750: {
3751: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
3755: ISDestroy(&pcbddc->user_primal_vertices);
3756: ISDestroy(&pcbddc->user_primal_vertices_local);
3757: ISDestroy(&pcbddc->NeumannBoundaries);
3758: ISDestroy(&pcbddc->NeumannBoundariesLocal);
3759: ISDestroy(&pcbddc->DirichletBoundaries);
3760: MatNullSpaceDestroy(&pcbddc->onearnullspace);
3761: PetscFree(pcbddc->onearnullvecs_state);
3762: ISDestroy(&pcbddc->DirichletBoundariesLocal);
3763: PCBDDCSetDofsSplitting(pc,0,NULL);
3764: PCBDDCSetDofsSplittingLocal(pc,0,NULL);
3765: return(0);
3766: }
3768: PetscErrorCode PCBDDCResetTopography(PC pc)
3769: {
3770: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
3771: PetscInt i;
3775: MatDestroy(&pcbddc->nedcG);
3776: ISDestroy(&pcbddc->nedclocal);
3777: MatDestroy(&pcbddc->discretegradient);
3778: MatDestroy(&pcbddc->user_ChangeOfBasisMatrix);
3779: MatDestroy(&pcbddc->ChangeOfBasisMatrix);
3780: MatDestroy(&pcbddc->switch_static_change);
3781: VecDestroy(&pcbddc->work_change);
3782: MatDestroy(&pcbddc->ConstraintMatrix);
3783: MatDestroy(&pcbddc->divudotp);
3784: ISDestroy(&pcbddc->divudotp_vl2l);
3785: PCBDDCGraphDestroy(&pcbddc->mat_graph);
3786: for (i=0;i<pcbddc->n_local_subs;i++) {
3787: ISDestroy(&pcbddc->local_subs[i]);
3788: }
3789: pcbddc->n_local_subs = 0;
3790: PetscFree(pcbddc->local_subs);
3791: PCBDDCSubSchursDestroy(&pcbddc->sub_schurs);
3792: pcbddc->graphanalyzed = PETSC_FALSE;
3793: pcbddc->recompute_topography = PETSC_TRUE;
3794: pcbddc->corner_selected = PETSC_FALSE;
3795: return(0);
3796: }
3798: PetscErrorCode PCBDDCResetSolvers(PC pc)
3799: {
3800: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
3804: VecDestroy(&pcbddc->coarse_vec);
3805: if (pcbddc->coarse_phi_B) {
3806: PetscScalar *array;
3807: MatDenseGetArray(pcbddc->coarse_phi_B,&array);
3808: PetscFree(array);
3809: }
3810: MatDestroy(&pcbddc->coarse_phi_B);
3811: MatDestroy(&pcbddc->coarse_phi_D);
3812: MatDestroy(&pcbddc->coarse_psi_B);
3813: MatDestroy(&pcbddc->coarse_psi_D);
3814: VecDestroy(&pcbddc->vec1_P);
3815: VecDestroy(&pcbddc->vec1_C);
3816: MatDestroy(&pcbddc->local_auxmat2);
3817: MatDestroy(&pcbddc->local_auxmat1);
3818: VecDestroy(&pcbddc->vec1_R);
3819: VecDestroy(&pcbddc->vec2_R);
3820: ISDestroy(&pcbddc->is_R_local);
3821: VecScatterDestroy(&pcbddc->R_to_B);
3822: VecScatterDestroy(&pcbddc->R_to_D);
3823: VecScatterDestroy(&pcbddc->coarse_loc_to_glob);
3824: KSPReset(pcbddc->ksp_D);
3825: KSPReset(pcbddc->ksp_R);
3826: KSPReset(pcbddc->coarse_ksp);
3827: MatDestroy(&pcbddc->local_mat);
3828: PetscFree(pcbddc->primal_indices_local_idxs);
3829: PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);
3830: PetscFree(pcbddc->global_primal_indices);
3831: ISDestroy(&pcbddc->coarse_subassembling);
3832: MatDestroy(&pcbddc->benign_change);
3833: VecDestroy(&pcbddc->benign_vec);
3834: PCBDDCBenignShellMat(pc,PETSC_TRUE);
3835: MatDestroy(&pcbddc->benign_B0);
3836: PetscSFDestroy(&pcbddc->benign_sf);
3837: if (pcbddc->benign_zerodiag_subs) {
3838: PetscInt i;
3839: for (i=0;i<pcbddc->benign_n;i++) {
3840: ISDestroy(&pcbddc->benign_zerodiag_subs[i]);
3841: }
3842: PetscFree(pcbddc->benign_zerodiag_subs);
3843: }
3844: PetscFree3(pcbddc->benign_p0_lidx,pcbddc->benign_p0_gidx,pcbddc->benign_p0);
3845: return(0);
3846: }
3848: PetscErrorCode PCBDDCSetUpLocalWorkVectors(PC pc)
3849: {
3850: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
3851: PC_IS *pcis = (PC_IS*)pc->data;
3852: VecType impVecType;
3853: PetscInt n_constraints,n_R,old_size;
3857: n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - pcbddc->n_vertices;
3858: n_R = pcis->n - pcbddc->n_vertices;
3859: VecGetType(pcis->vec1_N,&impVecType);
3860: /* local work vectors (try to avoid unneeded work)*/
3861: /* R nodes */
3862: old_size = -1;
3863: if (pcbddc->vec1_R) {
3864: VecGetSize(pcbddc->vec1_R,&old_size);
3865: }
3866: if (n_R != old_size) {
3867: VecDestroy(&pcbddc->vec1_R);
3868: VecDestroy(&pcbddc->vec2_R);
3869: VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_R);
3870: VecSetSizes(pcbddc->vec1_R,PETSC_DECIDE,n_R);
3871: VecSetType(pcbddc->vec1_R,impVecType);
3872: VecDuplicate(pcbddc->vec1_R,&pcbddc->vec2_R);
3873: }
3874: /* local primal dofs */
3875: old_size = -1;
3876: if (pcbddc->vec1_P) {
3877: VecGetSize(pcbddc->vec1_P,&old_size);
3878: }
3879: if (pcbddc->local_primal_size != old_size) {
3880: VecDestroy(&pcbddc->vec1_P);
3881: VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_P);
3882: VecSetSizes(pcbddc->vec1_P,PETSC_DECIDE,pcbddc->local_primal_size);
3883: VecSetType(pcbddc->vec1_P,impVecType);
3884: }
3885: /* local explicit constraints */
3886: old_size = -1;
3887: if (pcbddc->vec1_C) {
3888: VecGetSize(pcbddc->vec1_C,&old_size);
3889: }
3890: if (n_constraints && n_constraints != old_size) {
3891: VecDestroy(&pcbddc->vec1_C);
3892: VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&pcbddc->vec1_C);
3893: VecSetSizes(pcbddc->vec1_C,PETSC_DECIDE,n_constraints);
3894: VecSetType(pcbddc->vec1_C,impVecType);
3895: }
3896: return(0);
3897: }
3899: PetscErrorCode PCBDDCSetUpCorrection(PC pc, PetscScalar **coarse_submat_vals_n)
3900: {
3901: PetscErrorCode ierr;
3902: /* pointers to pcis and pcbddc */
3903: PC_IS* pcis = (PC_IS*)pc->data;
3904: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
3905: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
3906: /* submatrices of local problem */
3907: Mat A_RV,A_VR,A_VV,local_auxmat2_R;
3908: /* submatrices of local coarse problem */
3909: Mat S_VV,S_CV,S_VC,S_CC;
3910: /* working matrices */
3911: Mat C_CR;
3912: /* additional working stuff */
3913: PC pc_R;
3914: Mat F,Brhs = NULL;
3915: Vec dummy_vec;
3916: PetscBool isLU,isCHOL,need_benign_correction,sparserhs;
3917: PetscScalar *coarse_submat_vals; /* TODO: use a PETSc matrix */
3918: PetscScalar *work;
3919: PetscInt *idx_V_B;
3920: PetscInt lda_rhs,n,n_vertices,n_constraints,*p0_lidx_I;
3921: PetscInt i,n_R,n_D,n_B;
3922: PetscScalar one=1.0,m_one=-1.0;
3925: if (!pcbddc->symmetric_primal && pcbddc->benign_n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Non-symmetric primal basis computation with benign trick not yet implemented");
3926: PetscLogEventBegin(PC_BDDC_CorrectionSetUp[pcbddc->current_level],pc,0,0,0);
3928: /* Set Non-overlapping dimensions */
3929: n_vertices = pcbddc->n_vertices;
3930: n_constraints = pcbddc->local_primal_size - pcbddc->benign_n - n_vertices;
3931: n_B = pcis->n_B;
3932: n_D = pcis->n - n_B;
3933: n_R = pcis->n - n_vertices;
3935: /* vertices in boundary numbering */
3936: PetscMalloc1(n_vertices,&idx_V_B);
3937: ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,n_vertices,pcbddc->local_primal_ref_node,&i,idx_V_B);
3938: if (i != n_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D",n_vertices,i);
3940: /* Subdomain contribution (Non-overlapping) to coarse matrix */
3941: PetscCalloc1(pcbddc->local_primal_size*pcbddc->local_primal_size,&coarse_submat_vals);
3942: MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_vertices,coarse_submat_vals,&S_VV);
3943: MatSeqDenseSetLDA(S_VV,pcbddc->local_primal_size);
3944: MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_vertices,coarse_submat_vals+n_vertices,&S_CV);
3945: MatSeqDenseSetLDA(S_CV,pcbddc->local_primal_size);
3946: MatCreateSeqDense(PETSC_COMM_SELF,n_vertices,n_constraints,coarse_submat_vals+pcbddc->local_primal_size*n_vertices,&S_VC);
3947: MatSeqDenseSetLDA(S_VC,pcbddc->local_primal_size);
3948: MatCreateSeqDense(PETSC_COMM_SELF,n_constraints,n_constraints,coarse_submat_vals+(pcbddc->local_primal_size+1)*n_vertices,&S_CC);
3949: MatSeqDenseSetLDA(S_CC,pcbddc->local_primal_size);
3951: /* determine if can use MatSolve routines instead of calling KSPSolve on ksp_R */
3952: KSPGetPC(pcbddc->ksp_R,&pc_R);
3953: PCSetUp(pc_R);
3954: PetscObjectTypeCompare((PetscObject)pc_R,PCLU,&isLU);
3955: PetscObjectTypeCompare((PetscObject)pc_R,PCCHOLESKY,&isCHOL);
3956: lda_rhs = n_R;
3957: need_benign_correction = PETSC_FALSE;
3958: if (isLU || isCHOL) {
3959: PCFactorGetMatrix(pc_R,&F);
3960: } else if (sub_schurs && sub_schurs->reuse_solver) {
3961: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
3962: MatFactorType type;
3964: F = reuse_solver->F;
3965: MatGetFactorType(F,&type);
3966: if (type == MAT_FACTOR_CHOLESKY) isCHOL = PETSC_TRUE;
3967: if (type == MAT_FACTOR_LU) isLU = PETSC_TRUE;
3968: MatGetSize(F,&lda_rhs,NULL);
3969: need_benign_correction = (PetscBool)(!!reuse_solver->benign_n);
3970: } else F = NULL;
3972: /* determine if we can use a sparse right-hand side */
3973: sparserhs = PETSC_FALSE;
3974: if (F) {
3975: MatSolverType solver;
3977: MatFactorGetSolverType(F,&solver);
3978: PetscStrcmp(solver,MATSOLVERMUMPS,&sparserhs);
3979: }
3981: /* allocate workspace */
3982: n = 0;
3983: if (n_constraints) {
3984: n += lda_rhs*n_constraints;
3985: }
3986: if (n_vertices) {
3987: n = PetscMax(2*lda_rhs*n_vertices,n);
3988: n = PetscMax((lda_rhs+n_B)*n_vertices,n);
3989: }
3990: if (!pcbddc->symmetric_primal) {
3991: n = PetscMax(2*lda_rhs*pcbddc->local_primal_size,n);
3992: }
3993: PetscMalloc1(n,&work);
3995: /* create dummy vector to modify rhs and sol of MatMatSolve (work array will never be used) */
3996: dummy_vec = NULL;
3997: if (need_benign_correction && lda_rhs != n_R && F) {
3998: VecCreate(PetscObjectComm((PetscObject)pcis->vec1_N),&dummy_vec);
3999: VecSetSizes(dummy_vec,lda_rhs,PETSC_DECIDE);
4000: VecSetType(dummy_vec,((PetscObject)pcis->vec1_N)->type_name);
4001: }
4003: MatDestroy(&pcbddc->local_auxmat1);
4004: MatDestroy(&pcbddc->local_auxmat2);
4006: /* Precompute stuffs needed for preprocessing and Section 1.5 Writing Application Codes with PETSc of BDDC*/
4007: if (n_constraints) {
4008: Mat M3,C_B;
4009: IS is_aux;
4010: PetscScalar *array,*array2;
4012: /* Extract constraints on R nodes: C_{CR} */
4013: ISCreateStride(PETSC_COMM_SELF,n_constraints,n_vertices,1,&is_aux);
4014: MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&C_CR);
4015: MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_aux,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);
4017: /* Assemble local_auxmat2_R = (- A_{RR}^{-1} C^T_{CR}) needed by BDDC setup */
4018: /* Assemble pcbddc->local_auxmat2 = R_to_B (- A_{RR}^{-1} C^T_{CR}) needed by BDDC Section 1.5 Writing Application Codes with PETSc */
4019: if (!sparserhs) {
4020: PetscArrayzero(work,lda_rhs*n_constraints);
4021: for (i=0;i<n_constraints;i++) {
4022: const PetscScalar *row_cmat_values;
4023: const PetscInt *row_cmat_indices;
4024: PetscInt size_of_constraint,j;
4026: MatGetRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);
4027: for (j=0;j<size_of_constraint;j++) {
4028: work[row_cmat_indices[j]+i*lda_rhs] = -row_cmat_values[j];
4029: }
4030: MatRestoreRow(C_CR,i,&size_of_constraint,&row_cmat_indices,&row_cmat_values);
4031: }
4032: MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&Brhs);
4033: } else {
4034: Mat tC_CR;
4036: MatScale(C_CR,-1.0);
4037: if (lda_rhs != n_R) {
4038: PetscScalar *aa;
4039: PetscInt r,*ii,*jj;
4040: PetscBool done;
4042: MatGetRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);
4043: if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed");
4044: MatSeqAIJGetArray(C_CR,&aa);
4045: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_constraints,lda_rhs,ii,jj,aa,&tC_CR);
4046: MatRestoreRowIJ(C_CR,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);
4047: if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed");
4048: } else {
4049: PetscObjectReference((PetscObject)C_CR);
4050: tC_CR = C_CR;
4051: }
4052: MatCreateTranspose(tC_CR,&Brhs);
4053: MatDestroy(&tC_CR);
4054: }
4055: MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,NULL,&local_auxmat2_R);
4056: if (F) {
4057: if (need_benign_correction) {
4058: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4060: /* rhs is already zero on interior dofs, no need to change the rhs */
4061: PetscArrayzero(reuse_solver->benign_save_vals,pcbddc->benign_n);
4062: }
4063: MatMatSolve(F,Brhs,local_auxmat2_R);
4064: if (need_benign_correction) {
4065: PetscScalar *marr;
4066: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4068: MatDenseGetArray(local_auxmat2_R,&marr);
4069: if (lda_rhs != n_R) {
4070: for (i=0;i<n_constraints;i++) {
4071: VecPlaceArray(dummy_vec,marr+i*lda_rhs);
4072: PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);
4073: VecResetArray(dummy_vec);
4074: }
4075: } else {
4076: for (i=0;i<n_constraints;i++) {
4077: VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);
4078: PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);
4079: VecResetArray(pcbddc->vec1_R);
4080: }
4081: }
4082: MatDenseRestoreArray(local_auxmat2_R,&marr);
4083: }
4084: } else {
4085: PetscScalar *marr;
4087: MatDenseGetArray(local_auxmat2_R,&marr);
4088: for (i=0;i<n_constraints;i++) {
4089: VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);
4090: VecPlaceArray(pcbddc->vec2_R,marr+i*lda_rhs);
4091: KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);
4092: KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);
4093: VecResetArray(pcbddc->vec1_R);
4094: VecResetArray(pcbddc->vec2_R);
4095: }
4096: MatDenseRestoreArray(local_auxmat2_R,&marr);
4097: }
4098: if (sparserhs) {
4099: MatScale(C_CR,-1.0);
4100: }
4101: MatDestroy(&Brhs);
4102: if (!pcbddc->switch_static) {
4103: MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_constraints,NULL,&pcbddc->local_auxmat2);
4104: MatDenseGetArray(pcbddc->local_auxmat2,&array);
4105: MatDenseGetArray(local_auxmat2_R,&array2);
4106: for (i=0;i<n_constraints;i++) {
4107: VecPlaceArray(pcbddc->vec1_R,array2+i*lda_rhs);
4108: VecPlaceArray(pcis->vec1_B,array+i*n_B);
4109: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4110: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4111: VecResetArray(pcis->vec1_B);
4112: VecResetArray(pcbddc->vec1_R);
4113: }
4114: MatDenseRestoreArray(local_auxmat2_R,&array2);
4115: MatDenseRestoreArray(pcbddc->local_auxmat2,&array);
4116: MatMatMult(C_B,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);
4117: } else {
4118: if (lda_rhs != n_R) {
4119: IS dummy;
4121: ISCreateStride(PETSC_COMM_SELF,n_R,0,1,&dummy);
4122: MatCreateSubMatrix(local_auxmat2_R,dummy,NULL,MAT_INITIAL_MATRIX,&pcbddc->local_auxmat2);
4123: ISDestroy(&dummy);
4124: } else {
4125: PetscObjectReference((PetscObject)local_auxmat2_R);
4126: pcbddc->local_auxmat2 = local_auxmat2_R;
4127: }
4128: MatMatMult(C_CR,pcbddc->local_auxmat2,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&M3);
4129: }
4130: ISDestroy(&is_aux);
4131: /* Assemble explicitly S_CC = ( C_{CR} A_{RR}^{-1} C^T_{CR} )^{-1} */
4132: MatScale(M3,m_one);
4133: if (isCHOL) {
4134: MatCholeskyFactor(M3,NULL,NULL);
4135: } else {
4136: MatLUFactor(M3,NULL,NULL,NULL);
4137: }
4138: MatSeqDenseInvertFactors_Private(M3);
4139: /* Assemble local_auxmat1 = S_CC*C_{CB} needed by BDDC Section 1.5 Writing Application Codes with PETSc in KSP and in preproc */
4140: MatMatMult(M3,C_B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->local_auxmat1);
4141: MatDestroy(&C_B);
4142: MatCopy(M3,S_CC,SAME_NONZERO_PATTERN); /* S_CC can have a different LDA, MatMatSolve doesn't support it */
4143: MatDestroy(&M3);
4144: }
4146: /* Get submatrices from subdomain matrix */
4147: if (n_vertices) {
4148: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
4149: PetscBool oldpin;
4150: #endif
4151: PetscBool isaij;
4152: IS is_aux;
4154: if (sub_schurs && sub_schurs->reuse_solver) { /* is_R_local is not sorted, ISComplement doesn't like it */
4155: IS tis;
4157: ISDuplicate(pcbddc->is_R_local,&tis);
4158: ISSort(tis);
4159: ISComplement(tis,0,pcis->n,&is_aux);
4160: ISDestroy(&tis);
4161: } else {
4162: ISComplement(pcbddc->is_R_local,0,pcis->n,&is_aux);
4163: }
4164: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
4165: oldpin = pcbddc->local_mat->boundtocpu;
4166: #endif
4167: MatBindToCPU(pcbddc->local_mat,PETSC_TRUE);
4168: MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,is_aux,MAT_INITIAL_MATRIX,&A_RV);
4169: MatCreateSubMatrix(pcbddc->local_mat,is_aux,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_VR);
4170: PetscObjectBaseTypeCompare((PetscObject)A_VR,MATSEQAIJ,&isaij);
4171: if (!isaij) { /* TODO REMOVE: MatMatMult(A_VR,A_RRmA_RV) below may raise an error */
4172: MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);
4173: }
4174: MatCreateSubMatrix(pcbddc->local_mat,is_aux,is_aux,MAT_INITIAL_MATRIX,&A_VV);
4175: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
4176: MatBindToCPU(pcbddc->local_mat,oldpin);
4177: #endif
4178: ISDestroy(&is_aux);
4179: }
4181: /* Matrix of coarse basis functions (local) */
4182: if (pcbddc->coarse_phi_B) {
4183: PetscInt on_B,on_primal,on_D=n_D;
4184: if (pcbddc->coarse_phi_D) {
4185: MatGetSize(pcbddc->coarse_phi_D,&on_D,NULL);
4186: }
4187: MatGetSize(pcbddc->coarse_phi_B,&on_B,&on_primal);
4188: if (on_B != n_B || on_primal != pcbddc->local_primal_size || on_D != n_D) {
4189: PetscScalar *marray;
4191: MatDenseGetArray(pcbddc->coarse_phi_B,&marray);
4192: PetscFree(marray);
4193: MatDestroy(&pcbddc->coarse_phi_B);
4194: MatDestroy(&pcbddc->coarse_psi_B);
4195: MatDestroy(&pcbddc->coarse_phi_D);
4196: MatDestroy(&pcbddc->coarse_psi_D);
4197: }
4198: }
4200: if (!pcbddc->coarse_phi_B) {
4201: PetscScalar *marr;
4203: /* memory size */
4204: n = n_B*pcbddc->local_primal_size;
4205: if (pcbddc->switch_static || pcbddc->dbg_flag) n += n_D*pcbddc->local_primal_size;
4206: if (!pcbddc->symmetric_primal) n *= 2;
4207: PetscCalloc1(n,&marr);
4208: MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_B);
4209: marr += n_B*pcbddc->local_primal_size;
4210: if (pcbddc->switch_static || pcbddc->dbg_flag) {
4211: MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_phi_D);
4212: marr += n_D*pcbddc->local_primal_size;
4213: }
4214: if (!pcbddc->symmetric_primal) {
4215: MatCreateSeqDense(PETSC_COMM_SELF,n_B,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_B);
4216: marr += n_B*pcbddc->local_primal_size;
4217: if (pcbddc->switch_static || pcbddc->dbg_flag) {
4218: MatCreateSeqDense(PETSC_COMM_SELF,n_D,pcbddc->local_primal_size,marr,&pcbddc->coarse_psi_D);
4219: }
4220: } else {
4221: PetscObjectReference((PetscObject)pcbddc->coarse_phi_B);
4222: pcbddc->coarse_psi_B = pcbddc->coarse_phi_B;
4223: if (pcbddc->switch_static || pcbddc->dbg_flag) {
4224: PetscObjectReference((PetscObject)pcbddc->coarse_phi_D);
4225: pcbddc->coarse_psi_D = pcbddc->coarse_phi_D;
4226: }
4227: }
4228: }
4230: /* We are now ready to evaluate coarse basis functions and subdomain contribution to coarse problem */
4231: p0_lidx_I = NULL;
4232: if (pcbddc->benign_n && (pcbddc->switch_static || pcbddc->dbg_flag)) {
4233: const PetscInt *idxs;
4235: ISGetIndices(pcis->is_I_local,&idxs);
4236: PetscMalloc1(pcbddc->benign_n,&p0_lidx_I);
4237: for (i=0;i<pcbddc->benign_n;i++) {
4238: PetscFindInt(pcbddc->benign_p0_lidx[i],pcis->n-pcis->n_B,idxs,&p0_lidx_I[i]);
4239: }
4240: ISRestoreIndices(pcis->is_I_local,&idxs);
4241: }
4243: /* vertices */
4244: if (n_vertices) {
4245: PetscBool restoreavr = PETSC_FALSE;
4247: MatConvert(A_VV,MATDENSE,MAT_INPLACE_MATRIX,&A_VV);
4249: if (n_R) {
4250: Mat A_RRmA_RV,A_RV_bcorr=NULL,S_VVt; /* S_VVt with LDA=N */
4251: PetscBLASInt B_N,B_one = 1;
4252: const PetscScalar *x;
4253: PetscScalar *y;
4255: MatScale(A_RV,m_one);
4256: if (need_benign_correction) {
4257: ISLocalToGlobalMapping RtoN;
4258: IS is_p0;
4259: PetscInt *idxs_p0,n;
4261: PetscMalloc1(pcbddc->benign_n,&idxs_p0);
4262: ISLocalToGlobalMappingCreateIS(pcbddc->is_R_local,&RtoN);
4263: ISGlobalToLocalMappingApply(RtoN,IS_GTOLM_DROP,pcbddc->benign_n,pcbddc->benign_p0_lidx,&n,idxs_p0);
4264: if (n != pcbddc->benign_n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in R numbering for benign p0! %D != %D",n,pcbddc->benign_n);
4265: ISLocalToGlobalMappingDestroy(&RtoN);
4266: ISCreateGeneral(PETSC_COMM_SELF,n,idxs_p0,PETSC_OWN_POINTER,&is_p0);
4267: MatCreateSubMatrix(A_RV,is_p0,NULL,MAT_INITIAL_MATRIX,&A_RV_bcorr);
4268: ISDestroy(&is_p0);
4269: }
4271: MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work,&A_RRmA_RV);
4272: if (!sparserhs || need_benign_correction) {
4273: if (lda_rhs == n_R) {
4274: MatConvert(A_RV,MATDENSE,MAT_INPLACE_MATRIX,&A_RV);
4275: } else {
4276: PetscScalar *av,*array;
4277: const PetscInt *xadj,*adjncy;
4278: PetscInt n;
4279: PetscBool flg_row;
4281: array = work+lda_rhs*n_vertices;
4282: PetscArrayzero(array,lda_rhs*n_vertices);
4283: MatConvert(A_RV,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_RV);
4284: MatGetRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);
4285: MatSeqAIJGetArray(A_RV,&av);
4286: for (i=0;i<n;i++) {
4287: PetscInt j;
4288: for (j=xadj[i];j<xadj[i+1];j++) array[lda_rhs*adjncy[j]+i] = av[j];
4289: }
4290: MatRestoreRowIJ(A_RV,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);
4291: MatDestroy(&A_RV);
4292: MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,array,&A_RV);
4293: }
4294: if (need_benign_correction) {
4295: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4296: PetscScalar *marr;
4298: MatDenseGetArray(A_RV,&marr);
4299: /* need \Phi^T A_RV = (I+L)A_RV, L given by
4301: | 0 0 0 | (V)
4302: L = | 0 0 -1 | (P-p0)
4303: | 0 0 -1 | (p0)
4305: */
4306: for (i=0;i<reuse_solver->benign_n;i++) {
4307: const PetscScalar *vals;
4308: const PetscInt *idxs,*idxs_zero;
4309: PetscInt n,j,nz;
4311: ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);
4312: ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);
4313: MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);
4314: for (j=0;j<n;j++) {
4315: PetscScalar val = vals[j];
4316: PetscInt k,col = idxs[j];
4317: for (k=0;k<nz;k++) marr[idxs_zero[k]+lda_rhs*col] -= val;
4318: }
4319: MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);
4320: ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);
4321: }
4322: MatDenseRestoreArray(A_RV,&marr);
4323: }
4324: PetscObjectReference((PetscObject)A_RV);
4325: Brhs = A_RV;
4326: } else {
4327: Mat tA_RVT,A_RVT;
4329: if (!pcbddc->symmetric_primal) {
4330: /* A_RV already scaled by -1 */
4331: MatTranspose(A_RV,MAT_INITIAL_MATRIX,&A_RVT);
4332: } else {
4333: restoreavr = PETSC_TRUE;
4334: MatScale(A_VR,-1.0);
4335: PetscObjectReference((PetscObject)A_VR);
4336: A_RVT = A_VR;
4337: }
4338: if (lda_rhs != n_R) {
4339: PetscScalar *aa;
4340: PetscInt r,*ii,*jj;
4341: PetscBool done;
4343: MatGetRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);
4344: if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"GetRowIJ failed");
4345: MatSeqAIJGetArray(A_RVT,&aa);
4346: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,n_vertices,lda_rhs,ii,jj,aa,&tA_RVT);
4347: MatRestoreRowIJ(A_RVT,0,PETSC_FALSE,PETSC_FALSE,&r,(const PetscInt**)&ii,(const PetscInt**)&jj,&done);
4348: if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"RestoreRowIJ failed");
4349: } else {
4350: PetscObjectReference((PetscObject)A_RVT);
4351: tA_RVT = A_RVT;
4352: }
4353: MatCreateTranspose(tA_RVT,&Brhs);
4354: MatDestroy(&tA_RVT);
4355: MatDestroy(&A_RVT);
4356: }
4357: if (F) {
4358: /* need to correct the rhs */
4359: if (need_benign_correction) {
4360: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4361: PetscScalar *marr;
4363: MatDenseGetArray(Brhs,&marr);
4364: if (lda_rhs != n_R) {
4365: for (i=0;i<n_vertices;i++) {
4366: VecPlaceArray(dummy_vec,marr+i*lda_rhs);
4367: PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_FALSE,PETSC_TRUE);
4368: VecResetArray(dummy_vec);
4369: }
4370: } else {
4371: for (i=0;i<n_vertices;i++) {
4372: VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);
4373: PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_FALSE,PETSC_TRUE);
4374: VecResetArray(pcbddc->vec1_R);
4375: }
4376: }
4377: MatDenseRestoreArray(Brhs,&marr);
4378: }
4379: MatMatSolve(F,Brhs,A_RRmA_RV);
4380: if (restoreavr) {
4381: MatScale(A_VR,-1.0);
4382: }
4383: /* need to correct the solution */
4384: if (need_benign_correction) {
4385: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4386: PetscScalar *marr;
4388: MatDenseGetArray(A_RRmA_RV,&marr);
4389: if (lda_rhs != n_R) {
4390: for (i=0;i<n_vertices;i++) {
4391: VecPlaceArray(dummy_vec,marr+i*lda_rhs);
4392: PCBDDCReuseSolversBenignAdapt(reuse_solver,dummy_vec,NULL,PETSC_TRUE,PETSC_TRUE);
4393: VecResetArray(dummy_vec);
4394: }
4395: } else {
4396: for (i=0;i<n_vertices;i++) {
4397: VecPlaceArray(pcbddc->vec1_R,marr+i*lda_rhs);
4398: PCBDDCReuseSolversBenignAdapt(reuse_solver,pcbddc->vec1_R,NULL,PETSC_TRUE,PETSC_TRUE);
4399: VecResetArray(pcbddc->vec1_R);
4400: }
4401: }
4402: MatDenseRestoreArray(A_RRmA_RV,&marr);
4403: }
4404: } else {
4405: MatDenseGetArray(Brhs,&y);
4406: for (i=0;i<n_vertices;i++) {
4407: VecPlaceArray(pcbddc->vec1_R,y+i*lda_rhs);
4408: VecPlaceArray(pcbddc->vec2_R,work+i*lda_rhs);
4409: KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);
4410: KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);
4411: VecResetArray(pcbddc->vec1_R);
4412: VecResetArray(pcbddc->vec2_R);
4413: }
4414: MatDenseRestoreArray(Brhs,&y);
4415: }
4416: MatDestroy(&A_RV);
4417: MatDestroy(&Brhs);
4418: /* S_VV and S_CV */
4419: if (n_constraints) {
4420: Mat B;
4422: PetscArrayzero(work+lda_rhs*n_vertices,n_B*n_vertices);
4423: for (i=0;i<n_vertices;i++) {
4424: VecPlaceArray(pcbddc->vec1_R,work+i*lda_rhs);
4425: VecPlaceArray(pcis->vec1_B,work+lda_rhs*n_vertices+i*n_B);
4426: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4427: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4428: VecResetArray(pcis->vec1_B);
4429: VecResetArray(pcbddc->vec1_R);
4430: }
4431: MatCreateSeqDense(PETSC_COMM_SELF,n_B,n_vertices,work+lda_rhs*n_vertices,&B);
4432: /* Reuse dense S_C = pcbddc->local_auxmat1 * B */
4433: MatProductCreateWithMat(pcbddc->local_auxmat1,B,NULL,S_CV);
4434: MatProductSetType(S_CV,MATPRODUCT_AB);
4435: MatProductSetFromOptions(S_CV);
4436: MatProductNumeric(S_CV);
4437: MatProductClear(S_CV);
4439: MatDestroy(&B);
4440: MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_vertices,work+lda_rhs*n_vertices,&B);
4441: /* Reuse B = local_auxmat2_R * S_CV */
4442: MatProductCreateWithMat(local_auxmat2_R,S_CV,NULL,B);
4443: MatProductSetType(B,MATPRODUCT_AB);
4444: MatProductSetFromOptions(B);
4445: MatProductNumeric(B);
4447: MatScale(S_CV,m_one);
4448: PetscBLASIntCast(lda_rhs*n_vertices,&B_N);
4449: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,work+lda_rhs*n_vertices,&B_one,work,&B_one));
4450: MatDestroy(&B);
4451: }
4452: if (lda_rhs != n_R) {
4453: MatDestroy(&A_RRmA_RV);
4454: MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,work,&A_RRmA_RV);
4455: MatSeqDenseSetLDA(A_RRmA_RV,lda_rhs);
4456: }
4457: MatMatMult(A_VR,A_RRmA_RV,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VVt);
4458: /* need A_VR * \Phi * A_RRmA_RV = A_VR * (I+L)^T * A_RRmA_RV, L given as before */
4459: if (need_benign_correction) {
4460: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
4461: PetscScalar *marr,*sums;
4463: PetscMalloc1(n_vertices,&sums);
4464: MatDenseGetArray(S_VVt,&marr);
4465: for (i=0;i<reuse_solver->benign_n;i++) {
4466: const PetscScalar *vals;
4467: const PetscInt *idxs,*idxs_zero;
4468: PetscInt n,j,nz;
4470: ISGetLocalSize(reuse_solver->benign_zerodiag_subs[i],&nz);
4471: ISGetIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);
4472: for (j=0;j<n_vertices;j++) {
4473: PetscInt k;
4474: sums[j] = 0.;
4475: for (k=0;k<nz;k++) sums[j] += work[idxs_zero[k]+j*lda_rhs];
4476: }
4477: MatGetRow(A_RV_bcorr,i,&n,&idxs,&vals);
4478: for (j=0;j<n;j++) {
4479: PetscScalar val = vals[j];
4480: PetscInt k;
4481: for (k=0;k<n_vertices;k++) {
4482: marr[idxs[j]+k*n_vertices] += val*sums[k];
4483: }
4484: }
4485: MatRestoreRow(A_RV_bcorr,i,&n,&idxs,&vals);
4486: ISRestoreIndices(reuse_solver->benign_zerodiag_subs[i],&idxs_zero);
4487: }
4488: PetscFree(sums);
4489: MatDenseRestoreArray(S_VVt,&marr);
4490: MatDestroy(&A_RV_bcorr);
4491: }
4492: MatDestroy(&A_RRmA_RV);
4493: PetscBLASIntCast(n_vertices*n_vertices,&B_N);
4494: MatDenseGetArrayRead(A_VV,&x);
4495: MatDenseGetArray(S_VVt,&y);
4496: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&B_N,&one,x,&B_one,y,&B_one));
4497: MatDenseRestoreArrayRead(A_VV,&x);
4498: MatDenseRestoreArray(S_VVt,&y);
4499: MatCopy(S_VVt,S_VV,SAME_NONZERO_PATTERN);
4500: MatDestroy(&S_VVt);
4501: } else {
4502: MatCopy(A_VV,S_VV,SAME_NONZERO_PATTERN);
4503: }
4504: MatDestroy(&A_VV);
4506: /* coarse basis functions */
4507: for (i=0;i<n_vertices;i++) {
4508: PetscScalar *y;
4510: VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);
4511: MatDenseGetArray(pcbddc->coarse_phi_B,&y);
4512: VecPlaceArray(pcis->vec1_B,y+n_B*i);
4513: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4514: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4515: y[n_B*i+idx_V_B[i]] = 1.0;
4516: MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);
4517: VecResetArray(pcis->vec1_B);
4519: if (pcbddc->switch_static || pcbddc->dbg_flag) {
4520: PetscInt j;
4522: MatDenseGetArray(pcbddc->coarse_phi_D,&y);
4523: VecPlaceArray(pcis->vec1_D,y+n_D*i);
4524: VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
4525: VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
4526: VecResetArray(pcis->vec1_D);
4527: for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0;
4528: MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);
4529: }
4530: VecResetArray(pcbddc->vec1_R);
4531: }
4532: /* if n_R == 0 the object is not destroyed */
4533: MatDestroy(&A_RV);
4534: }
4535: VecDestroy(&dummy_vec);
4537: if (n_constraints) {
4538: Mat B;
4540: MatCreateSeqDense(PETSC_COMM_SELF,lda_rhs,n_constraints,work,&B);
4541: MatScale(S_CC,m_one);
4542: MatProductCreateWithMat(local_auxmat2_R,S_CC,NULL,B);
4543: MatProductSetType(B,MATPRODUCT_AB);
4544: MatProductSetFromOptions(B);
4545: MatProductNumeric(B);
4547: MatScale(S_CC,m_one);
4548: if (n_vertices) {
4549: if (isCHOL || need_benign_correction) { /* if we can solve the interior problem with cholesky, we should also be fine with transposing here */
4550: MatTranspose(S_CV,MAT_REUSE_MATRIX,&S_VC);
4551: } else {
4552: Mat S_VCt;
4554: if (lda_rhs != n_R) {
4555: MatDestroy(&B);
4556: MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_constraints,work,&B);
4557: MatSeqDenseSetLDA(B,lda_rhs);
4558: }
4559: MatMatMult(A_VR,B,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&S_VCt);
4560: MatCopy(S_VCt,S_VC,SAME_NONZERO_PATTERN);
4561: MatDestroy(&S_VCt);
4562: }
4563: }
4564: MatDestroy(&B);
4565: /* coarse basis functions */
4566: for (i=0;i<n_constraints;i++) {
4567: PetscScalar *y;
4569: VecPlaceArray(pcbddc->vec1_R,work+lda_rhs*i);
4570: MatDenseGetArray(pcbddc->coarse_phi_B,&y);
4571: VecPlaceArray(pcis->vec1_B,y+n_B*(i+n_vertices));
4572: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4573: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4574: MatDenseRestoreArray(pcbddc->coarse_phi_B,&y);
4575: VecResetArray(pcis->vec1_B);
4576: if (pcbddc->switch_static || pcbddc->dbg_flag) {
4577: PetscInt j;
4579: MatDenseGetArray(pcbddc->coarse_phi_D,&y);
4580: VecPlaceArray(pcis->vec1_D,y+n_D*(i+n_vertices));
4581: VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
4582: VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
4583: VecResetArray(pcis->vec1_D);
4584: for (j=0;j<pcbddc->benign_n;j++) y[n_D*i+p0_lidx_I[j]] = 0.0;
4585: MatDenseRestoreArray(pcbddc->coarse_phi_D,&y);
4586: }
4587: VecResetArray(pcbddc->vec1_R);
4588: }
4589: }
4590: if (n_constraints) {
4591: MatDestroy(&local_auxmat2_R);
4592: }
4593: PetscFree(p0_lidx_I);
4595: /* coarse matrix entries relative to B_0 */
4596: if (pcbddc->benign_n) {
4597: Mat B0_B,B0_BPHI;
4598: IS is_dummy;
4599: const PetscScalar *data;
4600: PetscInt j;
4602: ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);
4603: MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);
4604: ISDestroy(&is_dummy);
4605: MatMatMult(B0_B,pcbddc->coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);
4606: MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);
4607: MatDenseGetArrayRead(B0_BPHI,&data);
4608: for (j=0;j<pcbddc->benign_n;j++) {
4609: PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j;
4610: for (i=0;i<pcbddc->local_primal_size;i++) {
4611: coarse_submat_vals[primal_idx*pcbddc->local_primal_size+i] = data[i*pcbddc->benign_n+j];
4612: coarse_submat_vals[i*pcbddc->local_primal_size+primal_idx] = data[i*pcbddc->benign_n+j];
4613: }
4614: }
4615: MatDenseRestoreArrayRead(B0_BPHI,&data);
4616: MatDestroy(&B0_B);
4617: MatDestroy(&B0_BPHI);
4618: }
4620: /* compute other basis functions for non-symmetric problems */
4621: if (!pcbddc->symmetric_primal) {
4622: Mat B_V=NULL,B_C=NULL;
4623: PetscScalar *marray;
4625: if (n_constraints) {
4626: Mat S_CCT,C_CRT;
4628: MatTranspose(C_CR,MAT_INITIAL_MATRIX,&C_CRT);
4629: MatTranspose(S_CC,MAT_INITIAL_MATRIX,&S_CCT);
4630: MatMatMult(C_CRT,S_CCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_C);
4631: MatDestroy(&S_CCT);
4632: if (n_vertices) {
4633: Mat S_VCT;
4635: MatTranspose(S_VC,MAT_INITIAL_MATRIX,&S_VCT);
4636: MatMatMult(C_CRT,S_VCT,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B_V);
4637: MatDestroy(&S_VCT);
4638: }
4639: MatDestroy(&C_CRT);
4640: } else {
4641: MatCreateSeqDense(PETSC_COMM_SELF,n_R,n_vertices,NULL,&B_V);
4642: }
4643: if (n_vertices && n_R) {
4644: PetscScalar *av,*marray;
4645: const PetscInt *xadj,*adjncy;
4646: PetscInt n;
4647: PetscBool flg_row;
4649: /* B_V = B_V - A_VR^T */
4650: MatConvert(A_VR,MATSEQAIJ,MAT_INPLACE_MATRIX,&A_VR);
4651: MatGetRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);
4652: MatSeqAIJGetArray(A_VR,&av);
4653: MatDenseGetArray(B_V,&marray);
4654: for (i=0;i<n;i++) {
4655: PetscInt j;
4656: for (j=xadj[i];j<xadj[i+1];j++) marray[i*n_R + adjncy[j]] -= av[j];
4657: }
4658: MatDenseRestoreArray(B_V,&marray);
4659: MatRestoreRowIJ(A_VR,0,PETSC_FALSE,PETSC_FALSE,&n,&xadj,&adjncy,&flg_row);
4660: MatDestroy(&A_VR);
4661: }
4663: /* currently there's no support for MatTransposeMatSolve(F,B,X) */
4664: if (n_vertices) {
4665: MatDenseGetArray(B_V,&marray);
4666: for (i=0;i<n_vertices;i++) {
4667: VecPlaceArray(pcbddc->vec1_R,marray+i*n_R);
4668: VecPlaceArray(pcbddc->vec2_R,work+i*n_R);
4669: KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);
4670: KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);
4671: VecResetArray(pcbddc->vec1_R);
4672: VecResetArray(pcbddc->vec2_R);
4673: }
4674: MatDenseRestoreArray(B_V,&marray);
4675: }
4676: if (B_C) {
4677: MatDenseGetArray(B_C,&marray);
4678: for (i=n_vertices;i<n_constraints+n_vertices;i++) {
4679: VecPlaceArray(pcbddc->vec1_R,marray+(i-n_vertices)*n_R);
4680: VecPlaceArray(pcbddc->vec2_R,work+i*n_R);
4681: KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec2_R);
4682: KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);
4683: VecResetArray(pcbddc->vec1_R);
4684: VecResetArray(pcbddc->vec2_R);
4685: }
4686: MatDenseRestoreArray(B_C,&marray);
4687: }
4688: /* coarse basis functions */
4689: for (i=0;i<pcbddc->local_primal_size;i++) {
4690: PetscScalar *y;
4692: VecPlaceArray(pcbddc->vec1_R,work+i*n_R);
4693: MatDenseGetArray(pcbddc->coarse_psi_B,&y);
4694: VecPlaceArray(pcis->vec1_B,y+n_B*i);
4695: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4696: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
4697: if (i<n_vertices) {
4698: y[n_B*i+idx_V_B[i]] = 1.0;
4699: }
4700: MatDenseRestoreArray(pcbddc->coarse_psi_B,&y);
4701: VecResetArray(pcis->vec1_B);
4703: if (pcbddc->switch_static || pcbddc->dbg_flag) {
4704: MatDenseGetArray(pcbddc->coarse_psi_D,&y);
4705: VecPlaceArray(pcis->vec1_D,y+n_D*i);
4706: VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
4707: VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
4708: VecResetArray(pcis->vec1_D);
4709: MatDenseRestoreArray(pcbddc->coarse_psi_D,&y);
4710: }
4711: VecResetArray(pcbddc->vec1_R);
4712: }
4713: MatDestroy(&B_V);
4714: MatDestroy(&B_C);
4715: }
4717: /* free memory */
4718: PetscFree(idx_V_B);
4719: MatDestroy(&S_VV);
4720: MatDestroy(&S_CV);
4721: MatDestroy(&S_VC);
4722: MatDestroy(&S_CC);
4723: PetscFree(work);
4724: if (n_vertices) {
4725: MatDestroy(&A_VR);
4726: }
4727: if (n_constraints) {
4728: MatDestroy(&C_CR);
4729: }
4730: PetscLogEventEnd(PC_BDDC_CorrectionSetUp[pcbddc->current_level],pc,0,0,0);
4732: /* Checking coarse_sub_mat and coarse basis functios */
4733: /* Symmetric case : It should be \Phi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
4734: /* Non-symmetric case : It should be \Psi^{(j)^T} A^{(j)} \Phi^{(j)}=coarse_sub_mat */
4735: if (pcbddc->dbg_flag) {
4736: Mat coarse_sub_mat;
4737: Mat AUXMAT,TM1,TM2,TM3,TM4;
4738: Mat coarse_phi_D,coarse_phi_B;
4739: Mat coarse_psi_D,coarse_psi_B;
4740: Mat A_II,A_BB,A_IB,A_BI;
4741: Mat C_B,CPHI;
4742: IS is_dummy;
4743: Vec mones;
4744: MatType checkmattype=MATSEQAIJ;
4745: PetscReal real_value;
4747: if (pcbddc->benign_n && !pcbddc->benign_change_explicit) {
4748: Mat A;
4749: PCBDDCBenignProject(pc,NULL,NULL,&A);
4750: MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_II);
4751: MatCreateSubMatrix(A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_IB);
4752: MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&A_BI);
4753: MatCreateSubMatrix(A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&A_BB);
4754: MatDestroy(&A);
4755: } else {
4756: MatConvert(pcis->A_II,checkmattype,MAT_INITIAL_MATRIX,&A_II);
4757: MatConvert(pcis->A_IB,checkmattype,MAT_INITIAL_MATRIX,&A_IB);
4758: MatConvert(pcis->A_BI,checkmattype,MAT_INITIAL_MATRIX,&A_BI);
4759: MatConvert(pcis->A_BB,checkmattype,MAT_INITIAL_MATRIX,&A_BB);
4760: }
4761: MatConvert(pcbddc->coarse_phi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_D);
4762: MatConvert(pcbddc->coarse_phi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_phi_B);
4763: if (!pcbddc->symmetric_primal) {
4764: MatConvert(pcbddc->coarse_psi_D,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_D);
4765: MatConvert(pcbddc->coarse_psi_B,checkmattype,MAT_INITIAL_MATRIX,&coarse_psi_B);
4766: }
4767: MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_sub_mat);
4769: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");
4770: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse sub mat computation (symmetric %d)\n",pcbddc->symmetric_primal);
4771: PetscViewerFlush(pcbddc->dbg_viewer);
4772: if (!pcbddc->symmetric_primal) {
4773: MatMatMult(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);
4774: MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM1);
4775: MatDestroy(&AUXMAT);
4776: MatMatMult(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);
4777: MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM2);
4778: MatDestroy(&AUXMAT);
4779: MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);
4780: MatTransposeMatMult(coarse_psi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);
4781: MatDestroy(&AUXMAT);
4782: MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);
4783: MatTransposeMatMult(coarse_psi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);
4784: MatDestroy(&AUXMAT);
4785: } else {
4786: MatPtAP(A_II,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&TM1);
4787: MatPtAP(A_BB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&TM2);
4788: MatMatMult(A_IB,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&AUXMAT);
4789: MatTransposeMatMult(coarse_phi_D,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM3);
4790: MatDestroy(&AUXMAT);
4791: MatMatMult(A_BI,coarse_phi_D,MAT_INITIAL_MATRIX,1.0,&AUXMAT);
4792: MatTransposeMatMult(coarse_phi_B,AUXMAT,MAT_INITIAL_MATRIX,1.0,&TM4);
4793: MatDestroy(&AUXMAT);
4794: }
4795: MatAXPY(TM1,one,TM2,DIFFERENT_NONZERO_PATTERN);
4796: MatAXPY(TM1,one,TM3,DIFFERENT_NONZERO_PATTERN);
4797: MatAXPY(TM1,one,TM4,DIFFERENT_NONZERO_PATTERN);
4798: MatConvert(TM1,MATSEQDENSE,MAT_INPLACE_MATRIX,&TM1);
4799: if (pcbddc->benign_n) {
4800: Mat B0_B,B0_BPHI;
4801: const PetscScalar *data2;
4802: PetscScalar *data;
4803: PetscInt j;
4805: ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&is_dummy);
4806: MatCreateSubMatrix(pcbddc->benign_B0,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);
4807: MatMatMult(B0_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&B0_BPHI);
4808: MatConvert(B0_BPHI,MATSEQDENSE,MAT_INPLACE_MATRIX,&B0_BPHI);
4809: MatDenseGetArray(TM1,&data);
4810: MatDenseGetArrayRead(B0_BPHI,&data2);
4811: for (j=0;j<pcbddc->benign_n;j++) {
4812: PetscInt primal_idx = pcbddc->local_primal_size - pcbddc->benign_n + j;
4813: for (i=0;i<pcbddc->local_primal_size;i++) {
4814: data[primal_idx*pcbddc->local_primal_size+i] += data2[i*pcbddc->benign_n+j];
4815: data[i*pcbddc->local_primal_size+primal_idx] += data2[i*pcbddc->benign_n+j];
4816: }
4817: }
4818: MatDenseRestoreArray(TM1,&data);
4819: MatDenseRestoreArrayRead(B0_BPHI,&data2);
4820: MatDestroy(&B0_B);
4821: ISDestroy(&is_dummy);
4822: MatDestroy(&B0_BPHI);
4823: }
4824: #if 0
4825: {
4826: PetscViewer viewer;
4827: char filename[256];
4828: sprintf(filename,"details_local_coarse_mat%d_level%d.m",PetscGlobalRank,pcbddc->current_level);
4829: PetscViewerASCIIOpen(PETSC_COMM_SELF,filename,&viewer);
4830: PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);
4831: PetscObjectSetName((PetscObject)coarse_sub_mat,"computed");
4832: MatView(coarse_sub_mat,viewer);
4833: PetscObjectSetName((PetscObject)TM1,"projected");
4834: MatView(TM1,viewer);
4835: if (pcbddc->coarse_phi_B) {
4836: PetscObjectSetName((PetscObject)pcbddc->coarse_phi_B,"phi_B");
4837: MatView(pcbddc->coarse_phi_B,viewer);
4838: }
4839: if (pcbddc->coarse_phi_D) {
4840: PetscObjectSetName((PetscObject)pcbddc->coarse_phi_D,"phi_D");
4841: MatView(pcbddc->coarse_phi_D,viewer);
4842: }
4843: if (pcbddc->coarse_psi_B) {
4844: PetscObjectSetName((PetscObject)pcbddc->coarse_psi_B,"psi_B");
4845: MatView(pcbddc->coarse_psi_B,viewer);
4846: }
4847: if (pcbddc->coarse_psi_D) {
4848: PetscObjectSetName((PetscObject)pcbddc->coarse_psi_D,"psi_D");
4849: MatView(pcbddc->coarse_psi_D,viewer);
4850: }
4851: PetscObjectSetName((PetscObject)pcbddc->local_mat,"A");
4852: MatView(pcbddc->local_mat,viewer);
4853: PetscObjectSetName((PetscObject)pcbddc->ConstraintMatrix,"C");
4854: MatView(pcbddc->ConstraintMatrix,viewer);
4855: PetscObjectSetName((PetscObject)pcis->is_I_local,"I");
4856: ISView(pcis->is_I_local,viewer);
4857: PetscObjectSetName((PetscObject)pcis->is_B_local,"B");
4858: ISView(pcis->is_B_local,viewer);
4859: PetscObjectSetName((PetscObject)pcbddc->is_R_local,"R");
4860: ISView(pcbddc->is_R_local,viewer);
4861: PetscViewerDestroy(&viewer);
4862: }
4863: #endif
4864: MatAXPY(TM1,m_one,coarse_sub_mat,DIFFERENT_NONZERO_PATTERN);
4865: MatNorm(TM1,NORM_FROBENIUS,&real_value);
4866: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
4867: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d matrix error % 1.14e\n",PetscGlobalRank,real_value);
4869: /* check constraints */
4870: ISCreateStride(PETSC_COMM_SELF,pcbddc->local_primal_size-pcbddc->benign_n,0,1,&is_dummy);
4871: MatCreateSubMatrix(pcbddc->ConstraintMatrix,is_dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&C_B);
4872: if (!pcbddc->benign_n) { /* TODO: add benign case */
4873: MatMatMult(C_B,coarse_phi_B,MAT_INITIAL_MATRIX,1.0,&CPHI);
4874: } else {
4875: PetscScalar *data;
4876: Mat tmat;
4877: MatDenseGetArray(pcbddc->coarse_phi_B,&data);
4878: MatCreateSeqDense(PETSC_COMM_SELF,pcis->n_B,pcbddc->local_primal_size-pcbddc->benign_n,data,&tmat);
4879: MatDenseRestoreArray(pcbddc->coarse_phi_B,&data);
4880: MatMatMult(C_B,tmat,MAT_INITIAL_MATRIX,1.0,&CPHI);
4881: MatDestroy(&tmat);
4882: }
4883: MatCreateVecs(CPHI,&mones,NULL);
4884: VecSet(mones,-1.0);
4885: MatDiagonalSet(CPHI,mones,ADD_VALUES);
4886: MatNorm(CPHI,NORM_FROBENIUS,&real_value);
4887: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d phi constraints error % 1.14e\n",PetscGlobalRank,real_value);
4888: if (!pcbddc->symmetric_primal) {
4889: MatMatMult(C_B,coarse_psi_B,MAT_REUSE_MATRIX,1.0,&CPHI);
4890: VecSet(mones,-1.0);
4891: MatDiagonalSet(CPHI,mones,ADD_VALUES);
4892: MatNorm(CPHI,NORM_FROBENIUS,&real_value);
4893: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d psi constraints error % 1.14e\n",PetscGlobalRank,real_value);
4894: }
4895: MatDestroy(&C_B);
4896: MatDestroy(&CPHI);
4897: ISDestroy(&is_dummy);
4898: VecDestroy(&mones);
4899: PetscViewerFlush(pcbddc->dbg_viewer);
4900: MatDestroy(&A_II);
4901: MatDestroy(&A_BB);
4902: MatDestroy(&A_IB);
4903: MatDestroy(&A_BI);
4904: MatDestroy(&TM1);
4905: MatDestroy(&TM2);
4906: MatDestroy(&TM3);
4907: MatDestroy(&TM4);
4908: MatDestroy(&coarse_phi_D);
4909: MatDestroy(&coarse_phi_B);
4910: if (!pcbddc->symmetric_primal) {
4911: MatDestroy(&coarse_psi_D);
4912: MatDestroy(&coarse_psi_B);
4913: }
4914: MatDestroy(&coarse_sub_mat);
4915: }
4916: /* FINAL CUDA support (we cannot currently mix viennacl and cuda vectors */
4917: {
4918: PetscBool gpu;
4920: PetscObjectTypeCompare((PetscObject)pcis->vec1_N,VECSEQCUDA,&gpu);
4921: if (gpu) {
4922: if (pcbddc->local_auxmat1) {
4923: MatConvert(pcbddc->local_auxmat1,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->local_auxmat1);
4924: }
4925: if (pcbddc->local_auxmat2) {
4926: MatConvert(pcbddc->local_auxmat2,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->local_auxmat2);
4927: }
4928: if (pcbddc->coarse_phi_B) {
4929: MatConvert(pcbddc->coarse_phi_B,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_phi_B);
4930: }
4931: if (pcbddc->coarse_phi_D) {
4932: MatConvert(pcbddc->coarse_phi_D,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_phi_D);
4933: }
4934: if (pcbddc->coarse_psi_B) {
4935: MatConvert(pcbddc->coarse_psi_B,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_psi_B);
4936: }
4937: if (pcbddc->coarse_psi_D) {
4938: MatConvert(pcbddc->coarse_psi_D,MATSEQDENSECUDA,MAT_INPLACE_MATRIX,&pcbddc->coarse_psi_D);
4939: }
4940: }
4941: }
4942: /* get back data */
4943: *coarse_submat_vals_n = coarse_submat_vals;
4944: return(0);
4945: }
4947: PetscErrorCode MatCreateSubMatrixUnsorted(Mat A, IS isrow, IS iscol, Mat* B)
4948: {
4949: Mat *work_mat;
4950: IS isrow_s,iscol_s;
4951: PetscBool rsorted,csorted;
4952: PetscInt rsize,*idxs_perm_r=NULL,csize,*idxs_perm_c=NULL;
4956: ISSorted(isrow,&rsorted);
4957: ISSorted(iscol,&csorted);
4958: ISGetLocalSize(isrow,&rsize);
4959: ISGetLocalSize(iscol,&csize);
4961: if (!rsorted) {
4962: const PetscInt *idxs;
4963: PetscInt *idxs_sorted,i;
4965: PetscMalloc1(rsize,&idxs_perm_r);
4966: PetscMalloc1(rsize,&idxs_sorted);
4967: for (i=0;i<rsize;i++) {
4968: idxs_perm_r[i] = i;
4969: }
4970: ISGetIndices(isrow,&idxs);
4971: PetscSortIntWithPermutation(rsize,idxs,idxs_perm_r);
4972: for (i=0;i<rsize;i++) {
4973: idxs_sorted[i] = idxs[idxs_perm_r[i]];
4974: }
4975: ISRestoreIndices(isrow,&idxs);
4976: ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_sorted,PETSC_OWN_POINTER,&isrow_s);
4977: } else {
4978: PetscObjectReference((PetscObject)isrow);
4979: isrow_s = isrow;
4980: }
4982: if (!csorted) {
4983: if (isrow == iscol) {
4984: PetscObjectReference((PetscObject)isrow_s);
4985: iscol_s = isrow_s;
4986: } else {
4987: const PetscInt *idxs;
4988: PetscInt *idxs_sorted,i;
4990: PetscMalloc1(csize,&idxs_perm_c);
4991: PetscMalloc1(csize,&idxs_sorted);
4992: for (i=0;i<csize;i++) {
4993: idxs_perm_c[i] = i;
4994: }
4995: ISGetIndices(iscol,&idxs);
4996: PetscSortIntWithPermutation(csize,idxs,idxs_perm_c);
4997: for (i=0;i<csize;i++) {
4998: idxs_sorted[i] = idxs[idxs_perm_c[i]];
4999: }
5000: ISRestoreIndices(iscol,&idxs);
5001: ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_sorted,PETSC_OWN_POINTER,&iscol_s);
5002: }
5003: } else {
5004: PetscObjectReference((PetscObject)iscol);
5005: iscol_s = iscol;
5006: }
5008: MatCreateSubMatrices(A,1,&isrow_s,&iscol_s,MAT_INITIAL_MATRIX,&work_mat);
5010: if (!rsorted || !csorted) {
5011: Mat new_mat;
5012: IS is_perm_r,is_perm_c;
5014: if (!rsorted) {
5015: PetscInt *idxs_r,i;
5016: PetscMalloc1(rsize,&idxs_r);
5017: for (i=0;i<rsize;i++) {
5018: idxs_r[idxs_perm_r[i]] = i;
5019: }
5020: PetscFree(idxs_perm_r);
5021: ISCreateGeneral(PETSC_COMM_SELF,rsize,idxs_r,PETSC_OWN_POINTER,&is_perm_r);
5022: } else {
5023: ISCreateStride(PETSC_COMM_SELF,rsize,0,1,&is_perm_r);
5024: }
5025: ISSetPermutation(is_perm_r);
5027: if (!csorted) {
5028: if (isrow_s == iscol_s) {
5029: PetscObjectReference((PetscObject)is_perm_r);
5030: is_perm_c = is_perm_r;
5031: } else {
5032: PetscInt *idxs_c,i;
5033: if (!idxs_perm_c) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Permutation array not present");
5034: PetscMalloc1(csize,&idxs_c);
5035: for (i=0;i<csize;i++) {
5036: idxs_c[idxs_perm_c[i]] = i;
5037: }
5038: PetscFree(idxs_perm_c);
5039: ISCreateGeneral(PETSC_COMM_SELF,csize,idxs_c,PETSC_OWN_POINTER,&is_perm_c);
5040: }
5041: } else {
5042: ISCreateStride(PETSC_COMM_SELF,csize,0,1,&is_perm_c);
5043: }
5044: ISSetPermutation(is_perm_c);
5046: MatPermute(work_mat[0],is_perm_r,is_perm_c,&new_mat);
5047: MatDestroy(&work_mat[0]);
5048: work_mat[0] = new_mat;
5049: ISDestroy(&is_perm_r);
5050: ISDestroy(&is_perm_c);
5051: }
5053: PetscObjectReference((PetscObject)work_mat[0]);
5054: *B = work_mat[0];
5055: MatDestroyMatrices(1,&work_mat);
5056: ISDestroy(&isrow_s);
5057: ISDestroy(&iscol_s);
5058: return(0);
5059: }
5061: PetscErrorCode PCBDDCComputeLocalMatrix(PC pc, Mat ChangeOfBasisMatrix)
5062: {
5063: Mat_IS* matis = (Mat_IS*)pc->pmat->data;
5064: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
5065: Mat new_mat,lA;
5066: IS is_local,is_global;
5067: PetscInt local_size;
5068: PetscBool isseqaij;
5072: MatDestroy(&pcbddc->local_mat);
5073: MatGetSize(matis->A,&local_size,NULL);
5074: ISCreateStride(PetscObjectComm((PetscObject)matis->A),local_size,0,1,&is_local);
5075: ISLocalToGlobalMappingApplyIS(pc->pmat->rmap->mapping,is_local,&is_global);
5076: ISDestroy(&is_local);
5077: MatCreateSubMatrixUnsorted(ChangeOfBasisMatrix,is_global,is_global,&new_mat);
5078: ISDestroy(&is_global);
5080: if (pcbddc->dbg_flag) {
5081: Vec x,x_change;
5082: PetscReal error;
5084: MatCreateVecs(ChangeOfBasisMatrix,&x,&x_change);
5085: VecSetRandom(x,NULL);
5086: MatMult(ChangeOfBasisMatrix,x,x_change);
5087: VecScatterBegin(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);
5088: VecScatterEnd(matis->cctx,x,matis->x,INSERT_VALUES,SCATTER_FORWARD);
5089: MatMult(new_mat,matis->x,matis->y);
5090: if (!pcbddc->change_interior) {
5091: const PetscScalar *x,*y,*v;
5092: PetscReal lerror = 0.;
5093: PetscInt i;
5095: VecGetArrayRead(matis->x,&x);
5096: VecGetArrayRead(matis->y,&y);
5097: VecGetArrayRead(matis->counter,&v);
5098: for (i=0;i<local_size;i++)
5099: if (PetscRealPart(v[i]) < 1.5 && PetscAbsScalar(x[i]-y[i]) > lerror)
5100: lerror = PetscAbsScalar(x[i]-y[i]);
5101: VecRestoreArrayRead(matis->x,&x);
5102: VecRestoreArrayRead(matis->y,&y);
5103: VecRestoreArrayRead(matis->counter,&v);
5104: MPIU_Allreduce(&lerror,&error,1,MPIU_REAL,MPI_MAX,PetscObjectComm((PetscObject)pc));
5105: if (error > PETSC_SMALL) {
5106: if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) {
5107: SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on I: %1.6e",error);
5108: } else {
5109: SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on I: %1.6e",error);
5110: }
5111: }
5112: }
5113: VecScatterBegin(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);
5114: VecScatterEnd(matis->rctx,matis->y,x,INSERT_VALUES,SCATTER_REVERSE);
5115: VecAXPY(x,-1.0,x_change);
5116: VecNorm(x,NORM_INFINITY,&error);
5117: if (error > PETSC_SMALL) {
5118: if (!pcbddc->user_ChangeOfBasisMatrix || pcbddc->current_level) {
5119: SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e",error);
5120: } else {
5121: SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_USER,"Error global vs local change on N: %1.6e",error);
5122: }
5123: }
5124: VecDestroy(&x);
5125: VecDestroy(&x_change);
5126: }
5128: /* lA is present if we are setting up an inner BDDC for a saddle point FETI-DP */
5129: PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject*)&lA);
5131: /* TODO: HOW TO WORK WITH BAIJ and SBAIJ and SEQDENSE? */
5132: PetscObjectBaseTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);
5133: if (isseqaij) {
5134: MatDestroy(&pcbddc->local_mat);
5135: MatPtAP(matis->A,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);
5136: if (lA) {
5137: Mat work;
5138: MatPtAP(lA,new_mat,MAT_INITIAL_MATRIX,2.0,&work);
5139: PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);
5140: MatDestroy(&work);
5141: }
5142: } else {
5143: Mat work_mat;
5145: MatDestroy(&pcbddc->local_mat);
5146: MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);
5147: MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&pcbddc->local_mat);
5148: MatDestroy(&work_mat);
5149: if (lA) {
5150: Mat work;
5151: MatConvert(lA,MATSEQAIJ,MAT_INITIAL_MATRIX,&work_mat);
5152: MatPtAP(work_mat,new_mat,MAT_INITIAL_MATRIX,2.0,&work);
5153: PetscObjectCompose((PetscObject)pc,"__KSPFETIDP_lA" ,(PetscObject)work);
5154: MatDestroy(&work);
5155: }
5156: }
5157: if (matis->A->symmetric_set) {
5158: MatSetOption(pcbddc->local_mat,MAT_SYMMETRIC,matis->A->symmetric);
5159: #if !defined(PETSC_USE_COMPLEX)
5160: MatSetOption(pcbddc->local_mat,MAT_HERMITIAN,matis->A->symmetric);
5161: #endif
5162: }
5163: MatDestroy(&new_mat);
5164: return(0);
5165: }
5167: PetscErrorCode PCBDDCSetUpLocalScatters(PC pc)
5168: {
5169: PC_IS* pcis = (PC_IS*)(pc->data);
5170: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
5171: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
5172: PetscInt *idx_R_local=NULL;
5173: PetscInt n_vertices,i,j,n_R,n_D,n_B;
5174: PetscInt vbs,bs;
5175: PetscBT bitmask=NULL;
5176: PetscErrorCode ierr;
5179: /*
5180: No need to setup local scatters if
5181: - primal space is unchanged
5182: AND
5183: - we actually have locally some primal dofs (could not be true in multilevel or for isolated subdomains)
5184: AND
5185: - we are not in debugging mode (this is needed since there are Synchronized prints at the end of the subroutine
5186: */
5187: if (!pcbddc->new_primal_space_local && pcbddc->local_primal_size && !pcbddc->dbg_flag) {
5188: return(0);
5189: }
5190: /* destroy old objects */
5191: ISDestroy(&pcbddc->is_R_local);
5192: VecScatterDestroy(&pcbddc->R_to_B);
5193: VecScatterDestroy(&pcbddc->R_to_D);
5194: /* Set Non-overlapping dimensions */
5195: n_B = pcis->n_B;
5196: n_D = pcis->n - n_B;
5197: n_vertices = pcbddc->n_vertices;
5199: /* Dohrmann's notation: dofs splitted in R (Remaining: all dofs but the vertices) and V (Vertices) */
5201: /* create auxiliary bitmask and allocate workspace */
5202: if (!sub_schurs || !sub_schurs->reuse_solver) {
5203: PetscMalloc1(pcis->n-n_vertices,&idx_R_local);
5204: PetscBTCreate(pcis->n,&bitmask);
5205: for (i=0;i<n_vertices;i++) {
5206: PetscBTSet(bitmask,pcbddc->local_primal_ref_node[i]);
5207: }
5209: for (i=0, n_R=0; i<pcis->n; i++) {
5210: if (!PetscBTLookup(bitmask,i)) {
5211: idx_R_local[n_R++] = i;
5212: }
5213: }
5214: } else { /* A different ordering (already computed) is present if we are reusing the Schur solver */
5215: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5217: ISGetIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);
5218: ISGetLocalSize(reuse_solver->is_R,&n_R);
5219: }
5221: /* Block code */
5222: vbs = 1;
5223: MatGetBlockSize(pcbddc->local_mat,&bs);
5224: if (bs>1 && !(n_vertices%bs)) {
5225: PetscBool is_blocked = PETSC_TRUE;
5226: PetscInt *vary;
5227: if (!sub_schurs || !sub_schurs->reuse_solver) {
5228: PetscMalloc1(pcis->n/bs,&vary);
5229: PetscArrayzero(vary,pcis->n/bs);
5230: /* Verify that the vertex indices correspond to each element in a block (code taken from sbaij2.c) */
5231: /* it is ok to check this way since local_primal_ref_node are always sorted by local numbering and idx_R_local is obtained as a complement */
5232: for (i=0; i<n_vertices; i++) vary[pcbddc->local_primal_ref_node[i]/bs]++;
5233: for (i=0; i<pcis->n/bs; i++) {
5234: if (vary[i]!=0 && vary[i]!=bs) {
5235: is_blocked = PETSC_FALSE;
5236: break;
5237: }
5238: }
5239: PetscFree(vary);
5240: } else {
5241: /* Verify directly the R set */
5242: for (i=0; i<n_R/bs; i++) {
5243: PetscInt j,node=idx_R_local[bs*i];
5244: for (j=1; j<bs; j++) {
5245: if (node != idx_R_local[bs*i+j]-j) {
5246: is_blocked = PETSC_FALSE;
5247: break;
5248: }
5249: }
5250: }
5251: }
5252: if (is_blocked) { /* build compressed IS for R nodes (complement of vertices) */
5253: vbs = bs;
5254: for (i=0;i<n_R/vbs;i++) {
5255: idx_R_local[i] = idx_R_local[vbs*i]/vbs;
5256: }
5257: }
5258: }
5259: ISCreateBlock(PETSC_COMM_SELF,vbs,n_R/vbs,idx_R_local,PETSC_COPY_VALUES,&pcbddc->is_R_local);
5260: if (sub_schurs && sub_schurs->reuse_solver) {
5261: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5263: ISRestoreIndices(reuse_solver->is_R,(const PetscInt**)&idx_R_local);
5264: ISDestroy(&reuse_solver->is_R);
5265: PetscObjectReference((PetscObject)pcbddc->is_R_local);
5266: reuse_solver->is_R = pcbddc->is_R_local;
5267: } else {
5268: PetscFree(idx_R_local);
5269: }
5271: /* print some info if requested */
5272: if (pcbddc->dbg_flag) {
5273: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");
5274: PetscViewerFlush(pcbddc->dbg_viewer);
5275: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
5276: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d local dimensions\n",PetscGlobalRank);
5277: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_size = %D, dirichlet_size = %D, boundary_size = %D\n",pcis->n,n_D,n_B);
5278: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"r_size = %D, v_size = %D, constraints = %D, local_primal_size = %D\n",n_R,n_vertices,pcbddc->local_primal_size-n_vertices-pcbddc->benign_n,pcbddc->local_primal_size);
5279: PetscViewerFlush(pcbddc->dbg_viewer);
5280: }
5282: /* VecScatters pcbddc->R_to_B and (optionally) pcbddc->R_to_D */
5283: if (!sub_schurs || !sub_schurs->reuse_solver) {
5284: IS is_aux1,is_aux2;
5285: PetscInt *aux_array1,*aux_array2,*is_indices,*idx_R_local;
5287: ISGetIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);
5288: PetscMalloc1(pcis->n_B-n_vertices,&aux_array1);
5289: PetscMalloc1(pcis->n_B-n_vertices,&aux_array2);
5290: ISGetIndices(pcis->is_I_local,(const PetscInt**)&is_indices);
5291: for (i=0; i<n_D; i++) {
5292: PetscBTSet(bitmask,is_indices[i]);
5293: }
5294: ISRestoreIndices(pcis->is_I_local,(const PetscInt**)&is_indices);
5295: for (i=0, j=0; i<n_R; i++) {
5296: if (!PetscBTLookup(bitmask,idx_R_local[i])) {
5297: aux_array1[j++] = i;
5298: }
5299: }
5300: ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);
5301: ISGetIndices(pcis->is_B_local,(const PetscInt**)&is_indices);
5302: for (i=0, j=0; i<n_B; i++) {
5303: if (!PetscBTLookup(bitmask,is_indices[i])) {
5304: aux_array2[j++] = i;
5305: }
5306: }
5307: ISRestoreIndices(pcis->is_B_local,(const PetscInt**)&is_indices);
5308: ISCreateGeneral(PETSC_COMM_SELF,j,aux_array2,PETSC_OWN_POINTER,&is_aux2);
5309: VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_B,is_aux2,&pcbddc->R_to_B);
5310: ISDestroy(&is_aux1);
5311: ISDestroy(&is_aux2);
5313: if (pcbddc->switch_static || pcbddc->dbg_flag) {
5314: PetscMalloc1(n_D,&aux_array1);
5315: for (i=0, j=0; i<n_R; i++) {
5316: if (PetscBTLookup(bitmask,idx_R_local[i])) {
5317: aux_array1[j++] = i;
5318: }
5319: }
5320: ISCreateGeneral(PETSC_COMM_SELF,j,aux_array1,PETSC_OWN_POINTER,&is_aux1);
5321: VecScatterCreate(pcbddc->vec1_R,is_aux1,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);
5322: ISDestroy(&is_aux1);
5323: }
5324: PetscBTDestroy(&bitmask);
5325: ISRestoreIndices(pcbddc->is_R_local,(const PetscInt**)&idx_R_local);
5326: } else {
5327: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5328: IS tis;
5329: PetscInt schur_size;
5331: ISGetLocalSize(reuse_solver->is_B,&schur_size);
5332: ISCreateStride(PETSC_COMM_SELF,schur_size,n_D,1,&tis);
5333: VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_B,reuse_solver->is_B,&pcbddc->R_to_B);
5334: ISDestroy(&tis);
5335: if (pcbddc->switch_static || pcbddc->dbg_flag) {
5336: ISCreateStride(PETSC_COMM_SELF,n_D,0,1,&tis);
5337: VecScatterCreate(pcbddc->vec1_R,tis,pcis->vec1_D,(IS)0,&pcbddc->R_to_D);
5338: ISDestroy(&tis);
5339: }
5340: }
5341: return(0);
5342: }
5344: static PetscErrorCode MatNullSpacePropagateAny_Private(Mat A, IS is, Mat B)
5345: {
5346: MatNullSpace NullSpace;
5347: Mat dmat;
5348: const Vec *nullvecs;
5349: Vec v,v2,*nullvecs2;
5350: VecScatter sct = NULL;
5351: PetscContainer c;
5352: PetscScalar *ddata;
5353: PetscInt k,nnsp_size,bsiz,bsiz2,n,N,bs;
5354: PetscBool nnsp_has_cnst;
5358: if (!is && !B) { /* MATIS */
5359: Mat_IS* matis = (Mat_IS*)A->data;
5361: if (!B) {
5362: MatISGetLocalMat(A,&B);
5363: }
5364: sct = matis->cctx;
5365: PetscObjectReference((PetscObject)sct);
5366: } else {
5367: MatGetNullSpace(B,&NullSpace);
5368: if (!NullSpace) {
5369: MatGetNearNullSpace(B,&NullSpace);
5370: }
5371: if (NullSpace) return(0);
5372: }
5373: MatGetNullSpace(A,&NullSpace);
5374: if (!NullSpace) {
5375: MatGetNearNullSpace(A,&NullSpace);
5376: }
5377: if (!NullSpace) return(0);
5379: MatCreateVecs(A,&v,NULL);
5380: MatCreateVecs(B,&v2,NULL);
5381: if (!sct) {
5382: VecScatterCreate(v,is,v2,NULL,&sct);
5383: }
5384: MatNullSpaceGetVecs(NullSpace,&nnsp_has_cnst,&nnsp_size,(const Vec**)&nullvecs);
5385: bsiz = bsiz2 = nnsp_size+!!nnsp_has_cnst;
5386: PetscMalloc1(bsiz,&nullvecs2);
5387: VecGetBlockSize(v2,&bs);
5388: VecGetSize(v2,&N);
5389: VecGetLocalSize(v2,&n);
5390: PetscMalloc1(n*bsiz,&ddata);
5391: for (k=0;k<nnsp_size;k++) {
5392: VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),bs,n,N,ddata + n*k,&nullvecs2[k]);
5393: VecScatterBegin(sct,nullvecs[k],nullvecs2[k],INSERT_VALUES,SCATTER_FORWARD);
5394: VecScatterEnd(sct,nullvecs[k],nullvecs2[k],INSERT_VALUES,SCATTER_FORWARD);
5395: }
5396: if (nnsp_has_cnst) {
5397: VecCreateMPIWithArray(PetscObjectComm((PetscObject)B),bs,n,N,ddata + n*nnsp_size,&nullvecs2[nnsp_size]);
5398: VecSet(nullvecs2[nnsp_size],1.0);
5399: }
5400: PCBDDCOrthonormalizeVecs(&bsiz2,nullvecs2);
5401: MatNullSpaceCreate(PetscObjectComm((PetscObject)B),PETSC_FALSE,bsiz2,nullvecs2,&NullSpace);
5403: MatCreateDense(PetscObjectComm((PetscObject)B),n,PETSC_DECIDE,N,bsiz2,ddata,&dmat);
5404: PetscContainerCreate(PetscObjectComm((PetscObject)B),&c);
5405: PetscContainerSetPointer(c,ddata);
5406: PetscContainerSetUserDestroy(c,PetscContainerUserDestroyDefault);
5407: PetscObjectCompose((PetscObject)dmat,"_PBDDC_Null_dmat_arr",(PetscObject)c);
5408: PetscContainerDestroy(&c);
5409: PetscObjectCompose((PetscObject)NullSpace,"_PBDDC_Null_dmat",(PetscObject)dmat);
5410: MatDestroy(&dmat);
5412: for (k=0;k<bsiz;k++) {
5413: VecDestroy(&nullvecs2[k]);
5414: }
5415: PetscFree(nullvecs2);
5416: MatSetNearNullSpace(B,NullSpace);
5417: MatNullSpaceDestroy(&NullSpace);
5418: VecDestroy(&v);
5419: VecDestroy(&v2);
5420: VecScatterDestroy(&sct);
5421: return(0);
5422: }
5424: PetscErrorCode PCBDDCSetUpLocalSolvers(PC pc, PetscBool dirichlet, PetscBool neumann)
5425: {
5426: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
5427: PC_IS *pcis = (PC_IS*)pc->data;
5428: PC pc_temp;
5429: Mat A_RR;
5430: MatNullSpace nnsp;
5431: MatReuse reuse;
5432: PetscScalar m_one = -1.0;
5433: PetscReal value;
5434: PetscInt n_D,n_R;
5435: PetscBool issbaij,opts;
5437: void (*f)(void) = 0;
5438: char dir_prefix[256],neu_prefix[256],str_level[16];
5439: size_t len;
5442: PetscLogEventBegin(PC_BDDC_LocalSolvers[pcbddc->current_level],pc,0,0,0);
5443: /* approximate solver, propagate NearNullSpace if needed */
5444: if (!pc->setupcalled && (pcbddc->NullSpace_corr[0] || pcbddc->NullSpace_corr[2])) {
5445: MatNullSpace gnnsp1,gnnsp2;
5446: PetscBool lhas,ghas;
5448: MatGetNearNullSpace(pcbddc->local_mat,&nnsp);
5449: MatGetNearNullSpace(pc->pmat,&gnnsp1);
5450: MatGetNullSpace(pc->pmat,&gnnsp2);
5451: lhas = nnsp ? PETSC_TRUE : PETSC_FALSE;
5452: MPIU_Allreduce(&lhas,&ghas,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));
5453: if (!ghas && (gnnsp1 || gnnsp2)) {
5454: MatNullSpacePropagateAny_Private(pc->pmat,NULL,NULL);
5455: }
5456: }
5458: /* compute prefixes */
5459: PetscStrcpy(dir_prefix,"");
5460: PetscStrcpy(neu_prefix,"");
5461: if (!pcbddc->current_level) {
5462: PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,sizeof(dir_prefix));
5463: PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,sizeof(neu_prefix));
5464: PetscStrlcat(dir_prefix,"pc_bddc_dirichlet_",sizeof(dir_prefix));
5465: PetscStrlcat(neu_prefix,"pc_bddc_neumann_",sizeof(neu_prefix));
5466: } else {
5467: PetscSNPrintf(str_level,sizeof(str_level),"l%d_",(int)(pcbddc->current_level));
5468: PetscStrlen(((PetscObject)pc)->prefix,&len);
5469: len -= 15; /* remove "pc_bddc_coarse_" */
5470: if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */
5471: if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */
5472: /* Nonstandard use of PetscStrncpy() to only copy a portion of the input string */
5473: PetscStrncpy(dir_prefix,((PetscObject)pc)->prefix,len+1);
5474: PetscStrncpy(neu_prefix,((PetscObject)pc)->prefix,len+1);
5475: PetscStrlcat(dir_prefix,"pc_bddc_dirichlet_",sizeof(dir_prefix));
5476: PetscStrlcat(neu_prefix,"pc_bddc_neumann_",sizeof(neu_prefix));
5477: PetscStrlcat(dir_prefix,str_level,sizeof(dir_prefix));
5478: PetscStrlcat(neu_prefix,str_level,sizeof(neu_prefix));
5479: }
5481: /* DIRICHLET PROBLEM */
5482: if (dirichlet) {
5483: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
5484: if (pcbddc->benign_n && !pcbddc->benign_change_explicit) {
5485: if (!sub_schurs || !sub_schurs->reuse_solver) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet implemented");
5486: if (pcbddc->dbg_flag) {
5487: Mat A_IIn;
5489: PCBDDCBenignProject(pc,pcis->is_I_local,pcis->is_I_local,&A_IIn);
5490: MatDestroy(&pcis->A_II);
5491: pcis->A_II = A_IIn;
5492: }
5493: }
5494: if (pcbddc->local_mat->symmetric_set) {
5495: MatSetOption(pcis->A_II,MAT_SYMMETRIC,pcbddc->local_mat->symmetric);
5496: }
5497: /* Matrix for Dirichlet problem is pcis->A_II */
5498: n_D = pcis->n - pcis->n_B;
5499: opts = PETSC_FALSE;
5500: if (!pcbddc->ksp_D) { /* create object if not yet build */
5501: opts = PETSC_TRUE;
5502: KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_D);
5503: PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_D,(PetscObject)pc,1);
5504: /* default */
5505: KSPSetType(pcbddc->ksp_D,KSPPREONLY);
5506: KSPSetOptionsPrefix(pcbddc->ksp_D,dir_prefix);
5507: PetscObjectTypeCompare((PetscObject)pcis->pA_II,MATSEQSBAIJ,&issbaij);
5508: KSPGetPC(pcbddc->ksp_D,&pc_temp);
5509: if (issbaij) {
5510: PCSetType(pc_temp,PCCHOLESKY);
5511: } else {
5512: PCSetType(pc_temp,PCLU);
5513: }
5514: KSPSetErrorIfNotConverged(pcbddc->ksp_D,pc->erroriffailure);
5515: }
5516: MatSetOptionsPrefix(pcis->pA_II,((PetscObject)pcbddc->ksp_D)->prefix);
5517: KSPSetOperators(pcbddc->ksp_D,pcis->A_II,pcis->pA_II);
5518: /* Allow user's customization */
5519: if (opts) {
5520: KSPSetFromOptions(pcbddc->ksp_D);
5521: }
5522: MatGetNearNullSpace(pcis->pA_II,&nnsp);
5523: if (pcbddc->NullSpace_corr[0] && !nnsp) { /* approximate solver, propagate NearNullSpace */
5524: MatNullSpacePropagateAny_Private(pcbddc->local_mat,pcis->is_I_local,pcis->pA_II);
5525: }
5526: MatGetNearNullSpace(pcis->pA_II,&nnsp);
5527: KSPGetPC(pcbddc->ksp_D,&pc_temp);
5528: PetscObjectQueryFunction((PetscObject)pc_temp,"PCSetCoordinates_C",&f);
5529: if (f && pcbddc->mat_graph->cloc && !nnsp) {
5530: PetscReal *coords = pcbddc->mat_graph->coords,*scoords;
5531: const PetscInt *idxs;
5532: PetscInt cdim = pcbddc->mat_graph->cdim,nl,i,d;
5534: ISGetLocalSize(pcis->is_I_local,&nl);
5535: ISGetIndices(pcis->is_I_local,&idxs);
5536: PetscMalloc1(nl*cdim,&scoords);
5537: for (i=0;i<nl;i++) {
5538: for (d=0;d<cdim;d++) {
5539: scoords[i*cdim+d] = coords[idxs[i]*cdim+d];
5540: }
5541: }
5542: ISRestoreIndices(pcis->is_I_local,&idxs);
5543: PCSetCoordinates(pc_temp,cdim,nl,scoords);
5544: PetscFree(scoords);
5545: }
5546: if (sub_schurs && sub_schurs->reuse_solver) {
5547: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5549: KSPSetPC(pcbddc->ksp_D,reuse_solver->interior_solver);
5550: }
5552: /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
5553: if (!n_D) {
5554: KSPGetPC(pcbddc->ksp_D,&pc_temp);
5555: PCSetType(pc_temp,PCNONE);
5556: }
5557: KSPSetUp(pcbddc->ksp_D);
5558: /* set ksp_D into pcis data */
5559: PetscObjectReference((PetscObject)pcbddc->ksp_D);
5560: KSPDestroy(&pcis->ksp_D);
5561: pcis->ksp_D = pcbddc->ksp_D;
5562: }
5564: /* NEUMANN PROBLEM */
5565: A_RR = 0;
5566: if (neumann) {
5567: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
5568: PetscInt ibs,mbs;
5569: PetscBool issbaij, reuse_neumann_solver;
5570: Mat_IS* matis = (Mat_IS*)pc->pmat->data;
5572: reuse_neumann_solver = PETSC_FALSE;
5573: if (sub_schurs && sub_schurs->reuse_solver) {
5574: IS iP;
5576: reuse_neumann_solver = PETSC_TRUE;
5577: PetscObjectQuery((PetscObject)sub_schurs->A,"__KSPFETIDP_iP",(PetscObject*)&iP);
5578: if (iP) reuse_neumann_solver = PETSC_FALSE;
5579: }
5580: /* Matrix for Neumann problem is A_RR -> we need to create/reuse it at this point */
5581: ISGetSize(pcbddc->is_R_local,&n_R);
5582: if (pcbddc->ksp_R) { /* already created ksp */
5583: PetscInt nn_R;
5584: KSPGetOperators(pcbddc->ksp_R,NULL,&A_RR);
5585: PetscObjectReference((PetscObject)A_RR);
5586: MatGetSize(A_RR,&nn_R,NULL);
5587: if (nn_R != n_R) { /* old ksp is not reusable, so reset it */
5588: KSPReset(pcbddc->ksp_R);
5589: MatDestroy(&A_RR);
5590: reuse = MAT_INITIAL_MATRIX;
5591: } else { /* same sizes, but nonzero pattern depend on primal vertices so it can be changed */
5592: if (pcbddc->new_primal_space_local) { /* we are not sure the matrix will have the same nonzero pattern */
5593: MatDestroy(&A_RR);
5594: reuse = MAT_INITIAL_MATRIX;
5595: } else { /* safe to reuse the matrix */
5596: reuse = MAT_REUSE_MATRIX;
5597: }
5598: }
5599: /* last check */
5600: if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
5601: MatDestroy(&A_RR);
5602: reuse = MAT_INITIAL_MATRIX;
5603: }
5604: } else { /* first time, so we need to create the matrix */
5605: reuse = MAT_INITIAL_MATRIX;
5606: }
5607: /* convert pcbddc->local_mat if needed later in PCBDDCSetUpCorrection
5608: TODO: Get Rid of these conversions */
5609: MatGetBlockSize(pcbddc->local_mat,&mbs);
5610: ISGetBlockSize(pcbddc->is_R_local,&ibs);
5611: PetscObjectTypeCompare((PetscObject)pcbddc->local_mat,MATSEQSBAIJ,&issbaij);
5612: if (ibs != mbs) { /* need to convert to SEQAIJ to extract any submatrix with is_R_local */
5613: if (matis->A == pcbddc->local_mat) {
5614: MatDestroy(&pcbddc->local_mat);
5615: MatConvert(matis->A,MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);
5616: } else {
5617: MatConvert(pcbddc->local_mat,MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);
5618: }
5619: } else if (issbaij) { /* need to convert to BAIJ to get offdiagonal blocks */
5620: if (matis->A == pcbddc->local_mat) {
5621: MatDestroy(&pcbddc->local_mat);
5622: MatConvert(matis->A,mbs > 1 ? MATSEQBAIJ : MATSEQAIJ,MAT_INITIAL_MATRIX,&pcbddc->local_mat);
5623: } else {
5624: MatConvert(pcbddc->local_mat,mbs > 1 ? MATSEQBAIJ : MATSEQAIJ,MAT_INPLACE_MATRIX,&pcbddc->local_mat);
5625: }
5626: }
5627: /* extract A_RR */
5628: if (reuse_neumann_solver) {
5629: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5631: if (pcbddc->dbg_flag) { /* we need A_RR to test the solver later */
5632: MatDestroy(&A_RR);
5633: if (reuse_solver->benign_n) { /* we are not using the explicit change of basis on the pressures */
5634: PCBDDCBenignProject(pc,pcbddc->is_R_local,pcbddc->is_R_local,&A_RR);
5635: } else {
5636: MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,MAT_INITIAL_MATRIX,&A_RR);
5637: }
5638: } else {
5639: MatDestroy(&A_RR);
5640: PCGetOperators(reuse_solver->correction_solver,&A_RR,NULL);
5641: PetscObjectReference((PetscObject)A_RR);
5642: }
5643: } else { /* we have to build the neumann solver, so we need to extract the relevant matrix */
5644: MatCreateSubMatrix(pcbddc->local_mat,pcbddc->is_R_local,pcbddc->is_R_local,reuse,&A_RR);
5645: }
5646: if (pcbddc->local_mat->symmetric_set) {
5647: MatSetOption(A_RR,MAT_SYMMETRIC,pcbddc->local_mat->symmetric);
5648: }
5649: opts = PETSC_FALSE;
5650: if (!pcbddc->ksp_R) { /* create object if not present */
5651: opts = PETSC_TRUE;
5652: KSPCreate(PETSC_COMM_SELF,&pcbddc->ksp_R);
5653: PetscObjectIncrementTabLevel((PetscObject)pcbddc->ksp_R,(PetscObject)pc,1);
5654: /* default */
5655: KSPSetType(pcbddc->ksp_R,KSPPREONLY);
5656: KSPSetOptionsPrefix(pcbddc->ksp_R,neu_prefix);
5657: KSPGetPC(pcbddc->ksp_R,&pc_temp);
5658: PetscObjectTypeCompare((PetscObject)A_RR,MATSEQSBAIJ,&issbaij);
5659: if (issbaij) {
5660: PCSetType(pc_temp,PCCHOLESKY);
5661: } else {
5662: PCSetType(pc_temp,PCLU);
5663: }
5664: KSPSetErrorIfNotConverged(pcbddc->ksp_R,pc->erroriffailure);
5665: }
5666: KSPSetOperators(pcbddc->ksp_R,A_RR,A_RR);
5667: MatSetOptionsPrefix(A_RR,((PetscObject)pcbddc->ksp_R)->prefix);
5668: if (opts) { /* Allow user's customization once */
5669: KSPSetFromOptions(pcbddc->ksp_R);
5670: }
5671: MatGetNearNullSpace(A_RR,&nnsp);
5672: if (pcbddc->NullSpace_corr[2] && !nnsp) { /* approximate solver, propagate NearNullSpace */
5673: MatNullSpacePropagateAny_Private(pcbddc->local_mat,pcbddc->is_R_local,A_RR);
5674: }
5675: MatGetNearNullSpace(A_RR,&nnsp);
5676: KSPGetPC(pcbddc->ksp_R,&pc_temp);
5677: PetscObjectQueryFunction((PetscObject)pc_temp,"PCSetCoordinates_C",&f);
5678: if (f && pcbddc->mat_graph->cloc && !nnsp) {
5679: PetscReal *coords = pcbddc->mat_graph->coords,*scoords;
5680: const PetscInt *idxs;
5681: PetscInt cdim = pcbddc->mat_graph->cdim,nl,i,d;
5683: ISGetLocalSize(pcbddc->is_R_local,&nl);
5684: ISGetIndices(pcbddc->is_R_local,&idxs);
5685: PetscMalloc1(nl*cdim,&scoords);
5686: for (i=0;i<nl;i++) {
5687: for (d=0;d<cdim;d++) {
5688: scoords[i*cdim+d] = coords[idxs[i]*cdim+d];
5689: }
5690: }
5691: ISRestoreIndices(pcbddc->is_R_local,&idxs);
5692: PCSetCoordinates(pc_temp,cdim,nl,scoords);
5693: PetscFree(scoords);
5694: }
5696: /* umfpack interface has a bug when matrix dimension is zero. TODO solve from umfpack interface */
5697: if (!n_R) {
5698: KSPGetPC(pcbddc->ksp_R,&pc_temp);
5699: PCSetType(pc_temp,PCNONE);
5700: }
5701: /* Reuse solver if it is present */
5702: if (reuse_neumann_solver) {
5703: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5705: KSPSetPC(pcbddc->ksp_R,reuse_solver->correction_solver);
5706: }
5707: KSPSetUp(pcbddc->ksp_R);
5708: }
5710: if (pcbddc->dbg_flag) {
5711: PetscViewerFlush(pcbddc->dbg_viewer);
5712: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
5713: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");
5714: }
5715: PetscLogEventEnd(PC_BDDC_LocalSolvers[pcbddc->current_level],pc,0,0,0);
5717: /* adapt Dirichlet and Neumann solvers if a nullspace correction has been requested */
5718: if (pcbddc->NullSpace_corr[0]) {
5719: PCBDDCSetUseExactDirichlet(pc,PETSC_FALSE);
5720: }
5721: if (dirichlet && pcbddc->NullSpace_corr[0] && !pcbddc->switch_static) {
5722: PCBDDCNullSpaceAssembleCorrection(pc,PETSC_TRUE,pcbddc->NullSpace_corr[1]);
5723: }
5724: if (neumann && pcbddc->NullSpace_corr[2]) {
5725: PCBDDCNullSpaceAssembleCorrection(pc,PETSC_FALSE,pcbddc->NullSpace_corr[3]);
5726: }
5727: /* check Dirichlet and Neumann solvers */
5728: if (pcbddc->dbg_flag) {
5729: if (dirichlet) { /* Dirichlet */
5730: VecSetRandom(pcis->vec1_D,NULL);
5731: MatMult(pcis->A_II,pcis->vec1_D,pcis->vec2_D);
5732: KSPSolve(pcbddc->ksp_D,pcis->vec2_D,pcis->vec2_D);
5733: KSPCheckSolve(pcbddc->ksp_D,pc,pcis->vec2_D);
5734: VecAXPY(pcis->vec1_D,m_one,pcis->vec2_D);
5735: VecNorm(pcis->vec1_D,NORM_INFINITY,&value);
5736: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Dirichlet solve (%s) = % 1.14e \n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_D))->prefix,value);
5737: PetscViewerFlush(pcbddc->dbg_viewer);
5738: }
5739: if (neumann) { /* Neumann */
5740: VecSetRandom(pcbddc->vec1_R,NULL);
5741: MatMult(A_RR,pcbddc->vec1_R,pcbddc->vec2_R);
5742: KSPSolve(pcbddc->ksp_R,pcbddc->vec2_R,pcbddc->vec2_R);
5743: KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec2_R);
5744: VecAXPY(pcbddc->vec1_R,m_one,pcbddc->vec2_R);
5745: VecNorm(pcbddc->vec1_R,NORM_INFINITY,&value);
5746: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d infinity error for Neumann solve (%s) = % 1.14e\n",PetscGlobalRank,((PetscObject)(pcbddc->ksp_R))->prefix,value);
5747: PetscViewerFlush(pcbddc->dbg_viewer);
5748: }
5749: }
5750: /* free Neumann problem's matrix */
5751: MatDestroy(&A_RR);
5752: return(0);
5753: }
5755: static PetscErrorCode PCBDDCSolveSubstructureCorrection(PC pc, Vec inout_B, Vec inout_D, PetscBool applytranspose)
5756: {
5757: PetscErrorCode ierr;
5758: PC_BDDC* pcbddc = (PC_BDDC*)(pc->data);
5759: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
5760: PetscBool reuse_solver = sub_schurs ? ( sub_schurs->reuse_solver ? PETSC_TRUE : PETSC_FALSE ) : PETSC_FALSE;
5763: if (!reuse_solver) {
5764: VecSet(pcbddc->vec1_R,0.);
5765: }
5766: if (!pcbddc->switch_static) {
5767: if (applytranspose && pcbddc->local_auxmat1) {
5768: MatMultTranspose(pcbddc->local_auxmat2,inout_B,pcbddc->vec1_C);
5769: MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);
5770: }
5771: if (!reuse_solver) {
5772: VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5773: VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5774: } else {
5775: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5777: VecScatterBegin(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);
5778: VecScatterEnd(reuse_solver->correction_scatter_B,inout_B,reuse_solver->rhs_B,INSERT_VALUES,SCATTER_FORWARD);
5779: }
5780: } else {
5781: VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5782: VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5783: VecScatterBegin(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5784: VecScatterEnd(pcbddc->R_to_D,inout_D,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5785: if (applytranspose && pcbddc->local_auxmat1) {
5786: MatMultTranspose(pcbddc->local_auxmat2,pcbddc->vec1_R,pcbddc->vec1_C);
5787: MatMultTransposeAdd(pcbddc->local_auxmat1,pcbddc->vec1_C,inout_B,inout_B);
5788: VecScatterBegin(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5789: VecScatterEnd(pcbddc->R_to_B,inout_B,pcbddc->vec1_R,INSERT_VALUES,SCATTER_REVERSE);
5790: }
5791: }
5792: if (!reuse_solver || pcbddc->switch_static) {
5793: if (applytranspose) {
5794: KSPSolveTranspose(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);
5795: } else {
5796: KSPSolve(pcbddc->ksp_R,pcbddc->vec1_R,pcbddc->vec1_R);
5797: }
5798: KSPCheckSolve(pcbddc->ksp_R,pc,pcbddc->vec1_R);
5799: } else {
5800: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5802: if (applytranspose) {
5803: MatFactorSolveSchurComplementTranspose(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);
5804: } else {
5805: MatFactorSolveSchurComplement(reuse_solver->F,reuse_solver->rhs_B,reuse_solver->sol_B);
5806: }
5807: }
5808: VecSet(inout_B,0.);
5809: if (!pcbddc->switch_static) {
5810: if (!reuse_solver) {
5811: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);
5812: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);
5813: } else {
5814: PCBDDCReuseSolvers reuse_solver = sub_schurs->reuse_solver;
5816: VecScatterBegin(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);
5817: VecScatterEnd(reuse_solver->correction_scatter_B,reuse_solver->sol_B,inout_B,INSERT_VALUES,SCATTER_REVERSE);
5818: }
5819: if (!applytranspose && pcbddc->local_auxmat1) {
5820: MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);
5821: MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,inout_B,inout_B);
5822: }
5823: } else {
5824: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);
5825: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);
5826: VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);
5827: VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);
5828: if (!applytranspose && pcbddc->local_auxmat1) {
5829: MatMult(pcbddc->local_auxmat1,inout_B,pcbddc->vec1_C);
5830: MatMultAdd(pcbddc->local_auxmat2,pcbddc->vec1_C,pcbddc->vec1_R,pcbddc->vec1_R);
5831: }
5832: VecScatterBegin(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);
5833: VecScatterEnd(pcbddc->R_to_B,pcbddc->vec1_R,inout_B,INSERT_VALUES,SCATTER_FORWARD);
5834: VecScatterBegin(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);
5835: VecScatterEnd(pcbddc->R_to_D,pcbddc->vec1_R,inout_D,INSERT_VALUES,SCATTER_FORWARD);
5836: }
5837: return(0);
5838: }
5840: /* parameter apply transpose determines if the interface preconditioner should be applied transposed or not */
5841: PetscErrorCode PCBDDCApplyInterfacePreconditioner(PC pc, PetscBool applytranspose)
5842: {
5844: PC_BDDC* pcbddc = (PC_BDDC*)(pc->data);
5845: PC_IS* pcis = (PC_IS*) (pc->data);
5846: const PetscScalar zero = 0.0;
5849: /* Application of PSI^T or PHI^T (depending on applytranspose, see comment above) */
5850: if (!pcbddc->benign_apply_coarse_only) {
5851: if (applytranspose) {
5852: MatMultTranspose(pcbddc->coarse_phi_B,pcis->vec1_B,pcbddc->vec1_P);
5853: if (pcbddc->switch_static) { MatMultTransposeAdd(pcbddc->coarse_phi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P); }
5854: } else {
5855: MatMultTranspose(pcbddc->coarse_psi_B,pcis->vec1_B,pcbddc->vec1_P);
5856: if (pcbddc->switch_static) { MatMultTransposeAdd(pcbddc->coarse_psi_D,pcis->vec1_D,pcbddc->vec1_P,pcbddc->vec1_P); }
5857: }
5858: } else {
5859: VecSet(pcbddc->vec1_P,zero);
5860: }
5862: /* add p0 to the last value of vec1_P holding the coarse dof relative to p0 */
5863: if (pcbddc->benign_n) {
5864: PetscScalar *array;
5865: PetscInt j;
5867: VecGetArray(pcbddc->vec1_P,&array);
5868: for (j=0;j<pcbddc->benign_n;j++) array[pcbddc->local_primal_size-pcbddc->benign_n+j] += pcbddc->benign_p0[j];
5869: VecRestoreArray(pcbddc->vec1_P,&array);
5870: }
5872: /* start communications from local primal nodes to rhs of coarse solver */
5873: VecSet(pcbddc->coarse_vec,zero);
5874: PCBDDCScatterCoarseDataBegin(pc,ADD_VALUES,SCATTER_FORWARD);
5875: PCBDDCScatterCoarseDataEnd(pc,ADD_VALUES,SCATTER_FORWARD);
5877: /* Coarse solution -> rhs and sol updated inside PCBDDCScattarCoarseDataBegin/End */
5878: if (pcbddc->coarse_ksp) {
5879: Mat coarse_mat;
5880: Vec rhs,sol;
5881: MatNullSpace nullsp;
5882: PetscBool isbddc = PETSC_FALSE;
5884: if (pcbddc->benign_have_null) {
5885: PC coarse_pc;
5887: KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);
5888: PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);
5889: /* we need to propagate to coarser levels the need for a possible benign correction */
5890: if (isbddc && pcbddc->benign_apply_coarse_only && !pcbddc->benign_skip_correction) {
5891: PC_BDDC* coarsepcbddc = (PC_BDDC*)(coarse_pc->data);
5892: coarsepcbddc->benign_skip_correction = PETSC_FALSE;
5893: coarsepcbddc->benign_apply_coarse_only = PETSC_TRUE;
5894: }
5895: }
5896: KSPGetRhs(pcbddc->coarse_ksp,&rhs);
5897: KSPGetSolution(pcbddc->coarse_ksp,&sol);
5898: KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);
5899: if (applytranspose) {
5900: if (pcbddc->benign_apply_coarse_only) SETERRQ(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),PETSC_ERR_SUP,"Not yet implemented");
5901: KSPSolveTranspose(pcbddc->coarse_ksp,rhs,sol);
5902: KSPCheckSolve(pcbddc->coarse_ksp,pc,sol);
5903: MatGetTransposeNullSpace(coarse_mat,&nullsp);
5904: if (nullsp) {
5905: MatNullSpaceRemove(nullsp,sol);
5906: }
5907: } else {
5908: MatGetNullSpace(coarse_mat,&nullsp);
5909: if (pcbddc->benign_apply_coarse_only && isbddc) { /* need just to apply the coarse preconditioner during presolve */
5910: PC coarse_pc;
5912: if (nullsp) {
5913: MatNullSpaceRemove(nullsp,rhs);
5914: }
5915: KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);
5916: PCPreSolve(coarse_pc,pcbddc->coarse_ksp);
5917: PCBDDCBenignRemoveInterior(coarse_pc,rhs,sol);
5918: PCPostSolve(coarse_pc,pcbddc->coarse_ksp);
5919: } else {
5920: KSPSolve(pcbddc->coarse_ksp,rhs,sol);
5921: KSPCheckSolve(pcbddc->coarse_ksp,pc,sol);
5922: if (nullsp) {
5923: MatNullSpaceRemove(nullsp,sol);
5924: }
5925: }
5926: }
5927: /* we don't need the benign correction at coarser levels anymore */
5928: if (pcbddc->benign_have_null && isbddc) {
5929: PC coarse_pc;
5930: PC_BDDC* coarsepcbddc;
5932: KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);
5933: coarsepcbddc = (PC_BDDC*)(coarse_pc->data);
5934: coarsepcbddc->benign_skip_correction = PETSC_TRUE;
5935: coarsepcbddc->benign_apply_coarse_only = PETSC_FALSE;
5936: }
5937: }
5939: /* Local solution on R nodes */
5940: if (pcis->n && !pcbddc->benign_apply_coarse_only) {
5941: PCBDDCSolveSubstructureCorrection(pc,pcis->vec1_B,pcis->vec1_D,applytranspose);
5942: }
5943: /* communications from coarse sol to local primal nodes */
5944: PCBDDCScatterCoarseDataBegin(pc,INSERT_VALUES,SCATTER_REVERSE);
5945: PCBDDCScatterCoarseDataEnd(pc,INSERT_VALUES,SCATTER_REVERSE);
5947: /* Sum contributions from the two levels */
5948: if (!pcbddc->benign_apply_coarse_only) {
5949: if (applytranspose) {
5950: MatMultAdd(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);
5951: if (pcbddc->switch_static) { MatMultAdd(pcbddc->coarse_psi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D); }
5952: } else {
5953: MatMultAdd(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B,pcis->vec1_B);
5954: if (pcbddc->switch_static) { MatMultAdd(pcbddc->coarse_phi_D,pcbddc->vec1_P,pcis->vec1_D,pcis->vec1_D); }
5955: }
5956: /* store p0 */
5957: if (pcbddc->benign_n) {
5958: PetscScalar *array;
5959: PetscInt j;
5961: VecGetArray(pcbddc->vec1_P,&array);
5962: for (j=0;j<pcbddc->benign_n;j++) pcbddc->benign_p0[j] = array[pcbddc->local_primal_size-pcbddc->benign_n+j];
5963: VecRestoreArray(pcbddc->vec1_P,&array);
5964: }
5965: } else { /* expand the coarse solution */
5966: if (applytranspose) {
5967: MatMult(pcbddc->coarse_psi_B,pcbddc->vec1_P,pcis->vec1_B);
5968: } else {
5969: MatMult(pcbddc->coarse_phi_B,pcbddc->vec1_P,pcis->vec1_B);
5970: }
5971: }
5972: return(0);
5973: }
5975: PetscErrorCode PCBDDCScatterCoarseDataBegin(PC pc,InsertMode imode, ScatterMode smode)
5976: {
5977: PC_BDDC* pcbddc = (PC_BDDC*)(pc->data);
5978: Vec from,to;
5979: const PetscScalar *array;
5980: PetscErrorCode ierr;
5983: if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
5984: from = pcbddc->coarse_vec;
5985: to = pcbddc->vec1_P;
5986: if (pcbddc->coarse_ksp) { /* get array from coarse processes */
5987: Vec tvec;
5989: KSPGetRhs(pcbddc->coarse_ksp,&tvec);
5990: VecResetArray(tvec);
5991: KSPGetSolution(pcbddc->coarse_ksp,&tvec);
5992: VecGetArrayRead(tvec,&array);
5993: VecPlaceArray(from,array);
5994: VecRestoreArrayRead(tvec,&array);
5995: }
5996: } else { /* from local to global -> put data in coarse right hand side */
5997: from = pcbddc->vec1_P;
5998: to = pcbddc->coarse_vec;
5999: }
6000: VecScatterBegin(pcbddc->coarse_loc_to_glob,from,to,imode,smode);
6001: return(0);
6002: }
6004: PetscErrorCode PCBDDCScatterCoarseDataEnd(PC pc, InsertMode imode, ScatterMode smode)
6005: {
6006: PC_BDDC* pcbddc = (PC_BDDC*)(pc->data);
6007: Vec from,to;
6008: const PetscScalar *array;
6009: PetscErrorCode ierr;
6012: if (smode == SCATTER_REVERSE) { /* from global to local -> get data from coarse solution */
6013: from = pcbddc->coarse_vec;
6014: to = pcbddc->vec1_P;
6015: } else { /* from local to global -> put data in coarse right hand side */
6016: from = pcbddc->vec1_P;
6017: to = pcbddc->coarse_vec;
6018: }
6019: VecScatterEnd(pcbddc->coarse_loc_to_glob,from,to,imode,smode);
6020: if (smode == SCATTER_FORWARD) {
6021: if (pcbddc->coarse_ksp) { /* get array from coarse processes */
6022: Vec tvec;
6024: KSPGetRhs(pcbddc->coarse_ksp,&tvec);
6025: VecGetArrayRead(to,&array);
6026: VecPlaceArray(tvec,array);
6027: VecRestoreArrayRead(to,&array);
6028: }
6029: } else {
6030: if (pcbddc->coarse_ksp) { /* restore array of pcbddc->coarse_vec */
6031: VecResetArray(from);
6032: }
6033: }
6034: return(0);
6035: }
6037: PetscErrorCode PCBDDCConstraintsSetUp(PC pc)
6038: {
6039: PetscErrorCode ierr;
6040: PC_IS* pcis = (PC_IS*)(pc->data);
6041: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
6042: Mat_IS* matis = (Mat_IS*)pc->pmat->data;
6043: /* one and zero */
6044: PetscScalar one=1.0,zero=0.0;
6045: /* space to store constraints and their local indices */
6046: PetscScalar *constraints_data;
6047: PetscInt *constraints_idxs,*constraints_idxs_B;
6048: PetscInt *constraints_idxs_ptr,*constraints_data_ptr;
6049: PetscInt *constraints_n;
6050: /* iterators */
6051: PetscInt i,j,k,total_counts,total_counts_cc,cum;
6052: /* BLAS integers */
6053: PetscBLASInt lwork,lierr;
6054: PetscBLASInt Blas_N,Blas_M,Blas_K,Blas_one=1;
6055: PetscBLASInt Blas_LDA,Blas_LDB,Blas_LDC;
6056: /* reuse */
6057: PetscInt olocal_primal_size,olocal_primal_size_cc;
6058: PetscInt *olocal_primal_ref_node,*olocal_primal_ref_mult;
6059: /* change of basis */
6060: PetscBool qr_needed;
6061: PetscBT change_basis,qr_needed_idx;
6062: /* auxiliary stuff */
6063: PetscInt *nnz,*is_indices;
6064: PetscInt ncc;
6065: /* some quantities */
6066: PetscInt n_vertices,total_primal_vertices,valid_constraints;
6067: PetscInt size_of_constraint,max_size_of_constraint=0,max_constraints,temp_constraints;
6068: PetscReal tol; /* tolerance for retaining eigenmodes */
6071: tol = PetscSqrtReal(PETSC_SMALL);
6072: /* Destroy Mat objects computed previously */
6073: MatDestroy(&pcbddc->ChangeOfBasisMatrix);
6074: MatDestroy(&pcbddc->ConstraintMatrix);
6075: MatDestroy(&pcbddc->switch_static_change);
6076: /* save info on constraints from previous setup (if any) */
6077: olocal_primal_size = pcbddc->local_primal_size;
6078: olocal_primal_size_cc = pcbddc->local_primal_size_cc;
6079: PetscMalloc2(olocal_primal_size_cc,&olocal_primal_ref_node,olocal_primal_size_cc,&olocal_primal_ref_mult);
6080: PetscArraycpy(olocal_primal_ref_node,pcbddc->local_primal_ref_node,olocal_primal_size_cc);
6081: PetscArraycpy(olocal_primal_ref_mult,pcbddc->local_primal_ref_mult,olocal_primal_size_cc);
6082: PetscFree2(pcbddc->local_primal_ref_node,pcbddc->local_primal_ref_mult);
6083: PetscFree(pcbddc->primal_indices_local_idxs);
6085: if (!pcbddc->adaptive_selection) {
6086: IS ISForVertices,*ISForFaces,*ISForEdges;
6087: MatNullSpace nearnullsp;
6088: const Vec *nearnullvecs;
6089: Vec *localnearnullsp;
6090: PetscScalar *array;
6091: PetscInt n_ISForFaces,n_ISForEdges,nnsp_size;
6092: PetscBool nnsp_has_cnst;
6093: /* LAPACK working arrays for SVD or POD */
6094: PetscBool skip_lapack,boolforchange;
6095: PetscScalar *work;
6096: PetscReal *singular_vals;
6097: #if defined(PETSC_USE_COMPLEX)
6098: PetscReal *rwork;
6099: #endif
6100: PetscScalar *temp_basis = NULL,*correlation_mat = NULL;
6101: PetscBLASInt dummy_int=1;
6102: PetscScalar dummy_scalar=1.;
6103: PetscBool use_pod = PETSC_FALSE;
6105: /* MKL SVD with same input gives different results on different processes! */
6106: #if defined(PETSC_MISSING_LAPACK_GESVD) || defined(PETSC_HAVE_MKL)
6107: use_pod = PETSC_TRUE;
6108: #endif
6109: /* Get index sets for faces, edges and vertices from graph */
6110: PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,&n_ISForFaces,&ISForFaces,&n_ISForEdges,&ISForEdges,&ISForVertices);
6111: /* print some info */
6112: if (pcbddc->dbg_flag && (!pcbddc->sub_schurs || pcbddc->sub_schurs_rebuild)) {
6113: PetscInt nv;
6115: PCBDDCGraphASCIIView(pcbddc->mat_graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);
6116: ISGetSize(ISForVertices,&nv);
6117: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
6118: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");
6119: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%D)\n",PetscGlobalRank,nv,pcbddc->use_vertices);
6120: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%D)\n",PetscGlobalRank,n_ISForEdges,pcbddc->use_edges);
6121: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%D)\n",PetscGlobalRank,n_ISForFaces,pcbddc->use_faces);
6122: PetscViewerFlush(pcbddc->dbg_viewer);
6123: PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);
6124: }
6126: /* free unneeded index sets */
6127: if (!pcbddc->use_vertices) {
6128: ISDestroy(&ISForVertices);
6129: }
6130: if (!pcbddc->use_edges) {
6131: for (i=0;i<n_ISForEdges;i++) {
6132: ISDestroy(&ISForEdges[i]);
6133: }
6134: PetscFree(ISForEdges);
6135: n_ISForEdges = 0;
6136: }
6137: if (!pcbddc->use_faces) {
6138: for (i=0;i<n_ISForFaces;i++) {
6139: ISDestroy(&ISForFaces[i]);
6140: }
6141: PetscFree(ISForFaces);
6142: n_ISForFaces = 0;
6143: }
6145: /* check if near null space is attached to global mat */
6146: if (pcbddc->use_nnsp) {
6147: MatGetNearNullSpace(pc->pmat,&nearnullsp);
6148: } else nearnullsp = NULL;
6150: if (nearnullsp) {
6151: MatNullSpaceGetVecs(nearnullsp,&nnsp_has_cnst,&nnsp_size,&nearnullvecs);
6152: /* remove any stored info */
6153: MatNullSpaceDestroy(&pcbddc->onearnullspace);
6154: PetscFree(pcbddc->onearnullvecs_state);
6155: /* store information for BDDC solver reuse */
6156: PetscObjectReference((PetscObject)nearnullsp);
6157: pcbddc->onearnullspace = nearnullsp;
6158: PetscMalloc1(nnsp_size,&pcbddc->onearnullvecs_state);
6159: for (i=0;i<nnsp_size;i++) {
6160: PetscObjectStateGet((PetscObject)nearnullvecs[i],&pcbddc->onearnullvecs_state[i]);
6161: }
6162: } else { /* if near null space is not provided BDDC uses constants by default */
6163: nnsp_size = 0;
6164: nnsp_has_cnst = PETSC_TRUE;
6165: }
6166: /* get max number of constraints on a single cc */
6167: max_constraints = nnsp_size;
6168: if (nnsp_has_cnst) max_constraints++;
6170: /*
6171: Evaluate maximum storage size needed by the procedure
6172: - Indices for connected component i stored at "constraints_idxs + constraints_idxs_ptr[i]"
6173: - Values for constraints on connected component i stored at "constraints_data + constraints_data_ptr[i]"
6174: There can be multiple constraints per connected component
6175: */
6176: n_vertices = 0;
6177: if (ISForVertices) {
6178: ISGetSize(ISForVertices,&n_vertices);
6179: }
6180: ncc = n_vertices+n_ISForFaces+n_ISForEdges;
6181: PetscMalloc3(ncc+1,&constraints_idxs_ptr,ncc+1,&constraints_data_ptr,ncc,&constraints_n);
6183: total_counts = n_ISForFaces+n_ISForEdges;
6184: total_counts *= max_constraints;
6185: total_counts += n_vertices;
6186: PetscBTCreate(total_counts,&change_basis);
6188: total_counts = 0;
6189: max_size_of_constraint = 0;
6190: for (i=0;i<n_ISForEdges+n_ISForFaces;i++) {
6191: IS used_is;
6192: if (i<n_ISForEdges) {
6193: used_is = ISForEdges[i];
6194: } else {
6195: used_is = ISForFaces[i-n_ISForEdges];
6196: }
6197: ISGetSize(used_is,&j);
6198: total_counts += j;
6199: max_size_of_constraint = PetscMax(j,max_size_of_constraint);
6200: }
6201: PetscMalloc3(total_counts*max_constraints+n_vertices,&constraints_data,total_counts+n_vertices,&constraints_idxs,total_counts+n_vertices,&constraints_idxs_B);
6203: /* get local part of global near null space vectors */
6204: PetscMalloc1(nnsp_size,&localnearnullsp);
6205: for (k=0;k<nnsp_size;k++) {
6206: VecDuplicate(pcis->vec1_N,&localnearnullsp[k]);
6207: VecScatterBegin(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);
6208: VecScatterEnd(matis->rctx,nearnullvecs[k],localnearnullsp[k],INSERT_VALUES,SCATTER_FORWARD);
6209: }
6211: /* whether or not to skip lapack calls */
6212: skip_lapack = PETSC_TRUE;
6213: if (n_ISForFaces+n_ISForEdges && max_constraints > 1 && !pcbddc->use_nnsp_true) skip_lapack = PETSC_FALSE;
6215: /* First we issue queries to allocate optimal workspace for LAPACKgesvd (or LAPACKsyev if SVD is missing) */
6216: if (!skip_lapack) {
6217: PetscScalar temp_work;
6219: if (use_pod) {
6220: /* Proper Orthogonal Decomposition (POD) using the snapshot method */
6221: PetscMalloc1(max_constraints*max_constraints,&correlation_mat);
6222: PetscMalloc1(max_constraints,&singular_vals);
6223: PetscMalloc1(max_size_of_constraint*max_constraints,&temp_basis);
6224: #if defined(PETSC_USE_COMPLEX)
6225: PetscMalloc1(3*max_constraints,&rwork);
6226: #endif
6227: /* now we evaluate the optimal workspace using query with lwork=-1 */
6228: PetscBLASIntCast(max_constraints,&Blas_N);
6229: PetscBLASIntCast(max_constraints,&Blas_LDA);
6230: lwork = -1;
6231: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6232: #if !defined(PETSC_USE_COMPLEX)
6233: PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,&lierr));
6234: #else
6235: PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,&temp_work,&lwork,rwork,&lierr));
6236: #endif
6237: PetscFPTrapPop();
6238: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to SYEV Lapack routine %d",(int)lierr);
6239: } else {
6240: #if !defined(PETSC_MISSING_LAPACK_GESVD)
6241: /* SVD */
6242: PetscInt max_n,min_n;
6243: max_n = max_size_of_constraint;
6244: min_n = max_constraints;
6245: if (max_size_of_constraint < max_constraints) {
6246: min_n = max_size_of_constraint;
6247: max_n = max_constraints;
6248: }
6249: PetscMalloc1(min_n,&singular_vals);
6250: #if defined(PETSC_USE_COMPLEX)
6251: PetscMalloc1(5*min_n,&rwork);
6252: #endif
6253: /* now we evaluate the optimal workspace using query with lwork=-1 */
6254: lwork = -1;
6255: PetscBLASIntCast(max_n,&Blas_M);
6256: PetscBLASIntCast(min_n,&Blas_N);
6257: PetscBLASIntCast(max_n,&Blas_LDA);
6258: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6259: #if !defined(PETSC_USE_COMPLEX)
6260: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,&lierr));
6261: #else
6262: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,&constraints_data[0],&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,&temp_work,&lwork,rwork,&lierr));
6263: #endif
6264: PetscFPTrapPop();
6265: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GESVD Lapack routine %d",(int)lierr);
6266: #else
6267: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"This should not happen");
6268: #endif /* on missing GESVD */
6269: }
6270: /* Allocate optimal workspace */
6271: PetscBLASIntCast((PetscInt)PetscRealPart(temp_work),&lwork);
6272: PetscMalloc1(lwork,&work);
6273: }
6274: /* Now we can loop on constraining sets */
6275: total_counts = 0;
6276: constraints_idxs_ptr[0] = 0;
6277: constraints_data_ptr[0] = 0;
6278: /* vertices */
6279: if (n_vertices) {
6280: ISGetIndices(ISForVertices,(const PetscInt**)&is_indices);
6281: PetscArraycpy(constraints_idxs,is_indices,n_vertices);
6282: for (i=0;i<n_vertices;i++) {
6283: constraints_n[total_counts] = 1;
6284: constraints_data[total_counts] = 1.0;
6285: constraints_idxs_ptr[total_counts+1] = constraints_idxs_ptr[total_counts]+1;
6286: constraints_data_ptr[total_counts+1] = constraints_data_ptr[total_counts]+1;
6287: total_counts++;
6288: }
6289: ISRestoreIndices(ISForVertices,(const PetscInt**)&is_indices);
6290: n_vertices = total_counts;
6291: }
6293: /* edges and faces */
6294: total_counts_cc = total_counts;
6295: for (ncc=0;ncc<n_ISForEdges+n_ISForFaces;ncc++) {
6296: IS used_is;
6297: PetscBool idxs_copied = PETSC_FALSE;
6299: if (ncc<n_ISForEdges) {
6300: used_is = ISForEdges[ncc];
6301: boolforchange = pcbddc->use_change_of_basis; /* change or not the basis on the edge */
6302: } else {
6303: used_is = ISForFaces[ncc-n_ISForEdges];
6304: boolforchange = (PetscBool)(pcbddc->use_change_of_basis && pcbddc->use_change_on_faces); /* change or not the basis on the face */
6305: }
6306: temp_constraints = 0; /* zero the number of constraints I have on this conn comp */
6308: ISGetSize(used_is,&size_of_constraint);
6309: ISGetIndices(used_is,(const PetscInt**)&is_indices);
6310: /* change of basis should not be performed on local periodic nodes */
6311: if (pcbddc->mat_graph->mirrors && pcbddc->mat_graph->mirrors[is_indices[0]]) boolforchange = PETSC_FALSE;
6312: if (nnsp_has_cnst) {
6313: PetscScalar quad_value;
6315: PetscArraycpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint);
6316: idxs_copied = PETSC_TRUE;
6318: if (!pcbddc->use_nnsp_true) {
6319: quad_value = (PetscScalar)(1.0/PetscSqrtReal((PetscReal)size_of_constraint));
6320: } else {
6321: quad_value = 1.0;
6322: }
6323: for (j=0;j<size_of_constraint;j++) {
6324: constraints_data[constraints_data_ptr[total_counts_cc]+j] = quad_value;
6325: }
6326: temp_constraints++;
6327: total_counts++;
6328: }
6329: for (k=0;k<nnsp_size;k++) {
6330: PetscReal real_value;
6331: PetscScalar *ptr_to_data;
6333: VecGetArrayRead(localnearnullsp[k],(const PetscScalar**)&array);
6334: ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]+temp_constraints*size_of_constraint];
6335: for (j=0;j<size_of_constraint;j++) {
6336: ptr_to_data[j] = array[is_indices[j]];
6337: }
6338: VecRestoreArrayRead(localnearnullsp[k],(const PetscScalar**)&array);
6339: /* check if array is null on the connected component */
6340: PetscBLASIntCast(size_of_constraint,&Blas_N);
6341: PetscStackCallBLAS("BLASasum",real_value = BLASasum_(&Blas_N,ptr_to_data,&Blas_one));
6342: if (real_value > tol*size_of_constraint) { /* keep indices and values */
6343: temp_constraints++;
6344: total_counts++;
6345: if (!idxs_copied) {
6346: PetscArraycpy(constraints_idxs + constraints_idxs_ptr[total_counts_cc],is_indices,size_of_constraint);
6347: idxs_copied = PETSC_TRUE;
6348: }
6349: }
6350: }
6351: ISRestoreIndices(used_is,(const PetscInt**)&is_indices);
6352: valid_constraints = temp_constraints;
6353: if (!pcbddc->use_nnsp_true && temp_constraints) {
6354: if (temp_constraints == 1) { /* just normalize the constraint */
6355: PetscScalar norm,*ptr_to_data;
6357: ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]];
6358: PetscBLASIntCast(size_of_constraint,&Blas_N);
6359: PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,ptr_to_data,&Blas_one,ptr_to_data,&Blas_one));
6360: norm = 1.0/PetscSqrtReal(PetscRealPart(norm));
6361: PetscStackCallBLAS("BLASscal",BLASscal_(&Blas_N,&norm,ptr_to_data,&Blas_one));
6362: } else { /* perform SVD */
6363: PetscScalar *ptr_to_data = &constraints_data[constraints_data_ptr[total_counts_cc]];
6365: if (use_pod) {
6366: /* SVD: Y = U*S*V^H -> U (eigenvectors of Y*Y^H) = Y*V*(S)^\dag
6367: POD: Y^H*Y = V*D*V^H, D = S^H*S -> U = Y*V*D^(-1/2)
6368: -> When PETSC_USE_COMPLEX and PETSC_MISSING_LAPACK_GESVD are defined
6369: the constraints basis will differ (by a complex factor with absolute value equal to 1)
6370: from that computed using LAPACKgesvd
6371: -> This is due to a different computation of eigenvectors in LAPACKheev
6372: -> The quality of the POD-computed basis will be the same */
6373: PetscArrayzero(correlation_mat,temp_constraints*temp_constraints);
6374: /* Store upper triangular part of correlation matrix */
6375: PetscBLASIntCast(size_of_constraint,&Blas_N);
6376: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6377: for (j=0;j<temp_constraints;j++) {
6378: for (k=0;k<j+1;k++) {
6379: PetscStackCallBLAS("BLASdot",correlation_mat[j*temp_constraints+k] = BLASdot_(&Blas_N,ptr_to_data+k*size_of_constraint,&Blas_one,ptr_to_data+j*size_of_constraint,&Blas_one));
6380: }
6381: }
6382: /* compute eigenvalues and eigenvectors of correlation matrix */
6383: PetscBLASIntCast(temp_constraints,&Blas_N);
6384: PetscBLASIntCast(temp_constraints,&Blas_LDA);
6385: #if !defined(PETSC_USE_COMPLEX)
6386: PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,&lierr));
6387: #else
6388: PetscStackCallBLAS("LAPACKsyev",LAPACKsyev_("V","U",&Blas_N,correlation_mat,&Blas_LDA,singular_vals,work,&lwork,rwork,&lierr));
6389: #endif
6390: PetscFPTrapPop();
6391: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in SYEV Lapack routine %d",(int)lierr);
6392: /* retain eigenvalues greater than tol: note that LAPACKsyev gives eigs in ascending order */
6393: j = 0;
6394: while (j < temp_constraints && singular_vals[j]/singular_vals[temp_constraints-1] < tol) j++;
6395: total_counts = total_counts-j;
6396: valid_constraints = temp_constraints-j;
6397: /* scale and copy POD basis into used quadrature memory */
6398: PetscBLASIntCast(size_of_constraint,&Blas_M);
6399: PetscBLASIntCast(temp_constraints,&Blas_N);
6400: PetscBLASIntCast(temp_constraints,&Blas_K);
6401: PetscBLASIntCast(size_of_constraint,&Blas_LDA);
6402: PetscBLASIntCast(temp_constraints,&Blas_LDB);
6403: PetscBLASIntCast(size_of_constraint,&Blas_LDC);
6404: if (j<temp_constraints) {
6405: PetscInt ii;
6406: for (k=j;k<temp_constraints;k++) singular_vals[k] = 1.0/PetscSqrtReal(singular_vals[k]);
6407: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6408: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,ptr_to_data,&Blas_LDA,correlation_mat,&Blas_LDB,&zero,temp_basis,&Blas_LDC));
6409: PetscFPTrapPop();
6410: for (k=0;k<temp_constraints-j;k++) {
6411: for (ii=0;ii<size_of_constraint;ii++) {
6412: ptr_to_data[k*size_of_constraint+ii] = singular_vals[temp_constraints-1-k]*temp_basis[(temp_constraints-1-k)*size_of_constraint+ii];
6413: }
6414: }
6415: }
6416: } else {
6417: #if !defined(PETSC_MISSING_LAPACK_GESVD)
6418: PetscBLASIntCast(size_of_constraint,&Blas_M);
6419: PetscBLASIntCast(temp_constraints,&Blas_N);
6420: PetscBLASIntCast(size_of_constraint,&Blas_LDA);
6421: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6422: #if !defined(PETSC_USE_COMPLEX)
6423: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,&lierr));
6424: #else
6425: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&Blas_M,&Blas_N,ptr_to_data,&Blas_LDA,singular_vals,&dummy_scalar,&dummy_int,&dummy_scalar,&dummy_int,work,&lwork,rwork,&lierr));
6426: #endif
6427: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GESVD Lapack routine %d",(int)lierr);
6428: PetscFPTrapPop();
6429: /* retain eigenvalues greater than tol: note that LAPACKgesvd gives eigs in descending order */
6430: k = temp_constraints;
6431: if (k > size_of_constraint) k = size_of_constraint;
6432: j = 0;
6433: while (j < k && singular_vals[k-j-1]/singular_vals[0] < tol) j++;
6434: valid_constraints = k-j;
6435: total_counts = total_counts-temp_constraints+valid_constraints;
6436: #else
6437: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_LIB,"This should not happen");
6438: #endif /* on missing GESVD */
6439: }
6440: }
6441: }
6442: /* update pointers information */
6443: if (valid_constraints) {
6444: constraints_n[total_counts_cc] = valid_constraints;
6445: constraints_idxs_ptr[total_counts_cc+1] = constraints_idxs_ptr[total_counts_cc]+size_of_constraint;
6446: constraints_data_ptr[total_counts_cc+1] = constraints_data_ptr[total_counts_cc]+size_of_constraint*valid_constraints;
6447: /* set change_of_basis flag */
6448: if (boolforchange) {
6449: PetscBTSet(change_basis,total_counts_cc);
6450: }
6451: total_counts_cc++;
6452: }
6453: }
6454: /* free workspace */
6455: if (!skip_lapack) {
6456: PetscFree(work);
6457: #if defined(PETSC_USE_COMPLEX)
6458: PetscFree(rwork);
6459: #endif
6460: PetscFree(singular_vals);
6461: PetscFree(correlation_mat);
6462: PetscFree(temp_basis);
6463: }
6464: for (k=0;k<nnsp_size;k++) {
6465: VecDestroy(&localnearnullsp[k]);
6466: }
6467: PetscFree(localnearnullsp);
6468: /* free index sets of faces, edges and vertices */
6469: for (i=0;i<n_ISForFaces;i++) {
6470: ISDestroy(&ISForFaces[i]);
6471: }
6472: if (n_ISForFaces) {
6473: PetscFree(ISForFaces);
6474: }
6475: for (i=0;i<n_ISForEdges;i++) {
6476: ISDestroy(&ISForEdges[i]);
6477: }
6478: if (n_ISForEdges) {
6479: PetscFree(ISForEdges);
6480: }
6481: ISDestroy(&ISForVertices);
6482: } else {
6483: PCBDDCSubSchurs sub_schurs = pcbddc->sub_schurs;
6485: total_counts = 0;
6486: n_vertices = 0;
6487: if (sub_schurs->is_vertices && pcbddc->use_vertices) {
6488: ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);
6489: }
6490: max_constraints = 0;
6491: total_counts_cc = 0;
6492: for (i=0;i<sub_schurs->n_subs+n_vertices;i++) {
6493: total_counts += pcbddc->adaptive_constraints_n[i];
6494: if (pcbddc->adaptive_constraints_n[i]) total_counts_cc++;
6495: max_constraints = PetscMax(max_constraints,pcbddc->adaptive_constraints_n[i]);
6496: }
6497: constraints_idxs_ptr = pcbddc->adaptive_constraints_idxs_ptr;
6498: constraints_data_ptr = pcbddc->adaptive_constraints_data_ptr;
6499: constraints_idxs = pcbddc->adaptive_constraints_idxs;
6500: constraints_data = pcbddc->adaptive_constraints_data;
6501: /* constraints_n differs from pcbddc->adaptive_constraints_n */
6502: PetscMalloc1(total_counts_cc,&constraints_n);
6503: total_counts_cc = 0;
6504: for (i=0;i<sub_schurs->n_subs+n_vertices;i++) {
6505: if (pcbddc->adaptive_constraints_n[i]) {
6506: constraints_n[total_counts_cc++] = pcbddc->adaptive_constraints_n[i];
6507: }
6508: }
6510: max_size_of_constraint = 0;
6511: for (i=0;i<total_counts_cc;i++) max_size_of_constraint = PetscMax(max_size_of_constraint,constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i]);
6512: PetscMalloc1(constraints_idxs_ptr[total_counts_cc],&constraints_idxs_B);
6513: /* Change of basis */
6514: PetscBTCreate(total_counts_cc,&change_basis);
6515: if (pcbddc->use_change_of_basis) {
6516: for (i=0;i<sub_schurs->n_subs;i++) {
6517: if (PetscBTLookup(sub_schurs->is_edge,i) || pcbddc->use_change_on_faces) {
6518: PetscBTSet(change_basis,i+n_vertices);
6519: }
6520: }
6521: }
6522: }
6523: pcbddc->local_primal_size = total_counts;
6524: PetscMalloc1(pcbddc->local_primal_size+pcbddc->benign_n,&pcbddc->primal_indices_local_idxs);
6526: /* map constraints_idxs in boundary numbering */
6527: ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,constraints_idxs_ptr[total_counts_cc],constraints_idxs,&i,constraints_idxs_B);
6528: if (i != constraints_idxs_ptr[total_counts_cc]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for constraints indices %D != %D",constraints_idxs_ptr[total_counts_cc],i);
6530: /* Create constraint matrix */
6531: MatCreate(PETSC_COMM_SELF,&pcbddc->ConstraintMatrix);
6532: MatSetType(pcbddc->ConstraintMatrix,MATAIJ);
6533: MatSetSizes(pcbddc->ConstraintMatrix,pcbddc->local_primal_size,pcis->n,pcbddc->local_primal_size,pcis->n);
6535: /* find primal_dofs: subdomain corners plus dofs selected as primal after change of basis */
6536: /* determine if a QR strategy is needed for change of basis */
6537: qr_needed = pcbddc->use_qr_single;
6538: PetscBTCreate(total_counts_cc,&qr_needed_idx);
6539: total_primal_vertices=0;
6540: pcbddc->local_primal_size_cc = 0;
6541: for (i=0;i<total_counts_cc;i++) {
6542: size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
6543: if (size_of_constraint == 1 && pcbddc->mat_graph->custom_minimal_size) {
6544: pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]];
6545: pcbddc->local_primal_size_cc += 1;
6546: } else if (PetscBTLookup(change_basis,i)) {
6547: for (k=0;k<constraints_n[i];k++) {
6548: pcbddc->primal_indices_local_idxs[total_primal_vertices++] = constraints_idxs[constraints_idxs_ptr[i]+k];
6549: }
6550: pcbddc->local_primal_size_cc += constraints_n[i];
6551: if (constraints_n[i] > 1 || pcbddc->use_qr_single) {
6552: PetscBTSet(qr_needed_idx,i);
6553: qr_needed = PETSC_TRUE;
6554: }
6555: } else {
6556: pcbddc->local_primal_size_cc += 1;
6557: }
6558: }
6559: /* note that the local variable n_vertices used below stores the number of pointwise constraints */
6560: pcbddc->n_vertices = total_primal_vertices;
6561: /* permute indices in order to have a sorted set of vertices */
6562: PetscSortInt(total_primal_vertices,pcbddc->primal_indices_local_idxs);
6563: PetscMalloc2(pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_node,pcbddc->local_primal_size_cc+pcbddc->benign_n,&pcbddc->local_primal_ref_mult);
6564: PetscArraycpy(pcbddc->local_primal_ref_node,pcbddc->primal_indices_local_idxs,total_primal_vertices);
6565: for (i=0;i<total_primal_vertices;i++) pcbddc->local_primal_ref_mult[i] = 1;
6567: /* nonzero structure of constraint matrix */
6568: /* and get reference dof for local constraints */
6569: PetscMalloc1(pcbddc->local_primal_size,&nnz);
6570: for (i=0;i<total_primal_vertices;i++) nnz[i] = 1;
6572: j = total_primal_vertices;
6573: total_counts = total_primal_vertices;
6574: cum = total_primal_vertices;
6575: for (i=n_vertices;i<total_counts_cc;i++) {
6576: if (!PetscBTLookup(change_basis,i)) {
6577: pcbddc->local_primal_ref_node[cum] = constraints_idxs[constraints_idxs_ptr[i]];
6578: pcbddc->local_primal_ref_mult[cum] = constraints_n[i];
6579: cum++;
6580: size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
6581: for (k=0;k<constraints_n[i];k++) {
6582: pcbddc->primal_indices_local_idxs[total_counts++] = constraints_idxs[constraints_idxs_ptr[i]+k];
6583: nnz[j+k] = size_of_constraint;
6584: }
6585: j += constraints_n[i];
6586: }
6587: }
6588: MatSeqAIJSetPreallocation(pcbddc->ConstraintMatrix,0,nnz);
6589: MatSetOption(pcbddc->ConstraintMatrix,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
6590: PetscFree(nnz);
6592: /* set values in constraint matrix */
6593: for (i=0;i<total_primal_vertices;i++) {
6594: MatSetValue(pcbddc->ConstraintMatrix,i,pcbddc->local_primal_ref_node[i],1.0,INSERT_VALUES);
6595: }
6596: total_counts = total_primal_vertices;
6597: for (i=n_vertices;i<total_counts_cc;i++) {
6598: if (!PetscBTLookup(change_basis,i)) {
6599: PetscInt *cols;
6601: size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
6602: cols = constraints_idxs+constraints_idxs_ptr[i];
6603: for (k=0;k<constraints_n[i];k++) {
6604: PetscInt row = total_counts+k;
6605: PetscScalar *vals;
6607: vals = constraints_data+constraints_data_ptr[i]+k*size_of_constraint;
6608: MatSetValues(pcbddc->ConstraintMatrix,1,&row,size_of_constraint,cols,vals,INSERT_VALUES);
6609: }
6610: total_counts += constraints_n[i];
6611: }
6612: }
6613: /* assembling */
6614: MatAssemblyBegin(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);
6615: MatAssemblyEnd(pcbddc->ConstraintMatrix,MAT_FINAL_ASSEMBLY);
6616: MatViewFromOptions(pcbddc->ConstraintMatrix,(PetscObject)pc,"-pc_bddc_constraint_mat_view");
6618: /* Create matrix for change of basis. We don't need it in case pcbddc->use_change_of_basis is FALSE */
6619: if (pcbddc->use_change_of_basis) {
6620: /* dual and primal dofs on a single cc */
6621: PetscInt dual_dofs,primal_dofs;
6622: /* working stuff for GEQRF */
6623: PetscScalar *qr_basis = NULL,*qr_tau = NULL,*qr_work = NULL,lqr_work_t;
6624: PetscBLASInt lqr_work;
6625: /* working stuff for UNGQR */
6626: PetscScalar *gqr_work = NULL,lgqr_work_t=0.0;
6627: PetscBLASInt lgqr_work;
6628: /* working stuff for TRTRS */
6629: PetscScalar *trs_rhs = NULL;
6630: PetscBLASInt Blas_NRHS;
6631: /* pointers for values insertion into change of basis matrix */
6632: PetscInt *start_rows,*start_cols;
6633: PetscScalar *start_vals;
6634: /* working stuff for values insertion */
6635: PetscBT is_primal;
6636: PetscInt *aux_primal_numbering_B;
6637: /* matrix sizes */
6638: PetscInt global_size,local_size;
6639: /* temporary change of basis */
6640: Mat localChangeOfBasisMatrix;
6641: /* extra space for debugging */
6642: PetscScalar *dbg_work = NULL;
6644: /* local temporary change of basis acts on local interfaces -> dimension is n_B x n_B */
6645: MatCreate(PETSC_COMM_SELF,&localChangeOfBasisMatrix);
6646: MatSetType(localChangeOfBasisMatrix,MATAIJ);
6647: MatSetSizes(localChangeOfBasisMatrix,pcis->n,pcis->n,pcis->n,pcis->n);
6648: /* nonzeros for local mat */
6649: PetscMalloc1(pcis->n,&nnz);
6650: if (!pcbddc->benign_change || pcbddc->fake_change) {
6651: for (i=0;i<pcis->n;i++) nnz[i]=1;
6652: } else {
6653: const PetscInt *ii;
6654: PetscInt n;
6655: PetscBool flg_row;
6656: MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);
6657: for (i=0;i<n;i++) nnz[i] = ii[i+1]-ii[i];
6658: MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,NULL,&flg_row);
6659: }
6660: for (i=n_vertices;i<total_counts_cc;i++) {
6661: if (PetscBTLookup(change_basis,i)) {
6662: size_of_constraint = constraints_idxs_ptr[i+1]-constraints_idxs_ptr[i];
6663: if (PetscBTLookup(qr_needed_idx,i)) {
6664: for (j=0;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = size_of_constraint;
6665: } else {
6666: nnz[constraints_idxs[constraints_idxs_ptr[i]]] = size_of_constraint;
6667: for (j=1;j<size_of_constraint;j++) nnz[constraints_idxs[constraints_idxs_ptr[i]+j]] = 2;
6668: }
6669: }
6670: }
6671: MatSeqAIJSetPreallocation(localChangeOfBasisMatrix,0,nnz);
6672: MatSetOption(localChangeOfBasisMatrix,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
6673: PetscFree(nnz);
6674: /* Set interior change in the matrix */
6675: if (!pcbddc->benign_change || pcbddc->fake_change) {
6676: for (i=0;i<pcis->n;i++) {
6677: MatSetValue(localChangeOfBasisMatrix,i,i,1.0,INSERT_VALUES);
6678: }
6679: } else {
6680: const PetscInt *ii,*jj;
6681: PetscScalar *aa;
6682: PetscInt n;
6683: PetscBool flg_row;
6684: MatGetRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);
6685: MatSeqAIJGetArray(pcbddc->benign_change,&aa);
6686: for (i=0;i<n;i++) {
6687: MatSetValues(localChangeOfBasisMatrix,1,&i,ii[i+1]-ii[i],jj+ii[i],aa+ii[i],INSERT_VALUES);
6688: }
6689: MatSeqAIJRestoreArray(pcbddc->benign_change,&aa);
6690: MatRestoreRowIJ(pcbddc->benign_change,0,PETSC_FALSE,PETSC_FALSE,&n,&ii,&jj,&flg_row);
6691: }
6693: if (pcbddc->dbg_flag) {
6694: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");
6695: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Checking change of basis computation for subdomain %04d\n",PetscGlobalRank);
6696: }
6699: /* Now we loop on the constraints which need a change of basis */
6700: /*
6701: Change of basis matrix is evaluated similarly to the FIRST APPROACH in
6702: Klawonn and Widlund, Dual-primal FETI-DP methods for linear elasticity, (see Sect 6.2.1)
6704: Basic blocks of change of basis matrix T computed by
6706: - Using the following block transformation if there is only a primal dof on the cc (and -pc_bddc_use_qr_single is not specified)
6708: | 1 0 ... 0 s_1/S |
6709: | 0 1 ... 0 s_2/S |
6710: | ... |
6711: | 0 ... 1 s_{n-1}/S |
6712: | -s_1/s_n ... -s_{n-1}/s_n s_n/S |
6714: with S = \sum_{i=1}^n s_i^2
6715: NOTE: in the above example, the primal dof is the last one of the edge in LOCAL ordering
6716: in the current implementation, the primal dof is the first one of the edge in GLOBAL ordering
6718: - QR decomposition of constraints otherwise
6719: */
6720: if (qr_needed && max_size_of_constraint) {
6721: /* space to store Q */
6722: PetscMalloc1(max_size_of_constraint*max_size_of_constraint,&qr_basis);
6723: /* array to store scaling factors for reflectors */
6724: PetscMalloc1(max_constraints,&qr_tau);
6725: /* first we issue queries for optimal work */
6726: PetscBLASIntCast(max_size_of_constraint,&Blas_M);
6727: PetscBLASIntCast(max_constraints,&Blas_N);
6728: PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);
6729: lqr_work = -1;
6730: PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,&lqr_work_t,&lqr_work,&lierr));
6731: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to GEQRF Lapack routine %d",(int)lierr);
6732: PetscBLASIntCast((PetscInt)PetscRealPart(lqr_work_t),&lqr_work);
6733: PetscMalloc1((PetscInt)PetscRealPart(lqr_work_t),&qr_work);
6734: lgqr_work = -1;
6735: PetscBLASIntCast(max_size_of_constraint,&Blas_M);
6736: PetscBLASIntCast(max_size_of_constraint,&Blas_N);
6737: PetscBLASIntCast(max_constraints,&Blas_K);
6738: PetscBLASIntCast(max_size_of_constraint,&Blas_LDA);
6739: if (Blas_K>Blas_M) Blas_K=Blas_M; /* adjust just for computing optimal work */
6740: PetscStackCallBLAS("LAPACKorgqr",LAPACKorgqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,&lgqr_work_t,&lgqr_work,&lierr));
6741: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in query to ORGQR/UNGQR Lapack routine %d",(int)lierr);
6742: PetscBLASIntCast((PetscInt)PetscRealPart(lgqr_work_t),&lgqr_work);
6743: PetscMalloc1((PetscInt)PetscRealPart(lgqr_work_t),&gqr_work);
6744: /* array to store rhs and solution of triangular solver */
6745: PetscMalloc1(max_constraints*max_constraints,&trs_rhs);
6746: /* allocating workspace for check */
6747: if (pcbddc->dbg_flag) {
6748: PetscMalloc1(max_size_of_constraint*(max_constraints+max_size_of_constraint),&dbg_work);
6749: }
6750: }
6751: /* array to store whether a node is primal or not */
6752: PetscBTCreate(pcis->n_B,&is_primal);
6753: PetscMalloc1(total_primal_vertices,&aux_primal_numbering_B);
6754: ISGlobalToLocalMappingApply(pcis->BtoNmap,IS_GTOLM_DROP,total_primal_vertices,pcbddc->local_primal_ref_node,&i,aux_primal_numbering_B);
6755: if (i != total_primal_vertices) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error in boundary numbering for BDDC vertices! %D != %D",total_primal_vertices,i);
6756: for (i=0;i<total_primal_vertices;i++) {
6757: PetscBTSet(is_primal,aux_primal_numbering_B[i]);
6758: }
6759: PetscFree(aux_primal_numbering_B);
6761: /* loop on constraints and see whether or not they need a change of basis and compute it */
6762: for (total_counts=n_vertices;total_counts<total_counts_cc;total_counts++) {
6763: size_of_constraint = constraints_idxs_ptr[total_counts+1]-constraints_idxs_ptr[total_counts];
6764: if (PetscBTLookup(change_basis,total_counts)) {
6765: /* get constraint info */
6766: primal_dofs = constraints_n[total_counts];
6767: dual_dofs = size_of_constraint-primal_dofs;
6769: if (pcbddc->dbg_flag) {
6770: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraints %D: %D need a change of basis (size %D)\n",total_counts,primal_dofs,size_of_constraint);
6771: }
6773: if (PetscBTLookup(qr_needed_idx,total_counts)) { /* QR */
6775: /* copy quadrature constraints for change of basis check */
6776: if (pcbddc->dbg_flag) {
6777: PetscArraycpy(dbg_work,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs);
6778: }
6779: /* copy temporary constraints into larger work vector (in order to store all columns of Q) */
6780: PetscArraycpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs);
6782: /* compute QR decomposition of constraints */
6783: PetscBLASIntCast(size_of_constraint,&Blas_M);
6784: PetscBLASIntCast(primal_dofs,&Blas_N);
6785: PetscBLASIntCast(size_of_constraint,&Blas_LDA);
6786: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6787: PetscStackCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&Blas_M,&Blas_N,qr_basis,&Blas_LDA,qr_tau,qr_work,&lqr_work,&lierr));
6788: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in GEQRF Lapack routine %d",(int)lierr);
6789: PetscFPTrapPop();
6791: /* explictly compute R^-T */
6792: PetscArrayzero(trs_rhs,primal_dofs*primal_dofs);
6793: for (j=0;j<primal_dofs;j++) trs_rhs[j*(primal_dofs+1)] = 1.0;
6794: PetscBLASIntCast(primal_dofs,&Blas_N);
6795: PetscBLASIntCast(primal_dofs,&Blas_NRHS);
6796: PetscBLASIntCast(size_of_constraint,&Blas_LDA);
6797: PetscBLASIntCast(primal_dofs,&Blas_LDB);
6798: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6799: PetscStackCallBLAS("LAPACKtrtrs",LAPACKtrtrs_("U","T","N",&Blas_N,&Blas_NRHS,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&lierr));
6800: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in TRTRS Lapack routine %d",(int)lierr);
6801: PetscFPTrapPop();
6803: /* explicitly compute all columns of Q (Q = [Q1 | Q2] ) overwriting QR factorization in qr_basis */
6804: PetscBLASIntCast(size_of_constraint,&Blas_M);
6805: PetscBLASIntCast(size_of_constraint,&Blas_N);
6806: PetscBLASIntCast(primal_dofs,&Blas_K);
6807: PetscBLASIntCast(size_of_constraint,&Blas_LDA);
6808: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6809: PetscStackCallBLAS("LAPACKorgqr",LAPACKorgqr_(&Blas_M,&Blas_N,&Blas_K,qr_basis,&Blas_LDA,qr_tau,gqr_work,&lgqr_work,&lierr));
6810: if (lierr) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Error in ORGQR/UNGQR Lapack routine %d",(int)lierr);
6811: PetscFPTrapPop();
6813: /* first primal_dofs columns of Q need to be re-scaled in order to be unitary w.r.t constraints
6814: i.e. C_{pxn}*Q_{nxn} should be equal to [I_pxp | 0_pxd] (see check below)
6815: where n=size_of_constraint, p=primal_dofs, d=dual_dofs (n=p+d), I and 0 identity and null matrix resp. */
6816: PetscBLASIntCast(size_of_constraint,&Blas_M);
6817: PetscBLASIntCast(primal_dofs,&Blas_N);
6818: PetscBLASIntCast(primal_dofs,&Blas_K);
6819: PetscBLASIntCast(size_of_constraint,&Blas_LDA);
6820: PetscBLASIntCast(primal_dofs,&Blas_LDB);
6821: PetscBLASIntCast(size_of_constraint,&Blas_LDC);
6822: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6823: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&Blas_M,&Blas_N,&Blas_K,&one,qr_basis,&Blas_LDA,trs_rhs,&Blas_LDB,&zero,constraints_data+constraints_data_ptr[total_counts],&Blas_LDC));
6824: PetscFPTrapPop();
6825: PetscArraycpy(qr_basis,&constraints_data[constraints_data_ptr[total_counts]],size_of_constraint*primal_dofs);
6827: /* insert values in change of basis matrix respecting global ordering of new primal dofs */
6828: start_rows = &constraints_idxs[constraints_idxs_ptr[total_counts]];
6829: /* insert cols for primal dofs */
6830: for (j=0;j<primal_dofs;j++) {
6831: start_vals = &qr_basis[j*size_of_constraint];
6832: start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+j];
6833: MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);
6834: }
6835: /* insert cols for dual dofs */
6836: for (j=0,k=0;j<dual_dofs;k++) {
6837: if (!PetscBTLookup(is_primal,constraints_idxs_B[constraints_idxs_ptr[total_counts]+k])) {
6838: start_vals = &qr_basis[(primal_dofs+j)*size_of_constraint];
6839: start_cols = &constraints_idxs[constraints_idxs_ptr[total_counts]+k];
6840: MatSetValues(localChangeOfBasisMatrix,size_of_constraint,start_rows,1,start_cols,start_vals,INSERT_VALUES);
6841: j++;
6842: }
6843: }
6845: /* check change of basis */
6846: if (pcbddc->dbg_flag) {
6847: PetscInt ii,jj;
6848: PetscBool valid_qr=PETSC_TRUE;
6849: PetscBLASIntCast(primal_dofs,&Blas_M);
6850: PetscBLASIntCast(size_of_constraint,&Blas_N);
6851: PetscBLASIntCast(size_of_constraint,&Blas_K);
6852: PetscBLASIntCast(size_of_constraint,&Blas_LDA);
6853: PetscBLASIntCast(size_of_constraint,&Blas_LDB);
6854: PetscBLASIntCast(primal_dofs,&Blas_LDC);
6855: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
6856: PetscStackCallBLAS("BLASgemm",BLASgemm_("T","N",&Blas_M,&Blas_N,&Blas_K,&one,dbg_work,&Blas_LDA,qr_basis,&Blas_LDB,&zero,&dbg_work[size_of_constraint*primal_dofs],&Blas_LDC));
6857: PetscFPTrapPop();
6858: for (jj=0;jj<size_of_constraint;jj++) {
6859: for (ii=0;ii<primal_dofs;ii++) {
6860: if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) valid_qr = PETSC_FALSE;
6861: if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-(PetscReal)1) > 1.e-12) valid_qr = PETSC_FALSE;
6862: }
6863: }
6864: if (!valid_qr) {
6865: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> wrong change of basis!\n");
6866: for (jj=0;jj<size_of_constraint;jj++) {
6867: for (ii=0;ii<primal_dofs;ii++) {
6868: if (ii != jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]) > 1.e-12) {
6869: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %D is not orthogonal to constraint %D (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]));
6870: }
6871: if (ii == jj && PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]-(PetscReal)1) > 1.e-12) {
6872: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\tQr basis function %D is not unitary w.r.t constraint %D (%1.14e)!\n",jj,ii,PetscAbsScalar(dbg_work[size_of_constraint*primal_dofs+jj*primal_dofs+ii]));
6873: }
6874: }
6875: }
6876: } else {
6877: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> right change of basis!\n");
6878: }
6879: }
6880: } else { /* simple transformation block */
6881: PetscInt row,col;
6882: PetscScalar val,norm;
6884: PetscBLASIntCast(size_of_constraint,&Blas_N);
6885: PetscStackCallBLAS("BLASdot",norm = BLASdot_(&Blas_N,constraints_data+constraints_data_ptr[total_counts],&Blas_one,constraints_data+constraints_data_ptr[total_counts],&Blas_one));
6886: for (j=0;j<size_of_constraint;j++) {
6887: PetscInt row_B = constraints_idxs_B[constraints_idxs_ptr[total_counts]+j];
6888: row = constraints_idxs[constraints_idxs_ptr[total_counts]+j];
6889: if (!PetscBTLookup(is_primal,row_B)) {
6890: col = constraints_idxs[constraints_idxs_ptr[total_counts]];
6891: MatSetValue(localChangeOfBasisMatrix,row,row,1.0,INSERT_VALUES);
6892: MatSetValue(localChangeOfBasisMatrix,row,col,constraints_data[constraints_data_ptr[total_counts]+j]/norm,INSERT_VALUES);
6893: } else {
6894: for (k=0;k<size_of_constraint;k++) {
6895: col = constraints_idxs[constraints_idxs_ptr[total_counts]+k];
6896: if (row != col) {
6897: val = -constraints_data[constraints_data_ptr[total_counts]+k]/constraints_data[constraints_data_ptr[total_counts]];
6898: } else {
6899: val = constraints_data[constraints_data_ptr[total_counts]]/norm;
6900: }
6901: MatSetValue(localChangeOfBasisMatrix,row,col,val,INSERT_VALUES);
6902: }
6903: }
6904: }
6905: if (pcbddc->dbg_flag) {
6906: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"\t-> using standard change of basis\n");
6907: }
6908: }
6909: } else {
6910: if (pcbddc->dbg_flag) {
6911: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Constraint %D does not need a change of basis (size %D)\n",total_counts,size_of_constraint);
6912: }
6913: }
6914: }
6916: /* free workspace */
6917: if (qr_needed) {
6918: if (pcbddc->dbg_flag) {
6919: PetscFree(dbg_work);
6920: }
6921: PetscFree(trs_rhs);
6922: PetscFree(qr_tau);
6923: PetscFree(qr_work);
6924: PetscFree(gqr_work);
6925: PetscFree(qr_basis);
6926: }
6927: PetscBTDestroy(&is_primal);
6928: MatAssemblyBegin(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);
6929: MatAssemblyEnd(localChangeOfBasisMatrix,MAT_FINAL_ASSEMBLY);
6931: /* assembling of global change of variable */
6932: if (!pcbddc->fake_change) {
6933: Mat tmat;
6934: PetscInt bs;
6936: VecGetSize(pcis->vec1_global,&global_size);
6937: VecGetLocalSize(pcis->vec1_global,&local_size);
6938: MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&tmat);
6939: MatISSetLocalMat(tmat,localChangeOfBasisMatrix);
6940: MatAssemblyBegin(tmat,MAT_FINAL_ASSEMBLY);
6941: MatAssemblyEnd(tmat,MAT_FINAL_ASSEMBLY);
6942: MatCreate(PetscObjectComm((PetscObject)pc),&pcbddc->ChangeOfBasisMatrix);
6943: MatSetType(pcbddc->ChangeOfBasisMatrix,MATAIJ);
6944: MatGetBlockSize(pc->pmat,&bs);
6945: MatSetBlockSize(pcbddc->ChangeOfBasisMatrix,bs);
6946: MatSetSizes(pcbddc->ChangeOfBasisMatrix,local_size,local_size,global_size,global_size);
6947: MatISSetMPIXAIJPreallocation_Private(tmat,pcbddc->ChangeOfBasisMatrix,PETSC_TRUE);
6948: MatConvert(tmat,MATAIJ,MAT_REUSE_MATRIX,&pcbddc->ChangeOfBasisMatrix);
6949: MatDestroy(&tmat);
6950: VecSet(pcis->vec1_global,0.0);
6951: VecSet(pcis->vec1_N,1.0);
6952: VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
6953: VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
6954: VecReciprocal(pcis->vec1_global);
6955: MatDiagonalScale(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,NULL);
6957: /* check */
6958: if (pcbddc->dbg_flag) {
6959: PetscReal error;
6960: Vec x,x_change;
6962: VecDuplicate(pcis->vec1_global,&x);
6963: VecDuplicate(pcis->vec1_global,&x_change);
6964: VecSetRandom(x,NULL);
6965: VecCopy(x,pcis->vec1_global);
6966: VecScatterBegin(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);
6967: VecScatterEnd(matis->rctx,x,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);
6968: MatMult(localChangeOfBasisMatrix,pcis->vec1_N,pcis->vec2_N);
6969: VecScatterBegin(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);
6970: VecScatterEnd(matis->rctx,pcis->vec2_N,x,INSERT_VALUES,SCATTER_REVERSE);
6971: MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_global,x_change);
6972: VecAXPY(x,-1.0,x_change);
6973: VecNorm(x,NORM_INFINITY,&error);
6974: if (error > PETSC_SMALL) {
6975: SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"Error global vs local change on N: %1.6e",error);
6976: }
6977: VecDestroy(&x);
6978: VecDestroy(&x_change);
6979: }
6980: /* adapt sub_schurs computed (if any) */
6981: if (pcbddc->use_deluxe_scaling) {
6982: PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs;
6984: if (pcbddc->use_change_of_basis && pcbddc->adaptive_userdefined) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_SUP,"Cannot mix automatic change of basis, adaptive selection and user-defined constraints");
6985: if (sub_schurs && sub_schurs->S_Ej_all) {
6986: Mat S_new,tmat;
6987: IS is_all_N,is_V_Sall = NULL;
6989: ISLocalToGlobalMappingApplyIS(pcis->BtoNmap,sub_schurs->is_Ej_all,&is_all_N);
6990: MatCreateSubMatrix(localChangeOfBasisMatrix,is_all_N,is_all_N,MAT_INITIAL_MATRIX,&tmat);
6991: if (pcbddc->deluxe_zerorows) {
6992: ISLocalToGlobalMapping NtoSall;
6993: IS is_V;
6994: ISCreateGeneral(PETSC_COMM_SELF,pcbddc->n_vertices,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&is_V);
6995: ISLocalToGlobalMappingCreateIS(is_all_N,&NtoSall);
6996: ISGlobalToLocalMappingApplyIS(NtoSall,IS_GTOLM_DROP,is_V,&is_V_Sall);
6997: ISLocalToGlobalMappingDestroy(&NtoSall);
6998: ISDestroy(&is_V);
6999: }
7000: ISDestroy(&is_all_N);
7001: MatPtAP(sub_schurs->S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);
7002: MatDestroy(&sub_schurs->S_Ej_all);
7003: PetscObjectReference((PetscObject)S_new);
7004: if (pcbddc->deluxe_zerorows) {
7005: const PetscScalar *array;
7006: const PetscInt *idxs_V,*idxs_all;
7007: PetscInt i,n_V;
7009: MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);
7010: ISGetLocalSize(is_V_Sall,&n_V);
7011: ISGetIndices(is_V_Sall,&idxs_V);
7012: ISGetIndices(sub_schurs->is_Ej_all,&idxs_all);
7013: VecGetArrayRead(pcis->D,&array);
7014: for (i=0;i<n_V;i++) {
7015: PetscScalar val;
7016: PetscInt idx;
7018: idx = idxs_V[i];
7019: val = array[idxs_all[idxs_V[i]]];
7020: MatSetValue(S_new,idx,idx,val,INSERT_VALUES);
7021: }
7022: MatAssemblyBegin(S_new,MAT_FINAL_ASSEMBLY);
7023: MatAssemblyEnd(S_new,MAT_FINAL_ASSEMBLY);
7024: VecRestoreArrayRead(pcis->D,&array);
7025: ISRestoreIndices(sub_schurs->is_Ej_all,&idxs_all);
7026: ISRestoreIndices(is_V_Sall,&idxs_V);
7027: }
7028: sub_schurs->S_Ej_all = S_new;
7029: MatDestroy(&S_new);
7030: if (sub_schurs->sum_S_Ej_all) {
7031: MatPtAP(sub_schurs->sum_S_Ej_all,tmat,MAT_INITIAL_MATRIX,1.0,&S_new);
7032: MatDestroy(&sub_schurs->sum_S_Ej_all);
7033: PetscObjectReference((PetscObject)S_new);
7034: if (pcbddc->deluxe_zerorows) {
7035: MatZeroRowsColumnsIS(S_new,is_V_Sall,1.,NULL,NULL);
7036: }
7037: sub_schurs->sum_S_Ej_all = S_new;
7038: MatDestroy(&S_new);
7039: }
7040: ISDestroy(&is_V_Sall);
7041: MatDestroy(&tmat);
7042: }
7043: /* destroy any change of basis context in sub_schurs */
7044: if (sub_schurs && sub_schurs->change) {
7045: PetscInt i;
7047: for (i=0;i<sub_schurs->n_subs;i++) {
7048: KSPDestroy(&sub_schurs->change[i]);
7049: }
7050: PetscFree(sub_schurs->change);
7051: }
7052: }
7053: if (pcbddc->switch_static) { /* need to save the local change */
7054: pcbddc->switch_static_change = localChangeOfBasisMatrix;
7055: } else {
7056: MatDestroy(&localChangeOfBasisMatrix);
7057: }
7058: /* determine if any process has changed the pressures locally */
7059: pcbddc->change_interior = pcbddc->benign_have_null;
7060: } else { /* fake change (get back change of basis into ConstraintMatrix and info on qr) */
7061: MatDestroy(&pcbddc->ConstraintMatrix);
7062: pcbddc->ConstraintMatrix = localChangeOfBasisMatrix;
7063: pcbddc->use_qr_single = qr_needed;
7064: }
7065: } else if (pcbddc->user_ChangeOfBasisMatrix || pcbddc->benign_saddle_point) {
7066: if (!pcbddc->benign_have_null && pcbddc->user_ChangeOfBasisMatrix) {
7067: PetscObjectReference((PetscObject)pcbddc->user_ChangeOfBasisMatrix);
7068: pcbddc->ChangeOfBasisMatrix = pcbddc->user_ChangeOfBasisMatrix;
7069: } else {
7070: Mat benign_global = NULL;
7071: if (pcbddc->benign_have_null) {
7072: Mat M;
7074: pcbddc->change_interior = PETSC_TRUE;
7075: VecCopy(matis->counter,pcis->vec1_N);
7076: VecReciprocal(pcis->vec1_N);
7077: MatDuplicate(pc->pmat,MAT_DO_NOT_COPY_VALUES,&benign_global);
7078: if (pcbddc->benign_change) {
7079: MatDuplicate(pcbddc->benign_change,MAT_COPY_VALUES,&M);
7080: MatDiagonalScale(M,pcis->vec1_N,NULL);
7081: } else {
7082: MatCreateSeqAIJ(PETSC_COMM_SELF,pcis->n,pcis->n,1,NULL,&M);
7083: MatDiagonalSet(M,pcis->vec1_N,INSERT_VALUES);
7084: }
7085: MatISSetLocalMat(benign_global,M);
7086: MatDestroy(&M);
7087: MatAssemblyBegin(benign_global,MAT_FINAL_ASSEMBLY);
7088: MatAssemblyEnd(benign_global,MAT_FINAL_ASSEMBLY);
7089: }
7090: if (pcbddc->user_ChangeOfBasisMatrix) {
7091: MatMatMult(pcbddc->user_ChangeOfBasisMatrix,benign_global,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&pcbddc->ChangeOfBasisMatrix);
7092: MatDestroy(&benign_global);
7093: } else if (pcbddc->benign_have_null) {
7094: pcbddc->ChangeOfBasisMatrix = benign_global;
7095: }
7096: }
7097: if (pcbddc->switch_static && pcbddc->ChangeOfBasisMatrix) { /* need to save the local change */
7098: IS is_global;
7099: const PetscInt *gidxs;
7101: ISLocalToGlobalMappingGetIndices(pc->pmat->rmap->mapping,&gidxs);
7102: ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcis->n,gidxs,PETSC_COPY_VALUES,&is_global);
7103: ISLocalToGlobalMappingRestoreIndices(pc->pmat->rmap->mapping,&gidxs);
7104: MatCreateSubMatrixUnsorted(pcbddc->ChangeOfBasisMatrix,is_global,is_global,&pcbddc->switch_static_change);
7105: ISDestroy(&is_global);
7106: }
7107: }
7108: if (!pcbddc->fake_change && pcbddc->ChangeOfBasisMatrix && !pcbddc->work_change) {
7109: VecDuplicate(pcis->vec1_global,&pcbddc->work_change);
7110: }
7112: if (!pcbddc->fake_change) {
7113: /* add pressure dofs to set of primal nodes for numbering purposes */
7114: for (i=0;i<pcbddc->benign_n;i++) {
7115: pcbddc->local_primal_ref_node[pcbddc->local_primal_size_cc] = pcbddc->benign_p0_lidx[i];
7116: pcbddc->primal_indices_local_idxs[pcbddc->local_primal_size] = pcbddc->benign_p0_lidx[i];
7117: pcbddc->local_primal_ref_mult[pcbddc->local_primal_size_cc] = 1;
7118: pcbddc->local_primal_size_cc++;
7119: pcbddc->local_primal_size++;
7120: }
7122: /* check if a new primal space has been introduced (also take into account benign trick) */
7123: pcbddc->new_primal_space_local = PETSC_TRUE;
7124: if (olocal_primal_size == pcbddc->local_primal_size) {
7125: PetscArraycmp(pcbddc->local_primal_ref_node,olocal_primal_ref_node,olocal_primal_size_cc,&pcbddc->new_primal_space_local);
7126: pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local);
7127: if (!pcbddc->new_primal_space_local) {
7128: PetscArraycmp(pcbddc->local_primal_ref_mult,olocal_primal_ref_mult,olocal_primal_size_cc,&pcbddc->new_primal_space_local);
7129: pcbddc->new_primal_space_local = (PetscBool)(!pcbddc->new_primal_space_local);
7130: }
7131: }
7132: /* new_primal_space will be used for numbering of coarse dofs, so it should be the same across all subdomains */
7133: MPIU_Allreduce(&pcbddc->new_primal_space_local,&pcbddc->new_primal_space,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));
7134: }
7135: PetscFree2(olocal_primal_ref_node,olocal_primal_ref_mult);
7137: /* flush dbg viewer */
7138: if (pcbddc->dbg_flag) {
7139: PetscViewerFlush(pcbddc->dbg_viewer);
7140: }
7142: /* free workspace */
7143: PetscBTDestroy(&qr_needed_idx);
7144: PetscBTDestroy(&change_basis);
7145: if (!pcbddc->adaptive_selection) {
7146: PetscFree3(constraints_idxs_ptr,constraints_data_ptr,constraints_n);
7147: PetscFree3(constraints_data,constraints_idxs,constraints_idxs_B);
7148: } else {
7149: PetscFree5(pcbddc->adaptive_constraints_n,
7150: pcbddc->adaptive_constraints_idxs_ptr,
7151: pcbddc->adaptive_constraints_data_ptr,
7152: pcbddc->adaptive_constraints_idxs,
7153: pcbddc->adaptive_constraints_data);
7154: PetscFree(constraints_n);
7155: PetscFree(constraints_idxs_B);
7156: }
7157: return(0);
7158: }
7160: PetscErrorCode PCBDDCAnalyzeInterface(PC pc)
7161: {
7162: ISLocalToGlobalMapping map;
7163: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
7164: Mat_IS *matis = (Mat_IS*)pc->pmat->data;
7165: PetscInt i,N;
7166: PetscBool rcsr = PETSC_FALSE;
7167: PetscErrorCode ierr;
7170: if (pcbddc->recompute_topography) {
7171: pcbddc->graphanalyzed = PETSC_FALSE;
7172: /* Reset previously computed graph */
7173: PCBDDCGraphReset(pcbddc->mat_graph);
7174: /* Init local Graph struct */
7175: MatGetSize(pc->pmat,&N,NULL);
7176: MatGetLocalToGlobalMapping(pc->pmat,&map,NULL);
7177: PCBDDCGraphInit(pcbddc->mat_graph,map,N,pcbddc->graphmaxcount);
7179: if (pcbddc->user_primal_vertices_local && !pcbddc->user_primal_vertices) {
7180: PCBDDCConsistencyCheckIS(pc,MPI_LOR,&pcbddc->user_primal_vertices_local);
7181: }
7182: /* Check validity of the csr graph passed in by the user */
7183: if (pcbddc->mat_graph->nvtxs_csr && pcbddc->mat_graph->nvtxs_csr != pcbddc->mat_graph->nvtxs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid size of local CSR graph! Found %D, expected %D",pcbddc->mat_graph->nvtxs_csr,pcbddc->mat_graph->nvtxs);
7185: /* Set default CSR adjacency of local dofs if not provided by the user with PCBDDCSetLocalAdjacencyGraph */
7186: if (!pcbddc->mat_graph->xadj && pcbddc->use_local_adj) {
7187: PetscInt *xadj,*adjncy;
7188: PetscInt nvtxs;
7189: PetscBool flg_row=PETSC_FALSE;
7191: MatGetRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);
7192: if (flg_row) {
7193: PCBDDCSetLocalAdjacencyGraph(pc,nvtxs,xadj,adjncy,PETSC_COPY_VALUES);
7194: pcbddc->computed_rowadj = PETSC_TRUE;
7195: }
7196: MatRestoreRowIJ(matis->A,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&flg_row);
7197: rcsr = PETSC_TRUE;
7198: }
7199: if (pcbddc->dbg_flag) {
7200: PetscViewerFlush(pcbddc->dbg_viewer);
7201: }
7203: if (pcbddc->mat_graph->cdim && !pcbddc->mat_graph->cloc) {
7204: PetscReal *lcoords;
7205: PetscInt n;
7206: MPI_Datatype dimrealtype;
7208: /* TODO: support for blocked */
7209: if (pcbddc->mat_graph->cnloc != pc->pmat->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid number of local coordinates! Got %D, expected %D",pcbddc->mat_graph->cnloc,pc->pmat->rmap->n);
7210: MatGetLocalSize(matis->A,&n,NULL);
7211: PetscMalloc1(pcbddc->mat_graph->cdim*n,&lcoords);
7212: MPI_Type_contiguous(pcbddc->mat_graph->cdim,MPIU_REAL,&dimrealtype);
7213: MPI_Type_commit(&dimrealtype);
7214: PetscSFBcastBegin(matis->sf,dimrealtype,pcbddc->mat_graph->coords,lcoords);
7215: PetscSFBcastEnd(matis->sf,dimrealtype,pcbddc->mat_graph->coords,lcoords);
7216: MPI_Type_free(&dimrealtype);
7217: PetscFree(pcbddc->mat_graph->coords);
7219: pcbddc->mat_graph->coords = lcoords;
7220: pcbddc->mat_graph->cloc = PETSC_TRUE;
7221: pcbddc->mat_graph->cnloc = n;
7222: }
7223: if (pcbddc->mat_graph->cnloc && pcbddc->mat_graph->cnloc != pcbddc->mat_graph->nvtxs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_USER,"Invalid number of local subdomain coordinates! Got %D, expected %D",pcbddc->mat_graph->cnloc,pcbddc->mat_graph->nvtxs);
7224: pcbddc->mat_graph->active_coords = (PetscBool)(pcbddc->corner_selection && !pcbddc->corner_selected);
7226: /* Setup of Graph */
7227: pcbddc->mat_graph->commsizelimit = 0; /* don't use the COMM_SELF variant of the graph */
7228: PCBDDCGraphSetUp(pcbddc->mat_graph,pcbddc->vertex_size,pcbddc->NeumannBoundariesLocal,pcbddc->DirichletBoundariesLocal,pcbddc->n_ISForDofsLocal,pcbddc->ISForDofsLocal,pcbddc->user_primal_vertices_local);
7230: /* attach info on disconnected subdomains if present */
7231: if (pcbddc->n_local_subs) {
7232: PetscInt *local_subs,n,totn;
7234: MatGetLocalSize(matis->A,&n,NULL);
7235: PetscMalloc1(n,&local_subs);
7236: for (i=0;i<n;i++) local_subs[i] = pcbddc->n_local_subs;
7237: for (i=0;i<pcbddc->n_local_subs;i++) {
7238: const PetscInt *idxs;
7239: PetscInt nl,j;
7241: ISGetLocalSize(pcbddc->local_subs[i],&nl);
7242: ISGetIndices(pcbddc->local_subs[i],&idxs);
7243: for (j=0;j<nl;j++) local_subs[idxs[j]] = i;
7244: ISRestoreIndices(pcbddc->local_subs[i],&idxs);
7245: }
7246: for (i=0,totn=0;i<n;i++) totn = PetscMax(totn,local_subs[i]);
7247: pcbddc->mat_graph->n_local_subs = totn + 1;
7248: pcbddc->mat_graph->local_subs = local_subs;
7249: }
7250: }
7252: if (!pcbddc->graphanalyzed) {
7253: /* Graph's connected components analysis */
7254: PCBDDCGraphComputeConnectedComponents(pcbddc->mat_graph);
7255: pcbddc->graphanalyzed = PETSC_TRUE;
7256: pcbddc->corner_selected = pcbddc->corner_selection;
7257: }
7258: if (rcsr) pcbddc->mat_graph->nvtxs_csr = 0;
7259: return(0);
7260: }
7262: PetscErrorCode PCBDDCOrthonormalizeVecs(PetscInt *nio, Vec vecs[])
7263: {
7264: PetscInt i,j,n;
7265: PetscScalar *alphas;
7266: PetscReal norm,*onorms;
7270: n = *nio;
7271: if (!n) return(0);
7272: PetscMalloc2(n,&alphas,n,&onorms);
7273: VecNormalize(vecs[0],&norm);
7274: if (norm < PETSC_SMALL) {
7275: onorms[0] = 0.0;
7276: VecSet(vecs[0],0.0);
7277: } else {
7278: onorms[0] = norm;
7279: }
7281: for (i=1;i<n;i++) {
7282: VecMDot(vecs[i],i,vecs,alphas);
7283: for (j=0;j<i;j++) alphas[j] = PetscConj(-alphas[j]);
7284: VecMAXPY(vecs[i],i,alphas,vecs);
7285: VecNormalize(vecs[i],&norm);
7286: if (norm < PETSC_SMALL) {
7287: onorms[i] = 0.0;
7288: VecSet(vecs[i],0.0);
7289: } else {
7290: onorms[i] = norm;
7291: }
7292: }
7293: /* push nonzero vectors at the beginning */
7294: for (i=0;i<n;i++) {
7295: if (onorms[i] == 0.0) {
7296: for (j=i+1;j<n;j++) {
7297: if (onorms[j] != 0.0) {
7298: VecCopy(vecs[j],vecs[i]);
7299: onorms[j] = 0.0;
7300: }
7301: }
7302: }
7303: }
7304: for (i=0,*nio=0;i<n;i++) *nio += onorms[i] != 0.0 ? 1 : 0;
7305: PetscFree2(alphas,onorms);
7306: return(0);
7307: }
7309: PetscErrorCode PCBDDCMatISGetSubassemblingPattern(Mat mat, PetscInt *n_subdomains, PetscInt redprocs, IS* is_sends, PetscBool *have_void)
7310: {
7311: Mat A;
7312: PetscInt n_neighs,*neighs,*n_shared,**shared;
7313: PetscMPIInt size,rank,color;
7314: PetscInt *xadj,*adjncy;
7315: PetscInt *adjncy_wgt,*v_wgt,*ranks_send_to_idx;
7316: PetscInt im_active,active_procs,N,n,i,j,threshold = 2;
7317: PetscInt void_procs,*procs_candidates = NULL;
7318: PetscInt xadj_count,*count;
7319: PetscBool ismatis,use_vwgt=PETSC_FALSE;
7320: PetscSubcomm psubcomm;
7321: MPI_Comm subcomm;
7326: PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);
7327: if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",PETSC_FUNCTION_NAME);
7330: if (*n_subdomains <=0) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONG,"Invalid number of subdomains requested %D",*n_subdomains);
7332: if (have_void) *have_void = PETSC_FALSE;
7333: MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
7334: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
7335: MatISGetLocalMat(mat,&A);
7336: MatGetLocalSize(A,&n,NULL);
7337: im_active = !!n;
7338: MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mat));
7339: void_procs = size - active_procs;
7340: /* get ranks of of non-active processes in mat communicator */
7341: if (void_procs) {
7342: PetscInt ncand;
7344: if (have_void) *have_void = PETSC_TRUE;
7345: PetscMalloc1(size,&procs_candidates);
7346: MPI_Allgather(&im_active,1,MPIU_INT,procs_candidates,1,MPIU_INT,PetscObjectComm((PetscObject)mat));
7347: for (i=0,ncand=0;i<size;i++) {
7348: if (!procs_candidates[i]) {
7349: procs_candidates[ncand++] = i;
7350: }
7351: }
7352: /* force n_subdomains to be not greater that the number of non-active processes */
7353: *n_subdomains = PetscMin(void_procs,*n_subdomains);
7354: }
7356: /* number of subdomains requested greater than active processes or matrix size -> just shift the matrix
7357: number of subdomains requested 1 -> send to master or first candidate in voids */
7358: MatGetSize(mat,&N,NULL);
7359: if (active_procs < *n_subdomains || *n_subdomains == 1 || N <= *n_subdomains) {
7360: PetscInt issize,isidx,dest;
7361: if (*n_subdomains == 1) dest = 0;
7362: else dest = rank;
7363: if (im_active) {
7364: issize = 1;
7365: if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */
7366: isidx = procs_candidates[dest];
7367: } else {
7368: isidx = dest;
7369: }
7370: } else {
7371: issize = 0;
7372: isidx = -1;
7373: }
7374: if (*n_subdomains != 1) *n_subdomains = active_procs;
7375: ISCreateGeneral(PetscObjectComm((PetscObject)mat),issize,&isidx,PETSC_COPY_VALUES,is_sends);
7376: PetscFree(procs_candidates);
7377: return(0);
7378: }
7379: PetscOptionsGetBool(NULL,NULL,"-matis_partitioning_use_vwgt",&use_vwgt,NULL);
7380: PetscOptionsGetInt(NULL,NULL,"-matis_partitioning_threshold",&threshold,NULL);
7381: threshold = PetscMax(threshold,2);
7383: /* Get info on mapping */
7384: ISLocalToGlobalMappingGetInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);
7386: /* build local CSR graph of subdomains' connectivity */
7387: PetscMalloc1(2,&xadj);
7388: xadj[0] = 0;
7389: xadj[1] = PetscMax(n_neighs-1,0);
7390: PetscMalloc1(xadj[1],&adjncy);
7391: PetscMalloc1(xadj[1],&adjncy_wgt);
7392: PetscCalloc1(n,&count);
7393: for (i=1;i<n_neighs;i++)
7394: for (j=0;j<n_shared[i];j++)
7395: count[shared[i][j]] += 1;
7397: xadj_count = 0;
7398: for (i=1;i<n_neighs;i++) {
7399: for (j=0;j<n_shared[i];j++) {
7400: if (count[shared[i][j]] < threshold) {
7401: adjncy[xadj_count] = neighs[i];
7402: adjncy_wgt[xadj_count] = n_shared[i];
7403: xadj_count++;
7404: break;
7405: }
7406: }
7407: }
7408: xadj[1] = xadj_count;
7409: PetscFree(count);
7410: ISLocalToGlobalMappingRestoreInfo(mat->rmap->mapping,&n_neighs,&neighs,&n_shared,&shared);
7411: PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);
7413: PetscMalloc1(1,&ranks_send_to_idx);
7415: /* Restrict work on active processes only */
7416: PetscMPIIntCast(im_active,&color);
7417: if (void_procs) {
7418: PetscSubcommCreate(PetscObjectComm((PetscObject)mat),&psubcomm);
7419: PetscSubcommSetNumber(psubcomm,2); /* 2 groups, active process and not active processes */
7420: PetscSubcommSetTypeGeneral(psubcomm,color,rank);
7421: subcomm = PetscSubcommChild(psubcomm);
7422: } else {
7423: psubcomm = NULL;
7424: subcomm = PetscObjectComm((PetscObject)mat);
7425: }
7427: v_wgt = NULL;
7428: if (!color) {
7429: PetscFree(xadj);
7430: PetscFree(adjncy);
7431: PetscFree(adjncy_wgt);
7432: } else {
7433: Mat subdomain_adj;
7434: IS new_ranks,new_ranks_contig;
7435: MatPartitioning partitioner;
7436: PetscInt rstart=0,rend=0;
7437: PetscInt *is_indices,*oldranks;
7438: PetscMPIInt size;
7439: PetscBool aggregate;
7441: MPI_Comm_size(subcomm,&size);
7442: if (void_procs) {
7443: PetscInt prank = rank;
7444: PetscMalloc1(size,&oldranks);
7445: MPI_Allgather(&prank,1,MPIU_INT,oldranks,1,MPIU_INT,subcomm);
7446: for (i=0;i<xadj[1];i++) {
7447: PetscFindInt(adjncy[i],size,oldranks,&adjncy[i]);
7448: }
7449: PetscSortIntWithArray(xadj[1],adjncy,adjncy_wgt);
7450: } else {
7451: oldranks = NULL;
7452: }
7453: aggregate = ((redprocs > 0 && redprocs < size) ? PETSC_TRUE : PETSC_FALSE);
7454: if (aggregate) { /* TODO: all this part could be made more efficient */
7455: PetscInt lrows,row,ncols,*cols;
7456: PetscMPIInt nrank;
7457: PetscScalar *vals;
7459: MPI_Comm_rank(subcomm,&nrank);
7460: lrows = 0;
7461: if (nrank<redprocs) {
7462: lrows = size/redprocs;
7463: if (nrank<size%redprocs) lrows++;
7464: }
7465: MatCreateAIJ(subcomm,lrows,lrows,size,size,50,NULL,50,NULL,&subdomain_adj);
7466: MatGetOwnershipRange(subdomain_adj,&rstart,&rend);
7467: MatSetOption(subdomain_adj,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
7468: MatSetOption(subdomain_adj,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
7469: row = nrank;
7470: ncols = xadj[1]-xadj[0];
7471: cols = adjncy;
7472: PetscMalloc1(ncols,&vals);
7473: for (i=0;i<ncols;i++) vals[i] = adjncy_wgt[i];
7474: MatSetValues(subdomain_adj,1,&row,ncols,cols,vals,INSERT_VALUES);
7475: MatAssemblyBegin(subdomain_adj,MAT_FINAL_ASSEMBLY);
7476: MatAssemblyEnd(subdomain_adj,MAT_FINAL_ASSEMBLY);
7477: PetscFree(xadj);
7478: PetscFree(adjncy);
7479: PetscFree(adjncy_wgt);
7480: PetscFree(vals);
7481: if (use_vwgt) {
7482: Vec v;
7483: const PetscScalar *array;
7484: PetscInt nl;
7486: MatCreateVecs(subdomain_adj,&v,NULL);
7487: VecSetValue(v,row,(PetscScalar)n,INSERT_VALUES);
7488: VecAssemblyBegin(v);
7489: VecAssemblyEnd(v);
7490: VecGetLocalSize(v,&nl);
7491: VecGetArrayRead(v,&array);
7492: PetscMalloc1(nl,&v_wgt);
7493: for (i=0;i<nl;i++) v_wgt[i] = (PetscInt)PetscRealPart(array[i]);
7494: VecRestoreArrayRead(v,&array);
7495: VecDestroy(&v);
7496: }
7497: } else {
7498: MatCreateMPIAdj(subcomm,1,(PetscInt)size,xadj,adjncy,adjncy_wgt,&subdomain_adj);
7499: if (use_vwgt) {
7500: PetscMalloc1(1,&v_wgt);
7501: v_wgt[0] = n;
7502: }
7503: }
7504: /* MatView(subdomain_adj,0); */
7506: /* Partition */
7507: MatPartitioningCreate(subcomm,&partitioner);
7508: #if defined(PETSC_HAVE_PTSCOTCH)
7509: MatPartitioningSetType(partitioner,MATPARTITIONINGPTSCOTCH);
7510: #elif defined(PETSC_HAVE_PARMETIS)
7511: MatPartitioningSetType(partitioner,MATPARTITIONINGPARMETIS);
7512: #else
7513: MatPartitioningSetType(partitioner,MATPARTITIONINGAVERAGE);
7514: #endif
7515: MatPartitioningSetAdjacency(partitioner,subdomain_adj);
7516: if (v_wgt) {
7517: MatPartitioningSetVertexWeights(partitioner,v_wgt);
7518: }
7519: *n_subdomains = PetscMin((PetscInt)size,*n_subdomains);
7520: MatPartitioningSetNParts(partitioner,*n_subdomains);
7521: MatPartitioningSetFromOptions(partitioner);
7522: MatPartitioningApply(partitioner,&new_ranks);
7523: /* MatPartitioningView(partitioner,0); */
7525: /* renumber new_ranks to avoid "holes" in new set of processors */
7526: ISRenumber(new_ranks,NULL,NULL,&new_ranks_contig);
7527: ISDestroy(&new_ranks);
7528: ISGetIndices(new_ranks_contig,(const PetscInt**)&is_indices);
7529: if (!aggregate) {
7530: if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */
7531: #if defined(PETSC_USE_DEBUG)
7532: if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");
7533: #endif
7534: ranks_send_to_idx[0] = procs_candidates[oldranks[is_indices[0]]];
7535: } else if (oldranks) {
7536: ranks_send_to_idx[0] = oldranks[is_indices[0]];
7537: } else {
7538: ranks_send_to_idx[0] = is_indices[0];
7539: }
7540: } else {
7541: PetscInt idx = 0;
7542: PetscMPIInt tag;
7543: MPI_Request *reqs;
7545: PetscObjectGetNewTag((PetscObject)subdomain_adj,&tag);
7546: PetscMalloc1(rend-rstart,&reqs);
7547: for (i=rstart;i<rend;i++) {
7548: MPI_Isend(is_indices+i-rstart,1,MPIU_INT,i,tag,subcomm,&reqs[i-rstart]);
7549: }
7550: MPI_Recv(&idx,1,MPIU_INT,MPI_ANY_SOURCE,tag,subcomm,MPI_STATUS_IGNORE);
7551: MPI_Waitall(rend-rstart,reqs,MPI_STATUSES_IGNORE);
7552: PetscFree(reqs);
7553: if (procs_candidates) { /* shift the pattern on non-active candidates (if any) */
7554: #if defined(PETSC_USE_DEBUG)
7555: if (!oldranks) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"This should not happen");
7556: #endif
7557: ranks_send_to_idx[0] = procs_candidates[oldranks[idx]];
7558: } else if (oldranks) {
7559: ranks_send_to_idx[0] = oldranks[idx];
7560: } else {
7561: ranks_send_to_idx[0] = idx;
7562: }
7563: }
7564: ISRestoreIndices(new_ranks_contig,(const PetscInt**)&is_indices);
7565: /* clean up */
7566: PetscFree(oldranks);
7567: ISDestroy(&new_ranks_contig);
7568: MatDestroy(&subdomain_adj);
7569: MatPartitioningDestroy(&partitioner);
7570: }
7571: PetscSubcommDestroy(&psubcomm);
7572: PetscFree(procs_candidates);
7574: /* assemble parallel IS for sends */
7575: i = 1;
7576: if (!color) i=0;
7577: ISCreateGeneral(PetscObjectComm((PetscObject)mat),i,ranks_send_to_idx,PETSC_OWN_POINTER,is_sends);
7578: return(0);
7579: }
7581: typedef enum {MATDENSE_PRIVATE=0,MATAIJ_PRIVATE,MATBAIJ_PRIVATE,MATSBAIJ_PRIVATE}MatTypePrivate;
7583: PetscErrorCode PCBDDCMatISSubassemble(Mat mat, IS is_sends, PetscInt n_subdomains, PetscBool restrict_comm, PetscBool restrict_full, PetscBool reuse, Mat *mat_n, PetscInt nis, IS isarray[], PetscInt nvecs, Vec nnsp_vec[])
7584: {
7585: Mat local_mat;
7586: IS is_sends_internal;
7587: PetscInt rows,cols,new_local_rows;
7588: PetscInt i,bs,buf_size_idxs,buf_size_idxs_is,buf_size_vals,buf_size_vecs;
7589: PetscBool ismatis,isdense,newisdense,destroy_mat;
7590: ISLocalToGlobalMapping l2gmap;
7591: PetscInt* l2gmap_indices;
7592: const PetscInt* is_indices;
7593: MatType new_local_type;
7594: /* buffers */
7595: PetscInt *ptr_idxs,*send_buffer_idxs,*recv_buffer_idxs;
7596: PetscInt *ptr_idxs_is,*send_buffer_idxs_is,*recv_buffer_idxs_is;
7597: PetscInt *recv_buffer_idxs_local;
7598: PetscScalar *ptr_vals,*recv_buffer_vals;
7599: const PetscScalar *send_buffer_vals;
7600: PetscScalar *ptr_vecs,*send_buffer_vecs,*recv_buffer_vecs;
7601: /* MPI */
7602: MPI_Comm comm,comm_n;
7603: PetscSubcomm subcomm;
7604: PetscMPIInt n_sends,n_recvs,size;
7605: PetscMPIInt *iflags,*ilengths_idxs,*ilengths_vals,*ilengths_idxs_is;
7606: PetscMPIInt *onodes,*onodes_is,*olengths_idxs,*olengths_idxs_is,*olengths_vals;
7607: PetscMPIInt len,tag_idxs,tag_idxs_is,tag_vals,tag_vecs,source_dest;
7608: MPI_Request *send_req_idxs,*send_req_idxs_is,*send_req_vals,*send_req_vecs;
7609: MPI_Request *recv_req_idxs,*recv_req_idxs_is,*recv_req_vals,*recv_req_vecs;
7610: PetscErrorCode ierr;
7614: PetscObjectTypeCompare((PetscObject)mat,MATIS,&ismatis);
7615: if (!ismatis) SETERRQ1(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot use %s on a matrix object which is not of type MATIS",PETSC_FUNCTION_NAME);
7622: if (nvecs) {
7623: if (nvecs > 1) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Just 1 vector supported");
7625: }
7626: /* further checks */
7627: MatISGetLocalMat(mat,&local_mat);
7628: PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&isdense);
7629: if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Currently cannot subassemble MATIS when local matrix type is not of type SEQDENSE");
7630: MatGetSize(local_mat,&rows,&cols);
7631: if (rows != cols) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Local MATIS matrices should be square");
7632: if (reuse && *mat_n) {
7633: PetscInt mrows,mcols,mnrows,mncols;
7635: PetscObjectTypeCompare((PetscObject)*mat_n,MATIS,&ismatis);
7636: if (!ismatis) SETERRQ(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_SUP,"Cannot reuse a matrix which is not of type MATIS");
7637: MatGetSize(mat,&mrows,&mcols);
7638: MatGetSize(*mat_n,&mnrows,&mncols);
7639: if (mrows != mnrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of rows %D != %D",mrows,mnrows);
7640: if (mcols != mncols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix! Wrong number of cols %D != %D",mcols,mncols);
7641: }
7642: MatGetBlockSize(local_mat,&bs);
7645: /* prepare IS for sending if not provided */
7646: if (!is_sends) {
7647: if (!n_subdomains) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"You should specify either an IS or a target number of subdomains");
7648: PCBDDCMatISGetSubassemblingPattern(mat,&n_subdomains,0,&is_sends_internal,NULL);
7649: } else {
7650: PetscObjectReference((PetscObject)is_sends);
7651: is_sends_internal = is_sends;
7652: }
7654: /* get comm */
7655: PetscObjectGetComm((PetscObject)mat,&comm);
7657: /* compute number of sends */
7658: ISGetLocalSize(is_sends_internal,&i);
7659: PetscMPIIntCast(i,&n_sends);
7661: /* compute number of receives */
7662: MPI_Comm_size(comm,&size);
7663: PetscMalloc1(size,&iflags);
7664: PetscArrayzero(iflags,size);
7665: ISGetIndices(is_sends_internal,&is_indices);
7666: for (i=0;i<n_sends;i++) iflags[is_indices[i]] = 1;
7667: PetscGatherNumberOfMessages(comm,iflags,NULL,&n_recvs);
7668: PetscFree(iflags);
7670: /* restrict comm if requested */
7671: subcomm = 0;
7672: destroy_mat = PETSC_FALSE;
7673: if (restrict_comm) {
7674: PetscMPIInt color,subcommsize;
7676: color = 0;
7677: if (restrict_full) {
7678: if (!n_recvs) color = 1; /* processes not receiving anything will not partecipate in new comm (full restriction) */
7679: } else {
7680: if (!n_recvs && n_sends) color = 1; /* just those processes that are sending but not receiving anything will not partecipate in new comm */
7681: }
7682: MPIU_Allreduce(&color,&subcommsize,1,MPI_INT,MPI_SUM,comm);
7683: subcommsize = size - subcommsize;
7684: /* check if reuse has been requested */
7685: if (reuse) {
7686: if (*mat_n) {
7687: PetscMPIInt subcommsize2;
7688: MPI_Comm_size(PetscObjectComm((PetscObject)*mat_n),&subcommsize2);
7689: if (subcommsize != subcommsize2) SETERRQ2(PetscObjectComm((PetscObject)*mat_n),PETSC_ERR_PLIB,"Cannot reuse matrix! wrong subcomm size %d != %d",subcommsize,subcommsize2);
7690: comm_n = PetscObjectComm((PetscObject)*mat_n);
7691: } else {
7692: comm_n = PETSC_COMM_SELF;
7693: }
7694: } else { /* MAT_INITIAL_MATRIX */
7695: PetscMPIInt rank;
7697: MPI_Comm_rank(comm,&rank);
7698: PetscSubcommCreate(comm,&subcomm);
7699: PetscSubcommSetNumber(subcomm,2);
7700: PetscSubcommSetTypeGeneral(subcomm,color,rank);
7701: comm_n = PetscSubcommChild(subcomm);
7702: }
7703: /* flag to destroy *mat_n if not significative */
7704: if (color) destroy_mat = PETSC_TRUE;
7705: } else {
7706: comm_n = comm;
7707: }
7709: /* prepare send/receive buffers */
7710: PetscMalloc1(size,&ilengths_idxs);
7711: PetscArrayzero(ilengths_idxs,size);
7712: PetscMalloc1(size,&ilengths_vals);
7713: PetscArrayzero(ilengths_vals,size);
7714: if (nis) {
7715: PetscCalloc1(size,&ilengths_idxs_is);
7716: }
7718: /* Get data from local matrices */
7719: if (!isdense) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Subassembling of AIJ local matrices not yet implemented");
7720: /* TODO: See below some guidelines on how to prepare the local buffers */
7721: /*
7722: send_buffer_vals should contain the raw values of the local matrix
7723: send_buffer_idxs should contain:
7724: - MatType_PRIVATE type
7725: - PetscInt size_of_l2gmap
7726: - PetscInt global_row_indices[size_of_l2gmap]
7727: - PetscInt all_other_info_which_is_needed_to_compute_preallocation_and_set_values
7728: */
7729: else {
7730: MatDenseGetArrayRead(local_mat,&send_buffer_vals);
7731: ISLocalToGlobalMappingGetSize(mat->rmap->mapping,&i);
7732: PetscMalloc1(i+2,&send_buffer_idxs);
7733: send_buffer_idxs[0] = (PetscInt)MATDENSE_PRIVATE;
7734: send_buffer_idxs[1] = i;
7735: ISLocalToGlobalMappingGetIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);
7736: PetscArraycpy(&send_buffer_idxs[2],ptr_idxs,i);
7737: ISLocalToGlobalMappingRestoreIndices(mat->rmap->mapping,(const PetscInt**)&ptr_idxs);
7738: PetscMPIIntCast(i,&len);
7739: for (i=0;i<n_sends;i++) {
7740: ilengths_vals[is_indices[i]] = len*len;
7741: ilengths_idxs[is_indices[i]] = len+2;
7742: }
7743: }
7744: PetscGatherMessageLengths2(comm,n_sends,n_recvs,ilengths_idxs,ilengths_vals,&onodes,&olengths_idxs,&olengths_vals);
7745: /* additional is (if any) */
7746: if (nis) {
7747: PetscMPIInt psum;
7748: PetscInt j;
7749: for (j=0,psum=0;j<nis;j++) {
7750: PetscInt plen;
7751: ISGetLocalSize(isarray[j],&plen);
7752: PetscMPIIntCast(plen,&len);
7753: psum += len+1; /* indices + lenght */
7754: }
7755: PetscMalloc1(psum,&send_buffer_idxs_is);
7756: for (j=0,psum=0;j<nis;j++) {
7757: PetscInt plen;
7758: const PetscInt *is_array_idxs;
7759: ISGetLocalSize(isarray[j],&plen);
7760: send_buffer_idxs_is[psum] = plen;
7761: ISGetIndices(isarray[j],&is_array_idxs);
7762: PetscArraycpy(&send_buffer_idxs_is[psum+1],is_array_idxs,plen);
7763: ISRestoreIndices(isarray[j],&is_array_idxs);
7764: psum += plen+1; /* indices + lenght */
7765: }
7766: for (i=0;i<n_sends;i++) {
7767: ilengths_idxs_is[is_indices[i]] = psum;
7768: }
7769: PetscGatherMessageLengths(comm,n_sends,n_recvs,ilengths_idxs_is,&onodes_is,&olengths_idxs_is);
7770: }
7771: MatISRestoreLocalMat(mat,&local_mat);
7773: buf_size_idxs = 0;
7774: buf_size_vals = 0;
7775: buf_size_idxs_is = 0;
7776: buf_size_vecs = 0;
7777: for (i=0;i<n_recvs;i++) {
7778: buf_size_idxs += (PetscInt)olengths_idxs[i];
7779: buf_size_vals += (PetscInt)olengths_vals[i];
7780: if (nis) buf_size_idxs_is += (PetscInt)olengths_idxs_is[i];
7781: if (nvecs) buf_size_vecs += (PetscInt)olengths_idxs[i];
7782: }
7783: PetscMalloc1(buf_size_idxs,&recv_buffer_idxs);
7784: PetscMalloc1(buf_size_vals,&recv_buffer_vals);
7785: PetscMalloc1(buf_size_idxs_is,&recv_buffer_idxs_is);
7786: PetscMalloc1(buf_size_vecs,&recv_buffer_vecs);
7788: /* get new tags for clean communications */
7789: PetscObjectGetNewTag((PetscObject)mat,&tag_idxs);
7790: PetscObjectGetNewTag((PetscObject)mat,&tag_vals);
7791: PetscObjectGetNewTag((PetscObject)mat,&tag_idxs_is);
7792: PetscObjectGetNewTag((PetscObject)mat,&tag_vecs);
7794: /* allocate for requests */
7795: PetscMalloc1(n_sends,&send_req_idxs);
7796: PetscMalloc1(n_sends,&send_req_vals);
7797: PetscMalloc1(n_sends,&send_req_idxs_is);
7798: PetscMalloc1(n_sends,&send_req_vecs);
7799: PetscMalloc1(n_recvs,&recv_req_idxs);
7800: PetscMalloc1(n_recvs,&recv_req_vals);
7801: PetscMalloc1(n_recvs,&recv_req_idxs_is);
7802: PetscMalloc1(n_recvs,&recv_req_vecs);
7804: /* communications */
7805: ptr_idxs = recv_buffer_idxs;
7806: ptr_vals = recv_buffer_vals;
7807: ptr_idxs_is = recv_buffer_idxs_is;
7808: ptr_vecs = recv_buffer_vecs;
7809: for (i=0;i<n_recvs;i++) {
7810: source_dest = onodes[i];
7811: MPI_Irecv(ptr_idxs,olengths_idxs[i],MPIU_INT,source_dest,tag_idxs,comm,&recv_req_idxs[i]);
7812: MPI_Irecv(ptr_vals,olengths_vals[i],MPIU_SCALAR,source_dest,tag_vals,comm,&recv_req_vals[i]);
7813: ptr_idxs += olengths_idxs[i];
7814: ptr_vals += olengths_vals[i];
7815: if (nis) {
7816: source_dest = onodes_is[i];
7817: MPI_Irecv(ptr_idxs_is,olengths_idxs_is[i],MPIU_INT,source_dest,tag_idxs_is,comm,&recv_req_idxs_is[i]);
7818: ptr_idxs_is += olengths_idxs_is[i];
7819: }
7820: if (nvecs) {
7821: source_dest = onodes[i];
7822: MPI_Irecv(ptr_vecs,olengths_idxs[i]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&recv_req_vecs[i]);
7823: ptr_vecs += olengths_idxs[i]-2;
7824: }
7825: }
7826: for (i=0;i<n_sends;i++) {
7827: PetscMPIIntCast(is_indices[i],&source_dest);
7828: MPI_Isend(send_buffer_idxs,ilengths_idxs[source_dest],MPIU_INT,source_dest,tag_idxs,comm,&send_req_idxs[i]);
7829: MPI_Isend((PetscScalar*)send_buffer_vals,ilengths_vals[source_dest],MPIU_SCALAR,source_dest,tag_vals,comm,&send_req_vals[i]);
7830: if (nis) {
7831: MPI_Isend(send_buffer_idxs_is,ilengths_idxs_is[source_dest],MPIU_INT,source_dest,tag_idxs_is,comm,&send_req_idxs_is[i]);
7832: }
7833: if (nvecs) {
7834: VecGetArray(nnsp_vec[0],&send_buffer_vecs);
7835: MPI_Isend(send_buffer_vecs,ilengths_idxs[source_dest]-2,MPIU_SCALAR,source_dest,tag_vecs,comm,&send_req_vecs[i]);
7836: }
7837: }
7838: ISRestoreIndices(is_sends_internal,&is_indices);
7839: ISDestroy(&is_sends_internal);
7841: /* assemble new l2g map */
7842: MPI_Waitall(n_recvs,recv_req_idxs,MPI_STATUSES_IGNORE);
7843: ptr_idxs = recv_buffer_idxs;
7844: new_local_rows = 0;
7845: for (i=0;i<n_recvs;i++) {
7846: new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */
7847: ptr_idxs += olengths_idxs[i];
7848: }
7849: PetscMalloc1(new_local_rows,&l2gmap_indices);
7850: ptr_idxs = recv_buffer_idxs;
7851: new_local_rows = 0;
7852: for (i=0;i<n_recvs;i++) {
7853: PetscArraycpy(&l2gmap_indices[new_local_rows],ptr_idxs+2,*(ptr_idxs+1));
7854: new_local_rows += *(ptr_idxs+1); /* second element is the local size of the l2gmap */
7855: ptr_idxs += olengths_idxs[i];
7856: }
7857: PetscSortRemoveDupsInt(&new_local_rows,l2gmap_indices);
7858: ISLocalToGlobalMappingCreate(comm_n,1,new_local_rows,l2gmap_indices,PETSC_COPY_VALUES,&l2gmap);
7859: PetscFree(l2gmap_indices);
7861: /* infer new local matrix type from received local matrices type */
7862: /* currently if all local matrices are of type X, then the resulting matrix will be of type X, except for the dense case */
7863: /* it also assumes that if the block size is set, than it is the same among all local matrices (see checks at the beginning of the function) */
7864: if (n_recvs) {
7865: MatTypePrivate new_local_type_private = (MatTypePrivate)send_buffer_idxs[0];
7866: ptr_idxs = recv_buffer_idxs;
7867: for (i=0;i<n_recvs;i++) {
7868: if ((PetscInt)new_local_type_private != *ptr_idxs) {
7869: new_local_type_private = MATAIJ_PRIVATE;
7870: break;
7871: }
7872: ptr_idxs += olengths_idxs[i];
7873: }
7874: switch (new_local_type_private) {
7875: case MATDENSE_PRIVATE:
7876: new_local_type = MATSEQAIJ;
7877: bs = 1;
7878: break;
7879: case MATAIJ_PRIVATE:
7880: new_local_type = MATSEQAIJ;
7881: bs = 1;
7882: break;
7883: case MATBAIJ_PRIVATE:
7884: new_local_type = MATSEQBAIJ;
7885: break;
7886: case MATSBAIJ_PRIVATE:
7887: new_local_type = MATSEQSBAIJ;
7888: break;
7889: default:
7890: SETERRQ2(comm,PETSC_ERR_SUP,"Unsupported private type %d in %s",new_local_type_private,PETSC_FUNCTION_NAME);
7891: break;
7892: }
7893: } else { /* by default, new_local_type is seqaij */
7894: new_local_type = MATSEQAIJ;
7895: bs = 1;
7896: }
7898: /* create MATIS object if needed */
7899: if (!reuse) {
7900: MatGetSize(mat,&rows,&cols);
7901: MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);
7902: } else {
7903: /* it also destroys the local matrices */
7904: if (*mat_n) {
7905: MatSetLocalToGlobalMapping(*mat_n,l2gmap,l2gmap);
7906: } else { /* this is a fake object */
7907: MatCreateIS(comm_n,bs,PETSC_DECIDE,PETSC_DECIDE,rows,cols,l2gmap,NULL,mat_n);
7908: }
7909: }
7910: MatISGetLocalMat(*mat_n,&local_mat);
7911: MatSetType(local_mat,new_local_type);
7913: MPI_Waitall(n_recvs,recv_req_vals,MPI_STATUSES_IGNORE);
7915: /* Global to local map of received indices */
7916: PetscMalloc1(buf_size_idxs,&recv_buffer_idxs_local); /* needed for values insertion */
7917: ISGlobalToLocalMappingApply(l2gmap,IS_GTOLM_MASK,buf_size_idxs,recv_buffer_idxs,&i,recv_buffer_idxs_local);
7918: ISLocalToGlobalMappingDestroy(&l2gmap);
7920: /* restore attributes -> type of incoming data and its size */
7921: buf_size_idxs = 0;
7922: for (i=0;i<n_recvs;i++) {
7923: recv_buffer_idxs_local[buf_size_idxs] = recv_buffer_idxs[buf_size_idxs];
7924: recv_buffer_idxs_local[buf_size_idxs+1] = recv_buffer_idxs[buf_size_idxs+1];
7925: buf_size_idxs += (PetscInt)olengths_idxs[i];
7926: }
7927: PetscFree(recv_buffer_idxs);
7929: /* set preallocation */
7930: PetscObjectTypeCompare((PetscObject)local_mat,MATSEQDENSE,&newisdense);
7931: if (!newisdense) {
7932: PetscInt *new_local_nnz=0;
7934: ptr_idxs = recv_buffer_idxs_local;
7935: if (n_recvs) {
7936: PetscCalloc1(new_local_rows,&new_local_nnz);
7937: }
7938: for (i=0;i<n_recvs;i++) {
7939: PetscInt j;
7940: if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* preallocation provided for dense case only */
7941: for (j=0;j<*(ptr_idxs+1);j++) {
7942: new_local_nnz[*(ptr_idxs+2+j)] += *(ptr_idxs+1);
7943: }
7944: } else {
7945: /* TODO */
7946: }
7947: ptr_idxs += olengths_idxs[i];
7948: }
7949: if (new_local_nnz) {
7950: for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMin(new_local_nnz[i],new_local_rows);
7951: MatSeqAIJSetPreallocation(local_mat,0,new_local_nnz);
7952: for (i=0;i<new_local_rows;i++) new_local_nnz[i] /= bs;
7953: MatSeqBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);
7954: for (i=0;i<new_local_rows;i++) new_local_nnz[i] = PetscMax(new_local_nnz[i]-i,0);
7955: MatSeqSBAIJSetPreallocation(local_mat,bs,0,new_local_nnz);
7956: } else {
7957: MatSetUp(local_mat);
7958: }
7959: PetscFree(new_local_nnz);
7960: } else {
7961: MatSetUp(local_mat);
7962: }
7964: /* set values */
7965: ptr_vals = recv_buffer_vals;
7966: ptr_idxs = recv_buffer_idxs_local;
7967: for (i=0;i<n_recvs;i++) {
7968: if (*ptr_idxs == (PetscInt)MATDENSE_PRIVATE) { /* values insertion provided for dense case only */
7969: MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_FALSE);
7970: MatSetValues(local_mat,*(ptr_idxs+1),ptr_idxs+2,*(ptr_idxs+1),ptr_idxs+2,ptr_vals,ADD_VALUES);
7971: MatAssemblyBegin(local_mat,MAT_FLUSH_ASSEMBLY);
7972: MatAssemblyEnd(local_mat,MAT_FLUSH_ASSEMBLY);
7973: MatSetOption(local_mat,MAT_ROW_ORIENTED,PETSC_TRUE);
7974: } else {
7975: /* TODO */
7976: }
7977: ptr_idxs += olengths_idxs[i];
7978: ptr_vals += olengths_vals[i];
7979: }
7980: MatAssemblyBegin(local_mat,MAT_FINAL_ASSEMBLY);
7981: MatAssemblyEnd(local_mat,MAT_FINAL_ASSEMBLY);
7982: MatISRestoreLocalMat(*mat_n,&local_mat);
7983: MatAssemblyBegin(*mat_n,MAT_FINAL_ASSEMBLY);
7984: MatAssemblyEnd(*mat_n,MAT_FINAL_ASSEMBLY);
7985: PetscFree(recv_buffer_vals);
7987: #if 0
7988: if (!restrict_comm) { /* check */
7989: Vec lvec,rvec;
7990: PetscReal infty_error;
7992: MatCreateVecs(mat,&rvec,&lvec);
7993: VecSetRandom(rvec,NULL);
7994: MatMult(mat,rvec,lvec);
7995: VecScale(lvec,-1.0);
7996: MatMultAdd(*mat_n,rvec,lvec,lvec);
7997: VecNorm(lvec,NORM_INFINITY,&infty_error);
7998: PetscPrintf(PetscObjectComm((PetscObject)mat),"Infinity error subassembling %1.6e\n",infty_error);
7999: VecDestroy(&rvec);
8000: VecDestroy(&lvec);
8001: }
8002: #endif
8004: /* assemble new additional is (if any) */
8005: if (nis) {
8006: PetscInt **temp_idxs,*count_is,j,psum;
8008: MPI_Waitall(n_recvs,recv_req_idxs_is,MPI_STATUSES_IGNORE);
8009: PetscCalloc1(nis,&count_is);
8010: ptr_idxs = recv_buffer_idxs_is;
8011: psum = 0;
8012: for (i=0;i<n_recvs;i++) {
8013: for (j=0;j<nis;j++) {
8014: PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
8015: count_is[j] += plen; /* increment counting of buffer for j-th IS */
8016: psum += plen;
8017: ptr_idxs += plen+1; /* shift pointer to received data */
8018: }
8019: }
8020: PetscMalloc1(nis,&temp_idxs);
8021: PetscMalloc1(psum,&temp_idxs[0]);
8022: for (i=1;i<nis;i++) {
8023: temp_idxs[i] = temp_idxs[i-1]+count_is[i-1];
8024: }
8025: PetscArrayzero(count_is,nis);
8026: ptr_idxs = recv_buffer_idxs_is;
8027: for (i=0;i<n_recvs;i++) {
8028: for (j=0;j<nis;j++) {
8029: PetscInt plen = *(ptr_idxs); /* first element is the local size of IS's indices */
8030: PetscArraycpy(&temp_idxs[j][count_is[j]],ptr_idxs+1,plen);
8031: count_is[j] += plen; /* increment starting point of buffer for j-th IS */
8032: ptr_idxs += plen+1; /* shift pointer to received data */
8033: }
8034: }
8035: for (i=0;i<nis;i++) {
8036: ISDestroy(&isarray[i]);
8037: PetscSortRemoveDupsInt(&count_is[i],temp_idxs[i]);
8038: ISCreateGeneral(comm_n,count_is[i],temp_idxs[i],PETSC_COPY_VALUES,&isarray[i]);
8039: }
8040: PetscFree(count_is);
8041: PetscFree(temp_idxs[0]);
8042: PetscFree(temp_idxs);
8043: }
8044: /* free workspace */
8045: PetscFree(recv_buffer_idxs_is);
8046: MPI_Waitall(n_sends,send_req_idxs,MPI_STATUSES_IGNORE);
8047: PetscFree(send_buffer_idxs);
8048: MPI_Waitall(n_sends,send_req_vals,MPI_STATUSES_IGNORE);
8049: if (isdense) {
8050: MatISGetLocalMat(mat,&local_mat);
8051: MatDenseRestoreArrayRead(local_mat,&send_buffer_vals);
8052: MatISRestoreLocalMat(mat,&local_mat);
8053: } else {
8054: /* PetscFree(send_buffer_vals); */
8055: }
8056: if (nis) {
8057: MPI_Waitall(n_sends,send_req_idxs_is,MPI_STATUSES_IGNORE);
8058: PetscFree(send_buffer_idxs_is);
8059: }
8061: if (nvecs) {
8062: MPI_Waitall(n_recvs,recv_req_vecs,MPI_STATUSES_IGNORE);
8063: MPI_Waitall(n_sends,send_req_vecs,MPI_STATUSES_IGNORE);
8064: VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);
8065: VecDestroy(&nnsp_vec[0]);
8066: VecCreate(comm_n,&nnsp_vec[0]);
8067: VecSetSizes(nnsp_vec[0],new_local_rows,PETSC_DECIDE);
8068: VecSetType(nnsp_vec[0],VECSTANDARD);
8069: /* set values */
8070: ptr_vals = recv_buffer_vecs;
8071: ptr_idxs = recv_buffer_idxs_local;
8072: VecGetArray(nnsp_vec[0],&send_buffer_vecs);
8073: for (i=0;i<n_recvs;i++) {
8074: PetscInt j;
8075: for (j=0;j<*(ptr_idxs+1);j++) {
8076: send_buffer_vecs[*(ptr_idxs+2+j)] += *(ptr_vals + j);
8077: }
8078: ptr_idxs += olengths_idxs[i];
8079: ptr_vals += olengths_idxs[i]-2;
8080: }
8081: VecRestoreArray(nnsp_vec[0],&send_buffer_vecs);
8082: VecAssemblyBegin(nnsp_vec[0]);
8083: VecAssemblyEnd(nnsp_vec[0]);
8084: }
8086: PetscFree(recv_buffer_vecs);
8087: PetscFree(recv_buffer_idxs_local);
8088: PetscFree(recv_req_idxs);
8089: PetscFree(recv_req_vals);
8090: PetscFree(recv_req_vecs);
8091: PetscFree(recv_req_idxs_is);
8092: PetscFree(send_req_idxs);
8093: PetscFree(send_req_vals);
8094: PetscFree(send_req_vecs);
8095: PetscFree(send_req_idxs_is);
8096: PetscFree(ilengths_vals);
8097: PetscFree(ilengths_idxs);
8098: PetscFree(olengths_vals);
8099: PetscFree(olengths_idxs);
8100: PetscFree(onodes);
8101: if (nis) {
8102: PetscFree(ilengths_idxs_is);
8103: PetscFree(olengths_idxs_is);
8104: PetscFree(onodes_is);
8105: }
8106: PetscSubcommDestroy(&subcomm);
8107: if (destroy_mat) { /* destroy mat is true only if restrict comm is true and process will not partecipate */
8108: MatDestroy(mat_n);
8109: for (i=0;i<nis;i++) {
8110: ISDestroy(&isarray[i]);
8111: }
8112: if (nvecs) { /* need to match VecDestroy nnsp_vec called in the other code path */
8113: VecDestroy(&nnsp_vec[0]);
8114: }
8115: *mat_n = NULL;
8116: }
8117: return(0);
8118: }
8120: /* temporary hack into ksp private data structure */
8121: #include <petsc/private/kspimpl.h>
8123: PetscErrorCode PCBDDCSetUpCoarseSolver(PC pc,PetscScalar* coarse_submat_vals)
8124: {
8125: PC_BDDC *pcbddc = (PC_BDDC*)pc->data;
8126: PC_IS *pcis = (PC_IS*)pc->data;
8127: Mat coarse_mat,coarse_mat_is,coarse_submat_dense;
8128: Mat coarsedivudotp = NULL;
8129: Mat coarseG,t_coarse_mat_is;
8130: MatNullSpace CoarseNullSpace = NULL;
8131: ISLocalToGlobalMapping coarse_islg;
8132: IS coarse_is,*isarray,corners;
8133: PetscInt i,im_active=-1,active_procs=-1;
8134: PetscInt nis,nisdofs,nisneu,nisvert;
8135: PetscInt coarse_eqs_per_proc;
8136: PC pc_temp;
8137: PCType coarse_pc_type;
8138: KSPType coarse_ksp_type;
8139: PetscBool multilevel_requested,multilevel_allowed;
8140: PetscBool coarse_reuse;
8141: PetscInt ncoarse,nedcfield;
8142: PetscBool compute_vecs = PETSC_FALSE;
8143: PetscScalar *array;
8144: MatReuse coarse_mat_reuse;
8145: PetscBool restr, full_restr, have_void;
8146: PetscMPIInt size;
8147: PetscErrorCode ierr;
8150: PetscLogEventBegin(PC_BDDC_CoarseSetUp[pcbddc->current_level],pc,0,0,0);
8151: /* Assign global numbering to coarse dofs */
8152: if (pcbddc->new_primal_space || pcbddc->coarse_size == -1) { /* a new primal space is present or it is the first initialization, so recompute global numbering */
8153: PetscInt ocoarse_size;
8154: compute_vecs = PETSC_TRUE;
8156: pcbddc->new_primal_space = PETSC_TRUE;
8157: ocoarse_size = pcbddc->coarse_size;
8158: PetscFree(pcbddc->global_primal_indices);
8159: PCBDDCComputePrimalNumbering(pc,&pcbddc->coarse_size,&pcbddc->global_primal_indices);
8160: /* see if we can avoid some work */
8161: if (pcbddc->coarse_ksp) { /* coarse ksp has already been created */
8162: /* if the coarse size is different or we are using adaptive selection, better to not reuse the coarse matrix */
8163: if (ocoarse_size != pcbddc->coarse_size || pcbddc->adaptive_selection) {
8164: KSPReset(pcbddc->coarse_ksp);
8165: coarse_reuse = PETSC_FALSE;
8166: } else { /* we can safely reuse already computed coarse matrix */
8167: coarse_reuse = PETSC_TRUE;
8168: }
8169: } else { /* there's no coarse ksp, so we need to create the coarse matrix too */
8170: coarse_reuse = PETSC_FALSE;
8171: }
8172: /* reset any subassembling information */
8173: if (!coarse_reuse || pcbddc->recompute_topography) {
8174: ISDestroy(&pcbddc->coarse_subassembling);
8175: }
8176: } else { /* primal space is unchanged, so we can reuse coarse matrix */
8177: coarse_reuse = PETSC_TRUE;
8178: }
8179: if (coarse_reuse && pcbddc->coarse_ksp) {
8180: KSPGetOperators(pcbddc->coarse_ksp,&coarse_mat,NULL);
8181: PetscObjectReference((PetscObject)coarse_mat);
8182: coarse_mat_reuse = MAT_REUSE_MATRIX;
8183: } else {
8184: coarse_mat = NULL;
8185: coarse_mat_reuse = MAT_INITIAL_MATRIX;
8186: }
8188: /* creates temporary l2gmap and IS for coarse indexes */
8189: ISCreateGeneral(PetscObjectComm((PetscObject)pc),pcbddc->local_primal_size,pcbddc->global_primal_indices,PETSC_COPY_VALUES,&coarse_is);
8190: ISLocalToGlobalMappingCreateIS(coarse_is,&coarse_islg);
8192: /* creates temporary MATIS object for coarse matrix */
8193: MatCreateSeqDense(PETSC_COMM_SELF,pcbddc->local_primal_size,pcbddc->local_primal_size,coarse_submat_vals,&coarse_submat_dense);
8194: MatCreateIS(PetscObjectComm((PetscObject)pc),1,PETSC_DECIDE,PETSC_DECIDE,pcbddc->coarse_size,pcbddc->coarse_size,coarse_islg,NULL,&t_coarse_mat_is);
8195: MatISSetLocalMat(t_coarse_mat_is,coarse_submat_dense);
8196: MatAssemblyBegin(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);
8197: MatAssemblyEnd(t_coarse_mat_is,MAT_FINAL_ASSEMBLY);
8198: MatDestroy(&coarse_submat_dense);
8200: /* count "active" (i.e. with positive local size) and "void" processes */
8201: im_active = !!(pcis->n);
8202: MPIU_Allreduce(&im_active,&active_procs,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));
8204: /* determine number of processes partecipating to coarse solver and compute subassembling pattern */
8205: /* restr : whether we want to exclude senders (which are not receivers) from the subassembling pattern */
8206: /* full_restr : just use the receivers from the subassembling pattern */
8207: MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);
8208: coarse_mat_is = NULL;
8209: multilevel_allowed = PETSC_FALSE;
8210: multilevel_requested = PETSC_FALSE;
8211: coarse_eqs_per_proc = PetscMin(PetscMax(pcbddc->coarse_size,1),pcbddc->coarse_eqs_per_proc);
8212: if (coarse_eqs_per_proc < 0) coarse_eqs_per_proc = pcbddc->coarse_size;
8213: if (pcbddc->current_level < pcbddc->max_levels) multilevel_requested = PETSC_TRUE;
8214: if (pcbddc->coarse_size <= pcbddc->coarse_eqs_limit) multilevel_requested = PETSC_FALSE;
8215: if (multilevel_requested) {
8216: ncoarse = active_procs/pcbddc->coarsening_ratio;
8217: restr = PETSC_FALSE;
8218: full_restr = PETSC_FALSE;
8219: } else {
8220: ncoarse = pcbddc->coarse_size/coarse_eqs_per_proc + !!(pcbddc->coarse_size%coarse_eqs_per_proc);
8221: restr = PETSC_TRUE;
8222: full_restr = PETSC_TRUE;
8223: }
8224: if (!pcbddc->coarse_size || size == 1) multilevel_allowed = multilevel_requested = restr = full_restr = PETSC_FALSE;
8225: ncoarse = PetscMax(1,ncoarse);
8226: if (!pcbddc->coarse_subassembling) {
8227: if (pcbddc->coarsening_ratio > 1) {
8228: if (multilevel_requested) {
8229: PCBDDCMatISGetSubassemblingPattern(pc->pmat,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);
8230: } else {
8231: PCBDDCMatISGetSubassemblingPattern(t_coarse_mat_is,&ncoarse,pcbddc->coarse_adj_red,&pcbddc->coarse_subassembling,&have_void);
8232: }
8233: } else {
8234: PetscMPIInt rank;
8236: MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);
8237: have_void = (active_procs == (PetscInt)size) ? PETSC_FALSE : PETSC_TRUE;
8238: ISCreateStride(PetscObjectComm((PetscObject)pc),1,rank,1,&pcbddc->coarse_subassembling);
8239: }
8240: } else { /* if a subassembling pattern exists, then we can reuse the coarse ksp and compute the number of process involved */
8241: PetscInt psum;
8242: if (pcbddc->coarse_ksp) psum = 1;
8243: else psum = 0;
8244: MPIU_Allreduce(&psum,&ncoarse,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));
8245: have_void = ncoarse < size ? PETSC_TRUE : PETSC_FALSE;
8246: }
8247: /* determine if we can go multilevel */
8248: if (multilevel_requested) {
8249: if (ncoarse > 1) multilevel_allowed = PETSC_TRUE; /* found enough processes */
8250: else restr = full_restr = PETSC_TRUE; /* 1 subdomain, use a direct solver */
8251: }
8252: if (multilevel_allowed && have_void) restr = PETSC_TRUE;
8254: /* dump subassembling pattern */
8255: if (pcbddc->dbg_flag && multilevel_allowed) {
8256: ISView(pcbddc->coarse_subassembling,pcbddc->dbg_viewer);
8257: }
8258: /* compute dofs splitting and neumann boundaries for coarse dofs */
8259: nedcfield = -1;
8260: corners = NULL;
8261: if (multilevel_allowed && !coarse_reuse && (pcbddc->n_ISForDofsLocal || pcbddc->NeumannBoundariesLocal || pcbddc->nedclocal || pcbddc->corner_selected)) { /* protects from unneeded computations */
8262: PetscInt *tidxs,*tidxs2,nout,tsize,i;
8263: const PetscInt *idxs;
8264: ISLocalToGlobalMapping tmap;
8266: /* create map between primal indices (in local representative ordering) and local primal numbering */
8267: ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,1,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,PETSC_COPY_VALUES,&tmap);
8268: /* allocate space for temporary storage */
8269: PetscMalloc1(pcbddc->local_primal_size,&tidxs);
8270: PetscMalloc1(pcbddc->local_primal_size,&tidxs2);
8271: /* allocate for IS array */
8272: nisdofs = pcbddc->n_ISForDofsLocal;
8273: if (pcbddc->nedclocal) {
8274: if (pcbddc->nedfield > -1) {
8275: nedcfield = pcbddc->nedfield;
8276: } else {
8277: nedcfield = 0;
8278: if (nisdofs) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"This should not happen (%D)",nisdofs);
8279: nisdofs = 1;
8280: }
8281: }
8282: nisneu = !!pcbddc->NeumannBoundariesLocal;
8283: nisvert = 0; /* nisvert is not used */
8284: nis = nisdofs + nisneu + nisvert;
8285: PetscMalloc1(nis,&isarray);
8286: /* dofs splitting */
8287: for (i=0;i<nisdofs;i++) {
8288: /* ISView(pcbddc->ISForDofsLocal[i],0); */
8289: if (nedcfield != i) {
8290: ISGetLocalSize(pcbddc->ISForDofsLocal[i],&tsize);
8291: ISGetIndices(pcbddc->ISForDofsLocal[i],&idxs);
8292: ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);
8293: ISRestoreIndices(pcbddc->ISForDofsLocal[i],&idxs);
8294: } else {
8295: ISGetLocalSize(pcbddc->nedclocal,&tsize);
8296: ISGetIndices(pcbddc->nedclocal,&idxs);
8297: ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);
8298: if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping coarse nedelec field! %D != %D",tsize,nout);
8299: ISRestoreIndices(pcbddc->nedclocal,&idxs);
8300: }
8301: ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);
8302: ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[i]);
8303: /* ISView(isarray[i],0); */
8304: }
8305: /* neumann boundaries */
8306: if (pcbddc->NeumannBoundariesLocal) {
8307: /* ISView(pcbddc->NeumannBoundariesLocal,0); */
8308: ISGetLocalSize(pcbddc->NeumannBoundariesLocal,&tsize);
8309: ISGetIndices(pcbddc->NeumannBoundariesLocal,&idxs);
8310: ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);
8311: ISRestoreIndices(pcbddc->NeumannBoundariesLocal,&idxs);
8312: ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);
8313: ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&isarray[nisdofs]);
8314: /* ISView(isarray[nisdofs],0); */
8315: }
8316: /* coordinates */
8317: if (pcbddc->corner_selected) {
8318: PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&corners);
8319: ISGetLocalSize(corners,&tsize);
8320: ISGetIndices(corners,&idxs);
8321: ISGlobalToLocalMappingApply(tmap,IS_GTOLM_DROP,tsize,idxs,&nout,tidxs);
8322: if (tsize != nout) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Failed when mapping corners! %D != %D",tsize,nout);
8323: ISRestoreIndices(corners,&idxs);
8324: PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&corners);
8325: ISLocalToGlobalMappingApply(coarse_islg,nout,tidxs,tidxs2);
8326: ISCreateGeneral(PetscObjectComm((PetscObject)pc),nout,tidxs2,PETSC_COPY_VALUES,&corners);
8327: }
8328: PetscFree(tidxs);
8329: PetscFree(tidxs2);
8330: ISLocalToGlobalMappingDestroy(&tmap);
8331: } else {
8332: nis = 0;
8333: nisdofs = 0;
8334: nisneu = 0;
8335: nisvert = 0;
8336: isarray = NULL;
8337: }
8338: /* destroy no longer needed map */
8339: ISLocalToGlobalMappingDestroy(&coarse_islg);
8341: /* subassemble */
8342: if (multilevel_allowed) {
8343: Vec vp[1];
8344: PetscInt nvecs = 0;
8345: PetscBool reuse,reuser;
8347: if (coarse_mat) reuse = PETSC_TRUE;
8348: else reuse = PETSC_FALSE;
8349: MPIU_Allreduce(&reuse,&reuser,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));
8350: vp[0] = NULL;
8351: if (pcbddc->benign_have_null) { /* propagate no-net-flux quadrature to coarser level */
8352: VecCreate(PetscObjectComm((PetscObject)pc),&vp[0]);
8353: VecSetSizes(vp[0],pcbddc->local_primal_size,PETSC_DECIDE);
8354: VecSetType(vp[0],VECSTANDARD);
8355: nvecs = 1;
8357: if (pcbddc->divudotp) {
8358: Mat B,loc_divudotp;
8359: Vec v,p;
8360: IS dummy;
8361: PetscInt np;
8363: MatISGetLocalMat(pcbddc->divudotp,&loc_divudotp);
8364: MatGetSize(loc_divudotp,&np,NULL);
8365: ISCreateStride(PETSC_COMM_SELF,np,0,1,&dummy);
8366: MatCreateSubMatrix(loc_divudotp,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B);
8367: MatCreateVecs(B,&v,&p);
8368: VecSet(p,1.);
8369: MatMultTranspose(B,p,v);
8370: VecDestroy(&p);
8371: MatDestroy(&B);
8372: VecGetArray(vp[0],&array);
8373: VecPlaceArray(pcbddc->vec1_P,array);
8374: VecRestoreArray(vp[0],&array);
8375: MatMultTranspose(pcbddc->coarse_phi_B,v,pcbddc->vec1_P);
8376: VecResetArray(pcbddc->vec1_P);
8377: ISDestroy(&dummy);
8378: VecDestroy(&v);
8379: }
8380: }
8381: if (reuser) {
8382: PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_TRUE,&coarse_mat,nis,isarray,nvecs,vp);
8383: } else {
8384: PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,nis,isarray,nvecs,vp);
8385: }
8386: if (vp[0]) { /* vp[0] could have been placed on a different set of processes */
8387: PetscScalar *arraym;
8388: const PetscScalar *arrayv;
8389: PetscInt nl;
8390: VecGetLocalSize(vp[0],&nl);
8391: MatCreateSeqDense(PETSC_COMM_SELF,1,nl,NULL,&coarsedivudotp);
8392: MatDenseGetArray(coarsedivudotp,&arraym);
8393: VecGetArrayRead(vp[0],&arrayv);
8394: PetscArraycpy(arraym,arrayv,nl);
8395: VecRestoreArrayRead(vp[0],&arrayv);
8396: MatDenseRestoreArray(coarsedivudotp,&arraym);
8397: VecDestroy(&vp[0]);
8398: } else {
8399: MatCreateSeqAIJ(PETSC_COMM_SELF,0,0,1,NULL,&coarsedivudotp);
8400: }
8401: } else {
8402: PCBDDCMatISSubassemble(t_coarse_mat_is,pcbddc->coarse_subassembling,0,restr,full_restr,PETSC_FALSE,&coarse_mat_is,0,NULL,0,NULL);
8403: }
8404: if (coarse_mat_is || coarse_mat) {
8405: if (!multilevel_allowed) {
8406: MatConvert(coarse_mat_is,MATAIJ,coarse_mat_reuse,&coarse_mat);
8407: } else {
8408: /* if this matrix is present, it means we are not reusing the coarse matrix */
8409: if (coarse_mat_is) {
8410: if (coarse_mat) SETERRQ(PetscObjectComm((PetscObject)coarse_mat_is),PETSC_ERR_PLIB,"This should not happen");
8411: PetscObjectReference((PetscObject)coarse_mat_is);
8412: coarse_mat = coarse_mat_is;
8413: }
8414: }
8415: }
8416: MatDestroy(&t_coarse_mat_is);
8417: MatDestroy(&coarse_mat_is);
8419: /* create local to global scatters for coarse problem */
8420: if (compute_vecs) {
8421: PetscInt lrows;
8422: VecDestroy(&pcbddc->coarse_vec);
8423: if (coarse_mat) {
8424: MatGetLocalSize(coarse_mat,&lrows,NULL);
8425: } else {
8426: lrows = 0;
8427: }
8428: VecCreate(PetscObjectComm((PetscObject)pc),&pcbddc->coarse_vec);
8429: VecSetSizes(pcbddc->coarse_vec,lrows,PETSC_DECIDE);
8430: VecSetType(pcbddc->coarse_vec,coarse_mat ? coarse_mat->defaultvectype : VECSTANDARD);
8431: VecScatterDestroy(&pcbddc->coarse_loc_to_glob);
8432: VecScatterCreate(pcbddc->vec1_P,NULL,pcbddc->coarse_vec,coarse_is,&pcbddc->coarse_loc_to_glob);
8433: }
8434: ISDestroy(&coarse_is);
8436: /* set defaults for coarse KSP and PC */
8437: if (multilevel_allowed) {
8438: coarse_ksp_type = KSPRICHARDSON;
8439: coarse_pc_type = PCBDDC;
8440: } else {
8441: coarse_ksp_type = KSPPREONLY;
8442: coarse_pc_type = PCREDUNDANT;
8443: }
8445: /* print some info if requested */
8446: if (pcbddc->dbg_flag) {
8447: if (!multilevel_allowed) {
8448: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");
8449: if (multilevel_requested) {
8450: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Not enough active processes on level %D (active processes %D, coarsening ratio %D)\n",pcbddc->current_level,active_procs,pcbddc->coarsening_ratio);
8451: } else if (pcbddc->max_levels) {
8452: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Maximum number of requested levels reached (%D)\n",pcbddc->max_levels);
8453: }
8454: PetscViewerFlush(pcbddc->dbg_viewer);
8455: }
8456: }
8458: /* communicate coarse discrete gradient */
8459: coarseG = NULL;
8460: if (pcbddc->nedcG && multilevel_allowed) {
8461: MPI_Comm ccomm;
8462: if (coarse_mat) {
8463: ccomm = PetscObjectComm((PetscObject)coarse_mat);
8464: } else {
8465: ccomm = MPI_COMM_NULL;
8466: }
8467: MatMPIAIJRestrict(pcbddc->nedcG,ccomm,&coarseG);
8468: }
8470: /* create the coarse KSP object only once with defaults */
8471: if (coarse_mat) {
8472: PetscBool isredundant,isbddc,force,valid;
8473: PetscViewer dbg_viewer = NULL;
8475: if (pcbddc->dbg_flag) {
8476: dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)coarse_mat));
8477: PetscViewerASCIIAddTab(dbg_viewer,2*pcbddc->current_level);
8478: }
8479: if (!pcbddc->coarse_ksp) {
8480: char prefix[256],str_level[16];
8481: size_t len;
8483: KSPCreate(PetscObjectComm((PetscObject)coarse_mat),&pcbddc->coarse_ksp);
8484: KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,pc->erroriffailure);
8485: PetscObjectIncrementTabLevel((PetscObject)pcbddc->coarse_ksp,(PetscObject)pc,1);
8486: KSPSetTolerances(pcbddc->coarse_ksp,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT,1);
8487: KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);
8488: KSPSetType(pcbddc->coarse_ksp,coarse_ksp_type);
8489: KSPSetNormType(pcbddc->coarse_ksp,KSP_NORM_NONE);
8490: KSPGetPC(pcbddc->coarse_ksp,&pc_temp);
8491: /* TODO is this logic correct? should check for coarse_mat type */
8492: PCSetType(pc_temp,coarse_pc_type);
8493: /* prefix */
8494: PetscStrcpy(prefix,"");
8495: PetscStrcpy(str_level,"");
8496: if (!pcbddc->current_level) {
8497: PetscStrncpy(prefix,((PetscObject)pc)->prefix,sizeof(prefix));
8498: PetscStrlcat(prefix,"pc_bddc_coarse_",sizeof(prefix));
8499: } else {
8500: PetscStrlen(((PetscObject)pc)->prefix,&len);
8501: if (pcbddc->current_level>1) len -= 3; /* remove "lX_" with X level number */
8502: if (pcbddc->current_level>10) len -= 1; /* remove another char from level number */
8503: /* Nonstandard use of PetscStrncpy() to copy only a portion of the string */
8504: PetscStrncpy(prefix,((PetscObject)pc)->prefix,len+1);
8505: PetscSNPrintf(str_level,sizeof(str_level),"l%d_",(int)(pcbddc->current_level));
8506: PetscStrlcat(prefix,str_level,sizeof(prefix));
8507: }
8508: KSPSetOptionsPrefix(pcbddc->coarse_ksp,prefix);
8509: /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
8510: PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);
8511: PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);
8512: PCBDDCSetLevels(pc_temp,pcbddc->max_levels);
8513: /* allow user customization */
8514: KSPSetFromOptions(pcbddc->coarse_ksp);
8515: /* get some info after set from options */
8516: KSPGetPC(pcbddc->coarse_ksp,&pc_temp);
8517: /* multilevel cannot be done with coarse PC different from BDDC, NN, HPDDM, unless forced to */
8518: force = PETSC_FALSE;
8519: PetscOptionsGetBool(NULL,((PetscObject)pc_temp)->prefix,"-pc_type_forced",&force,NULL);
8520: PetscObjectTypeCompareAny((PetscObject)pc_temp,&valid,PCBDDC,PCNN,PCHPDDM,"");
8521: PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);
8522: if (multilevel_allowed && !force && !valid) {
8523: isbddc = PETSC_TRUE;
8524: PCSetType(pc_temp,PCBDDC);
8525: PCBDDCSetLevel(pc_temp,pcbddc->current_level+1);
8526: PCBDDCSetCoarseningRatio(pc_temp,pcbddc->coarsening_ratio);
8527: PCBDDCSetLevels(pc_temp,pcbddc->max_levels);
8528: if (pc_temp->ops->setfromoptions) { /* need to setfromoptions again, skipping the pc_type */
8529: PetscObjectOptionsBegin((PetscObject)pc_temp);
8530: (*pc_temp->ops->setfromoptions)(PetscOptionsObject,pc_temp);
8531: PetscObjectProcessOptionsHandlers(PetscOptionsObject,(PetscObject)pc_temp);
8532: PetscOptionsEnd();
8533: pc_temp->setfromoptionscalled++;
8534: }
8535: }
8536: }
8537: /* propagate BDDC info to the next level (these are dummy calls if pc_temp is not of type PCBDDC) */
8538: KSPGetPC(pcbddc->coarse_ksp,&pc_temp);
8539: if (nisdofs) {
8540: PCBDDCSetDofsSplitting(pc_temp,nisdofs,isarray);
8541: for (i=0;i<nisdofs;i++) {
8542: ISDestroy(&isarray[i]);
8543: }
8544: }
8545: if (nisneu) {
8546: PCBDDCSetNeumannBoundaries(pc_temp,isarray[nisdofs]);
8547: ISDestroy(&isarray[nisdofs]);
8548: }
8549: if (nisvert) {
8550: PCBDDCSetPrimalVerticesIS(pc_temp,isarray[nis-1]);
8551: ISDestroy(&isarray[nis-1]);
8552: }
8553: if (coarseG) {
8554: PCBDDCSetDiscreteGradient(pc_temp,coarseG,1,nedcfield,PETSC_FALSE,PETSC_TRUE);
8555: }
8557: /* get some info after set from options */
8558: PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);
8560: /* multilevel can only be requested via -pc_bddc_levels or PCBDDCSetLevels */
8561: if (isbddc && !multilevel_allowed) {
8562: PCSetType(pc_temp,coarse_pc_type);
8563: }
8564: /* multilevel cannot be done with coarse PC different from BDDC, NN, HPDDM, unless forced to */
8565: force = PETSC_FALSE;
8566: PetscOptionsGetBool(NULL,((PetscObject)pc_temp)->prefix,"-pc_type_forced",&force,NULL);
8567: PetscObjectTypeCompareAny((PetscObject)pc_temp,&valid,PCBDDC,PCNN,PCHPDDM,"");
8568: if (multilevel_requested && multilevel_allowed && !valid && !force) {
8569: PCSetType(pc_temp,PCBDDC);
8570: }
8571: PetscObjectTypeCompare((PetscObject)pc_temp,PCREDUNDANT,&isredundant);
8572: if (isredundant) {
8573: KSP inner_ksp;
8574: PC inner_pc;
8576: PCRedundantGetKSP(pc_temp,&inner_ksp);
8577: KSPGetPC(inner_ksp,&inner_pc);
8578: }
8580: /* parameters which miss an API */
8581: PetscObjectTypeCompare((PetscObject)pc_temp,PCBDDC,&isbddc);
8582: if (isbddc) {
8583: PC_BDDC* pcbddc_coarse = (PC_BDDC*)pc_temp->data;
8585: pcbddc_coarse->detect_disconnected = PETSC_TRUE;
8586: pcbddc_coarse->coarse_eqs_per_proc = pcbddc->coarse_eqs_per_proc;
8587: pcbddc_coarse->coarse_eqs_limit = pcbddc->coarse_eqs_limit;
8588: pcbddc_coarse->benign_saddle_point = pcbddc->benign_have_null;
8589: if (pcbddc_coarse->benign_saddle_point) {
8590: Mat coarsedivudotp_is;
8591: ISLocalToGlobalMapping l2gmap,rl2g,cl2g;
8592: IS row,col;
8593: const PetscInt *gidxs;
8594: PetscInt n,st,M,N;
8596: MatGetSize(coarsedivudotp,&n,NULL);
8597: MPI_Scan(&n,&st,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)coarse_mat));
8598: st = st-n;
8599: ISCreateStride(PetscObjectComm((PetscObject)coarse_mat),1,st,1,&row);
8600: MatGetLocalToGlobalMapping(coarse_mat,&l2gmap,NULL);
8601: ISLocalToGlobalMappingGetSize(l2gmap,&n);
8602: ISLocalToGlobalMappingGetIndices(l2gmap,&gidxs);
8603: ISCreateGeneral(PetscObjectComm((PetscObject)coarse_mat),n,gidxs,PETSC_COPY_VALUES,&col);
8604: ISLocalToGlobalMappingRestoreIndices(l2gmap,&gidxs);
8605: ISLocalToGlobalMappingCreateIS(row,&rl2g);
8606: ISLocalToGlobalMappingCreateIS(col,&cl2g);
8607: ISGetSize(row,&M);
8608: MatGetSize(coarse_mat,&N,NULL);
8609: ISDestroy(&row);
8610: ISDestroy(&col);
8611: MatCreate(PetscObjectComm((PetscObject)coarse_mat),&coarsedivudotp_is);
8612: MatSetType(coarsedivudotp_is,MATIS);
8613: MatSetSizes(coarsedivudotp_is,PETSC_DECIDE,PETSC_DECIDE,M,N);
8614: MatSetLocalToGlobalMapping(coarsedivudotp_is,rl2g,cl2g);
8615: ISLocalToGlobalMappingDestroy(&rl2g);
8616: ISLocalToGlobalMappingDestroy(&cl2g);
8617: MatISSetLocalMat(coarsedivudotp_is,coarsedivudotp);
8618: MatDestroy(&coarsedivudotp);
8619: PCBDDCSetDivergenceMat(pc_temp,coarsedivudotp_is,PETSC_FALSE,NULL);
8620: MatDestroy(&coarsedivudotp_is);
8621: pcbddc_coarse->adaptive_userdefined = PETSC_TRUE;
8622: if (pcbddc->adaptive_threshold[0] == 0.0) pcbddc_coarse->deluxe_zerorows = PETSC_TRUE;
8623: }
8624: }
8626: /* propagate symmetry info of coarse matrix */
8627: MatSetOption(coarse_mat,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);
8628: if (pc->pmat->symmetric_set) {
8629: MatSetOption(coarse_mat,MAT_SYMMETRIC,pc->pmat->symmetric);
8630: }
8631: if (pc->pmat->hermitian_set) {
8632: MatSetOption(coarse_mat,MAT_HERMITIAN,pc->pmat->hermitian);
8633: }
8634: if (pc->pmat->spd_set) {
8635: MatSetOption(coarse_mat,MAT_SPD,pc->pmat->spd);
8636: }
8637: if (pcbddc->benign_saddle_point && !pcbddc->benign_have_null) {
8638: MatSetOption(coarse_mat,MAT_SPD,PETSC_TRUE);
8639: }
8640: /* set operators */
8641: MatViewFromOptions(coarse_mat,(PetscObject)pc,"-pc_bddc_coarse_mat_view");
8642: MatSetOptionsPrefix(coarse_mat,((PetscObject)pcbddc->coarse_ksp)->prefix);
8643: KSPSetOperators(pcbddc->coarse_ksp,coarse_mat,coarse_mat);
8644: if (pcbddc->dbg_flag) {
8645: PetscViewerASCIISubtractTab(dbg_viewer,2*pcbddc->current_level);
8646: }
8647: }
8648: MatDestroy(&coarseG);
8649: PetscFree(isarray);
8650: #if 0
8651: {
8652: PetscViewer viewer;
8653: char filename[256];
8654: sprintf(filename,"coarse_mat_level%d.m",pcbddc->current_level);
8655: PetscViewerASCIIOpen(PetscObjectComm((PetscObject)coarse_mat),filename,&viewer);
8656: PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);
8657: MatView(coarse_mat,viewer);
8658: PetscViewerPopFormat(viewer);
8659: PetscViewerDestroy(&viewer);
8660: }
8661: #endif
8663: if (corners) {
8664: Vec gv;
8665: IS is;
8666: const PetscInt *idxs;
8667: PetscInt i,d,N,n,cdim = pcbddc->mat_graph->cdim;
8668: PetscScalar *coords;
8670: if (!pcbddc->mat_graph->cloc) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing local coordinates");
8671: VecGetSize(pcbddc->coarse_vec,&N);
8672: VecGetLocalSize(pcbddc->coarse_vec,&n);
8673: VecCreate(PetscObjectComm((PetscObject)pcbddc->coarse_vec),&gv);
8674: VecSetBlockSize(gv,cdim);
8675: VecSetSizes(gv,n*cdim,N*cdim);
8676: VecSetType(gv,VECSTANDARD);
8677: VecSetFromOptions(gv);
8678: VecSet(gv,PETSC_MAX_REAL); /* we only propagate coordinates from vertices constraints */
8680: PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&is);
8681: ISGetLocalSize(is,&n);
8682: ISGetIndices(is,&idxs);
8683: PetscMalloc1(n*cdim,&coords);
8684: for (i=0;i<n;i++) {
8685: for (d=0;d<cdim;d++) {
8686: coords[cdim*i+d] = pcbddc->mat_graph->coords[cdim*idxs[i]+d];
8687: }
8688: }
8689: ISRestoreIndices(is,&idxs);
8690: PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&is);
8692: ISGetLocalSize(corners,&n);
8693: ISGetIndices(corners,&idxs);
8694: VecSetValuesBlocked(gv,n,idxs,coords,INSERT_VALUES);
8695: ISRestoreIndices(corners,&idxs);
8696: PetscFree(coords);
8697: VecAssemblyBegin(gv);
8698: VecAssemblyEnd(gv);
8699: VecGetArray(gv,&coords);
8700: if (pcbddc->coarse_ksp) {
8701: PC coarse_pc;
8702: PetscBool isbddc;
8704: KSPGetPC(pcbddc->coarse_ksp,&coarse_pc);
8705: PetscObjectTypeCompare((PetscObject)coarse_pc,PCBDDC,&isbddc);
8706: if (isbddc) { /* coarse coordinates have PETSC_MAX_REAL, specific for BDDC */
8707: PetscReal *realcoords;
8709: VecGetLocalSize(gv,&n);
8710: #if defined(PETSC_USE_COMPLEX)
8711: PetscMalloc1(n,&realcoords);
8712: for (i=0;i<n;i++) realcoords[i] = PetscRealPart(coords[i]);
8713: #else
8714: realcoords = coords;
8715: #endif
8716: PCSetCoordinates(coarse_pc,cdim,n/cdim,realcoords);
8717: #if defined(PETSC_USE_COMPLEX)
8718: PetscFree(realcoords);
8719: #endif
8720: }
8721: }
8722: VecRestoreArray(gv,&coords);
8723: VecDestroy(&gv);
8724: }
8725: ISDestroy(&corners);
8727: if (pcbddc->coarse_ksp) {
8728: Vec crhs,csol;
8730: KSPGetSolution(pcbddc->coarse_ksp,&csol);
8731: KSPGetRhs(pcbddc->coarse_ksp,&crhs);
8732: if (!csol) {
8733: MatCreateVecs(coarse_mat,&((pcbddc->coarse_ksp)->vec_sol),NULL);
8734: }
8735: if (!crhs) {
8736: MatCreateVecs(coarse_mat,NULL,&((pcbddc->coarse_ksp)->vec_rhs));
8737: }
8738: }
8739: MatDestroy(&coarsedivudotp);
8741: /* compute null space for coarse solver if the benign trick has been requested */
8742: if (pcbddc->benign_null) {
8744: VecSet(pcbddc->vec1_P,0.);
8745: for (i=0;i<pcbddc->benign_n;i++) {
8746: VecSetValue(pcbddc->vec1_P,pcbddc->local_primal_size-pcbddc->benign_n+i,1.0,INSERT_VALUES);
8747: }
8748: VecAssemblyBegin(pcbddc->vec1_P);
8749: VecAssemblyEnd(pcbddc->vec1_P);
8750: VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);
8751: VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,INSERT_VALUES,SCATTER_FORWARD);
8752: if (coarse_mat) {
8753: Vec nullv;
8754: PetscScalar *array,*array2;
8755: PetscInt nl;
8757: MatCreateVecs(coarse_mat,&nullv,NULL);
8758: VecGetLocalSize(nullv,&nl);
8759: VecGetArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);
8760: VecGetArray(nullv,&array2);
8761: PetscArraycpy(array2,array,nl);
8762: VecRestoreArray(nullv,&array2);
8763: VecRestoreArrayRead(pcbddc->coarse_vec,(const PetscScalar**)&array);
8764: VecNormalize(nullv,NULL);
8765: MatNullSpaceCreate(PetscObjectComm((PetscObject)coarse_mat),PETSC_FALSE,1,&nullv,&CoarseNullSpace);
8766: VecDestroy(&nullv);
8767: }
8768: }
8769: PetscLogEventEnd(PC_BDDC_CoarseSetUp[pcbddc->current_level],pc,0,0,0);
8771: PetscLogEventBegin(PC_BDDC_CoarseSolver[pcbddc->current_level],pc,0,0,0);
8772: if (pcbddc->coarse_ksp) {
8773: PetscBool ispreonly;
8775: if (CoarseNullSpace) {
8776: PetscBool isnull;
8777: MatNullSpaceTest(CoarseNullSpace,coarse_mat,&isnull);
8778: if (isnull) {
8779: MatSetNullSpace(coarse_mat,CoarseNullSpace);
8780: }
8781: /* TODO: add local nullspaces (if any) */
8782: }
8783: /* setup coarse ksp */
8784: KSPSetUp(pcbddc->coarse_ksp);
8785: /* Check coarse problem if in debug mode or if solving with an iterative method */
8786: PetscObjectTypeCompare((PetscObject)pcbddc->coarse_ksp,KSPPREONLY,&ispreonly);
8787: if (pcbddc->dbg_flag || (!ispreonly && pcbddc->use_coarse_estimates) ) {
8788: KSP check_ksp;
8789: KSPType check_ksp_type;
8790: PC check_pc;
8791: Vec check_vec,coarse_vec;
8792: PetscReal abs_infty_error,infty_error,lambda_min=1.0,lambda_max=1.0;
8793: PetscInt its;
8794: PetscBool compute_eigs;
8795: PetscReal *eigs_r,*eigs_c;
8796: PetscInt neigs;
8797: const char *prefix;
8799: /* Create ksp object suitable for estimation of extreme eigenvalues */
8800: KSPCreate(PetscObjectComm((PetscObject)pcbddc->coarse_ksp),&check_ksp);
8801: PetscObjectIncrementTabLevel((PetscObject)check_ksp,(PetscObject)pcbddc->coarse_ksp,0);
8802: KSPSetErrorIfNotConverged(pcbddc->coarse_ksp,PETSC_FALSE);
8803: KSPSetOperators(check_ksp,coarse_mat,coarse_mat);
8804: KSPSetTolerances(check_ksp,1.e-12,1.e-12,PETSC_DEFAULT,pcbddc->coarse_size);
8805: /* prevent from setup unneeded object */
8806: KSPGetPC(check_ksp,&check_pc);
8807: PCSetType(check_pc,PCNONE);
8808: if (ispreonly) {
8809: check_ksp_type = KSPPREONLY;
8810: compute_eigs = PETSC_FALSE;
8811: } else {
8812: check_ksp_type = KSPGMRES;
8813: compute_eigs = PETSC_TRUE;
8814: }
8815: KSPSetType(check_ksp,check_ksp_type);
8816: KSPSetComputeSingularValues(check_ksp,compute_eigs);
8817: KSPSetComputeEigenvalues(check_ksp,compute_eigs);
8818: KSPGMRESSetRestart(check_ksp,pcbddc->coarse_size+1);
8819: KSPGetOptionsPrefix(pcbddc->coarse_ksp,&prefix);
8820: KSPSetOptionsPrefix(check_ksp,prefix);
8821: KSPAppendOptionsPrefix(check_ksp,"check_");
8822: KSPSetFromOptions(check_ksp);
8823: KSPSetUp(check_ksp);
8824: KSPGetPC(pcbddc->coarse_ksp,&check_pc);
8825: KSPSetPC(check_ksp,check_pc);
8826: /* create random vec */
8827: MatCreateVecs(coarse_mat,&coarse_vec,&check_vec);
8828: VecSetRandom(check_vec,NULL);
8829: MatMult(coarse_mat,check_vec,coarse_vec);
8830: /* solve coarse problem */
8831: KSPSolve(check_ksp,coarse_vec,coarse_vec);
8832: KSPCheckSolve(check_ksp,pc,coarse_vec);
8833: /* set eigenvalue estimation if preonly has not been requested */
8834: if (compute_eigs) {
8835: PetscMalloc1(pcbddc->coarse_size+1,&eigs_r);
8836: PetscMalloc1(pcbddc->coarse_size+1,&eigs_c);
8837: KSPComputeEigenvalues(check_ksp,pcbddc->coarse_size+1,eigs_r,eigs_c,&neigs);
8838: if (neigs) {
8839: lambda_max = eigs_r[neigs-1];
8840: lambda_min = eigs_r[0];
8841: if (pcbddc->use_coarse_estimates) {
8842: if (lambda_max>=lambda_min) { /* using PETSC_SMALL since lambda_max == lambda_min is not allowed by KSPChebyshevSetEigenvalues */
8843: KSPChebyshevSetEigenvalues(pcbddc->coarse_ksp,lambda_max+PETSC_SMALL,lambda_min);
8844: KSPRichardsonSetScale(pcbddc->coarse_ksp,2.0/(lambda_max+lambda_min));
8845: }
8846: }
8847: }
8848: }
8850: /* check coarse problem residual error */
8851: if (pcbddc->dbg_flag) {
8852: PetscViewer dbg_viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)pcbddc->coarse_ksp));
8853: PetscViewerASCIIAddTab(dbg_viewer,2*(pcbddc->current_level+1));
8854: VecAXPY(check_vec,-1.0,coarse_vec);
8855: VecNorm(check_vec,NORM_INFINITY,&infty_error);
8856: MatMult(coarse_mat,check_vec,coarse_vec);
8857: VecNorm(coarse_vec,NORM_INFINITY,&abs_infty_error);
8858: PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem details (use estimates %d)\n",pcbddc->use_coarse_estimates);
8859: PetscObjectPrintClassNamePrefixType((PetscObject)(pcbddc->coarse_ksp),dbg_viewer);
8860: PetscObjectPrintClassNamePrefixType((PetscObject)(check_pc),dbg_viewer);
8861: PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem exact infty_error : %1.6e\n",infty_error);
8862: PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem residual infty_error: %1.6e\n",abs_infty_error);
8863: if (CoarseNullSpace) {
8864: PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem is singular\n");
8865: }
8866: if (compute_eigs) {
8867: PetscReal lambda_max_s,lambda_min_s;
8868: KSPConvergedReason reason;
8869: KSPGetType(check_ksp,&check_ksp_type);
8870: KSPGetIterationNumber(check_ksp,&its);
8871: KSPGetConvergedReason(check_ksp,&reason);
8872: KSPComputeExtremeSingularValues(check_ksp,&lambda_max_s,&lambda_min_s);
8873: PetscViewerASCIIPrintf(dbg_viewer,"Coarse problem eigenvalues (estimated with %d iterations of %s, conv reason %d): %1.6e %1.6e (%1.6e %1.6e)\n",its,check_ksp_type,reason,lambda_min,lambda_max,lambda_min_s,lambda_max_s);
8874: for (i=0;i<neigs;i++) {
8875: PetscViewerASCIIPrintf(dbg_viewer,"%1.6e %1.6ei\n",eigs_r[i],eigs_c[i]);
8876: }
8877: }
8878: PetscViewerFlush(dbg_viewer);
8879: PetscViewerASCIISubtractTab(dbg_viewer,2*(pcbddc->current_level+1));
8880: }
8881: VecDestroy(&check_vec);
8882: VecDestroy(&coarse_vec);
8883: KSPDestroy(&check_ksp);
8884: if (compute_eigs) {
8885: PetscFree(eigs_r);
8886: PetscFree(eigs_c);
8887: }
8888: }
8889: }
8890: MatNullSpaceDestroy(&CoarseNullSpace);
8891: /* print additional info */
8892: if (pcbddc->dbg_flag) {
8893: /* waits until all processes reaches this point */
8894: PetscBarrier((PetscObject)pc);
8895: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Coarse solver setup completed at level %D\n",pcbddc->current_level);
8896: PetscViewerFlush(pcbddc->dbg_viewer);
8897: }
8899: /* free memory */
8900: MatDestroy(&coarse_mat);
8901: PetscLogEventEnd(PC_BDDC_CoarseSolver[pcbddc->current_level],pc,0,0,0);
8902: return(0);
8903: }
8905: PetscErrorCode PCBDDCComputePrimalNumbering(PC pc,PetscInt* coarse_size_n,PetscInt** local_primal_indices_n)
8906: {
8907: PC_BDDC* pcbddc = (PC_BDDC*)pc->data;
8908: PC_IS* pcis = (PC_IS*)pc->data;
8909: Mat_IS* matis = (Mat_IS*)pc->pmat->data;
8910: IS subset,subset_mult,subset_n;
8911: PetscInt local_size,coarse_size=0;
8912: PetscInt *local_primal_indices=NULL;
8913: const PetscInt *t_local_primal_indices;
8917: /* Compute global number of coarse dofs */
8918: if (pcbddc->local_primal_size && !pcbddc->local_primal_ref_node) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"BDDC ConstraintsSetUp should be called first");
8919: ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_node,PETSC_COPY_VALUES,&subset_n);
8920: ISLocalToGlobalMappingApplyIS(pcis->mapping,subset_n,&subset);
8921: ISDestroy(&subset_n);
8922: ISCreateGeneral(PetscObjectComm((PetscObject)(pc->pmat)),pcbddc->local_primal_size_cc,pcbddc->local_primal_ref_mult,PETSC_COPY_VALUES,&subset_mult);
8923: ISRenumber(subset,subset_mult,&coarse_size,&subset_n);
8924: ISDestroy(&subset);
8925: ISDestroy(&subset_mult);
8926: ISGetLocalSize(subset_n,&local_size);
8927: if (local_size != pcbddc->local_primal_size) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Invalid number of local primal indices computed %D != %D",local_size,pcbddc->local_primal_size);
8928: PetscMalloc1(local_size,&local_primal_indices);
8929: ISGetIndices(subset_n,&t_local_primal_indices);
8930: PetscArraycpy(local_primal_indices,t_local_primal_indices,local_size);
8931: ISRestoreIndices(subset_n,&t_local_primal_indices);
8932: ISDestroy(&subset_n);
8934: /* check numbering */
8935: if (pcbddc->dbg_flag) {
8936: PetscScalar coarsesum,*array,*array2;
8937: PetscInt i;
8938: PetscBool set_error = PETSC_FALSE,set_error_reduced = PETSC_FALSE;
8940: PetscViewerFlush(pcbddc->dbg_viewer);
8941: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"--------------------------------------------------\n");
8942: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Check coarse indices\n");
8943: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
8944: /* counter */
8945: VecSet(pcis->vec1_global,0.0);
8946: VecSet(pcis->vec1_N,1.0);
8947: VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
8948: VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
8949: VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);
8950: VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);
8951: VecSet(pcis->vec1_N,0.0);
8952: for (i=0;i<pcbddc->local_primal_size;i++) {
8953: VecSetValue(pcis->vec1_N,pcbddc->primal_indices_local_idxs[i],1.0,INSERT_VALUES);
8954: }
8955: VecAssemblyBegin(pcis->vec1_N);
8956: VecAssemblyEnd(pcis->vec1_N);
8957: VecSet(pcis->vec1_global,0.0);
8958: VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
8959: VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
8960: VecScatterBegin(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);
8961: VecScatterEnd(matis->rctx,pcis->vec1_global,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);
8962: VecGetArray(pcis->vec1_N,&array);
8963: VecGetArray(pcis->vec2_N,&array2);
8964: for (i=0;i<pcis->n;i++) {
8965: if (array[i] != 0.0 && array[i] != array2[i]) {
8966: PetscInt owned = (PetscInt)PetscRealPart(array[i]),gi;
8967: PetscInt neigh = (PetscInt)PetscRealPart(array2[i]);
8968: set_error = PETSC_TRUE;
8969: ISLocalToGlobalMappingApply(pcis->mapping,1,&i,&gi);
8970: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d: local index %D (gid %D) owned by %D processes instead of %D!\n",PetscGlobalRank,i,gi,owned,neigh);
8971: }
8972: }
8973: VecRestoreArray(pcis->vec2_N,&array2);
8974: MPIU_Allreduce(&set_error,&set_error_reduced,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));
8975: PetscViewerFlush(pcbddc->dbg_viewer);
8976: for (i=0;i<pcis->n;i++) {
8977: if (PetscRealPart(array[i]) > 0.0) array[i] = 1.0/PetscRealPart(array[i]);
8978: }
8979: VecRestoreArray(pcis->vec1_N,&array);
8980: VecSet(pcis->vec1_global,0.0);
8981: VecScatterBegin(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
8982: VecScatterEnd(matis->rctx,pcis->vec1_N,pcis->vec1_global,ADD_VALUES,SCATTER_REVERSE);
8983: VecSum(pcis->vec1_global,&coarsesum);
8984: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Size of coarse problem is %D (%lf)\n",coarse_size,PetscRealPart(coarsesum));
8985: if (pcbddc->dbg_flag > 1 || set_error_reduced) {
8986: PetscInt *gidxs;
8988: PetscMalloc1(pcbddc->local_primal_size,&gidxs);
8989: ISLocalToGlobalMappingApply(pcis->mapping,pcbddc->local_primal_size,pcbddc->primal_indices_local_idxs,gidxs);
8990: PetscViewerASCIIPrintf(pcbddc->dbg_viewer,"Distribution of local primal indices\n");
8991: PetscViewerFlush(pcbddc->dbg_viewer);
8992: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d\n",PetscGlobalRank);
8993: for (i=0;i<pcbddc->local_primal_size;i++) {
8994: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"local_primal_indices[%D]=%D (%D,%D)\n",i,local_primal_indices[i],pcbddc->primal_indices_local_idxs[i],gidxs[i]);
8995: }
8996: PetscViewerFlush(pcbddc->dbg_viewer);
8997: PetscFree(gidxs);
8998: }
8999: PetscViewerFlush(pcbddc->dbg_viewer);
9000: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
9001: if (set_error_reduced) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_PLIB,"BDDC Numbering of coarse dofs failed");
9002: }
9004: /* get back data */
9005: *coarse_size_n = coarse_size;
9006: *local_primal_indices_n = local_primal_indices;
9007: return(0);
9008: }
9010: PetscErrorCode PCBDDCGlobalToLocal(VecScatter g2l_ctx,Vec gwork, Vec lwork, IS globalis, IS* localis)
9011: {
9012: IS localis_t;
9013: PetscInt i,lsize,*idxs,n;
9014: PetscScalar *vals;
9018: /* get indices in local ordering exploiting local to global map */
9019: ISGetLocalSize(globalis,&lsize);
9020: PetscMalloc1(lsize,&vals);
9021: for (i=0;i<lsize;i++) vals[i] = 1.0;
9022: ISGetIndices(globalis,(const PetscInt**)&idxs);
9023: VecSet(gwork,0.0);
9024: VecSet(lwork,0.0);
9025: if (idxs) { /* multilevel guard */
9026: VecSetOption(gwork,VEC_IGNORE_NEGATIVE_INDICES,PETSC_TRUE);
9027: VecSetValues(gwork,lsize,idxs,vals,INSERT_VALUES);
9028: }
9029: VecAssemblyBegin(gwork);
9030: ISRestoreIndices(globalis,(const PetscInt**)&idxs);
9031: PetscFree(vals);
9032: VecAssemblyEnd(gwork);
9033: /* now compute set in local ordering */
9034: VecScatterBegin(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);
9035: VecScatterEnd(g2l_ctx,gwork,lwork,INSERT_VALUES,SCATTER_FORWARD);
9036: VecGetArrayRead(lwork,(const PetscScalar**)&vals);
9037: VecGetSize(lwork,&n);
9038: for (i=0,lsize=0;i<n;i++) {
9039: if (PetscRealPart(vals[i]) > 0.5) {
9040: lsize++;
9041: }
9042: }
9043: PetscMalloc1(lsize,&idxs);
9044: for (i=0,lsize=0;i<n;i++) {
9045: if (PetscRealPart(vals[i]) > 0.5) {
9046: idxs[lsize++] = i;
9047: }
9048: }
9049: VecRestoreArrayRead(lwork,(const PetscScalar**)&vals);
9050: ISCreateGeneral(PetscObjectComm((PetscObject)gwork),lsize,idxs,PETSC_OWN_POINTER,&localis_t);
9051: *localis = localis_t;
9052: return(0);
9053: }
9055: PetscErrorCode PCBDDCSetUpSubSchurs(PC pc)
9056: {
9057: PC_IS *pcis=(PC_IS*)pc->data;
9058: PC_BDDC *pcbddc=(PC_BDDC*)pc->data;
9059: PCBDDCSubSchurs sub_schurs=pcbddc->sub_schurs;
9060: Mat S_j;
9061: PetscInt *used_xadj,*used_adjncy;
9062: PetscBool free_used_adj;
9063: PetscErrorCode ierr;
9066: PetscLogEventBegin(PC_BDDC_Schurs[pcbddc->current_level],pc,0,0,0);
9067: /* decide the adjacency to be used for determining internal problems for local schur on subsets */
9068: free_used_adj = PETSC_FALSE;
9069: if (pcbddc->sub_schurs_layers == -1) {
9070: used_xadj = NULL;
9071: used_adjncy = NULL;
9072: } else {
9073: if (pcbddc->sub_schurs_use_useradj && pcbddc->mat_graph->xadj) {
9074: used_xadj = pcbddc->mat_graph->xadj;
9075: used_adjncy = pcbddc->mat_graph->adjncy;
9076: } else if (pcbddc->computed_rowadj) {
9077: used_xadj = pcbddc->mat_graph->xadj;
9078: used_adjncy = pcbddc->mat_graph->adjncy;
9079: } else {
9080: PetscBool flg_row=PETSC_FALSE;
9081: const PetscInt *xadj,*adjncy;
9082: PetscInt nvtxs;
9084: MatGetRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);
9085: if (flg_row) {
9086: PetscMalloc2(nvtxs+1,&used_xadj,xadj[nvtxs],&used_adjncy);
9087: PetscArraycpy(used_xadj,xadj,nvtxs+1);
9088: PetscArraycpy(used_adjncy,adjncy,xadj[nvtxs]);
9089: free_used_adj = PETSC_TRUE;
9090: } else {
9091: pcbddc->sub_schurs_layers = -1;
9092: used_xadj = NULL;
9093: used_adjncy = NULL;
9094: }
9095: MatRestoreRowIJ(pcbddc->local_mat,0,PETSC_TRUE,PETSC_FALSE,&nvtxs,&xadj,&adjncy,&flg_row);
9096: }
9097: }
9099: /* setup sub_schurs data */
9100: MatCreateSchurComplement(pcis->A_II,pcis->pA_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);
9101: if (!sub_schurs->schur_explicit) {
9102: /* pcbddc->ksp_D up to date only if not using MatFactor with Schur complement support */
9103: MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);
9104: PCBDDCSubSchursSetUp(sub_schurs,NULL,S_j,PETSC_FALSE,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,NULL,pcbddc->adaptive_selection,PETSC_FALSE,PETSC_FALSE,0,NULL,NULL,NULL,NULL);
9105: } else {
9106: Mat change = NULL;
9107: Vec scaling = NULL;
9108: IS change_primal = NULL, iP;
9109: PetscInt benign_n;
9110: PetscBool reuse_solvers = (PetscBool)!pcbddc->use_change_of_basis;
9111: PetscBool need_change = PETSC_FALSE;
9112: PetscBool discrete_harmonic = PETSC_FALSE;
9114: if (!pcbddc->use_vertices && reuse_solvers) {
9115: PetscInt n_vertices;
9117: ISGetLocalSize(sub_schurs->is_vertices,&n_vertices);
9118: reuse_solvers = (PetscBool)!n_vertices;
9119: }
9120: if (!pcbddc->benign_change_explicit) {
9121: benign_n = pcbddc->benign_n;
9122: } else {
9123: benign_n = 0;
9124: }
9125: /* sub_schurs->change is a local object; instead, PCBDDCConstraintsSetUp and the quantities used in the test below are logically collective on pc.
9126: We need a global reduction to avoid possible deadlocks.
9127: We assume that sub_schurs->change is created once, and then reused for different solves, unless the topography has been recomputed */
9128: if (pcbddc->adaptive_userdefined || (pcbddc->deluxe_zerorows && !pcbddc->use_change_of_basis)) {
9129: PetscBool have_loc_change = (PetscBool)(!!sub_schurs->change);
9130: MPIU_Allreduce(&have_loc_change,&need_change,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)pc));
9131: need_change = (PetscBool)(!need_change);
9132: }
9133: /* If the user defines additional constraints, we import them here.
9134: We need to compute the change of basis according to the quadrature weights attached to pmat via MatSetNearNullSpace, and this could not be done (at the moment) without some hacking */
9135: if (need_change) {
9136: PC_IS *pcisf;
9137: PC_BDDC *pcbddcf;
9138: PC pcf;
9140: if (pcbddc->sub_schurs_rebuild) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot compute change of basis with a different graph");
9141: PCCreate(PetscObjectComm((PetscObject)pc),&pcf);
9142: PCSetOperators(pcf,pc->mat,pc->pmat);
9143: PCSetType(pcf,PCBDDC);
9145: /* hacks */
9146: pcisf = (PC_IS*)pcf->data;
9147: pcisf->is_B_local = pcis->is_B_local;
9148: pcisf->vec1_N = pcis->vec1_N;
9149: pcisf->BtoNmap = pcis->BtoNmap;
9150: pcisf->n = pcis->n;
9151: pcisf->n_B = pcis->n_B;
9152: pcbddcf = (PC_BDDC*)pcf->data;
9153: PetscFree(pcbddcf->mat_graph);
9154: pcbddcf->mat_graph = pcbddc->mat_graph;
9155: pcbddcf->use_faces = PETSC_TRUE;
9156: pcbddcf->use_change_of_basis = PETSC_TRUE;
9157: pcbddcf->use_change_on_faces = PETSC_TRUE;
9158: pcbddcf->use_qr_single = PETSC_TRUE;
9159: pcbddcf->fake_change = PETSC_TRUE;
9161: /* setup constraints so that we can get information on primal vertices and change of basis (in local numbering) */
9162: PCBDDCConstraintsSetUp(pcf);
9163: sub_schurs->change_with_qr = pcbddcf->use_qr_single;
9164: ISCreateGeneral(PETSC_COMM_SELF,pcbddcf->n_vertices,pcbddcf->local_primal_ref_node,PETSC_COPY_VALUES,&change_primal);
9165: change = pcbddcf->ConstraintMatrix;
9166: pcbddcf->ConstraintMatrix = NULL;
9168: /* free unneeded memory allocated in PCBDDCConstraintsSetUp */
9169: PetscFree(pcbddcf->sub_schurs);
9170: MatNullSpaceDestroy(&pcbddcf->onearnullspace);
9171: PetscFree2(pcbddcf->local_primal_ref_node,pcbddcf->local_primal_ref_mult);
9172: PetscFree(pcbddcf->primal_indices_local_idxs);
9173: PetscFree(pcbddcf->onearnullvecs_state);
9174: PetscFree(pcf->data);
9175: pcf->ops->destroy = NULL;
9176: pcf->ops->reset = NULL;
9177: PCDestroy(&pcf);
9178: }
9179: if (!pcbddc->use_deluxe_scaling) scaling = pcis->D;
9181: PetscObjectQuery((PetscObject)pc,"__KSPFETIDP_iP",(PetscObject*)&iP);
9182: if (iP) {
9183: PetscOptionsBegin(PetscObjectComm((PetscObject)iP),sub_schurs->prefix,"BDDC sub_schurs options","PC");
9184: PetscOptionsBool("-sub_schurs_discrete_harmonic",NULL,NULL,discrete_harmonic,&discrete_harmonic,NULL);
9185: PetscOptionsEnd();
9186: }
9187: if (discrete_harmonic) {
9188: Mat A;
9189: MatDuplicate(pcbddc->local_mat,MAT_COPY_VALUES,&A);
9190: MatZeroRowsColumnsIS(A,iP,1.0,NULL,NULL);
9191: PetscObjectCompose((PetscObject)A,"__KSPFETIDP_iP",(PetscObject)iP);
9192: PCBDDCSubSchursSetUp(sub_schurs,A,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);
9193: MatDestroy(&A);
9194: } else {
9195: PCBDDCSubSchursSetUp(sub_schurs,pcbddc->local_mat,S_j,pcbddc->sub_schurs_exact_schur,used_xadj,used_adjncy,pcbddc->sub_schurs_layers,scaling,pcbddc->adaptive_selection,reuse_solvers,pcbddc->benign_saddle_point,benign_n,pcbddc->benign_p0_lidx,pcbddc->benign_zerodiag_subs,change,change_primal);
9196: }
9197: MatDestroy(&change);
9198: ISDestroy(&change_primal);
9199: }
9200: MatDestroy(&S_j);
9202: /* free adjacency */
9203: if (free_used_adj) {
9204: PetscFree2(used_xadj,used_adjncy);
9205: }
9206: PetscLogEventEnd(PC_BDDC_Schurs[pcbddc->current_level],pc,0,0,0);
9207: return(0);
9208: }
9210: PetscErrorCode PCBDDCInitSubSchurs(PC pc)
9211: {
9212: PC_IS *pcis=(PC_IS*)pc->data;
9213: PC_BDDC *pcbddc=(PC_BDDC*)pc->data;
9214: PCBDDCGraph graph;
9215: PetscErrorCode ierr;
9218: /* attach interface graph for determining subsets */
9219: if (pcbddc->sub_schurs_rebuild) { /* in case rebuild has been requested, it uses a graph generated only by the neighbouring information */
9220: IS verticesIS,verticescomm;
9221: PetscInt vsize,*idxs;
9223: PCBDDCGraphGetCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);
9224: ISGetSize(verticesIS,&vsize);
9225: ISGetIndices(verticesIS,(const PetscInt**)&idxs);
9226: ISCreateGeneral(PetscObjectComm((PetscObject)pc),vsize,idxs,PETSC_COPY_VALUES,&verticescomm);
9227: ISRestoreIndices(verticesIS,(const PetscInt**)&idxs);
9228: PCBDDCGraphRestoreCandidatesIS(pcbddc->mat_graph,NULL,NULL,NULL,NULL,&verticesIS);
9229: PCBDDCGraphCreate(&graph);
9230: PCBDDCGraphInit(graph,pcbddc->mat_graph->l2gmap,pcbddc->mat_graph->nvtxs_global,pcbddc->graphmaxcount);
9231: PCBDDCGraphSetUp(graph,pcbddc->mat_graph->custom_minimal_size,NULL,pcbddc->DirichletBoundariesLocal,0,NULL,verticescomm);
9232: ISDestroy(&verticescomm);
9233: PCBDDCGraphComputeConnectedComponents(graph);
9234: } else {
9235: graph = pcbddc->mat_graph;
9236: }
9237: /* print some info */
9238: if (pcbddc->dbg_flag && !pcbddc->sub_schurs_rebuild) {
9239: IS vertices;
9240: PetscInt nv,nedges,nfaces;
9241: PCBDDCGraphASCIIView(graph,pcbddc->dbg_flag,pcbddc->dbg_viewer);
9242: PCBDDCGraphGetCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);
9243: ISGetSize(vertices,&nv);
9244: PetscViewerASCIIPushSynchronized(pcbddc->dbg_viewer);
9245: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"--------------------------------------------------------------\n");
9246: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate vertices (%D)\n",PetscGlobalRank,(int)nv,pcbddc->use_vertices);
9247: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate edges (%D)\n",PetscGlobalRank,(int)nedges,pcbddc->use_edges);
9248: PetscViewerASCIISynchronizedPrintf(pcbddc->dbg_viewer,"Subdomain %04d got %02d local candidate faces (%D)\n",PetscGlobalRank,(int)nfaces,pcbddc->use_faces);
9249: PetscViewerFlush(pcbddc->dbg_viewer);
9250: PetscViewerASCIIPopSynchronized(pcbddc->dbg_viewer);
9251: PCBDDCGraphRestoreCandidatesIS(graph,&nfaces,NULL,&nedges,NULL,&vertices);
9252: }
9254: /* sub_schurs init */
9255: if (!pcbddc->sub_schurs) {
9256: PCBDDCSubSchursCreate(&pcbddc->sub_schurs);
9257: }
9258: PCBDDCSubSchursInit(pcbddc->sub_schurs,((PetscObject)pc)->prefix,pcis->is_I_local,pcis->is_B_local,graph,pcis->BtoNmap,pcbddc->sub_schurs_rebuild);
9260: /* free graph struct */
9261: if (pcbddc->sub_schurs_rebuild) {
9262: PCBDDCGraphDestroy(&graph);
9263: }
9264: return(0);
9265: }
9267: PetscErrorCode PCBDDCCheckOperator(PC pc)
9268: {
9269: PC_IS *pcis=(PC_IS*)pc->data;
9270: PC_BDDC *pcbddc=(PC_BDDC*)pc->data;
9271: PetscErrorCode ierr;
9274: if (pcbddc->n_vertices == pcbddc->local_primal_size) {
9275: IS zerodiag = NULL;
9276: Mat S_j,B0_B=NULL;
9277: Vec dummy_vec=NULL,vec_check_B,vec_scale_P;
9278: PetscScalar *p0_check,*array,*array2;
9279: PetscReal norm;
9280: PetscInt i;
9282: /* B0 and B0_B */
9283: if (zerodiag) {
9284: IS dummy;
9286: ISCreateStride(PETSC_COMM_SELF,pcbddc->benign_n,0,1,&dummy);
9287: MatCreateSubMatrix(pcbddc->benign_B0,dummy,pcis->is_B_local,MAT_INITIAL_MATRIX,&B0_B);
9288: MatCreateVecs(B0_B,NULL,&dummy_vec);
9289: ISDestroy(&dummy);
9290: }
9291: /* I need a primal vector to scale primal nodes since BDDC sums contibutions */
9292: VecDuplicate(pcbddc->vec1_P,&vec_scale_P);
9293: VecSet(pcbddc->vec1_P,1.0);
9294: VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);
9295: VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);
9296: VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);
9297: VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,vec_scale_P,INSERT_VALUES,SCATTER_REVERSE);
9298: VecReciprocal(vec_scale_P);
9299: /* S_j */
9300: MatCreateSchurComplement(pcis->A_II,pcis->pA_II,pcis->A_IB,pcis->A_BI,pcis->A_BB,&S_j);
9301: MatSchurComplementSetKSP(S_j,pcbddc->ksp_D);
9303: /* mimic vector in \widetilde{W}_\Gamma */
9304: VecSetRandom(pcis->vec1_N,NULL);
9305: /* continuous in primal space */
9306: VecSetRandom(pcbddc->coarse_vec,NULL);
9307: VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);
9308: VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);
9309: VecGetArray(pcbddc->vec1_P,&array);
9310: PetscCalloc1(pcbddc->benign_n,&p0_check);
9311: for (i=0;i<pcbddc->benign_n;i++) p0_check[i] = array[pcbddc->local_primal_size-pcbddc->benign_n+i];
9312: VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);
9313: VecRestoreArray(pcbddc->vec1_P,&array);
9314: VecAssemblyBegin(pcis->vec1_N);
9315: VecAssemblyEnd(pcis->vec1_N);
9316: VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);
9317: VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec2_B,INSERT_VALUES,SCATTER_FORWARD);
9318: VecDuplicate(pcis->vec2_B,&vec_check_B);
9319: VecCopy(pcis->vec2_B,vec_check_B);
9321: /* assemble rhs for coarse problem */
9322: /* widetilde{S}_\Gamma w_\Gamma + \widetilde{B0}^T_B p0 */
9323: /* local with Schur */
9324: MatMult(S_j,pcis->vec2_B,pcis->vec1_B);
9325: if (zerodiag) {
9326: VecGetArray(dummy_vec,&array);
9327: for (i=0;i<pcbddc->benign_n;i++) array[i] = p0_check[i];
9328: VecRestoreArray(dummy_vec,&array);
9329: MatMultTransposeAdd(B0_B,dummy_vec,pcis->vec1_B,pcis->vec1_B);
9330: }
9331: /* sum on primal nodes the local contributions */
9332: VecScatterBegin(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);
9333: VecScatterEnd(pcis->N_to_B,pcis->vec1_B,pcis->vec1_N,INSERT_VALUES,SCATTER_REVERSE);
9334: VecGetArray(pcis->vec1_N,&array);
9335: VecGetArray(pcbddc->vec1_P,&array2);
9336: for (i=0;i<pcbddc->local_primal_size;i++) array2[i] = array[pcbddc->local_primal_ref_node[i]];
9337: VecRestoreArray(pcbddc->vec1_P,&array2);
9338: VecRestoreArray(pcis->vec1_N,&array);
9339: VecSet(pcbddc->coarse_vec,0.);
9340: VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);
9341: VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->vec1_P,pcbddc->coarse_vec,ADD_VALUES,SCATTER_FORWARD);
9342: VecScatterBegin(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);
9343: VecScatterEnd(pcbddc->coarse_loc_to_glob,pcbddc->coarse_vec,pcbddc->vec1_P,INSERT_VALUES,SCATTER_REVERSE);
9344: VecGetArray(pcbddc->vec1_P,&array);
9345: /* scale primal nodes (BDDC sums contibutions) */
9346: VecPointwiseMult(pcbddc->vec1_P,vec_scale_P,pcbddc->vec1_P);
9347: VecSetValues(pcis->vec1_N,pcbddc->local_primal_size,pcbddc->local_primal_ref_node,array,INSERT_VALUES);
9348: VecRestoreArray(pcbddc->vec1_P,&array);
9349: VecAssemblyBegin(pcis->vec1_N);
9350: VecAssemblyEnd(pcis->vec1_N);
9351: VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
9352: VecScatterEnd(pcis->N_to_B,pcis->vec1_N,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
9353: /* global: \widetilde{B0}_B w_\Gamma */
9354: if (zerodiag) {
9355: MatMult(B0_B,pcis->vec2_B,dummy_vec);
9356: VecGetArray(dummy_vec,&array);
9357: for (i=0;i<pcbddc->benign_n;i++) pcbddc->benign_p0[i] = array[i];
9358: VecRestoreArray(dummy_vec,&array);
9359: }
9360: /* BDDC */
9361: VecSet(pcis->vec1_D,0.);
9362: PCBDDCApplyInterfacePreconditioner(pc,PETSC_FALSE);
9364: VecCopy(pcis->vec1_B,pcis->vec2_B);
9365: VecAXPY(pcis->vec1_B,-1.0,vec_check_B);
9366: VecNorm(pcis->vec1_B,NORM_INFINITY,&norm);
9367: PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC local error is %1.4e\n",PetscGlobalRank,norm);
9368: for (i=0;i<pcbddc->benign_n;i++) {
9369: PetscPrintf(PETSC_COMM_SELF,"[%d] BDDC p0[%D] error is %1.4e\n",PetscGlobalRank,i,PetscAbsScalar(pcbddc->benign_p0[i]-p0_check[i]));
9370: }
9371: PetscFree(p0_check);
9372: VecDestroy(&vec_scale_P);
9373: VecDestroy(&vec_check_B);
9374: VecDestroy(&dummy_vec);
9375: MatDestroy(&S_j);
9376: MatDestroy(&B0_B);
9377: }
9378: return(0);
9379: }
9381: #include <../src/mat/impls/aij/mpi/mpiaij.h>
9382: PetscErrorCode MatMPIAIJRestrict(Mat A, MPI_Comm ccomm, Mat *B)
9383: {
9384: Mat At;
9385: IS rows;
9386: PetscInt rst,ren;
9388: PetscLayout rmap;
9391: rst = ren = 0;
9392: if (ccomm != MPI_COMM_NULL) {
9393: PetscLayoutCreate(ccomm,&rmap);
9394: PetscLayoutSetSize(rmap,A->rmap->N);
9395: PetscLayoutSetBlockSize(rmap,1);
9396: PetscLayoutSetUp(rmap);
9397: PetscLayoutGetRange(rmap,&rst,&ren);
9398: }
9399: ISCreateStride(PetscObjectComm((PetscObject)A),ren-rst,rst,1,&rows);
9400: MatCreateSubMatrix(A,rows,NULL,MAT_INITIAL_MATRIX,&At);
9401: ISDestroy(&rows);
9403: if (ccomm != MPI_COMM_NULL) {
9404: Mat_MPIAIJ *a,*b;
9405: IS from,to;
9406: Vec gvec;
9407: PetscInt lsize;
9409: MatCreate(ccomm,B);
9410: MatSetSizes(*B,ren-rst,PETSC_DECIDE,PETSC_DECIDE,At->cmap->N);
9411: MatSetType(*B,MATAIJ);
9412: PetscLayoutDestroy(&((*B)->rmap));
9413: PetscLayoutSetUp((*B)->cmap);
9414: a = (Mat_MPIAIJ*)At->data;
9415: b = (Mat_MPIAIJ*)(*B)->data;
9416: MPI_Comm_size(ccomm,&b->size);
9417: MPI_Comm_rank(ccomm,&b->rank);
9418: PetscObjectReference((PetscObject)a->A);
9419: PetscObjectReference((PetscObject)a->B);
9420: b->A = a->A;
9421: b->B = a->B;
9423: b->donotstash = a->donotstash;
9424: b->roworiented = a->roworiented;
9425: b->rowindices = 0;
9426: b->rowvalues = 0;
9427: b->getrowactive = PETSC_FALSE;
9429: (*B)->rmap = rmap;
9430: (*B)->factortype = A->factortype;
9431: (*B)->assembled = PETSC_TRUE;
9432: (*B)->insertmode = NOT_SET_VALUES;
9433: (*B)->preallocated = PETSC_TRUE;
9435: if (a->colmap) {
9436: #if defined(PETSC_USE_CTABLE)
9437: PetscTableCreateCopy(a->colmap,&b->colmap);
9438: #else
9439: PetscMalloc1(At->cmap->N,&b->colmap);
9440: PetscLogObjectMemory((PetscObject)*B,At->cmap->N*sizeof(PetscInt));
9441: PetscArraycpy(b->colmap,a->colmap,At->cmap->N);
9442: #endif
9443: } else b->colmap = 0;
9444: if (a->garray) {
9445: PetscInt len;
9446: len = a->B->cmap->n;
9447: PetscMalloc1(len+1,&b->garray);
9448: PetscLogObjectMemory((PetscObject)(*B),len*sizeof(PetscInt));
9449: if (len) { PetscArraycpy(b->garray,a->garray,len); }
9450: } else b->garray = 0;
9452: PetscObjectReference((PetscObject)a->lvec);
9453: b->lvec = a->lvec;
9454: PetscLogObjectParent((PetscObject)*B,(PetscObject)b->lvec);
9456: /* cannot use VecScatterCopy */
9457: VecGetLocalSize(b->lvec,&lsize);
9458: ISCreateGeneral(ccomm,lsize,b->garray,PETSC_USE_POINTER,&from);
9459: ISCreateStride(PETSC_COMM_SELF,lsize,0,1,&to);
9460: MatCreateVecs(*B,&gvec,NULL);
9461: VecScatterCreate(gvec,from,b->lvec,to,&b->Mvctx);
9462: PetscLogObjectParent((PetscObject)*B,(PetscObject)b->Mvctx);
9463: ISDestroy(&from);
9464: ISDestroy(&to);
9465: VecDestroy(&gvec);
9466: }
9467: MatDestroy(&At);
9468: return(0);
9469: }