Actual source code: redundant.c
petsc-3.10.5 2019-03-28
2: /*
3: This file defines a "solve the problem redundantly on each subgroup of processor" preconditioner.
4: */
5: #include <petsc/private/pcimpl.h>
6: #include <petscksp.h>
8: typedef struct {
9: KSP ksp;
10: PC pc; /* actual preconditioner used on each processor */
11: Vec xsub,ysub; /* vectors of a subcommunicator to hold parallel vectors of PetscObjectComm((PetscObject)pc) */
12: Vec xdup,ydup; /* parallel vector that congregates xsub or ysub facilitating vector scattering */
13: Mat pmats; /* matrix and optional preconditioner matrix belong to a subcommunicator */
14: VecScatter scatterin,scatterout; /* scatter used to move all values to each processor group (subcommunicator) */
15: PetscBool useparallelmat;
16: PetscSubcomm psubcomm;
17: PetscInt nsubcomm; /* num of data structure PetscSubcomm */
18: PetscBool shifttypeset;
19: MatFactorShiftType shifttype;
20: } PC_Redundant;
22: PetscErrorCode PCFactorSetShiftType_Redundant(PC pc,MatFactorShiftType shifttype)
23: {
24: PC_Redundant *red = (PC_Redundant*)pc->data;
28: if (red->ksp) {
29: PC pc;
30: KSPGetPC(red->ksp,&pc);
31: PCFactorSetShiftType(pc,shifttype);
32: } else {
33: red->shifttypeset = PETSC_TRUE;
34: red->shifttype = shifttype;
35: }
36: return(0);
37: }
39: static PetscErrorCode PCView_Redundant(PC pc,PetscViewer viewer)
40: {
41: PC_Redundant *red = (PC_Redundant*)pc->data;
43: PetscBool iascii,isstring;
44: PetscViewer subviewer;
47: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
48: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSTRING,&isstring);
49: if (iascii) {
50: if (!red->psubcomm) {
51: PetscViewerASCIIPrintf(viewer," Not yet setup\n");
52: } else {
53: PetscViewerASCIIPrintf(viewer," First (color=0) of %D PCs follows\n",red->nsubcomm);
54: PetscViewerGetSubViewer(viewer,((PetscObject)red->pc)->comm,&subviewer);
55: if (!red->psubcomm->color) { /* only view first redundant pc */
56: PetscViewerASCIIPushTab(subviewer);
57: KSPView(red->ksp,subviewer);
58: PetscViewerASCIIPopTab(subviewer);
59: }
60: PetscViewerRestoreSubViewer(viewer,((PetscObject)red->pc)->comm,&subviewer);
61: }
62: } else if (isstring) {
63: PetscViewerStringSPrintf(viewer," Redundant solver preconditioner");
64: }
65: return(0);
66: }
68: #include <../src/mat/impls/aij/mpi/mpiaij.h>
69: static PetscErrorCode PCSetUp_Redundant(PC pc)
70: {
71: PC_Redundant *red = (PC_Redundant*)pc->data;
73: PetscInt mstart,mend,mlocal,M;
74: PetscMPIInt size;
75: MPI_Comm comm,subcomm;
76: Vec x;
79: PetscObjectGetComm((PetscObject)pc,&comm);
81: /* if pmatrix set by user is sequential then we do not need to gather the parallel matrix */
82: MPI_Comm_size(comm,&size);
83: if (size == 1) red->useparallelmat = PETSC_FALSE;
85: if (!pc->setupcalled) {
86: PetscInt mloc_sub;
87: if (!red->psubcomm) { /* create red->psubcomm, new ksp and pc over subcomm */
88: KSP ksp;
89: PCRedundantGetKSP(pc,&ksp);
90: }
91: subcomm = PetscSubcommChild(red->psubcomm);
93: if (red->useparallelmat) {
94: /* grab the parallel matrix and put it into processors of a subcomminicator */
95: MatCreateRedundantMatrix(pc->pmat,red->psubcomm->n,subcomm,MAT_INITIAL_MATRIX,&red->pmats);
97: MPI_Comm_size(subcomm,&size);
98: if (size > 1) {
99: PetscBool foundpack;
100: MatGetFactorAvailable(red->pmats,NULL,MAT_FACTOR_LU,&foundpack);
101: if (!foundpack) { /* reset default ksp and pc */
102: KSPSetType(red->ksp,KSPGMRES);
103: PCSetType(red->pc,PCBJACOBI);
104: } else {
105: PCFactorSetMatSolverType(red->pc,NULL);
106: }
107: }
109: KSPSetOperators(red->ksp,red->pmats,red->pmats);
111: /* get working vectors xsub and ysub */
112: MatCreateVecs(red->pmats,&red->xsub,&red->ysub);
114: /* create working vectors xdup and ydup.
115: xdup concatenates all xsub's contigously to form a mpi vector over dupcomm (see PetscSubcommCreate_interlaced())
116: ydup concatenates all ysub and has empty local arrays because ysub's arrays will be place into it.
117: Note: we use communicator dupcomm, not PetscObjectComm((PetscObject)pc)! */
118: MatGetLocalSize(red->pmats,&mloc_sub,NULL);
119: VecCreateMPI(PetscSubcommContiguousParent(red->psubcomm),mloc_sub,PETSC_DECIDE,&red->xdup);
120: VecCreateMPIWithArray(PetscSubcommContiguousParent(red->psubcomm),1,mloc_sub,PETSC_DECIDE,NULL,&red->ydup);
122: /* create vecscatters */
123: if (!red->scatterin) { /* efficiency of scatterin is independent from psubcomm_type! */
124: IS is1,is2;
125: PetscInt *idx1,*idx2,i,j,k;
127: MatCreateVecs(pc->pmat,&x,0);
128: VecGetSize(x,&M);
129: VecGetOwnershipRange(x,&mstart,&mend);
130: mlocal = mend - mstart;
131: PetscMalloc2(red->psubcomm->n*mlocal,&idx1,red->psubcomm->n*mlocal,&idx2);
132: j = 0;
133: for (k=0; k<red->psubcomm->n; k++) {
134: for (i=mstart; i<mend; i++) {
135: idx1[j] = i;
136: idx2[j++] = i + M*k;
137: }
138: }
139: ISCreateGeneral(comm,red->psubcomm->n*mlocal,idx1,PETSC_COPY_VALUES,&is1);
140: ISCreateGeneral(comm,red->psubcomm->n*mlocal,idx2,PETSC_COPY_VALUES,&is2);
141: VecScatterCreate(x,is1,red->xdup,is2,&red->scatterin);
142: ISDestroy(&is1);
143: ISDestroy(&is2);
145: /* Impl below is good for PETSC_SUBCOMM_INTERLACED (no inter-process communication) and PETSC_SUBCOMM_CONTIGUOUS (communication within subcomm) */
146: ISCreateStride(comm,mlocal,mstart+ red->psubcomm->color*M,1,&is1);
147: ISCreateStride(comm,mlocal,mstart,1,&is2);
148: VecScatterCreate(red->xdup,is1,x,is2,&red->scatterout);
149: ISDestroy(&is1);
150: ISDestroy(&is2);
151: PetscFree2(idx1,idx2);
152: VecDestroy(&x);
153: }
154: } else { /* !red->useparallelmat */
155: KSPSetOperators(red->ksp,pc->mat,pc->pmat);
156: }
157: } else { /* pc->setupcalled */
158: if (red->useparallelmat) {
159: MatReuse reuse;
160: /* grab the parallel matrix and put it into processors of a subcomminicator */
161: /*--------------------------------------------------------------------------*/
162: if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
163: /* destroy old matrices */
164: MatDestroy(&red->pmats);
165: reuse = MAT_INITIAL_MATRIX;
166: } else {
167: reuse = MAT_REUSE_MATRIX;
168: }
169: MatCreateRedundantMatrix(pc->pmat,red->psubcomm->n,PetscSubcommChild(red->psubcomm),reuse,&red->pmats);
170: KSPSetOperators(red->ksp,red->pmats,red->pmats);
171: } else { /* !red->useparallelmat */
172: KSPSetOperators(red->ksp,pc->mat,pc->pmat);
173: }
174: }
176: if (pc->setfromoptionscalled) {
177: KSPSetFromOptions(red->ksp);
178: }
179: KSPSetUp(red->ksp);
180: return(0);
181: }
183: static PetscErrorCode PCApply_Redundant(PC pc,Vec x,Vec y)
184: {
185: PC_Redundant *red = (PC_Redundant*)pc->data;
187: PetscScalar *array;
190: if (!red->useparallelmat) {
191: KSPSolve(red->ksp,x,y);
192: return(0);
193: }
195: /* scatter x to xdup */
196: VecScatterBegin(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);
197: VecScatterEnd(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);
199: /* place xdup's local array into xsub */
200: VecGetArray(red->xdup,&array);
201: VecPlaceArray(red->xsub,(const PetscScalar*)array);
203: /* apply preconditioner on each processor */
204: KSPSolve(red->ksp,red->xsub,red->ysub);
205: VecResetArray(red->xsub);
206: VecRestoreArray(red->xdup,&array);
208: /* place ysub's local array into ydup */
209: VecGetArray(red->ysub,&array);
210: VecPlaceArray(red->ydup,(const PetscScalar*)array);
212: /* scatter ydup to y */
213: VecScatterBegin(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);
214: VecScatterEnd(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);
215: VecResetArray(red->ydup);
216: VecRestoreArray(red->ysub,&array);
217: return(0);
218: }
220: static PetscErrorCode PCApplyTranspose_Redundant(PC pc,Vec x,Vec y)
221: {
222: PC_Redundant *red = (PC_Redundant*)pc->data;
224: PetscScalar *array;
227: if (!red->useparallelmat) {
228: KSPSolveTranspose(red->ksp,x,y);
229: return(0);
230: }
232: /* scatter x to xdup */
233: VecScatterBegin(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);
234: VecScatterEnd(red->scatterin,x,red->xdup,INSERT_VALUES,SCATTER_FORWARD);
236: /* place xdup's local array into xsub */
237: VecGetArray(red->xdup,&array);
238: VecPlaceArray(red->xsub,(const PetscScalar*)array);
240: /* apply preconditioner on each processor */
241: KSPSolveTranspose(red->ksp,red->xsub,red->ysub);
242: VecResetArray(red->xsub);
243: VecRestoreArray(red->xdup,&array);
245: /* place ysub's local array into ydup */
246: VecGetArray(red->ysub,&array);
247: VecPlaceArray(red->ydup,(const PetscScalar*)array);
249: /* scatter ydup to y */
250: VecScatterBegin(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);
251: VecScatterEnd(red->scatterout,red->ydup,y,INSERT_VALUES,SCATTER_FORWARD);
252: VecResetArray(red->ydup);
253: VecRestoreArray(red->ysub,&array);
254: return(0);
255: }
257: static PetscErrorCode PCReset_Redundant(PC pc)
258: {
259: PC_Redundant *red = (PC_Redundant*)pc->data;
263: if (red->useparallelmat) {
264: VecScatterDestroy(&red->scatterin);
265: VecScatterDestroy(&red->scatterout);
266: VecDestroy(&red->ysub);
267: VecDestroy(&red->xsub);
268: VecDestroy(&red->xdup);
269: VecDestroy(&red->ydup);
270: }
271: MatDestroy(&red->pmats);
272: KSPReset(red->ksp);
273: return(0);
274: }
276: static PetscErrorCode PCDestroy_Redundant(PC pc)
277: {
278: PC_Redundant *red = (PC_Redundant*)pc->data;
282: PCReset_Redundant(pc);
283: KSPDestroy(&red->ksp);
284: PetscSubcommDestroy(&red->psubcomm);
285: PetscFree(pc->data);
286: return(0);
287: }
289: static PetscErrorCode PCSetFromOptions_Redundant(PetscOptionItems *PetscOptionsObject,PC pc)
290: {
292: PC_Redundant *red = (PC_Redundant*)pc->data;
295: PetscOptionsHead(PetscOptionsObject,"Redundant options");
296: PetscOptionsInt("-pc_redundant_number","Number of redundant pc","PCRedundantSetNumber",red->nsubcomm,&red->nsubcomm,0);
297: PetscOptionsTail();
298: return(0);
299: }
301: static PetscErrorCode PCRedundantSetNumber_Redundant(PC pc,PetscInt nreds)
302: {
303: PC_Redundant *red = (PC_Redundant*)pc->data;
306: red->nsubcomm = nreds;
307: return(0);
308: }
310: /*@
311: PCRedundantSetNumber - Sets the number of redundant preconditioner contexts.
313: Logically Collective on PC
315: Input Parameters:
316: + pc - the preconditioner context
317: - nredundant - number of redundant preconditioner contexts; for example if you are using 64 MPI processes and
318: use an nredundant of 4 there will be 4 parallel solves each on 16 = 64/4 processes.
320: Level: advanced
322: .keywords: PC, redundant solve
323: @*/
324: PetscErrorCode PCRedundantSetNumber(PC pc,PetscInt nredundant)
325: {
330: if (nredundant <= 0) SETERRQ1(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_WRONG, "num of redundant pc %D must be positive",nredundant);
331: PetscTryMethod(pc,"PCRedundantSetNumber_C",(PC,PetscInt),(pc,nredundant));
332: return(0);
333: }
335: static PetscErrorCode PCRedundantSetScatter_Redundant(PC pc,VecScatter in,VecScatter out)
336: {
337: PC_Redundant *red = (PC_Redundant*)pc->data;
341: PetscObjectReference((PetscObject)in);
342: VecScatterDestroy(&red->scatterin);
344: red->scatterin = in;
346: PetscObjectReference((PetscObject)out);
347: VecScatterDestroy(&red->scatterout);
348: red->scatterout = out;
349: return(0);
350: }
352: /*@
353: PCRedundantSetScatter - Sets the scatter used to copy values into the
354: redundant local solve and the scatter to move them back into the global
355: vector.
357: Logically Collective on PC
359: Input Parameters:
360: + pc - the preconditioner context
361: . in - the scatter to move the values in
362: - out - the scatter to move them out
364: Level: advanced
366: .keywords: PC, redundant solve
367: @*/
368: PetscErrorCode PCRedundantSetScatter(PC pc,VecScatter in,VecScatter out)
369: {
376: PetscTryMethod(pc,"PCRedundantSetScatter_C",(PC,VecScatter,VecScatter),(pc,in,out));
377: return(0);
378: }
380: static PetscErrorCode PCRedundantGetKSP_Redundant(PC pc,KSP *innerksp)
381: {
383: PC_Redundant *red = (PC_Redundant*)pc->data;
384: MPI_Comm comm,subcomm;
385: const char *prefix;
388: if (!red->psubcomm) {
389: PCGetOptionsPrefix(pc,&prefix);
391: PetscObjectGetComm((PetscObject)pc,&comm);
392: PetscSubcommCreate(comm,&red->psubcomm);
393: PetscSubcommSetNumber(red->psubcomm,red->nsubcomm);
394: PetscSubcommSetType(red->psubcomm,PETSC_SUBCOMM_CONTIGUOUS);
396: PetscSubcommSetOptionsPrefix(red->psubcomm,prefix);
397: PetscSubcommSetFromOptions(red->psubcomm);
398: PetscLogObjectMemory((PetscObject)pc,sizeof(PetscSubcomm));
400: /* create a new PC that processors in each subcomm have copy of */
401: subcomm = PetscSubcommChild(red->psubcomm);
403: KSPCreate(subcomm,&red->ksp);
404: KSPSetErrorIfNotConverged(red->ksp,pc->erroriffailure);
405: PetscObjectIncrementTabLevel((PetscObject)red->ksp,(PetscObject)pc,1);
406: PetscLogObjectParent((PetscObject)pc,(PetscObject)red->ksp);
407: KSPSetType(red->ksp,KSPPREONLY);
408: KSPGetPC(red->ksp,&red->pc);
409: PCSetType(red->pc,PCLU);
410: if (red->shifttypeset) {
411: PCFactorSetShiftType(red->pc,red->shifttype);
412: red->shifttypeset = PETSC_FALSE;
413: }
414: KSPSetOptionsPrefix(red->ksp,prefix);
415: KSPAppendOptionsPrefix(red->ksp,"redundant_");
416: }
417: *innerksp = red->ksp;
418: return(0);
419: }
421: /*@
422: PCRedundantGetKSP - Gets the less parallel KSP created by the redundant PC.
424: Not Collective
426: Input Parameter:
427: . pc - the preconditioner context
429: Output Parameter:
430: . innerksp - the KSP on the smaller set of processes
432: Level: advanced
434: .keywords: PC, redundant solve
435: @*/
436: PetscErrorCode PCRedundantGetKSP(PC pc,KSP *innerksp)
437: {
443: PetscUseMethod(pc,"PCRedundantGetKSP_C",(PC,KSP*),(pc,innerksp));
444: return(0);
445: }
447: static PetscErrorCode PCRedundantGetOperators_Redundant(PC pc,Mat *mat,Mat *pmat)
448: {
449: PC_Redundant *red = (PC_Redundant*)pc->data;
452: if (mat) *mat = red->pmats;
453: if (pmat) *pmat = red->pmats;
454: return(0);
455: }
457: /*@
458: PCRedundantGetOperators - gets the sequential matrix and preconditioner matrix
460: Not Collective
462: Input Parameter:
463: . pc - the preconditioner context
465: Output Parameters:
466: + mat - the matrix
467: - pmat - the (possibly different) preconditioner matrix
469: Level: advanced
471: .keywords: PC, redundant solve
472: @*/
473: PetscErrorCode PCRedundantGetOperators(PC pc,Mat *mat,Mat *pmat)
474: {
481: PetscUseMethod(pc,"PCRedundantGetOperators_C",(PC,Mat*,Mat*),(pc,mat,pmat));
482: return(0);
483: }
485: /* -------------------------------------------------------------------------------------*/
486: /*MC
487: PCREDUNDANT - Runs a KSP solver with preconditioner for the entire problem on subgroups of processors
489: Options for the redundant preconditioners can be set with -redundant_pc_xxx for the redundant KSP with -redundant_ksp_xxx
491: Options Database:
492: . -pc_redundant_number <n> - number of redundant solves, for example if you are using 64 MPI processes and
493: use an n of 4 there will be 4 parallel solves each on 16 = 64/4 processes.
495: Level: intermediate
497: Notes:
498: The default KSP is preonly and the default PC is LU.
500: PCFactorSetShiftType() applied to this PC will convey they shift type into the inner PC if it is factorization based.
502: Developer Notes:
503: Note that PCSetInitialGuessNonzero() is not used by this class but likely should be.
505: .seealso: PCCreate(), PCSetType(), PCType (for list of available types), PCRedundantSetScatter(),
506: PCRedundantGetKSP(), PCRedundantGetOperators(), PCRedundantSetNumber()
507: M*/
509: PETSC_EXTERN PetscErrorCode PCCreate_Redundant(PC pc)
510: {
512: PC_Redundant *red;
513: PetscMPIInt size;
516: PetscNewLog(pc,&red);
517: MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);
519: red->nsubcomm = size;
520: red->useparallelmat = PETSC_TRUE;
521: pc->data = (void*)red;
523: pc->ops->apply = PCApply_Redundant;
524: pc->ops->applytranspose = PCApplyTranspose_Redundant;
525: pc->ops->setup = PCSetUp_Redundant;
526: pc->ops->destroy = PCDestroy_Redundant;
527: pc->ops->reset = PCReset_Redundant;
528: pc->ops->setfromoptions = PCSetFromOptions_Redundant;
529: pc->ops->view = PCView_Redundant;
531: PetscObjectComposeFunction((PetscObject)pc,"PCRedundantSetScatter_C",PCRedundantSetScatter_Redundant);
532: PetscObjectComposeFunction((PetscObject)pc,"PCRedundantSetNumber_C",PCRedundantSetNumber_Redundant);
533: PetscObjectComposeFunction((PetscObject)pc,"PCRedundantGetKSP_C",PCRedundantGetKSP_Redundant);
534: PetscObjectComposeFunction((PetscObject)pc,"PCRedundantGetOperators_C",PCRedundantGetOperators_Redundant);
535: PetscObjectComposeFunction((PetscObject)pc,"PCFactorSetShiftType_C",PCFactorSetShiftType_Redundant);
536: return(0);
537: }