Actual source code: ciss.c
slepc-3.13.1 2020-04-12
1: /*
2: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3: SLEPc - Scalable Library for Eigenvalue Problem Computations
4: Copyright (c) 2002-2020, Universitat Politecnica de Valencia, Spain
6: This file is part of SLEPc.
7: SLEPc is distributed under a 2-clause BSD license (see LICENSE).
8: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
9: */
10: /*
11: SLEPc eigensolver: "ciss"
13: Method: Contour Integral Spectral Slicing
15: Algorithm:
17: Contour integral based on Sakurai-Sugiura method to construct a
18: subspace, with various eigenpair extractions (Rayleigh-Ritz,
19: explicit moment).
21: Based on code contributed by Y. Maeda, T. Sakurai.
23: References:
25: [1] T. Sakurai and H. Sugiura, "A projection method for generalized
26: eigenvalue problems", J. Comput. Appl. Math. 159:119-128, 2003.
28: [2] T. Sakurai and H. Tadano, "CIRR: a Rayleigh-Ritz type method with
29: contour integral for generalized eigenvalue problems", Hokkaido
30: Math. J. 36:745-757, 2007.
31: */
33: #include <slepc/private/epsimpl.h> /*I "slepceps.h" I*/
34: #include <slepcblaslapack.h>
36: typedef struct {
37: /* parameters */
38: PetscInt N; /* number of integration points (32) */
39: PetscInt L; /* block size (16) */
40: PetscInt M; /* moment degree (N/4 = 4) */
41: PetscReal delta; /* threshold of singular value (1e-12) */
42: PetscInt L_max; /* maximum number of columns of the source matrix V */
43: PetscReal spurious_threshold; /* discard spurious eigenpairs */
44: PetscBool isreal; /* A and B are real */
45: PetscInt npart; /* number of partitions */
46: PetscInt refine_inner;
47: PetscInt refine_blocksize;
48: /* private data */
49: PetscReal *sigma; /* threshold for numerical rank */
50: PetscInt subcomm_id;
51: PetscInt num_solve_point;
52: PetscScalar *weight;
53: PetscScalar *omega;
54: PetscScalar *pp;
55: BV V;
56: BV S;
57: BV pV;
58: BV Y;
59: Vec xsub;
60: Vec xdup;
61: KSP *ksp; /* ksp array for storing factorizations at integration points */
62: PetscBool useconj;
63: PetscReal est_eig;
64: VecScatter scatterin;
65: Mat pA,pB;
66: PetscSubcomm subcomm;
67: PetscBool usest;
68: PetscBool usest_set; /* whether the user set the usest flag or not */
69: EPSCISSQuadRule quad;
70: EPSCISSExtraction extraction;
71: } EPS_CISS;
73: /* destroy KSP objects when the number of solve points changes */
74: PETSC_STATIC_INLINE PetscErrorCode EPSCISSResetSolvers(EPS eps)
75: {
77: PetscInt i;
78: EPS_CISS *ctx = (EPS_CISS*)eps->data;
81: if (ctx->ksp) {
82: for (i=0;i<ctx->num_solve_point;i++) {
83: KSPDestroy(&ctx->ksp[i]);
84: }
85: PetscFree(ctx->ksp);
86: }
87: return(0);
88: }
90: /* clean PetscSubcomm object when the number of partitions changes */
91: PETSC_STATIC_INLINE PetscErrorCode EPSCISSResetSubcomm(EPS eps)
92: {
94: EPS_CISS *ctx = (EPS_CISS*)eps->data;
97: EPSCISSResetSolvers(eps);
98: PetscSubcommDestroy(&ctx->subcomm);
99: return(0);
100: }
102: /* determine whether half of integration points can be avoided (use its conjugates);
103: depends on isreal and the center of the region */
104: PETSC_STATIC_INLINE PetscErrorCode EPSCISSSetUseConj(EPS eps,PetscBool *useconj)
105: {
107: PetscScalar center;
108: PetscReal c,d;
109: PetscBool isellipse,isinterval;
110: #if defined(PETSC_USE_COMPLEX)
111: EPS_CISS *ctx = (EPS_CISS*)eps->data;
112: #endif
115: *useconj = PETSC_FALSE;
116: PetscObjectTypeCompare((PetscObject)eps->rg,RGELLIPSE,&isellipse);
117: if (isellipse) {
118: RGEllipseGetParameters(eps->rg,¢er,NULL,NULL);
119: #if defined(PETSC_USE_COMPLEX)
120: *useconj = (ctx->isreal && PetscImaginaryPart(center) == 0.0)? PETSC_TRUE: PETSC_FALSE;
121: #endif
122: } else {
123: PetscObjectTypeCompare((PetscObject)eps->rg,RGINTERVAL,&isinterval);
124: if (isinterval) {
125: RGIntervalGetEndpoints(eps->rg,NULL,NULL,&c,&d);
126: #if defined(PETSC_USE_COMPLEX)
127: *useconj = (ctx->isreal && c==d)? PETSC_TRUE: PETSC_FALSE;
128: #endif
129: }
130: }
131: return(0);
132: }
134: /* create PetscSubcomm object and determine num_solve_point (depends on npart, N, useconj) */
135: PETSC_STATIC_INLINE PetscErrorCode EPSCISSSetUpSubComm(EPS eps,PetscInt *num_solve_point)
136: {
138: EPS_CISS *ctx = (EPS_CISS*)eps->data;
139: PetscInt N = ctx->N;
142: PetscSubcommCreate(PetscObjectComm((PetscObject)eps),&ctx->subcomm);
143: PetscSubcommSetNumber(ctx->subcomm,ctx->npart);
144: PetscSubcommSetType(ctx->subcomm,PETSC_SUBCOMM_INTERLACED);
145: PetscLogObjectMemory((PetscObject)eps,sizeof(PetscSubcomm));
146: ctx->subcomm_id = ctx->subcomm->color;
147: EPSCISSSetUseConj(eps,&ctx->useconj);
148: if (ctx->useconj) N = N/2;
149: *num_solve_point = N / ctx->npart;
150: if (N%ctx->npart > ctx->subcomm_id) (*num_solve_point)++;
151: return(0);
152: }
154: static PetscErrorCode CISSRedundantMat(EPS eps)
155: {
157: EPS_CISS *ctx = (EPS_CISS*)eps->data;
158: Mat A,B;
159: PetscInt nmat;
162: STGetNumMatrices(eps->st,&nmat);
163: if (ctx->subcomm->n != 1) {
164: STGetMatrix(eps->st,0,&A);
165: MatDestroy(&ctx->pA);
166: MatCreateRedundantMatrix(A,ctx->subcomm->n,PetscSubcommChild(ctx->subcomm),MAT_INITIAL_MATRIX,&ctx->pA);
167: if (nmat>1) {
168: STGetMatrix(eps->st,1,&B);
169: MatDestroy(&ctx->pB);
170: MatCreateRedundantMatrix(B,ctx->subcomm->n,PetscSubcommChild(ctx->subcomm),MAT_INITIAL_MATRIX,&ctx->pB);
171: } else ctx->pB = NULL;
172: } else {
173: ctx->pA = NULL;
174: ctx->pB = NULL;
175: }
176: return(0);
177: }
179: static PetscErrorCode CISSScatterVec(EPS eps)
180: {
182: EPS_CISS *ctx = (EPS_CISS*)eps->data;
183: IS is1,is2;
184: Vec v0;
185: PetscInt i,j,k,mstart,mend,mlocal;
186: PetscInt *idx1,*idx2,mloc_sub;
189: VecDestroy(&ctx->xsub);
190: MatCreateVecs(ctx->pA,&ctx->xsub,NULL);
192: VecDestroy(&ctx->xdup);
193: MatGetLocalSize(ctx->pA,&mloc_sub,NULL);
194: VecCreateMPI(PetscSubcommContiguousParent(ctx->subcomm),mloc_sub,PETSC_DECIDE,&ctx->xdup);
196: VecScatterDestroy(&ctx->scatterin);
197: BVGetColumn(ctx->V,0,&v0);
198: VecGetOwnershipRange(v0,&mstart,&mend);
199: mlocal = mend - mstart;
200: PetscMalloc2(ctx->subcomm->n*mlocal,&idx1,ctx->subcomm->n*mlocal,&idx2);
201: j = 0;
202: for (k=0;k<ctx->subcomm->n;k++) {
203: for (i=mstart;i<mend;i++) {
204: idx1[j] = i;
205: idx2[j++] = i + eps->n*k;
206: }
207: }
208: ISCreateGeneral(PetscObjectComm((PetscObject)eps),ctx->subcomm->n*mlocal,idx1,PETSC_COPY_VALUES,&is1);
209: ISCreateGeneral(PetscObjectComm((PetscObject)eps),ctx->subcomm->n*mlocal,idx2,PETSC_COPY_VALUES,&is2);
210: VecScatterCreate(v0,is1,ctx->xdup,is2,&ctx->scatterin);
211: ISDestroy(&is1);
212: ISDestroy(&is2);
213: PetscFree2(idx1,idx2);
214: BVRestoreColumn(ctx->V,0,&v0);
215: return(0);
216: }
218: static PetscErrorCode SetPathParameter(EPS eps)
219: {
221: EPS_CISS *ctx = (EPS_CISS*)eps->data;
222: PetscInt i,j;
223: PetscScalar center=0.0,tmp,tmp2,*omegai;
224: PetscReal theta,radius=1.0,vscale,a,b,c,d,max_w=0.0,rgscale;
225: #if defined(PETSC_USE_COMPLEX)
226: PetscReal start_ang,end_ang;
227: #endif
228: PetscBool isring=PETSC_FALSE,isellipse=PETSC_FALSE,isinterval=PETSC_FALSE;
231: PetscObjectTypeCompare((PetscObject)eps->rg,RGELLIPSE,&isellipse);
232: PetscObjectTypeCompare((PetscObject)eps->rg,RGRING,&isring);
233: PetscObjectTypeCompare((PetscObject)eps->rg,RGINTERVAL,&isinterval);
234: RGGetScale(eps->rg,&rgscale);
235: PetscMalloc1(ctx->N+1l,&omegai);
236: RGComputeContour(eps->rg,ctx->N,ctx->omega,omegai);
237: if (isellipse) {
238: RGEllipseGetParameters(eps->rg,¢er,&radius,&vscale);
239: for (i=0;i<ctx->N;i++) {
240: #if defined(PETSC_USE_COMPLEX)
241: theta = 2.0*PETSC_PI*(i+0.5)/ctx->N;
242: ctx->pp[i] = PetscCMPLX(PetscCosReal(theta),vscale*PetscSinReal(theta));
243: ctx->weight[i] = rgscale*radius*(PetscCMPLX(vscale*PetscCosReal(theta),PetscSinReal(theta)))/(PetscReal)ctx->N;
244: #else
245: theta = (PETSC_PI/ctx->N)*(i+0.5);
246: ctx->pp[i] = PetscCosReal(theta);
247: ctx->weight[i] = PetscCosReal((ctx->N-1)*theta)/ctx->N;
248: ctx->omega[i] = rgscale*(center + radius*ctx->pp[i]);
249: #endif
250: }
251: } else if (ctx->quad == EPS_CISS_QUADRULE_CHEBYSHEV) {
252: for (i=0;i<ctx->N;i++) {
253: theta = (PETSC_PI/ctx->N)*(i+0.5);
254: ctx->pp[i] = PetscCosReal(theta);
255: ctx->weight[i] = PetscCosReal((ctx->N-1)*theta)/ctx->N;
256: }
257: if (isinterval) {
258: RGIntervalGetEndpoints(eps->rg,&a,&b,&c,&d);
259: if ((c!=d || c!=0.0) && (a!=b || a!=0.0)) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"Endpoints of the imaginary axis or the real axis must be both zero");
260: for (i=0;i<ctx->N;i++) {
261: if (c==d) ctx->omega[i] = ((b-a)*(ctx->pp[i]+1.0)/2.0+a)*rgscale;
262: if (a==b) {
263: #if defined(PETSC_USE_COMPLEX)
264: ctx->omega[i] = ((d-c)*(ctx->pp[i]+1.0)/2.0+c)*rgscale*PETSC_i;
265: #else
266: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Integration points on a vertical line require complex arithmetic");
267: #endif
268: }
269: }
270: }
271: if (isring) { /* only supported in complex scalars */
272: #if defined(PETSC_USE_COMPLEX)
273: RGRingGetParameters(eps->rg,¢er,&radius,&vscale,&start_ang,&end_ang,NULL);
274: for (i=0;i<ctx->N;i++) {
275: theta = (start_ang*2.0+(end_ang-start_ang)*(PetscRealPart(ctx->pp[i])+1.0))*PETSC_PI;
276: ctx->omega[i] = rgscale*(center + radius*PetscCMPLX(PetscCosReal(theta),vscale*PetscSinReal(theta)));
277: }
278: #endif
279: }
280: } else {
281: if (isinterval) {
282: RGIntervalGetEndpoints(eps->rg,&a,&b,&c,&d);
283: center = rgscale*((b+a)/2.0+(d+c)/2.0*PETSC_PI);
284: radius = PetscSqrtReal(PetscPowRealInt(rgscale*(b-a)/2.0,2)+PetscPowRealInt(rgscale*(d-c)/2.0,2));
285: } else if (isring) {
286: RGRingGetParameters(eps->rg,¢er,&radius,NULL,NULL,NULL,NULL);
287: center *= rgscale;
288: radius *= rgscale;
289: }
290: for (i=0;i<ctx->N;i++) {
291: ctx->pp[i] = (ctx->omega[i]-center)/radius;
292: tmp = 1; tmp2 = 1;
293: for (j=0;j<ctx->N;j++) {
294: tmp *= ctx->omega[j];
295: if (i != j) tmp2 *= ctx->omega[j]-ctx->omega[i];
296: }
297: ctx->weight[i] = tmp/tmp2;
298: max_w = PetscMax(PetscAbsScalar(ctx->weight[i]),max_w);
299: }
300: for (i=0;i<ctx->N;i++) ctx->weight[i] /= (PetscScalar)max_w;
301: }
302: PetscFree(omegai);
303: return(0);
304: }
306: static PetscErrorCode CISSVecSetRandom(BV V,PetscInt i0,PetscInt i1)
307: {
309: PetscInt i,j,nlocal;
310: PetscScalar *vdata;
311: Vec x;
314: BVGetSizes(V,&nlocal,NULL,NULL);
315: for (i=i0;i<i1;i++) {
316: BVSetRandomColumn(V,i);
317: BVGetColumn(V,i,&x);
318: VecGetArray(x,&vdata);
319: for (j=0;j<nlocal;j++) {
320: vdata[j] = PetscRealPart(vdata[j]);
321: if (PetscRealPart(vdata[j]) < 0.5) vdata[j] = -1.0;
322: else vdata[j] = 1.0;
323: }
324: VecRestoreArray(x,&vdata);
325: BVRestoreColumn(V,i,&x);
326: }
327: return(0);
328: }
330: static PetscErrorCode VecScatterVecs(EPS eps,BV Vin,PetscInt n)
331: {
332: PetscErrorCode ierr;
333: EPS_CISS *ctx = (EPS_CISS*)eps->data;
334: PetscInt i;
335: Vec vi,pvi;
336: const PetscScalar *array;
339: for (i=0;i<n;i++) {
340: BVGetColumn(Vin,i,&vi);
341: VecScatterBegin(ctx->scatterin,vi,ctx->xdup,INSERT_VALUES,SCATTER_FORWARD);
342: VecScatterEnd(ctx->scatterin,vi,ctx->xdup,INSERT_VALUES,SCATTER_FORWARD);
343: BVRestoreColumn(Vin,i,&vi);
344: VecGetArrayRead(ctx->xdup,&array);
345: VecPlaceArray(ctx->xsub,array);
346: BVGetColumn(ctx->pV,i,&pvi);
347: VecCopy(ctx->xsub,pvi);
348: BVRestoreColumn(ctx->pV,i,&pvi);
349: VecResetArray(ctx->xsub);
350: VecRestoreArrayRead(ctx->xdup,&array);
351: }
352: return(0);
353: }
355: static PetscErrorCode SolveLinearSystem(EPS eps,Mat A,Mat B,BV V,PetscInt L_start,PetscInt L_end,PetscBool initksp)
356: {
358: EPS_CISS *ctx = (EPS_CISS*)eps->data;
359: PetscInt i,j,p_id;
360: Mat Fz,kspMat;
361: Vec Bvj,vj,yj;
362: KSP ksp;
365: if (!ctx->ksp) { EPSCISSGetKSPs(eps,&ctx->num_solve_point,&ctx->ksp); }
366: BVCreateVec(V,&Bvj);
367: if (ctx->usest) {
368: MatDuplicate(A,MAT_DO_NOT_COPY_VALUES,&Fz);
369: }
370: for (i=0;i<ctx->num_solve_point;i++) {
371: p_id = i*ctx->subcomm->n + ctx->subcomm_id;
372: if (!ctx->usest && initksp) {
373: MatDuplicate(A,MAT_COPY_VALUES,&kspMat);
374: if (B) {
375: MatAXPY(kspMat,-ctx->omega[p_id],B,DIFFERENT_NONZERO_PATTERN);
376: } else {
377: MatShift(kspMat,-ctx->omega[p_id]);
378: }
379: KSPSetOperators(ctx->ksp[i],kspMat,kspMat);
380: MatDestroy(&kspMat);
381: } else if (ctx->usest) {
382: STSetShift(eps->st,ctx->omega[p_id]);
383: STGetKSP(eps->st,&ksp);
384: }
385: for (j=L_start;j<L_end;j++) {
386: BVGetColumn(V,j,&vj);
387: BVGetColumn(ctx->Y,i*ctx->L_max+j,&yj);
388: if (B) {
389: MatMult(B,vj,Bvj);
390: if (ctx->usest) {
391: KSPSolve(ksp,Bvj,yj);
392: } else {
393: KSPSolve(ctx->ksp[i],Bvj,yj);
394: }
395: } else {
396: if (ctx->usest) {
397: KSPSolve(ksp,vj,yj);
398: } else {
399: KSPSolve(ctx->ksp[i],vj,yj);
400: }
401: }
402: BVRestoreColumn(V,j,&vj);
403: BVRestoreColumn(ctx->Y,i*ctx->L_max+j,&yj);
404: }
405: if (ctx->usest && i<ctx->num_solve_point-1) { KSPReset(ksp); }
406: }
407: if (ctx->usest) { MatDestroy(&Fz); }
408: VecDestroy(&Bvj);
409: return(0);
410: }
412: #if defined(PETSC_USE_COMPLEX)
413: static PetscErrorCode EstimateNumberEigs(EPS eps,PetscInt *L_add)
414: {
416: EPS_CISS *ctx = (EPS_CISS*)eps->data;
417: PetscInt i,j,p_id;
418: PetscScalar tmp,m = 1,sum = 0.0;
419: PetscReal eta;
420: Vec v,vtemp,vj,yj;
423: BVGetColumn(ctx->Y,0,&yj);
424: VecDuplicate(yj,&v);
425: BVRestoreColumn(ctx->Y,0,&yj);
426: BVCreateVec(ctx->V,&vtemp);
427: for (j=0;j<ctx->L;j++) {
428: VecSet(v,0);
429: for (i=0;i<ctx->num_solve_point; i++) {
430: p_id = i*ctx->subcomm->n + ctx->subcomm_id;
431: BVSetActiveColumns(ctx->Y,i*ctx->L_max+j,i*ctx->L_max+j+1);
432: BVMultVec(ctx->Y,ctx->weight[p_id],1,v,&m);
433: }
434: BVGetColumn(ctx->V,j,&vj);
435: if (ctx->pA) {
436: VecSet(vtemp,0);
437: VecScatterBegin(ctx->scatterin,v,vtemp,ADD_VALUES,SCATTER_REVERSE);
438: VecScatterEnd(ctx->scatterin,v,vtemp,ADD_VALUES,SCATTER_REVERSE);
439: VecDot(vj,vtemp,&tmp);
440: } else {
441: VecDot(vj,v,&tmp);
442: }
443: BVRestoreColumn(ctx->V,j,&vj);
444: if (ctx->useconj) sum += PetscRealPart(tmp)*2;
445: else sum += tmp;
446: }
447: ctx->est_eig = PetscAbsScalar(sum/(PetscReal)ctx->L);
448: eta = PetscPowReal(10.0,-PetscLog10Real(eps->tol)/ctx->N);
449: PetscInfo1(eps,"Estimation_#Eig %f\n",(double)ctx->est_eig);
450: *L_add = (PetscInt)PetscCeilReal((ctx->est_eig*eta)/ctx->M) - ctx->L;
451: if (*L_add < 0) *L_add = 0;
452: if (*L_add>ctx->L_max-ctx->L) {
453: PetscInfo(eps,"Number of eigenvalues around the contour path may be too large\n");
454: *L_add = ctx->L_max-ctx->L;
455: }
456: VecDestroy(&v);
457: VecDestroy(&vtemp);
458: return(0);
459: }
460: #endif
462: static PetscErrorCode CalcMu(EPS eps,PetscScalar *Mu)
463: {
465: PetscMPIInt sub_size,len;
466: PetscInt i,j,k,s;
467: PetscScalar *m,*temp,*temp2,*ppk,alp;
468: EPS_CISS *ctx = (EPS_CISS*)eps->data;
469: Mat M;
472: MPI_Comm_size(PetscSubcommChild(ctx->subcomm),&sub_size);
473: PetscMalloc3(ctx->num_solve_point*ctx->L*(ctx->L+1),&temp,2*ctx->M*ctx->L*ctx->L,&temp2,ctx->num_solve_point,&ppk);
474: MatCreateSeqDense(PETSC_COMM_SELF,ctx->L,ctx->L_max*ctx->num_solve_point,NULL,&M);
475: for (i=0;i<2*ctx->M*ctx->L*ctx->L;i++) temp2[i] = 0;
476: BVSetActiveColumns(ctx->Y,0,ctx->L_max*ctx->num_solve_point);
477: if (ctx->pA) {
478: BVSetActiveColumns(ctx->pV,0,ctx->L);
479: BVDot(ctx->Y,ctx->pV,M);
480: } else {
481: BVSetActiveColumns(ctx->V,0,ctx->L);
482: BVDot(ctx->Y,ctx->V,M);
483: }
484: MatDenseGetArray(M,&m);
485: for (i=0;i<ctx->num_solve_point;i++) {
486: for (j=0;j<ctx->L;j++) {
487: for (k=0;k<ctx->L;k++) {
488: temp[k+j*ctx->L+i*ctx->L*ctx->L]=m[k+j*ctx->L+i*ctx->L*ctx->L_max];
489: }
490: }
491: }
492: MatDenseRestoreArray(M,&m);
493: for (i=0;i<ctx->num_solve_point;i++) ppk[i] = 1;
494: for (k=0;k<2*ctx->M;k++) {
495: for (j=0;j<ctx->L;j++) {
496: for (i=0;i<ctx->num_solve_point;i++) {
497: alp = ppk[i]*ctx->weight[i*ctx->subcomm->n + ctx->subcomm_id];
498: for (s=0;s<ctx->L;s++) {
499: if (ctx->useconj) temp2[s+(j+k*ctx->L)*ctx->L] += PetscRealPart(alp*temp[s+(j+i*ctx->L)*ctx->L])*2;
500: else temp2[s+(j+k*ctx->L)*ctx->L] += alp*temp[s+(j+i*ctx->L)*ctx->L];
501: }
502: }
503: }
504: for (i=0;i<ctx->num_solve_point;i++)
505: ppk[i] *= ctx->pp[i*ctx->subcomm->n + ctx->subcomm_id];
506: }
507: for (i=0;i<2*ctx->M*ctx->L*ctx->L;i++) temp2[i] /= sub_size;
508: PetscMPIIntCast(2*ctx->M*ctx->L*ctx->L,&len);
509: MPI_Allreduce(temp2,Mu,len,MPIU_SCALAR,MPIU_SUM,PetscObjectComm((PetscObject)eps));
510: PetscFree3(temp,temp2,ppk);
511: MatDestroy(&M);
512: return(0);
513: }
515: static PetscErrorCode BlockHankel(EPS eps,PetscScalar *Mu,PetscInt s,PetscScalar *H)
516: {
517: EPS_CISS *ctx = (EPS_CISS*)eps->data;
518: PetscInt i,j,k,L=ctx->L,M=ctx->M;
521: for (k=0;k<L*M;k++)
522: for (j=0;j<M;j++)
523: for (i=0;i<L;i++)
524: H[j*L+i+k*L*M] = Mu[i+k*L+(j+s)*L*L];
525: return(0);
526: }
528: static PetscErrorCode SVD_H0(EPS eps,PetscScalar *S,PetscInt *K)
529: {
531: EPS_CISS *ctx = (EPS_CISS*)eps->data;
532: PetscInt i,ml=ctx->L*ctx->M;
533: PetscBLASInt m,n,lda,ldu,ldvt,lwork,info;
534: PetscScalar *work;
535: #if defined(PETSC_USE_COMPLEX)
536: PetscReal *rwork;
537: #endif
540: PetscMalloc1(5*ml,&work);
541: #if defined(PETSC_USE_COMPLEX)
542: PetscMalloc1(5*ml,&rwork);
543: #endif
544: PetscBLASIntCast(ml,&m);
545: n = m; lda = m; ldu = m; ldvt = m; lwork = 5*m;
546: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
547: #if defined(PETSC_USE_COMPLEX)
548: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("N","N",&m,&n,S,&lda,ctx->sigma,NULL,&ldu,NULL,&ldvt,work,&lwork,rwork,&info));
549: #else
550: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("N","N",&m,&n,S,&lda,ctx->sigma,NULL,&ldu,NULL,&ldvt,work,&lwork,&info));
551: #endif
552: SlepcCheckLapackInfo("gesvd",info);
553: PetscFPTrapPop();
554: (*K) = 0;
555: for (i=0;i<ml;i++) {
556: if (ctx->sigma[i]/PetscMax(ctx->sigma[0],1)>ctx->delta) (*K)++;
557: }
558: PetscFree(work);
559: #if defined(PETSC_USE_COMPLEX)
560: PetscFree(rwork);
561: #endif
562: return(0);
563: }
565: static PetscErrorCode ConstructS(EPS eps)
566: {
568: EPS_CISS *ctx = (EPS_CISS*)eps->data;
569: PetscInt i,j,k,vec_local_size,p_id;
570: Vec v,sj,yj;
571: PetscScalar *ppk, *v_data, m = 1;
574: BVGetSizes(ctx->Y,&vec_local_size,NULL,NULL);
575: PetscMalloc1(ctx->num_solve_point,&ppk);
576: for (i=0;i<ctx->num_solve_point;i++) ppk[i] = 1;
577: BVGetColumn(ctx->Y,0,&yj);
578: VecDuplicate(yj,&v);
579: BVRestoreColumn(ctx->Y,0,&yj);
580: for (k=0;k<ctx->M;k++) {
581: for (j=0;j<ctx->L;j++) {
582: VecSet(v,0);
583: for (i=0;i<ctx->num_solve_point;i++) {
584: p_id = i*ctx->subcomm->n + ctx->subcomm_id;
585: BVSetActiveColumns(ctx->Y,i*ctx->L_max+j,i*ctx->L_max+j+1);
586: BVMultVec(ctx->Y,ppk[i]*ctx->weight[p_id],1.0,v,&m);
587: }
588: if (ctx->useconj) {
589: VecGetArray(v,&v_data);
590: for (i=0;i<vec_local_size;i++) v_data[i] = PetscRealPart(v_data[i])*2;
591: VecRestoreArray(v,&v_data);
592: }
593: BVGetColumn(ctx->S,k*ctx->L+j,&sj);
594: if (ctx->pA) {
595: VecSet(sj,0);
596: VecScatterBegin(ctx->scatterin,v,sj,ADD_VALUES,SCATTER_REVERSE);
597: VecScatterEnd(ctx->scatterin,v,sj,ADD_VALUES,SCATTER_REVERSE);
598: } else {
599: VecCopy(v,sj);
600: }
601: BVRestoreColumn(ctx->S,k*ctx->L+j,&sj);
602: }
603: for (i=0;i<ctx->num_solve_point;i++) {
604: p_id = i*ctx->subcomm->n + ctx->subcomm_id;
605: ppk[i] *= ctx->pp[p_id];
606: }
607: }
608: PetscFree(ppk);
609: VecDestroy(&v);
610: return(0);
611: }
613: static PetscErrorCode SVD_S(BV S,PetscInt ml,PetscReal delta,PetscReal *sigma,PetscInt *K)
614: {
616: PetscInt i,j,k,local_size;
617: PetscMPIInt len;
618: PetscScalar *work,*temp,*B,*tempB,*s_data,*Q1,*Q2,*temp2,alpha=1,beta=0;
619: PetscBLASInt l,m,n,lda,ldu,ldvt,lwork,info,ldb,ldc;
620: #if defined(PETSC_USE_COMPLEX)
621: PetscReal *rwork;
622: #endif
625: BVGetSizes(S,&local_size,NULL,NULL);
626: BVGetArray(S,&s_data);
627: PetscMalloc7(ml*ml,&temp,ml*ml,&temp2,local_size*ml,&Q1,local_size*ml,&Q2,ml*ml,&B,ml*ml,&tempB,5*ml,&work);
628: PetscArrayzero(B,ml*ml);
629: #if defined(PETSC_USE_COMPLEX)
630: PetscMalloc1(5*ml,&rwork);
631: #endif
632: PetscFPTrapPush(PETSC_FP_TRAP_OFF);
634: for (i=0;i<ml;i++) B[i*ml+i]=1;
636: for (k=0;k<2;k++) {
637: PetscBLASIntCast(local_size,&m);
638: PetscBLASIntCast(ml,&l);
639: n = l; lda = m; ldb = m; ldc = l;
640: if (k == 0) {
641: PetscStackCallBLAS("BLASgemm",BLASgemm_("C","N",&l,&n,&m,&alpha,s_data,&lda,s_data,&ldb,&beta,temp,&ldc));
642: } else if ((k%2)==1) {
643: PetscStackCallBLAS("BLASgemm",BLASgemm_("C","N",&l,&n,&m,&alpha,Q1,&lda,Q1,&ldb,&beta,temp,&ldc));
644: } else {
645: PetscStackCallBLAS("BLASgemm",BLASgemm_("C","N",&l,&n,&m,&alpha,Q2,&lda,Q2,&ldb,&beta,temp,&ldc));
646: }
647: PetscArrayzero(temp2,ml*ml);
648: PetscMPIIntCast(ml*ml,&len);
649: MPI_Allreduce(temp,temp2,len,MPIU_SCALAR,MPIU_SUM,PetscObjectComm((PetscObject)S));
651: PetscBLASIntCast(ml,&m);
652: n = m; lda = m; lwork = 5*m, ldu = 1; ldvt = 1;
653: #if defined(PETSC_USE_COMPLEX)
654: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&m,&n,temp2,&lda,sigma,NULL,&ldu,NULL,&ldvt,work,&lwork,rwork,&info));
655: #else
656: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("O","N",&m,&n,temp2,&lda,sigma,NULL,&ldu,NULL,&ldvt,work,&lwork,&info));
657: #endif
658: SlepcCheckLapackInfo("gesvd",info);
660: PetscBLASIntCast(local_size,&l);
661: PetscBLASIntCast(ml,&n);
662: m = n; lda = l; ldb = m; ldc = l;
663: if (k==0) {
664: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&l,&n,&m,&alpha,s_data,&lda,temp2,&ldb,&beta,Q1,&ldc));
665: } else if ((k%2)==1) {
666: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&l,&n,&m,&alpha,Q1,&lda,temp2,&ldb,&beta,Q2,&ldc));
667: } else {
668: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&l,&n,&m,&alpha,Q2,&lda,temp2,&ldb,&beta,Q1,&ldc));
669: }
671: PetscBLASIntCast(ml,&l);
672: m = l; n = l; lda = l; ldb = m; ldc = l;
673: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","N",&l,&n,&m,&alpha,B,&lda,temp2,&ldb,&beta,tempB,&ldc));
674: for (i=0;i<ml;i++) {
675: sigma[i] = sqrt(sigma[i]);
676: for (j=0;j<local_size;j++) {
677: if ((k%2)==1) Q2[j+i*local_size]/=sigma[i];
678: else Q1[j+i*local_size]/=sigma[i];
679: }
680: for (j=0;j<ml;j++) {
681: B[j+i*ml]=tempB[j+i*ml]*sigma[i];
682: }
683: }
684: }
686: PetscBLASIntCast(ml,&m);
687: n = m; lda = m; ldu=1; ldvt=1;
688: #if defined(PETSC_USE_COMPLEX)
689: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("N","O",&m,&n,B,&lda,sigma,NULL,&ldu,NULL,&ldvt,work,&lwork,rwork,&info));
690: #else
691: PetscStackCallBLAS("LAPACKgesvd",LAPACKgesvd_("N","O",&m,&n,B,&lda,sigma,NULL,&ldu,NULL,&ldvt,work,&lwork,&info));
692: #endif
693: SlepcCheckLapackInfo("gesvd",info);
695: PetscBLASIntCast(local_size,&l);
696: PetscBLASIntCast(ml,&n);
697: m = n; lda = l; ldb = m; ldc = l;
698: if ((k%2)==1) {
699: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","T",&l,&n,&m,&alpha,Q1,&lda,B,&ldb,&beta,s_data,&ldc));
700: } else {
701: PetscStackCallBLAS("BLASgemm",BLASgemm_("N","T",&l,&n,&m,&alpha,Q2,&lda,B,&ldb,&beta,s_data,&ldc));
702: }
704: PetscFPTrapPop();
705: BVRestoreArray(S,&s_data);
707: (*K) = 0;
708: for (i=0;i<ml;i++) {
709: if (sigma[i]/PetscMax(sigma[0],1)>delta) (*K)++;
710: }
711: PetscFree7(temp,temp2,Q1,Q2,B,tempB,work);
712: #if defined(PETSC_USE_COMPLEX)
713: PetscFree(rwork);
714: #endif
715: return(0);
716: }
718: static PetscErrorCode isGhost(EPS eps,PetscInt ld,PetscInt nv,PetscBool *fl)
719: {
721: EPS_CISS *ctx = (EPS_CISS*)eps->data;
722: PetscInt i,j;
723: PetscScalar *pX;
724: PetscReal *tau,s1,s2,tau_max=0.0;
727: PetscMalloc1(nv,&tau);
728: DSVectors(eps->ds,DS_MAT_X,NULL,NULL);
729: DSGetArray(eps->ds,DS_MAT_X,&pX);
731: for (i=0;i<nv;i++) {
732: s1 = 0;
733: s2 = 0;
734: for (j=0;j<nv;j++) {
735: s1 += PetscAbsScalar(PetscPowScalarInt(pX[i*ld+j],2));
736: s2 += PetscPowRealInt(PetscAbsScalar(pX[i*ld+j]),2)/ctx->sigma[j];
737: }
738: tau[i] = s1/s2;
739: tau_max = PetscMax(tau_max,tau[i]);
740: }
741: DSRestoreArray(eps->ds,DS_MAT_X,&pX);
742: for (i=0;i<nv;i++) {
743: tau[i] /= tau_max;
744: }
745: for (i=0;i<nv;i++) {
746: if (tau[i]>=ctx->spurious_threshold) fl[i] = PETSC_TRUE;
747: else fl[i] = PETSC_FALSE;
748: }
749: PetscFree(tau);
750: return(0);
751: }
753: static PetscErrorCode rescale_eig(EPS eps,PetscInt nv)
754: {
756: EPS_CISS *ctx = (EPS_CISS*)eps->data;
757: PetscInt i;
758: PetscScalar center;
759: PetscReal radius,a,b,c,d,rgscale;
760: #if defined(PETSC_USE_COMPLEX)
761: PetscReal start_ang,end_ang,vscale,theta;
762: #endif
763: PetscBool isring,isellipse,isinterval;
766: PetscObjectTypeCompare((PetscObject)eps->rg,RGELLIPSE,&isellipse);
767: PetscObjectTypeCompare((PetscObject)eps->rg,RGRING,&isring);
768: PetscObjectTypeCompare((PetscObject)eps->rg,RGINTERVAL,&isinterval);
769: RGGetScale(eps->rg,&rgscale);
770: if (isinterval) {
771: RGIntervalGetEndpoints(eps->rg,NULL,NULL,&c,&d);
772: if (c==d) {
773: for (i=0;i<nv;i++) {
774: #if defined(PETSC_USE_COMPLEX)
775: eps->eigr[i] = PetscRealPart(eps->eigr[i]);
776: #else
777: eps->eigi[i] = 0;
778: #endif
779: }
780: }
781: }
782: if (ctx->extraction == EPS_CISS_EXTRACTION_HANKEL) {
783: if (isellipse) {
784: RGEllipseGetParameters(eps->rg,¢er,&radius,NULL);
785: for (i=0;i<nv;i++) eps->eigr[i] = rgscale*(center + radius*eps->eigr[i]);
786: } else if (isinterval) {
787: RGIntervalGetEndpoints(eps->rg,&a,&b,&c,&d);
788: if (ctx->quad == EPS_CISS_QUADRULE_CHEBYSHEV) {
789: for (i=0;i<nv;i++) {
790: if (c==d) eps->eigr[i] = ((b-a)*(eps->eigr[i]+1.0)/2.0+a)*rgscale;
791: if (a==b) {
792: #if defined(PETSC_USE_COMPLEX)
793: eps->eigr[i] = ((d-c)*(eps->eigr[i]+1.0)/2.0+c)*rgscale*PETSC_i;
794: #else
795: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Integration points on a vertical line require complex arithmetic");
796: #endif
797: }
798: }
799: } else {
800: center = (b+a)/2.0+(d+c)/2.0*PETSC_PI;
801: radius = PetscSqrtReal(PetscPowRealInt((b-a)/2.0,2)+PetscPowRealInt((d-c)/2.0,2));
802: for (i=0;i<nv;i++) eps->eigr[i] = center + radius*eps->eigr[i];
803: }
804: } else if (isring) { /* only supported in complex scalars */
805: #if defined(PETSC_USE_COMPLEX)
806: RGRingGetParameters(eps->rg,¢er,&radius,&vscale,&start_ang,&end_ang,NULL);
807: if (ctx->quad == EPS_CISS_QUADRULE_CHEBYSHEV) {
808: for (i=0;i<nv;i++) {
809: theta = (start_ang*2.0+(end_ang-start_ang)*(PetscRealPart(eps->eigr[i])+1.0))*PETSC_PI;
810: eps->eigr[i] = rgscale*center + (rgscale*radius+PetscImaginaryPart(eps->eigr[i]))*PetscCMPLX(PetscCosReal(theta),vscale*PetscSinReal(theta));
811: }
812: } else {
813: for (i=0;i<nv;i++) eps->eigr[i] = rgscale*(center + radius*eps->eigr[i]);
814: }
815: #endif
816: }
817: }
818: return(0);
819: }
821: PetscErrorCode EPSSetUp_CISS(EPS eps)
822: {
824: EPS_CISS *ctx = (EPS_CISS*)eps->data;
825: PetscBool issinvert,istrivial,isring,isellipse,isinterval,flg,useconj;
826: PetscReal c,d;
827: Mat A;
830: if (!eps->ncv) {
831: eps->ncv = ctx->L_max*ctx->M;
832: if (eps->ncv>eps->n) {
833: eps->ncv = eps->n;
834: ctx->L_max = eps->ncv/ctx->M;
835: if (!ctx->L_max) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"Cannot adjust solver parameters, try setting a smaller value of M (moment size)");
836: }
837: } else {
838: EPSSetDimensions_Default(eps,eps->nev,&eps->ncv,&eps->mpd);
839: ctx->L_max = eps->ncv/ctx->M;
840: if (!ctx->L_max) {
841: ctx->L_max = 1;
842: eps->ncv = ctx->L_max*ctx->M;
843: }
844: }
845: ctx->L = PetscMin(ctx->L,ctx->L_max);
846: if (!eps->max_it) eps->max_it = 1;
847: if (!eps->mpd) eps->mpd = eps->ncv;
848: if (!eps->which) eps->which = EPS_ALL;
849: if (!eps->extraction) { EPSSetExtraction(eps,EPS_RITZ); }
850: else if (eps->extraction!=EPS_RITZ) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"Unsupported extraction type");
851: if (eps->arbitrary) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"Arbitrary selection of eigenpairs not supported in this solver");
852: if (eps->stopping!=EPSStoppingBasic) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"This solver does not support user-defined stopping test");
854: /* check region */
855: RGIsTrivial(eps->rg,&istrivial);
856: if (istrivial) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"CISS requires a nontrivial region, e.g. -rg_type ellipse ...");
857: RGGetComplement(eps->rg,&flg);
858: if (flg) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"A region with complement flag set is not allowed");
859: PetscObjectTypeCompare((PetscObject)eps->rg,RGELLIPSE,&isellipse);
860: PetscObjectTypeCompare((PetscObject)eps->rg,RGRING,&isring);
861: PetscObjectTypeCompare((PetscObject)eps->rg,RGINTERVAL,&isinterval);
862: if (!isellipse && !isring && !isinterval) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"Currently only implemented for interval, elliptic or ring regions");
863: /* if useconj has changed, then reset subcomm data */
864: EPSCISSSetUseConj(eps,&useconj);
865: if (useconj!=ctx->useconj) { EPSCISSResetSubcomm(eps); }
867: #if !defined(PETSC_USE_COMPLEX)
868: if (isring) {
869: SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"Ring region only supported for complex scalars");
870: }
871: #endif
872: if (isinterval) {
873: RGIntervalGetEndpoints(eps->rg,NULL,NULL,&c,&d);
874: #if !defined(PETSC_USE_COMPLEX)
875: if (c!=d || c!=0.0) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"In real scalars, endpoints of the imaginary axis must be both zero");
876: #endif
877: if (!ctx->quad && c==d) ctx->quad = EPS_CISS_QUADRULE_CHEBYSHEV;
878: }
879: if (!ctx->quad) ctx->quad = EPS_CISS_QUADRULE_TRAPEZOIDAL;
881: /* create split comm */
882: if (!ctx->subcomm) { EPSCISSSetUpSubComm(eps,&ctx->num_solve_point); }
884: EPSAllocateSolution(eps,0);
885: if (ctx->weight) { PetscFree4(ctx->weight,ctx->omega,ctx->pp,ctx->sigma); }
886: PetscMalloc4(ctx->N,&ctx->weight,ctx->N+1,&ctx->omega,ctx->N,&ctx->pp,ctx->L_max*ctx->M,&ctx->sigma);
887: PetscLogObjectMemory((PetscObject)eps,3*ctx->N*sizeof(PetscScalar)+ctx->L_max*ctx->N*sizeof(PetscReal));
889: /* allocate basis vectors */
890: BVDestroy(&ctx->S);
891: BVDuplicateResize(eps->V,ctx->L_max*ctx->M,&ctx->S);
892: PetscLogObjectParent((PetscObject)eps,(PetscObject)ctx->S);
893: BVDestroy(&ctx->V);
894: BVDuplicateResize(eps->V,ctx->L_max,&ctx->V);
895: PetscLogObjectParent((PetscObject)eps,(PetscObject)ctx->V);
897: STGetMatrix(eps->st,0,&A);
898: PetscObjectTypeCompare((PetscObject)A,MATSHELL,&flg);
899: if (flg) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"Matrix type shell is not supported in this solver");
901: if (!ctx->usest_set) ctx->usest = (ctx->npart>1)? PETSC_FALSE: PETSC_TRUE;
902: if (ctx->usest && ctx->npart>1) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"The usest flag is not supported when partitions > 1");
904: CISSRedundantMat(eps);
905: if (ctx->pA) {
906: CISSScatterVec(eps);
907: BVDestroy(&ctx->pV);
908: BVCreate(PetscObjectComm((PetscObject)ctx->xsub),&ctx->pV);
909: BVSetSizesFromVec(ctx->pV,ctx->xsub,eps->n);
910: BVSetFromOptions(ctx->pV);
911: BVResize(ctx->pV,ctx->L_max,PETSC_FALSE);
912: PetscLogObjectParent((PetscObject)eps,(PetscObject)ctx->pV);
913: }
915: if (ctx->usest) {
916: PetscObjectTypeCompare((PetscObject)eps->st,STSINVERT,&issinvert);
917: if (!issinvert) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_SUP,"If the usest flag is set, you must select the STSINVERT spectral transformation");
918: }
920: BVDestroy(&ctx->Y);
921: if (ctx->pA) {
922: BVCreate(PetscObjectComm((PetscObject)ctx->xsub),&ctx->Y);
923: BVSetSizesFromVec(ctx->Y,ctx->xsub,eps->n);
924: BVSetFromOptions(ctx->Y);
925: BVResize(ctx->Y,ctx->num_solve_point*ctx->L_max,PETSC_FALSE);
926: } else {
927: BVDuplicateResize(eps->V,ctx->num_solve_point*ctx->L_max,&ctx->Y);
928: }
929: PetscLogObjectParent((PetscObject)eps,(PetscObject)ctx->Y);
931: if (ctx->extraction == EPS_CISS_EXTRACTION_HANKEL) {
932: DSSetType(eps->ds,DSGNHEP);
933: } else if (eps->isgeneralized) {
934: if (eps->ishermitian && eps->ispositive) {
935: DSSetType(eps->ds,DSGHEP);
936: } else {
937: DSSetType(eps->ds,DSGNHEP);
938: }
939: } else {
940: if (eps->ishermitian) {
941: DSSetType(eps->ds,DSHEP);
942: } else {
943: DSSetType(eps->ds,DSNHEP);
944: }
945: }
946: DSAllocate(eps->ds,eps->ncv);
947: EPSSetWorkVecs(eps,2);
949: #if !defined(PETSC_USE_COMPLEX)
950: if (!eps->ishermitian) { PetscInfo(eps,"Warning: complex eigenvalues are not calculated exactly without --with-scalar-type=complex in PETSc\n"); }
951: #endif
952: return(0);
953: }
955: PetscErrorCode EPSSolve_CISS(EPS eps)
956: {
958: EPS_CISS *ctx = (EPS_CISS*)eps->data;
959: Mat A,B,X,M,pA,pB;
960: PetscInt i,j,ld,nmat,L_add=0,nv=0,L_base=ctx->L,inner,nlocal,*inside;
961: PetscScalar *Mu,*H0,*H1=NULL,*rr,*temp;
962: PetscReal error,max_error;
963: PetscBool *fl1;
964: Vec si,w[3];
965: SlepcSC sc;
966: PetscRandom rand;
967: #if defined(PETSC_USE_COMPLEX)
968: PetscBool isellipse;
969: #endif
972: w[0] = eps->work[0];
973: w[1] = NULL;
974: w[2] = eps->work[1];
975: /* override SC settings */
976: DSGetSlepcSC(eps->ds,&sc);
977: sc->comparison = SlepcCompareLargestMagnitude;
978: sc->comparisonctx = NULL;
979: sc->map = NULL;
980: sc->mapobj = NULL;
981: VecGetLocalSize(w[0],&nlocal);
982: DSGetLeadingDimension(eps->ds,&ld);
983: STGetNumMatrices(eps->st,&nmat);
984: STGetMatrix(eps->st,0,&A);
985: if (nmat>1) { STGetMatrix(eps->st,1,&B); }
986: else B = NULL;
987: SetPathParameter(eps);
988: CISSVecSetRandom(ctx->V,0,ctx->L);
989: BVGetRandomContext(ctx->V,&rand);
991: if (ctx->pA) {
992: VecScatterVecs(eps,ctx->V,ctx->L);
993: SolveLinearSystem(eps,ctx->pA,ctx->pB,ctx->pV,0,ctx->L,PETSC_TRUE);
994: } else {
995: SolveLinearSystem(eps,A,B,ctx->V,0,ctx->L,PETSC_TRUE);
996: }
997: #if defined(PETSC_USE_COMPLEX)
998: PetscObjectTypeCompare((PetscObject)eps->rg,RGELLIPSE,&isellipse);
999: if (isellipse) {
1000: EstimateNumberEigs(eps,&L_add);
1001: } else {
1002: L_add = 0;
1003: }
1004: #else
1005: L_add = 0;
1006: #endif
1007: if (L_add>0) {
1008: PetscInfo2(eps,"Changing L %D -> %D by Estimate #Eig\n",ctx->L,ctx->L+L_add);
1009: CISSVecSetRandom(ctx->V,ctx->L,ctx->L+L_add);
1010: if (ctx->pA) {
1011: VecScatterVecs(eps,ctx->V,ctx->L+L_add);
1012: SolveLinearSystem(eps,ctx->pA,ctx->pB,ctx->pV,ctx->L,ctx->L+L_add,PETSC_FALSE);
1013: } else {
1014: SolveLinearSystem(eps,A,B,ctx->V,ctx->L,ctx->L+L_add,PETSC_FALSE);
1015: }
1016: ctx->L += L_add;
1017: }
1018: PetscMalloc2(ctx->L*ctx->L*ctx->M*2,&Mu,ctx->L*ctx->M*ctx->L*ctx->M,&H0);
1019: for (i=0;i<ctx->refine_blocksize;i++) {
1020: CalcMu(eps,Mu);
1021: BlockHankel(eps,Mu,0,H0);
1022: SVD_H0(eps,H0,&nv);
1023: if (ctx->sigma[0]<=ctx->delta || nv < ctx->L*ctx->M || ctx->L == ctx->L_max) break;
1024: L_add = L_base;
1025: if (ctx->L+L_add>ctx->L_max) L_add = ctx->L_max-ctx->L;
1026: PetscInfo2(eps,"Changing L %D -> %D by SVD(H0)\n",ctx->L,ctx->L+L_add);
1027: CISSVecSetRandom(ctx->V,ctx->L,ctx->L+L_add);
1028: if (ctx->pA) {
1029: VecScatterVecs(eps,ctx->V,ctx->L+L_add);
1030: SolveLinearSystem(eps,ctx->pA,ctx->pB,ctx->pV,ctx->L,ctx->L+L_add,PETSC_FALSE);
1031: } else {
1032: SolveLinearSystem(eps,A,B,ctx->V,ctx->L,ctx->L+L_add,PETSC_FALSE);
1033: }
1034: ctx->L += L_add;
1035: }
1036: if (ctx->extraction == EPS_CISS_EXTRACTION_HANKEL) {
1037: PetscMalloc1(ctx->L*ctx->M*ctx->L*ctx->M,&H1);
1038: }
1040: while (eps->reason == EPS_CONVERGED_ITERATING) {
1041: eps->its++;
1042: for (inner=0;inner<=ctx->refine_inner;inner++) {
1043: if (ctx->extraction == EPS_CISS_EXTRACTION_HANKEL) {
1044: CalcMu(eps,Mu);
1045: BlockHankel(eps,Mu,0,H0);
1046: SVD_H0(eps,H0,&nv);
1047: break;
1048: } else {
1049: ConstructS(eps);
1050: BVSetActiveColumns(ctx->S,0,ctx->L);
1051: BVCopy(ctx->S,ctx->V);
1052: SVD_S(ctx->S,ctx->L*ctx->M,ctx->delta,ctx->sigma,&nv);
1053: if (ctx->sigma[0]>ctx->delta && nv==ctx->L*ctx->M && inner!=ctx->refine_inner) {
1054: if (ctx->pA) {
1055: VecScatterVecs(eps,ctx->V,ctx->L);
1056: SolveLinearSystem(eps,ctx->pA,ctx->pB,ctx->pV,0,ctx->L,PETSC_FALSE);
1057: } else {
1058: SolveLinearSystem(eps,A,B,ctx->V,0,ctx->L,PETSC_FALSE);
1059: }
1060: } else break;
1061: }
1062: }
1063: eps->nconv = 0;
1064: if (nv == 0) eps->reason = EPS_CONVERGED_TOL;
1065: else {
1066: DSSetDimensions(eps->ds,nv,0,0,0);
1067: DSSetState(eps->ds,DS_STATE_RAW);
1069: if (ctx->extraction == EPS_CISS_EXTRACTION_HANKEL) {
1070: BlockHankel(eps,Mu,0,H0);
1071: BlockHankel(eps,Mu,1,H1);
1072: DSGetArray(eps->ds,DS_MAT_A,&temp);
1073: for (j=0;j<nv;j++) {
1074: for (i=0;i<nv;i++) {
1075: temp[i+j*ld] = H1[i+j*ctx->L*ctx->M];
1076: }
1077: }
1078: DSRestoreArray(eps->ds,DS_MAT_A,&temp);
1079: DSGetArray(eps->ds,DS_MAT_B,&temp);
1080: for (j=0;j<nv;j++) {
1081: for (i=0;i<nv;i++) {
1082: temp[i+j*ld] = H0[i+j*ctx->L*ctx->M];
1083: }
1084: }
1085: DSRestoreArray(eps->ds,DS_MAT_B,&temp);
1086: } else {
1087: BVSetActiveColumns(ctx->S,0,nv);
1088: DSGetMat(eps->ds,DS_MAT_A,&pA);
1089: MatZeroEntries(pA);
1090: BVMatProject(ctx->S,A,ctx->S,pA);
1091: DSRestoreMat(eps->ds,DS_MAT_A,&pA);
1092: if (B) {
1093: DSGetMat(eps->ds,DS_MAT_B,&pB);
1094: MatZeroEntries(pB);
1095: BVMatProject(ctx->S,B,ctx->S,pB);
1096: DSRestoreMat(eps->ds,DS_MAT_B,&pB);
1097: }
1098: }
1100: DSSolve(eps->ds,eps->eigr,eps->eigi);
1101: DSSynchronize(eps->ds,eps->eigr,eps->eigi);
1103: PetscMalloc3(nv,&fl1,nv,&inside,nv,&rr);
1104: rescale_eig(eps,nv);
1105: isGhost(eps,ld,nv,fl1);
1106: RGCheckInside(eps->rg,nv,eps->eigr,eps->eigi,inside);
1107: for (i=0;i<nv;i++) {
1108: if (fl1[i] && inside[i]>=0) {
1109: rr[i] = 1.0;
1110: eps->nconv++;
1111: } else rr[i] = 0.0;
1112: }
1113: DSSort(eps->ds,eps->eigr,eps->eigi,rr,NULL,&eps->nconv);
1114: DSSynchronize(eps->ds,eps->eigr,eps->eigi);
1115: rescale_eig(eps,nv);
1116: PetscFree3(fl1,inside,rr);
1117: BVSetActiveColumns(eps->V,0,nv);
1118: if (ctx->extraction == EPS_CISS_EXTRACTION_HANKEL) {
1119: ConstructS(eps);
1120: BVSetActiveColumns(ctx->S,0,ctx->L);
1121: BVCopy(ctx->S,ctx->V);
1122: BVSetActiveColumns(ctx->S,0,nv);
1123: }
1124: BVCopy(ctx->S,eps->V);
1126: DSVectors(eps->ds,DS_MAT_X,NULL,NULL);
1127: DSGetMat(eps->ds,DS_MAT_X,&X);
1128: BVMultInPlace(ctx->S,X,0,eps->nconv);
1129: if (eps->ishermitian) {
1130: BVMultInPlace(eps->V,X,0,eps->nconv);
1131: }
1132: MatDestroy(&X);
1133: max_error = 0.0;
1134: for (i=0;i<eps->nconv;i++) {
1135: BVGetColumn(ctx->S,i,&si);
1136: EPSComputeResidualNorm_Private(eps,PETSC_FALSE,eps->eigr[i],eps->eigi[i],si,NULL,w,&error);
1137: (*eps->converged)(eps,eps->eigr[i],eps->eigi[i],error,&error,eps->convergedctx);
1138: BVRestoreColumn(ctx->S,i,&si);
1139: max_error = PetscMax(max_error,error);
1140: }
1142: if (max_error <= eps->tol) eps->reason = EPS_CONVERGED_TOL;
1143: else if (eps->its >= eps->max_it) eps->reason = EPS_DIVERGED_ITS;
1144: else {
1145: if (eps->nconv > ctx->L) {
1146: MatCreateSeqDense(PETSC_COMM_SELF,eps->nconv,ctx->L,NULL,&M);
1147: MatDenseGetArray(M,&temp);
1148: for (i=0;i<ctx->L*eps->nconv;i++) {
1149: PetscRandomGetValue(rand,&temp[i]);
1150: temp[i] = PetscRealPart(temp[i]);
1151: }
1152: MatDenseRestoreArray(M,&temp);
1153: BVSetActiveColumns(ctx->S,0,eps->nconv);
1154: BVMultInPlace(ctx->S,M,0,ctx->L);
1155: MatDestroy(&M);
1156: BVSetActiveColumns(ctx->S,0,ctx->L);
1157: BVCopy(ctx->S,ctx->V);
1158: }
1159: if (ctx->pA) {
1160: VecScatterVecs(eps,ctx->V,ctx->L);
1161: SolveLinearSystem(eps,ctx->pA,ctx->pB,ctx->pV,0,ctx->L,PETSC_FALSE);
1162: } else {
1163: SolveLinearSystem(eps,A,B,ctx->V,0,ctx->L,PETSC_FALSE);
1164: }
1165: }
1166: }
1167: }
1168: if (ctx->extraction == EPS_CISS_EXTRACTION_HANKEL) {
1169: PetscFree(H1);
1170: }
1171: PetscFree2(Mu,H0);
1172: return(0);
1173: }
1175: static PetscErrorCode EPSCISSSetSizes_CISS(EPS eps,PetscInt ip,PetscInt bs,PetscInt ms,PetscInt npart,PetscInt bsmax,PetscBool realmats)
1176: {
1178: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1179: PetscInt oN,onpart;
1182: oN = ctx->N;
1183: if (ip == PETSC_DECIDE || ip == PETSC_DEFAULT) {
1184: if (ctx->N!=32) { ctx->N =32; ctx->M = ctx->N/4; }
1185: } else {
1186: if (ip<1) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The ip argument must be > 0");
1187: if (ip%2) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The ip argument must be an even number");
1188: if (ctx->N!=ip) { ctx->N = ip; ctx->M = ctx->N/4; }
1189: }
1190: if (bs == PETSC_DECIDE || bs == PETSC_DEFAULT) {
1191: ctx->L = 16;
1192: } else {
1193: if (bs<1) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The bs argument must be > 0");
1194: ctx->L = bs;
1195: }
1196: if (ms == PETSC_DECIDE || ms == PETSC_DEFAULT) {
1197: ctx->M = ctx->N/4;
1198: } else {
1199: if (ms<1) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The ms argument must be > 0");
1200: if (ms>ctx->N) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The ms argument must be less than or equal to the number of integration points");
1201: ctx->M = ms;
1202: }
1203: onpart = ctx->npart;
1204: if (npart == PETSC_DECIDE || npart == PETSC_DEFAULT) {
1205: ctx->npart = 1;
1206: } else {
1207: if (npart<1) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The npart argument must be > 0");
1208: ctx->npart = npart;
1209: }
1210: if (bsmax == PETSC_DECIDE || bsmax == PETSC_DEFAULT) {
1211: ctx->L_max = 64;
1212: } else {
1213: if (bsmax<1) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The bsmax argument must be > 0");
1214: ctx->L_max = PetscMax(bsmax,ctx->L);
1215: }
1216: if (onpart != ctx->npart || oN != ctx->N || realmats != ctx->isreal) { EPSCISSResetSubcomm(eps); }
1217: ctx->isreal = realmats;
1218: eps->state = EPS_STATE_INITIAL;
1219: return(0);
1220: }
1222: /*@
1223: EPSCISSSetSizes - Sets the values of various size parameters in the CISS solver.
1225: Logically Collective on eps
1227: Input Parameters:
1228: + eps - the eigenproblem solver context
1229: . ip - number of integration points
1230: . bs - block size
1231: . ms - moment size
1232: . npart - number of partitions when splitting the communicator
1233: . bsmax - max block size
1234: - realmats - A and B are real
1236: Options Database Keys:
1237: + -eps_ciss_integration_points - Sets the number of integration points
1238: . -eps_ciss_blocksize - Sets the block size
1239: . -eps_ciss_moments - Sets the moment size
1240: . -eps_ciss_partitions - Sets the number of partitions
1241: . -eps_ciss_maxblocksize - Sets the maximum block size
1242: - -eps_ciss_realmats - A and B are real
1244: Note:
1245: The default number of partitions is 1. This means the internal KSP object is shared
1246: among all processes of the EPS communicator. Otherwise, the communicator is split
1247: into npart communicators, so that npart KSP solves proceed simultaneously.
1249: Level: advanced
1251: .seealso: EPSCISSGetSizes()
1252: @*/
1253: PetscErrorCode EPSCISSSetSizes(EPS eps,PetscInt ip,PetscInt bs,PetscInt ms,PetscInt npart,PetscInt bsmax,PetscBool realmats)
1254: {
1265: PetscTryMethod(eps,"EPSCISSSetSizes_C",(EPS,PetscInt,PetscInt,PetscInt,PetscInt,PetscInt,PetscBool),(eps,ip,bs,ms,npart,bsmax,realmats));
1266: return(0);
1267: }
1269: static PetscErrorCode EPSCISSGetSizes_CISS(EPS eps,PetscInt *ip,PetscInt *bs,PetscInt *ms,PetscInt *npart,PetscInt *bsmax,PetscBool *realmats)
1270: {
1271: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1274: if (ip) *ip = ctx->N;
1275: if (bs) *bs = ctx->L;
1276: if (ms) *ms = ctx->M;
1277: if (npart) *npart = ctx->npart;
1278: if (bsmax) *bsmax = ctx->L_max;
1279: if (realmats) *realmats = ctx->isreal;
1280: return(0);
1281: }
1283: /*@
1284: EPSCISSGetSizes - Gets the values of various size parameters in the CISS solver.
1286: Not Collective
1288: Input Parameter:
1289: . eps - the eigenproblem solver context
1291: Output Parameters:
1292: + ip - number of integration points
1293: . bs - block size
1294: . ms - moment size
1295: . npart - number of partitions when splitting the communicator
1296: . bsmax - max block size
1297: - realmats - A and B are real
1299: Level: advanced
1301: .seealso: EPSCISSSetSizes()
1302: @*/
1303: PetscErrorCode EPSCISSGetSizes(EPS eps,PetscInt *ip,PetscInt *bs,PetscInt *ms,PetscInt *npart,PetscInt *bsmax,PetscBool *realmats)
1304: {
1309: PetscUseMethod(eps,"EPSCISSGetSizes_C",(EPS,PetscInt*,PetscInt*,PetscInt*,PetscInt*,PetscInt*,PetscBool*),(eps,ip,bs,ms,npart,bsmax,realmats));
1310: return(0);
1311: }
1313: static PetscErrorCode EPSCISSSetThreshold_CISS(EPS eps,PetscReal delta,PetscReal spur)
1314: {
1315: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1318: if (delta == PETSC_DEFAULT) {
1319: ctx->delta = 1e-12;
1320: } else {
1321: if (delta<=0.0) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The delta argument must be > 0.0");
1322: ctx->delta = delta;
1323: }
1324: if (spur == PETSC_DEFAULT) {
1325: ctx->spurious_threshold = 1e-4;
1326: } else {
1327: if (spur<=0.0) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The spurious threshold argument must be > 0.0");
1328: ctx->spurious_threshold = spur;
1329: }
1330: return(0);
1331: }
1333: /*@
1334: EPSCISSSetThreshold - Sets the values of various threshold parameters in
1335: the CISS solver.
1337: Logically Collective on eps
1339: Input Parameters:
1340: + eps - the eigenproblem solver context
1341: . delta - threshold for numerical rank
1342: - spur - spurious threshold (to discard spurious eigenpairs)
1344: Options Database Keys:
1345: + -eps_ciss_delta - Sets the delta
1346: - -eps_ciss_spurious_threshold - Sets the spurious threshold
1348: Level: advanced
1350: .seealso: EPSCISSGetThreshold()
1351: @*/
1352: PetscErrorCode EPSCISSSetThreshold(EPS eps,PetscReal delta,PetscReal spur)
1353: {
1360: PetscTryMethod(eps,"EPSCISSSetThreshold_C",(EPS,PetscReal,PetscReal),(eps,delta,spur));
1361: return(0);
1362: }
1364: static PetscErrorCode EPSCISSGetThreshold_CISS(EPS eps,PetscReal *delta,PetscReal *spur)
1365: {
1366: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1369: if (delta) *delta = ctx->delta;
1370: if (spur) *spur = ctx->spurious_threshold;
1371: return(0);
1372: }
1374: /*@
1375: EPSCISSGetThreshold - Gets the values of various threshold parameters
1376: in the CISS solver.
1378: Not Collective
1380: Input Parameter:
1381: . eps - the eigenproblem solver context
1383: Output Parameters:
1384: + delta - threshold for numerical rank
1385: - spur - spurious threshold (to discard spurious eigenpairs)
1387: Level: advanced
1389: .seealso: EPSCISSSetThreshold()
1390: @*/
1391: PetscErrorCode EPSCISSGetThreshold(EPS eps,PetscReal *delta,PetscReal *spur)
1392: {
1397: PetscUseMethod(eps,"EPSCISSGetThreshold_C",(EPS,PetscReal*,PetscReal*),(eps,delta,spur));
1398: return(0);
1399: }
1401: static PetscErrorCode EPSCISSSetRefinement_CISS(EPS eps,PetscInt inner,PetscInt blsize)
1402: {
1403: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1406: if (inner == PETSC_DEFAULT) {
1407: ctx->refine_inner = 0;
1408: } else {
1409: if (inner<0) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The refine inner argument must be >= 0");
1410: ctx->refine_inner = inner;
1411: }
1412: if (blsize == PETSC_DEFAULT) {
1413: ctx->refine_blocksize = 0;
1414: } else {
1415: if (blsize<0) SETERRQ(PetscObjectComm((PetscObject)eps),PETSC_ERR_ARG_OUTOFRANGE,"The refine blocksize argument must be >= 0");
1416: ctx->refine_blocksize = blsize;
1417: }
1418: return(0);
1419: }
1421: /*@
1422: EPSCISSSetRefinement - Sets the values of various refinement parameters
1423: in the CISS solver.
1425: Logically Collective on eps
1427: Input Parameters:
1428: + eps - the eigenproblem solver context
1429: . inner - number of iterative refinement iterations (inner loop)
1430: - blsize - number of iterative refinement iterations (blocksize loop)
1432: Options Database Keys:
1433: + -eps_ciss_refine_inner - Sets number of inner iterations
1434: - -eps_ciss_refine_blocksize - Sets number of blocksize iterations
1436: Level: advanced
1438: .seealso: EPSCISSGetRefinement()
1439: @*/
1440: PetscErrorCode EPSCISSSetRefinement(EPS eps,PetscInt inner,PetscInt blsize)
1441: {
1448: PetscTryMethod(eps,"EPSCISSSetRefinement_C",(EPS,PetscInt,PetscInt),(eps,inner,blsize));
1449: return(0);
1450: }
1452: static PetscErrorCode EPSCISSGetRefinement_CISS(EPS eps,PetscInt *inner,PetscInt *blsize)
1453: {
1454: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1457: if (inner) *inner = ctx->refine_inner;
1458: if (blsize) *blsize = ctx->refine_blocksize;
1459: return(0);
1460: }
1462: /*@
1463: EPSCISSGetRefinement - Gets the values of various refinement parameters
1464: in the CISS solver.
1466: Not Collective
1468: Input Parameter:
1469: . eps - the eigenproblem solver context
1471: Output Parameters:
1472: + inner - number of iterative refinement iterations (inner loop)
1473: - blsize - number of iterative refinement iterations (blocksize loop)
1475: Level: advanced
1477: .seealso: EPSCISSSetRefinement()
1478: @*/
1479: PetscErrorCode EPSCISSGetRefinement(EPS eps, PetscInt *inner, PetscInt *blsize)
1480: {
1485: PetscUseMethod(eps,"EPSCISSGetRefinement_C",(EPS,PetscInt*,PetscInt*),(eps,inner,blsize));
1486: return(0);
1487: }
1489: static PetscErrorCode EPSCISSSetUseST_CISS(EPS eps,PetscBool usest)
1490: {
1491: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1494: ctx->usest = usest;
1495: ctx->usest_set = PETSC_TRUE;
1496: eps->state = EPS_STATE_INITIAL;
1497: return(0);
1498: }
1500: /*@
1501: EPSCISSSetUseST - Sets a flag indicating that the CISS solver will
1502: use the ST object for the linear solves.
1504: Logically Collective on eps
1506: Input Parameters:
1507: + eps - the eigenproblem solver context
1508: - usest - boolean flag to use the ST object or not
1510: Options Database Keys:
1511: . -eps_ciss_usest <bool> - whether the ST object will be used or not
1513: Level: advanced
1515: .seealso: EPSCISSGetUseST()
1516: @*/
1517: PetscErrorCode EPSCISSSetUseST(EPS eps,PetscBool usest)
1518: {
1524: PetscTryMethod(eps,"EPSCISSSetUseST_C",(EPS,PetscBool),(eps,usest));
1525: return(0);
1526: }
1528: static PetscErrorCode EPSCISSGetUseST_CISS(EPS eps,PetscBool *usest)
1529: {
1530: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1533: *usest = ctx->usest;
1534: return(0);
1535: }
1537: /*@
1538: EPSCISSGetUseST - Gets the flag for using the ST object
1539: in the CISS solver.
1541: Not Collective
1543: Input Parameter:
1544: . eps - the eigenproblem solver context
1546: Output Parameters:
1547: . usest - boolean flag indicating if the ST object is being used
1549: Level: advanced
1551: .seealso: EPSCISSSetUseST()
1552: @*/
1553: PetscErrorCode EPSCISSGetUseST(EPS eps,PetscBool *usest)
1554: {
1560: PetscUseMethod(eps,"EPSCISSGetUseST_C",(EPS,PetscBool*),(eps,usest));
1561: return(0);
1562: }
1564: static PetscErrorCode EPSCISSSetQuadRule_CISS(EPS eps,EPSCISSQuadRule quad)
1565: {
1566: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1569: ctx->quad = quad;
1570: return(0);
1571: }
1573: /*@
1574: EPSCISSSetQuadRule - Sets the quadrature rule used in the CISS solver.
1576: Logically Collective on eps
1578: Input Parameters:
1579: + eps - the eigenproblem solver context
1580: - quad - the quadrature rule
1582: Options Database Key:
1583: . -eps_ciss_quadrule - Sets the quadrature rule (either 'trapezoidal' or
1584: 'chebyshev')
1586: Notes:
1587: By default, the trapezoidal rule is used (EPS_CISS_QUADRULE_TRAPEZOIDAL).
1589: If the 'chebyshev' option is specified (EPS_CISS_QUADRULE_CHEBYSHEV), then
1590: Chebyshev points are used as quadrature points.
1592: Level: advanced
1594: .seealso: EPSCISSGetQuadRule(), EPSCISSQuadRule
1595: @*/
1596: PetscErrorCode EPSCISSSetQuadRule(EPS eps,EPSCISSQuadRule quad)
1597: {
1603: PetscTryMethod(eps,"EPSCISSSetQuadRule_C",(EPS,EPSCISSQuadRule),(eps,quad));
1604: return(0);
1605: }
1607: static PetscErrorCode EPSCISSGetQuadRule_CISS(EPS eps,EPSCISSQuadRule *quad)
1608: {
1609: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1612: *quad = ctx->quad;
1613: return(0);
1614: }
1616: /*@
1617: EPSCISSGetQuadRule - Gets the quadrature rule used in the CISS solver.
1619: Not Collective
1621: Input Parameter:
1622: . eps - the eigenproblem solver context
1624: Output Parameters:
1625: . quad - quadrature rule
1627: Level: advanced
1629: .seealso: EPSCISSSetQuadRule() EPSCISSQuadRule
1630: @*/
1631: PetscErrorCode EPSCISSGetQuadRule(EPS eps,EPSCISSQuadRule *quad)
1632: {
1638: PetscUseMethod(eps,"EPSCISSGetQuadRule_C",(EPS,EPSCISSQuadRule*),(eps,quad));
1639: return(0);
1640: }
1642: static PetscErrorCode EPSCISSSetExtraction_CISS(EPS eps,EPSCISSExtraction extraction)
1643: {
1644: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1647: ctx->extraction = extraction;
1648: return(0);
1649: }
1651: /*@
1652: EPSCISSSetExtraction - Sets the extraction technique used in the CISS solver.
1654: Logically Collective on eps
1656: Input Parameters:
1657: + eps - the eigenproblem solver context
1658: - extraction - the extraction technique
1660: Options Database Key:
1661: . -eps_ciss_extraction - Sets the extraction technique (either 'ritz' or
1662: 'hankel')
1664: Notes:
1665: By default, the Rayleigh-Ritz extraction is used (EPS_CISS_EXTRACTION_RITZ).
1667: If the 'hankel' option is specified (EPS_CISS_EXTRACTION_HANKEL), then
1668: the Block Hankel method is used for extracting eigenpairs.
1670: Level: advanced
1672: .seealso: EPSCISSGetExtraction(), EPSCISSExtraction
1673: @*/
1674: PetscErrorCode EPSCISSSetExtraction(EPS eps,EPSCISSExtraction extraction)
1675: {
1681: PetscTryMethod(eps,"EPSCISSSetExtraction_C",(EPS,EPSCISSExtraction),(eps,extraction));
1682: return(0);
1683: }
1685: static PetscErrorCode EPSCISSGetExtraction_CISS(EPS eps,EPSCISSExtraction *extraction)
1686: {
1687: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1690: *extraction = ctx->extraction;
1691: return(0);
1692: }
1694: /*@
1695: EPSCISSGetExtraction - Gets the extraction technique used in the CISS solver.
1697: Not Collective
1699: Input Parameter:
1700: . eps - the eigenproblem solver context
1702: Output Parameters:
1703: + extraction - extraction technique
1705: Level: advanced
1707: .seealso: EPSCISSSetExtraction() EPSCISSExtraction
1708: @*/
1709: PetscErrorCode EPSCISSGetExtraction(EPS eps,EPSCISSExtraction *extraction)
1710: {
1716: PetscUseMethod(eps,"EPSCISSGetExtraction_C",(EPS,EPSCISSExtraction*),(eps,extraction));
1717: return(0);
1718: }
1720: static PetscErrorCode EPSCISSGetKSPs_CISS(EPS eps,PetscInt *nsolve,KSP **ksp)
1721: {
1723: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1724: PetscInt i;
1725: PC pc;
1728: if (!ctx->ksp) {
1729: if (!ctx->subcomm) { /* initialize subcomm first */
1730: EPSCISSSetUseConj(eps,&ctx->useconj);
1731: EPSCISSSetUpSubComm(eps,&ctx->num_solve_point);
1732: }
1733: PetscMalloc1(ctx->num_solve_point,&ctx->ksp);
1734: for (i=0;i<ctx->num_solve_point;i++) {
1735: KSPCreate(PetscSubcommChild(ctx->subcomm),&ctx->ksp[i]);
1736: PetscObjectIncrementTabLevel((PetscObject)ctx->ksp[i],(PetscObject)eps,1);
1737: KSPSetOptionsPrefix(ctx->ksp[i],((PetscObject)eps)->prefix);
1738: KSPAppendOptionsPrefix(ctx->ksp[i],"eps_ciss_");
1739: PetscLogObjectParent((PetscObject)eps,(PetscObject)ctx->ksp[i]);
1740: PetscObjectSetOptions((PetscObject)ctx->ksp[i],((PetscObject)eps)->options);
1741: KSPSetErrorIfNotConverged(ctx->ksp[i],PETSC_TRUE);
1742: KSPSetTolerances(ctx->ksp[i],SLEPC_DEFAULT_TOL,PETSC_DEFAULT,PETSC_DEFAULT,PETSC_DEFAULT);
1743: KSPGetPC(ctx->ksp[i],&pc);
1744: KSPSetType(ctx->ksp[i],KSPPREONLY);
1745: PCSetType(pc,PCLU);
1746: }
1747: }
1748: if (nsolve) *nsolve = ctx->num_solve_point;
1749: if (ksp) *ksp = ctx->ksp;
1750: return(0);
1751: }
1753: /*@C
1754: EPSCISSGetKSPs - Retrieve the array of linear solver objects associated with
1755: the CISS solver.
1757: Not Collective
1759: Input Parameter:
1760: . eps - the eigenproblem solver solver
1762: Output Parameters:
1763: + nsolve - number of solver objects
1764: - ksp - array of linear solver object
1766: Notes:
1767: The number of KSP solvers is equal to the number of integration points divided by
1768: the number of partitions. This value is halved in the case of real matrices with
1769: a region centered at the real axis.
1771: Level: advanced
1773: .seealso: EPSCISSSetSizes()
1774: @*/
1775: PetscErrorCode EPSCISSGetKSPs(EPS eps,PetscInt *nsolve,KSP **ksp)
1776: {
1781: PetscUseMethod(eps,"EPSCISSGetKSPs_C",(EPS,PetscInt*,KSP**),(eps,nsolve,ksp));
1782: return(0);
1783: }
1785: PetscErrorCode EPSReset_CISS(EPS eps)
1786: {
1788: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1789: PetscInt i;
1792: BVDestroy(&ctx->S);
1793: BVDestroy(&ctx->V);
1794: BVDestroy(&ctx->Y);
1795: if (!ctx->usest) {
1796: for (i=0;i<ctx->num_solve_point;i++) {
1797: KSPReset(ctx->ksp[i]);
1798: }
1799: }
1800: VecScatterDestroy(&ctx->scatterin);
1801: VecDestroy(&ctx->xsub);
1802: VecDestroy(&ctx->xdup);
1803: if (ctx->pA) {
1804: MatDestroy(&ctx->pA);
1805: MatDestroy(&ctx->pB);
1806: BVDestroy(&ctx->pV);
1807: }
1808: return(0);
1809: }
1811: PetscErrorCode EPSSetFromOptions_CISS(PetscOptionItems *PetscOptionsObject,EPS eps)
1812: {
1813: PetscErrorCode ierr;
1814: PetscReal r3,r4;
1815: PetscInt i,i1,i2,i3,i4,i5,i6,i7;
1816: PetscBool b1,b2,flg;
1817: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1818: EPSCISSQuadRule quad;
1819: EPSCISSExtraction extraction;
1822: PetscOptionsHead(PetscOptionsObject,"EPS CISS Options");
1824: EPSCISSGetSizes(eps,&i1,&i2,&i3,&i4,&i5,&b1);
1825: PetscOptionsInt("-eps_ciss_integration_points","Number of integration points","EPSCISSSetSizes",i1,&i1,NULL);
1826: PetscOptionsInt("-eps_ciss_blocksize","Block size","EPSCISSSetSizes",i2,&i2,NULL);
1827: PetscOptionsInt("-eps_ciss_moments","Moment size","EPSCISSSetSizes",i3,&i3,NULL);
1828: PetscOptionsInt("-eps_ciss_partitions","Number of partitions","EPSCISSSetSizes",i4,&i4,NULL);
1829: PetscOptionsInt("-eps_ciss_maxblocksize","Maximum block size","EPSCISSSetSizes",i5,&i5,NULL);
1830: PetscOptionsBool("-eps_ciss_realmats","True if A and B are real","EPSCISSSetSizes",b1,&b1,NULL);
1831: EPSCISSSetSizes(eps,i1,i2,i3,i4,i5,b1);
1833: EPSCISSGetThreshold(eps,&r3,&r4);
1834: PetscOptionsReal("-eps_ciss_delta","Threshold for numerical rank","EPSCISSSetThreshold",r3,&r3,NULL);
1835: PetscOptionsReal("-eps_ciss_spurious_threshold","Threshold for the spurious eigenpairs","EPSCISSSetThreshold",r4,&r4,NULL);
1836: EPSCISSSetThreshold(eps,r3,r4);
1838: EPSCISSGetRefinement(eps,&i6,&i7);
1839: PetscOptionsInt("-eps_ciss_refine_inner","Number of inner iterative refinement iterations","EPSCISSSetRefinement",i6,&i6,NULL);
1840: PetscOptionsInt("-eps_ciss_refine_blocksize","Number of blocksize iterative refinement iterations","EPSCISSSetRefinement",i7,&i7,NULL);
1841: EPSCISSSetRefinement(eps,i6,i7);
1843: EPSCISSGetUseST(eps,&b2);
1844: PetscOptionsBool("-eps_ciss_usest","Use ST for linear solves","EPSCISSSetUseST",b2,&b2,&flg);
1845: if (flg) { EPSCISSSetUseST(eps,b2); }
1847: PetscOptionsEnum("-eps_ciss_quadrule","Quadrature rule","EPSCISSSetQuadRule",EPSCISSQuadRules,(PetscEnum)ctx->quad,(PetscEnum*)&quad,&flg);
1848: if (flg) { EPSCISSSetQuadRule(eps,quad); }
1850: PetscOptionsEnum("-eps_ciss_extraction","Extraction technique","EPSCISSSetExtraction",EPSCISSExtractions,(PetscEnum)ctx->extraction,(PetscEnum*)&extraction,&flg);
1851: if (flg) { EPSCISSSetExtraction(eps,extraction); }
1853: PetscOptionsTail();
1855: if (!eps->rg) { EPSGetRG(eps,&eps->rg); }
1856: RGSetFromOptions(eps->rg); /* this is necessary here to set useconj */
1857: if (!ctx->ksp) { EPSCISSGetKSPs(eps,&ctx->num_solve_point,&ctx->ksp); }
1858: for (i=0;i<ctx->num_solve_point;i++) {
1859: KSPSetFromOptions(ctx->ksp[i]);
1860: }
1861: PetscSubcommSetFromOptions(ctx->subcomm);
1862: return(0);
1863: }
1865: PetscErrorCode EPSDestroy_CISS(EPS eps)
1866: {
1868: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1871: EPSCISSResetSubcomm(eps);
1872: PetscFree4(ctx->weight,ctx->omega,ctx->pp,ctx->sigma);
1873: PetscFree(eps->data);
1874: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetSizes_C",NULL);
1875: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetSizes_C",NULL);
1876: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetThreshold_C",NULL);
1877: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetThreshold_C",NULL);
1878: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetRefinement_C",NULL);
1879: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetRefinement_C",NULL);
1880: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetUseST_C",NULL);
1881: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetUseST_C",NULL);
1882: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetQuadRule_C",NULL);
1883: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetQuadRule_C",NULL);
1884: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetExtraction_C",NULL);
1885: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetExtraction_C",NULL);
1886: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetKSPs_C",NULL);
1887: return(0);
1888: }
1890: PetscErrorCode EPSView_CISS(EPS eps,PetscViewer viewer)
1891: {
1893: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1894: PetscBool isascii;
1897: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);
1898: if (isascii) {
1899: PetscViewerASCIIPrintf(viewer," sizes { integration points: %D, block size: %D, moment size: %D, partitions: %D, maximum block size: %D }\n",ctx->N,ctx->L,ctx->M,ctx->npart,ctx->L_max);
1900: if (ctx->isreal) {
1901: PetscViewerASCIIPrintf(viewer," exploiting symmetry of integration points\n");
1902: }
1903: PetscViewerASCIIPrintf(viewer," threshold { delta: %g, spurious threshold: %g }\n",(double)ctx->delta,(double)ctx->spurious_threshold);
1904: PetscViewerASCIIPrintf(viewer," iterative refinement { inner: %D, blocksize: %D }\n",ctx->refine_inner, ctx->refine_blocksize);
1905: PetscViewerASCIIPrintf(viewer," extraction: %s\n",EPSCISSExtractions[ctx->extraction]);
1906: PetscViewerASCIIPrintf(viewer," quadrature rule: %s\n",EPSCISSQuadRules[ctx->quad]);
1907: if (ctx->usest) {
1908: PetscViewerASCIIPrintf(viewer," using ST for linear solves\n");
1909: } else {
1910: if (!ctx->ksp) { EPSCISSGetKSPs(eps,&ctx->num_solve_point,&ctx->ksp); }
1911: PetscViewerASCIIPushTab(viewer);
1912: KSPView(ctx->ksp[0],viewer);
1913: PetscViewerASCIIPopTab(viewer);
1914: }
1915: }
1916: return(0);
1917: }
1919: PetscErrorCode EPSSetDefaultST_CISS(EPS eps)
1920: {
1922: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1923: PetscBool usest = ctx->usest;
1926: if (!((PetscObject)eps->st)->type_name) {
1927: if (!ctx->usest_set) usest = (ctx->npart>1)? PETSC_FALSE: PETSC_TRUE;
1928: if (usest) {
1929: STSetType(eps->st,STSINVERT);
1930: } else {
1931: /* we are not going to use ST, so avoid factorizing the matrix */
1932: STSetType(eps->st,STSHIFT);
1933: }
1934: }
1935: return(0);
1936: }
1938: SLEPC_EXTERN PetscErrorCode EPSCreate_CISS(EPS eps)
1939: {
1941: EPS_CISS *ctx = (EPS_CISS*)eps->data;
1944: PetscNewLog(eps,&ctx);
1945: eps->data = ctx;
1947: eps->useds = PETSC_TRUE;
1948: eps->categ = EPS_CATEGORY_CONTOUR;
1950: eps->ops->solve = EPSSolve_CISS;
1951: eps->ops->setup = EPSSetUp_CISS;
1952: eps->ops->setfromoptions = EPSSetFromOptions_CISS;
1953: eps->ops->destroy = EPSDestroy_CISS;
1954: eps->ops->reset = EPSReset_CISS;
1955: eps->ops->view = EPSView_CISS;
1956: eps->ops->computevectors = EPSComputeVectors_Schur;
1957: eps->ops->setdefaultst = EPSSetDefaultST_CISS;
1959: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetSizes_C",EPSCISSSetSizes_CISS);
1960: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetSizes_C",EPSCISSGetSizes_CISS);
1961: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetThreshold_C",EPSCISSSetThreshold_CISS);
1962: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetThreshold_C",EPSCISSGetThreshold_CISS);
1963: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetRefinement_C",EPSCISSSetRefinement_CISS);
1964: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetRefinement_C",EPSCISSGetRefinement_CISS);
1965: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetUseST_C",EPSCISSSetUseST_CISS);
1966: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetUseST_C",EPSCISSGetUseST_CISS);
1967: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetQuadRule_C",EPSCISSSetQuadRule_CISS);
1968: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetQuadRule_C",EPSCISSGetQuadRule_CISS);
1969: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSSetExtraction_C",EPSCISSSetExtraction_CISS);
1970: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetExtraction_C",EPSCISSGetExtraction_CISS);
1971: PetscObjectComposeFunction((PetscObject)eps,"EPSCISSGetKSPs_C",EPSCISSGetKSPs_CISS);
1972: /* set default values of parameters */
1973: ctx->N = 32;
1974: ctx->L = 16;
1975: ctx->M = ctx->N/4;
1976: ctx->delta = 1e-12;
1977: ctx->L_max = 64;
1978: ctx->spurious_threshold = 1e-4;
1979: ctx->usest = PETSC_TRUE;
1980: ctx->usest_set = PETSC_FALSE;
1981: ctx->isreal = PETSC_FALSE;
1982: ctx->refine_inner = 0;
1983: ctx->refine_blocksize = 0;
1984: ctx->npart = 1;
1985: ctx->quad = (EPSCISSQuadRule)0;
1986: ctx->extraction = EPS_CISS_EXTRACTION_RITZ;
1987: return(0);
1988: }