Actual source code: pjd.c
slepc-3.19.0 2023-03-31
1: /*
2: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3: SLEPc - Scalable Library for Eigenvalue Problem Computations
4: Copyright (c) 2002-, 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 polynomial eigensolver: "jd"
13: Method: Jacobi-Davidson
15: Algorithm:
17: Jacobi-Davidson for polynomial eigenvalue problems.
19: References:
21: [1] C. Campos and J.E. Roman, "A polynomial Jacobi-Davidson solver
22: with support for non-monomial bases and deflation", BIT Numer.
23: Math. 60:295-318, 2020.
25: [2] G.L.G. Sleijpen et al., "Jacobi-Davidson type methods for
26: generalized eigenproblems and polynomial eigenproblems", BIT
27: 36(3):595-633, 1996.
29: [3] Feng-Nan Hwang, Zih-Hao Wei, Tsung-Ming Huang, Weichung Wang,
30: "A Parallel Additive Schwarz Preconditioned Jacobi-Davidson
31: Algorithm for Polynomial Eigenvalue Problems in Quantum Dot
32: Simulation", J. Comput. Phys. 229(8):2932-2947, 2010.
33: */
35: #include <slepc/private/pepimpl.h>
36: #include <slepcblaslapack.h>
38: static PetscBool cited = PETSC_FALSE;
39: static const char citation[] =
40: "@Article{slepc-slice-qep,\n"
41: " author = \"C. Campos and J. E. Roman\",\n"
42: " title = \"A polynomial {Jacobi-Davidson} solver with support for non-monomial bases and deflation\",\n"
43: " journal = \"{BIT} Numer. Math.\",\n"
44: " volume = \"60\",\n"
45: " pages = \"295--318\",\n"
46: " year = \"2020,\"\n"
47: " doi = \"https://doi.org/10.1007/s10543-019-00778-z\"\n"
48: "}\n";
50: typedef struct {
51: PetscReal keep; /* restart parameter */
52: PetscReal fix; /* fix parameter */
53: PetscBool reusepc; /* flag indicating whether pc is rebuilt or not */
54: BV V; /* work basis vectors to store the search space */
55: BV W; /* work basis vectors to store the test space */
56: BV *TV; /* work basis vectors to store T*V (each TV[i] is the coefficient for \lambda^i of T*V for the extended T) */
57: BV *AX; /* work basis vectors to store A_i*X for locked eigenvectors */
58: BV N[2]; /* auxiliary work BVs */
59: BV X; /* locked eigenvectors */
60: PetscScalar *T; /* matrix of the invariant pair */
61: PetscScalar *Tj; /* matrix containing the powers of the invariant pair matrix */
62: PetscScalar *XpX; /* X^H*X */
63: PetscInt ld; /* leading dimension for Tj and XpX */
64: PC pcshell; /* preconditioner including basic precond+projector */
65: Mat Pshell; /* auxiliary shell matrix */
66: PetscInt nlock; /* number of locked vectors in the invariant pair */
67: Vec vtempl; /* reference nested vector */
68: PetscInt midx; /* minimality index */
69: PetscInt mmidx; /* maximum allowed minimality index */
70: PEPJDProjection proj; /* projection type (orthogonal, harmonic) */
71: } PEP_JD;
73: typedef struct {
74: PEP pep;
75: PC pc; /* basic preconditioner */
76: Vec Bp[2]; /* preconditioned residual of derivative polynomial, B\p */
77: Vec u[2]; /* Ritz vector */
78: PetscScalar gamma[2]; /* precomputed scalar u'*B\p */
79: PetscScalar theta;
80: PetscScalar *M;
81: PetscScalar *ps;
82: PetscInt ld;
83: Vec *work;
84: Mat PPr;
85: BV X;
86: PetscInt n;
87: } PEP_JD_PCSHELL;
89: typedef struct {
90: Mat Pr,Pi; /* matrix polynomial evaluated at theta */
91: PEP pep;
92: Vec *work;
93: PetscScalar theta[2];
94: } PEP_JD_MATSHELL;
96: /*
97: Duplicate and resize auxiliary basis
98: */
99: static PetscErrorCode PEPJDDuplicateBasis(PEP pep,BV *basis)
100: {
101: PEP_JD *pjd = (PEP_JD*)pep->data;
102: PetscInt nloc,m;
103: BVType type;
104: BVOrthogType otype;
105: BVOrthogRefineType oref;
106: PetscReal oeta;
107: BVOrthogBlockType oblock;
109: PetscFunctionBegin;
110: if (pjd->ld>1) {
111: PetscCall(BVCreate(PetscObjectComm((PetscObject)pep),basis));
112: PetscCall(BVGetSizes(pep->V,&nloc,NULL,&m));
113: nloc += pjd->ld-1;
114: PetscCall(BVSetSizes(*basis,nloc,PETSC_DECIDE,m));
115: PetscCall(BVGetType(pep->V,&type));
116: PetscCall(BVSetType(*basis,type));
117: PetscCall(BVGetOrthogonalization(pep->V,&otype,&oref,&oeta,&oblock));
118: PetscCall(BVSetOrthogonalization(*basis,otype,oref,oeta,oblock));
119: PetscCall(PetscObjectStateIncrease((PetscObject)*basis));
120: } else PetscCall(BVDuplicate(pep->V,basis));
121: PetscFunctionReturn(PETSC_SUCCESS);
122: }
124: PetscErrorCode PEPSetUp_JD(PEP pep)
125: {
126: PEP_JD *pjd = (PEP_JD*)pep->data;
127: PetscBool isprecond,flg;
128: PetscRandom rand;
129: PetscInt i;
131: PetscFunctionBegin;
132: PetscCall(PEPSetDimensions_Default(pep,pep->nev,&pep->ncv,&pep->mpd));
133: if (pep->max_it==PETSC_DEFAULT) pep->max_it = PetscMax(100,2*pep->n/pep->ncv);
134: if (!pep->which) pep->which = PEP_TARGET_MAGNITUDE;
135: PetscCheck(pep->which==PEP_TARGET_MAGNITUDE || pep->which==PEP_TARGET_REAL || pep->which==PEP_TARGET_IMAGINARY,PetscObjectComm((PetscObject)pep),PETSC_ERR_SUP,"The JD solver supports only target which, see PEPSetWhichEigenpairs()");
137: PetscCall(PetscObjectTypeCompare((PetscObject)pep->st,STPRECOND,&isprecond));
138: PetscCheck(isprecond,PetscObjectComm((PetscObject)pep),PETSC_ERR_SUP,"The JD solver only works with PRECOND spectral transformation");
140: PetscCall(STGetTransform(pep->st,&flg));
141: PetscCheck(!flg,PetscObjectComm((PetscObject)pep),PETSC_ERR_SUP,"The JD solver requires the ST transform flag unset, see STSetTransform()");
142: PEPCheckIgnored(pep,PEP_FEATURE_EXTRACT);
144: if (!pjd->mmidx) pjd->mmidx = pep->nmat-1;
145: pjd->mmidx = PetscMin(pjd->mmidx,pep->nmat-1);
146: if (!pjd->keep) pjd->keep = 0.5;
147: PetscCall(PEPBasisCoefficients(pep,pep->pbc));
148: PetscCall(PEPAllocateSolution(pep,0));
149: PetscCall(BVGetRandomContext(pep->V,&rand)); /* make sure the random context is available when duplicating */
150: PetscCall(PEPSetWorkVecs(pep,5));
151: pjd->ld = pep->nev;
152: #if !defined (PETSC_USE_COMPLEX)
153: pjd->ld++;
154: #endif
155: PetscCall(PetscMalloc2(pep->nmat,&pjd->TV,pep->nmat,&pjd->AX));
156: for (i=0;i<pep->nmat;i++) PetscCall(PEPJDDuplicateBasis(pep,pjd->TV+i));
157: if (pjd->ld>1) {
158: PetscCall(PEPJDDuplicateBasis(pep,&pjd->V));
159: PetscCall(BVSetFromOptions(pjd->V));
160: for (i=0;i<pep->nmat;i++) PetscCall(BVDuplicateResize(pep->V,pjd->ld-1,pjd->AX+i));
161: PetscCall(BVDuplicateResize(pep->V,pjd->ld-1,pjd->N));
162: PetscCall(BVDuplicateResize(pep->V,pjd->ld-1,pjd->N+1));
163: pjd->X = pep->V;
164: PetscCall(PetscCalloc3((pjd->ld)*(pjd->ld),&pjd->XpX,pep->ncv*pep->ncv,&pjd->T,pjd->ld*pjd->ld*pep->nmat,&pjd->Tj));
165: } else pjd->V = pep->V;
166: if (pjd->proj==PEP_JD_PROJECTION_HARMONIC) PetscCall(PEPJDDuplicateBasis(pep,&pjd->W));
167: else pjd->W = pjd->V;
168: PetscCall(DSSetType(pep->ds,DSPEP));
169: PetscCall(DSPEPSetDegree(pep->ds,pep->nmat-1));
170: if (pep->basis!=PEP_BASIS_MONOMIAL) PetscCall(DSPEPSetCoefficients(pep->ds,pep->pbc));
171: PetscCall(DSAllocate(pep->ds,pep->ncv));
172: PetscFunctionReturn(PETSC_SUCCESS);
173: }
175: /*
176: Updates columns (low to (high-1)) of TV[i]
177: */
178: static PetscErrorCode PEPJDUpdateTV(PEP pep,PetscInt low,PetscInt high,Vec *w)
179: {
180: PEP_JD *pjd = (PEP_JD*)pep->data;
181: PetscInt pp,col,i,nloc,nconv;
182: Vec v1,v2,t1,t2;
183: PetscScalar *array1,*array2,*x2,*xx,*N,*Np,*y2=NULL,zero=0.0,sone=1.0,*pT,fact,*psc;
184: PetscReal *cg,*ca,*cb;
185: PetscMPIInt rk,np;
186: PetscBLASInt n_,ld_,one=1;
187: Mat T;
188: BV pbv;
190: PetscFunctionBegin;
191: ca = pep->pbc; cb = ca+pep->nmat; cg = cb + pep->nmat;
192: nconv = pjd->nlock;
193: PetscCall(PetscMalloc5(nconv,&x2,nconv,&xx,nconv*nconv,&pT,nconv*nconv,&N,nconv*nconv,&Np));
194: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)pep),&rk));
195: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pep),&np));
196: PetscCall(BVGetSizes(pep->V,&nloc,NULL,NULL));
197: t1 = w[0];
198: t2 = w[1];
199: PetscCall(PetscBLASIntCast(pjd->nlock,&n_));
200: PetscCall(PetscBLASIntCast(pjd->ld,&ld_));
201: if (nconv) {
202: for (i=0;i<nconv;i++) PetscCall(PetscArraycpy(pT+i*nconv,pjd->T+i*pep->ncv,nconv));
203: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nconv,nconv,pT,&T));
204: }
205: for (col=low;col<high;col++) {
206: PetscCall(BVGetColumn(pjd->V,col,&v1));
207: PetscCall(VecGetArray(v1,&array1));
208: if (nconv>0) {
209: for (i=0;i<nconv;i++) x2[i] = array1[nloc+i]* PetscSqrtReal(np);
210: }
211: PetscCall(VecPlaceArray(t1,array1));
212: if (nconv) {
213: PetscCall(BVSetActiveColumns(pjd->N[0],0,nconv));
214: PetscCall(BVSetActiveColumns(pjd->N[1],0,nconv));
215: PetscCall(BVDotVec(pjd->X,t1,xx));
216: }
217: for (pp=pep->nmat-1;pp>=0;pp--) {
218: PetscCall(BVGetColumn(pjd->TV[pp],col,&v2));
219: PetscCall(VecGetArray(v2,&array2));
220: PetscCall(VecPlaceArray(t2,array2));
221: PetscCall(MatMult(pep->A[pp],t1,t2));
222: if (nconv) {
223: if (pp<pep->nmat-3) {
224: PetscCall(BVMult(pjd->N[0],1.0,-cg[pp+2],pjd->AX[pp+1],NULL));
225: PetscCall(MatShift(T,-cb[pp+1]));
226: PetscCall(BVMult(pjd->N[0],1.0/ca[pp],1.0/ca[pp],pjd->N[1],T));
227: pbv = pjd->N[0]; pjd->N[0] = pjd->N[1]; pjd->N[1] = pbv;
228: PetscCall(BVMultVec(pjd->N[1],1.0,1.0,t2,x2));
229: PetscCall(MatShift(T,cb[pp+1]));
230: } else if (pp==pep->nmat-3) {
231: PetscCall(BVCopy(pjd->AX[pp+2],pjd->N[0]));
232: PetscCall(BVScale(pjd->N[0],1/ca[pp+1]));
233: PetscCall(BVCopy(pjd->AX[pp+1],pjd->N[1]));
234: PetscCall(MatShift(T,-cb[pp+1]));
235: PetscCall(BVMult(pjd->N[1],1.0/ca[pp],1.0/ca[pp],pjd->N[0],T));
236: PetscCall(BVMultVec(pjd->N[1],1.0,1.0,t2,x2));
237: PetscCall(MatShift(T,cb[pp+1]));
238: } else if (pp==pep->nmat-2) PetscCall(BVMultVec(pjd->AX[pp+1],1.0/ca[pp],1.0,t2,x2));
239: if (pp<pjd->midx) {
240: y2 = array2+nloc;
241: PetscCallBLAS("BLASgemv",BLASgemv_("C",&n_,&n_,&sone,pjd->Tj+pjd->ld*pjd->ld*pp,&ld_,xx,&one,&zero,y2,&one));
242: if (pp<pjd->midx-2) {
243: fact = -cg[pp+2];
244: PetscCallBLAS("BLASgemm",BLASgemm_("C","N",&n_,&n_,&n_,&sone,pjd->Tj+(pp+1)*pjd->ld*pjd->ld,&ld_,pjd->XpX,&ld_,&fact,Np,&n_));
245: fact = 1/ca[pp];
246: PetscCall(MatShift(T,-cb[pp+1]));
247: PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&n_,&n_,&n_,&fact,N,&n_,pT,&n_,&fact,Np,&n_));
248: PetscCall(MatShift(T,cb[pp+1]));
249: psc = Np; Np = N; N = psc;
250: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n_,&n_,&sone,N,&n_,x2,&one,&sone,y2,&one));
251: } else if (pp==pjd->midx-2) {
252: fact = 1/ca[pp];
253: PetscCallBLAS("BLASgemm",BLASgemm_("C","N",&n_,&n_,&n_,&fact,pjd->Tj+(pp+1)*pjd->ld*pjd->ld,&ld_,pjd->XpX,&ld_,&zero,N,&n_));
254: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n_,&n_,&sone,N,&n_,x2,&one,&sone,y2,&one));
255: } else if (pp==pjd->midx-1) PetscCall(PetscArrayzero(Np,nconv*nconv));
256: }
257: for (i=0;i<nconv;i++) array2[nloc+i] /= PetscSqrtReal(np);
258: }
259: PetscCall(VecResetArray(t2));
260: PetscCall(VecRestoreArray(v2,&array2));
261: PetscCall(BVRestoreColumn(pjd->TV[pp],col,&v2));
262: }
263: PetscCall(VecResetArray(t1));
264: PetscCall(VecRestoreArray(v1,&array1));
265: PetscCall(BVRestoreColumn(pjd->V,col,&v1));
266: }
267: if (nconv) PetscCall(MatDestroy(&T));
268: PetscCall(PetscFree5(x2,xx,pT,N,Np));
269: PetscFunctionReturn(PETSC_SUCCESS);
270: }
272: /*
273: RRQR of X. Xin*P=Xou*R. Rank of R is rk
274: */
275: static PetscErrorCode PEPJDOrthogonalize(PetscInt row,PetscInt col,PetscScalar *X,PetscInt ldx,PetscInt *rk,PetscInt *P,PetscScalar *R,PetscInt ldr)
276: {
277: PetscInt i,j,n,r;
278: PetscBLASInt row_,col_,ldx_,*p,lwork,info,n_;
279: PetscScalar *tau,*work;
280: PetscReal tol,*rwork;
282: PetscFunctionBegin;
283: PetscCall(PetscBLASIntCast(row,&row_));
284: PetscCall(PetscBLASIntCast(col,&col_));
285: PetscCall(PetscBLASIntCast(ldx,&ldx_));
286: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
287: n = PetscMin(row,col);
288: PetscCall(PetscBLASIntCast(n,&n_));
289: lwork = 3*col_+1;
290: PetscCall(PetscMalloc4(col,&p,n,&tau,lwork,&work,2*col,&rwork));
291: for (i=1;i<col;i++) p[i] = 0;
292: p[0] = 1;
294: /* rank revealing QR */
295: #if defined(PETSC_USE_COMPLEX)
296: PetscCallBLAS("LAPACKgeqp3",LAPACKgeqp3_(&row_,&col_,X,&ldx_,p,tau,work,&lwork,rwork,&info));
297: #else
298: PetscCallBLAS("LAPACKgeqp3",LAPACKgeqp3_(&row_,&col_,X,&ldx_,p,tau,work,&lwork,&info));
299: #endif
300: SlepcCheckLapackInfo("geqp3",info);
301: if (P) for (i=0;i<col;i++) P[i] = p[i]-1;
303: /* rank computation */
304: tol = PetscMax(row,col)*PETSC_MACHINE_EPSILON*PetscAbsScalar(X[0]);
305: r = 1;
306: for (i=1;i<n;i++) {
307: if (PetscAbsScalar(X[i+ldx*i])>tol) r++;
308: else break;
309: }
310: if (rk) *rk=r;
312: /* copy upper triangular matrix if requested */
313: if (R) {
314: for (i=0;i<r;i++) {
315: PetscCall(PetscArrayzero(R+i*ldr,r));
316: for (j=0;j<=i;j++) R[i*ldr+j] = X[i*ldx+j];
317: }
318: }
319: PetscCallBLAS("LAPACKorgqr",LAPACKorgqr_(&row_,&n_,&n_,X,&ldx_,tau,work,&lwork,&info));
320: SlepcCheckLapackInfo("orgqr",info);
321: PetscCall(PetscFPTrapPop());
322: PetscCall(PetscFree4(p,tau,work,rwork));
323: PetscFunctionReturn(PETSC_SUCCESS);
324: }
326: /*
327: Application of extended preconditioner
328: */
329: static PetscErrorCode PEPJDExtendedPCApply(PC pc,Vec x,Vec y)
330: {
331: PetscInt i,j,nloc,n,ld=0;
332: PetscMPIInt np;
333: Vec tx,ty;
334: PEP_JD_PCSHELL *ctx;
335: const PetscScalar *array1;
336: PetscScalar *x2=NULL,*t=NULL,*ps=NULL,*array2,zero=0.0,sone=1.0;
337: PetscBLASInt one=1,ld_,n_,ncv_;
338: PEP_JD *pjd=NULL;
340: PetscFunctionBegin;
341: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pc),&np));
342: PetscCall(PCShellGetContext(pc,&ctx));
343: n = ctx->n;
344: if (n) {
345: pjd = (PEP_JD*)ctx->pep->data;
346: ps = ctx->ps;
347: ld = pjd->ld;
348: PetscCall(PetscMalloc2(n,&x2,n,&t));
349: PetscCall(VecGetLocalSize(ctx->work[0],&nloc));
350: PetscCall(VecGetArrayRead(x,&array1));
351: for (i=0;i<n;i++) x2[i] = array1[nloc+i]* PetscSqrtReal(np);
352: PetscCall(VecRestoreArrayRead(x,&array1));
353: }
355: /* y = B\x apply PC */
356: tx = ctx->work[0];
357: ty = ctx->work[1];
358: PetscCall(VecGetArrayRead(x,&array1));
359: PetscCall(VecPlaceArray(tx,array1));
360: PetscCall(VecGetArray(y,&array2));
361: PetscCall(VecPlaceArray(ty,array2));
362: PetscCall(PCApply(ctx->pc,tx,ty));
363: if (n) {
364: PetscCall(PetscBLASIntCast(ld,&ld_));
365: PetscCall(PetscBLASIntCast(n,&n_));
366: for (i=0;i<n;i++) {
367: t[i] = 0.0;
368: for (j=0;j<n;j++) t[i] += ctx->M[i+j*ld]*x2[j];
369: }
370: if (pjd->midx==1) {
371: PetscCall(PetscBLASIntCast(ctx->pep->ncv,&ncv_));
372: for (i=0;i<n;i++) pjd->T[i*(1+ctx->pep->ncv)] -= ctx->theta;
373: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n_,&n_,&sone,pjd->T,&ncv_,t,&one,&zero,x2,&one));
374: for (i=0;i<n;i++) pjd->T[i*(1+ctx->pep->ncv)] += ctx->theta;
375: for (i=0;i<n;i++) array2[nloc+i] = x2[i];
376: for (i=0;i<n;i++) x2[i] = -t[i];
377: } else {
378: for (i=0;i<n;i++) array2[nloc+i] = t[i];
379: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n_,&n_,&sone,ps,&ld_,t,&one,&zero,x2,&one));
380: }
381: for (i=0;i<n;i++) array2[nloc+i] /= PetscSqrtReal(np);
382: PetscCall(BVSetActiveColumns(pjd->X,0,n));
383: PetscCall(BVMultVec(pjd->X,-1.0,1.0,ty,x2));
384: PetscCall(PetscFree2(x2,t));
385: }
386: PetscCall(VecResetArray(tx));
387: PetscCall(VecResetArray(ty));
388: PetscCall(VecRestoreArrayRead(x,&array1));
389: PetscCall(VecRestoreArray(y,&array2));
390: PetscFunctionReturn(PETSC_SUCCESS);
391: }
393: /*
394: Application of shell preconditioner:
395: y = B\x - eta*B\p, with eta = (u'*B\x)/(u'*B\p)
396: */
397: static PetscErrorCode PCShellApply_PEPJD(PC pc,Vec x,Vec y)
398: {
399: PetscScalar rr,eta;
400: PEP_JD_PCSHELL *ctx;
401: PetscInt sz;
402: const Vec *xs,*ys;
403: #if !defined(PETSC_USE_COMPLEX)
404: PetscScalar rx,xr,xx;
405: #endif
407: PetscFunctionBegin;
408: PetscCall(PCShellGetContext(pc,&ctx));
409: PetscCall(VecCompGetSubVecs(x,&sz,&xs));
410: PetscCall(VecCompGetSubVecs(y,NULL,&ys));
411: /* y = B\x apply extended PC */
412: PetscCall(PEPJDExtendedPCApply(pc,xs[0],ys[0]));
413: #if !defined(PETSC_USE_COMPLEX)
414: if (sz==2) PetscCall(PEPJDExtendedPCApply(pc,xs[1],ys[1]));
415: #endif
417: /* Compute eta = u'*y / u'*Bp */
418: PetscCall(VecDot(ys[0],ctx->u[0],&rr));
419: eta = -rr*ctx->gamma[0];
420: #if !defined(PETSC_USE_COMPLEX)
421: if (sz==2) {
422: PetscCall(VecDot(ys[0],ctx->u[1],&xr));
423: PetscCall(VecDot(ys[1],ctx->u[0],&rx));
424: PetscCall(VecDot(ys[1],ctx->u[1],&xx));
425: eta += -ctx->gamma[0]*xx-ctx->gamma[1]*(-xr+rx);
426: }
427: #endif
428: eta /= ctx->gamma[0]*ctx->gamma[0]+ctx->gamma[1]*ctx->gamma[1];
430: /* y = y - eta*Bp */
431: PetscCall(VecAXPY(ys[0],eta,ctx->Bp[0]));
432: #if !defined(PETSC_USE_COMPLEX)
433: if (sz==2) {
434: PetscCall(VecAXPY(ys[1],eta,ctx->Bp[1]));
435: eta = -ctx->gamma[1]*(rr+xx)+ctx->gamma[0]*(-xr+rx);
436: eta /= ctx->gamma[0]*ctx->gamma[0]+ctx->gamma[1]*ctx->gamma[1];
437: PetscCall(VecAXPY(ys[0],eta,ctx->Bp[1]));
438: PetscCall(VecAXPY(ys[1],-eta,ctx->Bp[0]));
439: }
440: #endif
441: PetscFunctionReturn(PETSC_SUCCESS);
442: }
444: static PetscErrorCode PEPJDCopyToExtendedVec(PEP pep,Vec v,PetscScalar *a,PetscInt na,PetscInt off,Vec vex,PetscBool back)
445: {
446: PetscMPIInt np,rk,count;
447: PetscScalar *array1,*array2;
448: PetscInt nloc;
450: PetscFunctionBegin;
451: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)pep),&rk));
452: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pep),&np));
453: PetscCall(BVGetSizes(pep->V,&nloc,NULL,NULL));
454: if (v) {
455: PetscCall(VecGetArray(v,&array1));
456: PetscCall(VecGetArray(vex,&array2));
457: if (back) PetscCall(PetscArraycpy(array1,array2,nloc));
458: else PetscCall(PetscArraycpy(array2,array1,nloc));
459: PetscCall(VecRestoreArray(v,&array1));
460: PetscCall(VecRestoreArray(vex,&array2));
461: }
462: if (a) {
463: PetscCall(VecGetArray(vex,&array2));
464: if (back) {
465: PetscCall(PetscArraycpy(a,array2+nloc+off,na));
466: PetscCall(PetscMPIIntCast(na,&count));
467: PetscCallMPI(MPI_Bcast(a,count,MPIU_SCALAR,np-1,PetscObjectComm((PetscObject)pep)));
468: } else {
469: PetscCall(PetscArraycpy(array2+nloc+off,a,na));
470: PetscCall(PetscMPIIntCast(na,&count));
471: PetscCallMPI(MPI_Bcast(array2+nloc+off,count,MPIU_SCALAR,np-1,PetscObjectComm((PetscObject)pep)));
472: }
473: PetscCall(VecRestoreArray(vex,&array2));
474: }
475: PetscFunctionReturn(PETSC_SUCCESS);
476: }
478: /* Computes Phi^hat(lambda) times a vector or its derivative (depends on beval)
479: if no vector is provided returns a matrix
480: */
481: static PetscErrorCode PEPJDEvaluateHatBasis(PEP pep,PetscInt n,PetscScalar *H,PetscInt ldh,PetscScalar *beval,PetscScalar *t,PetscInt idx,PetscScalar *qpp,PetscScalar *qp,PetscScalar *q)
482: {
483: PetscInt j,i;
484: PetscBLASInt n_,ldh_,one=1;
485: PetscReal *a,*b,*g;
486: PetscScalar sone=1.0,zero=0.0;
488: PetscFunctionBegin;
489: a = pep->pbc; b=a+pep->nmat; g=b+pep->nmat;
490: PetscCall(PetscBLASIntCast(n,&n_));
491: PetscCall(PetscBLASIntCast(ldh,&ldh_));
492: if (idx<1) PetscCall(PetscArrayzero(q,t?n:n*n));
493: else if (idx==1) {
494: if (t) {for (j=0;j<n;j++) q[j] = t[j]*beval[idx-1]/a[0];}
495: else {
496: PetscCall(PetscArrayzero(q,n*n));
497: for (j=0;j<n;j++) q[(j+1)*n] = beval[idx-1]/a[0];
498: }
499: } else {
500: if (t) {
501: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n_,&n_,&sone,H,&ldh_,qp,&one,&zero,q,&one));
502: for (j=0;j<n;j++) {
503: q[j] += beval[idx-1]*t[j]-b[idx-1]*qp[j]-g[idx-1]*qpp[j];
504: q[j] /= a[idx-1];
505: }
506: } else {
507: PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&n_,&n_,&n_,&sone,H,&ldh_,qp,&n_,&zero,q,&n_));
508: for (j=0;j<n;j++) {
509: q[j+n*j] += beval[idx-1];
510: for (i=0;i<n;i++) {
511: q[i+n*j] += -b[idx-1]*qp[j*n+i]-g[idx-1]*qpp[j*n+i];
512: q[i+n*j] /= a[idx-1];
513: }
514: }
515: }
516: }
517: PetscFunctionReturn(PETSC_SUCCESS);
518: }
520: static PetscErrorCode PEPJDComputeResidual(PEP pep,PetscBool derivative,PetscInt sz,Vec *u,PetscScalar *theta,Vec *p,Vec *work)
521: {
522: PEP_JD *pjd = (PEP_JD*)pep->data;
523: PetscMPIInt rk,np,count;
524: Vec tu,tp,w;
525: PetscScalar *dval,*dvali,*array1,*array2,*x2=NULL,*y2,*qj=NULL,*tt=NULL,*xx=NULL,*xxi=NULL,sone=1.0;
526: PetscInt i,j,nconv,nloc;
527: PetscBLASInt n,ld,one=1;
528: #if !defined(PETSC_USE_COMPLEX)
529: Vec tui=NULL,tpi=NULL;
530: PetscScalar *x2i=NULL,*qji=NULL,*qq,*y2i,*arrayi1,*arrayi2;
531: #endif
533: PetscFunctionBegin;
534: nconv = pjd->nlock;
535: if (!nconv) PetscCall(PetscMalloc1(2*sz*pep->nmat,&dval));
536: else {
537: PetscCall(PetscMalloc5(2*pep->nmat,&dval,2*nconv,&xx,nconv,&tt,sz*nconv,&x2,(sz==2?3:1)*nconv*pep->nmat,&qj));
538: PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)pep),&rk));
539: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pep),&np));
540: PetscCall(BVGetSizes(pep->V,&nloc,NULL,NULL));
541: PetscCall(VecGetArray(u[0],&array1));
542: for (i=0;i<nconv;i++) x2[i] = array1[nloc+i]*PetscSqrtReal(np);
543: PetscCall(VecRestoreArray(u[0],&array1));
544: #if !defined(PETSC_USE_COMPLEX)
545: if (sz==2) {
546: x2i = x2+nconv;
547: PetscCall(VecGetArray(u[1],&arrayi1));
548: for (i=0;i<nconv;i++) x2i[i] = arrayi1[nloc+i]*PetscSqrtReal(np);
549: PetscCall(VecRestoreArray(u[1],&arrayi1));
550: }
551: #endif
552: }
553: dvali = dval+pep->nmat;
554: tu = work[0];
555: tp = work[1];
556: w = work[2];
557: PetscCall(VecGetArray(u[0],&array1));
558: PetscCall(VecPlaceArray(tu,array1));
559: PetscCall(VecGetArray(p[0],&array2));
560: PetscCall(VecPlaceArray(tp,array2));
561: PetscCall(VecSet(tp,0.0));
562: #if !defined(PETSC_USE_COMPLEX)
563: if (sz==2) {
564: tui = work[3];
565: tpi = work[4];
566: PetscCall(VecGetArray(u[1],&arrayi1));
567: PetscCall(VecPlaceArray(tui,arrayi1));
568: PetscCall(VecGetArray(p[1],&arrayi2));
569: PetscCall(VecPlaceArray(tpi,arrayi2));
570: PetscCall(VecSet(tpi,0.0));
571: }
572: #endif
573: if (derivative) PetscCall(PEPEvaluateBasisDerivative(pep,theta[0],theta[1],dval,dvali));
574: else PetscCall(PEPEvaluateBasis(pep,theta[0],theta[1],dval,dvali));
575: for (i=derivative?1:0;i<pep->nmat;i++) {
576: PetscCall(MatMult(pep->A[i],tu,w));
577: PetscCall(VecAXPY(tp,dval[i],w));
578: #if !defined(PETSC_USE_COMPLEX)
579: if (sz==2) {
580: PetscCall(VecAXPY(tpi,dvali[i],w));
581: PetscCall(MatMult(pep->A[i],tui,w));
582: PetscCall(VecAXPY(tpi,dval[i],w));
583: PetscCall(VecAXPY(tp,-dvali[i],w));
584: }
585: #endif
586: }
587: if (nconv) {
588: for (i=0;i<pep->nmat;i++) PetscCall(PEPJDEvaluateHatBasis(pep,nconv,pjd->T,pep->ncv,dval,x2,i,i>1?qj+(i-2)*nconv:NULL,i>0?qj+(i-1)*nconv:NULL,qj+i*nconv));
589: #if !defined(PETSC_USE_COMPLEX)
590: if (sz==2) {
591: qji = qj+nconv*pep->nmat;
592: qq = qji+nconv*pep->nmat;
593: for (i=0;i<pep->nmat;i++) PetscCall(PEPJDEvaluateHatBasis(pep,nconv,pjd->T,pep->ncv,dvali,x2i,i,i>1?qji+(i-2)*nconv:NULL,i>0?qji+(i-1)*nconv:NULL,qji+i*nconv));
594: for (i=0;i<nconv*pep->nmat;i++) qj[i] -= qji[i];
595: for (i=0;i<pep->nmat;i++) {
596: PetscCall(PEPJDEvaluateHatBasis(pep,nconv,pjd->T,pep->ncv,dval,x2i,i,i>1?qji+(i-2)*nconv:NULL,i>0?qji+(i-1)*nconv:NULL,qji+i*nconv));
597: PetscCall(PEPJDEvaluateHatBasis(pep,nconv,pjd->T,pep->ncv,dvali,x2,i,i>1?qq+(i-2)*nconv:NULL,i>0?qq+(i-1)*nconv:NULL,qq+i*nconv));
598: }
599: for (i=0;i<nconv*pep->nmat;i++) qji[i] += qq[i];
600: for (i=derivative?2:1;i<pep->nmat;i++) PetscCall(BVMultVec(pjd->AX[i],1.0,1.0,tpi,qji+i*nconv));
601: }
602: #endif
603: for (i=derivative?2:1;i<pep->nmat;i++) PetscCall(BVMultVec(pjd->AX[i],1.0,1.0,tp,qj+i*nconv));
605: /* extended vector part */
606: PetscCall(BVSetActiveColumns(pjd->X,0,nconv));
607: PetscCall(BVDotVec(pjd->X,tu,xx));
608: xxi = xx+nconv;
609: #if !defined(PETSC_USE_COMPLEX)
610: if (sz==2) PetscCall(BVDotVec(pjd->X,tui,xxi));
611: #endif
612: if (sz==1) PetscCall(PetscArrayzero(xxi,nconv));
613: if (rk==np-1) {
614: PetscCall(PetscBLASIntCast(nconv,&n));
615: PetscCall(PetscBLASIntCast(pjd->ld,&ld));
616: y2 = array2+nloc;
617: PetscCall(PetscArrayzero(y2,nconv));
618: for (j=derivative?1:0;j<pjd->midx;j++) {
619: for (i=0;i<nconv;i++) tt[i] = dval[j]*xx[i]-dvali[j]*xxi[i];
620: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&sone,pjd->XpX,&ld,qj+j*nconv,&one,&sone,tt,&one));
621: PetscCallBLAS("BLASgemv",BLASgemv_("C",&n,&n,&sone,pjd->Tj+j*ld*ld,&ld,tt,&one,&sone,y2,&one));
622: }
623: for (i=0;i<nconv;i++) array2[nloc+i] /= PetscSqrtReal(np);
624: #if !defined(PETSC_USE_COMPLEX)
625: if (sz==2) {
626: y2i = arrayi2+nloc;
627: PetscCall(PetscArrayzero(y2i,nconv));
628: for (j=derivative?1:0;j<pjd->midx;j++) {
629: for (i=0;i<nconv;i++) tt[i] = dval[j]*xxi[i]+dvali[j]*xx[i];
630: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&sone,pjd->XpX,&ld,qji+j*nconv,&one,&sone,tt,&one));
631: PetscCallBLAS("BLASgemv",BLASgemv_("C",&n,&n,&sone,pjd->Tj+j*ld*ld,&ld,tt,&one,&sone,y2i,&one));
632: }
633: for (i=0;i<nconv;i++) arrayi2[nloc+i] /= PetscSqrtReal(np);
634: }
635: #endif
636: }
637: PetscCall(PetscMPIIntCast(nconv,&count));
638: PetscCallMPI(MPI_Bcast(array2+nloc,count,MPIU_SCALAR,np-1,PetscObjectComm((PetscObject)pep)));
639: #if !defined(PETSC_USE_COMPLEX)
640: if (sz==2) PetscCallMPI(MPI_Bcast(arrayi2+nloc,count,MPIU_SCALAR,np-1,PetscObjectComm((PetscObject)pep)));
641: #endif
642: }
643: if (nconv) PetscCall(PetscFree5(dval,xx,tt,x2,qj));
644: else PetscCall(PetscFree(dval));
645: PetscCall(VecResetArray(tu));
646: PetscCall(VecRestoreArray(u[0],&array1));
647: PetscCall(VecResetArray(tp));
648: PetscCall(VecRestoreArray(p[0],&array2));
649: #if !defined(PETSC_USE_COMPLEX)
650: if (sz==2) {
651: PetscCall(VecResetArray(tui));
652: PetscCall(VecRestoreArray(u[1],&arrayi1));
653: PetscCall(VecResetArray(tpi));
654: PetscCall(VecRestoreArray(p[1],&arrayi2));
655: }
656: #endif
657: PetscFunctionReturn(PETSC_SUCCESS);
658: }
660: static PetscErrorCode PEPJDProcessInitialSpace(PEP pep,Vec *w)
661: {
662: PEP_JD *pjd = (PEP_JD*)pep->data;
663: PetscScalar *tt,target[2];
664: Vec vg,wg;
665: PetscInt i;
666: PetscReal norm;
668: PetscFunctionBegin;
669: PetscCall(PetscMalloc1(pjd->ld-1,&tt));
670: PetscCheck(pep->nini==0,PETSC_COMM_SELF,PETSC_ERR_SUP,"Support for initial vectors not implemented yet");
671: PetscCall(BVSetRandomColumn(pjd->V,0));
672: for (i=0;i<pjd->ld-1;i++) tt[i] = 0.0;
673: PetscCall(BVGetColumn(pjd->V,0,&vg));
674: PetscCall(PEPJDCopyToExtendedVec(pep,NULL,tt,pjd->ld-1,0,vg,PETSC_FALSE));
675: PetscCall(BVRestoreColumn(pjd->V,0,&vg));
676: PetscCall(BVNormColumn(pjd->V,0,NORM_2,&norm));
677: PetscCall(BVScaleColumn(pjd->V,0,1.0/norm));
678: if (pjd->proj==PEP_JD_PROJECTION_HARMONIC) {
679: PetscCall(BVGetColumn(pjd->V,0,&vg));
680: PetscCall(BVGetColumn(pjd->W,0,&wg));
681: PetscCall(VecSet(wg,0.0));
682: target[0] = pep->target; target[1] = 0.0;
683: PetscCall(PEPJDComputeResidual(pep,PETSC_TRUE,1,&vg,target,&wg,w));
684: PetscCall(BVRestoreColumn(pjd->W,0,&wg));
685: PetscCall(BVRestoreColumn(pjd->V,0,&vg));
686: PetscCall(BVNormColumn(pjd->W,0,NORM_2,&norm));
687: PetscCall(BVScaleColumn(pjd->W,0,1.0/norm));
688: }
689: PetscCall(PetscFree(tt));
690: PetscFunctionReturn(PETSC_SUCCESS);
691: }
693: static PetscErrorCode MatMult_PEPJD(Mat P,Vec x,Vec y)
694: {
695: PEP_JD_MATSHELL *matctx;
696: PEP_JD *pjd;
697: PetscInt i,j,nconv,nloc,nmat,ldt,ncv,sz;
698: Vec tx,ty;
699: const Vec *xs,*ys;
700: PetscScalar *array1,*array2,*x2=NULL,*y2,*tt=NULL,*xx=NULL,*xxi,theta[2],sone=1.0,*qj,*val,*vali=NULL;
701: PetscBLASInt n,ld,one=1;
702: PetscMPIInt np;
703: #if !defined(PETSC_USE_COMPLEX)
704: Vec txi=NULL,tyi=NULL;
705: PetscScalar *x2i=NULL,*qji=NULL,*qq,*y2i,*arrayi1,*arrayi2;
706: #endif
708: PetscFunctionBegin;
709: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)P),&np));
710: PetscCall(MatShellGetContext(P,&matctx));
711: pjd = (PEP_JD*)(matctx->pep->data);
712: nconv = pjd->nlock;
713: nmat = matctx->pep->nmat;
714: ncv = matctx->pep->ncv;
715: ldt = pjd->ld;
716: PetscCall(VecCompGetSubVecs(x,&sz,&xs));
717: PetscCall(VecCompGetSubVecs(y,NULL,&ys));
718: theta[0] = matctx->theta[0];
719: theta[1] = (sz==2)?matctx->theta[1]:0.0;
720: if (nconv>0) {
721: PetscCall(PetscMalloc5(nconv,&tt,sz*nconv,&x2,(sz==2?3:1)*nconv*nmat,&qj,2*nconv,&xx,2*nmat,&val));
722: PetscCall(BVGetSizes(matctx->pep->V,&nloc,NULL,NULL));
723: PetscCall(VecGetArray(xs[0],&array1));
724: for (i=0;i<nconv;i++) x2[i] = array1[nloc+i]* PetscSqrtReal(np);
725: PetscCall(VecRestoreArray(xs[0],&array1));
726: #if !defined(PETSC_USE_COMPLEX)
727: if (sz==2) {
728: x2i = x2+nconv;
729: PetscCall(VecGetArray(xs[1],&arrayi1));
730: for (i=0;i<nconv;i++) x2i[i] = arrayi1[nloc+i]* PetscSqrtReal(np);
731: PetscCall(VecRestoreArray(xs[1],&arrayi1));
732: }
733: #endif
734: vali = val+nmat;
735: }
736: tx = matctx->work[0];
737: ty = matctx->work[1];
738: PetscCall(VecGetArray(xs[0],&array1));
739: PetscCall(VecPlaceArray(tx,array1));
740: PetscCall(VecGetArray(ys[0],&array2));
741: PetscCall(VecPlaceArray(ty,array2));
742: PetscCall(MatMult(matctx->Pr,tx,ty));
743: #if !defined(PETSC_USE_COMPLEX)
744: if (sz==2) {
745: txi = matctx->work[2];
746: tyi = matctx->work[3];
747: PetscCall(VecGetArray(xs[1],&arrayi1));
748: PetscCall(VecPlaceArray(txi,arrayi1));
749: PetscCall(VecGetArray(ys[1],&arrayi2));
750: PetscCall(VecPlaceArray(tyi,arrayi2));
751: PetscCall(MatMult(matctx->Pr,txi,tyi));
752: if (theta[1]!=0.0) {
753: PetscCall(MatMult(matctx->Pi,txi,matctx->work[4]));
754: PetscCall(VecAXPY(ty,-1.0,matctx->work[4]));
755: PetscCall(MatMult(matctx->Pi,tx,matctx->work[4]));
756: PetscCall(VecAXPY(tyi,1.0,matctx->work[4]));
757: }
758: }
759: #endif
760: if (nconv>0) {
761: PetscCall(PEPEvaluateBasis(matctx->pep,theta[0],theta[1],val,vali));
762: for (i=0;i<nmat;i++) PetscCall(PEPJDEvaluateHatBasis(matctx->pep,nconv,pjd->T,ncv,val,x2,i,i>1?qj+(i-2)*nconv:NULL,i>0?qj+(i-1)*nconv:NULL,qj+i*nconv));
763: #if !defined(PETSC_USE_COMPLEX)
764: if (sz==2) {
765: qji = qj+nconv*nmat;
766: qq = qji+nconv*nmat;
767: for (i=0;i<nmat;i++) PetscCall(PEPJDEvaluateHatBasis(matctx->pep,nconv,pjd->T,matctx->pep->ncv,vali,x2i,i,i>1?qji+(i-2)*nconv:NULL,i>0?qji+(i-1)*nconv:NULL,qji+i*nconv));
768: for (i=0;i<nconv*nmat;i++) qj[i] -= qji[i];
769: for (i=0;i<nmat;i++) {
770: PetscCall(PEPJDEvaluateHatBasis(matctx->pep,nconv,pjd->T,matctx->pep->ncv,val,x2i,i,i>1?qji+(i-2)*nconv:NULL,i>0?qji+(i-1)*nconv:NULL,qji+i*nconv));
771: PetscCall(PEPJDEvaluateHatBasis(matctx->pep,nconv,pjd->T,matctx->pep->ncv,vali,x2,i,i>1?qq+(i-2)*nconv:NULL,i>0?qq+(i-1)*nconv:NULL,qq+i*nconv));
772: }
773: for (i=0;i<nconv*nmat;i++) qji[i] += qq[i];
774: for (i=1;i<matctx->pep->nmat;i++) PetscCall(BVMultVec(pjd->AX[i],1.0,1.0,tyi,qji+i*nconv));
775: }
776: #endif
777: for (i=1;i<nmat;i++) PetscCall(BVMultVec(pjd->AX[i],1.0,1.0,ty,qj+i*nconv));
779: /* extended vector part */
780: PetscCall(BVSetActiveColumns(pjd->X,0,nconv));
781: PetscCall(BVDotVec(pjd->X,tx,xx));
782: xxi = xx+nconv;
783: #if !defined(PETSC_USE_COMPLEX)
784: if (sz==2) PetscCall(BVDotVec(pjd->X,txi,xxi));
785: #endif
786: if (sz==1) PetscCall(PetscArrayzero(xxi,nconv));
787: PetscCall(PetscBLASIntCast(pjd->nlock,&n));
788: PetscCall(PetscBLASIntCast(ldt,&ld));
789: y2 = array2+nloc;
790: PetscCall(PetscArrayzero(y2,nconv));
791: for (j=0;j<pjd->midx;j++) {
792: for (i=0;i<nconv;i++) tt[i] = val[j]*xx[i]-vali[j]*xxi[i];
793: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&sone,pjd->XpX,&ld,qj+j*nconv,&one,&sone,tt,&one));
794: PetscCallBLAS("BLASgemv",BLASgemv_("C",&n,&n,&sone,pjd->Tj+j*ld*ld,&ld,tt,&one,&sone,y2,&one));
795: }
796: #if !defined(PETSC_USE_COMPLEX)
797: if (sz==2) {
798: y2i = arrayi2+nloc;
799: PetscCall(PetscArrayzero(y2i,nconv));
800: for (j=0;j<pjd->midx;j++) {
801: for (i=0;i<nconv;i++) tt[i] = val[j]*xxi[i]+vali[j]*xx[i];
802: PetscCallBLAS("BLASgemv",BLASgemv_("N",&n,&n,&sone,pjd->XpX,&ld,qji+j*nconv,&one,&sone,tt,&one));
803: PetscCallBLAS("BLASgemv",BLASgemv_("C",&n,&n,&sone,pjd->Tj+j*ld*ld,&ld,tt,&one,&sone,y2i,&one));
804: }
805: for (i=0;i<nconv;i++) arrayi2[nloc+i] /= PetscSqrtReal(np);
806: }
807: #endif
808: for (i=0;i<nconv;i++) array2[nloc+i] /= PetscSqrtReal(np);
809: PetscCall(PetscFree5(tt,x2,qj,xx,val));
810: }
811: PetscCall(VecResetArray(tx));
812: PetscCall(VecRestoreArray(xs[0],&array1));
813: PetscCall(VecResetArray(ty));
814: PetscCall(VecRestoreArray(ys[0],&array2));
815: #if !defined(PETSC_USE_COMPLEX)
816: if (sz==2) {
817: PetscCall(VecResetArray(txi));
818: PetscCall(VecRestoreArray(xs[1],&arrayi1));
819: PetscCall(VecResetArray(tyi));
820: PetscCall(VecRestoreArray(ys[1],&arrayi2));
821: }
822: #endif
823: PetscFunctionReturn(PETSC_SUCCESS);
824: }
826: static PetscErrorCode MatCreateVecs_PEPJD(Mat A,Vec *right,Vec *left)
827: {
828: PEP_JD_MATSHELL *matctx;
829: PEP_JD *pjd;
830: PetscInt kspsf=1,i;
831: Vec v[2];
833: PetscFunctionBegin;
834: PetscCall(MatShellGetContext(A,&matctx));
835: pjd = (PEP_JD*)(matctx->pep->data);
836: #if !defined (PETSC_USE_COMPLEX)
837: kspsf = 2;
838: #endif
839: for (i=0;i<kspsf;i++) PetscCall(BVCreateVec(pjd->V,v+i));
840: if (right) PetscCall(VecCreateCompWithVecs(v,kspsf,pjd->vtempl,right));
841: if (left) PetscCall(VecCreateCompWithVecs(v,kspsf,pjd->vtempl,left));
842: for (i=0;i<kspsf;i++) PetscCall(VecDestroy(&v[i]));
843: PetscFunctionReturn(PETSC_SUCCESS);
844: }
846: static PetscErrorCode PEPJDUpdateExtendedPC(PEP pep,PetscScalar theta)
847: {
848: PEP_JD *pjd = (PEP_JD*)pep->data;
849: PEP_JD_PCSHELL *pcctx;
850: PetscInt i,j,k,n=pjd->nlock,ld=pjd->ld,deg=pep->nmat-1;
851: PetscScalar *M,*ps,*work,*U,*V,*S,*Sp,*Spp,snone=-1.0,sone=1.0,zero=0.0,*val;
852: PetscReal tol,maxeig=0.0,*sg,*rwork;
853: PetscBLASInt n_,info,ld_,*p,lw_,rk=0;
855: PetscFunctionBegin;
856: if (n) {
857: PetscCall(PCShellGetContext(pjd->pcshell,&pcctx));
858: pcctx->theta = theta;
859: pcctx->n = n;
860: M = pcctx->M;
861: PetscCall(PetscBLASIntCast(n,&n_));
862: PetscCall(PetscBLASIntCast(ld,&ld_));
863: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
864: if (pjd->midx==1) {
865: PetscCall(PetscArraycpy(M,pjd->XpX,ld*ld));
866: PetscCall(PetscCalloc2(10*n,&work,n,&p));
867: } else {
868: ps = pcctx->ps;
869: PetscCall(PetscCalloc7(2*n*n,&U,3*n*n,&S,n,&sg,10*n,&work,5*n,&rwork,n,&p,deg+1,&val));
870: V = U+n*n;
871: /* pseudo-inverse */
872: for (j=0;j<n;j++) {
873: for (i=0;i<n;i++) S[n*j+i] = -pjd->T[pep->ncv*j+i];
874: S[n*j+j] += theta;
875: }
876: lw_ = 10*n_;
877: #if !defined (PETSC_USE_COMPLEX)
878: PetscCallBLAS("LAPACKgesvd",LAPACKgesvd_("S","S",&n_,&n_,S,&n_,sg,U,&n_,V,&n_,work,&lw_,&info));
879: #else
880: PetscCallBLAS("LAPACKgesvd",LAPACKgesvd_("S","S",&n_,&n_,S,&n_,sg,U,&n_,V,&n_,work,&lw_,rwork,&info));
881: #endif
882: SlepcCheckLapackInfo("gesvd",info);
883: for (i=0;i<n;i++) maxeig = PetscMax(maxeig,sg[i]);
884: tol = 10*PETSC_MACHINE_EPSILON*n*maxeig;
885: for (j=0;j<n;j++) {
886: if (sg[j]>tol) {
887: for (i=0;i<n;i++) U[j*n+i] /= sg[j];
888: rk++;
889: } else break;
890: }
891: PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&n_,&n_,&rk,&sone,U,&n_,V,&n_,&zero,ps,&ld_));
893: /* compute M */
894: PetscCall(PEPEvaluateBasis(pep,theta,0.0,val,NULL));
895: PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&n_,&n_,&n_,&snone,pjd->XpX,&ld_,ps,&ld_,&zero,M,&ld_));
896: PetscCall(PetscArrayzero(S,2*n*n));
897: Sp = S+n*n;
898: for (j=0;j<n;j++) S[j*(n+1)] = 1.0;
899: for (k=1;k<pjd->midx;k++) {
900: for (j=0;j<n;j++) for (i=0;i<n;i++) V[j*n+i] = S[j*n+i] - ps[j*ld+i]*val[k];
901: PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&n_,&n_,&n_,&sone,pjd->XpX,&ld_,V,&n_,&zero,U,&n_));
902: PetscCallBLAS("BLASgemm",BLASgemm_("C","N",&n_,&n_,&n_,&sone,pjd->Tj+k*ld*ld,&ld_,U,&n_,&sone,M,&ld_));
903: Spp = Sp; Sp = S;
904: PetscCall(PEPJDEvaluateHatBasis(pep,n,pjd->T,pep->ncv,val,NULL,k+1,Spp,Sp,S));
905: }
906: }
907: /* inverse */
908: PetscCallBLAS("LAPACKgetrf",LAPACKgetrf_(&n_,&n_,M,&ld_,p,&info));
909: SlepcCheckLapackInfo("getrf",info);
910: PetscCallBLAS("LAPACKgetri",LAPACKgetri_(&n_,M,&ld_,p,work,&n_,&info));
911: SlepcCheckLapackInfo("getri",info);
912: PetscCall(PetscFPTrapPop());
913: if (pjd->midx==1) PetscCall(PetscFree2(work,p));
914: else PetscCall(PetscFree7(U,S,sg,work,rwork,p,val));
915: }
916: PetscFunctionReturn(PETSC_SUCCESS);
917: }
919: static PetscErrorCode PEPJDMatSetUp(PEP pep,PetscInt sz,PetscScalar *theta)
920: {
921: PEP_JD *pjd = (PEP_JD*)pep->data;
922: PEP_JD_MATSHELL *matctx;
923: PEP_JD_PCSHELL *pcctx;
924: MatStructure str;
925: PetscScalar *vals,*valsi;
926: PetscBool skipmat=PETSC_FALSE;
927: PetscInt i;
928: Mat Pr=NULL;
930: PetscFunctionBegin;
931: if (sz==2 && theta[1]==0.0) sz = 1;
932: PetscCall(MatShellGetContext(pjd->Pshell,&matctx));
933: PetscCall(PCShellGetContext(pjd->pcshell,&pcctx));
934: if (matctx->Pr && matctx->theta[0]==theta[0] && ((!matctx->Pi && sz==1) || (sz==2 && matctx->theta[1]==theta[1]))) {
935: if (pcctx->n == pjd->nlock) PetscFunctionReturn(PETSC_SUCCESS);
936: skipmat = PETSC_TRUE;
937: }
938: if (!skipmat) {
939: PetscCall(PetscMalloc2(pep->nmat,&vals,pep->nmat,&valsi));
940: PetscCall(STGetMatStructure(pep->st,&str));
941: PetscCall(PEPEvaluateBasis(pep,theta[0],theta[1],vals,valsi));
942: if (!matctx->Pr) PetscCall(MatDuplicate(pep->A[0],MAT_COPY_VALUES,&matctx->Pr));
943: else PetscCall(MatCopy(pep->A[0],matctx->Pr,str));
944: for (i=1;i<pep->nmat;i++) PetscCall(MatAXPY(matctx->Pr,vals[i],pep->A[i],str));
945: if (!pjd->reusepc) {
946: if (pcctx->PPr && sz==2) {
947: PetscCall(MatCopy(matctx->Pr,pcctx->PPr,str));
948: Pr = pcctx->PPr;
949: } else Pr = matctx->Pr;
950: }
951: matctx->theta[0] = theta[0];
952: #if !defined(PETSC_USE_COMPLEX)
953: if (sz==2) {
954: if (!matctx->Pi) PetscCall(MatDuplicate(pep->A[0],MAT_COPY_VALUES,&matctx->Pi));
955: else PetscCall(MatCopy(pep->A[1],matctx->Pi,str));
956: PetscCall(MatScale(matctx->Pi,valsi[1]));
957: for (i=2;i<pep->nmat;i++) PetscCall(MatAXPY(matctx->Pi,valsi[i],pep->A[i],str));
958: matctx->theta[1] = theta[1];
959: }
960: #endif
961: PetscCall(PetscFree2(vals,valsi));
962: }
963: if (!pjd->reusepc) {
964: if (!skipmat) {
965: PetscCall(PCSetOperators(pcctx->pc,Pr,Pr));
966: PetscCall(PCSetUp(pcctx->pc));
967: }
968: PetscCall(PEPJDUpdateExtendedPC(pep,theta[0]));
969: }
970: PetscFunctionReturn(PETSC_SUCCESS);
971: }
973: static PetscErrorCode PEPJDCreateShellPC(PEP pep,Vec *ww)
974: {
975: PEP_JD *pjd = (PEP_JD*)pep->data;
976: PEP_JD_PCSHELL *pcctx;
977: PEP_JD_MATSHELL *matctx;
978: KSP ksp;
979: PetscInt nloc,mloc,kspsf=1;
980: Vec v[2];
981: PetscScalar target[2];
982: Mat Pr;
984: PetscFunctionBegin;
985: /* Create the reference vector */
986: PetscCall(BVGetColumn(pjd->V,0,&v[0]));
987: v[1] = v[0];
988: #if !defined (PETSC_USE_COMPLEX)
989: kspsf = 2;
990: #endif
991: PetscCall(VecCreateCompWithVecs(v,kspsf,NULL,&pjd->vtempl));
992: PetscCall(BVRestoreColumn(pjd->V,0,&v[0]));
994: /* Replace preconditioner with one containing projectors */
995: PetscCall(PCCreate(PetscObjectComm((PetscObject)pep),&pjd->pcshell));
996: PetscCall(PCSetType(pjd->pcshell,PCSHELL));
997: PetscCall(PCShellSetName(pjd->pcshell,"PCPEPJD"));
998: PetscCall(PCShellSetApply(pjd->pcshell,PCShellApply_PEPJD));
999: PetscCall(PetscNew(&pcctx));
1000: PetscCall(PCShellSetContext(pjd->pcshell,pcctx));
1001: PetscCall(STGetKSP(pep->st,&ksp));
1002: PetscCall(BVCreateVec(pjd->V,&pcctx->Bp[0]));
1003: PetscCall(VecDuplicate(pcctx->Bp[0],&pcctx->Bp[1]));
1004: PetscCall(KSPGetPC(ksp,&pcctx->pc));
1005: PetscCall(PetscObjectReference((PetscObject)pcctx->pc));
1006: PetscCall(MatGetLocalSize(pep->A[0],&mloc,&nloc));
1007: if (pjd->ld>1) {
1008: nloc += pjd->ld-1; mloc += pjd->ld-1;
1009: }
1010: PetscCall(PetscNew(&matctx));
1011: PetscCall(MatCreateShell(PetscObjectComm((PetscObject)pep),kspsf*nloc,kspsf*mloc,PETSC_DETERMINE,PETSC_DETERMINE,matctx,&pjd->Pshell));
1012: PetscCall(MatShellSetOperation(pjd->Pshell,MATOP_MULT,(void(*)(void))MatMult_PEPJD));
1013: PetscCall(MatShellSetOperation(pjd->Pshell,MATOP_CREATE_VECS,(void(*)(void))MatCreateVecs_PEPJD));
1014: matctx->pep = pep;
1015: target[0] = pep->target; target[1] = 0.0;
1016: PetscCall(PEPJDMatSetUp(pep,1,target));
1017: Pr = matctx->Pr;
1018: pcctx->PPr = NULL;
1019: #if !defined(PETSC_USE_COMPLEX)
1020: if (!pjd->reusepc) {
1021: PetscCall(MatDuplicate(matctx->Pr,MAT_COPY_VALUES,&pcctx->PPr));
1022: Pr = pcctx->PPr;
1023: }
1024: #endif
1025: PetscCall(PCSetOperators(pcctx->pc,Pr,Pr));
1026: PetscCall(PCSetErrorIfFailure(pcctx->pc,PETSC_TRUE));
1027: PetscCall(KSPSetPC(ksp,pjd->pcshell));
1028: if (pjd->reusepc) {
1029: PetscCall(PCSetReusePreconditioner(pcctx->pc,PETSC_TRUE));
1030: PetscCall(KSPSetReusePreconditioner(ksp,PETSC_TRUE));
1031: }
1032: PetscCall(PEP_KSPSetOperators(ksp,pjd->Pshell,pjd->Pshell));
1033: PetscCall(KSPSetUp(ksp));
1034: if (pjd->ld>1) {
1035: PetscCall(PetscMalloc2(pjd->ld*pjd->ld,&pcctx->M,pjd->ld*pjd->ld,&pcctx->ps));
1036: pcctx->pep = pep;
1037: }
1038: matctx->work = ww;
1039: pcctx->work = ww;
1040: PetscFunctionReturn(PETSC_SUCCESS);
1041: }
1043: static PetscErrorCode PEPJDEigenvectors(PEP pep)
1044: {
1045: PEP_JD *pjd = (PEP_JD*)pep->data;
1046: PetscBLASInt ld,nconv,info,nc;
1047: PetscScalar *Z;
1048: PetscReal *wr;
1049: Mat U;
1050: #if defined(PETSC_USE_COMPLEX)
1051: PetscScalar *w;
1052: #endif
1054: PetscFunctionBegin;
1055: PetscCall(PetscBLASIntCast(pep->ncv,&ld));
1056: PetscCall(PetscBLASIntCast(pep->nconv,&nconv));
1057: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
1058: #if !defined(PETSC_USE_COMPLEX)
1059: PetscCall(PetscMalloc2(pep->nconv*pep->nconv,&Z,3*pep->ncv,&wr));
1060: PetscCallBLAS("LAPACKtrevc",LAPACKtrevc_("R","A",NULL,&nconv,pjd->T,&ld,NULL,&nconv,Z,&nconv,&nconv,&nc,wr,&info));
1061: #else
1062: PetscCall(PetscMalloc3(pep->nconv*pep->nconv,&Z,3*pep->ncv,&wr,2*pep->ncv,&w));
1063: PetscCallBLAS("LAPACKtrevc",LAPACKtrevc_("R","A",NULL,&nconv,pjd->T,&ld,NULL,&nconv,Z,&nconv,&nconv,&nc,w,wr,&info));
1064: #endif
1065: PetscCall(PetscFPTrapPop());
1066: SlepcCheckLapackInfo("trevc",info);
1067: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nconv,nconv,Z,&U));
1068: PetscCall(BVSetActiveColumns(pjd->X,0,pep->nconv));
1069: PetscCall(BVMultInPlace(pjd->X,U,0,pep->nconv));
1070: PetscCall(BVNormalize(pjd->X,pep->eigi));
1071: PetscCall(MatDestroy(&U));
1072: #if !defined(PETSC_USE_COMPLEX)
1073: PetscCall(PetscFree2(Z,wr));
1074: #else
1075: PetscCall(PetscFree3(Z,wr,w));
1076: #endif
1077: PetscFunctionReturn(PETSC_SUCCESS);
1078: }
1080: static PetscErrorCode PEPJDLockConverged(PEP pep,PetscInt *nv,PetscInt sz)
1081: {
1082: PEP_JD *pjd = (PEP_JD*)pep->data;
1083: PetscInt j,i,*P,ldds,rk=0,nvv=*nv;
1084: Vec v,x,w;
1085: PetscScalar *R,*r,*pX,target[2];
1086: Mat X;
1087: PetscBLASInt sz_,rk_,nv_,info;
1088: PetscMPIInt np;
1090: PetscFunctionBegin;
1091: /* update AX and XpX */
1092: for (i=sz;i>0;i--) {
1093: PetscCall(BVGetColumn(pjd->X,pjd->nlock-i,&x));
1094: for (j=0;j<pep->nmat;j++) {
1095: PetscCall(BVGetColumn(pjd->AX[j],pjd->nlock-i,&v));
1096: PetscCall(MatMult(pep->A[j],x,v));
1097: PetscCall(BVRestoreColumn(pjd->AX[j],pjd->nlock-i,&v));
1098: PetscCall(BVSetActiveColumns(pjd->AX[j],0,pjd->nlock-i+1));
1099: }
1100: PetscCall(BVRestoreColumn(pjd->X,pjd->nlock-i,&x));
1101: PetscCall(BVDotColumn(pjd->X,(pjd->nlock-i),pjd->XpX+(pjd->nlock-i)*(pjd->ld)));
1102: pjd->XpX[(pjd->nlock-i)*(1+pjd->ld)] = 1.0;
1103: for (j=0;j<pjd->nlock-i;j++) pjd->XpX[j*(pjd->ld)+pjd->nlock-i] = PetscConj(pjd->XpX[(pjd->nlock-i)*(pjd->ld)+j]);
1104: }
1106: /* minimality index */
1107: pjd->midx = PetscMin(pjd->mmidx,pjd->nlock);
1109: /* evaluate the polynomial basis in T */
1110: PetscCall(PetscArrayzero(pjd->Tj,pjd->ld*pjd->ld*pep->nmat));
1111: for (j=0;j<pep->nmat;j++) PetscCall(PEPEvaluateBasisMat(pep,pjd->nlock,pjd->T,pep->ncv,j,(j>1)?pjd->Tj+(j-2)*pjd->ld*pjd->ld:NULL,pjd->ld,j?pjd->Tj+(j-1)*pjd->ld*pjd->ld:NULL,pjd->ld,pjd->Tj+j*pjd->ld*pjd->ld,pjd->ld));
1113: /* Extend search space */
1114: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pep),&np));
1115: PetscCall(PetscCalloc3(nvv,&P,nvv*nvv,&R,nvv*sz,&r));
1116: PetscCall(DSGetLeadingDimension(pep->ds,&ldds));
1117: PetscCall(DSGetArray(pep->ds,DS_MAT_X,&pX));
1118: PetscCall(PEPJDOrthogonalize(nvv,nvv,pX,ldds,&rk,P,R,nvv));
1119: for (j=0;j<sz;j++) {
1120: for (i=0;i<rk;i++) r[i*sz+j] = PetscConj(R[nvv*i+j]*pep->eigr[P[i]]); /* first row scaled with permuted diagonal */
1121: }
1122: PetscCall(PetscBLASIntCast(rk,&rk_));
1123: PetscCall(PetscBLASIntCast(sz,&sz_));
1124: PetscCall(PetscBLASIntCast(nvv,&nv_));
1125: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
1126: PetscCallBLAS("LAPACKtrtri",LAPACKtrtri_("U","N",&rk_,R,&nv_,&info));
1127: PetscCall(PetscFPTrapPop());
1128: SlepcCheckLapackInfo("trtri",info);
1129: for (i=0;i<sz;i++) PetscCallBLAS("BLAStrmv",BLAStrmv_("U","C","N",&rk_,R,&nv_,r+i,&sz_));
1130: for (i=0;i<sz*rk;i++) r[i] = PetscConj(r[i])/PetscSqrtReal(np); /* revert */
1131: PetscCall(BVSetActiveColumns(pjd->V,0,nvv));
1132: rk -= sz;
1133: for (j=0;j<rk;j++) PetscCall(PetscArraycpy(R+j*nvv,pX+(j+sz)*ldds,nvv));
1134: PetscCall(DSRestoreArray(pep->ds,DS_MAT_X,&pX));
1135: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,nvv,rk,R,&X));
1136: PetscCall(BVMultInPlace(pjd->V,X,0,rk));
1137: PetscCall(MatDestroy(&X));
1138: PetscCall(BVSetActiveColumns(pjd->V,0,rk));
1139: for (j=0;j<rk;j++) {
1140: PetscCall(BVGetColumn(pjd->V,j,&v));
1141: PetscCall(PEPJDCopyToExtendedVec(pep,NULL,r+sz*(j+sz),sz,pjd->nlock-sz,v,PETSC_FALSE));
1142: PetscCall(BVRestoreColumn(pjd->V,j,&v));
1143: }
1144: PetscCall(BVOrthogonalize(pjd->V,NULL));
1146: if (pjd->proj==PEP_JD_PROJECTION_HARMONIC) {
1147: for (j=0;j<rk;j++) {
1148: /* W = P(target)*V */
1149: PetscCall(BVGetColumn(pjd->W,j,&w));
1150: PetscCall(BVGetColumn(pjd->V,j,&v));
1151: target[0] = pep->target; target[1] = 0.0;
1152: PetscCall(PEPJDComputeResidual(pep,PETSC_FALSE,1,&v,target,&w,pep->work));
1153: PetscCall(BVRestoreColumn(pjd->V,j,&v));
1154: PetscCall(BVRestoreColumn(pjd->W,j,&w));
1155: }
1156: PetscCall(BVSetActiveColumns(pjd->W,0,rk));
1157: PetscCall(BVOrthogonalize(pjd->W,NULL));
1158: }
1159: *nv = rk;
1160: PetscCall(PetscFree3(P,R,r));
1161: PetscFunctionReturn(PETSC_SUCCESS);
1162: }
1164: PetscErrorCode PEPJDSystemSetUp(PEP pep,PetscInt sz,PetscScalar *theta,Vec *u,Vec *p,Vec *ww)
1165: {
1166: PEP_JD *pjd = (PEP_JD*)pep->data;
1167: PEP_JD_PCSHELL *pcctx;
1168: #if !defined(PETSC_USE_COMPLEX)
1169: PetscScalar s[2];
1170: #endif
1172: PetscFunctionBegin;
1173: PetscCall(PCShellGetContext(pjd->pcshell,&pcctx));
1174: PetscCall(PEPJDMatSetUp(pep,sz,theta));
1175: pcctx->u[0] = u[0]; pcctx->u[1] = u[1];
1176: /* Compute r'. p is a work space vector */
1177: PetscCall(PEPJDComputeResidual(pep,PETSC_TRUE,sz,u,theta,p,ww));
1178: PetscCall(PEPJDExtendedPCApply(pjd->pcshell,p[0],pcctx->Bp[0]));
1179: PetscCall(VecDot(pcctx->Bp[0],u[0],pcctx->gamma));
1180: #if !defined(PETSC_USE_COMPLEX)
1181: if (sz==2) {
1182: PetscCall(PEPJDExtendedPCApply(pjd->pcshell,p[1],pcctx->Bp[1]));
1183: PetscCall(VecDot(pcctx->Bp[0],u[1],pcctx->gamma+1));
1184: PetscCall(VecMDot(pcctx->Bp[1],2,u,s));
1185: pcctx->gamma[0] += s[1];
1186: pcctx->gamma[1] = -pcctx->gamma[1]+s[0];
1187: }
1188: #endif
1189: if (sz==1) {
1190: PetscCall(VecZeroEntries(pcctx->Bp[1]));
1191: pcctx->gamma[1] = 0.0;
1192: }
1193: PetscFunctionReturn(PETSC_SUCCESS);
1194: }
1196: PetscErrorCode PEPSolve_JD(PEP pep)
1197: {
1198: PEP_JD *pjd = (PEP_JD*)pep->data;
1199: PetscInt k,nv,nvc,ld,minv,dim,bupdated=0,sz=1,kspsf=1,idx,off,maxits,nloc;
1200: PetscMPIInt np,count;
1201: PetscScalar theta[2]={0.0,0.0},ritz[2]={0.0,0.0},*pX,*eig,*eigi,*array;
1202: PetscReal norm,*res,tol=0.0,rtol,abstol, dtol;
1203: PetscBool lindep,ini=PETSC_TRUE;
1204: Vec tc,t[2]={NULL,NULL},u[2]={NULL,NULL},p[2]={NULL,NULL};
1205: Vec rc,rr[2],r[2]={NULL,NULL},*ww=pep->work,v[2];
1206: Mat G,X,Y;
1207: KSP ksp;
1208: PEP_JD_PCSHELL *pcctx;
1209: PEP_JD_MATSHELL *matctx;
1210: #if !defined(PETSC_USE_COMPLEX)
1211: PetscReal norm1;
1212: #endif
1214: PetscFunctionBegin;
1215: PetscCall(PetscCitationsRegister(citation,&cited));
1216: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)pep),&np));
1217: PetscCall(BVGetSizes(pep->V,&nloc,NULL,NULL));
1218: PetscCall(DSGetLeadingDimension(pep->ds,&ld));
1219: PetscCall(PetscCalloc3(pep->ncv+pep->nev,&eig,pep->ncv+pep->nev,&eigi,pep->ncv+pep->nev,&res));
1220: pjd->nlock = 0;
1221: PetscCall(STGetKSP(pep->st,&ksp));
1222: PetscCall(KSPGetTolerances(ksp,&rtol,&abstol,&dtol,&maxits));
1223: #if !defined (PETSC_USE_COMPLEX)
1224: kspsf = 2;
1225: #endif
1226: PetscCall(PEPJDProcessInitialSpace(pep,ww));
1227: nv = (pep->nini)?pep->nini:1;
1229: /* Replace preconditioner with one containing projectors */
1230: PetscCall(PEPJDCreateShellPC(pep,ww));
1231: PetscCall(PCShellGetContext(pjd->pcshell,&pcctx));
1233: /* Create auxiliary vectors */
1234: PetscCall(BVCreateVec(pjd->V,&u[0]));
1235: PetscCall(VecDuplicate(u[0],&p[0]));
1236: PetscCall(VecDuplicate(u[0],&r[0]));
1237: #if !defined (PETSC_USE_COMPLEX)
1238: PetscCall(VecDuplicate(u[0],&u[1]));
1239: PetscCall(VecDuplicate(u[0],&p[1]));
1240: PetscCall(VecDuplicate(u[0],&r[1]));
1241: #endif
1243: /* Restart loop */
1244: while (pep->reason == PEP_CONVERGED_ITERATING) {
1245: pep->its++;
1246: PetscCall(DSSetDimensions(pep->ds,nv,0,0));
1247: PetscCall(BVSetActiveColumns(pjd->V,bupdated,nv));
1248: PetscCall(PEPJDUpdateTV(pep,bupdated,nv,ww));
1249: if (pjd->proj==PEP_JD_PROJECTION_HARMONIC) PetscCall(BVSetActiveColumns(pjd->W,bupdated,nv));
1250: for (k=0;k<pep->nmat;k++) {
1251: PetscCall(BVSetActiveColumns(pjd->TV[k],bupdated,nv));
1252: PetscCall(DSGetMat(pep->ds,DSMatExtra[k],&G));
1253: PetscCall(BVMatProject(pjd->TV[k],NULL,pjd->W,G));
1254: PetscCall(DSRestoreMat(pep->ds,DSMatExtra[k],&G));
1255: }
1256: PetscCall(BVSetActiveColumns(pjd->V,0,nv));
1257: PetscCall(BVSetActiveColumns(pjd->W,0,nv));
1259: /* Solve projected problem */
1260: PetscCall(DSSetState(pep->ds,DS_STATE_RAW));
1261: PetscCall(DSSolve(pep->ds,pep->eigr,pep->eigi));
1262: PetscCall(DSSort(pep->ds,pep->eigr,pep->eigi,NULL,NULL,NULL));
1263: PetscCall(DSSynchronize(pep->ds,pep->eigr,pep->eigi));
1264: idx = 0;
1265: do {
1266: ritz[0] = pep->eigr[idx];
1267: #if !defined(PETSC_USE_COMPLEX)
1268: ritz[1] = pep->eigi[idx];
1269: sz = (ritz[1]==0.0)?1:2;
1270: #endif
1271: /* Compute Ritz vector u=V*X(:,1) */
1272: PetscCall(DSGetArray(pep->ds,DS_MAT_X,&pX));
1273: PetscCall(BVSetActiveColumns(pjd->V,0,nv));
1274: PetscCall(BVMultVec(pjd->V,1.0,0.0,u[0],pX+idx*ld));
1275: #if !defined(PETSC_USE_COMPLEX)
1276: if (sz==2) PetscCall(BVMultVec(pjd->V,1.0,0.0,u[1],pX+(idx+1)*ld));
1277: #endif
1278: PetscCall(DSRestoreArray(pep->ds,DS_MAT_X,&pX));
1279: PetscCall(PEPJDComputeResidual(pep,PETSC_FALSE,sz,u,ritz,r,ww));
1280: /* Check convergence */
1281: PetscCall(VecNorm(r[0],NORM_2,&norm));
1282: #if !defined(PETSC_USE_COMPLEX)
1283: if (sz==2) {
1284: PetscCall(VecNorm(r[1],NORM_2,&norm1));
1285: norm = SlepcAbs(norm,norm1);
1286: }
1287: #endif
1288: PetscCall((*pep->converged)(pep,ritz[0],ritz[1],norm,&pep->errest[pep->nconv],pep->convergedctx));
1289: if (sz==2) pep->errest[pep->nconv+1] = pep->errest[pep->nconv];
1290: if (ini) {
1291: tol = PetscMin(.1,pep->errest[pep->nconv]); ini = PETSC_FALSE;
1292: } else tol = PetscMin(pep->errest[pep->nconv],tol/2);
1293: PetscCall((*pep->stopping)(pep,pep->its,pep->max_it,(pep->errest[pep->nconv]<pep->tol)?pep->nconv+sz:pep->nconv,pep->nev,&pep->reason,pep->stoppingctx));
1294: if (pep->errest[pep->nconv]<pep->tol) {
1295: /* Ritz pair converged */
1296: ini = PETSC_TRUE;
1297: minv = PetscMin(nv,(PetscInt)(pjd->keep*pep->ncv));
1298: if (pjd->ld>1) {
1299: PetscCall(BVGetColumn(pjd->X,pep->nconv,&v[0]));
1300: PetscCall(PEPJDCopyToExtendedVec(pep,v[0],pjd->T+pep->ncv*pep->nconv,pjd->ld-1,0,u[0],PETSC_TRUE));
1301: PetscCall(BVRestoreColumn(pjd->X,pep->nconv,&v[0]));
1302: PetscCall(BVSetActiveColumns(pjd->X,0,pep->nconv+1));
1303: PetscCall(BVNormColumn(pjd->X,pep->nconv,NORM_2,&norm));
1304: PetscCall(BVScaleColumn(pjd->X,pep->nconv,1.0/norm));
1305: for (k=0;k<pep->nconv;k++) pjd->T[pep->ncv*pep->nconv+k] *= PetscSqrtReal(np)/norm;
1306: pjd->T[(pep->ncv+1)*pep->nconv] = ritz[0];
1307: eig[pep->nconv] = ritz[0];
1308: idx++;
1309: #if !defined(PETSC_USE_COMPLEX)
1310: if (sz==2) {
1311: PetscCall(BVGetColumn(pjd->X,pep->nconv+1,&v[0]));
1312: PetscCall(PEPJDCopyToExtendedVec(pep,v[0],pjd->T+pep->ncv*(pep->nconv+1),pjd->ld-1,0,u[1],PETSC_TRUE));
1313: PetscCall(BVRestoreColumn(pjd->X,pep->nconv+1,&v[0]));
1314: PetscCall(BVSetActiveColumns(pjd->X,0,pep->nconv+2));
1315: PetscCall(BVNormColumn(pjd->X,pep->nconv+1,NORM_2,&norm1));
1316: PetscCall(BVScaleColumn(pjd->X,pep->nconv+1,1.0/norm1));
1317: for (k=0;k<pep->nconv;k++) pjd->T[pep->ncv*(pep->nconv+1)+k] *= PetscSqrtReal(np)/norm1;
1318: pjd->T[(pep->ncv+1)*(pep->nconv+1)] = ritz[0];
1319: pjd->T[(pep->ncv+1)*pep->nconv+1] = -ritz[1]*norm1/norm;
1320: pjd->T[(pep->ncv+1)*(pep->nconv+1)-1] = ritz[1]*norm/norm1;
1321: eig[pep->nconv+1] = ritz[0];
1322: eigi[pep->nconv] = ritz[1]; eigi[pep->nconv+1] = -ritz[1];
1323: idx++;
1324: }
1325: #endif
1326: } else PetscCall(BVInsertVec(pep->V,pep->nconv,u[0]));
1327: pep->nconv += sz;
1328: }
1329: } while (pep->errest[pep->nconv]<pep->tol && pep->nconv<nv);
1331: if (pep->reason==PEP_CONVERGED_ITERATING) {
1332: nvc = nv;
1333: if (idx) {
1334: pjd->nlock +=idx;
1335: PetscCall(PEPJDLockConverged(pep,&nv,idx));
1336: }
1337: if (nv+sz>=pep->ncv-1) {
1338: /* Basis full, force restart */
1339: minv = PetscMin(nv,(PetscInt)(pjd->keep*pep->ncv));
1340: PetscCall(DSGetDimensions(pep->ds,&dim,NULL,NULL,NULL));
1341: PetscCall(DSGetArray(pep->ds,DS_MAT_X,&pX));
1342: PetscCall(PEPJDOrthogonalize(dim,minv,pX,ld,&minv,NULL,NULL,ld));
1343: PetscCall(DSRestoreArray(pep->ds,DS_MAT_X,&pX));
1344: PetscCall(DSGetArray(pep->ds,DS_MAT_Y,&pX));
1345: PetscCall(PEPJDOrthogonalize(dim,minv,pX,ld,&minv,NULL,NULL,ld));
1346: PetscCall(DSRestoreArray(pep->ds,DS_MAT_Y,&pX));
1347: PetscCall(DSGetMat(pep->ds,DS_MAT_X,&X));
1348: PetscCall(BVMultInPlace(pjd->V,X,0,minv));
1349: PetscCall(DSRestoreMat(pep->ds,DS_MAT_X,&X));
1350: if (pjd->proj==PEP_JD_PROJECTION_HARMONIC) {
1351: PetscCall(DSGetMat(pep->ds,DS_MAT_Y,&Y));
1352: PetscCall(BVMultInPlace(pjd->W,Y,pep->nconv,minv));
1353: PetscCall(DSRestoreMat(pep->ds,DS_MAT_Y,&Y));
1354: }
1355: nv = minv;
1356: bupdated = 0;
1357: } else {
1358: if (!idx && pep->errest[pep->nconv]<pjd->fix) {theta[0] = ritz[0]; theta[1] = ritz[1];}
1359: else {theta[0] = pep->target; theta[1] = 0.0;}
1360: /* Update system mat */
1361: PetscCall(PEPJDSystemSetUp(pep,sz,theta,u,p,ww));
1362: /* Solve correction equation to expand basis */
1363: PetscCall(BVGetColumn(pjd->V,nv,&t[0]));
1364: rr[0] = r[0];
1365: if (sz==2) {
1366: PetscCall(BVGetColumn(pjd->V,nv+1,&t[1]));
1367: rr[1] = r[1];
1368: } else {
1369: t[1] = NULL;
1370: rr[1] = NULL;
1371: }
1372: PetscCall(VecCreateCompWithVecs(t,kspsf,pjd->vtempl,&tc));
1373: PetscCall(VecCreateCompWithVecs(rr,kspsf,pjd->vtempl,&rc));
1374: PetscCall(VecCompSetSubVecs(pjd->vtempl,sz,NULL));
1375: tol = PetscMax(rtol,tol/2);
1376: PetscCall(KSPSetTolerances(ksp,tol,abstol,dtol,maxits));
1377: PetscCall(KSPSolve(ksp,rc,tc));
1378: PetscCall(VecDestroy(&tc));
1379: PetscCall(VecDestroy(&rc));
1380: PetscCall(VecGetArray(t[0],&array));
1381: PetscCall(PetscMPIIntCast(pep->nconv,&count));
1382: PetscCallMPI(MPI_Bcast(array+nloc,count,MPIU_SCALAR,np-1,PetscObjectComm((PetscObject)pep)));
1383: PetscCall(VecRestoreArray(t[0],&array));
1384: PetscCall(BVRestoreColumn(pjd->V,nv,&t[0]));
1385: PetscCall(BVOrthogonalizeColumn(pjd->V,nv,NULL,&norm,&lindep));
1386: if (lindep || norm==0.0) {
1387: PetscCheck(sz!=1,PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"Linearly dependent continuation vector");
1388: off = 1;
1389: } else {
1390: off = 0;
1391: PetscCall(BVScaleColumn(pjd->V,nv,1.0/norm));
1392: }
1393: #if !defined(PETSC_USE_COMPLEX)
1394: if (sz==2) {
1395: PetscCall(VecGetArray(t[1],&array));
1396: PetscCallMPI(MPI_Bcast(array+nloc,count,MPIU_SCALAR,np-1,PetscObjectComm((PetscObject)pep)));
1397: PetscCall(VecRestoreArray(t[1],&array));
1398: PetscCall(BVRestoreColumn(pjd->V,nv+1,&t[1]));
1399: if (off) PetscCall(BVCopyColumn(pjd->V,nv+1,nv));
1400: PetscCall(BVOrthogonalizeColumn(pjd->V,nv+1-off,NULL,&norm,&lindep));
1401: if (lindep || norm==0.0) {
1402: PetscCheck(off==0,PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"Linearly dependent continuation vector");
1403: off = 1;
1404: } else PetscCall(BVScaleColumn(pjd->V,nv+1-off,1.0/norm));
1405: }
1406: #endif
1407: if (pjd->proj==PEP_JD_PROJECTION_HARMONIC) {
1408: PetscCall(BVInsertVec(pjd->W,nv,r[0]));
1409: if (sz==2 && !off) PetscCall(BVInsertVec(pjd->W,nv+1,r[1]));
1410: PetscCall(BVOrthogonalizeColumn(pjd->W,nv,NULL,&norm,&lindep));
1411: PetscCheck(!lindep && norm>0.0,PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"Linearly dependent continuation vector");
1412: PetscCall(BVScaleColumn(pjd->W,nv,1.0/norm));
1413: if (sz==2 && !off) {
1414: PetscCall(BVOrthogonalizeColumn(pjd->W,nv+1,NULL,&norm,&lindep));
1415: PetscCheck(!lindep && norm>0.0,PETSC_COMM_SELF,PETSC_ERR_CONV_FAILED,"Linearly dependent continuation vector");
1416: PetscCall(BVScaleColumn(pjd->W,nv+1,1.0/norm));
1417: }
1418: }
1419: bupdated = idx?0:nv;
1420: nv += sz-off;
1421: }
1422: for (k=0;k<nvc;k++) {
1423: eig[pep->nconv-idx+k] = pep->eigr[k];
1424: #if !defined(PETSC_USE_COMPLEX)
1425: eigi[pep->nconv-idx+k] = pep->eigi[k];
1426: #endif
1427: }
1428: PetscCall(PEPMonitor(pep,pep->its,pep->nconv,eig,eigi,pep->errest,pep->nconv+1));
1429: }
1430: }
1431: if (pjd->ld>1) {
1432: for (k=0;k<pep->nconv;k++) {
1433: pep->eigr[k] = eig[k];
1434: pep->eigi[k] = eigi[k];
1435: }
1436: if (pep->nconv>0) PetscCall(PEPJDEigenvectors(pep));
1437: PetscCall(PetscFree2(pcctx->M,pcctx->ps));
1438: }
1439: PetscCall(VecDestroy(&u[0]));
1440: PetscCall(VecDestroy(&r[0]));
1441: PetscCall(VecDestroy(&p[0]));
1442: #if !defined (PETSC_USE_COMPLEX)
1443: PetscCall(VecDestroy(&u[1]));
1444: PetscCall(VecDestroy(&r[1]));
1445: PetscCall(VecDestroy(&p[1]));
1446: #endif
1447: PetscCall(KSPSetTolerances(ksp,rtol,abstol,dtol,maxits));
1448: PetscCall(KSPSetPC(ksp,pcctx->pc));
1449: PetscCall(VecDestroy(&pcctx->Bp[0]));
1450: PetscCall(VecDestroy(&pcctx->Bp[1]));
1451: PetscCall(MatShellGetContext(pjd->Pshell,&matctx));
1452: PetscCall(MatDestroy(&matctx->Pr));
1453: PetscCall(MatDestroy(&matctx->Pi));
1454: PetscCall(MatDestroy(&pjd->Pshell));
1455: PetscCall(MatDestroy(&pcctx->PPr));
1456: PetscCall(PCDestroy(&pcctx->pc));
1457: PetscCall(PetscFree(pcctx));
1458: PetscCall(PetscFree(matctx));
1459: PetscCall(PCDestroy(&pjd->pcshell));
1460: PetscCall(PetscFree3(eig,eigi,res));
1461: PetscCall(VecDestroy(&pjd->vtempl));
1462: PetscFunctionReturn(PETSC_SUCCESS);
1463: }
1465: PetscErrorCode PEPJDSetRestart_JD(PEP pep,PetscReal keep)
1466: {
1467: PEP_JD *pjd = (PEP_JD*)pep->data;
1469: PetscFunctionBegin;
1470: if (keep==(PetscReal)PETSC_DEFAULT) pjd->keep = 0.5;
1471: else {
1472: PetscCheck(keep>=0.1 && keep<=0.9,PetscObjectComm((PetscObject)pep),PETSC_ERR_ARG_OUTOFRANGE,"The keep argument must be in the range [0.1,0.9]");
1473: pjd->keep = keep;
1474: }
1475: PetscFunctionReturn(PETSC_SUCCESS);
1476: }
1478: /*@
1479: PEPJDSetRestart - Sets the restart parameter for the Jacobi-Davidson
1480: method, in particular the proportion of basis vectors that must be kept
1481: after restart.
1483: Logically Collective
1485: Input Parameters:
1486: + pep - the eigenproblem solver context
1487: - keep - the number of vectors to be kept at restart
1489: Options Database Key:
1490: . -pep_jd_restart - Sets the restart parameter
1492: Notes:
1493: Allowed values are in the range [0.1,0.9]. The default is 0.5.
1495: Level: advanced
1497: .seealso: PEPJDGetRestart()
1498: @*/
1499: PetscErrorCode PEPJDSetRestart(PEP pep,PetscReal keep)
1500: {
1501: PetscFunctionBegin;
1504: PetscTryMethod(pep,"PEPJDSetRestart_C",(PEP,PetscReal),(pep,keep));
1505: PetscFunctionReturn(PETSC_SUCCESS);
1506: }
1508: PetscErrorCode PEPJDGetRestart_JD(PEP pep,PetscReal *keep)
1509: {
1510: PEP_JD *pjd = (PEP_JD*)pep->data;
1512: PetscFunctionBegin;
1513: *keep = pjd->keep;
1514: PetscFunctionReturn(PETSC_SUCCESS);
1515: }
1517: /*@
1518: PEPJDGetRestart - Gets the restart parameter used in the Jacobi-Davidson method.
1520: Not Collective
1522: Input Parameter:
1523: . pep - the eigenproblem solver context
1525: Output Parameter:
1526: . keep - the restart parameter
1528: Level: advanced
1530: .seealso: PEPJDSetRestart()
1531: @*/
1532: PetscErrorCode PEPJDGetRestart(PEP pep,PetscReal *keep)
1533: {
1534: PetscFunctionBegin;
1537: PetscUseMethod(pep,"PEPJDGetRestart_C",(PEP,PetscReal*),(pep,keep));
1538: PetscFunctionReturn(PETSC_SUCCESS);
1539: }
1541: PetscErrorCode PEPJDSetFix_JD(PEP pep,PetscReal fix)
1542: {
1543: PEP_JD *pjd = (PEP_JD*)pep->data;
1545: PetscFunctionBegin;
1546: if (fix == (PetscReal)PETSC_DEFAULT || fix == (PetscReal)PETSC_DECIDE) pjd->fix = 0.01;
1547: else {
1548: PetscCheck(fix>=0.0,PetscObjectComm((PetscObject)pep),PETSC_ERR_ARG_OUTOFRANGE,"Invalid fix value, must be >0");
1549: pjd->fix = fix;
1550: }
1551: PetscFunctionReturn(PETSC_SUCCESS);
1552: }
1554: /*@
1555: PEPJDSetFix - Sets the threshold for changing the target in the correction
1556: equation.
1558: Logically Collective
1560: Input Parameters:
1561: + pep - the eigenproblem solver context
1562: - fix - threshold for changing the target
1564: Options Database Key:
1565: . -pep_jd_fix - the fix value
1567: Note:
1568: The target in the correction equation is fixed at the first iterations.
1569: When the norm of the residual vector is lower than the fix value,
1570: the target is set to the corresponding eigenvalue.
1572: Level: advanced
1574: .seealso: PEPJDGetFix()
1575: @*/
1576: PetscErrorCode PEPJDSetFix(PEP pep,PetscReal fix)
1577: {
1578: PetscFunctionBegin;
1581: PetscTryMethod(pep,"PEPJDSetFix_C",(PEP,PetscReal),(pep,fix));
1582: PetscFunctionReturn(PETSC_SUCCESS);
1583: }
1585: PetscErrorCode PEPJDGetFix_JD(PEP pep,PetscReal *fix)
1586: {
1587: PEP_JD *pjd = (PEP_JD*)pep->data;
1589: PetscFunctionBegin;
1590: *fix = pjd->fix;
1591: PetscFunctionReturn(PETSC_SUCCESS);
1592: }
1594: /*@
1595: PEPJDGetFix - Returns the threshold for changing the target in the correction
1596: equation.
1598: Not Collective
1600: Input Parameter:
1601: . pep - the eigenproblem solver context
1603: Output Parameter:
1604: . fix - threshold for changing the target
1606: Note:
1607: The target in the correction equation is fixed at the first iterations.
1608: When the norm of the residual vector is lower than the fix value,
1609: the target is set to the corresponding eigenvalue.
1611: Level: advanced
1613: .seealso: PEPJDSetFix()
1614: @*/
1615: PetscErrorCode PEPJDGetFix(PEP pep,PetscReal *fix)
1616: {
1617: PetscFunctionBegin;
1620: PetscUseMethod(pep,"PEPJDGetFix_C",(PEP,PetscReal*),(pep,fix));
1621: PetscFunctionReturn(PETSC_SUCCESS);
1622: }
1624: PetscErrorCode PEPJDSetReusePreconditioner_JD(PEP pep,PetscBool reusepc)
1625: {
1626: PEP_JD *pjd = (PEP_JD*)pep->data;
1628: PetscFunctionBegin;
1629: pjd->reusepc = reusepc;
1630: PetscFunctionReturn(PETSC_SUCCESS);
1631: }
1633: /*@
1634: PEPJDSetReusePreconditioner - Sets a flag indicating whether the preconditioner
1635: must be reused or not.
1637: Logically Collective
1639: Input Parameters:
1640: + pep - the eigenproblem solver context
1641: - reusepc - the reuse flag
1643: Options Database Key:
1644: . -pep_jd_reuse_preconditioner - the reuse flag
1646: Note:
1647: The default value is False. If set to True, the preconditioner is built
1648: only at the beginning, using the target value. Otherwise, it may be rebuilt
1649: (depending on the fix parameter) at each iteration from the Ritz value.
1651: Level: advanced
1653: .seealso: PEPJDGetReusePreconditioner(), PEPJDSetFix()
1654: @*/
1655: PetscErrorCode PEPJDSetReusePreconditioner(PEP pep,PetscBool reusepc)
1656: {
1657: PetscFunctionBegin;
1660: PetscTryMethod(pep,"PEPJDSetReusePreconditioner_C",(PEP,PetscBool),(pep,reusepc));
1661: PetscFunctionReturn(PETSC_SUCCESS);
1662: }
1664: PetscErrorCode PEPJDGetReusePreconditioner_JD(PEP pep,PetscBool *reusepc)
1665: {
1666: PEP_JD *pjd = (PEP_JD*)pep->data;
1668: PetscFunctionBegin;
1669: *reusepc = pjd->reusepc;
1670: PetscFunctionReturn(PETSC_SUCCESS);
1671: }
1673: /*@
1674: PEPJDGetReusePreconditioner - Returns the flag for reusing the preconditioner.
1676: Not Collective
1678: Input Parameter:
1679: . pep - the eigenproblem solver context
1681: Output Parameter:
1682: . reusepc - the reuse flag
1684: Level: advanced
1686: .seealso: PEPJDSetReusePreconditioner()
1687: @*/
1688: PetscErrorCode PEPJDGetReusePreconditioner(PEP pep,PetscBool *reusepc)
1689: {
1690: PetscFunctionBegin;
1693: PetscUseMethod(pep,"PEPJDGetReusePreconditioner_C",(PEP,PetscBool*),(pep,reusepc));
1694: PetscFunctionReturn(PETSC_SUCCESS);
1695: }
1697: PetscErrorCode PEPJDSetMinimalityIndex_JD(PEP pep,PetscInt mmidx)
1698: {
1699: PEP_JD *pjd = (PEP_JD*)pep->data;
1701: PetscFunctionBegin;
1702: if (mmidx == PETSC_DEFAULT || mmidx == PETSC_DECIDE) {
1703: if (pjd->mmidx != 1) pep->state = PEP_STATE_INITIAL;
1704: pjd->mmidx = 1;
1705: } else {
1706: PetscCheck(mmidx>0,PetscObjectComm((PetscObject)pep),PETSC_ERR_ARG_OUTOFRANGE,"Invalid mmidx value, should be >0");
1707: if (pjd->mmidx != mmidx) pep->state = PEP_STATE_INITIAL;
1708: pjd->mmidx = mmidx;
1709: }
1710: PetscFunctionReturn(PETSC_SUCCESS);
1711: }
1713: /*@
1714: PEPJDSetMinimalityIndex - Sets the maximum allowed value for the minimality index.
1716: Logically Collective
1718: Input Parameters:
1719: + pep - the eigenproblem solver context
1720: - mmidx - maximum minimality index
1722: Options Database Key:
1723: . -pep_jd_minimality_index - the minimality index value
1725: Note:
1726: The default value is equal to the degree of the polynomial. A smaller value
1727: can be used if the wanted eigenvectors are known to be linearly independent.
1729: Level: advanced
1731: .seealso: PEPJDGetMinimalityIndex()
1732: @*/
1733: PetscErrorCode PEPJDSetMinimalityIndex(PEP pep,PetscInt mmidx)
1734: {
1735: PetscFunctionBegin;
1738: PetscTryMethod(pep,"PEPJDSetMinimalityIndex_C",(PEP,PetscInt),(pep,mmidx));
1739: PetscFunctionReturn(PETSC_SUCCESS);
1740: }
1742: PetscErrorCode PEPJDGetMinimalityIndex_JD(PEP pep,PetscInt *mmidx)
1743: {
1744: PEP_JD *pjd = (PEP_JD*)pep->data;
1746: PetscFunctionBegin;
1747: *mmidx = pjd->mmidx;
1748: PetscFunctionReturn(PETSC_SUCCESS);
1749: }
1751: /*@
1752: PEPJDGetMinimalityIndex - Returns the maximum allowed value of the minimality
1753: index.
1755: Not Collective
1757: Input Parameter:
1758: . pep - the eigenproblem solver context
1760: Output Parameter:
1761: . mmidx - minimality index
1763: Level: advanced
1765: .seealso: PEPJDSetMinimalityIndex()
1766: @*/
1767: PetscErrorCode PEPJDGetMinimalityIndex(PEP pep,PetscInt *mmidx)
1768: {
1769: PetscFunctionBegin;
1772: PetscUseMethod(pep,"PEPJDGetMinimalityIndex_C",(PEP,PetscInt*),(pep,mmidx));
1773: PetscFunctionReturn(PETSC_SUCCESS);
1774: }
1776: PetscErrorCode PEPJDSetProjection_JD(PEP pep,PEPJDProjection proj)
1777: {
1778: PEP_JD *pjd = (PEP_JD*)pep->data;
1780: PetscFunctionBegin;
1781: switch (proj) {
1782: case PEP_JD_PROJECTION_HARMONIC:
1783: case PEP_JD_PROJECTION_ORTHOGONAL:
1784: if (pjd->proj != proj) {
1785: pep->state = PEP_STATE_INITIAL;
1786: pjd->proj = proj;
1787: }
1788: break;
1789: default:
1790: SETERRQ(PetscObjectComm((PetscObject)pep),PETSC_ERR_ARG_OUTOFRANGE,"Invalid 'proj' value");
1791: }
1792: PetscFunctionReturn(PETSC_SUCCESS);
1793: }
1795: /*@
1796: PEPJDSetProjection - Sets the type of projection to be used in the Jacobi-Davidson solver.
1798: Logically Collective
1800: Input Parameters:
1801: + pep - the eigenproblem solver context
1802: - proj - the type of projection
1804: Options Database Key:
1805: . -pep_jd_projection - the projection type, either orthogonal or harmonic
1807: Level: advanced
1809: .seealso: PEPJDGetProjection()
1810: @*/
1811: PetscErrorCode PEPJDSetProjection(PEP pep,PEPJDProjection proj)
1812: {
1813: PetscFunctionBegin;
1816: PetscTryMethod(pep,"PEPJDSetProjection_C",(PEP,PEPJDProjection),(pep,proj));
1817: PetscFunctionReturn(PETSC_SUCCESS);
1818: }
1820: PetscErrorCode PEPJDGetProjection_JD(PEP pep,PEPJDProjection *proj)
1821: {
1822: PEP_JD *pjd = (PEP_JD*)pep->data;
1824: PetscFunctionBegin;
1825: *proj = pjd->proj;
1826: PetscFunctionReturn(PETSC_SUCCESS);
1827: }
1829: /*@
1830: PEPJDGetProjection - Returns the type of projection used by the Jacobi-Davidson solver.
1832: Not Collective
1834: Input Parameter:
1835: . pep - the eigenproblem solver context
1837: Output Parameter:
1838: . proj - the type of projection
1840: Level: advanced
1842: .seealso: PEPJDSetProjection()
1843: @*/
1844: PetscErrorCode PEPJDGetProjection(PEP pep,PEPJDProjection *proj)
1845: {
1846: PetscFunctionBegin;
1849: PetscUseMethod(pep,"PEPJDGetProjection_C",(PEP,PEPJDProjection*),(pep,proj));
1850: PetscFunctionReturn(PETSC_SUCCESS);
1851: }
1853: PetscErrorCode PEPSetFromOptions_JD(PEP pep,PetscOptionItems *PetscOptionsObject)
1854: {
1855: PetscBool flg,b1;
1856: PetscReal r1;
1857: PetscInt i1;
1858: PEPJDProjection proj;
1860: PetscFunctionBegin;
1861: PetscOptionsHeadBegin(PetscOptionsObject,"PEP JD Options");
1863: PetscCall(PetscOptionsReal("-pep_jd_restart","Proportion of vectors kept after restart","PEPJDSetRestart",0.5,&r1,&flg));
1864: if (flg) PetscCall(PEPJDSetRestart(pep,r1));
1866: PetscCall(PetscOptionsReal("-pep_jd_fix","Tolerance for changing the target in the correction equation","PEPJDSetFix",0.01,&r1,&flg));
1867: if (flg) PetscCall(PEPJDSetFix(pep,r1));
1869: PetscCall(PetscOptionsBool("-pep_jd_reuse_preconditioner","Whether to reuse the preconditioner","PEPJDSetReusePreconditoiner",PETSC_FALSE,&b1,&flg));
1870: if (flg) PetscCall(PEPJDSetReusePreconditioner(pep,b1));
1872: PetscCall(PetscOptionsInt("-pep_jd_minimality_index","Maximum allowed minimality index","PEPJDSetMinimalityIndex",1,&i1,&flg));
1873: if (flg) PetscCall(PEPJDSetMinimalityIndex(pep,i1));
1875: PetscCall(PetscOptionsEnum("-pep_jd_projection","Type of projection","PEPJDSetProjection",PEPJDProjectionTypes,(PetscEnum)PEP_JD_PROJECTION_HARMONIC,(PetscEnum*)&proj,&flg));
1876: if (flg) PetscCall(PEPJDSetProjection(pep,proj));
1878: PetscOptionsHeadEnd();
1879: PetscFunctionReturn(PETSC_SUCCESS);
1880: }
1882: PetscErrorCode PEPView_JD(PEP pep,PetscViewer viewer)
1883: {
1884: PEP_JD *pjd = (PEP_JD*)pep->data;
1885: PetscBool isascii;
1887: PetscFunctionBegin;
1888: PetscCall(PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii));
1889: if (isascii) {
1890: PetscCall(PetscViewerASCIIPrintf(viewer," %d%% of basis vectors kept after restart\n",(int)(100*pjd->keep)));
1891: PetscCall(PetscViewerASCIIPrintf(viewer," threshold for changing the target in the correction equation (fix): %g\n",(double)pjd->fix));
1892: PetscCall(PetscViewerASCIIPrintf(viewer," projection type: %s\n",PEPJDProjectionTypes[pjd->proj]));
1893: PetscCall(PetscViewerASCIIPrintf(viewer," maximum allowed minimality index: %" PetscInt_FMT "\n",pjd->mmidx));
1894: if (pjd->reusepc) PetscCall(PetscViewerASCIIPrintf(viewer," reusing the preconditioner\n"));
1895: }
1896: PetscFunctionReturn(PETSC_SUCCESS);
1897: }
1899: PetscErrorCode PEPSetDefaultST_JD(PEP pep)
1900: {
1901: KSP ksp;
1903: PetscFunctionBegin;
1904: if (!((PetscObject)pep->st)->type_name) {
1905: PetscCall(STSetType(pep->st,STPRECOND));
1906: PetscCall(STPrecondSetKSPHasMat(pep->st,PETSC_TRUE));
1907: }
1908: PetscCall(STSetTransform(pep->st,PETSC_FALSE));
1909: PetscCall(STGetKSP(pep->st,&ksp));
1910: if (!((PetscObject)ksp)->type_name) {
1911: PetscCall(KSPSetType(ksp,KSPBCGSL));
1912: PetscCall(KSPSetTolerances(ksp,1e-5,PETSC_DEFAULT,PETSC_DEFAULT,100));
1913: }
1914: PetscFunctionReturn(PETSC_SUCCESS);
1915: }
1917: PetscErrorCode PEPReset_JD(PEP pep)
1918: {
1919: PEP_JD *pjd = (PEP_JD*)pep->data;
1920: PetscInt i;
1922: PetscFunctionBegin;
1923: for (i=0;i<pep->nmat;i++) PetscCall(BVDestroy(pjd->TV+i));
1924: if (pjd->proj==PEP_JD_PROJECTION_HARMONIC) PetscCall(BVDestroy(&pjd->W));
1925: if (pjd->ld>1) {
1926: PetscCall(BVDestroy(&pjd->V));
1927: for (i=0;i<pep->nmat;i++) PetscCall(BVDestroy(pjd->AX+i));
1928: PetscCall(BVDestroy(&pjd->N[0]));
1929: PetscCall(BVDestroy(&pjd->N[1]));
1930: PetscCall(PetscFree3(pjd->XpX,pjd->T,pjd->Tj));
1931: }
1932: PetscCall(PetscFree2(pjd->TV,pjd->AX));
1933: PetscFunctionReturn(PETSC_SUCCESS);
1934: }
1936: PetscErrorCode PEPDestroy_JD(PEP pep)
1937: {
1938: PetscFunctionBegin;
1939: PetscCall(PetscFree(pep->data));
1940: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetRestart_C",NULL));
1941: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetRestart_C",NULL));
1942: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetFix_C",NULL));
1943: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetFix_C",NULL));
1944: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetReusePreconditioner_C",NULL));
1945: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetReusePreconditioner_C",NULL));
1946: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetMinimalityIndex_C",NULL));
1947: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetMinimalityIndex_C",NULL));
1948: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetProjection_C",NULL));
1949: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetProjection_C",NULL));
1950: PetscFunctionReturn(PETSC_SUCCESS);
1951: }
1953: SLEPC_EXTERN PetscErrorCode PEPCreate_JD(PEP pep)
1954: {
1955: PEP_JD *pjd;
1957: PetscFunctionBegin;
1958: PetscCall(PetscNew(&pjd));
1959: pep->data = (void*)pjd;
1961: pep->lineariz = PETSC_FALSE;
1962: pjd->fix = 0.01;
1963: pjd->mmidx = 0;
1965: pep->ops->solve = PEPSolve_JD;
1966: pep->ops->setup = PEPSetUp_JD;
1967: pep->ops->setfromoptions = PEPSetFromOptions_JD;
1968: pep->ops->destroy = PEPDestroy_JD;
1969: pep->ops->reset = PEPReset_JD;
1970: pep->ops->view = PEPView_JD;
1971: pep->ops->setdefaultst = PEPSetDefaultST_JD;
1973: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetRestart_C",PEPJDSetRestart_JD));
1974: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetRestart_C",PEPJDGetRestart_JD));
1975: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetFix_C",PEPJDSetFix_JD));
1976: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetFix_C",PEPJDGetFix_JD));
1977: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetReusePreconditioner_C",PEPJDSetReusePreconditioner_JD));
1978: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetReusePreconditioner_C",PEPJDGetReusePreconditioner_JD));
1979: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetMinimalityIndex_C",PEPJDSetMinimalityIndex_JD));
1980: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetMinimalityIndex_C",PEPJDGetMinimalityIndex_JD));
1981: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDSetProjection_C",PEPJDSetProjection_JD));
1982: PetscCall(PetscObjectComposeFunction((PetscObject)pep,"PEPJDGetProjection_C",PEPJDGetProjection_JD));
1983: PetscFunctionReturn(PETSC_SUCCESS);
1984: }