Line data Source code
1 : /*
2 : - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3 : SLEPc - Scalable Library for Eigenvalue Problem Computations
4 : Copyright (c) 2002-, Universitat Politecnica de Valencia, Spain
5 :
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: "davidson"
12 :
13 : Step: improve the eigenvectors X
14 : */
15 :
16 : #include "davidson.h"
17 : #include <slepcblaslapack.h>
18 :
19 : /**** JD update step (I - Kfg'/(g'Kf)) K(A - sB) (I - Kfg'/(g'Kf)) t = (I - Kfg'/(g'Kf))r ****/
20 :
21 : typedef struct {
22 : PetscInt size_X;
23 : KSP ksp; /* correction equation solver */
24 : Vec friends; /* reference vector for composite vectors */
25 : PetscScalar theta[4],thetai[2]; /* the shifts used in the correction eq. */
26 : PetscInt maxits; /* maximum number of iterations */
27 : PetscInt r_s,r_e; /* the selected eigenpairs to improve */
28 : PetscInt ksp_max_size; /* the ksp maximum subvectors size */
29 : PetscReal tol; /* the maximum solution tolerance */
30 : PetscReal lastTol; /* last tol for dynamic stopping criterion */
31 : PetscReal fix; /* tolerance for using the approx. eigenvalue */
32 : PetscBool dynamic; /* if the dynamic stopping criterion is applied */
33 : dvdDashboard *d; /* the current dvdDashboard reference */
34 : PC old_pc; /* old pc in ksp */
35 : BV KZ; /* KZ vecs for the projector KZ*inv(X'*KZ)*X' */
36 : BV U; /* new X vectors */
37 : PetscScalar *XKZ; /* X'*KZ */
38 : PetscScalar *iXKZ; /* inverse of XKZ */
39 : PetscInt ldXKZ; /* leading dimension of XKZ */
40 : PetscInt size_iXKZ; /* size of iXKZ */
41 : PetscInt ldiXKZ; /* leading dimension of iXKZ */
42 : PetscInt size_cX; /* last value of d->size_cX */
43 : PetscInt old_size_X; /* last number of improved vectors */
44 : PetscBLASInt *iXKZPivots; /* array of pivots */
45 : } dvdImprovex_jd;
46 :
47 : /*
48 : Compute (I - KZ*iXKZ*X')*V where,
49 : V, the vectors to apply the projector,
50 : cV, the number of vectors in V,
51 : */
52 66908 : static PetscErrorCode dvd_improvex_apply_proj(dvdDashboard *d,Vec *V,PetscInt cV)
53 : {
54 66908 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
55 66908 : PetscInt i,ldh,k,l;
56 66908 : PetscScalar *h;
57 66908 : PetscBLASInt cV_,n,info,ld;
58 : #if defined(PETSC_USE_COMPLEX)
59 66908 : PetscInt j;
60 : #endif
61 :
62 66908 : PetscFunctionBegin;
63 66908 : PetscAssert(cV<=2,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Consistency broken");
64 :
65 : /* h <- X'*V */
66 66908 : PetscCall(PetscMalloc1(data->size_iXKZ*cV,&h));
67 66908 : ldh = data->size_iXKZ;
68 66908 : PetscCall(BVGetActiveColumns(data->U,&l,&k));
69 66908 : PetscAssert(ldh==k,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Consistency broken");
70 66908 : PetscCall(BVSetActiveColumns(data->U,0,k));
71 133816 : for (i=0;i<cV;i++) {
72 66908 : PetscCall(BVDotVec(data->U,V[i],&h[ldh*i]));
73 : #if defined(PETSC_USE_COMPLEX)
74 133816 : for (j=0; j<k; j++) h[ldh*i+j] = PetscConj(h[ldh*i+j]);
75 : #endif
76 : }
77 66908 : PetscCall(BVSetActiveColumns(data->U,l,k));
78 :
79 : /* h <- iXKZ\h */
80 66908 : PetscCall(PetscBLASIntCast(cV,&cV_));
81 66908 : PetscCall(PetscBLASIntCast(data->size_iXKZ,&n));
82 66908 : PetscCall(PetscBLASIntCast(data->ldiXKZ,&ld));
83 66908 : PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
84 66908 : PetscCallBLAS("LAPACKgetrs",LAPACKgetrs_("N",&n,&cV_,data->iXKZ,&ld,data->iXKZPivots,h,&n,&info));
85 66908 : PetscCall(PetscFPTrapPop());
86 66908 : SlepcCheckLapackInfo("getrs",info);
87 :
88 : /* V <- V - KZ*h */
89 66908 : PetscCall(BVSetActiveColumns(data->KZ,0,k));
90 133816 : for (i=0;i<cV;i++) PetscCall(BVMultVec(data->KZ,-1.0,1.0,V[i],&h[ldh*i]));
91 66908 : PetscCall(BVSetActiveColumns(data->KZ,l,k));
92 66908 : PetscCall(PetscFree(h));
93 66908 : PetscFunctionReturn(PETSC_SUCCESS);
94 : }
95 :
96 : /*
97 : Compute (I - X*iXKZ*KZ')*V where,
98 : V, the vectors to apply the projector,
99 : cV, the number of vectors in V,
100 : */
101 1172 : static PetscErrorCode dvd_improvex_applytrans_proj(dvdDashboard *d,Vec *V,PetscInt cV)
102 : {
103 1172 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
104 1172 : PetscInt i,ldh,k,l;
105 1172 : PetscScalar *h;
106 1172 : PetscBLASInt cV_, n, info, ld;
107 : #if defined(PETSC_USE_COMPLEX)
108 1172 : PetscInt j;
109 : #endif
110 :
111 1172 : PetscFunctionBegin;
112 1172 : PetscAssert(cV<=2,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Consistency broken");
113 :
114 : /* h <- KZ'*V */
115 1172 : PetscCall(PetscMalloc1(data->size_iXKZ*cV,&h));
116 1172 : ldh = data->size_iXKZ;
117 1172 : PetscCall(BVGetActiveColumns(data->U,&l,&k));
118 1172 : PetscAssert(ldh==k,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Consistency broken");
119 1172 : PetscCall(BVSetActiveColumns(data->KZ,0,k));
120 2344 : for (i=0;i<cV;i++) {
121 1172 : PetscCall(BVDotVec(data->KZ,V[i],&h[ldh*i]));
122 : #if defined(PETSC_USE_COMPLEX)
123 2344 : for (j=0;j<k;j++) h[ldh*i+j] = PetscConj(h[ldh*i+j]);
124 : #endif
125 : }
126 1172 : PetscCall(BVSetActiveColumns(data->KZ,l,k));
127 :
128 : /* h <- iXKZ\h */
129 1172 : PetscCall(PetscBLASIntCast(cV,&cV_));
130 1172 : PetscCall(PetscBLASIntCast(data->size_iXKZ,&n));
131 1172 : PetscCall(PetscBLASIntCast(data->ldiXKZ,&ld));
132 1172 : PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
133 1172 : PetscCallBLAS("LAPACKgetrs",LAPACKgetrs_("C",&n,&cV_,data->iXKZ,&ld,data->iXKZPivots,h,&n,&info));
134 1172 : PetscCall(PetscFPTrapPop());
135 1172 : SlepcCheckLapackInfo("getrs",info);
136 :
137 : /* V <- V - U*h */
138 1172 : PetscCall(BVSetActiveColumns(data->U,0,k));
139 2344 : for (i=0;i<cV;i++) PetscCall(BVMultVec(data->U,-1.0,1.0,V[i],&h[ldh*i]));
140 1172 : PetscCall(BVSetActiveColumns(data->U,l,k));
141 1172 : PetscCall(PetscFree(h));
142 1172 : PetscFunctionReturn(PETSC_SUCCESS);
143 : }
144 :
145 80 : static PetscErrorCode dvd_improvex_jd_end(dvdDashboard *d)
146 : {
147 80 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
148 :
149 80 : PetscFunctionBegin;
150 80 : PetscCall(VecDestroy(&data->friends));
151 :
152 : /* Restore the pc of ksp */
153 80 : if (data->old_pc) {
154 18 : PetscCall(KSPSetPC(data->ksp, data->old_pc));
155 18 : PetscCall(PCDestroy(&data->old_pc));
156 : }
157 80 : PetscFunctionReturn(PETSC_SUCCESS);
158 : }
159 :
160 80 : static PetscErrorCode dvd_improvex_jd_d(dvdDashboard *d)
161 : {
162 80 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
163 :
164 80 : PetscFunctionBegin;
165 : /* Free local data and objects */
166 80 : PetscCall(PetscFree(data->XKZ));
167 80 : PetscCall(PetscFree(data->iXKZ));
168 80 : PetscCall(PetscFree(data->iXKZPivots));
169 80 : PetscCall(BVDestroy(&data->KZ));
170 80 : PetscCall(BVDestroy(&data->U));
171 80 : PetscCall(PetscFree(data));
172 80 : PetscFunctionReturn(PETSC_SUCCESS);
173 : }
174 :
175 : /*
176 : y <- theta[1]A*x - theta[0]*B*x
177 : auxV, two auxiliary vectors
178 : */
179 75642 : static inline PetscErrorCode dvd_aux_matmult(dvdImprovex_jd *data,const Vec *x,const Vec *y)
180 : {
181 75642 : PetscInt n,i;
182 75642 : const Vec *Bx;
183 75642 : Vec *auxV;
184 :
185 75642 : PetscFunctionBegin;
186 75642 : n = data->r_e - data->r_s;
187 151284 : for (i=0;i<n;i++) PetscCall(MatMult(data->d->A,x[i],y[i]));
188 :
189 75642 : PetscCall(SlepcVecPoolGetVecs(data->d->auxV,2,&auxV));
190 151284 : for (i=0;i<n;i++) {
191 : #if !defined(PETSC_USE_COMPLEX)
192 : if (PetscUnlikely(data->d->eigi[data->r_s+i] != 0.0)) {
193 : if (data->d->B) {
194 : PetscCall(MatMult(data->d->B,x[i],auxV[0]));
195 : PetscCall(MatMult(data->d->B,x[i+1],auxV[1]));
196 : Bx = auxV;
197 : } else Bx = &x[i];
198 :
199 : /* y_i <- [ t_2i+1*A*x_i - t_2i*Bx_i + ti_i*Bx_i+1;
200 : y_i+1 t_2i+1*A*x_i+1 - ti_i*Bx_i - t_2i*Bx_i+1 ] */
201 : PetscCall(VecAXPBYPCZ(y[i],-data->theta[2*i],data->thetai[i],data->theta[2*i+1],Bx[0],Bx[1]));
202 : PetscCall(VecAXPBYPCZ(y[i+1],-data->thetai[i],-data->theta[2*i],data->theta[2*i+1],Bx[0],Bx[1]));
203 : i++;
204 : } else
205 : #endif
206 : {
207 75642 : if (data->d->B) {
208 15636 : PetscCall(MatMult(data->d->B,x[i],auxV[0]));
209 15636 : Bx = auxV;
210 60006 : } else Bx = &x[i];
211 75642 : PetscCall(VecAXPBY(y[i],-data->theta[i*2],data->theta[i*2+1],Bx[0]));
212 : }
213 : }
214 75642 : PetscCall(SlepcVecPoolRestoreVecs(data->d->auxV,2,&auxV));
215 75642 : PetscFunctionReturn(PETSC_SUCCESS);
216 : }
217 :
218 : /*
219 : y <- theta[1]'*A'*x - theta[0]'*B'*x
220 : */
221 1123 : static inline PetscErrorCode dvd_aux_matmulttrans(dvdImprovex_jd *data,const Vec *x,const Vec *y)
222 : {
223 1123 : PetscInt n,i;
224 1123 : const Vec *Bx;
225 1123 : Vec *auxV;
226 :
227 1123 : PetscFunctionBegin;
228 1123 : n = data->r_e - data->r_s;
229 2246 : for (i=0;i<n;i++) PetscCall(MatMultTranspose(data->d->A,x[i],y[i]));
230 :
231 1123 : PetscCall(SlepcVecPoolGetVecs(data->d->auxV,2,&auxV));
232 2246 : for (i=0;i<n;i++) {
233 : #if !defined(PETSC_USE_COMPLEX)
234 : if (data->d->eigi[data->r_s+i] != 0.0) {
235 : if (data->d->B) {
236 : PetscCall(MatMultTranspose(data->d->B,x[i],auxV[0]));
237 : PetscCall(MatMultTranspose(data->d->B,x[i+1],auxV[1]));
238 : Bx = auxV;
239 : } else Bx = &x[i];
240 :
241 : /* y_i <- [ t_2i+1*A*x_i - t_2i*Bx_i - ti_i*Bx_i+1;
242 : y_i+1 t_2i+1*A*x_i+1 + ti_i*Bx_i - t_2i*Bx_i+1 ] */
243 : PetscCall(VecAXPBYPCZ(y[i],-data->theta[2*i],-data->thetai[i],data->theta[2*i+1],Bx[0],Bx[1]));
244 : PetscCall(VecAXPBYPCZ(y[i+1],data->thetai[i],-data->theta[2*i],data->theta[2*i+1],Bx[0],Bx[1]));
245 : i++;
246 : } else
247 : #endif
248 : {
249 1123 : if (data->d->B) {
250 0 : PetscCall(MatMultTranspose(data->d->B,x[i],auxV[0]));
251 0 : Bx = auxV;
252 1123 : } else Bx = &x[i];
253 1123 : PetscCall(VecAXPBY(y[i],PetscConj(-data->theta[i*2]),PetscConj(data->theta[i*2+1]),Bx[0]));
254 : }
255 : }
256 1123 : PetscCall(SlepcVecPoolRestoreVecs(data->d->auxV,2,&auxV));
257 1123 : PetscFunctionReturn(PETSC_SUCCESS);
258 : }
259 :
260 60656 : static PetscErrorCode PCApplyBA_dvd(PC pc,PCSide side,Vec in,Vec out,Vec w)
261 : {
262 60656 : dvdImprovex_jd *data;
263 60656 : PetscInt n,i;
264 60656 : const Vec *inx,*outx,*wx;
265 60656 : Vec *auxV;
266 60656 : Mat A;
267 :
268 60656 : PetscFunctionBegin;
269 60656 : PetscCall(PCGetOperators(pc,&A,NULL));
270 60656 : PetscCall(MatShellGetContext(A,&data));
271 60656 : PetscCall(VecCompGetSubVecs(in,NULL,&inx));
272 60656 : PetscCall(VecCompGetSubVecs(out,NULL,&outx));
273 60656 : PetscCall(VecCompGetSubVecs(w,NULL,&wx));
274 60656 : n = data->r_e - data->r_s;
275 60656 : PetscCall(SlepcVecPoolGetVecs(data->d->auxV,n,&auxV));
276 60656 : switch (side) {
277 60656 : case PC_LEFT:
278 : /* aux <- theta[1]A*in - theta[0]*B*in */
279 60656 : PetscCall(dvd_aux_matmult(data,inx,auxV));
280 :
281 : /* out <- K * aux */
282 121312 : for (i=0;i<n;i++) PetscCall(data->d->improvex_precond(data->d,data->r_s+i,auxV[i],outx[i]));
283 : break;
284 : case PC_RIGHT:
285 : /* aux <- K * in */
286 0 : for (i=0;i<n;i++) PetscCall(data->d->improvex_precond(data->d,data->r_s+i,inx[i],auxV[i]));
287 :
288 : /* out <- theta[1]A*auxV - theta[0]*B*auxV */
289 0 : PetscCall(dvd_aux_matmult(data,auxV,outx));
290 : break;
291 : case PC_SYMMETRIC:
292 : /* aux <- K^{1/2} * in */
293 0 : for (i=0;i<n;i++) PetscCall(PCApplySymmetricRight(data->old_pc,inx[i],auxV[i]));
294 :
295 : /* wx <- theta[1]A*auxV - theta[0]*B*auxV */
296 0 : PetscCall(dvd_aux_matmult(data,auxV,wx));
297 :
298 : /* aux <- K^{1/2} * in */
299 0 : for (i=0;i<n;i++) PetscCall(PCApplySymmetricLeft(data->old_pc,wx[i],outx[i]));
300 : break;
301 0 : default:
302 0 : SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unsupported KSP side");
303 : }
304 : /* out <- out - v*(u'*out) */
305 60656 : PetscCall(dvd_improvex_apply_proj(data->d,(Vec*)outx,n));
306 60656 : PetscCall(SlepcVecPoolRestoreVecs(data->d->auxV,n,&auxV));
307 60656 : PetscFunctionReturn(PETSC_SUCCESS);
308 : }
309 :
310 1680 : static PetscErrorCode PCApply_dvd(PC pc,Vec in,Vec out)
311 : {
312 1680 : dvdImprovex_jd *data;
313 1680 : PetscInt n,i;
314 1680 : const Vec *inx, *outx;
315 1680 : Mat A;
316 :
317 1680 : PetscFunctionBegin;
318 1680 : PetscCall(PCGetOperators(pc,&A,NULL));
319 1680 : PetscCall(MatShellGetContext(A,&data));
320 1680 : PetscCall(VecCompGetSubVecs(in,NULL,&inx));
321 1680 : PetscCall(VecCompGetSubVecs(out,NULL,&outx));
322 1680 : n = data->r_e - data->r_s;
323 : /* out <- K * in */
324 3360 : for (i=0;i<n;i++) PetscCall(data->d->improvex_precond(data->d,data->r_s+i,inx[i],outx[i]));
325 : /* out <- out - v*(u'*out) */
326 1680 : PetscCall(dvd_improvex_apply_proj(data->d,(Vec*)outx,n));
327 1680 : PetscFunctionReturn(PETSC_SUCCESS);
328 : }
329 :
330 1172 : static PetscErrorCode PCApplyTranspose_dvd(PC pc,Vec in,Vec out)
331 : {
332 1172 : dvdImprovex_jd *data;
333 1172 : PetscInt n,i;
334 1172 : const Vec *inx, *outx;
335 1172 : Vec *auxV;
336 1172 : Mat A;
337 :
338 1172 : PetscFunctionBegin;
339 1172 : PetscCall(PCGetOperators(pc,&A,NULL));
340 1172 : PetscCall(MatShellGetContext(A,&data));
341 1172 : PetscCall(VecCompGetSubVecs(in,NULL,&inx));
342 1172 : PetscCall(VecCompGetSubVecs(out,NULL,&outx));
343 1172 : n = data->r_e - data->r_s;
344 1172 : PetscCall(SlepcVecPoolGetVecs(data->d->auxV,n,&auxV));
345 : /* auxV <- in */
346 2344 : for (i=0;i<n;i++) PetscCall(VecCopy(inx[i],auxV[i]));
347 : /* auxV <- auxV - u*(v'*auxV) */
348 1172 : PetscCall(dvd_improvex_applytrans_proj(data->d,auxV,n));
349 : /* out <- K' * aux */
350 2344 : for (i=0;i<n;i++) PetscCall(PCApplyTranspose(data->old_pc,auxV[i],outx[i]));
351 1172 : PetscCall(SlepcVecPoolRestoreVecs(data->d->auxV,n,&auxV));
352 1172 : PetscFunctionReturn(PETSC_SUCCESS);
353 : }
354 :
355 14986 : static PetscErrorCode MatMult_dvd_jd(Mat A,Vec in,Vec out)
356 : {
357 14986 : dvdImprovex_jd *data;
358 14986 : PetscInt n;
359 14986 : const Vec *inx, *outx;
360 14986 : PCSide side;
361 :
362 14986 : PetscFunctionBegin;
363 14986 : PetscCall(MatShellGetContext(A,&data));
364 14986 : PetscCall(VecCompGetSubVecs(in,NULL,&inx));
365 14986 : PetscCall(VecCompGetSubVecs(out,NULL,&outx));
366 14986 : n = data->r_e - data->r_s;
367 : /* out <- theta[1]A*in - theta[0]*B*in */
368 14986 : PetscCall(dvd_aux_matmult(data,inx,outx));
369 14986 : PetscCall(KSPGetPCSide(data->ksp,&side));
370 14986 : if (side == PC_RIGHT) {
371 : /* out <- out - v*(u'*out) */
372 0 : PetscCall(dvd_improvex_apply_proj(data->d,(Vec*)outx,n));
373 : }
374 14986 : PetscFunctionReturn(PETSC_SUCCESS);
375 : }
376 :
377 1123 : static PetscErrorCode MatMultTranspose_dvd_jd(Mat A,Vec in,Vec out)
378 : {
379 1123 : dvdImprovex_jd *data;
380 1123 : PetscInt n,i;
381 1123 : const Vec *inx,*outx,*r;
382 1123 : Vec *auxV;
383 1123 : PCSide side;
384 :
385 1123 : PetscFunctionBegin;
386 1123 : PetscCall(MatShellGetContext(A,&data));
387 1123 : PetscCall(VecCompGetSubVecs(in,NULL,&inx));
388 1123 : PetscCall(VecCompGetSubVecs(out,NULL,&outx));
389 1123 : n = data->r_e - data->r_s;
390 1123 : PetscCall(KSPGetPCSide(data->ksp,&side));
391 1123 : if (side == PC_RIGHT) {
392 : /* auxV <- in */
393 0 : PetscCall(SlepcVecPoolGetVecs(data->d->auxV,n,&auxV));
394 0 : for (i=0;i<n;i++) PetscCall(VecCopy(inx[i],auxV[i]));
395 : /* auxV <- auxV - v*(u'*auxV) */
396 0 : PetscCall(dvd_improvex_applytrans_proj(data->d,auxV,n));
397 0 : r = auxV;
398 1123 : } else r = inx;
399 : /* out <- theta[1]A*r - theta[0]*B*r */
400 1123 : PetscCall(dvd_aux_matmulttrans(data,r,outx));
401 1123 : if (side == PC_RIGHT) PetscCall(SlepcVecPoolRestoreVecs(data->d->auxV,n,&auxV));
402 1123 : PetscFunctionReturn(PETSC_SUCCESS);
403 : }
404 :
405 24 : static PetscErrorCode MatCreateVecs_dvd_jd(Mat A,Vec *right,Vec *left)
406 : {
407 24 : Vec *r,*l;
408 24 : dvdImprovex_jd *data;
409 24 : PetscInt n,i;
410 :
411 24 : PetscFunctionBegin;
412 24 : PetscCall(MatShellGetContext(A,&data));
413 24 : n = data->ksp_max_size;
414 24 : if (right) PetscCall(PetscMalloc1(n,&r));
415 24 : if (left) PetscCall(PetscMalloc1(n,&l));
416 48 : for (i=0;i<n;i++) PetscCall(MatCreateVecs(data->d->A,right?&r[i]:NULL,left?&l[i]:NULL));
417 24 : if (right) {
418 24 : PetscCall(VecCreateCompWithVecs(r,n,data->friends,right));
419 48 : for (i=0;i<n;i++) PetscCall(VecDestroy(&r[i]));
420 : }
421 24 : if (left) {
422 0 : PetscCall(VecCreateCompWithVecs(l,n,data->friends,left));
423 0 : for (i=0;i<n;i++) PetscCall(VecDestroy(&l[i]));
424 : }
425 :
426 24 : if (right) PetscCall(PetscFree(r));
427 24 : if (left) PetscCall(PetscFree(l));
428 24 : PetscFunctionReturn(PETSC_SUCCESS);
429 : }
430 :
431 80 : static PetscErrorCode dvd_improvex_jd_start(dvdDashboard *d)
432 : {
433 80 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
434 80 : PetscInt rA, cA, rlA, clA;
435 80 : Mat A;
436 80 : PetscBool t;
437 80 : PC pc;
438 80 : Vec v0[2];
439 :
440 80 : PetscFunctionBegin;
441 80 : data->size_cX = data->old_size_X = 0;
442 80 : data->lastTol = data->dynamic?0.5:0.0;
443 :
444 : /* Setup the ksp */
445 80 : if (data->ksp) {
446 : /* Create the reference vector */
447 32 : PetscCall(BVGetColumn(d->eps->V,0,&v0[0]));
448 32 : v0[1] = v0[0];
449 32 : PetscCall(VecCreateCompWithVecs(v0,data->ksp_max_size,NULL,&data->friends));
450 32 : PetscCall(BVRestoreColumn(d->eps->V,0,&v0[0]));
451 :
452 : /* Save the current pc and set a PCNONE */
453 32 : PetscCall(KSPGetPC(data->ksp, &data->old_pc));
454 32 : PetscCall(PetscObjectTypeCompare((PetscObject)data->old_pc,PCNONE,&t));
455 32 : data->lastTol = 0.5;
456 32 : if (t) data->old_pc = NULL;
457 : else {
458 18 : PetscCall(PetscObjectReference((PetscObject)data->old_pc));
459 18 : PetscCall(PCCreate(PetscObjectComm((PetscObject)d->eps),&pc));
460 18 : PetscCall(PCSetType(pc,PCSHELL));
461 18 : PetscCall(PCSetOperators(pc,d->A,d->A));
462 18 : PetscCall(PCSetReusePreconditioner(pc,PETSC_TRUE));
463 18 : PetscCall(PCShellSetApply(pc,PCApply_dvd));
464 18 : PetscCall(PCShellSetApplyBA(pc,PCApplyBA_dvd));
465 18 : PetscCall(PCShellSetApplyTranspose(pc,PCApplyTranspose_dvd));
466 18 : PetscCall(KSPSetPC(data->ksp,pc));
467 18 : PetscCall(PCDestroy(&pc));
468 : }
469 :
470 : /* Create the (I-v*u')*K*(A-s*B) matrix */
471 32 : PetscCall(MatGetSize(d->A,&rA,&cA));
472 32 : PetscCall(MatGetLocalSize(d->A,&rlA,&clA));
473 32 : PetscCall(MatCreateShell(PetscObjectComm((PetscObject)d->A),rlA*data->ksp_max_size,clA*data->ksp_max_size,rA*data->ksp_max_size,cA*data->ksp_max_size,data,&A));
474 32 : PetscCall(MatShellSetOperation(A,MATOP_MULT,(void(*)(void))MatMult_dvd_jd));
475 32 : PetscCall(MatShellSetOperation(A,MATOP_MULT_TRANSPOSE,(void(*)(void))MatMultTranspose_dvd_jd));
476 32 : PetscCall(MatShellSetOperation(A,MATOP_CREATE_VECS,(void(*)(void))MatCreateVecs_dvd_jd));
477 :
478 : /* Try to avoid KSPReset */
479 32 : PetscCall(KSPGetOperatorsSet(data->ksp,&t,NULL));
480 32 : if (t) {
481 24 : Mat M;
482 24 : PetscInt rM;
483 24 : PetscCall(KSPGetOperators(data->ksp,&M,NULL));
484 24 : PetscCall(MatGetSize(M,&rM,NULL));
485 24 : if (rM != rA*data->ksp_max_size) PetscCall(KSPReset(data->ksp));
486 : }
487 32 : PetscCall(EPS_KSPSetOperators(data->ksp,A,A));
488 32 : PetscCall(KSPSetReusePreconditioner(data->ksp,PETSC_TRUE));
489 32 : PetscCall(KSPSetUp(data->ksp));
490 32 : PetscCall(MatDestroy(&A));
491 : } else {
492 48 : data->old_pc = NULL;
493 48 : data->friends = NULL;
494 : }
495 80 : PetscCall(BVSetActiveColumns(data->KZ,0,0));
496 80 : PetscCall(BVSetActiveColumns(data->U,0,0));
497 80 : PetscFunctionReturn(PETSC_SUCCESS);
498 : }
499 :
500 : /*
501 : Compute: u <- X, v <- K*(theta[0]*A+theta[1]*B)*X,
502 : kr <- K^{-1}*(A-eig*B)*X, being X <- V*pX[i_s..i_e-1], Y <- W*pY[i_s..i_e-1]
503 : where
504 : pX,pY, the right and left eigenvectors of the projected system
505 : ld, the leading dimension of pX and pY
506 : */
507 5607 : static PetscErrorCode dvd_improvex_jd_proj_cuv(dvdDashboard *d,PetscInt i_s,PetscInt i_e,Vec *kr,PetscScalar *theta,PetscScalar *thetai,PetscScalar *pX,PetscScalar *pY,PetscInt ld)
508 : {
509 5607 : PetscInt n=i_e-i_s,size_KZ,V_new,rm,i,lv,kv,lKZ,kKZ;
510 5607 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
511 5607 : const PetscScalar *array;
512 5607 : Mat M;
513 5607 : Vec u[2],v[2];
514 5607 : PetscBLASInt s,ldXKZ,info;
515 :
516 5607 : PetscFunctionBegin;
517 : /* Check consistency */
518 5607 : PetscCall(BVGetActiveColumns(d->eps->V,&lv,&kv));
519 5607 : V_new = lv - data->size_cX;
520 5607 : PetscAssert(V_new<=data->old_size_X,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Consistency broken");
521 5607 : data->old_size_X = n;
522 5607 : data->size_cX = lv;
523 :
524 : /* KZ <- KZ(rm:rm+max_cX-1) */
525 5607 : PetscCall(BVGetActiveColumns(data->KZ,&lKZ,&kKZ));
526 5607 : rm = PetscMax(V_new+lKZ,0);
527 5607 : if (rm > 0) {
528 183 : for (i=0;i<lKZ;i++) {
529 0 : PetscCall(BVCopyColumn(data->KZ,i+rm,i));
530 0 : PetscCall(BVCopyColumn(data->U,i+rm,i));
531 : }
532 : }
533 :
534 : /* XKZ <- XKZ(rm:rm+max_cX-1,rm:rm+max_cX-1) */
535 : if (rm > 0) {
536 183 : for (i=0;i<lKZ;i++) PetscCall(PetscArraycpy(&data->XKZ[i*data->ldXKZ+i],&data->XKZ[(i+rm)*data->ldXKZ+i+rm],lKZ));
537 : }
538 5607 : lKZ = PetscMin(0,lKZ+V_new);
539 5607 : PetscCall(BVSetActiveColumns(data->KZ,lKZ,lKZ+n));
540 5607 : PetscCall(BVSetActiveColumns(data->U,lKZ,lKZ+n));
541 :
542 : /* Compute X, KZ and KR */
543 5607 : PetscCall(BVGetColumn(data->U,lKZ,u));
544 5607 : if (n>1) PetscCall(BVGetColumn(data->U,lKZ+1,&u[1]));
545 5607 : PetscCall(BVGetColumn(data->KZ,lKZ,v));
546 5607 : if (n>1) PetscCall(BVGetColumn(data->KZ,lKZ+1,&v[1]));
547 5607 : PetscCall(d->improvex_jd_proj_uv(d,i_s,i_e,u,v,kr,theta,thetai,pX,pY,ld));
548 5607 : PetscCall(BVRestoreColumn(data->U,lKZ,u));
549 5607 : if (n>1) PetscCall(BVRestoreColumn(data->U,lKZ+1,&u[1]));
550 5607 : PetscCall(BVRestoreColumn(data->KZ,lKZ,v));
551 5607 : if (n>1) PetscCall(BVRestoreColumn(data->KZ,lKZ+1,&v[1]));
552 :
553 : /* XKZ <- U'*KZ */
554 5607 : PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,lKZ+n,lKZ+n,NULL,&M));
555 5607 : PetscCall(BVMatProject(data->KZ,NULL,data->U,M));
556 5607 : PetscCall(MatDenseGetArrayRead(M,&array));
557 11214 : for (i=lKZ;i<lKZ+n;i++) { /* upper part */
558 5607 : PetscCall(PetscArraycpy(&data->XKZ[data->ldXKZ*i],&array[i*(lKZ+n)],lKZ));
559 : }
560 11214 : for (i=0;i<lKZ+n;i++) { /* lower part */
561 5607 : PetscCall(PetscArraycpy(&data->XKZ[data->ldXKZ*i+lKZ],&array[i*(lKZ+n)+lKZ],n));
562 : }
563 5607 : PetscCall(MatDenseRestoreArrayRead(M,&array));
564 5607 : PetscCall(MatDestroy(&M));
565 :
566 : /* iXKZ <- inv(XKZ) */
567 5607 : size_KZ = lKZ+n;
568 5607 : PetscCall(PetscBLASIntCast(lKZ+n,&s));
569 5607 : data->ldiXKZ = data->size_iXKZ = size_KZ;
570 11214 : for (i=0;i<size_KZ;i++) PetscCall(PetscArraycpy(&data->iXKZ[data->ldiXKZ*i],&data->XKZ[data->ldXKZ*i],size_KZ));
571 5607 : PetscCall(PetscBLASIntCast(data->ldiXKZ,&ldXKZ));
572 5607 : PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
573 5607 : PetscCallBLAS("LAPACKgetrf",LAPACKgetrf_(&s,&s,data->iXKZ,&ldXKZ,data->iXKZPivots,&info));
574 5607 : PetscCall(PetscFPTrapPop());
575 5607 : SlepcCheckLapackInfo("getrf",info);
576 5607 : PetscFunctionReturn(PETSC_SUCCESS);
577 : }
578 :
579 5206 : static PetscErrorCode dvd_improvex_jd_gen(dvdDashboard *d,PetscInt r_s,PetscInt r_e,PetscInt *size_D)
580 : {
581 5206 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
582 5206 : PetscInt i,j,n,maxits,maxits0,lits,s,ld,k,max_size_D,lV,kV;
583 5206 : PetscScalar *pX,*pY;
584 5206 : PetscReal tol,tol0;
585 5206 : Vec *kr,kr_comp,D_comp,D[2],kr0[2];
586 5206 : PetscBool odd_situation = PETSC_FALSE;
587 :
588 5206 : PetscFunctionBegin;
589 5206 : PetscCall(BVGetActiveColumns(d->eps->V,&lV,&kV));
590 5206 : max_size_D = d->eps->ncv-kV;
591 : /* Quick exit */
592 5206 : if ((max_size_D == 0) || r_e-r_s <= 0) {
593 0 : *size_D = 0;
594 0 : PetscFunctionReturn(PETSC_SUCCESS);
595 : }
596 :
597 5206 : n = PetscMin(PetscMin(data->size_X, max_size_D), r_e-r_s);
598 5206 : PetscAssert(n>0,PETSC_COMM_SELF,PETSC_ERR_PLIB,"n == 0");
599 5206 : PetscAssert(data->size_X>=r_e-r_s,PETSC_COMM_SELF,PETSC_ERR_PLIB,"size_X < r_e-r_s");
600 :
601 5206 : PetscCall(DSGetLeadingDimension(d->eps->ds,&ld));
602 :
603 : /* Restart lastTol if a new pair converged */
604 5206 : if (data->dynamic && data->size_cX < lV)
605 1 : data->lastTol = 0.5;
606 :
607 10550 : for (i=0,s=0;i<n;i+=s) {
608 : /* If the selected eigenvalue is complex, but the arithmetic is real... */
609 : #if !defined(PETSC_USE_COMPLEX)
610 : if (d->eigi[r_s+i] != 0.0) {
611 : if (i+2 <= max_size_D) s=2;
612 : else break;
613 : } else
614 : #endif
615 5607 : s=1;
616 :
617 5607 : data->r_s = r_s+i;
618 5607 : data->r_e = r_s+i+s;
619 5607 : PetscCall(SlepcVecPoolGetVecs(d->auxV,s,&kr));
620 :
621 : /* Compute theta, maximum iterations and tolerance */
622 : maxits = 0;
623 : tol = 1;
624 11214 : for (j=0;j<s;j++) {
625 5607 : PetscCall(d->improvex_jd_lit(d,r_s+i+j,&data->theta[2*j],&data->thetai[j],&maxits0,&tol0));
626 5607 : maxits += maxits0;
627 5607 : tol *= tol0;
628 : }
629 5607 : maxits/= s;
630 5607 : tol = data->dynamic?data->lastTol:PetscExpReal(PetscLogReal(tol)/s);
631 :
632 : /* Compute u, v and kr */
633 5607 : k = r_s+i;
634 5607 : PetscCall(DSVectors(d->eps->ds,DS_MAT_X,&k,NULL));
635 5607 : k = r_s+i;
636 5607 : PetscCall(DSVectors(d->eps->ds,DS_MAT_Y,&k,NULL));
637 5607 : PetscCall(DSGetArray(d->eps->ds,DS_MAT_X,&pX));
638 5607 : PetscCall(DSGetArray(d->eps->ds,DS_MAT_Y,&pY));
639 5607 : PetscCall(dvd_improvex_jd_proj_cuv(d,r_s+i,r_s+i+s,kr,data->theta,data->thetai,pX,pY,ld));
640 5607 : PetscCall(DSRestoreArray(d->eps->ds,DS_MAT_X,&pX));
641 5607 : PetscCall(DSRestoreArray(d->eps->ds,DS_MAT_Y,&pY));
642 :
643 : /* Check if the first eigenpairs are converged */
644 5607 : if (i == 0) {
645 5206 : PetscInt oldnpreconv = d->npreconv;
646 5206 : PetscCall(d->preTestConv(d,0,r_s+s,r_s+s,&d->npreconv));
647 5206 : if (d->npreconv > oldnpreconv) break;
648 : }
649 :
650 : /* Test the odd situation of solving Ax=b with A=I */
651 : #if !defined(PETSC_USE_COMPLEX)
652 : odd_situation = (data->ksp && data->theta[0] == 1. && data->theta[1] == 0. && data->thetai[0] == 0. && d->B == NULL)? PETSC_TRUE: PETSC_FALSE;
653 : #else
654 5344 : odd_situation = (data->ksp && data->theta[0] == 1. && data->theta[1] == 0. && d->B == NULL)? PETSC_TRUE: PETSC_FALSE;
655 : #endif
656 : /* If JD */
657 5344 : if (data->ksp && !odd_situation) {
658 : /* kr <- -kr */
659 1544 : for (j=0;j<s;j++) PetscCall(VecScale(kr[j],-1.0));
660 :
661 : /* Compose kr and D */
662 772 : kr0[0] = kr[0];
663 772 : kr0[1] = (s==2 ? kr[1] : NULL);
664 772 : PetscCall(VecCreateCompWithVecs(kr0,data->ksp_max_size,data->friends,&kr_comp));
665 772 : PetscCall(BVGetColumn(d->eps->V,kV+i,&D[0]));
666 772 : if (s==2) PetscCall(BVGetColumn(d->eps->V,kV+i+1,&D[1]));
667 772 : else D[1] = NULL;
668 772 : PetscCall(VecCreateCompWithVecs(D,data->ksp_max_size,data->friends,&D_comp));
669 772 : PetscCall(VecCompSetSubVecs(data->friends,s,NULL));
670 :
671 : /* Solve the correction equation */
672 772 : PetscCall(KSPSetTolerances(data->ksp,tol,PETSC_DEFAULT,PETSC_DEFAULT,maxits));
673 772 : PetscCall(KSPSolve(data->ksp,kr_comp,D_comp));
674 772 : PetscCall(KSPGetIterationNumber(data->ksp,&lits));
675 :
676 : /* Destroy the composed ks and D */
677 772 : PetscCall(VecDestroy(&kr_comp));
678 772 : PetscCall(VecDestroy(&D_comp));
679 772 : PetscCall(BVRestoreColumn(d->eps->V,kV+i,&D[0]));
680 : if (s==2) PetscCall(BVRestoreColumn(d->eps->V,kV+i+1,&D[1]));
681 :
682 : /* If GD */
683 : } else {
684 4572 : PetscCall(BVGetColumn(d->eps->V,kV+i,&D[0]));
685 : if (s==2) PetscCall(BVGetColumn(d->eps->V,kV+i+1,&D[1]));
686 9144 : for (j=0;j<s;j++) PetscCall(d->improvex_precond(d,r_s+i+j,kr[j],D[j]));
687 4572 : PetscCall(dvd_improvex_apply_proj(d,D,s));
688 4572 : PetscCall(BVRestoreColumn(d->eps->V,kV+i,&D[0]));
689 5344 : if (s==2) PetscCall(BVRestoreColumn(d->eps->V,kV+i+1,&D[1]));
690 : }
691 : /* Prevent that short vectors are discarded in the orthogonalization */
692 5344 : if (i == 0 && d->eps->errest[d->nconv+r_s] > PETSC_MACHINE_EPSILON && d->eps->errest[d->nconv+r_s] < PETSC_MAX_REAL) {
693 9886 : for (j=0;j<s;j++) PetscCall(BVScaleColumn(d->eps->V,kV+i+j,1.0/d->eps->errest[d->nconv+r_s]));
694 : }
695 5344 : PetscCall(SlepcVecPoolRestoreVecs(d->auxV,s,&kr));
696 : }
697 5206 : *size_D = i;
698 5300 : if (data->dynamic) data->lastTol = PetscMax(data->lastTol/2.0,PETSC_MACHINE_EPSILON*10.0);
699 5206 : PetscFunctionReturn(PETSC_SUCCESS);
700 : }
701 :
702 240 : PetscErrorCode dvd_improvex_jd(dvdDashboard *d,dvdBlackboard *b,KSP ksp,PetscInt max_bs,PetscBool dynamic)
703 : {
704 240 : dvdImprovex_jd *data;
705 240 : PetscBool useGD;
706 240 : PC pc;
707 240 : PetscInt size_P;
708 :
709 240 : PetscFunctionBegin;
710 : /* Setting configuration constrains */
711 240 : PetscCall(PetscObjectTypeCompare((PetscObject)ksp,KSPPREONLY,&useGD));
712 :
713 : /* If the arithmetic is real and the problem is not Hermitian, then
714 : the block size is incremented in one */
715 : #if !defined(PETSC_USE_COMPLEX)
716 : if (!DVD_IS(d->sEP,DVD_EP_HERMITIAN)) {
717 : max_bs++;
718 : b->max_size_P = PetscMax(b->max_size_P,2);
719 : } else
720 : #endif
721 : {
722 240 : b->max_size_P = PetscMax(b->max_size_P,1);
723 : }
724 240 : b->max_size_X = PetscMax(b->max_size_X,max_bs);
725 240 : size_P = b->max_size_P;
726 :
727 : /* Setup the preconditioner */
728 240 : if (ksp) {
729 240 : PetscCall(KSPGetPC(ksp,&pc));
730 240 : PetscCall(dvd_static_precond_PC(d,b,pc));
731 0 : } else PetscCall(dvd_static_precond_PC(d,b,NULL));
732 :
733 : /* Setup the step */
734 240 : if (b->state >= DVD_STATE_CONF) {
735 80 : PetscCall(PetscNew(&data));
736 80 : data->dynamic = dynamic;
737 80 : PetscCall(PetscMalloc1(size_P*size_P,&data->XKZ));
738 80 : PetscCall(PetscMalloc1(size_P*size_P,&data->iXKZ));
739 80 : PetscCall(PetscMalloc1(size_P,&data->iXKZPivots));
740 80 : data->ldXKZ = size_P;
741 80 : data->size_X = b->max_size_X;
742 80 : d->improveX_data = data;
743 80 : data->ksp = useGD? NULL: ksp;
744 80 : data->d = d;
745 80 : d->improveX = dvd_improvex_jd_gen;
746 : #if !defined(PETSC_USE_COMPLEX)
747 : if (!DVD_IS(d->sEP,DVD_EP_HERMITIAN)) data->ksp_max_size = 2;
748 : else
749 : #endif
750 80 : data->ksp_max_size = 1;
751 : /* Create various vector basis */
752 80 : PetscCall(BVDuplicateResize(d->eps->V,size_P,&data->KZ));
753 80 : PetscCall(BVSetMatrix(data->KZ,NULL,PETSC_FALSE));
754 80 : PetscCall(BVDuplicate(data->KZ,&data->U));
755 :
756 80 : PetscCall(EPSDavidsonFLAdd(&d->startList,dvd_improvex_jd_start));
757 80 : PetscCall(EPSDavidsonFLAdd(&d->endList,dvd_improvex_jd_end));
758 80 : PetscCall(EPSDavidsonFLAdd(&d->destroyList,dvd_improvex_jd_d));
759 : }
760 240 : PetscFunctionReturn(PETSC_SUCCESS);
761 : }
762 :
763 : #if !defined(PETSC_USE_COMPLEX)
764 : static inline PetscErrorCode dvd_complex_rayleigh_quotient(Vec ur,Vec ui,Vec Axr,Vec Axi,Vec Bxr,Vec Bxi,PetscScalar *eigr,PetscScalar *eigi)
765 : {
766 : PetscScalar rAr,iAr,rAi,iAi,rBr,iBr,rBi,iBi,b0,b2,b4,b6,b7;
767 :
768 : PetscFunctionBegin;
769 : /* eigr = [(rAr+iAi)*(rBr+iBi) + (rAi-iAr)*(rBi-iBr)]/k
770 : eigi = [(rAi-iAr)*(rBr+iBi) - (rAr+iAi)*(rBi-iBr)]/k
771 : k = (rBr+iBi)*(rBr+iBi) + (rBi-iBr)*(rBi-iBr) */
772 : PetscCall(VecDotBegin(Axr,ur,&rAr)); /* r*A*r */
773 : PetscCall(VecDotBegin(Axr,ui,&iAr)); /* i*A*r */
774 : PetscCall(VecDotBegin(Axi,ur,&rAi)); /* r*A*i */
775 : PetscCall(VecDotBegin(Axi,ui,&iAi)); /* i*A*i */
776 : PetscCall(VecDotBegin(Bxr,ur,&rBr)); /* r*B*r */
777 : PetscCall(VecDotBegin(Bxr,ui,&iBr)); /* i*B*r */
778 : PetscCall(VecDotBegin(Bxi,ur,&rBi)); /* r*B*i */
779 : PetscCall(VecDotBegin(Bxi,ui,&iBi)); /* i*B*i */
780 : PetscCall(VecDotEnd(Axr,ur,&rAr)); /* r*A*r */
781 : PetscCall(VecDotEnd(Axr,ui,&iAr)); /* i*A*r */
782 : PetscCall(VecDotEnd(Axi,ur,&rAi)); /* r*A*i */
783 : PetscCall(VecDotEnd(Axi,ui,&iAi)); /* i*A*i */
784 : PetscCall(VecDotEnd(Bxr,ur,&rBr)); /* r*B*r */
785 : PetscCall(VecDotEnd(Bxr,ui,&iBr)); /* i*B*r */
786 : PetscCall(VecDotEnd(Bxi,ur,&rBi)); /* r*B*i */
787 : PetscCall(VecDotEnd(Bxi,ui,&iBi)); /* i*B*i */
788 : b0 = rAr+iAi; /* rAr+iAi */
789 : b2 = rAi-iAr; /* rAi-iAr */
790 : b4 = rBr+iBi; /* rBr+iBi */
791 : b6 = rBi-iBr; /* rBi-iBr */
792 : b7 = b4*b4 + b6*b6; /* k */
793 : *eigr = (b0*b4 + b2*b6) / b7; /* eig_r */
794 : *eigi = (b2*b4 - b0*b6) / b7; /* eig_i */
795 : PetscFunctionReturn(PETSC_SUCCESS);
796 : }
797 : #endif
798 :
799 5607 : static inline PetscErrorCode dvd_compute_n_rr(PetscInt i_s,PetscInt n,PetscScalar *eigr,PetscScalar *eigi,Vec *u,Vec *Ax,Vec *Bx)
800 : {
801 5607 : PetscInt i;
802 5607 : PetscScalar b0,b1;
803 :
804 5607 : PetscFunctionBegin;
805 11214 : for (i=0; i<n; i++) {
806 : #if !defined(PETSC_USE_COMPLEX)
807 : if (eigi[i_s+i] != 0.0) {
808 : PetscScalar eigr0=0.0,eigi0=0.0;
809 : PetscCall(dvd_complex_rayleigh_quotient(u[i],u[i+1],Ax[i],Ax[i+1],Bx[i],Bx[i+1],&eigr0,&eigi0));
810 : if (PetscAbsScalar(eigr[i_s+i]-eigr0)/PetscAbsScalar(eigr[i_s+i]) > 1e-10 || PetscAbsScalar(eigi[i_s+i]-eigi0)/PetscAbsScalar(eigi[i_s+i]) > 1e-10) PetscCall(PetscInfo(u[0],"The eigenvalue %g%+gi is far from its Rayleigh quotient value %g%+gi\n",(double)eigr[i_s+i],(double)eigi[i_s+i],(double)eigr0,(double)eigi0));
811 : i++;
812 : } else
813 : #endif
814 : {
815 5607 : PetscCall(VecDotBegin(Ax[i],u[i],&b0));
816 5607 : PetscCall(VecDotBegin(Bx[i],u[i],&b1));
817 5607 : PetscCall(VecDotEnd(Ax[i],u[i],&b0));
818 5607 : PetscCall(VecDotEnd(Bx[i],u[i],&b1));
819 5607 : b0 = b0/b1;
820 5607 : if (PetscAbsScalar(eigr[i_s+i]-b0)/PetscAbsScalar(eigr[i_s+i]) > 1e-10) PetscCall(PetscInfo(u[0],"The eigenvalue %g+%g is far from its Rayleigh quotient value %g+%g\n",(double)PetscRealPart(eigr[i_s+i]),(double)PetscImaginaryPart(eigr[i_s+i]),(double)PetscRealPart(b0),(double)PetscImaginaryPart(b0)));
821 : }
822 : }
823 5607 : PetscFunctionReturn(PETSC_SUCCESS);
824 : }
825 :
826 : /*
827 : Compute: u <- X, v <- K*(theta[0]*A+theta[1]*B)*X,
828 : kr <- K^{-1}*(A-eig*B)*X, being X <- V*pX[i_s..i_e-1], Y <- W*pY[i_s..i_e-1]
829 : where
830 : pX,pY, the right and left eigenvectors of the projected system
831 : ld, the leading dimension of pX and pY
832 : */
833 5607 : static PetscErrorCode dvd_improvex_jd_proj_uv_KZX(dvdDashboard *d,PetscInt i_s,PetscInt i_e,Vec *u,Vec *v,Vec *kr,PetscScalar *theta,PetscScalar *thetai,PetscScalar *pX,PetscScalar *pY,PetscInt ld)
834 : {
835 5607 : PetscInt n = i_e-i_s,i;
836 5607 : PetscScalar *b;
837 5607 : Vec *Ax,*Bx,*r;
838 5607 : Mat M;
839 5607 : BV X;
840 :
841 5607 : PetscFunctionBegin;
842 5607 : PetscCall(BVDuplicateResize(d->eps->V,4,&X));
843 5607 : PetscCall(MatCreateSeqDense(PETSC_COMM_SELF,4,4,NULL,&M));
844 : /* u <- X(i) */
845 5607 : PetscCall(dvd_improvex_compute_X(d,i_s,i_e,u,pX,ld));
846 :
847 : /* v <- theta[0]A*u + theta[1]*B*u */
848 :
849 : /* Bx <- B*X(i) */
850 5607 : Bx = kr;
851 5607 : if (d->BX) {
852 2462 : for (i=i_s; i<i_e; ++i) PetscCall(BVMultVec(d->BX,1.0,0.0,Bx[i-i_s],&pX[ld*i]));
853 : } else {
854 8752 : for (i=0;i<n;i++) {
855 4376 : if (d->B) PetscCall(MatMult(d->B, u[i], Bx[i]));
856 4376 : else PetscCall(VecCopy(u[i], Bx[i]));
857 : }
858 : }
859 :
860 : /* Ax <- A*X(i) */
861 5607 : PetscCall(SlepcVecPoolGetVecs(d->auxV,i_e-i_s,&r));
862 5607 : Ax = r;
863 11214 : for (i=i_s; i<i_e; ++i) PetscCall(BVMultVec(d->AX,1.0,0.0,Ax[i-i_s],&pX[ld*i]));
864 :
865 : /* v <- Y(i) */
866 11214 : for (i=i_s; i<i_e; ++i) PetscCall(BVMultVec(d->W?d->W:d->eps->V,1.0,0.0,v[i-i_s],&pY[ld*i]));
867 :
868 : /* Recompute the eigenvalue */
869 5607 : PetscCall(dvd_compute_n_rr(i_s,n,d->eigr,d->eigi,v,Ax,Bx));
870 :
871 11214 : for (i=0;i<n;i++) {
872 : #if !defined(PETSC_USE_COMPLEX)
873 : if (d->eigi[i_s+i] != 0.0) {
874 : /* [r_i r_i+1 kr_i kr_i+1]*= [ theta_2i' 0 1 0
875 : 0 theta_2i' 0 1
876 : theta_2i+1 -thetai_i -eigr_i -eigi_i
877 : thetai_i theta_2i+1 eigi_i -eigr_i ] */
878 : PetscCall(MatDenseGetArrayWrite(M,&b));
879 : b[0] = b[5] = PetscConj(theta[2*i]);
880 : b[2] = b[7] = -theta[2*i+1];
881 : b[6] = -(b[3] = thetai[i]);
882 : b[1] = b[4] = 0.0;
883 : b[8] = b[13] = 1.0/d->nX[i_s+i];
884 : b[10] = b[15] = -d->eigr[i_s+i]/d->nX[i_s+i];
885 : b[14] = -(b[11] = d->eigi[i_s+i]/d->nX[i_s+i]);
886 : b[9] = b[12] = 0.0;
887 : PetscCall(MatDenseRestoreArrayWrite(M,&b));
888 : PetscCall(BVInsertVec(X,0,Ax[i]));
889 : PetscCall(BVInsertVec(X,1,Ax[i+1]));
890 : PetscCall(BVInsertVec(X,2,Bx[i]));
891 : PetscCall(BVInsertVec(X,3,Bx[i+1]));
892 : PetscCall(BVSetActiveColumns(X,0,4));
893 : PetscCall(BVMultInPlace(X,M,0,4));
894 : PetscCall(BVCopyVec(X,0,Ax[i]));
895 : PetscCall(BVCopyVec(X,1,Ax[i+1]));
896 : PetscCall(BVCopyVec(X,2,Bx[i]));
897 : PetscCall(BVCopyVec(X,3,Bx[i+1]));
898 : i++;
899 : } else
900 : #endif
901 : {
902 : /* [Ax_i Bx_i]*= [ theta_2i' 1/nX_i
903 : theta_2i+1 -eig_i/nX_i ] */
904 5607 : PetscCall(MatDenseGetArrayWrite(M,&b));
905 5607 : b[0] = PetscConj(theta[i*2]);
906 5607 : b[1] = theta[i*2+1];
907 5607 : b[4] = 1.0/d->nX[i_s+i];
908 5607 : b[5] = -d->eigr[i_s+i]/d->nX[i_s+i];
909 5607 : PetscCall(MatDenseRestoreArrayWrite(M,&b));
910 5607 : PetscCall(BVInsertVec(X,0,Ax[i]));
911 5607 : PetscCall(BVInsertVec(X,1,Bx[i]));
912 5607 : PetscCall(BVSetActiveColumns(X,0,2));
913 5607 : PetscCall(BVMultInPlace(X,M,0,2));
914 5607 : PetscCall(BVCopyVec(X,0,Ax[i]));
915 5607 : PetscCall(BVCopyVec(X,1,Bx[i]));
916 : }
917 : }
918 11214 : for (i=0; i<n; i++) d->nX[i_s+i] = 1.0;
919 :
920 : /* v <- K^{-1} r = K^{-1}(theta_2i'*Ax + theta_2i+1*Bx) */
921 11214 : for (i=0;i<n;i++) PetscCall(d->improvex_precond(d,i_s+i,r[i],v[i]));
922 5607 : PetscCall(SlepcVecPoolRestoreVecs(d->auxV,i_e-i_s,&r));
923 :
924 : /* kr <- P*(Ax - eig_i*Bx) */
925 5607 : PetscCall(d->calcpairs_proj_res(d,i_s,i_e,kr));
926 5607 : PetscCall(BVDestroy(&X));
927 5607 : PetscCall(MatDestroy(&M));
928 5607 : PetscFunctionReturn(PETSC_SUCCESS);
929 : }
930 :
931 5607 : static PetscErrorCode dvd_improvex_jd_lit_const_0(dvdDashboard *d,PetscInt i,PetscScalar* theta,PetscScalar* thetai,PetscInt *maxits,PetscReal *tol)
932 : {
933 5607 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
934 5607 : PetscReal a;
935 :
936 5607 : PetscFunctionBegin;
937 5607 : a = SlepcAbsEigenvalue(d->eigr[i],d->eigi[i]);
938 :
939 5607 : if (d->nR[i] < data->fix*a) {
940 578 : theta[0] = d->eigr[i];
941 578 : theta[1] = 1.0;
942 : #if !defined(PETSC_USE_COMPLEX)
943 : *thetai = d->eigi[i];
944 : #endif
945 : } else {
946 5029 : theta[0] = d->target[0];
947 5029 : theta[1] = d->target[1];
948 : #if !defined(PETSC_USE_COMPLEX)
949 : *thetai = 0.0;
950 : #endif
951 : }
952 :
953 : #if defined(PETSC_USE_COMPLEX)
954 5607 : if (thetai) *thetai = 0.0;
955 : #endif
956 5607 : *maxits = data->maxits;
957 5607 : *tol = data->tol;
958 5607 : PetscFunctionReturn(PETSC_SUCCESS);
959 : }
960 :
961 240 : PetscErrorCode dvd_improvex_jd_lit_const(dvdDashboard *d,dvdBlackboard *b,PetscInt maxits,PetscReal tol,PetscReal fix)
962 : {
963 240 : dvdImprovex_jd *data = (dvdImprovex_jd*)d->improveX_data;
964 :
965 240 : PetscFunctionBegin;
966 : /* Setup the step */
967 240 : if (b->state >= DVD_STATE_CONF) {
968 80 : data->maxits = maxits;
969 80 : data->tol = tol;
970 80 : data->fix = fix;
971 80 : d->improvex_jd_lit = dvd_improvex_jd_lit_const_0;
972 : }
973 240 : PetscFunctionReturn(PETSC_SUCCESS);
974 : }
975 :
976 240 : PetscErrorCode dvd_improvex_jd_proj_uv(dvdDashboard *d,dvdBlackboard *b)
977 : {
978 240 : PetscFunctionBegin;
979 : /* Setup the step */
980 240 : if (b->state >= DVD_STATE_CONF) {
981 80 : d->improvex_jd_proj_uv = dvd_improvex_jd_proj_uv_KZX;
982 : }
983 240 : PetscFunctionReturn(PETSC_SUCCESS);
984 : }
985 :
986 25613 : PetscErrorCode dvd_improvex_compute_X(dvdDashboard *d,PetscInt i_s,PetscInt i_e,Vec *u_,PetscScalar *pX,PetscInt ld)
987 : {
988 25613 : PetscInt n = i_e - i_s,i;
989 25613 : Vec *u;
990 :
991 25613 : PetscFunctionBegin;
992 25613 : if (u_) u = u_;
993 2601 : else if (d->correctXnorm) PetscCall(SlepcVecPoolGetVecs(d->auxV,i_e-i_s,&u));
994 25613 : if (u_ || d->correctXnorm) {
995 46310 : for (i=0;i<n;i++) PetscCall(BVMultVec(d->eps->V,1.0,0.0,u[i],&pX[ld*(i+i_s)]));
996 : }
997 : /* nX(i) <- ||X(i)|| */
998 25613 : if (d->correctXnorm) {
999 2478 : for (i=0;i<n;i++) PetscCall(VecNormBegin(u[i],NORM_2,&d->nX[i_s+i]));
1000 2478 : for (i=0;i<n;i++) PetscCall(VecNormEnd(u[i],NORM_2,&d->nX[i_s+i]));
1001 : #if !defined(PETSC_USE_COMPLEX)
1002 : for (i=0;i<n;i++) {
1003 : if (d->eigi[i_s+i] != 0.0) {
1004 : d->nX[i_s+i] = d->nX[i_s+i+1] = PetscSqrtScalar(d->nX[i_s+i]*d->nX[i_s+i]+d->nX[i_s+i+1]*d->nX[i_s+i+1]);
1005 : i++;
1006 : }
1007 : }
1008 : #endif
1009 : } else {
1010 48748 : for (i=0;i<n;i++) d->nX[i_s+i] = 1.0;
1011 : }
1012 25613 : if (d->correctXnorm && !u_) PetscCall(SlepcVecPoolRestoreVecs(d->auxV,i_e-i_s,&u));
1013 25613 : PetscFunctionReturn(PETSC_SUCCESS);
1014 : }
|
