Actual source code: vecs.c
slepc-3.13.3 2020-06-14
1: /*
2: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
3: SLEPc - Scalable Library for Eigenvalue Problem Computations
4: Copyright (c) 2002-2020, Universitat Politecnica de Valencia, Spain
6: This file is part of SLEPc.
7: SLEPc is distributed under a 2-clause BSD license (see LICENSE).
8: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
9: */
10: /*
11: BV implemented as an array of independent Vecs
12: */
14: #include <slepc/private/bvimpl.h>
16: typedef struct {
17: Vec *V;
18: PetscInt vmip; /* Version of BVMultInPlace:
19: 0: memory-efficient version, uses VecGetArray (default in CPU)
20: 1: version that allocates (e-s) work vectors in every call (default in GPU) */
21: } BV_VECS;
23: PetscErrorCode BVMult_Vecs(BV Y,PetscScalar alpha,PetscScalar beta,BV X,Mat Q)
24: {
26: BV_VECS *y = (BV_VECS*)Y->data,*x = (BV_VECS*)X->data;
27: PetscScalar *q,*s=NULL;
28: PetscInt i,j,ldq;
31: if (Q) {
32: MatGetSize(Q,&ldq,NULL);
33: if (alpha!=1.0) {
34: BVAllocateWork_Private(Y,X->k-X->l);
35: s = Y->work;
36: }
37: MatDenseGetArray(Q,&q);
38: for (j=Y->l;j<Y->k;j++) {
39: VecScale(y->V[Y->nc+j],beta);
40: if (alpha!=1.0) {
41: for (i=X->l;i<X->k;i++) s[i-X->l] = alpha*q[i+j*ldq];
42: } else s = q+j*ldq+X->l;
43: VecMAXPY(y->V[Y->nc+j],X->k-X->l,s,x->V+X->nc+X->l);
44: }
45: MatDenseRestoreArray(Q,&q);
46: } else {
47: for (j=0;j<Y->k-Y->l;j++) {
48: VecScale(y->V[Y->nc+Y->l+j],beta);
49: VecAXPY(y->V[Y->nc+Y->l+j],alpha,x->V[X->nc+X->l+j]);
50: }
51: }
52: return(0);
53: }
55: PetscErrorCode BVMultVec_Vecs(BV X,PetscScalar alpha,PetscScalar beta,Vec y,PetscScalar *q)
56: {
58: BV_VECS *x = (BV_VECS*)X->data;
59: PetscScalar *s=NULL,*qq=q;
60: PetscInt i;
63: if (alpha!=1.0) {
64: BVAllocateWork_Private(X,X->k-X->l);
65: s = X->work;
66: }
67: if (!q) { VecGetArray(X->buffer,&qq); }
68: VecScale(y,beta);
69: if (alpha!=1.0) {
70: for (i=0;i<X->k-X->l;i++) s[i] = alpha*qq[i];
71: } else s = qq;
72: VecMAXPY(y,X->k-X->l,s,x->V+X->nc+X->l);
73: if (!q) { VecRestoreArray(X->buffer,&qq); }
74: return(0);
75: }
77: /*
78: BVMultInPlace_Vecs_ME - V(:,s:e-1) = V*Q(:,s:e-1) for regular vectors.
80: Memory-efficient version, uses VecGetArray (default in CPU)
82: Writing V = [ V1 V2 V3 ] and Q(:,s:e-1) = [ Q1 Q2 Q3 ]', where V2
83: corresponds to the columns s:e-1, the computation is done as
84: V2 := V2*Q2 + V1*Q1 + V3*Q3
85: */
86: PetscErrorCode BVMultInPlace_Vecs_ME(BV V,Mat Q,PetscInt s,PetscInt e)
87: {
89: BV_VECS *ctx = (BV_VECS*)V->data;
90: PetscScalar *q;
91: PetscInt i,ldq;
94: MatGetSize(Q,&ldq,NULL);
95: MatDenseGetArray(Q,&q);
96: /* V2 := V2*Q2 */
97: BVMultInPlace_Vecs_Private(V,V->n,e-s,ldq,ctx->V+V->nc+s,q+s*ldq+s,PETSC_FALSE);
98: /* V2 += V1*Q1 + V3*Q3 */
99: for (i=s;i<e;i++) {
100: if (s>V->l) {
101: VecMAXPY(ctx->V[V->nc+i],s-V->l,q+i*ldq+V->l,ctx->V+V->nc+V->l);
102: }
103: if (V->k>e) {
104: VecMAXPY(ctx->V[V->nc+i],V->k-e,q+i*ldq+e,ctx->V+V->nc+e);
105: }
106: }
107: MatDenseRestoreArray(Q,&q);
108: return(0);
109: }
111: /*
112: BVMultInPlace_Vecs_Alloc - V(:,s:e-1) = V*Q(:,s:e-1) for regular vectors.
114: Version that allocates (e-s) work vectors in every call (default in GPU)
115: */
116: PetscErrorCode BVMultInPlace_Vecs_Alloc(BV V,Mat Q,PetscInt s,PetscInt e)
117: {
119: BV_VECS *ctx = (BV_VECS*)V->data;
120: PetscScalar *q;
121: PetscInt i,ldq;
122: Vec *W;
125: MatGetSize(Q,&ldq,NULL);
126: MatDenseGetArray(Q,&q);
127: VecDuplicateVecs(V->t,e-s,&W);
128: for (i=s;i<e;i++) {
129: VecMAXPY(W[i-s],V->k-V->l,q+i*ldq+V->l,ctx->V+V->nc+V->l);
130: }
131: for (i=s;i<e;i++) {
132: VecCopy(W[i-s],ctx->V[V->nc+i]);
133: }
134: VecDestroyVecs(e-s,&W);
135: MatDenseRestoreArray(Q,&q);
136: return(0);
137: }
139: /*
140: BVMultInPlaceTranspose_Vecs - V(:,s:e-1) = V*Q'(:,s:e-1) for regular vectors.
141: */
142: PetscErrorCode BVMultInPlaceTranspose_Vecs(BV V,Mat Q,PetscInt s,PetscInt e)
143: {
145: BV_VECS *ctx = (BV_VECS*)V->data;
146: PetscScalar *q;
147: PetscInt i,j,ldq,n;
150: MatGetSize(Q,&ldq,&n);
151: MatDenseGetArray(Q,&q);
152: /* V2 := V2*Q2' */
153: BVMultInPlace_Vecs_Private(V,V->n,e-s,ldq,ctx->V+V->nc+s,q+s*ldq+s,PETSC_TRUE);
154: /* V2 += V1*Q1' + V3*Q3' */
155: for (i=s;i<e;i++) {
156: for (j=V->l;j<s;j++) {
157: VecAXPY(ctx->V[V->nc+i],q[i+j*ldq],ctx->V[V->nc+j]);
158: }
159: for (j=e;j<n;j++) {
160: VecAXPY(ctx->V[V->nc+i],q[i+j*ldq],ctx->V[V->nc+j]);
161: }
162: }
163: MatDenseRestoreArray(Q,&q);
164: return(0);
165: }
167: PetscErrorCode BVDot_Vecs(BV X,BV Y,Mat M)
168: {
170: BV_VECS *x = (BV_VECS*)X->data,*y = (BV_VECS*)Y->data;
171: PetscScalar *m;
172: PetscInt j,ldm;
175: MatGetSize(M,&ldm,NULL);
176: MatDenseGetArray(M,&m);
177: for (j=X->l;j<X->k;j++) {
178: VecMDot(x->V[X->nc+j],Y->k-Y->l,y->V+Y->nc+Y->l,m+j*ldm+Y->l);
179: }
180: MatDenseRestoreArray(M,&m);
181: return(0);
182: }
184: PetscErrorCode BVDotVec_Vecs(BV X,Vec y,PetscScalar *q)
185: {
187: BV_VECS *x = (BV_VECS*)X->data;
188: Vec z = y;
189: PetscScalar *qq=q;
192: if (X->matrix) {
193: BV_IPMatMult(X,y);
194: z = X->Bx;
195: }
196: if (!q) { VecGetArray(X->buffer,&qq); }
197: VecMDot(z,X->k-X->l,x->V+X->nc+X->l,qq);
198: if (!q) { VecRestoreArray(X->buffer,&qq); }
199: return(0);
200: }
202: PetscErrorCode BVDotVec_Begin_Vecs(BV X,Vec y,PetscScalar *m)
203: {
205: BV_VECS *x = (BV_VECS*)X->data;
206: Vec z = y;
209: if (X->matrix) {
210: BV_IPMatMult(X,y);
211: z = X->Bx;
212: }
213: VecMDotBegin(z,X->k-X->l,x->V+X->nc+X->l,m);
214: return(0);
215: }
217: PetscErrorCode BVDotVec_End_Vecs(BV X,Vec y,PetscScalar *m)
218: {
220: BV_VECS *x = (BV_VECS*)X->data;
223: VecMDotEnd(y,X->k-X->l,x->V+X->nc+X->l,m);
224: return(0);
225: }
227: PetscErrorCode BVScale_Vecs(BV bv,PetscInt j,PetscScalar alpha)
228: {
230: PetscInt i;
231: BV_VECS *ctx = (BV_VECS*)bv->data;
234: if (j<0) {
235: for (i=bv->l;i<bv->k;i++) {
236: VecScale(ctx->V[bv->nc+i],alpha);
237: }
238: } else {
239: VecScale(ctx->V[bv->nc+j],alpha);
240: }
241: return(0);
242: }
244: PetscErrorCode BVNorm_Vecs(BV bv,PetscInt j,NormType type,PetscReal *val)
245: {
247: PetscInt i;
248: PetscReal nrm;
249: BV_VECS *ctx = (BV_VECS*)bv->data;
252: if (j<0) {
253: if (type==NORM_FROBENIUS) {
254: *val = 0.0;
255: for (i=bv->l;i<bv->k;i++) {
256: VecNorm(ctx->V[bv->nc+i],NORM_2,&nrm);
257: *val += nrm*nrm;
258: }
259: *val = PetscSqrtReal(*val);
260: } else SETERRQ(PetscObjectComm((PetscObject)bv),PETSC_ERR_SUP,"Requested norm not implemented in BVVECS");
261: } else {
262: VecNorm(ctx->V[bv->nc+j],type,val);
263: }
264: return(0);
265: }
267: PetscErrorCode BVNorm_Begin_Vecs(BV bv,PetscInt j,NormType type,PetscReal *val)
268: {
270: BV_VECS *ctx = (BV_VECS*)bv->data;
273: if (j<0) SETERRQ(PetscObjectComm((PetscObject)bv),PETSC_ERR_SUP,"Requested norm not implemented in BVVECS");
274: else {
275: VecNormBegin(ctx->V[bv->nc+j],type,val);
276: }
277: return(0);
278: }
280: PetscErrorCode BVNorm_End_Vecs(BV bv,PetscInt j,NormType type,PetscReal *val)
281: {
283: BV_VECS *ctx = (BV_VECS*)bv->data;
286: if (j<0) SETERRQ(PetscObjectComm((PetscObject)bv),PETSC_ERR_SUP,"Requested norm not implemented in BVVECS");
287: else {
288: VecNormEnd(ctx->V[bv->nc+j],type,val);
289: }
290: return(0);
291: }
293: PetscErrorCode BVMatMult_Vecs(BV V,Mat A,BV W)
294: {
296: BV_VECS *v = (BV_VECS*)V->data,*w = (BV_VECS*)W->data;
297: PetscInt j;
298: PetscBool flg;
299: Mat Vmat,Wmat;
302: MatHasOperation(A,MATOP_MAT_MULT,&flg);
303: if (V->vmm && flg) {
304: BVGetMat(V,&Vmat);
305: BVGetMat(W,&Wmat);
306: MatProductCreateWithMat(A,Vmat,NULL,Wmat);
307: MatProductSetType(Wmat,MATPRODUCT_AB);
308: MatProductSetFromOptions(Wmat);
309: MatProductSymbolic(Wmat);
310: MatProductNumeric(Wmat);
311: MatProductClear(Wmat);
312: BVRestoreMat(V,&Vmat);
313: BVRestoreMat(W,&Wmat);
314: } else {
315: for (j=0;j<V->k-V->l;j++) {
316: MatMult(A,v->V[V->nc+V->l+j],w->V[W->nc+W->l+j]);
317: }
318: }
319: return(0);
320: }
322: PetscErrorCode BVCopy_Vecs(BV V,BV W)
323: {
325: BV_VECS *v = (BV_VECS*)V->data,*w = (BV_VECS*)W->data;
326: PetscInt j;
329: for (j=0;j<V->k-V->l;j++) {
330: VecCopy(v->V[V->nc+V->l+j],w->V[W->nc+W->l+j]);
331: }
332: return(0);
333: }
335: PetscErrorCode BVCopyColumn_Vecs(BV V,PetscInt j,PetscInt i)
336: {
338: BV_VECS *v = (BV_VECS*)V->data;
341: VecCopy(v->V[V->nc+j],v->V[V->nc+i]);
342: return(0);
343: }
345: PetscErrorCode BVResize_Vecs(BV bv,PetscInt m,PetscBool copy)
346: {
348: BV_VECS *ctx = (BV_VECS*)bv->data;
349: Vec *newV;
350: PetscInt j,lsplit;
351: char str[50];
352: BV parent;
355: if (bv->issplit==2) {
356: parent = bv->splitparent;
357: lsplit = parent->lsplit;
358: ctx->V = ((BV_VECS*)parent->data)->V+lsplit;
359: } else if (!bv->issplit) {
360: VecDuplicateVecs(bv->t,m,&newV);
361: PetscLogObjectParents(bv,m,newV);
362: if (((PetscObject)bv)->name) {
363: for (j=0;j<m;j++) {
364: PetscSNPrintf(str,50,"%s_%D",((PetscObject)bv)->name,j);
365: PetscObjectSetName((PetscObject)newV[j],str);
366: }
367: }
368: if (copy) {
369: for (j=0;j<PetscMin(m,bv->m);j++) {
370: VecCopy(ctx->V[j],newV[j]);
371: }
372: }
373: VecDestroyVecs(bv->m,&ctx->V);
374: ctx->V = newV;
375: }
376: return(0);
377: }
379: PetscErrorCode BVGetColumn_Vecs(BV bv,PetscInt j,Vec *v)
380: {
381: BV_VECS *ctx = (BV_VECS*)bv->data;
382: PetscInt l;
385: l = BVAvailableVec;
386: bv->cv[l] = ctx->V[bv->nc+j];
387: return(0);
388: }
390: PetscErrorCode BVGetArray_Vecs(BV bv,PetscScalar **a)
391: {
392: PetscErrorCode ierr;
393: BV_VECS *ctx = (BV_VECS*)bv->data;
394: PetscInt j;
395: const PetscScalar *p;
398: PetscMalloc1((bv->nc+bv->m)*bv->n,a);
399: for (j=0;j<bv->nc+bv->m;j++) {
400: VecGetArrayRead(ctx->V[j],&p);
401: PetscArraycpy(*a+j*bv->n,p,bv->n);
402: VecRestoreArrayRead(ctx->V[j],&p);
403: }
404: return(0);
405: }
407: PetscErrorCode BVRestoreArray_Vecs(BV bv,PetscScalar **a)
408: {
410: BV_VECS *ctx = (BV_VECS*)bv->data;
411: PetscInt j;
412: PetscScalar *p;
415: for (j=0;j<bv->nc+bv->m;j++) {
416: VecGetArray(ctx->V[j],&p);
417: PetscArraycpy(p,*a+j*bv->n,bv->n);
418: VecRestoreArray(ctx->V[j],&p);
419: }
420: PetscFree(*a);
421: return(0);
422: }
424: PetscErrorCode BVGetArrayRead_Vecs(BV bv,const PetscScalar **a)
425: {
426: PetscErrorCode ierr;
427: BV_VECS *ctx = (BV_VECS*)bv->data;
428: PetscInt j;
429: const PetscScalar *p;
432: PetscMalloc1((bv->nc+bv->m)*bv->n,(PetscScalar**)a);
433: for (j=0;j<bv->nc+bv->m;j++) {
434: VecGetArrayRead(ctx->V[j],&p);
435: PetscArraycpy((PetscScalar*)*a+j*bv->n,p,bv->n);
436: VecRestoreArrayRead(ctx->V[j],&p);
437: }
438: return(0);
439: }
441: PetscErrorCode BVRestoreArrayRead_Vecs(BV bv,const PetscScalar **a)
442: {
446: PetscFree(*a);
447: return(0);
448: }
450: /*
451: Sets the value of vmip flag and resets ops->multinplace accordingly
452: */
453: PETSC_STATIC_INLINE PetscErrorCode BVVecsSetVmip(BV bv,PetscInt vmip)
454: {
455: typedef PetscErrorCode (*fmultinplace)(BV,Mat,PetscInt,PetscInt);
456: fmultinplace multinplace[2] = {BVMultInPlace_Vecs_ME, BVMultInPlace_Vecs_Alloc};
457: BV_VECS *ctx = (BV_VECS*)bv->data;
460: ctx->vmip = vmip;
461: bv->ops->multinplace = multinplace[vmip];
462: return(0);
463: }
465: PetscErrorCode BVSetFromOptions_Vecs(PetscOptionItems *PetscOptionsObject,BV bv)
466: {
468: BV_VECS *ctx = (BV_VECS*)bv->data;
471: PetscOptionsHead(PetscOptionsObject,"BV Vecs Options");
473: PetscOptionsRangeInt("-bv_vecs_vmip","Version of BVMultInPlace operation","",ctx->vmip,&ctx->vmip,NULL,0,1);
474: BVVecsSetVmip(bv,ctx->vmip);
476: PetscOptionsTail();
477: return(0);
478: }
480: PetscErrorCode BVView_Vecs(BV bv,PetscViewer viewer)
481: {
482: PetscErrorCode ierr;
483: BV_VECS *ctx = (BV_VECS*)bv->data;
484: PetscInt j;
485: PetscViewerFormat format;
486: PetscBool isascii,ismatlab=PETSC_FALSE;
487: const char *bvname,*name;
490: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&isascii);
491: if (isascii) {
492: PetscViewerGetFormat(viewer,&format);
493: if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) return(0);
494: if (format == PETSC_VIEWER_ASCII_MATLAB) ismatlab = PETSC_TRUE;
495: }
496: if (ismatlab) {
497: PetscObjectGetName((PetscObject)bv,&bvname);
498: PetscViewerASCIIPrintf(viewer,"%s=[];\n",bvname);
499: }
500: for (j=bv->nc;j<bv->nc+bv->m;j++) {
501: VecView(ctx->V[j],viewer);
502: if (ismatlab) {
503: PetscObjectGetName((PetscObject)ctx->V[j],&name);
504: PetscViewerASCIIPrintf(viewer,"%s=[%s,%s];clear %s\n",bvname,bvname,name,name);
505: }
506: }
507: return(0);
508: }
510: PetscErrorCode BVDestroy_Vecs(BV bv)
511: {
513: BV_VECS *ctx = (BV_VECS*)bv->data;
516: if (!bv->issplit) { VecDestroyVecs(bv->nc+bv->m,&ctx->V); }
517: PetscFree(bv->data);
518: return(0);
519: }
521: PetscErrorCode BVDuplicate_Vecs(BV V,BV W)
522: {
524: BV_VECS *ctx = (BV_VECS*)V->data;
527: BVVecsSetVmip(W,ctx->vmip);
528: return(0);
529: }
531: SLEPC_EXTERN PetscErrorCode BVCreate_Vecs(BV bv)
532: {
534: BV_VECS *ctx;
535: PetscInt j,lsplit;
536: PetscBool isgpu;
537: char str[50];
538: BV parent;
539: Vec *Vpar;
542: PetscNewLog(bv,&ctx);
543: bv->data = (void*)ctx;
545: if (bv->issplit) {
546: /* split BV: share the Vecs of the parent BV */
547: parent = bv->splitparent;
548: lsplit = parent->lsplit;
549: Vpar = ((BV_VECS*)parent->data)->V;
550: ctx->V = (bv->issplit==1)? Vpar: Vpar+lsplit;
551: } else {
552: /* regular BV: create array of Vecs to store the BV columns */
553: VecDuplicateVecs(bv->t,bv->m,&ctx->V);
554: PetscLogObjectParents(bv,bv->m,ctx->V);
555: if (((PetscObject)bv)->name) {
556: for (j=0;j<bv->m;j++) {
557: PetscSNPrintf(str,50,"%s_%D",((PetscObject)bv)->name,j);
558: PetscObjectSetName((PetscObject)ctx->V[j],str);
559: }
560: }
561: }
563: if (bv->Acreate) {
564: for (j=0;j<bv->m;j++) {
565: MatGetColumnVector(bv->Acreate,ctx->V[j],j);
566: }
567: MatDestroy(&bv->Acreate);
568: }
570: /* Default version of BVMultInPlace */
571: PetscObjectTypeCompareAny((PetscObject)bv->t,&isgpu,VECSEQCUDA,VECMPICUDA,"");
572: ctx->vmip = isgpu? 1: 0;
574: /* Default BVMatMult method */
575: bv->vmm = BV_MATMULT_VECS;
577: /* Deferred call to setfromoptions */
578: if (bv->defersfo) {
579: PetscObjectOptionsBegin((PetscObject)bv);
580: BVSetFromOptions_Vecs(PetscOptionsObject,bv);
581: PetscOptionsEnd();
582: }
583: BVVecsSetVmip(bv,ctx->vmip);
585: bv->ops->mult = BVMult_Vecs;
586: bv->ops->multvec = BVMultVec_Vecs;
587: bv->ops->multinplacetrans = BVMultInPlaceTranspose_Vecs;
588: bv->ops->dot = BVDot_Vecs;
589: bv->ops->dotvec = BVDotVec_Vecs;
590: bv->ops->dotvec_begin = BVDotVec_Begin_Vecs;
591: bv->ops->dotvec_end = BVDotVec_End_Vecs;
592: bv->ops->scale = BVScale_Vecs;
593: bv->ops->norm = BVNorm_Vecs;
594: bv->ops->norm_begin = BVNorm_Begin_Vecs;
595: bv->ops->norm_end = BVNorm_End_Vecs;
596: bv->ops->matmult = BVMatMult_Vecs;
597: bv->ops->copy = BVCopy_Vecs;
598: bv->ops->copycolumn = BVCopyColumn_Vecs;
599: bv->ops->resize = BVResize_Vecs;
600: bv->ops->getcolumn = BVGetColumn_Vecs;
601: bv->ops->getarray = BVGetArray_Vecs;
602: bv->ops->restorearray = BVRestoreArray_Vecs;
603: bv->ops->getarrayread = BVGetArrayRead_Vecs;
604: bv->ops->restorearrayread = BVRestoreArrayRead_Vecs;
605: bv->ops->destroy = BVDestroy_Vecs;
606: bv->ops->duplicate = BVDuplicate_Vecs;
607: bv->ops->setfromoptions = BVSetFromOptions_Vecs;
608: bv->ops->view = BVView_Vecs;
609: return(0);
610: }