Actual source code: bvimpl.h

slepc-3.13.2 2020-05-12
Report Typos and Errors
  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: */

 11: #if !defined(SLEPCBVIMPL_H)
 12: #define SLEPCBVIMPL_H

 14:  #include <slepcbv.h>
 15:  #include <slepc/private/slepcimpl.h>

 17: SLEPC_EXTERN PetscBool BVRegisterAllCalled;
 18: SLEPC_EXTERN PetscErrorCode BVRegisterAll(void);

 20: SLEPC_EXTERN PetscLogEvent BV_Create,BV_Copy,BV_Mult,BV_MultVec,BV_MultInPlace,BV_Dot,BV_DotVec,BV_Orthogonalize,BV_OrthogonalizeVec,BV_Scale,BV_Norm,BV_NormVec,BV_SetRandom,BV_MatMult,BV_MatMultVec,BV_MatProject;

 22: typedef struct _BVOps *BVOps;

 24: struct _BVOps {
 25:   PetscErrorCode (*mult)(BV,PetscScalar,PetscScalar,BV,Mat);
 26:   PetscErrorCode (*multvec)(BV,PetscScalar,PetscScalar,Vec,PetscScalar*);
 27:   PetscErrorCode (*multinplace)(BV,Mat,PetscInt,PetscInt);
 28:   PetscErrorCode (*multinplacetrans)(BV,Mat,PetscInt,PetscInt);
 29:   PetscErrorCode (*dot)(BV,BV,Mat);
 30:   PetscErrorCode (*dotvec)(BV,Vec,PetscScalar*);
 31:   PetscErrorCode (*dotvec_local)(BV,Vec,PetscScalar*);
 32:   PetscErrorCode (*dotvec_begin)(BV,Vec,PetscScalar*);
 33:   PetscErrorCode (*dotvec_end)(BV,Vec,PetscScalar*);
 34:   PetscErrorCode (*scale)(BV,PetscInt,PetscScalar);
 35:   PetscErrorCode (*norm)(BV,PetscInt,NormType,PetscReal*);
 36:   PetscErrorCode (*norm_local)(BV,PetscInt,NormType,PetscReal*);
 37:   PetscErrorCode (*norm_begin)(BV,PetscInt,NormType,PetscReal*);
 38:   PetscErrorCode (*norm_end)(BV,PetscInt,NormType,PetscReal*);
 39:   PetscErrorCode (*matmult)(BV,Mat,BV);
 40:   PetscErrorCode (*copy)(BV,BV);
 41:   PetscErrorCode (*copycolumn)(BV,PetscInt,PetscInt);
 42:   PetscErrorCode (*resize)(BV,PetscInt,PetscBool);
 43:   PetscErrorCode (*getcolumn)(BV,PetscInt,Vec*);
 44:   PetscErrorCode (*restorecolumn)(BV,PetscInt,Vec*);
 45:   PetscErrorCode (*getarray)(BV,PetscScalar**);
 46:   PetscErrorCode (*restorearray)(BV,PetscScalar**);
 47:   PetscErrorCode (*getarrayread)(BV,const PetscScalar**);
 48:   PetscErrorCode (*restorearrayread)(BV,const PetscScalar**);
 49:   PetscErrorCode (*restoresplit)(BV,BV*,BV*);
 50:   PetscErrorCode (*gramschmidt)(BV,PetscInt,Vec,PetscBool*,PetscScalar*,PetscScalar*,PetscReal*,PetscReal*);
 51:   PetscErrorCode (*duplicate)(BV,BV);
 52:   PetscErrorCode (*create)(BV);
 53:   PetscErrorCode (*setfromoptions)(PetscOptionItems*,BV);
 54:   PetscErrorCode (*view)(BV,PetscViewer);
 55:   PetscErrorCode (*destroy)(BV);
 56: };

 58: struct _p_BV {
 59:   PETSCHEADER(struct _BVOps);
 60:   /*------------------------- User parameters --------------------------*/
 61:   Vec                t;            /* template vector */
 62:   PetscInt           n,N;          /* dimensions of vectors (local, global) */
 63:   PetscInt           m;            /* number of vectors */
 64:   PetscInt           l;            /* number of leading columns */
 65:   PetscInt           k;            /* number of active columns */
 66:   PetscInt           nc;           /* number of constraints */
 67:   BVOrthogType       orthog_type;  /* the method of vector orthogonalization */
 68:   BVOrthogRefineType orthog_ref;   /* refinement method */
 69:   PetscReal          orthog_eta;   /* refinement threshold */
 70:   BVOrthogBlockType  orthog_block; /* the method of block orthogonalization */
 71:   Mat                matrix;       /* inner product matrix */
 72:   PetscBool          indef;        /* matrix is indefinite */
 73:   BVMatMultType      vmm;          /* version of matmult operation */
 74:   PetscBool          rrandom;      /* reproducible random vectors */

 76:   /*---------------------- Cached data and workspace -------------------*/
 77:   Vec                buffer;       /* buffer vector used in orthogonalization */
 78:   Mat                Abuffer;      /* auxiliary seqdense matrix that wraps the buffer */
 79:   Vec                Bx;           /* result of matrix times a vector x */
 80:   PetscObjectId      xid;          /* object id of vector x */
 81:   PetscObjectState   xstate;       /* state of vector x */
 82:   Vec                cv[2];        /* column vectors obtained with BVGetColumn() */
 83:   PetscInt           ci[2];        /* column indices of obtained vectors */
 84:   PetscObjectState   st[2];        /* state of obtained vectors */
 85:   PetscObjectId      id[2];        /* object id of obtained vectors */
 86:   PetscScalar        *h,*c;        /* orthogonalization coefficients */
 87:   Vec                omega;        /* signature matrix values for indefinite case */
 88:   PetscBool          defersfo;     /* deferred call to setfromoptions */
 89:   BV                 cached;       /* cached BV to store result of matrix times BV */
 90:   PetscObjectState   bvstate;      /* state of BV when BVApplyMatrixBV() was called */
 91:   BV                 L,R;          /* BV objects obtained with BVGetSplit() */
 92:   PetscObjectState   lstate,rstate;/* state of L and R when BVGetSplit() was called */
 93:   PetscInt           lsplit;       /* the value of l when BVGetSplit() was called */
 94:   PetscInt           issplit;      /* >0 if this BV has been created by splitting (1=left, 2=right) */
 95:   BV                 splitparent;  /* my parent if I am a split BV */
 96:   PetscRandom        rand;         /* random number generator */
 97:   Mat                Acreate;      /* matrix given at BVCreateFromMat() */
 98:   Mat                Aget;         /* matrix returned for BVGetMat() */
 99:   PetscBool          cuda;         /* true if GPU must be used in SVEC */
100:   PetscScalar        *work;
101:   PetscInt           lwork;
102:   void               *data;
103: };

105: /*
106:   BV_SafeSqrt - Computes the square root of a scalar value alpha, which is
107:   assumed to be z'*B*z. The result is
108:     if definite inner product:     res = sqrt(alpha)
109:     if indefinite inner product:   res = sgn(alpha)*sqrt(abs(alpha))
110: */
111: PETSC_STATIC_INLINE PetscErrorCode BV_SafeSqrt(BV bv,PetscScalar alpha,PetscReal *res)
112: {
114:   PetscReal      absal,realp;

117:   absal = PetscAbsScalar(alpha);
118:   realp = PetscRealPart(alpha);
119:   if (absal<PETSC_MACHINE_EPSILON) {
120:     PetscInfo(bv,"Zero norm, either the vector is zero or a semi-inner product is being used\n");
121:   }
122: #if defined(PETSC_USE_COMPLEX)
123:   if (PetscAbsReal(PetscImaginaryPart(alpha))>10*PETSC_MACHINE_EPSILON && PetscAbsReal(PetscImaginaryPart(alpha))/absal>100*PETSC_MACHINE_EPSILON) SETERRQ1(PetscObjectComm((PetscObject)bv),1,"The inner product is not well defined: nonzero imaginary part %g",PetscImaginaryPart(alpha));
124: #endif
125:   if (bv->indef) {
126:     *res = (realp<0.0)? -PetscSqrtReal(-realp): PetscSqrtReal(realp);
127:   } else {
128:     if (realp<-10*PETSC_MACHINE_EPSILON) SETERRQ(PetscObjectComm((PetscObject)bv),1,"The inner product is not well defined: indefinite matrix");
129:     *res = (realp<0.0)? 0.0: PetscSqrtReal(realp);
130:   }
131:   return(0);
132: }

134: /*
135:   BV_IPMatMult - Multiply a vector x by the inner-product matrix, cache the
136:   result in Bx.
137: */
138: PETSC_STATIC_INLINE PetscErrorCode BV_IPMatMult(BV bv,Vec x)
139: {

143:   if (((PetscObject)x)->id != bv->xid || ((PetscObject)x)->state != bv->xstate) {
144:     if (!bv->Bx) {
145:       MatCreateVecs(bv->matrix,&bv->Bx,NULL);
146:       PetscLogObjectParent((PetscObject)bv,(PetscObject)bv->Bx);
147:     }
148:     MatMult(bv->matrix,x,bv->Bx);
149:     PetscObjectGetId((PetscObject)x,&bv->xid);
150:     PetscObjectStateGet((PetscObject)x,&bv->xstate);
151:   }
152:   return(0);
153: }

155: /*
156:   BV_IPMatMultBV - Multiply BV by the inner-product matrix, cache the
157:   result internally in bv->cached.
158: */
159: PETSC_STATIC_INLINE PetscErrorCode BV_IPMatMultBV(BV bv)
160: {

164:   BVGetCachedBV(bv,&bv->cached);
165:   if (((PetscObject)bv)->state != bv->bvstate || bv->l != bv->cached->l || bv->k != bv->cached->k) {
166:     BVSetActiveColumns(bv->cached,bv->l,bv->k);
167:     if (bv->matrix) {
168:       BVMatMult(bv,bv->matrix,bv->cached);
169:     } else {
170:       BVCopy(bv,bv->cached);
171:     }
172:     bv->bvstate = ((PetscObject)bv)->state;
173:   }
174:   return(0);
175: }

177: /*
178:   BV_AllocateCoeffs - Allocate orthogonalization coefficients if not done already.
179: */
180: PETSC_STATIC_INLINE PetscErrorCode BV_AllocateCoeffs(BV bv)
181: {

185:   if (!bv->h) {
186:     PetscMalloc2(bv->nc+bv->m,&bv->h,bv->nc+bv->m,&bv->c);
187:     PetscLogObjectMemory((PetscObject)bv,2*bv->m*sizeof(PetscScalar));
188:   }
189:   return(0);
190: }

192: /*
193:   BV_AllocateSignature - Allocate signature coefficients if not done already.
194: */
195: PETSC_STATIC_INLINE PetscErrorCode BV_AllocateSignature(BV bv)
196: {

200:   if (bv->indef && !bv->omega) {
201:     if (bv->cuda) {
202: #if defined(PETSC_HAVE_CUDA)
203:       VecCreateSeqCUDA(PETSC_COMM_SELF,bv->nc+bv->m,&bv->omega);
204: #else
205:       SETERRQ(PetscObjectComm((PetscObject)bv),1,"Something wrong happened");
206: #endif
207:     } else {
208:       VecCreateSeq(PETSC_COMM_SELF,bv->nc+bv->m,&bv->omega);
209:     }
210:     PetscLogObjectParent((PetscObject)bv,(PetscObject)bv->omega);
211:     VecSet(bv->omega,1.0);
212:   }
213:   return(0);
214: }

216: /*
217:   BVAvailableVec: First (0) or second (1) vector available for
218:   getcolumn operation (or -1 if both vectors already fetched).
219: */
220: #define BVAvailableVec (((bv->ci[0]==-bv->nc-1)? 0: (bv->ci[1]==-bv->nc-1)? 1: -1))

222: /*
223:     Macros to test valid BV arguments
224: */
225: #if !defined(PETSC_USE_DEBUG)

227: #define BVCheckSizes(h,arg) do {} while (0)
228: #define BVCheckOp(h,arg,op) do {} while (0)

230: #else

232: #define BVCheckSizes(h,arg) \
233:   do { \
234:     if (!(h)->m) SETERRQ1(PetscObjectComm((PetscObject)(h)),PETSC_ERR_ARG_WRONGSTATE,"BV sizes have not been defined: Parameter #%d",arg); \
235:   } while (0)

237: #define BVCheckOp(h,arg,op) \
238:   do { \
239:     if (!(h)->ops->op) SETERRQ1(PetscObjectComm((PetscObject)(h)),PETSC_ERR_SUP,"Operation not implemented in this BV type: Parameter #%d",arg); \
240:   } while (0)

242: #endif

244: SLEPC_INTERN PetscErrorCode BVView_Vecs(BV,PetscViewer);

246: SLEPC_INTERN PetscErrorCode BVAllocateWork_Private(BV,PetscInt);

248: SLEPC_INTERN PetscErrorCode BVMult_BLAS_Private(BV,PetscInt,PetscInt,PetscInt,PetscInt,PetscScalar,const PetscScalar*,const PetscScalar*,PetscScalar,PetscScalar*);
249: SLEPC_INTERN PetscErrorCode BVMultVec_BLAS_Private(BV,PetscInt,PetscInt,PetscScalar,const PetscScalar*,const PetscScalar*,PetscScalar,PetscScalar*);
250: SLEPC_INTERN PetscErrorCode BVMultInPlace_BLAS_Private(BV,PetscInt,PetscInt,PetscInt,PetscInt,PetscInt,PetscScalar*,const PetscScalar*,PetscBool);
251: SLEPC_INTERN PetscErrorCode BVMultInPlace_Vecs_Private(BV,PetscInt,PetscInt,PetscInt,Vec*,const PetscScalar*,PetscBool);
252: SLEPC_INTERN PetscErrorCode BVAXPY_BLAS_Private(BV,PetscInt,PetscInt,PetscScalar,const PetscScalar*,PetscScalar,PetscScalar*);
253: SLEPC_INTERN PetscErrorCode BVDot_BLAS_Private(BV,PetscInt,PetscInt,PetscInt,PetscInt,const PetscScalar*,const PetscScalar*,PetscScalar*,PetscBool);
254: SLEPC_INTERN PetscErrorCode BVDotVec_BLAS_Private(BV,PetscInt,PetscInt,const PetscScalar*,const PetscScalar*,PetscScalar*,PetscBool);
255: SLEPC_INTERN PetscErrorCode BVScale_BLAS_Private(BV,PetscInt,PetscScalar*,PetscScalar);
256: SLEPC_INTERN PetscErrorCode BVNorm_LAPACK_Private(BV,PetscInt,PetscInt,const PetscScalar*,NormType,PetscReal*,PetscBool);
257: SLEPC_INTERN PetscErrorCode BVMatCholInv_LAPACK_Private(BV,Mat,Mat);
258: SLEPC_INTERN PetscErrorCode BVMatTriInv_LAPACK_Private(BV,Mat,Mat);
259: SLEPC_INTERN PetscErrorCode BVMatSVQB_LAPACK_Private(BV,Mat,Mat);
260: SLEPC_INTERN PetscErrorCode BVOrthogonalize_LAPACK_TSQR(BV,PetscInt,PetscInt,PetscScalar*,PetscScalar*,PetscInt);
261: SLEPC_INTERN PetscErrorCode BVOrthogonalize_LAPACK_TSQR_OnlyR(BV,PetscInt,PetscInt,PetscScalar*,PetscScalar*,PetscInt);

263: /* reduction operation used in BVOrthogonalize */
264: SLEPC_EXTERN MPI_Op MPIU_TSQR;
265: SLEPC_EXTERN void MPIAPI SlepcGivensPacked(void*,void*,PetscMPIInt*,MPI_Datatype*);

267: #if defined(PETSC_HAVE_CUDA)

269: /* complex single */
270: #if defined(PETSC_USE_COMPLEX)
271: #if defined(PETSC_USE_REAL_SINGLE)
272: #define cublasXgemm(a,b,c,d,e,f,g,h,i,j,k,l,m,n) cublasCgemm((a),(b),(c),(d),(e),(f),(cuComplex*)(g),(cuComplex*)(h),(i),(cuComplex*)(j),(k),(cuComplex*)(l),(cuComplex*)(m),(n))
273: #define cublasXgemv(a,b,c,d,e,f,g,h,i,j,k,l) cublasCgemv((a),(b),(c),(d),(cuComplex*)(e),(cuComplex*)(f),(g),(cuComplex*)(h),(i),(cuComplex*)(j),(cuComplex*)(k),(l))
274: #define cublasXscal(a,b,c,d,e) cublasCscal(a,b,(const cuComplex*)(c),(cuComplex*)(d),e)
275: #define cublasXnrm2(a,b,c,d,e) cublasScnrm2(a,b,(const cuComplex*)(c),d,e)
276: #define cublasXaxpy(a,b,c,d,e,f,g) cublasCaxpy((a),(b),(cuComplex*)(c),(cuComplex*)(d),(e),(cuComplex*)(f),(g))
277: #define cublasXdotc(a,b,c,d,e,f,g) cublasCdotc((a),(b),(const cuComplex *)(c),(d),(const cuComplex *)(e),(f),(cuComplex *)(g))
278: #else /* complex double */
279: #define cublasXgemm(a,b,c,d,e,f,g,h,i,j,k,l,m,n) cublasZgemm((a),(b),(c),(d),(e),(f),(cuDoubleComplex*)(g),(cuDoubleComplex*)(h),(i),(cuDoubleComplex*)(j),(k),(cuDoubleComplex*)(l),(cuDoubleComplex*)(m),(n))
280: #define cublasXgemv(a,b,c,d,e,f,g,h,i,j,k,l) cublasZgemv((a),(b),(c),(d),(cuDoubleComplex*)(e),(cuDoubleComplex*)(f),(g),(cuDoubleComplex*)(h),(i),(cuDoubleComplex*)(j),(cuDoubleComplex*)(k),(l))
281: #define cublasXscal(a,b,c,d,e) cublasZscal(a,b,(const cuDoubleComplex*)(c),(cuDoubleComplex*)(d),e)
282: #define cublasXnrm2(a,b,c,d,e) cublasDznrm2(a,b,(const cuDoubleComplex*)(c),d,e)
283: #define cublasXaxpy(a,b,c,d,e,f,g) cublasZaxpy((a),(b),(cuDoubleComplex*)(c),(cuDoubleComplex*)(d),(e),(cuDoubleComplex*)(f),(g))
284: #define cublasXdotc(a,b,c,d,e,f,g) cublasZdotc((a),(b),(const cuDoubleComplex *)(c),(d),(const cuDoubleComplex *)(e),(f),(cuDoubleComplex *)(g))
285: #endif
286: #else /* real single */
287: #if defined(PETSC_USE_REAL_SINGLE)
288: #define cublasXgemm cublasSgemm
289: #define cublasXgemv cublasSgemv
290: #define cublasXscal cublasSscal
291: #define cublasXnrm2 cublasSnrm2
292: #define cublasXaxpy cublasSaxpy
293: #define cublasXdotc cublasSdot
294: #else /* real double */
295: #define cublasXgemm cublasDgemm
296: #define cublasXgemv cublasDgemv
297: #define cublasXscal cublasDscal
298: #define cublasXnrm2 cublasDnrm2
299: #define cublasXaxpy cublasDaxpy
300: #define cublasXdotc cublasDdot
301: #endif
302: #endif

304: SLEPC_INTERN PetscErrorCode BV_CleanCoefficients_CUDA(BV,PetscInt,PetscScalar*);
305: SLEPC_INTERN PetscErrorCode BV_AddCoefficients_CUDA(BV,PetscInt,PetscScalar*,PetscScalar*);
306: SLEPC_INTERN PetscErrorCode BV_SetValue_CUDA(BV,PetscInt,PetscInt,PetscScalar*,PetscScalar);
307: SLEPC_INTERN PetscErrorCode BV_SquareSum_CUDA(BV,PetscInt,PetscScalar*,PetscReal*);
308: SLEPC_INTERN PetscErrorCode BV_ApplySignature_CUDA(BV,PetscInt,PetscScalar*,PetscBool);
309: SLEPC_INTERN PetscErrorCode BV_SquareRoot_CUDA(BV,PetscInt,PetscScalar*,PetscReal*);
310: SLEPC_INTERN PetscErrorCode BV_StoreCoefficients_CUDA(BV,PetscInt,PetscScalar*,PetscScalar*);

312: #endif /* PETSC_HAVE_CUDA */

314: /*
315:    BV_CleanCoefficients_Default - Sets to zero all entries of column j of the bv buffer
316: */
317: PETSC_STATIC_INLINE PetscErrorCode BV_CleanCoefficients_Default(BV bv,PetscInt j,PetscScalar *h)
318: {
320:   PetscScalar    *hh=h,*a;
321:   PetscInt       i;

324:   if (!h) {
325:     VecGetArray(bv->buffer,&a);
326:     hh = a + j*(bv->nc+bv->m);
327:   }
328:   for (i=0;i<bv->nc+j;i++) hh[i] = 0.0;
329:   if (!h) { VecRestoreArray(bv->buffer,&a); }
330:   return(0);
331: }

333: /*
334:    BV_AddCoefficients_Default - Add the contents of the scratch (0-th column) of the bv buffer
335:    into column j of the bv buffer
336: */
337: PETSC_STATIC_INLINE PetscErrorCode BV_AddCoefficients_Default(BV bv,PetscInt j,PetscScalar *h,PetscScalar *c)
338: {
340:   PetscScalar    *hh=h,*cc=c;
341:   PetscInt       i;

344:   if (!h) {
345:     VecGetArray(bv->buffer,&cc);
346:     hh = cc + j*(bv->nc+bv->m);
347:   }
348:   for (i=0;i<bv->nc+j;i++) hh[i] += cc[i];
349:   if (!h) { VecRestoreArray(bv->buffer,&cc); }
350:   return(0);
351: }

353: /*
354:    BV_SetValue_Default - Sets value in row j (counted after the constraints) of column k
355:    of the coefficients array
356: */
357: PETSC_STATIC_INLINE PetscErrorCode BV_SetValue_Default(BV bv,PetscInt j,PetscInt k,PetscScalar *h,PetscScalar value)
358: {
360:   PetscScalar    *hh=h,*a;

363:   if (!h) {
364:     VecGetArray(bv->buffer,&a);
365:     hh = a + k*(bv->nc+bv->m);
366:   }
367:   hh[bv->nc+j] = value;
368:   if (!h) { VecRestoreArray(bv->buffer,&a); }
369:   return(0);
370: }

372: /*
373:    BV_SquareSum_Default - Returns the value h'*h, where h represents the contents of the
374:    coefficients array (up to position j)
375: */
376: PETSC_STATIC_INLINE PetscErrorCode BV_SquareSum_Default(BV bv,PetscInt j,PetscScalar *h,PetscReal *sum)
377: {
379:   PetscScalar    *hh=h;
380:   PetscInt       i;

383:   *sum = 0.0;
384:   if (!h) { VecGetArray(bv->buffer,&hh); }
385:   for (i=0;i<bv->nc+j;i++) *sum += PetscRealPart(hh[i]*PetscConj(hh[i]));
386:   if (!h) { VecRestoreArray(bv->buffer,&hh); }
387:   return(0);
388: }

390: /*
391:    BV_ApplySignature_Default - Computes the pointwise product h*omega, where h represents
392:    the contents of the coefficients array (up to position j) and omega is the signature;
393:    if inverse=TRUE then the operation is h/omega
394: */
395: PETSC_STATIC_INLINE PetscErrorCode BV_ApplySignature_Default(BV bv,PetscInt j,PetscScalar *h,PetscBool inverse)
396: {
397:   PetscErrorCode    ierr;
398:   PetscScalar       *hh=h;
399:   PetscInt          i;
400:   const PetscScalar *omega;

403:   if (!(bv->nc+j)) return(0);
404:   if (!h) { VecGetArray(bv->buffer,&hh); }
405:   VecGetArrayRead(bv->omega,&omega);
406:   if (inverse) for (i=0;i<bv->nc+j;i++) hh[i] /= PetscRealPart(omega[i]);
407:   else for (i=0;i<bv->nc+j;i++) hh[i] *= PetscRealPart(omega[i]);
408:   VecRestoreArrayRead(bv->omega,&omega);
409:   if (!h) { VecRestoreArray(bv->buffer,&hh); }
410:   return(0);
411: }

413: /*
414:    BV_SquareRoot_Default - Returns the square root of position j (counted after the constraints)
415:    of the coefficients array
416: */
417: PETSC_STATIC_INLINE PetscErrorCode BV_SquareRoot_Default(BV bv,PetscInt j,PetscScalar *h,PetscReal *beta)
418: {
420:   PetscScalar    *hh=h;

423:   if (!h) { VecGetArray(bv->buffer,&hh); }
424:   BV_SafeSqrt(bv,hh[bv->nc+j],beta);
425:   if (!h) { VecRestoreArray(bv->buffer,&hh); }
426:   return(0);
427: }

429: /*
430:    BV_StoreCoefficients_Default - Copy the contents of the coefficients array to an array dest
431:    provided by the caller (only values from l to j are copied)
432: */
433: PETSC_STATIC_INLINE PetscErrorCode BV_StoreCoefficients_Default(BV bv,PetscInt j,PetscScalar *h,PetscScalar *dest)
434: {
436:   PetscScalar    *hh=h,*a;
437:   PetscInt       i;

440:   if (!h) {
441:     VecGetArray(bv->buffer,&a);
442:     hh = a + j*(bv->nc+bv->m);
443:   }
444:   for (i=bv->l;i<j;i++) dest[i-bv->l] = hh[bv->nc+i];
445:   if (!h) { VecRestoreArray(bv->buffer,&a); }
446:   return(0);
447: }

449: /*
450:   BV_GetEigenvector - retrieves k-th eigenvector from basis vectors V.
451:   The argument eigi is the imaginary part of the corresponding eigenvalue.
452: */
453: PETSC_STATIC_INLINE PetscErrorCode BV_GetEigenvector(BV V,PetscInt k,PetscScalar eigi,Vec Vr,Vec Vi)
454: {

458: #if defined(PETSC_USE_COMPLEX)
459:   if (Vr) { BVCopyVec(V,k,Vr); }
460:   if (Vi) { VecSet(Vi,0.0); }
461: #else
462:   if (eigi > 0.0) { /* first value of conjugate pair */
463:     if (Vr) { BVCopyVec(V,k,Vr); }
464:     if (Vi) { BVCopyVec(V,k+1,Vi); }
465:   } else if (eigi < 0.0) { /* second value of conjugate pair */
466:     if (Vr) { BVCopyVec(V,k-1,Vr); }
467:     if (Vi) {
468:       BVCopyVec(V,k,Vi);
469:       VecScale(Vi,-1.0);
470:     }
471:   } else { /* real eigenvalue */
472:     if (Vr) { BVCopyVec(V,k,Vr); }
473:     if (Vi) { VecSet(Vi,0.0); }
474:   }
475: #endif
476:   return(0);
477: }

479: /*
480:    BV_OrthogonalizeColumn_Safe - this is intended for cases where we know that
481:    the resulting vector is going to be numerically zero, so normalization or
482:    iterative refinement may cause problems in parallel (collective operation
483:    not being called by all processes)
484: */
485: PETSC_STATIC_INLINE PetscErrorCode BV_OrthogonalizeColumn_Safe(BV bv,PetscInt j,PetscScalar *H,PetscReal *norm,PetscBool *lindep)
486: {
487:   PetscErrorCode     ierr;
488:   BVOrthogRefineType orthog_ref;

491:   PetscInfo1(bv,"Orthogonalizing column %D without refinement\n",j);
492:   orthog_ref     = bv->orthog_ref;
493:   bv->orthog_ref = BV_ORTHOG_REFINE_NEVER;  /* avoid refinement */
494:   BVOrthogonalizeColumn(bv,j,H,NULL,NULL);
495:   bv->orthog_ref = orthog_ref;  /* restore refinement setting */
496:   if (norm)   *norm  = 0.0;
497:   if (lindep) *lindep = PETSC_TRUE;
498:   return(0);
499: }

501: #if defined(PETSC_HAVE_CUDA)
502: #define BV_CleanCoefficients(a,b,c)   ((a)->cuda?BV_CleanCoefficients_CUDA:BV_CleanCoefficients_Default)((a),(b),(c))
503: #define BV_AddCoefficients(a,b,c,d)   ((a)->cuda?BV_AddCoefficients_CUDA:BV_AddCoefficients_Default)((a),(b),(c),(d))
504: #define BV_SetValue(a,b,c,d,e)        ((a)->cuda?BV_SetValue_CUDA:BV_SetValue_Default)((a),(b),(c),(d),(e))
505: #define BV_SquareSum(a,b,c,d)         ((a)->cuda?BV_SquareSum_CUDA:BV_SquareSum_Default)((a),(b),(c),(d))
506: #define BV_ApplySignature(a,b,c,d)    ((a)->cuda?BV_ApplySignature_CUDA:BV_ApplySignature_Default)((a),(b),(c),(d))
507: #define BV_SquareRoot(a,b,c,d)        ((a)->cuda?BV_SquareRoot_CUDA:BV_SquareRoot_Default)((a),(b),(c),(d))
508: #define BV_StoreCoefficients(a,b,c,d) ((a)->cuda?BV_StoreCoefficients_CUDA:BV_StoreCoefficients_Default)((a),(b),(c),(d))
509: #else
510: #define BV_CleanCoefficients(a,b,c)   BV_CleanCoefficients_Default((a),(b),(c))
511: #define BV_AddCoefficients(a,b,c,d)   BV_AddCoefficients_Default((a),(b),(c),(d))
512: #define BV_SetValue(a,b,c,d,e)        BV_SetValue_Default((a),(b),(c),(d),(e))
513: #define BV_SquareSum(a,b,c,d)         BV_SquareSum_Default((a),(b),(c),(d))
514: #define BV_ApplySignature(a,b,c,d)    BV_ApplySignature_Default((a),(b),(c),(d))
515: #define BV_SquareRoot(a,b,c,d)        BV_SquareRoot_Default((a),(b),(c),(d))
516: #define BV_StoreCoefficients(a,b,c,d) BV_StoreCoefficients_Default((a),(b),(c),(d))
517: #endif /* PETSC_HAVE_CUDA */

519: #endif