xref: /petsc/src/mat/impls/lrc/lrc.c (revision 2f6eced2a19e978d64f27de66fbc6c26cc5c7934)
1 
2 #include <petsc/private/matimpl.h>          /*I "petscmat.h" I*/
3 
4 typedef struct {
5   Mat         A;           /* sparse matrix */
6   Mat         U,V;         /* dense tall-skinny matrices */
7   Vec         c;           /* sequential vector containing the diagonal of C */
8   Vec         work1,work2; /* sequential vectors that hold partial products */
9   PetscMPIInt nwork;       /* length of work vectors */
10   Vec         xl,yl;       /* auxiliary sequential vectors for matmult operation */
11 } Mat_LRC;
12 
13 
14 PetscErrorCode MatMult_LRC(Mat N,Vec x,Vec y)
15 {
16   Mat_LRC           *Na = (Mat_LRC*)N->data;
17   PetscErrorCode    ierr;
18   PetscScalar       *w1,*w2;
19   const PetscScalar *a;
20 
21   PetscFunctionBegin;
22   ierr = VecGetArrayRead(x,&a);CHKERRQ(ierr);
23   ierr = VecPlaceArray(Na->xl,a);CHKERRQ(ierr);
24   ierr = VecGetLocalVector(y,Na->yl);CHKERRQ(ierr);
25 
26   /* multiply the local part of V with the local part of x */
27 #if defined(PETSC_USE_COMPLEX)
28   ierr = MatMultHermitianTranspose(Na->V,Na->xl,Na->work1);CHKERRQ(ierr);
29 #else
30   ierr = MatMultTranspose(Na->V,Na->xl,Na->work1);CHKERRQ(ierr);
31 #endif
32 
33   /* form the sum of all the local multiplies: this is work2 = V'*x =
34      sum_{all processors} work1 */
35   ierr = VecGetArray(Na->work1,&w1);CHKERRQ(ierr);
36   ierr = VecGetArray(Na->work2,&w2);CHKERRQ(ierr);
37   ierr = MPIU_Allreduce(w1,w2,Na->nwork,MPIU_SCALAR,MPIU_SUM,PetscObjectComm((PetscObject)N));CHKERRQ(ierr);
38   ierr = VecRestoreArray(Na->work1,&w1);CHKERRQ(ierr);
39   ierr = VecRestoreArray(Na->work2,&w2);CHKERRQ(ierr);
40 
41   if (Na->c) {  /* work2 = C*work2 */
42     ierr = VecPointwiseMult(Na->work2,Na->c,Na->work2);CHKERRQ(ierr);
43   }
44 
45   if (Na->A) {
46     /* form y = A*x */
47     ierr = MatMult(Na->A,x,y);CHKERRQ(ierr);
48     /* multiply-add y = y + U*work2 */
49     ierr = MatMultAdd(Na->U,Na->work2,Na->yl,Na->yl);CHKERRQ(ierr);
50   } else {
51     /* multiply y = U*work2 */
52     ierr = MatMult(Na->U,Na->work2,Na->yl);CHKERRQ(ierr);
53   }
54 
55   ierr = VecRestoreArrayRead(x,&a);CHKERRQ(ierr);
56   ierr = VecResetArray(Na->xl);CHKERRQ(ierr);
57   ierr = VecRestoreLocalVector(y,Na->yl);CHKERRQ(ierr);
58   PetscFunctionReturn(0);
59 }
60 
61 PetscErrorCode MatDestroy_LRC(Mat N)
62 {
63   Mat_LRC        *Na = (Mat_LRC*)N->data;
64   PetscErrorCode ierr;
65 
66   PetscFunctionBegin;
67   ierr = MatDestroy(&Na->A);CHKERRQ(ierr);
68   ierr = MatDestroy(&Na->U);CHKERRQ(ierr);
69   ierr = MatDestroy(&Na->V);CHKERRQ(ierr);
70   ierr = VecDestroy(&Na->c);CHKERRQ(ierr);
71   ierr = VecDestroy(&Na->work1);CHKERRQ(ierr);
72   ierr = VecDestroy(&Na->work2);CHKERRQ(ierr);
73   ierr = VecDestroy(&Na->xl);CHKERRQ(ierr);
74   ierr = VecDestroy(&Na->yl);CHKERRQ(ierr);
75   ierr = PetscFree(N->data);CHKERRQ(ierr);
76   ierr = PetscObjectComposeFunction((PetscObject)N,"MatLRCGetMats_C",0);CHKERRQ(ierr);
77   PetscFunctionReturn(0);
78 }
79 
80 PetscErrorCode MatLRCGetMats_LRC(Mat N,Mat *A,Mat *U,Vec *c,Mat *V)
81 {
82   Mat_LRC *Na = (Mat_LRC*)N->data;
83 
84   PetscFunctionBegin;
85   if (A) *A = Na->A;
86   if (U) *U = Na->U;
87   if (c) *c = Na->c;
88   if (V) *V = Na->V;
89   PetscFunctionReturn(0);
90 }
91 
92 /*@
93    MatLRCGetMats - Returns the constituents of an LRC matrix
94 
95    Collective on Mat
96 
97    Input Parameter:
98 .  N    - matrix of type LRC
99 
100    Output Parameters:
101 +  A    - the (sparse) matrix
102 .  U, V - two dense rectangular (tall and skinny) matrices
103 -  c    - a sequential vector containing the diagonal of C
104 
105    Note:
106    The returned matrices need not be destroyed by the caller.
107 
108    Level: intermediate
109 
110 .seealso: MatCreateLRC()
111 @*/
112 PetscErrorCode  MatLRCGetMats(Mat N,Mat *A,Mat *U,Vec *c,Mat *V)
113 {
114   PetscErrorCode ierr;
115 
116   PetscFunctionBegin;
117   ierr = PetscUseMethod(N,"MatLRCGetMats_C",(Mat,Mat*,Mat*,Vec*,Mat*),(N,A,U,c,V));CHKERRQ(ierr);
118   PetscFunctionReturn(0);
119 }
120 
121 /*@
122    MatCreateLRC - Creates a new matrix object that behaves like A + U*C*V'
123 
124    Collective on Mat
125 
126    Input Parameters:
127 +  A    - the (sparse) matrix (can be NULL)
128 .  U, V - two dense rectangular (tall and skinny) matrices
129 -  c    - a sequential vector containing the diagonal of C (can be NULL)
130 
131    Output Parameter:
132 .  N    - the matrix that represents A + U*C*V'
133 
134    Notes:
135    The matrix A + U*C*V' is not formed! Rather the new matrix
136    object performs the matrix-vector product by first multiplying by
137    A and then adding the other term.
138 
139    C is a diagonal matrix (represented as a vector) of order k,
140    where k is the number of columns of both U and V.
141 
142    If A is NULL then the new object behaves like a low-rank matrix U*C*V'.
143 
144    Use V=U (or V=NULL) for a symmetric low-rank correction, A + U*C*U'.
145 
146    If c is NULL then the low-rank correction is just U*V'.
147 
148    Level: intermediate
149 
150 .seealso: MatLRCGetMats()
151 @*/
152 PetscErrorCode MatCreateLRC(Mat A,Mat U,Vec c,Mat V,Mat *N)
153 {
154   PetscErrorCode ierr;
155   PetscBool      match;
156   PetscInt       m,n,k,m1,n1,k1;
157   Mat_LRC        *Na;
158 
159   PetscFunctionBegin;
160   if (A) PetscValidHeaderSpecific(A,MAT_CLASSID,1);
161   PetscValidHeaderSpecific(U,MAT_CLASSID,2);
162   if (c) PetscValidHeaderSpecific(c,VEC_CLASSID,3);
163   if (V) PetscValidHeaderSpecific(V,MAT_CLASSID,4);
164   else V=U;
165   if (A) PetscCheckSameComm(A,1,U,2);
166   PetscCheckSameComm(U,2,V,4);
167 
168   ierr = PetscObjectTypeCompareAny((PetscObject)U,&match,MATSEQDENSE,MATMPIDENSE,"");CHKERRQ(ierr);
169   if (!match) SETERRQ(PetscObjectComm((PetscObject)U),PETSC_ERR_SUP,"Matrix U must be of type dense");
170   if (V) {
171     ierr = PetscObjectTypeCompareAny((PetscObject)V,&match,MATSEQDENSE,MATMPIDENSE,"");CHKERRQ(ierr);
172     if (!match) SETERRQ(PetscObjectComm((PetscObject)V),PETSC_ERR_SUP,"Matrix V must be of type dense");
173   }
174 
175   ierr = MatGetSize(U,NULL,&k);CHKERRQ(ierr);
176   ierr = MatGetSize(V,NULL,&k1);CHKERRQ(ierr);
177   if (k!=k1) SETERRQ2(PetscObjectComm((PetscObject)U),PETSC_ERR_ARG_INCOMP,"U and V have different number of columns (%D vs %D)",k,k1);
178   ierr = MatGetLocalSize(U,&m,NULL);CHKERRQ(ierr);
179   ierr = MatGetLocalSize(V,&n,NULL);CHKERRQ(ierr);
180   if (A) {
181     ierr = MatGetLocalSize(A,&m1,&n1);CHKERRQ(ierr);
182     if (m!=m1) SETERRQ2(PetscObjectComm((PetscObject)U),PETSC_ERR_ARG_INCOMP,"Local dimensions of U %D and A %D do not match",m,m1);
183     if (n!=n1) SETERRQ2(PetscObjectComm((PetscObject)V),PETSC_ERR_ARG_INCOMP,"Local dimensions of V %D and A %D do not match",n,n1);
184   }
185   if (c) {
186     ierr = VecGetSize(c,&k1);CHKERRQ(ierr);
187     if (k!=k1) SETERRQ2(PetscObjectComm((PetscObject)c),PETSC_ERR_ARG_INCOMP,"The length of c %D does not match the number of columns of U and V (%D)",k1,k);
188     ierr = VecGetLocalSize(c,&k1);CHKERRQ(ierr);
189     if (k!=k1) SETERRQ(PetscObjectComm((PetscObject)c),PETSC_ERR_ARG_INCOMP,"c must be a sequential vector");
190   }
191 
192   ierr = MatCreate(PetscObjectComm((PetscObject)U),N);CHKERRQ(ierr);
193   ierr = MatSetSizes(*N,m,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
194   ierr = PetscObjectChangeTypeName((PetscObject)*N,MATLRC);CHKERRQ(ierr);
195 
196   ierr       = PetscNewLog(*N,&Na);CHKERRQ(ierr);
197   (*N)->data = (void*)Na;
198   Na->A      = A;
199   Na->c      = c;
200 
201   ierr = MatDenseGetLocalMatrix(U,&Na->U);CHKERRQ(ierr);
202   ierr = MatDenseGetLocalMatrix(V,&Na->V);CHKERRQ(ierr);
203   if (A) { ierr = PetscObjectReference((PetscObject)A);CHKERRQ(ierr); }
204   ierr = PetscObjectReference((PetscObject)Na->U);CHKERRQ(ierr);
205   ierr = PetscObjectReference((PetscObject)Na->V);CHKERRQ(ierr);
206   if (c) { ierr = PetscObjectReference((PetscObject)c);CHKERRQ(ierr); }
207 
208   ierr = VecCreateSeq(PETSC_COMM_SELF,U->cmap->N,&Na->work1);CHKERRQ(ierr);
209   ierr = VecDuplicate(Na->work1,&Na->work2);CHKERRQ(ierr);
210   ierr = PetscMPIIntCast(U->cmap->N,&Na->nwork);CHKERRQ(ierr);
211 
212   ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,V->rmap->n,NULL,&Na->xl);CHKERRQ(ierr);
213   ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,U->rmap->n,NULL,&Na->yl);CHKERRQ(ierr);
214 
215   (*N)->ops->destroy = MatDestroy_LRC;
216   (*N)->ops->mult    = MatMult_LRC;
217   (*N)->assembled    = PETSC_TRUE;
218   (*N)->preallocated = PETSC_TRUE;
219   (*N)->cmap->N      = V->rmap->N;
220   (*N)->rmap->N      = U->rmap->N;
221   (*N)->cmap->n      = V->rmap->n;
222   (*N)->rmap->n      = U->rmap->n;
223 
224   ierr = PetscObjectComposeFunction((PetscObject)(*N),"MatLRCGetMats_C",MatLRCGetMats_LRC);CHKERRQ(ierr);
225   PetscFunctionReturn(0);
226 }
227 
228