xref: /petsc/src/mat/impls/aij/seq/matptap.c (revision 503c0ea9b45bcfbcebbb1ea5341243bbc69f0bea)
1 
2 /*
3   Defines projective product routines where A is a SeqAIJ matrix
4           C = P^T * A * P
5 */
6 
7 #include <../src/mat/impls/aij/seq/aij.h>   /*I "petscmat.h" I*/
8 #include <../src/mat/utils/freespace.h>
9 #include <petscbt.h>
10 #include <petsctime.h>
11 
12 #if defined(PETSC_HAVE_HYPRE)
13 PETSC_INTERN PetscErrorCode MatPtAPSymbolic_AIJ_AIJ_wHYPRE(Mat,Mat,PetscReal,Mat);
14 #endif
15 
16 PetscErrorCode MatProductSymbolic_PtAP_SeqAIJ_SeqAIJ(Mat C)
17 {
18   Mat_Product         *product = C->product;
19   Mat                 A=product->A,P=product->B;
20   MatProductAlgorithm alg=product->alg;
21   PetscReal           fill=product->fill;
22   PetscBool           flg;
23   Mat                 Pt;
24 
25   PetscFunctionBegin;
26   /* "scalable" */
27   PetscCall(PetscStrcmp(alg,"scalable",&flg));
28   if (flg) {
29     PetscCall(MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(A,P,fill,C));
30     C->ops->productnumeric = MatProductNumeric_PtAP;
31     PetscFunctionReturn(0);
32   }
33 
34   /* "rap" */
35   PetscCall(PetscStrcmp(alg,"rap",&flg));
36   if (flg) {
37     Mat_MatTransMatMult *atb;
38 
39     PetscCall(PetscNew(&atb));
40     PetscCall(MatTranspose_SeqAIJ(P,MAT_INITIAL_MATRIX,&Pt));
41     PetscCall(MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ(Pt,A,P,fill,C));
42 
43     atb->At                = Pt;
44     atb->data              = C->product->data;
45     atb->destroy           = C->product->destroy;
46     C->product->data       = atb;
47     C->product->destroy    = MatDestroy_SeqAIJ_MatTransMatMult;
48     C->ops->ptapnumeric    = MatPtAPNumeric_SeqAIJ_SeqAIJ;
49     C->ops->productnumeric = MatProductNumeric_PtAP;
50     PetscFunctionReturn(0);
51   }
52 
53   /* hypre */
54 #if defined(PETSC_HAVE_HYPRE)
55   PetscCall(PetscStrcmp(alg,"hypre",&flg));
56   if (flg) {
57     PetscCall(MatPtAPSymbolic_AIJ_AIJ_wHYPRE(A,P,fill,C));
58     PetscFunctionReturn(0);
59   }
60 #endif
61 
62   SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatProductType is not supported");
63 }
64 
65 PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(Mat A,Mat P,PetscReal fill,Mat C)
66 {
67   PetscFreeSpaceList free_space=NULL,current_space=NULL;
68   Mat_SeqAIJ         *a        = (Mat_SeqAIJ*)A->data,*p = (Mat_SeqAIJ*)P->data,*c;
69   PetscInt           *pti,*ptj,*ptJ,*ai=a->i,*aj=a->j,*ajj,*pi=p->i,*pj=p->j,*pjj;
70   PetscInt           *ci,*cj,*ptadenserow,*ptasparserow,*ptaj,nspacedouble=0;
71   PetscInt           an=A->cmap->N,am=A->rmap->N,pn=P->cmap->N,pm=P->rmap->N;
72   PetscInt           i,j,k,ptnzi,arow,anzj,ptanzi,prow,pnzj,cnzi,nlnk,*lnk;
73   MatScalar          *ca;
74   PetscBT            lnkbt;
75   PetscReal          afill;
76 
77   PetscFunctionBegin;
78   /* Get ij structure of P^T */
79   PetscCall(MatGetSymbolicTranspose_SeqAIJ(P,&pti,&ptj));
80   ptJ  = ptj;
81 
82   /* Allocate ci array, arrays for fill computation and */
83   /* free space for accumulating nonzero column info */
84   PetscCall(PetscMalloc1(pn+1,&ci));
85   ci[0] = 0;
86 
87   PetscCall(PetscCalloc1(2*an+1,&ptadenserow));
88   ptasparserow = ptadenserow  + an;
89 
90   /* create and initialize a linked list */
91   nlnk = pn+1;
92   PetscCall(PetscLLCreate(pn,pn,nlnk,lnk,lnkbt));
93 
94   /* Set initial free space to be fill*(nnz(A)+ nnz(P)) */
95   PetscCall(PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ai[am],pi[pm])),&free_space));
96   current_space = free_space;
97 
98   /* Determine symbolic info for each row of C: */
99   for (i=0; i<pn; i++) {
100     ptnzi  = pti[i+1] - pti[i];
101     ptanzi = 0;
102     /* Determine symbolic row of PtA: */
103     for (j=0; j<ptnzi; j++) {
104       arow = *ptJ++;
105       anzj = ai[arow+1] - ai[arow];
106       ajj  = aj + ai[arow];
107       for (k=0; k<anzj; k++) {
108         if (!ptadenserow[ajj[k]]) {
109           ptadenserow[ajj[k]]    = -1;
110           ptasparserow[ptanzi++] = ajj[k];
111         }
112       }
113     }
114     /* Using symbolic info for row of PtA, determine symbolic info for row of C: */
115     ptaj = ptasparserow;
116     cnzi = 0;
117     for (j=0; j<ptanzi; j++) {
118       prow = *ptaj++;
119       pnzj = pi[prow+1] - pi[prow];
120       pjj  = pj + pi[prow];
121       /* add non-zero cols of P into the sorted linked list lnk */
122       PetscCall(PetscLLAddSorted(pnzj,pjj,pn,&nlnk,lnk,lnkbt));
123       cnzi += nlnk;
124     }
125 
126     /* If free space is not available, make more free space */
127     /* Double the amount of total space in the list */
128     if (current_space->local_remaining<cnzi) {
129       PetscCall(PetscFreeSpaceGet(PetscIntSumTruncate(cnzi,current_space->total_array_size),&current_space));
130       nspacedouble++;
131     }
132 
133     /* Copy data into free space, and zero out denserows */
134     PetscCall(PetscLLClean(pn,pn,cnzi,lnk,current_space->array,lnkbt));
135 
136     current_space->array           += cnzi;
137     current_space->local_used      += cnzi;
138     current_space->local_remaining -= cnzi;
139 
140     for (j=0; j<ptanzi; j++) ptadenserow[ptasparserow[j]] = 0;
141 
142     /* Aside: Perhaps we should save the pta info for the numerical factorization. */
143     /*        For now, we will recompute what is needed. */
144     ci[i+1] = ci[i] + cnzi;
145   }
146   /* nnz is now stored in ci[ptm], column indices are in the list of free space */
147   /* Allocate space for cj, initialize cj, and */
148   /* destroy list of free space and other temporary array(s) */
149   PetscCall(PetscMalloc1(ci[pn]+1,&cj));
150   PetscCall(PetscFreeSpaceContiguous(&free_space,cj));
151   PetscCall(PetscFree(ptadenserow));
152   PetscCall(PetscLLDestroy(lnk,lnkbt));
153 
154   PetscCall(PetscCalloc1(ci[pn]+1,&ca));
155 
156   /* put together the new matrix */
157   PetscCall(MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),pn,pn,ci,cj,ca,((PetscObject)A)->type_name,C));
158   PetscCall(MatSetBlockSizes(C,PetscAbs(P->cmap->bs),PetscAbs(P->cmap->bs)));
159 
160   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
161   /* Since these are PETSc arrays, change flags to free them as necessary. */
162   c          = (Mat_SeqAIJ*)((C)->data);
163   c->free_a  = PETSC_TRUE;
164   c->free_ij = PETSC_TRUE;
165   c->nonew   = 0;
166 
167   C->ops->ptapnumeric = MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy;
168 
169   /* set MatInfo */
170   afill = (PetscReal)ci[pn]/(ai[am]+pi[pm] + 1.e-5);
171   if (afill < 1.0) afill = 1.0;
172   C->info.mallocs           = nspacedouble;
173   C->info.fill_ratio_given  = fill;
174   C->info.fill_ratio_needed = afill;
175 
176   /* Clean up. */
177   PetscCall(MatRestoreSymbolicTranspose_SeqAIJ(P,&pti,&ptj));
178 #if defined(PETSC_USE_INFO)
179   if (ci[pn] != 0) {
180     PetscCall(PetscInfo(C,"Reallocs %" PetscInt_FMT "; Fill ratio: given %g needed %g.\n",nspacedouble,(double)fill,(double)afill));
181     PetscCall(PetscInfo(C,"Use MatPtAP(A,P,MatReuse,%g,&C) for best performance.\n",(double)afill));
182   } else {
183     PetscCall(PetscInfo(C,"Empty matrix product\n"));
184   }
185 #endif
186   PetscFunctionReturn(0);
187 }
188 
189 PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy(Mat A,Mat P,Mat C)
190 {
191   Mat_SeqAIJ     *a = (Mat_SeqAIJ*) A->data;
192   Mat_SeqAIJ     *p = (Mat_SeqAIJ*) P->data;
193   Mat_SeqAIJ     *c = (Mat_SeqAIJ*) C->data;
194   PetscInt       *ai=a->i,*aj=a->j,*apj,*apjdense,*pi=p->i,*pj=p->j,*pJ=p->j,*pjj;
195   PetscInt       *ci=c->i,*cj=c->j,*cjj;
196   PetscInt       am =A->rmap->N,cn=C->cmap->N,cm=C->rmap->N;
197   PetscInt       i,j,k,anzi,pnzi,apnzj,nextap,pnzj,prow,crow;
198   MatScalar      *aa,*apa,*pa,*pA,*paj,*ca,*caj;
199 
200   PetscFunctionBegin;
201   /* Allocate temporary array for storage of one row of A*P (cn: non-scalable) */
202   PetscCall(PetscCalloc2(cn,&apa,cn,&apjdense));
203   PetscCall(PetscMalloc1(cn,&apj));
204   /* trigger CPU copies if needed and flag CPU mask for C */
205 #if defined(PETSC_HAVE_DEVICE)
206   {
207     const PetscScalar *dummy;
208     PetscCall(MatSeqAIJGetArrayRead(A,&dummy));
209     PetscCall(MatSeqAIJRestoreArrayRead(A,&dummy));
210     PetscCall(MatSeqAIJGetArrayRead(P,&dummy));
211     PetscCall(MatSeqAIJRestoreArrayRead(P,&dummy));
212     if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU;
213   }
214 #endif
215   aa = a->a;
216   pa = p->a;
217   pA = p->a;
218   ca = c->a;
219 
220   /* Clear old values in C */
221   PetscCall(PetscArrayzero(ca,ci[cm]));
222 
223   for (i=0; i<am; i++) {
224     /* Form sparse row of A*P */
225     anzi  = ai[i+1] - ai[i];
226     apnzj = 0;
227     for (j=0; j<anzi; j++) {
228       prow = *aj++;
229       pnzj = pi[prow+1] - pi[prow];
230       pjj  = pj + pi[prow];
231       paj  = pa + pi[prow];
232       for (k=0; k<pnzj; k++) {
233         if (!apjdense[pjj[k]]) {
234           apjdense[pjj[k]] = -1;
235           apj[apnzj++]     = pjj[k];
236         }
237         apa[pjj[k]] += (*aa)*paj[k];
238       }
239       PetscCall(PetscLogFlops(2.0*pnzj));
240       aa++;
241     }
242 
243     /* Sort the j index array for quick sparse axpy. */
244     /* Note: a array does not need sorting as it is in dense storage locations. */
245     PetscCall(PetscSortInt(apnzj,apj));
246 
247     /* Compute P^T*A*P using outer product (P^T)[:,j]*(A*P)[j,:]. */
248     pnzi = pi[i+1] - pi[i];
249     for (j=0; j<pnzi; j++) {
250       nextap = 0;
251       crow   = *pJ++;
252       cjj    = cj + ci[crow];
253       caj    = ca + ci[crow];
254       /* Perform sparse axpy operation.  Note cjj includes apj. */
255       for (k=0; nextap<apnzj; k++) {
256         PetscAssert(k < ci[crow+1] - ci[crow],PETSC_COMM_SELF,PETSC_ERR_PLIB,"k too large k %" PetscInt_FMT ", crow %" PetscInt_FMT,k,crow);
257         if (cjj[k]==apj[nextap]) {
258           caj[k] += (*pA)*apa[apj[nextap++]];
259         }
260       }
261       PetscCall(PetscLogFlops(2.0*apnzj));
262       pA++;
263     }
264 
265     /* Zero the current row info for A*P */
266     for (j=0; j<apnzj; j++) {
267       apa[apj[j]]      = 0.;
268       apjdense[apj[j]] = 0;
269     }
270   }
271 
272   /* Assemble the final matrix and clean up */
273   PetscCall(MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY));
274   PetscCall(MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY));
275 
276   PetscCall(PetscFree2(apa,apjdense));
277   PetscCall(PetscFree(apj));
278   PetscFunctionReturn(0);
279 }
280 
281 PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat A,Mat P,Mat C)
282 {
283   Mat_MatTransMatMult *atb;
284 
285   PetscFunctionBegin;
286   MatCheckProduct(C,3);
287   atb  = (Mat_MatTransMatMult*)C->product->data;
288   PetscCheck(atb,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Missing data structure");
289   PetscCall(MatTranspose_SeqAIJ(P,MAT_REUSE_MATRIX,&atb->At));
290   PetscCheck(C->ops->matmultnumeric,PetscObjectComm((PetscObject)C),PETSC_ERR_PLIB,"Missing numeric operation");
291   /* when using rap, MatMatMatMultSymbolic used a different data */
292   if (atb->data) C->product->data = atb->data;
293   PetscCall((*C->ops->matmatmultnumeric)(atb->At,A,P,C));
294   C->product->data = atb;
295   PetscFunctionReturn(0);
296 }
297