xref: /petsc/src/mat/impls/aij/mpi/mmaij.c (revision c8a8475e04bcaa43590892a5c3e60c6f87bc31f7)
1 /*$Id: mmaij.c,v 1.59 2001/08/07 03:02:49 balay Exp $*/
2 
3 /*
4    Support for the parallel AIJ matrix vector multiply
5 */
6 #include "src/mat/impls/aij/mpi/mpiaij.h"
7 #include "src/vec/vecimpl.h"
8 
9 #undef __FUNCT__
10 #define __FUNCT__ "MatSetUpMultiply_MPIAIJ"
11 int MatSetUpMultiply_MPIAIJ(Mat mat)
12 {
13   Mat_MPIAIJ         *aij = (Mat_MPIAIJ*)mat->data;
14   Mat_SeqAIJ         *B = (Mat_SeqAIJ*)(aij->B->data);
15   int                N = mat->N,i,j,*indices,*aj = B->j,ierr,ec = 0,*garray;
16   int                shift = B->indexshift;
17   IS                 from,to;
18   Vec                gvec;
19 #if defined (PETSC_USE_CTABLE)
20   PetscTable         gid1_lid1;
21   PetscTablePosition tpos;
22   int                gid,lid;
23 #endif
24 
25   PetscFunctionBegin;
26 
27 #if defined (PETSC_USE_CTABLE)
28   /* use a table - Mark Adams (this has not been tested with "shift") */
29   ierr = PetscTableCreate(aij->B->m,&gid1_lid1);CHKERRQ(ierr);
30   for (i=0; i<aij->B->m; i++) {
31     for (j=0; j<B->ilen[i]; j++) {
32       int data,gid1 = aj[B->i[i] + shift + j] + 1 + shift;
33       ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr);
34       if (!data) {
35         /* one based table */
36         ierr = PetscTableAdd(gid1_lid1,gid1,++ec);CHKERRQ(ierr);
37       }
38     }
39   }
40   /* form array of columns we need */
41   ierr = PetscMalloc((ec+1)*sizeof(int),&garray);CHKERRQ(ierr);
42   ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
43   while (tpos) {
44     ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr);
45     gid--;
46     lid--;
47     garray[lid] = gid;
48   }
49   ierr = PetscSortInt(ec,garray);CHKERRQ(ierr); /* sort, and rebuild */
50   ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
51   for (i=0; i<ec; i++) {
52     ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1);CHKERRQ(ierr);
53   }
54   /* compact out the extra columns in B */
55   for (i=0; i<aij->B->m; i++) {
56     for (j=0; j<B->ilen[i]; j++) {
57       int gid1 = aj[B->i[i] + shift + j] + 1 + shift;
58       ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr);
59       lid --;
60       aj[B->i[i] + shift + j]  = lid - shift;
61     }
62   }
63   aij->B->n = aij->B->N = ec;
64   ierr = PetscTableDelete(gid1_lid1);CHKERRQ(ierr);
65   /* Mark Adams */
66 #else
67   /* For the first stab we make an array as long as the number of columns */
68   /* mark those columns that are in aij->B */
69   ierr = PetscMalloc((N+1)*sizeof(int),&indices);CHKERRQ(ierr);
70   ierr = PetscMemzero(indices,N*sizeof(int));CHKERRQ(ierr);
71   for (i=0; i<aij->B->m; i++) {
72     for (j=0; j<B->ilen[i]; j++) {
73       if (!indices[aj[B->i[i] +shift + j] + shift]) ec++;
74       indices[aj[B->i[i] + shift + j] + shift] = 1;
75     }
76   }
77 
78   /* form array of columns we need */
79   ierr = PetscMalloc((ec+1)*sizeof(int),&garray);CHKERRQ(ierr);
80   ec = 0;
81   for (i=0; i<N; i++) {
82     if (indices[i]) garray[ec++] = i;
83   }
84 
85   /* make indices now point into garray */
86   for (i=0; i<ec; i++) {
87     indices[garray[i]] = i-shift;
88   }
89 
90   /* compact out the extra columns in B */
91   for (i=0; i<aij->B->m; i++) {
92     for (j=0; j<B->ilen[i]; j++) {
93       aj[B->i[i] + shift + j] = indices[aj[B->i[i] + shift + j]+shift];
94     }
95   }
96   aij->B->n = aij->B->N = ec;
97   ierr = PetscFree(indices);CHKERRQ(ierr);
98 #endif
99   /* create local vector that is used to scatter into */
100   ierr = VecCreateSeq(PETSC_COMM_SELF,ec,&aij->lvec);CHKERRQ(ierr);
101 
102   /* create two temporary Index sets for build scatter gather */
103   ierr = ISCreateGeneral(mat->comm,ec,garray,&from);CHKERRQ(ierr);
104   ierr = ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);CHKERRQ(ierr);
105 
106   /* create temporary global vector to generate scatter context */
107   /* this is inefficient, but otherwise we must do either
108      1) save garray until the first actual scatter when the vector is known or
109      2) have another way of generating a scatter context without a vector.*/
110   ierr = VecCreateMPI(mat->comm,mat->n,mat->N,&gvec);CHKERRQ(ierr);
111 
112   /* generate the scatter context */
113   ierr = VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);CHKERRQ(ierr);
114   PetscLogObjectParent(mat,aij->Mvctx);
115   PetscLogObjectParent(mat,aij->lvec);
116   PetscLogObjectParent(mat,from);
117   PetscLogObjectParent(mat,to);
118   aij->garray = garray;
119   PetscLogObjectMemory(mat,(ec+1)*sizeof(int));
120   ierr = ISDestroy(from);CHKERRQ(ierr);
121   ierr = ISDestroy(to);CHKERRQ(ierr);
122   ierr = VecDestroy(gvec);CHKERRQ(ierr);
123   PetscFunctionReturn(0);
124 }
125 
126 
127 #undef __FUNCT__
128 #define __FUNCT__ "DisAssemble_MPIAIJ"
129 /*
130      Takes the local part of an already assembled MPIAIJ matrix
131    and disassembles it. This is to allow new nonzeros into the matrix
132    that require more communication in the matrix vector multiply.
133    Thus certain data-structures must be rebuilt.
134 
135    Kind of slow! But that's what application programmers get when
136    they are sloppy.
137 */
138 int DisAssemble_MPIAIJ(Mat A)
139 {
140   Mat_MPIAIJ   *aij = (Mat_MPIAIJ*)A->data;
141   Mat          B = aij->B,Bnew;
142   Mat_SeqAIJ   *Baij = (Mat_SeqAIJ*)B->data;
143   int          ierr,i,j,m = B->m,n = A->N,col,ct = 0,*garray = aij->garray;
144   int          *nz,ec,shift = Baij->indexshift;
145   PetscScalar  v;
146 
147   PetscFunctionBegin;
148   /* free stuff related to matrix-vec multiply */
149   ierr = VecGetSize(aij->lvec,&ec);CHKERRQ(ierr); /* needed for PetscLogObjectMemory below */
150   ierr = VecDestroy(aij->lvec);CHKERRQ(ierr); aij->lvec = 0;
151   ierr = VecScatterDestroy(aij->Mvctx);CHKERRQ(ierr); aij->Mvctx = 0;
152   if (aij->colmap) {
153 #if defined (PETSC_USE_CTABLE)
154     ierr = PetscTableDelete(aij->colmap);CHKERRQ(ierr);
155     aij->colmap = 0;
156 #else
157     ierr = PetscFree(aij->colmap);CHKERRQ(ierr);
158     aij->colmap = 0;
159     PetscLogObjectMemory(A,-aij->B->n*sizeof(int));
160 #endif
161   }
162 
163   /* make sure that B is assembled so we can access its values */
164   ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
165   ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
166 
167   /* invent new B and copy stuff over */
168   ierr = PetscMalloc((m+1)*sizeof(int),&nz);CHKERRQ(ierr);
169   for (i=0; i<m; i++) {
170     nz[i] = Baij->i[i+1] - Baij->i[i];
171   }
172   ierr = MatCreateSeqAIJ(PETSC_COMM_SELF,m,n,0,nz,&Bnew);CHKERRQ(ierr);
173   ierr = PetscFree(nz);CHKERRQ(ierr);
174   for (i=0; i<m; i++) {
175     for (j=Baij->i[i]+shift; j<Baij->i[i+1]+shift; j++) {
176       col  = garray[Baij->j[ct]+shift];
177       v    = Baij->a[ct++];
178       ierr = MatSetValues(Bnew,1,&i,1,&col,&v,B->insertmode);CHKERRQ(ierr);
179     }
180   }
181   ierr = PetscFree(aij->garray);CHKERRQ(ierr);
182   aij->garray = 0;
183   PetscLogObjectMemory(A,-ec*sizeof(int));
184   ierr = MatDestroy(B);CHKERRQ(ierr);
185   PetscLogObjectParent(A,Bnew);
186   aij->B = Bnew;
187   A->was_assembled = PETSC_FALSE;
188   PetscFunctionReturn(0);
189 }
190 
191 /*      ugly stuff added for Glenn someday we should fix this up */
192 
193 static int *auglyrmapd = 0,*auglyrmapo = 0;  /* mapping from the local ordering to the "diagonal" and "off-diagonal"
194                                       parts of the local matrix */
195 static Vec auglydd = 0,auglyoo = 0;   /* work vectors used to scale the two parts of the local matrix */
196 
197 
198 #undef __FUNCT__
199 #define __FUNCT__ "MatMPIAIJDiagonalScaleLocalSetUp"
200 int MatMPIAIJDiagonalScaleLocalSetUp(Mat inA,Vec scale)
201 {
202   Mat_MPIAIJ  *ina = (Mat_MPIAIJ*) inA->data; /*access private part of matrix */
203   int          ierr,i,n,nt,cstart,cend,no,*garray = ina->garray,*lindices;
204   int          *r_rmapd,*r_rmapo;
205 
206   PetscFunctionBegin;
207   ierr = MatGetOwnershipRange(inA,&cstart,&cend);CHKERRQ(ierr);
208   ierr = MatGetSize(ina->A,PETSC_NULL,&n);CHKERRQ(ierr);
209   ierr = PetscMalloc((inA->mapping->n+1)*sizeof(int),&r_rmapd);CHKERRQ(ierr);
210   ierr = PetscMemzero(r_rmapd,inA->mapping->n*sizeof(int));CHKERRQ(ierr);
211   nt   = 0;
212   for (i=0; i<inA->mapping->n; i++) {
213     if (inA->mapping->indices[i] >= cstart && inA->mapping->indices[i] < cend) {
214       nt++;
215       r_rmapd[i] = inA->mapping->indices[i] + 1;
216     }
217   }
218   if (nt != n) SETERRQ2(1,"Hmm nt %d n %d",nt,n);
219   ierr = PetscMalloc((n+1)*sizeof(int),&auglyrmapd);CHKERRQ(ierr);
220   for (i=0; i<inA->mapping->n; i++) {
221     if (r_rmapd[i]){
222       auglyrmapd[(r_rmapd[i]-1)-cstart] = i;
223     }
224   }
225   ierr = PetscFree(r_rmapd);CHKERRQ(ierr);
226   ierr = VecCreateSeq(PETSC_COMM_SELF,n,&auglydd);CHKERRQ(ierr);
227 
228   ierr = PetscMalloc((inA->N+1)*sizeof(int),&lindices);CHKERRQ(ierr);
229   ierr = PetscMemzero(lindices,inA->N*sizeof(int));CHKERRQ(ierr);
230   for (i=0; i<ina->B->n; i++) {
231     lindices[garray[i]] = i+1;
232   }
233   no   = inA->mapping->n - nt;
234   ierr = PetscMalloc((inA->mapping->n+1)*sizeof(int),&r_rmapo);CHKERRQ(ierr);
235   ierr = PetscMemzero(r_rmapo,inA->mapping->n*sizeof(int));CHKERRQ(ierr);
236   nt   = 0;
237   for (i=0; i<inA->mapping->n; i++) {
238     if (lindices[inA->mapping->indices[i]]) {
239       nt++;
240       r_rmapo[i] = lindices[inA->mapping->indices[i]];
241     }
242   }
243   if (nt > no) SETERRQ2(1,"Hmm nt %d no %d",nt,n);
244   ierr = PetscFree(lindices);CHKERRQ(ierr);
245   ierr = PetscMalloc((nt+1)*sizeof(int),&auglyrmapo);CHKERRQ(ierr);
246   for (i=0; i<inA->mapping->n; i++) {
247     if (r_rmapo[i]){
248       auglyrmapo[(r_rmapo[i]-1)] = i;
249     }
250   }
251   ierr = PetscFree(r_rmapo);CHKERRQ(ierr);
252   ierr = VecCreateSeq(PETSC_COMM_SELF,nt,&auglyoo);CHKERRQ(ierr);
253 
254   PetscFunctionReturn(0);
255 }
256 
257 #undef __FUNCT__
258 #define __FUNCT__ "MatMPIAIJDiagonalScaleLocal"
259 int MatMPIAIJDiagonalScaleLocal(Mat A,Vec scale)
260 {
261   Mat_MPIAIJ  *a = (Mat_MPIAIJ*) A->data; /*access private part of matrix */
262   int         ierr,n,i;
263   PetscScalar *d,*o,*s;
264 
265   PetscFunctionBegin;
266   if (!auglyrmapd) {
267     ierr = MatMPIAIJDiagonalScaleLocalSetUp(A,scale);CHKERRQ(ierr);
268   }
269 
270   ierr = VecGetArray(scale,&s);CHKERRQ(ierr);
271 
272   ierr = VecGetLocalSize(auglydd,&n);CHKERRQ(ierr);
273   ierr = VecGetArray(auglydd,&d);CHKERRQ(ierr);
274   for (i=0; i<n; i++) {
275     d[i] = s[auglyrmapd[i]]; /* copy "diagonal" (true local) portion of scale into dd vector */
276   }
277   ierr = VecRestoreArray(auglydd,&d);CHKERRQ(ierr);
278   /* column scale "diagonal" portion of local matrix */
279   ierr = MatDiagonalScale(a->A,PETSC_NULL,auglydd);CHKERRQ(ierr);
280 
281   ierr = VecGetLocalSize(auglyoo,&n);CHKERRQ(ierr);
282   ierr = VecGetArray(auglyoo,&o);CHKERRQ(ierr);
283   for (i=0; i<n; i++) {
284     o[i] = s[auglyrmapo[i]]; /* copy "off-diagonal" portion of scale into oo vector */
285   }
286   ierr = VecRestoreArray(scale,&s);CHKERRQ(ierr);
287   ierr = VecRestoreArray(auglyoo,&o);CHKERRQ(ierr);
288   /* column scale "off-diagonal" portion of local matrix */
289   ierr = MatDiagonalScale(a->B,PETSC_NULL,auglyoo);CHKERRQ(ierr);
290 
291   PetscFunctionReturn(0);
292 }
293 
294 
295 
296 
297