Actual source code: mmaij.c
petsc-3.3-p7 2013-05-11
2: /*
3: Support for the parallel AIJ matrix vector multiply
4: */
5: #include <../src/mat/impls/aij/mpi/mpiaij.h>
9: PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat mat)
10: {
11: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
12: Mat_SeqAIJ *B = (Mat_SeqAIJ*)(aij->B->data);
13: PetscErrorCode ierr;
14: PetscInt i,j,*aj = B->j,ec = 0,*garray;
15: IS from,to;
16: Vec gvec;
17: #if defined (PETSC_USE_CTABLE)
18: PetscTable gid1_lid1;
19: PetscTablePosition tpos;
20: PetscInt gid,lid;
21: #else
22: PetscInt N = mat->cmap->N,*indices;
23: #endif
27: #if defined (PETSC_USE_CTABLE)
28: /* use a table */
29: PetscTableCreate(aij->B->rmap->n,mat->cmap->N+1,&gid1_lid1);
30: for (i=0; i<aij->B->rmap->n; i++) {
31: for (j=0; j<B->ilen[i]; j++) {
32: PetscInt data,gid1 = aj[B->i[i] + j] + 1;
33: PetscTableFind(gid1_lid1,gid1,&data);
34: if (!data) {
35: /* one based table */
36: PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);
37: }
38: }
39: }
40: /* form array of columns we need */
41: PetscMalloc((ec+1)*sizeof(PetscInt),&garray);
42: PetscTableGetHeadPosition(gid1_lid1,&tpos);
43: while (tpos) {
44: PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);
45: gid--;
46: lid--;
47: garray[lid] = gid;
48: }
49: PetscSortInt(ec,garray); /* sort, and rebuild */
50: PetscTableRemoveAll(gid1_lid1);
51: for (i=0; i<ec; i++) {
52: PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);
53: }
54: /* compact out the extra columns in B */
55: for (i=0; i<aij->B->rmap->n; i++) {
56: for (j=0; j<B->ilen[i]; j++) {
57: PetscInt gid1 = aj[B->i[i] + j] + 1;
58: PetscTableFind(gid1_lid1,gid1,&lid);
59: lid --;
60: aj[B->i[i] + j] = lid;
61: }
62: }
63: aij->B->cmap->n = aij->B->cmap->N = ec;
64: PetscLayoutSetUp((aij->B->cmap));
65: PetscTableDestroy(&gid1_lid1);
66: #else
67: /* Make an array as long as the number of columns */
68: /* mark those columns that are in aij->B */
69: PetscMalloc((N+1)*sizeof(PetscInt),&indices);
70: PetscMemzero(indices,N*sizeof(PetscInt));
71: for (i=0; i<aij->B->rmap->n; i++) {
72: for (j=0; j<B->ilen[i]; j++) {
73: if (!indices[aj[B->i[i] + j] ]) ec++;
74: indices[aj[B->i[i] + j] ] = 1;
75: }
76: }
78: /* form array of columns we need */
79: PetscMalloc((ec+1)*sizeof(PetscInt),&garray);
80: ec = 0;
81: for (i=0; i<N; i++) {
82: if (indices[i]) garray[ec++] = i;
83: }
85: /* make indices now point into garray */
86: for (i=0; i<ec; i++) {
87: indices[garray[i]] = i;
88: }
90: /* compact out the extra columns in B */
91: for (i=0; i<aij->B->rmap->n; i++) {
92: for (j=0; j<B->ilen[i]; j++) {
93: aj[B->i[i] + j] = indices[aj[B->i[i] + j]];
94: }
95: }
96: aij->B->cmap->n = aij->B->cmap->N = ec;
97: PetscLayoutSetUp((aij->B->cmap));
98: PetscFree(indices);
99: #endif
100: /* create local vector that is used to scatter into */
101: VecCreateSeq(PETSC_COMM_SELF,ec,&aij->lvec);
103: /* create two temporary Index sets for build scatter gather */
104: ISCreateGeneral(((PetscObject)mat)->comm,ec,garray,PETSC_COPY_VALUES,&from);
106: ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);
108: /* create temporary global vector to generate scatter context */
109: /* This does not allocate the array's memory so is efficient */
110: VecCreateMPIWithArray(((PetscObject)mat)->comm,1,mat->cmap->n,mat->cmap->N,PETSC_NULL,&gvec);
112: /* generate the scatter context */
113: VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);
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(PetscInt));
120: ISDestroy(&from);
121: ISDestroy(&to);
122: VecDestroy(&gvec);
123: return(0);
124: }
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.
135: Kind of slow! But that's what application programmers get when
136: they are sloppy.
137: */
138: PetscErrorCode MatDisAssemble_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;
144: PetscInt i,j,m = B->rmap->n,n = A->cmap->N,col,ct = 0,*garray = aij->garray,*nz,ec;
145: PetscScalar v;
148: /* free stuff related to matrix-vec multiply */
149: VecGetSize(aij->lvec,&ec); /* needed for PetscLogObjectMemory below */
150: VecDestroy(&aij->lvec);
151: VecScatterDestroy(&aij->Mvctx);
152: if (aij->colmap) {
153: #if defined (PETSC_USE_CTABLE)
154: PetscTableDestroy(&aij->colmap);
155: #else
156: PetscFree(aij->colmap);
157: PetscLogObjectMemory(A,-aij->B->cmap->n*sizeof(PetscInt));
158: #endif
159: }
161: /* make sure that B is assembled so we can access its values */
162: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
163: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
165: /* invent new B and copy stuff over */
166: PetscMalloc((m+1)*sizeof(PetscInt),&nz);
167: for (i=0; i<m; i++) {
168: nz[i] = Baij->i[i+1] - Baij->i[i];
169: }
170: MatCreate(PETSC_COMM_SELF,&Bnew);
171: MatSetSizes(Bnew,m,n,m,n);
172: MatSetBlockSizes(Bnew,A->rmap->bs,A->cmap->bs);
173: MatSetType(Bnew,((PetscObject)B)->type_name);
174: MatSeqAIJSetPreallocation(Bnew,0,nz);
175: ((Mat_SeqAIJ*)Bnew->data)->nonew = Baij->nonew; /* Inherit insertion error options. */
176: PetscFree(nz);
177: for (i=0; i<m; i++) {
178: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
179: col = garray[Baij->j[ct]];
180: v = Baij->a[ct++];
181: MatSetValues(Bnew,1,&i,1,&col,&v,B->insertmode);
182: }
183: }
184: PetscFree(aij->garray);
185: PetscLogObjectMemory(A,-ec*sizeof(PetscInt));
186: MatDestroy(&B);
187: PetscLogObjectParent(A,Bnew);
188: aij->B = Bnew;
189: A->was_assembled = PETSC_FALSE;
190: return(0);
191: }
193: /* ugly stuff added for Glenn someday we should fix this up */
195: static PetscInt *auglyrmapd = 0,*auglyrmapo = 0; /* mapping from the local ordering to the "diagonal" and "off-diagonal"
196: parts of the local matrix */
197: static Vec auglydd = 0,auglyoo = 0; /* work vectors used to scale the two parts of the local matrix */
202: PetscErrorCode MatMPIAIJDiagonalScaleLocalSetUp(Mat inA,Vec scale)
203: {
204: Mat_MPIAIJ *ina = (Mat_MPIAIJ*) inA->data; /*access private part of matrix */
206: PetscInt i,n,nt,cstart,cend,no,*garray = ina->garray,*lindices;
207: PetscInt *r_rmapd,*r_rmapo;
208:
210: MatGetOwnershipRange(inA,&cstart,&cend);
211: MatGetSize(ina->A,PETSC_NULL,&n);
212: PetscMalloc((inA->rmap->mapping->n+1)*sizeof(PetscInt),&r_rmapd);
213: PetscMemzero(r_rmapd,inA->rmap->mapping->n*sizeof(PetscInt));
214: nt = 0;
215: for (i=0; i<inA->rmap->mapping->n; i++) {
216: if (inA->rmap->mapping->indices[i] >= cstart && inA->rmap->mapping->indices[i] < cend) {
217: nt++;
218: r_rmapd[i] = inA->rmap->mapping->indices[i] + 1;
219: }
220: }
221: if (nt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Hmm nt %D n %D",nt,n);
222: PetscMalloc((n+1)*sizeof(PetscInt),&auglyrmapd);
223: for (i=0; i<inA->rmap->mapping->n; i++) {
224: if (r_rmapd[i]){
225: auglyrmapd[(r_rmapd[i]-1)-cstart] = i;
226: }
227: }
228: PetscFree(r_rmapd);
229: VecCreateSeq(PETSC_COMM_SELF,n,&auglydd);
231: PetscMalloc((inA->cmap->N+1)*sizeof(PetscInt),&lindices);
232: PetscMemzero(lindices,inA->cmap->N*sizeof(PetscInt));
233: for (i=0; i<ina->B->cmap->n; i++) {
234: lindices[garray[i]] = i+1;
235: }
236: no = inA->rmap->mapping->n - nt;
237: PetscMalloc((inA->rmap->mapping->n+1)*sizeof(PetscInt),&r_rmapo);
238: PetscMemzero(r_rmapo,inA->rmap->mapping->n*sizeof(PetscInt));
239: nt = 0;
240: for (i=0; i<inA->rmap->mapping->n; i++) {
241: if (lindices[inA->rmap->mapping->indices[i]]) {
242: nt++;
243: r_rmapo[i] = lindices[inA->rmap->mapping->indices[i]];
244: }
245: }
246: if (nt > no) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Hmm nt %D no %D",nt,n);
247: PetscFree(lindices);
248: PetscMalloc((nt+1)*sizeof(PetscInt),&auglyrmapo);
249: for (i=0; i<inA->rmap->mapping->n; i++) {
250: if (r_rmapo[i]){
251: auglyrmapo[(r_rmapo[i]-1)] = i;
252: }
253: }
254: PetscFree(r_rmapo);
255: VecCreateSeq(PETSC_COMM_SELF,nt,&auglyoo);
257: return(0);
258: }
262: PetscErrorCode MatMPIAIJDiagonalScaleLocal(Mat A,Vec scale)
263: {
264: /* This routine should really be abandoned as it duplicates MatDiagonalScaleLocal */
268: PetscTryMethod(A,"MatDiagonalScaleLocal_C",(Mat,Vec),(A,scale));
269: return(0);
270: }
272: EXTERN_C_BEGIN
275: PetscErrorCode MatDiagonalScaleLocal_MPIAIJ(Mat A,Vec scale)
276: {
277: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; /*access private part of matrix */
279: PetscInt n,i;
280: PetscScalar *d,*o,*s;
281:
283: if (!auglyrmapd) {
284: MatMPIAIJDiagonalScaleLocalSetUp(A,scale);
285: }
287: VecGetArray(scale,&s);
288:
289: VecGetLocalSize(auglydd,&n);
290: VecGetArray(auglydd,&d);
291: for (i=0; i<n; i++) {
292: d[i] = s[auglyrmapd[i]]; /* copy "diagonal" (true local) portion of scale into dd vector */
293: }
294: VecRestoreArray(auglydd,&d);
295: /* column scale "diagonal" portion of local matrix */
296: MatDiagonalScale(a->A,PETSC_NULL,auglydd);
298: VecGetLocalSize(auglyoo,&n);
299: VecGetArray(auglyoo,&o);
300: for (i=0; i<n; i++) {
301: o[i] = s[auglyrmapo[i]]; /* copy "off-diagonal" portion of scale into oo vector */
302: }
303: VecRestoreArray(scale,&s);
304: VecRestoreArray(auglyoo,&o);
305: /* column scale "off-diagonal" portion of local matrix */
306: MatDiagonalScale(a->B,PETSC_NULL,auglyoo);
308: return(0);
309: }
310: EXTERN_C_END