Actual source code: mmaij.c
2: /*
3: Support for the parallel AIJ matrix vector multiply
4: */
5: #include <../src/mat/impls/aij/mpi/mpiaij.h>
6: #include <petsc/private/vecimpl.h>
7: #include <petsc/private/isimpl.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);
14: PetscInt i,j,*aj = B->j,*garray;
15: PetscInt ec = 0; /* Number of nonzero external columns */
16: IS from,to;
17: Vec gvec;
18: #if defined(PETSC_USE_CTABLE)
19: PetscTable gid1_lid1;
20: PetscTablePosition tpos;
21: PetscInt gid,lid;
22: #else
23: PetscInt N = mat->cmap->N,*indices;
24: #endif
27: if (!aij->garray) {
28: if (!aij->B) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing B mat");
29: #if defined(PETSC_USE_CTABLE)
30: /* use a table */
31: PetscTableCreate(aij->B->rmap->n,mat->cmap->N+1,&gid1_lid1);
32: for (i=0; i<aij->B->rmap->n; i++) {
33: for (j=0; j<B->ilen[i]; j++) {
34: PetscInt data,gid1 = aj[B->i[i] + j] + 1;
35: PetscTableFind(gid1_lid1,gid1,&data);
36: if (!data) {
37: /* one based table */
38: PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);
39: }
40: }
41: }
42: /* form array of columns we need */
43: PetscMalloc1(ec,&garray);
44: PetscTableGetHeadPosition(gid1_lid1,&tpos);
45: while (tpos) {
46: PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);
47: gid--;
48: lid--;
49: garray[lid] = gid;
50: }
51: PetscSortInt(ec,garray); /* sort, and rebuild */
52: PetscTableRemoveAll(gid1_lid1);
53: for (i=0; i<ec; i++) {
54: PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);
55: }
56: /* compact out the extra columns in B */
57: for (i=0; i<aij->B->rmap->n; i++) {
58: for (j=0; j<B->ilen[i]; j++) {
59: PetscInt gid1 = aj[B->i[i] + j] + 1;
60: PetscTableFind(gid1_lid1,gid1,&lid);
61: lid--;
62: aj[B->i[i] + j] = lid;
63: }
64: }
65: PetscLayoutDestroy(&aij->B->cmap);
66: PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)aij->B),ec,ec,1,&aij->B->cmap);
67: PetscTableDestroy(&gid1_lid1);
68: #else
69: /* Make an array as long as the number of columns */
70: /* mark those columns that are in aij->B */
71: PetscCalloc1(N,&indices);
72: for (i=0; i<aij->B->rmap->n; i++) {
73: for (j=0; j<B->ilen[i]; j++) {
74: if (!indices[aj[B->i[i] + j]]) ec++;
75: indices[aj[B->i[i] + j]] = 1;
76: }
77: }
79: /* form array of columns we need */
80: PetscMalloc1(ec,&garray);
81: ec = 0;
82: for (i=0; i<N; i++) {
83: if (indices[i]) garray[ec++] = i;
84: }
86: /* make indices now point into garray */
87: for (i=0; i<ec; i++) {
88: indices[garray[i]] = i;
89: }
91: /* compact out the extra columns in B */
92: for (i=0; i<aij->B->rmap->n; i++) {
93: for (j=0; j<B->ilen[i]; j++) {
94: aj[B->i[i] + j] = indices[aj[B->i[i] + j]];
95: }
96: }
97: PetscLayoutDestroy(&aij->B->cmap);
98: PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)aij->B),ec,ec,1,&aij->B->cmap);
99: PetscFree(indices);
100: #endif
101: } else {
102: garray = aij->garray;
103: }
105: if (!aij->lvec) {
106: if (!aij->B) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing B mat");
107: MatCreateVecs(aij->B,&aij->lvec,NULL);
108: }
109: VecGetSize(aij->lvec,&ec);
111: /* create two temporary Index sets for build scatter gather */
112: ISCreateGeneral(PETSC_COMM_SELF,ec,garray,PETSC_COPY_VALUES,&from);
113: ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);
115: /* create temporary global vector to generate scatter context */
116: /* This does not allocate the array's memory so is efficient */
117: VecCreateMPIWithArray(PetscObjectComm((PetscObject)mat),1,mat->cmap->n,mat->cmap->N,NULL,&gvec);
119: /* generate the scatter context */
120: VecScatterDestroy(&aij->Mvctx);
121: VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);
122: PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->Mvctx);
123: PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->lvec);
124: PetscLogObjectMemory((PetscObject)mat,ec*sizeof(PetscInt));
125: aij->garray = garray;
127: PetscLogObjectParent((PetscObject)mat,(PetscObject)from);
128: PetscLogObjectParent((PetscObject)mat,(PetscObject)to);
130: ISDestroy(&from);
131: ISDestroy(&to);
132: VecDestroy(&gvec);
133: return(0);
134: }
136: /*
137: Takes the local part of an already assembled MPIAIJ matrix
138: and disassembles it. This is to allow new nonzeros into the matrix
139: that require more communication in the matrix vector multiply.
140: Thus certain data-structures must be rebuilt.
142: Kind of slow! But that's what application programmers get when
143: they are sloppy.
144: */
145: PetscErrorCode MatDisAssemble_MPIAIJ(Mat A)
146: {
147: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data;
148: Mat B = aij->B,Bnew;
149: Mat_SeqAIJ *Baij = (Mat_SeqAIJ*)B->data;
151: PetscInt i,j,m = B->rmap->n,n = A->cmap->N,col,ct = 0,*garray = aij->garray,*nz,ec;
152: PetscScalar v;
155: /* free stuff related to matrix-vec multiply */
156: VecGetSize(aij->lvec,&ec); /* needed for PetscLogObjectMemory below */
157: VecDestroy(&aij->lvec);
158: if (aij->colmap) {
159: #if defined(PETSC_USE_CTABLE)
160: PetscTableDestroy(&aij->colmap);
161: #else
162: PetscFree(aij->colmap);
163: PetscLogObjectMemory((PetscObject)A,-aij->B->cmap->n*sizeof(PetscInt));
164: #endif
165: }
167: /* make sure that B is assembled so we can access its values */
168: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
169: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
171: /* invent new B and copy stuff over */
172: PetscMalloc1(m+1,&nz);
173: for (i=0; i<m; i++) {
174: nz[i] = Baij->i[i+1] - Baij->i[i];
175: }
176: MatCreate(PETSC_COMM_SELF,&Bnew);
177: MatSetSizes(Bnew,m,n,m,n);
178: MatSetBlockSizesFromMats(Bnew,A,A);
179: MatSetType(Bnew,((PetscObject)B)->type_name);
180: MatSeqAIJSetPreallocation(Bnew,0,nz);
182: if (Baij->nonew >= 0) { /* Inherit insertion error options (if positive). */
183: ((Mat_SeqAIJ*)Bnew->data)->nonew = Baij->nonew;
184: }
186: /*
187: Ensure that B's nonzerostate is monotonically increasing.
188: Or should this follow the MatSetValues() loop to preserve B's nonzerstate across a MatDisAssemble() call?
189: */
190: Bnew->nonzerostate = B->nonzerostate;
192: PetscFree(nz);
193: for (i=0; i<m; i++) {
194: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
195: col = garray[Baij->j[ct]];
196: v = Baij->a[ct++];
197: MatSetValues(Bnew,1,&i,1,&col,&v,B->insertmode);
198: }
199: }
200: PetscFree(aij->garray);
201: PetscLogObjectMemory((PetscObject)A,-ec*sizeof(PetscInt));
202: MatDestroy(&B);
203: PetscLogObjectParent((PetscObject)A,(PetscObject)Bnew);
205: aij->B = Bnew;
206: A->was_assembled = PETSC_FALSE;
207: return(0);
208: }
210: /* ugly stuff added for Glenn someday we should fix this up */
212: static PetscInt *auglyrmapd = NULL,*auglyrmapo = NULL; /* mapping from the local ordering to the "diagonal" and "off-diagonal" parts of the local matrix */
213: static Vec auglydd = NULL,auglyoo = NULL; /* work vectors used to scale the two parts of the local matrix */
215: PetscErrorCode MatMPIAIJDiagonalScaleLocalSetUp(Mat inA,Vec scale)
216: {
217: Mat_MPIAIJ *ina = (Mat_MPIAIJ*) inA->data; /*access private part of matrix */
219: PetscInt i,n,nt,cstart,cend,no,*garray = ina->garray,*lindices;
220: PetscInt *r_rmapd,*r_rmapo;
223: MatGetOwnershipRange(inA,&cstart,&cend);
224: MatGetSize(ina->A,NULL,&n);
225: PetscCalloc1(inA->rmap->mapping->n+1,&r_rmapd);
226: nt = 0;
227: for (i=0; i<inA->rmap->mapping->n; i++) {
228: if (inA->rmap->mapping->indices[i] >= cstart && inA->rmap->mapping->indices[i] < cend) {
229: nt++;
230: r_rmapd[i] = inA->rmap->mapping->indices[i] + 1;
231: }
232: }
233: if (nt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Hmm nt %D n %D",nt,n);
234: PetscMalloc1(n+1,&auglyrmapd);
235: for (i=0; i<inA->rmap->mapping->n; i++) {
236: if (r_rmapd[i]) {
237: auglyrmapd[(r_rmapd[i]-1)-cstart] = i;
238: }
239: }
240: PetscFree(r_rmapd);
241: VecCreateSeq(PETSC_COMM_SELF,n,&auglydd);
243: PetscCalloc1(inA->cmap->N+1,&lindices);
244: for (i=0; i<ina->B->cmap->n; i++) {
245: lindices[garray[i]] = i+1;
246: }
247: no = inA->rmap->mapping->n - nt;
248: PetscCalloc1(inA->rmap->mapping->n+1,&r_rmapo);
249: nt = 0;
250: for (i=0; i<inA->rmap->mapping->n; i++) {
251: if (lindices[inA->rmap->mapping->indices[i]]) {
252: nt++;
253: r_rmapo[i] = lindices[inA->rmap->mapping->indices[i]];
254: }
255: }
256: if (nt > no) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Hmm nt %D no %D",nt,n);
257: PetscFree(lindices);
258: PetscMalloc1(nt+1,&auglyrmapo);
259: for (i=0; i<inA->rmap->mapping->n; i++) {
260: if (r_rmapo[i]) {
261: auglyrmapo[(r_rmapo[i]-1)] = i;
262: }
263: }
264: PetscFree(r_rmapo);
265: VecCreateSeq(PETSC_COMM_SELF,nt,&auglyoo);
266: return(0);
267: }
269: PetscErrorCode MatMPIAIJDiagonalScaleLocal(Mat A,Vec scale)
270: {
271: /* This routine should really be abandoned as it duplicates MatDiagonalScaleLocal */
275: PetscTryMethod(A,"MatDiagonalScaleLocal_C",(Mat,Vec),(A,scale));
276: return(0);
277: }
279: PetscErrorCode MatDiagonalScaleLocal_MPIAIJ(Mat A,Vec scale)
280: {
281: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; /*access private part of matrix */
282: PetscErrorCode ierr;
283: PetscInt n,i;
284: PetscScalar *d,*o;
285: const PetscScalar *s;
288: if (!auglyrmapd) {
289: MatMPIAIJDiagonalScaleLocalSetUp(A,scale);
290: }
292: VecGetArrayRead(scale,&s);
294: VecGetLocalSize(auglydd,&n);
295: VecGetArray(auglydd,&d);
296: for (i=0; i<n; i++) {
297: d[i] = s[auglyrmapd[i]]; /* copy "diagonal" (true local) portion of scale into dd vector */
298: }
299: VecRestoreArray(auglydd,&d);
300: /* column scale "diagonal" portion of local matrix */
301: MatDiagonalScale(a->A,NULL,auglydd);
303: VecGetLocalSize(auglyoo,&n);
304: VecGetArray(auglyoo,&o);
305: for (i=0; i<n; i++) {
306: o[i] = s[auglyrmapo[i]]; /* copy "off-diagonal" portion of scale into oo vector */
307: }
308: VecRestoreArrayRead(scale,&s);
309: VecRestoreArray(auglyoo,&o);
310: /* column scale "off-diagonal" portion of local matrix */
311: MatDiagonalScale(a->B,NULL,auglyoo);
312: return(0);
313: }