Actual source code: mmaij.c
petsc-3.9.1 2018-04-29
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,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
26: if (!aij->garray) {
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: PetscMalloc1(ec+1,&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: aij->B->cmap->bs = 1;
66: PetscLayoutSetUp((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+1,&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+1,&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: aij->B->cmap->n = aij->B->cmap->N = ec;
98: aij->B->cmap->bs = 1;
100: PetscLayoutSetUp((aij->B->cmap));
101: PetscFree(indices);
102: #endif
103: } else {
104: garray = aij->garray;
105: }
107: if (!aij->lvec) {
108: /* create local vector that is used to scatter into */
109: VecCreateSeq(PETSC_COMM_SELF,ec,&aij->lvec);
110: } else {
111: VecGetSize(aij->lvec,&ec);
112: }
114: /* create two temporary Index sets for build scatter gather */
115: ISCreateGeneral(((PetscObject)mat)->comm,ec,garray,PETSC_COPY_VALUES,&from);
117: ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);
119: /* create temporary global vector to generate scatter context */
120: /* This does not allocate the array's memory so is efficient */
121: VecCreateMPIWithArray(PetscObjectComm((PetscObject)mat),1,mat->cmap->n,mat->cmap->N,NULL,&gvec);
123: /* generate the scatter context */
124: if (aij->Mvctx_mpi1_flg) {
125: VecScatterDestroy(&aij->Mvctx_mpi1);
126: VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx_mpi1);
127: VecScatterSetType(aij->Mvctx_mpi1,VECSCATTERMPI1);
128: PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->Mvctx_mpi1);
129: } else {
130: VecScatterDestroy(&aij->Mvctx);
131: VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);
132: PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->Mvctx);
133: PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->lvec);
134: PetscLogObjectMemory((PetscObject)mat,(ec+1)*sizeof(PetscInt));
135: }
136: aij->garray = garray;
138: PetscLogObjectParent((PetscObject)mat,(PetscObject)from);
139: PetscLogObjectParent((PetscObject)mat,(PetscObject)to);
141: ISDestroy(&from);
142: ISDestroy(&to);
143: VecDestroy(&gvec);
144: return(0);
145: }
147: /*
148: Takes the local part of an already assembled MPIAIJ matrix
149: and disassembles it. This is to allow new nonzeros into the matrix
150: that require more communication in the matrix vector multiply.
151: Thus certain data-structures must be rebuilt.
153: Kind of slow! But that's what application programmers get when
154: they are sloppy.
155: */
156: PetscErrorCode MatDisAssemble_MPIAIJ(Mat A)
157: {
158: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data;
159: Mat B = aij->B,Bnew;
160: Mat_SeqAIJ *Baij = (Mat_SeqAIJ*)B->data;
162: PetscInt i,j,m = B->rmap->n,n = A->cmap->N,col,ct = 0,*garray = aij->garray,*nz,ec;
163: PetscScalar v;
166: /* free stuff related to matrix-vec multiply */
167: VecGetSize(aij->lvec,&ec); /* needed for PetscLogObjectMemory below */
168: VecDestroy(&aij->lvec);
169: if (aij->colmap) {
170: #if defined(PETSC_USE_CTABLE)
171: PetscTableDestroy(&aij->colmap);
172: #else
173: PetscFree(aij->colmap);
174: PetscLogObjectMemory((PetscObject)A,-aij->B->cmap->n*sizeof(PetscInt));
175: #endif
176: }
178: /* make sure that B is assembled so we can access its values */
179: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
180: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
182: /* invent new B and copy stuff over */
183: PetscMalloc1(m+1,&nz);
184: for (i=0; i<m; i++) {
185: nz[i] = Baij->i[i+1] - Baij->i[i];
186: }
187: MatCreate(PETSC_COMM_SELF,&Bnew);
188: MatSetSizes(Bnew,m,n,m,n);
189: MatSetBlockSizesFromMats(Bnew,A,A);
190: MatSetType(Bnew,((PetscObject)B)->type_name);
191: MatSeqAIJSetPreallocation(Bnew,0,nz);
193: ((Mat_SeqAIJ*)Bnew->data)->nonew = Baij->nonew; /* Inherit insertion error options. */
194: /*
195: Ensure that B's nonzerostate is monotonically increasing.
196: Or should this follow the MatSetValues() loop to preserve B's nonzerstate across a MatDisAssemble() call?
197: */
198: Bnew->nonzerostate = B->nonzerostate;
200: PetscFree(nz);
201: for (i=0; i<m; i++) {
202: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
203: col = garray[Baij->j[ct]];
204: v = Baij->a[ct++];
205: MatSetValues(Bnew,1,&i,1,&col,&v,B->insertmode);
206: }
207: }
208: PetscFree(aij->garray);
209: PetscLogObjectMemory((PetscObject)A,-ec*sizeof(PetscInt));
210: MatDestroy(&B);
211: PetscLogObjectParent((PetscObject)A,(PetscObject)Bnew);
213: aij->B = Bnew;
214: A->was_assembled = PETSC_FALSE;
215: return(0);
216: }
218: /* ugly stuff added for Glenn someday we should fix this up */
220: static PetscInt *auglyrmapd = 0,*auglyrmapo = 0; /* mapping from the local ordering to the "diagonal" and "off-diagonal" parts of the local matrix */
221: static Vec auglydd = 0,auglyoo = 0; /* work vectors used to scale the two parts of the local matrix */
224: PetscErrorCode MatMPIAIJDiagonalScaleLocalSetUp(Mat inA,Vec scale)
225: {
226: Mat_MPIAIJ *ina = (Mat_MPIAIJ*) inA->data; /*access private part of matrix */
228: PetscInt i,n,nt,cstart,cend,no,*garray = ina->garray,*lindices;
229: PetscInt *r_rmapd,*r_rmapo;
232: MatGetOwnershipRange(inA,&cstart,&cend);
233: MatGetSize(ina->A,NULL,&n);
234: PetscCalloc1(inA->rmap->mapping->n+1,&r_rmapd);
235: nt = 0;
236: for (i=0; i<inA->rmap->mapping->n; i++) {
237: if (inA->rmap->mapping->indices[i] >= cstart && inA->rmap->mapping->indices[i] < cend) {
238: nt++;
239: r_rmapd[i] = inA->rmap->mapping->indices[i] + 1;
240: }
241: }
242: if (nt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Hmm nt %D n %D",nt,n);
243: PetscMalloc1(n+1,&auglyrmapd);
244: for (i=0; i<inA->rmap->mapping->n; i++) {
245: if (r_rmapd[i]) {
246: auglyrmapd[(r_rmapd[i]-1)-cstart] = i;
247: }
248: }
249: PetscFree(r_rmapd);
250: VecCreateSeq(PETSC_COMM_SELF,n,&auglydd);
252: PetscCalloc1(inA->cmap->N+1,&lindices);
253: for (i=0; i<ina->B->cmap->n; i++) {
254: lindices[garray[i]] = i+1;
255: }
256: no = inA->rmap->mapping->n - nt;
257: PetscCalloc1(inA->rmap->mapping->n+1,&r_rmapo);
258: nt = 0;
259: for (i=0; i<inA->rmap->mapping->n; i++) {
260: if (lindices[inA->rmap->mapping->indices[i]]) {
261: nt++;
262: r_rmapo[i] = lindices[inA->rmap->mapping->indices[i]];
263: }
264: }
265: if (nt > no) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Hmm nt %D no %D",nt,n);
266: PetscFree(lindices);
267: PetscMalloc1(nt+1,&auglyrmapo);
268: for (i=0; i<inA->rmap->mapping->n; i++) {
269: if (r_rmapo[i]) {
270: auglyrmapo[(r_rmapo[i]-1)] = i;
271: }
272: }
273: PetscFree(r_rmapo);
274: VecCreateSeq(PETSC_COMM_SELF,nt,&auglyoo);
275: return(0);
276: }
278: PetscErrorCode MatMPIAIJDiagonalScaleLocal(Mat A,Vec scale)
279: {
280: /* This routine should really be abandoned as it duplicates MatDiagonalScaleLocal */
284: PetscTryMethod(A,"MatDiagonalScaleLocal_C",(Mat,Vec),(A,scale));
285: return(0);
286: }
288: PetscErrorCode MatDiagonalScaleLocal_MPIAIJ(Mat A,Vec scale)
289: {
290: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; /*access private part of matrix */
291: PetscErrorCode ierr;
292: PetscInt n,i;
293: PetscScalar *d,*o;
294: const PetscScalar *s;
297: if (!auglyrmapd) {
298: MatMPIAIJDiagonalScaleLocalSetUp(A,scale);
299: }
301: VecGetArrayRead(scale,&s);
303: VecGetLocalSize(auglydd,&n);
304: VecGetArray(auglydd,&d);
305: for (i=0; i<n; i++) {
306: d[i] = s[auglyrmapd[i]]; /* copy "diagonal" (true local) portion of scale into dd vector */
307: }
308: VecRestoreArray(auglydd,&d);
309: /* column scale "diagonal" portion of local matrix */
310: MatDiagonalScale(a->A,NULL,auglydd);
312: VecGetLocalSize(auglyoo,&n);
313: VecGetArray(auglyoo,&o);
314: for (i=0; i<n; i++) {
315: o[i] = s[auglyrmapo[i]]; /* copy "off-diagonal" portion of scale into oo vector */
316: }
317: VecRestoreArrayRead(scale,&s);
318: VecRestoreArray(auglyoo,&o);
319: /* column scale "off-diagonal" portion of local matrix */
320: MatDiagonalScale(a->B,NULL,auglyoo);
321: return(0);
322: }