Actual source code: aij.c
petsc-3.8.4 2018-03-24
2: /*
3: Defines the basic matrix operations for the AIJ (compressed row)
4: matrix storage format.
5: */
8: #include <../src/mat/impls/aij/seq/aij.h>
9: #include <petscblaslapack.h>
10: #include <petscbt.h>
11: #include <petsc/private/kernels/blocktranspose.h>
13: PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A)
14: {
15: PetscErrorCode ierr;
16: PetscBool flg;
17: char type[256];
20: PetscObjectOptionsBegin((PetscObject)A);
21: PetscOptionsFList("-mat_seqaij_type","Matrix SeqAIJ type","MatSeqAIJSetType",MatSeqAIJList,"seqaij",type,256,&flg);
22: if (flg) {
23: MatSeqAIJSetType(A,type);
24: }
25: PetscOptionsEnd();
26: return(0);
27: }
29: PetscErrorCode MatGetColumnNorms_SeqAIJ(Mat A,NormType type,PetscReal *norms)
30: {
32: PetscInt i,m,n;
33: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
36: MatGetSize(A,&m,&n);
37: PetscMemzero(norms,n*sizeof(PetscReal));
38: if (type == NORM_2) {
39: for (i=0; i<aij->i[m]; i++) {
40: norms[aij->j[i]] += PetscAbsScalar(aij->a[i]*aij->a[i]);
41: }
42: } else if (type == NORM_1) {
43: for (i=0; i<aij->i[m]; i++) {
44: norms[aij->j[i]] += PetscAbsScalar(aij->a[i]);
45: }
46: } else if (type == NORM_INFINITY) {
47: for (i=0; i<aij->i[m]; i++) {
48: norms[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]),norms[aij->j[i]]);
49: }
50: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown NormType");
52: if (type == NORM_2) {
53: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
54: }
55: return(0);
56: }
58: PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A,IS *is)
59: {
60: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
61: PetscInt i,m=A->rmap->n,cnt = 0, bs = A->rmap->bs;
62: const PetscInt *jj = a->j,*ii = a->i;
63: PetscInt *rows;
64: PetscErrorCode ierr;
67: for (i=0; i<m; i++) {
68: if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) {
69: cnt++;
70: }
71: }
72: PetscMalloc1(cnt,&rows);
73: cnt = 0;
74: for (i=0; i<m; i++) {
75: if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) {
76: rows[cnt] = i;
77: cnt++;
78: }
79: }
80: ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,is);
81: return(0);
82: }
84: PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A,PetscInt *nrows,PetscInt **zrows)
85: {
86: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
87: const MatScalar *aa = a->a;
88: PetscInt i,m=A->rmap->n,cnt = 0;
89: const PetscInt *ii = a->i,*jj = a->j,*diag;
90: PetscInt *rows;
91: PetscErrorCode ierr;
94: MatMarkDiagonal_SeqAIJ(A);
95: diag = a->diag;
96: for (i=0; i<m; i++) {
97: if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) {
98: cnt++;
99: }
100: }
101: PetscMalloc1(cnt,&rows);
102: cnt = 0;
103: for (i=0; i<m; i++) {
104: if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) {
105: rows[cnt++] = i;
106: }
107: }
108: *nrows = cnt;
109: *zrows = rows;
110: return(0);
111: }
113: PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A,IS *zrows)
114: {
115: PetscInt nrows,*rows;
119: *zrows = NULL;
120: MatFindZeroDiagonals_SeqAIJ_Private(A,&nrows,&rows);
121: ISCreateGeneral(PetscObjectComm((PetscObject)A),nrows,rows,PETSC_OWN_POINTER,zrows);
122: return(0);
123: }
125: PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A,IS *keptrows)
126: {
127: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
128: const MatScalar *aa;
129: PetscInt m=A->rmap->n,cnt = 0;
130: const PetscInt *ii;
131: PetscInt n,i,j,*rows;
132: PetscErrorCode ierr;
135: *keptrows = 0;
136: ii = a->i;
137: for (i=0; i<m; i++) {
138: n = ii[i+1] - ii[i];
139: if (!n) {
140: cnt++;
141: goto ok1;
142: }
143: aa = a->a + ii[i];
144: for (j=0; j<n; j++) {
145: if (aa[j] != 0.0) goto ok1;
146: }
147: cnt++;
148: ok1:;
149: }
150: if (!cnt) return(0);
151: PetscMalloc1(A->rmap->n-cnt,&rows);
152: cnt = 0;
153: for (i=0; i<m; i++) {
154: n = ii[i+1] - ii[i];
155: if (!n) continue;
156: aa = a->a + ii[i];
157: for (j=0; j<n; j++) {
158: if (aa[j] != 0.0) {
159: rows[cnt++] = i;
160: break;
161: }
162: }
163: }
164: ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,keptrows);
165: return(0);
166: }
168: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y,Vec D,InsertMode is)
169: {
170: PetscErrorCode ierr;
171: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) Y->data;
172: PetscInt i,m = Y->rmap->n;
173: const PetscInt *diag;
174: MatScalar *aa = aij->a;
175: const PetscScalar *v;
176: PetscBool missing;
179: if (Y->assembled) {
180: MatMissingDiagonal_SeqAIJ(Y,&missing,NULL);
181: if (!missing) {
182: diag = aij->diag;
183: VecGetArrayRead(D,&v);
184: if (is == INSERT_VALUES) {
185: for (i=0; i<m; i++) {
186: aa[diag[i]] = v[i];
187: }
188: } else {
189: for (i=0; i<m; i++) {
190: aa[diag[i]] += v[i];
191: }
192: }
193: VecRestoreArrayRead(D,&v);
194: return(0);
195: }
196: MatSeqAIJInvalidateDiagonal(Y);
197: }
198: MatDiagonalSet_Default(Y,D,is);
199: return(0);
200: }
202: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *m,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
203: {
204: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
206: PetscInt i,ishift;
209: *m = A->rmap->n;
210: if (!ia) return(0);
211: ishift = 0;
212: if (symmetric && !A->structurally_symmetric) {
213: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,ishift,oshift,(PetscInt**)ia,(PetscInt**)ja);
214: } else if (oshift == 1) {
215: PetscInt *tia;
216: PetscInt nz = a->i[A->rmap->n];
217: /* malloc space and add 1 to i and j indices */
218: PetscMalloc1(A->rmap->n+1,&tia);
219: for (i=0; i<A->rmap->n+1; i++) tia[i] = a->i[i] + 1;
220: *ia = tia;
221: if (ja) {
222: PetscInt *tja;
223: PetscMalloc1(nz+1,&tja);
224: for (i=0; i<nz; i++) tja[i] = a->j[i] + 1;
225: *ja = tja;
226: }
227: } else {
228: *ia = a->i;
229: if (ja) *ja = a->j;
230: }
231: return(0);
232: }
234: PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
235: {
239: if (!ia) return(0);
240: if ((symmetric && !A->structurally_symmetric) || oshift == 1) {
241: PetscFree(*ia);
242: if (ja) {PetscFree(*ja);}
243: }
244: return(0);
245: }
247: PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
248: {
249: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
251: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
252: PetscInt nz = a->i[m],row,*jj,mr,col;
255: *nn = n;
256: if (!ia) return(0);
257: if (symmetric) {
258: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,0,oshift,(PetscInt**)ia,(PetscInt**)ja);
259: } else {
260: PetscCalloc1(n+1,&collengths);
261: PetscMalloc1(n+1,&cia);
262: PetscMalloc1(nz+1,&cja);
263: jj = a->j;
264: for (i=0; i<nz; i++) {
265: collengths[jj[i]]++;
266: }
267: cia[0] = oshift;
268: for (i=0; i<n; i++) {
269: cia[i+1] = cia[i] + collengths[i];
270: }
271: PetscMemzero(collengths,n*sizeof(PetscInt));
272: jj = a->j;
273: for (row=0; row<m; row++) {
274: mr = a->i[row+1] - a->i[row];
275: for (i=0; i<mr; i++) {
276: col = *jj++;
278: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
279: }
280: }
281: PetscFree(collengths);
282: *ia = cia; *ja = cja;
283: }
284: return(0);
285: }
287: PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
288: {
292: if (!ia) return(0);
294: PetscFree(*ia);
295: PetscFree(*ja);
296: return(0);
297: }
299: /*
300: MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from
301: MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output
302: spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ()
303: */
304: PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
305: {
306: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
308: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
309: PetscInt nz = a->i[m],row,*jj,mr,col;
310: PetscInt *cspidx;
313: *nn = n;
314: if (!ia) return(0);
316: PetscCalloc1(n+1,&collengths);
317: PetscMalloc1(n+1,&cia);
318: PetscMalloc1(nz+1,&cja);
319: PetscMalloc1(nz+1,&cspidx);
320: jj = a->j;
321: for (i=0; i<nz; i++) {
322: collengths[jj[i]]++;
323: }
324: cia[0] = oshift;
325: for (i=0; i<n; i++) {
326: cia[i+1] = cia[i] + collengths[i];
327: }
328: PetscMemzero(collengths,n*sizeof(PetscInt));
329: jj = a->j;
330: for (row=0; row<m; row++) {
331: mr = a->i[row+1] - a->i[row];
332: for (i=0; i<mr; i++) {
333: col = *jj++;
334: cspidx[cia[col] + collengths[col] - oshift] = a->i[row] + i; /* index of a->j */
335: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
336: }
337: }
338: PetscFree(collengths);
339: *ia = cia; *ja = cja;
340: *spidx = cspidx;
341: return(0);
342: }
344: PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
345: {
349: MatRestoreColumnIJ_SeqAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);
350: PetscFree(*spidx);
351: return(0);
352: }
354: PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A,PetscInt row,const PetscScalar v[])
355: {
356: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
357: PetscInt *ai = a->i;
361: PetscMemcpy(a->a+ai[row],v,(ai[row+1]-ai[row])*sizeof(PetscScalar));
362: return(0);
363: }
365: /*
366: MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions
368: - a single row of values is set with each call
369: - no row or column indices are negative or (in error) larger than the number of rows or columns
370: - the values are always added to the matrix, not set
371: - no new locations are introduced in the nonzero structure of the matrix
373: This does NOT assume the global column indices are sorted
375: */
377: #include <petsc/private/isimpl.h>
378: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
379: {
380: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
381: PetscInt low,high,t,row,nrow,i,col,l;
382: const PetscInt *rp,*ai = a->i,*ailen = a->ilen,*aj = a->j;
383: PetscInt lastcol = -1;
384: MatScalar *ap,value,*aa = a->a;
385: const PetscInt *ridx = A->rmap->mapping->indices,*cidx = A->cmap->mapping->indices;
387: row = ridx[im[0]];
388: rp = aj + ai[row];
389: ap = aa + ai[row];
390: nrow = ailen[row];
391: low = 0;
392: high = nrow;
393: for (l=0; l<n; l++) { /* loop over added columns */
394: col = cidx[in[l]];
395: value = v[l];
397: if (col <= lastcol) low = 0;
398: else high = nrow;
399: lastcol = col;
400: while (high-low > 5) {
401: t = (low+high)/2;
402: if (rp[t] > col) high = t;
403: else low = t;
404: }
405: for (i=low; i<high; i++) {
406: if (rp[i] == col) {
407: ap[i] += value;
408: low = i + 1;
409: break;
410: }
411: }
412: }
413: return 0;
414: }
416: PetscErrorCode MatSetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
417: {
418: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
419: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
420: PetscInt *imax = a->imax,*ai = a->i,*ailen = a->ilen;
422: PetscInt *aj = a->j,nonew = a->nonew,lastcol = -1;
423: MatScalar *ap=NULL,value=0.0,*aa = a->a;
424: PetscBool ignorezeroentries = a->ignorezeroentries;
425: PetscBool roworiented = a->roworiented;
428: for (k=0; k<m; k++) { /* loop over added rows */
429: row = im[k];
430: if (row < 0) continue;
431: #if defined(PETSC_USE_DEBUG)
432: if (row >= A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->n-1);
433: #endif
434: rp = aj + ai[row];
435: if (!A->structure_only) ap = aa + ai[row];
436: rmax = imax[row]; nrow = ailen[row];
437: low = 0;
438: high = nrow;
439: for (l=0; l<n; l++) { /* loop over added columns */
440: if (in[l] < 0) continue;
441: #if defined(PETSC_USE_DEBUG)
442: if (in[l] >= A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1);
443: #endif
444: col = in[l];
445: if (!A->structure_only) {
446: if (roworiented) {
447: value = v[l + k*n];
448: } else {
449: value = v[k + l*m];
450: }
451: } else { /* A->structure_only */
452: value = 1; /* avoid 'continue' below? */
453: }
454: if ((value == 0.0 && ignorezeroentries) && (is == ADD_VALUES) && row != col) continue;
456: if (col <= lastcol) low = 0;
457: else high = nrow;
458: lastcol = col;
459: while (high-low > 5) {
460: t = (low+high)/2;
461: if (rp[t] > col) high = t;
462: else low = t;
463: }
464: for (i=low; i<high; i++) {
465: if (rp[i] > col) break;
466: if (rp[i] == col) {
467: if (!A->structure_only) {
468: if (is == ADD_VALUES) ap[i] += value;
469: else ap[i] = value;
470: }
471: low = i + 1;
472: goto noinsert;
473: }
474: }
475: if (value == 0.0 && ignorezeroentries && row != col) goto noinsert;
476: if (nonew == 1) goto noinsert;
477: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at (%D,%D) in the matrix",row,col);
478: if (A->structure_only) {
479: MatSeqXAIJReallocateAIJ_structure_only(A,A->rmap->n,1,nrow,row,col,rmax,ai,aj,rp,imax,nonew,MatScalar);
480: } else {
481: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
482: }
483: N = nrow++ - 1; a->nz++; high++;
484: /* shift up all the later entries in this row */
485: for (ii=N; ii>=i; ii--) {
486: rp[ii+1] = rp[ii];
487: if (!A->structure_only) ap[ii+1] = ap[ii];
488: }
489: rp[i] = col;
490: if (!A->structure_only) ap[i] = value;
491: low = i + 1;
492: A->nonzerostate++;
493: noinsert:;
494: }
495: ailen[row] = nrow;
496: }
497: return(0);
498: }
501: PetscErrorCode MatGetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[])
502: {
503: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
504: PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j;
505: PetscInt *ai = a->i,*ailen = a->ilen;
506: MatScalar *ap,*aa = a->a;
509: for (k=0; k<m; k++) { /* loop over rows */
510: row = im[k];
511: if (row < 0) {v += n; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",row); */
512: if (row >= A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->n-1);
513: rp = aj + ai[row]; ap = aa + ai[row];
514: nrow = ailen[row];
515: for (l=0; l<n; l++) { /* loop over columns */
516: if (in[l] < 0) {v++; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",in[l]); */
517: if (in[l] >= A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1);
518: col = in[l];
519: high = nrow; low = 0; /* assume unsorted */
520: while (high-low > 5) {
521: t = (low+high)/2;
522: if (rp[t] > col) high = t;
523: else low = t;
524: }
525: for (i=low; i<high; i++) {
526: if (rp[i] > col) break;
527: if (rp[i] == col) {
528: *v++ = ap[i];
529: goto finished;
530: }
531: }
532: *v++ = 0.0;
533: finished:;
534: }
535: }
536: return(0);
537: }
540: PetscErrorCode MatView_SeqAIJ_Binary(Mat A,PetscViewer viewer)
541: {
542: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
544: PetscInt i,*col_lens;
545: int fd;
546: FILE *file;
549: PetscViewerBinaryGetDescriptor(viewer,&fd);
550: PetscMalloc1(4+A->rmap->n,&col_lens);
552: col_lens[0] = MAT_FILE_CLASSID;
553: col_lens[1] = A->rmap->n;
554: col_lens[2] = A->cmap->n;
555: col_lens[3] = a->nz;
557: /* store lengths of each row and write (including header) to file */
558: for (i=0; i<A->rmap->n; i++) {
559: col_lens[4+i] = a->i[i+1] - a->i[i];
560: }
561: PetscBinaryWrite(fd,col_lens,4+A->rmap->n,PETSC_INT,PETSC_TRUE);
562: PetscFree(col_lens);
564: /* store column indices (zero start index) */
565: PetscBinaryWrite(fd,a->j,a->nz,PETSC_INT,PETSC_FALSE);
567: /* store nonzero values */
568: PetscBinaryWrite(fd,a->a,a->nz,PETSC_SCALAR,PETSC_FALSE);
570: PetscViewerBinaryGetInfoPointer(viewer,&file);
571: if (file) {
572: fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(A->rmap->bs));
573: }
574: return(0);
575: }
577: static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A,PetscViewer viewer)
578: {
580: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
581: PetscInt i,k,m=A->rmap->N;
584: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
585: for (i=0; i<m; i++) {
586: PetscViewerASCIIPrintf(viewer,"row %D:",i);
587: for (k=a->i[i]; k<a->i[i+1]; k++) {
588: PetscViewerASCIIPrintf(viewer," (%D) ",a->j[k]);
589: }
590: PetscViewerASCIIPrintf(viewer,"\n");
591: }
592: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
593: return(0);
594: }
596: extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat,PetscViewer);
598: PetscErrorCode MatView_SeqAIJ_ASCII(Mat A,PetscViewer viewer)
599: {
600: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
601: PetscErrorCode ierr;
602: PetscInt i,j,m = A->rmap->n;
603: const char *name;
604: PetscViewerFormat format;
607: if (A->structure_only) {
608: MatView_SeqAIJ_ASCII_structonly(A,viewer);
609: return(0);
610: }
612: PetscViewerGetFormat(viewer,&format);
613: if (format == PETSC_VIEWER_ASCII_MATLAB) {
614: PetscInt nofinalvalue = 0;
615: if (m && ((a->i[m] == a->i[m-1]) || (a->j[a->nz-1] != A->cmap->n-1))) {
616: /* Need a dummy value to ensure the dimension of the matrix. */
617: nofinalvalue = 1;
618: }
619: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
620: PetscViewerASCIIPrintf(viewer,"%% Size = %D %D \n",m,A->cmap->n);
621: PetscViewerASCIIPrintf(viewer,"%% Nonzeros = %D \n",a->nz);
622: #if defined(PETSC_USE_COMPLEX)
623: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,4);\n",a->nz+nofinalvalue);
624: #else
625: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,3);\n",a->nz+nofinalvalue);
626: #endif
627: PetscViewerASCIIPrintf(viewer,"zzz = [\n");
629: for (i=0; i<m; i++) {
630: for (j=a->i[i]; j<a->i[i+1]; j++) {
631: #if defined(PETSC_USE_COMPLEX)
632: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",i+1,a->j[j]+1,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
633: #else
634: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",i+1,a->j[j]+1,(double)a->a[j]);
635: #endif
636: }
637: }
638: if (nofinalvalue) {
639: #if defined(PETSC_USE_COMPLEX)
640: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",m,A->cmap->n,0.,0.);
641: #else
642: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",m,A->cmap->n,0.0);
643: #endif
644: }
645: PetscObjectGetName((PetscObject)A,&name);
646: PetscViewerASCIIPrintf(viewer,"];\n %s = spconvert(zzz);\n",name);
647: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
648: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO) {
649: return(0);
650: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
651: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
652: for (i=0; i<m; i++) {
653: PetscViewerASCIIPrintf(viewer,"row %D:",i);
654: for (j=a->i[i]; j<a->i[i+1]; j++) {
655: #if defined(PETSC_USE_COMPLEX)
656: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
657: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
658: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
659: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
660: } else if (PetscRealPart(a->a[j]) != 0.0) {
661: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
662: }
663: #else
664: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);}
665: #endif
666: }
667: PetscViewerASCIIPrintf(viewer,"\n");
668: }
669: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
670: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
671: PetscInt nzd=0,fshift=1,*sptr;
672: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
673: PetscMalloc1(m+1,&sptr);
674: for (i=0; i<m; i++) {
675: sptr[i] = nzd+1;
676: for (j=a->i[i]; j<a->i[i+1]; j++) {
677: if (a->j[j] >= i) {
678: #if defined(PETSC_USE_COMPLEX)
679: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++;
680: #else
681: if (a->a[j] != 0.0) nzd++;
682: #endif
683: }
684: }
685: }
686: sptr[m] = nzd+1;
687: PetscViewerASCIIPrintf(viewer," %D %D\n\n",m,nzd);
688: for (i=0; i<m+1; i+=6) {
689: if (i+4<m) {
690: PetscViewerASCIIPrintf(viewer," %D %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4],sptr[i+5]);
691: } else if (i+3<m) {
692: PetscViewerASCIIPrintf(viewer," %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4]);
693: } else if (i+2<m) {
694: PetscViewerASCIIPrintf(viewer," %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3]);
695: } else if (i+1<m) {
696: PetscViewerASCIIPrintf(viewer," %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2]);
697: } else if (i<m) {
698: PetscViewerASCIIPrintf(viewer," %D %D\n",sptr[i],sptr[i+1]);
699: } else {
700: PetscViewerASCIIPrintf(viewer," %D\n",sptr[i]);
701: }
702: }
703: PetscViewerASCIIPrintf(viewer,"\n");
704: PetscFree(sptr);
705: for (i=0; i<m; i++) {
706: for (j=a->i[i]; j<a->i[i+1]; j++) {
707: if (a->j[j] >= i) {PetscViewerASCIIPrintf(viewer," %D ",a->j[j]+fshift);}
708: }
709: PetscViewerASCIIPrintf(viewer,"\n");
710: }
711: PetscViewerASCIIPrintf(viewer,"\n");
712: for (i=0; i<m; i++) {
713: for (j=a->i[i]; j<a->i[i+1]; j++) {
714: if (a->j[j] >= i) {
715: #if defined(PETSC_USE_COMPLEX)
716: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) {
717: PetscViewerASCIIPrintf(viewer," %18.16e %18.16e ",(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
718: }
719: #else
720: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," %18.16e ",(double)a->a[j]);}
721: #endif
722: }
723: }
724: PetscViewerASCIIPrintf(viewer,"\n");
725: }
726: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
727: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
728: PetscInt cnt = 0,jcnt;
729: PetscScalar value;
730: #if defined(PETSC_USE_COMPLEX)
731: PetscBool realonly = PETSC_TRUE;
733: for (i=0; i<a->i[m]; i++) {
734: if (PetscImaginaryPart(a->a[i]) != 0.0) {
735: realonly = PETSC_FALSE;
736: break;
737: }
738: }
739: #endif
741: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
742: for (i=0; i<m; i++) {
743: jcnt = 0;
744: for (j=0; j<A->cmap->n; j++) {
745: if (jcnt < a->i[i+1]-a->i[i] && j == a->j[cnt]) {
746: value = a->a[cnt++];
747: jcnt++;
748: } else {
749: value = 0.0;
750: }
751: #if defined(PETSC_USE_COMPLEX)
752: if (realonly) {
753: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)PetscRealPart(value));
754: } else {
755: PetscViewerASCIIPrintf(viewer," %7.5e+%7.5e i ",(double)PetscRealPart(value),(double)PetscImaginaryPart(value));
756: }
757: #else
758: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)value);
759: #endif
760: }
761: PetscViewerASCIIPrintf(viewer,"\n");
762: }
763: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
764: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
765: PetscInt fshift=1;
766: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
767: #if defined(PETSC_USE_COMPLEX)
768: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate complex general\n");
769: #else
770: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate real general\n");
771: #endif
772: PetscViewerASCIIPrintf(viewer,"%D %D %D\n", m, A->cmap->n, a->nz);
773: for (i=0; i<m; i++) {
774: for (j=a->i[i]; j<a->i[i+1]; j++) {
775: #if defined(PETSC_USE_COMPLEX)
776: PetscViewerASCIIPrintf(viewer,"%D %D %g %g\n", i+fshift,a->j[j]+fshift,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
777: #else
778: PetscViewerASCIIPrintf(viewer,"%D %D %g\n", i+fshift, a->j[j]+fshift, (double)a->a[j]);
779: #endif
780: }
781: }
782: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
783: } else {
784: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
785: if (A->factortype) {
786: for (i=0; i<m; i++) {
787: PetscViewerASCIIPrintf(viewer,"row %D:",i);
788: /* L part */
789: for (j=a->i[i]; j<a->i[i+1]; j++) {
790: #if defined(PETSC_USE_COMPLEX)
791: if (PetscImaginaryPart(a->a[j]) > 0.0) {
792: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
793: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
794: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
795: } else {
796: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
797: }
798: #else
799: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
800: #endif
801: }
802: /* diagonal */
803: j = a->diag[i];
804: #if defined(PETSC_USE_COMPLEX)
805: if (PetscImaginaryPart(a->a[j]) > 0.0) {
806: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)PetscImaginaryPart(1.0/a->a[j]));
807: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
808: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)(-PetscImaginaryPart(1.0/a->a[j])));
809: } else {
810: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(1.0/a->a[j]));
811: }
812: #else
813: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)(1.0/a->a[j]));
814: #endif
816: /* U part */
817: for (j=a->diag[i+1]+1; j<a->diag[i]; j++) {
818: #if defined(PETSC_USE_COMPLEX)
819: if (PetscImaginaryPart(a->a[j]) > 0.0) {
820: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
821: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
822: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
823: } else {
824: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
825: }
826: #else
827: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
828: #endif
829: }
830: PetscViewerASCIIPrintf(viewer,"\n");
831: }
832: } else {
833: for (i=0; i<m; i++) {
834: PetscViewerASCIIPrintf(viewer,"row %D:",i);
835: for (j=a->i[i]; j<a->i[i+1]; j++) {
836: #if defined(PETSC_USE_COMPLEX)
837: if (PetscImaginaryPart(a->a[j]) > 0.0) {
838: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
839: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
840: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
841: } else {
842: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
843: }
844: #else
845: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
846: #endif
847: }
848: PetscViewerASCIIPrintf(viewer,"\n");
849: }
850: }
851: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
852: }
853: PetscViewerFlush(viewer);
854: return(0);
855: }
857: #include <petscdraw.h>
858: PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw,void *Aa)
859: {
860: Mat A = (Mat) Aa;
861: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
862: PetscErrorCode ierr;
863: PetscInt i,j,m = A->rmap->n;
864: int color;
865: PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r;
866: PetscViewer viewer;
867: PetscViewerFormat format;
870: PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);
871: PetscViewerGetFormat(viewer,&format);
872: PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);
874: /* loop over matrix elements drawing boxes */
876: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
877: PetscDrawCollectiveBegin(draw);
878: /* Blue for negative, Cyan for zero and Red for positive */
879: color = PETSC_DRAW_BLUE;
880: for (i=0; i<m; i++) {
881: y_l = m - i - 1.0; y_r = y_l + 1.0;
882: for (j=a->i[i]; j<a->i[i+1]; j++) {
883: x_l = a->j[j]; x_r = x_l + 1.0;
884: if (PetscRealPart(a->a[j]) >= 0.) continue;
885: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
886: }
887: }
888: color = PETSC_DRAW_CYAN;
889: for (i=0; i<m; i++) {
890: y_l = m - i - 1.0; y_r = y_l + 1.0;
891: for (j=a->i[i]; j<a->i[i+1]; j++) {
892: x_l = a->j[j]; x_r = x_l + 1.0;
893: if (a->a[j] != 0.) continue;
894: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
895: }
896: }
897: color = PETSC_DRAW_RED;
898: for (i=0; i<m; i++) {
899: y_l = m - i - 1.0; y_r = y_l + 1.0;
900: for (j=a->i[i]; j<a->i[i+1]; j++) {
901: x_l = a->j[j]; x_r = x_l + 1.0;
902: if (PetscRealPart(a->a[j]) <= 0.) continue;
903: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
904: }
905: }
906: PetscDrawCollectiveEnd(draw);
907: } else {
908: /* use contour shading to indicate magnitude of values */
909: /* first determine max of all nonzero values */
910: PetscReal minv = 0.0, maxv = 0.0;
911: PetscInt nz = a->nz, count = 0;
912: PetscDraw popup;
914: for (i=0; i<nz; i++) {
915: if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]);
916: }
917: if (minv >= maxv) maxv = minv + PETSC_SMALL;
918: PetscDrawGetPopup(draw,&popup);
919: PetscDrawScalePopup(popup,minv,maxv);
921: PetscDrawCollectiveBegin(draw);
922: for (i=0; i<m; i++) {
923: y_l = m - i - 1.0;
924: y_r = y_l + 1.0;
925: for (j=a->i[i]; j<a->i[i+1]; j++) {
926: x_l = a->j[j];
927: x_r = x_l + 1.0;
928: color = PetscDrawRealToColor(PetscAbsScalar(a->a[count]),minv,maxv);
929: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
930: count++;
931: }
932: }
933: PetscDrawCollectiveEnd(draw);
934: }
935: return(0);
936: }
938: #include <petscdraw.h>
939: PetscErrorCode MatView_SeqAIJ_Draw(Mat A,PetscViewer viewer)
940: {
942: PetscDraw draw;
943: PetscReal xr,yr,xl,yl,h,w;
944: PetscBool isnull;
947: PetscViewerDrawGetDraw(viewer,0,&draw);
948: PetscDrawIsNull(draw,&isnull);
949: if (isnull) return(0);
951: xr = A->cmap->n; yr = A->rmap->n; h = yr/10.0; w = xr/10.0;
952: xr += w; yr += h; xl = -w; yl = -h;
953: PetscDrawSetCoordinates(draw,xl,yl,xr,yr);
954: PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);
955: PetscDrawZoom(draw,MatView_SeqAIJ_Draw_Zoom,A);
956: PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);
957: PetscDrawSave(draw);
958: return(0);
959: }
961: PetscErrorCode MatView_SeqAIJ(Mat A,PetscViewer viewer)
962: {
964: PetscBool iascii,isbinary,isdraw;
967: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
968: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
969: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
970: if (iascii) {
971: MatView_SeqAIJ_ASCII(A,viewer);
972: } else if (isbinary) {
973: MatView_SeqAIJ_Binary(A,viewer);
974: } else if (isdraw) {
975: MatView_SeqAIJ_Draw(A,viewer);
976: }
977: MatView_SeqAIJ_Inode(A,viewer);
978: return(0);
979: }
981: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A,MatAssemblyType mode)
982: {
983: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
985: PetscInt fshift = 0,i,j,*ai = a->i,*aj = a->j,*imax = a->imax;
986: PetscInt m = A->rmap->n,*ip,N,*ailen = a->ilen,rmax = 0;
987: MatScalar *aa = a->a,*ap;
988: PetscReal ratio = 0.6;
991: if (mode == MAT_FLUSH_ASSEMBLY) return(0);
993: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
994: for (i=1; i<m; i++) {
995: /* move each row back by the amount of empty slots (fshift) before it*/
996: fshift += imax[i-1] - ailen[i-1];
997: rmax = PetscMax(rmax,ailen[i]);
998: if (fshift) {
999: ip = aj + ai[i];
1000: ap = aa + ai[i];
1001: N = ailen[i];
1002: for (j=0; j<N; j++) {
1003: ip[j-fshift] = ip[j];
1004: if (!A->structure_only) ap[j-fshift] = ap[j];
1005: }
1006: }
1007: ai[i] = ai[i-1] + ailen[i-1];
1008: }
1009: if (m) {
1010: fshift += imax[m-1] - ailen[m-1];
1011: ai[m] = ai[m-1] + ailen[m-1];
1012: }
1014: /* reset ilen and imax for each row */
1015: a->nonzerorowcnt = 0;
1016: if (A->structure_only) {
1017: PetscFree2(a->imax,a->ilen);
1018: } else { /* !A->structure_only */
1019: for (i=0; i<m; i++) {
1020: ailen[i] = imax[i] = ai[i+1] - ai[i];
1021: a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0);
1022: }
1023: }
1024: a->nz = ai[m];
1025: if (fshift && a->nounused == -1) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB, "Unused space detected in matrix: %D X %D, %D unneeded", m, A->cmap->n, fshift);
1027: MatMarkDiagonal_SeqAIJ(A);
1028: PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap->n,fshift,a->nz);
1029: PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);
1030: PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);
1032: A->info.mallocs += a->reallocs;
1033: a->reallocs = 0;
1034: A->info.nz_unneeded = (PetscReal)fshift;
1035: a->rmax = rmax;
1037: if (!A->structure_only) {
1038: MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,m,ratio);
1039: }
1040: MatAssemblyEnd_SeqAIJ_Inode(A,mode);
1041: MatSeqAIJInvalidateDiagonal(A);
1042: return(0);
1043: }
1045: PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1046: {
1047: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1048: PetscInt i,nz = a->nz;
1049: MatScalar *aa = a->a;
1053: for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]);
1054: MatSeqAIJInvalidateDiagonal(A);
1055: return(0);
1056: }
1058: PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1059: {
1060: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1061: PetscInt i,nz = a->nz;
1062: MatScalar *aa = a->a;
1066: for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1067: MatSeqAIJInvalidateDiagonal(A);
1068: return(0);
1069: }
1071: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1072: {
1073: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1077: PetscMemzero(a->a,(a->i[A->rmap->n])*sizeof(PetscScalar));
1078: MatSeqAIJInvalidateDiagonal(A);
1079: return(0);
1080: }
1082: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1083: {
1084: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1088: #if defined(PETSC_USE_LOG)
1089: PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->n,A->cmap->n,a->nz);
1090: #endif
1091: MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);
1092: ISDestroy(&a->row);
1093: ISDestroy(&a->col);
1094: PetscFree(a->diag);
1095: PetscFree(a->ibdiag);
1096: PetscFree2(a->imax,a->ilen);
1097: PetscFree3(a->idiag,a->mdiag,a->ssor_work);
1098: PetscFree(a->solve_work);
1099: ISDestroy(&a->icol);
1100: PetscFree(a->saved_values);
1101: ISColoringDestroy(&a->coloring);
1102: PetscFree2(a->compressedrow.i,a->compressedrow.rindex);
1103: PetscFree(a->matmult_abdense);
1105: MatDestroy_SeqAIJ_Inode(A);
1106: PetscFree(A->data);
1108: PetscObjectChangeTypeName((PetscObject)A,0);
1109: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetColumnIndices_C",NULL);
1110: PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);
1111: PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);
1112: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsbaij_C",NULL);
1113: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqbaij_C",NULL);
1114: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijperm_C",NULL);
1115: #if defined(PETSC_HAVE_ELEMENTAL)
1116: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_elemental_C",NULL);
1117: #endif
1118: #if defined(PETSC_HAVE_HYPRE)
1119: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_hypre_C",NULL);
1120: PetscObjectComposeFunction((PetscObject)A,"MatMatMatMult_transpose_seqaij_seqaij_C",NULL);
1121: #endif
1122: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqdense_C",NULL);
1123: PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);
1124: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);
1125: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);
1126: PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);
1127: return(0);
1128: }
1130: PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg)
1131: {
1132: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1136: switch (op) {
1137: case MAT_ROW_ORIENTED:
1138: a->roworiented = flg;
1139: break;
1140: case MAT_KEEP_NONZERO_PATTERN:
1141: a->keepnonzeropattern = flg;
1142: break;
1143: case MAT_NEW_NONZERO_LOCATIONS:
1144: a->nonew = (flg ? 0 : 1);
1145: break;
1146: case MAT_NEW_NONZERO_LOCATION_ERR:
1147: a->nonew = (flg ? -1 : 0);
1148: break;
1149: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1150: a->nonew = (flg ? -2 : 0);
1151: break;
1152: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1153: a->nounused = (flg ? -1 : 0);
1154: break;
1155: case MAT_IGNORE_ZERO_ENTRIES:
1156: a->ignorezeroentries = flg;
1157: break;
1158: case MAT_SPD:
1159: case MAT_SYMMETRIC:
1160: case MAT_STRUCTURALLY_SYMMETRIC:
1161: case MAT_HERMITIAN:
1162: case MAT_SYMMETRY_ETERNAL:
1163: case MAT_STRUCTURE_ONLY:
1164: /* These options are handled directly by MatSetOption() */
1165: break;
1166: case MAT_NEW_DIAGONALS:
1167: case MAT_IGNORE_OFF_PROC_ENTRIES:
1168: case MAT_USE_HASH_TABLE:
1169: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1170: break;
1171: case MAT_USE_INODES:
1172: /* Not an error because MatSetOption_SeqAIJ_Inode handles this one */
1173: break;
1174: case MAT_SUBMAT_SINGLEIS:
1175: A->submat_singleis = flg;
1176: break;
1177: default:
1178: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1179: }
1180: MatSetOption_SeqAIJ_Inode(A,op,flg);
1181: return(0);
1182: }
1184: PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v)
1185: {
1186: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1188: PetscInt i,j,n,*ai=a->i,*aj=a->j,nz;
1189: PetscScalar *aa=a->a,*x,zero=0.0;
1192: VecGetLocalSize(v,&n);
1193: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
1195: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1196: PetscInt *diag=a->diag;
1197: VecGetArray(v,&x);
1198: for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]];
1199: VecRestoreArray(v,&x);
1200: return(0);
1201: }
1203: VecSet(v,zero);
1204: VecGetArray(v,&x);
1205: for (i=0; i<n; i++) {
1206: nz = ai[i+1] - ai[i];
1207: if (!nz) x[i] = 0.0;
1208: for (j=ai[i]; j<ai[i+1]; j++) {
1209: if (aj[j] == i) {
1210: x[i] = aa[j];
1211: break;
1212: }
1213: }
1214: }
1215: VecRestoreArray(v,&x);
1216: return(0);
1217: }
1219: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1220: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy)
1221: {
1222: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1223: PetscScalar *y;
1224: const PetscScalar *x;
1225: PetscErrorCode ierr;
1226: PetscInt m = A->rmap->n;
1227: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1228: const MatScalar *v;
1229: PetscScalar alpha;
1230: PetscInt n,i,j;
1231: const PetscInt *idx,*ii,*ridx=NULL;
1232: Mat_CompressedRow cprow = a->compressedrow;
1233: PetscBool usecprow = cprow.use;
1234: #endif
1237: if (zz != yy) {VecCopy(zz,yy);}
1238: VecGetArrayRead(xx,&x);
1239: VecGetArray(yy,&y);
1241: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1242: fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y);
1243: #else
1244: if (usecprow) {
1245: m = cprow.nrows;
1246: ii = cprow.i;
1247: ridx = cprow.rindex;
1248: } else {
1249: ii = a->i;
1250: }
1251: for (i=0; i<m; i++) {
1252: idx = a->j + ii[i];
1253: v = a->a + ii[i];
1254: n = ii[i+1] - ii[i];
1255: if (usecprow) {
1256: alpha = x[ridx[i]];
1257: } else {
1258: alpha = x[i];
1259: }
1260: for (j=0; j<n; j++) y[idx[j]] += alpha*v[j];
1261: }
1262: #endif
1263: PetscLogFlops(2.0*a->nz);
1264: VecRestoreArrayRead(xx,&x);
1265: VecRestoreArray(yy,&y);
1266: return(0);
1267: }
1269: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy)
1270: {
1274: VecSet(yy,0.0);
1275: MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);
1276: return(0);
1277: }
1279: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1281: PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy)
1282: {
1283: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1284: PetscScalar *y;
1285: const PetscScalar *x;
1286: const MatScalar *aa;
1287: PetscErrorCode ierr;
1288: PetscInt m=A->rmap->n;
1289: const PetscInt *aj,*ii,*ridx=NULL;
1290: PetscInt n,i;
1291: PetscScalar sum;
1292: PetscBool usecprow=a->compressedrow.use;
1294: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1295: #pragma disjoint(*x,*y,*aa)
1296: #endif
1299: VecGetArrayRead(xx,&x);
1300: VecGetArray(yy,&y);
1301: ii = a->i;
1302: if (usecprow) { /* use compressed row format */
1303: PetscMemzero(y,m*sizeof(PetscScalar));
1304: m = a->compressedrow.nrows;
1305: ii = a->compressedrow.i;
1306: ridx = a->compressedrow.rindex;
1307: for (i=0; i<m; i++) {
1308: n = ii[i+1] - ii[i];
1309: aj = a->j + ii[i];
1310: aa = a->a + ii[i];
1311: sum = 0.0;
1312: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1313: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1314: y[*ridx++] = sum;
1315: }
1316: } else { /* do not use compressed row format */
1317: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1318: aj = a->j;
1319: aa = a->a;
1320: fortranmultaij_(&m,x,ii,aj,aa,y);
1321: #else
1322: for (i=0; i<m; i++) {
1323: n = ii[i+1] - ii[i];
1324: aj = a->j + ii[i];
1325: aa = a->a + ii[i];
1326: sum = 0.0;
1327: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1328: y[i] = sum;
1329: }
1330: #endif
1331: }
1332: PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);
1333: VecRestoreArrayRead(xx,&x);
1334: VecRestoreArray(yy,&y);
1335: return(0);
1336: }
1338: PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy)
1339: {
1340: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1341: PetscScalar *y;
1342: const PetscScalar *x;
1343: const MatScalar *aa;
1344: PetscErrorCode ierr;
1345: PetscInt m=A->rmap->n;
1346: const PetscInt *aj,*ii,*ridx=NULL;
1347: PetscInt n,i,nonzerorow=0;
1348: PetscScalar sum;
1349: PetscBool usecprow=a->compressedrow.use;
1351: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1352: #pragma disjoint(*x,*y,*aa)
1353: #endif
1356: VecGetArrayRead(xx,&x);
1357: VecGetArray(yy,&y);
1358: if (usecprow) { /* use compressed row format */
1359: m = a->compressedrow.nrows;
1360: ii = a->compressedrow.i;
1361: ridx = a->compressedrow.rindex;
1362: for (i=0; i<m; i++) {
1363: n = ii[i+1] - ii[i];
1364: aj = a->j + ii[i];
1365: aa = a->a + ii[i];
1366: sum = 0.0;
1367: nonzerorow += (n>0);
1368: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1369: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1370: y[*ridx++] = sum;
1371: }
1372: } else { /* do not use compressed row format */
1373: ii = a->i;
1374: for (i=0; i<m; i++) {
1375: n = ii[i+1] - ii[i];
1376: aj = a->j + ii[i];
1377: aa = a->a + ii[i];
1378: sum = 0.0;
1379: nonzerorow += (n>0);
1380: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1381: y[i] = sum;
1382: }
1383: }
1384: PetscLogFlops(2.0*a->nz - nonzerorow);
1385: VecRestoreArrayRead(xx,&x);
1386: VecRestoreArray(yy,&y);
1387: return(0);
1388: }
1390: PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1391: {
1392: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1393: PetscScalar *y,*z;
1394: const PetscScalar *x;
1395: const MatScalar *aa;
1396: PetscErrorCode ierr;
1397: PetscInt m = A->rmap->n,*aj,*ii;
1398: PetscInt n,i,*ridx=NULL;
1399: PetscScalar sum;
1400: PetscBool usecprow=a->compressedrow.use;
1403: VecGetArrayRead(xx,&x);
1404: VecGetArrayPair(yy,zz,&y,&z);
1405: if (usecprow) { /* use compressed row format */
1406: if (zz != yy) {
1407: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1408: }
1409: m = a->compressedrow.nrows;
1410: ii = a->compressedrow.i;
1411: ridx = a->compressedrow.rindex;
1412: for (i=0; i<m; i++) {
1413: n = ii[i+1] - ii[i];
1414: aj = a->j + ii[i];
1415: aa = a->a + ii[i];
1416: sum = y[*ridx];
1417: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1418: z[*ridx++] = sum;
1419: }
1420: } else { /* do not use compressed row format */
1421: ii = a->i;
1422: for (i=0; i<m; i++) {
1423: n = ii[i+1] - ii[i];
1424: aj = a->j + ii[i];
1425: aa = a->a + ii[i];
1426: sum = y[i];
1427: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1428: z[i] = sum;
1429: }
1430: }
1431: PetscLogFlops(2.0*a->nz);
1432: VecRestoreArrayRead(xx,&x);
1433: VecRestoreArrayPair(yy,zz,&y,&z);
1434: return(0);
1435: }
1437: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1438: PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1439: {
1440: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1441: PetscScalar *y,*z;
1442: const PetscScalar *x;
1443: const MatScalar *aa;
1444: PetscErrorCode ierr;
1445: const PetscInt *aj,*ii,*ridx=NULL;
1446: PetscInt m = A->rmap->n,n,i;
1447: PetscScalar sum;
1448: PetscBool usecprow=a->compressedrow.use;
1451: VecGetArrayRead(xx,&x);
1452: VecGetArrayPair(yy,zz,&y,&z);
1453: if (usecprow) { /* use compressed row format */
1454: if (zz != yy) {
1455: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1456: }
1457: m = a->compressedrow.nrows;
1458: ii = a->compressedrow.i;
1459: ridx = a->compressedrow.rindex;
1460: for (i=0; i<m; i++) {
1461: n = ii[i+1] - ii[i];
1462: aj = a->j + ii[i];
1463: aa = a->a + ii[i];
1464: sum = y[*ridx];
1465: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1466: z[*ridx++] = sum;
1467: }
1468: } else { /* do not use compressed row format */
1469: ii = a->i;
1470: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1471: aj = a->j;
1472: aa = a->a;
1473: fortranmultaddaij_(&m,x,ii,aj,aa,y,z);
1474: #else
1475: for (i=0; i<m; i++) {
1476: n = ii[i+1] - ii[i];
1477: aj = a->j + ii[i];
1478: aa = a->a + ii[i];
1479: sum = y[i];
1480: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1481: z[i] = sum;
1482: }
1483: #endif
1484: }
1485: PetscLogFlops(2.0*a->nz);
1486: VecRestoreArrayRead(xx,&x);
1487: VecRestoreArrayPair(yy,zz,&y,&z);
1488: #if defined(PETSC_HAVE_CUSP)
1489: /*
1490: VecView(xx,0);
1491: VecView(zz,0);
1492: MatView(A,0);
1493: */
1494: #endif
1495: return(0);
1496: }
1498: /*
1499: Adds diagonal pointers to sparse matrix structure.
1500: */
1501: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1502: {
1503: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1505: PetscInt i,j,m = A->rmap->n;
1508: if (!a->diag) {
1509: PetscMalloc1(m,&a->diag);
1510: PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));
1511: }
1512: for (i=0; i<A->rmap->n; i++) {
1513: a->diag[i] = a->i[i+1];
1514: for (j=a->i[i]; j<a->i[i+1]; j++) {
1515: if (a->j[j] == i) {
1516: a->diag[i] = j;
1517: break;
1518: }
1519: }
1520: }
1521: return(0);
1522: }
1524: /*
1525: Checks for missing diagonals
1526: */
1527: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d)
1528: {
1529: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1530: PetscInt *diag,*ii = a->i,i;
1533: *missing = PETSC_FALSE;
1534: if (A->rmap->n > 0 && !ii) {
1535: *missing = PETSC_TRUE;
1536: if (d) *d = 0;
1537: PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");
1538: } else {
1539: diag = a->diag;
1540: for (i=0; i<A->rmap->n; i++) {
1541: if (diag[i] >= ii[i+1]) {
1542: *missing = PETSC_TRUE;
1543: if (d) *d = i;
1544: PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);
1545: break;
1546: }
1547: }
1548: }
1549: return(0);
1550: }
1552: /*
1553: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1554: */
1555: PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift)
1556: {
1557: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
1559: PetscInt i,*diag,m = A->rmap->n;
1560: MatScalar *v = a->a;
1561: PetscScalar *idiag,*mdiag;
1564: if (a->idiagvalid) return(0);
1565: MatMarkDiagonal_SeqAIJ(A);
1566: diag = a->diag;
1567: if (!a->idiag) {
1568: PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);
1569: PetscLogObjectMemory((PetscObject)A, 3*m*sizeof(PetscScalar));
1570: v = a->a;
1571: }
1572: mdiag = a->mdiag;
1573: idiag = a->idiag;
1575: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1576: for (i=0; i<m; i++) {
1577: mdiag[i] = v[diag[i]];
1578: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1579: if (PetscRealPart(fshift)) {
1580: PetscInfo1(A,"Zero diagonal on row %D\n",i);
1581: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1582: A->factorerror_zeropivot_value = 0.0;
1583: A->factorerror_zeropivot_row = i;
1584: } SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i);
1585: }
1586: idiag[i] = 1.0/v[diag[i]];
1587: }
1588: PetscLogFlops(m);
1589: } else {
1590: for (i=0; i<m; i++) {
1591: mdiag[i] = v[diag[i]];
1592: idiag[i] = omega/(fshift + v[diag[i]]);
1593: }
1594: PetscLogFlops(2.0*m);
1595: }
1596: a->idiagvalid = PETSC_TRUE;
1597: return(0);
1598: }
1600: #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h>
1601: PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1602: {
1603: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1604: PetscScalar *x,d,sum,*t,scale;
1605: const MatScalar *v,*idiag=0,*mdiag;
1606: const PetscScalar *b, *bs,*xb, *ts;
1607: PetscErrorCode ierr;
1608: PetscInt n,m = A->rmap->n,i;
1609: const PetscInt *idx,*diag;
1612: its = its*lits;
1614: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1615: if (!a->idiagvalid) {MatInvertDiagonal_SeqAIJ(A,omega,fshift);}
1616: a->fshift = fshift;
1617: a->omega = omega;
1619: diag = a->diag;
1620: t = a->ssor_work;
1621: idiag = a->idiag;
1622: mdiag = a->mdiag;
1624: VecGetArray(xx,&x);
1625: VecGetArrayRead(bb,&b);
1626: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1627: if (flag == SOR_APPLY_UPPER) {
1628: /* apply (U + D/omega) to the vector */
1629: bs = b;
1630: for (i=0; i<m; i++) {
1631: d = fshift + mdiag[i];
1632: n = a->i[i+1] - diag[i] - 1;
1633: idx = a->j + diag[i] + 1;
1634: v = a->a + diag[i] + 1;
1635: sum = b[i]*d/omega;
1636: PetscSparseDensePlusDot(sum,bs,v,idx,n);
1637: x[i] = sum;
1638: }
1639: VecRestoreArray(xx,&x);
1640: VecRestoreArrayRead(bb,&b);
1641: PetscLogFlops(a->nz);
1642: return(0);
1643: }
1645: if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented");
1646: else if (flag & SOR_EISENSTAT) {
1647: /* Let A = L + U + D; where L is lower trianglar,
1648: U is upper triangular, E = D/omega; This routine applies
1650: (L + E)^{-1} A (U + E)^{-1}
1652: to a vector efficiently using Eisenstat's trick.
1653: */
1654: scale = (2.0/omega) - 1.0;
1656: /* x = (E + U)^{-1} b */
1657: for (i=m-1; i>=0; i--) {
1658: n = a->i[i+1] - diag[i] - 1;
1659: idx = a->j + diag[i] + 1;
1660: v = a->a + diag[i] + 1;
1661: sum = b[i];
1662: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1663: x[i] = sum*idiag[i];
1664: }
1666: /* t = b - (2*E - D)x */
1667: v = a->a;
1668: for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i];
1670: /* t = (E + L)^{-1}t */
1671: ts = t;
1672: diag = a->diag;
1673: for (i=0; i<m; i++) {
1674: n = diag[i] - a->i[i];
1675: idx = a->j + a->i[i];
1676: v = a->a + a->i[i];
1677: sum = t[i];
1678: PetscSparseDenseMinusDot(sum,ts,v,idx,n);
1679: t[i] = sum*idiag[i];
1680: /* x = x + t */
1681: x[i] += t[i];
1682: }
1684: PetscLogFlops(6.0*m-1 + 2.0*a->nz);
1685: VecRestoreArray(xx,&x);
1686: VecRestoreArrayRead(bb,&b);
1687: return(0);
1688: }
1689: if (flag & SOR_ZERO_INITIAL_GUESS) {
1690: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1691: for (i=0; i<m; i++) {
1692: n = diag[i] - a->i[i];
1693: idx = a->j + a->i[i];
1694: v = a->a + a->i[i];
1695: sum = b[i];
1696: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1697: t[i] = sum;
1698: x[i] = sum*idiag[i];
1699: }
1700: xb = t;
1701: PetscLogFlops(a->nz);
1702: } else xb = b;
1703: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1704: for (i=m-1; i>=0; i--) {
1705: n = a->i[i+1] - diag[i] - 1;
1706: idx = a->j + diag[i] + 1;
1707: v = a->a + diag[i] + 1;
1708: sum = xb[i];
1709: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1710: if (xb == b) {
1711: x[i] = sum*idiag[i];
1712: } else {
1713: x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1714: }
1715: }
1716: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1717: }
1718: its--;
1719: }
1720: while (its--) {
1721: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1722: for (i=0; i<m; i++) {
1723: /* lower */
1724: n = diag[i] - a->i[i];
1725: idx = a->j + a->i[i];
1726: v = a->a + a->i[i];
1727: sum = b[i];
1728: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1729: t[i] = sum; /* save application of the lower-triangular part */
1730: /* upper */
1731: n = a->i[i+1] - diag[i] - 1;
1732: idx = a->j + diag[i] + 1;
1733: v = a->a + diag[i] + 1;
1734: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1735: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1736: }
1737: xb = t;
1738: PetscLogFlops(2.0*a->nz);
1739: } else xb = b;
1740: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1741: for (i=m-1; i>=0; i--) {
1742: sum = xb[i];
1743: if (xb == b) {
1744: /* whole matrix (no checkpointing available) */
1745: n = a->i[i+1] - a->i[i];
1746: idx = a->j + a->i[i];
1747: v = a->a + a->i[i];
1748: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1749: x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i];
1750: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
1751: n = a->i[i+1] - diag[i] - 1;
1752: idx = a->j + diag[i] + 1;
1753: v = a->a + diag[i] + 1;
1754: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1755: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1756: }
1757: }
1758: if (xb == b) {
1759: PetscLogFlops(2.0*a->nz);
1760: } else {
1761: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1762: }
1763: }
1764: }
1765: VecRestoreArray(xx,&x);
1766: VecRestoreArrayRead(bb,&b);
1767: return(0);
1768: }
1771: PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info)
1772: {
1773: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1776: info->block_size = 1.0;
1777: info->nz_allocated = (double)a->maxnz;
1778: info->nz_used = (double)a->nz;
1779: info->nz_unneeded = (double)(a->maxnz - a->nz);
1780: info->assemblies = (double)A->num_ass;
1781: info->mallocs = (double)A->info.mallocs;
1782: info->memory = ((PetscObject)A)->mem;
1783: if (A->factortype) {
1784: info->fill_ratio_given = A->info.fill_ratio_given;
1785: info->fill_ratio_needed = A->info.fill_ratio_needed;
1786: info->factor_mallocs = A->info.factor_mallocs;
1787: } else {
1788: info->fill_ratio_given = 0;
1789: info->fill_ratio_needed = 0;
1790: info->factor_mallocs = 0;
1791: }
1792: return(0);
1793: }
1795: PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1796: {
1797: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1798: PetscInt i,m = A->rmap->n - 1;
1799: PetscErrorCode ierr;
1800: const PetscScalar *xx;
1801: PetscScalar *bb;
1802: PetscInt d = 0;
1805: if (x && b) {
1806: VecGetArrayRead(x,&xx);
1807: VecGetArray(b,&bb);
1808: for (i=0; i<N; i++) {
1809: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1810: bb[rows[i]] = diag*xx[rows[i]];
1811: }
1812: VecRestoreArrayRead(x,&xx);
1813: VecRestoreArray(b,&bb);
1814: }
1816: if (a->keepnonzeropattern) {
1817: for (i=0; i<N; i++) {
1818: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1819: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1820: }
1821: if (diag != 0.0) {
1822: for (i=0; i<N; i++) {
1823: d = rows[i];
1824: if (a->diag[d] >= a->i[d+1]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in the zeroed row %D",d);
1825: }
1826: for (i=0; i<N; i++) {
1827: a->a[a->diag[rows[i]]] = diag;
1828: }
1829: }
1830: } else {
1831: if (diag != 0.0) {
1832: for (i=0; i<N; i++) {
1833: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1834: if (a->ilen[rows[i]] > 0) {
1835: a->ilen[rows[i]] = 1;
1836: a->a[a->i[rows[i]]] = diag;
1837: a->j[a->i[rows[i]]] = rows[i];
1838: } else { /* in case row was completely empty */
1839: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
1840: }
1841: }
1842: } else {
1843: for (i=0; i<N; i++) {
1844: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1845: a->ilen[rows[i]] = 0;
1846: }
1847: }
1848: A->nonzerostate++;
1849: }
1850: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
1851: return(0);
1852: }
1854: PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1855: {
1856: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1857: PetscInt i,j,m = A->rmap->n - 1,d = 0;
1858: PetscErrorCode ierr;
1859: PetscBool missing,*zeroed,vecs = PETSC_FALSE;
1860: const PetscScalar *xx;
1861: PetscScalar *bb;
1864: if (x && b) {
1865: VecGetArrayRead(x,&xx);
1866: VecGetArray(b,&bb);
1867: vecs = PETSC_TRUE;
1868: }
1869: PetscCalloc1(A->rmap->n,&zeroed);
1870: for (i=0; i<N; i++) {
1871: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1872: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1874: zeroed[rows[i]] = PETSC_TRUE;
1875: }
1876: for (i=0; i<A->rmap->n; i++) {
1877: if (!zeroed[i]) {
1878: for (j=a->i[i]; j<a->i[i+1]; j++) {
1879: if (zeroed[a->j[j]]) {
1880: if (vecs) bb[i] -= a->a[j]*xx[a->j[j]];
1881: a->a[j] = 0.0;
1882: }
1883: }
1884: } else if (vecs) bb[i] = diag*xx[i];
1885: }
1886: if (x && b) {
1887: VecRestoreArrayRead(x,&xx);
1888: VecRestoreArray(b,&bb);
1889: }
1890: PetscFree(zeroed);
1891: if (diag != 0.0) {
1892: MatMissingDiagonal_SeqAIJ(A,&missing,&d);
1893: if (missing) {
1894: if (a->nonew) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D",d);
1895: else {
1896: for (i=0; i<N; i++) {
1897: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
1898: }
1899: }
1900: } else {
1901: for (i=0; i<N; i++) {
1902: a->a[a->diag[rows[i]]] = diag;
1903: }
1904: }
1905: }
1906: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
1907: return(0);
1908: }
1910: PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1911: {
1912: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1913: PetscInt *itmp;
1916: if (row < 0 || row >= A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range",row);
1918: *nz = a->i[row+1] - a->i[row];
1919: if (v) *v = a->a + a->i[row];
1920: if (idx) {
1921: itmp = a->j + a->i[row];
1922: if (*nz) *idx = itmp;
1923: else *idx = 0;
1924: }
1925: return(0);
1926: }
1928: /* remove this function? */
1929: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1930: {
1932: return(0);
1933: }
1935: PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm)
1936: {
1937: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1938: MatScalar *v = a->a;
1939: PetscReal sum = 0.0;
1941: PetscInt i,j;
1944: if (type == NORM_FROBENIUS) {
1945: #if defined(PETSC_USE_REAL___FP16)
1946: PetscBLASInt one = 1,nz = a->nz;
1947: *nrm = BLASnrm2_(&nz,v,&one);
1948: #else
1949: for (i=0; i<a->nz; i++) {
1950: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1951: }
1952: *nrm = PetscSqrtReal(sum);
1953: #endif
1954: PetscLogFlops(2*a->nz);
1955: } else if (type == NORM_1) {
1956: PetscReal *tmp;
1957: PetscInt *jj = a->j;
1958: PetscCalloc1(A->cmap->n+1,&tmp);
1959: *nrm = 0.0;
1960: for (j=0; j<a->nz; j++) {
1961: tmp[*jj++] += PetscAbsScalar(*v); v++;
1962: }
1963: for (j=0; j<A->cmap->n; j++) {
1964: if (tmp[j] > *nrm) *nrm = tmp[j];
1965: }
1966: PetscFree(tmp);
1967: PetscLogFlops(PetscMax(a->nz-1,0));
1968: } else if (type == NORM_INFINITY) {
1969: *nrm = 0.0;
1970: for (j=0; j<A->rmap->n; j++) {
1971: v = a->a + a->i[j];
1972: sum = 0.0;
1973: for (i=0; i<a->i[j+1]-a->i[j]; i++) {
1974: sum += PetscAbsScalar(*v); v++;
1975: }
1976: if (sum > *nrm) *nrm = sum;
1977: }
1978: PetscLogFlops(PetscMax(a->nz-1,0));
1979: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm");
1980: return(0);
1981: }
1983: /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */
1984: PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B)
1985: {
1987: PetscInt i,j,anzj;
1988: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b;
1989: PetscInt an=A->cmap->N,am=A->rmap->N;
1990: PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j;
1993: /* Allocate space for symbolic transpose info and work array */
1994: PetscCalloc1(an+1,&ati);
1995: PetscMalloc1(ai[am],&atj);
1996: PetscMalloc1(an,&atfill);
1998: /* Walk through aj and count ## of non-zeros in each row of A^T. */
1999: /* Note: offset by 1 for fast conversion into csr format. */
2000: for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1;
2001: /* Form ati for csr format of A^T. */
2002: for (i=0;i<an;i++) ati[i+1] += ati[i];
2004: /* Copy ati into atfill so we have locations of the next free space in atj */
2005: PetscMemcpy(atfill,ati,an*sizeof(PetscInt));
2007: /* Walk through A row-wise and mark nonzero entries of A^T. */
2008: for (i=0;i<am;i++) {
2009: anzj = ai[i+1] - ai[i];
2010: for (j=0;j<anzj;j++) {
2011: atj[atfill[*aj]] = i;
2012: atfill[*aj++] += 1;
2013: }
2014: }
2016: /* Clean up temporary space and complete requests. */
2017: PetscFree(atfill);
2018: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);
2019: MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2021: b = (Mat_SeqAIJ*)((*B)->data);
2022: b->free_a = PETSC_FALSE;
2023: b->free_ij = PETSC_TRUE;
2024: b->nonew = 0;
2025: return(0);
2026: }
2028: PetscErrorCode MatTranspose_SeqAIJ(Mat A,MatReuse reuse,Mat *B)
2029: {
2030: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2031: Mat C;
2033: PetscInt i,*aj = a->j,*ai = a->i,m = A->rmap->n,len,*col;
2034: MatScalar *array = a->a;
2037: if (reuse == MAT_INPLACE_MATRIX && m != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
2039: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
2040: PetscCalloc1(1+A->cmap->n,&col);
2042: for (i=0; i<ai[m]; i++) col[aj[i]] += 1;
2043: MatCreate(PetscObjectComm((PetscObject)A),&C);
2044: MatSetSizes(C,A->cmap->n,m,A->cmap->n,m);
2045: MatSetBlockSizes(C,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2046: MatSetType(C,((PetscObject)A)->type_name);
2047: MatSeqAIJSetPreallocation_SeqAIJ(C,0,col);
2048: PetscFree(col);
2049: } else {
2050: C = *B;
2051: }
2053: for (i=0; i<m; i++) {
2054: len = ai[i+1]-ai[i];
2055: MatSetValues_SeqAIJ(C,len,aj,1,&i,array,INSERT_VALUES);
2056: array += len;
2057: aj += len;
2058: }
2059: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2060: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2062: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2063: *B = C;
2064: } else {
2065: MatHeaderMerge(A,&C);
2066: }
2067: return(0);
2068: }
2070: PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2071: {
2072: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2073: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2074: MatScalar *va,*vb;
2076: PetscInt ma,na,mb,nb, i;
2079: MatGetSize(A,&ma,&na);
2080: MatGetSize(B,&mb,&nb);
2081: if (ma!=nb || na!=mb) {
2082: *f = PETSC_FALSE;
2083: return(0);
2084: }
2085: aii = aij->i; bii = bij->i;
2086: adx = aij->j; bdx = bij->j;
2087: va = aij->a; vb = bij->a;
2088: PetscMalloc1(ma,&aptr);
2089: PetscMalloc1(mb,&bptr);
2090: for (i=0; i<ma; i++) aptr[i] = aii[i];
2091: for (i=0; i<mb; i++) bptr[i] = bii[i];
2093: *f = PETSC_TRUE;
2094: for (i=0; i<ma; i++) {
2095: while (aptr[i]<aii[i+1]) {
2096: PetscInt idc,idr;
2097: PetscScalar vc,vr;
2098: /* column/row index/value */
2099: idc = adx[aptr[i]];
2100: idr = bdx[bptr[idc]];
2101: vc = va[aptr[i]];
2102: vr = vb[bptr[idc]];
2103: if (i!=idr || PetscAbsScalar(vc-vr) > tol) {
2104: *f = PETSC_FALSE;
2105: goto done;
2106: } else {
2107: aptr[i]++;
2108: if (B || i!=idc) bptr[idc]++;
2109: }
2110: }
2111: }
2112: done:
2113: PetscFree(aptr);
2114: PetscFree(bptr);
2115: return(0);
2116: }
2118: PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2119: {
2120: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2121: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2122: MatScalar *va,*vb;
2124: PetscInt ma,na,mb,nb, i;
2127: MatGetSize(A,&ma,&na);
2128: MatGetSize(B,&mb,&nb);
2129: if (ma!=nb || na!=mb) {
2130: *f = PETSC_FALSE;
2131: return(0);
2132: }
2133: aii = aij->i; bii = bij->i;
2134: adx = aij->j; bdx = bij->j;
2135: va = aij->a; vb = bij->a;
2136: PetscMalloc1(ma,&aptr);
2137: PetscMalloc1(mb,&bptr);
2138: for (i=0; i<ma; i++) aptr[i] = aii[i];
2139: for (i=0; i<mb; i++) bptr[i] = bii[i];
2141: *f = PETSC_TRUE;
2142: for (i=0; i<ma; i++) {
2143: while (aptr[i]<aii[i+1]) {
2144: PetscInt idc,idr;
2145: PetscScalar vc,vr;
2146: /* column/row index/value */
2147: idc = adx[aptr[i]];
2148: idr = bdx[bptr[idc]];
2149: vc = va[aptr[i]];
2150: vr = vb[bptr[idc]];
2151: if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) {
2152: *f = PETSC_FALSE;
2153: goto done;
2154: } else {
2155: aptr[i]++;
2156: if (B || i!=idc) bptr[idc]++;
2157: }
2158: }
2159: }
2160: done:
2161: PetscFree(aptr);
2162: PetscFree(bptr);
2163: return(0);
2164: }
2166: PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2167: {
2171: MatIsTranspose_SeqAIJ(A,A,tol,f);
2172: return(0);
2173: }
2175: PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2176: {
2180: MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);
2181: return(0);
2182: }
2184: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr)
2185: {
2186: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2187: const PetscScalar *l,*r;
2188: PetscScalar x;
2189: MatScalar *v;
2190: PetscErrorCode ierr;
2191: PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz;
2192: const PetscInt *jj;
2195: if (ll) {
2196: /* The local size is used so that VecMPI can be passed to this routine
2197: by MatDiagonalScale_MPIAIJ */
2198: VecGetLocalSize(ll,&m);
2199: if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length");
2200: VecGetArrayRead(ll,&l);
2201: v = a->a;
2202: for (i=0; i<m; i++) {
2203: x = l[i];
2204: M = a->i[i+1] - a->i[i];
2205: for (j=0; j<M; j++) (*v++) *= x;
2206: }
2207: VecRestoreArrayRead(ll,&l);
2208: PetscLogFlops(nz);
2209: }
2210: if (rr) {
2211: VecGetLocalSize(rr,&n);
2212: if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length");
2213: VecGetArrayRead(rr,&r);
2214: v = a->a; jj = a->j;
2215: for (i=0; i<nz; i++) (*v++) *= r[*jj++];
2216: VecRestoreArrayRead(rr,&r);
2217: PetscLogFlops(nz);
2218: }
2219: MatSeqAIJInvalidateDiagonal(A);
2220: return(0);
2221: }
2223: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B)
2224: {
2225: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c;
2227: PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens;
2228: PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi;
2229: const PetscInt *irow,*icol;
2230: PetscInt nrows,ncols;
2231: PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen;
2232: MatScalar *a_new,*mat_a;
2233: Mat C;
2234: PetscBool stride;
2238: ISGetIndices(isrow,&irow);
2239: ISGetLocalSize(isrow,&nrows);
2240: ISGetLocalSize(iscol,&ncols);
2242: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);
2243: if (stride) {
2244: ISStrideGetInfo(iscol,&first,&step);
2245: } else {
2246: first = 0;
2247: step = 0;
2248: }
2249: if (stride && step == 1) {
2250: /* special case of contiguous rows */
2251: PetscMalloc2(nrows,&lens,nrows,&starts);
2252: /* loop over new rows determining lens and starting points */
2253: for (i=0; i<nrows; i++) {
2254: kstart = ai[irow[i]];
2255: kend = kstart + ailen[irow[i]];
2256: starts[i] = kstart;
2257: for (k=kstart; k<kend; k++) {
2258: if (aj[k] >= first) {
2259: starts[i] = k;
2260: break;
2261: }
2262: }
2263: sum = 0;
2264: while (k < kend) {
2265: if (aj[k++] >= first+ncols) break;
2266: sum++;
2267: }
2268: lens[i] = sum;
2269: }
2270: /* create submatrix */
2271: if (scall == MAT_REUSE_MATRIX) {
2272: PetscInt n_cols,n_rows;
2273: MatGetSize(*B,&n_rows,&n_cols);
2274: if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
2275: MatZeroEntries(*B);
2276: C = *B;
2277: } else {
2278: PetscInt rbs,cbs;
2279: MatCreate(PetscObjectComm((PetscObject)A),&C);
2280: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2281: ISGetBlockSize(isrow,&rbs);
2282: ISGetBlockSize(iscol,&cbs);
2283: MatSetBlockSizes(C,rbs,cbs);
2284: MatSetType(C,((PetscObject)A)->type_name);
2285: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2286: }
2287: c = (Mat_SeqAIJ*)C->data;
2289: /* loop over rows inserting into submatrix */
2290: a_new = c->a;
2291: j_new = c->j;
2292: i_new = c->i;
2294: for (i=0; i<nrows; i++) {
2295: ii = starts[i];
2296: lensi = lens[i];
2297: for (k=0; k<lensi; k++) {
2298: *j_new++ = aj[ii+k] - first;
2299: }
2300: PetscMemcpy(a_new,a->a + starts[i],lensi*sizeof(PetscScalar));
2301: a_new += lensi;
2302: i_new[i+1] = i_new[i] + lensi;
2303: c->ilen[i] = lensi;
2304: }
2305: PetscFree2(lens,starts);
2306: } else {
2307: ISGetIndices(iscol,&icol);
2308: PetscCalloc1(oldcols,&smap);
2309: PetscMalloc1(1+nrows,&lens);
2310: for (i=0; i<ncols; i++) {
2311: #if defined(PETSC_USE_DEBUG)
2312: if (icol[i] >= oldcols) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Requesting column beyond largest column icol[%D] %D <= A->cmap->n %D",i,icol[i],oldcols);
2313: #endif
2314: smap[icol[i]] = i+1;
2315: }
2317: /* determine lens of each row */
2318: for (i=0; i<nrows; i++) {
2319: kstart = ai[irow[i]];
2320: kend = kstart + a->ilen[irow[i]];
2321: lens[i] = 0;
2322: for (k=kstart; k<kend; k++) {
2323: if (smap[aj[k]]) {
2324: lens[i]++;
2325: }
2326: }
2327: }
2328: /* Create and fill new matrix */
2329: if (scall == MAT_REUSE_MATRIX) {
2330: PetscBool equal;
2332: c = (Mat_SeqAIJ*)((*B)->data);
2333: if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
2334: PetscMemcmp(c->ilen,lens,(*B)->rmap->n*sizeof(PetscInt),&equal);
2335: if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
2336: PetscMemzero(c->ilen,(*B)->rmap->n*sizeof(PetscInt));
2337: C = *B;
2338: } else {
2339: PetscInt rbs,cbs;
2340: MatCreate(PetscObjectComm((PetscObject)A),&C);
2341: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2342: ISGetBlockSize(isrow,&rbs);
2343: ISGetBlockSize(iscol,&cbs);
2344: MatSetBlockSizes(C,rbs,cbs);
2345: MatSetType(C,((PetscObject)A)->type_name);
2346: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2347: }
2348: c = (Mat_SeqAIJ*)(C->data);
2349: for (i=0; i<nrows; i++) {
2350: row = irow[i];
2351: kstart = ai[row];
2352: kend = kstart + a->ilen[row];
2353: mat_i = c->i[i];
2354: mat_j = c->j + mat_i;
2355: mat_a = c->a + mat_i;
2356: mat_ilen = c->ilen + i;
2357: for (k=kstart; k<kend; k++) {
2358: if ((tcol=smap[a->j[k]])) {
2359: *mat_j++ = tcol - 1;
2360: *mat_a++ = a->a[k];
2361: (*mat_ilen)++;
2363: }
2364: }
2365: }
2366: /* Free work space */
2367: ISRestoreIndices(iscol,&icol);
2368: PetscFree(smap);
2369: PetscFree(lens);
2370: /* sort */
2371: for (i = 0; i < nrows; i++) {
2372: PetscInt ilen;
2374: mat_i = c->i[i];
2375: mat_j = c->j + mat_i;
2376: mat_a = c->a + mat_i;
2377: ilen = c->ilen[i];
2378: PetscSortIntWithScalarArray(ilen,mat_j,mat_a);
2379: }
2380: }
2381: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2382: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2384: ISRestoreIndices(isrow,&irow);
2385: *B = C;
2386: return(0);
2387: }
2389: PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat)
2390: {
2392: Mat B;
2395: if (scall == MAT_INITIAL_MATRIX) {
2396: MatCreate(subComm,&B);
2397: MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);
2398: MatSetBlockSizesFromMats(B,mat,mat);
2399: MatSetType(B,MATSEQAIJ);
2400: MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);
2401: *subMat = B;
2402: } else {
2403: MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);
2404: }
2405: return(0);
2406: }
2408: PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info)
2409: {
2410: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2412: Mat outA;
2413: PetscBool row_identity,col_identity;
2416: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu");
2418: ISIdentity(row,&row_identity);
2419: ISIdentity(col,&col_identity);
2421: outA = inA;
2422: outA->factortype = MAT_FACTOR_LU;
2423: PetscFree(inA->solvertype);
2424: PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);
2426: PetscObjectReference((PetscObject)row);
2427: ISDestroy(&a->row);
2429: a->row = row;
2431: PetscObjectReference((PetscObject)col);
2432: ISDestroy(&a->col);
2434: a->col = col;
2436: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2437: ISDestroy(&a->icol);
2438: ISInvertPermutation(col,PETSC_DECIDE,&a->icol);
2439: PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);
2441: if (!a->solve_work) { /* this matrix may have been factored before */
2442: PetscMalloc1(inA->rmap->n+1,&a->solve_work);
2443: PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));
2444: }
2446: MatMarkDiagonal_SeqAIJ(inA);
2447: if (row_identity && col_identity) {
2448: MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);
2449: } else {
2450: MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);
2451: }
2452: return(0);
2453: }
2455: PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha)
2456: {
2457: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2458: PetscScalar oalpha = alpha;
2460: PetscBLASInt one = 1,bnz;
2463: PetscBLASIntCast(a->nz,&bnz);
2464: PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&oalpha,a->a,&one));
2465: PetscLogFlops(a->nz);
2466: MatSeqAIJInvalidateDiagonal(inA);
2467: return(0);
2468: }
2470: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2471: {
2473: PetscInt i;
2476: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2477: PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);
2479: for (i=0; i<submatj->nrqr; ++i) {
2480: PetscFree(submatj->sbuf2[i]);
2481: }
2482: PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);
2484: if (submatj->rbuf1) {
2485: PetscFree(submatj->rbuf1[0]);
2486: PetscFree(submatj->rbuf1);
2487: }
2489: for (i=0; i<submatj->nrqs; ++i) {
2490: PetscFree(submatj->rbuf3[i]);
2491: }
2492: PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);
2493: PetscFree(submatj->pa);
2494: }
2496: #if defined(PETSC_USE_CTABLE)
2497: PetscTableDestroy((PetscTable*)&submatj->rmap);
2498: if (submatj->cmap_loc) {PetscFree(submatj->cmap_loc);}
2499: PetscFree(submatj->rmap_loc);
2500: #else
2501: PetscFree(submatj->rmap);
2502: #endif
2504: if (!submatj->allcolumns) {
2505: #if defined(PETSC_USE_CTABLE)
2506: PetscTableDestroy((PetscTable*)&submatj->cmap);
2507: #else
2508: PetscFree(submatj->cmap);
2509: #endif
2510: }
2511: PetscFree(submatj->row2proc);
2513: PetscFree(submatj);
2514: return(0);
2515: }
2517: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2518: {
2520: Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data;
2521: Mat_SubSppt *submatj = c->submatis1;
2524: submatj->destroy(C);
2525: MatDestroySubMatrix_Private(submatj);
2526: return(0);
2527: }
2529: PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n,Mat *mat[])
2530: {
2532: PetscInt i;
2533: Mat C;
2534: Mat_SeqAIJ *c;
2535: Mat_SubSppt *submatj;
2538: for (i=0; i<n; i++) {
2539: C = (*mat)[i];
2540: c = (Mat_SeqAIJ*)C->data;
2541: submatj = c->submatis1;
2542: if (submatj) {
2543: if (--((PetscObject)C)->refct <= 0) {
2544: (submatj->destroy)(C);
2545: MatDestroySubMatrix_Private(submatj);
2546: PetscLayoutDestroy(&C->rmap);
2547: PetscLayoutDestroy(&C->cmap);
2548: PetscHeaderDestroy(&C);
2549: }
2550: } else {
2551: MatDestroy(&C);
2552: }
2553: }
2555: PetscFree(*mat);
2556: return(0);
2557: }
2559: PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
2560: {
2562: PetscInt i;
2565: if (scall == MAT_INITIAL_MATRIX) {
2566: PetscCalloc1(n+1,B);
2567: }
2569: for (i=0; i<n; i++) {
2570: MatCreateSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);
2571: }
2572: return(0);
2573: }
2575: PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov)
2576: {
2577: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2579: PetscInt row,i,j,k,l,m,n,*nidx,isz,val;
2580: const PetscInt *idx;
2581: PetscInt start,end,*ai,*aj;
2582: PetscBT table;
2585: m = A->rmap->n;
2586: ai = a->i;
2587: aj = a->j;
2589: if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used");
2591: PetscMalloc1(m+1,&nidx);
2592: PetscBTCreate(m,&table);
2594: for (i=0; i<is_max; i++) {
2595: /* Initialize the two local arrays */
2596: isz = 0;
2597: PetscBTMemzero(m,table);
2599: /* Extract the indices, assume there can be duplicate entries */
2600: ISGetIndices(is[i],&idx);
2601: ISGetLocalSize(is[i],&n);
2603: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2604: for (j=0; j<n; ++j) {
2605: if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j];
2606: }
2607: ISRestoreIndices(is[i],&idx);
2608: ISDestroy(&is[i]);
2610: k = 0;
2611: for (j=0; j<ov; j++) { /* for each overlap */
2612: n = isz;
2613: for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */
2614: row = nidx[k];
2615: start = ai[row];
2616: end = ai[row+1];
2617: for (l = start; l<end; l++) {
2618: val = aj[l];
2619: if (!PetscBTLookupSet(table,val)) nidx[isz++] = val;
2620: }
2621: }
2622: }
2623: ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));
2624: }
2625: PetscBTDestroy(&table);
2626: PetscFree(nidx);
2627: return(0);
2628: }
2630: /* -------------------------------------------------------------- */
2631: PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B)
2632: {
2633: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2635: PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n;
2636: const PetscInt *row,*col;
2637: PetscInt *cnew,j,*lens;
2638: IS icolp,irowp;
2639: PetscInt *cwork = NULL;
2640: PetscScalar *vwork = NULL;
2643: ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);
2644: ISGetIndices(irowp,&row);
2645: ISInvertPermutation(colp,PETSC_DECIDE,&icolp);
2646: ISGetIndices(icolp,&col);
2648: /* determine lengths of permuted rows */
2649: PetscMalloc1(m+1,&lens);
2650: for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i];
2651: MatCreate(PetscObjectComm((PetscObject)A),B);
2652: MatSetSizes(*B,m,n,m,n);
2653: MatSetBlockSizesFromMats(*B,A,A);
2654: MatSetType(*B,((PetscObject)A)->type_name);
2655: MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);
2656: PetscFree(lens);
2658: PetscMalloc1(n,&cnew);
2659: for (i=0; i<m; i++) {
2660: MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2661: for (j=0; j<nz; j++) cnew[j] = col[cwork[j]];
2662: MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);
2663: MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2664: }
2665: PetscFree(cnew);
2667: (*B)->assembled = PETSC_FALSE;
2669: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
2670: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
2671: ISRestoreIndices(irowp,&row);
2672: ISRestoreIndices(icolp,&col);
2673: ISDestroy(&irowp);
2674: ISDestroy(&icolp);
2675: return(0);
2676: }
2678: PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str)
2679: {
2683: /* If the two matrices have the same copy implementation, use fast copy. */
2684: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2685: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2686: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
2688: if (a->i[A->rmap->n] != b->i[B->rmap->n]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number of nonzeros in two matrices are different");
2689: PetscMemcpy(b->a,a->a,(a->i[A->rmap->n])*sizeof(PetscScalar));
2690: PetscObjectStateIncrease((PetscObject)B);
2691: } else {
2692: MatCopy_Basic(A,B,str);
2693: }
2694: return(0);
2695: }
2697: PetscErrorCode MatSetUp_SeqAIJ(Mat A)
2698: {
2702: MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,0);
2703: return(0);
2704: }
2706: PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[])
2707: {
2708: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2711: *array = a->a;
2712: return(0);
2713: }
2715: PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[])
2716: {
2718: return(0);
2719: }
2721: /*
2722: Computes the number of nonzeros per row needed for preallocation when X and Y
2723: have different nonzero structure.
2724: */
2725: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz)
2726: {
2727: PetscInt i,j,k,nzx,nzy;
2730: /* Set the number of nonzeros in the new matrix */
2731: for (i=0; i<m; i++) {
2732: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2733: nzx = xi[i+1] - xi[i];
2734: nzy = yi[i+1] - yi[i];
2735: nnz[i] = 0;
2736: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2737: for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */
2738: if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */
2739: nnz[i]++;
2740: }
2741: for (; k<nzy; k++) nnz[i]++;
2742: }
2743: return(0);
2744: }
2746: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz)
2747: {
2748: PetscInt m = Y->rmap->N;
2749: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2750: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2754: /* Set the number of nonzeros in the new matrix */
2755: MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);
2756: return(0);
2757: }
2759: PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2760: {
2762: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data;
2763: PetscBLASInt one=1,bnz;
2766: PetscBLASIntCast(x->nz,&bnz);
2767: if (str == SAME_NONZERO_PATTERN) {
2768: PetscScalar alpha = a;
2769: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2770: MatSeqAIJInvalidateDiagonal(Y);
2771: PetscObjectStateIncrease((PetscObject)Y);
2772: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2773: MatAXPY_Basic(Y,a,X,str);
2774: } else {
2775: Mat B;
2776: PetscInt *nnz;
2777: PetscMalloc1(Y->rmap->N,&nnz);
2778: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2779: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2780: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2781: MatSetBlockSizesFromMats(B,Y,Y);
2782: MatSetType(B,(MatType) ((PetscObject)Y)->type_name);
2783: MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);
2784: MatSeqAIJSetPreallocation(B,0,nnz);
2785: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2786: MatHeaderReplace(Y,&B);
2787: PetscFree(nnz);
2788: }
2789: return(0);
2790: }
2792: PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
2793: {
2794: #if defined(PETSC_USE_COMPLEX)
2795: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
2796: PetscInt i,nz;
2797: PetscScalar *a;
2800: nz = aij->nz;
2801: a = aij->a;
2802: for (i=0; i<nz; i++) a[i] = PetscConj(a[i]);
2803: #else
2805: #endif
2806: return(0);
2807: }
2809: PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2810: {
2811: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2813: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2814: PetscReal atmp;
2815: PetscScalar *x;
2816: MatScalar *aa;
2819: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2820: aa = a->a;
2821: ai = a->i;
2822: aj = a->j;
2824: VecSet(v,0.0);
2825: VecGetArray(v,&x);
2826: VecGetLocalSize(v,&n);
2827: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2828: for (i=0; i<m; i++) {
2829: ncols = ai[1] - ai[0]; ai++;
2830: x[i] = 0.0;
2831: for (j=0; j<ncols; j++) {
2832: atmp = PetscAbsScalar(*aa);
2833: if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2834: aa++; aj++;
2835: }
2836: }
2837: VecRestoreArray(v,&x);
2838: return(0);
2839: }
2841: PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2842: {
2843: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2845: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2846: PetscScalar *x;
2847: MatScalar *aa;
2850: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2851: aa = a->a;
2852: ai = a->i;
2853: aj = a->j;
2855: VecSet(v,0.0);
2856: VecGetArray(v,&x);
2857: VecGetLocalSize(v,&n);
2858: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2859: for (i=0; i<m; i++) {
2860: ncols = ai[1] - ai[0]; ai++;
2861: if (ncols == A->cmap->n) { /* row is dense */
2862: x[i] = *aa; if (idx) idx[i] = 0;
2863: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
2864: x[i] = 0.0;
2865: if (idx) {
2866: idx[i] = 0; /* in case ncols is zero */
2867: for (j=0;j<ncols;j++) { /* find first implicit 0.0 in the row */
2868: if (aj[j] > j) {
2869: idx[i] = j;
2870: break;
2871: }
2872: }
2873: }
2874: }
2875: for (j=0; j<ncols; j++) {
2876: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
2877: aa++; aj++;
2878: }
2879: }
2880: VecRestoreArray(v,&x);
2881: return(0);
2882: }
2884: PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2885: {
2886: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2888: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2889: PetscReal atmp;
2890: PetscScalar *x;
2891: MatScalar *aa;
2894: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2895: aa = a->a;
2896: ai = a->i;
2897: aj = a->j;
2899: VecSet(v,0.0);
2900: VecGetArray(v,&x);
2901: VecGetLocalSize(v,&n);
2902: if (n != A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector, %D vs. %D rows", A->rmap->n, n);
2903: for (i=0; i<m; i++) {
2904: ncols = ai[1] - ai[0]; ai++;
2905: if (ncols) {
2906: /* Get first nonzero */
2907: for (j = 0; j < ncols; j++) {
2908: atmp = PetscAbsScalar(aa[j]);
2909: if (atmp > 1.0e-12) {
2910: x[i] = atmp;
2911: if (idx) idx[i] = aj[j];
2912: break;
2913: }
2914: }
2915: if (j == ncols) {x[i] = PetscAbsScalar(*aa); if (idx) idx[i] = *aj;}
2916: } else {
2917: x[i] = 0.0; if (idx) idx[i] = 0;
2918: }
2919: for (j = 0; j < ncols; j++) {
2920: atmp = PetscAbsScalar(*aa);
2921: if (atmp > 1.0e-12 && PetscAbsScalar(x[i]) > atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2922: aa++; aj++;
2923: }
2924: }
2925: VecRestoreArray(v,&x);
2926: return(0);
2927: }
2929: PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2930: {
2931: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2932: PetscErrorCode ierr;
2933: PetscInt i,j,m = A->rmap->n,ncols,n;
2934: const PetscInt *ai,*aj;
2935: PetscScalar *x;
2936: const MatScalar *aa;
2939: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2940: aa = a->a;
2941: ai = a->i;
2942: aj = a->j;
2944: VecSet(v,0.0);
2945: VecGetArray(v,&x);
2946: VecGetLocalSize(v,&n);
2947: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2948: for (i=0; i<m; i++) {
2949: ncols = ai[1] - ai[0]; ai++;
2950: if (ncols == A->cmap->n) { /* row is dense */
2951: x[i] = *aa; if (idx) idx[i] = 0;
2952: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
2953: x[i] = 0.0;
2954: if (idx) { /* find first implicit 0.0 in the row */
2955: idx[i] = 0; /* in case ncols is zero */
2956: for (j=0; j<ncols; j++) {
2957: if (aj[j] > j) {
2958: idx[i] = j;
2959: break;
2960: }
2961: }
2962: }
2963: }
2964: for (j=0; j<ncols; j++) {
2965: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
2966: aa++; aj++;
2967: }
2968: }
2969: VecRestoreArray(v,&x);
2970: return(0);
2971: }
2973: #include <petscblaslapack.h>
2974: #include <petsc/private/kernels/blockinvert.h>
2976: PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values)
2977: {
2978: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
2980: PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j;
2981: MatScalar *diag,work[25],*v_work;
2982: PetscReal shift = 0.0;
2983: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
2986: allowzeropivot = PetscNot(A->erroriffailure);
2987: if (a->ibdiagvalid) {
2988: if (values) *values = a->ibdiag;
2989: return(0);
2990: }
2991: MatMarkDiagonal_SeqAIJ(A);
2992: if (!a->ibdiag) {
2993: PetscMalloc1(bs2*mbs,&a->ibdiag);
2994: PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));
2995: }
2996: diag = a->ibdiag;
2997: if (values) *values = a->ibdiag;
2998: /* factor and invert each block */
2999: switch (bs) {
3000: case 1:
3001: for (i=0; i<mbs; i++) {
3002: MatGetValues(A,1,&i,1,&i,diag+i);
3003: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3004: if (allowzeropivot) {
3005: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3006: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3007: A->factorerror_zeropivot_row = i;
3008: PetscInfo3(A,"Zero pivot, row %D pivot %g tolerance %g\n",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);
3009: } else SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D pivot %g tolerance %g",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);
3010: }
3011: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3012: }
3013: break;
3014: case 2:
3015: for (i=0; i<mbs; i++) {
3016: ij[0] = 2*i; ij[1] = 2*i + 1;
3017: MatGetValues(A,2,ij,2,ij,diag);
3018: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
3019: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3020: PetscKernel_A_gets_transpose_A_2(diag);
3021: diag += 4;
3022: }
3023: break;
3024: case 3:
3025: for (i=0; i<mbs; i++) {
3026: ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2;
3027: MatGetValues(A,3,ij,3,ij,diag);
3028: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
3029: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3030: PetscKernel_A_gets_transpose_A_3(diag);
3031: diag += 9;
3032: }
3033: break;
3034: case 4:
3035: for (i=0; i<mbs; i++) {
3036: ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3;
3037: MatGetValues(A,4,ij,4,ij,diag);
3038: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
3039: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3040: PetscKernel_A_gets_transpose_A_4(diag);
3041: diag += 16;
3042: }
3043: break;
3044: case 5:
3045: for (i=0; i<mbs; i++) {
3046: ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4;
3047: MatGetValues(A,5,ij,5,ij,diag);
3048: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
3049: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3050: PetscKernel_A_gets_transpose_A_5(diag);
3051: diag += 25;
3052: }
3053: break;
3054: case 6:
3055: for (i=0; i<mbs; i++) {
3056: ij[0] = 6*i; ij[1] = 6*i + 1; ij[2] = 6*i + 2; ij[3] = 6*i + 3; ij[4] = 6*i + 4; ij[5] = 6*i + 5;
3057: MatGetValues(A,6,ij,6,ij,diag);
3058: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
3059: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3060: PetscKernel_A_gets_transpose_A_6(diag);
3061: diag += 36;
3062: }
3063: break;
3064: case 7:
3065: for (i=0; i<mbs; i++) {
3066: ij[0] = 7*i; ij[1] = 7*i + 1; ij[2] = 7*i + 2; ij[3] = 7*i + 3; ij[4] = 7*i + 4; ij[5] = 7*i + 5; ij[5] = 7*i + 6;
3067: MatGetValues(A,7,ij,7,ij,diag);
3068: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
3069: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3070: PetscKernel_A_gets_transpose_A_7(diag);
3071: diag += 49;
3072: }
3073: break;
3074: default:
3075: PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);
3076: for (i=0; i<mbs; i++) {
3077: for (j=0; j<bs; j++) {
3078: IJ[j] = bs*i + j;
3079: }
3080: MatGetValues(A,bs,IJ,bs,IJ,diag);
3081: PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
3082: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3083: PetscKernel_A_gets_transpose_A_N(diag,bs);
3084: diag += bs2;
3085: }
3086: PetscFree3(v_work,v_pivots,IJ);
3087: }
3088: a->ibdiagvalid = PETSC_TRUE;
3089: return(0);
3090: }
3092: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx)
3093: {
3095: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3096: PetscScalar a;
3097: PetscInt m,n,i,j,col;
3100: if (!x->assembled) {
3101: MatGetSize(x,&m,&n);
3102: for (i=0; i<m; i++) {
3103: for (j=0; j<aij->imax[i]; j++) {
3104: PetscRandomGetValue(rctx,&a);
3105: col = (PetscInt)(n*PetscRealPart(a));
3106: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3107: }
3108: }
3109: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet coded");
3110: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3111: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3112: return(0);
3113: }
3115: PetscErrorCode MatShift_SeqAIJ(Mat Y,PetscScalar a)
3116: {
3118: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)Y->data;
3121: if (!Y->preallocated || !aij->nz) {
3122: MatSeqAIJSetPreallocation(Y,1,NULL);
3123: }
3124: MatShift_Basic(Y,a);
3125: return(0);
3126: }
3128: /* -------------------------------------------------------------------*/
3129: static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ,
3130: MatGetRow_SeqAIJ,
3131: MatRestoreRow_SeqAIJ,
3132: MatMult_SeqAIJ,
3133: /* 4*/ MatMultAdd_SeqAIJ,
3134: MatMultTranspose_SeqAIJ,
3135: MatMultTransposeAdd_SeqAIJ,
3136: 0,
3137: 0,
3138: 0,
3139: /* 10*/ 0,
3140: MatLUFactor_SeqAIJ,
3141: 0,
3142: MatSOR_SeqAIJ,
3143: MatTranspose_SeqAIJ,
3144: /*1 5*/ MatGetInfo_SeqAIJ,
3145: MatEqual_SeqAIJ,
3146: MatGetDiagonal_SeqAIJ,
3147: MatDiagonalScale_SeqAIJ,
3148: MatNorm_SeqAIJ,
3149: /* 20*/ 0,
3150: MatAssemblyEnd_SeqAIJ,
3151: MatSetOption_SeqAIJ,
3152: MatZeroEntries_SeqAIJ,
3153: /* 24*/ MatZeroRows_SeqAIJ,
3154: 0,
3155: 0,
3156: 0,
3157: 0,
3158: /* 29*/ MatSetUp_SeqAIJ,
3159: 0,
3160: 0,
3161: 0,
3162: 0,
3163: /* 34*/ MatDuplicate_SeqAIJ,
3164: 0,
3165: 0,
3166: MatILUFactor_SeqAIJ,
3167: 0,
3168: /* 39*/ MatAXPY_SeqAIJ,
3169: MatCreateSubMatrices_SeqAIJ,
3170: MatIncreaseOverlap_SeqAIJ,
3171: MatGetValues_SeqAIJ,
3172: MatCopy_SeqAIJ,
3173: /* 44*/ MatGetRowMax_SeqAIJ,
3174: MatScale_SeqAIJ,
3175: MatShift_SeqAIJ,
3176: MatDiagonalSet_SeqAIJ,
3177: MatZeroRowsColumns_SeqAIJ,
3178: /* 49*/ MatSetRandom_SeqAIJ,
3179: MatGetRowIJ_SeqAIJ,
3180: MatRestoreRowIJ_SeqAIJ,
3181: MatGetColumnIJ_SeqAIJ,
3182: MatRestoreColumnIJ_SeqAIJ,
3183: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3184: 0,
3185: 0,
3186: MatPermute_SeqAIJ,
3187: 0,
3188: /* 59*/ 0,
3189: MatDestroy_SeqAIJ,
3190: MatView_SeqAIJ,
3191: 0,
3192: MatMatMatMult_SeqAIJ_SeqAIJ_SeqAIJ,
3193: /* 64*/ MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ,
3194: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3195: 0,
3196: 0,
3197: 0,
3198: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3199: MatGetRowMinAbs_SeqAIJ,
3200: 0,
3201: 0,
3202: 0,
3203: /* 74*/ 0,
3204: MatFDColoringApply_AIJ,
3205: 0,
3206: 0,
3207: 0,
3208: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3209: 0,
3210: 0,
3211: 0,
3212: MatLoad_SeqAIJ,
3213: /* 84*/ MatIsSymmetric_SeqAIJ,
3214: MatIsHermitian_SeqAIJ,
3215: 0,
3216: 0,
3217: 0,
3218: /* 89*/ MatMatMult_SeqAIJ_SeqAIJ,
3219: MatMatMultSymbolic_SeqAIJ_SeqAIJ,
3220: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3221: MatPtAP_SeqAIJ_SeqAIJ,
3222: MatPtAPSymbolic_SeqAIJ_SeqAIJ_DenseAxpy,
3223: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ,
3224: MatMatTransposeMult_SeqAIJ_SeqAIJ,
3225: MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ,
3226: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3227: 0,
3228: /* 99*/ 0,
3229: 0,
3230: 0,
3231: MatConjugate_SeqAIJ,
3232: 0,
3233: /*104*/ MatSetValuesRow_SeqAIJ,
3234: MatRealPart_SeqAIJ,
3235: MatImaginaryPart_SeqAIJ,
3236: 0,
3237: 0,
3238: /*109*/ MatMatSolve_SeqAIJ,
3239: 0,
3240: MatGetRowMin_SeqAIJ,
3241: 0,
3242: MatMissingDiagonal_SeqAIJ,
3243: /*114*/ 0,
3244: 0,
3245: 0,
3246: 0,
3247: 0,
3248: /*119*/ 0,
3249: 0,
3250: 0,
3251: 0,
3252: MatGetMultiProcBlock_SeqAIJ,
3253: /*124*/ MatFindNonzeroRows_SeqAIJ,
3254: MatGetColumnNorms_SeqAIJ,
3255: MatInvertBlockDiagonal_SeqAIJ,
3256: 0,
3257: 0,
3258: /*129*/ 0,
3259: MatTransposeMatMult_SeqAIJ_SeqAIJ,
3260: MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ,
3261: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3262: MatTransposeColoringCreate_SeqAIJ,
3263: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3264: MatTransColoringApplyDenToSp_SeqAIJ,
3265: MatRARt_SeqAIJ_SeqAIJ,
3266: MatRARtSymbolic_SeqAIJ_SeqAIJ,
3267: MatRARtNumeric_SeqAIJ_SeqAIJ,
3268: /*139*/0,
3269: 0,
3270: 0,
3271: MatFDColoringSetUp_SeqXAIJ,
3272: MatFindOffBlockDiagonalEntries_SeqAIJ,
3273: /*144*/MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3274: MatDestroySubMatrices_SeqAIJ
3275: };
3277: PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices)
3278: {
3279: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3280: PetscInt i,nz,n;
3283: nz = aij->maxnz;
3284: n = mat->rmap->n;
3285: for (i=0; i<nz; i++) {
3286: aij->j[i] = indices[i];
3287: }
3288: aij->nz = nz;
3289: for (i=0; i<n; i++) {
3290: aij->ilen[i] = aij->imax[i];
3291: }
3292: return(0);
3293: }
3295: /*@
3296: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3297: in the matrix.
3299: Input Parameters:
3300: + mat - the SeqAIJ matrix
3301: - indices - the column indices
3303: Level: advanced
3305: Notes:
3306: This can be called if you have precomputed the nonzero structure of the
3307: matrix and want to provide it to the matrix object to improve the performance
3308: of the MatSetValues() operation.
3310: You MUST have set the correct numbers of nonzeros per row in the call to
3311: MatCreateSeqAIJ(), and the columns indices MUST be sorted.
3313: MUST be called before any calls to MatSetValues();
3315: The indices should start with zero, not one.
3317: @*/
3318: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices)
3319: {
3325: PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));
3326: return(0);
3327: }
3329: /* ----------------------------------------------------------------------------------------*/
3331: PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3332: {
3333: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3335: size_t nz = aij->i[mat->rmap->n];
3338: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3340: /* allocate space for values if not already there */
3341: if (!aij->saved_values) {
3342: PetscMalloc1(nz+1,&aij->saved_values);
3343: PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));
3344: }
3346: /* copy values over */
3347: PetscMemcpy(aij->saved_values,aij->a,nz*sizeof(PetscScalar));
3348: return(0);
3349: }
3351: /*@
3352: MatStoreValues - Stashes a copy of the matrix values; this allows, for
3353: example, reuse of the linear part of a Jacobian, while recomputing the
3354: nonlinear portion.
3356: Collect on Mat
3358: Input Parameters:
3359: . mat - the matrix (currently only AIJ matrices support this option)
3361: Level: advanced
3363: Common Usage, with SNESSolve():
3364: $ Create Jacobian matrix
3365: $ Set linear terms into matrix
3366: $ Apply boundary conditions to matrix, at this time matrix must have
3367: $ final nonzero structure (i.e. setting the nonlinear terms and applying
3368: $ boundary conditions again will not change the nonzero structure
3369: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3370: $ MatStoreValues(mat);
3371: $ Call SNESSetJacobian() with matrix
3372: $ In your Jacobian routine
3373: $ MatRetrieveValues(mat);
3374: $ Set nonlinear terms in matrix
3376: Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself:
3377: $ // build linear portion of Jacobian
3378: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3379: $ MatStoreValues(mat);
3380: $ loop over nonlinear iterations
3381: $ MatRetrieveValues(mat);
3382: $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3383: $ // call MatAssemblyBegin/End() on matrix
3384: $ Solve linear system with Jacobian
3385: $ endloop
3387: Notes:
3388: Matrix must already be assemblied before calling this routine
3389: Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before
3390: calling this routine.
3392: When this is called multiple times it overwrites the previous set of stored values
3393: and does not allocated additional space.
3395: .seealso: MatRetrieveValues()
3397: @*/
3398: PetscErrorCode MatStoreValues(Mat mat)
3399: {
3404: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3405: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3406: PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));
3407: return(0);
3408: }
3410: PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3411: {
3412: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3414: PetscInt nz = aij->i[mat->rmap->n];
3417: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3418: if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first");
3419: /* copy values over */
3420: PetscMemcpy(aij->a,aij->saved_values,nz*sizeof(PetscScalar));
3421: return(0);
3422: }
3424: /*@
3425: MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for
3426: example, reuse of the linear part of a Jacobian, while recomputing the
3427: nonlinear portion.
3429: Collect on Mat
3431: Input Parameters:
3432: . mat - the matrix (currently only AIJ matrices support this option)
3434: Level: advanced
3436: .seealso: MatStoreValues()
3438: @*/
3439: PetscErrorCode MatRetrieveValues(Mat mat)
3440: {
3445: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3446: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3447: PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));
3448: return(0);
3449: }
3452: /* --------------------------------------------------------------------------------*/
3453: /*@C
3454: MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format
3455: (the default parallel PETSc format). For good matrix assembly performance
3456: the user should preallocate the matrix storage by setting the parameter nz
3457: (or the array nnz). By setting these parameters accurately, performance
3458: during matrix assembly can be increased by more than a factor of 50.
3460: Collective on MPI_Comm
3462: Input Parameters:
3463: + comm - MPI communicator, set to PETSC_COMM_SELF
3464: . m - number of rows
3465: . n - number of columns
3466: . nz - number of nonzeros per row (same for all rows)
3467: - nnz - array containing the number of nonzeros in the various rows
3468: (possibly different for each row) or NULL
3470: Output Parameter:
3471: . A - the matrix
3473: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3474: MatXXXXSetPreallocation() paradgm instead of this routine directly.
3475: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3477: Notes:
3478: If nnz is given then nz is ignored
3480: The AIJ format (also called the Yale sparse matrix format or
3481: compressed row storage), is fully compatible with standard Fortran 77
3482: storage. That is, the stored row and column indices can begin at
3483: either one (as in Fortran) or zero. See the users' manual for details.
3485: Specify the preallocated storage with either nz or nnz (not both).
3486: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3487: allocation. For large problems you MUST preallocate memory or you
3488: will get TERRIBLE performance, see the users' manual chapter on matrices.
3490: By default, this format uses inodes (identical nodes) when possible, to
3491: improve numerical efficiency of matrix-vector products and solves. We
3492: search for consecutive rows with the same nonzero structure, thereby
3493: reusing matrix information to achieve increased efficiency.
3495: Options Database Keys:
3496: + -mat_no_inode - Do not use inodes
3497: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3499: Level: intermediate
3501: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays()
3503: @*/
3504: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A)
3505: {
3509: MatCreate(comm,A);
3510: MatSetSizes(*A,m,n,m,n);
3511: MatSetType(*A,MATSEQAIJ);
3512: MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);
3513: return(0);
3514: }
3516: /*@C
3517: MatSeqAIJSetPreallocation - For good matrix assembly performance
3518: the user should preallocate the matrix storage by setting the parameter nz
3519: (or the array nnz). By setting these parameters accurately, performance
3520: during matrix assembly can be increased by more than a factor of 50.
3522: Collective on MPI_Comm
3524: Input Parameters:
3525: + B - The matrix
3526: . nz - number of nonzeros per row (same for all rows)
3527: - nnz - array containing the number of nonzeros in the various rows
3528: (possibly different for each row) or NULL
3530: Notes:
3531: If nnz is given then nz is ignored
3533: The AIJ format (also called the Yale sparse matrix format or
3534: compressed row storage), is fully compatible with standard Fortran 77
3535: storage. That is, the stored row and column indices can begin at
3536: either one (as in Fortran) or zero. See the users' manual for details.
3538: Specify the preallocated storage with either nz or nnz (not both).
3539: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3540: allocation. For large problems you MUST preallocate memory or you
3541: will get TERRIBLE performance, see the users' manual chapter on matrices.
3543: You can call MatGetInfo() to get information on how effective the preallocation was;
3544: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3545: You can also run with the option -info and look for messages with the string
3546: malloc in them to see if additional memory allocation was needed.
3548: Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix
3549: entries or columns indices
3551: By default, this format uses inodes (identical nodes) when possible, to
3552: improve numerical efficiency of matrix-vector products and solves. We
3553: search for consecutive rows with the same nonzero structure, thereby
3554: reusing matrix information to achieve increased efficiency.
3556: Options Database Keys:
3557: + -mat_no_inode - Do not use inodes
3558: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3560: Level: intermediate
3562: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo()
3564: @*/
3565: PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[])
3566: {
3572: PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));
3573: return(0);
3574: }
3576: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz)
3577: {
3578: Mat_SeqAIJ *b;
3579: PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE;
3581: PetscInt i;
3584: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3585: if (nz == MAT_SKIP_ALLOCATION) {
3586: skipallocation = PETSC_TRUE;
3587: nz = 0;
3588: }
3589: PetscLayoutSetUp(B->rmap);
3590: PetscLayoutSetUp(B->cmap);
3592: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3593: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz);
3594: if (nnz) {
3595: for (i=0; i<B->rmap->n; i++) {
3596: if (nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be less than 0: local row %D value %D",i,nnz[i]);
3597: if (nnz[i] > B->cmap->n) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be greater than row length: local row %D value %d rowlength %D",i,nnz[i],B->cmap->n);
3598: }
3599: }
3601: B->preallocated = PETSC_TRUE;
3603: b = (Mat_SeqAIJ*)B->data;
3605: if (!skipallocation) {
3606: if (!b->imax) {
3607: PetscMalloc2(B->rmap->n,&b->imax,B->rmap->n,&b->ilen);
3608: PetscLogObjectMemory((PetscObject)B,2*B->rmap->n*sizeof(PetscInt));
3609: }
3610: if (!nnz) {
3611: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3612: else if (nz < 0) nz = 1;
3613: for (i=0; i<B->rmap->n; i++) b->imax[i] = nz;
3614: nz = nz*B->rmap->n;
3615: } else {
3616: nz = 0;
3617: for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz += nnz[i];}
3618: }
3619: /* b->ilen will count nonzeros in each row so far. */
3620: for (i=0; i<B->rmap->n; i++) b->ilen[i] = 0;
3622: /* allocate the matrix space */
3623: /* FIXME: should B's old memory be unlogged? */
3624: MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);
3625: if (B->structure_only) {
3626: PetscMalloc1(nz,&b->j);
3627: PetscMalloc1(B->rmap->n+1,&b->i);
3628: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));
3629: } else {
3630: PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);
3631: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));
3632: }
3633: b->i[0] = 0;
3634: for (i=1; i<B->rmap->n+1; i++) {
3635: b->i[i] = b->i[i-1] + b->imax[i-1];
3636: }
3637: if (B->structure_only) {
3638: b->singlemalloc = PETSC_FALSE;
3639: b->free_a = PETSC_FALSE;
3640: } else {
3641: b->singlemalloc = PETSC_TRUE;
3642: b->free_a = PETSC_TRUE;
3643: }
3644: b->free_ij = PETSC_TRUE;
3645: } else {
3646: b->free_a = PETSC_FALSE;
3647: b->free_ij = PETSC_FALSE;
3648: }
3650: b->nz = 0;
3651: b->maxnz = nz;
3652: B->info.nz_unneeded = (double)b->maxnz;
3653: if (realalloc) {
3654: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
3655: }
3656: B->was_assembled = PETSC_FALSE;
3657: B->assembled = PETSC_FALSE;
3658: return(0);
3659: }
3661: /*@
3662: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format.
3664: Input Parameters:
3665: + B - the matrix
3666: . i - the indices into j for the start of each row (starts with zero)
3667: . j - the column indices for each row (starts with zero) these must be sorted for each row
3668: - v - optional values in the matrix
3670: Level: developer
3672: The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays()
3674: .keywords: matrix, aij, compressed row, sparse, sequential
3676: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), SeqAIJ
3677: @*/
3678: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[])
3679: {
3685: PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3686: return(0);
3687: }
3689: PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3690: {
3691: PetscInt i;
3692: PetscInt m,n;
3693: PetscInt nz;
3694: PetscInt *nnz, nz_max = 0;
3695: PetscScalar *values;
3699: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]);
3701: PetscLayoutSetUp(B->rmap);
3702: PetscLayoutSetUp(B->cmap);
3704: MatGetSize(B, &m, &n);
3705: PetscMalloc1(m+1, &nnz);
3706: for (i = 0; i < m; i++) {
3707: nz = Ii[i+1]- Ii[i];
3708: nz_max = PetscMax(nz_max, nz);
3709: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz);
3710: nnz[i] = nz;
3711: }
3712: MatSeqAIJSetPreallocation(B, 0, nnz);
3713: PetscFree(nnz);
3715: if (v) {
3716: values = (PetscScalar*) v;
3717: } else {
3718: PetscCalloc1(nz_max, &values);
3719: }
3721: for (i = 0; i < m; i++) {
3722: nz = Ii[i+1] - Ii[i];
3723: MatSetValues_SeqAIJ(B, 1, &i, nz, J+Ii[i], values + (v ? Ii[i] : 0), INSERT_VALUES);
3724: }
3726: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3727: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3729: if (!v) {
3730: PetscFree(values);
3731: }
3732: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3733: return(0);
3734: }
3736: #include <../src/mat/impls/dense/seq/dense.h>
3737: #include <petsc/private/kernels/petscaxpy.h>
3739: /*
3740: Computes (B'*A')' since computing B*A directly is untenable
3742: n p p
3743: ( ) ( ) ( )
3744: m ( A ) * n ( B ) = m ( C )
3745: ( ) ( ) ( )
3747: */
3748: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C)
3749: {
3750: PetscErrorCode ierr;
3751: Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data;
3752: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data;
3753: Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data;
3754: PetscInt i,n,m,q,p;
3755: const PetscInt *ii,*idx;
3756: const PetscScalar *b,*a,*a_q;
3757: PetscScalar *c,*c_q;
3760: m = A->rmap->n;
3761: n = A->cmap->n;
3762: p = B->cmap->n;
3763: a = sub_a->v;
3764: b = sub_b->a;
3765: c = sub_c->v;
3766: PetscMemzero(c,m*p*sizeof(PetscScalar));
3768: ii = sub_b->i;
3769: idx = sub_b->j;
3770: for (i=0; i<n; i++) {
3771: q = ii[i+1] - ii[i];
3772: while (q-->0) {
3773: c_q = c + m*(*idx);
3774: a_q = a + m*i;
3775: PetscKernelAXPY(c_q,*b,a_q,m);
3776: idx++;
3777: b++;
3778: }
3779: }
3780: return(0);
3781: }
3783: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
3784: {
3786: PetscInt m=A->rmap->n,n=B->cmap->n;
3787: Mat Cmat;
3790: if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %D != B->rmap->n %D\n",A->cmap->n,B->rmap->n);
3791: MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
3792: MatSetSizes(Cmat,m,n,m,n);
3793: MatSetBlockSizesFromMats(Cmat,A,B);
3794: MatSetType(Cmat,MATSEQDENSE);
3795: MatSeqDenseSetPreallocation(Cmat,NULL);
3797: Cmat->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
3799: *C = Cmat;
3800: return(0);
3801: }
3803: /* ----------------------------------------------------------------*/
3804: PETSC_INTERN PetscErrorCode MatMatMult_SeqDense_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3805: {
3809: if (scall == MAT_INITIAL_MATRIX) {
3810: PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
3811: MatMatMultSymbolic_SeqDense_SeqAIJ(A,B,fill,C);
3812: PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
3813: }
3814: PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
3815: MatMatMultNumeric_SeqDense_SeqAIJ(A,B,*C);
3816: PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
3817: return(0);
3818: }
3821: /*MC
3822: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
3823: based on compressed sparse row format.
3825: Options Database Keys:
3826: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
3828: Level: beginner
3830: .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType
3831: M*/
3833: /*MC
3834: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
3836: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
3837: and MATMPIAIJ otherwise. As a result, for single process communicators,
3838: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
3839: for communicators controlling multiple processes. It is recommended that you call both of
3840: the above preallocation routines for simplicity.
3842: Options Database Keys:
3843: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
3845: Developer Notes: Subclasses include MATAIJCUSP, MATAIJPERM, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
3846: enough exist.
3848: Level: beginner
3850: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
3851: M*/
3853: /*MC
3854: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
3856: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
3857: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
3858: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
3859: for communicators controlling multiple processes. It is recommended that you call both of
3860: the above preallocation routines for simplicity.
3862: Options Database Keys:
3863: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
3865: Level: beginner
3867: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
3868: M*/
3870: /*@C
3871: MatSeqAIJGetArray - gives access to the array where the data for a MATSEQAIJ matrix is stored
3873: Not Collective
3875: Input Parameter:
3876: . mat - a MATSEQAIJ matrix
3878: Output Parameter:
3879: . array - pointer to the data
3881: Level: intermediate
3883: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
3884: @*/
3885: PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array)
3886: {
3890: PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));
3891: return(0);
3892: }
3894: /*@C
3895: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
3897: Not Collective
3899: Input Parameter:
3900: . mat - a MATSEQAIJ matrix
3902: Output Parameter:
3903: . nz - the maximum number of nonzeros in any row
3905: Level: intermediate
3907: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
3908: @*/
3909: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz)
3910: {
3911: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
3914: *nz = aij->rmax;
3915: return(0);
3916: }
3918: /*@C
3919: MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray()
3921: Not Collective
3923: Input Parameters:
3924: . mat - a MATSEQAIJ matrix
3925: . array - pointer to the data
3927: Level: intermediate
3929: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90()
3930: @*/
3931: PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array)
3932: {
3936: PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));
3937: return(0);
3938: }
3940: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
3941: {
3942: Mat_SeqAIJ *b;
3944: PetscMPIInt size;
3947: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
3948: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1");
3950: PetscNewLog(B,&b);
3952: B->data = (void*)b;
3954: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
3956: b->row = 0;
3957: b->col = 0;
3958: b->icol = 0;
3959: b->reallocs = 0;
3960: b->ignorezeroentries = PETSC_FALSE;
3961: b->roworiented = PETSC_TRUE;
3962: b->nonew = 0;
3963: b->diag = 0;
3964: b->solve_work = 0;
3965: B->spptr = 0;
3966: b->saved_values = 0;
3967: b->idiag = 0;
3968: b->mdiag = 0;
3969: b->ssor_work = 0;
3970: b->omega = 1.0;
3971: b->fshift = 0.0;
3972: b->idiagvalid = PETSC_FALSE;
3973: b->ibdiagvalid = PETSC_FALSE;
3974: b->keepnonzeropattern = PETSC_FALSE;
3976: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
3977: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);
3978: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);
3980: #if defined(PETSC_HAVE_MATLAB_ENGINE)
3981: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);
3982: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);
3983: #endif
3985: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);
3986: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);
3987: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);
3988: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);
3989: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);
3990: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);
3991: #if defined(PETSC_HAVE_MKL_SPARSE)
3992: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijmkl_C",MatConvert_SeqAIJ_SeqAIJMKL);
3993: #endif
3994: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);
3995: #if defined(PETSC_HAVE_ELEMENTAL)
3996: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);
3997: #endif
3998: #if defined(PETSC_HAVE_HYPRE)
3999: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_hypre_C",MatConvert_AIJ_HYPRE);
4000: PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_seqaij_seqaij_C",MatMatMatMult_Transpose_AIJ_AIJ);
4001: #endif
4002: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);
4003: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);
4004: PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);
4005: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);
4006: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);
4007: PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);
4008: PetscObjectComposeFunction((PetscObject)B,"MatMatMult_seqdense_seqaij_C",MatMatMult_SeqDense_SeqAIJ);
4009: PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_seqdense_seqaij_C",MatMatMultSymbolic_SeqDense_SeqAIJ);
4010: PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_seqdense_seqaij_C",MatMatMultNumeric_SeqDense_SeqAIJ);
4011: MatCreate_SeqAIJ_Inode(B);
4012: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4013: MatSeqAIJSetTypeFromOptions(B); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4014: return(0);
4015: }
4017: /*
4018: Given a matrix generated with MatGetFactor() duplicates all the information in A into B
4019: */
4020: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace)
4021: {
4022: Mat_SeqAIJ *c,*a = (Mat_SeqAIJ*)A->data;
4024: PetscInt i,m = A->rmap->n;
4027: c = (Mat_SeqAIJ*)C->data;
4029: C->factortype = A->factortype;
4030: c->row = 0;
4031: c->col = 0;
4032: c->icol = 0;
4033: c->reallocs = 0;
4035: C->assembled = PETSC_TRUE;
4037: PetscLayoutReference(A->rmap,&C->rmap);
4038: PetscLayoutReference(A->cmap,&C->cmap);
4040: PetscMalloc2(m,&c->imax,m,&c->ilen);
4041: PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));
4042: for (i=0; i<m; i++) {
4043: c->imax[i] = a->imax[i];
4044: c->ilen[i] = a->ilen[i];
4045: }
4047: /* allocate the matrix space */
4048: if (mallocmatspace) {
4049: PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);
4050: PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));
4052: c->singlemalloc = PETSC_TRUE;
4054: PetscMemcpy(c->i,a->i,(m+1)*sizeof(PetscInt));
4055: if (m > 0) {
4056: PetscMemcpy(c->j,a->j,(a->i[m])*sizeof(PetscInt));
4057: if (cpvalues == MAT_COPY_VALUES) {
4058: PetscMemcpy(c->a,a->a,(a->i[m])*sizeof(PetscScalar));
4059: } else {
4060: PetscMemzero(c->a,(a->i[m])*sizeof(PetscScalar));
4061: }
4062: }
4063: }
4065: c->ignorezeroentries = a->ignorezeroentries;
4066: c->roworiented = a->roworiented;
4067: c->nonew = a->nonew;
4068: if (a->diag) {
4069: PetscMalloc1(m+1,&c->diag);
4070: PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));
4071: for (i=0; i<m; i++) {
4072: c->diag[i] = a->diag[i];
4073: }
4074: } else c->diag = 0;
4076: c->solve_work = 0;
4077: c->saved_values = 0;
4078: c->idiag = 0;
4079: c->ssor_work = 0;
4080: c->keepnonzeropattern = a->keepnonzeropattern;
4081: c->free_a = PETSC_TRUE;
4082: c->free_ij = PETSC_TRUE;
4084: c->rmax = a->rmax;
4085: c->nz = a->nz;
4086: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4087: C->preallocated = PETSC_TRUE;
4089: c->compressedrow.use = a->compressedrow.use;
4090: c->compressedrow.nrows = a->compressedrow.nrows;
4091: if (a->compressedrow.use) {
4092: i = a->compressedrow.nrows;
4093: PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);
4094: PetscMemcpy(c->compressedrow.i,a->compressedrow.i,(i+1)*sizeof(PetscInt));
4095: PetscMemcpy(c->compressedrow.rindex,a->compressedrow.rindex,i*sizeof(PetscInt));
4096: } else {
4097: c->compressedrow.use = PETSC_FALSE;
4098: c->compressedrow.i = NULL;
4099: c->compressedrow.rindex = NULL;
4100: }
4101: c->nonzerorowcnt = a->nonzerorowcnt;
4102: C->nonzerostate = A->nonzerostate;
4104: MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);
4105: PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);
4106: return(0);
4107: }
4109: PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B)
4110: {
4114: MatCreate(PetscObjectComm((PetscObject)A),B);
4115: MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);
4116: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) {
4117: MatSetBlockSizesFromMats(*B,A,A);
4118: }
4119: MatSetType(*B,((PetscObject)A)->type_name);
4120: MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);
4121: return(0);
4122: }
4124: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4125: {
4126: Mat_SeqAIJ *a;
4128: PetscInt i,sum,nz,header[4],*rowlengths = 0,M,N,rows,cols;
4129: int fd;
4130: PetscMPIInt size;
4131: MPI_Comm comm;
4132: PetscInt bs = newMat->rmap->bs;
4135: /* force binary viewer to load .info file if it has not yet done so */
4136: PetscViewerSetUp(viewer);
4137: PetscObjectGetComm((PetscObject)viewer,&comm);
4138: MPI_Comm_size(comm,&size);
4139: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"view must have one processor");
4141: PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");
4142: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
4143: PetscOptionsEnd();
4144: if (bs < 0) bs = 1;
4145: MatSetBlockSize(newMat,bs);
4147: PetscViewerBinaryGetDescriptor(viewer,&fd);
4148: PetscBinaryRead(fd,header,4,PETSC_INT);
4149: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object in file");
4150: M = header[1]; N = header[2]; nz = header[3];
4152: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as SeqAIJ");
4154: /* read in row lengths */
4155: PetscMalloc1(M,&rowlengths);
4156: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
4158: /* check if sum of rowlengths is same as nz */
4159: for (i=0,sum=0; i< M; i++) sum +=rowlengths[i];
4160: if (sum != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_READ,"Inconsistant matrix data in file. no-nonzeros = %dD, sum-row-lengths = %D\n",nz,sum);
4162: /* set global size if not set already*/
4163: if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) {
4164: MatSetSizes(newMat,PETSC_DECIDE,PETSC_DECIDE,M,N);
4165: } else {
4166: /* if sizes and type are already set, check if the matrix global sizes are correct */
4167: MatGetSize(newMat,&rows,&cols);
4168: if (rows < 0 && cols < 0) { /* user might provide local size instead of global size */
4169: MatGetLocalSize(newMat,&rows,&cols);
4170: }
4171: if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different length (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols);
4172: }
4173: MatSeqAIJSetPreallocation_SeqAIJ(newMat,0,rowlengths);
4174: a = (Mat_SeqAIJ*)newMat->data;
4176: PetscBinaryRead(fd,a->j,nz,PETSC_INT);
4178: /* read in nonzero values */
4179: PetscBinaryRead(fd,a->a,nz,PETSC_SCALAR);
4181: /* set matrix "i" values */
4182: a->i[0] = 0;
4183: for (i=1; i<= M; i++) {
4184: a->i[i] = a->i[i-1] + rowlengths[i-1];
4185: a->ilen[i-1] = rowlengths[i-1];
4186: }
4187: PetscFree(rowlengths);
4189: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
4190: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
4191: return(0);
4192: }
4194: PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg)
4195: {
4196: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
4198: #if defined(PETSC_USE_COMPLEX)
4199: PetscInt k;
4200: #endif
4203: /* If the matrix dimensions are not equal,or no of nonzeros */
4204: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) {
4205: *flg = PETSC_FALSE;
4206: return(0);
4207: }
4209: /* if the a->i are the same */
4210: PetscMemcmp(a->i,b->i,(A->rmap->n+1)*sizeof(PetscInt),flg);
4211: if (!*flg) return(0);
4213: /* if a->j are the same */
4214: PetscMemcmp(a->j,b->j,(a->nz)*sizeof(PetscInt),flg);
4215: if (!*flg) return(0);
4217: /* if a->a are the same */
4218: #if defined(PETSC_USE_COMPLEX)
4219: for (k=0; k<a->nz; k++) {
4220: if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) {
4221: *flg = PETSC_FALSE;
4222: return(0);
4223: }
4224: }
4225: #else
4226: PetscMemcmp(a->a,b->a,(a->nz)*sizeof(PetscScalar),flg);
4227: #endif
4228: return(0);
4229: }
4231: /*@
4232: MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format)
4233: provided by the user.
4235: Collective on MPI_Comm
4237: Input Parameters:
4238: + comm - must be an MPI communicator of size 1
4239: . m - number of rows
4240: . n - number of columns
4241: . i - row indices
4242: . j - column indices
4243: - a - matrix values
4245: Output Parameter:
4246: . mat - the matrix
4248: Level: intermediate
4250: Notes:
4251: The i, j, and a arrays are not copied by this routine, the user must free these arrays
4252: once the matrix is destroyed and not before
4254: You cannot set new nonzero locations into this matrix, that will generate an error.
4256: The i and j indices are 0 based
4258: The format which is used for the sparse matrix input, is equivalent to a
4259: row-major ordering.. i.e for the following matrix, the input data expected is
4260: as shown
4262: $ 1 0 0
4263: $ 2 0 3
4264: $ 4 5 6
4265: $
4266: $ i = {0,1,3,6} [size = nrow+1 = 3+1]
4267: $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
4268: $ v = {1,2,3,4,5,6} [size = 6]
4271: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4273: @*/
4274: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat)
4275: {
4277: PetscInt ii;
4278: Mat_SeqAIJ *aij;
4279: #if defined(PETSC_USE_DEBUG)
4280: PetscInt jj;
4281: #endif
4284: if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4285: MatCreate(comm,mat);
4286: MatSetSizes(*mat,m,n,m,n);
4287: /* MatSetBlockSizes(*mat,,); */
4288: MatSetType(*mat,MATSEQAIJ);
4289: MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,0);
4290: aij = (Mat_SeqAIJ*)(*mat)->data;
4291: PetscMalloc2(m,&aij->imax,m,&aij->ilen);
4293: aij->i = i;
4294: aij->j = j;
4295: aij->a = a;
4296: aij->singlemalloc = PETSC_FALSE;
4297: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
4298: aij->free_a = PETSC_FALSE;
4299: aij->free_ij = PETSC_FALSE;
4301: for (ii=0; ii<m; ii++) {
4302: aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii];
4303: #if defined(PETSC_USE_DEBUG)
4304: if (i[ii+1] - i[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row length in i (row indices) row = %D length = %D",ii,i[ii+1] - i[ii]);
4305: for (jj=i[ii]+1; jj<i[ii+1]; jj++) {
4306: if (j[jj] < j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual colum %D) in row %D is not sorted",jj-i[ii],j[jj],ii);
4307: if (j[jj] == j[jj]-1) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual colum %D) in row %D is identical to previous entry",jj-i[ii],j[jj],ii);
4308: }
4309: #endif
4310: }
4311: #if defined(PETSC_USE_DEBUG)
4312: for (ii=0; ii<aij->i[m]; ii++) {
4313: if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]);
4314: if (j[ii] > n - 1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column index to large at location = %D index = %D",ii,j[ii]);
4315: }
4316: #endif
4318: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4319: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4320: return(0);
4321: }
4322: /*@C
4323: MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format)
4324: provided by the user.
4326: Collective on MPI_Comm
4328: Input Parameters:
4329: + comm - must be an MPI communicator of size 1
4330: . m - number of rows
4331: . n - number of columns
4332: . i - row indices
4333: . j - column indices
4334: . a - matrix values
4335: . nz - number of nonzeros
4336: - idx - 0 or 1 based
4338: Output Parameter:
4339: . mat - the matrix
4341: Level: intermediate
4343: Notes:
4344: The i and j indices are 0 based
4346: The format which is used for the sparse matrix input, is equivalent to a
4347: row-major ordering.. i.e for the following matrix, the input data expected is
4348: as shown:
4350: 1 0 0
4351: 2 0 3
4352: 4 5 6
4354: i = {0,1,1,2,2,2}
4355: j = {0,0,2,0,1,2}
4356: v = {1,2,3,4,5,6}
4359: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4361: @*/
4362: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat,PetscInt nz,PetscBool idx)
4363: {
4365: PetscInt ii, *nnz, one = 1,row,col;
4369: PetscCalloc1(m,&nnz);
4370: for (ii = 0; ii < nz; ii++) {
4371: nnz[i[ii] - !!idx] += 1;
4372: }
4373: MatCreate(comm,mat);
4374: MatSetSizes(*mat,m,n,m,n);
4375: MatSetType(*mat,MATSEQAIJ);
4376: MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);
4377: for (ii = 0; ii < nz; ii++) {
4378: if (idx) {
4379: row = i[ii] - 1;
4380: col = j[ii] - 1;
4381: } else {
4382: row = i[ii];
4383: col = j[ii];
4384: }
4385: MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);
4386: }
4387: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4388: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4389: PetscFree(nnz);
4390: return(0);
4391: }
4393: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
4394: {
4395: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data;
4399: a->idiagvalid = PETSC_FALSE;
4400: a->ibdiagvalid = PETSC_FALSE;
4402: MatSeqAIJInvalidateDiagonal_Inode(A);
4403: return(0);
4404: }
4406: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4407: {
4409: PetscMPIInt size;
4412: MPI_Comm_size(comm,&size);
4413: if (size == 1) {
4414: if (scall == MAT_INITIAL_MATRIX) {
4415: MatDuplicate(inmat,MAT_COPY_VALUES,outmat);
4416: } else {
4417: MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);
4418: }
4419: } else {
4420: MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);
4421: }
4422: return(0);
4423: }
4425: /*
4426: Permute A into C's *local* index space using rowemb,colemb.
4427: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
4428: of [0,m), colemb is in [0,n).
4429: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
4430: */
4431: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B)
4432: {
4433: /* If making this function public, change the error returned in this function away from _PLIB. */
4435: Mat_SeqAIJ *Baij;
4436: PetscBool seqaij;
4437: PetscInt m,n,*nz,i,j,count;
4438: PetscScalar v;
4439: const PetscInt *rowindices,*colindices;
4442: if (!B) return(0);
4443: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
4444: PetscObjectBaseTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);
4445: if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type");
4446: if (rowemb) {
4447: ISGetLocalSize(rowemb,&m);
4448: if (m != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Row IS of size %D is incompatible with matrix row size %D",m,B->rmap->n);
4449: } else {
4450: if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix");
4451: }
4452: if (colemb) {
4453: ISGetLocalSize(colemb,&n);
4454: if (n != B->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag col IS of size %D is incompatible with input matrix col size %D",n,B->cmap->n);
4455: } else {
4456: if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix");
4457: }
4459: Baij = (Mat_SeqAIJ*)(B->data);
4460: if (pattern == DIFFERENT_NONZERO_PATTERN) {
4461: PetscMalloc1(B->rmap->n,&nz);
4462: for (i=0; i<B->rmap->n; i++) {
4463: nz[i] = Baij->i[i+1] - Baij->i[i];
4464: }
4465: MatSeqAIJSetPreallocation(C,0,nz);
4466: PetscFree(nz);
4467: }
4468: if (pattern == SUBSET_NONZERO_PATTERN) {
4469: MatZeroEntries(C);
4470: }
4471: count = 0;
4472: rowindices = NULL;
4473: colindices = NULL;
4474: if (rowemb) {
4475: ISGetIndices(rowemb,&rowindices);
4476: }
4477: if (colemb) {
4478: ISGetIndices(colemb,&colindices);
4479: }
4480: for (i=0; i<B->rmap->n; i++) {
4481: PetscInt row;
4482: row = i;
4483: if (rowindices) row = rowindices[i];
4484: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
4485: PetscInt col;
4486: col = Baij->j[count];
4487: if (colindices) col = colindices[col];
4488: v = Baij->a[count];
4489: MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);
4490: ++count;
4491: }
4492: }
4493: /* FIXME: set C's nonzerostate correctly. */
4494: /* Assembly for C is necessary. */
4495: C->preallocated = PETSC_TRUE;
4496: C->assembled = PETSC_TRUE;
4497: C->was_assembled = PETSC_FALSE;
4498: return(0);
4499: }
4501: PetscFunctionList MatSeqAIJList = NULL;
4503: /*@C
4504: MatSeqAIJSetType - Converts a MATSEQAIJ matrix to a subtype
4506: Collective on Mat
4508: Input Parameters:
4509: + mat - the matrix object
4510: - matype - matrix type
4512: Options Database Key:
4513: . -mat_seqai_type <method> - for example seqaijcrl
4516: Level: intermediate
4518: .keywords: Mat, MatType, set, method
4520: .seealso: PCSetType(), VecSetType(), MatCreate(), MatType, Mat
4521: @*/
4522: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
4523: {
4524: PetscErrorCode ierr,(*r)(Mat,const MatType,MatReuse,Mat*);
4525: PetscBool sametype;
4529: PetscObjectTypeCompare((PetscObject)mat,matype,&sametype);
4530: if (sametype) return(0);
4532: PetscFunctionListFind(MatSeqAIJList,matype,&r);
4533: if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown Mat type given: %s",matype);
4534: (*r)(mat,matype,MAT_INPLACE_MATRIX,&mat);
4535: return(0);
4536: }
4538:
4539: /*@C
4540: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential AIJ matrices
4542: Not Collective
4544: Input Parameters:
4545: + name - name of a new user-defined matrix type, for example MATSEQAIJCRL
4546: - function - routine to convert to subtype
4548: Notes:
4549: MatSeqAIJRegister() may be called multiple times to add several user-defined solvers.
4552: Then, your matrix can be chosen with the procedural interface at runtime via the option
4553: $ -mat_seqaij_type my_mat
4555: Level: advanced
4557: .keywords: Mat, register
4559: .seealso: MatSeqAIJRegisterAll()
4562: Level: advanced
4563: @*/
4564: PetscErrorCode MatSeqAIJRegister(const char sname[],PetscErrorCode (*function)(Mat,MatType,MatReuse,Mat *))
4565: {
4569: PetscFunctionListAdd(&MatSeqAIJList,sname,function);
4570: return(0);
4571: }
4573: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
4575: /*@C
4576: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of SeqAIJ
4578: Not Collective
4580: Level: advanced
4582: Developers Note: CUSP and CUSPARSE do not yet support the MatConvert_SeqAIJ..() paradigm and thus cannot be registered here
4584: .keywords: KSP, register, all
4586: .seealso: MatRegisterAll(), MatSeqAIJRegister()
4587: @*/
4588: PetscErrorCode MatSeqAIJRegisterAll(void)
4589: {
4593: if (MatSeqAIJRegisterAllCalled) return(0);
4594: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
4596: MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);
4597: MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);
4598: #if defined(PETSC_HAVE_MKL_SPARSE)
4599: MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);
4600: #endif
4601: #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA)
4602: MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);
4603: #endif
4604: return(0);
4605: }
4607: /*
4608: Special version for direct calls from Fortran
4609: */
4610: #include <petsc/private/fortranimpl.h>
4611: #if defined(PETSC_HAVE_FORTRAN_CAPS)
4612: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
4613: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
4614: #define matsetvaluesseqaij_ matsetvaluesseqaij
4615: #endif
4617: /* Change these macros so can be used in void function */
4618: #undef CHKERRQ
4619: #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr)
4620: #undef SETERRQ2
4621: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
4622: #undef SETERRQ3
4623: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
4625: PETSC_EXTERN void PETSC_STDCALL matsetvaluesseqaij_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[],InsertMode *isis, PetscErrorCode *_ierr)
4626: {
4627: Mat A = *AA;
4628: PetscInt m = *mm, n = *nn;
4629: InsertMode is = *isis;
4630: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
4631: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
4632: PetscInt *imax,*ai,*ailen;
4634: PetscInt *aj,nonew = a->nonew,lastcol = -1;
4635: MatScalar *ap,value,*aa;
4636: PetscBool ignorezeroentries = a->ignorezeroentries;
4637: PetscBool roworiented = a->roworiented;
4640: MatCheckPreallocated(A,1);
4641: imax = a->imax;
4642: ai = a->i;
4643: ailen = a->ilen;
4644: aj = a->j;
4645: aa = a->a;
4647: for (k=0; k<m; k++) { /* loop over added rows */
4648: row = im[k];
4649: if (row < 0) continue;
4650: #if defined(PETSC_USE_DEBUG)
4651: if (row >= A->rmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
4652: #endif
4653: rp = aj + ai[row]; ap = aa + ai[row];
4654: rmax = imax[row]; nrow = ailen[row];
4655: low = 0;
4656: high = nrow;
4657: for (l=0; l<n; l++) { /* loop over added columns */
4658: if (in[l] < 0) continue;
4659: #if defined(PETSC_USE_DEBUG)
4660: if (in[l] >= A->cmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
4661: #endif
4662: col = in[l];
4663: if (roworiented) value = v[l + k*n];
4664: else value = v[k + l*m];
4666: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
4668: if (col <= lastcol) low = 0;
4669: else high = nrow;
4670: lastcol = col;
4671: while (high-low > 5) {
4672: t = (low+high)/2;
4673: if (rp[t] > col) high = t;
4674: else low = t;
4675: }
4676: for (i=low; i<high; i++) {
4677: if (rp[i] > col) break;
4678: if (rp[i] == col) {
4679: if (is == ADD_VALUES) ap[i] += value;
4680: else ap[i] = value;
4681: goto noinsert;
4682: }
4683: }
4684: if (value == 0.0 && ignorezeroentries) goto noinsert;
4685: if (nonew == 1) goto noinsert;
4686: if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix");
4687: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
4688: N = nrow++ - 1; a->nz++; high++;
4689: /* shift up all the later entries in this row */
4690: for (ii=N; ii>=i; ii--) {
4691: rp[ii+1] = rp[ii];
4692: ap[ii+1] = ap[ii];
4693: }
4694: rp[i] = col;
4695: ap[i] = value;
4696: A->nonzerostate++;
4697: noinsert:;
4698: low = i + 1;
4699: }
4700: ailen[row] = nrow;
4701: }
4702: PetscFunctionReturnVoid();
4703: }