Actual source code: aij.c
petsc-3.9.1 2018-04-29
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: PetscFree(a->ipre);
1098: PetscFree3(a->idiag,a->mdiag,a->ssor_work);
1099: PetscFree(a->solve_work);
1100: ISDestroy(&a->icol);
1101: PetscFree(a->saved_values);
1102: ISColoringDestroy(&a->coloring);
1103: PetscFree2(a->compressedrow.i,a->compressedrow.rindex);
1104: PetscFree(a->matmult_abdense);
1106: MatDestroy_SeqAIJ_Inode(A);
1107: PetscFree(A->data);
1109: PetscObjectChangeTypeName((PetscObject)A,0);
1110: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetColumnIndices_C",NULL);
1111: PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);
1112: PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);
1113: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsbaij_C",NULL);
1114: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqbaij_C",NULL);
1115: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijperm_C",NULL);
1116: #if defined(PETSC_HAVE_ELEMENTAL)
1117: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_elemental_C",NULL);
1118: #endif
1119: #if defined(PETSC_HAVE_HYPRE)
1120: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_hypre_C",NULL);
1121: PetscObjectComposeFunction((PetscObject)A,"MatMatMatMult_transpose_seqaij_seqaij_C",NULL);
1122: #endif
1123: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqdense_C",NULL);
1124: PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);
1125: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);
1126: PetscObjectComposeFunction((PetscObject)A,"MatResetPreallocation_C",NULL);
1127: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);
1128: PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);
1129: return(0);
1130: }
1132: PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg)
1133: {
1134: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1138: switch (op) {
1139: case MAT_ROW_ORIENTED:
1140: a->roworiented = flg;
1141: break;
1142: case MAT_KEEP_NONZERO_PATTERN:
1143: a->keepnonzeropattern = flg;
1144: break;
1145: case MAT_NEW_NONZERO_LOCATIONS:
1146: a->nonew = (flg ? 0 : 1);
1147: break;
1148: case MAT_NEW_NONZERO_LOCATION_ERR:
1149: a->nonew = (flg ? -1 : 0);
1150: break;
1151: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1152: a->nonew = (flg ? -2 : 0);
1153: break;
1154: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1155: a->nounused = (flg ? -1 : 0);
1156: break;
1157: case MAT_IGNORE_ZERO_ENTRIES:
1158: a->ignorezeroentries = flg;
1159: break;
1160: case MAT_SPD:
1161: case MAT_SYMMETRIC:
1162: case MAT_STRUCTURALLY_SYMMETRIC:
1163: case MAT_HERMITIAN:
1164: case MAT_SYMMETRY_ETERNAL:
1165: case MAT_STRUCTURE_ONLY:
1166: /* These options are handled directly by MatSetOption() */
1167: break;
1168: case MAT_NEW_DIAGONALS:
1169: case MAT_IGNORE_OFF_PROC_ENTRIES:
1170: case MAT_USE_HASH_TABLE:
1171: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1172: break;
1173: case MAT_USE_INODES:
1174: /* Not an error because MatSetOption_SeqAIJ_Inode handles this one */
1175: break;
1176: case MAT_SUBMAT_SINGLEIS:
1177: A->submat_singleis = flg;
1178: break;
1179: default:
1180: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1181: }
1182: MatSetOption_SeqAIJ_Inode(A,op,flg);
1183: return(0);
1184: }
1186: PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v)
1187: {
1188: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1190: PetscInt i,j,n,*ai=a->i,*aj=a->j,nz;
1191: PetscScalar *aa=a->a,*x,zero=0.0;
1194: VecGetLocalSize(v,&n);
1195: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
1197: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1198: PetscInt *diag=a->diag;
1199: VecGetArray(v,&x);
1200: for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]];
1201: VecRestoreArray(v,&x);
1202: return(0);
1203: }
1205: VecSet(v,zero);
1206: VecGetArray(v,&x);
1207: for (i=0; i<n; i++) {
1208: nz = ai[i+1] - ai[i];
1209: if (!nz) x[i] = 0.0;
1210: for (j=ai[i]; j<ai[i+1]; j++) {
1211: if (aj[j] == i) {
1212: x[i] = aa[j];
1213: break;
1214: }
1215: }
1216: }
1217: VecRestoreArray(v,&x);
1218: return(0);
1219: }
1221: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1222: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy)
1223: {
1224: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1225: PetscScalar *y;
1226: const PetscScalar *x;
1227: PetscErrorCode ierr;
1228: PetscInt m = A->rmap->n;
1229: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1230: const MatScalar *v;
1231: PetscScalar alpha;
1232: PetscInt n,i,j;
1233: const PetscInt *idx,*ii,*ridx=NULL;
1234: Mat_CompressedRow cprow = a->compressedrow;
1235: PetscBool usecprow = cprow.use;
1236: #endif
1239: if (zz != yy) {VecCopy(zz,yy);}
1240: VecGetArrayRead(xx,&x);
1241: VecGetArray(yy,&y);
1243: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1244: fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y);
1245: #else
1246: if (usecprow) {
1247: m = cprow.nrows;
1248: ii = cprow.i;
1249: ridx = cprow.rindex;
1250: } else {
1251: ii = a->i;
1252: }
1253: for (i=0; i<m; i++) {
1254: idx = a->j + ii[i];
1255: v = a->a + ii[i];
1256: n = ii[i+1] - ii[i];
1257: if (usecprow) {
1258: alpha = x[ridx[i]];
1259: } else {
1260: alpha = x[i];
1261: }
1262: for (j=0; j<n; j++) y[idx[j]] += alpha*v[j];
1263: }
1264: #endif
1265: PetscLogFlops(2.0*a->nz);
1266: VecRestoreArrayRead(xx,&x);
1267: VecRestoreArray(yy,&y);
1268: return(0);
1269: }
1271: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy)
1272: {
1276: VecSet(yy,0.0);
1277: MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);
1278: return(0);
1279: }
1281: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1283: PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy)
1284: {
1285: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1286: PetscScalar *y;
1287: const PetscScalar *x;
1288: const MatScalar *aa;
1289: PetscErrorCode ierr;
1290: PetscInt m=A->rmap->n;
1291: const PetscInt *aj,*ii,*ridx=NULL;
1292: PetscInt n,i;
1293: PetscScalar sum;
1294: PetscBool usecprow=a->compressedrow.use;
1296: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1297: #pragma disjoint(*x,*y,*aa)
1298: #endif
1301: VecGetArrayRead(xx,&x);
1302: VecGetArray(yy,&y);
1303: ii = a->i;
1304: if (usecprow) { /* use compressed row format */
1305: PetscMemzero(y,m*sizeof(PetscScalar));
1306: m = a->compressedrow.nrows;
1307: ii = a->compressedrow.i;
1308: ridx = a->compressedrow.rindex;
1309: for (i=0; i<m; i++) {
1310: n = ii[i+1] - ii[i];
1311: aj = a->j + ii[i];
1312: aa = a->a + ii[i];
1313: sum = 0.0;
1314: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1315: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1316: y[*ridx++] = sum;
1317: }
1318: } else { /* do not use compressed row format */
1319: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1320: aj = a->j;
1321: aa = a->a;
1322: fortranmultaij_(&m,x,ii,aj,aa,y);
1323: #else
1324: for (i=0; i<m; i++) {
1325: n = ii[i+1] - ii[i];
1326: aj = a->j + ii[i];
1327: aa = a->a + ii[i];
1328: sum = 0.0;
1329: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1330: y[i] = sum;
1331: }
1332: #endif
1333: }
1334: PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);
1335: VecRestoreArrayRead(xx,&x);
1336: VecRestoreArray(yy,&y);
1337: return(0);
1338: }
1340: PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy)
1341: {
1342: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1343: PetscScalar *y;
1344: const PetscScalar *x;
1345: const MatScalar *aa;
1346: PetscErrorCode ierr;
1347: PetscInt m=A->rmap->n;
1348: const PetscInt *aj,*ii,*ridx=NULL;
1349: PetscInt n,i,nonzerorow=0;
1350: PetscScalar sum;
1351: PetscBool usecprow=a->compressedrow.use;
1353: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1354: #pragma disjoint(*x,*y,*aa)
1355: #endif
1358: VecGetArrayRead(xx,&x);
1359: VecGetArray(yy,&y);
1360: if (usecprow) { /* use compressed row format */
1361: m = a->compressedrow.nrows;
1362: ii = a->compressedrow.i;
1363: ridx = a->compressedrow.rindex;
1364: for (i=0; i<m; i++) {
1365: n = ii[i+1] - ii[i];
1366: aj = a->j + ii[i];
1367: aa = a->a + ii[i];
1368: sum = 0.0;
1369: nonzerorow += (n>0);
1370: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1371: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1372: y[*ridx++] = sum;
1373: }
1374: } else { /* do not use compressed row format */
1375: ii = a->i;
1376: for (i=0; i<m; i++) {
1377: n = ii[i+1] - ii[i];
1378: aj = a->j + ii[i];
1379: aa = a->a + ii[i];
1380: sum = 0.0;
1381: nonzerorow += (n>0);
1382: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1383: y[i] = sum;
1384: }
1385: }
1386: PetscLogFlops(2.0*a->nz - nonzerorow);
1387: VecRestoreArrayRead(xx,&x);
1388: VecRestoreArray(yy,&y);
1389: return(0);
1390: }
1392: PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1393: {
1394: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1395: PetscScalar *y,*z;
1396: const PetscScalar *x;
1397: const MatScalar *aa;
1398: PetscErrorCode ierr;
1399: PetscInt m = A->rmap->n,*aj,*ii;
1400: PetscInt n,i,*ridx=NULL;
1401: PetscScalar sum;
1402: PetscBool usecprow=a->compressedrow.use;
1405: VecGetArrayRead(xx,&x);
1406: VecGetArrayPair(yy,zz,&y,&z);
1407: if (usecprow) { /* use compressed row format */
1408: if (zz != yy) {
1409: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1410: }
1411: m = a->compressedrow.nrows;
1412: ii = a->compressedrow.i;
1413: ridx = a->compressedrow.rindex;
1414: for (i=0; i<m; i++) {
1415: n = ii[i+1] - ii[i];
1416: aj = a->j + ii[i];
1417: aa = a->a + ii[i];
1418: sum = y[*ridx];
1419: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1420: z[*ridx++] = sum;
1421: }
1422: } else { /* do not use compressed row format */
1423: ii = a->i;
1424: for (i=0; i<m; i++) {
1425: n = ii[i+1] - ii[i];
1426: aj = a->j + ii[i];
1427: aa = a->a + ii[i];
1428: sum = y[i];
1429: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1430: z[i] = sum;
1431: }
1432: }
1433: PetscLogFlops(2.0*a->nz);
1434: VecRestoreArrayRead(xx,&x);
1435: VecRestoreArrayPair(yy,zz,&y,&z);
1436: return(0);
1437: }
1439: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1440: PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1441: {
1442: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1443: PetscScalar *y,*z;
1444: const PetscScalar *x;
1445: const MatScalar *aa;
1446: PetscErrorCode ierr;
1447: const PetscInt *aj,*ii,*ridx=NULL;
1448: PetscInt m = A->rmap->n,n,i;
1449: PetscScalar sum;
1450: PetscBool usecprow=a->compressedrow.use;
1453: VecGetArrayRead(xx,&x);
1454: VecGetArrayPair(yy,zz,&y,&z);
1455: if (usecprow) { /* use compressed row format */
1456: if (zz != yy) {
1457: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1458: }
1459: m = a->compressedrow.nrows;
1460: ii = a->compressedrow.i;
1461: ridx = a->compressedrow.rindex;
1462: for (i=0; i<m; i++) {
1463: n = ii[i+1] - ii[i];
1464: aj = a->j + ii[i];
1465: aa = a->a + ii[i];
1466: sum = y[*ridx];
1467: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1468: z[*ridx++] = sum;
1469: }
1470: } else { /* do not use compressed row format */
1471: ii = a->i;
1472: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1473: aj = a->j;
1474: aa = a->a;
1475: fortranmultaddaij_(&m,x,ii,aj,aa,y,z);
1476: #else
1477: for (i=0; i<m; i++) {
1478: n = ii[i+1] - ii[i];
1479: aj = a->j + ii[i];
1480: aa = a->a + ii[i];
1481: sum = y[i];
1482: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1483: z[i] = sum;
1484: }
1485: #endif
1486: }
1487: PetscLogFlops(2.0*a->nz);
1488: VecRestoreArrayRead(xx,&x);
1489: VecRestoreArrayPair(yy,zz,&y,&z);
1490: return(0);
1491: }
1493: /*
1494: Adds diagonal pointers to sparse matrix structure.
1495: */
1496: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1497: {
1498: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1500: PetscInt i,j,m = A->rmap->n;
1503: if (!a->diag) {
1504: PetscMalloc1(m,&a->diag);
1505: PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));
1506: }
1507: for (i=0; i<A->rmap->n; i++) {
1508: a->diag[i] = a->i[i+1];
1509: for (j=a->i[i]; j<a->i[i+1]; j++) {
1510: if (a->j[j] == i) {
1511: a->diag[i] = j;
1512: break;
1513: }
1514: }
1515: }
1516: return(0);
1517: }
1519: /*
1520: Checks for missing diagonals
1521: */
1522: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d)
1523: {
1524: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1525: PetscInt *diag,*ii = a->i,i;
1528: *missing = PETSC_FALSE;
1529: if (A->rmap->n > 0 && !ii) {
1530: *missing = PETSC_TRUE;
1531: if (d) *d = 0;
1532: PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");
1533: } else {
1534: diag = a->diag;
1535: for (i=0; i<A->rmap->n; i++) {
1536: if (diag[i] >= ii[i+1]) {
1537: *missing = PETSC_TRUE;
1538: if (d) *d = i;
1539: PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);
1540: break;
1541: }
1542: }
1543: }
1544: return(0);
1545: }
1547: /*
1548: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1549: */
1550: PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift)
1551: {
1552: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
1554: PetscInt i,*diag,m = A->rmap->n;
1555: MatScalar *v = a->a;
1556: PetscScalar *idiag,*mdiag;
1559: if (a->idiagvalid) return(0);
1560: MatMarkDiagonal_SeqAIJ(A);
1561: diag = a->diag;
1562: if (!a->idiag) {
1563: PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);
1564: PetscLogObjectMemory((PetscObject)A, 3*m*sizeof(PetscScalar));
1565: v = a->a;
1566: }
1567: mdiag = a->mdiag;
1568: idiag = a->idiag;
1570: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1571: for (i=0; i<m; i++) {
1572: mdiag[i] = v[diag[i]];
1573: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1574: if (PetscRealPart(fshift)) {
1575: PetscInfo1(A,"Zero diagonal on row %D\n",i);
1576: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1577: A->factorerror_zeropivot_value = 0.0;
1578: A->factorerror_zeropivot_row = i;
1579: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i);
1580: }
1581: idiag[i] = 1.0/v[diag[i]];
1582: }
1583: PetscLogFlops(m);
1584: } else {
1585: for (i=0; i<m; i++) {
1586: mdiag[i] = v[diag[i]];
1587: idiag[i] = omega/(fshift + v[diag[i]]);
1588: }
1589: PetscLogFlops(2.0*m);
1590: }
1591: a->idiagvalid = PETSC_TRUE;
1592: return(0);
1593: }
1595: #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h>
1596: PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1597: {
1598: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1599: PetscScalar *x,d,sum,*t,scale;
1600: const MatScalar *v,*idiag=0,*mdiag;
1601: const PetscScalar *b, *bs,*xb, *ts;
1602: PetscErrorCode ierr;
1603: PetscInt n,m = A->rmap->n,i;
1604: const PetscInt *idx,*diag;
1607: its = its*lits;
1609: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1610: if (!a->idiagvalid) {MatInvertDiagonal_SeqAIJ(A,omega,fshift);}
1611: a->fshift = fshift;
1612: a->omega = omega;
1614: diag = a->diag;
1615: t = a->ssor_work;
1616: idiag = a->idiag;
1617: mdiag = a->mdiag;
1619: VecGetArray(xx,&x);
1620: VecGetArrayRead(bb,&b);
1621: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1622: if (flag == SOR_APPLY_UPPER) {
1623: /* apply (U + D/omega) to the vector */
1624: bs = b;
1625: for (i=0; i<m; i++) {
1626: d = fshift + mdiag[i];
1627: n = a->i[i+1] - diag[i] - 1;
1628: idx = a->j + diag[i] + 1;
1629: v = a->a + diag[i] + 1;
1630: sum = b[i]*d/omega;
1631: PetscSparseDensePlusDot(sum,bs,v,idx,n);
1632: x[i] = sum;
1633: }
1634: VecRestoreArray(xx,&x);
1635: VecRestoreArrayRead(bb,&b);
1636: PetscLogFlops(a->nz);
1637: return(0);
1638: }
1640: if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented");
1641: else if (flag & SOR_EISENSTAT) {
1642: /* Let A = L + U + D; where L is lower trianglar,
1643: U is upper triangular, E = D/omega; This routine applies
1645: (L + E)^{-1} A (U + E)^{-1}
1647: to a vector efficiently using Eisenstat's trick.
1648: */
1649: scale = (2.0/omega) - 1.0;
1651: /* x = (E + U)^{-1} b */
1652: for (i=m-1; i>=0; i--) {
1653: n = a->i[i+1] - diag[i] - 1;
1654: idx = a->j + diag[i] + 1;
1655: v = a->a + diag[i] + 1;
1656: sum = b[i];
1657: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1658: x[i] = sum*idiag[i];
1659: }
1661: /* t = b - (2*E - D)x */
1662: v = a->a;
1663: for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i];
1665: /* t = (E + L)^{-1}t */
1666: ts = t;
1667: diag = a->diag;
1668: for (i=0; i<m; i++) {
1669: n = diag[i] - a->i[i];
1670: idx = a->j + a->i[i];
1671: v = a->a + a->i[i];
1672: sum = t[i];
1673: PetscSparseDenseMinusDot(sum,ts,v,idx,n);
1674: t[i] = sum*idiag[i];
1675: /* x = x + t */
1676: x[i] += t[i];
1677: }
1679: PetscLogFlops(6.0*m-1 + 2.0*a->nz);
1680: VecRestoreArray(xx,&x);
1681: VecRestoreArrayRead(bb,&b);
1682: return(0);
1683: }
1684: if (flag & SOR_ZERO_INITIAL_GUESS) {
1685: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1686: for (i=0; i<m; i++) {
1687: n = diag[i] - a->i[i];
1688: idx = a->j + a->i[i];
1689: v = a->a + a->i[i];
1690: sum = b[i];
1691: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1692: t[i] = sum;
1693: x[i] = sum*idiag[i];
1694: }
1695: xb = t;
1696: PetscLogFlops(a->nz);
1697: } else xb = b;
1698: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1699: for (i=m-1; i>=0; i--) {
1700: n = a->i[i+1] - diag[i] - 1;
1701: idx = a->j + diag[i] + 1;
1702: v = a->a + diag[i] + 1;
1703: sum = xb[i];
1704: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1705: if (xb == b) {
1706: x[i] = sum*idiag[i];
1707: } else {
1708: x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1709: }
1710: }
1711: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1712: }
1713: its--;
1714: }
1715: while (its--) {
1716: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1717: for (i=0; i<m; i++) {
1718: /* lower */
1719: n = diag[i] - a->i[i];
1720: idx = a->j + a->i[i];
1721: v = a->a + a->i[i];
1722: sum = b[i];
1723: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1724: t[i] = sum; /* save application of the lower-triangular part */
1725: /* upper */
1726: n = a->i[i+1] - diag[i] - 1;
1727: idx = a->j + diag[i] + 1;
1728: v = a->a + diag[i] + 1;
1729: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1730: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1731: }
1732: xb = t;
1733: PetscLogFlops(2.0*a->nz);
1734: } else xb = b;
1735: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1736: for (i=m-1; i>=0; i--) {
1737: sum = xb[i];
1738: if (xb == b) {
1739: /* whole matrix (no checkpointing available) */
1740: n = a->i[i+1] - a->i[i];
1741: idx = a->j + a->i[i];
1742: v = a->a + a->i[i];
1743: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1744: x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i];
1745: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
1746: n = a->i[i+1] - diag[i] - 1;
1747: idx = a->j + diag[i] + 1;
1748: v = a->a + diag[i] + 1;
1749: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1750: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1751: }
1752: }
1753: if (xb == b) {
1754: PetscLogFlops(2.0*a->nz);
1755: } else {
1756: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1757: }
1758: }
1759: }
1760: VecRestoreArray(xx,&x);
1761: VecRestoreArrayRead(bb,&b);
1762: return(0);
1763: }
1766: PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info)
1767: {
1768: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1771: info->block_size = 1.0;
1772: info->nz_allocated = (double)a->maxnz;
1773: info->nz_used = (double)a->nz;
1774: info->nz_unneeded = (double)(a->maxnz - a->nz);
1775: info->assemblies = (double)A->num_ass;
1776: info->mallocs = (double)A->info.mallocs;
1777: info->memory = ((PetscObject)A)->mem;
1778: if (A->factortype) {
1779: info->fill_ratio_given = A->info.fill_ratio_given;
1780: info->fill_ratio_needed = A->info.fill_ratio_needed;
1781: info->factor_mallocs = A->info.factor_mallocs;
1782: } else {
1783: info->fill_ratio_given = 0;
1784: info->fill_ratio_needed = 0;
1785: info->factor_mallocs = 0;
1786: }
1787: return(0);
1788: }
1790: PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1791: {
1792: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1793: PetscInt i,m = A->rmap->n - 1;
1794: PetscErrorCode ierr;
1795: const PetscScalar *xx;
1796: PetscScalar *bb;
1797: PetscInt d = 0;
1800: if (x && b) {
1801: VecGetArrayRead(x,&xx);
1802: VecGetArray(b,&bb);
1803: for (i=0; i<N; i++) {
1804: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1805: bb[rows[i]] = diag*xx[rows[i]];
1806: }
1807: VecRestoreArrayRead(x,&xx);
1808: VecRestoreArray(b,&bb);
1809: }
1811: if (a->keepnonzeropattern) {
1812: for (i=0; i<N; i++) {
1813: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1814: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1815: }
1816: if (diag != 0.0) {
1817: for (i=0; i<N; i++) {
1818: d = rows[i];
1819: 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);
1820: }
1821: for (i=0; i<N; i++) {
1822: a->a[a->diag[rows[i]]] = diag;
1823: }
1824: }
1825: } else {
1826: if (diag != 0.0) {
1827: for (i=0; i<N; i++) {
1828: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1829: if (a->ilen[rows[i]] > 0) {
1830: a->ilen[rows[i]] = 1;
1831: a->a[a->i[rows[i]]] = diag;
1832: a->j[a->i[rows[i]]] = rows[i];
1833: } else { /* in case row was completely empty */
1834: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
1835: }
1836: }
1837: } else {
1838: for (i=0; i<N; i++) {
1839: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1840: a->ilen[rows[i]] = 0;
1841: }
1842: }
1843: A->nonzerostate++;
1844: }
1845: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
1846: return(0);
1847: }
1849: PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1850: {
1851: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1852: PetscInt i,j,m = A->rmap->n - 1,d = 0;
1853: PetscErrorCode ierr;
1854: PetscBool missing,*zeroed,vecs = PETSC_FALSE;
1855: const PetscScalar *xx;
1856: PetscScalar *bb;
1859: if (x && b) {
1860: VecGetArrayRead(x,&xx);
1861: VecGetArray(b,&bb);
1862: vecs = PETSC_TRUE;
1863: }
1864: PetscCalloc1(A->rmap->n,&zeroed);
1865: for (i=0; i<N; i++) {
1866: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1867: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1869: zeroed[rows[i]] = PETSC_TRUE;
1870: }
1871: for (i=0; i<A->rmap->n; i++) {
1872: if (!zeroed[i]) {
1873: for (j=a->i[i]; j<a->i[i+1]; j++) {
1874: if (zeroed[a->j[j]]) {
1875: if (vecs) bb[i] -= a->a[j]*xx[a->j[j]];
1876: a->a[j] = 0.0;
1877: }
1878: }
1879: } else if (vecs) bb[i] = diag*xx[i];
1880: }
1881: if (x && b) {
1882: VecRestoreArrayRead(x,&xx);
1883: VecRestoreArray(b,&bb);
1884: }
1885: PetscFree(zeroed);
1886: if (diag != 0.0) {
1887: MatMissingDiagonal_SeqAIJ(A,&missing,&d);
1888: if (missing) {
1889: if (a->nonew) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D",d);
1890: else {
1891: for (i=0; i<N; i++) {
1892: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
1893: }
1894: }
1895: } else {
1896: for (i=0; i<N; i++) {
1897: a->a[a->diag[rows[i]]] = diag;
1898: }
1899: }
1900: }
1901: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
1902: return(0);
1903: }
1905: PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1906: {
1907: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1908: PetscInt *itmp;
1911: if (row < 0 || row >= A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range",row);
1913: *nz = a->i[row+1] - a->i[row];
1914: if (v) *v = a->a + a->i[row];
1915: if (idx) {
1916: itmp = a->j + a->i[row];
1917: if (*nz) *idx = itmp;
1918: else *idx = 0;
1919: }
1920: return(0);
1921: }
1923: /* remove this function? */
1924: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1925: {
1927: return(0);
1928: }
1930: PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm)
1931: {
1932: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1933: MatScalar *v = a->a;
1934: PetscReal sum = 0.0;
1936: PetscInt i,j;
1939: if (type == NORM_FROBENIUS) {
1940: #if defined(PETSC_USE_REAL___FP16)
1941: PetscBLASInt one = 1,nz = a->nz;
1942: *nrm = BLASnrm2_(&nz,v,&one);
1943: #else
1944: for (i=0; i<a->nz; i++) {
1945: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1946: }
1947: *nrm = PetscSqrtReal(sum);
1948: #endif
1949: PetscLogFlops(2*a->nz);
1950: } else if (type == NORM_1) {
1951: PetscReal *tmp;
1952: PetscInt *jj = a->j;
1953: PetscCalloc1(A->cmap->n+1,&tmp);
1954: *nrm = 0.0;
1955: for (j=0; j<a->nz; j++) {
1956: tmp[*jj++] += PetscAbsScalar(*v); v++;
1957: }
1958: for (j=0; j<A->cmap->n; j++) {
1959: if (tmp[j] > *nrm) *nrm = tmp[j];
1960: }
1961: PetscFree(tmp);
1962: PetscLogFlops(PetscMax(a->nz-1,0));
1963: } else if (type == NORM_INFINITY) {
1964: *nrm = 0.0;
1965: for (j=0; j<A->rmap->n; j++) {
1966: v = a->a + a->i[j];
1967: sum = 0.0;
1968: for (i=0; i<a->i[j+1]-a->i[j]; i++) {
1969: sum += PetscAbsScalar(*v); v++;
1970: }
1971: if (sum > *nrm) *nrm = sum;
1972: }
1973: PetscLogFlops(PetscMax(a->nz-1,0));
1974: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm");
1975: return(0);
1976: }
1978: /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */
1979: PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B)
1980: {
1982: PetscInt i,j,anzj;
1983: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b;
1984: PetscInt an=A->cmap->N,am=A->rmap->N;
1985: PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j;
1988: /* Allocate space for symbolic transpose info and work array */
1989: PetscCalloc1(an+1,&ati);
1990: PetscMalloc1(ai[am],&atj);
1991: PetscMalloc1(an,&atfill);
1993: /* Walk through aj and count ## of non-zeros in each row of A^T. */
1994: /* Note: offset by 1 for fast conversion into csr format. */
1995: for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1;
1996: /* Form ati for csr format of A^T. */
1997: for (i=0;i<an;i++) ati[i+1] += ati[i];
1999: /* Copy ati into atfill so we have locations of the next free space in atj */
2000: PetscMemcpy(atfill,ati,an*sizeof(PetscInt));
2002: /* Walk through A row-wise and mark nonzero entries of A^T. */
2003: for (i=0;i<am;i++) {
2004: anzj = ai[i+1] - ai[i];
2005: for (j=0;j<anzj;j++) {
2006: atj[atfill[*aj]] = i;
2007: atfill[*aj++] += 1;
2008: }
2009: }
2011: /* Clean up temporary space and complete requests. */
2012: PetscFree(atfill);
2013: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);
2014: MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2016: b = (Mat_SeqAIJ*)((*B)->data);
2017: b->free_a = PETSC_FALSE;
2018: b->free_ij = PETSC_TRUE;
2019: b->nonew = 0;
2020: return(0);
2021: }
2023: PetscErrorCode MatTranspose_SeqAIJ(Mat A,MatReuse reuse,Mat *B)
2024: {
2025: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2026: Mat C;
2028: PetscInt i,*aj = a->j,*ai = a->i,m = A->rmap->n,len,*col;
2029: MatScalar *array = a->a;
2032: if (reuse == MAT_INPLACE_MATRIX && m != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
2034: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
2035: PetscCalloc1(1+A->cmap->n,&col);
2037: for (i=0; i<ai[m]; i++) col[aj[i]] += 1;
2038: MatCreate(PetscObjectComm((PetscObject)A),&C);
2039: MatSetSizes(C,A->cmap->n,m,A->cmap->n,m);
2040: MatSetBlockSizes(C,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2041: MatSetType(C,((PetscObject)A)->type_name);
2042: MatSeqAIJSetPreallocation_SeqAIJ(C,0,col);
2043: PetscFree(col);
2044: } else {
2045: C = *B;
2046: }
2048: for (i=0; i<m; i++) {
2049: len = ai[i+1]-ai[i];
2050: MatSetValues_SeqAIJ(C,len,aj,1,&i,array,INSERT_VALUES);
2051: array += len;
2052: aj += len;
2053: }
2054: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2055: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2057: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2058: *B = C;
2059: } else {
2060: MatHeaderMerge(A,&C);
2061: }
2062: return(0);
2063: }
2065: PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2066: {
2067: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2068: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2069: MatScalar *va,*vb;
2071: PetscInt ma,na,mb,nb, i;
2074: MatGetSize(A,&ma,&na);
2075: MatGetSize(B,&mb,&nb);
2076: if (ma!=nb || na!=mb) {
2077: *f = PETSC_FALSE;
2078: return(0);
2079: }
2080: aii = aij->i; bii = bij->i;
2081: adx = aij->j; bdx = bij->j;
2082: va = aij->a; vb = bij->a;
2083: PetscMalloc1(ma,&aptr);
2084: PetscMalloc1(mb,&bptr);
2085: for (i=0; i<ma; i++) aptr[i] = aii[i];
2086: for (i=0; i<mb; i++) bptr[i] = bii[i];
2088: *f = PETSC_TRUE;
2089: for (i=0; i<ma; i++) {
2090: while (aptr[i]<aii[i+1]) {
2091: PetscInt idc,idr;
2092: PetscScalar vc,vr;
2093: /* column/row index/value */
2094: idc = adx[aptr[i]];
2095: idr = bdx[bptr[idc]];
2096: vc = va[aptr[i]];
2097: vr = vb[bptr[idc]];
2098: if (i!=idr || PetscAbsScalar(vc-vr) > tol) {
2099: *f = PETSC_FALSE;
2100: goto done;
2101: } else {
2102: aptr[i]++;
2103: if (B || i!=idc) bptr[idc]++;
2104: }
2105: }
2106: }
2107: done:
2108: PetscFree(aptr);
2109: PetscFree(bptr);
2110: return(0);
2111: }
2113: PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2114: {
2115: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2116: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2117: MatScalar *va,*vb;
2119: PetscInt ma,na,mb,nb, i;
2122: MatGetSize(A,&ma,&na);
2123: MatGetSize(B,&mb,&nb);
2124: if (ma!=nb || na!=mb) {
2125: *f = PETSC_FALSE;
2126: return(0);
2127: }
2128: aii = aij->i; bii = bij->i;
2129: adx = aij->j; bdx = bij->j;
2130: va = aij->a; vb = bij->a;
2131: PetscMalloc1(ma,&aptr);
2132: PetscMalloc1(mb,&bptr);
2133: for (i=0; i<ma; i++) aptr[i] = aii[i];
2134: for (i=0; i<mb; i++) bptr[i] = bii[i];
2136: *f = PETSC_TRUE;
2137: for (i=0; i<ma; i++) {
2138: while (aptr[i]<aii[i+1]) {
2139: PetscInt idc,idr;
2140: PetscScalar vc,vr;
2141: /* column/row index/value */
2142: idc = adx[aptr[i]];
2143: idr = bdx[bptr[idc]];
2144: vc = va[aptr[i]];
2145: vr = vb[bptr[idc]];
2146: if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) {
2147: *f = PETSC_FALSE;
2148: goto done;
2149: } else {
2150: aptr[i]++;
2151: if (B || i!=idc) bptr[idc]++;
2152: }
2153: }
2154: }
2155: done:
2156: PetscFree(aptr);
2157: PetscFree(bptr);
2158: return(0);
2159: }
2161: PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2162: {
2166: MatIsTranspose_SeqAIJ(A,A,tol,f);
2167: return(0);
2168: }
2170: PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2171: {
2175: MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);
2176: return(0);
2177: }
2179: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr)
2180: {
2181: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2182: const PetscScalar *l,*r;
2183: PetscScalar x;
2184: MatScalar *v;
2185: PetscErrorCode ierr;
2186: PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz;
2187: const PetscInt *jj;
2190: if (ll) {
2191: /* The local size is used so that VecMPI can be passed to this routine
2192: by MatDiagonalScale_MPIAIJ */
2193: VecGetLocalSize(ll,&m);
2194: if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length");
2195: VecGetArrayRead(ll,&l);
2196: v = a->a;
2197: for (i=0; i<m; i++) {
2198: x = l[i];
2199: M = a->i[i+1] - a->i[i];
2200: for (j=0; j<M; j++) (*v++) *= x;
2201: }
2202: VecRestoreArrayRead(ll,&l);
2203: PetscLogFlops(nz);
2204: }
2205: if (rr) {
2206: VecGetLocalSize(rr,&n);
2207: if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length");
2208: VecGetArrayRead(rr,&r);
2209: v = a->a; jj = a->j;
2210: for (i=0; i<nz; i++) (*v++) *= r[*jj++];
2211: VecRestoreArrayRead(rr,&r);
2212: PetscLogFlops(nz);
2213: }
2214: MatSeqAIJInvalidateDiagonal(A);
2215: return(0);
2216: }
2218: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B)
2219: {
2220: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c;
2222: PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens;
2223: PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi;
2224: const PetscInt *irow,*icol;
2225: PetscInt nrows,ncols;
2226: PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen;
2227: MatScalar *a_new,*mat_a;
2228: Mat C;
2229: PetscBool stride;
2233: ISGetIndices(isrow,&irow);
2234: ISGetLocalSize(isrow,&nrows);
2235: ISGetLocalSize(iscol,&ncols);
2237: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);
2238: if (stride) {
2239: ISStrideGetInfo(iscol,&first,&step);
2240: } else {
2241: first = 0;
2242: step = 0;
2243: }
2244: if (stride && step == 1) {
2245: /* special case of contiguous rows */
2246: PetscMalloc2(nrows,&lens,nrows,&starts);
2247: /* loop over new rows determining lens and starting points */
2248: for (i=0; i<nrows; i++) {
2249: kstart = ai[irow[i]];
2250: kend = kstart + ailen[irow[i]];
2251: starts[i] = kstart;
2252: for (k=kstart; k<kend; k++) {
2253: if (aj[k] >= first) {
2254: starts[i] = k;
2255: break;
2256: }
2257: }
2258: sum = 0;
2259: while (k < kend) {
2260: if (aj[k++] >= first+ncols) break;
2261: sum++;
2262: }
2263: lens[i] = sum;
2264: }
2265: /* create submatrix */
2266: if (scall == MAT_REUSE_MATRIX) {
2267: PetscInt n_cols,n_rows;
2268: MatGetSize(*B,&n_rows,&n_cols);
2269: if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
2270: MatZeroEntries(*B);
2271: C = *B;
2272: } else {
2273: PetscInt rbs,cbs;
2274: MatCreate(PetscObjectComm((PetscObject)A),&C);
2275: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2276: ISGetBlockSize(isrow,&rbs);
2277: ISGetBlockSize(iscol,&cbs);
2278: MatSetBlockSizes(C,rbs,cbs);
2279: MatSetType(C,((PetscObject)A)->type_name);
2280: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2281: }
2282: c = (Mat_SeqAIJ*)C->data;
2284: /* loop over rows inserting into submatrix */
2285: a_new = c->a;
2286: j_new = c->j;
2287: i_new = c->i;
2289: for (i=0; i<nrows; i++) {
2290: ii = starts[i];
2291: lensi = lens[i];
2292: for (k=0; k<lensi; k++) {
2293: *j_new++ = aj[ii+k] - first;
2294: }
2295: PetscMemcpy(a_new,a->a + starts[i],lensi*sizeof(PetscScalar));
2296: a_new += lensi;
2297: i_new[i+1] = i_new[i] + lensi;
2298: c->ilen[i] = lensi;
2299: }
2300: PetscFree2(lens,starts);
2301: } else {
2302: ISGetIndices(iscol,&icol);
2303: PetscCalloc1(oldcols,&smap);
2304: PetscMalloc1(1+nrows,&lens);
2305: for (i=0; i<ncols; i++) {
2306: #if defined(PETSC_USE_DEBUG)
2307: 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);
2308: #endif
2309: smap[icol[i]] = i+1;
2310: }
2312: /* determine lens of each row */
2313: for (i=0; i<nrows; i++) {
2314: kstart = ai[irow[i]];
2315: kend = kstart + a->ilen[irow[i]];
2316: lens[i] = 0;
2317: for (k=kstart; k<kend; k++) {
2318: if (smap[aj[k]]) {
2319: lens[i]++;
2320: }
2321: }
2322: }
2323: /* Create and fill new matrix */
2324: if (scall == MAT_REUSE_MATRIX) {
2325: PetscBool equal;
2327: c = (Mat_SeqAIJ*)((*B)->data);
2328: if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
2329: PetscMemcmp(c->ilen,lens,(*B)->rmap->n*sizeof(PetscInt),&equal);
2330: if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
2331: PetscMemzero(c->ilen,(*B)->rmap->n*sizeof(PetscInt));
2332: C = *B;
2333: } else {
2334: PetscInt rbs,cbs;
2335: MatCreate(PetscObjectComm((PetscObject)A),&C);
2336: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2337: ISGetBlockSize(isrow,&rbs);
2338: ISGetBlockSize(iscol,&cbs);
2339: MatSetBlockSizes(C,rbs,cbs);
2340: MatSetType(C,((PetscObject)A)->type_name);
2341: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2342: }
2343: c = (Mat_SeqAIJ*)(C->data);
2344: for (i=0; i<nrows; i++) {
2345: row = irow[i];
2346: kstart = ai[row];
2347: kend = kstart + a->ilen[row];
2348: mat_i = c->i[i];
2349: mat_j = c->j + mat_i;
2350: mat_a = c->a + mat_i;
2351: mat_ilen = c->ilen + i;
2352: for (k=kstart; k<kend; k++) {
2353: if ((tcol=smap[a->j[k]])) {
2354: *mat_j++ = tcol - 1;
2355: *mat_a++ = a->a[k];
2356: (*mat_ilen)++;
2358: }
2359: }
2360: }
2361: /* Free work space */
2362: ISRestoreIndices(iscol,&icol);
2363: PetscFree(smap);
2364: PetscFree(lens);
2365: /* sort */
2366: for (i = 0; i < nrows; i++) {
2367: PetscInt ilen;
2369: mat_i = c->i[i];
2370: mat_j = c->j + mat_i;
2371: mat_a = c->a + mat_i;
2372: ilen = c->ilen[i];
2373: PetscSortIntWithScalarArray(ilen,mat_j,mat_a);
2374: }
2375: }
2376: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2377: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2379: ISRestoreIndices(isrow,&irow);
2380: *B = C;
2381: return(0);
2382: }
2384: PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat)
2385: {
2387: Mat B;
2390: if (scall == MAT_INITIAL_MATRIX) {
2391: MatCreate(subComm,&B);
2392: MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);
2393: MatSetBlockSizesFromMats(B,mat,mat);
2394: MatSetType(B,MATSEQAIJ);
2395: MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);
2396: *subMat = B;
2397: } else {
2398: MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);
2399: }
2400: return(0);
2401: }
2403: PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info)
2404: {
2405: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2407: Mat outA;
2408: PetscBool row_identity,col_identity;
2411: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu");
2413: ISIdentity(row,&row_identity);
2414: ISIdentity(col,&col_identity);
2416: outA = inA;
2417: outA->factortype = MAT_FACTOR_LU;
2418: PetscFree(inA->solvertype);
2419: PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);
2421: PetscObjectReference((PetscObject)row);
2422: ISDestroy(&a->row);
2424: a->row = row;
2426: PetscObjectReference((PetscObject)col);
2427: ISDestroy(&a->col);
2429: a->col = col;
2431: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2432: ISDestroy(&a->icol);
2433: ISInvertPermutation(col,PETSC_DECIDE,&a->icol);
2434: PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);
2436: if (!a->solve_work) { /* this matrix may have been factored before */
2437: PetscMalloc1(inA->rmap->n+1,&a->solve_work);
2438: PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));
2439: }
2441: MatMarkDiagonal_SeqAIJ(inA);
2442: if (row_identity && col_identity) {
2443: MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);
2444: } else {
2445: MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);
2446: }
2447: return(0);
2448: }
2450: PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha)
2451: {
2452: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2453: PetscScalar oalpha = alpha;
2455: PetscBLASInt one = 1,bnz;
2458: PetscBLASIntCast(a->nz,&bnz);
2459: PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&oalpha,a->a,&one));
2460: PetscLogFlops(a->nz);
2461: MatSeqAIJInvalidateDiagonal(inA);
2462: return(0);
2463: }
2465: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2466: {
2468: PetscInt i;
2471: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2472: PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);
2474: for (i=0; i<submatj->nrqr; ++i) {
2475: PetscFree(submatj->sbuf2[i]);
2476: }
2477: PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);
2479: if (submatj->rbuf1) {
2480: PetscFree(submatj->rbuf1[0]);
2481: PetscFree(submatj->rbuf1);
2482: }
2484: for (i=0; i<submatj->nrqs; ++i) {
2485: PetscFree(submatj->rbuf3[i]);
2486: }
2487: PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);
2488: PetscFree(submatj->pa);
2489: }
2491: #if defined(PETSC_USE_CTABLE)
2492: PetscTableDestroy((PetscTable*)&submatj->rmap);
2493: if (submatj->cmap_loc) {PetscFree(submatj->cmap_loc);}
2494: PetscFree(submatj->rmap_loc);
2495: #else
2496: PetscFree(submatj->rmap);
2497: #endif
2499: if (!submatj->allcolumns) {
2500: #if defined(PETSC_USE_CTABLE)
2501: PetscTableDestroy((PetscTable*)&submatj->cmap);
2502: #else
2503: PetscFree(submatj->cmap);
2504: #endif
2505: }
2506: PetscFree(submatj->row2proc);
2508: PetscFree(submatj);
2509: return(0);
2510: }
2512: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2513: {
2515: Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data;
2516: Mat_SubSppt *submatj = c->submatis1;
2519: submatj->destroy(C);
2520: MatDestroySubMatrix_Private(submatj);
2521: return(0);
2522: }
2524: PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n,Mat *mat[])
2525: {
2527: PetscInt i;
2528: Mat C;
2529: Mat_SeqAIJ *c;
2530: Mat_SubSppt *submatj;
2533: for (i=0; i<n; i++) {
2534: C = (*mat)[i];
2535: c = (Mat_SeqAIJ*)C->data;
2536: submatj = c->submatis1;
2537: if (submatj) {
2538: if (--((PetscObject)C)->refct <= 0) {
2539: (submatj->destroy)(C);
2540: MatDestroySubMatrix_Private(submatj);
2541: PetscLayoutDestroy(&C->rmap);
2542: PetscLayoutDestroy(&C->cmap);
2543: PetscHeaderDestroy(&C);
2544: }
2545: } else {
2546: MatDestroy(&C);
2547: }
2548: }
2550: PetscFree(*mat);
2551: return(0);
2552: }
2554: PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
2555: {
2557: PetscInt i;
2560: if (scall == MAT_INITIAL_MATRIX) {
2561: PetscCalloc1(n+1,B);
2562: }
2564: for (i=0; i<n; i++) {
2565: MatCreateSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);
2566: }
2567: return(0);
2568: }
2570: PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov)
2571: {
2572: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2574: PetscInt row,i,j,k,l,m,n,*nidx,isz,val;
2575: const PetscInt *idx;
2576: PetscInt start,end,*ai,*aj;
2577: PetscBT table;
2580: m = A->rmap->n;
2581: ai = a->i;
2582: aj = a->j;
2584: if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used");
2586: PetscMalloc1(m+1,&nidx);
2587: PetscBTCreate(m,&table);
2589: for (i=0; i<is_max; i++) {
2590: /* Initialize the two local arrays */
2591: isz = 0;
2592: PetscBTMemzero(m,table);
2594: /* Extract the indices, assume there can be duplicate entries */
2595: ISGetIndices(is[i],&idx);
2596: ISGetLocalSize(is[i],&n);
2598: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2599: for (j=0; j<n; ++j) {
2600: if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j];
2601: }
2602: ISRestoreIndices(is[i],&idx);
2603: ISDestroy(&is[i]);
2605: k = 0;
2606: for (j=0; j<ov; j++) { /* for each overlap */
2607: n = isz;
2608: for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */
2609: row = nidx[k];
2610: start = ai[row];
2611: end = ai[row+1];
2612: for (l = start; l<end; l++) {
2613: val = aj[l];
2614: if (!PetscBTLookupSet(table,val)) nidx[isz++] = val;
2615: }
2616: }
2617: }
2618: ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));
2619: }
2620: PetscBTDestroy(&table);
2621: PetscFree(nidx);
2622: return(0);
2623: }
2625: /* -------------------------------------------------------------- */
2626: PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B)
2627: {
2628: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2630: PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n;
2631: const PetscInt *row,*col;
2632: PetscInt *cnew,j,*lens;
2633: IS icolp,irowp;
2634: PetscInt *cwork = NULL;
2635: PetscScalar *vwork = NULL;
2638: ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);
2639: ISGetIndices(irowp,&row);
2640: ISInvertPermutation(colp,PETSC_DECIDE,&icolp);
2641: ISGetIndices(icolp,&col);
2643: /* determine lengths of permuted rows */
2644: PetscMalloc1(m+1,&lens);
2645: for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i];
2646: MatCreate(PetscObjectComm((PetscObject)A),B);
2647: MatSetSizes(*B,m,n,m,n);
2648: MatSetBlockSizesFromMats(*B,A,A);
2649: MatSetType(*B,((PetscObject)A)->type_name);
2650: MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);
2651: PetscFree(lens);
2653: PetscMalloc1(n,&cnew);
2654: for (i=0; i<m; i++) {
2655: MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2656: for (j=0; j<nz; j++) cnew[j] = col[cwork[j]];
2657: MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);
2658: MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2659: }
2660: PetscFree(cnew);
2662: (*B)->assembled = PETSC_FALSE;
2664: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
2665: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
2666: ISRestoreIndices(irowp,&row);
2667: ISRestoreIndices(icolp,&col);
2668: ISDestroy(&irowp);
2669: ISDestroy(&icolp);
2670: return(0);
2671: }
2673: PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str)
2674: {
2678: /* If the two matrices have the same copy implementation, use fast copy. */
2679: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2680: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2681: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
2683: 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");
2684: PetscMemcpy(b->a,a->a,(a->i[A->rmap->n])*sizeof(PetscScalar));
2685: PetscObjectStateIncrease((PetscObject)B);
2686: } else {
2687: MatCopy_Basic(A,B,str);
2688: }
2689: return(0);
2690: }
2692: PetscErrorCode MatSetUp_SeqAIJ(Mat A)
2693: {
2697: MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,0);
2698: return(0);
2699: }
2701: PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[])
2702: {
2703: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2706: *array = a->a;
2707: return(0);
2708: }
2710: PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[])
2711: {
2713: return(0);
2714: }
2716: /*
2717: Computes the number of nonzeros per row needed for preallocation when X and Y
2718: have different nonzero structure.
2719: */
2720: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz)
2721: {
2722: PetscInt i,j,k,nzx,nzy;
2725: /* Set the number of nonzeros in the new matrix */
2726: for (i=0; i<m; i++) {
2727: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2728: nzx = xi[i+1] - xi[i];
2729: nzy = yi[i+1] - yi[i];
2730: nnz[i] = 0;
2731: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2732: for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */
2733: if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */
2734: nnz[i]++;
2735: }
2736: for (; k<nzy; k++) nnz[i]++;
2737: }
2738: return(0);
2739: }
2741: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz)
2742: {
2743: PetscInt m = Y->rmap->N;
2744: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2745: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2749: /* Set the number of nonzeros in the new matrix */
2750: MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);
2751: return(0);
2752: }
2754: PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2755: {
2757: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data;
2758: PetscBLASInt one=1,bnz;
2761: PetscBLASIntCast(x->nz,&bnz);
2762: if (str == SAME_NONZERO_PATTERN) {
2763: PetscScalar alpha = a;
2764: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2765: MatSeqAIJInvalidateDiagonal(Y);
2766: PetscObjectStateIncrease((PetscObject)Y);
2767: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2768: MatAXPY_Basic(Y,a,X,str);
2769: } else {
2770: Mat B;
2771: PetscInt *nnz;
2772: PetscMalloc1(Y->rmap->N,&nnz);
2773: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2774: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2775: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2776: MatSetBlockSizesFromMats(B,Y,Y);
2777: MatSetType(B,(MatType) ((PetscObject)Y)->type_name);
2778: MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);
2779: MatSeqAIJSetPreallocation(B,0,nnz);
2780: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2781: MatHeaderReplace(Y,&B);
2782: PetscFree(nnz);
2783: }
2784: return(0);
2785: }
2787: PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
2788: {
2789: #if defined(PETSC_USE_COMPLEX)
2790: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
2791: PetscInt i,nz;
2792: PetscScalar *a;
2795: nz = aij->nz;
2796: a = aij->a;
2797: for (i=0; i<nz; i++) a[i] = PetscConj(a[i]);
2798: #else
2800: #endif
2801: return(0);
2802: }
2804: PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2805: {
2806: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2808: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2809: PetscReal atmp;
2810: PetscScalar *x;
2811: MatScalar *aa;
2814: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2815: aa = a->a;
2816: ai = a->i;
2817: aj = a->j;
2819: VecSet(v,0.0);
2820: VecGetArray(v,&x);
2821: VecGetLocalSize(v,&n);
2822: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2823: for (i=0; i<m; i++) {
2824: ncols = ai[1] - ai[0]; ai++;
2825: x[i] = 0.0;
2826: for (j=0; j<ncols; j++) {
2827: atmp = PetscAbsScalar(*aa);
2828: if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2829: aa++; aj++;
2830: }
2831: }
2832: VecRestoreArray(v,&x);
2833: return(0);
2834: }
2836: PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2837: {
2838: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2840: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2841: PetscScalar *x;
2842: MatScalar *aa;
2845: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2846: aa = a->a;
2847: ai = a->i;
2848: aj = a->j;
2850: VecSet(v,0.0);
2851: VecGetArray(v,&x);
2852: VecGetLocalSize(v,&n);
2853: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2854: for (i=0; i<m; i++) {
2855: ncols = ai[1] - ai[0]; ai++;
2856: if (ncols == A->cmap->n) { /* row is dense */
2857: x[i] = *aa; if (idx) idx[i] = 0;
2858: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
2859: x[i] = 0.0;
2860: if (idx) {
2861: idx[i] = 0; /* in case ncols is zero */
2862: for (j=0;j<ncols;j++) { /* find first implicit 0.0 in the row */
2863: if (aj[j] > j) {
2864: idx[i] = j;
2865: break;
2866: }
2867: }
2868: }
2869: }
2870: for (j=0; j<ncols; j++) {
2871: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
2872: aa++; aj++;
2873: }
2874: }
2875: VecRestoreArray(v,&x);
2876: return(0);
2877: }
2879: PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2880: {
2881: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2883: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2884: PetscReal atmp;
2885: PetscScalar *x;
2886: MatScalar *aa;
2889: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2890: aa = a->a;
2891: ai = a->i;
2892: aj = a->j;
2894: VecSet(v,0.0);
2895: VecGetArray(v,&x);
2896: VecGetLocalSize(v,&n);
2897: 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);
2898: for (i=0; i<m; i++) {
2899: ncols = ai[1] - ai[0]; ai++;
2900: if (ncols) {
2901: /* Get first nonzero */
2902: for (j = 0; j < ncols; j++) {
2903: atmp = PetscAbsScalar(aa[j]);
2904: if (atmp > 1.0e-12) {
2905: x[i] = atmp;
2906: if (idx) idx[i] = aj[j];
2907: break;
2908: }
2909: }
2910: if (j == ncols) {x[i] = PetscAbsScalar(*aa); if (idx) idx[i] = *aj;}
2911: } else {
2912: x[i] = 0.0; if (idx) idx[i] = 0;
2913: }
2914: for (j = 0; j < ncols; j++) {
2915: atmp = PetscAbsScalar(*aa);
2916: if (atmp > 1.0e-12 && PetscAbsScalar(x[i]) > atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2917: aa++; aj++;
2918: }
2919: }
2920: VecRestoreArray(v,&x);
2921: return(0);
2922: }
2924: PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2925: {
2926: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2927: PetscErrorCode ierr;
2928: PetscInt i,j,m = A->rmap->n,ncols,n;
2929: const PetscInt *ai,*aj;
2930: PetscScalar *x;
2931: const MatScalar *aa;
2934: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2935: aa = a->a;
2936: ai = a->i;
2937: aj = a->j;
2939: VecSet(v,0.0);
2940: VecGetArray(v,&x);
2941: VecGetLocalSize(v,&n);
2942: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2943: for (i=0; i<m; i++) {
2944: ncols = ai[1] - ai[0]; ai++;
2945: if (ncols == A->cmap->n) { /* row is dense */
2946: x[i] = *aa; if (idx) idx[i] = 0;
2947: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
2948: x[i] = 0.0;
2949: if (idx) { /* find first implicit 0.0 in the row */
2950: idx[i] = 0; /* in case ncols is zero */
2951: for (j=0; j<ncols; j++) {
2952: if (aj[j] > j) {
2953: idx[i] = j;
2954: break;
2955: }
2956: }
2957: }
2958: }
2959: for (j=0; j<ncols; j++) {
2960: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
2961: aa++; aj++;
2962: }
2963: }
2964: VecRestoreArray(v,&x);
2965: return(0);
2966: }
2968: #include <petscblaslapack.h>
2969: #include <petsc/private/kernels/blockinvert.h>
2971: PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values)
2972: {
2973: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
2975: PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j;
2976: MatScalar *diag,work[25],*v_work;
2977: PetscReal shift = 0.0;
2978: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
2981: allowzeropivot = PetscNot(A->erroriffailure);
2982: if (a->ibdiagvalid) {
2983: if (values) *values = a->ibdiag;
2984: return(0);
2985: }
2986: MatMarkDiagonal_SeqAIJ(A);
2987: if (!a->ibdiag) {
2988: PetscMalloc1(bs2*mbs,&a->ibdiag);
2989: PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));
2990: }
2991: diag = a->ibdiag;
2992: if (values) *values = a->ibdiag;
2993: /* factor and invert each block */
2994: switch (bs) {
2995: case 1:
2996: for (i=0; i<mbs; i++) {
2997: MatGetValues(A,1,&i,1,&i,diag+i);
2998: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
2999: if (allowzeropivot) {
3000: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3001: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3002: A->factorerror_zeropivot_row = i;
3003: PetscInfo3(A,"Zero pivot, row %D pivot %g tolerance %g\n",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);
3004: } 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);
3005: }
3006: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3007: }
3008: break;
3009: case 2:
3010: for (i=0; i<mbs; i++) {
3011: ij[0] = 2*i; ij[1] = 2*i + 1;
3012: MatGetValues(A,2,ij,2,ij,diag);
3013: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
3014: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3015: PetscKernel_A_gets_transpose_A_2(diag);
3016: diag += 4;
3017: }
3018: break;
3019: case 3:
3020: for (i=0; i<mbs; i++) {
3021: ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2;
3022: MatGetValues(A,3,ij,3,ij,diag);
3023: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
3024: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3025: PetscKernel_A_gets_transpose_A_3(diag);
3026: diag += 9;
3027: }
3028: break;
3029: case 4:
3030: for (i=0; i<mbs; i++) {
3031: ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3;
3032: MatGetValues(A,4,ij,4,ij,diag);
3033: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
3034: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3035: PetscKernel_A_gets_transpose_A_4(diag);
3036: diag += 16;
3037: }
3038: break;
3039: case 5:
3040: for (i=0; i<mbs; i++) {
3041: ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4;
3042: MatGetValues(A,5,ij,5,ij,diag);
3043: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
3044: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3045: PetscKernel_A_gets_transpose_A_5(diag);
3046: diag += 25;
3047: }
3048: break;
3049: case 6:
3050: for (i=0; i<mbs; i++) {
3051: 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;
3052: MatGetValues(A,6,ij,6,ij,diag);
3053: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
3054: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3055: PetscKernel_A_gets_transpose_A_6(diag);
3056: diag += 36;
3057: }
3058: break;
3059: case 7:
3060: for (i=0; i<mbs; i++) {
3061: 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;
3062: MatGetValues(A,7,ij,7,ij,diag);
3063: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
3064: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3065: PetscKernel_A_gets_transpose_A_7(diag);
3066: diag += 49;
3067: }
3068: break;
3069: default:
3070: PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);
3071: for (i=0; i<mbs; i++) {
3072: for (j=0; j<bs; j++) {
3073: IJ[j] = bs*i + j;
3074: }
3075: MatGetValues(A,bs,IJ,bs,IJ,diag);
3076: PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
3077: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3078: PetscKernel_A_gets_transpose_A_N(diag,bs);
3079: diag += bs2;
3080: }
3081: PetscFree3(v_work,v_pivots,IJ);
3082: }
3083: a->ibdiagvalid = PETSC_TRUE;
3084: return(0);
3085: }
3087: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx)
3088: {
3090: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3091: PetscScalar a;
3092: PetscInt m,n,i,j,col;
3095: if (!x->assembled) {
3096: MatGetSize(x,&m,&n);
3097: for (i=0; i<m; i++) {
3098: for (j=0; j<aij->imax[i]; j++) {
3099: PetscRandomGetValue(rctx,&a);
3100: col = (PetscInt)(n*PetscRealPart(a));
3101: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3102: }
3103: }
3104: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet coded");
3105: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3106: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3107: return(0);
3108: }
3110: PetscErrorCode MatShift_SeqAIJ(Mat Y,PetscScalar a)
3111: {
3113: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)Y->data;
3116: if (!Y->preallocated || !aij->nz) {
3117: MatSeqAIJSetPreallocation(Y,1,NULL);
3118: }
3119: MatShift_Basic(Y,a);
3120: return(0);
3121: }
3123: /* -------------------------------------------------------------------*/
3124: static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ,
3125: MatGetRow_SeqAIJ,
3126: MatRestoreRow_SeqAIJ,
3127: MatMult_SeqAIJ,
3128: /* 4*/ MatMultAdd_SeqAIJ,
3129: MatMultTranspose_SeqAIJ,
3130: MatMultTransposeAdd_SeqAIJ,
3131: 0,
3132: 0,
3133: 0,
3134: /* 10*/ 0,
3135: MatLUFactor_SeqAIJ,
3136: 0,
3137: MatSOR_SeqAIJ,
3138: MatTranspose_SeqAIJ,
3139: /*1 5*/ MatGetInfo_SeqAIJ,
3140: MatEqual_SeqAIJ,
3141: MatGetDiagonal_SeqAIJ,
3142: MatDiagonalScale_SeqAIJ,
3143: MatNorm_SeqAIJ,
3144: /* 20*/ 0,
3145: MatAssemblyEnd_SeqAIJ,
3146: MatSetOption_SeqAIJ,
3147: MatZeroEntries_SeqAIJ,
3148: /* 24*/ MatZeroRows_SeqAIJ,
3149: 0,
3150: 0,
3151: 0,
3152: 0,
3153: /* 29*/ MatSetUp_SeqAIJ,
3154: 0,
3155: 0,
3156: 0,
3157: 0,
3158: /* 34*/ MatDuplicate_SeqAIJ,
3159: 0,
3160: 0,
3161: MatILUFactor_SeqAIJ,
3162: 0,
3163: /* 39*/ MatAXPY_SeqAIJ,
3164: MatCreateSubMatrices_SeqAIJ,
3165: MatIncreaseOverlap_SeqAIJ,
3166: MatGetValues_SeqAIJ,
3167: MatCopy_SeqAIJ,
3168: /* 44*/ MatGetRowMax_SeqAIJ,
3169: MatScale_SeqAIJ,
3170: MatShift_SeqAIJ,
3171: MatDiagonalSet_SeqAIJ,
3172: MatZeroRowsColumns_SeqAIJ,
3173: /* 49*/ MatSetRandom_SeqAIJ,
3174: MatGetRowIJ_SeqAIJ,
3175: MatRestoreRowIJ_SeqAIJ,
3176: MatGetColumnIJ_SeqAIJ,
3177: MatRestoreColumnIJ_SeqAIJ,
3178: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3179: 0,
3180: 0,
3181: MatPermute_SeqAIJ,
3182: 0,
3183: /* 59*/ 0,
3184: MatDestroy_SeqAIJ,
3185: MatView_SeqAIJ,
3186: 0,
3187: MatMatMatMult_SeqAIJ_SeqAIJ_SeqAIJ,
3188: /* 64*/ MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ,
3189: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3190: 0,
3191: 0,
3192: 0,
3193: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3194: MatGetRowMinAbs_SeqAIJ,
3195: 0,
3196: 0,
3197: 0,
3198: /* 74*/ 0,
3199: MatFDColoringApply_AIJ,
3200: 0,
3201: 0,
3202: 0,
3203: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3204: 0,
3205: 0,
3206: 0,
3207: MatLoad_SeqAIJ,
3208: /* 84*/ MatIsSymmetric_SeqAIJ,
3209: MatIsHermitian_SeqAIJ,
3210: 0,
3211: 0,
3212: 0,
3213: /* 89*/ MatMatMult_SeqAIJ_SeqAIJ,
3214: MatMatMultSymbolic_SeqAIJ_SeqAIJ,
3215: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3216: MatPtAP_SeqAIJ_SeqAIJ,
3217: MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy,
3218: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3219: MatMatTransposeMult_SeqAIJ_SeqAIJ,
3220: MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ,
3221: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3222: 0,
3223: /* 99*/ 0,
3224: 0,
3225: 0,
3226: MatConjugate_SeqAIJ,
3227: 0,
3228: /*104*/ MatSetValuesRow_SeqAIJ,
3229: MatRealPart_SeqAIJ,
3230: MatImaginaryPart_SeqAIJ,
3231: 0,
3232: 0,
3233: /*109*/ MatMatSolve_SeqAIJ,
3234: 0,
3235: MatGetRowMin_SeqAIJ,
3236: 0,
3237: MatMissingDiagonal_SeqAIJ,
3238: /*114*/ 0,
3239: 0,
3240: 0,
3241: 0,
3242: 0,
3243: /*119*/ 0,
3244: 0,
3245: 0,
3246: 0,
3247: MatGetMultiProcBlock_SeqAIJ,
3248: /*124*/ MatFindNonzeroRows_SeqAIJ,
3249: MatGetColumnNorms_SeqAIJ,
3250: MatInvertBlockDiagonal_SeqAIJ,
3251: 0,
3252: 0,
3253: /*129*/ 0,
3254: MatTransposeMatMult_SeqAIJ_SeqAIJ,
3255: MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ,
3256: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3257: MatTransposeColoringCreate_SeqAIJ,
3258: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3259: MatTransColoringApplyDenToSp_SeqAIJ,
3260: MatRARt_SeqAIJ_SeqAIJ,
3261: MatRARtSymbolic_SeqAIJ_SeqAIJ,
3262: MatRARtNumeric_SeqAIJ_SeqAIJ,
3263: /*139*/0,
3264: 0,
3265: 0,
3266: MatFDColoringSetUp_SeqXAIJ,
3267: MatFindOffBlockDiagonalEntries_SeqAIJ,
3268: /*144*/MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3269: MatDestroySubMatrices_SeqAIJ
3270: };
3272: PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices)
3273: {
3274: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3275: PetscInt i,nz,n;
3278: nz = aij->maxnz;
3279: n = mat->rmap->n;
3280: for (i=0; i<nz; i++) {
3281: aij->j[i] = indices[i];
3282: }
3283: aij->nz = nz;
3284: for (i=0; i<n; i++) {
3285: aij->ilen[i] = aij->imax[i];
3286: }
3287: return(0);
3288: }
3290: /*@
3291: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3292: in the matrix.
3294: Input Parameters:
3295: + mat - the SeqAIJ matrix
3296: - indices - the column indices
3298: Level: advanced
3300: Notes:
3301: This can be called if you have precomputed the nonzero structure of the
3302: matrix and want to provide it to the matrix object to improve the performance
3303: of the MatSetValues() operation.
3305: You MUST have set the correct numbers of nonzeros per row in the call to
3306: MatCreateSeqAIJ(), and the columns indices MUST be sorted.
3308: MUST be called before any calls to MatSetValues();
3310: The indices should start with zero, not one.
3312: @*/
3313: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices)
3314: {
3320: PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));
3321: return(0);
3322: }
3324: /* ----------------------------------------------------------------------------------------*/
3326: PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3327: {
3328: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3330: size_t nz = aij->i[mat->rmap->n];
3333: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3335: /* allocate space for values if not already there */
3336: if (!aij->saved_values) {
3337: PetscMalloc1(nz+1,&aij->saved_values);
3338: PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));
3339: }
3341: /* copy values over */
3342: PetscMemcpy(aij->saved_values,aij->a,nz*sizeof(PetscScalar));
3343: return(0);
3344: }
3346: /*@
3347: MatStoreValues - Stashes a copy of the matrix values; this allows, for
3348: example, reuse of the linear part of a Jacobian, while recomputing the
3349: nonlinear portion.
3351: Collect on Mat
3353: Input Parameters:
3354: . mat - the matrix (currently only AIJ matrices support this option)
3356: Level: advanced
3358: Common Usage, with SNESSolve():
3359: $ Create Jacobian matrix
3360: $ Set linear terms into matrix
3361: $ Apply boundary conditions to matrix, at this time matrix must have
3362: $ final nonzero structure (i.e. setting the nonlinear terms and applying
3363: $ boundary conditions again will not change the nonzero structure
3364: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3365: $ MatStoreValues(mat);
3366: $ Call SNESSetJacobian() with matrix
3367: $ In your Jacobian routine
3368: $ MatRetrieveValues(mat);
3369: $ Set nonlinear terms in matrix
3371: Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself:
3372: $ // build linear portion of Jacobian
3373: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3374: $ MatStoreValues(mat);
3375: $ loop over nonlinear iterations
3376: $ MatRetrieveValues(mat);
3377: $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3378: $ // call MatAssemblyBegin/End() on matrix
3379: $ Solve linear system with Jacobian
3380: $ endloop
3382: Notes:
3383: Matrix must already be assemblied before calling this routine
3384: Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before
3385: calling this routine.
3387: When this is called multiple times it overwrites the previous set of stored values
3388: and does not allocated additional space.
3390: .seealso: MatRetrieveValues()
3392: @*/
3393: PetscErrorCode MatStoreValues(Mat mat)
3394: {
3399: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3400: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3401: PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));
3402: return(0);
3403: }
3405: PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3406: {
3407: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3409: PetscInt nz = aij->i[mat->rmap->n];
3412: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3413: if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first");
3414: /* copy values over */
3415: PetscMemcpy(aij->a,aij->saved_values,nz*sizeof(PetscScalar));
3416: return(0);
3417: }
3419: /*@
3420: MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for
3421: example, reuse of the linear part of a Jacobian, while recomputing the
3422: nonlinear portion.
3424: Collect on Mat
3426: Input Parameters:
3427: . mat - the matrix (currently only AIJ matrices support this option)
3429: Level: advanced
3431: .seealso: MatStoreValues()
3433: @*/
3434: PetscErrorCode MatRetrieveValues(Mat mat)
3435: {
3440: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3441: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3442: PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));
3443: return(0);
3444: }
3447: /* --------------------------------------------------------------------------------*/
3448: /*@C
3449: MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format
3450: (the default parallel PETSc format). For good matrix assembly performance
3451: the user should preallocate the matrix storage by setting the parameter nz
3452: (or the array nnz). By setting these parameters accurately, performance
3453: during matrix assembly can be increased by more than a factor of 50.
3455: Collective on MPI_Comm
3457: Input Parameters:
3458: + comm - MPI communicator, set to PETSC_COMM_SELF
3459: . m - number of rows
3460: . n - number of columns
3461: . nz - number of nonzeros per row (same for all rows)
3462: - nnz - array containing the number of nonzeros in the various rows
3463: (possibly different for each row) or NULL
3465: Output Parameter:
3466: . A - the matrix
3468: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3469: MatXXXXSetPreallocation() paradgm instead of this routine directly.
3470: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3472: Notes:
3473: If nnz is given then nz is ignored
3475: The AIJ format (also called the Yale sparse matrix format or
3476: compressed row storage), is fully compatible with standard Fortran 77
3477: storage. That is, the stored row and column indices can begin at
3478: either one (as in Fortran) or zero. See the users' manual for details.
3480: Specify the preallocated storage with either nz or nnz (not both).
3481: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3482: allocation. For large problems you MUST preallocate memory or you
3483: will get TERRIBLE performance, see the users' manual chapter on matrices.
3485: By default, this format uses inodes (identical nodes) when possible, to
3486: improve numerical efficiency of matrix-vector products and solves. We
3487: search for consecutive rows with the same nonzero structure, thereby
3488: reusing matrix information to achieve increased efficiency.
3490: Options Database Keys:
3491: + -mat_no_inode - Do not use inodes
3492: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3494: Level: intermediate
3496: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays()
3498: @*/
3499: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A)
3500: {
3504: MatCreate(comm,A);
3505: MatSetSizes(*A,m,n,m,n);
3506: MatSetType(*A,MATSEQAIJ);
3507: MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);
3508: return(0);
3509: }
3511: /*@C
3512: MatSeqAIJSetPreallocation - For good matrix assembly performance
3513: the user should preallocate the matrix storage by setting the parameter nz
3514: (or the array nnz). By setting these parameters accurately, performance
3515: during matrix assembly can be increased by more than a factor of 50.
3517: Collective on MPI_Comm
3519: Input Parameters:
3520: + B - The matrix
3521: . nz - number of nonzeros per row (same for all rows)
3522: - nnz - array containing the number of nonzeros in the various rows
3523: (possibly different for each row) or NULL
3525: Notes:
3526: If nnz is given then nz is ignored
3528: The AIJ format (also called the Yale sparse matrix format or
3529: compressed row storage), is fully compatible with standard Fortran 77
3530: storage. That is, the stored row and column indices can begin at
3531: either one (as in Fortran) or zero. See the users' manual for details.
3533: Specify the preallocated storage with either nz or nnz (not both).
3534: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3535: allocation. For large problems you MUST preallocate memory or you
3536: will get TERRIBLE performance, see the users' manual chapter on matrices.
3538: You can call MatGetInfo() to get information on how effective the preallocation was;
3539: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3540: You can also run with the option -info and look for messages with the string
3541: malloc in them to see if additional memory allocation was needed.
3543: Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix
3544: entries or columns indices
3546: By default, this format uses inodes (identical nodes) when possible, to
3547: improve numerical efficiency of matrix-vector products and solves. We
3548: search for consecutive rows with the same nonzero structure, thereby
3549: reusing matrix information to achieve increased efficiency.
3551: Options Database Keys:
3552: + -mat_no_inode - Do not use inodes
3553: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3555: Level: intermediate
3557: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo()
3559: @*/
3560: PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[])
3561: {
3567: PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));
3568: return(0);
3569: }
3571: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz)
3572: {
3573: Mat_SeqAIJ *b;
3574: PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE;
3576: PetscInt i;
3579: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3580: if (nz == MAT_SKIP_ALLOCATION) {
3581: skipallocation = PETSC_TRUE;
3582: nz = 0;
3583: }
3584: PetscLayoutSetUp(B->rmap);
3585: PetscLayoutSetUp(B->cmap);
3587: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3588: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz);
3589: if (nnz) {
3590: for (i=0; i<B->rmap->n; i++) {
3591: 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]);
3592: 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);
3593: }
3594: }
3596: B->preallocated = PETSC_TRUE;
3598: b = (Mat_SeqAIJ*)B->data;
3600: if (!skipallocation) {
3601: if (!b->imax) {
3602: PetscMalloc2(B->rmap->n,&b->imax,B->rmap->n,&b->ilen);
3603: PetscLogObjectMemory((PetscObject)B,2*B->rmap->n*sizeof(PetscInt));
3604: }
3605: if (!b->ipre) {
3606: PetscMalloc1(B->rmap->n,&b->ipre);
3607: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
3608: }
3609: if (!nnz) {
3610: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3611: else if (nz < 0) nz = 1;
3612: for (i=0; i<B->rmap->n; i++) b->imax[i] = nz;
3613: nz = nz*B->rmap->n;
3614: } else {
3615: nz = 0;
3616: for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz += nnz[i];}
3617: }
3618: /* b->ilen will count nonzeros in each row so far. */
3619: for (i=0; i<B->rmap->n; i++) b->ilen[i] = 0;
3621: /* allocate the matrix space */
3622: /* FIXME: should B's old memory be unlogged? */
3623: MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);
3624: if (B->structure_only) {
3625: PetscMalloc1(nz,&b->j);
3626: PetscMalloc1(B->rmap->n+1,&b->i);
3627: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));
3628: } else {
3629: PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);
3630: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));
3631: }
3632: b->i[0] = 0;
3633: for (i=1; i<B->rmap->n+1; i++) {
3634: b->i[i] = b->i[i-1] + b->imax[i-1];
3635: }
3636: if (B->structure_only) {
3637: b->singlemalloc = PETSC_FALSE;
3638: b->free_a = PETSC_FALSE;
3639: } else {
3640: b->singlemalloc = PETSC_TRUE;
3641: b->free_a = PETSC_TRUE;
3642: }
3643: b->free_ij = PETSC_TRUE;
3644: } else {
3645: b->free_a = PETSC_FALSE;
3646: b->free_ij = PETSC_FALSE;
3647: }
3649: if (b->ipre && nnz != b->ipre && b->imax) {
3650: /* reserve user-requested sparsity */
3651: PetscMemcpy(b->ipre,b->imax,B->rmap->n*sizeof(PetscInt));
3652: }
3655: b->nz = 0;
3656: b->maxnz = nz;
3657: B->info.nz_unneeded = (double)b->maxnz;
3658: if (realalloc) {
3659: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
3660: }
3661: B->was_assembled = PETSC_FALSE;
3662: B->assembled = PETSC_FALSE;
3663: return(0);
3664: }
3667: PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
3668: {
3669: Mat_SeqAIJ *a;
3670: PetscInt i;
3675: a = (Mat_SeqAIJ*)A->data;
3676: /* if no saved info, we error out */
3677: if (!a->ipre) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_ARG_NULL,"No saved preallocation info \n");
3679: if (!a->i || !a->j || !a->a || !a->imax || !a->ilen) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_ARG_NULL,"Memory info is incomplete, and can not reset preallocation \n");
3681: PetscMemcpy(a->imax,a->ipre,A->rmap->n*sizeof(PetscInt));
3682: PetscMemzero(a->ilen,A->rmap->n*sizeof(PetscInt));
3683: a->i[0] = 0;
3684: for (i=1; i<A->rmap->n+1; i++) {
3685: a->i[i] = a->i[i-1] + a->imax[i-1];
3686: }
3687: A->preallocated = PETSC_TRUE;
3688: a->nz = 0;
3689: a->maxnz = a->i[A->rmap->n];
3690: A->info.nz_unneeded = (double)a->maxnz;
3691: A->was_assembled = PETSC_FALSE;
3692: A->assembled = PETSC_FALSE;
3693: return(0);
3694: }
3696: /*@
3697: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format.
3699: Input Parameters:
3700: + B - the matrix
3701: . i - the indices into j for the start of each row (starts with zero)
3702: . j - the column indices for each row (starts with zero) these must be sorted for each row
3703: - v - optional values in the matrix
3705: Level: developer
3707: The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays()
3709: .keywords: matrix, aij, compressed row, sparse, sequential
3711: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), SeqAIJ
3712: @*/
3713: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[])
3714: {
3720: PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3721: return(0);
3722: }
3724: PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3725: {
3726: PetscInt i;
3727: PetscInt m,n;
3728: PetscInt nz;
3729: PetscInt *nnz, nz_max = 0;
3730: PetscScalar *values;
3734: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]);
3736: PetscLayoutSetUp(B->rmap);
3737: PetscLayoutSetUp(B->cmap);
3739: MatGetSize(B, &m, &n);
3740: PetscMalloc1(m+1, &nnz);
3741: for (i = 0; i < m; i++) {
3742: nz = Ii[i+1]- Ii[i];
3743: nz_max = PetscMax(nz_max, nz);
3744: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz);
3745: nnz[i] = nz;
3746: }
3747: MatSeqAIJSetPreallocation(B, 0, nnz);
3748: PetscFree(nnz);
3750: if (v) {
3751: values = (PetscScalar*) v;
3752: } else {
3753: PetscCalloc1(nz_max, &values);
3754: }
3756: for (i = 0; i < m; i++) {
3757: nz = Ii[i+1] - Ii[i];
3758: MatSetValues_SeqAIJ(B, 1, &i, nz, J+Ii[i], values + (v ? Ii[i] : 0), INSERT_VALUES);
3759: }
3761: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3762: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3764: if (!v) {
3765: PetscFree(values);
3766: }
3767: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3768: return(0);
3769: }
3771: #include <../src/mat/impls/dense/seq/dense.h>
3772: #include <petsc/private/kernels/petscaxpy.h>
3774: /*
3775: Computes (B'*A')' since computing B*A directly is untenable
3777: n p p
3778: ( ) ( ) ( )
3779: m ( A ) * n ( B ) = m ( C )
3780: ( ) ( ) ( )
3782: */
3783: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C)
3784: {
3785: PetscErrorCode ierr;
3786: Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data;
3787: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data;
3788: Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data;
3789: PetscInt i,n,m,q,p;
3790: const PetscInt *ii,*idx;
3791: const PetscScalar *b,*a,*a_q;
3792: PetscScalar *c,*c_q;
3795: m = A->rmap->n;
3796: n = A->cmap->n;
3797: p = B->cmap->n;
3798: a = sub_a->v;
3799: b = sub_b->a;
3800: c = sub_c->v;
3801: PetscMemzero(c,m*p*sizeof(PetscScalar));
3803: ii = sub_b->i;
3804: idx = sub_b->j;
3805: for (i=0; i<n; i++) {
3806: q = ii[i+1] - ii[i];
3807: while (q-->0) {
3808: c_q = c + m*(*idx);
3809: a_q = a + m*i;
3810: PetscKernelAXPY(c_q,*b,a_q,m);
3811: idx++;
3812: b++;
3813: }
3814: }
3815: return(0);
3816: }
3818: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
3819: {
3821: PetscInt m=A->rmap->n,n=B->cmap->n;
3822: Mat Cmat;
3825: 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);
3826: MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
3827: MatSetSizes(Cmat,m,n,m,n);
3828: MatSetBlockSizesFromMats(Cmat,A,B);
3829: MatSetType(Cmat,MATSEQDENSE);
3830: MatSeqDenseSetPreallocation(Cmat,NULL);
3832: Cmat->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
3834: *C = Cmat;
3835: return(0);
3836: }
3838: /* ----------------------------------------------------------------*/
3839: PETSC_INTERN PetscErrorCode MatMatMult_SeqDense_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3840: {
3844: if (scall == MAT_INITIAL_MATRIX) {
3845: PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
3846: MatMatMultSymbolic_SeqDense_SeqAIJ(A,B,fill,C);
3847: PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
3848: }
3849: PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
3850: MatMatMultNumeric_SeqDense_SeqAIJ(A,B,*C);
3851: PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
3852: return(0);
3853: }
3856: /*MC
3857: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
3858: based on compressed sparse row format.
3860: Options Database Keys:
3861: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
3863: Level: beginner
3865: .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType
3866: M*/
3868: /*MC
3869: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
3871: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
3872: and MATMPIAIJ otherwise. As a result, for single process communicators,
3873: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
3874: for communicators controlling multiple processes. It is recommended that you call both of
3875: the above preallocation routines for simplicity.
3877: Options Database Keys:
3878: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
3880: Developer Notes: Subclasses include MATAIJCUSPARSE, MATAIJPERM, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
3881: enough exist.
3883: Level: beginner
3885: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
3886: M*/
3888: /*MC
3889: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
3891: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
3892: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
3893: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
3894: for communicators controlling multiple processes. It is recommended that you call both of
3895: the above preallocation routines for simplicity.
3897: Options Database Keys:
3898: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
3900: Level: beginner
3902: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
3903: M*/
3905: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*);
3906: #if defined(PETSC_HAVE_ELEMENTAL)
3907: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
3908: #endif
3909: #if defined(PETSC_HAVE_HYPRE)
3910: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A,MatType,MatReuse,Mat*);
3911: PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*);
3912: #endif
3913: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*);
3915: #if defined(PETSC_HAVE_MATLAB_ENGINE)
3916: PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject,void*);
3917: PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject,void*);
3918: #endif
3920: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat,MatType,MatReuse,Mat*);
3922: /*@C
3923: MatSeqAIJGetArray - gives access to the array where the data for a MATSEQAIJ matrix is stored
3925: Not Collective
3927: Input Parameter:
3928: . mat - a MATSEQAIJ matrix
3930: Output Parameter:
3931: . array - pointer to the data
3933: Level: intermediate
3935: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
3936: @*/
3937: PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array)
3938: {
3942: PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));
3943: return(0);
3944: }
3946: /*@C
3947: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
3949: Not Collective
3951: Input Parameter:
3952: . mat - a MATSEQAIJ matrix
3954: Output Parameter:
3955: . nz - the maximum number of nonzeros in any row
3957: Level: intermediate
3959: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
3960: @*/
3961: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz)
3962: {
3963: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
3966: *nz = aij->rmax;
3967: return(0);
3968: }
3970: /*@C
3971: MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray()
3973: Not Collective
3975: Input Parameters:
3976: . mat - a MATSEQAIJ matrix
3977: . array - pointer to the data
3979: Level: intermediate
3981: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90()
3982: @*/
3983: PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array)
3984: {
3988: PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));
3989: return(0);
3990: }
3992: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
3993: {
3994: Mat_SeqAIJ *b;
3996: PetscMPIInt size;
3999: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
4000: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1");
4002: PetscNewLog(B,&b);
4004: B->data = (void*)b;
4006: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
4008: b->row = 0;
4009: b->col = 0;
4010: b->icol = 0;
4011: b->reallocs = 0;
4012: b->ignorezeroentries = PETSC_FALSE;
4013: b->roworiented = PETSC_TRUE;
4014: b->nonew = 0;
4015: b->diag = 0;
4016: b->solve_work = 0;
4017: B->spptr = 0;
4018: b->saved_values = 0;
4019: b->idiag = 0;
4020: b->mdiag = 0;
4021: b->ssor_work = 0;
4022: b->omega = 1.0;
4023: b->fshift = 0.0;
4024: b->idiagvalid = PETSC_FALSE;
4025: b->ibdiagvalid = PETSC_FALSE;
4026: b->keepnonzeropattern = PETSC_FALSE;
4028: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4029: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);
4030: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);
4032: #if defined(PETSC_HAVE_MATLAB_ENGINE)
4033: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);
4034: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);
4035: #endif
4037: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);
4038: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);
4039: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);
4040: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);
4041: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);
4042: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);
4043: #if defined(PETSC_HAVE_MKL_SPARSE)
4044: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijmkl_C",MatConvert_SeqAIJ_SeqAIJMKL);
4045: #endif
4046: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);
4047: #if defined(PETSC_HAVE_ELEMENTAL)
4048: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);
4049: #endif
4050: #if defined(PETSC_HAVE_HYPRE)
4051: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_hypre_C",MatConvert_AIJ_HYPRE);
4052: PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_seqaij_seqaij_C",MatMatMatMult_Transpose_AIJ_AIJ);
4053: #endif
4054: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);
4055: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsell_C",MatConvert_SeqAIJ_SeqSELL);
4056: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);
4057: PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);
4058: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);
4059: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_SeqAIJ);
4060: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);
4061: PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);
4062: PetscObjectComposeFunction((PetscObject)B,"MatMatMult_seqdense_seqaij_C",MatMatMult_SeqDense_SeqAIJ);
4063: PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_seqdense_seqaij_C",MatMatMultSymbolic_SeqDense_SeqAIJ);
4064: PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_seqdense_seqaij_C",MatMatMultNumeric_SeqDense_SeqAIJ);
4065: MatCreate_SeqAIJ_Inode(B);
4066: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4067: MatSeqAIJSetTypeFromOptions(B); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4068: return(0);
4069: }
4071: /*
4072: Given a matrix generated with MatGetFactor() duplicates all the information in A into B
4073: */
4074: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace)
4075: {
4076: Mat_SeqAIJ *c,*a = (Mat_SeqAIJ*)A->data;
4078: PetscInt i,m = A->rmap->n;
4081: c = (Mat_SeqAIJ*)C->data;
4083: C->factortype = A->factortype;
4084: c->row = 0;
4085: c->col = 0;
4086: c->icol = 0;
4087: c->reallocs = 0;
4089: C->assembled = PETSC_TRUE;
4091: PetscLayoutReference(A->rmap,&C->rmap);
4092: PetscLayoutReference(A->cmap,&C->cmap);
4094: PetscMalloc2(m,&c->imax,m,&c->ilen);
4095: PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));
4096: for (i=0; i<m; i++) {
4097: c->imax[i] = a->imax[i];
4098: c->ilen[i] = a->ilen[i];
4099: }
4101: /* allocate the matrix space */
4102: if (mallocmatspace) {
4103: PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);
4104: PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));
4106: c->singlemalloc = PETSC_TRUE;
4108: PetscMemcpy(c->i,a->i,(m+1)*sizeof(PetscInt));
4109: if (m > 0) {
4110: PetscMemcpy(c->j,a->j,(a->i[m])*sizeof(PetscInt));
4111: if (cpvalues == MAT_COPY_VALUES) {
4112: PetscMemcpy(c->a,a->a,(a->i[m])*sizeof(PetscScalar));
4113: } else {
4114: PetscMemzero(c->a,(a->i[m])*sizeof(PetscScalar));
4115: }
4116: }
4117: }
4119: c->ignorezeroentries = a->ignorezeroentries;
4120: c->roworiented = a->roworiented;
4121: c->nonew = a->nonew;
4122: if (a->diag) {
4123: PetscMalloc1(m+1,&c->diag);
4124: PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));
4125: for (i=0; i<m; i++) {
4126: c->diag[i] = a->diag[i];
4127: }
4128: } else c->diag = 0;
4130: c->solve_work = 0;
4131: c->saved_values = 0;
4132: c->idiag = 0;
4133: c->ssor_work = 0;
4134: c->keepnonzeropattern = a->keepnonzeropattern;
4135: c->free_a = PETSC_TRUE;
4136: c->free_ij = PETSC_TRUE;
4138: c->rmax = a->rmax;
4139: c->nz = a->nz;
4140: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4141: C->preallocated = PETSC_TRUE;
4143: c->compressedrow.use = a->compressedrow.use;
4144: c->compressedrow.nrows = a->compressedrow.nrows;
4145: if (a->compressedrow.use) {
4146: i = a->compressedrow.nrows;
4147: PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);
4148: PetscMemcpy(c->compressedrow.i,a->compressedrow.i,(i+1)*sizeof(PetscInt));
4149: PetscMemcpy(c->compressedrow.rindex,a->compressedrow.rindex,i*sizeof(PetscInt));
4150: } else {
4151: c->compressedrow.use = PETSC_FALSE;
4152: c->compressedrow.i = NULL;
4153: c->compressedrow.rindex = NULL;
4154: }
4155: c->nonzerorowcnt = a->nonzerorowcnt;
4156: C->nonzerostate = A->nonzerostate;
4158: MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);
4159: PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);
4160: return(0);
4161: }
4163: PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B)
4164: {
4168: MatCreate(PetscObjectComm((PetscObject)A),B);
4169: MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);
4170: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) {
4171: MatSetBlockSizesFromMats(*B,A,A);
4172: }
4173: MatSetType(*B,((PetscObject)A)->type_name);
4174: MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);
4175: return(0);
4176: }
4178: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4179: {
4180: Mat_SeqAIJ *a;
4182: PetscInt i,sum,nz,header[4],*rowlengths = 0,M,N,rows,cols;
4183: int fd;
4184: PetscMPIInt size;
4185: MPI_Comm comm;
4186: PetscInt bs = newMat->rmap->bs;
4189: /* force binary viewer to load .info file if it has not yet done so */
4190: PetscViewerSetUp(viewer);
4191: PetscObjectGetComm((PetscObject)viewer,&comm);
4192: MPI_Comm_size(comm,&size);
4193: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"view must have one processor");
4195: PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");
4196: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
4197: PetscOptionsEnd();
4198: if (bs < 0) bs = 1;
4199: MatSetBlockSize(newMat,bs);
4201: PetscViewerBinaryGetDescriptor(viewer,&fd);
4202: PetscBinaryRead(fd,header,4,PETSC_INT);
4203: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object in file");
4204: M = header[1]; N = header[2]; nz = header[3];
4206: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as SeqAIJ");
4208: /* read in row lengths */
4209: PetscMalloc1(M,&rowlengths);
4210: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
4212: /* check if sum of rowlengths is same as nz */
4213: for (i=0,sum=0; i< M; i++) sum +=rowlengths[i];
4214: 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);
4216: /* set global size if not set already*/
4217: if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) {
4218: MatSetSizes(newMat,PETSC_DECIDE,PETSC_DECIDE,M,N);
4219: } else {
4220: /* if sizes and type are already set, check if the matrix global sizes are correct */
4221: MatGetSize(newMat,&rows,&cols);
4222: if (rows < 0 && cols < 0) { /* user might provide local size instead of global size */
4223: MatGetLocalSize(newMat,&rows,&cols);
4224: }
4225: 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);
4226: }
4227: MatSeqAIJSetPreallocation_SeqAIJ(newMat,0,rowlengths);
4228: a = (Mat_SeqAIJ*)newMat->data;
4230: PetscBinaryRead(fd,a->j,nz,PETSC_INT);
4232: /* read in nonzero values */
4233: PetscBinaryRead(fd,a->a,nz,PETSC_SCALAR);
4235: /* set matrix "i" values */
4236: a->i[0] = 0;
4237: for (i=1; i<= M; i++) {
4238: a->i[i] = a->i[i-1] + rowlengths[i-1];
4239: a->ilen[i-1] = rowlengths[i-1];
4240: }
4241: PetscFree(rowlengths);
4243: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
4244: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
4245: return(0);
4246: }
4248: PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg)
4249: {
4250: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
4252: #if defined(PETSC_USE_COMPLEX)
4253: PetscInt k;
4254: #endif
4257: /* If the matrix dimensions are not equal,or no of nonzeros */
4258: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) {
4259: *flg = PETSC_FALSE;
4260: return(0);
4261: }
4263: /* if the a->i are the same */
4264: PetscMemcmp(a->i,b->i,(A->rmap->n+1)*sizeof(PetscInt),flg);
4265: if (!*flg) return(0);
4267: /* if a->j are the same */
4268: PetscMemcmp(a->j,b->j,(a->nz)*sizeof(PetscInt),flg);
4269: if (!*flg) return(0);
4271: /* if a->a are the same */
4272: #if defined(PETSC_USE_COMPLEX)
4273: for (k=0; k<a->nz; k++) {
4274: if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) {
4275: *flg = PETSC_FALSE;
4276: return(0);
4277: }
4278: }
4279: #else
4280: PetscMemcmp(a->a,b->a,(a->nz)*sizeof(PetscScalar),flg);
4281: #endif
4282: return(0);
4283: }
4285: /*@
4286: MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format)
4287: provided by the user.
4289: Collective on MPI_Comm
4291: Input Parameters:
4292: + comm - must be an MPI communicator of size 1
4293: . m - number of rows
4294: . n - number of columns
4295: . i - row indices
4296: . j - column indices
4297: - a - matrix values
4299: Output Parameter:
4300: . mat - the matrix
4302: Level: intermediate
4304: Notes:
4305: The i, j, and a arrays are not copied by this routine, the user must free these arrays
4306: once the matrix is destroyed and not before
4308: You cannot set new nonzero locations into this matrix, that will generate an error.
4310: The i and j indices are 0 based
4312: The format which is used for the sparse matrix input, is equivalent to a
4313: row-major ordering.. i.e for the following matrix, the input data expected is
4314: as shown
4316: $ 1 0 0
4317: $ 2 0 3
4318: $ 4 5 6
4319: $
4320: $ i = {0,1,3,6} [size = nrow+1 = 3+1]
4321: $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
4322: $ v = {1,2,3,4,5,6} [size = 6]
4325: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4327: @*/
4328: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat)
4329: {
4331: PetscInt ii;
4332: Mat_SeqAIJ *aij;
4333: #if defined(PETSC_USE_DEBUG)
4334: PetscInt jj;
4335: #endif
4338: if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4339: MatCreate(comm,mat);
4340: MatSetSizes(*mat,m,n,m,n);
4341: /* MatSetBlockSizes(*mat,,); */
4342: MatSetType(*mat,MATSEQAIJ);
4343: MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,0);
4344: aij = (Mat_SeqAIJ*)(*mat)->data;
4345: PetscMalloc2(m,&aij->imax,m,&aij->ilen);
4347: aij->i = i;
4348: aij->j = j;
4349: aij->a = a;
4350: aij->singlemalloc = PETSC_FALSE;
4351: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
4352: aij->free_a = PETSC_FALSE;
4353: aij->free_ij = PETSC_FALSE;
4355: for (ii=0; ii<m; ii++) {
4356: aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii];
4357: #if defined(PETSC_USE_DEBUG)
4358: 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]);
4359: for (jj=i[ii]+1; jj<i[ii+1]; jj++) {
4360: 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);
4361: 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);
4362: }
4363: #endif
4364: }
4365: #if defined(PETSC_USE_DEBUG)
4366: for (ii=0; ii<aij->i[m]; ii++) {
4367: if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]);
4368: 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]);
4369: }
4370: #endif
4372: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4373: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4374: return(0);
4375: }
4376: /*@C
4377: MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format)
4378: provided by the user.
4380: Collective on MPI_Comm
4382: Input Parameters:
4383: + comm - must be an MPI communicator of size 1
4384: . m - number of rows
4385: . n - number of columns
4386: . i - row indices
4387: . j - column indices
4388: . a - matrix values
4389: . nz - number of nonzeros
4390: - idx - 0 or 1 based
4392: Output Parameter:
4393: . mat - the matrix
4395: Level: intermediate
4397: Notes:
4398: The i and j indices are 0 based
4400: The format which is used for the sparse matrix input, is equivalent to a
4401: row-major ordering.. i.e for the following matrix, the input data expected is
4402: as shown:
4404: 1 0 0
4405: 2 0 3
4406: 4 5 6
4408: i = {0,1,1,2,2,2}
4409: j = {0,0,2,0,1,2}
4410: v = {1,2,3,4,5,6}
4413: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4415: @*/
4416: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat,PetscInt nz,PetscBool idx)
4417: {
4419: PetscInt ii, *nnz, one = 1,row,col;
4423: PetscCalloc1(m,&nnz);
4424: for (ii = 0; ii < nz; ii++) {
4425: nnz[i[ii] - !!idx] += 1;
4426: }
4427: MatCreate(comm,mat);
4428: MatSetSizes(*mat,m,n,m,n);
4429: MatSetType(*mat,MATSEQAIJ);
4430: MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);
4431: for (ii = 0; ii < nz; ii++) {
4432: if (idx) {
4433: row = i[ii] - 1;
4434: col = j[ii] - 1;
4435: } else {
4436: row = i[ii];
4437: col = j[ii];
4438: }
4439: MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);
4440: }
4441: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4442: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4443: PetscFree(nnz);
4444: return(0);
4445: }
4447: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
4448: {
4449: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data;
4453: a->idiagvalid = PETSC_FALSE;
4454: a->ibdiagvalid = PETSC_FALSE;
4456: MatSeqAIJInvalidateDiagonal_Inode(A);
4457: return(0);
4458: }
4460: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4461: {
4463: PetscMPIInt size;
4466: MPI_Comm_size(comm,&size);
4467: if (size == 1) {
4468: if (scall == MAT_INITIAL_MATRIX) {
4469: MatDuplicate(inmat,MAT_COPY_VALUES,outmat);
4470: } else {
4471: MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);
4472: }
4473: } else {
4474: MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);
4475: }
4476: return(0);
4477: }
4479: /*
4480: Permute A into C's *local* index space using rowemb,colemb.
4481: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
4482: of [0,m), colemb is in [0,n).
4483: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
4484: */
4485: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B)
4486: {
4487: /* If making this function public, change the error returned in this function away from _PLIB. */
4489: Mat_SeqAIJ *Baij;
4490: PetscBool seqaij;
4491: PetscInt m,n,*nz,i,j,count;
4492: PetscScalar v;
4493: const PetscInt *rowindices,*colindices;
4496: if (!B) return(0);
4497: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
4498: PetscObjectBaseTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);
4499: if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type");
4500: if (rowemb) {
4501: ISGetLocalSize(rowemb,&m);
4502: 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);
4503: } else {
4504: if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix");
4505: }
4506: if (colemb) {
4507: ISGetLocalSize(colemb,&n);
4508: 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);
4509: } else {
4510: if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix");
4511: }
4513: Baij = (Mat_SeqAIJ*)(B->data);
4514: if (pattern == DIFFERENT_NONZERO_PATTERN) {
4515: PetscMalloc1(B->rmap->n,&nz);
4516: for (i=0; i<B->rmap->n; i++) {
4517: nz[i] = Baij->i[i+1] - Baij->i[i];
4518: }
4519: MatSeqAIJSetPreallocation(C,0,nz);
4520: PetscFree(nz);
4521: }
4522: if (pattern == SUBSET_NONZERO_PATTERN) {
4523: MatZeroEntries(C);
4524: }
4525: count = 0;
4526: rowindices = NULL;
4527: colindices = NULL;
4528: if (rowemb) {
4529: ISGetIndices(rowemb,&rowindices);
4530: }
4531: if (colemb) {
4532: ISGetIndices(colemb,&colindices);
4533: }
4534: for (i=0; i<B->rmap->n; i++) {
4535: PetscInt row;
4536: row = i;
4537: if (rowindices) row = rowindices[i];
4538: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
4539: PetscInt col;
4540: col = Baij->j[count];
4541: if (colindices) col = colindices[col];
4542: v = Baij->a[count];
4543: MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);
4544: ++count;
4545: }
4546: }
4547: /* FIXME: set C's nonzerostate correctly. */
4548: /* Assembly for C is necessary. */
4549: C->preallocated = PETSC_TRUE;
4550: C->assembled = PETSC_TRUE;
4551: C->was_assembled = PETSC_FALSE;
4552: return(0);
4553: }
4555: PetscFunctionList MatSeqAIJList = NULL;
4557: /*@C
4558: MatSeqAIJSetType - Converts a MATSEQAIJ matrix to a subtype
4560: Collective on Mat
4562: Input Parameters:
4563: + mat - the matrix object
4564: - matype - matrix type
4566: Options Database Key:
4567: . -mat_seqai_type <method> - for example seqaijcrl
4570: Level: intermediate
4572: .keywords: Mat, MatType, set, method
4574: .seealso: PCSetType(), VecSetType(), MatCreate(), MatType, Mat
4575: @*/
4576: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
4577: {
4578: PetscErrorCode ierr,(*r)(Mat,const MatType,MatReuse,Mat*);
4579: PetscBool sametype;
4583: PetscObjectTypeCompare((PetscObject)mat,matype,&sametype);
4584: if (sametype) return(0);
4586: PetscFunctionListFind(MatSeqAIJList,matype,&r);
4587: if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown Mat type given: %s",matype);
4588: (*r)(mat,matype,MAT_INPLACE_MATRIX,&mat);
4589: return(0);
4590: }
4593: /*@C
4594: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential AIJ matrices
4596: Not Collective
4598: Input Parameters:
4599: + name - name of a new user-defined matrix type, for example MATSEQAIJCRL
4600: - function - routine to convert to subtype
4602: Notes:
4603: MatSeqAIJRegister() may be called multiple times to add several user-defined solvers.
4606: Then, your matrix can be chosen with the procedural interface at runtime via the option
4607: $ -mat_seqaij_type my_mat
4609: Level: advanced
4611: .keywords: Mat, register
4613: .seealso: MatSeqAIJRegisterAll()
4616: Level: advanced
4617: @*/
4618: PetscErrorCode MatSeqAIJRegister(const char sname[],PetscErrorCode (*function)(Mat,MatType,MatReuse,Mat *))
4619: {
4623: PetscFunctionListAdd(&MatSeqAIJList,sname,function);
4624: return(0);
4625: }
4627: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
4629: /*@C
4630: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of SeqAIJ
4632: Not Collective
4634: Level: advanced
4636: Developers Note: CUSP and CUSPARSE do not yet support the MatConvert_SeqAIJ..() paradigm and thus cannot be registered here
4638: .keywords: KSP, register, all
4640: .seealso: MatRegisterAll(), MatSeqAIJRegister()
4641: @*/
4642: PetscErrorCode MatSeqAIJRegisterAll(void)
4643: {
4647: if (MatSeqAIJRegisterAllCalled) return(0);
4648: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
4650: MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);
4651: MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);
4652: #if defined(PETSC_HAVE_MKL_SPARSE)
4653: MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);
4654: #endif
4655: #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA)
4656: MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);
4657: #endif
4658: return(0);
4659: }
4661: /*
4662: Special version for direct calls from Fortran
4663: */
4664: #include <petsc/private/fortranimpl.h>
4665: #if defined(PETSC_HAVE_FORTRAN_CAPS)
4666: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
4667: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
4668: #define matsetvaluesseqaij_ matsetvaluesseqaij
4669: #endif
4671: /* Change these macros so can be used in void function */
4672: #undef CHKERRQ
4673: #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr)
4674: #undef SETERRQ2
4675: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
4676: #undef SETERRQ3
4677: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
4679: 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)
4680: {
4681: Mat A = *AA;
4682: PetscInt m = *mm, n = *nn;
4683: InsertMode is = *isis;
4684: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
4685: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
4686: PetscInt *imax,*ai,*ailen;
4688: PetscInt *aj,nonew = a->nonew,lastcol = -1;
4689: MatScalar *ap,value,*aa;
4690: PetscBool ignorezeroentries = a->ignorezeroentries;
4691: PetscBool roworiented = a->roworiented;
4694: MatCheckPreallocated(A,1);
4695: imax = a->imax;
4696: ai = a->i;
4697: ailen = a->ilen;
4698: aj = a->j;
4699: aa = a->a;
4701: for (k=0; k<m; k++) { /* loop over added rows */
4702: row = im[k];
4703: if (row < 0) continue;
4704: #if defined(PETSC_USE_DEBUG)
4705: if (row >= A->rmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
4706: #endif
4707: rp = aj + ai[row]; ap = aa + ai[row];
4708: rmax = imax[row]; nrow = ailen[row];
4709: low = 0;
4710: high = nrow;
4711: for (l=0; l<n; l++) { /* loop over added columns */
4712: if (in[l] < 0) continue;
4713: #if defined(PETSC_USE_DEBUG)
4714: if (in[l] >= A->cmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
4715: #endif
4716: col = in[l];
4717: if (roworiented) value = v[l + k*n];
4718: else value = v[k + l*m];
4720: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
4722: if (col <= lastcol) low = 0;
4723: else high = nrow;
4724: lastcol = col;
4725: while (high-low > 5) {
4726: t = (low+high)/2;
4727: if (rp[t] > col) high = t;
4728: else low = t;
4729: }
4730: for (i=low; i<high; i++) {
4731: if (rp[i] > col) break;
4732: if (rp[i] == col) {
4733: if (is == ADD_VALUES) ap[i] += value;
4734: else ap[i] = value;
4735: goto noinsert;
4736: }
4737: }
4738: if (value == 0.0 && ignorezeroentries) goto noinsert;
4739: if (nonew == 1) goto noinsert;
4740: if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix");
4741: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
4742: N = nrow++ - 1; a->nz++; high++;
4743: /* shift up all the later entries in this row */
4744: for (ii=N; ii>=i; ii--) {
4745: rp[ii+1] = rp[ii];
4746: ap[ii+1] = ap[ii];
4747: }
4748: rp[i] = col;
4749: ap[i] = value;
4750: A->nonzerostate++;
4751: noinsert:;
4752: low = i + 1;
4753: }
4754: ailen[row] = nrow;
4755: }
4756: PetscFunctionReturnVoid();
4757: }