Actual source code: aij.c
petsc-3.10.3 2018-12-18
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,"MatConvert_seqaij_seqsell_C",NULL);
1125: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_is_C",NULL);
1126: PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);
1127: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);
1128: PetscObjectComposeFunction((PetscObject)A,"MatResetPreallocation_C",NULL);
1129: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);
1130: PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);
1131: PetscObjectComposeFunction((PetscObject)A,"MatPtAP_is_seqaij_C",NULL);
1132: return(0);
1133: }
1135: PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg)
1136: {
1137: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1141: switch (op) {
1142: case MAT_ROW_ORIENTED:
1143: a->roworiented = flg;
1144: break;
1145: case MAT_KEEP_NONZERO_PATTERN:
1146: a->keepnonzeropattern = flg;
1147: break;
1148: case MAT_NEW_NONZERO_LOCATIONS:
1149: a->nonew = (flg ? 0 : 1);
1150: break;
1151: case MAT_NEW_NONZERO_LOCATION_ERR:
1152: a->nonew = (flg ? -1 : 0);
1153: break;
1154: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1155: a->nonew = (flg ? -2 : 0);
1156: break;
1157: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1158: a->nounused = (flg ? -1 : 0);
1159: break;
1160: case MAT_IGNORE_ZERO_ENTRIES:
1161: a->ignorezeroentries = flg;
1162: break;
1163: case MAT_SPD:
1164: case MAT_SYMMETRIC:
1165: case MAT_STRUCTURALLY_SYMMETRIC:
1166: case MAT_HERMITIAN:
1167: case MAT_SYMMETRY_ETERNAL:
1168: case MAT_STRUCTURE_ONLY:
1169: /* These options are handled directly by MatSetOption() */
1170: break;
1171: case MAT_NEW_DIAGONALS:
1172: case MAT_IGNORE_OFF_PROC_ENTRIES:
1173: case MAT_USE_HASH_TABLE:
1174: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1175: break;
1176: case MAT_USE_INODES:
1177: /* Not an error because MatSetOption_SeqAIJ_Inode handles this one */
1178: break;
1179: case MAT_SUBMAT_SINGLEIS:
1180: A->submat_singleis = flg;
1181: break;
1182: default:
1183: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1184: }
1185: MatSetOption_SeqAIJ_Inode(A,op,flg);
1186: return(0);
1187: }
1189: PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v)
1190: {
1191: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1193: PetscInt i,j,n,*ai=a->i,*aj=a->j,nz;
1194: PetscScalar *aa=a->a,*x,zero=0.0;
1197: VecGetLocalSize(v,&n);
1198: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
1200: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1201: PetscInt *diag=a->diag;
1202: VecGetArray(v,&x);
1203: for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]];
1204: VecRestoreArray(v,&x);
1205: return(0);
1206: }
1208: VecSet(v,zero);
1209: VecGetArray(v,&x);
1210: for (i=0; i<n; i++) {
1211: nz = ai[i+1] - ai[i];
1212: if (!nz) x[i] = 0.0;
1213: for (j=ai[i]; j<ai[i+1]; j++) {
1214: if (aj[j] == i) {
1215: x[i] = aa[j];
1216: break;
1217: }
1218: }
1219: }
1220: VecRestoreArray(v,&x);
1221: return(0);
1222: }
1224: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1225: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy)
1226: {
1227: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1228: PetscScalar *y;
1229: const PetscScalar *x;
1230: PetscErrorCode ierr;
1231: PetscInt m = A->rmap->n;
1232: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1233: const MatScalar *v;
1234: PetscScalar alpha;
1235: PetscInt n,i,j;
1236: const PetscInt *idx,*ii,*ridx=NULL;
1237: Mat_CompressedRow cprow = a->compressedrow;
1238: PetscBool usecprow = cprow.use;
1239: #endif
1242: if (zz != yy) {VecCopy(zz,yy);}
1243: VecGetArrayRead(xx,&x);
1244: VecGetArray(yy,&y);
1246: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1247: fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y);
1248: #else
1249: if (usecprow) {
1250: m = cprow.nrows;
1251: ii = cprow.i;
1252: ridx = cprow.rindex;
1253: } else {
1254: ii = a->i;
1255: }
1256: for (i=0; i<m; i++) {
1257: idx = a->j + ii[i];
1258: v = a->a + ii[i];
1259: n = ii[i+1] - ii[i];
1260: if (usecprow) {
1261: alpha = x[ridx[i]];
1262: } else {
1263: alpha = x[i];
1264: }
1265: for (j=0; j<n; j++) y[idx[j]] += alpha*v[j];
1266: }
1267: #endif
1268: PetscLogFlops(2.0*a->nz);
1269: VecRestoreArrayRead(xx,&x);
1270: VecRestoreArray(yy,&y);
1271: return(0);
1272: }
1274: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy)
1275: {
1279: VecSet(yy,0.0);
1280: MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);
1281: return(0);
1282: }
1284: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1286: PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy)
1287: {
1288: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1289: PetscScalar *y;
1290: const PetscScalar *x;
1291: const MatScalar *aa;
1292: PetscErrorCode ierr;
1293: PetscInt m=A->rmap->n;
1294: const PetscInt *aj,*ii,*ridx=NULL;
1295: PetscInt n,i;
1296: PetscScalar sum;
1297: PetscBool usecprow=a->compressedrow.use;
1299: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1300: #pragma disjoint(*x,*y,*aa)
1301: #endif
1304: VecGetArrayRead(xx,&x);
1305: VecGetArray(yy,&y);
1306: ii = a->i;
1307: if (usecprow) { /* use compressed row format */
1308: PetscMemzero(y,m*sizeof(PetscScalar));
1309: m = a->compressedrow.nrows;
1310: ii = a->compressedrow.i;
1311: ridx = a->compressedrow.rindex;
1312: for (i=0; i<m; i++) {
1313: n = ii[i+1] - ii[i];
1314: aj = a->j + ii[i];
1315: aa = a->a + ii[i];
1316: sum = 0.0;
1317: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1318: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1319: y[*ridx++] = sum;
1320: }
1321: } else { /* do not use compressed row format */
1322: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1323: aj = a->j;
1324: aa = a->a;
1325: fortranmultaij_(&m,x,ii,aj,aa,y);
1326: #else
1327: for (i=0; i<m; i++) {
1328: n = ii[i+1] - ii[i];
1329: aj = a->j + ii[i];
1330: aa = a->a + ii[i];
1331: sum = 0.0;
1332: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1333: y[i] = sum;
1334: }
1335: #endif
1336: }
1337: PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);
1338: VecRestoreArrayRead(xx,&x);
1339: VecRestoreArray(yy,&y);
1340: return(0);
1341: }
1343: PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy)
1344: {
1345: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1346: PetscScalar *y;
1347: const PetscScalar *x;
1348: const MatScalar *aa;
1349: PetscErrorCode ierr;
1350: PetscInt m=A->rmap->n;
1351: const PetscInt *aj,*ii,*ridx=NULL;
1352: PetscInt n,i,nonzerorow=0;
1353: PetscScalar sum;
1354: PetscBool usecprow=a->compressedrow.use;
1356: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1357: #pragma disjoint(*x,*y,*aa)
1358: #endif
1361: VecGetArrayRead(xx,&x);
1362: VecGetArray(yy,&y);
1363: if (usecprow) { /* use compressed row format */
1364: m = a->compressedrow.nrows;
1365: ii = a->compressedrow.i;
1366: ridx = a->compressedrow.rindex;
1367: for (i=0; i<m; i++) {
1368: n = ii[i+1] - ii[i];
1369: aj = a->j + ii[i];
1370: aa = a->a + ii[i];
1371: sum = 0.0;
1372: nonzerorow += (n>0);
1373: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1374: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1375: y[*ridx++] = sum;
1376: }
1377: } else { /* do not use compressed row format */
1378: ii = a->i;
1379: for (i=0; i<m; i++) {
1380: n = ii[i+1] - ii[i];
1381: aj = a->j + ii[i];
1382: aa = a->a + ii[i];
1383: sum = 0.0;
1384: nonzerorow += (n>0);
1385: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1386: y[i] = sum;
1387: }
1388: }
1389: PetscLogFlops(2.0*a->nz - nonzerorow);
1390: VecRestoreArrayRead(xx,&x);
1391: VecRestoreArray(yy,&y);
1392: return(0);
1393: }
1395: PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1396: {
1397: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1398: PetscScalar *y,*z;
1399: const PetscScalar *x;
1400: const MatScalar *aa;
1401: PetscErrorCode ierr;
1402: PetscInt m = A->rmap->n,*aj,*ii;
1403: PetscInt n,i,*ridx=NULL;
1404: PetscScalar sum;
1405: PetscBool usecprow=a->compressedrow.use;
1408: VecGetArrayRead(xx,&x);
1409: VecGetArrayPair(yy,zz,&y,&z);
1410: if (usecprow) { /* use compressed row format */
1411: if (zz != yy) {
1412: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1413: }
1414: m = a->compressedrow.nrows;
1415: ii = a->compressedrow.i;
1416: ridx = a->compressedrow.rindex;
1417: for (i=0; i<m; i++) {
1418: n = ii[i+1] - ii[i];
1419: aj = a->j + ii[i];
1420: aa = a->a + ii[i];
1421: sum = y[*ridx];
1422: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1423: z[*ridx++] = sum;
1424: }
1425: } else { /* do not use compressed row format */
1426: ii = a->i;
1427: for (i=0; i<m; i++) {
1428: n = ii[i+1] - ii[i];
1429: aj = a->j + ii[i];
1430: aa = a->a + ii[i];
1431: sum = y[i];
1432: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1433: z[i] = sum;
1434: }
1435: }
1436: PetscLogFlops(2.0*a->nz);
1437: VecRestoreArrayRead(xx,&x);
1438: VecRestoreArrayPair(yy,zz,&y,&z);
1439: return(0);
1440: }
1442: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1443: PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1444: {
1445: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1446: PetscScalar *y,*z;
1447: const PetscScalar *x;
1448: const MatScalar *aa;
1449: PetscErrorCode ierr;
1450: const PetscInt *aj,*ii,*ridx=NULL;
1451: PetscInt m = A->rmap->n,n,i;
1452: PetscScalar sum;
1453: PetscBool usecprow=a->compressedrow.use;
1456: VecGetArrayRead(xx,&x);
1457: VecGetArrayPair(yy,zz,&y,&z);
1458: if (usecprow) { /* use compressed row format */
1459: if (zz != yy) {
1460: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1461: }
1462: m = a->compressedrow.nrows;
1463: ii = a->compressedrow.i;
1464: ridx = a->compressedrow.rindex;
1465: for (i=0; i<m; i++) {
1466: n = ii[i+1] - ii[i];
1467: aj = a->j + ii[i];
1468: aa = a->a + ii[i];
1469: sum = y[*ridx];
1470: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1471: z[*ridx++] = sum;
1472: }
1473: } else { /* do not use compressed row format */
1474: ii = a->i;
1475: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1476: aj = a->j;
1477: aa = a->a;
1478: fortranmultaddaij_(&m,x,ii,aj,aa,y,z);
1479: #else
1480: for (i=0; i<m; i++) {
1481: n = ii[i+1] - ii[i];
1482: aj = a->j + ii[i];
1483: aa = a->a + ii[i];
1484: sum = y[i];
1485: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1486: z[i] = sum;
1487: }
1488: #endif
1489: }
1490: PetscLogFlops(2.0*a->nz);
1491: VecRestoreArrayRead(xx,&x);
1492: VecRestoreArrayPair(yy,zz,&y,&z);
1493: return(0);
1494: }
1496: /*
1497: Adds diagonal pointers to sparse matrix structure.
1498: */
1499: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1500: {
1501: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1503: PetscInt i,j,m = A->rmap->n;
1506: if (!a->diag) {
1507: PetscMalloc1(m,&a->diag);
1508: PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));
1509: }
1510: for (i=0; i<A->rmap->n; i++) {
1511: a->diag[i] = a->i[i+1];
1512: for (j=a->i[i]; j<a->i[i+1]; j++) {
1513: if (a->j[j] == i) {
1514: a->diag[i] = j;
1515: break;
1516: }
1517: }
1518: }
1519: return(0);
1520: }
1522: /*
1523: Checks for missing diagonals
1524: */
1525: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d)
1526: {
1527: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1528: PetscInt *diag,*ii = a->i,i;
1532: *missing = PETSC_FALSE;
1533: if (A->rmap->n > 0 && !ii) {
1534: *missing = PETSC_TRUE;
1535: if (d) *d = 0;
1536: PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");
1537: } else {
1538: diag = a->diag;
1539: for (i=0; i<A->rmap->n; i++) {
1540: if (diag[i] >= ii[i+1]) {
1541: *missing = PETSC_TRUE;
1542: if (d) *d = i;
1543: PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);
1544: break;
1545: }
1546: }
1547: }
1548: return(0);
1549: }
1551: #include <petscblaslapack.h>
1552: #include <petsc/private/kernels/blockinvert.h>
1554: /*
1555: Note that values is allocated externally by the PC and then passed into this routine
1556: */
1557: PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
1558: {
1559: PetscErrorCode ierr;
1560: PetscInt n = A->rmap->n, i, ncnt = 0, *indx,j,bsizemax = 0,*v_pivots;
1561: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
1562: const PetscReal shift = 0.0;
1563: PetscInt ipvt[5];
1564: PetscScalar work[25],*v_work;
1567: allowzeropivot = PetscNot(A->erroriffailure);
1568: for (i=0; i<nblocks; i++) ncnt += bsizes[i];
1569: if (ncnt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Total blocksizes %D doesn't match number matrix rows %D",ncnt,n);
1570: for (i=0; i<nblocks; i++) {
1571: bsizemax = PetscMax(bsizemax,bsizes[i]);
1572: }
1573: PetscMalloc1(bsizemax,&indx);
1574: if (bsizemax > 7) {
1575: PetscMalloc2(bsizemax,&v_work,bsizemax,&v_pivots);
1576: }
1577: ncnt = 0;
1578: for (i=0; i<nblocks; i++) {
1579: for (j=0; j<bsizes[i]; j++) indx[j] = ncnt+j;
1580: MatGetValues(A,bsizes[i],indx,bsizes[i],indx,diag);
1581: switch (bsizes[i]) {
1582: case 1:
1583: *diag = 1.0/(*diag);
1584: break;
1585: case 2:
1586: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
1587: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1588: PetscKernel_A_gets_transpose_A_2(diag);
1589: break;
1590: case 3:
1591: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
1592: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1593: PetscKernel_A_gets_transpose_A_3(diag);
1594: break;
1595: case 4:
1596: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
1597: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1598: PetscKernel_A_gets_transpose_A_4(diag);
1599: break;
1600: case 5:
1601: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
1602: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1603: PetscKernel_A_gets_transpose_A_5(diag);
1604: break;
1605: case 6:
1606: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
1607: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1608: PetscKernel_A_gets_transpose_A_6(diag);
1609: break;
1610: case 7:
1611: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
1612: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1613: PetscKernel_A_gets_transpose_A_7(diag);
1614: break;
1615: default:
1616: PetscKernel_A_gets_inverse_A(bsizes[i],diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
1617: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1618: PetscKernel_A_gets_transpose_A_N(diag,bsizes[i]);
1619: }
1620: ncnt += bsizes[i];
1621: diag += bsizes[i]*bsizes[i];
1622: }
1623: if (bsizemax > 7) {
1624: PetscFree2(v_work,v_pivots);
1625: }
1626: PetscFree(indx);
1627: return(0);
1628: }
1630: /*
1631: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1632: */
1633: PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift)
1634: {
1635: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
1637: PetscInt i,*diag,m = A->rmap->n;
1638: MatScalar *v = a->a;
1639: PetscScalar *idiag,*mdiag;
1642: if (a->idiagvalid) return(0);
1643: MatMarkDiagonal_SeqAIJ(A);
1644: diag = a->diag;
1645: if (!a->idiag) {
1646: PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);
1647: PetscLogObjectMemory((PetscObject)A, 3*m*sizeof(PetscScalar));
1648: v = a->a;
1649: }
1650: mdiag = a->mdiag;
1651: idiag = a->idiag;
1653: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1654: for (i=0; i<m; i++) {
1655: mdiag[i] = v[diag[i]];
1656: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1657: if (PetscRealPart(fshift)) {
1658: PetscInfo1(A,"Zero diagonal on row %D\n",i);
1659: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1660: A->factorerror_zeropivot_value = 0.0;
1661: A->factorerror_zeropivot_row = i;
1662: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i);
1663: }
1664: idiag[i] = 1.0/v[diag[i]];
1665: }
1666: PetscLogFlops(m);
1667: } else {
1668: for (i=0; i<m; i++) {
1669: mdiag[i] = v[diag[i]];
1670: idiag[i] = omega/(fshift + v[diag[i]]);
1671: }
1672: PetscLogFlops(2.0*m);
1673: }
1674: a->idiagvalid = PETSC_TRUE;
1675: return(0);
1676: }
1678: #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h>
1679: PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1680: {
1681: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1682: PetscScalar *x,d,sum,*t,scale;
1683: const MatScalar *v,*idiag=0,*mdiag;
1684: const PetscScalar *b, *bs,*xb, *ts;
1685: PetscErrorCode ierr;
1686: PetscInt n,m = A->rmap->n,i;
1687: const PetscInt *idx,*diag;
1690: its = its*lits;
1692: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1693: if (!a->idiagvalid) {MatInvertDiagonal_SeqAIJ(A,omega,fshift);}
1694: a->fshift = fshift;
1695: a->omega = omega;
1697: diag = a->diag;
1698: t = a->ssor_work;
1699: idiag = a->idiag;
1700: mdiag = a->mdiag;
1702: VecGetArray(xx,&x);
1703: VecGetArrayRead(bb,&b);
1704: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1705: if (flag == SOR_APPLY_UPPER) {
1706: /* apply (U + D/omega) to the vector */
1707: bs = b;
1708: for (i=0; i<m; i++) {
1709: d = fshift + mdiag[i];
1710: n = a->i[i+1] - diag[i] - 1;
1711: idx = a->j + diag[i] + 1;
1712: v = a->a + diag[i] + 1;
1713: sum = b[i]*d/omega;
1714: PetscSparseDensePlusDot(sum,bs,v,idx,n);
1715: x[i] = sum;
1716: }
1717: VecRestoreArray(xx,&x);
1718: VecRestoreArrayRead(bb,&b);
1719: PetscLogFlops(a->nz);
1720: return(0);
1721: }
1723: if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented");
1724: else if (flag & SOR_EISENSTAT) {
1725: /* Let A = L + U + D; where L is lower trianglar,
1726: U is upper triangular, E = D/omega; This routine applies
1728: (L + E)^{-1} A (U + E)^{-1}
1730: to a vector efficiently using Eisenstat's trick.
1731: */
1732: scale = (2.0/omega) - 1.0;
1734: /* x = (E + U)^{-1} b */
1735: for (i=m-1; i>=0; i--) {
1736: n = a->i[i+1] - diag[i] - 1;
1737: idx = a->j + diag[i] + 1;
1738: v = a->a + diag[i] + 1;
1739: sum = b[i];
1740: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1741: x[i] = sum*idiag[i];
1742: }
1744: /* t = b - (2*E - D)x */
1745: v = a->a;
1746: for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i];
1748: /* t = (E + L)^{-1}t */
1749: ts = t;
1750: diag = a->diag;
1751: for (i=0; i<m; i++) {
1752: n = diag[i] - a->i[i];
1753: idx = a->j + a->i[i];
1754: v = a->a + a->i[i];
1755: sum = t[i];
1756: PetscSparseDenseMinusDot(sum,ts,v,idx,n);
1757: t[i] = sum*idiag[i];
1758: /* x = x + t */
1759: x[i] += t[i];
1760: }
1762: PetscLogFlops(6.0*m-1 + 2.0*a->nz);
1763: VecRestoreArray(xx,&x);
1764: VecRestoreArrayRead(bb,&b);
1765: return(0);
1766: }
1767: if (flag & SOR_ZERO_INITIAL_GUESS) {
1768: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1769: for (i=0; i<m; i++) {
1770: n = diag[i] - a->i[i];
1771: idx = a->j + a->i[i];
1772: v = a->a + a->i[i];
1773: sum = b[i];
1774: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1775: t[i] = sum;
1776: x[i] = sum*idiag[i];
1777: }
1778: xb = t;
1779: PetscLogFlops(a->nz);
1780: } else xb = b;
1781: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1782: for (i=m-1; i>=0; i--) {
1783: n = a->i[i+1] - diag[i] - 1;
1784: idx = a->j + diag[i] + 1;
1785: v = a->a + diag[i] + 1;
1786: sum = xb[i];
1787: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1788: if (xb == b) {
1789: x[i] = sum*idiag[i];
1790: } else {
1791: x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1792: }
1793: }
1794: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1795: }
1796: its--;
1797: }
1798: while (its--) {
1799: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1800: for (i=0; i<m; i++) {
1801: /* lower */
1802: n = diag[i] - a->i[i];
1803: idx = a->j + a->i[i];
1804: v = a->a + a->i[i];
1805: sum = b[i];
1806: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1807: t[i] = sum; /* save application of the lower-triangular part */
1808: /* upper */
1809: n = a->i[i+1] - diag[i] - 1;
1810: idx = a->j + diag[i] + 1;
1811: v = a->a + diag[i] + 1;
1812: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1813: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1814: }
1815: xb = t;
1816: PetscLogFlops(2.0*a->nz);
1817: } else xb = b;
1818: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1819: for (i=m-1; i>=0; i--) {
1820: sum = xb[i];
1821: if (xb == b) {
1822: /* whole matrix (no checkpointing available) */
1823: n = a->i[i+1] - a->i[i];
1824: idx = a->j + a->i[i];
1825: v = a->a + a->i[i];
1826: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1827: x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i];
1828: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
1829: n = a->i[i+1] - diag[i] - 1;
1830: idx = a->j + diag[i] + 1;
1831: v = a->a + diag[i] + 1;
1832: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1833: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1834: }
1835: }
1836: if (xb == b) {
1837: PetscLogFlops(2.0*a->nz);
1838: } else {
1839: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1840: }
1841: }
1842: }
1843: VecRestoreArray(xx,&x);
1844: VecRestoreArrayRead(bb,&b);
1845: return(0);
1846: }
1849: PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info)
1850: {
1851: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1854: info->block_size = 1.0;
1855: info->nz_allocated = (double)a->maxnz;
1856: info->nz_used = (double)a->nz;
1857: info->nz_unneeded = (double)(a->maxnz - a->nz);
1858: info->assemblies = (double)A->num_ass;
1859: info->mallocs = (double)A->info.mallocs;
1860: info->memory = ((PetscObject)A)->mem;
1861: if (A->factortype) {
1862: info->fill_ratio_given = A->info.fill_ratio_given;
1863: info->fill_ratio_needed = A->info.fill_ratio_needed;
1864: info->factor_mallocs = A->info.factor_mallocs;
1865: } else {
1866: info->fill_ratio_given = 0;
1867: info->fill_ratio_needed = 0;
1868: info->factor_mallocs = 0;
1869: }
1870: return(0);
1871: }
1873: PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1874: {
1875: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1876: PetscInt i,m = A->rmap->n - 1;
1877: PetscErrorCode ierr;
1878: const PetscScalar *xx;
1879: PetscScalar *bb;
1880: PetscInt d = 0;
1883: if (x && b) {
1884: VecGetArrayRead(x,&xx);
1885: VecGetArray(b,&bb);
1886: for (i=0; i<N; i++) {
1887: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1888: bb[rows[i]] = diag*xx[rows[i]];
1889: }
1890: VecRestoreArrayRead(x,&xx);
1891: VecRestoreArray(b,&bb);
1892: }
1894: if (a->keepnonzeropattern) {
1895: for (i=0; i<N; i++) {
1896: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1897: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1898: }
1899: if (diag != 0.0) {
1900: for (i=0; i<N; i++) {
1901: d = rows[i];
1902: 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);
1903: }
1904: for (i=0; i<N; i++) {
1905: a->a[a->diag[rows[i]]] = diag;
1906: }
1907: }
1908: } else {
1909: if (diag != 0.0) {
1910: for (i=0; i<N; i++) {
1911: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1912: if (a->ilen[rows[i]] > 0) {
1913: a->ilen[rows[i]] = 1;
1914: a->a[a->i[rows[i]]] = diag;
1915: a->j[a->i[rows[i]]] = rows[i];
1916: } else { /* in case row was completely empty */
1917: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
1918: }
1919: }
1920: } else {
1921: for (i=0; i<N; i++) {
1922: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1923: a->ilen[rows[i]] = 0;
1924: }
1925: }
1926: A->nonzerostate++;
1927: }
1928: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
1929: return(0);
1930: }
1932: PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1933: {
1934: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1935: PetscInt i,j,m = A->rmap->n - 1,d = 0;
1936: PetscErrorCode ierr;
1937: PetscBool missing,*zeroed,vecs = PETSC_FALSE;
1938: const PetscScalar *xx;
1939: PetscScalar *bb;
1942: if (x && b) {
1943: VecGetArrayRead(x,&xx);
1944: VecGetArray(b,&bb);
1945: vecs = PETSC_TRUE;
1946: }
1947: PetscCalloc1(A->rmap->n,&zeroed);
1948: for (i=0; i<N; i++) {
1949: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1950: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1952: zeroed[rows[i]] = PETSC_TRUE;
1953: }
1954: for (i=0; i<A->rmap->n; i++) {
1955: if (!zeroed[i]) {
1956: for (j=a->i[i]; j<a->i[i+1]; j++) {
1957: if (zeroed[a->j[j]]) {
1958: if (vecs) bb[i] -= a->a[j]*xx[a->j[j]];
1959: a->a[j] = 0.0;
1960: }
1961: }
1962: } else if (vecs) bb[i] = diag*xx[i];
1963: }
1964: if (x && b) {
1965: VecRestoreArrayRead(x,&xx);
1966: VecRestoreArray(b,&bb);
1967: }
1968: PetscFree(zeroed);
1969: if (diag != 0.0) {
1970: MatMissingDiagonal_SeqAIJ(A,&missing,&d);
1971: if (missing) {
1972: if (a->nonew) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D",d);
1973: else {
1974: for (i=0; i<N; i++) {
1975: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
1976: }
1977: }
1978: } else {
1979: for (i=0; i<N; i++) {
1980: a->a[a->diag[rows[i]]] = diag;
1981: }
1982: }
1983: }
1984: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
1985: return(0);
1986: }
1988: PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1989: {
1990: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1991: PetscInt *itmp;
1994: if (row < 0 || row >= A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range",row);
1996: *nz = a->i[row+1] - a->i[row];
1997: if (v) *v = a->a + a->i[row];
1998: if (idx) {
1999: itmp = a->j + a->i[row];
2000: if (*nz) *idx = itmp;
2001: else *idx = 0;
2002: }
2003: return(0);
2004: }
2006: /* remove this function? */
2007: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
2008: {
2010: return(0);
2011: }
2013: PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm)
2014: {
2015: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2016: MatScalar *v = a->a;
2017: PetscReal sum = 0.0;
2019: PetscInt i,j;
2022: if (type == NORM_FROBENIUS) {
2023: #if defined(PETSC_USE_REAL___FP16)
2024: PetscBLASInt one = 1,nz = a->nz;
2025: *nrm = BLASnrm2_(&nz,v,&one);
2026: #else
2027: for (i=0; i<a->nz; i++) {
2028: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
2029: }
2030: *nrm = PetscSqrtReal(sum);
2031: #endif
2032: PetscLogFlops(2*a->nz);
2033: } else if (type == NORM_1) {
2034: PetscReal *tmp;
2035: PetscInt *jj = a->j;
2036: PetscCalloc1(A->cmap->n+1,&tmp);
2037: *nrm = 0.0;
2038: for (j=0; j<a->nz; j++) {
2039: tmp[*jj++] += PetscAbsScalar(*v); v++;
2040: }
2041: for (j=0; j<A->cmap->n; j++) {
2042: if (tmp[j] > *nrm) *nrm = tmp[j];
2043: }
2044: PetscFree(tmp);
2045: PetscLogFlops(PetscMax(a->nz-1,0));
2046: } else if (type == NORM_INFINITY) {
2047: *nrm = 0.0;
2048: for (j=0; j<A->rmap->n; j++) {
2049: v = a->a + a->i[j];
2050: sum = 0.0;
2051: for (i=0; i<a->i[j+1]-a->i[j]; i++) {
2052: sum += PetscAbsScalar(*v); v++;
2053: }
2054: if (sum > *nrm) *nrm = sum;
2055: }
2056: PetscLogFlops(PetscMax(a->nz-1,0));
2057: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm");
2058: return(0);
2059: }
2061: /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */
2062: PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B)
2063: {
2065: PetscInt i,j,anzj;
2066: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b;
2067: PetscInt an=A->cmap->N,am=A->rmap->N;
2068: PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j;
2071: /* Allocate space for symbolic transpose info and work array */
2072: PetscCalloc1(an+1,&ati);
2073: PetscMalloc1(ai[am],&atj);
2074: PetscMalloc1(an,&atfill);
2076: /* Walk through aj and count ## of non-zeros in each row of A^T. */
2077: /* Note: offset by 1 for fast conversion into csr format. */
2078: for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1;
2079: /* Form ati for csr format of A^T. */
2080: for (i=0;i<an;i++) ati[i+1] += ati[i];
2082: /* Copy ati into atfill so we have locations of the next free space in atj */
2083: PetscMemcpy(atfill,ati,an*sizeof(PetscInt));
2085: /* Walk through A row-wise and mark nonzero entries of A^T. */
2086: for (i=0;i<am;i++) {
2087: anzj = ai[i+1] - ai[i];
2088: for (j=0;j<anzj;j++) {
2089: atj[atfill[*aj]] = i;
2090: atfill[*aj++] += 1;
2091: }
2092: }
2094: /* Clean up temporary space and complete requests. */
2095: PetscFree(atfill);
2096: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);
2097: MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2099: b = (Mat_SeqAIJ*)((*B)->data);
2100: b->free_a = PETSC_FALSE;
2101: b->free_ij = PETSC_TRUE;
2102: b->nonew = 0;
2103: return(0);
2104: }
2106: PetscErrorCode MatTranspose_SeqAIJ(Mat A,MatReuse reuse,Mat *B)
2107: {
2108: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2109: Mat C;
2111: PetscInt i,*aj = a->j,*ai = a->i,m = A->rmap->n,len,*col;
2112: MatScalar *array = a->a;
2115: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_INPLACE_MATRIX) {
2116: PetscCalloc1(1+A->cmap->n,&col);
2118: for (i=0; i<ai[m]; i++) col[aj[i]] += 1;
2119: MatCreate(PetscObjectComm((PetscObject)A),&C);
2120: MatSetSizes(C,A->cmap->n,m,A->cmap->n,m);
2121: MatSetBlockSizes(C,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2122: MatSetType(C,((PetscObject)A)->type_name);
2123: MatSeqAIJSetPreallocation_SeqAIJ(C,0,col);
2124: PetscFree(col);
2125: } else {
2126: C = *B;
2127: }
2129: for (i=0; i<m; i++) {
2130: len = ai[i+1]-ai[i];
2131: MatSetValues_SeqAIJ(C,len,aj,1,&i,array,INSERT_VALUES);
2132: array += len;
2133: aj += len;
2134: }
2135: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2136: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2138: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2139: *B = C;
2140: } else {
2141: MatHeaderMerge(A,&C);
2142: }
2143: return(0);
2144: }
2146: PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2147: {
2148: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2149: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2150: MatScalar *va,*vb;
2152: PetscInt ma,na,mb,nb, i;
2155: MatGetSize(A,&ma,&na);
2156: MatGetSize(B,&mb,&nb);
2157: if (ma!=nb || na!=mb) {
2158: *f = PETSC_FALSE;
2159: return(0);
2160: }
2161: aii = aij->i; bii = bij->i;
2162: adx = aij->j; bdx = bij->j;
2163: va = aij->a; vb = bij->a;
2164: PetscMalloc1(ma,&aptr);
2165: PetscMalloc1(mb,&bptr);
2166: for (i=0; i<ma; i++) aptr[i] = aii[i];
2167: for (i=0; i<mb; i++) bptr[i] = bii[i];
2169: *f = PETSC_TRUE;
2170: for (i=0; i<ma; i++) {
2171: while (aptr[i]<aii[i+1]) {
2172: PetscInt idc,idr;
2173: PetscScalar vc,vr;
2174: /* column/row index/value */
2175: idc = adx[aptr[i]];
2176: idr = bdx[bptr[idc]];
2177: vc = va[aptr[i]];
2178: vr = vb[bptr[idc]];
2179: if (i!=idr || PetscAbsScalar(vc-vr) > tol) {
2180: *f = PETSC_FALSE;
2181: goto done;
2182: } else {
2183: aptr[i]++;
2184: if (B || i!=idc) bptr[idc]++;
2185: }
2186: }
2187: }
2188: done:
2189: PetscFree(aptr);
2190: PetscFree(bptr);
2191: return(0);
2192: }
2194: PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2195: {
2196: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2197: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2198: MatScalar *va,*vb;
2200: PetscInt ma,na,mb,nb, i;
2203: MatGetSize(A,&ma,&na);
2204: MatGetSize(B,&mb,&nb);
2205: if (ma!=nb || na!=mb) {
2206: *f = PETSC_FALSE;
2207: return(0);
2208: }
2209: aii = aij->i; bii = bij->i;
2210: adx = aij->j; bdx = bij->j;
2211: va = aij->a; vb = bij->a;
2212: PetscMalloc1(ma,&aptr);
2213: PetscMalloc1(mb,&bptr);
2214: for (i=0; i<ma; i++) aptr[i] = aii[i];
2215: for (i=0; i<mb; i++) bptr[i] = bii[i];
2217: *f = PETSC_TRUE;
2218: for (i=0; i<ma; i++) {
2219: while (aptr[i]<aii[i+1]) {
2220: PetscInt idc,idr;
2221: PetscScalar vc,vr;
2222: /* column/row index/value */
2223: idc = adx[aptr[i]];
2224: idr = bdx[bptr[idc]];
2225: vc = va[aptr[i]];
2226: vr = vb[bptr[idc]];
2227: if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) {
2228: *f = PETSC_FALSE;
2229: goto done;
2230: } else {
2231: aptr[i]++;
2232: if (B || i!=idc) bptr[idc]++;
2233: }
2234: }
2235: }
2236: done:
2237: PetscFree(aptr);
2238: PetscFree(bptr);
2239: return(0);
2240: }
2242: PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2243: {
2247: MatIsTranspose_SeqAIJ(A,A,tol,f);
2248: return(0);
2249: }
2251: PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2252: {
2256: MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);
2257: return(0);
2258: }
2260: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr)
2261: {
2262: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2263: const PetscScalar *l,*r;
2264: PetscScalar x;
2265: MatScalar *v;
2266: PetscErrorCode ierr;
2267: PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz;
2268: const PetscInt *jj;
2271: if (ll) {
2272: /* The local size is used so that VecMPI can be passed to this routine
2273: by MatDiagonalScale_MPIAIJ */
2274: VecGetLocalSize(ll,&m);
2275: if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length");
2276: VecGetArrayRead(ll,&l);
2277: v = a->a;
2278: for (i=0; i<m; i++) {
2279: x = l[i];
2280: M = a->i[i+1] - a->i[i];
2281: for (j=0; j<M; j++) (*v++) *= x;
2282: }
2283: VecRestoreArrayRead(ll,&l);
2284: PetscLogFlops(nz);
2285: }
2286: if (rr) {
2287: VecGetLocalSize(rr,&n);
2288: if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length");
2289: VecGetArrayRead(rr,&r);
2290: v = a->a; jj = a->j;
2291: for (i=0; i<nz; i++) (*v++) *= r[*jj++];
2292: VecRestoreArrayRead(rr,&r);
2293: PetscLogFlops(nz);
2294: }
2295: MatSeqAIJInvalidateDiagonal(A);
2296: return(0);
2297: }
2299: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B)
2300: {
2301: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c;
2303: PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens;
2304: PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi;
2305: const PetscInt *irow,*icol;
2306: PetscInt nrows,ncols;
2307: PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen;
2308: MatScalar *a_new,*mat_a;
2309: Mat C;
2310: PetscBool stride;
2314: ISGetIndices(isrow,&irow);
2315: ISGetLocalSize(isrow,&nrows);
2316: ISGetLocalSize(iscol,&ncols);
2318: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);
2319: if (stride) {
2320: ISStrideGetInfo(iscol,&first,&step);
2321: } else {
2322: first = 0;
2323: step = 0;
2324: }
2325: if (stride && step == 1) {
2326: /* special case of contiguous rows */
2327: PetscMalloc2(nrows,&lens,nrows,&starts);
2328: /* loop over new rows determining lens and starting points */
2329: for (i=0; i<nrows; i++) {
2330: kstart = ai[irow[i]];
2331: kend = kstart + ailen[irow[i]];
2332: starts[i] = kstart;
2333: for (k=kstart; k<kend; k++) {
2334: if (aj[k] >= first) {
2335: starts[i] = k;
2336: break;
2337: }
2338: }
2339: sum = 0;
2340: while (k < kend) {
2341: if (aj[k++] >= first+ncols) break;
2342: sum++;
2343: }
2344: lens[i] = sum;
2345: }
2346: /* create submatrix */
2347: if (scall == MAT_REUSE_MATRIX) {
2348: PetscInt n_cols,n_rows;
2349: MatGetSize(*B,&n_rows,&n_cols);
2350: if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
2351: MatZeroEntries(*B);
2352: C = *B;
2353: } else {
2354: PetscInt rbs,cbs;
2355: MatCreate(PetscObjectComm((PetscObject)A),&C);
2356: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2357: ISGetBlockSize(isrow,&rbs);
2358: ISGetBlockSize(iscol,&cbs);
2359: MatSetBlockSizes(C,rbs,cbs);
2360: MatSetType(C,((PetscObject)A)->type_name);
2361: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2362: }
2363: c = (Mat_SeqAIJ*)C->data;
2365: /* loop over rows inserting into submatrix */
2366: a_new = c->a;
2367: j_new = c->j;
2368: i_new = c->i;
2370: for (i=0; i<nrows; i++) {
2371: ii = starts[i];
2372: lensi = lens[i];
2373: for (k=0; k<lensi; k++) {
2374: *j_new++ = aj[ii+k] - first;
2375: }
2376: PetscMemcpy(a_new,a->a + starts[i],lensi*sizeof(PetscScalar));
2377: a_new += lensi;
2378: i_new[i+1] = i_new[i] + lensi;
2379: c->ilen[i] = lensi;
2380: }
2381: PetscFree2(lens,starts);
2382: } else {
2383: ISGetIndices(iscol,&icol);
2384: PetscCalloc1(oldcols,&smap);
2385: PetscMalloc1(1+nrows,&lens);
2386: for (i=0; i<ncols; i++) {
2387: #if defined(PETSC_USE_DEBUG)
2388: 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);
2389: #endif
2390: smap[icol[i]] = i+1;
2391: }
2393: /* determine lens of each row */
2394: for (i=0; i<nrows; i++) {
2395: kstart = ai[irow[i]];
2396: kend = kstart + a->ilen[irow[i]];
2397: lens[i] = 0;
2398: for (k=kstart; k<kend; k++) {
2399: if (smap[aj[k]]) {
2400: lens[i]++;
2401: }
2402: }
2403: }
2404: /* Create and fill new matrix */
2405: if (scall == MAT_REUSE_MATRIX) {
2406: PetscBool equal;
2408: c = (Mat_SeqAIJ*)((*B)->data);
2409: if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
2410: PetscMemcmp(c->ilen,lens,(*B)->rmap->n*sizeof(PetscInt),&equal);
2411: if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
2412: PetscMemzero(c->ilen,(*B)->rmap->n*sizeof(PetscInt));
2413: C = *B;
2414: } else {
2415: PetscInt rbs,cbs;
2416: MatCreate(PetscObjectComm((PetscObject)A),&C);
2417: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2418: ISGetBlockSize(isrow,&rbs);
2419: ISGetBlockSize(iscol,&cbs);
2420: MatSetBlockSizes(C,rbs,cbs);
2421: MatSetType(C,((PetscObject)A)->type_name);
2422: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2423: }
2424: c = (Mat_SeqAIJ*)(C->data);
2425: for (i=0; i<nrows; i++) {
2426: row = irow[i];
2427: kstart = ai[row];
2428: kend = kstart + a->ilen[row];
2429: mat_i = c->i[i];
2430: mat_j = c->j + mat_i;
2431: mat_a = c->a + mat_i;
2432: mat_ilen = c->ilen + i;
2433: for (k=kstart; k<kend; k++) {
2434: if ((tcol=smap[a->j[k]])) {
2435: *mat_j++ = tcol - 1;
2436: *mat_a++ = a->a[k];
2437: (*mat_ilen)++;
2439: }
2440: }
2441: }
2442: /* Free work space */
2443: ISRestoreIndices(iscol,&icol);
2444: PetscFree(smap);
2445: PetscFree(lens);
2446: /* sort */
2447: for (i = 0; i < nrows; i++) {
2448: PetscInt ilen;
2450: mat_i = c->i[i];
2451: mat_j = c->j + mat_i;
2452: mat_a = c->a + mat_i;
2453: ilen = c->ilen[i];
2454: PetscSortIntWithScalarArray(ilen,mat_j,mat_a);
2455: }
2456: }
2457: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2458: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2460: ISRestoreIndices(isrow,&irow);
2461: *B = C;
2462: return(0);
2463: }
2465: PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat)
2466: {
2468: Mat B;
2471: if (scall == MAT_INITIAL_MATRIX) {
2472: MatCreate(subComm,&B);
2473: MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);
2474: MatSetBlockSizesFromMats(B,mat,mat);
2475: MatSetType(B,MATSEQAIJ);
2476: MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);
2477: *subMat = B;
2478: } else {
2479: MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);
2480: }
2481: return(0);
2482: }
2484: PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info)
2485: {
2486: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2488: Mat outA;
2489: PetscBool row_identity,col_identity;
2492: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu");
2494: ISIdentity(row,&row_identity);
2495: ISIdentity(col,&col_identity);
2497: outA = inA;
2498: outA->factortype = MAT_FACTOR_LU;
2499: PetscFree(inA->solvertype);
2500: PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);
2502: PetscObjectReference((PetscObject)row);
2503: ISDestroy(&a->row);
2505: a->row = row;
2507: PetscObjectReference((PetscObject)col);
2508: ISDestroy(&a->col);
2510: a->col = col;
2512: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2513: ISDestroy(&a->icol);
2514: ISInvertPermutation(col,PETSC_DECIDE,&a->icol);
2515: PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);
2517: if (!a->solve_work) { /* this matrix may have been factored before */
2518: PetscMalloc1(inA->rmap->n+1,&a->solve_work);
2519: PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));
2520: }
2522: MatMarkDiagonal_SeqAIJ(inA);
2523: if (row_identity && col_identity) {
2524: MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);
2525: } else {
2526: MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);
2527: }
2528: return(0);
2529: }
2531: PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha)
2532: {
2533: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2534: PetscScalar oalpha = alpha;
2536: PetscBLASInt one = 1,bnz;
2539: PetscBLASIntCast(a->nz,&bnz);
2540: PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&oalpha,a->a,&one));
2541: PetscLogFlops(a->nz);
2542: MatSeqAIJInvalidateDiagonal(inA);
2543: return(0);
2544: }
2546: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2547: {
2549: PetscInt i;
2552: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2553: PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);
2555: for (i=0; i<submatj->nrqr; ++i) {
2556: PetscFree(submatj->sbuf2[i]);
2557: }
2558: PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);
2560: if (submatj->rbuf1) {
2561: PetscFree(submatj->rbuf1[0]);
2562: PetscFree(submatj->rbuf1);
2563: }
2565: for (i=0; i<submatj->nrqs; ++i) {
2566: PetscFree(submatj->rbuf3[i]);
2567: }
2568: PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);
2569: PetscFree(submatj->pa);
2570: }
2572: #if defined(PETSC_USE_CTABLE)
2573: PetscTableDestroy((PetscTable*)&submatj->rmap);
2574: if (submatj->cmap_loc) {PetscFree(submatj->cmap_loc);}
2575: PetscFree(submatj->rmap_loc);
2576: #else
2577: PetscFree(submatj->rmap);
2578: #endif
2580: if (!submatj->allcolumns) {
2581: #if defined(PETSC_USE_CTABLE)
2582: PetscTableDestroy((PetscTable*)&submatj->cmap);
2583: #else
2584: PetscFree(submatj->cmap);
2585: #endif
2586: }
2587: PetscFree(submatj->row2proc);
2589: PetscFree(submatj);
2590: return(0);
2591: }
2593: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2594: {
2596: Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data;
2597: Mat_SubSppt *submatj = c->submatis1;
2600: (*submatj->destroy)(C);
2601: MatDestroySubMatrix_Private(submatj);
2602: return(0);
2603: }
2605: PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n,Mat *mat[])
2606: {
2608: PetscInt i;
2609: Mat C;
2610: Mat_SeqAIJ *c;
2611: Mat_SubSppt *submatj;
2614: for (i=0; i<n; i++) {
2615: C = (*mat)[i];
2616: c = (Mat_SeqAIJ*)C->data;
2617: submatj = c->submatis1;
2618: if (submatj) {
2619: if (--((PetscObject)C)->refct <= 0) {
2620: (*submatj->destroy)(C);
2621: MatDestroySubMatrix_Private(submatj);
2622: PetscFree(C->defaultvectype);
2623: PetscLayoutDestroy(&C->rmap);
2624: PetscLayoutDestroy(&C->cmap);
2625: PetscHeaderDestroy(&C);
2626: }
2627: } else {
2628: MatDestroy(&C);
2629: }
2630: }
2632: PetscFree(*mat);
2633: return(0);
2634: }
2636: PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
2637: {
2639: PetscInt i;
2642: if (scall == MAT_INITIAL_MATRIX) {
2643: PetscCalloc1(n+1,B);
2644: }
2646: for (i=0; i<n; i++) {
2647: MatCreateSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);
2648: }
2649: return(0);
2650: }
2652: PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov)
2653: {
2654: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2656: PetscInt row,i,j,k,l,m,n,*nidx,isz,val;
2657: const PetscInt *idx;
2658: PetscInt start,end,*ai,*aj;
2659: PetscBT table;
2662: m = A->rmap->n;
2663: ai = a->i;
2664: aj = a->j;
2666: if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used");
2668: PetscMalloc1(m+1,&nidx);
2669: PetscBTCreate(m,&table);
2671: for (i=0; i<is_max; i++) {
2672: /* Initialize the two local arrays */
2673: isz = 0;
2674: PetscBTMemzero(m,table);
2676: /* Extract the indices, assume there can be duplicate entries */
2677: ISGetIndices(is[i],&idx);
2678: ISGetLocalSize(is[i],&n);
2680: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2681: for (j=0; j<n; ++j) {
2682: if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j];
2683: }
2684: ISRestoreIndices(is[i],&idx);
2685: ISDestroy(&is[i]);
2687: k = 0;
2688: for (j=0; j<ov; j++) { /* for each overlap */
2689: n = isz;
2690: for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */
2691: row = nidx[k];
2692: start = ai[row];
2693: end = ai[row+1];
2694: for (l = start; l<end; l++) {
2695: val = aj[l];
2696: if (!PetscBTLookupSet(table,val)) nidx[isz++] = val;
2697: }
2698: }
2699: }
2700: ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));
2701: }
2702: PetscBTDestroy(&table);
2703: PetscFree(nidx);
2704: return(0);
2705: }
2707: /* -------------------------------------------------------------- */
2708: PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B)
2709: {
2710: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2712: PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n;
2713: const PetscInt *row,*col;
2714: PetscInt *cnew,j,*lens;
2715: IS icolp,irowp;
2716: PetscInt *cwork = NULL;
2717: PetscScalar *vwork = NULL;
2720: ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);
2721: ISGetIndices(irowp,&row);
2722: ISInvertPermutation(colp,PETSC_DECIDE,&icolp);
2723: ISGetIndices(icolp,&col);
2725: /* determine lengths of permuted rows */
2726: PetscMalloc1(m+1,&lens);
2727: for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i];
2728: MatCreate(PetscObjectComm((PetscObject)A),B);
2729: MatSetSizes(*B,m,n,m,n);
2730: MatSetBlockSizesFromMats(*B,A,A);
2731: MatSetType(*B,((PetscObject)A)->type_name);
2732: MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);
2733: PetscFree(lens);
2735: PetscMalloc1(n,&cnew);
2736: for (i=0; i<m; i++) {
2737: MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2738: for (j=0; j<nz; j++) cnew[j] = col[cwork[j]];
2739: MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);
2740: MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2741: }
2742: PetscFree(cnew);
2744: (*B)->assembled = PETSC_FALSE;
2746: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
2747: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
2748: ISRestoreIndices(irowp,&row);
2749: ISRestoreIndices(icolp,&col);
2750: ISDestroy(&irowp);
2751: ISDestroy(&icolp);
2752: return(0);
2753: }
2755: PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str)
2756: {
2760: /* If the two matrices have the same copy implementation, use fast copy. */
2761: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2762: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2763: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
2765: 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");
2766: PetscMemcpy(b->a,a->a,(a->i[A->rmap->n])*sizeof(PetscScalar));
2767: PetscObjectStateIncrease((PetscObject)B);
2768: } else {
2769: MatCopy_Basic(A,B,str);
2770: }
2771: return(0);
2772: }
2774: PetscErrorCode MatSetUp_SeqAIJ(Mat A)
2775: {
2779: MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,0);
2780: return(0);
2781: }
2783: PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[])
2784: {
2785: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2788: *array = a->a;
2789: return(0);
2790: }
2792: PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[])
2793: {
2795: return(0);
2796: }
2798: /*
2799: Computes the number of nonzeros per row needed for preallocation when X and Y
2800: have different nonzero structure.
2801: */
2802: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz)
2803: {
2804: PetscInt i,j,k,nzx,nzy;
2807: /* Set the number of nonzeros in the new matrix */
2808: for (i=0; i<m; i++) {
2809: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2810: nzx = xi[i+1] - xi[i];
2811: nzy = yi[i+1] - yi[i];
2812: nnz[i] = 0;
2813: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2814: for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */
2815: if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */
2816: nnz[i]++;
2817: }
2818: for (; k<nzy; k++) nnz[i]++;
2819: }
2820: return(0);
2821: }
2823: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz)
2824: {
2825: PetscInt m = Y->rmap->N;
2826: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2827: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2831: /* Set the number of nonzeros in the new matrix */
2832: MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);
2833: return(0);
2834: }
2836: PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2837: {
2839: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data;
2840: PetscBLASInt one=1,bnz;
2843: PetscBLASIntCast(x->nz,&bnz);
2844: if (str == SAME_NONZERO_PATTERN) {
2845: PetscScalar alpha = a;
2846: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2847: MatSeqAIJInvalidateDiagonal(Y);
2848: PetscObjectStateIncrease((PetscObject)Y);
2849: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2850: MatAXPY_Basic(Y,a,X,str);
2851: } else {
2852: Mat B;
2853: PetscInt *nnz;
2854: PetscMalloc1(Y->rmap->N,&nnz);
2855: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2856: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2857: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2858: MatSetBlockSizesFromMats(B,Y,Y);
2859: MatSetType(B,(MatType) ((PetscObject)Y)->type_name);
2860: MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);
2861: MatSeqAIJSetPreallocation(B,0,nnz);
2862: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2863: MatHeaderReplace(Y,&B);
2864: PetscFree(nnz);
2865: }
2866: return(0);
2867: }
2869: PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
2870: {
2871: #if defined(PETSC_USE_COMPLEX)
2872: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
2873: PetscInt i,nz;
2874: PetscScalar *a;
2877: nz = aij->nz;
2878: a = aij->a;
2879: for (i=0; i<nz; i++) a[i] = PetscConj(a[i]);
2880: #else
2882: #endif
2883: return(0);
2884: }
2886: PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2887: {
2888: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2890: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2891: PetscReal atmp;
2892: PetscScalar *x;
2893: MatScalar *aa;
2896: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2897: aa = a->a;
2898: ai = a->i;
2899: aj = a->j;
2901: VecSet(v,0.0);
2902: VecGetArray(v,&x);
2903: VecGetLocalSize(v,&n);
2904: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2905: for (i=0; i<m; i++) {
2906: ncols = ai[1] - ai[0]; ai++;
2907: x[i] = 0.0;
2908: for (j=0; j<ncols; j++) {
2909: atmp = PetscAbsScalar(*aa);
2910: if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2911: aa++; aj++;
2912: }
2913: }
2914: VecRestoreArray(v,&x);
2915: return(0);
2916: }
2918: PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2919: {
2920: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2922: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2923: PetscScalar *x;
2924: MatScalar *aa;
2927: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2928: aa = a->a;
2929: ai = a->i;
2930: aj = a->j;
2932: VecSet(v,0.0);
2933: VecGetArray(v,&x);
2934: VecGetLocalSize(v,&n);
2935: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2936: for (i=0; i<m; i++) {
2937: ncols = ai[1] - ai[0]; ai++;
2938: if (ncols == A->cmap->n) { /* row is dense */
2939: x[i] = *aa; if (idx) idx[i] = 0;
2940: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
2941: x[i] = 0.0;
2942: if (idx) {
2943: idx[i] = 0; /* in case ncols is zero */
2944: for (j=0;j<ncols;j++) { /* find first implicit 0.0 in the row */
2945: if (aj[j] > j) {
2946: idx[i] = j;
2947: break;
2948: }
2949: }
2950: }
2951: }
2952: for (j=0; j<ncols; j++) {
2953: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
2954: aa++; aj++;
2955: }
2956: }
2957: VecRestoreArray(v,&x);
2958: return(0);
2959: }
2961: PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2962: {
2963: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2965: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2966: PetscReal atmp;
2967: PetscScalar *x;
2968: MatScalar *aa;
2971: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2972: aa = a->a;
2973: ai = a->i;
2974: aj = a->j;
2976: VecSet(v,0.0);
2977: VecGetArray(v,&x);
2978: VecGetLocalSize(v,&n);
2979: 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);
2980: for (i=0; i<m; i++) {
2981: ncols = ai[1] - ai[0]; ai++;
2982: if (ncols) {
2983: /* Get first nonzero */
2984: for (j = 0; j < ncols; j++) {
2985: atmp = PetscAbsScalar(aa[j]);
2986: if (atmp > 1.0e-12) {
2987: x[i] = atmp;
2988: if (idx) idx[i] = aj[j];
2989: break;
2990: }
2991: }
2992: if (j == ncols) {x[i] = PetscAbsScalar(*aa); if (idx) idx[i] = *aj;}
2993: } else {
2994: x[i] = 0.0; if (idx) idx[i] = 0;
2995: }
2996: for (j = 0; j < ncols; j++) {
2997: atmp = PetscAbsScalar(*aa);
2998: if (atmp > 1.0e-12 && PetscAbsScalar(x[i]) > atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2999: aa++; aj++;
3000: }
3001: }
3002: VecRestoreArray(v,&x);
3003: return(0);
3004: }
3006: PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3007: {
3008: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3009: PetscErrorCode ierr;
3010: PetscInt i,j,m = A->rmap->n,ncols,n;
3011: const PetscInt *ai,*aj;
3012: PetscScalar *x;
3013: const MatScalar *aa;
3016: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3017: aa = a->a;
3018: ai = a->i;
3019: aj = a->j;
3021: VecSet(v,0.0);
3022: VecGetArray(v,&x);
3023: VecGetLocalSize(v,&n);
3024: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
3025: for (i=0; i<m; i++) {
3026: ncols = ai[1] - ai[0]; ai++;
3027: if (ncols == A->cmap->n) { /* row is dense */
3028: x[i] = *aa; if (idx) idx[i] = 0;
3029: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3030: x[i] = 0.0;
3031: if (idx) { /* find first implicit 0.0 in the row */
3032: idx[i] = 0; /* in case ncols is zero */
3033: for (j=0; j<ncols; j++) {
3034: if (aj[j] > j) {
3035: idx[i] = j;
3036: break;
3037: }
3038: }
3039: }
3040: }
3041: for (j=0; j<ncols; j++) {
3042: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
3043: aa++; aj++;
3044: }
3045: }
3046: VecRestoreArray(v,&x);
3047: return(0);
3048: }
3050: PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values)
3051: {
3052: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
3053: PetscErrorCode ierr;
3054: PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j;
3055: MatScalar *diag,work[25],*v_work;
3056: const PetscReal shift = 0.0;
3057: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
3060: allowzeropivot = PetscNot(A->erroriffailure);
3061: if (a->ibdiagvalid) {
3062: if (values) *values = a->ibdiag;
3063: return(0);
3064: }
3065: MatMarkDiagonal_SeqAIJ(A);
3066: if (!a->ibdiag) {
3067: PetscMalloc1(bs2*mbs,&a->ibdiag);
3068: PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));
3069: }
3070: diag = a->ibdiag;
3071: if (values) *values = a->ibdiag;
3072: /* factor and invert each block */
3073: switch (bs) {
3074: case 1:
3075: for (i=0; i<mbs; i++) {
3076: MatGetValues(A,1,&i,1,&i,diag+i);
3077: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3078: if (allowzeropivot) {
3079: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3080: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3081: A->factorerror_zeropivot_row = i;
3082: PetscInfo3(A,"Zero pivot, row %D pivot %g tolerance %g\n",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);
3083: } 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);
3084: }
3085: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3086: }
3087: break;
3088: case 2:
3089: for (i=0; i<mbs; i++) {
3090: ij[0] = 2*i; ij[1] = 2*i + 1;
3091: MatGetValues(A,2,ij,2,ij,diag);
3092: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
3093: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3094: PetscKernel_A_gets_transpose_A_2(diag);
3095: diag += 4;
3096: }
3097: break;
3098: case 3:
3099: for (i=0; i<mbs; i++) {
3100: ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2;
3101: MatGetValues(A,3,ij,3,ij,diag);
3102: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
3103: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3104: PetscKernel_A_gets_transpose_A_3(diag);
3105: diag += 9;
3106: }
3107: break;
3108: case 4:
3109: for (i=0; i<mbs; i++) {
3110: ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3;
3111: MatGetValues(A,4,ij,4,ij,diag);
3112: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
3113: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3114: PetscKernel_A_gets_transpose_A_4(diag);
3115: diag += 16;
3116: }
3117: break;
3118: case 5:
3119: for (i=0; i<mbs; i++) {
3120: ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4;
3121: MatGetValues(A,5,ij,5,ij,diag);
3122: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
3123: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3124: PetscKernel_A_gets_transpose_A_5(diag);
3125: diag += 25;
3126: }
3127: break;
3128: case 6:
3129: for (i=0; i<mbs; i++) {
3130: 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;
3131: MatGetValues(A,6,ij,6,ij,diag);
3132: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
3133: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3134: PetscKernel_A_gets_transpose_A_6(diag);
3135: diag += 36;
3136: }
3137: break;
3138: case 7:
3139: for (i=0; i<mbs; i++) {
3140: 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;
3141: MatGetValues(A,7,ij,7,ij,diag);
3142: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
3143: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3144: PetscKernel_A_gets_transpose_A_7(diag);
3145: diag += 49;
3146: }
3147: break;
3148: default:
3149: PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);
3150: for (i=0; i<mbs; i++) {
3151: for (j=0; j<bs; j++) {
3152: IJ[j] = bs*i + j;
3153: }
3154: MatGetValues(A,bs,IJ,bs,IJ,diag);
3155: PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
3156: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3157: PetscKernel_A_gets_transpose_A_N(diag,bs);
3158: diag += bs2;
3159: }
3160: PetscFree3(v_work,v_pivots,IJ);
3161: }
3162: a->ibdiagvalid = PETSC_TRUE;
3163: return(0);
3164: }
3166: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx)
3167: {
3169: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3170: PetscScalar a;
3171: PetscInt m,n,i,j,col;
3174: if (!x->assembled) {
3175: MatGetSize(x,&m,&n);
3176: for (i=0; i<m; i++) {
3177: for (j=0; j<aij->imax[i]; j++) {
3178: PetscRandomGetValue(rctx,&a);
3179: col = (PetscInt)(n*PetscRealPart(a));
3180: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3181: }
3182: }
3183: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet coded");
3184: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3185: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3186: return(0);
3187: }
3189: PetscErrorCode MatShift_SeqAIJ(Mat Y,PetscScalar a)
3190: {
3192: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)Y->data;
3195: if (!Y->preallocated || !aij->nz) {
3196: MatSeqAIJSetPreallocation(Y,1,NULL);
3197: }
3198: MatShift_Basic(Y,a);
3199: return(0);
3200: }
3202: /* -------------------------------------------------------------------*/
3203: static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ,
3204: MatGetRow_SeqAIJ,
3205: MatRestoreRow_SeqAIJ,
3206: MatMult_SeqAIJ,
3207: /* 4*/ MatMultAdd_SeqAIJ,
3208: MatMultTranspose_SeqAIJ,
3209: MatMultTransposeAdd_SeqAIJ,
3210: 0,
3211: 0,
3212: 0,
3213: /* 10*/ 0,
3214: MatLUFactor_SeqAIJ,
3215: 0,
3216: MatSOR_SeqAIJ,
3217: MatTranspose_SeqAIJ_FAST,
3218: /*1 5*/ MatGetInfo_SeqAIJ,
3219: MatEqual_SeqAIJ,
3220: MatGetDiagonal_SeqAIJ,
3221: MatDiagonalScale_SeqAIJ,
3222: MatNorm_SeqAIJ,
3223: /* 20*/ 0,
3224: MatAssemblyEnd_SeqAIJ,
3225: MatSetOption_SeqAIJ,
3226: MatZeroEntries_SeqAIJ,
3227: /* 24*/ MatZeroRows_SeqAIJ,
3228: 0,
3229: 0,
3230: 0,
3231: 0,
3232: /* 29*/ MatSetUp_SeqAIJ,
3233: 0,
3234: 0,
3235: 0,
3236: 0,
3237: /* 34*/ MatDuplicate_SeqAIJ,
3238: 0,
3239: 0,
3240: MatILUFactor_SeqAIJ,
3241: 0,
3242: /* 39*/ MatAXPY_SeqAIJ,
3243: MatCreateSubMatrices_SeqAIJ,
3244: MatIncreaseOverlap_SeqAIJ,
3245: MatGetValues_SeqAIJ,
3246: MatCopy_SeqAIJ,
3247: /* 44*/ MatGetRowMax_SeqAIJ,
3248: MatScale_SeqAIJ,
3249: MatShift_SeqAIJ,
3250: MatDiagonalSet_SeqAIJ,
3251: MatZeroRowsColumns_SeqAIJ,
3252: /* 49*/ MatSetRandom_SeqAIJ,
3253: MatGetRowIJ_SeqAIJ,
3254: MatRestoreRowIJ_SeqAIJ,
3255: MatGetColumnIJ_SeqAIJ,
3256: MatRestoreColumnIJ_SeqAIJ,
3257: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3258: 0,
3259: 0,
3260: MatPermute_SeqAIJ,
3261: 0,
3262: /* 59*/ 0,
3263: MatDestroy_SeqAIJ,
3264: MatView_SeqAIJ,
3265: 0,
3266: MatMatMatMult_SeqAIJ_SeqAIJ_SeqAIJ,
3267: /* 64*/ MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ,
3268: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3269: 0,
3270: 0,
3271: 0,
3272: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3273: MatGetRowMinAbs_SeqAIJ,
3274: 0,
3275: 0,
3276: 0,
3277: /* 74*/ 0,
3278: MatFDColoringApply_AIJ,
3279: 0,
3280: 0,
3281: 0,
3282: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3283: 0,
3284: 0,
3285: 0,
3286: MatLoad_SeqAIJ,
3287: /* 84*/ MatIsSymmetric_SeqAIJ,
3288: MatIsHermitian_SeqAIJ,
3289: 0,
3290: 0,
3291: 0,
3292: /* 89*/ MatMatMult_SeqAIJ_SeqAIJ,
3293: MatMatMultSymbolic_SeqAIJ_SeqAIJ,
3294: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3295: MatPtAP_SeqAIJ_SeqAIJ,
3296: MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy,
3297: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3298: MatMatTransposeMult_SeqAIJ_SeqAIJ,
3299: MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ,
3300: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3301: 0,
3302: /* 99*/ 0,
3303: 0,
3304: 0,
3305: MatConjugate_SeqAIJ,
3306: 0,
3307: /*104*/ MatSetValuesRow_SeqAIJ,
3308: MatRealPart_SeqAIJ,
3309: MatImaginaryPart_SeqAIJ,
3310: 0,
3311: 0,
3312: /*109*/ MatMatSolve_SeqAIJ,
3313: 0,
3314: MatGetRowMin_SeqAIJ,
3315: 0,
3316: MatMissingDiagonal_SeqAIJ,
3317: /*114*/ 0,
3318: 0,
3319: 0,
3320: 0,
3321: 0,
3322: /*119*/ 0,
3323: 0,
3324: 0,
3325: 0,
3326: MatGetMultiProcBlock_SeqAIJ,
3327: /*124*/ MatFindNonzeroRows_SeqAIJ,
3328: MatGetColumnNorms_SeqAIJ,
3329: MatInvertBlockDiagonal_SeqAIJ,
3330: MatInvertVariableBlockDiagonal_SeqAIJ,
3331: 0,
3332: /*129*/ 0,
3333: MatTransposeMatMult_SeqAIJ_SeqAIJ,
3334: MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ,
3335: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3336: MatTransposeColoringCreate_SeqAIJ,
3337: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3338: MatTransColoringApplyDenToSp_SeqAIJ,
3339: MatRARt_SeqAIJ_SeqAIJ,
3340: MatRARtSymbolic_SeqAIJ_SeqAIJ,
3341: MatRARtNumeric_SeqAIJ_SeqAIJ,
3342: /*139*/0,
3343: 0,
3344: 0,
3345: MatFDColoringSetUp_SeqXAIJ,
3346: MatFindOffBlockDiagonalEntries_SeqAIJ,
3347: /*144*/MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3348: MatDestroySubMatrices_SeqAIJ
3349: };
3351: PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices)
3352: {
3353: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3354: PetscInt i,nz,n;
3357: nz = aij->maxnz;
3358: n = mat->rmap->n;
3359: for (i=0; i<nz; i++) {
3360: aij->j[i] = indices[i];
3361: }
3362: aij->nz = nz;
3363: for (i=0; i<n; i++) {
3364: aij->ilen[i] = aij->imax[i];
3365: }
3366: return(0);
3367: }
3369: /*@
3370: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3371: in the matrix.
3373: Input Parameters:
3374: + mat - the SeqAIJ matrix
3375: - indices - the column indices
3377: Level: advanced
3379: Notes:
3380: This can be called if you have precomputed the nonzero structure of the
3381: matrix and want to provide it to the matrix object to improve the performance
3382: of the MatSetValues() operation.
3384: You MUST have set the correct numbers of nonzeros per row in the call to
3385: MatCreateSeqAIJ(), and the columns indices MUST be sorted.
3387: MUST be called before any calls to MatSetValues();
3389: The indices should start with zero, not one.
3391: @*/
3392: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices)
3393: {
3399: PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));
3400: return(0);
3401: }
3403: /* ----------------------------------------------------------------------------------------*/
3405: PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3406: {
3407: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3409: size_t 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");
3414: /* allocate space for values if not already there */
3415: if (!aij->saved_values) {
3416: PetscMalloc1(nz+1,&aij->saved_values);
3417: PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));
3418: }
3420: /* copy values over */
3421: PetscMemcpy(aij->saved_values,aij->a,nz*sizeof(PetscScalar));
3422: return(0);
3423: }
3425: /*@
3426: MatStoreValues - Stashes a copy of the matrix values; this allows, for
3427: example, reuse of the linear part of a Jacobian, while recomputing the
3428: nonlinear portion.
3430: Collect on Mat
3432: Input Parameters:
3433: . mat - the matrix (currently only AIJ matrices support this option)
3435: Level: advanced
3437: Common Usage, with SNESSolve():
3438: $ Create Jacobian matrix
3439: $ Set linear terms into matrix
3440: $ Apply boundary conditions to matrix, at this time matrix must have
3441: $ final nonzero structure (i.e. setting the nonlinear terms and applying
3442: $ boundary conditions again will not change the nonzero structure
3443: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3444: $ MatStoreValues(mat);
3445: $ Call SNESSetJacobian() with matrix
3446: $ In your Jacobian routine
3447: $ MatRetrieveValues(mat);
3448: $ Set nonlinear terms in matrix
3450: Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself:
3451: $ // build linear portion of Jacobian
3452: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3453: $ MatStoreValues(mat);
3454: $ loop over nonlinear iterations
3455: $ MatRetrieveValues(mat);
3456: $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3457: $ // call MatAssemblyBegin/End() on matrix
3458: $ Solve linear system with Jacobian
3459: $ endloop
3461: Notes:
3462: Matrix must already be assemblied before calling this routine
3463: Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before
3464: calling this routine.
3466: When this is called multiple times it overwrites the previous set of stored values
3467: and does not allocated additional space.
3469: .seealso: MatRetrieveValues()
3471: @*/
3472: PetscErrorCode MatStoreValues(Mat mat)
3473: {
3478: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3479: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3480: PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));
3481: return(0);
3482: }
3484: PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3485: {
3486: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3488: PetscInt nz = aij->i[mat->rmap->n];
3491: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3492: if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first");
3493: /* copy values over */
3494: PetscMemcpy(aij->a,aij->saved_values,nz*sizeof(PetscScalar));
3495: return(0);
3496: }
3498: /*@
3499: MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for
3500: example, reuse of the linear part of a Jacobian, while recomputing the
3501: nonlinear portion.
3503: Collect on Mat
3505: Input Parameters:
3506: . mat - the matrix (currently only AIJ matrices support this option)
3508: Level: advanced
3510: .seealso: MatStoreValues()
3512: @*/
3513: PetscErrorCode MatRetrieveValues(Mat mat)
3514: {
3519: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3520: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3521: PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));
3522: return(0);
3523: }
3526: /* --------------------------------------------------------------------------------*/
3527: /*@C
3528: MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format
3529: (the default parallel PETSc format). For good matrix assembly performance
3530: the user should preallocate the matrix storage by setting the parameter nz
3531: (or the array nnz). By setting these parameters accurately, performance
3532: during matrix assembly can be increased by more than a factor of 50.
3534: Collective on MPI_Comm
3536: Input Parameters:
3537: + comm - MPI communicator, set to PETSC_COMM_SELF
3538: . m - number of rows
3539: . n - number of columns
3540: . nz - number of nonzeros per row (same for all rows)
3541: - nnz - array containing the number of nonzeros in the various rows
3542: (possibly different for each row) or NULL
3544: Output Parameter:
3545: . A - the matrix
3547: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3548: MatXXXXSetPreallocation() paradgm instead of this routine directly.
3549: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3551: Notes:
3552: If nnz is given then nz is ignored
3554: The AIJ format (also called the Yale sparse matrix format or
3555: compressed row storage), is fully compatible with standard Fortran 77
3556: storage. That is, the stored row and column indices can begin at
3557: either one (as in Fortran) or zero. See the users' manual for details.
3559: Specify the preallocated storage with either nz or nnz (not both).
3560: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3561: allocation. For large problems you MUST preallocate memory or you
3562: will get TERRIBLE performance, see the users' manual chapter on matrices.
3564: By default, this format uses inodes (identical nodes) when possible, to
3565: improve numerical efficiency of matrix-vector products and solves. We
3566: search for consecutive rows with the same nonzero structure, thereby
3567: reusing matrix information to achieve increased efficiency.
3569: Options Database Keys:
3570: + -mat_no_inode - Do not use inodes
3571: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3573: Level: intermediate
3575: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays()
3577: @*/
3578: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A)
3579: {
3583: MatCreate(comm,A);
3584: MatSetSizes(*A,m,n,m,n);
3585: MatSetType(*A,MATSEQAIJ);
3586: MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);
3587: return(0);
3588: }
3590: /*@C
3591: MatSeqAIJSetPreallocation - For good matrix assembly performance
3592: the user should preallocate the matrix storage by setting the parameter nz
3593: (or the array nnz). By setting these parameters accurately, performance
3594: during matrix assembly can be increased by more than a factor of 50.
3596: Collective on MPI_Comm
3598: Input Parameters:
3599: + B - The matrix
3600: . nz - number of nonzeros per row (same for all rows)
3601: - nnz - array containing the number of nonzeros in the various rows
3602: (possibly different for each row) or NULL
3604: Notes:
3605: If nnz is given then nz is ignored
3607: The AIJ format (also called the Yale sparse matrix format or
3608: compressed row storage), is fully compatible with standard Fortran 77
3609: storage. That is, the stored row and column indices can begin at
3610: either one (as in Fortran) or zero. See the users' manual for details.
3612: Specify the preallocated storage with either nz or nnz (not both).
3613: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3614: allocation. For large problems you MUST preallocate memory or you
3615: will get TERRIBLE performance, see the users' manual chapter on matrices.
3617: You can call MatGetInfo() to get information on how effective the preallocation was;
3618: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3619: You can also run with the option -info and look for messages with the string
3620: malloc in them to see if additional memory allocation was needed.
3622: Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix
3623: entries or columns indices
3625: By default, this format uses inodes (identical nodes) when possible, to
3626: improve numerical efficiency of matrix-vector products and solves. We
3627: search for consecutive rows with the same nonzero structure, thereby
3628: reusing matrix information to achieve increased efficiency.
3630: Options Database Keys:
3631: + -mat_no_inode - Do not use inodes
3632: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3634: Level: intermediate
3636: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo()
3638: @*/
3639: PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[])
3640: {
3646: PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));
3647: return(0);
3648: }
3650: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz)
3651: {
3652: Mat_SeqAIJ *b;
3653: PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE;
3655: PetscInt i;
3658: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3659: if (nz == MAT_SKIP_ALLOCATION) {
3660: skipallocation = PETSC_TRUE;
3661: nz = 0;
3662: }
3663: PetscLayoutSetUp(B->rmap);
3664: PetscLayoutSetUp(B->cmap);
3666: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3667: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz);
3668: if (nnz) {
3669: for (i=0; i<B->rmap->n; i++) {
3670: 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]);
3671: 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);
3672: }
3673: }
3675: B->preallocated = PETSC_TRUE;
3677: b = (Mat_SeqAIJ*)B->data;
3679: if (!skipallocation) {
3680: if (!b->imax) {
3681: PetscMalloc2(B->rmap->n,&b->imax,B->rmap->n,&b->ilen);
3682: PetscLogObjectMemory((PetscObject)B,2*B->rmap->n*sizeof(PetscInt));
3683: }
3684: if (!b->ipre) {
3685: PetscMalloc1(B->rmap->n,&b->ipre);
3686: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
3687: }
3688: if (!nnz) {
3689: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3690: else if (nz < 0) nz = 1;
3691: for (i=0; i<B->rmap->n; i++) b->imax[i] = nz;
3692: nz = nz*B->rmap->n;
3693: } else {
3694: nz = 0;
3695: for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz += nnz[i];}
3696: }
3697: /* b->ilen will count nonzeros in each row so far. */
3698: for (i=0; i<B->rmap->n; i++) b->ilen[i] = 0;
3700: /* allocate the matrix space */
3701: /* FIXME: should B's old memory be unlogged? */
3702: MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);
3703: if (B->structure_only) {
3704: PetscMalloc1(nz,&b->j);
3705: PetscMalloc1(B->rmap->n+1,&b->i);
3706: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));
3707: } else {
3708: PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);
3709: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));
3710: }
3711: b->i[0] = 0;
3712: for (i=1; i<B->rmap->n+1; i++) {
3713: b->i[i] = b->i[i-1] + b->imax[i-1];
3714: }
3715: if (B->structure_only) {
3716: b->singlemalloc = PETSC_FALSE;
3717: b->free_a = PETSC_FALSE;
3718: } else {
3719: b->singlemalloc = PETSC_TRUE;
3720: b->free_a = PETSC_TRUE;
3721: }
3722: b->free_ij = PETSC_TRUE;
3723: } else {
3724: b->free_a = PETSC_FALSE;
3725: b->free_ij = PETSC_FALSE;
3726: }
3728: if (b->ipre && nnz != b->ipre && b->imax) {
3729: /* reserve user-requested sparsity */
3730: PetscMemcpy(b->ipre,b->imax,B->rmap->n*sizeof(PetscInt));
3731: }
3734: b->nz = 0;
3735: b->maxnz = nz;
3736: B->info.nz_unneeded = (double)b->maxnz;
3737: if (realalloc) {
3738: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
3739: }
3740: B->was_assembled = PETSC_FALSE;
3741: B->assembled = PETSC_FALSE;
3742: return(0);
3743: }
3746: PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
3747: {
3748: Mat_SeqAIJ *a;
3749: PetscInt i;
3754: a = (Mat_SeqAIJ*)A->data;
3755: /* if no saved info, we error out */
3756: if (!a->ipre) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_ARG_NULL,"No saved preallocation info \n");
3758: 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");
3760: PetscMemcpy(a->imax,a->ipre,A->rmap->n*sizeof(PetscInt));
3761: PetscMemzero(a->ilen,A->rmap->n*sizeof(PetscInt));
3762: a->i[0] = 0;
3763: for (i=1; i<A->rmap->n+1; i++) {
3764: a->i[i] = a->i[i-1] + a->imax[i-1];
3765: }
3766: A->preallocated = PETSC_TRUE;
3767: a->nz = 0;
3768: a->maxnz = a->i[A->rmap->n];
3769: A->info.nz_unneeded = (double)a->maxnz;
3770: A->was_assembled = PETSC_FALSE;
3771: A->assembled = PETSC_FALSE;
3772: return(0);
3773: }
3775: /*@
3776: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format.
3778: Input Parameters:
3779: + B - the matrix
3780: . i - the indices into j for the start of each row (starts with zero)
3781: . j - the column indices for each row (starts with zero) these must be sorted for each row
3782: - v - optional values in the matrix
3784: Level: developer
3786: The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays()
3788: .keywords: matrix, aij, compressed row, sparse, sequential
3790: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), MATSEQAIJ
3791: @*/
3792: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[])
3793: {
3799: PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3800: return(0);
3801: }
3803: PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3804: {
3805: PetscInt i;
3806: PetscInt m,n;
3807: PetscInt nz;
3808: PetscInt *nnz, nz_max = 0;
3809: PetscScalar *values;
3813: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]);
3815: PetscLayoutSetUp(B->rmap);
3816: PetscLayoutSetUp(B->cmap);
3818: MatGetSize(B, &m, &n);
3819: PetscMalloc1(m+1, &nnz);
3820: for (i = 0; i < m; i++) {
3821: nz = Ii[i+1]- Ii[i];
3822: nz_max = PetscMax(nz_max, nz);
3823: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz);
3824: nnz[i] = nz;
3825: }
3826: MatSeqAIJSetPreallocation(B, 0, nnz);
3827: PetscFree(nnz);
3829: if (v) {
3830: values = (PetscScalar*) v;
3831: } else {
3832: PetscCalloc1(nz_max, &values);
3833: }
3835: for (i = 0; i < m; i++) {
3836: nz = Ii[i+1] - Ii[i];
3837: MatSetValues_SeqAIJ(B, 1, &i, nz, J+Ii[i], values + (v ? Ii[i] : 0), INSERT_VALUES);
3838: }
3840: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3841: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3843: if (!v) {
3844: PetscFree(values);
3845: }
3846: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3847: return(0);
3848: }
3850: #include <../src/mat/impls/dense/seq/dense.h>
3851: #include <petsc/private/kernels/petscaxpy.h>
3853: /*
3854: Computes (B'*A')' since computing B*A directly is untenable
3856: n p p
3857: ( ) ( ) ( )
3858: m ( A ) * n ( B ) = m ( C )
3859: ( ) ( ) ( )
3861: */
3862: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C)
3863: {
3864: PetscErrorCode ierr;
3865: Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data;
3866: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data;
3867: Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data;
3868: PetscInt i,n,m,q,p;
3869: const PetscInt *ii,*idx;
3870: const PetscScalar *b,*a,*a_q;
3871: PetscScalar *c,*c_q;
3874: m = A->rmap->n;
3875: n = A->cmap->n;
3876: p = B->cmap->n;
3877: a = sub_a->v;
3878: b = sub_b->a;
3879: c = sub_c->v;
3880: PetscMemzero(c,m*p*sizeof(PetscScalar));
3882: ii = sub_b->i;
3883: idx = sub_b->j;
3884: for (i=0; i<n; i++) {
3885: q = ii[i+1] - ii[i];
3886: while (q-->0) {
3887: c_q = c + m*(*idx);
3888: a_q = a + m*i;
3889: PetscKernelAXPY(c_q,*b,a_q,m);
3890: idx++;
3891: b++;
3892: }
3893: }
3894: return(0);
3895: }
3897: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
3898: {
3900: PetscInt m=A->rmap->n,n=B->cmap->n;
3901: Mat Cmat;
3904: 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);
3905: MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
3906: MatSetSizes(Cmat,m,n,m,n);
3907: MatSetBlockSizesFromMats(Cmat,A,B);
3908: MatSetType(Cmat,MATSEQDENSE);
3909: MatSeqDenseSetPreallocation(Cmat,NULL);
3911: Cmat->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
3913: *C = Cmat;
3914: return(0);
3915: }
3917: /* ----------------------------------------------------------------*/
3918: PETSC_INTERN PetscErrorCode MatMatMult_SeqDense_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3919: {
3923: if (scall == MAT_INITIAL_MATRIX) {
3924: PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
3925: MatMatMultSymbolic_SeqDense_SeqAIJ(A,B,fill,C);
3926: PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
3927: }
3928: PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
3929: MatMatMultNumeric_SeqDense_SeqAIJ(A,B,*C);
3930: PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
3931: return(0);
3932: }
3935: /*MC
3936: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
3937: based on compressed sparse row format.
3939: Options Database Keys:
3940: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
3942: Level: beginner
3944: .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType
3945: M*/
3947: /*MC
3948: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
3950: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
3951: and MATMPIAIJ otherwise. As a result, for single process communicators,
3952: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
3953: for communicators controlling multiple processes. It is recommended that you call both of
3954: the above preallocation routines for simplicity.
3956: Options Database Keys:
3957: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
3959: Developer Notes:
3960: Subclasses include MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
3961: enough exist.
3963: Level: beginner
3965: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
3966: M*/
3968: /*MC
3969: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
3971: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
3972: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
3973: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
3974: for communicators controlling multiple processes. It is recommended that you call both of
3975: the above preallocation routines for simplicity.
3977: Options Database Keys:
3978: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
3980: Level: beginner
3982: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
3983: M*/
3985: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*);
3986: #if defined(PETSC_HAVE_ELEMENTAL)
3987: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
3988: #endif
3989: #if defined(PETSC_HAVE_HYPRE)
3990: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A,MatType,MatReuse,Mat*);
3991: PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*);
3992: #endif
3993: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*);
3995: #if defined(PETSC_HAVE_MATLAB_ENGINE)
3996: PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject,void*);
3997: PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject,void*);
3998: #endif
4000: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat,MatType,MatReuse,Mat*);
4001: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
4002: PETSC_INTERN PetscErrorCode MatPtAP_IS_XAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
4004: /*@C
4005: MatSeqAIJGetArray - gives access to the array where the data for a MATSEQAIJ matrix is stored
4007: Not Collective
4009: Input Parameter:
4010: . mat - a MATSEQAIJ matrix
4012: Output Parameter:
4013: . array - pointer to the data
4015: Level: intermediate
4017: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
4018: @*/
4019: PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array)
4020: {
4024: PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));
4025: return(0);
4026: }
4028: /*@C
4029: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4031: Not Collective
4033: Input Parameter:
4034: . mat - a MATSEQAIJ matrix
4036: Output Parameter:
4037: . nz - the maximum number of nonzeros in any row
4039: Level: intermediate
4041: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
4042: @*/
4043: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz)
4044: {
4045: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
4048: *nz = aij->rmax;
4049: return(0);
4050: }
4052: /*@C
4053: MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray()
4055: Not Collective
4057: Input Parameters:
4058: . mat - a MATSEQAIJ matrix
4059: . array - pointer to the data
4061: Level: intermediate
4063: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90()
4064: @*/
4065: PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array)
4066: {
4070: PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));
4071: return(0);
4072: }
4074: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4075: {
4076: Mat_SeqAIJ *b;
4078: PetscMPIInt size;
4081: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
4082: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1");
4084: PetscNewLog(B,&b);
4086: B->data = (void*)b;
4088: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
4090: b->row = 0;
4091: b->col = 0;
4092: b->icol = 0;
4093: b->reallocs = 0;
4094: b->ignorezeroentries = PETSC_FALSE;
4095: b->roworiented = PETSC_TRUE;
4096: b->nonew = 0;
4097: b->diag = 0;
4098: b->solve_work = 0;
4099: B->spptr = 0;
4100: b->saved_values = 0;
4101: b->idiag = 0;
4102: b->mdiag = 0;
4103: b->ssor_work = 0;
4104: b->omega = 1.0;
4105: b->fshift = 0.0;
4106: b->idiagvalid = PETSC_FALSE;
4107: b->ibdiagvalid = PETSC_FALSE;
4108: b->keepnonzeropattern = PETSC_FALSE;
4110: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4111: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);
4112: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);
4114: #if defined(PETSC_HAVE_MATLAB_ENGINE)
4115: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);
4116: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);
4117: #endif
4119: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);
4120: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);
4121: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);
4122: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);
4123: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);
4124: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);
4125: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijsell_C",MatConvert_SeqAIJ_SeqAIJSELL);
4126: #if defined(PETSC_HAVE_MKL_SPARSE)
4127: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijmkl_C",MatConvert_SeqAIJ_SeqAIJMKL);
4128: #endif
4129: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);
4130: #if defined(PETSC_HAVE_ELEMENTAL)
4131: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);
4132: #endif
4133: #if defined(PETSC_HAVE_HYPRE)
4134: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_hypre_C",MatConvert_AIJ_HYPRE);
4135: PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_seqaij_seqaij_C",MatMatMatMult_Transpose_AIJ_AIJ);
4136: #endif
4137: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);
4138: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsell_C",MatConvert_SeqAIJ_SeqSELL);
4139: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_is_C",MatConvert_XAIJ_IS);
4140: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);
4141: PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);
4142: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);
4143: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_SeqAIJ);
4144: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);
4145: PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);
4146: PetscObjectComposeFunction((PetscObject)B,"MatMatMult_seqdense_seqaij_C",MatMatMult_SeqDense_SeqAIJ);
4147: PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_seqdense_seqaij_C",MatMatMultSymbolic_SeqDense_SeqAIJ);
4148: PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_seqdense_seqaij_C",MatMatMultNumeric_SeqDense_SeqAIJ);
4149: PetscObjectComposeFunction((PetscObject)B,"MatPtAP_is_seqaij_C",MatPtAP_IS_XAIJ);
4150: MatCreate_SeqAIJ_Inode(B);
4151: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4152: MatSeqAIJSetTypeFromOptions(B); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4153: return(0);
4154: }
4156: /*
4157: Given a matrix generated with MatGetFactor() duplicates all the information in A into B
4158: */
4159: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace)
4160: {
4161: Mat_SeqAIJ *c,*a = (Mat_SeqAIJ*)A->data;
4163: PetscInt i,m = A->rmap->n;
4166: c = (Mat_SeqAIJ*)C->data;
4168: C->factortype = A->factortype;
4169: c->row = 0;
4170: c->col = 0;
4171: c->icol = 0;
4172: c->reallocs = 0;
4174: C->assembled = PETSC_TRUE;
4176: PetscLayoutReference(A->rmap,&C->rmap);
4177: PetscLayoutReference(A->cmap,&C->cmap);
4179: PetscMalloc2(m,&c->imax,m,&c->ilen);
4180: PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));
4181: for (i=0; i<m; i++) {
4182: c->imax[i] = a->imax[i];
4183: c->ilen[i] = a->ilen[i];
4184: }
4186: /* allocate the matrix space */
4187: if (mallocmatspace) {
4188: PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);
4189: PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));
4191: c->singlemalloc = PETSC_TRUE;
4193: PetscMemcpy(c->i,a->i,(m+1)*sizeof(PetscInt));
4194: if (m > 0) {
4195: PetscMemcpy(c->j,a->j,(a->i[m])*sizeof(PetscInt));
4196: if (cpvalues == MAT_COPY_VALUES) {
4197: PetscMemcpy(c->a,a->a,(a->i[m])*sizeof(PetscScalar));
4198: } else {
4199: PetscMemzero(c->a,(a->i[m])*sizeof(PetscScalar));
4200: }
4201: }
4202: }
4204: c->ignorezeroentries = a->ignorezeroentries;
4205: c->roworiented = a->roworiented;
4206: c->nonew = a->nonew;
4207: if (a->diag) {
4208: PetscMalloc1(m+1,&c->diag);
4209: PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));
4210: for (i=0; i<m; i++) {
4211: c->diag[i] = a->diag[i];
4212: }
4213: } else c->diag = 0;
4215: c->solve_work = 0;
4216: c->saved_values = 0;
4217: c->idiag = 0;
4218: c->ssor_work = 0;
4219: c->keepnonzeropattern = a->keepnonzeropattern;
4220: c->free_a = PETSC_TRUE;
4221: c->free_ij = PETSC_TRUE;
4223: c->rmax = a->rmax;
4224: c->nz = a->nz;
4225: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4226: C->preallocated = PETSC_TRUE;
4228: c->compressedrow.use = a->compressedrow.use;
4229: c->compressedrow.nrows = a->compressedrow.nrows;
4230: if (a->compressedrow.use) {
4231: i = a->compressedrow.nrows;
4232: PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);
4233: PetscMemcpy(c->compressedrow.i,a->compressedrow.i,(i+1)*sizeof(PetscInt));
4234: PetscMemcpy(c->compressedrow.rindex,a->compressedrow.rindex,i*sizeof(PetscInt));
4235: } else {
4236: c->compressedrow.use = PETSC_FALSE;
4237: c->compressedrow.i = NULL;
4238: c->compressedrow.rindex = NULL;
4239: }
4240: c->nonzerorowcnt = a->nonzerorowcnt;
4241: C->nonzerostate = A->nonzerostate;
4243: MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);
4244: PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);
4245: return(0);
4246: }
4248: PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B)
4249: {
4253: MatCreate(PetscObjectComm((PetscObject)A),B);
4254: MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);
4255: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) {
4256: MatSetBlockSizesFromMats(*B,A,A);
4257: }
4258: MatSetType(*B,((PetscObject)A)->type_name);
4259: MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);
4260: return(0);
4261: }
4263: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4264: {
4265: Mat_SeqAIJ *a;
4267: PetscInt i,sum,nz,header[4],*rowlengths = 0,M,N,rows,cols;
4268: int fd;
4269: PetscMPIInt size;
4270: MPI_Comm comm;
4271: PetscInt bs = newMat->rmap->bs;
4274: /* force binary viewer to load .info file if it has not yet done so */
4275: PetscViewerSetUp(viewer);
4276: PetscObjectGetComm((PetscObject)viewer,&comm);
4277: MPI_Comm_size(comm,&size);
4278: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"view must have one processor");
4280: PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");
4281: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
4282: PetscOptionsEnd();
4283: if (bs < 0) bs = 1;
4284: MatSetBlockSize(newMat,bs);
4286: PetscViewerBinaryGetDescriptor(viewer,&fd);
4287: PetscBinaryRead(fd,header,4,PETSC_INT);
4288: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object in file");
4289: M = header[1]; N = header[2]; nz = header[3];
4291: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as SeqAIJ");
4293: /* read in row lengths */
4294: PetscMalloc1(M,&rowlengths);
4295: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
4297: /* check if sum of rowlengths is same as nz */
4298: for (i=0,sum=0; i< M; i++) sum +=rowlengths[i];
4299: 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);
4301: /* set global size if not set already*/
4302: if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) {
4303: MatSetSizes(newMat,PETSC_DECIDE,PETSC_DECIDE,M,N);
4304: } else {
4305: /* if sizes and type are already set, check if the matrix global sizes are correct */
4306: MatGetSize(newMat,&rows,&cols);
4307: if (rows < 0 && cols < 0) { /* user might provide local size instead of global size */
4308: MatGetLocalSize(newMat,&rows,&cols);
4309: }
4310: 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);
4311: }
4312: MatSeqAIJSetPreallocation_SeqAIJ(newMat,0,rowlengths);
4313: a = (Mat_SeqAIJ*)newMat->data;
4315: PetscBinaryRead(fd,a->j,nz,PETSC_INT);
4317: /* read in nonzero values */
4318: PetscBinaryRead(fd,a->a,nz,PETSC_SCALAR);
4320: /* set matrix "i" values */
4321: a->i[0] = 0;
4322: for (i=1; i<= M; i++) {
4323: a->i[i] = a->i[i-1] + rowlengths[i-1];
4324: a->ilen[i-1] = rowlengths[i-1];
4325: }
4326: PetscFree(rowlengths);
4328: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
4329: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
4330: return(0);
4331: }
4333: PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg)
4334: {
4335: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
4337: #if defined(PETSC_USE_COMPLEX)
4338: PetscInt k;
4339: #endif
4342: /* If the matrix dimensions are not equal,or no of nonzeros */
4343: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) {
4344: *flg = PETSC_FALSE;
4345: return(0);
4346: }
4348: /* if the a->i are the same */
4349: PetscMemcmp(a->i,b->i,(A->rmap->n+1)*sizeof(PetscInt),flg);
4350: if (!*flg) return(0);
4352: /* if a->j are the same */
4353: PetscMemcmp(a->j,b->j,(a->nz)*sizeof(PetscInt),flg);
4354: if (!*flg) return(0);
4356: /* if a->a are the same */
4357: #if defined(PETSC_USE_COMPLEX)
4358: for (k=0; k<a->nz; k++) {
4359: if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) {
4360: *flg = PETSC_FALSE;
4361: return(0);
4362: }
4363: }
4364: #else
4365: PetscMemcmp(a->a,b->a,(a->nz)*sizeof(PetscScalar),flg);
4366: #endif
4367: return(0);
4368: }
4370: /*@
4371: MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format)
4372: provided by the user.
4374: Collective on MPI_Comm
4376: Input Parameters:
4377: + comm - must be an MPI communicator of size 1
4378: . m - number of rows
4379: . n - number of columns
4380: . i - row indices
4381: . j - column indices
4382: - a - matrix values
4384: Output Parameter:
4385: . mat - the matrix
4387: Level: intermediate
4389: Notes:
4390: The i, j, and a arrays are not copied by this routine, the user must free these arrays
4391: once the matrix is destroyed and not before
4393: You cannot set new nonzero locations into this matrix, that will generate an error.
4395: The i and j indices are 0 based
4397: The format which is used for the sparse matrix input, is equivalent to a
4398: row-major ordering.. i.e for the following matrix, the input data expected is
4399: as shown
4401: $ 1 0 0
4402: $ 2 0 3
4403: $ 4 5 6
4404: $
4405: $ i = {0,1,3,6} [size = nrow+1 = 3+1]
4406: $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
4407: $ v = {1,2,3,4,5,6} [size = 6]
4410: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4412: @*/
4413: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat)
4414: {
4416: PetscInt ii;
4417: Mat_SeqAIJ *aij;
4418: #if defined(PETSC_USE_DEBUG)
4419: PetscInt jj;
4420: #endif
4423: if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4424: MatCreate(comm,mat);
4425: MatSetSizes(*mat,m,n,m,n);
4426: /* MatSetBlockSizes(*mat,,); */
4427: MatSetType(*mat,MATSEQAIJ);
4428: MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,0);
4429: aij = (Mat_SeqAIJ*)(*mat)->data;
4430: PetscMalloc2(m,&aij->imax,m,&aij->ilen);
4432: aij->i = i;
4433: aij->j = j;
4434: aij->a = a;
4435: aij->singlemalloc = PETSC_FALSE;
4436: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
4437: aij->free_a = PETSC_FALSE;
4438: aij->free_ij = PETSC_FALSE;
4440: for (ii=0; ii<m; ii++) {
4441: aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii];
4442: #if defined(PETSC_USE_DEBUG)
4443: 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]);
4444: for (jj=i[ii]+1; jj<i[ii+1]; jj++) {
4445: 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);
4446: 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);
4447: }
4448: #endif
4449: }
4450: #if defined(PETSC_USE_DEBUG)
4451: for (ii=0; ii<aij->i[m]; ii++) {
4452: if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]);
4453: 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]);
4454: }
4455: #endif
4457: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4458: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4459: return(0);
4460: }
4461: /*@C
4462: MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format)
4463: provided by the user.
4465: Collective on MPI_Comm
4467: Input Parameters:
4468: + comm - must be an MPI communicator of size 1
4469: . m - number of rows
4470: . n - number of columns
4471: . i - row indices
4472: . j - column indices
4473: . a - matrix values
4474: . nz - number of nonzeros
4475: - idx - 0 or 1 based
4477: Output Parameter:
4478: . mat - the matrix
4480: Level: intermediate
4482: Notes:
4483: The i and j indices are 0 based
4485: The format which is used for the sparse matrix input, is equivalent to a
4486: row-major ordering.. i.e for the following matrix, the input data expected is
4487: as shown:
4489: 1 0 0
4490: 2 0 3
4491: 4 5 6
4493: i = {0,1,1,2,2,2}
4494: j = {0,0,2,0,1,2}
4495: v = {1,2,3,4,5,6}
4498: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4500: @*/
4501: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat,PetscInt nz,PetscBool idx)
4502: {
4504: PetscInt ii, *nnz, one = 1,row,col;
4508: PetscCalloc1(m,&nnz);
4509: for (ii = 0; ii < nz; ii++) {
4510: nnz[i[ii] - !!idx] += 1;
4511: }
4512: MatCreate(comm,mat);
4513: MatSetSizes(*mat,m,n,m,n);
4514: MatSetType(*mat,MATSEQAIJ);
4515: MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);
4516: for (ii = 0; ii < nz; ii++) {
4517: if (idx) {
4518: row = i[ii] - 1;
4519: col = j[ii] - 1;
4520: } else {
4521: row = i[ii];
4522: col = j[ii];
4523: }
4524: MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);
4525: }
4526: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4527: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4528: PetscFree(nnz);
4529: return(0);
4530: }
4532: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
4533: {
4534: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data;
4538: a->idiagvalid = PETSC_FALSE;
4539: a->ibdiagvalid = PETSC_FALSE;
4541: MatSeqAIJInvalidateDiagonal_Inode(A);
4542: return(0);
4543: }
4545: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4546: {
4548: PetscMPIInt size;
4551: MPI_Comm_size(comm,&size);
4552: if (size == 1) {
4553: if (scall == MAT_INITIAL_MATRIX) {
4554: MatDuplicate(inmat,MAT_COPY_VALUES,outmat);
4555: } else {
4556: MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);
4557: }
4558: } else {
4559: MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);
4560: }
4561: return(0);
4562: }
4564: /*
4565: Permute A into C's *local* index space using rowemb,colemb.
4566: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
4567: of [0,m), colemb is in [0,n).
4568: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
4569: */
4570: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B)
4571: {
4572: /* If making this function public, change the error returned in this function away from _PLIB. */
4574: Mat_SeqAIJ *Baij;
4575: PetscBool seqaij;
4576: PetscInt m,n,*nz,i,j,count;
4577: PetscScalar v;
4578: const PetscInt *rowindices,*colindices;
4581: if (!B) return(0);
4582: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
4583: PetscObjectBaseTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);
4584: if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type");
4585: if (rowemb) {
4586: ISGetLocalSize(rowemb,&m);
4587: 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);
4588: } else {
4589: if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix");
4590: }
4591: if (colemb) {
4592: ISGetLocalSize(colemb,&n);
4593: 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);
4594: } else {
4595: if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix");
4596: }
4598: Baij = (Mat_SeqAIJ*)(B->data);
4599: if (pattern == DIFFERENT_NONZERO_PATTERN) {
4600: PetscMalloc1(B->rmap->n,&nz);
4601: for (i=0; i<B->rmap->n; i++) {
4602: nz[i] = Baij->i[i+1] - Baij->i[i];
4603: }
4604: MatSeqAIJSetPreallocation(C,0,nz);
4605: PetscFree(nz);
4606: }
4607: if (pattern == SUBSET_NONZERO_PATTERN) {
4608: MatZeroEntries(C);
4609: }
4610: count = 0;
4611: rowindices = NULL;
4612: colindices = NULL;
4613: if (rowemb) {
4614: ISGetIndices(rowemb,&rowindices);
4615: }
4616: if (colemb) {
4617: ISGetIndices(colemb,&colindices);
4618: }
4619: for (i=0; i<B->rmap->n; i++) {
4620: PetscInt row;
4621: row = i;
4622: if (rowindices) row = rowindices[i];
4623: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
4624: PetscInt col;
4625: col = Baij->j[count];
4626: if (colindices) col = colindices[col];
4627: v = Baij->a[count];
4628: MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);
4629: ++count;
4630: }
4631: }
4632: /* FIXME: set C's nonzerostate correctly. */
4633: /* Assembly for C is necessary. */
4634: C->preallocated = PETSC_TRUE;
4635: C->assembled = PETSC_TRUE;
4636: C->was_assembled = PETSC_FALSE;
4637: return(0);
4638: }
4640: PetscFunctionList MatSeqAIJList = NULL;
4642: /*@C
4643: MatSeqAIJSetType - Converts a MATSEQAIJ matrix to a subtype
4645: Collective on Mat
4647: Input Parameters:
4648: + mat - the matrix object
4649: - matype - matrix type
4651: Options Database Key:
4652: . -mat_seqai_type <method> - for example seqaijcrl
4655: Level: intermediate
4657: .keywords: Mat, MatType, set, method
4659: .seealso: PCSetType(), VecSetType(), MatCreate(), MatType, Mat
4660: @*/
4661: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
4662: {
4663: PetscErrorCode ierr,(*r)(Mat,MatType,MatReuse,Mat*);
4664: PetscBool sametype;
4668: PetscObjectTypeCompare((PetscObject)mat,matype,&sametype);
4669: if (sametype) return(0);
4671: PetscFunctionListFind(MatSeqAIJList,matype,&r);
4672: if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown Mat type given: %s",matype);
4673: (*r)(mat,matype,MAT_INPLACE_MATRIX,&mat);
4674: return(0);
4675: }
4678: /*@C
4679: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential AIJ matrices
4681: Not Collective
4683: Input Parameters:
4684: + name - name of a new user-defined matrix type, for example MATSEQAIJCRL
4685: - function - routine to convert to subtype
4687: Notes:
4688: MatSeqAIJRegister() may be called multiple times to add several user-defined solvers.
4691: Then, your matrix can be chosen with the procedural interface at runtime via the option
4692: $ -mat_seqaij_type my_mat
4694: Level: advanced
4696: .keywords: Mat, register
4698: .seealso: MatSeqAIJRegisterAll()
4701: Level: advanced
4702: @*/
4703: PetscErrorCode MatSeqAIJRegister(const char sname[],PetscErrorCode (*function)(Mat,MatType,MatReuse,Mat *))
4704: {
4708: MatInitializePackage();
4709: PetscFunctionListAdd(&MatSeqAIJList,sname,function);
4710: return(0);
4711: }
4713: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
4715: /*@C
4716: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of SeqAIJ
4718: Not Collective
4720: Level: advanced
4722: Developers Note: CUSP and CUSPARSE do not yet support the MatConvert_SeqAIJ..() paradigm and thus cannot be registered here
4724: .keywords: KSP, register, all
4726: .seealso: MatRegisterAll(), MatSeqAIJRegister()
4727: @*/
4728: PetscErrorCode MatSeqAIJRegisterAll(void)
4729: {
4733: if (MatSeqAIJRegisterAllCalled) return(0);
4734: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
4736: MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);
4737: MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);
4738: MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL);
4739: #if defined(PETSC_HAVE_MKL_SPARSE)
4740: MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);
4741: #endif
4742: #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA)
4743: MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);
4744: #endif
4745: return(0);
4746: }
4748: /*
4749: Special version for direct calls from Fortran
4750: */
4751: #include <petsc/private/fortranimpl.h>
4752: #if defined(PETSC_HAVE_FORTRAN_CAPS)
4753: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
4754: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
4755: #define matsetvaluesseqaij_ matsetvaluesseqaij
4756: #endif
4758: /* Change these macros so can be used in void function */
4759: #undef CHKERRQ
4760: #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr)
4761: #undef SETERRQ2
4762: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
4763: #undef SETERRQ3
4764: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
4766: 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)
4767: {
4768: Mat A = *AA;
4769: PetscInt m = *mm, n = *nn;
4770: InsertMode is = *isis;
4771: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
4772: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
4773: PetscInt *imax,*ai,*ailen;
4775: PetscInt *aj,nonew = a->nonew,lastcol = -1;
4776: MatScalar *ap,value,*aa;
4777: PetscBool ignorezeroentries = a->ignorezeroentries;
4778: PetscBool roworiented = a->roworiented;
4781: MatCheckPreallocated(A,1);
4782: imax = a->imax;
4783: ai = a->i;
4784: ailen = a->ilen;
4785: aj = a->j;
4786: aa = a->a;
4788: for (k=0; k<m; k++) { /* loop over added rows */
4789: row = im[k];
4790: if (row < 0) continue;
4791: #if defined(PETSC_USE_DEBUG)
4792: if (row >= A->rmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
4793: #endif
4794: rp = aj + ai[row]; ap = aa + ai[row];
4795: rmax = imax[row]; nrow = ailen[row];
4796: low = 0;
4797: high = nrow;
4798: for (l=0; l<n; l++) { /* loop over added columns */
4799: if (in[l] < 0) continue;
4800: #if defined(PETSC_USE_DEBUG)
4801: if (in[l] >= A->cmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
4802: #endif
4803: col = in[l];
4804: if (roworiented) value = v[l + k*n];
4805: else value = v[k + l*m];
4807: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
4809: if (col <= lastcol) low = 0;
4810: else high = nrow;
4811: lastcol = col;
4812: while (high-low > 5) {
4813: t = (low+high)/2;
4814: if (rp[t] > col) high = t;
4815: else low = t;
4816: }
4817: for (i=low; i<high; i++) {
4818: if (rp[i] > col) break;
4819: if (rp[i] == col) {
4820: if (is == ADD_VALUES) ap[i] += value;
4821: else ap[i] = value;
4822: goto noinsert;
4823: }
4824: }
4825: if (value == 0.0 && ignorezeroentries) goto noinsert;
4826: if (nonew == 1) goto noinsert;
4827: if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix");
4828: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
4829: N = nrow++ - 1; a->nz++; high++;
4830: /* shift up all the later entries in this row */
4831: for (ii=N; ii>=i; ii--) {
4832: rp[ii+1] = rp[ii];
4833: ap[ii+1] = ap[ii];
4834: }
4835: rp[i] = col;
4836: ap[i] = value;
4837: A->nonzerostate++;
4838: noinsert:;
4839: low = i + 1;
4840: }
4841: ailen[row] = nrow;
4842: }
4843: PetscFunctionReturnVoid();
4844: }