Actual source code: aij.c
petsc-3.14.6 2021-03-30
1: /*
2: Defines the basic matrix operations for the AIJ (compressed row)
3: matrix storage format.
4: */
7: #include <../src/mat/impls/aij/seq/aij.h>
8: #include <petscblaslapack.h>
9: #include <petscbt.h>
10: #include <petsc/private/kernels/blocktranspose.h>
12: PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A)
13: {
14: PetscErrorCode ierr;
15: PetscBool flg;
16: char type[256];
19: PetscObjectOptionsBegin((PetscObject)A);
20: PetscOptionsFList("-mat_seqaij_type","Matrix SeqAIJ type","MatSeqAIJSetType",MatSeqAIJList,"seqaij",type,256,&flg);
21: if (flg) {
22: MatSeqAIJSetType(A,type);
23: }
24: PetscOptionsEnd();
25: return(0);
26: }
28: PetscErrorCode MatGetColumnNorms_SeqAIJ(Mat A,NormType type,PetscReal *norms)
29: {
31: PetscInt i,m,n;
32: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
35: MatGetSize(A,&m,&n);
36: PetscArrayzero(norms,n);
37: if (type == NORM_2) {
38: for (i=0; i<aij->i[m]; i++) {
39: norms[aij->j[i]] += PetscAbsScalar(aij->a[i]*aij->a[i]);
40: }
41: } else if (type == NORM_1) {
42: for (i=0; i<aij->i[m]; i++) {
43: norms[aij->j[i]] += PetscAbsScalar(aij->a[i]);
44: }
45: } else if (type == NORM_INFINITY) {
46: for (i=0; i<aij->i[m]; i++) {
47: norms[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]),norms[aij->j[i]]);
48: }
49: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown NormType");
51: if (type == NORM_2) {
52: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
53: }
54: return(0);
55: }
57: PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A,IS *is)
58: {
59: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
60: PetscInt i,m=A->rmap->n,cnt = 0, bs = A->rmap->bs;
61: const PetscInt *jj = a->j,*ii = a->i;
62: PetscInt *rows;
63: PetscErrorCode ierr;
66: for (i=0; i<m; i++) {
67: if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) {
68: cnt++;
69: }
70: }
71: PetscMalloc1(cnt,&rows);
72: cnt = 0;
73: for (i=0; i<m; i++) {
74: if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) {
75: rows[cnt] = i;
76: cnt++;
77: }
78: }
79: ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,is);
80: return(0);
81: }
83: PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A,PetscInt *nrows,PetscInt **zrows)
84: {
85: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
86: const MatScalar *aa = a->a;
87: PetscInt i,m=A->rmap->n,cnt = 0;
88: const PetscInt *ii = a->i,*jj = a->j,*diag;
89: PetscInt *rows;
90: PetscErrorCode ierr;
93: MatMarkDiagonal_SeqAIJ(A);
94: diag = a->diag;
95: for (i=0; i<m; i++) {
96: if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) {
97: cnt++;
98: }
99: }
100: PetscMalloc1(cnt,&rows);
101: cnt = 0;
102: for (i=0; i<m; i++) {
103: if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) {
104: rows[cnt++] = i;
105: }
106: }
107: *nrows = cnt;
108: *zrows = rows;
109: return(0);
110: }
112: PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A,IS *zrows)
113: {
114: PetscInt nrows,*rows;
118: *zrows = NULL;
119: MatFindZeroDiagonals_SeqAIJ_Private(A,&nrows,&rows);
120: ISCreateGeneral(PetscObjectComm((PetscObject)A),nrows,rows,PETSC_OWN_POINTER,zrows);
121: return(0);
122: }
124: PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A,IS *keptrows)
125: {
126: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
127: const MatScalar *aa;
128: PetscInt m=A->rmap->n,cnt = 0;
129: const PetscInt *ii;
130: PetscInt n,i,j,*rows;
131: PetscErrorCode ierr;
134: *keptrows = NULL;
135: ii = a->i;
136: for (i=0; i<m; i++) {
137: n = ii[i+1] - ii[i];
138: if (!n) {
139: cnt++;
140: goto ok1;
141: }
142: aa = a->a + ii[i];
143: for (j=0; j<n; j++) {
144: if (aa[j] != 0.0) goto ok1;
145: }
146: cnt++;
147: ok1:;
148: }
149: if (!cnt) return(0);
150: PetscMalloc1(A->rmap->n-cnt,&rows);
151: cnt = 0;
152: for (i=0; i<m; i++) {
153: n = ii[i+1] - ii[i];
154: if (!n) continue;
155: aa = a->a + ii[i];
156: for (j=0; j<n; j++) {
157: if (aa[j] != 0.0) {
158: rows[cnt++] = i;
159: break;
160: }
161: }
162: }
163: ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,keptrows);
164: return(0);
165: }
167: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y,Vec D,InsertMode is)
168: {
169: PetscErrorCode ierr;
170: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) Y->data;
171: PetscInt i,m = Y->rmap->n;
172: const PetscInt *diag;
173: MatScalar *aa = aij->a;
174: const PetscScalar *v;
175: PetscBool missing;
176: #if defined(PETSC_HAVE_DEVICE)
177: PetscBool inserted = PETSC_FALSE;
178: #endif
181: if (Y->assembled) {
182: MatMissingDiagonal_SeqAIJ(Y,&missing,NULL);
183: if (!missing) {
184: diag = aij->diag;
185: VecGetArrayRead(D,&v);
186: if (is == INSERT_VALUES) {
187: #if defined(PETSC_HAVE_DEVICE)
188: inserted = PETSC_TRUE;
189: #endif
190: for (i=0; i<m; i++) {
191: aa[diag[i]] = v[i];
192: }
193: } else {
194: for (i=0; i<m; i++) {
195: #if defined(PETSC_HAVE_DEVICE)
196: if (v[i] != 0.0) inserted = PETSC_TRUE;
197: #endif
198: aa[diag[i]] += v[i];
199: }
200: }
201: #if defined(PETSC_HAVE_DEVICE)
202: if (inserted) Y->offloadmask = PETSC_OFFLOAD_CPU;
203: #endif
204: VecRestoreArrayRead(D,&v);
205: return(0);
206: }
207: MatSeqAIJInvalidateDiagonal(Y);
208: }
209: MatDiagonalSet_Default(Y,D,is);
210: return(0);
211: }
213: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *m,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
214: {
215: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
217: PetscInt i,ishift;
220: *m = A->rmap->n;
221: if (!ia) return(0);
222: ishift = 0;
223: if (symmetric && !A->structurally_symmetric) {
224: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,ishift,oshift,(PetscInt**)ia,(PetscInt**)ja);
225: } else if (oshift == 1) {
226: PetscInt *tia;
227: PetscInt nz = a->i[A->rmap->n];
228: /* malloc space and add 1 to i and j indices */
229: PetscMalloc1(A->rmap->n+1,&tia);
230: for (i=0; i<A->rmap->n+1; i++) tia[i] = a->i[i] + 1;
231: *ia = tia;
232: if (ja) {
233: PetscInt *tja;
234: PetscMalloc1(nz+1,&tja);
235: for (i=0; i<nz; i++) tja[i] = a->j[i] + 1;
236: *ja = tja;
237: }
238: } else {
239: *ia = a->i;
240: if (ja) *ja = a->j;
241: }
242: return(0);
243: }
245: PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
246: {
250: if (!ia) return(0);
251: if ((symmetric && !A->structurally_symmetric) || oshift == 1) {
252: PetscFree(*ia);
253: if (ja) {PetscFree(*ja);}
254: }
255: return(0);
256: }
258: PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
259: {
260: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
262: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
263: PetscInt nz = a->i[m],row,*jj,mr,col;
266: *nn = n;
267: if (!ia) return(0);
268: if (symmetric) {
269: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,0,oshift,(PetscInt**)ia,(PetscInt**)ja);
270: } else {
271: PetscCalloc1(n,&collengths);
272: PetscMalloc1(n+1,&cia);
273: PetscMalloc1(nz,&cja);
274: jj = a->j;
275: for (i=0; i<nz; i++) {
276: collengths[jj[i]]++;
277: }
278: cia[0] = oshift;
279: for (i=0; i<n; i++) {
280: cia[i+1] = cia[i] + collengths[i];
281: }
282: PetscArrayzero(collengths,n);
283: jj = a->j;
284: for (row=0; row<m; row++) {
285: mr = a->i[row+1] - a->i[row];
286: for (i=0; i<mr; i++) {
287: col = *jj++;
289: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
290: }
291: }
292: PetscFree(collengths);
293: *ia = cia; *ja = cja;
294: }
295: return(0);
296: }
298: PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
299: {
303: if (!ia) return(0);
305: PetscFree(*ia);
306: PetscFree(*ja);
307: return(0);
308: }
310: /*
311: MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from
312: MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output
313: spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ()
314: */
315: PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
316: {
317: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
319: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
320: PetscInt nz = a->i[m],row,mr,col,tmp;
321: PetscInt *cspidx;
322: const PetscInt *jj;
325: *nn = n;
326: if (!ia) return(0);
328: PetscCalloc1(n,&collengths);
329: PetscMalloc1(n+1,&cia);
330: PetscMalloc1(nz,&cja);
331: PetscMalloc1(nz,&cspidx);
332: jj = a->j;
333: for (i=0; i<nz; i++) {
334: collengths[jj[i]]++;
335: }
336: cia[0] = oshift;
337: for (i=0; i<n; i++) {
338: cia[i+1] = cia[i] + collengths[i];
339: }
340: PetscArrayzero(collengths,n);
341: jj = a->j;
342: for (row=0; row<m; row++) {
343: mr = a->i[row+1] - a->i[row];
344: for (i=0; i<mr; i++) {
345: col = *jj++;
346: tmp = cia[col] + collengths[col]++ - oshift;
347: cspidx[tmp] = a->i[row] + i; /* index of a->j */
348: cja[tmp] = row + oshift;
349: }
350: }
351: PetscFree(collengths);
352: *ia = cia;
353: *ja = cja;
354: *spidx = cspidx;
355: return(0);
356: }
358: PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
359: {
363: MatRestoreColumnIJ_SeqAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);
364: PetscFree(*spidx);
365: return(0);
366: }
368: PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A,PetscInt row,const PetscScalar v[])
369: {
370: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
371: PetscInt *ai = a->i;
375: PetscArraycpy(a->a+ai[row],v,ai[row+1]-ai[row]);
376: #if defined(PETSC_HAVE_DEVICE)
377: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && ai[row+1]-ai[row]) A->offloadmask = PETSC_OFFLOAD_CPU;
378: #endif
379: return(0);
380: }
382: /*
383: MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions
385: - a single row of values is set with each call
386: - no row or column indices are negative or (in error) larger than the number of rows or columns
387: - the values are always added to the matrix, not set
388: - no new locations are introduced in the nonzero structure of the matrix
390: This does NOT assume the global column indices are sorted
392: */
394: #include <petsc/private/isimpl.h>
395: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
396: {
397: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
398: PetscInt low,high,t,row,nrow,i,col,l;
399: const PetscInt *rp,*ai = a->i,*ailen = a->ilen,*aj = a->j;
400: PetscInt lastcol = -1;
401: MatScalar *ap,value,*aa = a->a;
402: const PetscInt *ridx = A->rmap->mapping->indices,*cidx = A->cmap->mapping->indices;
404: row = ridx[im[0]];
405: rp = aj + ai[row];
406: ap = aa + ai[row];
407: nrow = ailen[row];
408: low = 0;
409: high = nrow;
410: for (l=0; l<n; l++) { /* loop over added columns */
411: col = cidx[in[l]];
412: value = v[l];
414: if (col <= lastcol) low = 0;
415: else high = nrow;
416: lastcol = col;
417: while (high-low > 5) {
418: t = (low+high)/2;
419: if (rp[t] > col) high = t;
420: else low = t;
421: }
422: for (i=low; i<high; i++) {
423: if (rp[i] == col) {
424: ap[i] += value;
425: low = i + 1;
426: break;
427: }
428: }
429: }
430: #if defined(PETSC_HAVE_DEVICE)
431: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU;
432: #endif
433: return 0;
434: }
436: PetscErrorCode MatSetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
437: {
438: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
439: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
440: PetscInt *imax = a->imax,*ai = a->i,*ailen = a->ilen;
442: PetscInt *aj = a->j,nonew = a->nonew,lastcol = -1;
443: MatScalar *ap=NULL,value=0.0,*aa = a->a;
444: PetscBool ignorezeroentries = a->ignorezeroentries;
445: PetscBool roworiented = a->roworiented;
446: #if defined(PETSC_HAVE_DEVICE)
447: PetscBool inserted = PETSC_FALSE;
448: #endif
451: for (k=0; k<m; k++) { /* loop over added rows */
452: row = im[k];
453: if (row < 0) continue;
454: if (PetscUnlikelyDebug(row >= A->rmap->n)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->n-1);
455: rp = aj + ai[row];
456: if (!A->structure_only) ap = aa + ai[row];
457: rmax = imax[row]; nrow = ailen[row];
458: low = 0;
459: high = nrow;
460: for (l=0; l<n; l++) { /* loop over added columns */
461: if (in[l] < 0) continue;
462: if (PetscUnlikelyDebug(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);
463: col = in[l];
464: if (v && !A->structure_only) value = roworiented ? v[l + k*n] : v[k + l*m];
465: if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue;
467: if (col <= lastcol) low = 0;
468: else high = nrow;
469: lastcol = col;
470: while (high-low > 5) {
471: t = (low+high)/2;
472: if (rp[t] > col) high = t;
473: else low = t;
474: }
475: for (i=low; i<high; i++) {
476: if (rp[i] > col) break;
477: if (rp[i] == col) {
478: if (!A->structure_only) {
479: if (is == ADD_VALUES) {
480: ap[i] += value;
481: (void)PetscLogFlops(1.0);
482: }
483: else ap[i] = value;
484: #if defined(PETSC_HAVE_DEVICE)
485: inserted = PETSC_TRUE;
486: #endif
487: }
488: low = i + 1;
489: goto noinsert;
490: }
491: }
492: if (value == 0.0 && ignorezeroentries && row != col) goto noinsert;
493: if (nonew == 1) goto noinsert;
494: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at (%D,%D) in the matrix",row,col);
495: if (A->structure_only) {
496: MatSeqXAIJReallocateAIJ_structure_only(A,A->rmap->n,1,nrow,row,col,rmax,ai,aj,rp,imax,nonew,MatScalar);
497: } else {
498: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
499: }
500: N = nrow++ - 1; a->nz++; high++;
501: /* shift up all the later entries in this row */
502: PetscArraymove(rp+i+1,rp+i,N-i+1);
503: rp[i] = col;
504: if (!A->structure_only){
505: PetscArraymove(ap+i+1,ap+i,N-i+1);
506: ap[i] = value;
507: }
508: low = i + 1;
509: A->nonzerostate++;
510: #if defined(PETSC_HAVE_DEVICE)
511: inserted = PETSC_TRUE;
512: #endif
513: noinsert:;
514: }
515: ailen[row] = nrow;
516: }
517: #if defined(PETSC_HAVE_DEVICE)
518: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
519: #endif
520: return(0);
521: }
524: PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
525: {
526: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
527: PetscInt *rp,k,row;
528: PetscInt *ai = a->i;
530: PetscInt *aj = a->j;
531: MatScalar *aa = a->a,*ap;
534: if (A->was_assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot call on assembled matrix.");
535: if (m*n+a->nz > a->maxnz) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Number of entries in matrix will be larger than maximum nonzeros allocated for %D in MatSeqAIJSetTotalPreallocation()",a->maxnz);
536: for (k=0; k<m; k++) { /* loop over added rows */
537: row = im[k];
538: rp = aj + ai[row];
539: ap = aa + ai[row];
541: PetscMemcpy(rp,in,n*sizeof(PetscInt));
542: if (!A->structure_only) {
543: if (v) {
544: PetscMemcpy(ap,v,n*sizeof(PetscScalar));
545: v += n;
546: } else {
547: PetscMemzero(ap,n*sizeof(PetscScalar));
548: }
549: }
550: a->ilen[row] = n;
551: a->imax[row] = n;
552: a->i[row+1] = a->i[row]+n;
553: a->nz += n;
554: }
555: #if defined(PETSC_HAVE_DEVICE)
556: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU;
557: #endif
558: return(0);
559: }
561: /*@
562: MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix.
564: Input Parameters:
565: + A - the SeqAIJ matrix
566: - nztotal - bound on the number of nonzeros
568: Level: advanced
570: Notes:
571: This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row.
572: Simply call MatSetValues() after this call to provide the matrix entries in the usual manner. This matrix may be used
573: as always with multiple matrix assemblies.
575: .seealso: MatSetOption(), MAT_SORTED_FULL, MatSetValues(), MatSeqAIJSetPreallocation()
576: @*/
578: PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A,PetscInt nztotal)
579: {
581: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
584: PetscLayoutSetUp(A->rmap);
585: PetscLayoutSetUp(A->cmap);
586: a->maxnz = nztotal;
587: if (!a->imax) {
588: PetscMalloc1(A->rmap->n,&a->imax);
589: PetscLogObjectMemory((PetscObject)A,A->rmap->n*sizeof(PetscInt));
590: }
591: if (!a->ilen) {
592: PetscMalloc1(A->rmap->n,&a->ilen);
593: PetscLogObjectMemory((PetscObject)A,A->rmap->n*sizeof(PetscInt));
594: } else {
595: PetscMemzero(a->ilen,A->rmap->n*sizeof(PetscInt));
596: }
598: /* allocate the matrix space */
599: if (A->structure_only) {
600: PetscMalloc1(nztotal,&a->j);
601: PetscMalloc1(A->rmap->n+1,&a->i);
602: PetscLogObjectMemory((PetscObject)A,(A->rmap->n+1)*sizeof(PetscInt)+nztotal*sizeof(PetscInt));
603: } else {
604: PetscMalloc3(nztotal,&a->a,nztotal,&a->j,A->rmap->n+1,&a->i);
605: PetscLogObjectMemory((PetscObject)A,(A->rmap->n+1)*sizeof(PetscInt)+nztotal*(sizeof(PetscScalar)+sizeof(PetscInt)));
606: }
607: a->i[0] = 0;
608: if (A->structure_only) {
609: a->singlemalloc = PETSC_FALSE;
610: a->free_a = PETSC_FALSE;
611: } else {
612: a->singlemalloc = PETSC_TRUE;
613: a->free_a = PETSC_TRUE;
614: }
615: a->free_ij = PETSC_TRUE;
616: A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation;
617: A->preallocated = PETSC_TRUE;
618: return(0);
619: }
621: PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
622: {
623: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
624: PetscInt *rp,k,row;
625: PetscInt *ai = a->i,*ailen = a->ilen;
627: PetscInt *aj = a->j;
628: MatScalar *aa = a->a,*ap;
631: for (k=0; k<m; k++) { /* loop over added rows */
632: row = im[k];
633: if (PetscUnlikelyDebug(n > a->imax[row])) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Preallocation for row %D does not match number of columns provided",n);
634: rp = aj + ai[row];
635: ap = aa + ai[row];
636: if (!A->was_assembled) {
637: PetscMemcpy(rp,in,n*sizeof(PetscInt));
638: }
639: if (!A->structure_only) {
640: if (v) {
641: PetscMemcpy(ap,v,n*sizeof(PetscScalar));
642: v += n;
643: } else {
644: PetscMemzero(ap,n*sizeof(PetscScalar));
645: }
646: }
647: ailen[row] = n;
648: a->nz += n;
649: }
650: #if defined(PETSC_HAVE_DEVICE)
651: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU;
652: #endif
653: return(0);
654: }
657: PetscErrorCode MatGetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[])
658: {
659: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
660: PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j;
661: PetscInt *ai = a->i,*ailen = a->ilen;
662: MatScalar *ap,*aa = a->a;
665: for (k=0; k<m; k++) { /* loop over rows */
666: row = im[k];
667: if (row < 0) {v += n; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",row); */
668: 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);
669: rp = aj + ai[row]; ap = aa + ai[row];
670: nrow = ailen[row];
671: for (l=0; l<n; l++) { /* loop over columns */
672: if (in[l] < 0) {v++; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",in[l]); */
673: 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);
674: col = in[l];
675: high = nrow; low = 0; /* assume unsorted */
676: while (high-low > 5) {
677: t = (low+high)/2;
678: if (rp[t] > col) high = t;
679: else low = t;
680: }
681: for (i=low; i<high; i++) {
682: if (rp[i] > col) break;
683: if (rp[i] == col) {
684: *v++ = ap[i];
685: goto finished;
686: }
687: }
688: *v++ = 0.0;
689: finished:;
690: }
691: }
692: return(0);
693: }
695: PetscErrorCode MatView_SeqAIJ_Binary(Mat mat,PetscViewer viewer)
696: {
697: Mat_SeqAIJ *A = (Mat_SeqAIJ*)mat->data;
698: PetscInt header[4],M,N,m,nz,i;
699: PetscInt *rowlens;
703: PetscViewerSetUp(viewer);
705: M = mat->rmap->N;
706: N = mat->cmap->N;
707: m = mat->rmap->n;
708: nz = A->nz;
710: /* write matrix header */
711: header[0] = MAT_FILE_CLASSID;
712: header[1] = M; header[2] = N; header[3] = nz;
713: PetscViewerBinaryWrite(viewer,header,4,PETSC_INT);
715: /* fill in and store row lengths */
716: PetscMalloc1(m,&rowlens);
717: for (i=0; i<m; i++) rowlens[i] = A->i[i+1] - A->i[i];
718: PetscViewerBinaryWrite(viewer,rowlens,m,PETSC_INT);
719: PetscFree(rowlens);
720: /* store column indices */
721: PetscViewerBinaryWrite(viewer,A->j,nz,PETSC_INT);
722: /* store nonzero values */
723: PetscViewerBinaryWrite(viewer,A->a,nz,PETSC_SCALAR);
725: /* write block size option to the viewer's .info file */
726: MatView_Binary_BlockSizes(mat,viewer);
727: return(0);
728: }
730: static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A,PetscViewer viewer)
731: {
733: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
734: PetscInt i,k,m=A->rmap->N;
737: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
738: for (i=0; i<m; i++) {
739: PetscViewerASCIIPrintf(viewer,"row %D:",i);
740: for (k=a->i[i]; k<a->i[i+1]; k++) {
741: PetscViewerASCIIPrintf(viewer," (%D) ",a->j[k]);
742: }
743: PetscViewerASCIIPrintf(viewer,"\n");
744: }
745: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
746: return(0);
747: }
749: extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat,PetscViewer);
751: PetscErrorCode MatView_SeqAIJ_ASCII(Mat A,PetscViewer viewer)
752: {
753: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
754: PetscErrorCode ierr;
755: PetscInt i,j,m = A->rmap->n;
756: const char *name;
757: PetscViewerFormat format;
760: if (A->structure_only) {
761: MatView_SeqAIJ_ASCII_structonly(A,viewer);
762: return(0);
763: }
765: PetscViewerGetFormat(viewer,&format);
766: if (format == PETSC_VIEWER_ASCII_MATLAB) {
767: PetscInt nofinalvalue = 0;
768: if (m && ((a->i[m] == a->i[m-1]) || (a->j[a->nz-1] != A->cmap->n-1))) {
769: /* Need a dummy value to ensure the dimension of the matrix. */
770: nofinalvalue = 1;
771: }
772: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
773: PetscViewerASCIIPrintf(viewer,"%% Size = %D %D \n",m,A->cmap->n);
774: PetscViewerASCIIPrintf(viewer,"%% Nonzeros = %D \n",a->nz);
775: #if defined(PETSC_USE_COMPLEX)
776: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,4);\n",a->nz+nofinalvalue);
777: #else
778: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,3);\n",a->nz+nofinalvalue);
779: #endif
780: PetscViewerASCIIPrintf(viewer,"zzz = [\n");
782: for (i=0; i<m; i++) {
783: for (j=a->i[i]; j<a->i[i+1]; j++) {
784: #if defined(PETSC_USE_COMPLEX)
785: 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]));
786: #else
787: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",i+1,a->j[j]+1,(double)a->a[j]);
788: #endif
789: }
790: }
791: if (nofinalvalue) {
792: #if defined(PETSC_USE_COMPLEX)
793: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",m,A->cmap->n,0.,0.);
794: #else
795: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",m,A->cmap->n,0.0);
796: #endif
797: }
798: PetscObjectGetName((PetscObject)A,&name);
799: PetscViewerASCIIPrintf(viewer,"];\n %s = spconvert(zzz);\n",name);
800: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
801: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
802: return(0);
803: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
804: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
805: for (i=0; i<m; i++) {
806: PetscViewerASCIIPrintf(viewer,"row %D:",i);
807: for (j=a->i[i]; j<a->i[i+1]; j++) {
808: #if defined(PETSC_USE_COMPLEX)
809: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
810: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
811: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
812: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
813: } else if (PetscRealPart(a->a[j]) != 0.0) {
814: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
815: }
816: #else
817: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);}
818: #endif
819: }
820: PetscViewerASCIIPrintf(viewer,"\n");
821: }
822: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
823: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
824: PetscInt nzd=0,fshift=1,*sptr;
825: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
826: PetscMalloc1(m+1,&sptr);
827: for (i=0; i<m; i++) {
828: sptr[i] = nzd+1;
829: for (j=a->i[i]; j<a->i[i+1]; j++) {
830: if (a->j[j] >= i) {
831: #if defined(PETSC_USE_COMPLEX)
832: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++;
833: #else
834: if (a->a[j] != 0.0) nzd++;
835: #endif
836: }
837: }
838: }
839: sptr[m] = nzd+1;
840: PetscViewerASCIIPrintf(viewer," %D %D\n\n",m,nzd);
841: for (i=0; i<m+1; i+=6) {
842: if (i+4<m) {
843: 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]);
844: } else if (i+3<m) {
845: PetscViewerASCIIPrintf(viewer," %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4]);
846: } else if (i+2<m) {
847: PetscViewerASCIIPrintf(viewer," %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3]);
848: } else if (i+1<m) {
849: PetscViewerASCIIPrintf(viewer," %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2]);
850: } else if (i<m) {
851: PetscViewerASCIIPrintf(viewer," %D %D\n",sptr[i],sptr[i+1]);
852: } else {
853: PetscViewerASCIIPrintf(viewer," %D\n",sptr[i]);
854: }
855: }
856: PetscViewerASCIIPrintf(viewer,"\n");
857: PetscFree(sptr);
858: for (i=0; i<m; i++) {
859: for (j=a->i[i]; j<a->i[i+1]; j++) {
860: if (a->j[j] >= i) {PetscViewerASCIIPrintf(viewer," %D ",a->j[j]+fshift);}
861: }
862: PetscViewerASCIIPrintf(viewer,"\n");
863: }
864: PetscViewerASCIIPrintf(viewer,"\n");
865: for (i=0; i<m; i++) {
866: for (j=a->i[i]; j<a->i[i+1]; j++) {
867: if (a->j[j] >= i) {
868: #if defined(PETSC_USE_COMPLEX)
869: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) {
870: PetscViewerASCIIPrintf(viewer," %18.16e %18.16e ",(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
871: }
872: #else
873: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," %18.16e ",(double)a->a[j]);}
874: #endif
875: }
876: }
877: PetscViewerASCIIPrintf(viewer,"\n");
878: }
879: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
880: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
881: PetscInt cnt = 0,jcnt;
882: PetscScalar value;
883: #if defined(PETSC_USE_COMPLEX)
884: PetscBool realonly = PETSC_TRUE;
886: for (i=0; i<a->i[m]; i++) {
887: if (PetscImaginaryPart(a->a[i]) != 0.0) {
888: realonly = PETSC_FALSE;
889: break;
890: }
891: }
892: #endif
894: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
895: for (i=0; i<m; i++) {
896: jcnt = 0;
897: for (j=0; j<A->cmap->n; j++) {
898: if (jcnt < a->i[i+1]-a->i[i] && j == a->j[cnt]) {
899: value = a->a[cnt++];
900: jcnt++;
901: } else {
902: value = 0.0;
903: }
904: #if defined(PETSC_USE_COMPLEX)
905: if (realonly) {
906: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)PetscRealPart(value));
907: } else {
908: PetscViewerASCIIPrintf(viewer," %7.5e+%7.5e i ",(double)PetscRealPart(value),(double)PetscImaginaryPart(value));
909: }
910: #else
911: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)value);
912: #endif
913: }
914: PetscViewerASCIIPrintf(viewer,"\n");
915: }
916: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
917: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
918: PetscInt fshift=1;
919: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
920: #if defined(PETSC_USE_COMPLEX)
921: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate complex general\n");
922: #else
923: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate real general\n");
924: #endif
925: PetscViewerASCIIPrintf(viewer,"%D %D %D\n", m, A->cmap->n, a->nz);
926: for (i=0; i<m; i++) {
927: for (j=a->i[i]; j<a->i[i+1]; j++) {
928: #if defined(PETSC_USE_COMPLEX)
929: PetscViewerASCIIPrintf(viewer,"%D %D %g %g\n", i+fshift,a->j[j]+fshift,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
930: #else
931: PetscViewerASCIIPrintf(viewer,"%D %D %g\n", i+fshift, a->j[j]+fshift, (double)a->a[j]);
932: #endif
933: }
934: }
935: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
936: } else {
937: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
938: if (A->factortype) {
939: for (i=0; i<m; i++) {
940: PetscViewerASCIIPrintf(viewer,"row %D:",i);
941: /* L part */
942: for (j=a->i[i]; j<a->i[i+1]; j++) {
943: #if defined(PETSC_USE_COMPLEX)
944: if (PetscImaginaryPart(a->a[j]) > 0.0) {
945: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
946: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
947: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
948: } else {
949: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
950: }
951: #else
952: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
953: #endif
954: }
955: /* diagonal */
956: j = a->diag[i];
957: #if defined(PETSC_USE_COMPLEX)
958: if (PetscImaginaryPart(a->a[j]) > 0.0) {
959: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)PetscImaginaryPart(1.0/a->a[j]));
960: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
961: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)(-PetscImaginaryPart(1.0/a->a[j])));
962: } else {
963: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(1.0/a->a[j]));
964: }
965: #else
966: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)(1.0/a->a[j]));
967: #endif
969: /* U part */
970: for (j=a->diag[i+1]+1; j<a->diag[i]; j++) {
971: #if defined(PETSC_USE_COMPLEX)
972: if (PetscImaginaryPart(a->a[j]) > 0.0) {
973: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
974: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
975: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
976: } else {
977: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
978: }
979: #else
980: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
981: #endif
982: }
983: PetscViewerASCIIPrintf(viewer,"\n");
984: }
985: } else {
986: for (i=0; i<m; i++) {
987: PetscViewerASCIIPrintf(viewer,"row %D:",i);
988: for (j=a->i[i]; j<a->i[i+1]; j++) {
989: #if defined(PETSC_USE_COMPLEX)
990: if (PetscImaginaryPart(a->a[j]) > 0.0) {
991: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
992: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
993: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
994: } else {
995: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
996: }
997: #else
998: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
999: #endif
1000: }
1001: PetscViewerASCIIPrintf(viewer,"\n");
1002: }
1003: }
1004: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
1005: }
1006: PetscViewerFlush(viewer);
1007: return(0);
1008: }
1010: #include <petscdraw.h>
1011: PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw,void *Aa)
1012: {
1013: Mat A = (Mat) Aa;
1014: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1015: PetscErrorCode ierr;
1016: PetscInt i,j,m = A->rmap->n;
1017: int color;
1018: PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r;
1019: PetscViewer viewer;
1020: PetscViewerFormat format;
1023: PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);
1024: PetscViewerGetFormat(viewer,&format);
1025: PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);
1027: /* loop over matrix elements drawing boxes */
1029: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
1030: PetscDrawCollectiveBegin(draw);
1031: /* Blue for negative, Cyan for zero and Red for positive */
1032: color = PETSC_DRAW_BLUE;
1033: for (i=0; i<m; i++) {
1034: y_l = m - i - 1.0; y_r = y_l + 1.0;
1035: for (j=a->i[i]; j<a->i[i+1]; j++) {
1036: x_l = a->j[j]; x_r = x_l + 1.0;
1037: if (PetscRealPart(a->a[j]) >= 0.) continue;
1038: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1039: }
1040: }
1041: color = PETSC_DRAW_CYAN;
1042: for (i=0; i<m; i++) {
1043: y_l = m - i - 1.0; y_r = y_l + 1.0;
1044: for (j=a->i[i]; j<a->i[i+1]; j++) {
1045: x_l = a->j[j]; x_r = x_l + 1.0;
1046: if (a->a[j] != 0.) continue;
1047: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1048: }
1049: }
1050: color = PETSC_DRAW_RED;
1051: for (i=0; i<m; i++) {
1052: y_l = m - i - 1.0; y_r = y_l + 1.0;
1053: for (j=a->i[i]; j<a->i[i+1]; j++) {
1054: x_l = a->j[j]; x_r = x_l + 1.0;
1055: if (PetscRealPart(a->a[j]) <= 0.) continue;
1056: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1057: }
1058: }
1059: PetscDrawCollectiveEnd(draw);
1060: } else {
1061: /* use contour shading to indicate magnitude of values */
1062: /* first determine max of all nonzero values */
1063: PetscReal minv = 0.0, maxv = 0.0;
1064: PetscInt nz = a->nz, count = 0;
1065: PetscDraw popup;
1067: for (i=0; i<nz; i++) {
1068: if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]);
1069: }
1070: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1071: PetscDrawGetPopup(draw,&popup);
1072: PetscDrawScalePopup(popup,minv,maxv);
1074: PetscDrawCollectiveBegin(draw);
1075: for (i=0; i<m; i++) {
1076: y_l = m - i - 1.0;
1077: y_r = y_l + 1.0;
1078: for (j=a->i[i]; j<a->i[i+1]; j++) {
1079: x_l = a->j[j];
1080: x_r = x_l + 1.0;
1081: color = PetscDrawRealToColor(PetscAbsScalar(a->a[count]),minv,maxv);
1082: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1083: count++;
1084: }
1085: }
1086: PetscDrawCollectiveEnd(draw);
1087: }
1088: return(0);
1089: }
1091: #include <petscdraw.h>
1092: PetscErrorCode MatView_SeqAIJ_Draw(Mat A,PetscViewer viewer)
1093: {
1095: PetscDraw draw;
1096: PetscReal xr,yr,xl,yl,h,w;
1097: PetscBool isnull;
1100: PetscViewerDrawGetDraw(viewer,0,&draw);
1101: PetscDrawIsNull(draw,&isnull);
1102: if (isnull) return(0);
1104: xr = A->cmap->n; yr = A->rmap->n; h = yr/10.0; w = xr/10.0;
1105: xr += w; yr += h; xl = -w; yl = -h;
1106: PetscDrawSetCoordinates(draw,xl,yl,xr,yr);
1107: PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);
1108: PetscDrawZoom(draw,MatView_SeqAIJ_Draw_Zoom,A);
1109: PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);
1110: PetscDrawSave(draw);
1111: return(0);
1112: }
1114: PetscErrorCode MatView_SeqAIJ(Mat A,PetscViewer viewer)
1115: {
1117: PetscBool iascii,isbinary,isdraw;
1120: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1121: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1122: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1123: if (iascii) {
1124: MatView_SeqAIJ_ASCII(A,viewer);
1125: } else if (isbinary) {
1126: MatView_SeqAIJ_Binary(A,viewer);
1127: } else if (isdraw) {
1128: MatView_SeqAIJ_Draw(A,viewer);
1129: }
1130: MatView_SeqAIJ_Inode(A,viewer);
1131: return(0);
1132: }
1134: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A,MatAssemblyType mode)
1135: {
1136: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1138: PetscInt fshift = 0,i,*ai = a->i,*aj = a->j,*imax = a->imax;
1139: PetscInt m = A->rmap->n,*ip,N,*ailen = a->ilen,rmax = 0;
1140: MatScalar *aa = a->a,*ap;
1141: PetscReal ratio = 0.6;
1144: if (mode == MAT_FLUSH_ASSEMBLY) return(0);
1145: MatSeqAIJInvalidateDiagonal(A);
1146: if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) {
1147: /* we need to respect users asking to use or not the inodes routine in between matrix assemblies */
1148: MatAssemblyEnd_SeqAIJ_Inode(A,mode);
1149: return(0);
1150: }
1152: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
1153: for (i=1; i<m; i++) {
1154: /* move each row back by the amount of empty slots (fshift) before it*/
1155: fshift += imax[i-1] - ailen[i-1];
1156: rmax = PetscMax(rmax,ailen[i]);
1157: if (fshift) {
1158: ip = aj + ai[i];
1159: ap = aa + ai[i];
1160: N = ailen[i];
1161: PetscArraymove(ip-fshift,ip,N);
1162: if (!A->structure_only) {
1163: PetscArraymove(ap-fshift,ap,N);
1164: }
1165: }
1166: ai[i] = ai[i-1] + ailen[i-1];
1167: }
1168: if (m) {
1169: fshift += imax[m-1] - ailen[m-1];
1170: ai[m] = ai[m-1] + ailen[m-1];
1171: }
1173: /* reset ilen and imax for each row */
1174: a->nonzerorowcnt = 0;
1175: if (A->structure_only) {
1176: PetscFree(a->imax);
1177: PetscFree(a->ilen);
1178: } else { /* !A->structure_only */
1179: for (i=0; i<m; i++) {
1180: ailen[i] = imax[i] = ai[i+1] - ai[i];
1181: a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0);
1182: }
1183: }
1184: a->nz = ai[m];
1185: 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);
1187: MatMarkDiagonal_SeqAIJ(A);
1188: PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap->n,fshift,a->nz);
1189: PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);
1190: PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);
1192: A->info.mallocs += a->reallocs;
1193: a->reallocs = 0;
1194: A->info.nz_unneeded = (PetscReal)fshift;
1195: a->rmax = rmax;
1197: if (!A->structure_only) {
1198: MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,m,ratio);
1199: }
1200: MatAssemblyEnd_SeqAIJ_Inode(A,mode);
1201: return(0);
1202: }
1204: PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1205: {
1206: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1207: PetscInt i,nz = a->nz;
1208: MatScalar *aa = a->a;
1212: for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]);
1213: MatSeqAIJInvalidateDiagonal(A);
1214: #if defined(PETSC_HAVE_DEVICE)
1215: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
1216: #endif
1217: return(0);
1218: }
1220: PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1221: {
1222: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1223: PetscInt i,nz = a->nz;
1224: MatScalar *aa = a->a;
1228: for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1229: MatSeqAIJInvalidateDiagonal(A);
1230: #if defined(PETSC_HAVE_DEVICE)
1231: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
1232: #endif
1233: return(0);
1234: }
1236: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1237: {
1238: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1242: PetscArrayzero(a->a,a->i[A->rmap->n]);
1243: MatSeqAIJInvalidateDiagonal(A);
1244: #if defined(PETSC_HAVE_DEVICE)
1245: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
1246: #endif
1247: return(0);
1248: }
1250: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1251: {
1252: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1256: #if defined(PETSC_USE_LOG)
1257: PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->n,A->cmap->n,a->nz);
1258: #endif
1259: MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);
1260: ISDestroy(&a->row);
1261: ISDestroy(&a->col);
1262: PetscFree(a->diag);
1263: PetscFree(a->ibdiag);
1264: PetscFree(a->imax);
1265: PetscFree(a->ilen);
1266: PetscFree(a->ipre);
1267: PetscFree3(a->idiag,a->mdiag,a->ssor_work);
1268: PetscFree(a->solve_work);
1269: ISDestroy(&a->icol);
1270: PetscFree(a->saved_values);
1271: PetscFree2(a->compressedrow.i,a->compressedrow.rindex);
1273: MatDestroy_SeqAIJ_Inode(A);
1274: PetscFree(A->data);
1276: /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this.
1277: That function is so heavily used (sometimes in an hidden way through multnumeric function pointers)
1278: that is hard to properly add this data to the MatProduct data. We free it here to avoid
1279: users reusing the matrix object with different data to incur in obscure segmentation faults
1280: due to different matrix sizes */
1281: PetscObjectCompose((PetscObject)A,"__PETSc__ab_dense",NULL);
1283: PetscObjectChangeTypeName((PetscObject)A,NULL);
1284: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetColumnIndices_C",NULL);
1285: PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);
1286: PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);
1287: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsbaij_C",NULL);
1288: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqbaij_C",NULL);
1289: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijperm_C",NULL);
1290: #if defined(PETSC_HAVE_CUDA)
1291: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijcusparse_C",NULL);
1292: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqaijcusparse_seqaij_C",NULL);
1293: #endif
1294: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
1295: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijkokkos_C",NULL);
1296: #endif
1297: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijcrl_C",NULL);
1298: #if defined(PETSC_HAVE_ELEMENTAL)
1299: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_elemental_C",NULL);
1300: #endif
1301: #if defined(PETSC_HAVE_SCALAPACK)
1302: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_scalapack_C",NULL);
1303: #endif
1304: #if defined(PETSC_HAVE_HYPRE)
1305: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_hypre_C",NULL);
1306: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_transpose_seqaij_seqaij_C",NULL);
1307: #endif
1308: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqdense_C",NULL);
1309: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsell_C",NULL);
1310: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_is_C",NULL);
1311: PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);
1312: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);
1313: PetscObjectComposeFunction((PetscObject)A,"MatResetPreallocation_C",NULL);
1314: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);
1315: PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);
1316: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_is_seqaij_C",NULL);
1317: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqdense_seqaij_C",NULL);
1318: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqaij_seqaij_C",NULL);
1319: return(0);
1320: }
1322: PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg)
1323: {
1324: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1328: switch (op) {
1329: case MAT_ROW_ORIENTED:
1330: a->roworiented = flg;
1331: break;
1332: case MAT_KEEP_NONZERO_PATTERN:
1333: a->keepnonzeropattern = flg;
1334: break;
1335: case MAT_NEW_NONZERO_LOCATIONS:
1336: a->nonew = (flg ? 0 : 1);
1337: break;
1338: case MAT_NEW_NONZERO_LOCATION_ERR:
1339: a->nonew = (flg ? -1 : 0);
1340: break;
1341: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1342: a->nonew = (flg ? -2 : 0);
1343: break;
1344: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1345: a->nounused = (flg ? -1 : 0);
1346: break;
1347: case MAT_IGNORE_ZERO_ENTRIES:
1348: a->ignorezeroentries = flg;
1349: break;
1350: case MAT_SPD:
1351: case MAT_SYMMETRIC:
1352: case MAT_STRUCTURALLY_SYMMETRIC:
1353: case MAT_HERMITIAN:
1354: case MAT_SYMMETRY_ETERNAL:
1355: case MAT_STRUCTURE_ONLY:
1356: /* These options are handled directly by MatSetOption() */
1357: break;
1358: case MAT_NEW_DIAGONALS:
1359: case MAT_IGNORE_OFF_PROC_ENTRIES:
1360: case MAT_USE_HASH_TABLE:
1361: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1362: break;
1363: case MAT_USE_INODES:
1364: MatSetOption_SeqAIJ_Inode(A,MAT_USE_INODES,flg);
1365: break;
1366: case MAT_SUBMAT_SINGLEIS:
1367: A->submat_singleis = flg;
1368: break;
1369: case MAT_SORTED_FULL:
1370: if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
1371: else A->ops->setvalues = MatSetValues_SeqAIJ;
1372: break;
1373: default:
1374: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1375: }
1376: return(0);
1377: }
1379: PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v)
1380: {
1381: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1383: PetscInt i,j,n,*ai=a->i,*aj=a->j;
1384: PetscScalar *aa=a->a,*x;
1387: VecGetLocalSize(v,&n);
1388: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
1390: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1391: PetscInt *diag=a->diag;
1392: VecGetArrayWrite(v,&x);
1393: for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]];
1394: VecRestoreArrayWrite(v,&x);
1395: return(0);
1396: }
1398: VecGetArrayWrite(v,&x);
1399: for (i=0; i<n; i++) {
1400: x[i] = 0.0;
1401: for (j=ai[i]; j<ai[i+1]; j++) {
1402: if (aj[j] == i) {
1403: x[i] = aa[j];
1404: break;
1405: }
1406: }
1407: }
1408: VecRestoreArrayWrite(v,&x);
1409: return(0);
1410: }
1412: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1413: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy)
1414: {
1415: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1416: PetscScalar *y;
1417: const PetscScalar *x;
1418: PetscErrorCode ierr;
1419: PetscInt m = A->rmap->n;
1420: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1421: const MatScalar *v;
1422: PetscScalar alpha;
1423: PetscInt n,i,j;
1424: const PetscInt *idx,*ii,*ridx=NULL;
1425: Mat_CompressedRow cprow = a->compressedrow;
1426: PetscBool usecprow = cprow.use;
1427: #endif
1430: if (zz != yy) {VecCopy(zz,yy);}
1431: VecGetArrayRead(xx,&x);
1432: VecGetArray(yy,&y);
1434: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1435: fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y);
1436: #else
1437: if (usecprow) {
1438: m = cprow.nrows;
1439: ii = cprow.i;
1440: ridx = cprow.rindex;
1441: } else {
1442: ii = a->i;
1443: }
1444: for (i=0; i<m; i++) {
1445: idx = a->j + ii[i];
1446: v = a->a + ii[i];
1447: n = ii[i+1] - ii[i];
1448: if (usecprow) {
1449: alpha = x[ridx[i]];
1450: } else {
1451: alpha = x[i];
1452: }
1453: for (j=0; j<n; j++) y[idx[j]] += alpha*v[j];
1454: }
1455: #endif
1456: PetscLogFlops(2.0*a->nz);
1457: VecRestoreArrayRead(xx,&x);
1458: VecRestoreArray(yy,&y);
1459: return(0);
1460: }
1462: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy)
1463: {
1467: VecSet(yy,0.0);
1468: MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);
1469: return(0);
1470: }
1472: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1474: PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy)
1475: {
1476: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1477: PetscScalar *y;
1478: const PetscScalar *x;
1479: const MatScalar *aa;
1480: PetscErrorCode ierr;
1481: PetscInt m=A->rmap->n;
1482: const PetscInt *aj,*ii,*ridx=NULL;
1483: PetscInt n,i;
1484: PetscScalar sum;
1485: PetscBool usecprow=a->compressedrow.use;
1487: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1488: #pragma disjoint(*x,*y,*aa)
1489: #endif
1492: if (a->inode.use && a->inode.checked) {
1493: MatMult_SeqAIJ_Inode(A,xx,yy);
1494: return(0);
1495: }
1496: VecGetArrayRead(xx,&x);
1497: VecGetArray(yy,&y);
1498: ii = a->i;
1499: if (usecprow) { /* use compressed row format */
1500: PetscArrayzero(y,m);
1501: m = a->compressedrow.nrows;
1502: ii = a->compressedrow.i;
1503: ridx = a->compressedrow.rindex;
1504: for (i=0; i<m; i++) {
1505: n = ii[i+1] - ii[i];
1506: aj = a->j + ii[i];
1507: aa = a->a + ii[i];
1508: sum = 0.0;
1509: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1510: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1511: y[*ridx++] = sum;
1512: }
1513: } else { /* do not use compressed row format */
1514: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1515: aj = a->j;
1516: aa = a->a;
1517: fortranmultaij_(&m,x,ii,aj,aa,y);
1518: #else
1519: for (i=0; i<m; i++) {
1520: n = ii[i+1] - ii[i];
1521: aj = a->j + ii[i];
1522: aa = a->a + ii[i];
1523: sum = 0.0;
1524: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1525: y[i] = sum;
1526: }
1527: #endif
1528: }
1529: PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);
1530: VecRestoreArrayRead(xx,&x);
1531: VecRestoreArray(yy,&y);
1532: return(0);
1533: }
1535: PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy)
1536: {
1537: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1538: PetscScalar *y;
1539: const PetscScalar *x;
1540: const MatScalar *aa;
1541: PetscErrorCode ierr;
1542: PetscInt m=A->rmap->n;
1543: const PetscInt *aj,*ii,*ridx=NULL;
1544: PetscInt n,i,nonzerorow=0;
1545: PetscScalar sum;
1546: PetscBool usecprow=a->compressedrow.use;
1548: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1549: #pragma disjoint(*x,*y,*aa)
1550: #endif
1553: VecGetArrayRead(xx,&x);
1554: VecGetArray(yy,&y);
1555: if (usecprow) { /* use compressed row format */
1556: m = a->compressedrow.nrows;
1557: ii = a->compressedrow.i;
1558: ridx = a->compressedrow.rindex;
1559: for (i=0; i<m; i++) {
1560: n = ii[i+1] - ii[i];
1561: aj = a->j + ii[i];
1562: aa = a->a + ii[i];
1563: sum = 0.0;
1564: nonzerorow += (n>0);
1565: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1566: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1567: y[*ridx++] = sum;
1568: }
1569: } else { /* do not use compressed row format */
1570: ii = a->i;
1571: for (i=0; i<m; i++) {
1572: n = ii[i+1] - ii[i];
1573: aj = a->j + ii[i];
1574: aa = a->a + ii[i];
1575: sum = 0.0;
1576: nonzerorow += (n>0);
1577: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1578: y[i] = sum;
1579: }
1580: }
1581: PetscLogFlops(2.0*a->nz - nonzerorow);
1582: VecRestoreArrayRead(xx,&x);
1583: VecRestoreArray(yy,&y);
1584: return(0);
1585: }
1587: PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1588: {
1589: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1590: PetscScalar *y,*z;
1591: const PetscScalar *x;
1592: const MatScalar *aa;
1593: PetscErrorCode ierr;
1594: PetscInt m = A->rmap->n,*aj,*ii;
1595: PetscInt n,i,*ridx=NULL;
1596: PetscScalar sum;
1597: PetscBool usecprow=a->compressedrow.use;
1600: VecGetArrayRead(xx,&x);
1601: VecGetArrayPair(yy,zz,&y,&z);
1602: if (usecprow) { /* use compressed row format */
1603: if (zz != yy) {
1604: PetscArraycpy(z,y,m);
1605: }
1606: m = a->compressedrow.nrows;
1607: ii = a->compressedrow.i;
1608: ridx = a->compressedrow.rindex;
1609: for (i=0; i<m; i++) {
1610: n = ii[i+1] - ii[i];
1611: aj = a->j + ii[i];
1612: aa = a->a + ii[i];
1613: sum = y[*ridx];
1614: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1615: z[*ridx++] = sum;
1616: }
1617: } else { /* do not use compressed row format */
1618: ii = a->i;
1619: for (i=0; i<m; i++) {
1620: n = ii[i+1] - ii[i];
1621: aj = a->j + ii[i];
1622: aa = a->a + ii[i];
1623: sum = y[i];
1624: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1625: z[i] = sum;
1626: }
1627: }
1628: PetscLogFlops(2.0*a->nz);
1629: VecRestoreArrayRead(xx,&x);
1630: VecRestoreArrayPair(yy,zz,&y,&z);
1631: return(0);
1632: }
1634: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1635: PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1636: {
1637: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1638: PetscScalar *y,*z;
1639: const PetscScalar *x;
1640: const MatScalar *aa;
1641: PetscErrorCode ierr;
1642: const PetscInt *aj,*ii,*ridx=NULL;
1643: PetscInt m = A->rmap->n,n,i;
1644: PetscScalar sum;
1645: PetscBool usecprow=a->compressedrow.use;
1648: if (a->inode.use && a->inode.checked) {
1649: MatMultAdd_SeqAIJ_Inode(A,xx,yy,zz);
1650: return(0);
1651: }
1652: VecGetArrayRead(xx,&x);
1653: VecGetArrayPair(yy,zz,&y,&z);
1654: if (usecprow) { /* use compressed row format */
1655: if (zz != yy) {
1656: PetscArraycpy(z,y,m);
1657: }
1658: m = a->compressedrow.nrows;
1659: ii = a->compressedrow.i;
1660: ridx = a->compressedrow.rindex;
1661: for (i=0; i<m; i++) {
1662: n = ii[i+1] - ii[i];
1663: aj = a->j + ii[i];
1664: aa = a->a + ii[i];
1665: sum = y[*ridx];
1666: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1667: z[*ridx++] = sum;
1668: }
1669: } else { /* do not use compressed row format */
1670: ii = a->i;
1671: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1672: aj = a->j;
1673: aa = a->a;
1674: fortranmultaddaij_(&m,x,ii,aj,aa,y,z);
1675: #else
1676: for (i=0; i<m; i++) {
1677: n = ii[i+1] - ii[i];
1678: aj = a->j + ii[i];
1679: aa = a->a + ii[i];
1680: sum = y[i];
1681: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1682: z[i] = sum;
1683: }
1684: #endif
1685: }
1686: PetscLogFlops(2.0*a->nz);
1687: VecRestoreArrayRead(xx,&x);
1688: VecRestoreArrayPair(yy,zz,&y,&z);
1689: return(0);
1690: }
1692: /*
1693: Adds diagonal pointers to sparse matrix structure.
1694: */
1695: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1696: {
1697: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1699: PetscInt i,j,m = A->rmap->n;
1702: if (!a->diag) {
1703: PetscMalloc1(m,&a->diag);
1704: PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));
1705: }
1706: for (i=0; i<A->rmap->n; i++) {
1707: a->diag[i] = a->i[i+1];
1708: for (j=a->i[i]; j<a->i[i+1]; j++) {
1709: if (a->j[j] == i) {
1710: a->diag[i] = j;
1711: break;
1712: }
1713: }
1714: }
1715: return(0);
1716: }
1718: PetscErrorCode MatShift_SeqAIJ(Mat A,PetscScalar v)
1719: {
1720: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1721: const PetscInt *diag = (const PetscInt*)a->diag;
1722: const PetscInt *ii = (const PetscInt*) a->i;
1723: PetscInt i,*mdiag = NULL;
1724: PetscErrorCode ierr;
1725: PetscInt cnt = 0; /* how many diagonals are missing */
1728: if (!A->preallocated || !a->nz) {
1729: MatSeqAIJSetPreallocation(A,1,NULL);
1730: MatShift_Basic(A,v);
1731: return(0);
1732: }
1734: if (a->diagonaldense) {
1735: cnt = 0;
1736: } else {
1737: PetscCalloc1(A->rmap->n,&mdiag);
1738: for (i=0; i<A->rmap->n; i++) {
1739: if (diag[i] >= ii[i+1]) {
1740: cnt++;
1741: mdiag[i] = 1;
1742: }
1743: }
1744: }
1745: if (!cnt) {
1746: MatShift_Basic(A,v);
1747: } else {
1748: PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */
1749: PetscInt *oldj = a->j, *oldi = a->i;
1750: PetscBool singlemalloc = a->singlemalloc,free_a = a->free_a,free_ij = a->free_ij;
1752: a->a = NULL;
1753: a->j = NULL;
1754: a->i = NULL;
1755: /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */
1756: for (i=0; i<A->rmap->n; i++) {
1757: a->imax[i] += mdiag[i];
1758: a->imax[i] = PetscMin(a->imax[i],A->cmap->n);
1759: }
1760: MatSeqAIJSetPreallocation_SeqAIJ(A,0,a->imax);
1762: /* copy old values into new matrix data structure */
1763: for (i=0; i<A->rmap->n; i++) {
1764: MatSetValues(A,1,&i,a->imax[i] - mdiag[i],&oldj[oldi[i]],&olda[oldi[i]],ADD_VALUES);
1765: if (i < A->cmap->n) {
1766: MatSetValue(A,i,i,v,ADD_VALUES);
1767: }
1768: }
1769: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1770: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1771: if (singlemalloc) {
1772: PetscFree3(olda,oldj,oldi);
1773: } else {
1774: if (free_a) {PetscFree(olda);}
1775: if (free_ij) {PetscFree(oldj);}
1776: if (free_ij) {PetscFree(oldi);}
1777: }
1778: }
1779: PetscFree(mdiag);
1780: a->diagonaldense = PETSC_TRUE;
1781: return(0);
1782: }
1784: /*
1785: Checks for missing diagonals
1786: */
1787: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d)
1788: {
1789: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1790: PetscInt *diag,*ii = a->i,i;
1794: *missing = PETSC_FALSE;
1795: if (A->rmap->n > 0 && !ii) {
1796: *missing = PETSC_TRUE;
1797: if (d) *d = 0;
1798: PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");
1799: } else {
1800: PetscInt n;
1801: n = PetscMin(A->rmap->n, A->cmap->n);
1802: diag = a->diag;
1803: for (i=0; i<n; i++) {
1804: if (diag[i] >= ii[i+1]) {
1805: *missing = PETSC_TRUE;
1806: if (d) *d = i;
1807: PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);
1808: break;
1809: }
1810: }
1811: }
1812: return(0);
1813: }
1815: #include <petscblaslapack.h>
1816: #include <petsc/private/kernels/blockinvert.h>
1818: /*
1819: Note that values is allocated externally by the PC and then passed into this routine
1820: */
1821: PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
1822: {
1823: PetscErrorCode ierr;
1824: PetscInt n = A->rmap->n, i, ncnt = 0, *indx,j,bsizemax = 0,*v_pivots;
1825: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
1826: const PetscReal shift = 0.0;
1827: PetscInt ipvt[5];
1828: PetscScalar work[25],*v_work;
1831: allowzeropivot = PetscNot(A->erroriffailure);
1832: for (i=0; i<nblocks; i++) ncnt += bsizes[i];
1833: if (ncnt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Total blocksizes %D doesn't match number matrix rows %D",ncnt,n);
1834: for (i=0; i<nblocks; i++) {
1835: bsizemax = PetscMax(bsizemax,bsizes[i]);
1836: }
1837: PetscMalloc1(bsizemax,&indx);
1838: if (bsizemax > 7) {
1839: PetscMalloc2(bsizemax,&v_work,bsizemax,&v_pivots);
1840: }
1841: ncnt = 0;
1842: for (i=0; i<nblocks; i++) {
1843: for (j=0; j<bsizes[i]; j++) indx[j] = ncnt+j;
1844: MatGetValues(A,bsizes[i],indx,bsizes[i],indx,diag);
1845: switch (bsizes[i]) {
1846: case 1:
1847: *diag = 1.0/(*diag);
1848: break;
1849: case 2:
1850: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
1851: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1852: PetscKernel_A_gets_transpose_A_2(diag);
1853: break;
1854: case 3:
1855: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
1856: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1857: PetscKernel_A_gets_transpose_A_3(diag);
1858: break;
1859: case 4:
1860: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
1861: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1862: PetscKernel_A_gets_transpose_A_4(diag);
1863: break;
1864: case 5:
1865: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
1866: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1867: PetscKernel_A_gets_transpose_A_5(diag);
1868: break;
1869: case 6:
1870: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
1871: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1872: PetscKernel_A_gets_transpose_A_6(diag);
1873: break;
1874: case 7:
1875: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
1876: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1877: PetscKernel_A_gets_transpose_A_7(diag);
1878: break;
1879: default:
1880: PetscKernel_A_gets_inverse_A(bsizes[i],diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
1881: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1882: PetscKernel_A_gets_transpose_A_N(diag,bsizes[i]);
1883: }
1884: ncnt += bsizes[i];
1885: diag += bsizes[i]*bsizes[i];
1886: }
1887: if (bsizemax > 7) {
1888: PetscFree2(v_work,v_pivots);
1889: }
1890: PetscFree(indx);
1891: return(0);
1892: }
1894: /*
1895: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1896: */
1897: PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift)
1898: {
1899: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
1901: PetscInt i,*diag,m = A->rmap->n;
1902: MatScalar *v = a->a;
1903: PetscScalar *idiag,*mdiag;
1906: if (a->idiagvalid) return(0);
1907: MatMarkDiagonal_SeqAIJ(A);
1908: diag = a->diag;
1909: if (!a->idiag) {
1910: PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);
1911: PetscLogObjectMemory((PetscObject)A, 3*m*sizeof(PetscScalar));
1912: v = a->a;
1913: }
1914: mdiag = a->mdiag;
1915: idiag = a->idiag;
1917: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1918: for (i=0; i<m; i++) {
1919: mdiag[i] = v[diag[i]];
1920: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1921: if (PetscRealPart(fshift)) {
1922: PetscInfo1(A,"Zero diagonal on row %D\n",i);
1923: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1924: A->factorerror_zeropivot_value = 0.0;
1925: A->factorerror_zeropivot_row = i;
1926: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i);
1927: }
1928: idiag[i] = 1.0/v[diag[i]];
1929: }
1930: PetscLogFlops(m);
1931: } else {
1932: for (i=0; i<m; i++) {
1933: mdiag[i] = v[diag[i]];
1934: idiag[i] = omega/(fshift + v[diag[i]]);
1935: }
1936: PetscLogFlops(2.0*m);
1937: }
1938: a->idiagvalid = PETSC_TRUE;
1939: return(0);
1940: }
1942: #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h>
1943: PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1944: {
1945: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1946: PetscScalar *x,d,sum,*t,scale;
1947: const MatScalar *v,*idiag=NULL,*mdiag;
1948: const PetscScalar *b, *bs,*xb, *ts;
1949: PetscErrorCode ierr;
1950: PetscInt n,m = A->rmap->n,i;
1951: const PetscInt *idx,*diag;
1954: if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) {
1955: MatSOR_SeqAIJ_Inode(A,bb,omega,flag,fshift,its,lits,xx);
1956: return(0);
1957: }
1958: its = its*lits;
1960: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1961: if (!a->idiagvalid) {MatInvertDiagonal_SeqAIJ(A,omega,fshift);}
1962: a->fshift = fshift;
1963: a->omega = omega;
1965: diag = a->diag;
1966: t = a->ssor_work;
1967: idiag = a->idiag;
1968: mdiag = a->mdiag;
1970: VecGetArray(xx,&x);
1971: VecGetArrayRead(bb,&b);
1972: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1973: if (flag == SOR_APPLY_UPPER) {
1974: /* apply (U + D/omega) to the vector */
1975: bs = b;
1976: for (i=0; i<m; i++) {
1977: d = fshift + mdiag[i];
1978: n = a->i[i+1] - diag[i] - 1;
1979: idx = a->j + diag[i] + 1;
1980: v = a->a + diag[i] + 1;
1981: sum = b[i]*d/omega;
1982: PetscSparseDensePlusDot(sum,bs,v,idx,n);
1983: x[i] = sum;
1984: }
1985: VecRestoreArray(xx,&x);
1986: VecRestoreArrayRead(bb,&b);
1987: PetscLogFlops(a->nz);
1988: return(0);
1989: }
1991: if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented");
1992: else if (flag & SOR_EISENSTAT) {
1993: /* Let A = L + U + D; where L is lower triangular,
1994: U is upper triangular, E = D/omega; This routine applies
1996: (L + E)^{-1} A (U + E)^{-1}
1998: to a vector efficiently using Eisenstat's trick.
1999: */
2000: scale = (2.0/omega) - 1.0;
2002: /* x = (E + U)^{-1} b */
2003: for (i=m-1; i>=0; i--) {
2004: n = a->i[i+1] - diag[i] - 1;
2005: idx = a->j + diag[i] + 1;
2006: v = a->a + diag[i] + 1;
2007: sum = b[i];
2008: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2009: x[i] = sum*idiag[i];
2010: }
2012: /* t = b - (2*E - D)x */
2013: v = a->a;
2014: for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i];
2016: /* t = (E + L)^{-1}t */
2017: ts = t;
2018: diag = a->diag;
2019: for (i=0; i<m; i++) {
2020: n = diag[i] - a->i[i];
2021: idx = a->j + a->i[i];
2022: v = a->a + a->i[i];
2023: sum = t[i];
2024: PetscSparseDenseMinusDot(sum,ts,v,idx,n);
2025: t[i] = sum*idiag[i];
2026: /* x = x + t */
2027: x[i] += t[i];
2028: }
2030: PetscLogFlops(6.0*m-1 + 2.0*a->nz);
2031: VecRestoreArray(xx,&x);
2032: VecRestoreArrayRead(bb,&b);
2033: return(0);
2034: }
2035: if (flag & SOR_ZERO_INITIAL_GUESS) {
2036: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
2037: for (i=0; i<m; i++) {
2038: n = diag[i] - a->i[i];
2039: idx = a->j + a->i[i];
2040: v = a->a + a->i[i];
2041: sum = b[i];
2042: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2043: t[i] = sum;
2044: x[i] = sum*idiag[i];
2045: }
2046: xb = t;
2047: PetscLogFlops(a->nz);
2048: } else xb = b;
2049: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2050: for (i=m-1; i>=0; i--) {
2051: n = a->i[i+1] - diag[i] - 1;
2052: idx = a->j + diag[i] + 1;
2053: v = a->a + diag[i] + 1;
2054: sum = xb[i];
2055: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2056: if (xb == b) {
2057: x[i] = sum*idiag[i];
2058: } else {
2059: x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */
2060: }
2061: }
2062: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
2063: }
2064: its--;
2065: }
2066: while (its--) {
2067: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
2068: for (i=0; i<m; i++) {
2069: /* lower */
2070: n = diag[i] - a->i[i];
2071: idx = a->j + a->i[i];
2072: v = a->a + a->i[i];
2073: sum = b[i];
2074: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2075: t[i] = sum; /* save application of the lower-triangular part */
2076: /* upper */
2077: n = a->i[i+1] - diag[i] - 1;
2078: idx = a->j + diag[i] + 1;
2079: v = a->a + diag[i] + 1;
2080: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2081: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
2082: }
2083: xb = t;
2084: PetscLogFlops(2.0*a->nz);
2085: } else xb = b;
2086: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2087: for (i=m-1; i>=0; i--) {
2088: sum = xb[i];
2089: if (xb == b) {
2090: /* whole matrix (no checkpointing available) */
2091: n = a->i[i+1] - a->i[i];
2092: idx = a->j + a->i[i];
2093: v = a->a + a->i[i];
2094: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2095: x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i];
2096: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
2097: n = a->i[i+1] - diag[i] - 1;
2098: idx = a->j + diag[i] + 1;
2099: v = a->a + diag[i] + 1;
2100: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2101: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
2102: }
2103: }
2104: if (xb == b) {
2105: PetscLogFlops(2.0*a->nz);
2106: } else {
2107: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
2108: }
2109: }
2110: }
2111: VecRestoreArray(xx,&x);
2112: VecRestoreArrayRead(bb,&b);
2113: return(0);
2114: }
2117: PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info)
2118: {
2119: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2122: info->block_size = 1.0;
2123: info->nz_allocated = a->maxnz;
2124: info->nz_used = a->nz;
2125: info->nz_unneeded = (a->maxnz - a->nz);
2126: info->assemblies = A->num_ass;
2127: info->mallocs = A->info.mallocs;
2128: info->memory = ((PetscObject)A)->mem;
2129: if (A->factortype) {
2130: info->fill_ratio_given = A->info.fill_ratio_given;
2131: info->fill_ratio_needed = A->info.fill_ratio_needed;
2132: info->factor_mallocs = A->info.factor_mallocs;
2133: } else {
2134: info->fill_ratio_given = 0;
2135: info->fill_ratio_needed = 0;
2136: info->factor_mallocs = 0;
2137: }
2138: return(0);
2139: }
2141: PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
2142: {
2143: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2144: PetscInt i,m = A->rmap->n - 1;
2145: PetscErrorCode ierr;
2146: const PetscScalar *xx;
2147: PetscScalar *bb;
2148: PetscInt d = 0;
2151: if (x && b) {
2152: VecGetArrayRead(x,&xx);
2153: VecGetArray(b,&bb);
2154: for (i=0; i<N; i++) {
2155: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2156: if (rows[i] >= A->cmap->n) continue;
2157: bb[rows[i]] = diag*xx[rows[i]];
2158: }
2159: VecRestoreArrayRead(x,&xx);
2160: VecRestoreArray(b,&bb);
2161: }
2163: if (a->keepnonzeropattern) {
2164: for (i=0; i<N; i++) {
2165: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2166: PetscArrayzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]);
2167: }
2168: if (diag != 0.0) {
2169: for (i=0; i<N; i++) {
2170: d = rows[i];
2171: if (rows[i] >= A->cmap->n) continue;
2172: 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);
2173: }
2174: for (i=0; i<N; i++) {
2175: if (rows[i] >= A->cmap->n) continue;
2176: a->a[a->diag[rows[i]]] = diag;
2177: }
2178: }
2179: } else {
2180: if (diag != 0.0) {
2181: for (i=0; i<N; i++) {
2182: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2183: if (a->ilen[rows[i]] > 0) {
2184: if (rows[i] >= A->cmap->n) {
2185: a->ilen[rows[i]] = 0;
2186: } else {
2187: a->ilen[rows[i]] = 1;
2188: a->a[a->i[rows[i]]] = diag;
2189: a->j[a->i[rows[i]]] = rows[i];
2190: }
2191: } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */
2192: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
2193: }
2194: }
2195: } else {
2196: for (i=0; i<N; i++) {
2197: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2198: a->ilen[rows[i]] = 0;
2199: }
2200: }
2201: A->nonzerostate++;
2202: }
2203: #if defined(PETSC_HAVE_DEVICE)
2204: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
2205: #endif
2206: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
2207: return(0);
2208: }
2210: PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
2211: {
2212: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2213: PetscInt i,j,m = A->rmap->n - 1,d = 0;
2214: PetscErrorCode ierr;
2215: PetscBool missing,*zeroed,vecs = PETSC_FALSE;
2216: const PetscScalar *xx;
2217: PetscScalar *bb;
2220: if (x && b) {
2221: VecGetArrayRead(x,&xx);
2222: VecGetArray(b,&bb);
2223: vecs = PETSC_TRUE;
2224: }
2225: PetscCalloc1(A->rmap->n,&zeroed);
2226: for (i=0; i<N; i++) {
2227: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2228: PetscArrayzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]);
2230: zeroed[rows[i]] = PETSC_TRUE;
2231: }
2232: for (i=0; i<A->rmap->n; i++) {
2233: if (!zeroed[i]) {
2234: for (j=a->i[i]; j<a->i[i+1]; j++) {
2235: if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) {
2236: if (vecs) bb[i] -= a->a[j]*xx[a->j[j]];
2237: a->a[j] = 0.0;
2238: }
2239: }
2240: } else if (vecs && i < A->cmap->N) bb[i] = diag*xx[i];
2241: }
2242: if (x && b) {
2243: VecRestoreArrayRead(x,&xx);
2244: VecRestoreArray(b,&bb);
2245: }
2246: PetscFree(zeroed);
2247: if (diag != 0.0) {
2248: MatMissingDiagonal_SeqAIJ(A,&missing,&d);
2249: if (missing) {
2250: for (i=0; i<N; i++) {
2251: if (rows[i] >= A->cmap->N) continue;
2252: if (a->nonew && rows[i] >= d) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D (%D)",d,rows[i]);
2253: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
2254: }
2255: } else {
2256: for (i=0; i<N; i++) {
2257: a->a[a->diag[rows[i]]] = diag;
2258: }
2259: }
2260: }
2261: #if defined(PETSC_HAVE_DEVICE)
2262: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
2263: #endif
2264: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
2265: return(0);
2266: }
2268: PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
2269: {
2270: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2271: PetscInt *itmp;
2274: if (row < 0 || row >= A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range",row);
2276: *nz = a->i[row+1] - a->i[row];
2277: if (v) *v = a->a + a->i[row];
2278: if (idx) {
2279: itmp = a->j + a->i[row];
2280: if (*nz) *idx = itmp;
2281: else *idx = NULL;
2282: }
2283: return(0);
2284: }
2286: /* remove this function? */
2287: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
2288: {
2290: return(0);
2291: }
2293: PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm)
2294: {
2295: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2296: MatScalar *v = a->a;
2297: PetscReal sum = 0.0;
2299: PetscInt i,j;
2302: if (type == NORM_FROBENIUS) {
2303: #if defined(PETSC_USE_REAL___FP16)
2304: PetscBLASInt one = 1,nz = a->nz;
2305: *nrm = BLASnrm2_(&nz,v,&one);
2306: #else
2307: for (i=0; i<a->nz; i++) {
2308: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
2309: }
2310: *nrm = PetscSqrtReal(sum);
2311: #endif
2312: PetscLogFlops(2.0*a->nz);
2313: } else if (type == NORM_1) {
2314: PetscReal *tmp;
2315: PetscInt *jj = a->j;
2316: PetscCalloc1(A->cmap->n+1,&tmp);
2317: *nrm = 0.0;
2318: for (j=0; j<a->nz; j++) {
2319: tmp[*jj++] += PetscAbsScalar(*v); v++;
2320: }
2321: for (j=0; j<A->cmap->n; j++) {
2322: if (tmp[j] > *nrm) *nrm = tmp[j];
2323: }
2324: PetscFree(tmp);
2325: PetscLogFlops(PetscMax(a->nz-1,0));
2326: } else if (type == NORM_INFINITY) {
2327: *nrm = 0.0;
2328: for (j=0; j<A->rmap->n; j++) {
2329: v = a->a + a->i[j];
2330: sum = 0.0;
2331: for (i=0; i<a->i[j+1]-a->i[j]; i++) {
2332: sum += PetscAbsScalar(*v); v++;
2333: }
2334: if (sum > *nrm) *nrm = sum;
2335: }
2336: PetscLogFlops(PetscMax(a->nz-1,0));
2337: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm");
2338: return(0);
2339: }
2341: /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */
2342: PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B)
2343: {
2345: PetscInt i,j,anzj;
2346: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b;
2347: PetscInt an=A->cmap->N,am=A->rmap->N;
2348: PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j;
2351: /* Allocate space for symbolic transpose info and work array */
2352: PetscCalloc1(an+1,&ati);
2353: PetscMalloc1(ai[am],&atj);
2354: PetscMalloc1(an,&atfill);
2356: /* Walk through aj and count ## of non-zeros in each row of A^T. */
2357: /* Note: offset by 1 for fast conversion into csr format. */
2358: for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1;
2359: /* Form ati for csr format of A^T. */
2360: for (i=0;i<an;i++) ati[i+1] += ati[i];
2362: /* Copy ati into atfill so we have locations of the next free space in atj */
2363: PetscArraycpy(atfill,ati,an);
2365: /* Walk through A row-wise and mark nonzero entries of A^T. */
2366: for (i=0;i<am;i++) {
2367: anzj = ai[i+1] - ai[i];
2368: for (j=0;j<anzj;j++) {
2369: atj[atfill[*aj]] = i;
2370: atfill[*aj++] += 1;
2371: }
2372: }
2374: /* Clean up temporary space and complete requests. */
2375: PetscFree(atfill);
2376: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);
2377: MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2378: MatSetType(*B,((PetscObject)A)->type_name);
2380: b = (Mat_SeqAIJ*)((*B)->data);
2381: b->free_a = PETSC_FALSE;
2382: b->free_ij = PETSC_TRUE;
2383: b->nonew = 0;
2384: return(0);
2385: }
2387: PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2388: {
2389: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2390: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2391: MatScalar *va,*vb;
2393: PetscInt ma,na,mb,nb, i;
2396: MatGetSize(A,&ma,&na);
2397: MatGetSize(B,&mb,&nb);
2398: if (ma!=nb || na!=mb) {
2399: *f = PETSC_FALSE;
2400: return(0);
2401: }
2402: aii = aij->i; bii = bij->i;
2403: adx = aij->j; bdx = bij->j;
2404: va = aij->a; vb = bij->a;
2405: PetscMalloc1(ma,&aptr);
2406: PetscMalloc1(mb,&bptr);
2407: for (i=0; i<ma; i++) aptr[i] = aii[i];
2408: for (i=0; i<mb; i++) bptr[i] = bii[i];
2410: *f = PETSC_TRUE;
2411: for (i=0; i<ma; i++) {
2412: while (aptr[i]<aii[i+1]) {
2413: PetscInt idc,idr;
2414: PetscScalar vc,vr;
2415: /* column/row index/value */
2416: idc = adx[aptr[i]];
2417: idr = bdx[bptr[idc]];
2418: vc = va[aptr[i]];
2419: vr = vb[bptr[idc]];
2420: if (i!=idr || PetscAbsScalar(vc-vr) > tol) {
2421: *f = PETSC_FALSE;
2422: goto done;
2423: } else {
2424: aptr[i]++;
2425: if (B || i!=idc) bptr[idc]++;
2426: }
2427: }
2428: }
2429: done:
2430: PetscFree(aptr);
2431: PetscFree(bptr);
2432: return(0);
2433: }
2435: PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2436: {
2437: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2438: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2439: MatScalar *va,*vb;
2441: PetscInt ma,na,mb,nb, i;
2444: MatGetSize(A,&ma,&na);
2445: MatGetSize(B,&mb,&nb);
2446: if (ma!=nb || na!=mb) {
2447: *f = PETSC_FALSE;
2448: return(0);
2449: }
2450: aii = aij->i; bii = bij->i;
2451: adx = aij->j; bdx = bij->j;
2452: va = aij->a; vb = bij->a;
2453: PetscMalloc1(ma,&aptr);
2454: PetscMalloc1(mb,&bptr);
2455: for (i=0; i<ma; i++) aptr[i] = aii[i];
2456: for (i=0; i<mb; i++) bptr[i] = bii[i];
2458: *f = PETSC_TRUE;
2459: for (i=0; i<ma; i++) {
2460: while (aptr[i]<aii[i+1]) {
2461: PetscInt idc,idr;
2462: PetscScalar vc,vr;
2463: /* column/row index/value */
2464: idc = adx[aptr[i]];
2465: idr = bdx[bptr[idc]];
2466: vc = va[aptr[i]];
2467: vr = vb[bptr[idc]];
2468: if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) {
2469: *f = PETSC_FALSE;
2470: goto done;
2471: } else {
2472: aptr[i]++;
2473: if (B || i!=idc) bptr[idc]++;
2474: }
2475: }
2476: }
2477: done:
2478: PetscFree(aptr);
2479: PetscFree(bptr);
2480: return(0);
2481: }
2483: PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2484: {
2488: MatIsTranspose_SeqAIJ(A,A,tol,f);
2489: return(0);
2490: }
2492: PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2493: {
2497: MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);
2498: return(0);
2499: }
2501: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr)
2502: {
2503: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2504: const PetscScalar *l,*r;
2505: PetscScalar x;
2506: MatScalar *v;
2507: PetscErrorCode ierr;
2508: PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz;
2509: const PetscInt *jj;
2512: if (ll) {
2513: /* The local size is used so that VecMPI can be passed to this routine
2514: by MatDiagonalScale_MPIAIJ */
2515: VecGetLocalSize(ll,&m);
2516: if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length");
2517: VecGetArrayRead(ll,&l);
2518: v = a->a;
2519: for (i=0; i<m; i++) {
2520: x = l[i];
2521: M = a->i[i+1] - a->i[i];
2522: for (j=0; j<M; j++) (*v++) *= x;
2523: }
2524: VecRestoreArrayRead(ll,&l);
2525: PetscLogFlops(nz);
2526: }
2527: if (rr) {
2528: VecGetLocalSize(rr,&n);
2529: if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length");
2530: VecGetArrayRead(rr,&r);
2531: v = a->a; jj = a->j;
2532: for (i=0; i<nz; i++) (*v++) *= r[*jj++];
2533: VecRestoreArrayRead(rr,&r);
2534: PetscLogFlops(nz);
2535: }
2536: MatSeqAIJInvalidateDiagonal(A);
2537: #if defined(PETSC_HAVE_DEVICE)
2538: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
2539: #endif
2540: return(0);
2541: }
2543: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B)
2544: {
2545: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c;
2547: PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens;
2548: PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi;
2549: const PetscInt *irow,*icol;
2550: PetscInt nrows,ncols;
2551: PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen;
2552: MatScalar *a_new,*mat_a;
2553: Mat C;
2554: PetscBool stride;
2558: ISGetIndices(isrow,&irow);
2559: ISGetLocalSize(isrow,&nrows);
2560: ISGetLocalSize(iscol,&ncols);
2562: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);
2563: if (stride) {
2564: ISStrideGetInfo(iscol,&first,&step);
2565: } else {
2566: first = 0;
2567: step = 0;
2568: }
2569: if (stride && step == 1) {
2570: /* special case of contiguous rows */
2571: PetscMalloc2(nrows,&lens,nrows,&starts);
2572: /* loop over new rows determining lens and starting points */
2573: for (i=0; i<nrows; i++) {
2574: kstart = ai[irow[i]];
2575: kend = kstart + ailen[irow[i]];
2576: starts[i] = kstart;
2577: for (k=kstart; k<kend; k++) {
2578: if (aj[k] >= first) {
2579: starts[i] = k;
2580: break;
2581: }
2582: }
2583: sum = 0;
2584: while (k < kend) {
2585: if (aj[k++] >= first+ncols) break;
2586: sum++;
2587: }
2588: lens[i] = sum;
2589: }
2590: /* create submatrix */
2591: if (scall == MAT_REUSE_MATRIX) {
2592: PetscInt n_cols,n_rows;
2593: MatGetSize(*B,&n_rows,&n_cols);
2594: if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
2595: MatZeroEntries(*B);
2596: C = *B;
2597: } else {
2598: PetscInt rbs,cbs;
2599: MatCreate(PetscObjectComm((PetscObject)A),&C);
2600: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2601: ISGetBlockSize(isrow,&rbs);
2602: ISGetBlockSize(iscol,&cbs);
2603: MatSetBlockSizes(C,rbs,cbs);
2604: MatSetType(C,((PetscObject)A)->type_name);
2605: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2606: }
2607: c = (Mat_SeqAIJ*)C->data;
2609: /* loop over rows inserting into submatrix */
2610: a_new = c->a;
2611: j_new = c->j;
2612: i_new = c->i;
2614: for (i=0; i<nrows; i++) {
2615: ii = starts[i];
2616: lensi = lens[i];
2617: for (k=0; k<lensi; k++) {
2618: *j_new++ = aj[ii+k] - first;
2619: }
2620: PetscArraycpy(a_new,a->a + starts[i],lensi);
2621: a_new += lensi;
2622: i_new[i+1] = i_new[i] + lensi;
2623: c->ilen[i] = lensi;
2624: }
2625: PetscFree2(lens,starts);
2626: } else {
2627: ISGetIndices(iscol,&icol);
2628: PetscCalloc1(oldcols,&smap);
2629: PetscMalloc1(1+nrows,&lens);
2630: for (i=0; i<ncols; i++) {
2631: if (PetscUnlikelyDebug(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);
2632: smap[icol[i]] = i+1;
2633: }
2635: /* determine lens of each row */
2636: for (i=0; i<nrows; i++) {
2637: kstart = ai[irow[i]];
2638: kend = kstart + a->ilen[irow[i]];
2639: lens[i] = 0;
2640: for (k=kstart; k<kend; k++) {
2641: if (smap[aj[k]]) {
2642: lens[i]++;
2643: }
2644: }
2645: }
2646: /* Create and fill new matrix */
2647: if (scall == MAT_REUSE_MATRIX) {
2648: PetscBool equal;
2650: c = (Mat_SeqAIJ*)((*B)->data);
2651: if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
2652: PetscArraycmp(c->ilen,lens,(*B)->rmap->n,&equal);
2653: if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
2654: PetscArrayzero(c->ilen,(*B)->rmap->n);
2655: C = *B;
2656: } else {
2657: PetscInt rbs,cbs;
2658: MatCreate(PetscObjectComm((PetscObject)A),&C);
2659: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2660: ISGetBlockSize(isrow,&rbs);
2661: ISGetBlockSize(iscol,&cbs);
2662: MatSetBlockSizes(C,rbs,cbs);
2663: MatSetType(C,((PetscObject)A)->type_name);
2664: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2665: }
2666: c = (Mat_SeqAIJ*)(C->data);
2667: for (i=0; i<nrows; i++) {
2668: row = irow[i];
2669: kstart = ai[row];
2670: kend = kstart + a->ilen[row];
2671: mat_i = c->i[i];
2672: mat_j = c->j + mat_i;
2673: mat_a = c->a + mat_i;
2674: mat_ilen = c->ilen + i;
2675: for (k=kstart; k<kend; k++) {
2676: if ((tcol=smap[a->j[k]])) {
2677: *mat_j++ = tcol - 1;
2678: *mat_a++ = a->a[k];
2679: (*mat_ilen)++;
2681: }
2682: }
2683: }
2684: /* Free work space */
2685: ISRestoreIndices(iscol,&icol);
2686: PetscFree(smap);
2687: PetscFree(lens);
2688: /* sort */
2689: for (i = 0; i < nrows; i++) {
2690: PetscInt ilen;
2692: mat_i = c->i[i];
2693: mat_j = c->j + mat_i;
2694: mat_a = c->a + mat_i;
2695: ilen = c->ilen[i];
2696: PetscSortIntWithScalarArray(ilen,mat_j,mat_a);
2697: }
2698: }
2699: #if defined(PETSC_HAVE_DEVICE)
2700: MatBindToCPU(C,A->boundtocpu);
2701: #endif
2702: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2703: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2705: ISRestoreIndices(isrow,&irow);
2706: *B = C;
2707: return(0);
2708: }
2710: PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat)
2711: {
2713: Mat B;
2716: if (scall == MAT_INITIAL_MATRIX) {
2717: MatCreate(subComm,&B);
2718: MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);
2719: MatSetBlockSizesFromMats(B,mat,mat);
2720: MatSetType(B,MATSEQAIJ);
2721: MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);
2722: *subMat = B;
2723: } else {
2724: MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);
2725: }
2726: return(0);
2727: }
2729: PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info)
2730: {
2731: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2733: Mat outA;
2734: PetscBool row_identity,col_identity;
2737: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu");
2739: ISIdentity(row,&row_identity);
2740: ISIdentity(col,&col_identity);
2742: outA = inA;
2743: outA->factortype = MAT_FACTOR_LU;
2744: PetscFree(inA->solvertype);
2745: PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);
2747: PetscObjectReference((PetscObject)row);
2748: ISDestroy(&a->row);
2750: a->row = row;
2752: PetscObjectReference((PetscObject)col);
2753: ISDestroy(&a->col);
2755: a->col = col;
2757: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2758: ISDestroy(&a->icol);
2759: ISInvertPermutation(col,PETSC_DECIDE,&a->icol);
2760: PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);
2762: if (!a->solve_work) { /* this matrix may have been factored before */
2763: PetscMalloc1(inA->rmap->n+1,&a->solve_work);
2764: PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));
2765: }
2767: MatMarkDiagonal_SeqAIJ(inA);
2768: if (row_identity && col_identity) {
2769: MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);
2770: } else {
2771: MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);
2772: }
2773: return(0);
2774: }
2776: PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha)
2777: {
2778: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2779: PetscScalar oalpha = alpha;
2781: PetscBLASInt one = 1,bnz;
2784: PetscBLASIntCast(a->nz,&bnz);
2785: PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&oalpha,a->a,&one));
2786: PetscLogFlops(a->nz);
2787: MatSeqAIJInvalidateDiagonal(inA);
2788: #if defined(PETSC_HAVE_DEVICE)
2789: if (inA->offloadmask != PETSC_OFFLOAD_UNALLOCATED) inA->offloadmask = PETSC_OFFLOAD_CPU;
2790: #endif
2791: return(0);
2792: }
2794: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2795: {
2797: PetscInt i;
2800: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2801: PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);
2803: for (i=0; i<submatj->nrqr; ++i) {
2804: PetscFree(submatj->sbuf2[i]);
2805: }
2806: PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);
2808: if (submatj->rbuf1) {
2809: PetscFree(submatj->rbuf1[0]);
2810: PetscFree(submatj->rbuf1);
2811: }
2813: for (i=0; i<submatj->nrqs; ++i) {
2814: PetscFree(submatj->rbuf3[i]);
2815: }
2816: PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);
2817: PetscFree(submatj->pa);
2818: }
2820: #if defined(PETSC_USE_CTABLE)
2821: PetscTableDestroy((PetscTable*)&submatj->rmap);
2822: if (submatj->cmap_loc) {PetscFree(submatj->cmap_loc);}
2823: PetscFree(submatj->rmap_loc);
2824: #else
2825: PetscFree(submatj->rmap);
2826: #endif
2828: if (!submatj->allcolumns) {
2829: #if defined(PETSC_USE_CTABLE)
2830: PetscTableDestroy((PetscTable*)&submatj->cmap);
2831: #else
2832: PetscFree(submatj->cmap);
2833: #endif
2834: }
2835: PetscFree(submatj->row2proc);
2837: PetscFree(submatj);
2838: return(0);
2839: }
2841: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2842: {
2844: Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data;
2845: Mat_SubSppt *submatj = c->submatis1;
2848: (*submatj->destroy)(C);
2849: MatDestroySubMatrix_Private(submatj);
2850: return(0);
2851: }
2853: PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n,Mat *mat[])
2854: {
2856: PetscInt i;
2857: Mat C;
2858: Mat_SeqAIJ *c;
2859: Mat_SubSppt *submatj;
2862: for (i=0; i<n; i++) {
2863: C = (*mat)[i];
2864: c = (Mat_SeqAIJ*)C->data;
2865: submatj = c->submatis1;
2866: if (submatj) {
2867: if (--((PetscObject)C)->refct <= 0) {
2868: (*submatj->destroy)(C);
2869: MatDestroySubMatrix_Private(submatj);
2870: PetscFree(C->defaultvectype);
2871: PetscLayoutDestroy(&C->rmap);
2872: PetscLayoutDestroy(&C->cmap);
2873: PetscHeaderDestroy(&C);
2874: }
2875: } else {
2876: MatDestroy(&C);
2877: }
2878: }
2880: /* Destroy Dummy submatrices created for reuse */
2881: MatDestroySubMatrices_Dummy(n,mat);
2883: PetscFree(*mat);
2884: return(0);
2885: }
2887: PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
2888: {
2890: PetscInt i;
2893: if (scall == MAT_INITIAL_MATRIX) {
2894: PetscCalloc1(n+1,B);
2895: }
2897: for (i=0; i<n; i++) {
2898: MatCreateSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);
2899: }
2900: return(0);
2901: }
2903: PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov)
2904: {
2905: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2907: PetscInt row,i,j,k,l,m,n,*nidx,isz,val;
2908: const PetscInt *idx;
2909: PetscInt start,end,*ai,*aj;
2910: PetscBT table;
2913: m = A->rmap->n;
2914: ai = a->i;
2915: aj = a->j;
2917: if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used");
2919: PetscMalloc1(m+1,&nidx);
2920: PetscBTCreate(m,&table);
2922: for (i=0; i<is_max; i++) {
2923: /* Initialize the two local arrays */
2924: isz = 0;
2925: PetscBTMemzero(m,table);
2927: /* Extract the indices, assume there can be duplicate entries */
2928: ISGetIndices(is[i],&idx);
2929: ISGetLocalSize(is[i],&n);
2931: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2932: for (j=0; j<n; ++j) {
2933: if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j];
2934: }
2935: ISRestoreIndices(is[i],&idx);
2936: ISDestroy(&is[i]);
2938: k = 0;
2939: for (j=0; j<ov; j++) { /* for each overlap */
2940: n = isz;
2941: for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */
2942: row = nidx[k];
2943: start = ai[row];
2944: end = ai[row+1];
2945: for (l = start; l<end; l++) {
2946: val = aj[l];
2947: if (!PetscBTLookupSet(table,val)) nidx[isz++] = val;
2948: }
2949: }
2950: }
2951: ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));
2952: }
2953: PetscBTDestroy(&table);
2954: PetscFree(nidx);
2955: return(0);
2956: }
2958: /* -------------------------------------------------------------- */
2959: PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B)
2960: {
2961: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2963: PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n;
2964: const PetscInt *row,*col;
2965: PetscInt *cnew,j,*lens;
2966: IS icolp,irowp;
2967: PetscInt *cwork = NULL;
2968: PetscScalar *vwork = NULL;
2971: ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);
2972: ISGetIndices(irowp,&row);
2973: ISInvertPermutation(colp,PETSC_DECIDE,&icolp);
2974: ISGetIndices(icolp,&col);
2976: /* determine lengths of permuted rows */
2977: PetscMalloc1(m+1,&lens);
2978: for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i];
2979: MatCreate(PetscObjectComm((PetscObject)A),B);
2980: MatSetSizes(*B,m,n,m,n);
2981: MatSetBlockSizesFromMats(*B,A,A);
2982: MatSetType(*B,((PetscObject)A)->type_name);
2983: MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);
2984: PetscFree(lens);
2986: PetscMalloc1(n,&cnew);
2987: for (i=0; i<m; i++) {
2988: MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2989: for (j=0; j<nz; j++) cnew[j] = col[cwork[j]];
2990: MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);
2991: MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2992: }
2993: PetscFree(cnew);
2995: (*B)->assembled = PETSC_FALSE;
2997: #if defined(PETSC_HAVE_DEVICE)
2998: MatBindToCPU(*B,A->boundtocpu);
2999: #endif
3000: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
3001: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
3002: ISRestoreIndices(irowp,&row);
3003: ISRestoreIndices(icolp,&col);
3004: ISDestroy(&irowp);
3005: ISDestroy(&icolp);
3006: if (rowp == colp) {
3007: if (A->symmetric) {
3008: MatSetOption(*B,MAT_SYMMETRIC,PETSC_TRUE);
3009: }
3010: if (A->hermitian) {
3011: MatSetOption(*B,MAT_HERMITIAN,PETSC_TRUE);
3012: }
3013: }
3014: return(0);
3015: }
3017: PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str)
3018: {
3022: /* If the two matrices have the same copy implementation, use fast copy. */
3023: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
3024: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3025: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
3027: if (a->i[A->rmap->n] != b->i[B->rmap->n]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number of nonzeros in two matrices are different %D != %D",a->i[A->rmap->n],b->i[B->rmap->n]);
3028: PetscArraycpy(b->a,a->a,a->i[A->rmap->n]);
3029: PetscObjectStateIncrease((PetscObject)B);
3030: } else {
3031: MatCopy_Basic(A,B,str);
3032: }
3033: return(0);
3034: }
3036: PetscErrorCode MatSetUp_SeqAIJ(Mat A)
3037: {
3041: MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,NULL);
3042: return(0);
3043: }
3045: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[])
3046: {
3047: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3050: *array = a->a;
3051: return(0);
3052: }
3054: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[])
3055: {
3057: *array = NULL;
3058: return(0);
3059: }
3061: /*
3062: Computes the number of nonzeros per row needed for preallocation when X and Y
3063: have different nonzero structure.
3064: */
3065: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz)
3066: {
3067: PetscInt i,j,k,nzx,nzy;
3070: /* Set the number of nonzeros in the new matrix */
3071: for (i=0; i<m; i++) {
3072: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
3073: nzx = xi[i+1] - xi[i];
3074: nzy = yi[i+1] - yi[i];
3075: nnz[i] = 0;
3076: for (j=0,k=0; j<nzx; j++) { /* Point in X */
3077: for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */
3078: if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */
3079: nnz[i]++;
3080: }
3081: for (; k<nzy; k++) nnz[i]++;
3082: }
3083: return(0);
3084: }
3086: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz)
3087: {
3088: PetscInt m = Y->rmap->N;
3089: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
3090: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
3094: /* Set the number of nonzeros in the new matrix */
3095: MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);
3096: return(0);
3097: }
3099: PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
3100: {
3102: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data;
3105: if (str == DIFFERENT_NONZERO_PATTERN) {
3106: if (x->nz == y->nz) {
3107: PetscBool e;
3108: PetscArraycmp(x->i,y->i,Y->rmap->n+1,&e);
3109: if (e) {
3110: PetscArraycmp(x->j,y->j,y->nz,&e);
3111: if (e) {
3112: str = SAME_NONZERO_PATTERN;
3113: }
3114: }
3115: }
3116: }
3117: if (str == SAME_NONZERO_PATTERN) {
3118: PetscScalar alpha = a;
3119: PetscBLASInt one = 1,bnz;
3121: PetscBLASIntCast(x->nz,&bnz);
3122: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
3123: MatSeqAIJInvalidateDiagonal(Y);
3124: PetscObjectStateIncrease((PetscObject)Y);
3125: /* the MatAXPY_Basic* subroutines calls MatAssembly, so the matrix on the GPU will be updated */
3126: #if defined(PETSC_HAVE_DEVICE)
3127: if (Y->offloadmask != PETSC_OFFLOAD_UNALLOCATED) {
3128: Y->offloadmask = PETSC_OFFLOAD_CPU;
3129: }
3130: #endif
3131: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
3132: MatAXPY_Basic(Y,a,X,str);
3133: } else {
3134: Mat B;
3135: PetscInt *nnz;
3136: PetscMalloc1(Y->rmap->N,&nnz);
3137: MatCreate(PetscObjectComm((PetscObject)Y),&B);
3138: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
3139: MatSetLayouts(B,Y->rmap,Y->cmap);
3140: MatSetType(B,(MatType) ((PetscObject)Y)->type_name);
3141: MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);
3142: MatSeqAIJSetPreallocation(B,0,nnz);
3143: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
3144: MatHeaderReplace(Y,&B);
3145: PetscFree(nnz);
3146: }
3147: return(0);
3148: }
3150: PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
3151: {
3152: #if defined(PETSC_USE_COMPLEX)
3153: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3154: PetscInt i,nz;
3155: PetscScalar *a;
3158: nz = aij->nz;
3159: a = aij->a;
3160: for (i=0; i<nz; i++) a[i] = PetscConj(a[i]);
3161: #if defined(PETSC_HAVE_DEVICE)
3162: if (mat->offloadmask != PETSC_OFFLOAD_UNALLOCATED) mat->offloadmask = PETSC_OFFLOAD_CPU;
3163: #endif
3164: #else
3166: #endif
3167: return(0);
3168: }
3170: PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3171: {
3172: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3174: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
3175: PetscReal atmp;
3176: PetscScalar *x;
3177: MatScalar *aa;
3180: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3181: aa = a->a;
3182: ai = a->i;
3183: aj = a->j;
3185: VecSet(v,0.0);
3186: VecGetArrayWrite(v,&x);
3187: VecGetLocalSize(v,&n);
3188: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
3189: for (i=0; i<m; i++) {
3190: ncols = ai[1] - ai[0]; ai++;
3191: for (j=0; j<ncols; j++) {
3192: atmp = PetscAbsScalar(*aa);
3193: if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
3194: aa++; aj++;
3195: }
3196: }
3197: VecRestoreArrayWrite(v,&x);
3198: return(0);
3199: }
3201: PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3202: {
3203: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3205: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
3206: PetscScalar *x;
3207: MatScalar *aa;
3210: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3211: aa = a->a;
3212: ai = a->i;
3213: aj = a->j;
3215: VecSet(v,0.0);
3216: VecGetArrayWrite(v,&x);
3217: VecGetLocalSize(v,&n);
3218: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
3219: for (i=0; i<m; i++) {
3220: ncols = ai[1] - ai[0]; ai++;
3221: if (ncols == A->cmap->n) { /* row is dense */
3222: x[i] = *aa; if (idx) idx[i] = 0;
3223: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
3224: x[i] = 0.0;
3225: if (idx) {
3226: for (j=0; j<ncols; j++) { /* find first implicit 0.0 in the row */
3227: if (aj[j] > j) {
3228: idx[i] = j;
3229: break;
3230: }
3231: }
3232: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3233: if (j==ncols && j < A->cmap->n) idx[i] = j;
3234: }
3235: }
3236: for (j=0; j<ncols; j++) {
3237: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
3238: aa++; aj++;
3239: }
3240: }
3241: VecRestoreArrayWrite(v,&x);
3242: return(0);
3243: }
3245: PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3246: {
3247: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3249: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
3250: PetscScalar *x,*aa;
3253: aa = a->a;
3254: ai = a->i;
3255: aj = a->j;
3257: VecSet(v,0.0);
3258: VecGetArrayWrite(v,&x);
3259: VecGetLocalSize(v,&n);
3260: if (n != m) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector, %D vs. %D rows", m, n);
3261: for (i=0; i<m; i++) {
3262: ncols = ai[1] - ai[0]; ai++;
3263: if (ncols == A->cmap->n) { /* row is dense */
3264: x[i] = *aa; if (idx) idx[i] = 0;
3265: } else { /* row is sparse so already KNOW minimum is 0.0 or higher */
3266: x[i] = 0.0;
3267: if (idx) { /* find first implicit 0.0 in the row */
3268: for (j=0; j<ncols; j++) {
3269: if (aj[j] > j) {
3270: idx[i] = j;
3271: break;
3272: }
3273: }
3274: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3275: if (j==ncols && j < A->cmap->n) idx[i] = j;
3276: }
3277: }
3278: for (j=0; j<ncols; j++) {
3279: if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
3280: aa++; aj++;
3281: }
3282: }
3283: VecRestoreArrayWrite(v,&x);
3284: return(0);
3285: }
3287: PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3288: {
3289: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3290: PetscErrorCode ierr;
3291: PetscInt i,j,m = A->rmap->n,ncols,n;
3292: const PetscInt *ai,*aj;
3293: PetscScalar *x;
3294: const MatScalar *aa;
3297: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3298: aa = a->a;
3299: ai = a->i;
3300: aj = a->j;
3302: VecSet(v,0.0);
3303: VecGetArrayWrite(v,&x);
3304: VecGetLocalSize(v,&n);
3305: if (n != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
3306: for (i=0; i<m; i++) {
3307: ncols = ai[1] - ai[0]; ai++;
3308: if (ncols == A->cmap->n) { /* row is dense */
3309: x[i] = *aa; if (idx) idx[i] = 0;
3310: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3311: x[i] = 0.0;
3312: if (idx) { /* find first implicit 0.0 in the row */
3313: for (j=0; j<ncols; j++) {
3314: if (aj[j] > j) {
3315: idx[i] = j;
3316: break;
3317: }
3318: }
3319: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3320: if (j==ncols && j < A->cmap->n) idx[i] = j;
3321: }
3322: }
3323: for (j=0; j<ncols; j++) {
3324: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
3325: aa++; aj++;
3326: }
3327: }
3328: VecRestoreArrayWrite(v,&x);
3329: return(0);
3330: }
3332: PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values)
3333: {
3334: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
3335: PetscErrorCode ierr;
3336: PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j;
3337: MatScalar *diag,work[25],*v_work;
3338: const PetscReal shift = 0.0;
3339: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
3342: allowzeropivot = PetscNot(A->erroriffailure);
3343: if (a->ibdiagvalid) {
3344: if (values) *values = a->ibdiag;
3345: return(0);
3346: }
3347: MatMarkDiagonal_SeqAIJ(A);
3348: if (!a->ibdiag) {
3349: PetscMalloc1(bs2*mbs,&a->ibdiag);
3350: PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));
3351: }
3352: diag = a->ibdiag;
3353: if (values) *values = a->ibdiag;
3354: /* factor and invert each block */
3355: switch (bs) {
3356: case 1:
3357: for (i=0; i<mbs; i++) {
3358: MatGetValues(A,1,&i,1,&i,diag+i);
3359: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3360: if (allowzeropivot) {
3361: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3362: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3363: A->factorerror_zeropivot_row = i;
3364: PetscInfo3(A,"Zero pivot, row %D pivot %g tolerance %g\n",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);
3365: } 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);
3366: }
3367: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3368: }
3369: break;
3370: case 2:
3371: for (i=0; i<mbs; i++) {
3372: ij[0] = 2*i; ij[1] = 2*i + 1;
3373: MatGetValues(A,2,ij,2,ij,diag);
3374: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
3375: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3376: PetscKernel_A_gets_transpose_A_2(diag);
3377: diag += 4;
3378: }
3379: break;
3380: case 3:
3381: for (i=0; i<mbs; i++) {
3382: ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2;
3383: MatGetValues(A,3,ij,3,ij,diag);
3384: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
3385: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3386: PetscKernel_A_gets_transpose_A_3(diag);
3387: diag += 9;
3388: }
3389: break;
3390: case 4:
3391: for (i=0; i<mbs; i++) {
3392: ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3;
3393: MatGetValues(A,4,ij,4,ij,diag);
3394: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
3395: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3396: PetscKernel_A_gets_transpose_A_4(diag);
3397: diag += 16;
3398: }
3399: break;
3400: case 5:
3401: for (i=0; i<mbs; i++) {
3402: ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4;
3403: MatGetValues(A,5,ij,5,ij,diag);
3404: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
3405: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3406: PetscKernel_A_gets_transpose_A_5(diag);
3407: diag += 25;
3408: }
3409: break;
3410: case 6:
3411: for (i=0; i<mbs; i++) {
3412: 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;
3413: MatGetValues(A,6,ij,6,ij,diag);
3414: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
3415: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3416: PetscKernel_A_gets_transpose_A_6(diag);
3417: diag += 36;
3418: }
3419: break;
3420: case 7:
3421: for (i=0; i<mbs; i++) {
3422: 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;
3423: MatGetValues(A,7,ij,7,ij,diag);
3424: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
3425: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3426: PetscKernel_A_gets_transpose_A_7(diag);
3427: diag += 49;
3428: }
3429: break;
3430: default:
3431: PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);
3432: for (i=0; i<mbs; i++) {
3433: for (j=0; j<bs; j++) {
3434: IJ[j] = bs*i + j;
3435: }
3436: MatGetValues(A,bs,IJ,bs,IJ,diag);
3437: PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
3438: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3439: PetscKernel_A_gets_transpose_A_N(diag,bs);
3440: diag += bs2;
3441: }
3442: PetscFree3(v_work,v_pivots,IJ);
3443: }
3444: a->ibdiagvalid = PETSC_TRUE;
3445: return(0);
3446: }
3448: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx)
3449: {
3451: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3452: PetscScalar a;
3453: PetscInt m,n,i,j,col;
3456: if (!x->assembled) {
3457: MatGetSize(x,&m,&n);
3458: for (i=0; i<m; i++) {
3459: for (j=0; j<aij->imax[i]; j++) {
3460: PetscRandomGetValue(rctx,&a);
3461: col = (PetscInt)(n*PetscRealPart(a));
3462: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3463: }
3464: }
3465: } else {
3466: for (i=0; i<aij->nz; i++) {PetscRandomGetValue(rctx,aij->a+i);}
3467: }
3468: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3469: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3470: return(0);
3471: }
3473: /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */
3474: PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x,PetscInt low,PetscInt high,PetscRandom rctx)
3475: {
3477: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3478: PetscScalar a;
3479: PetscInt m,n,i,j,col,nskip;
3482: nskip = high - low;
3483: MatGetSize(x,&m,&n);
3484: n -= nskip; /* shrink number of columns where nonzeros can be set */
3485: for (i=0; i<m; i++) {
3486: for (j=0; j<aij->imax[i]; j++) {
3487: PetscRandomGetValue(rctx,&a);
3488: col = (PetscInt)(n*PetscRealPart(a));
3489: if (col >= low) col += nskip; /* shift col rightward to skip the hole */
3490: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3491: }
3492: }
3493: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3494: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3495: return(0);
3496: }
3499: /* -------------------------------------------------------------------*/
3500: static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ,
3501: MatGetRow_SeqAIJ,
3502: MatRestoreRow_SeqAIJ,
3503: MatMult_SeqAIJ,
3504: /* 4*/ MatMultAdd_SeqAIJ,
3505: MatMultTranspose_SeqAIJ,
3506: MatMultTransposeAdd_SeqAIJ,
3507: NULL,
3508: NULL,
3509: NULL,
3510: /* 10*/ NULL,
3511: MatLUFactor_SeqAIJ,
3512: NULL,
3513: MatSOR_SeqAIJ,
3514: MatTranspose_SeqAIJ,
3515: /*1 5*/ MatGetInfo_SeqAIJ,
3516: MatEqual_SeqAIJ,
3517: MatGetDiagonal_SeqAIJ,
3518: MatDiagonalScale_SeqAIJ,
3519: MatNorm_SeqAIJ,
3520: /* 20*/ NULL,
3521: MatAssemblyEnd_SeqAIJ,
3522: MatSetOption_SeqAIJ,
3523: MatZeroEntries_SeqAIJ,
3524: /* 24*/ MatZeroRows_SeqAIJ,
3525: NULL,
3526: NULL,
3527: NULL,
3528: NULL,
3529: /* 29*/ MatSetUp_SeqAIJ,
3530: NULL,
3531: NULL,
3532: NULL,
3533: NULL,
3534: /* 34*/ MatDuplicate_SeqAIJ,
3535: NULL,
3536: NULL,
3537: MatILUFactor_SeqAIJ,
3538: NULL,
3539: /* 39*/ MatAXPY_SeqAIJ,
3540: MatCreateSubMatrices_SeqAIJ,
3541: MatIncreaseOverlap_SeqAIJ,
3542: MatGetValues_SeqAIJ,
3543: MatCopy_SeqAIJ,
3544: /* 44*/ MatGetRowMax_SeqAIJ,
3545: MatScale_SeqAIJ,
3546: MatShift_SeqAIJ,
3547: MatDiagonalSet_SeqAIJ,
3548: MatZeroRowsColumns_SeqAIJ,
3549: /* 49*/ MatSetRandom_SeqAIJ,
3550: MatGetRowIJ_SeqAIJ,
3551: MatRestoreRowIJ_SeqAIJ,
3552: MatGetColumnIJ_SeqAIJ,
3553: MatRestoreColumnIJ_SeqAIJ,
3554: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3555: NULL,
3556: NULL,
3557: MatPermute_SeqAIJ,
3558: NULL,
3559: /* 59*/ NULL,
3560: MatDestroy_SeqAIJ,
3561: MatView_SeqAIJ,
3562: NULL,
3563: NULL,
3564: /* 64*/ NULL,
3565: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3566: NULL,
3567: NULL,
3568: NULL,
3569: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3570: MatGetRowMinAbs_SeqAIJ,
3571: NULL,
3572: NULL,
3573: NULL,
3574: /* 74*/ NULL,
3575: MatFDColoringApply_AIJ,
3576: NULL,
3577: NULL,
3578: NULL,
3579: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3580: NULL,
3581: NULL,
3582: NULL,
3583: MatLoad_SeqAIJ,
3584: /* 84*/ MatIsSymmetric_SeqAIJ,
3585: MatIsHermitian_SeqAIJ,
3586: NULL,
3587: NULL,
3588: NULL,
3589: /* 89*/ NULL,
3590: NULL,
3591: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3592: NULL,
3593: NULL,
3594: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3595: NULL,
3596: NULL,
3597: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3598: NULL,
3599: /* 99*/ MatProductSetFromOptions_SeqAIJ,
3600: NULL,
3601: NULL,
3602: MatConjugate_SeqAIJ,
3603: NULL,
3604: /*104*/ MatSetValuesRow_SeqAIJ,
3605: MatRealPart_SeqAIJ,
3606: MatImaginaryPart_SeqAIJ,
3607: NULL,
3608: NULL,
3609: /*109*/ MatMatSolve_SeqAIJ,
3610: NULL,
3611: MatGetRowMin_SeqAIJ,
3612: NULL,
3613: MatMissingDiagonal_SeqAIJ,
3614: /*114*/ NULL,
3615: NULL,
3616: NULL,
3617: NULL,
3618: NULL,
3619: /*119*/ NULL,
3620: NULL,
3621: NULL,
3622: NULL,
3623: MatGetMultiProcBlock_SeqAIJ,
3624: /*124*/ MatFindNonzeroRows_SeqAIJ,
3625: MatGetColumnNorms_SeqAIJ,
3626: MatInvertBlockDiagonal_SeqAIJ,
3627: MatInvertVariableBlockDiagonal_SeqAIJ,
3628: NULL,
3629: /*129*/ NULL,
3630: NULL,
3631: NULL,
3632: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3633: MatTransposeColoringCreate_SeqAIJ,
3634: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3635: MatTransColoringApplyDenToSp_SeqAIJ,
3636: NULL,
3637: NULL,
3638: MatRARtNumeric_SeqAIJ_SeqAIJ,
3639: /*139*/NULL,
3640: NULL,
3641: NULL,
3642: MatFDColoringSetUp_SeqXAIJ,
3643: MatFindOffBlockDiagonalEntries_SeqAIJ,
3644: MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3645: /*145*/MatDestroySubMatrices_SeqAIJ,
3646: NULL,
3647: NULL
3648: };
3650: PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices)
3651: {
3652: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3653: PetscInt i,nz,n;
3656: nz = aij->maxnz;
3657: n = mat->rmap->n;
3658: for (i=0; i<nz; i++) {
3659: aij->j[i] = indices[i];
3660: }
3661: aij->nz = nz;
3662: for (i=0; i<n; i++) {
3663: aij->ilen[i] = aij->imax[i];
3664: }
3665: return(0);
3666: }
3668: /*
3669: * When a sparse matrix has many zero columns, we should compact them out to save the space
3670: * This happens in MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable()
3671: * */
3672: PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping)
3673: {
3674: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3675: PetscTable gid1_lid1;
3676: PetscTablePosition tpos;
3677: PetscInt gid,lid,i,j,ncols,ec;
3678: PetscInt *garray;
3679: PetscErrorCode ierr;
3684: /* use a table */
3685: PetscTableCreate(mat->rmap->n,mat->cmap->N+1,&gid1_lid1);
3686: ec = 0;
3687: for (i=0; i<mat->rmap->n; i++) {
3688: ncols = aij->i[i+1] - aij->i[i];
3689: for (j=0; j<ncols; j++) {
3690: PetscInt data,gid1 = aij->j[aij->i[i] + j] + 1;
3691: PetscTableFind(gid1_lid1,gid1,&data);
3692: if (!data) {
3693: /* one based table */
3694: PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);
3695: }
3696: }
3697: }
3698: /* form array of columns we need */
3699: PetscMalloc1(ec+1,&garray);
3700: PetscTableGetHeadPosition(gid1_lid1,&tpos);
3701: while (tpos) {
3702: PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);
3703: gid--;
3704: lid--;
3705: garray[lid] = gid;
3706: }
3707: PetscSortInt(ec,garray); /* sort, and rebuild */
3708: PetscTableRemoveAll(gid1_lid1);
3709: for (i=0; i<ec; i++) {
3710: PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);
3711: }
3712: /* compact out the extra columns in B */
3713: for (i=0; i<mat->rmap->n; i++) {
3714: ncols = aij->i[i+1] - aij->i[i];
3715: for (j=0; j<ncols; j++) {
3716: PetscInt gid1 = aij->j[aij->i[i] + j] + 1;
3717: PetscTableFind(gid1_lid1,gid1,&lid);
3718: lid--;
3719: aij->j[aij->i[i] + j] = lid;
3720: }
3721: }
3722: PetscLayoutDestroy(&mat->cmap);
3723: PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat),ec,ec,1,&mat->cmap);
3724: PetscTableDestroy(&gid1_lid1);
3725: ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,mat->cmap->bs,mat->cmap->n,garray,PETSC_OWN_POINTER,mapping);
3726: ISLocalToGlobalMappingSetType(*mapping,ISLOCALTOGLOBALMAPPINGHASH);
3727: return(0);
3728: }
3730: /*@
3731: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3732: in the matrix.
3734: Input Parameters:
3735: + mat - the SeqAIJ matrix
3736: - indices - the column indices
3738: Level: advanced
3740: Notes:
3741: This can be called if you have precomputed the nonzero structure of the
3742: matrix and want to provide it to the matrix object to improve the performance
3743: of the MatSetValues() operation.
3745: You MUST have set the correct numbers of nonzeros per row in the call to
3746: MatCreateSeqAIJ(), and the columns indices MUST be sorted.
3748: MUST be called before any calls to MatSetValues();
3750: The indices should start with zero, not one.
3752: @*/
3753: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices)
3754: {
3760: PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));
3761: return(0);
3762: }
3764: /* ----------------------------------------------------------------------------------------*/
3766: PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3767: {
3768: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3770: size_t nz = aij->i[mat->rmap->n];
3773: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3775: /* allocate space for values if not already there */
3776: if (!aij->saved_values) {
3777: PetscMalloc1(nz+1,&aij->saved_values);
3778: PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));
3779: }
3781: /* copy values over */
3782: PetscArraycpy(aij->saved_values,aij->a,nz);
3783: return(0);
3784: }
3786: /*@
3787: MatStoreValues - Stashes a copy of the matrix values; this allows, for
3788: example, reuse of the linear part of a Jacobian, while recomputing the
3789: nonlinear portion.
3791: Collect on Mat
3793: Input Parameters:
3794: . mat - the matrix (currently only AIJ matrices support this option)
3796: Level: advanced
3798: Common Usage, with SNESSolve():
3799: $ Create Jacobian matrix
3800: $ Set linear terms into matrix
3801: $ Apply boundary conditions to matrix, at this time matrix must have
3802: $ final nonzero structure (i.e. setting the nonlinear terms and applying
3803: $ boundary conditions again will not change the nonzero structure
3804: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3805: $ MatStoreValues(mat);
3806: $ Call SNESSetJacobian() with matrix
3807: $ In your Jacobian routine
3808: $ MatRetrieveValues(mat);
3809: $ Set nonlinear terms in matrix
3811: Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself:
3812: $ // build linear portion of Jacobian
3813: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3814: $ MatStoreValues(mat);
3815: $ loop over nonlinear iterations
3816: $ MatRetrieveValues(mat);
3817: $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3818: $ // call MatAssemblyBegin/End() on matrix
3819: $ Solve linear system with Jacobian
3820: $ endloop
3822: Notes:
3823: Matrix must already be assemblied before calling this routine
3824: Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before
3825: calling this routine.
3827: When this is called multiple times it overwrites the previous set of stored values
3828: and does not allocated additional space.
3830: .seealso: MatRetrieveValues()
3832: @*/
3833: PetscErrorCode MatStoreValues(Mat mat)
3834: {
3839: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3840: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3841: PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));
3842: return(0);
3843: }
3845: PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3846: {
3847: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3849: PetscInt nz = aij->i[mat->rmap->n];
3852: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3853: if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first");
3854: /* copy values over */
3855: PetscArraycpy(aij->a,aij->saved_values,nz);
3856: return(0);
3857: }
3859: /*@
3860: MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for
3861: example, reuse of the linear part of a Jacobian, while recomputing the
3862: nonlinear portion.
3864: Collect on Mat
3866: Input Parameters:
3867: . mat - the matrix (currently only AIJ matrices support this option)
3869: Level: advanced
3871: .seealso: MatStoreValues()
3873: @*/
3874: PetscErrorCode MatRetrieveValues(Mat mat)
3875: {
3880: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3881: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3882: PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));
3883: return(0);
3884: }
3887: /* --------------------------------------------------------------------------------*/
3888: /*@C
3889: MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format
3890: (the default parallel PETSc format). For good matrix assembly performance
3891: the user should preallocate the matrix storage by setting the parameter nz
3892: (or the array nnz). By setting these parameters accurately, performance
3893: during matrix assembly can be increased by more than a factor of 50.
3895: Collective
3897: Input Parameters:
3898: + comm - MPI communicator, set to PETSC_COMM_SELF
3899: . m - number of rows
3900: . n - number of columns
3901: . nz - number of nonzeros per row (same for all rows)
3902: - nnz - array containing the number of nonzeros in the various rows
3903: (possibly different for each row) or NULL
3905: Output Parameter:
3906: . A - the matrix
3908: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3909: MatXXXXSetPreallocation() paradigm instead of this routine directly.
3910: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3912: Notes:
3913: If nnz is given then nz is ignored
3915: The AIJ format (also called the Yale sparse matrix format or
3916: compressed row storage), is fully compatible with standard Fortran 77
3917: storage. That is, the stored row and column indices can begin at
3918: either one (as in Fortran) or zero. See the users' manual for details.
3920: Specify the preallocated storage with either nz or nnz (not both).
3921: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3922: allocation. For large problems you MUST preallocate memory or you
3923: will get TERRIBLE performance, see the users' manual chapter on matrices.
3925: By default, this format uses inodes (identical nodes) when possible, to
3926: improve numerical efficiency of matrix-vector products and solves. We
3927: search for consecutive rows with the same nonzero structure, thereby
3928: reusing matrix information to achieve increased efficiency.
3930: Options Database Keys:
3931: + -mat_no_inode - Do not use inodes
3932: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3934: Level: intermediate
3936: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays()
3938: @*/
3939: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A)
3940: {
3944: MatCreate(comm,A);
3945: MatSetSizes(*A,m,n,m,n);
3946: MatSetType(*A,MATSEQAIJ);
3947: MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);
3948: return(0);
3949: }
3951: /*@C
3952: MatSeqAIJSetPreallocation - For good matrix assembly performance
3953: the user should preallocate the matrix storage by setting the parameter nz
3954: (or the array nnz). By setting these parameters accurately, performance
3955: during matrix assembly can be increased by more than a factor of 50.
3957: Collective
3959: Input Parameters:
3960: + B - The matrix
3961: . nz - number of nonzeros per row (same for all rows)
3962: - nnz - array containing the number of nonzeros in the various rows
3963: (possibly different for each row) or NULL
3965: Notes:
3966: If nnz is given then nz is ignored
3968: The AIJ format (also called the Yale sparse matrix format or
3969: compressed row storage), is fully compatible with standard Fortran 77
3970: storage. That is, the stored row and column indices can begin at
3971: either one (as in Fortran) or zero. See the users' manual for details.
3973: Specify the preallocated storage with either nz or nnz (not both).
3974: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3975: allocation. For large problems you MUST preallocate memory or you
3976: will get TERRIBLE performance, see the users' manual chapter on matrices.
3978: You can call MatGetInfo() to get information on how effective the preallocation was;
3979: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3980: You can also run with the option -info and look for messages with the string
3981: malloc in them to see if additional memory allocation was needed.
3983: Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix
3984: entries or columns indices
3986: By default, this format uses inodes (identical nodes) when possible, to
3987: improve numerical efficiency of matrix-vector products and solves. We
3988: search for consecutive rows with the same nonzero structure, thereby
3989: reusing matrix information to achieve increased efficiency.
3991: Options Database Keys:
3992: + -mat_no_inode - Do not use inodes
3993: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3995: Level: intermediate
3997: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo(),
3998: MatSeqAIJSetTotalPreallocation()
4000: @*/
4001: PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[])
4002: {
4008: PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));
4009: return(0);
4010: }
4012: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz)
4013: {
4014: Mat_SeqAIJ *b;
4015: PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE;
4017: PetscInt i;
4020: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
4021: if (nz == MAT_SKIP_ALLOCATION) {
4022: skipallocation = PETSC_TRUE;
4023: nz = 0;
4024: }
4025: PetscLayoutSetUp(B->rmap);
4026: PetscLayoutSetUp(B->cmap);
4028: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
4029: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz);
4030: if (PetscUnlikelyDebug(nnz)) {
4031: for (i=0; i<B->rmap->n; i++) {
4032: 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]);
4033: 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);
4034: }
4035: }
4037: B->preallocated = PETSC_TRUE;
4039: b = (Mat_SeqAIJ*)B->data;
4041: if (!skipallocation) {
4042: if (!b->imax) {
4043: PetscMalloc1(B->rmap->n,&b->imax);
4044: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
4045: }
4046: if (!b->ilen) {
4047: /* b->ilen will count nonzeros in each row so far. */
4048: PetscCalloc1(B->rmap->n,&b->ilen);
4049: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
4050: } else {
4051: PetscMemzero(b->ilen,B->rmap->n*sizeof(PetscInt));
4052: }
4053: if (!b->ipre) {
4054: PetscMalloc1(B->rmap->n,&b->ipre);
4055: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
4056: }
4057: if (!nnz) {
4058: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
4059: else if (nz < 0) nz = 1;
4060: nz = PetscMin(nz,B->cmap->n);
4061: for (i=0; i<B->rmap->n; i++) b->imax[i] = nz;
4062: nz = nz*B->rmap->n;
4063: } else {
4064: PetscInt64 nz64 = 0;
4065: for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz64 += nnz[i];}
4066: PetscIntCast(nz64,&nz);
4067: }
4069: /* allocate the matrix space */
4070: /* FIXME: should B's old memory be unlogged? */
4071: MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);
4072: if (B->structure_only) {
4073: PetscMalloc1(nz,&b->j);
4074: PetscMalloc1(B->rmap->n+1,&b->i);
4075: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));
4076: } else {
4077: PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);
4078: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));
4079: }
4080: b->i[0] = 0;
4081: for (i=1; i<B->rmap->n+1; i++) {
4082: b->i[i] = b->i[i-1] + b->imax[i-1];
4083: }
4084: if (B->structure_only) {
4085: b->singlemalloc = PETSC_FALSE;
4086: b->free_a = PETSC_FALSE;
4087: } else {
4088: b->singlemalloc = PETSC_TRUE;
4089: b->free_a = PETSC_TRUE;
4090: }
4091: b->free_ij = PETSC_TRUE;
4092: } else {
4093: b->free_a = PETSC_FALSE;
4094: b->free_ij = PETSC_FALSE;
4095: }
4097: if (b->ipre && nnz != b->ipre && b->imax) {
4098: /* reserve user-requested sparsity */
4099: PetscArraycpy(b->ipre,b->imax,B->rmap->n);
4100: }
4103: b->nz = 0;
4104: b->maxnz = nz;
4105: B->info.nz_unneeded = (double)b->maxnz;
4106: if (realalloc) {
4107: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
4108: }
4109: B->was_assembled = PETSC_FALSE;
4110: B->assembled = PETSC_FALSE;
4111: return(0);
4112: }
4115: PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
4116: {
4117: Mat_SeqAIJ *a;
4118: PetscInt i;
4124: /* Check local size. If zero, then return */
4125: if (!A->rmap->n) return(0);
4127: a = (Mat_SeqAIJ*)A->data;
4128: /* if no saved info, we error out */
4129: if (!a->ipre) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"No saved preallocation info \n");
4131: if (!a->i || !a->j || !a->a || !a->imax || !a->ilen) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"Memory info is incomplete, and can not reset preallocation \n");
4133: PetscArraycpy(a->imax,a->ipre,A->rmap->n);
4134: PetscArrayzero(a->ilen,A->rmap->n);
4135: a->i[0] = 0;
4136: for (i=1; i<A->rmap->n+1; i++) {
4137: a->i[i] = a->i[i-1] + a->imax[i-1];
4138: }
4139: A->preallocated = PETSC_TRUE;
4140: a->nz = 0;
4141: a->maxnz = a->i[A->rmap->n];
4142: A->info.nz_unneeded = (double)a->maxnz;
4143: A->was_assembled = PETSC_FALSE;
4144: A->assembled = PETSC_FALSE;
4145: return(0);
4146: }
4148: /*@
4149: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format.
4151: Input Parameters:
4152: + B - the matrix
4153: . i - the indices into j for the start of each row (starts with zero)
4154: . j - the column indices for each row (starts with zero) these must be sorted for each row
4155: - v - optional values in the matrix
4157: Level: developer
4159: Notes:
4160: The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays()
4162: This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero
4163: structure will be the union of all the previous nonzero structures.
4165: Developer Notes:
4166: An optimization could be added to the implementation where it checks if the i, and j are identical to the current i and j and
4167: then just copies the v values directly with PetscMemcpy().
4169: This routine could also take a PetscCopyMode argument to allow sharing the values instead of always copying them.
4171: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), MATSEQAIJ, MatResetPreallocation()
4172: @*/
4173: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[])
4174: {
4180: PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
4181: return(0);
4182: }
4184: PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
4185: {
4186: PetscInt i;
4187: PetscInt m,n;
4188: PetscInt nz;
4189: PetscInt *nnz;
4193: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]);
4195: PetscLayoutSetUp(B->rmap);
4196: PetscLayoutSetUp(B->cmap);
4198: MatGetSize(B, &m, &n);
4199: PetscMalloc1(m+1, &nnz);
4200: for (i = 0; i < m; i++) {
4201: nz = Ii[i+1]- Ii[i];
4202: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz);
4203: nnz[i] = nz;
4204: }
4205: MatSeqAIJSetPreallocation(B, 0, nnz);
4206: PetscFree(nnz);
4208: for (i = 0; i < m; i++) {
4209: MatSetValues_SeqAIJ(B, 1, &i, Ii[i+1] - Ii[i], J+Ii[i], v ? v + Ii[i] : NULL, INSERT_VALUES);
4210: }
4212: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4213: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4215: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
4216: return(0);
4217: }
4219: #include <../src/mat/impls/dense/seq/dense.h>
4220: #include <petsc/private/kernels/petscaxpy.h>
4222: /*
4223: Computes (B'*A')' since computing B*A directly is untenable
4225: n p p
4226: [ ] [ ] [ ]
4227: m [ A ] * n [ B ] = m [ C ]
4228: [ ] [ ] [ ]
4230: */
4231: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C)
4232: {
4233: PetscErrorCode ierr;
4234: Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data;
4235: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data;
4236: Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data;
4237: PetscInt i,j,n,m,q,p;
4238: const PetscInt *ii,*idx;
4239: const PetscScalar *b,*a,*a_q;
4240: PetscScalar *c,*c_q;
4241: PetscInt clda = sub_c->lda;
4242: PetscInt alda = sub_a->lda;
4245: m = A->rmap->n;
4246: n = A->cmap->n;
4247: p = B->cmap->n;
4248: a = sub_a->v;
4249: b = sub_b->a;
4250: c = sub_c->v;
4251: if (clda == m) {
4252: PetscArrayzero(c,m*p);
4253: } else {
4254: for (j=0;j<p;j++)
4255: for (i=0;i<m;i++)
4256: c[j*clda + i] = 0.0;
4257: }
4258: ii = sub_b->i;
4259: idx = sub_b->j;
4260: for (i=0; i<n; i++) {
4261: q = ii[i+1] - ii[i];
4262: while (q-->0) {
4263: c_q = c + clda*(*idx);
4264: a_q = a + alda*i;
4265: PetscKernelAXPY(c_q,*b,a_q,m);
4266: idx++;
4267: b++;
4268: }
4269: }
4270: return(0);
4271: }
4273: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat C)
4274: {
4276: PetscInt m=A->rmap->n,n=B->cmap->n;
4277: PetscBool cisdense;
4280: 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);
4281: MatSetSizes(C,m,n,m,n);
4282: MatSetBlockSizesFromMats(C,A,B);
4283: PetscObjectTypeCompareAny((PetscObject)C,&cisdense,MATSEQDENSE,MATSEQDENSECUDA,"");
4284: if (!cisdense) {
4285: MatSetType(C,MATDENSE);
4286: }
4287: MatSetUp(C);
4289: C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
4290: return(0);
4291: }
4293: /* ----------------------------------------------------------------*/
4294: /*MC
4295: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
4296: based on compressed sparse row format.
4298: Options Database Keys:
4299: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
4301: Level: beginner
4303: Notes:
4304: MatSetValues() may be called for this matrix type with a NULL argument for the numerical values,
4305: in this case the values associated with the rows and columns one passes in are set to zero
4306: in the matrix
4308: MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no
4309: space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored
4311: Developer Notes:
4312: It would be nice if all matrix formats supported passing NULL in for the numerical values
4314: .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType
4315: M*/
4317: /*MC
4318: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
4320: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
4321: and MATMPIAIJ otherwise. As a result, for single process communicators,
4322: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported
4323: for communicators controlling multiple processes. It is recommended that you call both of
4324: the above preallocation routines for simplicity.
4326: Options Database Keys:
4327: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
4329: Developer Notes:
4330: Subclasses include MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
4331: enough exist.
4333: Level: beginner
4335: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
4336: M*/
4338: /*MC
4339: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
4341: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
4342: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
4343: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
4344: for communicators controlling multiple processes. It is recommended that you call both of
4345: the above preallocation routines for simplicity.
4347: Options Database Keys:
4348: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
4350: Level: beginner
4352: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
4353: M*/
4355: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*);
4356: #if defined(PETSC_HAVE_ELEMENTAL)
4357: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
4358: #endif
4359: #if defined(PETSC_HAVE_SCALAPACK)
4360: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat,MatType,MatReuse,Mat*);
4361: #endif
4362: #if defined(PETSC_HAVE_HYPRE)
4363: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A,MatType,MatReuse,Mat*);
4364: #endif
4365: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*);
4367: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat,MatType,MatReuse,Mat*);
4368: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
4369: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
4371: /*@C
4372: MatSeqAIJGetArray - gives read/write access to the array where the data for a MATSEQAIJ matrix is stored
4374: Not Collective
4376: Input Parameter:
4377: . mat - a MATSEQAIJ matrix
4379: Output Parameter:
4380: . array - pointer to the data
4382: Level: intermediate
4384: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
4385: @*/
4386: PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array)
4387: {
4391: PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));
4392: return(0);
4393: }
4395: /*@C
4396: MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a MATSEQAIJ matrix is stored
4398: Not Collective
4400: Input Parameter:
4401: . mat - a MATSEQAIJ matrix
4403: Output Parameter:
4404: . array - pointer to the data
4406: Level: intermediate
4408: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayRead()
4409: @*/
4410: PetscErrorCode MatSeqAIJGetArrayRead(Mat A,const PetscScalar **array)
4411: {
4412: #if defined(PETSC_HAVE_DEVICE)
4413: PetscOffloadMask oval;
4414: #endif
4418: #if defined(PETSC_HAVE_DEVICE)
4419: oval = A->offloadmask;
4420: #endif
4421: MatSeqAIJGetArray(A,(PetscScalar**)array);
4422: #if defined(PETSC_HAVE_DEVICE)
4423: if (oval == PETSC_OFFLOAD_GPU || oval == PETSC_OFFLOAD_BOTH) A->offloadmask = PETSC_OFFLOAD_BOTH;
4424: #endif
4425: return(0);
4426: }
4428: /*@C
4429: MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from MatSeqAIJGetArrayRead
4431: Not Collective
4433: Input Parameter:
4434: . mat - a MATSEQAIJ matrix
4436: Output Parameter:
4437: . array - pointer to the data
4439: Level: intermediate
4441: .seealso: MatSeqAIJGetArray(), MatSeqAIJGetArrayRead()
4442: @*/
4443: PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A,const PetscScalar **array)
4444: {
4445: #if defined(PETSC_HAVE_DEVICE)
4446: PetscOffloadMask oval;
4447: #endif
4451: #if defined(PETSC_HAVE_DEVICE)
4452: oval = A->offloadmask;
4453: #endif
4454: MatSeqAIJRestoreArray(A,(PetscScalar**)array);
4455: #if defined(PETSC_HAVE_DEVICE)
4456: A->offloadmask = oval;
4457: #endif
4458: return(0);
4459: }
4461: /*@C
4462: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4464: Not Collective
4466: Input Parameter:
4467: . mat - a MATSEQAIJ matrix
4469: Output Parameter:
4470: . nz - the maximum number of nonzeros in any row
4472: Level: intermediate
4474: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
4475: @*/
4476: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz)
4477: {
4478: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
4481: *nz = aij->rmax;
4482: return(0);
4483: }
4485: /*@C
4486: MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray()
4488: Not Collective
4490: Input Parameters:
4491: + mat - a MATSEQAIJ matrix
4492: - array - pointer to the data
4494: Level: intermediate
4496: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90()
4497: @*/
4498: PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array)
4499: {
4503: PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));
4504: return(0);
4505: }
4507: #if defined(PETSC_HAVE_CUDA)
4508: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat);
4509: #endif
4510: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4511: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat);
4512: #endif
4514: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4515: {
4516: Mat_SeqAIJ *b;
4518: PetscMPIInt size;
4521: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
4522: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1");
4524: PetscNewLog(B,&b);
4526: B->data = (void*)b;
4528: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
4529: if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
4531: b->row = NULL;
4532: b->col = NULL;
4533: b->icol = NULL;
4534: b->reallocs = 0;
4535: b->ignorezeroentries = PETSC_FALSE;
4536: b->roworiented = PETSC_TRUE;
4537: b->nonew = 0;
4538: b->diag = NULL;
4539: b->solve_work = NULL;
4540: B->spptr = NULL;
4541: b->saved_values = NULL;
4542: b->idiag = NULL;
4543: b->mdiag = NULL;
4544: b->ssor_work = NULL;
4545: b->omega = 1.0;
4546: b->fshift = 0.0;
4547: b->idiagvalid = PETSC_FALSE;
4548: b->ibdiagvalid = PETSC_FALSE;
4549: b->keepnonzeropattern = PETSC_FALSE;
4551: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4552: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);
4553: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);
4555: #if defined(PETSC_HAVE_MATLAB_ENGINE)
4556: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);
4557: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);
4558: #endif
4560: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);
4561: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);
4562: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);
4563: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);
4564: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);
4565: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);
4566: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijsell_C",MatConvert_SeqAIJ_SeqAIJSELL);
4567: #if defined(PETSC_HAVE_MKL_SPARSE)
4568: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijmkl_C",MatConvert_SeqAIJ_SeqAIJMKL);
4569: #endif
4570: #if defined(PETSC_HAVE_CUDA)
4571: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcusparse_C",MatConvert_SeqAIJ_SeqAIJCUSPARSE);
4572: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqaijcusparse_seqaij_C",MatProductSetFromOptions_SeqAIJ);
4573: #endif
4574: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4575: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijkokkos_C",MatConvert_SeqAIJ_SeqAIJKokkos);
4576: #endif
4577: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);
4578: #if defined(PETSC_HAVE_ELEMENTAL)
4579: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);
4580: #endif
4581: #if defined(PETSC_HAVE_SCALAPACK)
4582: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_scalapack_C",MatConvert_AIJ_ScaLAPACK);
4583: #endif
4584: #if defined(PETSC_HAVE_HYPRE)
4585: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_hypre_C",MatConvert_AIJ_HYPRE);
4586: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_transpose_seqaij_seqaij_C",MatProductSetFromOptions_Transpose_AIJ_AIJ);
4587: #endif
4588: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);
4589: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsell_C",MatConvert_SeqAIJ_SeqSELL);
4590: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_is_C",MatConvert_XAIJ_IS);
4591: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);
4592: PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);
4593: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);
4594: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_SeqAIJ);
4595: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);
4596: PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);
4597: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_is_seqaij_C",MatProductSetFromOptions_IS_XAIJ);
4598: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqdense_seqaij_C",MatProductSetFromOptions_SeqDense_SeqAIJ);
4599: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqaij_seqaij_C",MatProductSetFromOptions_SeqAIJ);
4600: MatCreate_SeqAIJ_Inode(B);
4601: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4602: MatSeqAIJSetTypeFromOptions(B); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4603: return(0);
4604: }
4606: /*
4607: Given a matrix generated with MatGetFactor() duplicates all the information in A into B
4608: */
4609: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace)
4610: {
4611: Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data,*a = (Mat_SeqAIJ*)A->data;
4613: PetscInt m = A->rmap->n,i;
4616: if (!A->assembled && cpvalues!=MAT_DO_NOT_COPY_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot duplicate unassembled matrix");
4618: C->factortype = A->factortype;
4619: c->row = NULL;
4620: c->col = NULL;
4621: c->icol = NULL;
4622: c->reallocs = 0;
4624: C->assembled = PETSC_TRUE;
4626: PetscLayoutReference(A->rmap,&C->rmap);
4627: PetscLayoutReference(A->cmap,&C->cmap);
4629: PetscMalloc1(m,&c->imax);
4630: PetscMemcpy(c->imax,a->imax,m*sizeof(PetscInt));
4631: PetscMalloc1(m,&c->ilen);
4632: PetscMemcpy(c->ilen,a->ilen,m*sizeof(PetscInt));
4633: PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));
4635: /* allocate the matrix space */
4636: if (mallocmatspace) {
4637: PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);
4638: PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));
4640: c->singlemalloc = PETSC_TRUE;
4642: PetscArraycpy(c->i,a->i,m+1);
4643: if (m > 0) {
4644: PetscArraycpy(c->j,a->j,a->i[m]);
4645: if (cpvalues == MAT_COPY_VALUES) {
4646: PetscArraycpy(c->a,a->a,a->i[m]);
4647: } else {
4648: PetscArrayzero(c->a,a->i[m]);
4649: }
4650: }
4651: }
4653: c->ignorezeroentries = a->ignorezeroentries;
4654: c->roworiented = a->roworiented;
4655: c->nonew = a->nonew;
4656: if (a->diag) {
4657: PetscMalloc1(m+1,&c->diag);
4658: PetscMemcpy(c->diag,a->diag,m*sizeof(PetscInt));
4659: PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));
4660: } else c->diag = NULL;
4662: c->solve_work = NULL;
4663: c->saved_values = NULL;
4664: c->idiag = NULL;
4665: c->ssor_work = NULL;
4666: c->keepnonzeropattern = a->keepnonzeropattern;
4667: c->free_a = PETSC_TRUE;
4668: c->free_ij = PETSC_TRUE;
4670: c->rmax = a->rmax;
4671: c->nz = a->nz;
4672: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4673: C->preallocated = PETSC_TRUE;
4675: c->compressedrow.use = a->compressedrow.use;
4676: c->compressedrow.nrows = a->compressedrow.nrows;
4677: if (a->compressedrow.use) {
4678: i = a->compressedrow.nrows;
4679: PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);
4680: PetscArraycpy(c->compressedrow.i,a->compressedrow.i,i+1);
4681: PetscArraycpy(c->compressedrow.rindex,a->compressedrow.rindex,i);
4682: } else {
4683: c->compressedrow.use = PETSC_FALSE;
4684: c->compressedrow.i = NULL;
4685: c->compressedrow.rindex = NULL;
4686: }
4687: c->nonzerorowcnt = a->nonzerorowcnt;
4688: C->nonzerostate = A->nonzerostate;
4690: MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);
4691: PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);
4692: return(0);
4693: }
4695: PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B)
4696: {
4700: MatCreate(PetscObjectComm((PetscObject)A),B);
4701: MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);
4702: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) {
4703: MatSetBlockSizesFromMats(*B,A,A);
4704: }
4705: MatSetType(*B,((PetscObject)A)->type_name);
4706: MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);
4707: return(0);
4708: }
4710: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4711: {
4712: PetscBool isbinary, ishdf5;
4718: /* force binary viewer to load .info file if it has not yet done so */
4719: PetscViewerSetUp(viewer);
4720: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
4721: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);
4722: if (isbinary) {
4723: MatLoad_SeqAIJ_Binary(newMat,viewer);
4724: } else if (ishdf5) {
4725: #if defined(PETSC_HAVE_HDF5)
4726: MatLoad_AIJ_HDF5(newMat,viewer);
4727: #else
4728: SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
4729: #endif
4730: } else {
4731: SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name);
4732: }
4733: return(0);
4734: }
4736: PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
4737: {
4738: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->data;
4740: PetscInt header[4],*rowlens,M,N,nz,sum,rows,cols,i;
4743: PetscViewerSetUp(viewer);
4745: /* read in matrix header */
4746: PetscViewerBinaryRead(viewer,header,4,NULL,PETSC_INT);
4747: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Not a matrix object in file");
4748: M = header[1]; N = header[2]; nz = header[3];
4749: if (M < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix row size (%D) in file is negative",M);
4750: if (N < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix column size (%D) in file is negative",N);
4751: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk, cannot load as SeqAIJ");
4753: /* set block sizes from the viewer's .info file */
4754: MatLoad_Binary_BlockSizes(mat,viewer);
4755: /* set local and global sizes if not set already */
4756: if (mat->rmap->n < 0) mat->rmap->n = M;
4757: if (mat->cmap->n < 0) mat->cmap->n = N;
4758: if (mat->rmap->N < 0) mat->rmap->N = M;
4759: if (mat->cmap->N < 0) mat->cmap->N = N;
4760: PetscLayoutSetUp(mat->rmap);
4761: PetscLayoutSetUp(mat->cmap);
4763: /* check if the matrix sizes are correct */
4764: MatGetSize(mat,&rows,&cols);
4765: if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols);
4767: /* read in row lengths */
4768: PetscMalloc1(M,&rowlens);
4769: PetscViewerBinaryRead(viewer,rowlens,M,NULL,PETSC_INT);
4770: /* check if sum(rowlens) is same as nz */
4771: sum = 0; for (i=0; i<M; i++) sum += rowlens[i];
4772: if (sum != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent matrix data in file: nonzeros = %D, sum-row-lengths = %D\n",nz,sum);
4773: /* preallocate and check sizes */
4774: MatSeqAIJSetPreallocation_SeqAIJ(mat,0,rowlens);
4775: MatGetSize(mat,&rows,&cols);
4776: 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);
4777: /* store row lengths */
4778: PetscArraycpy(a->ilen,rowlens,M);
4779: PetscFree(rowlens);
4781: /* fill in "i" row pointers */
4782: a->i[0] = 0; for (i=0; i<M; i++) a->i[i+1] = a->i[i] + a->ilen[i];
4783: /* read in "j" column indices */
4784: PetscViewerBinaryRead(viewer,a->j,nz,NULL,PETSC_INT);
4785: /* read in "a" nonzero values */
4786: PetscViewerBinaryRead(viewer,a->a,nz,NULL,PETSC_SCALAR);
4788: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
4789: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
4790: return(0);
4791: }
4793: PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg)
4794: {
4795: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
4797: #if defined(PETSC_USE_COMPLEX)
4798: PetscInt k;
4799: #endif
4802: /* If the matrix dimensions are not equal,or no of nonzeros */
4803: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) {
4804: *flg = PETSC_FALSE;
4805: return(0);
4806: }
4808: /* if the a->i are the same */
4809: PetscArraycmp(a->i,b->i,A->rmap->n+1,flg);
4810: if (!*flg) return(0);
4812: /* if a->j are the same */
4813: PetscArraycmp(a->j,b->j,a->nz,flg);
4814: if (!*flg) return(0);
4816: /* if a->a are the same */
4817: #if defined(PETSC_USE_COMPLEX)
4818: for (k=0; k<a->nz; k++) {
4819: if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) {
4820: *flg = PETSC_FALSE;
4821: return(0);
4822: }
4823: }
4824: #else
4825: PetscArraycmp(a->a,b->a,a->nz,flg);
4826: #endif
4827: return(0);
4828: }
4830: /*@
4831: MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format)
4832: provided by the user.
4834: Collective
4836: Input Parameters:
4837: + comm - must be an MPI communicator of size 1
4838: . m - number of rows
4839: . n - number of columns
4840: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4841: . j - column indices
4842: - a - matrix values
4844: Output Parameter:
4845: . mat - the matrix
4847: Level: intermediate
4849: Notes:
4850: The i, j, and a arrays are not copied by this routine, the user must free these arrays
4851: once the matrix is destroyed and not before
4853: You cannot set new nonzero locations into this matrix, that will generate an error.
4855: The i and j indices are 0 based
4857: The format which is used for the sparse matrix input, is equivalent to a
4858: row-major ordering.. i.e for the following matrix, the input data expected is
4859: as shown
4861: $ 1 0 0
4862: $ 2 0 3
4863: $ 4 5 6
4864: $
4865: $ i = {0,1,3,6} [size = nrow+1 = 3+1]
4866: $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
4867: $ v = {1,2,3,4,5,6} [size = 6]
4870: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4872: @*/
4873: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat)
4874: {
4876: PetscInt ii;
4877: Mat_SeqAIJ *aij;
4878: PetscInt jj;
4881: if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4882: MatCreate(comm,mat);
4883: MatSetSizes(*mat,m,n,m,n);
4884: /* MatSetBlockSizes(*mat,,); */
4885: MatSetType(*mat,MATSEQAIJ);
4886: MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,NULL);
4887: aij = (Mat_SeqAIJ*)(*mat)->data;
4888: PetscMalloc1(m,&aij->imax);
4889: PetscMalloc1(m,&aij->ilen);
4891: aij->i = i;
4892: aij->j = j;
4893: aij->a = a;
4894: aij->singlemalloc = PETSC_FALSE;
4895: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
4896: aij->free_a = PETSC_FALSE;
4897: aij->free_ij = PETSC_FALSE;
4899: for (ii=0; ii<m; ii++) {
4900: aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii];
4901: if (PetscDefined(USE_DEBUG)) {
4902: 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]);
4903: for (jj=i[ii]+1; jj<i[ii+1]; jj++) {
4904: if (j[jj] < j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual column %D) in row %D is not sorted",jj-i[ii],j[jj],ii);
4905: if (j[jj] == j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual column %D) in row %D is identical to previous entry",jj-i[ii],j[jj],ii);
4906: }
4907: }
4908: }
4909: if (PetscDefined(USE_DEBUG)) {
4910: for (ii=0; ii<aij->i[m]; ii++) {
4911: if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]);
4912: 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]);
4913: }
4914: }
4916: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4917: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4918: return(0);
4919: }
4920: /*@C
4921: MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format)
4922: provided by the user.
4924: Collective
4926: Input Parameters:
4927: + comm - must be an MPI communicator of size 1
4928: . m - number of rows
4929: . n - number of columns
4930: . i - row indices
4931: . j - column indices
4932: . a - matrix values
4933: . nz - number of nonzeros
4934: - idx - 0 or 1 based
4936: Output Parameter:
4937: . mat - the matrix
4939: Level: intermediate
4941: Notes:
4942: The i and j indices are 0 based
4944: The format which is used for the sparse matrix input, is equivalent to a
4945: row-major ordering.. i.e for the following matrix, the input data expected is
4946: as shown:
4948: 1 0 0
4949: 2 0 3
4950: 4 5 6
4952: i = {0,1,1,2,2,2}
4953: j = {0,0,2,0,1,2}
4954: v = {1,2,3,4,5,6}
4957: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4959: @*/
4960: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat,PetscInt nz,PetscBool idx)
4961: {
4963: PetscInt ii, *nnz, one = 1,row,col;
4967: PetscCalloc1(m,&nnz);
4968: for (ii = 0; ii < nz; ii++) {
4969: nnz[i[ii] - !!idx] += 1;
4970: }
4971: MatCreate(comm,mat);
4972: MatSetSizes(*mat,m,n,m,n);
4973: MatSetType(*mat,MATSEQAIJ);
4974: MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);
4975: for (ii = 0; ii < nz; ii++) {
4976: if (idx) {
4977: row = i[ii] - 1;
4978: col = j[ii] - 1;
4979: } else {
4980: row = i[ii];
4981: col = j[ii];
4982: }
4983: MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);
4984: }
4985: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4986: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4987: PetscFree(nnz);
4988: return(0);
4989: }
4991: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
4992: {
4993: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data;
4997: a->idiagvalid = PETSC_FALSE;
4998: a->ibdiagvalid = PETSC_FALSE;
5000: MatSeqAIJInvalidateDiagonal_Inode(A);
5001: return(0);
5002: }
5004: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
5005: {
5007: PetscMPIInt size;
5010: MPI_Comm_size(comm,&size);
5011: if (size == 1) {
5012: if (scall == MAT_INITIAL_MATRIX) {
5013: MatDuplicate(inmat,MAT_COPY_VALUES,outmat);
5014: } else {
5015: MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);
5016: }
5017: } else {
5018: MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);
5019: }
5020: return(0);
5021: }
5023: /*
5024: Permute A into C's *local* index space using rowemb,colemb.
5025: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
5026: of [0,m), colemb is in [0,n).
5027: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
5028: */
5029: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B)
5030: {
5031: /* If making this function public, change the error returned in this function away from _PLIB. */
5033: Mat_SeqAIJ *Baij;
5034: PetscBool seqaij;
5035: PetscInt m,n,*nz,i,j,count;
5036: PetscScalar v;
5037: const PetscInt *rowindices,*colindices;
5040: if (!B) return(0);
5041: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
5042: PetscObjectBaseTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);
5043: if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type");
5044: if (rowemb) {
5045: ISGetLocalSize(rowemb,&m);
5046: 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);
5047: } else {
5048: if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix");
5049: }
5050: if (colemb) {
5051: ISGetLocalSize(colemb,&n);
5052: 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);
5053: } else {
5054: if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix");
5055: }
5057: Baij = (Mat_SeqAIJ*)(B->data);
5058: if (pattern == DIFFERENT_NONZERO_PATTERN) {
5059: PetscMalloc1(B->rmap->n,&nz);
5060: for (i=0; i<B->rmap->n; i++) {
5061: nz[i] = Baij->i[i+1] - Baij->i[i];
5062: }
5063: MatSeqAIJSetPreallocation(C,0,nz);
5064: PetscFree(nz);
5065: }
5066: if (pattern == SUBSET_NONZERO_PATTERN) {
5067: MatZeroEntries(C);
5068: }
5069: count = 0;
5070: rowindices = NULL;
5071: colindices = NULL;
5072: if (rowemb) {
5073: ISGetIndices(rowemb,&rowindices);
5074: }
5075: if (colemb) {
5076: ISGetIndices(colemb,&colindices);
5077: }
5078: for (i=0; i<B->rmap->n; i++) {
5079: PetscInt row;
5080: row = i;
5081: if (rowindices) row = rowindices[i];
5082: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
5083: PetscInt col;
5084: col = Baij->j[count];
5085: if (colindices) col = colindices[col];
5086: v = Baij->a[count];
5087: MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);
5088: ++count;
5089: }
5090: }
5091: /* FIXME: set C's nonzerostate correctly. */
5092: /* Assembly for C is necessary. */
5093: C->preallocated = PETSC_TRUE;
5094: C->assembled = PETSC_TRUE;
5095: C->was_assembled = PETSC_FALSE;
5096: return(0);
5097: }
5099: PetscFunctionList MatSeqAIJList = NULL;
5101: /*@C
5102: MatSeqAIJSetType - Converts a MATSEQAIJ matrix to a subtype
5104: Collective on Mat
5106: Input Parameters:
5107: + mat - the matrix object
5108: - matype - matrix type
5110: Options Database Key:
5111: . -mat_seqai_type <method> - for example seqaijcrl
5114: Level: intermediate
5116: .seealso: PCSetType(), VecSetType(), MatCreate(), MatType, Mat
5117: @*/
5118: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
5119: {
5120: PetscErrorCode ierr,(*r)(Mat,MatType,MatReuse,Mat*);
5121: PetscBool sametype;
5125: PetscObjectTypeCompare((PetscObject)mat,matype,&sametype);
5126: if (sametype) return(0);
5128: PetscFunctionListFind(MatSeqAIJList,matype,&r);
5129: if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown Mat type given: %s",matype);
5130: (*r)(mat,matype,MAT_INPLACE_MATRIX,&mat);
5131: return(0);
5132: }
5135: /*@C
5136: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential AIJ matrices
5138: Not Collective
5140: Input Parameters:
5141: + name - name of a new user-defined matrix type, for example MATSEQAIJCRL
5142: - function - routine to convert to subtype
5144: Notes:
5145: MatSeqAIJRegister() may be called multiple times to add several user-defined solvers.
5148: Then, your matrix can be chosen with the procedural interface at runtime via the option
5149: $ -mat_seqaij_type my_mat
5151: Level: advanced
5153: .seealso: MatSeqAIJRegisterAll()
5156: Level: advanced
5157: @*/
5158: PetscErrorCode MatSeqAIJRegister(const char sname[],PetscErrorCode (*function)(Mat,MatType,MatReuse,Mat *))
5159: {
5163: MatInitializePackage();
5164: PetscFunctionListAdd(&MatSeqAIJList,sname,function);
5165: return(0);
5166: }
5168: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
5170: /*@C
5171: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of SeqAIJ
5173: Not Collective
5175: Level: advanced
5177: Developers Note: CUSPARSE does not yet support the MatConvert_SeqAIJ..() paradigm and thus cannot be registered here
5179: .seealso: MatRegisterAll(), MatSeqAIJRegister()
5180: @*/
5181: PetscErrorCode MatSeqAIJRegisterAll(void)
5182: {
5186: if (MatSeqAIJRegisterAllCalled) return(0);
5187: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
5189: MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);
5190: MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);
5191: MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL);
5192: #if defined(PETSC_HAVE_MKL_SPARSE)
5193: MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);
5194: #endif
5195: #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA)
5196: MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);
5197: #endif
5198: return(0);
5199: }
5201: /*
5202: Special version for direct calls from Fortran
5203: */
5204: #include <petsc/private/fortranimpl.h>
5205: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5206: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
5207: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5208: #define matsetvaluesseqaij_ matsetvaluesseqaij
5209: #endif
5211: /* Change these macros so can be used in void function */
5212: #undef CHKERRQ
5213: #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr)
5214: #undef SETERRQ2
5215: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
5216: #undef SETERRQ3
5217: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
5219: PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[],InsertMode *isis, PetscErrorCode *_ierr)
5220: {
5221: Mat A = *AA;
5222: PetscInt m = *mm, n = *nn;
5223: InsertMode is = *isis;
5224: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
5225: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
5226: PetscInt *imax,*ai,*ailen;
5228: PetscInt *aj,nonew = a->nonew,lastcol = -1;
5229: MatScalar *ap,value,*aa;
5230: PetscBool ignorezeroentries = a->ignorezeroentries;
5231: PetscBool roworiented = a->roworiented;
5234: MatCheckPreallocated(A,1);
5235: imax = a->imax;
5236: ai = a->i;
5237: ailen = a->ilen;
5238: aj = a->j;
5239: aa = a->a;
5241: for (k=0; k<m; k++) { /* loop over added rows */
5242: row = im[k];
5243: if (row < 0) continue;
5244: if (PetscUnlikelyDebug(row >= A->rmap->n)) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
5245: rp = aj + ai[row]; ap = aa + ai[row];
5246: rmax = imax[row]; nrow = ailen[row];
5247: low = 0;
5248: high = nrow;
5249: for (l=0; l<n; l++) { /* loop over added columns */
5250: if (in[l] < 0) continue;
5251: if (PetscUnlikelyDebug(in[l] >= A->cmap->n)) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
5252: col = in[l];
5253: if (roworiented) value = v[l + k*n];
5254: else value = v[k + l*m];
5256: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
5258: if (col <= lastcol) low = 0;
5259: else high = nrow;
5260: lastcol = col;
5261: while (high-low > 5) {
5262: t = (low+high)/2;
5263: if (rp[t] > col) high = t;
5264: else low = t;
5265: }
5266: for (i=low; i<high; i++) {
5267: if (rp[i] > col) break;
5268: if (rp[i] == col) {
5269: if (is == ADD_VALUES) ap[i] += value;
5270: else ap[i] = value;
5271: goto noinsert;
5272: }
5273: }
5274: if (value == 0.0 && ignorezeroentries) goto noinsert;
5275: if (nonew == 1) goto noinsert;
5276: if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix");
5277: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
5278: N = nrow++ - 1; a->nz++; high++;
5279: /* shift up all the later entries in this row */
5280: for (ii=N; ii>=i; ii--) {
5281: rp[ii+1] = rp[ii];
5282: ap[ii+1] = ap[ii];
5283: }
5284: rp[i] = col;
5285: ap[i] = value;
5286: A->nonzerostate++;
5287: noinsert:;
5288: low = i + 1;
5289: }
5290: ailen[row] = nrow;
5291: }
5292: PetscFunctionReturnVoid();
5293: }