Actual source code: aij.c
petsc-main 2021-03-03
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: MatSeqAIJGetArrayRead(A,&aa);
135: *keptrows = NULL;
136: ii = a->i;
137: for (i=0; i<m; i++) {
138: n = ii[i+1] - ii[i];
139: if (!n) {
140: cnt++;
141: goto ok1;
142: }
143: for (j=ii[i]; j<ii[i+1]; j++) {
144: if (aa[j] != 0.0) goto ok1;
145: }
146: cnt++;
147: ok1:;
148: }
149: if (!cnt) {
150: MatSeqAIJRestoreArrayRead(A,&aa);
151: return(0);
152: }
153: PetscMalloc1(A->rmap->n-cnt,&rows);
154: cnt = 0;
155: for (i=0; i<m; i++) {
156: n = ii[i+1] - ii[i];
157: if (!n) continue;
158: for (j=ii[i]; j<ii[i+1]; j++) {
159: if (aa[j] != 0.0) {
160: rows[cnt++] = i;
161: break;
162: }
163: }
164: }
165: MatSeqAIJRestoreArrayRead(A,&aa);
166: ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,keptrows);
167: return(0);
168: }
170: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y,Vec D,InsertMode is)
171: {
172: PetscErrorCode ierr;
173: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) Y->data;
174: PetscInt i,m = Y->rmap->n;
175: const PetscInt *diag;
176: MatScalar *aa;
177: const PetscScalar *v;
178: PetscBool missing;
179: #if defined(PETSC_HAVE_DEVICE)
180: PetscBool inserted = PETSC_FALSE;
181: #endif
184: if (Y->assembled) {
185: MatMissingDiagonal_SeqAIJ(Y,&missing,NULL);
186: if (!missing) {
187: diag = aij->diag;
188: VecGetArrayRead(D,&v);
189: MatSeqAIJGetArray(Y,&aa);
190: if (is == INSERT_VALUES) {
191: #if defined(PETSC_HAVE_DEVICE)
192: inserted = PETSC_TRUE;
193: #endif
194: for (i=0; i<m; i++) {
195: aa[diag[i]] = v[i];
196: }
197: } else {
198: for (i=0; i<m; i++) {
199: #if defined(PETSC_HAVE_DEVICE)
200: if (v[i] != 0.0) inserted = PETSC_TRUE;
201: #endif
202: aa[diag[i]] += v[i];
203: }
204: }
205: MatSeqAIJRestoreArray(Y,&aa);
206: #if defined(PETSC_HAVE_DEVICE)
207: if (inserted) Y->offloadmask = PETSC_OFFLOAD_CPU;
208: #endif
209: VecRestoreArrayRead(D,&v);
210: return(0);
211: }
212: MatSeqAIJInvalidateDiagonal(Y);
213: }
214: MatDiagonalSet_Default(Y,D,is);
215: return(0);
216: }
218: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *m,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
219: {
220: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
222: PetscInt i,ishift;
225: *m = A->rmap->n;
226: if (!ia) return(0);
227: ishift = 0;
228: if (symmetric && !A->structurally_symmetric) {
229: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,ishift,oshift,(PetscInt**)ia,(PetscInt**)ja);
230: } else if (oshift == 1) {
231: PetscInt *tia;
232: PetscInt nz = a->i[A->rmap->n];
233: /* malloc space and add 1 to i and j indices */
234: PetscMalloc1(A->rmap->n+1,&tia);
235: for (i=0; i<A->rmap->n+1; i++) tia[i] = a->i[i] + 1;
236: *ia = tia;
237: if (ja) {
238: PetscInt *tja;
239: PetscMalloc1(nz+1,&tja);
240: for (i=0; i<nz; i++) tja[i] = a->j[i] + 1;
241: *ja = tja;
242: }
243: } else {
244: *ia = a->i;
245: if (ja) *ja = a->j;
246: }
247: return(0);
248: }
250: PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
251: {
255: if (!ia) return(0);
256: if ((symmetric && !A->structurally_symmetric) || oshift == 1) {
257: PetscFree(*ia);
258: if (ja) {PetscFree(*ja);}
259: }
260: return(0);
261: }
263: PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
264: {
265: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
267: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
268: PetscInt nz = a->i[m],row,*jj,mr,col;
271: *nn = n;
272: if (!ia) return(0);
273: if (symmetric) {
274: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,0,oshift,(PetscInt**)ia,(PetscInt**)ja);
275: } else {
276: PetscCalloc1(n,&collengths);
277: PetscMalloc1(n+1,&cia);
278: PetscMalloc1(nz,&cja);
279: jj = a->j;
280: for (i=0; i<nz; i++) {
281: collengths[jj[i]]++;
282: }
283: cia[0] = oshift;
284: for (i=0; i<n; i++) {
285: cia[i+1] = cia[i] + collengths[i];
286: }
287: PetscArrayzero(collengths,n);
288: jj = a->j;
289: for (row=0; row<m; row++) {
290: mr = a->i[row+1] - a->i[row];
291: for (i=0; i<mr; i++) {
292: col = *jj++;
294: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
295: }
296: }
297: PetscFree(collengths);
298: *ia = cia; *ja = cja;
299: }
300: return(0);
301: }
303: PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
304: {
308: if (!ia) return(0);
310: PetscFree(*ia);
311: PetscFree(*ja);
312: return(0);
313: }
315: /*
316: MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from
317: MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output
318: spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ()
319: */
320: PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
321: {
322: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
324: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
325: PetscInt nz = a->i[m],row,mr,col,tmp;
326: PetscInt *cspidx;
327: const PetscInt *jj;
330: *nn = n;
331: if (!ia) return(0);
333: PetscCalloc1(n,&collengths);
334: PetscMalloc1(n+1,&cia);
335: PetscMalloc1(nz,&cja);
336: PetscMalloc1(nz,&cspidx);
337: jj = a->j;
338: for (i=0; i<nz; i++) {
339: collengths[jj[i]]++;
340: }
341: cia[0] = oshift;
342: for (i=0; i<n; i++) {
343: cia[i+1] = cia[i] + collengths[i];
344: }
345: PetscArrayzero(collengths,n);
346: jj = a->j;
347: for (row=0; row<m; row++) {
348: mr = a->i[row+1] - a->i[row];
349: for (i=0; i<mr; i++) {
350: col = *jj++;
351: tmp = cia[col] + collengths[col]++ - oshift;
352: cspidx[tmp] = a->i[row] + i; /* index of a->j */
353: cja[tmp] = row + oshift;
354: }
355: }
356: PetscFree(collengths);
357: *ia = cia;
358: *ja = cja;
359: *spidx = cspidx;
360: return(0);
361: }
363: PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
364: {
368: MatRestoreColumnIJ_SeqAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);
369: PetscFree(*spidx);
370: return(0);
371: }
373: PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A,PetscInt row,const PetscScalar v[])
374: {
375: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
376: PetscInt *ai = a->i;
380: PetscArraycpy(a->a+ai[row],v,ai[row+1]-ai[row]);
381: #if defined(PETSC_HAVE_DEVICE)
382: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && ai[row+1]-ai[row]) A->offloadmask = PETSC_OFFLOAD_CPU;
383: #endif
384: return(0);
385: }
387: /*
388: MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions
390: - a single row of values is set with each call
391: - no row or column indices are negative or (in error) larger than the number of rows or columns
392: - the values are always added to the matrix, not set
393: - no new locations are introduced in the nonzero structure of the matrix
395: This does NOT assume the global column indices are sorted
397: */
399: #include <petsc/private/isimpl.h>
400: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
401: {
402: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
403: PetscInt low,high,t,row,nrow,i,col,l;
404: const PetscInt *rp,*ai = a->i,*ailen = a->ilen,*aj = a->j;
405: PetscInt lastcol = -1;
406: MatScalar *ap,value,*aa = a->a;
407: const PetscInt *ridx = A->rmap->mapping->indices,*cidx = A->cmap->mapping->indices;
409: row = ridx[im[0]];
410: rp = aj + ai[row];
411: ap = aa + ai[row];
412: nrow = ailen[row];
413: low = 0;
414: high = nrow;
415: for (l=0; l<n; l++) { /* loop over added columns */
416: col = cidx[in[l]];
417: value = v[l];
419: if (col <= lastcol) low = 0;
420: else high = nrow;
421: lastcol = col;
422: while (high-low > 5) {
423: t = (low+high)/2;
424: if (rp[t] > col) high = t;
425: else low = t;
426: }
427: for (i=low; i<high; i++) {
428: if (rp[i] == col) {
429: ap[i] += value;
430: low = i + 1;
431: break;
432: }
433: }
434: }
435: #if defined(PETSC_HAVE_DEVICE)
436: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU;
437: #endif
438: return 0;
439: }
441: PetscErrorCode MatSetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
442: {
443: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
444: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
445: PetscInt *imax = a->imax,*ai = a->i,*ailen = a->ilen;
447: PetscInt *aj = a->j,nonew = a->nonew,lastcol = -1;
448: MatScalar *ap=NULL,value=0.0,*aa;
449: PetscBool ignorezeroentries = a->ignorezeroentries;
450: PetscBool roworiented = a->roworiented;
451: #if defined(PETSC_HAVE_DEVICE)
452: PetscBool inserted = PETSC_FALSE;
453: #endif
456: #if defined(PETSC_HAVE_DEVICE)
457: if (A->offloadmask == PETSC_OFFLOAD_GPU) {
458: const PetscScalar *dummy;
459: MatSeqAIJGetArrayRead(A,&dummy);
460: MatSeqAIJRestoreArrayRead(A,&dummy);
461: }
462: #endif
463: aa = a->a;
464: for (k=0; k<m; k++) { /* loop over added rows */
465: row = im[k];
466: if (row < 0) continue;
467: 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);
468: rp = aj + ai[row];
469: if (!A->structure_only) ap = aa + ai[row];
470: rmax = imax[row]; nrow = ailen[row];
471: low = 0;
472: high = nrow;
473: for (l=0; l<n; l++) { /* loop over added columns */
474: if (in[l] < 0) continue;
475: 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);
476: col = in[l];
477: if (v && !A->structure_only) value = roworiented ? v[l + k*n] : v[k + l*m];
478: if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue;
480: if (col <= lastcol) low = 0;
481: else high = nrow;
482: lastcol = col;
483: while (high-low > 5) {
484: t = (low+high)/2;
485: if (rp[t] > col) high = t;
486: else low = t;
487: }
488: for (i=low; i<high; i++) {
489: if (rp[i] > col) break;
490: if (rp[i] == col) {
491: if (!A->structure_only) {
492: if (is == ADD_VALUES) {
493: ap[i] += value;
494: (void)PetscLogFlops(1.0);
495: }
496: else ap[i] = value;
497: #if defined(PETSC_HAVE_DEVICE)
498: inserted = PETSC_TRUE;
499: #endif
500: }
501: low = i + 1;
502: goto noinsert;
503: }
504: }
505: if (value == 0.0 && ignorezeroentries && row != col) goto noinsert;
506: if (nonew == 1) goto noinsert;
507: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at (%D,%D) in the matrix",row,col);
508: if (A->structure_only) {
509: MatSeqXAIJReallocateAIJ_structure_only(A,A->rmap->n,1,nrow,row,col,rmax,ai,aj,rp,imax,nonew,MatScalar);
510: } else {
511: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
512: }
513: N = nrow++ - 1; a->nz++; high++;
514: /* shift up all the later entries in this row */
515: PetscArraymove(rp+i+1,rp+i,N-i+1);
516: rp[i] = col;
517: if (!A->structure_only){
518: PetscArraymove(ap+i+1,ap+i,N-i+1);
519: ap[i] = value;
520: }
521: low = i + 1;
522: A->nonzerostate++;
523: #if defined(PETSC_HAVE_DEVICE)
524: inserted = PETSC_TRUE;
525: #endif
526: noinsert:;
527: }
528: ailen[row] = nrow;
529: }
530: #if defined(PETSC_HAVE_DEVICE)
531: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
532: #endif
533: return(0);
534: }
537: PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
538: {
539: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
540: PetscInt *rp,k,row;
541: PetscInt *ai = a->i;
543: PetscInt *aj = a->j;
544: MatScalar *aa = a->a,*ap;
547: if (A->was_assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot call on assembled matrix.");
548: 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);
549: for (k=0; k<m; k++) { /* loop over added rows */
550: row = im[k];
551: rp = aj + ai[row];
552: ap = aa + ai[row];
554: PetscMemcpy(rp,in,n*sizeof(PetscInt));
555: if (!A->structure_only) {
556: if (v) {
557: PetscMemcpy(ap,v,n*sizeof(PetscScalar));
558: v += n;
559: } else {
560: PetscMemzero(ap,n*sizeof(PetscScalar));
561: }
562: }
563: a->ilen[row] = n;
564: a->imax[row] = n;
565: a->i[row+1] = a->i[row]+n;
566: a->nz += n;
567: }
568: #if defined(PETSC_HAVE_DEVICE)
569: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU;
570: #endif
571: return(0);
572: }
574: /*@
575: MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix.
577: Input Parameters:
578: + A - the SeqAIJ matrix
579: - nztotal - bound on the number of nonzeros
581: Level: advanced
583: Notes:
584: This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row.
585: Simply call MatSetValues() after this call to provide the matrix entries in the usual manner. This matrix may be used
586: as always with multiple matrix assemblies.
588: .seealso: MatSetOption(), MAT_SORTED_FULL, MatSetValues(), MatSeqAIJSetPreallocation()
589: @*/
591: PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A,PetscInt nztotal)
592: {
594: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
597: PetscLayoutSetUp(A->rmap);
598: PetscLayoutSetUp(A->cmap);
599: a->maxnz = nztotal;
600: if (!a->imax) {
601: PetscMalloc1(A->rmap->n,&a->imax);
602: PetscLogObjectMemory((PetscObject)A,A->rmap->n*sizeof(PetscInt));
603: }
604: if (!a->ilen) {
605: PetscMalloc1(A->rmap->n,&a->ilen);
606: PetscLogObjectMemory((PetscObject)A,A->rmap->n*sizeof(PetscInt));
607: } else {
608: PetscMemzero(a->ilen,A->rmap->n*sizeof(PetscInt));
609: }
611: /* allocate the matrix space */
612: if (A->structure_only) {
613: PetscMalloc1(nztotal,&a->j);
614: PetscMalloc1(A->rmap->n+1,&a->i);
615: PetscLogObjectMemory((PetscObject)A,(A->rmap->n+1)*sizeof(PetscInt)+nztotal*sizeof(PetscInt));
616: } else {
617: PetscMalloc3(nztotal,&a->a,nztotal,&a->j,A->rmap->n+1,&a->i);
618: PetscLogObjectMemory((PetscObject)A,(A->rmap->n+1)*sizeof(PetscInt)+nztotal*(sizeof(PetscScalar)+sizeof(PetscInt)));
619: }
620: a->i[0] = 0;
621: if (A->structure_only) {
622: a->singlemalloc = PETSC_FALSE;
623: a->free_a = PETSC_FALSE;
624: } else {
625: a->singlemalloc = PETSC_TRUE;
626: a->free_a = PETSC_TRUE;
627: }
628: a->free_ij = PETSC_TRUE;
629: A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation;
630: A->preallocated = PETSC_TRUE;
631: return(0);
632: }
634: PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
635: {
636: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
637: PetscInt *rp,k,row;
638: PetscInt *ai = a->i,*ailen = a->ilen;
640: PetscInt *aj = a->j;
641: MatScalar *aa = a->a,*ap;
644: for (k=0; k<m; k++) { /* loop over added rows */
645: row = im[k];
646: 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);
647: rp = aj + ai[row];
648: ap = aa + ai[row];
649: if (!A->was_assembled) {
650: PetscMemcpy(rp,in,n*sizeof(PetscInt));
651: }
652: if (!A->structure_only) {
653: if (v) {
654: PetscMemcpy(ap,v,n*sizeof(PetscScalar));
655: v += n;
656: } else {
657: PetscMemzero(ap,n*sizeof(PetscScalar));
658: }
659: }
660: ailen[row] = n;
661: a->nz += n;
662: }
663: #if defined(PETSC_HAVE_DEVICE)
664: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU;
665: #endif
666: return(0);
667: }
670: PetscErrorCode MatGetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[])
671: {
672: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
673: PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j;
674: PetscInt *ai = a->i,*ailen = a->ilen;
675: MatScalar *ap,*aa = a->a;
678: for (k=0; k<m; k++) { /* loop over rows */
679: row = im[k];
680: if (row < 0) {v += n; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",row); */
681: 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);
682: rp = aj + ai[row]; ap = aa + ai[row];
683: nrow = ailen[row];
684: for (l=0; l<n; l++) { /* loop over columns */
685: if (in[l] < 0) {v++; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",in[l]); */
686: 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);
687: col = in[l];
688: high = nrow; low = 0; /* assume unsorted */
689: while (high-low > 5) {
690: t = (low+high)/2;
691: if (rp[t] > col) high = t;
692: else low = t;
693: }
694: for (i=low; i<high; i++) {
695: if (rp[i] > col) break;
696: if (rp[i] == col) {
697: *v++ = ap[i];
698: goto finished;
699: }
700: }
701: *v++ = 0.0;
702: finished:;
703: }
704: }
705: return(0);
706: }
708: PetscErrorCode MatView_SeqAIJ_Binary(Mat mat,PetscViewer viewer)
709: {
710: Mat_SeqAIJ *A = (Mat_SeqAIJ*)mat->data;
711: const PetscScalar *av;
712: PetscInt header[4],M,N,m,nz,i;
713: PetscInt *rowlens;
714: PetscErrorCode ierr;
717: PetscViewerSetUp(viewer);
719: M = mat->rmap->N;
720: N = mat->cmap->N;
721: m = mat->rmap->n;
722: nz = A->nz;
724: /* write matrix header */
725: header[0] = MAT_FILE_CLASSID;
726: header[1] = M; header[2] = N; header[3] = nz;
727: PetscViewerBinaryWrite(viewer,header,4,PETSC_INT);
729: /* fill in and store row lengths */
730: PetscMalloc1(m,&rowlens);
731: for (i=0; i<m; i++) rowlens[i] = A->i[i+1] - A->i[i];
732: PetscViewerBinaryWrite(viewer,rowlens,m,PETSC_INT);
733: PetscFree(rowlens);
734: /* store column indices */
735: PetscViewerBinaryWrite(viewer,A->j,nz,PETSC_INT);
736: /* store nonzero values */
737: MatSeqAIJGetArrayRead(mat,&av);
738: PetscViewerBinaryWrite(viewer,av,nz,PETSC_SCALAR);
739: MatSeqAIJRestoreArrayRead(mat,&av);
741: /* write block size option to the viewer's .info file */
742: MatView_Binary_BlockSizes(mat,viewer);
743: return(0);
744: }
746: static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A,PetscViewer viewer)
747: {
749: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
750: PetscInt i,k,m=A->rmap->N;
753: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
754: for (i=0; i<m; i++) {
755: PetscViewerASCIIPrintf(viewer,"row %D:",i);
756: for (k=a->i[i]; k<a->i[i+1]; k++) {
757: PetscViewerASCIIPrintf(viewer," (%D) ",a->j[k]);
758: }
759: PetscViewerASCIIPrintf(viewer,"\n");
760: }
761: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
762: return(0);
763: }
765: extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat,PetscViewer);
767: PetscErrorCode MatView_SeqAIJ_ASCII(Mat A,PetscViewer viewer)
768: {
769: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
770: const PetscScalar *av;
771: PetscErrorCode ierr;
772: PetscInt i,j,m = A->rmap->n;
773: const char *name;
774: PetscViewerFormat format;
777: if (A->structure_only) {
778: MatView_SeqAIJ_ASCII_structonly(A,viewer);
779: return(0);
780: }
782: PetscViewerGetFormat(viewer,&format);
783: if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) return(0);
785: /* trigger copy to CPU if needed */
786: MatSeqAIJGetArrayRead(A,&av);
787: MatSeqAIJRestoreArrayRead(A,&av);
788: if (format == PETSC_VIEWER_ASCII_MATLAB) {
789: PetscInt nofinalvalue = 0;
790: if (m && ((a->i[m] == a->i[m-1]) || (a->j[a->nz-1] != A->cmap->n-1))) {
791: /* Need a dummy value to ensure the dimension of the matrix. */
792: nofinalvalue = 1;
793: }
794: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
795: PetscViewerASCIIPrintf(viewer,"%% Size = %D %D \n",m,A->cmap->n);
796: PetscViewerASCIIPrintf(viewer,"%% Nonzeros = %D \n",a->nz);
797: #if defined(PETSC_USE_COMPLEX)
798: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,4);\n",a->nz+nofinalvalue);
799: #else
800: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,3);\n",a->nz+nofinalvalue);
801: #endif
802: PetscViewerASCIIPrintf(viewer,"zzz = [\n");
804: for (i=0; i<m; i++) {
805: for (j=a->i[i]; j<a->i[i+1]; j++) {
806: #if defined(PETSC_USE_COMPLEX)
807: 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]));
808: #else
809: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",i+1,a->j[j]+1,(double)a->a[j]);
810: #endif
811: }
812: }
813: if (nofinalvalue) {
814: #if defined(PETSC_USE_COMPLEX)
815: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",m,A->cmap->n,0.,0.);
816: #else
817: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",m,A->cmap->n,0.0);
818: #endif
819: }
820: PetscObjectGetName((PetscObject)A,&name);
821: PetscViewerASCIIPrintf(viewer,"];\n %s = spconvert(zzz);\n",name);
822: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
823: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
824: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
825: for (i=0; i<m; i++) {
826: PetscViewerASCIIPrintf(viewer,"row %D:",i);
827: for (j=a->i[i]; j<a->i[i+1]; j++) {
828: #if defined(PETSC_USE_COMPLEX)
829: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
830: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
831: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
832: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
833: } else if (PetscRealPart(a->a[j]) != 0.0) {
834: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
835: }
836: #else
837: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);}
838: #endif
839: }
840: PetscViewerASCIIPrintf(viewer,"\n");
841: }
842: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
843: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
844: PetscInt nzd=0,fshift=1,*sptr;
845: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
846: PetscMalloc1(m+1,&sptr);
847: for (i=0; i<m; i++) {
848: sptr[i] = nzd+1;
849: for (j=a->i[i]; j<a->i[i+1]; j++) {
850: if (a->j[j] >= i) {
851: #if defined(PETSC_USE_COMPLEX)
852: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++;
853: #else
854: if (a->a[j] != 0.0) nzd++;
855: #endif
856: }
857: }
858: }
859: sptr[m] = nzd+1;
860: PetscViewerASCIIPrintf(viewer," %D %D\n\n",m,nzd);
861: for (i=0; i<m+1; i+=6) {
862: if (i+4<m) {
863: 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]);
864: } else if (i+3<m) {
865: PetscViewerASCIIPrintf(viewer," %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4]);
866: } else if (i+2<m) {
867: PetscViewerASCIIPrintf(viewer," %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3]);
868: } else if (i+1<m) {
869: PetscViewerASCIIPrintf(viewer," %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2]);
870: } else if (i<m) {
871: PetscViewerASCIIPrintf(viewer," %D %D\n",sptr[i],sptr[i+1]);
872: } else {
873: PetscViewerASCIIPrintf(viewer," %D\n",sptr[i]);
874: }
875: }
876: PetscViewerASCIIPrintf(viewer,"\n");
877: PetscFree(sptr);
878: for (i=0; i<m; i++) {
879: for (j=a->i[i]; j<a->i[i+1]; j++) {
880: if (a->j[j] >= i) {PetscViewerASCIIPrintf(viewer," %D ",a->j[j]+fshift);}
881: }
882: PetscViewerASCIIPrintf(viewer,"\n");
883: }
884: PetscViewerASCIIPrintf(viewer,"\n");
885: for (i=0; i<m; i++) {
886: for (j=a->i[i]; j<a->i[i+1]; j++) {
887: if (a->j[j] >= i) {
888: #if defined(PETSC_USE_COMPLEX)
889: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) {
890: PetscViewerASCIIPrintf(viewer," %18.16e %18.16e ",(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
891: }
892: #else
893: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," %18.16e ",(double)a->a[j]);}
894: #endif
895: }
896: }
897: PetscViewerASCIIPrintf(viewer,"\n");
898: }
899: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
900: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
901: PetscInt cnt = 0,jcnt;
902: PetscScalar value;
903: #if defined(PETSC_USE_COMPLEX)
904: PetscBool realonly = PETSC_TRUE;
906: for (i=0; i<a->i[m]; i++) {
907: if (PetscImaginaryPart(a->a[i]) != 0.0) {
908: realonly = PETSC_FALSE;
909: break;
910: }
911: }
912: #endif
914: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
915: for (i=0; i<m; i++) {
916: jcnt = 0;
917: for (j=0; j<A->cmap->n; j++) {
918: if (jcnt < a->i[i+1]-a->i[i] && j == a->j[cnt]) {
919: value = a->a[cnt++];
920: jcnt++;
921: } else {
922: value = 0.0;
923: }
924: #if defined(PETSC_USE_COMPLEX)
925: if (realonly) {
926: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)PetscRealPart(value));
927: } else {
928: PetscViewerASCIIPrintf(viewer," %7.5e+%7.5e i ",(double)PetscRealPart(value),(double)PetscImaginaryPart(value));
929: }
930: #else
931: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)value);
932: #endif
933: }
934: PetscViewerASCIIPrintf(viewer,"\n");
935: }
936: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
937: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
938: PetscInt fshift=1;
939: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
940: #if defined(PETSC_USE_COMPLEX)
941: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate complex general\n");
942: #else
943: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate real general\n");
944: #endif
945: PetscViewerASCIIPrintf(viewer,"%D %D %D\n", m, A->cmap->n, a->nz);
946: for (i=0; i<m; i++) {
947: for (j=a->i[i]; j<a->i[i+1]; j++) {
948: #if defined(PETSC_USE_COMPLEX)
949: PetscViewerASCIIPrintf(viewer,"%D %D %g %g\n", i+fshift,a->j[j]+fshift,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
950: #else
951: PetscViewerASCIIPrintf(viewer,"%D %D %g\n", i+fshift, a->j[j]+fshift, (double)a->a[j]);
952: #endif
953: }
954: }
955: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
956: } else {
957: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
958: if (A->factortype) {
959: for (i=0; i<m; i++) {
960: PetscViewerASCIIPrintf(viewer,"row %D:",i);
961: /* L part */
962: for (j=a->i[i]; j<a->i[i+1]; j++) {
963: #if defined(PETSC_USE_COMPLEX)
964: if (PetscImaginaryPart(a->a[j]) > 0.0) {
965: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
966: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
967: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
968: } else {
969: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
970: }
971: #else
972: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
973: #endif
974: }
975: /* diagonal */
976: j = a->diag[i];
977: #if defined(PETSC_USE_COMPLEX)
978: if (PetscImaginaryPart(a->a[j]) > 0.0) {
979: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)PetscImaginaryPart(1.0/a->a[j]));
980: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
981: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)(-PetscImaginaryPart(1.0/a->a[j])));
982: } else {
983: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(1.0/a->a[j]));
984: }
985: #else
986: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)(1.0/a->a[j]));
987: #endif
989: /* U part */
990: for (j=a->diag[i+1]+1; j<a->diag[i]; j++) {
991: #if defined(PETSC_USE_COMPLEX)
992: 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 if (PetscImaginaryPart(a->a[j]) < 0.0) {
995: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
996: } else {
997: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
998: }
999: #else
1000: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
1001: #endif
1002: }
1003: PetscViewerASCIIPrintf(viewer,"\n");
1004: }
1005: } else {
1006: for (i=0; i<m; i++) {
1007: PetscViewerASCIIPrintf(viewer,"row %D:",i);
1008: for (j=a->i[i]; j<a->i[i+1]; j++) {
1009: #if defined(PETSC_USE_COMPLEX)
1010: if (PetscImaginaryPart(a->a[j]) > 0.0) {
1011: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
1012: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
1013: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
1014: } else {
1015: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
1016: }
1017: #else
1018: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
1019: #endif
1020: }
1021: PetscViewerASCIIPrintf(viewer,"\n");
1022: }
1023: }
1024: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
1025: }
1026: PetscViewerFlush(viewer);
1027: return(0);
1028: }
1030: #include <petscdraw.h>
1031: PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw,void *Aa)
1032: {
1033: Mat A = (Mat) Aa;
1034: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1035: PetscErrorCode ierr;
1036: PetscInt i,j,m = A->rmap->n;
1037: int color;
1038: PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r;
1039: PetscViewer viewer;
1040: PetscViewerFormat format;
1043: PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);
1044: PetscViewerGetFormat(viewer,&format);
1045: PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);
1047: /* loop over matrix elements drawing boxes */
1049: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
1050: PetscDrawCollectiveBegin(draw);
1051: /* Blue for negative, Cyan for zero and Red for positive */
1052: color = PETSC_DRAW_BLUE;
1053: for (i=0; i<m; i++) {
1054: y_l = m - i - 1.0; y_r = y_l + 1.0;
1055: for (j=a->i[i]; j<a->i[i+1]; j++) {
1056: x_l = a->j[j]; x_r = x_l + 1.0;
1057: if (PetscRealPart(a->a[j]) >= 0.) continue;
1058: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1059: }
1060: }
1061: color = PETSC_DRAW_CYAN;
1062: for (i=0; i<m; i++) {
1063: y_l = m - i - 1.0; y_r = y_l + 1.0;
1064: for (j=a->i[i]; j<a->i[i+1]; j++) {
1065: x_l = a->j[j]; x_r = x_l + 1.0;
1066: if (a->a[j] != 0.) continue;
1067: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1068: }
1069: }
1070: color = PETSC_DRAW_RED;
1071: for (i=0; i<m; i++) {
1072: y_l = m - i - 1.0; y_r = y_l + 1.0;
1073: for (j=a->i[i]; j<a->i[i+1]; j++) {
1074: x_l = a->j[j]; x_r = x_l + 1.0;
1075: if (PetscRealPart(a->a[j]) <= 0.) continue;
1076: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1077: }
1078: }
1079: PetscDrawCollectiveEnd(draw);
1080: } else {
1081: /* use contour shading to indicate magnitude of values */
1082: /* first determine max of all nonzero values */
1083: PetscReal minv = 0.0, maxv = 0.0;
1084: PetscInt nz = a->nz, count = 0;
1085: PetscDraw popup;
1087: for (i=0; i<nz; i++) {
1088: if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]);
1089: }
1090: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1091: PetscDrawGetPopup(draw,&popup);
1092: PetscDrawScalePopup(popup,minv,maxv);
1094: PetscDrawCollectiveBegin(draw);
1095: for (i=0; i<m; i++) {
1096: y_l = m - i - 1.0;
1097: y_r = y_l + 1.0;
1098: for (j=a->i[i]; j<a->i[i+1]; j++) {
1099: x_l = a->j[j];
1100: x_r = x_l + 1.0;
1101: color = PetscDrawRealToColor(PetscAbsScalar(a->a[count]),minv,maxv);
1102: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
1103: count++;
1104: }
1105: }
1106: PetscDrawCollectiveEnd(draw);
1107: }
1108: return(0);
1109: }
1111: #include <petscdraw.h>
1112: PetscErrorCode MatView_SeqAIJ_Draw(Mat A,PetscViewer viewer)
1113: {
1115: PetscDraw draw;
1116: PetscReal xr,yr,xl,yl,h,w;
1117: PetscBool isnull;
1120: PetscViewerDrawGetDraw(viewer,0,&draw);
1121: PetscDrawIsNull(draw,&isnull);
1122: if (isnull) return(0);
1124: xr = A->cmap->n; yr = A->rmap->n; h = yr/10.0; w = xr/10.0;
1125: xr += w; yr += h; xl = -w; yl = -h;
1126: PetscDrawSetCoordinates(draw,xl,yl,xr,yr);
1127: PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);
1128: PetscDrawZoom(draw,MatView_SeqAIJ_Draw_Zoom,A);
1129: PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);
1130: PetscDrawSave(draw);
1131: return(0);
1132: }
1134: PetscErrorCode MatView_SeqAIJ(Mat A,PetscViewer viewer)
1135: {
1137: PetscBool iascii,isbinary,isdraw;
1140: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1141: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1142: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1143: if (iascii) {
1144: MatView_SeqAIJ_ASCII(A,viewer);
1145: } else if (isbinary) {
1146: MatView_SeqAIJ_Binary(A,viewer);
1147: } else if (isdraw) {
1148: MatView_SeqAIJ_Draw(A,viewer);
1149: }
1150: MatView_SeqAIJ_Inode(A,viewer);
1151: return(0);
1152: }
1154: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A,MatAssemblyType mode)
1155: {
1156: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1158: PetscInt fshift = 0,i,*ai = a->i,*aj = a->j,*imax = a->imax;
1159: PetscInt m = A->rmap->n,*ip,N,*ailen = a->ilen,rmax = 0;
1160: MatScalar *aa = a->a,*ap;
1161: PetscReal ratio = 0.6;
1164: if (mode == MAT_FLUSH_ASSEMBLY) return(0);
1165: MatSeqAIJInvalidateDiagonal(A);
1166: if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) {
1167: /* we need to respect users asking to use or not the inodes routine in between matrix assemblies */
1168: MatAssemblyEnd_SeqAIJ_Inode(A,mode);
1169: return(0);
1170: }
1172: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
1173: for (i=1; i<m; i++) {
1174: /* move each row back by the amount of empty slots (fshift) before it*/
1175: fshift += imax[i-1] - ailen[i-1];
1176: rmax = PetscMax(rmax,ailen[i]);
1177: if (fshift) {
1178: ip = aj + ai[i];
1179: ap = aa + ai[i];
1180: N = ailen[i];
1181: PetscArraymove(ip-fshift,ip,N);
1182: if (!A->structure_only) {
1183: PetscArraymove(ap-fshift,ap,N);
1184: }
1185: }
1186: ai[i] = ai[i-1] + ailen[i-1];
1187: }
1188: if (m) {
1189: fshift += imax[m-1] - ailen[m-1];
1190: ai[m] = ai[m-1] + ailen[m-1];
1191: }
1193: /* reset ilen and imax for each row */
1194: a->nonzerorowcnt = 0;
1195: if (A->structure_only) {
1196: PetscFree(a->imax);
1197: PetscFree(a->ilen);
1198: } else { /* !A->structure_only */
1199: for (i=0; i<m; i++) {
1200: ailen[i] = imax[i] = ai[i+1] - ai[i];
1201: a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0);
1202: }
1203: }
1204: a->nz = ai[m];
1205: 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);
1207: MatMarkDiagonal_SeqAIJ(A);
1208: PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap->n,fshift,a->nz);
1209: PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);
1210: PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);
1212: A->info.mallocs += a->reallocs;
1213: a->reallocs = 0;
1214: A->info.nz_unneeded = (PetscReal)fshift;
1215: a->rmax = rmax;
1217: if (!A->structure_only) {
1218: MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,m,ratio);
1219: }
1220: MatAssemblyEnd_SeqAIJ_Inode(A,mode);
1221: return(0);
1222: }
1224: PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1225: {
1226: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1227: PetscInt i,nz = a->nz;
1228: MatScalar *aa;
1232: MatSeqAIJGetArray(A,&aa);
1233: for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]);
1234: MatSeqAIJRestoreArray(A,&aa);
1235: MatSeqAIJInvalidateDiagonal(A);
1236: #if defined(PETSC_HAVE_DEVICE)
1237: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
1238: #endif
1239: return(0);
1240: }
1242: PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1243: {
1244: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1245: PetscInt i,nz = a->nz;
1246: MatScalar *aa;
1250: MatSeqAIJGetArray(A,&aa);
1251: for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1252: MatSeqAIJRestoreArray(A,&aa);
1253: MatSeqAIJInvalidateDiagonal(A);
1254: #if defined(PETSC_HAVE_DEVICE)
1255: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
1256: #endif
1257: return(0);
1258: }
1260: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1261: {
1262: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1266: PetscArrayzero(a->a,a->i[A->rmap->n]);
1267: MatSeqAIJInvalidateDiagonal(A);
1268: #if defined(PETSC_HAVE_DEVICE)
1269: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
1270: #endif
1271: return(0);
1272: }
1274: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1275: {
1276: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1280: #if defined(PETSC_USE_LOG)
1281: PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->n,A->cmap->n,a->nz);
1282: #endif
1283: MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);
1284: ISDestroy(&a->row);
1285: ISDestroy(&a->col);
1286: PetscFree(a->diag);
1287: PetscFree(a->ibdiag);
1288: PetscFree(a->imax);
1289: PetscFree(a->ilen);
1290: PetscFree(a->ipre);
1291: PetscFree3(a->idiag,a->mdiag,a->ssor_work);
1292: PetscFree(a->solve_work);
1293: ISDestroy(&a->icol);
1294: PetscFree(a->saved_values);
1295: PetscFree2(a->compressedrow.i,a->compressedrow.rindex);
1297: MatDestroy_SeqAIJ_Inode(A);
1298: PetscFree(A->data);
1300: /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this.
1301: That function is so heavily used (sometimes in an hidden way through multnumeric function pointers)
1302: that is hard to properly add this data to the MatProduct data. We free it here to avoid
1303: users reusing the matrix object with different data to incur in obscure segmentation faults
1304: due to different matrix sizes */
1305: PetscObjectCompose((PetscObject)A,"__PETSc__ab_dense",NULL);
1307: PetscObjectChangeTypeName((PetscObject)A,NULL);
1308: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetColumnIndices_C",NULL);
1309: PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);
1310: PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);
1311: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsbaij_C",NULL);
1312: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqbaij_C",NULL);
1313: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijperm_C",NULL);
1314: #if defined(PETSC_HAVE_CUDA)
1315: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijcusparse_C",NULL);
1316: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqaijcusparse_seqaij_C",NULL);
1317: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqaij_seqaijcusparse_C",NULL);
1318: #endif
1319: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
1320: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijkokkos_C",NULL);
1321: #endif
1322: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijcrl_C",NULL);
1323: #if defined(PETSC_HAVE_ELEMENTAL)
1324: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_elemental_C",NULL);
1325: #endif
1326: #if defined(PETSC_HAVE_SCALAPACK)
1327: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_scalapack_C",NULL);
1328: #endif
1329: #if defined(PETSC_HAVE_HYPRE)
1330: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_hypre_C",NULL);
1331: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_transpose_seqaij_seqaij_C",NULL);
1332: #endif
1333: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqdense_C",NULL);
1334: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsell_C",NULL);
1335: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_is_C",NULL);
1336: PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);
1337: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);
1338: PetscObjectComposeFunction((PetscObject)A,"MatResetPreallocation_C",NULL);
1339: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);
1340: PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);
1341: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_is_seqaij_C",NULL);
1342: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqdense_seqaij_C",NULL);
1343: PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqaij_seqaij_C",NULL);
1344: return(0);
1345: }
1347: PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg)
1348: {
1349: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1353: switch (op) {
1354: case MAT_ROW_ORIENTED:
1355: a->roworiented = flg;
1356: break;
1357: case MAT_KEEP_NONZERO_PATTERN:
1358: a->keepnonzeropattern = flg;
1359: break;
1360: case MAT_NEW_NONZERO_LOCATIONS:
1361: a->nonew = (flg ? 0 : 1);
1362: break;
1363: case MAT_NEW_NONZERO_LOCATION_ERR:
1364: a->nonew = (flg ? -1 : 0);
1365: break;
1366: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1367: a->nonew = (flg ? -2 : 0);
1368: break;
1369: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1370: a->nounused = (flg ? -1 : 0);
1371: break;
1372: case MAT_IGNORE_ZERO_ENTRIES:
1373: a->ignorezeroentries = flg;
1374: break;
1375: case MAT_SPD:
1376: case MAT_SYMMETRIC:
1377: case MAT_STRUCTURALLY_SYMMETRIC:
1378: case MAT_HERMITIAN:
1379: case MAT_SYMMETRY_ETERNAL:
1380: case MAT_STRUCTURE_ONLY:
1381: /* These options are handled directly by MatSetOption() */
1382: break;
1383: case MAT_FORCE_DIAGONAL_ENTRIES:
1384: case MAT_IGNORE_OFF_PROC_ENTRIES:
1385: case MAT_USE_HASH_TABLE:
1386: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1387: break;
1388: case MAT_USE_INODES:
1389: MatSetOption_SeqAIJ_Inode(A,MAT_USE_INODES,flg);
1390: break;
1391: case MAT_SUBMAT_SINGLEIS:
1392: A->submat_singleis = flg;
1393: break;
1394: case MAT_SORTED_FULL:
1395: if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
1396: else A->ops->setvalues = MatSetValues_SeqAIJ;
1397: break;
1398: default:
1399: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1400: }
1401: return(0);
1402: }
1404: PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v)
1405: {
1406: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1407: PetscErrorCode ierr;
1408: PetscInt i,j,n,*ai=a->i,*aj=a->j;
1409: PetscScalar *x;
1410: const PetscScalar *aa;
1413: VecGetLocalSize(v,&n);
1414: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
1415: MatSeqAIJGetArrayRead(A,&aa);
1416: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1417: PetscInt *diag=a->diag;
1418: VecGetArrayWrite(v,&x);
1419: for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]];
1420: VecRestoreArrayWrite(v,&x);
1421: MatSeqAIJRestoreArrayRead(A,&aa);
1422: return(0);
1423: }
1425: VecGetArrayWrite(v,&x);
1426: for (i=0; i<n; i++) {
1427: x[i] = 0.0;
1428: for (j=ai[i]; j<ai[i+1]; j++) {
1429: if (aj[j] == i) {
1430: x[i] = aa[j];
1431: break;
1432: }
1433: }
1434: }
1435: VecRestoreArrayWrite(v,&x);
1436: MatSeqAIJRestoreArrayRead(A,&aa);
1437: return(0);
1438: }
1440: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1441: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy)
1442: {
1443: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1444: PetscScalar *y;
1445: const PetscScalar *x;
1446: PetscErrorCode ierr;
1447: PetscInt m = A->rmap->n;
1448: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1449: const MatScalar *v;
1450: PetscScalar alpha;
1451: PetscInt n,i,j;
1452: const PetscInt *idx,*ii,*ridx=NULL;
1453: Mat_CompressedRow cprow = a->compressedrow;
1454: PetscBool usecprow = cprow.use;
1455: #endif
1458: if (zz != yy) {VecCopy(zz,yy);}
1459: VecGetArrayRead(xx,&x);
1460: VecGetArray(yy,&y);
1462: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1463: fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y);
1464: #else
1465: if (usecprow) {
1466: m = cprow.nrows;
1467: ii = cprow.i;
1468: ridx = cprow.rindex;
1469: } else {
1470: ii = a->i;
1471: }
1472: for (i=0; i<m; i++) {
1473: idx = a->j + ii[i];
1474: v = a->a + ii[i];
1475: n = ii[i+1] - ii[i];
1476: if (usecprow) {
1477: alpha = x[ridx[i]];
1478: } else {
1479: alpha = x[i];
1480: }
1481: for (j=0; j<n; j++) y[idx[j]] += alpha*v[j];
1482: }
1483: #endif
1484: PetscLogFlops(2.0*a->nz);
1485: VecRestoreArrayRead(xx,&x);
1486: VecRestoreArray(yy,&y);
1487: return(0);
1488: }
1490: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy)
1491: {
1495: VecSet(yy,0.0);
1496: MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);
1497: return(0);
1498: }
1500: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1502: PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy)
1503: {
1504: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1505: PetscScalar *y;
1506: const PetscScalar *x;
1507: const MatScalar *aa;
1508: PetscErrorCode ierr;
1509: PetscInt m=A->rmap->n;
1510: const PetscInt *aj,*ii,*ridx=NULL;
1511: PetscInt n,i;
1512: PetscScalar sum;
1513: PetscBool usecprow=a->compressedrow.use;
1515: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1516: #pragma disjoint(*x,*y,*aa)
1517: #endif
1520: if (a->inode.use && a->inode.checked) {
1521: MatMult_SeqAIJ_Inode(A,xx,yy);
1522: return(0);
1523: }
1524: VecGetArrayRead(xx,&x);
1525: VecGetArray(yy,&y);
1526: ii = a->i;
1527: if (usecprow) { /* use compressed row format */
1528: PetscArrayzero(y,m);
1529: m = a->compressedrow.nrows;
1530: ii = a->compressedrow.i;
1531: ridx = a->compressedrow.rindex;
1532: for (i=0; i<m; i++) {
1533: n = ii[i+1] - ii[i];
1534: aj = a->j + ii[i];
1535: aa = a->a + ii[i];
1536: sum = 0.0;
1537: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1538: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1539: y[*ridx++] = sum;
1540: }
1541: } else { /* do not use compressed row format */
1542: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1543: aj = a->j;
1544: aa = a->a;
1545: fortranmultaij_(&m,x,ii,aj,aa,y);
1546: #else
1547: for (i=0; i<m; i++) {
1548: n = ii[i+1] - ii[i];
1549: aj = a->j + ii[i];
1550: aa = a->a + ii[i];
1551: sum = 0.0;
1552: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1553: y[i] = sum;
1554: }
1555: #endif
1556: }
1557: PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);
1558: VecRestoreArrayRead(xx,&x);
1559: VecRestoreArray(yy,&y);
1560: return(0);
1561: }
1563: PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy)
1564: {
1565: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1566: PetscScalar *y;
1567: const PetscScalar *x;
1568: const MatScalar *aa;
1569: PetscErrorCode ierr;
1570: PetscInt m=A->rmap->n;
1571: const PetscInt *aj,*ii,*ridx=NULL;
1572: PetscInt n,i,nonzerorow=0;
1573: PetscScalar sum;
1574: PetscBool usecprow=a->compressedrow.use;
1576: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1577: #pragma disjoint(*x,*y,*aa)
1578: #endif
1581: VecGetArrayRead(xx,&x);
1582: VecGetArray(yy,&y);
1583: if (usecprow) { /* use compressed row format */
1584: m = a->compressedrow.nrows;
1585: ii = a->compressedrow.i;
1586: ridx = a->compressedrow.rindex;
1587: for (i=0; i<m; i++) {
1588: n = ii[i+1] - ii[i];
1589: aj = a->j + ii[i];
1590: aa = a->a + ii[i];
1591: sum = 0.0;
1592: nonzerorow += (n>0);
1593: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1594: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1595: y[*ridx++] = sum;
1596: }
1597: } else { /* do not use compressed row format */
1598: ii = a->i;
1599: for (i=0; i<m; i++) {
1600: n = ii[i+1] - ii[i];
1601: aj = a->j + ii[i];
1602: aa = a->a + ii[i];
1603: sum = 0.0;
1604: nonzerorow += (n>0);
1605: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1606: y[i] = sum;
1607: }
1608: }
1609: PetscLogFlops(2.0*a->nz - nonzerorow);
1610: VecRestoreArrayRead(xx,&x);
1611: VecRestoreArray(yy,&y);
1612: return(0);
1613: }
1615: PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1616: {
1617: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1618: PetscScalar *y,*z;
1619: const PetscScalar *x;
1620: const MatScalar *aa;
1621: PetscErrorCode ierr;
1622: PetscInt m = A->rmap->n,*aj,*ii;
1623: PetscInt n,i,*ridx=NULL;
1624: PetscScalar sum;
1625: PetscBool usecprow=a->compressedrow.use;
1628: VecGetArrayRead(xx,&x);
1629: VecGetArrayPair(yy,zz,&y,&z);
1630: if (usecprow) { /* use compressed row format */
1631: if (zz != yy) {
1632: PetscArraycpy(z,y,m);
1633: }
1634: m = a->compressedrow.nrows;
1635: ii = a->compressedrow.i;
1636: ridx = a->compressedrow.rindex;
1637: for (i=0; i<m; i++) {
1638: n = ii[i+1] - ii[i];
1639: aj = a->j + ii[i];
1640: aa = a->a + ii[i];
1641: sum = y[*ridx];
1642: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1643: z[*ridx++] = sum;
1644: }
1645: } else { /* do not use compressed row format */
1646: ii = a->i;
1647: for (i=0; i<m; i++) {
1648: n = ii[i+1] - ii[i];
1649: aj = a->j + ii[i];
1650: aa = a->a + ii[i];
1651: sum = y[i];
1652: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1653: z[i] = sum;
1654: }
1655: }
1656: PetscLogFlops(2.0*a->nz);
1657: VecRestoreArrayRead(xx,&x);
1658: VecRestoreArrayPair(yy,zz,&y,&z);
1659: return(0);
1660: }
1662: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1663: PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1664: {
1665: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1666: PetscScalar *y,*z;
1667: const PetscScalar *x;
1668: const MatScalar *aa;
1669: PetscErrorCode ierr;
1670: const PetscInt *aj,*ii,*ridx=NULL;
1671: PetscInt m = A->rmap->n,n,i;
1672: PetscScalar sum;
1673: PetscBool usecprow=a->compressedrow.use;
1676: if (a->inode.use && a->inode.checked) {
1677: MatMultAdd_SeqAIJ_Inode(A,xx,yy,zz);
1678: return(0);
1679: }
1680: VecGetArrayRead(xx,&x);
1681: VecGetArrayPair(yy,zz,&y,&z);
1682: if (usecprow) { /* use compressed row format */
1683: if (zz != yy) {
1684: PetscArraycpy(z,y,m);
1685: }
1686: m = a->compressedrow.nrows;
1687: ii = a->compressedrow.i;
1688: ridx = a->compressedrow.rindex;
1689: for (i=0; i<m; i++) {
1690: n = ii[i+1] - ii[i];
1691: aj = a->j + ii[i];
1692: aa = a->a + ii[i];
1693: sum = y[*ridx];
1694: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1695: z[*ridx++] = sum;
1696: }
1697: } else { /* do not use compressed row format */
1698: ii = a->i;
1699: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1700: aj = a->j;
1701: aa = a->a;
1702: fortranmultaddaij_(&m,x,ii,aj,aa,y,z);
1703: #else
1704: for (i=0; i<m; i++) {
1705: n = ii[i+1] - ii[i];
1706: aj = a->j + ii[i];
1707: aa = a->a + ii[i];
1708: sum = y[i];
1709: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1710: z[i] = sum;
1711: }
1712: #endif
1713: }
1714: PetscLogFlops(2.0*a->nz);
1715: VecRestoreArrayRead(xx,&x);
1716: VecRestoreArrayPair(yy,zz,&y,&z);
1717: return(0);
1718: }
1720: /*
1721: Adds diagonal pointers to sparse matrix structure.
1722: */
1723: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1724: {
1725: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1727: PetscInt i,j,m = A->rmap->n;
1730: if (!a->diag) {
1731: PetscMalloc1(m,&a->diag);
1732: PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));
1733: }
1734: for (i=0; i<A->rmap->n; i++) {
1735: a->diag[i] = a->i[i+1];
1736: for (j=a->i[i]; j<a->i[i+1]; j++) {
1737: if (a->j[j] == i) {
1738: a->diag[i] = j;
1739: break;
1740: }
1741: }
1742: }
1743: return(0);
1744: }
1746: PetscErrorCode MatShift_SeqAIJ(Mat A,PetscScalar v)
1747: {
1748: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1749: const PetscInt *diag = (const PetscInt*)a->diag;
1750: const PetscInt *ii = (const PetscInt*) a->i;
1751: PetscInt i,*mdiag = NULL;
1752: PetscErrorCode ierr;
1753: PetscInt cnt = 0; /* how many diagonals are missing */
1756: if (!A->preallocated || !a->nz) {
1757: MatSeqAIJSetPreallocation(A,1,NULL);
1758: MatShift_Basic(A,v);
1759: return(0);
1760: }
1762: if (a->diagonaldense) {
1763: cnt = 0;
1764: } else {
1765: PetscCalloc1(A->rmap->n,&mdiag);
1766: for (i=0; i<A->rmap->n; i++) {
1767: if (diag[i] >= ii[i+1]) {
1768: cnt++;
1769: mdiag[i] = 1;
1770: }
1771: }
1772: }
1773: if (!cnt) {
1774: MatShift_Basic(A,v);
1775: } else {
1776: PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */
1777: PetscInt *oldj = a->j, *oldi = a->i;
1778: PetscBool singlemalloc = a->singlemalloc,free_a = a->free_a,free_ij = a->free_ij;
1780: a->a = NULL;
1781: a->j = NULL;
1782: a->i = NULL;
1783: /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */
1784: for (i=0; i<A->rmap->n; i++) {
1785: a->imax[i] += mdiag[i];
1786: a->imax[i] = PetscMin(a->imax[i],A->cmap->n);
1787: }
1788: MatSeqAIJSetPreallocation_SeqAIJ(A,0,a->imax);
1790: /* copy old values into new matrix data structure */
1791: for (i=0; i<A->rmap->n; i++) {
1792: MatSetValues(A,1,&i,a->imax[i] - mdiag[i],&oldj[oldi[i]],&olda[oldi[i]],ADD_VALUES);
1793: if (i < A->cmap->n) {
1794: MatSetValue(A,i,i,v,ADD_VALUES);
1795: }
1796: }
1797: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1798: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1799: if (singlemalloc) {
1800: PetscFree3(olda,oldj,oldi);
1801: } else {
1802: if (free_a) {PetscFree(olda);}
1803: if (free_ij) {PetscFree(oldj);}
1804: if (free_ij) {PetscFree(oldi);}
1805: }
1806: }
1807: PetscFree(mdiag);
1808: a->diagonaldense = PETSC_TRUE;
1809: return(0);
1810: }
1812: /*
1813: Checks for missing diagonals
1814: */
1815: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d)
1816: {
1817: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1818: PetscInt *diag,*ii = a->i,i;
1822: *missing = PETSC_FALSE;
1823: if (A->rmap->n > 0 && !ii) {
1824: *missing = PETSC_TRUE;
1825: if (d) *d = 0;
1826: PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");
1827: } else {
1828: PetscInt n;
1829: n = PetscMin(A->rmap->n, A->cmap->n);
1830: diag = a->diag;
1831: for (i=0; i<n; i++) {
1832: if (diag[i] >= ii[i+1]) {
1833: *missing = PETSC_TRUE;
1834: if (d) *d = i;
1835: PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);
1836: break;
1837: }
1838: }
1839: }
1840: return(0);
1841: }
1843: #include <petscblaslapack.h>
1844: #include <petsc/private/kernels/blockinvert.h>
1846: /*
1847: Note that values is allocated externally by the PC and then passed into this routine
1848: */
1849: PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
1850: {
1851: PetscErrorCode ierr;
1852: PetscInt n = A->rmap->n, i, ncnt = 0, *indx,j,bsizemax = 0,*v_pivots;
1853: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
1854: const PetscReal shift = 0.0;
1855: PetscInt ipvt[5];
1856: PetscScalar work[25],*v_work;
1859: allowzeropivot = PetscNot(A->erroriffailure);
1860: for (i=0; i<nblocks; i++) ncnt += bsizes[i];
1861: if (ncnt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Total blocksizes %D doesn't match number matrix rows %D",ncnt,n);
1862: for (i=0; i<nblocks; i++) {
1863: bsizemax = PetscMax(bsizemax,bsizes[i]);
1864: }
1865: PetscMalloc1(bsizemax,&indx);
1866: if (bsizemax > 7) {
1867: PetscMalloc2(bsizemax,&v_work,bsizemax,&v_pivots);
1868: }
1869: ncnt = 0;
1870: for (i=0; i<nblocks; i++) {
1871: for (j=0; j<bsizes[i]; j++) indx[j] = ncnt+j;
1872: MatGetValues(A,bsizes[i],indx,bsizes[i],indx,diag);
1873: switch (bsizes[i]) {
1874: case 1:
1875: *diag = 1.0/(*diag);
1876: break;
1877: case 2:
1878: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
1879: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1880: PetscKernel_A_gets_transpose_A_2(diag);
1881: break;
1882: case 3:
1883: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
1884: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1885: PetscKernel_A_gets_transpose_A_3(diag);
1886: break;
1887: case 4:
1888: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
1889: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1890: PetscKernel_A_gets_transpose_A_4(diag);
1891: break;
1892: case 5:
1893: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
1894: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1895: PetscKernel_A_gets_transpose_A_5(diag);
1896: break;
1897: case 6:
1898: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
1899: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1900: PetscKernel_A_gets_transpose_A_6(diag);
1901: break;
1902: case 7:
1903: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
1904: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1905: PetscKernel_A_gets_transpose_A_7(diag);
1906: break;
1907: default:
1908: PetscKernel_A_gets_inverse_A(bsizes[i],diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
1909: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1910: PetscKernel_A_gets_transpose_A_N(diag,bsizes[i]);
1911: }
1912: ncnt += bsizes[i];
1913: diag += bsizes[i]*bsizes[i];
1914: }
1915: if (bsizemax > 7) {
1916: PetscFree2(v_work,v_pivots);
1917: }
1918: PetscFree(indx);
1919: return(0);
1920: }
1922: /*
1923: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1924: */
1925: PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift)
1926: {
1927: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
1928: PetscErrorCode ierr;
1929: PetscInt i,*diag,m = A->rmap->n;
1930: const MatScalar *v;
1931: PetscScalar *idiag,*mdiag;
1934: if (a->idiagvalid) return(0);
1935: MatMarkDiagonal_SeqAIJ(A);
1936: diag = a->diag;
1937: if (!a->idiag) {
1938: PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);
1939: PetscLogObjectMemory((PetscObject)A,3*m*sizeof(PetscScalar));
1940: }
1942: mdiag = a->mdiag;
1943: idiag = a->idiag;
1944: MatSeqAIJGetArrayRead(A,&v);
1945: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1946: for (i=0; i<m; i++) {
1947: mdiag[i] = v[diag[i]];
1948: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1949: if (PetscRealPart(fshift)) {
1950: PetscInfo1(A,"Zero diagonal on row %D\n",i);
1951: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1952: A->factorerror_zeropivot_value = 0.0;
1953: A->factorerror_zeropivot_row = i;
1954: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i);
1955: }
1956: idiag[i] = 1.0/v[diag[i]];
1957: }
1958: PetscLogFlops(m);
1959: } else {
1960: for (i=0; i<m; i++) {
1961: mdiag[i] = v[diag[i]];
1962: idiag[i] = omega/(fshift + v[diag[i]]);
1963: }
1964: PetscLogFlops(2.0*m);
1965: }
1966: a->idiagvalid = PETSC_TRUE;
1967: MatSeqAIJRestoreArrayRead(A,&v);
1968: return(0);
1969: }
1971: #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h>
1972: PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1973: {
1974: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1975: PetscScalar *x,d,sum,*t,scale;
1976: const MatScalar *v,*idiag=NULL,*mdiag,*aa;
1977: const PetscScalar *b, *bs,*xb, *ts;
1978: PetscErrorCode ierr;
1979: PetscInt n,m = A->rmap->n,i;
1980: const PetscInt *idx,*diag;
1983: if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) {
1984: MatSOR_SeqAIJ_Inode(A,bb,omega,flag,fshift,its,lits,xx);
1985: return(0);
1986: }
1987: its = its*lits;
1989: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1990: if (!a->idiagvalid) {MatInvertDiagonal_SeqAIJ(A,omega,fshift);}
1991: a->fshift = fshift;
1992: a->omega = omega;
1994: diag = a->diag;
1995: t = a->ssor_work;
1996: idiag = a->idiag;
1997: mdiag = a->mdiag;
1999: MatSeqAIJGetArrayRead(A,&aa);
2000: VecGetArray(xx,&x);
2001: VecGetArrayRead(bb,&b);
2002: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
2003: if (flag == SOR_APPLY_UPPER) {
2004: /* apply (U + D/omega) to the vector */
2005: bs = b;
2006: for (i=0; i<m; i++) {
2007: d = fshift + mdiag[i];
2008: n = a->i[i+1] - diag[i] - 1;
2009: idx = a->j + diag[i] + 1;
2010: v = aa + diag[i] + 1;
2011: sum = b[i]*d/omega;
2012: PetscSparseDensePlusDot(sum,bs,v,idx,n);
2013: x[i] = sum;
2014: }
2015: VecRestoreArray(xx,&x);
2016: VecRestoreArrayRead(bb,&b);
2017: MatSeqAIJRestoreArrayRead(A,&aa);
2018: PetscLogFlops(a->nz);
2019: return(0);
2020: }
2022: if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented");
2023: else if (flag & SOR_EISENSTAT) {
2024: /* Let A = L + U + D; where L is lower triangular,
2025: U is upper triangular, E = D/omega; This routine applies
2027: (L + E)^{-1} A (U + E)^{-1}
2029: to a vector efficiently using Eisenstat's trick.
2030: */
2031: scale = (2.0/omega) - 1.0;
2033: /* x = (E + U)^{-1} b */
2034: for (i=m-1; i>=0; i--) {
2035: n = a->i[i+1] - diag[i] - 1;
2036: idx = a->j + diag[i] + 1;
2037: v = aa + diag[i] + 1;
2038: sum = b[i];
2039: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2040: x[i] = sum*idiag[i];
2041: }
2043: /* t = b - (2*E - D)x */
2044: v = aa;
2045: for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i];
2047: /* t = (E + L)^{-1}t */
2048: ts = t;
2049: diag = a->diag;
2050: for (i=0; i<m; i++) {
2051: n = diag[i] - a->i[i];
2052: idx = a->j + a->i[i];
2053: v = aa + a->i[i];
2054: sum = t[i];
2055: PetscSparseDenseMinusDot(sum,ts,v,idx,n);
2056: t[i] = sum*idiag[i];
2057: /* x = x + t */
2058: x[i] += t[i];
2059: }
2061: PetscLogFlops(6.0*m-1 + 2.0*a->nz);
2062: VecRestoreArray(xx,&x);
2063: VecRestoreArrayRead(bb,&b);
2064: return(0);
2065: }
2066: if (flag & SOR_ZERO_INITIAL_GUESS) {
2067: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
2068: for (i=0; i<m; i++) {
2069: n = diag[i] - a->i[i];
2070: idx = a->j + a->i[i];
2071: v = aa + a->i[i];
2072: sum = b[i];
2073: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2074: t[i] = sum;
2075: x[i] = sum*idiag[i];
2076: }
2077: xb = t;
2078: PetscLogFlops(a->nz);
2079: } else xb = b;
2080: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2081: for (i=m-1; i>=0; i--) {
2082: n = a->i[i+1] - diag[i] - 1;
2083: idx = a->j + diag[i] + 1;
2084: v = aa + diag[i] + 1;
2085: sum = xb[i];
2086: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2087: if (xb == b) {
2088: x[i] = sum*idiag[i];
2089: } else {
2090: x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */
2091: }
2092: }
2093: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
2094: }
2095: its--;
2096: }
2097: while (its--) {
2098: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
2099: for (i=0; i<m; i++) {
2100: /* lower */
2101: n = diag[i] - a->i[i];
2102: idx = a->j + a->i[i];
2103: v = aa + a->i[i];
2104: sum = b[i];
2105: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2106: t[i] = sum; /* save application of the lower-triangular part */
2107: /* upper */
2108: n = a->i[i+1] - diag[i] - 1;
2109: idx = a->j + diag[i] + 1;
2110: v = aa + diag[i] + 1;
2111: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2112: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
2113: }
2114: xb = t;
2115: PetscLogFlops(2.0*a->nz);
2116: } else xb = b;
2117: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2118: for (i=m-1; i>=0; i--) {
2119: sum = xb[i];
2120: if (xb == b) {
2121: /* whole matrix (no checkpointing available) */
2122: n = a->i[i+1] - a->i[i];
2123: idx = a->j + a->i[i];
2124: v = aa + a->i[i];
2125: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2126: x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i];
2127: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
2128: n = a->i[i+1] - diag[i] - 1;
2129: idx = a->j + diag[i] + 1;
2130: v = aa + diag[i] + 1;
2131: PetscSparseDenseMinusDot(sum,x,v,idx,n);
2132: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
2133: }
2134: }
2135: if (xb == b) {
2136: PetscLogFlops(2.0*a->nz);
2137: } else {
2138: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
2139: }
2140: }
2141: }
2142: MatSeqAIJRestoreArrayRead(A,&aa);
2143: VecRestoreArray(xx,&x);
2144: VecRestoreArrayRead(bb,&b);
2145: return(0);
2146: }
2149: PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info)
2150: {
2151: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2154: info->block_size = 1.0;
2155: info->nz_allocated = a->maxnz;
2156: info->nz_used = a->nz;
2157: info->nz_unneeded = (a->maxnz - a->nz);
2158: info->assemblies = A->num_ass;
2159: info->mallocs = A->info.mallocs;
2160: info->memory = ((PetscObject)A)->mem;
2161: if (A->factortype) {
2162: info->fill_ratio_given = A->info.fill_ratio_given;
2163: info->fill_ratio_needed = A->info.fill_ratio_needed;
2164: info->factor_mallocs = A->info.factor_mallocs;
2165: } else {
2166: info->fill_ratio_given = 0;
2167: info->fill_ratio_needed = 0;
2168: info->factor_mallocs = 0;
2169: }
2170: return(0);
2171: }
2173: PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
2174: {
2175: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2176: PetscInt i,m = A->rmap->n - 1;
2177: PetscErrorCode ierr;
2178: const PetscScalar *xx;
2179: PetscScalar *bb,*aa;
2180: PetscInt d = 0;
2183: if (x && b) {
2184: VecGetArrayRead(x,&xx);
2185: VecGetArray(b,&bb);
2186: for (i=0; i<N; i++) {
2187: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2188: if (rows[i] >= A->cmap->n) continue;
2189: bb[rows[i]] = diag*xx[rows[i]];
2190: }
2191: VecRestoreArrayRead(x,&xx);
2192: VecRestoreArray(b,&bb);
2193: }
2195: MatSeqAIJGetArray(A,&aa);
2196: if (a->keepnonzeropattern) {
2197: for (i=0; i<N; i++) {
2198: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2199: PetscArrayzero(&aa[a->i[rows[i]]],a->ilen[rows[i]]);
2200: }
2201: if (diag != 0.0) {
2202: for (i=0; i<N; i++) {
2203: d = rows[i];
2204: if (rows[i] >= A->cmap->n) continue;
2205: 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);
2206: }
2207: for (i=0; i<N; i++) {
2208: if (rows[i] >= A->cmap->n) continue;
2209: aa[a->diag[rows[i]]] = diag;
2210: }
2211: }
2212: } else {
2213: if (diag != 0.0) {
2214: for (i=0; i<N; i++) {
2215: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2216: if (a->ilen[rows[i]] > 0) {
2217: if (rows[i] >= A->cmap->n) {
2218: a->ilen[rows[i]] = 0;
2219: } else {
2220: a->ilen[rows[i]] = 1;
2221: aa[a->i[rows[i]]] = diag;
2222: a->j[a->i[rows[i]]] = rows[i];
2223: }
2224: } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */
2225: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
2226: }
2227: }
2228: } else {
2229: for (i=0; i<N; i++) {
2230: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2231: a->ilen[rows[i]] = 0;
2232: }
2233: }
2234: A->nonzerostate++;
2235: }
2236: MatSeqAIJRestoreArray(A,&aa);
2237: #if defined(PETSC_HAVE_DEVICE)
2238: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
2239: #endif
2240: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
2241: return(0);
2242: }
2244: PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
2245: {
2246: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2247: PetscInt i,j,m = A->rmap->n - 1,d = 0;
2248: PetscErrorCode ierr;
2249: PetscBool missing,*zeroed,vecs = PETSC_FALSE;
2250: const PetscScalar *xx;
2251: PetscScalar *bb,*aa;
2254: if (!N) return(0);
2255: MatSeqAIJGetArray(A,&aa);
2256: if (x && b) {
2257: VecGetArrayRead(x,&xx);
2258: VecGetArray(b,&bb);
2259: vecs = PETSC_TRUE;
2260: }
2261: PetscCalloc1(A->rmap->n,&zeroed);
2262: for (i=0; i<N; i++) {
2263: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
2264: PetscArrayzero(&aa[a->i[rows[i]]],a->ilen[rows[i]]);
2266: zeroed[rows[i]] = PETSC_TRUE;
2267: }
2268: for (i=0; i<A->rmap->n; i++) {
2269: if (!zeroed[i]) {
2270: for (j=a->i[i]; j<a->i[i+1]; j++) {
2271: if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) {
2272: if (vecs) bb[i] -= aa[j]*xx[a->j[j]];
2273: aa[j] = 0.0;
2274: }
2275: }
2276: } else if (vecs && i < A->cmap->N) bb[i] = diag*xx[i];
2277: }
2278: if (x && b) {
2279: VecRestoreArrayRead(x,&xx);
2280: VecRestoreArray(b,&bb);
2281: }
2282: PetscFree(zeroed);
2283: if (diag != 0.0) {
2284: MatMissingDiagonal_SeqAIJ(A,&missing,&d);
2285: if (missing) {
2286: for (i=0; i<N; i++) {
2287: if (rows[i] >= A->cmap->N) continue;
2288: 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]);
2289: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
2290: }
2291: } else {
2292: for (i=0; i<N; i++) {
2293: aa[a->diag[rows[i]]] = diag;
2294: }
2295: }
2296: }
2297: MatSeqAIJRestoreArray(A,&aa);
2298: #if defined(PETSC_HAVE_DEVICE)
2299: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
2300: #endif
2301: (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);
2302: return(0);
2303: }
2305: PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
2306: {
2307: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2308: const PetscScalar *aa = a->a;
2309: PetscInt *itmp;
2310: #if defined(PETSC_HAVE_DEVICE)
2311: PetscErrorCode ierr;
2312: PetscBool rest = PETSC_FALSE;
2313: #endif
2316: #if defined(PETSC_HAVE_DEVICE)
2317: if (v && A->offloadmask == PETSC_OFFLOAD_GPU) {
2318: /* triggers copy to CPU */
2319: rest = PETSC_TRUE;
2320: MatSeqAIJGetArrayRead(A,&aa);
2321: } else aa = a->a;
2322: #endif
2323: *nz = a->i[row+1] - a->i[row];
2324: if (v) *v = (PetscScalar*)(aa + a->i[row]);
2325: if (idx) {
2326: itmp = a->j + a->i[row];
2327: if (*nz) *idx = itmp;
2328: else *idx = NULL;
2329: }
2330: #if defined(PETSC_HAVE_DEVICE)
2331: if (rest) {
2332: MatSeqAIJRestoreArrayRead(A,&aa);
2333: }
2334: #endif
2335: return(0);
2336: }
2338: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
2339: {
2341: *nz = 0;
2342: if (idx) *idx = NULL;
2343: if (v) *v = NULL;
2344: return(0);
2345: }
2347: PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm)
2348: {
2349: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2350: const MatScalar *v;
2351: PetscReal sum = 0.0;
2352: PetscErrorCode ierr;
2353: PetscInt i,j;
2356: MatSeqAIJGetArrayRead(A,&v);
2357: if (type == NORM_FROBENIUS) {
2358: #if defined(PETSC_USE_REAL___FP16)
2359: PetscBLASInt one = 1,nz = a->nz;
2360: PetscStackCallBLAS("BLASnrm2",*nrm = BLASnrm2_(&nz,v,&one));
2361: #else
2362: for (i=0; i<a->nz; i++) {
2363: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
2364: }
2365: *nrm = PetscSqrtReal(sum);
2366: #endif
2367: PetscLogFlops(2.0*a->nz);
2368: } else if (type == NORM_1) {
2369: PetscReal *tmp;
2370: PetscInt *jj = a->j;
2371: PetscCalloc1(A->cmap->n+1,&tmp);
2372: *nrm = 0.0;
2373: for (j=0; j<a->nz; j++) {
2374: tmp[*jj++] += PetscAbsScalar(*v); v++;
2375: }
2376: for (j=0; j<A->cmap->n; j++) {
2377: if (tmp[j] > *nrm) *nrm = tmp[j];
2378: }
2379: PetscFree(tmp);
2380: PetscLogFlops(PetscMax(a->nz-1,0));
2381: } else if (type == NORM_INFINITY) {
2382: *nrm = 0.0;
2383: for (j=0; j<A->rmap->n; j++) {
2384: const PetscScalar *v2 = v + a->i[j];
2385: sum = 0.0;
2386: for (i=0; i<a->i[j+1]-a->i[j]; i++) {
2387: sum += PetscAbsScalar(*v2); v2++;
2388: }
2389: if (sum > *nrm) *nrm = sum;
2390: }
2391: PetscLogFlops(PetscMax(a->nz-1,0));
2392: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm");
2393: MatSeqAIJRestoreArrayRead(A,&v);
2394: return(0);
2395: }
2397: /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */
2398: PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B)
2399: {
2401: PetscInt i,j,anzj;
2402: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b;
2403: PetscInt an=A->cmap->N,am=A->rmap->N;
2404: PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j;
2407: /* Allocate space for symbolic transpose info and work array */
2408: PetscCalloc1(an+1,&ati);
2409: PetscMalloc1(ai[am],&atj);
2410: PetscMalloc1(an,&atfill);
2412: /* Walk through aj and count ## of non-zeros in each row of A^T. */
2413: /* Note: offset by 1 for fast conversion into csr format. */
2414: for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1;
2415: /* Form ati for csr format of A^T. */
2416: for (i=0;i<an;i++) ati[i+1] += ati[i];
2418: /* Copy ati into atfill so we have locations of the next free space in atj */
2419: PetscArraycpy(atfill,ati,an);
2421: /* Walk through A row-wise and mark nonzero entries of A^T. */
2422: for (i=0;i<am;i++) {
2423: anzj = ai[i+1] - ai[i];
2424: for (j=0;j<anzj;j++) {
2425: atj[atfill[*aj]] = i;
2426: atfill[*aj++] += 1;
2427: }
2428: }
2430: /* Clean up temporary space and complete requests. */
2431: PetscFree(atfill);
2432: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);
2433: MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2434: MatSetType(*B,((PetscObject)A)->type_name);
2436: b = (Mat_SeqAIJ*)((*B)->data);
2437: b->free_a = PETSC_FALSE;
2438: b->free_ij = PETSC_TRUE;
2439: b->nonew = 0;
2440: return(0);
2441: }
2443: PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2444: {
2445: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2446: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2447: const MatScalar *va,*vb;
2448: PetscErrorCode ierr;
2449: PetscInt ma,na,mb,nb, i;
2452: MatGetSize(A,&ma,&na);
2453: MatGetSize(B,&mb,&nb);
2454: if (ma!=nb || na!=mb) {
2455: *f = PETSC_FALSE;
2456: return(0);
2457: }
2458: MatSeqAIJGetArrayRead(A,&va);
2459: MatSeqAIJGetArrayRead(B,&vb);
2460: aii = aij->i; bii = bij->i;
2461: adx = aij->j; bdx = bij->j;
2462: PetscMalloc1(ma,&aptr);
2463: PetscMalloc1(mb,&bptr);
2464: for (i=0; i<ma; i++) aptr[i] = aii[i];
2465: for (i=0; i<mb; i++) bptr[i] = bii[i];
2467: *f = PETSC_TRUE;
2468: for (i=0; i<ma; i++) {
2469: while (aptr[i]<aii[i+1]) {
2470: PetscInt idc,idr;
2471: PetscScalar vc,vr;
2472: /* column/row index/value */
2473: idc = adx[aptr[i]];
2474: idr = bdx[bptr[idc]];
2475: vc = va[aptr[i]];
2476: vr = vb[bptr[idc]];
2477: if (i!=idr || PetscAbsScalar(vc-vr) > tol) {
2478: *f = PETSC_FALSE;
2479: goto done;
2480: } else {
2481: aptr[i]++;
2482: if (B || i!=idc) bptr[idc]++;
2483: }
2484: }
2485: }
2486: done:
2487: PetscFree(aptr);
2488: PetscFree(bptr);
2489: MatSeqAIJRestoreArrayRead(A,&va);
2490: MatSeqAIJRestoreArrayRead(B,&vb);
2491: return(0);
2492: }
2494: PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2495: {
2496: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2497: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2498: MatScalar *va,*vb;
2500: PetscInt ma,na,mb,nb, i;
2503: MatGetSize(A,&ma,&na);
2504: MatGetSize(B,&mb,&nb);
2505: if (ma!=nb || na!=mb) {
2506: *f = PETSC_FALSE;
2507: return(0);
2508: }
2509: aii = aij->i; bii = bij->i;
2510: adx = aij->j; bdx = bij->j;
2511: va = aij->a; vb = bij->a;
2512: PetscMalloc1(ma,&aptr);
2513: PetscMalloc1(mb,&bptr);
2514: for (i=0; i<ma; i++) aptr[i] = aii[i];
2515: for (i=0; i<mb; i++) bptr[i] = bii[i];
2517: *f = PETSC_TRUE;
2518: for (i=0; i<ma; i++) {
2519: while (aptr[i]<aii[i+1]) {
2520: PetscInt idc,idr;
2521: PetscScalar vc,vr;
2522: /* column/row index/value */
2523: idc = adx[aptr[i]];
2524: idr = bdx[bptr[idc]];
2525: vc = va[aptr[i]];
2526: vr = vb[bptr[idc]];
2527: if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) {
2528: *f = PETSC_FALSE;
2529: goto done;
2530: } else {
2531: aptr[i]++;
2532: if (B || i!=idc) bptr[idc]++;
2533: }
2534: }
2535: }
2536: done:
2537: PetscFree(aptr);
2538: PetscFree(bptr);
2539: return(0);
2540: }
2542: PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2543: {
2547: MatIsTranspose_SeqAIJ(A,A,tol,f);
2548: return(0);
2549: }
2551: PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2552: {
2556: MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);
2557: return(0);
2558: }
2560: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr)
2561: {
2562: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2563: const PetscScalar *l,*r;
2564: PetscScalar x;
2565: MatScalar *v;
2566: PetscErrorCode ierr;
2567: PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz;
2568: const PetscInt *jj;
2571: if (ll) {
2572: /* The local size is used so that VecMPI can be passed to this routine
2573: by MatDiagonalScale_MPIAIJ */
2574: VecGetLocalSize(ll,&m);
2575: if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length");
2576: VecGetArrayRead(ll,&l);
2577: MatSeqAIJGetArray(A,&v);
2578: for (i=0; i<m; i++) {
2579: x = l[i];
2580: M = a->i[i+1] - a->i[i];
2581: for (j=0; j<M; j++) (*v++) *= x;
2582: }
2583: VecRestoreArrayRead(ll,&l);
2584: PetscLogFlops(nz);
2585: MatSeqAIJRestoreArray(A,&v);
2586: }
2587: if (rr) {
2588: VecGetLocalSize(rr,&n);
2589: if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length");
2590: VecGetArrayRead(rr,&r);
2591: MatSeqAIJGetArray(A,&v);
2592: jj = a->j;
2593: for (i=0; i<nz; i++) (*v++) *= r[*jj++];
2594: MatSeqAIJRestoreArray(A,&v);
2595: VecRestoreArrayRead(rr,&r);
2596: PetscLogFlops(nz);
2597: }
2598: MatSeqAIJInvalidateDiagonal(A);
2599: #if defined(PETSC_HAVE_DEVICE)
2600: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
2601: #endif
2602: return(0);
2603: }
2605: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B)
2606: {
2607: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c;
2608: PetscErrorCode ierr;
2609: PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens;
2610: PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi;
2611: const PetscInt *irow,*icol;
2612: const PetscScalar *aa;
2613: PetscInt nrows,ncols;
2614: PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen;
2615: MatScalar *a_new,*mat_a;
2616: Mat C;
2617: PetscBool stride;
2620: ISGetIndices(isrow,&irow);
2621: ISGetLocalSize(isrow,&nrows);
2622: ISGetLocalSize(iscol,&ncols);
2624: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);
2625: if (stride) {
2626: ISStrideGetInfo(iscol,&first,&step);
2627: } else {
2628: first = 0;
2629: step = 0;
2630: }
2631: if (stride && step == 1) {
2632: /* special case of contiguous rows */
2633: PetscMalloc2(nrows,&lens,nrows,&starts);
2634: /* loop over new rows determining lens and starting points */
2635: for (i=0; i<nrows; i++) {
2636: kstart = ai[irow[i]];
2637: kend = kstart + ailen[irow[i]];
2638: starts[i] = kstart;
2639: for (k=kstart; k<kend; k++) {
2640: if (aj[k] >= first) {
2641: starts[i] = k;
2642: break;
2643: }
2644: }
2645: sum = 0;
2646: while (k < kend) {
2647: if (aj[k++] >= first+ncols) break;
2648: sum++;
2649: }
2650: lens[i] = sum;
2651: }
2652: /* create submatrix */
2653: if (scall == MAT_REUSE_MATRIX) {
2654: PetscInt n_cols,n_rows;
2655: MatGetSize(*B,&n_rows,&n_cols);
2656: if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
2657: MatZeroEntries(*B);
2658: C = *B;
2659: } else {
2660: PetscInt rbs,cbs;
2661: MatCreate(PetscObjectComm((PetscObject)A),&C);
2662: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2663: ISGetBlockSize(isrow,&rbs);
2664: ISGetBlockSize(iscol,&cbs);
2665: MatSetBlockSizes(C,rbs,cbs);
2666: MatSetType(C,((PetscObject)A)->type_name);
2667: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2668: }
2669: c = (Mat_SeqAIJ*)C->data;
2671: /* loop over rows inserting into submatrix */
2672: a_new = c->a;
2673: j_new = c->j;
2674: i_new = c->i;
2675: MatSeqAIJGetArrayRead(A,&aa);
2676: for (i=0; i<nrows; i++) {
2677: ii = starts[i];
2678: lensi = lens[i];
2679: for (k=0; k<lensi; k++) {
2680: *j_new++ = aj[ii+k] - first;
2681: }
2682: PetscArraycpy(a_new,aa + starts[i],lensi);
2683: a_new += lensi;
2684: i_new[i+1] = i_new[i] + lensi;
2685: c->ilen[i] = lensi;
2686: }
2687: MatSeqAIJRestoreArrayRead(A,&aa);
2688: PetscFree2(lens,starts);
2689: } else {
2690: ISGetIndices(iscol,&icol);
2691: PetscCalloc1(oldcols,&smap);
2692: PetscMalloc1(1+nrows,&lens);
2693: for (i=0; i<ncols; i++) {
2694: 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);
2695: smap[icol[i]] = i+1;
2696: }
2698: /* determine lens of each row */
2699: for (i=0; i<nrows; i++) {
2700: kstart = ai[irow[i]];
2701: kend = kstart + a->ilen[irow[i]];
2702: lens[i] = 0;
2703: for (k=kstart; k<kend; k++) {
2704: if (smap[aj[k]]) {
2705: lens[i]++;
2706: }
2707: }
2708: }
2709: /* Create and fill new matrix */
2710: if (scall == MAT_REUSE_MATRIX) {
2711: PetscBool equal;
2713: c = (Mat_SeqAIJ*)((*B)->data);
2714: if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
2715: PetscArraycmp(c->ilen,lens,(*B)->rmap->n,&equal);
2716: if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
2717: PetscArrayzero(c->ilen,(*B)->rmap->n);
2718: C = *B;
2719: } else {
2720: PetscInt rbs,cbs;
2721: MatCreate(PetscObjectComm((PetscObject)A),&C);
2722: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2723: ISGetBlockSize(isrow,&rbs);
2724: ISGetBlockSize(iscol,&cbs);
2725: MatSetBlockSizes(C,rbs,cbs);
2726: MatSetType(C,((PetscObject)A)->type_name);
2727: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2728: }
2729: MatSeqAIJGetArrayRead(A,&aa);
2730: c = (Mat_SeqAIJ*)(C->data);
2731: for (i=0; i<nrows; i++) {
2732: row = irow[i];
2733: kstart = ai[row];
2734: kend = kstart + a->ilen[row];
2735: mat_i = c->i[i];
2736: mat_j = c->j + mat_i;
2737: mat_a = c->a + mat_i;
2738: mat_ilen = c->ilen + i;
2739: for (k=kstart; k<kend; k++) {
2740: if ((tcol=smap[a->j[k]])) {
2741: *mat_j++ = tcol - 1;
2742: *mat_a++ = aa[k];
2743: (*mat_ilen)++;
2745: }
2746: }
2747: }
2748: MatSeqAIJRestoreArrayRead(A,&aa);
2749: /* Free work space */
2750: ISRestoreIndices(iscol,&icol);
2751: PetscFree(smap);
2752: PetscFree(lens);
2753: /* sort */
2754: for (i = 0; i < nrows; i++) {
2755: PetscInt ilen;
2757: mat_i = c->i[i];
2758: mat_j = c->j + mat_i;
2759: mat_a = c->a + mat_i;
2760: ilen = c->ilen[i];
2761: PetscSortIntWithScalarArray(ilen,mat_j,mat_a);
2762: }
2763: }
2764: #if defined(PETSC_HAVE_DEVICE)
2765: MatBindToCPU(C,A->boundtocpu);
2766: #endif
2767: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2768: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2770: ISRestoreIndices(isrow,&irow);
2771: *B = C;
2772: return(0);
2773: }
2775: PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat)
2776: {
2778: Mat B;
2781: if (scall == MAT_INITIAL_MATRIX) {
2782: MatCreate(subComm,&B);
2783: MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);
2784: MatSetBlockSizesFromMats(B,mat,mat);
2785: MatSetType(B,MATSEQAIJ);
2786: MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);
2787: *subMat = B;
2788: } else {
2789: MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);
2790: }
2791: return(0);
2792: }
2794: PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info)
2795: {
2796: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2798: Mat outA;
2799: PetscBool row_identity,col_identity;
2802: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu");
2804: ISIdentity(row,&row_identity);
2805: ISIdentity(col,&col_identity);
2807: outA = inA;
2808: outA->factortype = MAT_FACTOR_LU;
2809: PetscFree(inA->solvertype);
2810: PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);
2812: PetscObjectReference((PetscObject)row);
2813: ISDestroy(&a->row);
2815: a->row = row;
2817: PetscObjectReference((PetscObject)col);
2818: ISDestroy(&a->col);
2820: a->col = col;
2822: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2823: ISDestroy(&a->icol);
2824: ISInvertPermutation(col,PETSC_DECIDE,&a->icol);
2825: PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);
2827: if (!a->solve_work) { /* this matrix may have been factored before */
2828: PetscMalloc1(inA->rmap->n+1,&a->solve_work);
2829: PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));
2830: }
2832: MatMarkDiagonal_SeqAIJ(inA);
2833: if (row_identity && col_identity) {
2834: MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);
2835: } else {
2836: MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);
2837: }
2838: return(0);
2839: }
2841: PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha)
2842: {
2843: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2844: PetscScalar *v;
2846: PetscBLASInt one = 1,bnz;
2849: MatSeqAIJGetArray(inA,&v);
2850: PetscBLASIntCast(a->nz,&bnz);
2851: PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&alpha,v,&one));
2852: PetscLogFlops(a->nz);
2853: MatSeqAIJRestoreArray(inA,&v);
2854: MatSeqAIJInvalidateDiagonal(inA);
2855: return(0);
2856: }
2858: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2859: {
2861: PetscInt i;
2864: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2865: PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);
2867: for (i=0; i<submatj->nrqr; ++i) {
2868: PetscFree(submatj->sbuf2[i]);
2869: }
2870: PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);
2872: if (submatj->rbuf1) {
2873: PetscFree(submatj->rbuf1[0]);
2874: PetscFree(submatj->rbuf1);
2875: }
2877: for (i=0; i<submatj->nrqs; ++i) {
2878: PetscFree(submatj->rbuf3[i]);
2879: }
2880: PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);
2881: PetscFree(submatj->pa);
2882: }
2884: #if defined(PETSC_USE_CTABLE)
2885: PetscTableDestroy((PetscTable*)&submatj->rmap);
2886: if (submatj->cmap_loc) {PetscFree(submatj->cmap_loc);}
2887: PetscFree(submatj->rmap_loc);
2888: #else
2889: PetscFree(submatj->rmap);
2890: #endif
2892: if (!submatj->allcolumns) {
2893: #if defined(PETSC_USE_CTABLE)
2894: PetscTableDestroy((PetscTable*)&submatj->cmap);
2895: #else
2896: PetscFree(submatj->cmap);
2897: #endif
2898: }
2899: PetscFree(submatj->row2proc);
2901: PetscFree(submatj);
2902: return(0);
2903: }
2905: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2906: {
2908: Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data;
2909: Mat_SubSppt *submatj = c->submatis1;
2912: (*submatj->destroy)(C);
2913: MatDestroySubMatrix_Private(submatj);
2914: return(0);
2915: }
2917: PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n,Mat *mat[])
2918: {
2920: PetscInt i;
2921: Mat C;
2922: Mat_SeqAIJ *c;
2923: Mat_SubSppt *submatj;
2926: for (i=0; i<n; i++) {
2927: C = (*mat)[i];
2928: c = (Mat_SeqAIJ*)C->data;
2929: submatj = c->submatis1;
2930: if (submatj) {
2931: if (--((PetscObject)C)->refct <= 0) {
2932: (*submatj->destroy)(C);
2933: MatDestroySubMatrix_Private(submatj);
2934: PetscFree(C->defaultvectype);
2935: PetscLayoutDestroy(&C->rmap);
2936: PetscLayoutDestroy(&C->cmap);
2937: PetscHeaderDestroy(&C);
2938: }
2939: } else {
2940: MatDestroy(&C);
2941: }
2942: }
2944: /* Destroy Dummy submatrices created for reuse */
2945: MatDestroySubMatrices_Dummy(n,mat);
2947: PetscFree(*mat);
2948: return(0);
2949: }
2951: PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
2952: {
2954: PetscInt i;
2957: if (scall == MAT_INITIAL_MATRIX) {
2958: PetscCalloc1(n+1,B);
2959: }
2961: for (i=0; i<n; i++) {
2962: MatCreateSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);
2963: }
2964: return(0);
2965: }
2967: PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov)
2968: {
2969: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2971: PetscInt row,i,j,k,l,m,n,*nidx,isz,val;
2972: const PetscInt *idx;
2973: PetscInt start,end,*ai,*aj;
2974: PetscBT table;
2977: m = A->rmap->n;
2978: ai = a->i;
2979: aj = a->j;
2981: if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used");
2983: PetscMalloc1(m+1,&nidx);
2984: PetscBTCreate(m,&table);
2986: for (i=0; i<is_max; i++) {
2987: /* Initialize the two local arrays */
2988: isz = 0;
2989: PetscBTMemzero(m,table);
2991: /* Extract the indices, assume there can be duplicate entries */
2992: ISGetIndices(is[i],&idx);
2993: ISGetLocalSize(is[i],&n);
2995: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2996: for (j=0; j<n; ++j) {
2997: if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j];
2998: }
2999: ISRestoreIndices(is[i],&idx);
3000: ISDestroy(&is[i]);
3002: k = 0;
3003: for (j=0; j<ov; j++) { /* for each overlap */
3004: n = isz;
3005: for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */
3006: row = nidx[k];
3007: start = ai[row];
3008: end = ai[row+1];
3009: for (l = start; l<end; l++) {
3010: val = aj[l];
3011: if (!PetscBTLookupSet(table,val)) nidx[isz++] = val;
3012: }
3013: }
3014: }
3015: ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));
3016: }
3017: PetscBTDestroy(&table);
3018: PetscFree(nidx);
3019: return(0);
3020: }
3022: /* -------------------------------------------------------------- */
3023: PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B)
3024: {
3025: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3027: PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n;
3028: const PetscInt *row,*col;
3029: PetscInt *cnew,j,*lens;
3030: IS icolp,irowp;
3031: PetscInt *cwork = NULL;
3032: PetscScalar *vwork = NULL;
3035: ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);
3036: ISGetIndices(irowp,&row);
3037: ISInvertPermutation(colp,PETSC_DECIDE,&icolp);
3038: ISGetIndices(icolp,&col);
3040: /* determine lengths of permuted rows */
3041: PetscMalloc1(m+1,&lens);
3042: for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i];
3043: MatCreate(PetscObjectComm((PetscObject)A),B);
3044: MatSetSizes(*B,m,n,m,n);
3045: MatSetBlockSizesFromMats(*B,A,A);
3046: MatSetType(*B,((PetscObject)A)->type_name);
3047: MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);
3048: PetscFree(lens);
3050: PetscMalloc1(n,&cnew);
3051: for (i=0; i<m; i++) {
3052: MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
3053: for (j=0; j<nz; j++) cnew[j] = col[cwork[j]];
3054: MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);
3055: MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
3056: }
3057: PetscFree(cnew);
3059: (*B)->assembled = PETSC_FALSE;
3061: #if defined(PETSC_HAVE_DEVICE)
3062: MatBindToCPU(*B,A->boundtocpu);
3063: #endif
3064: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
3065: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
3066: ISRestoreIndices(irowp,&row);
3067: ISRestoreIndices(icolp,&col);
3068: ISDestroy(&irowp);
3069: ISDestroy(&icolp);
3070: if (rowp == colp) {
3071: MatPropagateSymmetryOptions(A,*B);
3072: }
3073: return(0);
3074: }
3076: PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str)
3077: {
3081: /* If the two matrices have the same copy implementation, use fast copy. */
3082: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
3083: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3084: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
3085: const PetscScalar *aa;
3087: MatSeqAIJGetArrayRead(A,&aa);
3088: 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]);
3089: PetscArraycpy(b->a,aa,a->i[A->rmap->n]);
3090: PetscObjectStateIncrease((PetscObject)B);
3091: MatSeqAIJRestoreArrayRead(A,&aa);
3092: } else {
3093: MatCopy_Basic(A,B,str);
3094: }
3095: return(0);
3096: }
3098: PetscErrorCode MatSetUp_SeqAIJ(Mat A)
3099: {
3103: MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,NULL);
3104: return(0);
3105: }
3107: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[])
3108: {
3109: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3112: *array = a->a;
3113: return(0);
3114: }
3116: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[])
3117: {
3119: *array = NULL;
3120: return(0);
3121: }
3123: /*
3124: Computes the number of nonzeros per row needed for preallocation when X and Y
3125: have different nonzero structure.
3126: */
3127: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz)
3128: {
3129: PetscInt i,j,k,nzx,nzy;
3132: /* Set the number of nonzeros in the new matrix */
3133: for (i=0; i<m; i++) {
3134: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
3135: nzx = xi[i+1] - xi[i];
3136: nzy = yi[i+1] - yi[i];
3137: nnz[i] = 0;
3138: for (j=0,k=0; j<nzx; j++) { /* Point in X */
3139: for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */
3140: if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */
3141: nnz[i]++;
3142: }
3143: for (; k<nzy; k++) nnz[i]++;
3144: }
3145: return(0);
3146: }
3148: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz)
3149: {
3150: PetscInt m = Y->rmap->N;
3151: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
3152: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
3156: /* Set the number of nonzeros in the new matrix */
3157: MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);
3158: return(0);
3159: }
3161: PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
3162: {
3164: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data;
3167: if (str == UNKNOWN_NONZERO_PATTERN && x->nz == y->nz) {
3168: PetscBool e;
3169: PetscArraycmp(x->i,y->i,Y->rmap->n+1,&e);
3170: if (e) {
3171: PetscArraycmp(x->j,y->j,y->nz,&e);
3172: if (e) {
3173: str = SAME_NONZERO_PATTERN;
3174: }
3175: }
3176: }
3177: if (str == SAME_NONZERO_PATTERN) {
3178: const PetscScalar *xa;
3179: PetscScalar *ya,alpha = a;
3180: PetscBLASInt one = 1,bnz;
3182: PetscBLASIntCast(x->nz,&bnz);
3183: MatSeqAIJGetArray(Y,&ya);
3184: MatSeqAIJGetArrayRead(X,&xa);
3185: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,xa,&one,ya,&one));
3186: MatSeqAIJRestoreArrayRead(X,&xa);
3187: MatSeqAIJRestoreArray(Y,&ya);
3188: PetscLogFlops(2.0*bnz);
3189: MatSeqAIJInvalidateDiagonal(Y);
3190: PetscObjectStateIncrease((PetscObject)Y);
3191: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
3192: MatAXPY_Basic(Y,a,X,str);
3193: } else {
3194: Mat B;
3195: PetscInt *nnz;
3196: PetscMalloc1(Y->rmap->N,&nnz);
3197: MatCreate(PetscObjectComm((PetscObject)Y),&B);
3198: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
3199: MatSetLayouts(B,Y->rmap,Y->cmap);
3200: MatSetType(B,((PetscObject)Y)->type_name);
3201: MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);
3202: MatSeqAIJSetPreallocation(B,0,nnz);
3203: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
3204: MatHeaderReplace(Y,&B);
3205: PetscFree(nnz);
3206: }
3207: return(0);
3208: }
3210: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
3211: {
3212: #if defined(PETSC_USE_COMPLEX)
3213: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3214: PetscInt i,nz;
3215: PetscScalar *a;
3219: nz = aij->nz;
3220: MatSeqAIJGetArray(mat,&a);
3221: for (i=0; i<nz; i++) a[i] = PetscConj(a[i]);
3222: MatSeqAIJRestoreArray(mat,&a);
3223: #else
3225: #endif
3226: return(0);
3227: }
3229: PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3230: {
3231: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3232: PetscErrorCode ierr;
3233: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
3234: PetscReal atmp;
3235: PetscScalar *x;
3236: const MatScalar *aa,*av;
3239: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3240: MatSeqAIJGetArrayRead(A,&av);
3241: aa = av;
3242: ai = a->i;
3243: aj = a->j;
3245: VecSet(v,0.0);
3246: VecGetArrayWrite(v,&x);
3247: VecGetLocalSize(v,&n);
3248: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
3249: for (i=0; i<m; i++) {
3250: ncols = ai[1] - ai[0]; ai++;
3251: for (j=0; j<ncols; j++) {
3252: atmp = PetscAbsScalar(*aa);
3253: if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
3254: aa++; aj++;
3255: }
3256: }
3257: VecRestoreArrayWrite(v,&x);
3258: MatSeqAIJRestoreArrayRead(A,&av);
3259: return(0);
3260: }
3262: PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3263: {
3264: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3265: PetscErrorCode ierr;
3266: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
3267: PetscScalar *x;
3268: const MatScalar *aa,*av;
3271: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3272: MatSeqAIJGetArrayRead(A,&av);
3273: aa = av;
3274: ai = a->i;
3275: aj = a->j;
3277: VecSet(v,0.0);
3278: VecGetArrayWrite(v,&x);
3279: VecGetLocalSize(v,&n);
3280: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
3281: for (i=0; i<m; i++) {
3282: ncols = ai[1] - ai[0]; ai++;
3283: if (ncols == A->cmap->n) { /* row is dense */
3284: x[i] = *aa; if (idx) idx[i] = 0;
3285: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
3286: x[i] = 0.0;
3287: if (idx) {
3288: for (j=0; j<ncols; j++) { /* find first implicit 0.0 in the row */
3289: if (aj[j] > j) {
3290: idx[i] = j;
3291: break;
3292: }
3293: }
3294: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3295: if (j==ncols && j < A->cmap->n) idx[i] = j;
3296: }
3297: }
3298: for (j=0; j<ncols; j++) {
3299: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
3300: aa++; aj++;
3301: }
3302: }
3303: VecRestoreArrayWrite(v,&x);
3304: MatSeqAIJRestoreArrayRead(A,&av);
3305: return(0);
3306: }
3308: PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3309: {
3310: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3311: PetscErrorCode ierr;
3312: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
3313: PetscScalar *x;
3314: const MatScalar *aa,*av;
3317: MatSeqAIJGetArrayRead(A,&av);
3318: aa = av;
3319: ai = a->i;
3320: aj = a->j;
3322: VecSet(v,0.0);
3323: VecGetArrayWrite(v,&x);
3324: VecGetLocalSize(v,&n);
3325: if (n != m) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector, %D vs. %D rows", m, n);
3326: for (i=0; i<m; i++) {
3327: ncols = ai[1] - ai[0]; ai++;
3328: if (ncols == A->cmap->n) { /* row is dense */
3329: x[i] = *aa; if (idx) idx[i] = 0;
3330: } else { /* row is sparse so already KNOW minimum is 0.0 or higher */
3331: x[i] = 0.0;
3332: if (idx) { /* find first implicit 0.0 in the row */
3333: for (j=0; j<ncols; j++) {
3334: if (aj[j] > j) {
3335: idx[i] = j;
3336: break;
3337: }
3338: }
3339: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3340: if (j==ncols && j < A->cmap->n) idx[i] = j;
3341: }
3342: }
3343: for (j=0; j<ncols; j++) {
3344: if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
3345: aa++; aj++;
3346: }
3347: }
3348: VecRestoreArrayWrite(v,&x);
3349: MatSeqAIJRestoreArrayRead(A,&av);
3350: return(0);
3351: }
3353: PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[])
3354: {
3355: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
3356: PetscErrorCode ierr;
3357: PetscInt i,j,m = A->rmap->n,ncols,n;
3358: const PetscInt *ai,*aj;
3359: PetscScalar *x;
3360: const MatScalar *aa,*av;
3363: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3364: MatSeqAIJGetArrayRead(A,&av);
3365: aa = av;
3366: ai = a->i;
3367: aj = a->j;
3369: VecSet(v,0.0);
3370: VecGetArrayWrite(v,&x);
3371: VecGetLocalSize(v,&n);
3372: if (n != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
3373: for (i=0; i<m; i++) {
3374: ncols = ai[1] - ai[0]; ai++;
3375: if (ncols == A->cmap->n) { /* row is dense */
3376: x[i] = *aa; if (idx) idx[i] = 0;
3377: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3378: x[i] = 0.0;
3379: if (idx) { /* find first implicit 0.0 in the row */
3380: for (j=0; j<ncols; j++) {
3381: if (aj[j] > j) {
3382: idx[i] = j;
3383: break;
3384: }
3385: }
3386: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3387: if (j==ncols && j < A->cmap->n) idx[i] = j;
3388: }
3389: }
3390: for (j=0; j<ncols; j++) {
3391: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
3392: aa++; aj++;
3393: }
3394: }
3395: VecRestoreArrayWrite(v,&x);
3396: MatSeqAIJRestoreArrayRead(A,&av);
3397: return(0);
3398: }
3400: PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values)
3401: {
3402: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
3403: PetscErrorCode ierr;
3404: PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j;
3405: MatScalar *diag,work[25],*v_work;
3406: const PetscReal shift = 0.0;
3407: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
3410: allowzeropivot = PetscNot(A->erroriffailure);
3411: if (a->ibdiagvalid) {
3412: if (values) *values = a->ibdiag;
3413: return(0);
3414: }
3415: MatMarkDiagonal_SeqAIJ(A);
3416: if (!a->ibdiag) {
3417: PetscMalloc1(bs2*mbs,&a->ibdiag);
3418: PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));
3419: }
3420: diag = a->ibdiag;
3421: if (values) *values = a->ibdiag;
3422: /* factor and invert each block */
3423: switch (bs) {
3424: case 1:
3425: for (i=0; i<mbs; i++) {
3426: MatGetValues(A,1,&i,1,&i,diag+i);
3427: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3428: if (allowzeropivot) {
3429: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3430: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3431: A->factorerror_zeropivot_row = i;
3432: PetscInfo3(A,"Zero pivot, row %D pivot %g tolerance %g\n",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);
3433: } 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);
3434: }
3435: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3436: }
3437: break;
3438: case 2:
3439: for (i=0; i<mbs; i++) {
3440: ij[0] = 2*i; ij[1] = 2*i + 1;
3441: MatGetValues(A,2,ij,2,ij,diag);
3442: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
3443: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3444: PetscKernel_A_gets_transpose_A_2(diag);
3445: diag += 4;
3446: }
3447: break;
3448: case 3:
3449: for (i=0; i<mbs; i++) {
3450: ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2;
3451: MatGetValues(A,3,ij,3,ij,diag);
3452: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
3453: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3454: PetscKernel_A_gets_transpose_A_3(diag);
3455: diag += 9;
3456: }
3457: break;
3458: case 4:
3459: for (i=0; i<mbs; i++) {
3460: ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3;
3461: MatGetValues(A,4,ij,4,ij,diag);
3462: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
3463: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3464: PetscKernel_A_gets_transpose_A_4(diag);
3465: diag += 16;
3466: }
3467: break;
3468: case 5:
3469: for (i=0; i<mbs; i++) {
3470: ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4;
3471: MatGetValues(A,5,ij,5,ij,diag);
3472: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
3473: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3474: PetscKernel_A_gets_transpose_A_5(diag);
3475: diag += 25;
3476: }
3477: break;
3478: case 6:
3479: for (i=0; i<mbs; i++) {
3480: 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;
3481: MatGetValues(A,6,ij,6,ij,diag);
3482: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
3483: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3484: PetscKernel_A_gets_transpose_A_6(diag);
3485: diag += 36;
3486: }
3487: break;
3488: case 7:
3489: for (i=0; i<mbs; i++) {
3490: 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;
3491: MatGetValues(A,7,ij,7,ij,diag);
3492: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
3493: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3494: PetscKernel_A_gets_transpose_A_7(diag);
3495: diag += 49;
3496: }
3497: break;
3498: default:
3499: PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);
3500: for (i=0; i<mbs; i++) {
3501: for (j=0; j<bs; j++) {
3502: IJ[j] = bs*i + j;
3503: }
3504: MatGetValues(A,bs,IJ,bs,IJ,diag);
3505: PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
3506: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3507: PetscKernel_A_gets_transpose_A_N(diag,bs);
3508: diag += bs2;
3509: }
3510: PetscFree3(v_work,v_pivots,IJ);
3511: }
3512: a->ibdiagvalid = PETSC_TRUE;
3513: return(0);
3514: }
3516: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx)
3517: {
3519: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3520: PetscScalar a;
3521: PetscInt m,n,i,j,col;
3524: if (!x->assembled) {
3525: MatGetSize(x,&m,&n);
3526: for (i=0; i<m; i++) {
3527: for (j=0; j<aij->imax[i]; j++) {
3528: PetscRandomGetValue(rctx,&a);
3529: col = (PetscInt)(n*PetscRealPart(a));
3530: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3531: }
3532: }
3533: } else {
3534: for (i=0; i<aij->nz; i++) {PetscRandomGetValue(rctx,aij->a+i);}
3535: }
3536: #if defined(PETSC_HAVE_DEVICE)
3537: if (x->offloadmask != PETSC_OFFLOAD_UNALLOCATED) x->offloadmask = PETSC_OFFLOAD_CPU;
3538: #endif
3539: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3540: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3541: return(0);
3542: }
3544: /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */
3545: PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x,PetscInt low,PetscInt high,PetscRandom rctx)
3546: {
3548: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3549: PetscScalar a;
3550: PetscInt m,n,i,j,col,nskip;
3553: nskip = high - low;
3554: MatGetSize(x,&m,&n);
3555: n -= nskip; /* shrink number of columns where nonzeros can be set */
3556: for (i=0; i<m; i++) {
3557: for (j=0; j<aij->imax[i]; j++) {
3558: PetscRandomGetValue(rctx,&a);
3559: col = (PetscInt)(n*PetscRealPart(a));
3560: if (col >= low) col += nskip; /* shift col rightward to skip the hole */
3561: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3562: }
3563: }
3564: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3565: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3566: return(0);
3567: }
3570: /* -------------------------------------------------------------------*/
3571: static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ,
3572: MatGetRow_SeqAIJ,
3573: MatRestoreRow_SeqAIJ,
3574: MatMult_SeqAIJ,
3575: /* 4*/ MatMultAdd_SeqAIJ,
3576: MatMultTranspose_SeqAIJ,
3577: MatMultTransposeAdd_SeqAIJ,
3578: NULL,
3579: NULL,
3580: NULL,
3581: /* 10*/ NULL,
3582: MatLUFactor_SeqAIJ,
3583: NULL,
3584: MatSOR_SeqAIJ,
3585: MatTranspose_SeqAIJ,
3586: /*1 5*/ MatGetInfo_SeqAIJ,
3587: MatEqual_SeqAIJ,
3588: MatGetDiagonal_SeqAIJ,
3589: MatDiagonalScale_SeqAIJ,
3590: MatNorm_SeqAIJ,
3591: /* 20*/ NULL,
3592: MatAssemblyEnd_SeqAIJ,
3593: MatSetOption_SeqAIJ,
3594: MatZeroEntries_SeqAIJ,
3595: /* 24*/ MatZeroRows_SeqAIJ,
3596: NULL,
3597: NULL,
3598: NULL,
3599: NULL,
3600: /* 29*/ MatSetUp_SeqAIJ,
3601: NULL,
3602: NULL,
3603: NULL,
3604: NULL,
3605: /* 34*/ MatDuplicate_SeqAIJ,
3606: NULL,
3607: NULL,
3608: MatILUFactor_SeqAIJ,
3609: NULL,
3610: /* 39*/ MatAXPY_SeqAIJ,
3611: MatCreateSubMatrices_SeqAIJ,
3612: MatIncreaseOverlap_SeqAIJ,
3613: MatGetValues_SeqAIJ,
3614: MatCopy_SeqAIJ,
3615: /* 44*/ MatGetRowMax_SeqAIJ,
3616: MatScale_SeqAIJ,
3617: MatShift_SeqAIJ,
3618: MatDiagonalSet_SeqAIJ,
3619: MatZeroRowsColumns_SeqAIJ,
3620: /* 49*/ MatSetRandom_SeqAIJ,
3621: MatGetRowIJ_SeqAIJ,
3622: MatRestoreRowIJ_SeqAIJ,
3623: MatGetColumnIJ_SeqAIJ,
3624: MatRestoreColumnIJ_SeqAIJ,
3625: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3626: NULL,
3627: NULL,
3628: MatPermute_SeqAIJ,
3629: NULL,
3630: /* 59*/ NULL,
3631: MatDestroy_SeqAIJ,
3632: MatView_SeqAIJ,
3633: NULL,
3634: NULL,
3635: /* 64*/ NULL,
3636: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3637: NULL,
3638: NULL,
3639: NULL,
3640: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3641: MatGetRowMinAbs_SeqAIJ,
3642: NULL,
3643: NULL,
3644: NULL,
3645: /* 74*/ NULL,
3646: MatFDColoringApply_AIJ,
3647: NULL,
3648: NULL,
3649: NULL,
3650: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3651: NULL,
3652: NULL,
3653: NULL,
3654: MatLoad_SeqAIJ,
3655: /* 84*/ MatIsSymmetric_SeqAIJ,
3656: MatIsHermitian_SeqAIJ,
3657: NULL,
3658: NULL,
3659: NULL,
3660: /* 89*/ NULL,
3661: NULL,
3662: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3663: NULL,
3664: NULL,
3665: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3666: NULL,
3667: NULL,
3668: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3669: NULL,
3670: /* 99*/ MatProductSetFromOptions_SeqAIJ,
3671: NULL,
3672: NULL,
3673: MatConjugate_SeqAIJ,
3674: NULL,
3675: /*104*/ MatSetValuesRow_SeqAIJ,
3676: MatRealPart_SeqAIJ,
3677: MatImaginaryPart_SeqAIJ,
3678: NULL,
3679: NULL,
3680: /*109*/ MatMatSolve_SeqAIJ,
3681: NULL,
3682: MatGetRowMin_SeqAIJ,
3683: NULL,
3684: MatMissingDiagonal_SeqAIJ,
3685: /*114*/ NULL,
3686: NULL,
3687: NULL,
3688: NULL,
3689: NULL,
3690: /*119*/ NULL,
3691: NULL,
3692: NULL,
3693: NULL,
3694: MatGetMultiProcBlock_SeqAIJ,
3695: /*124*/ MatFindNonzeroRows_SeqAIJ,
3696: MatGetColumnNorms_SeqAIJ,
3697: MatInvertBlockDiagonal_SeqAIJ,
3698: MatInvertVariableBlockDiagonal_SeqAIJ,
3699: NULL,
3700: /*129*/ NULL,
3701: NULL,
3702: NULL,
3703: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3704: MatTransposeColoringCreate_SeqAIJ,
3705: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3706: MatTransColoringApplyDenToSp_SeqAIJ,
3707: NULL,
3708: NULL,
3709: MatRARtNumeric_SeqAIJ_SeqAIJ,
3710: /*139*/NULL,
3711: NULL,
3712: NULL,
3713: MatFDColoringSetUp_SeqXAIJ,
3714: MatFindOffBlockDiagonalEntries_SeqAIJ,
3715: MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3716: /*145*/MatDestroySubMatrices_SeqAIJ,
3717: NULL,
3718: NULL
3719: };
3721: PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices)
3722: {
3723: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3724: PetscInt i,nz,n;
3727: nz = aij->maxnz;
3728: n = mat->rmap->n;
3729: for (i=0; i<nz; i++) {
3730: aij->j[i] = indices[i];
3731: }
3732: aij->nz = nz;
3733: for (i=0; i<n; i++) {
3734: aij->ilen[i] = aij->imax[i];
3735: }
3736: return(0);
3737: }
3739: /*
3740: * When a sparse matrix has many zero columns, we should compact them out to save the space
3741: * This happens in MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable()
3742: * */
3743: PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping)
3744: {
3745: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3746: PetscTable gid1_lid1;
3747: PetscTablePosition tpos;
3748: PetscInt gid,lid,i,ec,nz = aij->nz;
3749: PetscInt *garray,*jj = aij->j;
3750: PetscErrorCode ierr;
3755: /* use a table */
3756: PetscTableCreate(mat->rmap->n,mat->cmap->N+1,&gid1_lid1);
3757: ec = 0;
3758: for (i=0; i<nz; i++) {
3759: PetscInt data,gid1 = jj[i] + 1;
3760: PetscTableFind(gid1_lid1,gid1,&data);
3761: if (!data) {
3762: /* one based table */
3763: PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);
3764: }
3765: }
3766: /* form array of columns we need */
3767: PetscMalloc1(ec+1,&garray);
3768: PetscTableGetHeadPosition(gid1_lid1,&tpos);
3769: while (tpos) {
3770: PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);
3771: gid--;
3772: lid--;
3773: garray[lid] = gid;
3774: }
3775: PetscSortInt(ec,garray); /* sort, and rebuild */
3776: PetscTableRemoveAll(gid1_lid1);
3777: for (i=0; i<ec; i++) {
3778: PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);
3779: }
3780: /* compact out the extra columns in B */
3781: for (i=0; i<nz; i++) {
3782: PetscInt gid1 = jj[i] + 1;
3783: PetscTableFind(gid1_lid1,gid1,&lid);
3784: lid--;
3785: jj[i] = lid;
3786: }
3787: PetscLayoutDestroy(&mat->cmap);
3788: PetscTableDestroy(&gid1_lid1);
3789: PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat),ec,ec,1,&mat->cmap);
3790: ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,mat->cmap->bs,mat->cmap->n,garray,PETSC_OWN_POINTER,mapping);
3791: ISLocalToGlobalMappingSetType(*mapping,ISLOCALTOGLOBALMAPPINGHASH);
3792: return(0);
3793: }
3795: /*@
3796: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3797: in the matrix.
3799: Input Parameters:
3800: + mat - the SeqAIJ matrix
3801: - indices - the column indices
3803: Level: advanced
3805: Notes:
3806: This can be called if you have precomputed the nonzero structure of the
3807: matrix and want to provide it to the matrix object to improve the performance
3808: of the MatSetValues() operation.
3810: You MUST have set the correct numbers of nonzeros per row in the call to
3811: MatCreateSeqAIJ(), and the columns indices MUST be sorted.
3813: MUST be called before any calls to MatSetValues();
3815: The indices should start with zero, not one.
3817: @*/
3818: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices)
3819: {
3825: PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));
3826: return(0);
3827: }
3829: /* ----------------------------------------------------------------------------------------*/
3831: PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3832: {
3833: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3835: size_t nz = aij->i[mat->rmap->n];
3838: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3840: /* allocate space for values if not already there */
3841: if (!aij->saved_values) {
3842: PetscMalloc1(nz+1,&aij->saved_values);
3843: PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));
3844: }
3846: /* copy values over */
3847: PetscArraycpy(aij->saved_values,aij->a,nz);
3848: return(0);
3849: }
3851: /*@
3852: MatStoreValues - Stashes a copy of the matrix values; this allows, for
3853: example, reuse of the linear part of a Jacobian, while recomputing the
3854: nonlinear portion.
3856: Collect on Mat
3858: Input Parameters:
3859: . mat - the matrix (currently only AIJ matrices support this option)
3861: Level: advanced
3863: Common Usage, with SNESSolve():
3864: $ Create Jacobian matrix
3865: $ Set linear terms into matrix
3866: $ Apply boundary conditions to matrix, at this time matrix must have
3867: $ final nonzero structure (i.e. setting the nonlinear terms and applying
3868: $ boundary conditions again will not change the nonzero structure
3869: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3870: $ MatStoreValues(mat);
3871: $ Call SNESSetJacobian() with matrix
3872: $ In your Jacobian routine
3873: $ MatRetrieveValues(mat);
3874: $ Set nonlinear terms in matrix
3876: Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself:
3877: $ // build linear portion of Jacobian
3878: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3879: $ MatStoreValues(mat);
3880: $ loop over nonlinear iterations
3881: $ MatRetrieveValues(mat);
3882: $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3883: $ // call MatAssemblyBegin/End() on matrix
3884: $ Solve linear system with Jacobian
3885: $ endloop
3887: Notes:
3888: Matrix must already be assemblied before calling this routine
3889: Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before
3890: calling this routine.
3892: When this is called multiple times it overwrites the previous set of stored values
3893: and does not allocated additional space.
3895: .seealso: MatRetrieveValues()
3897: @*/
3898: PetscErrorCode MatStoreValues(Mat mat)
3899: {
3904: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3905: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3906: PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));
3907: return(0);
3908: }
3910: PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3911: {
3912: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3914: PetscInt nz = aij->i[mat->rmap->n];
3917: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3918: if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first");
3919: /* copy values over */
3920: PetscArraycpy(aij->a,aij->saved_values,nz);
3921: return(0);
3922: }
3924: /*@
3925: MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for
3926: example, reuse of the linear part of a Jacobian, while recomputing the
3927: nonlinear portion.
3929: Collect on Mat
3931: Input Parameters:
3932: . mat - the matrix (currently only AIJ matrices support this option)
3934: Level: advanced
3936: .seealso: MatStoreValues()
3938: @*/
3939: PetscErrorCode MatRetrieveValues(Mat mat)
3940: {
3945: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3946: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3947: PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));
3948: return(0);
3949: }
3952: /* --------------------------------------------------------------------------------*/
3953: /*@C
3954: MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format
3955: (the default parallel PETSc format). For good matrix assembly performance
3956: the user should preallocate the matrix storage by setting the parameter nz
3957: (or the array nnz). By setting these parameters accurately, performance
3958: during matrix assembly can be increased by more than a factor of 50.
3960: Collective
3962: Input Parameters:
3963: + comm - MPI communicator, set to PETSC_COMM_SELF
3964: . m - number of rows
3965: . n - number of columns
3966: . nz - number of nonzeros per row (same for all rows)
3967: - nnz - array containing the number of nonzeros in the various rows
3968: (possibly different for each row) or NULL
3970: Output Parameter:
3971: . A - the matrix
3973: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3974: MatXXXXSetPreallocation() paradigm instead of this routine directly.
3975: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3977: Notes:
3978: If nnz is given then nz is ignored
3980: The AIJ format (also called the Yale sparse matrix format or
3981: compressed row storage), is fully compatible with standard Fortran 77
3982: storage. That is, the stored row and column indices can begin at
3983: either one (as in Fortran) or zero. See the users' manual for details.
3985: Specify the preallocated storage with either nz or nnz (not both).
3986: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3987: allocation. For large problems you MUST preallocate memory or you
3988: will get TERRIBLE performance, see the users' manual chapter on matrices.
3990: By default, this format uses inodes (identical nodes) when possible, to
3991: improve numerical efficiency of matrix-vector products and solves. We
3992: search for consecutive rows with the same nonzero structure, thereby
3993: reusing matrix information to achieve increased efficiency.
3995: Options Database Keys:
3996: + -mat_no_inode - Do not use inodes
3997: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3999: Level: intermediate
4001: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays()
4003: @*/
4004: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A)
4005: {
4009: MatCreate(comm,A);
4010: MatSetSizes(*A,m,n,m,n);
4011: MatSetType(*A,MATSEQAIJ);
4012: MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);
4013: return(0);
4014: }
4016: /*@C
4017: MatSeqAIJSetPreallocation - For good matrix assembly performance
4018: the user should preallocate the matrix storage by setting the parameter nz
4019: (or the array nnz). By setting these parameters accurately, performance
4020: during matrix assembly can be increased by more than a factor of 50.
4022: Collective
4024: Input Parameters:
4025: + B - The matrix
4026: . nz - number of nonzeros per row (same for all rows)
4027: - nnz - array containing the number of nonzeros in the various rows
4028: (possibly different for each row) or NULL
4030: Notes:
4031: If nnz is given then nz is ignored
4033: The AIJ format (also called the Yale sparse matrix format or
4034: compressed row storage), is fully compatible with standard Fortran 77
4035: storage. That is, the stored row and column indices can begin at
4036: either one (as in Fortran) or zero. See the users' manual for details.
4038: Specify the preallocated storage with either nz or nnz (not both).
4039: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
4040: allocation. For large problems you MUST preallocate memory or you
4041: will get TERRIBLE performance, see the users' manual chapter on matrices.
4043: You can call MatGetInfo() to get information on how effective the preallocation was;
4044: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4045: You can also run with the option -info and look for messages with the string
4046: malloc in them to see if additional memory allocation was needed.
4048: Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix
4049: entries or columns indices
4051: By default, this format uses inodes (identical nodes) when possible, to
4052: improve numerical efficiency of matrix-vector products and solves. We
4053: search for consecutive rows with the same nonzero structure, thereby
4054: reusing matrix information to achieve increased efficiency.
4056: Options Database Keys:
4057: + -mat_no_inode - Do not use inodes
4058: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4060: Level: intermediate
4062: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo(),
4063: MatSeqAIJSetTotalPreallocation()
4065: @*/
4066: PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[])
4067: {
4073: PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));
4074: return(0);
4075: }
4077: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz)
4078: {
4079: Mat_SeqAIJ *b;
4080: PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE;
4082: PetscInt i;
4085: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
4086: if (nz == MAT_SKIP_ALLOCATION) {
4087: skipallocation = PETSC_TRUE;
4088: nz = 0;
4089: }
4090: PetscLayoutSetUp(B->rmap);
4091: PetscLayoutSetUp(B->cmap);
4093: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
4094: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz);
4095: if (PetscUnlikelyDebug(nnz)) {
4096: for (i=0; i<B->rmap->n; i++) {
4097: 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]);
4098: 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);
4099: }
4100: }
4102: B->preallocated = PETSC_TRUE;
4104: b = (Mat_SeqAIJ*)B->data;
4106: if (!skipallocation) {
4107: if (!b->imax) {
4108: PetscMalloc1(B->rmap->n,&b->imax);
4109: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
4110: }
4111: if (!b->ilen) {
4112: /* b->ilen will count nonzeros in each row so far. */
4113: PetscCalloc1(B->rmap->n,&b->ilen);
4114: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
4115: } else {
4116: PetscMemzero(b->ilen,B->rmap->n*sizeof(PetscInt));
4117: }
4118: if (!b->ipre) {
4119: PetscMalloc1(B->rmap->n,&b->ipre);
4120: PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));
4121: }
4122: if (!nnz) {
4123: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
4124: else if (nz < 0) nz = 1;
4125: nz = PetscMin(nz,B->cmap->n);
4126: for (i=0; i<B->rmap->n; i++) b->imax[i] = nz;
4127: nz = nz*B->rmap->n;
4128: } else {
4129: PetscInt64 nz64 = 0;
4130: for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz64 += nnz[i];}
4131: PetscIntCast(nz64,&nz);
4132: }
4134: /* allocate the matrix space */
4135: /* FIXME: should B's old memory be unlogged? */
4136: MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);
4137: if (B->structure_only) {
4138: PetscMalloc1(nz,&b->j);
4139: PetscMalloc1(B->rmap->n+1,&b->i);
4140: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));
4141: } else {
4142: PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);
4143: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));
4144: }
4145: b->i[0] = 0;
4146: for (i=1; i<B->rmap->n+1; i++) {
4147: b->i[i] = b->i[i-1] + b->imax[i-1];
4148: }
4149: if (B->structure_only) {
4150: b->singlemalloc = PETSC_FALSE;
4151: b->free_a = PETSC_FALSE;
4152: } else {
4153: b->singlemalloc = PETSC_TRUE;
4154: b->free_a = PETSC_TRUE;
4155: }
4156: b->free_ij = PETSC_TRUE;
4157: } else {
4158: b->free_a = PETSC_FALSE;
4159: b->free_ij = PETSC_FALSE;
4160: }
4162: if (b->ipre && nnz != b->ipre && b->imax) {
4163: /* reserve user-requested sparsity */
4164: PetscArraycpy(b->ipre,b->imax,B->rmap->n);
4165: }
4168: b->nz = 0;
4169: b->maxnz = nz;
4170: B->info.nz_unneeded = (double)b->maxnz;
4171: if (realalloc) {
4172: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
4173: }
4174: B->was_assembled = PETSC_FALSE;
4175: B->assembled = PETSC_FALSE;
4176: return(0);
4177: }
4180: PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
4181: {
4182: Mat_SeqAIJ *a;
4183: PetscInt i;
4189: /* Check local size. If zero, then return */
4190: if (!A->rmap->n) return(0);
4192: a = (Mat_SeqAIJ*)A->data;
4193: /* if no saved info, we error out */
4194: if (!a->ipre) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"No saved preallocation info \n");
4196: 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");
4198: PetscArraycpy(a->imax,a->ipre,A->rmap->n);
4199: PetscArrayzero(a->ilen,A->rmap->n);
4200: a->i[0] = 0;
4201: for (i=1; i<A->rmap->n+1; i++) {
4202: a->i[i] = a->i[i-1] + a->imax[i-1];
4203: }
4204: A->preallocated = PETSC_TRUE;
4205: a->nz = 0;
4206: a->maxnz = a->i[A->rmap->n];
4207: A->info.nz_unneeded = (double)a->maxnz;
4208: A->was_assembled = PETSC_FALSE;
4209: A->assembled = PETSC_FALSE;
4210: return(0);
4211: }
4213: /*@
4214: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format.
4216: Input Parameters:
4217: + B - the matrix
4218: . i - the indices into j for the start of each row (starts with zero)
4219: . j - the column indices for each row (starts with zero) these must be sorted for each row
4220: - v - optional values in the matrix
4222: Level: developer
4224: Notes:
4225: The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays()
4227: This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero
4228: structure will be the union of all the previous nonzero structures.
4230: Developer Notes:
4231: 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
4232: then just copies the v values directly with PetscMemcpy().
4234: This routine could also take a PetscCopyMode argument to allow sharing the values instead of always copying them.
4236: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), MATSEQAIJ, MatResetPreallocation()
4237: @*/
4238: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[])
4239: {
4245: PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
4246: return(0);
4247: }
4249: PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
4250: {
4251: PetscInt i;
4252: PetscInt m,n;
4253: PetscInt nz;
4254: PetscInt *nnz;
4258: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]);
4260: PetscLayoutSetUp(B->rmap);
4261: PetscLayoutSetUp(B->cmap);
4263: MatGetSize(B, &m, &n);
4264: PetscMalloc1(m+1, &nnz);
4265: for (i = 0; i < m; i++) {
4266: nz = Ii[i+1]- Ii[i];
4267: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz);
4268: nnz[i] = nz;
4269: }
4270: MatSeqAIJSetPreallocation(B, 0, nnz);
4271: PetscFree(nnz);
4273: for (i = 0; i < m; i++) {
4274: MatSetValues_SeqAIJ(B, 1, &i, Ii[i+1] - Ii[i], J+Ii[i], v ? v + Ii[i] : NULL, INSERT_VALUES);
4275: }
4277: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4278: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4280: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
4281: return(0);
4282: }
4284: #include <../src/mat/impls/dense/seq/dense.h>
4285: #include <petsc/private/kernels/petscaxpy.h>
4287: /*
4288: Computes (B'*A')' since computing B*A directly is untenable
4290: n p p
4291: [ ] [ ] [ ]
4292: m [ A ] * n [ B ] = m [ C ]
4293: [ ] [ ] [ ]
4295: */
4296: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C)
4297: {
4298: PetscErrorCode ierr;
4299: Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data;
4300: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data;
4301: Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data;
4302: PetscInt i,j,n,m,q,p;
4303: const PetscInt *ii,*idx;
4304: const PetscScalar *b,*a,*a_q;
4305: PetscScalar *c,*c_q;
4306: PetscInt clda = sub_c->lda;
4307: PetscInt alda = sub_a->lda;
4310: m = A->rmap->n;
4311: n = A->cmap->n;
4312: p = B->cmap->n;
4313: a = sub_a->v;
4314: b = sub_b->a;
4315: c = sub_c->v;
4316: if (clda == m) {
4317: PetscArrayzero(c,m*p);
4318: } else {
4319: for (j=0;j<p;j++)
4320: for (i=0;i<m;i++)
4321: c[j*clda + i] = 0.0;
4322: }
4323: ii = sub_b->i;
4324: idx = sub_b->j;
4325: for (i=0; i<n; i++) {
4326: q = ii[i+1] - ii[i];
4327: while (q-->0) {
4328: c_q = c + clda*(*idx);
4329: a_q = a + alda*i;
4330: PetscKernelAXPY(c_q,*b,a_q,m);
4331: idx++;
4332: b++;
4333: }
4334: }
4335: return(0);
4336: }
4338: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat C)
4339: {
4341: PetscInt m=A->rmap->n,n=B->cmap->n;
4342: PetscBool cisdense;
4345: 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);
4346: MatSetSizes(C,m,n,m,n);
4347: MatSetBlockSizesFromMats(C,A,B);
4348: PetscObjectTypeCompareAny((PetscObject)C,&cisdense,MATSEQDENSE,MATSEQDENSECUDA,"");
4349: if (!cisdense) {
4350: MatSetType(C,MATDENSE);
4351: }
4352: MatSetUp(C);
4354: C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
4355: return(0);
4356: }
4358: /* ----------------------------------------------------------------*/
4359: /*MC
4360: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
4361: based on compressed sparse row format.
4363: Options Database Keys:
4364: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
4366: Level: beginner
4368: Notes:
4369: MatSetValues() may be called for this matrix type with a NULL argument for the numerical values,
4370: in this case the values associated with the rows and columns one passes in are set to zero
4371: in the matrix
4373: MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no
4374: space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored
4376: Developer Notes:
4377: It would be nice if all matrix formats supported passing NULL in for the numerical values
4379: .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType
4380: M*/
4382: /*MC
4383: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
4385: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
4386: and MATMPIAIJ otherwise. As a result, for single process communicators,
4387: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported
4388: for communicators controlling multiple processes. It is recommended that you call both of
4389: the above preallocation routines for simplicity.
4391: Options Database Keys:
4392: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
4394: Developer Notes:
4395: Subclasses include MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
4396: enough exist.
4398: Level: beginner
4400: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
4401: M*/
4403: /*MC
4404: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
4406: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
4407: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
4408: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
4409: for communicators controlling multiple processes. It is recommended that you call both of
4410: the above preallocation routines for simplicity.
4412: Options Database Keys:
4413: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
4415: Level: beginner
4417: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
4418: M*/
4420: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*);
4421: #if defined(PETSC_HAVE_ELEMENTAL)
4422: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
4423: #endif
4424: #if defined(PETSC_HAVE_SCALAPACK)
4425: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat,MatType,MatReuse,Mat*);
4426: #endif
4427: #if defined(PETSC_HAVE_HYPRE)
4428: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A,MatType,MatReuse,Mat*);
4429: #endif
4431: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat,MatType,MatReuse,Mat*);
4432: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
4433: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
4435: /*@C
4436: MatSeqAIJGetArray - gives read/write access to the array where the data for a MATSEQAIJ matrix is stored
4438: Not Collective
4440: Input Parameter:
4441: . mat - a MATSEQAIJ matrix
4443: Output Parameter:
4444: . array - pointer to the data
4446: Level: intermediate
4448: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
4449: @*/
4450: PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array)
4451: {
4455: PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));
4456: #if defined(PETSC_HAVE_DEVICE)
4457: if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU;
4458: #endif
4459: return(0);
4460: }
4462: /*@C
4463: MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a MATSEQAIJ matrix is stored
4465: Not Collective
4467: Input Parameter:
4468: . mat - a MATSEQAIJ matrix
4470: Output Parameter:
4471: . array - pointer to the data
4473: Level: intermediate
4475: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayRead()
4476: @*/
4477: PetscErrorCode MatSeqAIJGetArrayRead(Mat A,const PetscScalar **array)
4478: {
4479: #if defined(PETSC_HAVE_DEVICE)
4480: PetscOffloadMask oval;
4481: #endif
4485: #if defined(PETSC_HAVE_DEVICE)
4486: oval = A->offloadmask;
4487: #endif
4488: MatSeqAIJGetArray(A,(PetscScalar**)array);
4489: #if defined(PETSC_HAVE_DEVICE)
4490: if (oval == PETSC_OFFLOAD_GPU || oval == PETSC_OFFLOAD_BOTH) A->offloadmask = PETSC_OFFLOAD_BOTH;
4491: #endif
4492: return(0);
4493: }
4495: /*@C
4496: MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from MatSeqAIJGetArrayRead
4498: Not Collective
4500: Input Parameter:
4501: . mat - a MATSEQAIJ matrix
4503: Output Parameter:
4504: . array - pointer to the data
4506: Level: intermediate
4508: .seealso: MatSeqAIJGetArray(), MatSeqAIJGetArrayRead()
4509: @*/
4510: PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A,const PetscScalar **array)
4511: {
4512: #if defined(PETSC_HAVE_DEVICE)
4513: PetscOffloadMask oval;
4514: #endif
4518: #if defined(PETSC_HAVE_DEVICE)
4519: oval = A->offloadmask;
4520: #endif
4521: MatSeqAIJRestoreArray(A,(PetscScalar**)array);
4522: #if defined(PETSC_HAVE_DEVICE)
4523: A->offloadmask = oval;
4524: #endif
4525: return(0);
4526: }
4528: /*@C
4529: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4531: Not Collective
4533: Input Parameter:
4534: . mat - a MATSEQAIJ matrix
4536: Output Parameter:
4537: . nz - the maximum number of nonzeros in any row
4539: Level: intermediate
4541: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
4542: @*/
4543: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz)
4544: {
4545: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
4548: *nz = aij->rmax;
4549: return(0);
4550: }
4552: /*@C
4553: MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray()
4555: Not Collective
4557: Input Parameters:
4558: + mat - a MATSEQAIJ matrix
4559: - array - pointer to the data
4561: Level: intermediate
4563: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90()
4564: @*/
4565: PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array)
4566: {
4570: PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));
4571: return(0);
4572: }
4574: #if defined(PETSC_HAVE_CUDA)
4575: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat);
4576: #endif
4577: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4578: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat);
4579: #endif
4581: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4582: {
4583: Mat_SeqAIJ *b;
4585: PetscMPIInt size;
4588: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
4589: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1");
4591: PetscNewLog(B,&b);
4593: B->data = (void*)b;
4595: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
4596: if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
4598: b->row = NULL;
4599: b->col = NULL;
4600: b->icol = NULL;
4601: b->reallocs = 0;
4602: b->ignorezeroentries = PETSC_FALSE;
4603: b->roworiented = PETSC_TRUE;
4604: b->nonew = 0;
4605: b->diag = NULL;
4606: b->solve_work = NULL;
4607: B->spptr = NULL;
4608: b->saved_values = NULL;
4609: b->idiag = NULL;
4610: b->mdiag = NULL;
4611: b->ssor_work = NULL;
4612: b->omega = 1.0;
4613: b->fshift = 0.0;
4614: b->idiagvalid = PETSC_FALSE;
4615: b->ibdiagvalid = PETSC_FALSE;
4616: b->keepnonzeropattern = PETSC_FALSE;
4618: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4619: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);
4620: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);
4622: #if defined(PETSC_HAVE_MATLAB_ENGINE)
4623: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);
4624: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);
4625: #endif
4627: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);
4628: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);
4629: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);
4630: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);
4631: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);
4632: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);
4633: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijsell_C",MatConvert_SeqAIJ_SeqAIJSELL);
4634: #if defined(PETSC_HAVE_MKL_SPARSE)
4635: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijmkl_C",MatConvert_SeqAIJ_SeqAIJMKL);
4636: #endif
4637: #if defined(PETSC_HAVE_CUDA)
4638: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcusparse_C",MatConvert_SeqAIJ_SeqAIJCUSPARSE);
4639: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqaijcusparse_seqaij_C",MatProductSetFromOptions_SeqAIJ);
4640: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqaij_seqaijcusparse_C",MatProductSetFromOptions_SeqAIJ);
4641: #endif
4642: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4643: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijkokkos_C",MatConvert_SeqAIJ_SeqAIJKokkos);
4644: #endif
4645: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);
4646: #if defined(PETSC_HAVE_ELEMENTAL)
4647: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);
4648: #endif
4649: #if defined(PETSC_HAVE_SCALAPACK)
4650: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_scalapack_C",MatConvert_AIJ_ScaLAPACK);
4651: #endif
4652: #if defined(PETSC_HAVE_HYPRE)
4653: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_hypre_C",MatConvert_AIJ_HYPRE);
4654: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_transpose_seqaij_seqaij_C",MatProductSetFromOptions_Transpose_AIJ_AIJ);
4655: #endif
4656: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);
4657: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsell_C",MatConvert_SeqAIJ_SeqSELL);
4658: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_is_C",MatConvert_XAIJ_IS);
4659: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);
4660: PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);
4661: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);
4662: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_SeqAIJ);
4663: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);
4664: PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);
4665: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_is_seqaij_C",MatProductSetFromOptions_IS_XAIJ);
4666: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqdense_seqaij_C",MatProductSetFromOptions_SeqDense_SeqAIJ);
4667: PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqaij_seqaij_C",MatProductSetFromOptions_SeqAIJ);
4668: MatCreate_SeqAIJ_Inode(B);
4669: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4670: MatSeqAIJSetTypeFromOptions(B); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4671: return(0);
4672: }
4674: /*
4675: Given a matrix generated with MatGetFactor() duplicates all the information in A into B
4676: */
4677: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace)
4678: {
4679: Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data,*a = (Mat_SeqAIJ*)A->data;
4681: PetscInt m = A->rmap->n,i;
4684: if (!A->assembled && cpvalues!=MAT_DO_NOT_COPY_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot duplicate unassembled matrix");
4686: C->factortype = A->factortype;
4687: c->row = NULL;
4688: c->col = NULL;
4689: c->icol = NULL;
4690: c->reallocs = 0;
4692: C->assembled = PETSC_TRUE;
4694: PetscLayoutReference(A->rmap,&C->rmap);
4695: PetscLayoutReference(A->cmap,&C->cmap);
4697: PetscMalloc1(m,&c->imax);
4698: PetscMemcpy(c->imax,a->imax,m*sizeof(PetscInt));
4699: PetscMalloc1(m,&c->ilen);
4700: PetscMemcpy(c->ilen,a->ilen,m*sizeof(PetscInt));
4701: PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));
4703: /* allocate the matrix space */
4704: if (mallocmatspace) {
4705: PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);
4706: PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));
4708: c->singlemalloc = PETSC_TRUE;
4710: PetscArraycpy(c->i,a->i,m+1);
4711: if (m > 0) {
4712: PetscArraycpy(c->j,a->j,a->i[m]);
4713: if (cpvalues == MAT_COPY_VALUES) {
4714: const PetscScalar *aa;
4716: MatSeqAIJGetArrayRead(A,&aa);
4717: PetscArraycpy(c->a,aa,a->i[m]);
4718: MatSeqAIJGetArrayRead(A,&aa);
4719: } else {
4720: PetscArrayzero(c->a,a->i[m]);
4721: }
4722: }
4723: }
4725: c->ignorezeroentries = a->ignorezeroentries;
4726: c->roworiented = a->roworiented;
4727: c->nonew = a->nonew;
4728: if (a->diag) {
4729: PetscMalloc1(m+1,&c->diag);
4730: PetscMemcpy(c->diag,a->diag,m*sizeof(PetscInt));
4731: PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));
4732: } else c->diag = NULL;
4734: c->solve_work = NULL;
4735: c->saved_values = NULL;
4736: c->idiag = NULL;
4737: c->ssor_work = NULL;
4738: c->keepnonzeropattern = a->keepnonzeropattern;
4739: c->free_a = PETSC_TRUE;
4740: c->free_ij = PETSC_TRUE;
4742: c->rmax = a->rmax;
4743: c->nz = a->nz;
4744: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4745: C->preallocated = PETSC_TRUE;
4747: c->compressedrow.use = a->compressedrow.use;
4748: c->compressedrow.nrows = a->compressedrow.nrows;
4749: if (a->compressedrow.use) {
4750: i = a->compressedrow.nrows;
4751: PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);
4752: PetscArraycpy(c->compressedrow.i,a->compressedrow.i,i+1);
4753: PetscArraycpy(c->compressedrow.rindex,a->compressedrow.rindex,i);
4754: } else {
4755: c->compressedrow.use = PETSC_FALSE;
4756: c->compressedrow.i = NULL;
4757: c->compressedrow.rindex = NULL;
4758: }
4759: c->nonzerorowcnt = a->nonzerorowcnt;
4760: C->nonzerostate = A->nonzerostate;
4762: MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);
4763: PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);
4764: return(0);
4765: }
4767: PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B)
4768: {
4772: MatCreate(PetscObjectComm((PetscObject)A),B);
4773: MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);
4774: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) {
4775: MatSetBlockSizesFromMats(*B,A,A);
4776: }
4777: MatSetType(*B,((PetscObject)A)->type_name);
4778: MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);
4779: return(0);
4780: }
4782: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4783: {
4784: PetscBool isbinary, ishdf5;
4790: /* force binary viewer to load .info file if it has not yet done so */
4791: PetscViewerSetUp(viewer);
4792: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
4793: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);
4794: if (isbinary) {
4795: MatLoad_SeqAIJ_Binary(newMat,viewer);
4796: } else if (ishdf5) {
4797: #if defined(PETSC_HAVE_HDF5)
4798: MatLoad_AIJ_HDF5(newMat,viewer);
4799: #else
4800: SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
4801: #endif
4802: } else {
4803: 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);
4804: }
4805: return(0);
4806: }
4808: PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
4809: {
4810: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->data;
4812: PetscInt header[4],*rowlens,M,N,nz,sum,rows,cols,i;
4815: PetscViewerSetUp(viewer);
4817: /* read in matrix header */
4818: PetscViewerBinaryRead(viewer,header,4,NULL,PETSC_INT);
4819: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Not a matrix object in file");
4820: M = header[1]; N = header[2]; nz = header[3];
4821: if (M < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix row size (%D) in file is negative",M);
4822: if (N < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix column size (%D) in file is negative",N);
4823: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk, cannot load as SeqAIJ");
4825: /* set block sizes from the viewer's .info file */
4826: MatLoad_Binary_BlockSizes(mat,viewer);
4827: /* set local and global sizes if not set already */
4828: if (mat->rmap->n < 0) mat->rmap->n = M;
4829: if (mat->cmap->n < 0) mat->cmap->n = N;
4830: if (mat->rmap->N < 0) mat->rmap->N = M;
4831: if (mat->cmap->N < 0) mat->cmap->N = N;
4832: PetscLayoutSetUp(mat->rmap);
4833: PetscLayoutSetUp(mat->cmap);
4835: /* check if the matrix sizes are correct */
4836: MatGetSize(mat,&rows,&cols);
4837: 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);
4839: /* read in row lengths */
4840: PetscMalloc1(M,&rowlens);
4841: PetscViewerBinaryRead(viewer,rowlens,M,NULL,PETSC_INT);
4842: /* check if sum(rowlens) is same as nz */
4843: sum = 0; for (i=0; i<M; i++) sum += rowlens[i];
4844: 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);
4845: /* preallocate and check sizes */
4846: MatSeqAIJSetPreallocation_SeqAIJ(mat,0,rowlens);
4847: MatGetSize(mat,&rows,&cols);
4848: 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);
4849: /* store row lengths */
4850: PetscArraycpy(a->ilen,rowlens,M);
4851: PetscFree(rowlens);
4853: /* fill in "i" row pointers */
4854: a->i[0] = 0; for (i=0; i<M; i++) a->i[i+1] = a->i[i] + a->ilen[i];
4855: /* read in "j" column indices */
4856: PetscViewerBinaryRead(viewer,a->j,nz,NULL,PETSC_INT);
4857: /* read in "a" nonzero values */
4858: PetscViewerBinaryRead(viewer,a->a,nz,NULL,PETSC_SCALAR);
4860: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
4861: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
4862: return(0);
4863: }
4865: PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg)
4866: {
4867: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
4869: #if defined(PETSC_USE_COMPLEX)
4870: PetscInt k;
4871: #endif
4874: /* If the matrix dimensions are not equal,or no of nonzeros */
4875: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) {
4876: *flg = PETSC_FALSE;
4877: return(0);
4878: }
4880: /* if the a->i are the same */
4881: PetscArraycmp(a->i,b->i,A->rmap->n+1,flg);
4882: if (!*flg) return(0);
4884: /* if a->j are the same */
4885: PetscArraycmp(a->j,b->j,a->nz,flg);
4886: if (!*flg) return(0);
4888: /* if a->a are the same */
4889: #if defined(PETSC_USE_COMPLEX)
4890: for (k=0; k<a->nz; k++) {
4891: if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) {
4892: *flg = PETSC_FALSE;
4893: return(0);
4894: }
4895: }
4896: #else
4897: PetscArraycmp(a->a,b->a,a->nz,flg);
4898: #endif
4899: return(0);
4900: }
4902: /*@
4903: MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format)
4904: provided by the user.
4906: Collective
4908: Input Parameters:
4909: + comm - must be an MPI communicator of size 1
4910: . m - number of rows
4911: . n - number of columns
4912: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4913: . j - column indices
4914: - a - matrix values
4916: Output Parameter:
4917: . mat - the matrix
4919: Level: intermediate
4921: Notes:
4922: The i, j, and a arrays are not copied by this routine, the user must free these arrays
4923: once the matrix is destroyed and not before
4925: You cannot set new nonzero locations into this matrix, that will generate an error.
4927: The i and j indices are 0 based
4929: The format which is used for the sparse matrix input, is equivalent to a
4930: row-major ordering.. i.e for the following matrix, the input data expected is
4931: as shown
4933: $ 1 0 0
4934: $ 2 0 3
4935: $ 4 5 6
4936: $
4937: $ i = {0,1,3,6} [size = nrow+1 = 3+1]
4938: $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
4939: $ v = {1,2,3,4,5,6} [size = 6]
4942: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4944: @*/
4945: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat)
4946: {
4948: PetscInt ii;
4949: Mat_SeqAIJ *aij;
4950: PetscInt jj;
4953: if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4954: MatCreate(comm,mat);
4955: MatSetSizes(*mat,m,n,m,n);
4956: /* MatSetBlockSizes(*mat,,); */
4957: MatSetType(*mat,MATSEQAIJ);
4958: MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,NULL);
4959: aij = (Mat_SeqAIJ*)(*mat)->data;
4960: PetscMalloc1(m,&aij->imax);
4961: PetscMalloc1(m,&aij->ilen);
4963: aij->i = i;
4964: aij->j = j;
4965: aij->a = a;
4966: aij->singlemalloc = PETSC_FALSE;
4967: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
4968: aij->free_a = PETSC_FALSE;
4969: aij->free_ij = PETSC_FALSE;
4971: for (ii=0; ii<m; ii++) {
4972: aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii];
4973: if (PetscDefined(USE_DEBUG)) {
4974: 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]);
4975: for (jj=i[ii]+1; jj<i[ii+1]; jj++) {
4976: 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);
4977: 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);
4978: }
4979: }
4980: }
4981: if (PetscDefined(USE_DEBUG)) {
4982: for (ii=0; ii<aij->i[m]; ii++) {
4983: if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]);
4984: 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]);
4985: }
4986: }
4988: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4989: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4990: return(0);
4991: }
4992: /*@C
4993: MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format)
4994: provided by the user.
4996: Collective
4998: Input Parameters:
4999: + comm - must be an MPI communicator of size 1
5000: . m - number of rows
5001: . n - number of columns
5002: . i - row indices
5003: . j - column indices
5004: . a - matrix values
5005: . nz - number of nonzeros
5006: - idx - 0 or 1 based
5008: Output Parameter:
5009: . mat - the matrix
5011: Level: intermediate
5013: Notes:
5014: The i and j indices are 0 based
5016: The format which is used for the sparse matrix input, is equivalent to a
5017: row-major ordering.. i.e for the following matrix, the input data expected is
5018: as shown:
5020: 1 0 0
5021: 2 0 3
5022: 4 5 6
5024: i = {0,1,1,2,2,2}
5025: j = {0,0,2,0,1,2}
5026: v = {1,2,3,4,5,6}
5029: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
5031: @*/
5032: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat,PetscInt nz,PetscBool idx)
5033: {
5035: PetscInt ii, *nnz, one = 1,row,col;
5039: PetscCalloc1(m,&nnz);
5040: for (ii = 0; ii < nz; ii++) {
5041: nnz[i[ii] - !!idx] += 1;
5042: }
5043: MatCreate(comm,mat);
5044: MatSetSizes(*mat,m,n,m,n);
5045: MatSetType(*mat,MATSEQAIJ);
5046: MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);
5047: for (ii = 0; ii < nz; ii++) {
5048: if (idx) {
5049: row = i[ii] - 1;
5050: col = j[ii] - 1;
5051: } else {
5052: row = i[ii];
5053: col = j[ii];
5054: }
5055: MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);
5056: }
5057: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
5058: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
5059: PetscFree(nnz);
5060: return(0);
5061: }
5063: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
5064: {
5065: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data;
5069: a->idiagvalid = PETSC_FALSE;
5070: a->ibdiagvalid = PETSC_FALSE;
5072: MatSeqAIJInvalidateDiagonal_Inode(A);
5073: return(0);
5074: }
5076: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
5077: {
5079: PetscMPIInt size;
5082: MPI_Comm_size(comm,&size);
5083: if (size == 1) {
5084: if (scall == MAT_INITIAL_MATRIX) {
5085: MatDuplicate(inmat,MAT_COPY_VALUES,outmat);
5086: } else {
5087: MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);
5088: }
5089: } else {
5090: MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);
5091: }
5092: return(0);
5093: }
5095: /*
5096: Permute A into C's *local* index space using rowemb,colemb.
5097: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
5098: of [0,m), colemb is in [0,n).
5099: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
5100: */
5101: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B)
5102: {
5103: /* If making this function public, change the error returned in this function away from _PLIB. */
5105: Mat_SeqAIJ *Baij;
5106: PetscBool seqaij;
5107: PetscInt m,n,*nz,i,j,count;
5108: PetscScalar v;
5109: const PetscInt *rowindices,*colindices;
5112: if (!B) return(0);
5113: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
5114: PetscObjectBaseTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);
5115: if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type");
5116: if (rowemb) {
5117: ISGetLocalSize(rowemb,&m);
5118: 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);
5119: } else {
5120: if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix");
5121: }
5122: if (colemb) {
5123: ISGetLocalSize(colemb,&n);
5124: 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);
5125: } else {
5126: if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix");
5127: }
5129: Baij = (Mat_SeqAIJ*)(B->data);
5130: if (pattern == DIFFERENT_NONZERO_PATTERN) {
5131: PetscMalloc1(B->rmap->n,&nz);
5132: for (i=0; i<B->rmap->n; i++) {
5133: nz[i] = Baij->i[i+1] - Baij->i[i];
5134: }
5135: MatSeqAIJSetPreallocation(C,0,nz);
5136: PetscFree(nz);
5137: }
5138: if (pattern == SUBSET_NONZERO_PATTERN) {
5139: MatZeroEntries(C);
5140: }
5141: count = 0;
5142: rowindices = NULL;
5143: colindices = NULL;
5144: if (rowemb) {
5145: ISGetIndices(rowemb,&rowindices);
5146: }
5147: if (colemb) {
5148: ISGetIndices(colemb,&colindices);
5149: }
5150: for (i=0; i<B->rmap->n; i++) {
5151: PetscInt row;
5152: row = i;
5153: if (rowindices) row = rowindices[i];
5154: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
5155: PetscInt col;
5156: col = Baij->j[count];
5157: if (colindices) col = colindices[col];
5158: v = Baij->a[count];
5159: MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);
5160: ++count;
5161: }
5162: }
5163: /* FIXME: set C's nonzerostate correctly. */
5164: /* Assembly for C is necessary. */
5165: C->preallocated = PETSC_TRUE;
5166: C->assembled = PETSC_TRUE;
5167: C->was_assembled = PETSC_FALSE;
5168: return(0);
5169: }
5171: PetscFunctionList MatSeqAIJList = NULL;
5173: /*@C
5174: MatSeqAIJSetType - Converts a MATSEQAIJ matrix to a subtype
5176: Collective on Mat
5178: Input Parameters:
5179: + mat - the matrix object
5180: - matype - matrix type
5182: Options Database Key:
5183: . -mat_seqai_type <method> - for example seqaijcrl
5186: Level: intermediate
5188: .seealso: PCSetType(), VecSetType(), MatCreate(), MatType, Mat
5189: @*/
5190: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
5191: {
5192: PetscErrorCode ierr,(*r)(Mat,MatType,MatReuse,Mat*);
5193: PetscBool sametype;
5197: PetscObjectTypeCompare((PetscObject)mat,matype,&sametype);
5198: if (sametype) return(0);
5200: PetscFunctionListFind(MatSeqAIJList,matype,&r);
5201: if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown Mat type given: %s",matype);
5202: (*r)(mat,matype,MAT_INPLACE_MATRIX,&mat);
5203: return(0);
5204: }
5207: /*@C
5208: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential AIJ matrices
5210: Not Collective
5212: Input Parameters:
5213: + name - name of a new user-defined matrix type, for example MATSEQAIJCRL
5214: - function - routine to convert to subtype
5216: Notes:
5217: MatSeqAIJRegister() may be called multiple times to add several user-defined solvers.
5220: Then, your matrix can be chosen with the procedural interface at runtime via the option
5221: $ -mat_seqaij_type my_mat
5223: Level: advanced
5225: .seealso: MatSeqAIJRegisterAll()
5228: Level: advanced
5229: @*/
5230: PetscErrorCode MatSeqAIJRegister(const char sname[],PetscErrorCode (*function)(Mat,MatType,MatReuse,Mat *))
5231: {
5235: MatInitializePackage();
5236: PetscFunctionListAdd(&MatSeqAIJList,sname,function);
5237: return(0);
5238: }
5240: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
5242: /*@C
5243: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of SeqAIJ
5245: Not Collective
5247: Level: advanced
5249: Developers Note: CUSPARSE does not yet support the MatConvert_SeqAIJ..() paradigm and thus cannot be registered here
5251: .seealso: MatRegisterAll(), MatSeqAIJRegister()
5252: @*/
5253: PetscErrorCode MatSeqAIJRegisterAll(void)
5254: {
5258: if (MatSeqAIJRegisterAllCalled) return(0);
5259: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
5261: MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);
5262: MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);
5263: MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL);
5264: #if defined(PETSC_HAVE_MKL_SPARSE)
5265: MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);
5266: #endif
5267: #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA)
5268: MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);
5269: #endif
5270: return(0);
5271: }
5273: /*
5274: Special version for direct calls from Fortran
5275: */
5276: #include <petsc/private/fortranimpl.h>
5277: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5278: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
5279: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5280: #define matsetvaluesseqaij_ matsetvaluesseqaij
5281: #endif
5283: /* Change these macros so can be used in void function */
5284: #undef CHKERRQ
5285: #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr)
5286: #undef SETERRQ2
5287: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
5288: #undef SETERRQ3
5289: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
5291: PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[],InsertMode *isis, PetscErrorCode *_ierr)
5292: {
5293: Mat A = *AA;
5294: PetscInt m = *mm, n = *nn;
5295: InsertMode is = *isis;
5296: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
5297: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
5298: PetscInt *imax,*ai,*ailen;
5300: PetscInt *aj,nonew = a->nonew,lastcol = -1;
5301: MatScalar *ap,value,*aa;
5302: PetscBool ignorezeroentries = a->ignorezeroentries;
5303: PetscBool roworiented = a->roworiented;
5306: MatCheckPreallocated(A,1);
5307: imax = a->imax;
5308: ai = a->i;
5309: ailen = a->ilen;
5310: aj = a->j;
5311: aa = a->a;
5313: for (k=0; k<m; k++) { /* loop over added rows */
5314: row = im[k];
5315: if (row < 0) continue;
5316: if (PetscUnlikelyDebug(row >= A->rmap->n)) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
5317: rp = aj + ai[row]; ap = aa + ai[row];
5318: rmax = imax[row]; nrow = ailen[row];
5319: low = 0;
5320: high = nrow;
5321: for (l=0; l<n; l++) { /* loop over added columns */
5322: if (in[l] < 0) continue;
5323: if (PetscUnlikelyDebug(in[l] >= A->cmap->n)) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
5324: col = in[l];
5325: if (roworiented) value = v[l + k*n];
5326: else value = v[k + l*m];
5328: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
5330: if (col <= lastcol) low = 0;
5331: else high = nrow;
5332: lastcol = col;
5333: while (high-low > 5) {
5334: t = (low+high)/2;
5335: if (rp[t] > col) high = t;
5336: else low = t;
5337: }
5338: for (i=low; i<high; i++) {
5339: if (rp[i] > col) break;
5340: if (rp[i] == col) {
5341: if (is == ADD_VALUES) ap[i] += value;
5342: else ap[i] = value;
5343: goto noinsert;
5344: }
5345: }
5346: if (value == 0.0 && ignorezeroentries) goto noinsert;
5347: if (nonew == 1) goto noinsert;
5348: if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix");
5349: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
5350: N = nrow++ - 1; a->nz++; high++;
5351: /* shift up all the later entries in this row */
5352: for (ii=N; ii>=i; ii--) {
5353: rp[ii+1] = rp[ii];
5354: ap[ii+1] = ap[ii];
5355: }
5356: rp[i] = col;
5357: ap[i] = value;
5358: A->nonzerostate++;
5359: noinsert:;
5360: low = i + 1;
5361: }
5362: ailen[row] = nrow;
5363: }
5364: PetscFunctionReturnVoid();
5365: }