Actual source code: aij.c
petsc-3.7.3 2016-08-01
2: /*
3: Defines the basic matrix operations for the AIJ (compressed row)
4: matrix storage format.
5: */
8: #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/
9: #include <petscblaslapack.h>
10: #include <petscbt.h>
11: #include <petsc/private/kernels/blocktranspose.h>
15: PetscErrorCode MatGetColumnNorms_SeqAIJ(Mat A,NormType type,PetscReal *norms)
16: {
18: PetscInt i,m,n;
19: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
22: MatGetSize(A,&m,&n);
23: PetscMemzero(norms,n*sizeof(PetscReal));
24: if (type == NORM_2) {
25: for (i=0; i<aij->i[m]; i++) {
26: norms[aij->j[i]] += PetscAbsScalar(aij->a[i]*aij->a[i]);
27: }
28: } else if (type == NORM_1) {
29: for (i=0; i<aij->i[m]; i++) {
30: norms[aij->j[i]] += PetscAbsScalar(aij->a[i]);
31: }
32: } else if (type == NORM_INFINITY) {
33: for (i=0; i<aij->i[m]; i++) {
34: norms[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]),norms[aij->j[i]]);
35: }
36: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown NormType");
38: if (type == NORM_2) {
39: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
40: }
41: return(0);
42: }
46: PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A,IS *is)
47: {
48: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
49: PetscInt i,m=A->rmap->n,cnt = 0, bs = A->rmap->bs;
50: const PetscInt *jj = a->j,*ii = a->i;
51: PetscInt *rows;
52: PetscErrorCode ierr;
55: for (i=0; i<m; i++) {
56: if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) {
57: cnt++;
58: }
59: }
60: PetscMalloc1(cnt,&rows);
61: cnt = 0;
62: for (i=0; i<m; i++) {
63: if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) {
64: rows[cnt] = i;
65: cnt++;
66: }
67: }
68: ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,is);
69: return(0);
70: }
74: PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A,PetscInt *nrows,PetscInt **zrows)
75: {
76: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
77: const MatScalar *aa = a->a;
78: PetscInt i,m=A->rmap->n,cnt = 0;
79: const PetscInt *jj = a->j,*diag;
80: PetscInt *rows;
81: PetscErrorCode ierr;
84: MatMarkDiagonal_SeqAIJ(A);
85: diag = a->diag;
86: for (i=0; i<m; i++) {
87: if ((jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) {
88: cnt++;
89: }
90: }
91: PetscMalloc1(cnt,&rows);
92: cnt = 0;
93: for (i=0; i<m; i++) {
94: if ((jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) {
95: rows[cnt++] = i;
96: }
97: }
98: *nrows = cnt;
99: *zrows = rows;
100: return(0);
101: }
105: PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A,IS *zrows)
106: {
107: PetscInt nrows,*rows;
111: *zrows = NULL;
112: MatFindZeroDiagonals_SeqAIJ_Private(A,&nrows,&rows);
113: ISCreateGeneral(PetscObjectComm((PetscObject)A),nrows,rows,PETSC_OWN_POINTER,zrows);
114: return(0);
115: }
119: PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A,IS *keptrows)
120: {
121: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
122: const MatScalar *aa;
123: PetscInt m=A->rmap->n,cnt = 0;
124: const PetscInt *ii;
125: PetscInt n,i,j,*rows;
126: PetscErrorCode ierr;
129: *keptrows = 0;
130: ii = a->i;
131: for (i=0; i<m; i++) {
132: n = ii[i+1] - ii[i];
133: if (!n) {
134: cnt++;
135: goto ok1;
136: }
137: aa = a->a + ii[i];
138: for (j=0; j<n; j++) {
139: if (aa[j] != 0.0) goto ok1;
140: }
141: cnt++;
142: ok1:;
143: }
144: if (!cnt) return(0);
145: PetscMalloc1(A->rmap->n-cnt,&rows);
146: cnt = 0;
147: for (i=0; i<m; i++) {
148: n = ii[i+1] - ii[i];
149: if (!n) continue;
150: aa = a->a + ii[i];
151: for (j=0; j<n; j++) {
152: if (aa[j] != 0.0) {
153: rows[cnt++] = i;
154: break;
155: }
156: }
157: }
158: ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,keptrows);
159: return(0);
160: }
164: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y,Vec D,InsertMode is)
165: {
166: PetscErrorCode ierr;
167: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) Y->data;
168: PetscInt i,m = Y->rmap->n;
169: const PetscInt *diag;
170: MatScalar *aa = aij->a;
171: const PetscScalar *v;
172: PetscBool missing;
175: if (Y->assembled) {
176: MatMissingDiagonal_SeqAIJ(Y,&missing,NULL);
177: if (!missing) {
178: diag = aij->diag;
179: VecGetArrayRead(D,&v);
180: if (is == INSERT_VALUES) {
181: for (i=0; i<m; i++) {
182: aa[diag[i]] = v[i];
183: }
184: } else {
185: for (i=0; i<m; i++) {
186: aa[diag[i]] += v[i];
187: }
188: }
189: VecRestoreArrayRead(D,&v);
190: return(0);
191: }
192: MatSeqAIJInvalidateDiagonal(Y);
193: }
194: MatDiagonalSet_Default(Y,D,is);
195: return(0);
196: }
200: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *m,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
201: {
202: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
204: PetscInt i,ishift;
207: *m = A->rmap->n;
208: if (!ia) return(0);
209: ishift = 0;
210: if (symmetric && !A->structurally_symmetric) {
211: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,ishift,oshift,(PetscInt**)ia,(PetscInt**)ja);
212: } else if (oshift == 1) {
213: PetscInt *tia;
214: PetscInt nz = a->i[A->rmap->n];
215: /* malloc space and add 1 to i and j indices */
216: PetscMalloc1(A->rmap->n+1,&tia);
217: for (i=0; i<A->rmap->n+1; i++) tia[i] = a->i[i] + 1;
218: *ia = tia;
219: if (ja) {
220: PetscInt *tja;
221: PetscMalloc1(nz+1,&tja);
222: for (i=0; i<nz; i++) tja[i] = a->j[i] + 1;
223: *ja = tja;
224: }
225: } else {
226: *ia = a->i;
227: if (ja) *ja = a->j;
228: }
229: return(0);
230: }
234: PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
235: {
239: if (!ia) return(0);
240: if ((symmetric && !A->structurally_symmetric) || oshift == 1) {
241: PetscFree(*ia);
242: if (ja) {PetscFree(*ja);}
243: }
244: return(0);
245: }
249: PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
250: {
251: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
253: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
254: PetscInt nz = a->i[m],row,*jj,mr,col;
257: *nn = n;
258: if (!ia) return(0);
259: if (symmetric) {
260: MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,0,oshift,(PetscInt**)ia,(PetscInt**)ja);
261: } else {
262: PetscCalloc1(n+1,&collengths);
263: PetscMalloc1(n+1,&cia);
264: PetscMalloc1(nz+1,&cja);
265: jj = a->j;
266: for (i=0; i<nz; i++) {
267: collengths[jj[i]]++;
268: }
269: cia[0] = oshift;
270: for (i=0; i<n; i++) {
271: cia[i+1] = cia[i] + collengths[i];
272: }
273: PetscMemzero(collengths,n*sizeof(PetscInt));
274: jj = a->j;
275: for (row=0; row<m; row++) {
276: mr = a->i[row+1] - a->i[row];
277: for (i=0; i<mr; i++) {
278: col = *jj++;
280: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
281: }
282: }
283: PetscFree(collengths);
284: *ia = cia; *ja = cja;
285: }
286: return(0);
287: }
291: PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done)
292: {
296: if (!ia) return(0);
298: PetscFree(*ia);
299: PetscFree(*ja);
300: return(0);
301: }
303: /*
304: MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from
305: MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output
306: spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ()
307: */
310: PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
311: {
312: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
314: PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n;
315: PetscInt nz = a->i[m],row,*jj,mr,col;
316: PetscInt *cspidx;
319: *nn = n;
320: if (!ia) return(0);
322: PetscCalloc1(n+1,&collengths);
323: PetscMalloc1(n+1,&cia);
324: PetscMalloc1(nz+1,&cja);
325: PetscMalloc1(nz+1,&cspidx);
326: jj = a->j;
327: for (i=0; i<nz; i++) {
328: collengths[jj[i]]++;
329: }
330: cia[0] = oshift;
331: for (i=0; i<n; i++) {
332: cia[i+1] = cia[i] + collengths[i];
333: }
334: PetscMemzero(collengths,n*sizeof(PetscInt));
335: jj = a->j;
336: for (row=0; row<m; row++) {
337: mr = a->i[row+1] - a->i[row];
338: for (i=0; i<mr; i++) {
339: col = *jj++;
340: cspidx[cia[col] + collengths[col] - oshift] = a->i[row] + i; /* index of a->j */
341: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
342: }
343: }
344: PetscFree(collengths);
345: *ia = cia; *ja = cja;
346: *spidx = cspidx;
347: return(0);
348: }
352: PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done)
353: {
357: MatRestoreColumnIJ_SeqAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);
358: PetscFree(*spidx);
359: return(0);
360: }
364: PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A,PetscInt row,const PetscScalar v[])
365: {
366: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
367: PetscInt *ai = a->i;
371: PetscMemcpy(a->a+ai[row],v,(ai[row+1]-ai[row])*sizeof(PetscScalar));
372: return(0);
373: }
375: /*
376: MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions
378: - a single row of values is set with each call
379: - no row or column indices are negative or (in error) larger than the number of rows or columns
380: - the values are always added to the matrix, not set
381: - no new locations are introduced in the nonzero structure of the matrix
383: This does NOT assume the global column indices are sorted
385: */
387: #include <petsc/private/isimpl.h>
390: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
391: {
392: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
393: PetscInt low,high,t,row,nrow,i,col,l;
394: const PetscInt *rp,*ai = a->i,*ailen = a->ilen,*aj = a->j;
395: PetscInt lastcol = -1;
396: MatScalar *ap,value,*aa = a->a;
397: const PetscInt *ridx = A->rmap->mapping->indices,*cidx = A->cmap->mapping->indices;
399: row = ridx[im[0]];
400: rp = aj + ai[row];
401: ap = aa + ai[row];
402: nrow = ailen[row];
403: low = 0;
404: high = nrow;
405: for (l=0; l<n; l++) { /* loop over added columns */
406: col = cidx[in[l]];
407: value = v[l];
409: if (col <= lastcol) low = 0;
410: else high = nrow;
411: lastcol = col;
412: while (high-low > 5) {
413: t = (low+high)/2;
414: if (rp[t] > col) high = t;
415: else low = t;
416: }
417: for (i=low; i<high; i++) {
418: if (rp[i] == col) {
419: ap[i] += value;
420: low = i + 1;
421: break;
422: }
423: }
424: }
425: return 0;
426: }
430: PetscErrorCode MatSetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
431: {
432: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
433: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
434: PetscInt *imax = a->imax,*ai = a->i,*ailen = a->ilen;
436: PetscInt *aj = a->j,nonew = a->nonew,lastcol = -1;
437: MatScalar *ap,value,*aa = a->a;
438: PetscBool ignorezeroentries = a->ignorezeroentries;
439: PetscBool roworiented = a->roworiented;
442: for (k=0; k<m; k++) { /* loop over added rows */
443: row = im[k];
444: if (row < 0) continue;
445: #if defined(PETSC_USE_DEBUG)
446: 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);
447: #endif
448: rp = aj + ai[row]; ap = aa + ai[row];
449: rmax = imax[row]; nrow = ailen[row];
450: low = 0;
451: high = nrow;
452: for (l=0; l<n; l++) { /* loop over added columns */
453: if (in[l] < 0) continue;
454: #if defined(PETSC_USE_DEBUG)
455: 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);
456: #endif
457: col = in[l];
458: if (roworiented) {
459: value = v[l + k*n];
460: } else {
461: value = v[k + l*m];
462: }
463: if ((value == 0.0 && ignorezeroentries) && (is == ADD_VALUES)) continue;
465: if (col <= lastcol) low = 0;
466: else high = nrow;
467: lastcol = col;
468: while (high-low > 5) {
469: t = (low+high)/2;
470: if (rp[t] > col) high = t;
471: else low = t;
472: }
473: for (i=low; i<high; i++) {
474: if (rp[i] > col) break;
475: if (rp[i] == col) {
476: if (is == ADD_VALUES) ap[i] += value;
477: else ap[i] = value;
478: low = i + 1;
479: goto noinsert;
480: }
481: }
482: if (value == 0.0 && ignorezeroentries) goto noinsert;
483: if (nonew == 1) goto noinsert;
484: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at (%D,%D) in the matrix",row,col);
485: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
486: N = nrow++ - 1; a->nz++; high++;
487: /* shift up all the later entries in this row */
488: for (ii=N; ii>=i; ii--) {
489: rp[ii+1] = rp[ii];
490: ap[ii+1] = ap[ii];
491: }
492: rp[i] = col;
493: ap[i] = value;
494: low = i + 1;
495: A->nonzerostate++;
496: noinsert:;
497: }
498: ailen[row] = nrow;
499: }
500: return(0);
501: }
506: PetscErrorCode MatGetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[])
507: {
508: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
509: PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j;
510: PetscInt *ai = a->i,*ailen = a->ilen;
511: MatScalar *ap,*aa = a->a;
514: for (k=0; k<m; k++) { /* loop over rows */
515: row = im[k];
516: if (row < 0) {v += n; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",row); */
517: 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);
518: rp = aj + ai[row]; ap = aa + ai[row];
519: nrow = ailen[row];
520: for (l=0; l<n; l++) { /* loop over columns */
521: if (in[l] < 0) {v++; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",in[l]); */
522: 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);
523: col = in[l];
524: high = nrow; low = 0; /* assume unsorted */
525: while (high-low > 5) {
526: t = (low+high)/2;
527: if (rp[t] > col) high = t;
528: else low = t;
529: }
530: for (i=low; i<high; i++) {
531: if (rp[i] > col) break;
532: if (rp[i] == col) {
533: *v++ = ap[i];
534: goto finished;
535: }
536: }
537: *v++ = 0.0;
538: finished:;
539: }
540: }
541: return(0);
542: }
547: PetscErrorCode MatView_SeqAIJ_Binary(Mat A,PetscViewer viewer)
548: {
549: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
551: PetscInt i,*col_lens;
552: int fd;
553: FILE *file;
556: PetscViewerBinaryGetDescriptor(viewer,&fd);
557: PetscMalloc1(4+A->rmap->n,&col_lens);
559: col_lens[0] = MAT_FILE_CLASSID;
560: col_lens[1] = A->rmap->n;
561: col_lens[2] = A->cmap->n;
562: col_lens[3] = a->nz;
564: /* store lengths of each row and write (including header) to file */
565: for (i=0; i<A->rmap->n; i++) {
566: col_lens[4+i] = a->i[i+1] - a->i[i];
567: }
568: PetscBinaryWrite(fd,col_lens,4+A->rmap->n,PETSC_INT,PETSC_TRUE);
569: PetscFree(col_lens);
571: /* store column indices (zero start index) */
572: PetscBinaryWrite(fd,a->j,a->nz,PETSC_INT,PETSC_FALSE);
574: /* store nonzero values */
575: PetscBinaryWrite(fd,a->a,a->nz,PETSC_SCALAR,PETSC_FALSE);
577: PetscViewerBinaryGetInfoPointer(viewer,&file);
578: if (file) {
579: fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(A->rmap->bs));
580: }
581: return(0);
582: }
584: extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat,PetscViewer);
588: PetscErrorCode MatView_SeqAIJ_ASCII(Mat A,PetscViewer viewer)
589: {
590: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
591: PetscErrorCode ierr;
592: PetscInt i,j,m = A->rmap->n;
593: const char *name;
594: PetscViewerFormat format;
597: PetscViewerGetFormat(viewer,&format);
598: if (format == PETSC_VIEWER_ASCII_MATLAB) {
599: PetscInt nofinalvalue = 0;
600: if (m && ((a->i[m] == a->i[m-1]) || (a->j[a->nz-1] != A->cmap->n-1))) {
601: /* Need a dummy value to ensure the dimension of the matrix. */
602: nofinalvalue = 1;
603: }
604: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
605: PetscViewerASCIIPrintf(viewer,"%% Size = %D %D \n",m,A->cmap->n);
606: PetscViewerASCIIPrintf(viewer,"%% Nonzeros = %D \n",a->nz);
607: #if defined(PETSC_USE_COMPLEX)
608: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,4);\n",a->nz+nofinalvalue);
609: #else
610: PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,3);\n",a->nz+nofinalvalue);
611: #endif
612: PetscViewerASCIIPrintf(viewer,"zzz = [\n");
614: for (i=0; i<m; i++) {
615: for (j=a->i[i]; j<a->i[i+1]; j++) {
616: #if defined(PETSC_USE_COMPLEX)
617: 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]));
618: #else
619: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",i+1,a->j[j]+1,(double)a->a[j]);
620: #endif
621: }
622: }
623: if (nofinalvalue) {
624: #if defined(PETSC_USE_COMPLEX)
625: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",m,A->cmap->n,0.,0.);
626: #else
627: PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",m,A->cmap->n,0.0);
628: #endif
629: }
630: PetscObjectGetName((PetscObject)A,&name);
631: PetscViewerASCIIPrintf(viewer,"];\n %s = spconvert(zzz);\n",name);
632: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
633: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO) {
634: return(0);
635: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
636: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
637: for (i=0; i<m; i++) {
638: PetscViewerASCIIPrintf(viewer,"row %D:",i);
639: for (j=a->i[i]; j<a->i[i+1]; j++) {
640: #if defined(PETSC_USE_COMPLEX)
641: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
642: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
643: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
644: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
645: } else if (PetscRealPart(a->a[j]) != 0.0) {
646: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
647: }
648: #else
649: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);}
650: #endif
651: }
652: PetscViewerASCIIPrintf(viewer,"\n");
653: }
654: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
655: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
656: PetscInt nzd=0,fshift=1,*sptr;
657: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
658: PetscMalloc1(m+1,&sptr);
659: for (i=0; i<m; i++) {
660: sptr[i] = nzd+1;
661: for (j=a->i[i]; j<a->i[i+1]; j++) {
662: if (a->j[j] >= i) {
663: #if defined(PETSC_USE_COMPLEX)
664: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++;
665: #else
666: if (a->a[j] != 0.0) nzd++;
667: #endif
668: }
669: }
670: }
671: sptr[m] = nzd+1;
672: PetscViewerASCIIPrintf(viewer," %D %D\n\n",m,nzd);
673: for (i=0; i<m+1; i+=6) {
674: if (i+4<m) {
675: 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]);
676: } else if (i+3<m) {
677: PetscViewerASCIIPrintf(viewer," %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4]);
678: } else if (i+2<m) {
679: PetscViewerASCIIPrintf(viewer," %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3]);
680: } else if (i+1<m) {
681: PetscViewerASCIIPrintf(viewer," %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2]);
682: } else if (i<m) {
683: PetscViewerASCIIPrintf(viewer," %D %D\n",sptr[i],sptr[i+1]);
684: } else {
685: PetscViewerASCIIPrintf(viewer," %D\n",sptr[i]);
686: }
687: }
688: PetscViewerASCIIPrintf(viewer,"\n");
689: PetscFree(sptr);
690: for (i=0; i<m; i++) {
691: for (j=a->i[i]; j<a->i[i+1]; j++) {
692: if (a->j[j] >= i) {PetscViewerASCIIPrintf(viewer," %D ",a->j[j]+fshift);}
693: }
694: PetscViewerASCIIPrintf(viewer,"\n");
695: }
696: PetscViewerASCIIPrintf(viewer,"\n");
697: for (i=0; i<m; i++) {
698: for (j=a->i[i]; j<a->i[i+1]; j++) {
699: if (a->j[j] >= i) {
700: #if defined(PETSC_USE_COMPLEX)
701: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) {
702: PetscViewerASCIIPrintf(viewer," %18.16e %18.16e ",(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
703: }
704: #else
705: if (a->a[j] != 0.0) {PetscViewerASCIIPrintf(viewer," %18.16e ",(double)a->a[j]);}
706: #endif
707: }
708: }
709: PetscViewerASCIIPrintf(viewer,"\n");
710: }
711: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
712: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
713: PetscInt cnt = 0,jcnt;
714: PetscScalar value;
715: #if defined(PETSC_USE_COMPLEX)
716: PetscBool realonly = PETSC_TRUE;
718: for (i=0; i<a->i[m]; i++) {
719: if (PetscImaginaryPart(a->a[i]) != 0.0) {
720: realonly = PETSC_FALSE;
721: break;
722: }
723: }
724: #endif
726: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
727: for (i=0; i<m; i++) {
728: jcnt = 0;
729: for (j=0; j<A->cmap->n; j++) {
730: if (jcnt < a->i[i+1]-a->i[i] && j == a->j[cnt]) {
731: value = a->a[cnt++];
732: jcnt++;
733: } else {
734: value = 0.0;
735: }
736: #if defined(PETSC_USE_COMPLEX)
737: if (realonly) {
738: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)PetscRealPart(value));
739: } else {
740: PetscViewerASCIIPrintf(viewer," %7.5e+%7.5e i ",(double)PetscRealPart(value),(double)PetscImaginaryPart(value));
741: }
742: #else
743: PetscViewerASCIIPrintf(viewer," %7.5e ",(double)value);
744: #endif
745: }
746: PetscViewerASCIIPrintf(viewer,"\n");
747: }
748: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
749: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
750: PetscInt fshift=1;
751: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
752: #if defined(PETSC_USE_COMPLEX)
753: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate complex general\n");
754: #else
755: PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate real general\n");
756: #endif
757: PetscViewerASCIIPrintf(viewer,"%D %D %D\n", m, A->cmap->n, a->nz);
758: for (i=0; i<m; i++) {
759: for (j=a->i[i]; j<a->i[i+1]; j++) {
760: #if defined(PETSC_USE_COMPLEX)
761: PetscViewerASCIIPrintf(viewer,"%D %D %g %g\n", i+fshift,a->j[j]+fshift,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
762: #else
763: PetscViewerASCIIPrintf(viewer,"%D %D %g\n", i+fshift, a->j[j]+fshift, (double)a->a[j]);
764: #endif
765: }
766: }
767: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
768: } else {
769: PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);
770: if (A->factortype) {
771: for (i=0; i<m; i++) {
772: PetscViewerASCIIPrintf(viewer,"row %D:",i);
773: /* L part */
774: for (j=a->i[i]; j<a->i[i+1]; j++) {
775: #if defined(PETSC_USE_COMPLEX)
776: if (PetscImaginaryPart(a->a[j]) > 0.0) {
777: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
778: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
779: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
780: } else {
781: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
782: }
783: #else
784: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
785: #endif
786: }
787: /* diagonal */
788: j = a->diag[i];
789: #if defined(PETSC_USE_COMPLEX)
790: if (PetscImaginaryPart(a->a[j]) > 0.0) {
791: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)PetscImaginaryPart(1.0/a->a[j]));
792: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
793: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)(-PetscImaginaryPart(1.0/a->a[j])));
794: } else {
795: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(1.0/a->a[j]));
796: }
797: #else
798: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)(1.0/a->a[j]));
799: #endif
801: /* U part */
802: for (j=a->diag[i+1]+1; j<a->diag[i]; j++) {
803: #if defined(PETSC_USE_COMPLEX)
804: if (PetscImaginaryPart(a->a[j]) > 0.0) {
805: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
806: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
807: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));
808: } else {
809: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
810: }
811: #else
812: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
813: #endif
814: }
815: PetscViewerASCIIPrintf(viewer,"\n");
816: }
817: } else {
818: for (i=0; i<m; i++) {
819: PetscViewerASCIIPrintf(viewer,"row %D:",i);
820: for (j=a->i[i]; j<a->i[i+1]; j++) {
821: #if defined(PETSC_USE_COMPLEX)
822: if (PetscImaginaryPart(a->a[j]) > 0.0) {
823: PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));
824: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
825: PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));
826: } else {
827: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));
828: }
829: #else
830: PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);
831: #endif
832: }
833: PetscViewerASCIIPrintf(viewer,"\n");
834: }
835: }
836: PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);
837: }
838: PetscViewerFlush(viewer);
839: return(0);
840: }
842: #include <petscdraw.h>
845: PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw,void *Aa)
846: {
847: Mat A = (Mat) Aa;
848: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
849: PetscErrorCode ierr;
850: PetscInt i,j,m = A->rmap->n;
851: int color;
852: PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r;
853: PetscViewer viewer;
854: PetscViewerFormat format;
857: PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);
858: PetscViewerGetFormat(viewer,&format);
859: PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);
861: /* loop over matrix elements drawing boxes */
863: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
864: PetscDrawCollectiveBegin(draw);
865: /* Blue for negative, Cyan for zero and Red for positive */
866: color = PETSC_DRAW_BLUE;
867: for (i=0; i<m; i++) {
868: y_l = m - i - 1.0; y_r = y_l + 1.0;
869: for (j=a->i[i]; j<a->i[i+1]; j++) {
870: x_l = a->j[j]; x_r = x_l + 1.0;
871: if (PetscRealPart(a->a[j]) >= 0.) continue;
872: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
873: }
874: }
875: color = PETSC_DRAW_CYAN;
876: for (i=0; i<m; i++) {
877: y_l = m - i - 1.0; y_r = y_l + 1.0;
878: for (j=a->i[i]; j<a->i[i+1]; j++) {
879: x_l = a->j[j]; x_r = x_l + 1.0;
880: if (a->a[j] != 0.) continue;
881: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
882: }
883: }
884: color = PETSC_DRAW_RED;
885: for (i=0; i<m; i++) {
886: y_l = m - i - 1.0; y_r = y_l + 1.0;
887: for (j=a->i[i]; j<a->i[i+1]; j++) {
888: x_l = a->j[j]; x_r = x_l + 1.0;
889: if (PetscRealPart(a->a[j]) <= 0.) continue;
890: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
891: }
892: }
893: PetscDrawCollectiveEnd(draw);
894: } else {
895: /* use contour shading to indicate magnitude of values */
896: /* first determine max of all nonzero values */
897: PetscReal minv = 0.0, maxv = 0.0;
898: PetscInt nz = a->nz, count = 0;
899: PetscDraw popup;
901: for (i=0; i<nz; i++) {
902: if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]);
903: }
904: if (minv >= maxv) maxv = minv + PETSC_SMALL;
905: PetscDrawGetPopup(draw,&popup);
906: PetscDrawScalePopup(popup,minv,maxv);
908: PetscDrawCollectiveBegin(draw);
909: for (i=0; i<m; i++) {
910: y_l = m - i - 1.0;
911: y_r = y_l + 1.0;
912: for (j=a->i[i]; j<a->i[i+1]; j++) {
913: x_l = a->j[j];
914: x_r = x_l + 1.0;
915: color = PetscDrawRealToColor(PetscAbsScalar(a->a[count]),minv,maxv);
916: PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);
917: count++;
918: }
919: }
920: PetscDrawCollectiveEnd(draw);
921: }
922: return(0);
923: }
925: #include <petscdraw.h>
928: PetscErrorCode MatView_SeqAIJ_Draw(Mat A,PetscViewer viewer)
929: {
931: PetscDraw draw;
932: PetscReal xr,yr,xl,yl,h,w;
933: PetscBool isnull;
936: PetscViewerDrawGetDraw(viewer,0,&draw);
937: PetscDrawIsNull(draw,&isnull);
938: if (isnull) return(0);
940: xr = A->cmap->n; yr = A->rmap->n; h = yr/10.0; w = xr/10.0;
941: xr += w; yr += h; xl = -w; yl = -h;
942: PetscDrawSetCoordinates(draw,xl,yl,xr,yr);
943: PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);
944: PetscDrawZoom(draw,MatView_SeqAIJ_Draw_Zoom,A);
945: PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);
946: PetscDrawSave(draw);
947: return(0);
948: }
952: PetscErrorCode MatView_SeqAIJ(Mat A,PetscViewer viewer)
953: {
955: PetscBool iascii,isbinary,isdraw;
958: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
959: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
960: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
961: if (iascii) {
962: MatView_SeqAIJ_ASCII(A,viewer);
963: } else if (isbinary) {
964: MatView_SeqAIJ_Binary(A,viewer);
965: } else if (isdraw) {
966: MatView_SeqAIJ_Draw(A,viewer);
967: }
968: MatView_SeqAIJ_Inode(A,viewer);
969: return(0);
970: }
974: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A,MatAssemblyType mode)
975: {
976: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
978: PetscInt fshift = 0,i,j,*ai = a->i,*aj = a->j,*imax = a->imax;
979: PetscInt m = A->rmap->n,*ip,N,*ailen = a->ilen,rmax = 0;
980: MatScalar *aa = a->a,*ap;
981: PetscReal ratio = 0.6;
984: if (mode == MAT_FLUSH_ASSEMBLY) return(0);
986: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
987: for (i=1; i<m; i++) {
988: /* move each row back by the amount of empty slots (fshift) before it*/
989: fshift += imax[i-1] - ailen[i-1];
990: rmax = PetscMax(rmax,ailen[i]);
991: if (fshift) {
992: ip = aj + ai[i];
993: ap = aa + ai[i];
994: N = ailen[i];
995: for (j=0; j<N; j++) {
996: ip[j-fshift] = ip[j];
997: ap[j-fshift] = ap[j];
998: }
999: }
1000: ai[i] = ai[i-1] + ailen[i-1];
1001: }
1002: if (m) {
1003: fshift += imax[m-1] - ailen[m-1];
1004: ai[m] = ai[m-1] + ailen[m-1];
1005: }
1007: /* reset ilen and imax for each row */
1008: a->nonzerorowcnt = 0;
1009: for (i=0; i<m; i++) {
1010: ailen[i] = imax[i] = ai[i+1] - ai[i];
1011: a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0);
1012: }
1013: a->nz = ai[m];
1014: 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);
1016: MatMarkDiagonal_SeqAIJ(A);
1017: PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap->n,fshift,a->nz);
1018: PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);
1019: PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);
1021: A->info.mallocs += a->reallocs;
1022: a->reallocs = 0;
1023: A->info.nz_unneeded = (PetscReal)fshift;
1024: a->rmax = rmax;
1026: MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,m,ratio);
1027: MatAssemblyEnd_SeqAIJ_Inode(A,mode);
1028: MatSeqAIJInvalidateDiagonal(A);
1029: return(0);
1030: }
1034: PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1035: {
1036: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1037: PetscInt i,nz = a->nz;
1038: MatScalar *aa = a->a;
1042: for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]);
1043: MatSeqAIJInvalidateDiagonal(A);
1044: return(0);
1045: }
1049: PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1050: {
1051: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1052: PetscInt i,nz = a->nz;
1053: MatScalar *aa = a->a;
1057: for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1058: MatSeqAIJInvalidateDiagonal(A);
1059: return(0);
1060: }
1064: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1065: {
1066: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1070: PetscMemzero(a->a,(a->i[A->rmap->n])*sizeof(PetscScalar));
1071: MatSeqAIJInvalidateDiagonal(A);
1072: return(0);
1073: }
1077: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1078: {
1079: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1083: #if defined(PETSC_USE_LOG)
1084: PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->n,A->cmap->n,a->nz);
1085: #endif
1086: MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);
1087: ISDestroy(&a->row);
1088: ISDestroy(&a->col);
1089: PetscFree(a->diag);
1090: PetscFree(a->ibdiag);
1091: PetscFree2(a->imax,a->ilen);
1092: PetscFree3(a->idiag,a->mdiag,a->ssor_work);
1093: PetscFree(a->solve_work);
1094: ISDestroy(&a->icol);
1095: PetscFree(a->saved_values);
1096: ISColoringDestroy(&a->coloring);
1097: PetscFree2(a->compressedrow.i,a->compressedrow.rindex);
1098: PetscFree(a->matmult_abdense);
1100: MatDestroy_SeqAIJ_Inode(A);
1101: PetscFree(A->data);
1103: PetscObjectChangeTypeName((PetscObject)A,0);
1104: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetColumnIndices_C",NULL);
1105: PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);
1106: PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);
1107: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsbaij_C",NULL);
1108: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqbaij_C",NULL);
1109: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijperm_C",NULL);
1110: #if defined(PETSC_HAVE_ELEMENTAL)
1111: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_elemental_C",NULL);
1112: #endif
1113: PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqdense_C",NULL);
1114: PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);
1115: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);
1116: PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);
1117: PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);
1118: return(0);
1119: }
1123: PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg)
1124: {
1125: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1129: switch (op) {
1130: case MAT_ROW_ORIENTED:
1131: a->roworiented = flg;
1132: break;
1133: case MAT_KEEP_NONZERO_PATTERN:
1134: a->keepnonzeropattern = flg;
1135: break;
1136: case MAT_NEW_NONZERO_LOCATIONS:
1137: a->nonew = (flg ? 0 : 1);
1138: break;
1139: case MAT_NEW_NONZERO_LOCATION_ERR:
1140: a->nonew = (flg ? -1 : 0);
1141: break;
1142: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1143: a->nonew = (flg ? -2 : 0);
1144: break;
1145: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1146: a->nounused = (flg ? -1 : 0);
1147: break;
1148: case MAT_IGNORE_ZERO_ENTRIES:
1149: a->ignorezeroentries = flg;
1150: break;
1151: case MAT_SPD:
1152: case MAT_SYMMETRIC:
1153: case MAT_STRUCTURALLY_SYMMETRIC:
1154: case MAT_HERMITIAN:
1155: case MAT_SYMMETRY_ETERNAL:
1156: /* These options are handled directly by MatSetOption() */
1157: break;
1158: case MAT_NEW_DIAGONALS:
1159: case MAT_IGNORE_OFF_PROC_ENTRIES:
1160: case MAT_USE_HASH_TABLE:
1161: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1162: break;
1163: case MAT_USE_INODES:
1164: /* Not an error because MatSetOption_SeqAIJ_Inode handles this one */
1165: break;
1166: default:
1167: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1168: }
1169: MatSetOption_SeqAIJ_Inode(A,op,flg);
1170: return(0);
1171: }
1175: PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v)
1176: {
1177: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1179: PetscInt i,j,n,*ai=a->i,*aj=a->j,nz;
1180: PetscScalar *aa=a->a,*x,zero=0.0;
1183: VecGetLocalSize(v,&n);
1184: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
1186: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1187: PetscInt *diag=a->diag;
1188: VecGetArray(v,&x);
1189: for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]];
1190: VecRestoreArray(v,&x);
1191: return(0);
1192: }
1194: VecSet(v,zero);
1195: VecGetArray(v,&x);
1196: for (i=0; i<n; i++) {
1197: nz = ai[i+1] - ai[i];
1198: if (!nz) x[i] = 0.0;
1199: for (j=ai[i]; j<ai[i+1]; j++) {
1200: if (aj[j] == i) {
1201: x[i] = aa[j];
1202: break;
1203: }
1204: }
1205: }
1206: VecRestoreArray(v,&x);
1207: return(0);
1208: }
1210: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1213: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy)
1214: {
1215: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1216: PetscScalar *y;
1217: const PetscScalar *x;
1218: PetscErrorCode ierr;
1219: PetscInt m = A->rmap->n;
1220: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1221: const MatScalar *v;
1222: PetscScalar alpha;
1223: PetscInt n,i,j;
1224: const PetscInt *idx,*ii,*ridx=NULL;
1225: Mat_CompressedRow cprow = a->compressedrow;
1226: PetscBool usecprow = cprow.use;
1227: #endif
1230: if (zz != yy) {VecCopy(zz,yy);}
1231: VecGetArrayRead(xx,&x);
1232: VecGetArray(yy,&y);
1234: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1235: fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y);
1236: #else
1237: if (usecprow) {
1238: m = cprow.nrows;
1239: ii = cprow.i;
1240: ridx = cprow.rindex;
1241: } else {
1242: ii = a->i;
1243: }
1244: for (i=0; i<m; i++) {
1245: idx = a->j + ii[i];
1246: v = a->a + ii[i];
1247: n = ii[i+1] - ii[i];
1248: if (usecprow) {
1249: alpha = x[ridx[i]];
1250: } else {
1251: alpha = x[i];
1252: }
1253: for (j=0; j<n; j++) y[idx[j]] += alpha*v[j];
1254: }
1255: #endif
1256: PetscLogFlops(2.0*a->nz);
1257: VecRestoreArrayRead(xx,&x);
1258: VecRestoreArray(yy,&y);
1259: return(0);
1260: }
1264: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy)
1265: {
1269: VecSet(yy,0.0);
1270: MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);
1271: return(0);
1272: }
1274: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1278: PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy)
1279: {
1280: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1281: PetscScalar *y;
1282: const PetscScalar *x;
1283: const MatScalar *aa;
1284: PetscErrorCode ierr;
1285: PetscInt m=A->rmap->n;
1286: const PetscInt *aj,*ii,*ridx=NULL;
1287: PetscInt n,i;
1288: PetscScalar sum;
1289: PetscBool usecprow=a->compressedrow.use;
1291: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1292: #pragma disjoint(*x,*y,*aa)
1293: #endif
1296: VecGetArrayRead(xx,&x);
1297: VecGetArray(yy,&y);
1298: ii = a->i;
1299: if (usecprow) { /* use compressed row format */
1300: PetscMemzero(y,m*sizeof(PetscScalar));
1301: m = a->compressedrow.nrows;
1302: ii = a->compressedrow.i;
1303: ridx = a->compressedrow.rindex;
1304: for (i=0; i<m; i++) {
1305: n = ii[i+1] - ii[i];
1306: aj = a->j + ii[i];
1307: aa = a->a + ii[i];
1308: sum = 0.0;
1309: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1310: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1311: y[*ridx++] = sum;
1312: }
1313: } else { /* do not use compressed row format */
1314: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1315: aj = a->j;
1316: aa = a->a;
1317: fortranmultaij_(&m,x,ii,aj,aa,y);
1318: #else
1319: for (i=0; i<m; i++) {
1320: n = ii[i+1] - ii[i];
1321: aj = a->j + ii[i];
1322: aa = a->a + ii[i];
1323: sum = 0.0;
1324: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1325: y[i] = sum;
1326: }
1327: #endif
1328: }
1329: PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);
1330: VecRestoreArrayRead(xx,&x);
1331: VecRestoreArray(yy,&y);
1332: return(0);
1333: }
1337: PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy)
1338: {
1339: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1340: PetscScalar *y;
1341: const PetscScalar *x;
1342: const MatScalar *aa;
1343: PetscErrorCode ierr;
1344: PetscInt m=A->rmap->n;
1345: const PetscInt *aj,*ii,*ridx=NULL;
1346: PetscInt n,i,nonzerorow=0;
1347: PetscScalar sum;
1348: PetscBool usecprow=a->compressedrow.use;
1350: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1351: #pragma disjoint(*x,*y,*aa)
1352: #endif
1355: VecGetArrayRead(xx,&x);
1356: VecGetArray(yy,&y);
1357: if (usecprow) { /* use compressed row format */
1358: m = a->compressedrow.nrows;
1359: ii = a->compressedrow.i;
1360: ridx = a->compressedrow.rindex;
1361: for (i=0; i<m; i++) {
1362: n = ii[i+1] - ii[i];
1363: aj = a->j + ii[i];
1364: aa = a->a + ii[i];
1365: sum = 0.0;
1366: nonzerorow += (n>0);
1367: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1368: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1369: y[*ridx++] = sum;
1370: }
1371: } else { /* do not use compressed row format */
1372: ii = a->i;
1373: for (i=0; i<m; i++) {
1374: n = ii[i+1] - ii[i];
1375: aj = a->j + ii[i];
1376: aa = a->a + ii[i];
1377: sum = 0.0;
1378: nonzerorow += (n>0);
1379: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1380: y[i] = sum;
1381: }
1382: }
1383: PetscLogFlops(2.0*a->nz - nonzerorow);
1384: VecRestoreArrayRead(xx,&x);
1385: VecRestoreArray(yy,&y);
1386: return(0);
1387: }
1391: PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1392: {
1393: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1394: PetscScalar *y,*z;
1395: const PetscScalar *x;
1396: const MatScalar *aa;
1397: PetscErrorCode ierr;
1398: PetscInt m = A->rmap->n,*aj,*ii;
1399: PetscInt n,i,*ridx=NULL;
1400: PetscScalar sum;
1401: PetscBool usecprow=a->compressedrow.use;
1404: VecGetArrayRead(xx,&x);
1405: VecGetArrayPair(yy,zz,&y,&z);
1406: if (usecprow) { /* use compressed row format */
1407: if (zz != yy) {
1408: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1409: }
1410: m = a->compressedrow.nrows;
1411: ii = a->compressedrow.i;
1412: ridx = a->compressedrow.rindex;
1413: for (i=0; i<m; i++) {
1414: n = ii[i+1] - ii[i];
1415: aj = a->j + ii[i];
1416: aa = a->a + ii[i];
1417: sum = y[*ridx];
1418: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1419: z[*ridx++] = sum;
1420: }
1421: } else { /* do not use compressed row format */
1422: ii = a->i;
1423: for (i=0; i<m; i++) {
1424: n = ii[i+1] - ii[i];
1425: aj = a->j + ii[i];
1426: aa = a->a + ii[i];
1427: sum = y[i];
1428: PetscSparseDenseMaxDot(sum,x,aa,aj,n);
1429: z[i] = sum;
1430: }
1431: }
1432: PetscLogFlops(2.0*a->nz);
1433: VecRestoreArrayRead(xx,&x);
1434: VecRestoreArrayPair(yy,zz,&y,&z);
1435: return(0);
1436: }
1438: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1441: PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1442: {
1443: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1444: PetscScalar *y,*z;
1445: const PetscScalar *x;
1446: const MatScalar *aa;
1447: PetscErrorCode ierr;
1448: const PetscInt *aj,*ii,*ridx=NULL;
1449: PetscInt m = A->rmap->n,n,i;
1450: PetscScalar sum;
1451: PetscBool usecprow=a->compressedrow.use;
1454: VecGetArrayRead(xx,&x);
1455: VecGetArrayPair(yy,zz,&y,&z);
1456: if (usecprow) { /* use compressed row format */
1457: if (zz != yy) {
1458: PetscMemcpy(z,y,m*sizeof(PetscScalar));
1459: }
1460: m = a->compressedrow.nrows;
1461: ii = a->compressedrow.i;
1462: ridx = a->compressedrow.rindex;
1463: for (i=0; i<m; i++) {
1464: n = ii[i+1] - ii[i];
1465: aj = a->j + ii[i];
1466: aa = a->a + ii[i];
1467: sum = y[*ridx];
1468: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1469: z[*ridx++] = sum;
1470: }
1471: } else { /* do not use compressed row format */
1472: ii = a->i;
1473: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1474: aj = a->j;
1475: aa = a->a;
1476: fortranmultaddaij_(&m,x,ii,aj,aa,y,z);
1477: #else
1478: for (i=0; i<m; i++) {
1479: n = ii[i+1] - ii[i];
1480: aj = a->j + ii[i];
1481: aa = a->a + ii[i];
1482: sum = y[i];
1483: PetscSparseDensePlusDot(sum,x,aa,aj,n);
1484: z[i] = sum;
1485: }
1486: #endif
1487: }
1488: PetscLogFlops(2.0*a->nz);
1489: VecRestoreArrayRead(xx,&x);
1490: VecRestoreArrayPair(yy,zz,&y,&z);
1491: #if defined(PETSC_HAVE_CUSP)
1492: /*
1493: VecView(xx,0);
1494: VecView(zz,0);
1495: MatView(A,0);
1496: */
1497: #endif
1498: return(0);
1499: }
1501: /*
1502: Adds diagonal pointers to sparse matrix structure.
1503: */
1506: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1507: {
1508: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1510: PetscInt i,j,m = A->rmap->n;
1513: if (!a->diag) {
1514: PetscMalloc1(m,&a->diag);
1515: PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));
1516: }
1517: for (i=0; i<A->rmap->n; i++) {
1518: a->diag[i] = a->i[i+1];
1519: for (j=a->i[i]; j<a->i[i+1]; j++) {
1520: if (a->j[j] == i) {
1521: a->diag[i] = j;
1522: break;
1523: }
1524: }
1525: }
1526: return(0);
1527: }
1529: /*
1530: Checks for missing diagonals
1531: */
1534: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d)
1535: {
1536: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1537: PetscInt *diag,*ii = a->i,i;
1540: *missing = PETSC_FALSE;
1541: if (A->rmap->n > 0 && !ii) {
1542: *missing = PETSC_TRUE;
1543: if (d) *d = 0;
1544: PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");
1545: } else {
1546: diag = a->diag;
1547: for (i=0; i<A->rmap->n; i++) {
1548: if (diag[i] >= ii[i+1]) {
1549: *missing = PETSC_TRUE;
1550: if (d) *d = i;
1551: PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);
1552: break;
1553: }
1554: }
1555: }
1556: return(0);
1557: }
1561: /*
1562: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1563: */
1564: PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift)
1565: {
1566: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
1568: PetscInt i,*diag,m = A->rmap->n;
1569: MatScalar *v = a->a;
1570: PetscScalar *idiag,*mdiag;
1573: if (a->idiagvalid) return(0);
1574: MatMarkDiagonal_SeqAIJ(A);
1575: diag = a->diag;
1576: if (!a->idiag) {
1577: PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);
1578: PetscLogObjectMemory((PetscObject)A, 3*m*sizeof(PetscScalar));
1579: v = a->a;
1580: }
1581: mdiag = a->mdiag;
1582: idiag = a->idiag;
1584: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1585: for (i=0; i<m; i++) {
1586: mdiag[i] = v[diag[i]];
1587: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1588: if (PetscRealPart(fshift)) {
1589: PetscInfo1(A,"Zero diagonal on row %D\n",i);
1590: A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1591: } else {
1592: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i);
1593: }
1594: }
1595: idiag[i] = 1.0/v[diag[i]];
1596: }
1597: PetscLogFlops(m);
1598: } else {
1599: for (i=0; i<m; i++) {
1600: mdiag[i] = v[diag[i]];
1601: idiag[i] = omega/(fshift + v[diag[i]]);
1602: }
1603: PetscLogFlops(2.0*m);
1604: }
1605: a->idiagvalid = PETSC_TRUE;
1606: return(0);
1607: }
1609: #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h>
1612: PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1613: {
1614: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1615: PetscScalar *x,d,sum,*t,scale;
1616: const MatScalar *v,*idiag=0,*mdiag;
1617: const PetscScalar *b, *bs,*xb, *ts;
1618: PetscErrorCode ierr;
1619: PetscInt n,m = A->rmap->n,i;
1620: const PetscInt *idx,*diag;
1623: its = its*lits;
1625: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1626: if (!a->idiagvalid) {MatInvertDiagonal_SeqAIJ(A,omega,fshift);}
1627: a->fshift = fshift;
1628: a->omega = omega;
1630: diag = a->diag;
1631: t = a->ssor_work;
1632: idiag = a->idiag;
1633: mdiag = a->mdiag;
1635: VecGetArray(xx,&x);
1636: VecGetArrayRead(bb,&b);
1637: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1638: if (flag == SOR_APPLY_UPPER) {
1639: /* apply (U + D/omega) to the vector */
1640: bs = b;
1641: for (i=0; i<m; i++) {
1642: d = fshift + mdiag[i];
1643: n = a->i[i+1] - diag[i] - 1;
1644: idx = a->j + diag[i] + 1;
1645: v = a->a + diag[i] + 1;
1646: sum = b[i]*d/omega;
1647: PetscSparseDensePlusDot(sum,bs,v,idx,n);
1648: x[i] = sum;
1649: }
1650: VecRestoreArray(xx,&x);
1651: VecRestoreArrayRead(bb,&b);
1652: PetscLogFlops(a->nz);
1653: return(0);
1654: }
1656: if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented");
1657: else if (flag & SOR_EISENSTAT) {
1658: /* Let A = L + U + D; where L is lower trianglar,
1659: U is upper triangular, E = D/omega; This routine applies
1661: (L + E)^{-1} A (U + E)^{-1}
1663: to a vector efficiently using Eisenstat's trick.
1664: */
1665: scale = (2.0/omega) - 1.0;
1667: /* x = (E + U)^{-1} b */
1668: for (i=m-1; i>=0; i--) {
1669: n = a->i[i+1] - diag[i] - 1;
1670: idx = a->j + diag[i] + 1;
1671: v = a->a + diag[i] + 1;
1672: sum = b[i];
1673: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1674: x[i] = sum*idiag[i];
1675: }
1677: /* t = b - (2*E - D)x */
1678: v = a->a;
1679: for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i];
1681: /* t = (E + L)^{-1}t */
1682: ts = t;
1683: diag = a->diag;
1684: for (i=0; i<m; i++) {
1685: n = diag[i] - a->i[i];
1686: idx = a->j + a->i[i];
1687: v = a->a + a->i[i];
1688: sum = t[i];
1689: PetscSparseDenseMinusDot(sum,ts,v,idx,n);
1690: t[i] = sum*idiag[i];
1691: /* x = x + t */
1692: x[i] += t[i];
1693: }
1695: PetscLogFlops(6.0*m-1 + 2.0*a->nz);
1696: VecRestoreArray(xx,&x);
1697: VecRestoreArrayRead(bb,&b);
1698: return(0);
1699: }
1700: if (flag & SOR_ZERO_INITIAL_GUESS) {
1701: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1702: for (i=0; i<m; i++) {
1703: n = diag[i] - a->i[i];
1704: idx = a->j + a->i[i];
1705: v = a->a + a->i[i];
1706: sum = b[i];
1707: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1708: t[i] = sum;
1709: x[i] = sum*idiag[i];
1710: }
1711: xb = t;
1712: PetscLogFlops(a->nz);
1713: } else xb = b;
1714: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1715: for (i=m-1; i>=0; i--) {
1716: n = a->i[i+1] - diag[i] - 1;
1717: idx = a->j + diag[i] + 1;
1718: v = a->a + diag[i] + 1;
1719: sum = xb[i];
1720: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1721: if (xb == b) {
1722: x[i] = sum*idiag[i];
1723: } else {
1724: x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1725: }
1726: }
1727: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1728: }
1729: its--;
1730: }
1731: while (its--) {
1732: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1733: for (i=0; i<m; i++) {
1734: /* lower */
1735: n = diag[i] - a->i[i];
1736: idx = a->j + a->i[i];
1737: v = a->a + a->i[i];
1738: sum = b[i];
1739: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1740: t[i] = sum; /* save application of the lower-triangular part */
1741: /* upper */
1742: n = a->i[i+1] - diag[i] - 1;
1743: idx = a->j + diag[i] + 1;
1744: v = a->a + diag[i] + 1;
1745: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1746: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1747: }
1748: xb = t;
1749: PetscLogFlops(2.0*a->nz);
1750: } else xb = b;
1751: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1752: for (i=m-1; i>=0; i--) {
1753: sum = xb[i];
1754: if (xb == b) {
1755: /* whole matrix (no checkpointing available) */
1756: n = a->i[i+1] - a->i[i];
1757: idx = a->j + a->i[i];
1758: v = a->a + a->i[i];
1759: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1760: x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i];
1761: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
1762: n = a->i[i+1] - diag[i] - 1;
1763: idx = a->j + diag[i] + 1;
1764: v = a->a + diag[i] + 1;
1765: PetscSparseDenseMinusDot(sum,x,v,idx,n);
1766: x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */
1767: }
1768: }
1769: if (xb == b) {
1770: PetscLogFlops(2.0*a->nz);
1771: } else {
1772: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1773: }
1774: }
1775: }
1776: VecRestoreArray(xx,&x);
1777: VecRestoreArrayRead(bb,&b);
1778: return(0);
1779: }
1784: PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info)
1785: {
1786: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1789: info->block_size = 1.0;
1790: info->nz_allocated = (double)a->maxnz;
1791: info->nz_used = (double)a->nz;
1792: info->nz_unneeded = (double)(a->maxnz - a->nz);
1793: info->assemblies = (double)A->num_ass;
1794: info->mallocs = (double)A->info.mallocs;
1795: info->memory = ((PetscObject)A)->mem;
1796: if (A->factortype) {
1797: info->fill_ratio_given = A->info.fill_ratio_given;
1798: info->fill_ratio_needed = A->info.fill_ratio_needed;
1799: info->factor_mallocs = A->info.factor_mallocs;
1800: } else {
1801: info->fill_ratio_given = 0;
1802: info->fill_ratio_needed = 0;
1803: info->factor_mallocs = 0;
1804: }
1805: return(0);
1806: }
1810: PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1811: {
1812: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1813: PetscInt i,m = A->rmap->n - 1,d = 0;
1814: PetscErrorCode ierr;
1815: const PetscScalar *xx;
1816: PetscScalar *bb;
1817: PetscBool missing;
1820: if (x && b) {
1821: VecGetArrayRead(x,&xx);
1822: VecGetArray(b,&bb);
1823: for (i=0; i<N; i++) {
1824: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1825: bb[rows[i]] = diag*xx[rows[i]];
1826: }
1827: VecRestoreArrayRead(x,&xx);
1828: VecRestoreArray(b,&bb);
1829: }
1831: if (a->keepnonzeropattern) {
1832: for (i=0; i<N; i++) {
1833: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1834: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1835: }
1836: if (diag != 0.0) {
1837: MatMissingDiagonal_SeqAIJ(A,&missing,&d);
1838: if (missing) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D",d);
1839: for (i=0; i<N; i++) {
1840: a->a[a->diag[rows[i]]] = diag;
1841: }
1842: }
1843: } else {
1844: if (diag != 0.0) {
1845: for (i=0; i<N; i++) {
1846: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1847: if (a->ilen[rows[i]] > 0) {
1848: a->ilen[rows[i]] = 1;
1849: a->a[a->i[rows[i]]] = diag;
1850: a->j[a->i[rows[i]]] = rows[i];
1851: } else { /* in case row was completely empty */
1852: MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);
1853: }
1854: }
1855: } else {
1856: for (i=0; i<N; i++) {
1857: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1858: a->ilen[rows[i]] = 0;
1859: }
1860: }
1861: A->nonzerostate++;
1862: }
1863: MatAssemblyEnd_SeqAIJ(A,MAT_FINAL_ASSEMBLY);
1864: return(0);
1865: }
1869: PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1870: {
1871: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1872: PetscInt i,j,m = A->rmap->n - 1,d = 0;
1873: PetscErrorCode ierr;
1874: PetscBool missing,*zeroed,vecs = PETSC_FALSE;
1875: const PetscScalar *xx;
1876: PetscScalar *bb;
1879: if (x && b) {
1880: VecGetArrayRead(x,&xx);
1881: VecGetArray(b,&bb);
1882: vecs = PETSC_TRUE;
1883: }
1884: PetscCalloc1(A->rmap->n,&zeroed);
1885: for (i=0; i<N; i++) {
1886: if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]);
1887: PetscMemzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]*sizeof(PetscScalar));
1889: zeroed[rows[i]] = PETSC_TRUE;
1890: }
1891: for (i=0; i<A->rmap->n; i++) {
1892: if (!zeroed[i]) {
1893: for (j=a->i[i]; j<a->i[i+1]; j++) {
1894: if (zeroed[a->j[j]]) {
1895: if (vecs) bb[i] -= a->a[j]*xx[a->j[j]];
1896: a->a[j] = 0.0;
1897: }
1898: }
1899: } else if (vecs) bb[i] = diag*xx[i];
1900: }
1901: if (x && b) {
1902: VecRestoreArrayRead(x,&xx);
1903: VecRestoreArray(b,&bb);
1904: }
1905: PetscFree(zeroed);
1906: if (diag != 0.0) {
1907: MatMissingDiagonal_SeqAIJ(A,&missing,&d);
1908: if (missing) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D",d);
1909: for (i=0; i<N; i++) {
1910: a->a[a->diag[rows[i]]] = diag;
1911: }
1912: }
1913: MatAssemblyEnd_SeqAIJ(A,MAT_FINAL_ASSEMBLY);
1914: return(0);
1915: }
1919: PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1920: {
1921: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1922: PetscInt *itmp;
1925: if (row < 0 || row >= A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range",row);
1927: *nz = a->i[row+1] - a->i[row];
1928: if (v) *v = a->a + a->i[row];
1929: if (idx) {
1930: itmp = a->j + a->i[row];
1931: if (*nz) *idx = itmp;
1932: else *idx = 0;
1933: }
1934: return(0);
1935: }
1937: /* remove this function? */
1940: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1941: {
1943: return(0);
1944: }
1948: PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm)
1949: {
1950: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
1951: MatScalar *v = a->a;
1952: PetscReal sum = 0.0;
1954: PetscInt i,j;
1957: if (type == NORM_FROBENIUS) {
1958: for (i=0; i<a->nz; i++) {
1959: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1960: }
1961: *nrm = PetscSqrtReal(sum);
1962: PetscLogFlops(2*a->nz);
1963: } else if (type == NORM_1) {
1964: PetscReal *tmp;
1965: PetscInt *jj = a->j;
1966: PetscCalloc1(A->cmap->n+1,&tmp);
1967: *nrm = 0.0;
1968: for (j=0; j<a->nz; j++) {
1969: tmp[*jj++] += PetscAbsScalar(*v); v++;
1970: }
1971: for (j=0; j<A->cmap->n; j++) {
1972: if (tmp[j] > *nrm) *nrm = tmp[j];
1973: }
1974: PetscFree(tmp);
1975: PetscLogFlops(PetscMax(a->nz-1,0));
1976: } else if (type == NORM_INFINITY) {
1977: *nrm = 0.0;
1978: for (j=0; j<A->rmap->n; j++) {
1979: v = a->a + a->i[j];
1980: sum = 0.0;
1981: for (i=0; i<a->i[j+1]-a->i[j]; i++) {
1982: sum += PetscAbsScalar(*v); v++;
1983: }
1984: if (sum > *nrm) *nrm = sum;
1985: }
1986: PetscLogFlops(PetscMax(a->nz-1,0));
1987: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm");
1988: return(0);
1989: }
1991: /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */
1994: PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B)
1995: {
1997: PetscInt i,j,anzj;
1998: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b;
1999: PetscInt an=A->cmap->N,am=A->rmap->N;
2000: PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j;
2003: /* Allocate space for symbolic transpose info and work array */
2004: PetscCalloc1(an+1,&ati);
2005: PetscMalloc1(ai[am],&atj);
2006: PetscMalloc1(an,&atfill);
2008: /* Walk through aj and count ## of non-zeros in each row of A^T. */
2009: /* Note: offset by 1 for fast conversion into csr format. */
2010: for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1;
2011: /* Form ati for csr format of A^T. */
2012: for (i=0;i<an;i++) ati[i+1] += ati[i];
2014: /* Copy ati into atfill so we have locations of the next free space in atj */
2015: PetscMemcpy(atfill,ati,an*sizeof(PetscInt));
2017: /* Walk through A row-wise and mark nonzero entries of A^T. */
2018: for (i=0;i<am;i++) {
2019: anzj = ai[i+1] - ai[i];
2020: for (j=0;j<anzj;j++) {
2021: atj[atfill[*aj]] = i;
2022: atfill[*aj++] += 1;
2023: }
2024: }
2026: /* Clean up temporary space and complete requests. */
2027: PetscFree(atfill);
2028: MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);
2029: MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2031: b = (Mat_SeqAIJ*)((*B)->data);
2032: b->free_a = PETSC_FALSE;
2033: b->free_ij = PETSC_TRUE;
2034: b->nonew = 0;
2035: return(0);
2036: }
2040: PetscErrorCode MatTranspose_SeqAIJ(Mat A,MatReuse reuse,Mat *B)
2041: {
2042: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2043: Mat C;
2045: PetscInt i,*aj = a->j,*ai = a->i,m = A->rmap->n,len,*col;
2046: MatScalar *array = a->a;
2049: if (reuse == MAT_REUSE_MATRIX && A == *B && m != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
2051: if (reuse == MAT_INITIAL_MATRIX || *B == A) {
2052: PetscCalloc1(1+A->cmap->n,&col);
2054: for (i=0; i<ai[m]; i++) col[aj[i]] += 1;
2055: MatCreate(PetscObjectComm((PetscObject)A),&C);
2056: MatSetSizes(C,A->cmap->n,m,A->cmap->n,m);
2057: MatSetBlockSizes(C,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2058: MatSetType(C,((PetscObject)A)->type_name);
2059: MatSeqAIJSetPreallocation_SeqAIJ(C,0,col);
2060: PetscFree(col);
2061: } else {
2062: C = *B;
2063: }
2065: for (i=0; i<m; i++) {
2066: len = ai[i+1]-ai[i];
2067: MatSetValues_SeqAIJ(C,len,aj,1,&i,array,INSERT_VALUES);
2068: array += len;
2069: aj += len;
2070: }
2071: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2072: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2074: if (reuse == MAT_INITIAL_MATRIX || *B != A) {
2075: *B = C;
2076: } else {
2077: MatHeaderMerge(A,&C);
2078: }
2079: return(0);
2080: }
2084: PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2085: {
2086: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2087: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2088: MatScalar *va,*vb;
2090: PetscInt ma,na,mb,nb, i;
2093: MatGetSize(A,&ma,&na);
2094: MatGetSize(B,&mb,&nb);
2095: if (ma!=nb || na!=mb) {
2096: *f = PETSC_FALSE;
2097: return(0);
2098: }
2099: aii = aij->i; bii = bij->i;
2100: adx = aij->j; bdx = bij->j;
2101: va = aij->a; vb = bij->a;
2102: PetscMalloc1(ma,&aptr);
2103: PetscMalloc1(mb,&bptr);
2104: for (i=0; i<ma; i++) aptr[i] = aii[i];
2105: for (i=0; i<mb; i++) bptr[i] = bii[i];
2107: *f = PETSC_TRUE;
2108: for (i=0; i<ma; i++) {
2109: while (aptr[i]<aii[i+1]) {
2110: PetscInt idc,idr;
2111: PetscScalar vc,vr;
2112: /* column/row index/value */
2113: idc = adx[aptr[i]];
2114: idr = bdx[bptr[idc]];
2115: vc = va[aptr[i]];
2116: vr = vb[bptr[idc]];
2117: if (i!=idr || PetscAbsScalar(vc-vr) > tol) {
2118: *f = PETSC_FALSE;
2119: goto done;
2120: } else {
2121: aptr[i]++;
2122: if (B || i!=idc) bptr[idc]++;
2123: }
2124: }
2125: }
2126: done:
2127: PetscFree(aptr);
2128: PetscFree(bptr);
2129: return(0);
2130: }
2134: PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f)
2135: {
2136: Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data;
2137: PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr;
2138: MatScalar *va,*vb;
2140: PetscInt ma,na,mb,nb, i;
2143: MatGetSize(A,&ma,&na);
2144: MatGetSize(B,&mb,&nb);
2145: if (ma!=nb || na!=mb) {
2146: *f = PETSC_FALSE;
2147: return(0);
2148: }
2149: aii = aij->i; bii = bij->i;
2150: adx = aij->j; bdx = bij->j;
2151: va = aij->a; vb = bij->a;
2152: PetscMalloc1(ma,&aptr);
2153: PetscMalloc1(mb,&bptr);
2154: for (i=0; i<ma; i++) aptr[i] = aii[i];
2155: for (i=0; i<mb; i++) bptr[i] = bii[i];
2157: *f = PETSC_TRUE;
2158: for (i=0; i<ma; i++) {
2159: while (aptr[i]<aii[i+1]) {
2160: PetscInt idc,idr;
2161: PetscScalar vc,vr;
2162: /* column/row index/value */
2163: idc = adx[aptr[i]];
2164: idr = bdx[bptr[idc]];
2165: vc = va[aptr[i]];
2166: vr = vb[bptr[idc]];
2167: if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) {
2168: *f = PETSC_FALSE;
2169: goto done;
2170: } else {
2171: aptr[i]++;
2172: if (B || i!=idc) bptr[idc]++;
2173: }
2174: }
2175: }
2176: done:
2177: PetscFree(aptr);
2178: PetscFree(bptr);
2179: return(0);
2180: }
2184: PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2185: {
2189: MatIsTranspose_SeqAIJ(A,A,tol,f);
2190: return(0);
2191: }
2195: PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f)
2196: {
2200: MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);
2201: return(0);
2202: }
2206: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr)
2207: {
2208: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2209: PetscScalar *l,*r,x;
2210: MatScalar *v;
2212: PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz,*jj;
2215: if (ll) {
2216: /* The local size is used so that VecMPI can be passed to this routine
2217: by MatDiagonalScale_MPIAIJ */
2218: VecGetLocalSize(ll,&m);
2219: if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length");
2220: VecGetArray(ll,&l);
2221: v = a->a;
2222: for (i=0; i<m; i++) {
2223: x = l[i];
2224: M = a->i[i+1] - a->i[i];
2225: for (j=0; j<M; j++) (*v++) *= x;
2226: }
2227: VecRestoreArray(ll,&l);
2228: PetscLogFlops(nz);
2229: }
2230: if (rr) {
2231: VecGetLocalSize(rr,&n);
2232: if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length");
2233: VecGetArray(rr,&r);
2234: v = a->a; jj = a->j;
2235: for (i=0; i<nz; i++) (*v++) *= r[*jj++];
2236: VecRestoreArray(rr,&r);
2237: PetscLogFlops(nz);
2238: }
2239: MatSeqAIJInvalidateDiagonal(A);
2240: return(0);
2241: }
2245: PetscErrorCode MatGetSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B)
2246: {
2247: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c;
2249: PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens;
2250: PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi;
2251: const PetscInt *irow,*icol;
2252: PetscInt nrows,ncols;
2253: PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen;
2254: MatScalar *a_new,*mat_a;
2255: Mat C;
2256: PetscBool stride;
2260: ISGetIndices(isrow,&irow);
2261: ISGetLocalSize(isrow,&nrows);
2262: ISGetLocalSize(iscol,&ncols);
2264: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);
2265: if (stride) {
2266: ISStrideGetInfo(iscol,&first,&step);
2267: } else {
2268: first = 0;
2269: step = 0;
2270: }
2271: if (stride && step == 1) {
2272: /* special case of contiguous rows */
2273: PetscMalloc2(nrows,&lens,nrows,&starts);
2274: /* loop over new rows determining lens and starting points */
2275: for (i=0; i<nrows; i++) {
2276: kstart = ai[irow[i]];
2277: kend = kstart + ailen[irow[i]];
2278: starts[i] = kstart;
2279: for (k=kstart; k<kend; k++) {
2280: if (aj[k] >= first) {
2281: starts[i] = k;
2282: break;
2283: }
2284: }
2285: sum = 0;
2286: while (k < kend) {
2287: if (aj[k++] >= first+ncols) break;
2288: sum++;
2289: }
2290: lens[i] = sum;
2291: }
2292: /* create submatrix */
2293: if (scall == MAT_REUSE_MATRIX) {
2294: PetscInt n_cols,n_rows;
2295: MatGetSize(*B,&n_rows,&n_cols);
2296: if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
2297: MatZeroEntries(*B);
2298: C = *B;
2299: } else {
2300: PetscInt rbs,cbs;
2301: MatCreate(PetscObjectComm((PetscObject)A),&C);
2302: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2303: ISGetBlockSize(isrow,&rbs);
2304: ISGetBlockSize(iscol,&cbs);
2305: MatSetBlockSizes(C,rbs,cbs);
2306: MatSetType(C,((PetscObject)A)->type_name);
2307: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2308: }
2309: c = (Mat_SeqAIJ*)C->data;
2311: /* loop over rows inserting into submatrix */
2312: a_new = c->a;
2313: j_new = c->j;
2314: i_new = c->i;
2316: for (i=0; i<nrows; i++) {
2317: ii = starts[i];
2318: lensi = lens[i];
2319: for (k=0; k<lensi; k++) {
2320: *j_new++ = aj[ii+k] - first;
2321: }
2322: PetscMemcpy(a_new,a->a + starts[i],lensi*sizeof(PetscScalar));
2323: a_new += lensi;
2324: i_new[i+1] = i_new[i] + lensi;
2325: c->ilen[i] = lensi;
2326: }
2327: PetscFree2(lens,starts);
2328: } else {
2329: ISGetIndices(iscol,&icol);
2330: PetscCalloc1(oldcols,&smap);
2331: PetscMalloc1(1+nrows,&lens);
2332: for (i=0; i<ncols; i++) {
2333: #if defined(PETSC_USE_DEBUG)
2334: if (icol[i] >= oldcols) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Requesting column beyond largest column icol[%D] %D <= A->cmap->n %D",i,icol[i],oldcols);
2335: #endif
2336: smap[icol[i]] = i+1;
2337: }
2339: /* determine lens of each row */
2340: for (i=0; i<nrows; i++) {
2341: kstart = ai[irow[i]];
2342: kend = kstart + a->ilen[irow[i]];
2343: lens[i] = 0;
2344: for (k=kstart; k<kend; k++) {
2345: if (smap[aj[k]]) {
2346: lens[i]++;
2347: }
2348: }
2349: }
2350: /* Create and fill new matrix */
2351: if (scall == MAT_REUSE_MATRIX) {
2352: PetscBool equal;
2354: c = (Mat_SeqAIJ*)((*B)->data);
2355: if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
2356: PetscMemcmp(c->ilen,lens,(*B)->rmap->n*sizeof(PetscInt),&equal);
2357: if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
2358: PetscMemzero(c->ilen,(*B)->rmap->n*sizeof(PetscInt));
2359: C = *B;
2360: } else {
2361: PetscInt rbs,cbs;
2362: MatCreate(PetscObjectComm((PetscObject)A),&C);
2363: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
2364: ISGetBlockSize(isrow,&rbs);
2365: ISGetBlockSize(iscol,&cbs);
2366: MatSetBlockSizes(C,rbs,cbs);
2367: MatSetType(C,((PetscObject)A)->type_name);
2368: MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);
2369: }
2370: c = (Mat_SeqAIJ*)(C->data);
2371: for (i=0; i<nrows; i++) {
2372: row = irow[i];
2373: kstart = ai[row];
2374: kend = kstart + a->ilen[row];
2375: mat_i = c->i[i];
2376: mat_j = c->j + mat_i;
2377: mat_a = c->a + mat_i;
2378: mat_ilen = c->ilen + i;
2379: for (k=kstart; k<kend; k++) {
2380: if ((tcol=smap[a->j[k]])) {
2381: *mat_j++ = tcol - 1;
2382: *mat_a++ = a->a[k];
2383: (*mat_ilen)++;
2385: }
2386: }
2387: }
2388: /* Free work space */
2389: ISRestoreIndices(iscol,&icol);
2390: PetscFree(smap);
2391: PetscFree(lens);
2392: /* sort */
2393: for (i = 0; i < nrows; i++) {
2394: PetscInt ilen;
2396: mat_i = c->i[i];
2397: mat_j = c->j + mat_i;
2398: mat_a = c->a + mat_i;
2399: ilen = c->ilen[i];
2400: PetscSortIntWithMatScalarArray(ilen,mat_j,mat_a);
2401: }
2402: }
2403: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2404: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2406: ISRestoreIndices(isrow,&irow);
2407: *B = C;
2408: return(0);
2409: }
2413: PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat)
2414: {
2416: Mat B;
2419: if (scall == MAT_INITIAL_MATRIX) {
2420: MatCreate(subComm,&B);
2421: MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);
2422: MatSetBlockSizesFromMats(B,mat,mat);
2423: MatSetType(B,MATSEQAIJ);
2424: MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);
2425: *subMat = B;
2426: } else {
2427: MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);
2428: }
2429: return(0);
2430: }
2434: PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info)
2435: {
2436: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2438: Mat outA;
2439: PetscBool row_identity,col_identity;
2442: if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu");
2444: ISIdentity(row,&row_identity);
2445: ISIdentity(col,&col_identity);
2447: outA = inA;
2448: outA->factortype = MAT_FACTOR_LU;
2449: PetscFree(inA->solvertype);
2450: PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);
2452: PetscObjectReference((PetscObject)row);
2453: ISDestroy(&a->row);
2455: a->row = row;
2457: PetscObjectReference((PetscObject)col);
2458: ISDestroy(&a->col);
2460: a->col = col;
2462: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2463: ISDestroy(&a->icol);
2464: ISInvertPermutation(col,PETSC_DECIDE,&a->icol);
2465: PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);
2467: if (!a->solve_work) { /* this matrix may have been factored before */
2468: PetscMalloc1(inA->rmap->n+1,&a->solve_work);
2469: PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));
2470: }
2472: MatMarkDiagonal_SeqAIJ(inA);
2473: if (row_identity && col_identity) {
2474: MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);
2475: } else {
2476: MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);
2477: }
2478: return(0);
2479: }
2483: PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha)
2484: {
2485: Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data;
2486: PetscScalar oalpha = alpha;
2488: PetscBLASInt one = 1,bnz;
2491: PetscBLASIntCast(a->nz,&bnz);
2492: PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&oalpha,a->a,&one));
2493: PetscLogFlops(a->nz);
2494: MatSeqAIJInvalidateDiagonal(inA);
2495: return(0);
2496: }
2500: PetscErrorCode MatGetSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[])
2501: {
2503: PetscInt i;
2506: if (scall == MAT_INITIAL_MATRIX) {
2507: PetscMalloc1(n+1,B);
2508: }
2510: for (i=0; i<n; i++) {
2511: MatGetSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);
2512: }
2513: return(0);
2514: }
2518: PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov)
2519: {
2520: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2522: PetscInt row,i,j,k,l,m,n,*nidx,isz,val;
2523: const PetscInt *idx;
2524: PetscInt start,end,*ai,*aj;
2525: PetscBT table;
2528: m = A->rmap->n;
2529: ai = a->i;
2530: aj = a->j;
2532: if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used");
2534: PetscMalloc1(m+1,&nidx);
2535: PetscBTCreate(m,&table);
2537: for (i=0; i<is_max; i++) {
2538: /* Initialize the two local arrays */
2539: isz = 0;
2540: PetscBTMemzero(m,table);
2542: /* Extract the indices, assume there can be duplicate entries */
2543: ISGetIndices(is[i],&idx);
2544: ISGetLocalSize(is[i],&n);
2546: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2547: for (j=0; j<n; ++j) {
2548: if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j];
2549: }
2550: ISRestoreIndices(is[i],&idx);
2551: ISDestroy(&is[i]);
2553: k = 0;
2554: for (j=0; j<ov; j++) { /* for each overlap */
2555: n = isz;
2556: for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */
2557: row = nidx[k];
2558: start = ai[row];
2559: end = ai[row+1];
2560: for (l = start; l<end; l++) {
2561: val = aj[l];
2562: if (!PetscBTLookupSet(table,val)) nidx[isz++] = val;
2563: }
2564: }
2565: }
2566: ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));
2567: }
2568: PetscBTDestroy(&table);
2569: PetscFree(nidx);
2570: return(0);
2571: }
2573: /* -------------------------------------------------------------- */
2576: PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B)
2577: {
2578: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2580: PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n;
2581: const PetscInt *row,*col;
2582: PetscInt *cnew,j,*lens;
2583: IS icolp,irowp;
2584: PetscInt *cwork = NULL;
2585: PetscScalar *vwork = NULL;
2588: ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);
2589: ISGetIndices(irowp,&row);
2590: ISInvertPermutation(colp,PETSC_DECIDE,&icolp);
2591: ISGetIndices(icolp,&col);
2593: /* determine lengths of permuted rows */
2594: PetscMalloc1(m+1,&lens);
2595: for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i];
2596: MatCreate(PetscObjectComm((PetscObject)A),B);
2597: MatSetSizes(*B,m,n,m,n);
2598: MatSetBlockSizesFromMats(*B,A,A);
2599: MatSetType(*B,((PetscObject)A)->type_name);
2600: MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);
2601: PetscFree(lens);
2603: PetscMalloc1(n,&cnew);
2604: for (i=0; i<m; i++) {
2605: MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2606: for (j=0; j<nz; j++) cnew[j] = col[cwork[j]];
2607: MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);
2608: MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);
2609: }
2610: PetscFree(cnew);
2612: (*B)->assembled = PETSC_FALSE;
2614: MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);
2615: MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);
2616: ISRestoreIndices(irowp,&row);
2617: ISRestoreIndices(icolp,&col);
2618: ISDestroy(&irowp);
2619: ISDestroy(&icolp);
2620: return(0);
2621: }
2625: PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str)
2626: {
2630: /* If the two matrices have the same copy implementation, use fast copy. */
2631: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2632: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2633: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
2635: if (a->i[A->rmap->n] != b->i[B->rmap->n]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number of nonzeros in two matrices are different");
2636: PetscMemcpy(b->a,a->a,(a->i[A->rmap->n])*sizeof(PetscScalar));
2637: } else {
2638: MatCopy_Basic(A,B,str);
2639: }
2640: return(0);
2641: }
2645: PetscErrorCode MatSetUp_SeqAIJ(Mat A)
2646: {
2650: MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,0);
2651: return(0);
2652: }
2656: PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[])
2657: {
2658: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2661: *array = a->a;
2662: return(0);
2663: }
2667: PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[])
2668: {
2670: return(0);
2671: }
2673: /*
2674: Computes the number of nonzeros per row needed for preallocation when X and Y
2675: have different nonzero structure.
2676: */
2679: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz)
2680: {
2681: PetscInt i,j,k,nzx,nzy;
2684: /* Set the number of nonzeros in the new matrix */
2685: for (i=0; i<m; i++) {
2686: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2687: nzx = xi[i+1] - xi[i];
2688: nzy = yi[i+1] - yi[i];
2689: nnz[i] = 0;
2690: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2691: for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */
2692: if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */
2693: nnz[i]++;
2694: }
2695: for (; k<nzy; k++) nnz[i]++;
2696: }
2697: return(0);
2698: }
2702: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz)
2703: {
2704: PetscInt m = Y->rmap->N;
2705: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2706: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2710: /* Set the number of nonzeros in the new matrix */
2711: MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);
2712: return(0);
2713: }
2717: PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2718: {
2720: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data;
2721: PetscBLASInt one=1,bnz;
2724: PetscBLASIntCast(x->nz,&bnz);
2725: if (str == SAME_NONZERO_PATTERN) {
2726: PetscScalar alpha = a;
2727: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2728: MatSeqAIJInvalidateDiagonal(Y);
2729: PetscObjectStateIncrease((PetscObject)Y);
2730: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2731: MatAXPY_Basic(Y,a,X,str);
2732: } else {
2733: Mat B;
2734: PetscInt *nnz;
2735: PetscMalloc1(Y->rmap->N,&nnz);
2736: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2737: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2738: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2739: MatSetBlockSizesFromMats(B,Y,Y);
2740: MatSetType(B,(MatType) ((PetscObject)Y)->type_name);
2741: MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);
2742: MatSeqAIJSetPreallocation(B,0,nnz);
2743: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2744: MatHeaderReplace(Y,&B);
2745: PetscFree(nnz);
2746: }
2747: return(0);
2748: }
2752: PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
2753: {
2754: #if defined(PETSC_USE_COMPLEX)
2755: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
2756: PetscInt i,nz;
2757: PetscScalar *a;
2760: nz = aij->nz;
2761: a = aij->a;
2762: for (i=0; i<nz; i++) a[i] = PetscConj(a[i]);
2763: #else
2765: #endif
2766: return(0);
2767: }
2771: PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2772: {
2773: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2775: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2776: PetscReal atmp;
2777: PetscScalar *x;
2778: MatScalar *aa;
2781: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2782: aa = a->a;
2783: ai = a->i;
2784: aj = a->j;
2786: VecSet(v,0.0);
2787: VecGetArray(v,&x);
2788: VecGetLocalSize(v,&n);
2789: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2790: for (i=0; i<m; i++) {
2791: ncols = ai[1] - ai[0]; ai++;
2792: x[i] = 0.0;
2793: for (j=0; j<ncols; j++) {
2794: atmp = PetscAbsScalar(*aa);
2795: if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2796: aa++; aj++;
2797: }
2798: }
2799: VecRestoreArray(v,&x);
2800: return(0);
2801: }
2805: PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2806: {
2807: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2809: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2810: PetscScalar *x;
2811: MatScalar *aa;
2814: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2815: aa = a->a;
2816: ai = a->i;
2817: aj = a->j;
2819: VecSet(v,0.0);
2820: VecGetArray(v,&x);
2821: VecGetLocalSize(v,&n);
2822: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2823: for (i=0; i<m; i++) {
2824: ncols = ai[1] - ai[0]; ai++;
2825: if (ncols == A->cmap->n) { /* row is dense */
2826: x[i] = *aa; if (idx) idx[i] = 0;
2827: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
2828: x[i] = 0.0;
2829: if (idx) {
2830: idx[i] = 0; /* in case ncols is zero */
2831: for (j=0;j<ncols;j++) { /* find first implicit 0.0 in the row */
2832: if (aj[j] > j) {
2833: idx[i] = j;
2834: break;
2835: }
2836: }
2837: }
2838: }
2839: for (j=0; j<ncols; j++) {
2840: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
2841: aa++; aj++;
2842: }
2843: }
2844: VecRestoreArray(v,&x);
2845: return(0);
2846: }
2850: PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2851: {
2852: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2854: PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n;
2855: PetscReal atmp;
2856: PetscScalar *x;
2857: MatScalar *aa;
2860: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2861: aa = a->a;
2862: ai = a->i;
2863: aj = a->j;
2865: VecSet(v,0.0);
2866: VecGetArray(v,&x);
2867: VecGetLocalSize(v,&n);
2868: if (n != A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector, %D vs. %D rows", A->rmap->n, n);
2869: for (i=0; i<m; i++) {
2870: ncols = ai[1] - ai[0]; ai++;
2871: if (ncols) {
2872: /* Get first nonzero */
2873: for (j = 0; j < ncols; j++) {
2874: atmp = PetscAbsScalar(aa[j]);
2875: if (atmp > 1.0e-12) {
2876: x[i] = atmp;
2877: if (idx) idx[i] = aj[j];
2878: break;
2879: }
2880: }
2881: if (j == ncols) {x[i] = PetscAbsScalar(*aa); if (idx) idx[i] = *aj;}
2882: } else {
2883: x[i] = 0.0; if (idx) idx[i] = 0;
2884: }
2885: for (j = 0; j < ncols; j++) {
2886: atmp = PetscAbsScalar(*aa);
2887: if (atmp > 1.0e-12 && PetscAbsScalar(x[i]) > atmp) {x[i] = atmp; if (idx) idx[i] = *aj;}
2888: aa++; aj++;
2889: }
2890: }
2891: VecRestoreArray(v,&x);
2892: return(0);
2893: }
2897: PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[])
2898: {
2899: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
2900: PetscErrorCode ierr;
2901: PetscInt i,j,m = A->rmap->n,ncols,n;
2902: const PetscInt *ai,*aj;
2903: PetscScalar *x;
2904: const MatScalar *aa;
2907: if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
2908: aa = a->a;
2909: ai = a->i;
2910: aj = a->j;
2912: VecSet(v,0.0);
2913: VecGetArray(v,&x);
2914: VecGetLocalSize(v,&n);
2915: if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector");
2916: for (i=0; i<m; i++) {
2917: ncols = ai[1] - ai[0]; ai++;
2918: if (ncols == A->cmap->n) { /* row is dense */
2919: x[i] = *aa; if (idx) idx[i] = 0;
2920: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
2921: x[i] = 0.0;
2922: if (idx) { /* find first implicit 0.0 in the row */
2923: idx[i] = 0; /* in case ncols is zero */
2924: for (j=0; j<ncols; j++) {
2925: if (aj[j] > j) {
2926: idx[i] = j;
2927: break;
2928: }
2929: }
2930: }
2931: }
2932: for (j=0; j<ncols; j++) {
2933: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;}
2934: aa++; aj++;
2935: }
2936: }
2937: VecRestoreArray(v,&x);
2938: return(0);
2939: }
2941: #include <petscblaslapack.h>
2942: #include <petsc/private/kernels/blockinvert.h>
2946: PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values)
2947: {
2948: Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data;
2950: PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j;
2951: MatScalar *diag,work[25],*v_work;
2952: PetscReal shift = 0.0;
2953: PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE;
2956: allowzeropivot = PetscNot(A->erroriffailure);
2957: if (a->ibdiagvalid) {
2958: if (values) *values = a->ibdiag;
2959: return(0);
2960: }
2961: MatMarkDiagonal_SeqAIJ(A);
2962: if (!a->ibdiag) {
2963: PetscMalloc1(bs2*mbs,&a->ibdiag);
2964: PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));
2965: }
2966: diag = a->ibdiag;
2967: if (values) *values = a->ibdiag;
2968: /* factor and invert each block */
2969: switch (bs) {
2970: case 1:
2971: for (i=0; i<mbs; i++) {
2972: MatGetValues(A,1,&i,1,&i,diag+i);
2973: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
2974: if (allowzeropivot) {
2975: A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
2976: PetscInfo1(A,"Zero pivot, row %D\n",i);
2977: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D",i);
2978: }
2979: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
2980: }
2981: break;
2982: case 2:
2983: for (i=0; i<mbs; i++) {
2984: ij[0] = 2*i; ij[1] = 2*i + 1;
2985: MatGetValues(A,2,ij,2,ij,diag);
2986: PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);
2987: if (zeropivotdetected) A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
2988: PetscKernel_A_gets_transpose_A_2(diag);
2989: diag += 4;
2990: }
2991: break;
2992: case 3:
2993: for (i=0; i<mbs; i++) {
2994: ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2;
2995: MatGetValues(A,3,ij,3,ij,diag);
2996: PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);
2997: if (zeropivotdetected) A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
2998: PetscKernel_A_gets_transpose_A_3(diag);
2999: diag += 9;
3000: }
3001: break;
3002: case 4:
3003: for (i=0; i<mbs; i++) {
3004: ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3;
3005: MatGetValues(A,4,ij,4,ij,diag);
3006: PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);
3007: if (zeropivotdetected) A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3008: PetscKernel_A_gets_transpose_A_4(diag);
3009: diag += 16;
3010: }
3011: break;
3012: case 5:
3013: for (i=0; i<mbs; i++) {
3014: ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4;
3015: MatGetValues(A,5,ij,5,ij,diag);
3016: PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);
3017: if (zeropivotdetected) A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3018: PetscKernel_A_gets_transpose_A_5(diag);
3019: diag += 25;
3020: }
3021: break;
3022: case 6:
3023: for (i=0; i<mbs; i++) {
3024: 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;
3025: MatGetValues(A,6,ij,6,ij,diag);
3026: PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);
3027: if (zeropivotdetected) A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3028: PetscKernel_A_gets_transpose_A_6(diag);
3029: diag += 36;
3030: }
3031: break;
3032: case 7:
3033: for (i=0; i<mbs; i++) {
3034: 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;
3035: MatGetValues(A,7,ij,7,ij,diag);
3036: PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);
3037: if (zeropivotdetected) A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3038: PetscKernel_A_gets_transpose_A_7(diag);
3039: diag += 49;
3040: }
3041: break;
3042: default:
3043: PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);
3044: for (i=0; i<mbs; i++) {
3045: for (j=0; j<bs; j++) {
3046: IJ[j] = bs*i + j;
3047: }
3048: MatGetValues(A,bs,IJ,bs,IJ,diag);
3049: PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);
3050: if (zeropivotdetected) A->errortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3051: PetscKernel_A_gets_transpose_A_N(diag,bs);
3052: diag += bs2;
3053: }
3054: PetscFree3(v_work,v_pivots,IJ);
3055: }
3056: a->ibdiagvalid = PETSC_TRUE;
3057: return(0);
3058: }
3062: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx)
3063: {
3065: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data;
3066: PetscScalar a;
3067: PetscInt m,n,i,j,col;
3070: if (!x->assembled) {
3071: MatGetSize(x,&m,&n);
3072: for (i=0; i<m; i++) {
3073: for (j=0; j<aij->imax[i]; j++) {
3074: PetscRandomGetValue(rctx,&a);
3075: col = (PetscInt)(n*PetscRealPart(a));
3076: MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);
3077: }
3078: }
3079: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not yet coded");
3080: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
3081: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
3082: return(0);
3083: }
3087: PetscErrorCode MatShift_SeqAIJ(Mat Y,PetscScalar a)
3088: {
3090: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)Y->data;
3093: if (!Y->preallocated || !aij->nz) {
3094: MatSeqAIJSetPreallocation(Y,1,NULL);
3095: }
3096: MatShift_Basic(Y,a);
3097: return(0);
3098: }
3100: /* -------------------------------------------------------------------*/
3101: static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ,
3102: MatGetRow_SeqAIJ,
3103: MatRestoreRow_SeqAIJ,
3104: MatMult_SeqAIJ,
3105: /* 4*/ MatMultAdd_SeqAIJ,
3106: MatMultTranspose_SeqAIJ,
3107: MatMultTransposeAdd_SeqAIJ,
3108: 0,
3109: 0,
3110: 0,
3111: /* 10*/ 0,
3112: MatLUFactor_SeqAIJ,
3113: 0,
3114: MatSOR_SeqAIJ,
3115: MatTranspose_SeqAIJ,
3116: /*1 5*/ MatGetInfo_SeqAIJ,
3117: MatEqual_SeqAIJ,
3118: MatGetDiagonal_SeqAIJ,
3119: MatDiagonalScale_SeqAIJ,
3120: MatNorm_SeqAIJ,
3121: /* 20*/ 0,
3122: MatAssemblyEnd_SeqAIJ,
3123: MatSetOption_SeqAIJ,
3124: MatZeroEntries_SeqAIJ,
3125: /* 24*/ MatZeroRows_SeqAIJ,
3126: 0,
3127: 0,
3128: 0,
3129: 0,
3130: /* 29*/ MatSetUp_SeqAIJ,
3131: 0,
3132: 0,
3133: 0,
3134: 0,
3135: /* 34*/ MatDuplicate_SeqAIJ,
3136: 0,
3137: 0,
3138: MatILUFactor_SeqAIJ,
3139: 0,
3140: /* 39*/ MatAXPY_SeqAIJ,
3141: MatGetSubMatrices_SeqAIJ,
3142: MatIncreaseOverlap_SeqAIJ,
3143: MatGetValues_SeqAIJ,
3144: MatCopy_SeqAIJ,
3145: /* 44*/ MatGetRowMax_SeqAIJ,
3146: MatScale_SeqAIJ,
3147: MatShift_SeqAIJ,
3148: MatDiagonalSet_SeqAIJ,
3149: MatZeroRowsColumns_SeqAIJ,
3150: /* 49*/ MatSetRandom_SeqAIJ,
3151: MatGetRowIJ_SeqAIJ,
3152: MatRestoreRowIJ_SeqAIJ,
3153: MatGetColumnIJ_SeqAIJ,
3154: MatRestoreColumnIJ_SeqAIJ,
3155: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3156: 0,
3157: 0,
3158: MatPermute_SeqAIJ,
3159: 0,
3160: /* 59*/ 0,
3161: MatDestroy_SeqAIJ,
3162: MatView_SeqAIJ,
3163: 0,
3164: MatMatMatMult_SeqAIJ_SeqAIJ_SeqAIJ,
3165: /* 64*/ MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ,
3166: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3167: 0,
3168: 0,
3169: 0,
3170: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3171: MatGetRowMinAbs_SeqAIJ,
3172: 0,
3173: MatSetColoring_SeqAIJ,
3174: 0,
3175: /* 74*/ MatSetValuesAdifor_SeqAIJ,
3176: MatFDColoringApply_AIJ,
3177: 0,
3178: 0,
3179: 0,
3180: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3181: 0,
3182: 0,
3183: 0,
3184: MatLoad_SeqAIJ,
3185: /* 84*/ MatIsSymmetric_SeqAIJ,
3186: MatIsHermitian_SeqAIJ,
3187: 0,
3188: 0,
3189: 0,
3190: /* 89*/ MatMatMult_SeqAIJ_SeqAIJ,
3191: MatMatMultSymbolic_SeqAIJ_SeqAIJ,
3192: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3193: MatPtAP_SeqAIJ_SeqAIJ,
3194: MatPtAPSymbolic_SeqAIJ_SeqAIJ_DenseAxpy,
3195: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ,
3196: MatMatTransposeMult_SeqAIJ_SeqAIJ,
3197: MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ,
3198: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3199: 0,
3200: /* 99*/ 0,
3201: 0,
3202: 0,
3203: MatConjugate_SeqAIJ,
3204: 0,
3205: /*104*/ MatSetValuesRow_SeqAIJ,
3206: MatRealPart_SeqAIJ,
3207: MatImaginaryPart_SeqAIJ,
3208: 0,
3209: 0,
3210: /*109*/ MatMatSolve_SeqAIJ,
3211: 0,
3212: MatGetRowMin_SeqAIJ,
3213: 0,
3214: MatMissingDiagonal_SeqAIJ,
3215: /*114*/ 0,
3216: 0,
3217: 0,
3218: 0,
3219: 0,
3220: /*119*/ 0,
3221: 0,
3222: 0,
3223: 0,
3224: MatGetMultiProcBlock_SeqAIJ,
3225: /*124*/ MatFindNonzeroRows_SeqAIJ,
3226: MatGetColumnNorms_SeqAIJ,
3227: MatInvertBlockDiagonal_SeqAIJ,
3228: 0,
3229: 0,
3230: /*129*/ 0,
3231: MatTransposeMatMult_SeqAIJ_SeqAIJ,
3232: MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ,
3233: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3234: MatTransposeColoringCreate_SeqAIJ,
3235: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3236: MatTransColoringApplyDenToSp_SeqAIJ,
3237: MatRARt_SeqAIJ_SeqAIJ,
3238: MatRARtSymbolic_SeqAIJ_SeqAIJ,
3239: MatRARtNumeric_SeqAIJ_SeqAIJ,
3240: /*139*/0,
3241: 0,
3242: 0,
3243: MatFDColoringSetUp_SeqXAIJ,
3244: MatFindOffBlockDiagonalEntries_SeqAIJ,
3245: /*144*/MatCreateMPIMatConcatenateSeqMat_SeqAIJ
3246: };
3250: PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices)
3251: {
3252: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3253: PetscInt i,nz,n;
3256: nz = aij->maxnz;
3257: n = mat->rmap->n;
3258: for (i=0; i<nz; i++) {
3259: aij->j[i] = indices[i];
3260: }
3261: aij->nz = nz;
3262: for (i=0; i<n; i++) {
3263: aij->ilen[i] = aij->imax[i];
3264: }
3265: return(0);
3266: }
3270: /*@
3271: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3272: in the matrix.
3274: Input Parameters:
3275: + mat - the SeqAIJ matrix
3276: - indices - the column indices
3278: Level: advanced
3280: Notes:
3281: This can be called if you have precomputed the nonzero structure of the
3282: matrix and want to provide it to the matrix object to improve the performance
3283: of the MatSetValues() operation.
3285: You MUST have set the correct numbers of nonzeros per row in the call to
3286: MatCreateSeqAIJ(), and the columns indices MUST be sorted.
3288: MUST be called before any calls to MatSetValues();
3290: The indices should start with zero, not one.
3292: @*/
3293: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices)
3294: {
3300: PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));
3301: return(0);
3302: }
3304: /* ----------------------------------------------------------------------------------------*/
3308: PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3309: {
3310: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3312: size_t nz = aij->i[mat->rmap->n];
3315: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3317: /* allocate space for values if not already there */
3318: if (!aij->saved_values) {
3319: PetscMalloc1(nz+1,&aij->saved_values);
3320: PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));
3321: }
3323: /* copy values over */
3324: PetscMemcpy(aij->saved_values,aij->a,nz*sizeof(PetscScalar));
3325: return(0);
3326: }
3330: /*@
3331: MatStoreValues - Stashes a copy of the matrix values; this allows, for
3332: example, reuse of the linear part of a Jacobian, while recomputing the
3333: nonlinear portion.
3335: Collect on Mat
3337: Input Parameters:
3338: . mat - the matrix (currently only AIJ matrices support this option)
3340: Level: advanced
3342: Common Usage, with SNESSolve():
3343: $ Create Jacobian matrix
3344: $ Set linear terms into matrix
3345: $ Apply boundary conditions to matrix, at this time matrix must have
3346: $ final nonzero structure (i.e. setting the nonlinear terms and applying
3347: $ boundary conditions again will not change the nonzero structure
3348: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3349: $ MatStoreValues(mat);
3350: $ Call SNESSetJacobian() with matrix
3351: $ In your Jacobian routine
3352: $ MatRetrieveValues(mat);
3353: $ Set nonlinear terms in matrix
3355: Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself:
3356: $ // build linear portion of Jacobian
3357: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3358: $ MatStoreValues(mat);
3359: $ loop over nonlinear iterations
3360: $ MatRetrieveValues(mat);
3361: $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3362: $ // call MatAssemblyBegin/End() on matrix
3363: $ Solve linear system with Jacobian
3364: $ endloop
3366: Notes:
3367: Matrix must already be assemblied before calling this routine
3368: Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before
3369: calling this routine.
3371: When this is called multiple times it overwrites the previous set of stored values
3372: and does not allocated additional space.
3374: .seealso: MatRetrieveValues()
3376: @*/
3377: PetscErrorCode MatStoreValues(Mat mat)
3378: {
3383: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3384: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3385: PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));
3386: return(0);
3387: }
3391: PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3392: {
3393: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data;
3395: PetscInt nz = aij->i[mat->rmap->n];
3398: if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3399: if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first");
3400: /* copy values over */
3401: PetscMemcpy(aij->a,aij->saved_values,nz*sizeof(PetscScalar));
3402: return(0);
3403: }
3407: /*@
3408: MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for
3409: example, reuse of the linear part of a Jacobian, while recomputing the
3410: nonlinear portion.
3412: Collect on Mat
3414: Input Parameters:
3415: . mat - the matrix (currently on AIJ matrices support this option)
3417: Level: advanced
3419: .seealso: MatStoreValues()
3421: @*/
3422: PetscErrorCode MatRetrieveValues(Mat mat)
3423: {
3428: if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
3429: if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3430: PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));
3431: return(0);
3432: }
3435: /* --------------------------------------------------------------------------------*/
3438: /*@C
3439: MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format
3440: (the default parallel PETSc format). For good matrix assembly performance
3441: the user should preallocate the matrix storage by setting the parameter nz
3442: (or the array nnz). By setting these parameters accurately, performance
3443: during matrix assembly can be increased by more than a factor of 50.
3445: Collective on MPI_Comm
3447: Input Parameters:
3448: + comm - MPI communicator, set to PETSC_COMM_SELF
3449: . m - number of rows
3450: . n - number of columns
3451: . nz - number of nonzeros per row (same for all rows)
3452: - nnz - array containing the number of nonzeros in the various rows
3453: (possibly different for each row) or NULL
3455: Output Parameter:
3456: . A - the matrix
3458: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3459: MatXXXXSetPreallocation() paradgm instead of this routine directly.
3460: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
3462: Notes:
3463: If nnz is given then nz is ignored
3465: The AIJ format (also called the Yale sparse matrix format or
3466: compressed row storage), is fully compatible with standard Fortran 77
3467: storage. That is, the stored row and column indices can begin at
3468: either one (as in Fortran) or zero. See the users' manual for details.
3470: Specify the preallocated storage with either nz or nnz (not both).
3471: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3472: allocation. For large problems you MUST preallocate memory or you
3473: will get TERRIBLE performance, see the users' manual chapter on matrices.
3475: By default, this format uses inodes (identical nodes) when possible, to
3476: improve numerical efficiency of matrix-vector products and solves. We
3477: search for consecutive rows with the same nonzero structure, thereby
3478: reusing matrix information to achieve increased efficiency.
3480: Options Database Keys:
3481: + -mat_no_inode - Do not use inodes
3482: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3484: Level: intermediate
3486: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays()
3488: @*/
3489: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A)
3490: {
3494: MatCreate(comm,A);
3495: MatSetSizes(*A,m,n,m,n);
3496: MatSetType(*A,MATSEQAIJ);
3497: MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);
3498: return(0);
3499: }
3503: /*@C
3504: MatSeqAIJSetPreallocation - For good matrix assembly performance
3505: the user should preallocate the matrix storage by setting the parameter nz
3506: (or the array nnz). By setting these parameters accurately, performance
3507: during matrix assembly can be increased by more than a factor of 50.
3509: Collective on MPI_Comm
3511: Input Parameters:
3512: + B - The matrix
3513: . nz - number of nonzeros per row (same for all rows)
3514: - nnz - array containing the number of nonzeros in the various rows
3515: (possibly different for each row) or NULL
3517: Notes:
3518: If nnz is given then nz is ignored
3520: The AIJ format (also called the Yale sparse matrix format or
3521: compressed row storage), is fully compatible with standard Fortran 77
3522: storage. That is, the stored row and column indices can begin at
3523: either one (as in Fortran) or zero. See the users' manual for details.
3525: Specify the preallocated storage with either nz or nnz (not both).
3526: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
3527: allocation. For large problems you MUST preallocate memory or you
3528: will get TERRIBLE performance, see the users' manual chapter on matrices.
3530: You can call MatGetInfo() to get information on how effective the preallocation was;
3531: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3532: You can also run with the option -info and look for messages with the string
3533: malloc in them to see if additional memory allocation was needed.
3535: Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix
3536: entries or columns indices
3538: By default, this format uses inodes (identical nodes) when possible, to
3539: improve numerical efficiency of matrix-vector products and solves. We
3540: search for consecutive rows with the same nonzero structure, thereby
3541: reusing matrix information to achieve increased efficiency.
3543: Options Database Keys:
3544: + -mat_no_inode - Do not use inodes
3545: . -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3546: - -mat_aij_oneindex - Internally use indexing starting at 1
3547: rather than 0. Note that when calling MatSetValues(),
3548: the user still MUST index entries starting at 0!
3550: Level: intermediate
3552: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo()
3554: @*/
3555: PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[])
3556: {
3562: PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));
3563: return(0);
3564: }
3568: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz)
3569: {
3570: Mat_SeqAIJ *b;
3571: PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE;
3573: PetscInt i;
3576: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3577: if (nz == MAT_SKIP_ALLOCATION) {
3578: skipallocation = PETSC_TRUE;
3579: nz = 0;
3580: }
3582: PetscLayoutSetUp(B->rmap);
3583: PetscLayoutSetUp(B->cmap);
3585: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3586: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz);
3587: if (nnz) {
3588: for (i=0; i<B->rmap->n; i++) {
3589: 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]);
3590: 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);
3591: }
3592: }
3594: B->preallocated = PETSC_TRUE;
3596: b = (Mat_SeqAIJ*)B->data;
3598: if (!skipallocation) {
3599: if (!b->imax) {
3600: PetscMalloc2(B->rmap->n,&b->imax,B->rmap->n,&b->ilen);
3601: PetscLogObjectMemory((PetscObject)B,2*B->rmap->n*sizeof(PetscInt));
3602: }
3603: if (!nnz) {
3604: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3605: else if (nz < 0) nz = 1;
3606: for (i=0; i<B->rmap->n; i++) b->imax[i] = nz;
3607: nz = nz*B->rmap->n;
3608: } else {
3609: nz = 0;
3610: for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz += nnz[i];}
3611: }
3612: /* b->ilen will count nonzeros in each row so far. */
3613: for (i=0; i<B->rmap->n; i++) b->ilen[i] = 0;
3615: /* allocate the matrix space */
3616: /* FIXME: should B's old memory be unlogged? */
3617: MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);
3618: PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);
3619: PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));
3620: b->i[0] = 0;
3621: for (i=1; i<B->rmap->n+1; i++) {
3622: b->i[i] = b->i[i-1] + b->imax[i-1];
3623: }
3624: b->singlemalloc = PETSC_TRUE;
3625: b->free_a = PETSC_TRUE;
3626: b->free_ij = PETSC_TRUE;
3627: } else {
3628: b->free_a = PETSC_FALSE;
3629: b->free_ij = PETSC_FALSE;
3630: }
3632: b->nz = 0;
3633: b->maxnz = nz;
3634: B->info.nz_unneeded = (double)b->maxnz;
3635: if (realalloc) {
3636: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
3637: }
3638: return(0);
3639: }
3641: #undef __FUNCT__
3643: /*@
3644: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format.
3646: Input Parameters:
3647: + B - the matrix
3648: . i - the indices into j for the start of each row (starts with zero)
3649: . j - the column indices for each row (starts with zero) these must be sorted for each row
3650: - v - optional values in the matrix
3652: Level: developer
3654: The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays()
3656: .keywords: matrix, aij, compressed row, sparse, sequential
3658: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), SeqAIJ
3659: @*/
3660: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[])
3661: {
3667: PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3668: return(0);
3669: }
3671: #undef __FUNCT__
3673: PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3674: {
3675: PetscInt i;
3676: PetscInt m,n;
3677: PetscInt nz;
3678: PetscInt *nnz, nz_max = 0;
3679: PetscScalar *values;
3683: if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]);
3685: PetscLayoutSetUp(B->rmap);
3686: PetscLayoutSetUp(B->cmap);
3688: MatGetSize(B, &m, &n);
3689: PetscMalloc1(m+1, &nnz);
3690: for (i = 0; i < m; i++) {
3691: nz = Ii[i+1]- Ii[i];
3692: nz_max = PetscMax(nz_max, nz);
3693: if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz);
3694: nnz[i] = nz;
3695: }
3696: MatSeqAIJSetPreallocation(B, 0, nnz);
3697: PetscFree(nnz);
3699: if (v) {
3700: values = (PetscScalar*) v;
3701: } else {
3702: PetscCalloc1(nz_max, &values);
3703: }
3705: for (i = 0; i < m; i++) {
3706: nz = Ii[i+1] - Ii[i];
3707: MatSetValues_SeqAIJ(B, 1, &i, nz, J+Ii[i], values + (v ? Ii[i] : 0), INSERT_VALUES);
3708: }
3710: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3711: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3713: if (!v) {
3714: PetscFree(values);
3715: }
3716: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3717: return(0);
3718: }
3720: #include <../src/mat/impls/dense/seq/dense.h>
3721: #include <petsc/private/kernels/petscaxpy.h>
3725: /*
3726: Computes (B'*A')' since computing B*A directly is untenable
3728: n p p
3729: ( ) ( ) ( )
3730: m ( A ) * n ( B ) = m ( C )
3731: ( ) ( ) ( )
3733: */
3734: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C)
3735: {
3736: PetscErrorCode ierr;
3737: Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data;
3738: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data;
3739: Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data;
3740: PetscInt i,n,m,q,p;
3741: const PetscInt *ii,*idx;
3742: const PetscScalar *b,*a,*a_q;
3743: PetscScalar *c,*c_q;
3746: m = A->rmap->n;
3747: n = A->cmap->n;
3748: p = B->cmap->n;
3749: a = sub_a->v;
3750: b = sub_b->a;
3751: c = sub_c->v;
3752: PetscMemzero(c,m*p*sizeof(PetscScalar));
3754: ii = sub_b->i;
3755: idx = sub_b->j;
3756: for (i=0; i<n; i++) {
3757: q = ii[i+1] - ii[i];
3758: while (q-->0) {
3759: c_q = c + m*(*idx);
3760: a_q = a + m*i;
3761: PetscKernelAXPY(c_q,*b,a_q,m);
3762: idx++;
3763: b++;
3764: }
3765: }
3766: return(0);
3767: }
3771: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
3772: {
3774: PetscInt m=A->rmap->n,n=B->cmap->n;
3775: Mat Cmat;
3778: 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);
3779: MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
3780: MatSetSizes(Cmat,m,n,m,n);
3781: MatSetBlockSizesFromMats(Cmat,A,B);
3782: MatSetType(Cmat,MATSEQDENSE);
3783: MatSeqDenseSetPreallocation(Cmat,NULL);
3785: Cmat->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
3787: *C = Cmat;
3788: return(0);
3789: }
3791: /* ----------------------------------------------------------------*/
3794: PETSC_INTERN PetscErrorCode MatMatMult_SeqDense_SeqAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
3795: {
3799: if (scall == MAT_INITIAL_MATRIX) {
3800: PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
3801: MatMatMultSymbolic_SeqDense_SeqAIJ(A,B,fill,C);
3802: PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
3803: }
3804: PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
3805: MatMatMultNumeric_SeqDense_SeqAIJ(A,B,*C);
3806: PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
3807: return(0);
3808: }
3811: /*MC
3812: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
3813: based on compressed sparse row format.
3815: Options Database Keys:
3816: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
3818: Level: beginner
3820: .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType
3821: M*/
3823: /*MC
3824: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
3826: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
3827: and MATMPIAIJ otherwise. As a result, for single process communicators,
3828: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
3829: for communicators controlling multiple processes. It is recommended that you call both of
3830: the above preallocation routines for simplicity.
3832: Options Database Keys:
3833: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
3835: Developer Notes: Subclasses include MATAIJCUSP, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when
3836: enough exist.
3838: Level: beginner
3840: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
3841: M*/
3843: /*MC
3844: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
3846: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
3847: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
3848: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
3849: for communicators controlling multiple processes. It is recommended that you call both of
3850: the above preallocation routines for simplicity.
3852: Options Database Keys:
3853: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
3855: Level: beginner
3857: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
3858: M*/
3860: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*);
3861: #if defined(PETSC_HAVE_ELEMENTAL)
3862: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
3863: #endif
3864: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*);
3866: #if defined(PETSC_HAVE_MATLAB_ENGINE)
3867: PETSC_EXTERN PetscErrorCode MatlabEnginePut_SeqAIJ(PetscObject,void*);
3868: PETSC_EXTERN PetscErrorCode MatlabEngineGet_SeqAIJ(PetscObject,void*);
3869: #endif
3874: /*@C
3875: MatSeqAIJGetArray - gives access to the array where the data for a SeqSeqAIJ matrix is stored
3877: Not Collective
3879: Input Parameter:
3880: . mat - a MATSEQAIJ matrix
3882: Output Parameter:
3883: . array - pointer to the data
3885: Level: intermediate
3887: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
3888: @*/
3889: PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array)
3890: {
3894: PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));
3895: return(0);
3896: }
3900: /*@C
3901: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
3903: Not Collective
3905: Input Parameter:
3906: . mat - a MATSEQAIJ matrix
3908: Output Parameter:
3909: . nz - the maximum number of nonzeros in any row
3911: Level: intermediate
3913: .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90()
3914: @*/
3915: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz)
3916: {
3917: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data;
3920: *nz = aij->rmax;
3921: return(0);
3922: }
3926: /*@C
3927: MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray()
3929: Not Collective
3931: Input Parameters:
3932: . mat - a MATSEQAIJ matrix
3933: . array - pointer to the data
3935: Level: intermediate
3937: .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90()
3938: @*/
3939: PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array)
3940: {
3944: PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));
3945: return(0);
3946: }
3950: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
3951: {
3952: Mat_SeqAIJ *b;
3954: PetscMPIInt size;
3957: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
3958: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1");
3960: PetscNewLog(B,&b);
3962: B->data = (void*)b;
3964: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
3966: b->row = 0;
3967: b->col = 0;
3968: b->icol = 0;
3969: b->reallocs = 0;
3970: b->ignorezeroentries = PETSC_FALSE;
3971: b->roworiented = PETSC_TRUE;
3972: b->nonew = 0;
3973: b->diag = 0;
3974: b->solve_work = 0;
3975: B->spptr = 0;
3976: b->saved_values = 0;
3977: b->idiag = 0;
3978: b->mdiag = 0;
3979: b->ssor_work = 0;
3980: b->omega = 1.0;
3981: b->fshift = 0.0;
3982: b->idiagvalid = PETSC_FALSE;
3983: b->ibdiagvalid = PETSC_FALSE;
3984: b->keepnonzeropattern = PETSC_FALSE;
3986: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
3987: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);
3988: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);
3990: #if defined(PETSC_HAVE_MATLAB_ENGINE)
3991: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);
3992: PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);
3993: #endif
3995: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);
3996: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);
3997: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);
3998: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);
3999: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);
4000: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);
4001: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);
4002: #if defined(PETSC_HAVE_ELEMENTAL)
4003: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);
4004: #endif
4005: PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);
4006: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);
4007: PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);
4008: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);
4009: PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);
4010: PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);
4011: PetscObjectComposeFunction((PetscObject)B,"MatMatMult_seqdense_seqaij_C",MatMatMult_SeqDense_SeqAIJ);
4012: PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_seqdense_seqaij_C",MatMatMultSymbolic_SeqDense_SeqAIJ);
4013: PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_seqdense_seqaij_C",MatMatMultNumeric_SeqDense_SeqAIJ);
4014: MatCreate_SeqAIJ_Inode(B);
4015: PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);
4016: return(0);
4017: }
4021: /*
4022: Given a matrix generated with MatGetFactor() duplicates all the information in A into B
4023: */
4024: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace)
4025: {
4026: Mat_SeqAIJ *c,*a = (Mat_SeqAIJ*)A->data;
4028: PetscInt i,m = A->rmap->n;
4031: c = (Mat_SeqAIJ*)C->data;
4033: C->factortype = A->factortype;
4034: c->row = 0;
4035: c->col = 0;
4036: c->icol = 0;
4037: c->reallocs = 0;
4039: C->assembled = PETSC_TRUE;
4041: PetscLayoutReference(A->rmap,&C->rmap);
4042: PetscLayoutReference(A->cmap,&C->cmap);
4044: PetscMalloc2(m,&c->imax,m,&c->ilen);
4045: PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));
4046: for (i=0; i<m; i++) {
4047: c->imax[i] = a->imax[i];
4048: c->ilen[i] = a->ilen[i];
4049: }
4051: /* allocate the matrix space */
4052: if (mallocmatspace) {
4053: PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);
4054: PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));
4056: c->singlemalloc = PETSC_TRUE;
4058: PetscMemcpy(c->i,a->i,(m+1)*sizeof(PetscInt));
4059: if (m > 0) {
4060: PetscMemcpy(c->j,a->j,(a->i[m])*sizeof(PetscInt));
4061: if (cpvalues == MAT_COPY_VALUES) {
4062: PetscMemcpy(c->a,a->a,(a->i[m])*sizeof(PetscScalar));
4063: } else {
4064: PetscMemzero(c->a,(a->i[m])*sizeof(PetscScalar));
4065: }
4066: }
4067: }
4069: c->ignorezeroentries = a->ignorezeroentries;
4070: c->roworiented = a->roworiented;
4071: c->nonew = a->nonew;
4072: if (a->diag) {
4073: PetscMalloc1(m+1,&c->diag);
4074: PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));
4075: for (i=0; i<m; i++) {
4076: c->diag[i] = a->diag[i];
4077: }
4078: } else c->diag = 0;
4080: c->solve_work = 0;
4081: c->saved_values = 0;
4082: c->idiag = 0;
4083: c->ssor_work = 0;
4084: c->keepnonzeropattern = a->keepnonzeropattern;
4085: c->free_a = PETSC_TRUE;
4086: c->free_ij = PETSC_TRUE;
4088: c->rmax = a->rmax;
4089: c->nz = a->nz;
4090: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4091: C->preallocated = PETSC_TRUE;
4093: c->compressedrow.use = a->compressedrow.use;
4094: c->compressedrow.nrows = a->compressedrow.nrows;
4095: if (a->compressedrow.use) {
4096: i = a->compressedrow.nrows;
4097: PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);
4098: PetscMemcpy(c->compressedrow.i,a->compressedrow.i,(i+1)*sizeof(PetscInt));
4099: PetscMemcpy(c->compressedrow.rindex,a->compressedrow.rindex,i*sizeof(PetscInt));
4100: } else {
4101: c->compressedrow.use = PETSC_FALSE;
4102: c->compressedrow.i = NULL;
4103: c->compressedrow.rindex = NULL;
4104: }
4105: c->nonzerorowcnt = a->nonzerorowcnt;
4106: C->nonzerostate = A->nonzerostate;
4108: MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);
4109: PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);
4110: return(0);
4111: }
4115: PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B)
4116: {
4120: MatCreate(PetscObjectComm((PetscObject)A),B);
4121: MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);
4122: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) {
4123: MatSetBlockSizesFromMats(*B,A,A);
4124: }
4125: MatSetType(*B,((PetscObject)A)->type_name);
4126: MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);
4127: return(0);
4128: }
4132: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4133: {
4134: Mat_SeqAIJ *a;
4136: PetscInt i,sum,nz,header[4],*rowlengths = 0,M,N,rows,cols;
4137: int fd;
4138: PetscMPIInt size;
4139: MPI_Comm comm;
4140: PetscInt bs = newMat->rmap->bs;
4143: /* force binary viewer to load .info file if it has not yet done so */
4144: PetscViewerSetUp(viewer);
4145: PetscObjectGetComm((PetscObject)viewer,&comm);
4146: MPI_Comm_size(comm,&size);
4147: if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"view must have one processor");
4149: PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");
4150: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
4151: PetscOptionsEnd();
4152: if (bs < 0) bs = 1;
4153: MatSetBlockSize(newMat,bs);
4155: PetscViewerBinaryGetDescriptor(viewer,&fd);
4156: PetscBinaryRead(fd,header,4,PETSC_INT);
4157: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object in file");
4158: M = header[1]; N = header[2]; nz = header[3];
4160: if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as SeqAIJ");
4162: /* read in row lengths */
4163: PetscMalloc1(M,&rowlengths);
4164: PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
4166: /* check if sum of rowlengths is same as nz */
4167: for (i=0,sum=0; i< M; i++) sum +=rowlengths[i];
4168: if (sum != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_READ,"Inconsistant matrix data in file. no-nonzeros = %dD, sum-row-lengths = %D\n",nz,sum);
4170: /* set global size if not set already*/
4171: if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) {
4172: MatSetSizes(newMat,PETSC_DECIDE,PETSC_DECIDE,M,N);
4173: } else {
4174: /* if sizes and type are already set, check if the matrix global sizes are correct */
4175: MatGetSize(newMat,&rows,&cols);
4176: if (rows < 0 && cols < 0) { /* user might provide local size instead of global size */
4177: MatGetLocalSize(newMat,&rows,&cols);
4178: }
4179: 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);
4180: }
4181: MatSeqAIJSetPreallocation_SeqAIJ(newMat,0,rowlengths);
4182: a = (Mat_SeqAIJ*)newMat->data;
4184: PetscBinaryRead(fd,a->j,nz,PETSC_INT);
4186: /* read in nonzero values */
4187: PetscBinaryRead(fd,a->a,nz,PETSC_SCALAR);
4189: /* set matrix "i" values */
4190: a->i[0] = 0;
4191: for (i=1; i<= M; i++) {
4192: a->i[i] = a->i[i-1] + rowlengths[i-1];
4193: a->ilen[i-1] = rowlengths[i-1];
4194: }
4195: PetscFree(rowlengths);
4197: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
4198: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
4199: return(0);
4200: }
4204: PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg)
4205: {
4206: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
4208: #if defined(PETSC_USE_COMPLEX)
4209: PetscInt k;
4210: #endif
4213: /* If the matrix dimensions are not equal,or no of nonzeros */
4214: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) {
4215: *flg = PETSC_FALSE;
4216: return(0);
4217: }
4219: /* if the a->i are the same */
4220: PetscMemcmp(a->i,b->i,(A->rmap->n+1)*sizeof(PetscInt),flg);
4221: if (!*flg) return(0);
4223: /* if a->j are the same */
4224: PetscMemcmp(a->j,b->j,(a->nz)*sizeof(PetscInt),flg);
4225: if (!*flg) return(0);
4227: /* if a->a are the same */
4228: #if defined(PETSC_USE_COMPLEX)
4229: for (k=0; k<a->nz; k++) {
4230: if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) {
4231: *flg = PETSC_FALSE;
4232: return(0);
4233: }
4234: }
4235: #else
4236: PetscMemcmp(a->a,b->a,(a->nz)*sizeof(PetscScalar),flg);
4237: #endif
4238: return(0);
4239: }
4243: /*@
4244: MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format)
4245: provided by the user.
4247: Collective on MPI_Comm
4249: Input Parameters:
4250: + comm - must be an MPI communicator of size 1
4251: . m - number of rows
4252: . n - number of columns
4253: . i - row indices
4254: . j - column indices
4255: - a - matrix values
4257: Output Parameter:
4258: . mat - the matrix
4260: Level: intermediate
4262: Notes:
4263: The i, j, and a arrays are not copied by this routine, the user must free these arrays
4264: once the matrix is destroyed and not before
4266: You cannot set new nonzero locations into this matrix, that will generate an error.
4268: The i and j indices are 0 based
4270: The format which is used for the sparse matrix input, is equivalent to a
4271: row-major ordering.. i.e for the following matrix, the input data expected is
4272: as shown
4274: $ 1 0 0
4275: $ 2 0 3
4276: $ 4 5 6
4277: $
4278: $ i = {0,1,3,6} [size = nrow+1 = 3+1]
4279: $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
4280: $ v = {1,2,3,4,5,6} [size = 6]
4283: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4285: @*/
4286: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt *i,PetscInt *j,PetscScalar *a,Mat *mat)
4287: {
4289: PetscInt ii;
4290: Mat_SeqAIJ *aij;
4291: #if defined(PETSC_USE_DEBUG)
4292: PetscInt jj;
4293: #endif
4296: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4297: MatCreate(comm,mat);
4298: MatSetSizes(*mat,m,n,m,n);
4299: /* MatSetBlockSizes(*mat,,); */
4300: MatSetType(*mat,MATSEQAIJ);
4301: MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,0);
4302: aij = (Mat_SeqAIJ*)(*mat)->data;
4303: PetscMalloc2(m,&aij->imax,m,&aij->ilen);
4305: aij->i = i;
4306: aij->j = j;
4307: aij->a = a;
4308: aij->singlemalloc = PETSC_FALSE;
4309: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
4310: aij->free_a = PETSC_FALSE;
4311: aij->free_ij = PETSC_FALSE;
4313: for (ii=0; ii<m; ii++) {
4314: aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii];
4315: #if defined(PETSC_USE_DEBUG)
4316: 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]);
4317: for (jj=i[ii]+1; jj<i[ii+1]; jj++) {
4318: if (j[jj] < j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual colum %D) in row %D is not sorted",jj-i[ii],j[jj],ii);
4319: if (j[jj] == j[jj]-1) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual colum %D) in row %D is identical to previous entry",jj-i[ii],j[jj],ii);
4320: }
4321: #endif
4322: }
4323: #if defined(PETSC_USE_DEBUG)
4324: for (ii=0; ii<aij->i[m]; ii++) {
4325: if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]);
4326: 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]);
4327: }
4328: #endif
4330: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4331: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4332: return(0);
4333: }
4336: /*@C
4337: MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format)
4338: provided by the user.
4340: Collective on MPI_Comm
4342: Input Parameters:
4343: + comm - must be an MPI communicator of size 1
4344: . m - number of rows
4345: . n - number of columns
4346: . i - row indices
4347: . j - column indices
4348: . a - matrix values
4349: . nz - number of nonzeros
4350: - idx - 0 or 1 based
4352: Output Parameter:
4353: . mat - the matrix
4355: Level: intermediate
4357: Notes:
4358: The i and j indices are 0 based
4360: The format which is used for the sparse matrix input, is equivalent to a
4361: row-major ordering.. i.e for the following matrix, the input data expected is
4362: as shown:
4364: 1 0 0
4365: 2 0 3
4366: 4 5 6
4368: i = {0,1,1,2,2,2}
4369: j = {0,0,2,0,1,2}
4370: v = {1,2,3,4,5,6}
4373: .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR()
4375: @*/
4376: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt *i,PetscInt *j,PetscScalar *a,Mat *mat,PetscInt nz,PetscBool idx)
4377: {
4379: PetscInt ii, *nnz, one = 1,row,col;
4383: PetscCalloc1(m,&nnz);
4384: for (ii = 0; ii < nz; ii++) {
4385: nnz[i[ii] - !!idx] += 1;
4386: }
4387: MatCreate(comm,mat);
4388: MatSetSizes(*mat,m,n,m,n);
4389: MatSetType(*mat,MATSEQAIJ);
4390: MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);
4391: for (ii = 0; ii < nz; ii++) {
4392: if (idx) {
4393: row = i[ii] - 1;
4394: col = j[ii] - 1;
4395: } else {
4396: row = i[ii];
4397: col = j[ii];
4398: }
4399: MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);
4400: }
4401: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4402: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4403: PetscFree(nnz);
4404: return(0);
4405: }
4409: PetscErrorCode MatSetColoring_SeqAIJ(Mat A,ISColoring coloring)
4410: {
4412: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
4415: if (coloring->ctype == IS_COLORING_GLOBAL) {
4416: ISColoringReference(coloring);
4417: a->coloring = coloring;
4418: } else if (coloring->ctype == IS_COLORING_GHOSTED) {
4419: PetscInt i,*larray;
4420: ISColoring ocoloring;
4421: ISColoringValue *colors;
4423: /* set coloring for diagonal portion */
4424: PetscMalloc1(A->cmap->n,&larray);
4425: for (i=0; i<A->cmap->n; i++) larray[i] = i;
4426: ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,A->cmap->n,larray,NULL,larray);
4427: PetscMalloc1(A->cmap->n,&colors);
4428: for (i=0; i<A->cmap->n; i++) colors[i] = coloring->colors[larray[i]];
4429: PetscFree(larray);
4430: ISColoringCreate(PETSC_COMM_SELF,coloring->n,A->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);
4431: a->coloring = ocoloring;
4432: }
4433: return(0);
4434: }
4438: PetscErrorCode MatSetValuesAdifor_SeqAIJ(Mat A,PetscInt nl,void *advalues)
4439: {
4440: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
4441: PetscInt m = A->rmap->n,*ii = a->i,*jj = a->j,nz,i,j;
4442: MatScalar *v = a->a;
4443: PetscScalar *values = (PetscScalar*)advalues;
4444: ISColoringValue *color;
4447: if (!a->coloring) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Coloring not set for matrix");
4448: color = a->coloring->colors;
4449: /* loop over rows */
4450: for (i=0; i<m; i++) {
4451: nz = ii[i+1] - ii[i];
4452: /* loop over columns putting computed value into matrix */
4453: for (j=0; j<nz; j++) *v++ = values[color[*jj++]];
4454: values += nl; /* jump to next row of derivatives */
4455: }
4456: return(0);
4457: }
4461: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
4462: {
4463: Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data;
4467: a->idiagvalid = PETSC_FALSE;
4468: a->ibdiagvalid = PETSC_FALSE;
4470: MatSeqAIJInvalidateDiagonal_Inode(A);
4471: return(0);
4472: }
4476: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4477: {
4481: MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);
4482: return(0);
4483: }
4485: /*
4486: Permute A into C's *local* index space using rowemb,colemb.
4487: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
4488: of [0,m), colemb is in [0,n).
4489: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
4490: */
4493: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B)
4494: {
4495: /* If making this function public, change the error returned in this function away from _PLIB. */
4497: Mat_SeqAIJ *Baij;
4498: PetscBool seqaij;
4499: PetscInt m,n,*nz,i,j,count;
4500: PetscScalar v;
4501: const PetscInt *rowindices,*colindices;
4504: if (!B) return(0);
4505: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
4506: PetscObjectTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);
4507: if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type");
4508: if (rowemb) {
4509: ISGetLocalSize(rowemb,&m);
4510: 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);
4511: } else {
4512: if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix");
4513: }
4514: if (colemb) {
4515: ISGetLocalSize(colemb,&n);
4516: 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);
4517: } else {
4518: if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix");
4519: }
4521: Baij = (Mat_SeqAIJ*)(B->data);
4522: if (pattern == DIFFERENT_NONZERO_PATTERN) {
4523: PetscMalloc1(B->rmap->n,&nz);
4524: for (i=0; i<B->rmap->n; i++) {
4525: nz[i] = Baij->i[i+1] - Baij->i[i];
4526: }
4527: MatSeqAIJSetPreallocation(C,0,nz);
4528: PetscFree(nz);
4529: }
4530: if (pattern == SUBSET_NONZERO_PATTERN) {
4531: MatZeroEntries(C);
4532: }
4533: count = 0;
4534: rowindices = NULL;
4535: colindices = NULL;
4536: if (rowemb) {
4537: ISGetIndices(rowemb,&rowindices);
4538: }
4539: if (colemb) {
4540: ISGetIndices(colemb,&colindices);
4541: }
4542: for (i=0; i<B->rmap->n; i++) {
4543: PetscInt row;
4544: row = i;
4545: if (rowindices) row = rowindices[i];
4546: for (j=Baij->i[i]; j<Baij->i[i+1]; j++) {
4547: PetscInt col;
4548: col = Baij->j[count];
4549: if (colindices) col = colindices[col];
4550: v = Baij->a[count];
4551: MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);
4552: ++count;
4553: }
4554: }
4555: /* FIXME: set C's nonzerostate correctly. */
4556: /* Assembly for C is necessary. */
4557: C->preallocated = PETSC_TRUE;
4558: C->assembled = PETSC_TRUE;
4559: C->was_assembled = PETSC_FALSE;
4560: return(0);
4561: }
4564: /*
4565: Special version for direct calls from Fortran
4566: */
4567: #include <petsc/private/fortranimpl.h>
4568: #if defined(PETSC_HAVE_FORTRAN_CAPS)
4569: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
4570: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
4571: #define matsetvaluesseqaij_ matsetvaluesseqaij
4572: #endif
4574: /* Change these macros so can be used in void function */
4575: #undef CHKERRQ
4576: #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr)
4577: #undef SETERRQ2
4578: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
4579: #undef SETERRQ3
4580: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
4584: PETSC_EXTERN void PETSC_STDCALL matsetvaluesseqaij_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[],InsertMode *isis, PetscErrorCode *_ierr)
4585: {
4586: Mat A = *AA;
4587: PetscInt m = *mm, n = *nn;
4588: InsertMode is = *isis;
4589: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
4590: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N;
4591: PetscInt *imax,*ai,*ailen;
4593: PetscInt *aj,nonew = a->nonew,lastcol = -1;
4594: MatScalar *ap,value,*aa;
4595: PetscBool ignorezeroentries = a->ignorezeroentries;
4596: PetscBool roworiented = a->roworiented;
4599: MatCheckPreallocated(A,1);
4600: imax = a->imax;
4601: ai = a->i;
4602: ailen = a->ilen;
4603: aj = a->j;
4604: aa = a->a;
4606: for (k=0; k<m; k++) { /* loop over added rows */
4607: row = im[k];
4608: if (row < 0) continue;
4609: #if defined(PETSC_USE_DEBUG)
4610: if (row >= A->rmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large");
4611: #endif
4612: rp = aj + ai[row]; ap = aa + ai[row];
4613: rmax = imax[row]; nrow = ailen[row];
4614: low = 0;
4615: high = nrow;
4616: for (l=0; l<n; l++) { /* loop over added columns */
4617: if (in[l] < 0) continue;
4618: #if defined(PETSC_USE_DEBUG)
4619: if (in[l] >= A->cmap->n) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large");
4620: #endif
4621: col = in[l];
4622: if (roworiented) value = v[l + k*n];
4623: else value = v[k + l*m];
4625: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
4627: if (col <= lastcol) low = 0;
4628: else high = nrow;
4629: lastcol = col;
4630: while (high-low > 5) {
4631: t = (low+high)/2;
4632: if (rp[t] > col) high = t;
4633: else low = t;
4634: }
4635: for (i=low; i<high; i++) {
4636: if (rp[i] > col) break;
4637: if (rp[i] == col) {
4638: if (is == ADD_VALUES) ap[i] += value;
4639: else ap[i] = value;
4640: goto noinsert;
4641: }
4642: }
4643: if (value == 0.0 && ignorezeroentries) goto noinsert;
4644: if (nonew == 1) goto noinsert;
4645: if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix");
4646: MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar);
4647: N = nrow++ - 1; a->nz++; high++;
4648: /* shift up all the later entries in this row */
4649: for (ii=N; ii>=i; ii--) {
4650: rp[ii+1] = rp[ii];
4651: ap[ii+1] = ap[ii];
4652: }
4653: rp[i] = col;
4654: ap[i] = value;
4655: A->nonzerostate++;
4656: noinsert:;
4657: low = i + 1;
4658: }
4659: ailen[row] = nrow;
4660: }
4661: PetscFunctionReturnVoid();
4662: }