Actual source code: blockmat.c
2: /*
3: This provides a matrix that consists of Mats
4: */
6: #include <petsc/private/matimpl.h>
7: #include <../src/mat/impls/baij/seq/baij.h>
9: typedef struct {
10: SEQAIJHEADER(Mat);
11: SEQBAIJHEADER;
12: Mat *diags;
14: Vec left,right,middle,workb; /* dummy vectors to perform local parts of product */
15: } Mat_BlockMat;
17: static PetscErrorCode MatSOR_BlockMat_Symmetric(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
18: {
19: Mat_BlockMat *a = (Mat_BlockMat*)A->data;
20: PetscScalar *x;
21: const Mat *v;
22: const PetscScalar *b;
23: PetscErrorCode ierr;
24: PetscInt n = A->cmap->n,i,mbs = n/A->rmap->bs,j,bs = A->rmap->bs;
25: const PetscInt *idx;
26: IS row,col;
27: MatFactorInfo info;
28: Vec left = a->left,right = a->right, middle = a->middle;
29: Mat *diag;
32: its = its*lits;
33: if (its <= 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
34: if (flag & SOR_EISENSTAT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for Eisenstat");
35: if (omega != 1.0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for omega not equal to 1.0");
36: if (fshift) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for fshift");
37: if ((flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) && !(flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP)) {
38: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Cannot do backward sweep without forward sweep");
39: }
41: if (!a->diags) {
42: PetscMalloc1(mbs,&a->diags);
43: MatFactorInfoInitialize(&info);
44: for (i=0; i<mbs; i++) {
45: MatGetOrdering(a->a[a->diag[i]], MATORDERINGND,&row,&col);
46: MatCholeskyFactorSymbolic(a->diags[i],a->a[a->diag[i]],row,&info);
47: MatCholeskyFactorNumeric(a->diags[i],a->a[a->diag[i]],&info);
48: ISDestroy(&row);
49: ISDestroy(&col);
50: }
51: VecDuplicate(bb,&a->workb);
52: }
53: diag = a->diags;
55: VecSet(xx,0.0);
56: VecGetArray(xx,&x);
57: /* copy right hand side because it must be modified during iteration */
58: VecCopy(bb,a->workb);
59: VecGetArrayRead(a->workb,&b);
61: /* need to add code for when initial guess is zero, see MatSOR_SeqAIJ */
62: while (its--) {
63: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
65: for (i=0; i<mbs; i++) {
66: n = a->i[i+1] - a->i[i] - 1;
67: idx = a->j + a->i[i] + 1;
68: v = a->a + a->i[i] + 1;
70: VecSet(left,0.0);
71: for (j=0; j<n; j++) {
72: VecPlaceArray(right,x + idx[j]*bs);
73: MatMultAdd(v[j],right,left,left);
74: VecResetArray(right);
75: }
76: VecPlaceArray(right,b + i*bs);
77: VecAYPX(left,-1.0,right);
78: VecResetArray(right);
80: VecPlaceArray(right,x + i*bs);
81: MatSolve(diag[i],left,right);
83: /* now adjust right hand side, see MatSOR_SeqSBAIJ */
84: for (j=0; j<n; j++) {
85: MatMultTranspose(v[j],right,left);
86: VecPlaceArray(middle,b + idx[j]*bs);
87: VecAXPY(middle,-1.0,left);
88: VecResetArray(middle);
89: }
90: VecResetArray(right);
92: }
93: }
94: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
96: for (i=mbs-1; i>=0; i--) {
97: n = a->i[i+1] - a->i[i] - 1;
98: idx = a->j + a->i[i] + 1;
99: v = a->a + a->i[i] + 1;
101: VecSet(left,0.0);
102: for (j=0; j<n; j++) {
103: VecPlaceArray(right,x + idx[j]*bs);
104: MatMultAdd(v[j],right,left,left);
105: VecResetArray(right);
106: }
107: VecPlaceArray(right,b + i*bs);
108: VecAYPX(left,-1.0,right);
109: VecResetArray(right);
111: VecPlaceArray(right,x + i*bs);
112: MatSolve(diag[i],left,right);
113: VecResetArray(right);
115: }
116: }
117: }
118: VecRestoreArray(xx,&x);
119: VecRestoreArrayRead(a->workb,&b);
120: return(0);
121: }
123: static PetscErrorCode MatSOR_BlockMat(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
124: {
125: Mat_BlockMat *a = (Mat_BlockMat*)A->data;
126: PetscScalar *x;
127: const Mat *v;
128: const PetscScalar *b;
129: PetscErrorCode ierr;
130: PetscInt n = A->cmap->n,i,mbs = n/A->rmap->bs,j,bs = A->rmap->bs;
131: const PetscInt *idx;
132: IS row,col;
133: MatFactorInfo info;
134: Vec left = a->left,right = a->right;
135: Mat *diag;
138: its = its*lits;
139: if (its <= 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);
140: if (flag & SOR_EISENSTAT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for Eisenstat");
141: if (omega != 1.0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for omega not equal to 1.0");
142: if (fshift) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support yet for fshift");
144: if (!a->diags) {
145: PetscMalloc1(mbs,&a->diags);
146: MatFactorInfoInitialize(&info);
147: for (i=0; i<mbs; i++) {
148: MatGetOrdering(a->a[a->diag[i]], MATORDERINGND,&row,&col);
149: MatLUFactorSymbolic(a->diags[i],a->a[a->diag[i]],row,col,&info);
150: MatLUFactorNumeric(a->diags[i],a->a[a->diag[i]],&info);
151: ISDestroy(&row);
152: ISDestroy(&col);
153: }
154: }
155: diag = a->diags;
157: VecSet(xx,0.0);
158: VecGetArray(xx,&x);
159: VecGetArrayRead(bb,&b);
161: /* need to add code for when initial guess is zero, see MatSOR_SeqAIJ */
162: while (its--) {
163: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
165: for (i=0; i<mbs; i++) {
166: n = a->i[i+1] - a->i[i];
167: idx = a->j + a->i[i];
168: v = a->a + a->i[i];
170: VecSet(left,0.0);
171: for (j=0; j<n; j++) {
172: if (idx[j] != i) {
173: VecPlaceArray(right,x + idx[j]*bs);
174: MatMultAdd(v[j],right,left,left);
175: VecResetArray(right);
176: }
177: }
178: VecPlaceArray(right,b + i*bs);
179: VecAYPX(left,-1.0,right);
180: VecResetArray(right);
182: VecPlaceArray(right,x + i*bs);
183: MatSolve(diag[i],left,right);
184: VecResetArray(right);
185: }
186: }
187: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
189: for (i=mbs-1; i>=0; i--) {
190: n = a->i[i+1] - a->i[i];
191: idx = a->j + a->i[i];
192: v = a->a + a->i[i];
194: VecSet(left,0.0);
195: for (j=0; j<n; j++) {
196: if (idx[j] != i) {
197: VecPlaceArray(right,x + idx[j]*bs);
198: MatMultAdd(v[j],right,left,left);
199: VecResetArray(right);
200: }
201: }
202: VecPlaceArray(right,b + i*bs);
203: VecAYPX(left,-1.0,right);
204: VecResetArray(right);
206: VecPlaceArray(right,x + i*bs);
207: MatSolve(diag[i],left,right);
208: VecResetArray(right);
210: }
211: }
212: }
213: VecRestoreArray(xx,&x);
214: VecRestoreArrayRead(bb,&b);
215: return(0);
216: }
218: static PetscErrorCode MatSetValues_BlockMat(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is)
219: {
220: Mat_BlockMat *a = (Mat_BlockMat*)A->data;
221: PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N,lastcol = -1;
222: PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen;
223: PetscInt *aj =a->j,nonew=a->nonew,bs=A->rmap->bs,brow,bcol;
225: PetscInt ridx,cidx;
226: PetscBool roworiented=a->roworiented;
227: MatScalar value;
228: Mat *ap,*aa = a->a;
231: for (k=0; k<m; k++) { /* loop over added rows */
232: row = im[k];
233: brow = row/bs;
234: if (row < 0) continue;
235: if (PetscUnlikelyDebug(row >= A->rmap->N)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->N-1);
236: rp = aj + ai[brow];
237: ap = aa + ai[brow];
238: rmax = imax[brow];
239: nrow = ailen[brow];
240: low = 0;
241: high = nrow;
242: for (l=0; l<n; l++) { /* loop over added columns */
243: if (in[l] < 0) continue;
244: if (PetscUnlikelyDebug(in[l] >= A->cmap->n)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1);
245: col = in[l]; bcol = col/bs;
246: if (A->symmetric && brow > bcol) continue;
247: ridx = row % bs; cidx = col % bs;
248: if (roworiented) value = v[l + k*n];
249: else value = v[k + l*m];
251: if (col <= lastcol) low = 0;
252: else high = nrow;
253: lastcol = col;
254: while (high-low > 7) {
255: t = (low+high)/2;
256: if (rp[t] > bcol) high = t;
257: else low = t;
258: }
259: for (i=low; i<high; i++) {
260: if (rp[i] > bcol) break;
261: if (rp[i] == bcol) goto noinsert1;
262: }
263: if (nonew == 1) goto noinsert1;
264: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) in the matrix", row, col);
265: MatSeqXAIJReallocateAIJ(A,a->mbs,1,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,imax,nonew,Mat);
266: N = nrow++ - 1; high++;
267: /* shift up all the later entries in this row */
268: for (ii=N; ii>=i; ii--) {
269: rp[ii+1] = rp[ii];
270: ap[ii+1] = ap[ii];
271: }
272: if (N>=i) ap[i] = NULL;
273: rp[i] = bcol;
274: a->nz++;
275: A->nonzerostate++;
276: noinsert1:;
277: if (!*(ap+i)) {
278: MatCreateSeqAIJ(PETSC_COMM_SELF,bs,bs,0,NULL,ap+i);
279: }
280: MatSetValues(ap[i],1,&ridx,1,&cidx,&value,is);
281: low = i;
282: }
283: ailen[brow] = nrow;
284: }
285: return(0);
286: }
288: static PetscErrorCode MatLoad_BlockMat(Mat newmat, PetscViewer viewer)
289: {
290: PetscErrorCode ierr;
291: Mat tmpA;
292: PetscInt i,j,m,n,bs = 1,ncols,*lens,currentcol,mbs,**ii,*ilens,nextcol,*llens,cnt = 0;
293: const PetscInt *cols;
294: const PetscScalar *values;
295: PetscBool flg = PETSC_FALSE,notdone;
296: Mat_SeqAIJ *a;
297: Mat_BlockMat *amat;
300: /* force binary viewer to load .info file if it has not yet done so */
301: PetscViewerSetUp(viewer);
302: MatCreate(PETSC_COMM_SELF,&tmpA);
303: MatSetType(tmpA,MATSEQAIJ);
304: MatLoad_SeqAIJ(tmpA,viewer);
306: MatGetLocalSize(tmpA,&m,&n);
307: PetscOptionsBegin(PETSC_COMM_SELF,NULL,"Options for loading BlockMat matrix 1","Mat");
308: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
309: PetscOptionsBool("-matload_symmetric","Store the matrix as symmetric","MatLoad",flg,&flg,NULL);
310: PetscOptionsEnd();
312: /* Determine number of nonzero blocks for each block row */
313: a = (Mat_SeqAIJ*) tmpA->data;
314: mbs = m/bs;
315: PetscMalloc3(mbs,&lens,bs,&ii,bs,&ilens);
316: PetscArrayzero(lens,mbs);
318: for (i=0; i<mbs; i++) {
319: for (j=0; j<bs; j++) {
320: ii[j] = a->j + a->i[i*bs + j];
321: ilens[j] = a->i[i*bs + j + 1] - a->i[i*bs + j];
322: }
324: currentcol = -1;
325: while (PETSC_TRUE) {
326: notdone = PETSC_FALSE;
327: nextcol = 1000000000;
328: for (j=0; j<bs; j++) {
329: while ((ilens[j] > 0 && ii[j][0]/bs <= currentcol)) {
330: ii[j]++;
331: ilens[j]--;
332: }
333: if (ilens[j] > 0) {
334: notdone = PETSC_TRUE;
335: nextcol = PetscMin(nextcol,ii[j][0]/bs);
336: }
337: }
338: if (!notdone) break;
339: if (!flg || (nextcol >= i)) lens[i]++;
340: currentcol = nextcol;
341: }
342: }
344: if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) {
345: MatSetSizes(newmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
346: }
347: MatBlockMatSetPreallocation(newmat,bs,0,lens);
348: if (flg) {
349: MatSetOption(newmat,MAT_SYMMETRIC,PETSC_TRUE);
350: }
351: amat = (Mat_BlockMat*)(newmat)->data;
353: /* preallocate the submatrices */
354: PetscMalloc1(bs,&llens);
355: for (i=0; i<mbs; i++) { /* loops for block rows */
356: for (j=0; j<bs; j++) {
357: ii[j] = a->j + a->i[i*bs + j];
358: ilens[j] = a->i[i*bs + j + 1] - a->i[i*bs + j];
359: }
361: currentcol = 1000000000;
362: for (j=0; j<bs; j++) { /* loop over rows in block finding first nonzero block */
363: if (ilens[j] > 0) {
364: currentcol = PetscMin(currentcol,ii[j][0]/bs);
365: }
366: }
368: while (PETSC_TRUE) { /* loops over blocks in block row */
369: notdone = PETSC_FALSE;
370: nextcol = 1000000000;
371: PetscArrayzero(llens,bs);
372: for (j=0; j<bs; j++) { /* loop over rows in block */
373: while ((ilens[j] > 0 && ii[j][0]/bs <= currentcol)) { /* loop over columns in row */
374: ii[j]++;
375: ilens[j]--;
376: llens[j]++;
377: }
378: if (ilens[j] > 0) {
379: notdone = PETSC_TRUE;
380: nextcol = PetscMin(nextcol,ii[j][0]/bs);
381: }
382: }
383: if (cnt >= amat->maxnz) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Number of blocks found greater than expected %D",cnt);
384: if (!flg || currentcol >= i) {
385: amat->j[cnt] = currentcol;
386: MatCreateSeqAIJ(PETSC_COMM_SELF,bs,bs,0,llens,amat->a+cnt++);
387: }
389: if (!notdone) break;
390: currentcol = nextcol;
391: }
392: amat->ilen[i] = lens[i];
393: }
395: PetscFree3(lens,ii,ilens);
396: PetscFree(llens);
398: /* copy over the matrix, one row at a time */
399: for (i=0; i<m; i++) {
400: MatGetRow(tmpA,i,&ncols,&cols,&values);
401: MatSetValues(newmat,1,&i,ncols,cols,values,INSERT_VALUES);
402: MatRestoreRow(tmpA,i,&ncols,&cols,&values);
403: }
404: MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);
405: MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);
406: return(0);
407: }
409: static PetscErrorCode MatView_BlockMat(Mat A,PetscViewer viewer)
410: {
411: Mat_BlockMat *a = (Mat_BlockMat*)A->data;
412: PetscErrorCode ierr;
413: const char *name;
414: PetscViewerFormat format;
417: PetscObjectGetName((PetscObject)A,&name);
418: PetscViewerGetFormat(viewer,&format);
419: if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO) {
420: PetscViewerASCIIPrintf(viewer,"Nonzero block matrices = %D \n",a->nz);
421: if (A->symmetric) {
422: PetscViewerASCIIPrintf(viewer,"Only upper triangular part of symmetric matrix is stored\n");
423: }
424: }
425: return(0);
426: }
428: static PetscErrorCode MatDestroy_BlockMat(Mat mat)
429: {
431: Mat_BlockMat *bmat = (Mat_BlockMat*)mat->data;
432: PetscInt i;
435: VecDestroy(&bmat->right);
436: VecDestroy(&bmat->left);
437: VecDestroy(&bmat->middle);
438: VecDestroy(&bmat->workb);
439: if (bmat->diags) {
440: for (i=0; i<mat->rmap->n/mat->rmap->bs; i++) {
441: MatDestroy(&bmat->diags[i]);
442: }
443: }
444: if (bmat->a) {
445: for (i=0; i<bmat->nz; i++) {
446: MatDestroy(&bmat->a[i]);
447: }
448: }
449: MatSeqXAIJFreeAIJ(mat,(PetscScalar**)&bmat->a,&bmat->j,&bmat->i);
450: PetscFree(mat->data);
451: return(0);
452: }
454: static PetscErrorCode MatMult_BlockMat(Mat A,Vec x,Vec y)
455: {
456: Mat_BlockMat *bmat = (Mat_BlockMat*)A->data;
458: PetscScalar *xx,*yy;
459: PetscInt *aj,i,*ii,jrow,m = A->rmap->n/A->rmap->bs,bs = A->rmap->bs,n,j;
460: Mat *aa;
463: /*
464: Standard CSR multiply except each entry is a Mat
465: */
466: VecGetArray(x,&xx);
468: VecSet(y,0.0);
469: VecGetArray(y,&yy);
470: aj = bmat->j;
471: aa = bmat->a;
472: ii = bmat->i;
473: for (i=0; i<m; i++) {
474: jrow = ii[i];
475: VecPlaceArray(bmat->left,yy + bs*i);
476: n = ii[i+1] - jrow;
477: for (j=0; j<n; j++) {
478: VecPlaceArray(bmat->right,xx + bs*aj[jrow]);
479: MatMultAdd(aa[jrow],bmat->right,bmat->left,bmat->left);
480: VecResetArray(bmat->right);
481: jrow++;
482: }
483: VecResetArray(bmat->left);
484: }
485: VecRestoreArray(x,&xx);
486: VecRestoreArray(y,&yy);
487: return(0);
488: }
490: PetscErrorCode MatMult_BlockMat_Symmetric(Mat A,Vec x,Vec y)
491: {
492: Mat_BlockMat *bmat = (Mat_BlockMat*)A->data;
494: PetscScalar *xx,*yy;
495: PetscInt *aj,i,*ii,jrow,m = A->rmap->n/A->rmap->bs,bs = A->rmap->bs,n,j;
496: Mat *aa;
499: /*
500: Standard CSR multiply except each entry is a Mat
501: */
502: VecGetArray(x,&xx);
504: VecSet(y,0.0);
505: VecGetArray(y,&yy);
506: aj = bmat->j;
507: aa = bmat->a;
508: ii = bmat->i;
509: for (i=0; i<m; i++) {
510: jrow = ii[i];
511: n = ii[i+1] - jrow;
512: VecPlaceArray(bmat->left,yy + bs*i);
513: VecPlaceArray(bmat->middle,xx + bs*i);
514: /* if we ALWAYS required a diagonal entry then could remove this if test */
515: if (aj[jrow] == i) {
516: VecPlaceArray(bmat->right,xx + bs*aj[jrow]);
517: MatMultAdd(aa[jrow],bmat->right,bmat->left,bmat->left);
518: VecResetArray(bmat->right);
519: jrow++;
520: n--;
521: }
522: for (j=0; j<n; j++) {
523: VecPlaceArray(bmat->right,xx + bs*aj[jrow]); /* upper triangular part */
524: MatMultAdd(aa[jrow],bmat->right,bmat->left,bmat->left);
525: VecResetArray(bmat->right);
527: VecPlaceArray(bmat->right,yy + bs*aj[jrow]); /* lower triangular part */
528: MatMultTransposeAdd(aa[jrow],bmat->middle,bmat->right,bmat->right);
529: VecResetArray(bmat->right);
530: jrow++;
531: }
532: VecResetArray(bmat->left);
533: VecResetArray(bmat->middle);
534: }
535: VecRestoreArray(x,&xx);
536: VecRestoreArray(y,&yy);
537: return(0);
538: }
540: static PetscErrorCode MatMultAdd_BlockMat(Mat A,Vec x,Vec y,Vec z)
541: {
543: return(0);
544: }
546: static PetscErrorCode MatMultTranspose_BlockMat(Mat A,Vec x,Vec y)
547: {
549: return(0);
550: }
552: static PetscErrorCode MatMultTransposeAdd_BlockMat(Mat A,Vec x,Vec y,Vec z)
553: {
555: return(0);
556: }
558: /*
559: Adds diagonal pointers to sparse matrix structure.
560: */
561: static PetscErrorCode MatMarkDiagonal_BlockMat(Mat A)
562: {
563: Mat_BlockMat *a = (Mat_BlockMat*)A->data;
565: PetscInt i,j,mbs = A->rmap->n/A->rmap->bs;
568: if (!a->diag) {
569: PetscMalloc1(mbs,&a->diag);
570: }
571: for (i=0; i<mbs; i++) {
572: a->diag[i] = a->i[i+1];
573: for (j=a->i[i]; j<a->i[i+1]; j++) {
574: if (a->j[j] == i) {
575: a->diag[i] = j;
576: break;
577: }
578: }
579: }
580: return(0);
581: }
583: static PetscErrorCode MatCreateSubMatrix_BlockMat(Mat A,IS isrow,IS iscol,MatReuse scall,Mat *B)
584: {
585: Mat_BlockMat *a = (Mat_BlockMat*)A->data;
586: Mat_SeqAIJ *c;
588: PetscInt i,k,first,step,lensi,nrows,ncols;
589: PetscInt *j_new,*i_new,*aj = a->j,*ailen = a->ilen;
590: PetscScalar *a_new;
591: Mat C,*aa = a->a;
592: PetscBool stride,equal;
595: ISEqual(isrow,iscol,&equal);
596: if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only for identical column and row indices");
597: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);
598: if (!stride) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only for stride indices");
599: ISStrideGetInfo(iscol,&first,&step);
600: if (step != A->rmap->bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Can only select one entry from each block");
602: ISGetLocalSize(isrow,&nrows);
603: ncols = nrows;
605: /* create submatrix */
606: if (scall == MAT_REUSE_MATRIX) {
607: PetscInt n_cols,n_rows;
608: C = *B;
609: MatGetSize(C,&n_rows,&n_cols);
610: if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size");
611: MatZeroEntries(C);
612: } else {
613: MatCreate(PetscObjectComm((PetscObject)A),&C);
614: MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);
615: if (A->symmetric) {
616: MatSetType(C,MATSEQSBAIJ);
617: } else {
618: MatSetType(C,MATSEQAIJ);
619: }
620: MatSeqAIJSetPreallocation(C,0,ailen);
621: MatSeqSBAIJSetPreallocation(C,1,0,ailen);
622: }
623: c = (Mat_SeqAIJ*)C->data;
625: /* loop over rows inserting into submatrix */
626: a_new = c->a;
627: j_new = c->j;
628: i_new = c->i;
630: for (i=0; i<nrows; i++) {
631: lensi = ailen[i];
632: for (k=0; k<lensi; k++) {
633: *j_new++ = *aj++;
634: MatGetValue(*aa++,first,first,a_new++);
635: }
636: i_new[i+1] = i_new[i] + lensi;
637: c->ilen[i] = lensi;
638: }
640: MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
641: MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
642: *B = C;
643: return(0);
644: }
646: static PetscErrorCode MatAssemblyEnd_BlockMat(Mat A,MatAssemblyType mode)
647: {
648: Mat_BlockMat *a = (Mat_BlockMat*)A->data;
650: PetscInt fshift = 0,i,j,*ai = a->i,*aj = a->j,*imax = a->imax;
651: PetscInt m = a->mbs,*ip,N,*ailen = a->ilen,rmax = 0;
652: Mat *aa = a->a,*ap;
655: if (mode == MAT_FLUSH_ASSEMBLY) return(0);
657: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
658: for (i=1; i<m; i++) {
659: /* move each row back by the amount of empty slots (fshift) before it*/
660: fshift += imax[i-1] - ailen[i-1];
661: rmax = PetscMax(rmax,ailen[i]);
662: if (fshift) {
663: ip = aj + ai[i];
664: ap = aa + ai[i];
665: N = ailen[i];
666: for (j=0; j<N; j++) {
667: ip[j-fshift] = ip[j];
668: ap[j-fshift] = ap[j];
669: }
670: }
671: ai[i] = ai[i-1] + ailen[i-1];
672: }
673: if (m) {
674: fshift += imax[m-1] - ailen[m-1];
675: ai[m] = ai[m-1] + ailen[m-1];
676: }
677: /* reset ilen and imax for each row */
678: for (i=0; i<m; i++) {
679: ailen[i] = imax[i] = ai[i+1] - ai[i];
680: }
681: a->nz = ai[m];
682: for (i=0; i<a->nz; i++) {
683: if (PetscUnlikelyDebug(!aa[i])) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Null matrix at location %D column %D nz %D",i,aj[i],a->nz);
684: MatAssemblyBegin(aa[i],MAT_FINAL_ASSEMBLY);
685: MatAssemblyEnd(aa[i],MAT_FINAL_ASSEMBLY);
686: }
687: PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap->n/A->cmap->bs,fshift,a->nz);
688: PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);
689: PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);
691: A->info.mallocs += a->reallocs;
692: a->reallocs = 0;
693: A->info.nz_unneeded = (double)fshift;
694: a->rmax = rmax;
695: MatMarkDiagonal_BlockMat(A);
696: return(0);
697: }
699: static PetscErrorCode MatSetOption_BlockMat(Mat A,MatOption opt,PetscBool flg)
700: {
703: if (opt == MAT_SYMMETRIC && flg) {
704: A->ops->sor = MatSOR_BlockMat_Symmetric;
705: A->ops->mult = MatMult_BlockMat_Symmetric;
706: } else {
707: PetscInfo1(A,"Unused matrix option %s\n",MatOptions[opt]);
708: }
709: return(0);
710: }
712: static struct _MatOps MatOps_Values = {MatSetValues_BlockMat,
713: NULL,
714: NULL,
715: MatMult_BlockMat,
716: /* 4*/ MatMultAdd_BlockMat,
717: MatMultTranspose_BlockMat,
718: MatMultTransposeAdd_BlockMat,
719: NULL,
720: NULL,
721: NULL,
722: /* 10*/ NULL,
723: NULL,
724: NULL,
725: MatSOR_BlockMat,
726: NULL,
727: /* 15*/ NULL,
728: NULL,
729: NULL,
730: NULL,
731: NULL,
732: /* 20*/ NULL,
733: MatAssemblyEnd_BlockMat,
734: MatSetOption_BlockMat,
735: NULL,
736: /* 24*/ NULL,
737: NULL,
738: NULL,
739: NULL,
740: NULL,
741: /* 29*/ NULL,
742: NULL,
743: NULL,
744: NULL,
745: NULL,
746: /* 34*/ NULL,
747: NULL,
748: NULL,
749: NULL,
750: NULL,
751: /* 39*/ NULL,
752: NULL,
753: NULL,
754: NULL,
755: NULL,
756: /* 44*/ NULL,
757: NULL,
758: MatShift_Basic,
759: NULL,
760: NULL,
761: /* 49*/ NULL,
762: NULL,
763: NULL,
764: NULL,
765: NULL,
766: /* 54*/ NULL,
767: NULL,
768: NULL,
769: NULL,
770: NULL,
771: /* 59*/ MatCreateSubMatrix_BlockMat,
772: MatDestroy_BlockMat,
773: MatView_BlockMat,
774: NULL,
775: NULL,
776: /* 64*/ NULL,
777: NULL,
778: NULL,
779: NULL,
780: NULL,
781: /* 69*/ NULL,
782: NULL,
783: NULL,
784: NULL,
785: NULL,
786: /* 74*/ NULL,
787: NULL,
788: NULL,
789: NULL,
790: NULL,
791: /* 79*/ NULL,
792: NULL,
793: NULL,
794: NULL,
795: MatLoad_BlockMat,
796: /* 84*/ NULL,
797: NULL,
798: NULL,
799: NULL,
800: NULL,
801: /* 89*/ NULL,
802: NULL,
803: NULL,
804: NULL,
805: NULL,
806: /* 94*/ NULL,
807: NULL,
808: NULL,
809: NULL,
810: NULL,
811: /* 99*/ NULL,
812: NULL,
813: NULL,
814: NULL,
815: NULL,
816: /*104*/ NULL,
817: NULL,
818: NULL,
819: NULL,
820: NULL,
821: /*109*/ NULL,
822: NULL,
823: NULL,
824: NULL,
825: NULL,
826: /*114*/ NULL,
827: NULL,
828: NULL,
829: NULL,
830: NULL,
831: /*119*/ NULL,
832: NULL,
833: NULL,
834: NULL,
835: NULL,
836: /*124*/ NULL,
837: NULL,
838: NULL,
839: NULL,
840: NULL,
841: /*129*/ NULL,
842: NULL,
843: NULL,
844: NULL,
845: NULL,
846: /*134*/ NULL,
847: NULL,
848: NULL,
849: NULL,
850: NULL,
851: /*139*/ NULL,
852: NULL,
853: NULL
854: };
856: /*@C
857: MatBlockMatSetPreallocation - For good matrix assembly performance
858: the user should preallocate the matrix storage by setting the parameter nz
859: (or the array nnz). By setting these parameters accurately, performance
860: during matrix assembly can be increased by more than a factor of 50.
862: Collective
864: Input Parameters:
865: + B - The matrix
866: . bs - size of each block in matrix
867: . nz - number of nonzeros per block row (same for all rows)
868: - nnz - array containing the number of nonzeros in the various block rows
869: (possibly different for each row) or NULL
871: Notes:
872: If nnz is given then nz is ignored
874: Specify the preallocated storage with either nz or nnz (not both).
875: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
876: allocation. For large problems you MUST preallocate memory or you
877: will get TERRIBLE performance, see the users' manual chapter on matrices.
879: Level: intermediate
881: .seealso: MatCreate(), MatCreateBlockMat(), MatSetValues()
883: @*/
884: PetscErrorCode MatBlockMatSetPreallocation(Mat B,PetscInt bs,PetscInt nz,const PetscInt nnz[])
885: {
889: PetscTryMethod(B,"MatBlockMatSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[]),(B,bs,nz,nnz));
890: return(0);
891: }
893: static PetscErrorCode MatBlockMatSetPreallocation_BlockMat(Mat A,PetscInt bs,PetscInt nz,PetscInt *nnz)
894: {
895: Mat_BlockMat *bmat = (Mat_BlockMat*)A->data;
897: PetscInt i;
900: PetscLayoutSetBlockSize(A->rmap,bs);
901: PetscLayoutSetBlockSize(A->cmap,bs);
902: PetscLayoutSetUp(A->rmap);
903: PetscLayoutSetUp(A->cmap);
904: PetscLayoutGetBlockSize(A->rmap,&bs);
906: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
907: if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %d",nz);
908: if (nnz) {
909: for (i=0; i<A->rmap->n/bs; i++) {
910: 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]);
911: if (nnz[i] > A->cmap->n/bs) 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],A->cmap->n/bs);
912: }
913: }
914: bmat->mbs = A->rmap->n/bs;
916: VecCreateSeqWithArray(PETSC_COMM_SELF,1,bs,NULL,&bmat->right);
917: VecCreateSeqWithArray(PETSC_COMM_SELF,1,bs,NULL,&bmat->middle);
918: VecCreateSeq(PETSC_COMM_SELF,bs,&bmat->left);
920: if (!bmat->imax) {
921: PetscMalloc2(A->rmap->n,&bmat->imax,A->rmap->n,&bmat->ilen);
922: PetscLogObjectMemory((PetscObject)A,2*A->rmap->n*sizeof(PetscInt));
923: }
924: if (nnz) {
925: nz = 0;
926: for (i=0; i<A->rmap->n/A->rmap->bs; i++) {
927: bmat->imax[i] = nnz[i];
928: nz += nnz[i];
929: }
930: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Currently requires block row by row preallocation");
932: /* bmat->ilen will count nonzeros in each row so far. */
933: for (i=0; i<bmat->mbs; i++) bmat->ilen[i] = 0;
935: /* allocate the matrix space */
936: MatSeqXAIJFreeAIJ(A,(PetscScalar**)&bmat->a,&bmat->j,&bmat->i);
937: PetscMalloc3(nz,&bmat->a,nz,&bmat->j,A->rmap->n+1,&bmat->i);
938: PetscLogObjectMemory((PetscObject)A,(A->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));
939: bmat->i[0] = 0;
940: for (i=1; i<bmat->mbs+1; i++) {
941: bmat->i[i] = bmat->i[i-1] + bmat->imax[i-1];
942: }
943: bmat->singlemalloc = PETSC_TRUE;
944: bmat->free_a = PETSC_TRUE;
945: bmat->free_ij = PETSC_TRUE;
947: bmat->nz = 0;
948: bmat->maxnz = nz;
949: A->info.nz_unneeded = (double)bmat->maxnz;
950: MatSetOption(A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
951: return(0);
952: }
954: /*MC
955: MATBLOCKMAT - A matrix that is defined by a set of Mat's that represents a sparse block matrix
956: consisting of (usually) sparse blocks.
958: Level: advanced
960: .seealso: MatCreateBlockMat()
962: M*/
964: PETSC_EXTERN PetscErrorCode MatCreate_BlockMat(Mat A)
965: {
966: Mat_BlockMat *b;
970: PetscNewLog(A,&b);
971: A->data = (void*)b;
972: PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));
974: A->assembled = PETSC_TRUE;
975: A->preallocated = PETSC_FALSE;
976: PetscObjectChangeTypeName((PetscObject)A,MATBLOCKMAT);
978: PetscObjectComposeFunction((PetscObject)A,"MatBlockMatSetPreallocation_C",MatBlockMatSetPreallocation_BlockMat);
979: return(0);
980: }
982: /*@C
983: MatCreateBlockMat - Creates a new matrix in which each block contains a uniform-size sequential Mat object
985: Collective
987: Input Parameters:
988: + comm - MPI communicator
989: . m - number of rows
990: . n - number of columns
991: . bs - size of each submatrix
992: . nz - expected maximum number of nonzero blocks in row (use PETSC_DEFAULT if not known)
993: - nnz - expected number of nonzers per block row if known (use NULL otherwise)
995: Output Parameter:
996: . A - the matrix
998: Level: intermediate
1000: Notes:
1001: Matrices of this type are nominally-sparse matrices in which each "entry" is a Mat object. Each Mat must
1002: have the same size and be sequential. The local and global sizes must be compatible with this decomposition.
1004: For matrices containing parallel submatrices and variable block sizes, see MATNEST.
1006: .seealso: MATBLOCKMAT, MatCreateNest()
1007: @*/
1008: PetscErrorCode MatCreateBlockMat(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt bs,PetscInt nz,PetscInt *nnz, Mat *A)
1009: {
1013: MatCreate(comm,A);
1014: MatSetSizes(*A,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
1015: MatSetType(*A,MATBLOCKMAT);
1016: MatBlockMatSetPreallocation(*A,bs,nz,nnz);
1017: return(0);
1018: }